Description
Urban economics is broadly the economic study of urban areas; as such, it involves using the tools of economics to analyze urban issues such as crime, education, public transit, housing, and local government finance.
ABSTRACT
Title of Dissertation: TWO EMPIRICAL ESSAYS IN
ENVIRONMENTAL AND URBAN
ECONOMICS
Yi J iang
Doctor of Philosophy, 2008
Directed By: ProfessorsMaureen Cropper and William Evans
Externalities associated with automobile use have long been an important
topic in environmental and urban economics. Air pollution and traffic congestion
constitute two main external costs of driving (Parry, Walls and Harrington 2007).
Because pricing approaches such as higher fuel taxes and road pricing are unpopular,
varioustravel demand management (TDM) programs aiming to control vehicle travel
demand through non-pricing approaches have been adopted by government agencies
across the country. These programs provide public information, use persuasion,
subsidize transit riding, and promote carpooling and telecommuting. However,
whether these programs generate incentives for people to reduce driving remains an
open question.
I address this question with respect to two types of TDM strategies:
telecommuting and public information provision. The first essay examines whether
telecommutingopportunities lead employees to have longer commute lengths.
Because telecommuting is often jointly chosen with commuting patterns and no
single dataset contains sufficient information to solve the endogeneity problem, I use
a two-sample instrumental variables technique to estimate the causal impact of
telecommuting on commute length. The data for the project are assembled from the
May 2001 Current Population Survey (CPS) and the 2000 Census 5% Public Use
Micro-data Series (PUMS). The results suggest that telecommuting increases married
female workers’ one-way commute time by 9 – 12 minutes, but the effect on male
workers’ commute lengthis not precisely estimated. Although telecommuting may
still cut down total commute miles, it is less effective than expected, in particular for
married women.
The second essay assesses the effectiveness of the Air Quality Action Days
program in the Baltimore metropolitan area in getting cars off the road on high ozone
days. The program asks people to reduce vehicle trips on code red days when the
ozone level is forecast to exceed the EPA’s standard. I look at traffic volumes on
highways in the Baltimore area, and using aregression discontinuity design, measure
the extent that traffic is lower due to the announcement. I find that the program
generally has little effect except that it reduces morning inbound traffic by 4-5
percent. Evening outbound traffic declines correspondingly.
TWO EMPIRICAL ESSAYS IN ENVIRONMENTAL AND URBAN ECONOMICS
By
Yi J iang
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park, in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2008
Advisory Committee:
Dr. Maureen Cropper, Chair
Dr. William Evans, Co-chair
Dr. Mahlon Straszheim
Dr. J udith Hellerstein
Dr. Anna Alberini
©Copyright by
Yi J iang
2008
ii
Dedication
This work is dedicated to my parents, who have always been the source of my
aspiration, energy and strength, and to my wife Ruoyun (Rebecca), for her constant
support, encouragement and love.
iii
Acknowledgements
I am pleased that I chose environmental economics as my major field and am
extremely fortunate to have Maureen Cropper as my advisor. She has spent
uncounted hours instructing and mentoring me. Beyond this dissertation, Maureen’s
help and support has been complete and unconditional: from obtaining external
funding to recommending me for a Resources for the Futureinternship; from eliciting
information from knowledgeable researchers for my projects to supporting my
participation in academic workshops and conferences. I think the best way to return
her great support is to be a sound environmental economist, which I will strive to be.
I am grateful to Bill Evans for his guidance and employment advice. His
ability to creatively combine empirical methods and data to answer a significant
question will always be an inspiration to me. I also benefited tremendously from his
Applied Microeconomics course, which is one of the best prepared and most
enjoyable classes I have ever taken.
I thank Mahlon Straszheim and J udy Hellerstein for their constructive insights
on the papers herein as well as their service on my dissertation committee. I thank
Anna Alberini for serving on the committee and for collaborating with Maureen and
me on a project on episodic ozone control. I am also indebted to Dave Evans, J onah
Gelbach, Christopher McKelvey, Peter Nelson, J eff Smith, Elena Safirova, Margaret
Walls, and to several seminar participants for helpful suggestions and comments.
Finally, special thanks go to J ennifer Desimone at the Metropolitan
Washington Council of Government, Charles Piety at the Department of Atmospheric
and Oceanic Science, University of Maryland, and Michael Pack and the staff of the
iv
Center for Advanced Transportation Technology, University of Maryland for
providing data and background information for the second essay.
v
Table of Contents
Dedication.....................................................................................................................ii
Acknowledgements......................................................................................................iii
List of Tables............................................................................................................... vi
List of Figures.............................................................................................................vii
1 Introduction........................................................................................................... 1
1.1 Objective of the dissertation......................................................................... 1
1.2 Contribution to the Literature....................................................................... 4
1.3 Plan of the Dissertation................................................................................. 6
2 The Impact of Telecommuting on the J ourney to Work: A Two-Sample
Instrumental Variables Approach................................................................................. 7
2.1 Introduction................................................................................................... 7
2.2 Urban Problems and Telecommuting......................................................... 11
2.3 Theory and Empirical Liteature.................................................................. 14
2.4 The NHTS and an Empirical Baseline........................................................ 16
2.5 Empirical Strategy...................................................................................... 19
2.5.1 Instrumental Variable.............................................................................. 19
2.5.2 The Two-Sample Instrumental Variables (TSIV) Method..................... 22
2.5.3 CPS and PUMS Samples........................................................................ 26
2.6 The First-Stage Estimates........................................................................... 29
2.7 Reduced-Form and TSIV Estimates........................................................... 33
2.7.1 Reduced-Form Estimates........................................................................ 33
2.7.2 TSIV Estimates of the Effects on Commute Length.............................. 35
2.7.3 Effects of Telecommuting on Commute Mode...................................... 36
2.7.4 Sensitivity Analysis................................................................................ 37
2.8 Discussion................................................................................................... 39
2.9 Conclusion.................................................................................................. 41
Figures and Tables for Chapter 2............................................................................ 43
3 Do People Drive Less on Code Red Days?........................................................ 59
3.1 Introduction................................................................................................. 59
3.2 AQAD Program in Baltimore Area............................................................ 62
3.3 Theory......................................................................................................... 65
3.4 Empirical Strategy...................................................................................... 67
3.5 Data............................................................................................................. 71
3.6 Results......................................................................................................... 76
3.7 Discussion and Conclusion......................................................................... 80
Figures and Tables for Chapter 3............................................................................ 82
4 Concluding Comments........................................................................................ 92
4.1 Summary of Results.................................................................................... 92
4.2 Directions for Future Research................................................................... 93
Appendices.................................................................................................................. 95
Appendix 1 A Monocentric City Model with Commuters and Telecommuters..... 95
Appendix 2 Imputation of Top-Coded Commuting Time in the PUMS................ 98
Bibliography............................................................................................................. 105
vi
List of Tables
Table 2-1. Descriptive Statistics of NHTS Sample.................................................... 45
Table 2-2. OLS Estimates of the "Effect" of Telecommuting on Commute Lengths
and Travel Mode, 2001 NHTS.................................................................................... 46
Table 2-3. Summary Statistics of CPS Sample by Gender and Telecommuting Status
..................................................................................................................................... 47
Table 2-4. First-Stage Estimates of Telecommuting Models, May 2001 CPS........... 49
Table 2-5. Reduced-Form Estimates of Commute Time Model, 2000 PUMS........... 52
Table 2-6. TSIV Estimates of the Effects of Telecommuting on Commute Time and
Mode........................................................................................................................... 55
Table 2-7. Robustness Check of the TSIV Estimates................................................. 56
Table 2-8. Projection of Commute Time (Minute) onto Commute Distance (Mile).. 58
Table 3-1. One-hour Ozone Level and Code Colors in Baltimore Area.................... 85
Table 3-2. Description of Detectors and Traffic in Baltimore Area........................... 86
Table 3-3. Correlation between Missing Time Block and Code Red Day................. 87
Table 3-4. Summary Statistics and Difference in Selected Covariates Between Code
Red Days and Other Days........................................................................................... 88
Table 3-5. Impact of Code Red Day Announcement on Traffic Volumes by Time of
Day.............................................................................................................................. 89
Table 3-6. Robustness Check...................................................................................... 90
Table 3-7. Impact of CRDs in Sequence on Traffic Volumes by Time of Day.......... 91
Table A1. CPS and NHTS Sample Construction...................................................... 100
Table A2. Distributions by Commute Mode across Cities....................................... 101
Table A3. Full Results of Regression Discontinuity Models (Code red day
coefficients correspond to Columns (4)-(6) in Table 3-5)........................................ 102
vii
List of Figures
Figure 2-1. Distributions of Telecommuters and Commuters by Commute Time and
Distance....................................................................................................................... 43
Figure 2-2. Internet Penetration by 2-Digit Occupation and MSA size (2001 CPS).. 44
Figure 3-1. Illustration of Biased Estimates with Linear Forecast Ozone Level........ 82
Figure 3-2. Map of Baltimore Region Major Freeways and Maryland SHA’s Traffic
Detectors..................................................................................................................... 83
Figure 3-3. Similarity of Covariates around Code Red Day Cutoff Point.................. 84
Figure A1. Bid Rent Curves in a Monocentric City with Telecommuters and
Commuters.................................................................................................................. 99
1
1 Introduction
Automobile use generates significant negative externalities that can be quite
costly to the society. Some economists estimate that external costs amount to 5 cents per
mile in the form of congestion, 2 cents per mile in air pollution and 3 cents per mile in
accidents (Parry, Walls and Harrington 2007). The externalities associated with driving
are not an easy problem to solve. For instance, road capacity expansion has not proved
effective in mitigating congestion due to latent demand (e.g. Small 1992a), and it
exacerbates the air pollution problem by encouraging more driving. Economic theory
suggests that excess driving can be reduced by adding the extra social cost caused by
vehicle driving to a driver's personal cost calculation. This can be accomplished with
instruments like higher fuel taxes and/or a congestion toll. However, in the US where car
ownership is prevalent and demand for vehicle travel is price inelastic, it is not surprising
that these pricing approaches havelittle political support. Meanwhile, government
agencies such as urban planning boards have been interested in controlling the demand
for vehicle travel directly. They have developed various travel demand management
(TDM) programs that provide public information, use persuasion, subsidize transit riding,
and promote carpooling and telecommuting. However, whether these programs generate
incentives for people to reduce driving remains an open question.
1.1 Objective of the dissertation
This dissertation evaluates two distinct TDM strategies: telecommuting and a
public information program. The focus of Chapter 2 is the impact of telecommuting on
2
total commute miles. Telecommuting refers to working from home instead of traveling to
work at least two days every month. In addition to being a favorable arrangement
between employers and employees, telecommuting is increasingly used as a TDM
strategy in order to mitigate urban congestion and air pollution attributable to commute
trips. However, to what extent telecommuting reduces vehicle miles traveled (VMTs)
depends on how it affects commute VMTs and non-commute VMTs. The impact of
telecommuting on commute VMTs depends on the effect of telecommuting on one-way
commuting length as well as telecommuting frequency. The naïve conclusion that
telecommuting reduces total commute miles in the same proportion as it reduces
commuting frequency could be wrong because telecommuting may actually increase one-
way commuting length.
The goal of Chapter 2 is to empirically measure the impacts of telecommuting on
one-way commute length and the probability of driving to work. One of the difficulties is
that telecommuting is not a random choice. Individuals who telecommute could be
systematically different from non-telecommuters in unobserved ways that are also
correlated with commuting behavior. I employ instrumental variable methodsto obtain
consistent estimates, and develop an instrumental variable that captures variation in
telecommuting opportunity across occupation and city size. Another difficulty is that
there exists no single data set that contains telecommuting, commuting and the
instrumental variable. The problem is tackled by using the two-sample instrumental
variable technique, initially developed in Angrist and Krueger (1992). The two samples
3
come from the May 2001 CPS and 2000 PUMS, both of which are nationally
representative.
Chapter 3 looks at a specific public information program---the Air Quality Action
Days program in the Baltimore metropolitan area---which features a code red day alert
when ground-level ozone is forecast to exceed the EPA’s standard. The program not only
warns the public of high ozone levels but also tries to lower ozone concentrations on
those days bypersuading people not to drive. The program can be viewed as a TDM
initiative that aims to influence vehicle travel episodically for environmental purposes. It
is unclear, however, whether an individual would forego driving on code red dayseven if
he/she internalizes the environmental cost resulting from his/her driving. Driving might
still be the optimal mode choice because a person has a smaller risk of being exposed to
bad air when driving than when walking to transit. This chapter conducts an evaluation of
effectiveness of the program in reducing on-road vehicles. A regression discontinuity
design is employed to overcome potential omitted variable bias, since the code red day is
a discontinuous function of an observed continuous variable, forecast ozone level.
In sum, the dissertation studies two popular TDM strategies from an economic
perspective. Potential behavioral responses are taken into account and state-of-the-art
econometric techniques are used in the analysis to carefully examine the effectiveness of
the strategies in lowering vehicle travel. Both positive and negative findings should be
useful to economists. The former may lead us to re-think TDM programs and consider
4
combining pricing approaches with TDM. A negative finding could strengthen the
argument for a complete pricing strategy.
1.2 Contribution to the Literature
Most empirical studies about telecommuting are designed to identifywhat factors
explain the choice of telecommuting. Relatively few studies examine the impacts of
telecommuting on commute length or total VMTs. The results from previous studies are
mixed, with earlier ones suggesting a reduction in total VMTs and later ones showing
positive effects on one-way commute distance and total commute miles. But most studies
are subject to two critical shortcomings. First, telecommuting is assumed to be exogenous
in explaining commute distance or miles. Second, the datasets used in the analyses are
usually small and not representative.
The main contribution of Chapter 2 is that I tackle the endogeneity issue with an
instrumental variable procedure and use large, nationally representative samples
assembled from the 2001 CPS and 2000 PUMS. The instrumental variable measures the
internet penetration for working from home across different occupation by city size cells,
which should capture inexogenous telecommuting opportunity for individuals. Because
of the data constraints mentioned in the previous section, I apply the two-sample
instrumental variable technique of Angrist and Krueger (1992) to information on
telecommuting from the May 2001 CPS and information on commuting from the 2000
PUMS. As both datasets are national samples, the results have implications for different
regions and working groups.
5
The main findings are that telecommuting leadsmarried female workers’ one-way
commute time to increase by 9 – 12 minutes and has a smaller, positive, but statistically
insignificant effect on men’s commute time. Exploring heterogeneous impacts between
married women, single women and men is an innovation to the literature. It is plausible to
expect that married women are more responsive to lower commuting costs as they are
more often the secondary earner in a two-earner household and are likely to be more
constrained in workplace locations. The results confirm these expectations.
Another contribution is to apply the same method to the probability of driving to
work. The OLS estimates using the 2001 National Household Travel Survey (NHTS)
data show that the propensity of driving to work declines among telecommuters. This
result implies that telecommuting provides an extra bonus by changing commute modes
in favor of transit. However, the two-sample instrumental variables estimates indicate
that telecommuting has a positive but statistically insignificant effect on commute mode
choice. They suggest that the OLS estimates could be misleading without correcting the
endogeneity problem. Overall, the chapter suggests that telecommuting is less effective in
lowering total commute miles than people think it is, in particular for married women.
Chapter 3 contributes to the literature by examining a public information program
in Baltimore, andexploiting the program’s institutional features to identify the traffic
reduction attributable to the program. Similar advisory programs have been implemented
in other urban areas that have ozone problem. Earlier studies either rely on survey
respondents’ stated information or are unable to detect a change in traffic caused by the
6
program. Only the recent study by Cutter and Neidell (2007) uses the same econometric
technique and reaches similar conclusions as my study does, but instead looks at the San
Francisco Bay area. I find that the code red day announcements result in morning traffic
reductions by 3 – 5%. This reduction occurs only for inbound traffic in the morning and
outbound traffic in the evening.
1.3 Plan of the Dissertation
Chapters 2 and 3 are both self-contained essays. Chapter 2 examines the impact of
telecommuting on one-way commuting length and on the probability of driving to work.
Chapter 3 estimates the effect of code red day announcements on traffic volumes in the
Baltimore Metropolitan area. Each chapter starts with an introduction of the research
question, methodology and main findings. More detailed research or institutional
background is provided next, and followed by a simple theoretical model to convey the
key intuitions. Empirical methods as well as the data used are described. The main results
and sensitivity checks are presented in subsequent sections. Each essay ends with a
further discussion of the findings and conclusions. Chapter 4 summarizes the main
findings in both essays and draws some common conclusions. It also discusses the
questions stemming from the study that deserve future research.
7
2 The Impact of Telecommuting on the J ourney to Work: A Two-
Sample Instrumental Variables Approach
2.1 Introduction
Telecommuting reduces both the monetary and psychological costs of commuting.
Employers, by allowing workers to telecommute, can recruit and retain valued employees
and possibly reduce the costs of office space and administrative support. More
importantly, telecommuting is increasingly suggested as a solution to traffic congestion
and air pollution in urban areas. For instance, the Connecticut Department of
Transportation established a statewide initiative "Telecommute Connecticut!" to help
employers within the state set up and run telecommuting programs.
1
In May 2006, the
U.S. Department of Transportation announced its new National Strategy to Reduce
Congestion on America's Transportation Network, which highlights "Four Ts" – tolling,
transit, telecommuting and technology – as an approach to reducing traffic congestion.
From the perspective of reducing congestion and pollution caused by vehicle miles
traveled, a key policy question is what impact telecommuting has on total commute miles
traveled.
2
At first blush it would appear that greater telecommuting should decrease
1
See their website http://www.telecommutect.com for more information about "Telecommute
Connecticut!".
2
The impact of telecommuting on non-commute VMTs is another important topic but is beyond the scope
of the dissertation. In theory, telecommuting could affect non-commute VMTs in multiple ways. For
example, flexible working schedules allow telecommuters to go shopping or run errands more often. If the
individual changes home location in response to telecommuting, her/his demand for non-commute travel is
likely to change as well. Walls and Safirova (2004) review a series of telecommuting papers and find no
8
commute miles. However, since telecommuting decreases the cost of commuting, it is
plausible that telecommuting actually induces workers to work farther from home. For
example, a woman who works at home one day a week reduces her commuting costs by
20% compared to a non-telecommuter. The decline in commuting costs provides an
incentive for the woman to live farther away from her workplace or work farther from
home.
Telecommuting may not achieve its policy objectives if it leads to a longer journey
to work. However, it is not easy to obtain a consistent estimate of the causal impact of
telecommuting on commute length. For research purposes, the ideal situation would be to
randomly assign the opportunity to telecommute to a panel of workers and then examine
how often they telecommute and the length of their commutes before and after the
intervention. However, this type of experiments have never been performed. Because
commute length and the decision to telecommute are jointly determined, estimates of the
impact of telecommuting on travel time may be biased. Yet, the direction of the bias is
unclear. On one hand, workers who have longer commute distances may be more likely
to telecommute. At the same time, people who have a distaste for commuting would, all
else equal, live closer to work as well as welcome a telecommuting opportunity.
In this paper, I examine the impact of telecommuting on total commute miles
traveled while controlling for the endogeneity of the telecommuting decision. Because
study show evidence of significant increase in non-commute travel for telecommuters. However, a common
shortcoming of those studies is that they are based on small samples of workers.
9
information on whether an individual telecommutes and the length of his commute are
not contained in one data set, I utilize the two-sample instrumental variables (TSIV)
technique developed by Angrist and Krueger (1992). The key data sets include the work
schedule supplement to the May 2001 Current Population Survey (CPS) that contains
telecommuting data, and the 5-percent Public Use Micro-data Series (PUMS) of the 2000
Census that contains information about one-way commute time and mode. An instrument
is developed from the CPS sample that measures internet utilization for working at home
for each 2-digit occupation and MSA-size combination. The instrument exploits the fact
that certain occupations and MSA combinations are more open to telecommunication
technology than others. These differences are by and large determined by job
characteristics and internet infrastructure distribution, which, once I control for MSA and
occupation fixed-effects, should be orthogonal to individuals' commutes. I also examine
the effect of telecommuting on travel mode choice using the same method.
It is well documented in the literature that men and women exhibit distinct
commuting patterns (White 1977, 1986), especially with respect to marital status and
family composition. I conjecture that telecommuting might have differential effects on
married women and single women for the following reasons. First, in a dual-earner
household, the woman is more often the secondary earner rather than the primary. She is
more likely than her husband to have a part-time or lower-paying job. Therefore,
commuting costs, which will be reduced by telecommuting, may be more important to
her workplace location than to her husband’s. Second, the husband’s job situation is
10
likely to dominate the residential location of the household and affect the workplace
location of the wife. Married women restrict the geographic ranges of their job search and
often work closer to home than their husbands. People who are in occupations where
telecommuting is an option will consider a larger range of workplace location than those
who are not. Since married women may be more constrained in their job search than
single women, telecommuting may have a larger impact on married women choosing
workplace locations than on single women. Therefore, I estimate each model for men,
married women and single women separately to explore the heterogeneity in the response
across these demographic groups.
3
TSIV estimates demonstrate that telecommuting has a large positive effect on
commute length for married female workers: Married women tend to work farther from
home when they can substitute working at home for commuting. Being able to
telecommute causes married women to increase their one-way commute an additional 9-
12 minutes. This finding is consistent with the fact that married female workers have
short commuteswhen telecommuting is not an option. The effect for male workers is
smaller and statistically insignificant. For an average married women who works from
home two out of five days a week, telecommuting reducestotal commute miles, but not
by 40 percent. My analysis also suggests that telecommuting is unlikely to affect the
probability of a worker driving to work.
3
It could be argued that women with children are more constrained in their choice of workplace location
than women without children, regardless of marital status. When, however, the sample is split between
women with and without children, the instrumental variable does not have enough explanatory power.
11
The rest of the paper is organized as follows: Section 2.2 defines telecommuting and
provides background information about telecommuting and relevant studies. Section 2.3
presents baseline estimates of the "effect" of telecommuting on journey-to-work from
OLS analysis of the 2001 Nationwide Household Transportation Survey (NHTS). Section
2.4 describes the identification strategy and the data. TSIV estimation results appear in
Sections 2.5 and 2.6, with discussion and conclusions following in Sections 2.7 and 2.8.
2.2 Urban Problems and Telecommuting
Traffic congestion is a problem for many urban areas in the US and around the
world. The social costs of having millions of cars stuck in traffic are high. The Texas
Transportation Institute estimates that, in 2003, congestion in the 85 largest urban areas
in the US caused 3.7 billion vehicle-hours of delay, resulting in a cost of $63 billion.
According to Lomax and Schrank (2005), each rush hour traveler pays an annual
congestion tax of $800 to $1,600 in lost time and fuel in the 10 most congested areas of
the US. The costs of congestion extend to the environment as well. Automobile emissions
are an important source of ozone precursors—nitrogen oxides (NOx) and volatile organic
compounds (VOCs). In 2003 more than 100 million people lived in counties that violated
the federal ozone standard (EPA, 2004). This is a serious public health problem since it is
well established that ozone can induce respiratory symptoms, and cause decrements in
lung function and inflammation of the airways (EPA, 2003).
While pricing instruments such as congestion tolls and gasoline taxes are a way to
internalize the external costs of driving, they are unpopular in the US. More attention has
12
therefore been devoted to non-pricing strategies that control the demand for automobile
travel directly. A subset of these strategies, Commute Trip Reduction (CTR) programs,
focuses on commute trips, the largest contributor to rush hour traffic and one of the main
contributors to the total vehicle miles traveled (VMT). These programs, often
implemented through cooperation agreements between government authorities,
employers and individuals, provide persuasion (e.g., Earth Day fairs), incentives (e.g.,
transit subsidies) and/or facilitate carpooling. Telecommuting is one of the most popular
components of these programs (Pollution Probe 2001).
The literature has not settled on a consistent definition of telecommuting. Some
studies include as telecommuters people who take work home and never substitute
working from home for commuting on a work day. I refer to these people as teleworkers.
Some research includes the self-employed who work at home sometimes as
telecommuters. As a result, counts of telecommuters vary dramatically across studies.
Mokhtarian et al. (2005) reviewed a number of papers using various data sets and
concludes that the percentage of telecommuters in the late 1990s ranged from 3% to 20%.
The latter figure includes the home-based self-employed and all teleworkers.
In this paper, I define a telecommuter as an employee who works at home instead of
traveling to a workplace at least one day every two weeks. People who commute every
day even though they sometimes work from home, as well as those who telecommute
infrequently are not counted as telecommuters in my definition. My definition also
excludes the self-employed since they are not the target population of TDM policy.
13
Finally, telecommuting does not require that the individual use information and
communication technology (ICT) when working at home, although technology (ICT)
plays a significant role in enhancing telecommuting opportunities.
The May 2001 CPS supplemental survey collected information about work schedules
and working at home from 51,000 working adults from approximately 47,000
households. The final CPS sample in this analysis consists of 29,147workers who lived
in an MSA and were not self-employed in their main jobs.
4
Among them 1,138 were
telecommuters, accounting for 4 percent of the sample. This figure falls at the low end of
the range identified in Mokhtarian et al. (2005).
Many studies of telecommuting have examined who telecommutes or why people
telecommute.
5
For instance, Drucker and Khattak (2000) found in the 1995 Nationwide
Personal Transportation Survey sample that ceteris paribus, males, older people, those
with more education, those with higher incomes, parents of young children, those in rural
areas and those with inferior access to transit are more likely to telecommute. They also
found that one-way commute distance negatively impacts the propensity to telecommute.
Popuri and Bhat (2003) and Walls et al. (2007) analyzed large data sets from New York
and Southern California, respectively. They confirmed the role of the aforementioned
demographic characteristics in determining telecommuting status. In addition, they found
4
Table A1 provides information on sample construction for both the May 2001 CPS sample and the
2001 NHTS sample.
5
The literature contains various definitions of telecommuting. For a comprehensive review, see Walls and
Safirova (2004).
14
that job types and employer characteristics such as employer size and industry have
significant power in explaining telecommuting adoption. However, some variables such
as home location and job tenure may be affected by telecommuting status as well. Using
them directly as explanatory variables yields biased model estimates in these studies.
2.3 Theory and Empirical Liteature
The question of interest here is what effects telecommuting has on workers' journey-
to-work, and, in particular, on commute length. A monocentric-city framework as
described in Brueckner (2001) can be utilized to convey some simple intuition about the
likely impact of telecommuting on commute length. Suppose two types of workers,
commuters and telecommuters, live in a city where all employment is concentrated in the
central business district (CBD). Telecommuters travel to the CBD for work only part of
the week while commuters go five days a week. Because telecommuters have lower
commuting costs than commuters, all else equal, they bid less for homes close to the
CBD and more for homes in suburban areas than commuters. In equilibrium, commuters
live close to the CBD and telecommuters sort into the surrounding region with longer
commutes (see Appendix A for a formal exposition).
The monocentric model, though simple and stylized, predicts that telecommuting
results in a longer commute distance due to a reduction in the marginal cost of
commuting. In a more realistic model that features cities with multiple employment
centers (Glaeser and Kahn 2001), the result may not be so straightforward. In a
polycentric city, employers who are located farther from regions where potential
15
qualified employees live may use telecommuting as a tool in the recruitment (e.g.,
Prystash 1995, Guimaraes and Dallow 1999). This would attract individuals who would
choose to work near their homes if they had to commute everyday. This seems
particularly likely for married women who are more often the secondary earner of the
family and, on average, have shorter commutes than their husbands. Thus, telecommuters
could have longer commutes than non-telecommuters in a polycentric city if they choose
an employer located farther from their home who offerstelecommuting.
The preceding discussion suggests that the impact of telecommuting on commute
length is an empirical question. The difficulty of testing the hypothesis that
telecommuting increases one-way commute distance lies in that telecommuting choice is
unlikely to be exogenous to commuting preference and/or behavior. If original longer
commute encourages an individual to work from home when allowed, a regression of
commute length on telecommuting status will overestimate the effect of telecommuting.
On the contrary, telecommuters could be those who feel more pressures from traffic.
They would have shorter commutes in the absence of telecommuting opportunities. This
unobserved selection will lead to a downward base in the regression estimates. The
existing literature has started to notice the policy significance of the question, but has not
addressed it satisfactorily.
Earlier studies (e.g., Kitamura et al. 1991; Koening et al. 1996; Henderson and
Mokhtarian 1996) found that telecommuting led to a large reduction in total VMTs.
These studies all treat the decision to telecommute as exogenous. Among recent studies,
16
Mokhtarian et al. (2004) analyzedretrospective data from a survey of 218 California state
government employees regarding their telecommuting and commuting behavior over a
ten-year period, from 1988 to 1998. The authors found that telecommuters had higher
one-way commuting lengths than non-telecommuters. Again, assuming telecommuting is
an exogenous choice, the study was unableto tell whether longer commuting distances
encouraged telecommuting or telecommuting facilitated residential relocation farther
from work. Ellen and Hempstead (2002) examinedthe correlation between
telecommuting and city size using the work schedule supplement to the May 1997 CPS.
Their results showed that telecommuters were more likely to live in large, high-density
metropolitan areas. As the authors acknowledge, these results fail to shed light on a
causal relationship: telecommuting opportunities were more likely to appear in
information-intensive service businesses, which tend to concentrate in large, dense
metropolitan areas.
2.4 The NHTS and an Empirical Baseline
The NHTS is a survey of the daily and long-distance travel behavior of the American
public conducted periodically by the Federal Highway Administration (FHWA) since
1969. In the 2001 NHTS, 69,817 households wereinterviewed. The survey collected
detailed information about travel of all sorts including the journey to work. A
shortcoming of the NHTS data is that it does not have much information about a
respondent's job, so that it is difficult to instrument for telecommuting as I do below. I
instead use the NHTS to generate a conditional correlation between telecommuting and
17
commute length, which sets a baseline for comparison with the two-sample instrumental
variables estimates I obtain from the combined CPS and PUMS samples.
The sample constructed from the 2001 NHTS includes individuals who lived in a
Metropolitan Statistical Areas (MSA) and had a job at the time of the survey.
Unfortunately, the NHTS did not ask whether the individual was self-employed. The
problem is mitigated by excluding those who always work at home or have no fixed
workplace. A small portion of respondents with outlier values for commute length or
speed are also removed from the sample.
6
The final sample contains 47,730 individuals
from 33,326 households. I treat as telecommuters those who substitute working from
home for traveling to their usual workplace once every month or more. In this case,
telecommuters constitute of 7.1 percent of the sample. This figure is higher than in the
2001 CPS because the self-employed who work in a fixed place outside the home some
days and at home other days are counted as telecommuters.
7
Table 2-1 reports means and standard deviations of key variables for telecommuters
and non-telecommuters in the NHTS sample. It is clear that the two groups of workers
differ considerably in demographic and socioeconomic characteristics. Telecommuters
6
As there is no way to identify whether outliers are due to misreporting, I employ conservative thresholds
on commute length and speed in sample selection. Individuals reporting one-way commute time greater
than 180 minutes, commute distance longer than 180 miles, or speed lower than 0.01 or greater than 1.5
miles per minute are removed from the sample, which results in 189 exclusions.
7
Due to data constraint, the minimum frequency requirement (one day every month) in the NHTS
definition is lower than that (one day every two weeks) of the CPS. Counting the self-employed, the
percentage of telecommuters in the CPS goes up to 6.2. The difference in the definitions may explain part
of the remaining gap.
18
are, on average, more likely to be male, white, older, better educated, more likely to be
married, have young children and have higher household incomes compared to non-
telecommuters. Telecommuters are more concentrated in professional, managerial, or
technical occupations than non-telecommuters. In terms of commuting patterns, an
average telecommuter spends about 3.5 more minutes and travels an additional 2.6 miles
for a one-way trip to his workplace than an average non-telecommuter. Figures 2-1graph
the distributions of commuters and telecommuters across groups defined by commuting
distance or commuting time. The proportions of telecommuters that fall in groups with
longer commutes are higher than the proportions of commuters in those groups. The last
row of Table 2-1 shows that driving is the main travel mode for 92 percent of commuters.
The proportion of workers commuting by car is 3 percentage points lower among
telecommuters.
A naïve approach to examining how telecommuting impacts the journey to work is to
estimate a single-equation regression model with a commuting variable (i.e. length or
travel mode) as the dependent variable and telecommuting status, together with other
relevant variables, as the explanatory variables. Table 2-2 reports the OLS coefficient
estimates of the telecommuting dummy. Telecommuting has a large positive effect on
both commute time and commute distance for married women. A married female
telecommuter is estimated to travel 3 minutes or 3 miles longer to work than a married
female commuter, ceteris paribus. The estimates for single women and men are smaller
and statistically indistinguishable from zero. In terms of travel mode, telecommuters,
19
except for single women, are less likely by 4 – 5 percentage points—to drive to work
than the average commuter. However, none of these results should be interpreted as the
causal effects of telecommuting as it is likely that people choose telecommuting based on
how far and by which means they commute. The confounding factors would cause OLS
estimates to be biased and the direction of the bias is unclear. To obtain consistent
estimates of the impacts of telecommuting on the journey-to-work, we need to instrument
for telecommuting choice.
The coefficient estimates for other variables indicate that commute lengths as well as
probability of driving to work increase with age (at a decreasing rate), education, and
household income across different population groups. Black workers commute longer
than white, Hispanic and Asian workers, as is documented in the spatial mismatch
literature (e.g., Kain 1968). Married men commute longer than single men. The variables
have qualitatively the same effects on the probability of driving to work.
2.5 Empirical Strategy
2.5.1 Instrumental Variable
The opportunities for teleworking and telecommuting vary substantially from job to
job because of the variation in the relativeproductivity of working from home to working
on-site, which is generally determined by the need for face-to-face communication with
colleagues and customers, as well as the need for team-work. Theapplication of
telecommunication technology during teleworking could alter the substitutability of
teleworking for face-to-face contact. For some jobs, internet technology maintains or
20
even increases the productivity of employees working from home, while for others, it
appears less helpful. The employees in the former case are likely to have more options for
teleworking and telecommuting. While a variable measuring the occupational technology
penetration for teleworking may explain individual’s telecommuting choice, some
unobserved occupational characteristics that affect commute length might be correlated
with that variable. For instance, a high school teacher uses the internet less often when
she works at home than a college professor does. Furthermore, there are more high
schools geographically scattered in a city than colleges. An instrumental variable that
shows that a high school teacher has fewer telecommuting opportunities may also capture
the difference in the geographical distributions of the two jobs if the latter is not well
measured or controlled for in the model.
In the early 2000s internet services, and in particular the broadband capacity, were
not evenly distributed across the country. Some studies show that internet infrastructure
investment or city accessibility to the internet was biased toward larger metropolitan
areas and a group of midsized urban areas (e.g. Malecki, 2002; Grubesic and O'Kelly,
2002). Consequently, the competitiveness of the broadband market varied considerably
across regions. The Federal Communications Commissions (2002) shows that 40.5% of
zip codes had none or one broadband line, in contrast to 27.6% of zip codes with four or
more high-speed lines by J une 2001. The number of broadband providers increased with
population density (Grubesic and Murray, 2004), and rural and smaller metropolitan
areas failed to attract significant levels of competition. The spatial variation in the
21
internet and broadband markets could have led to spatial differences in technology
options for teleworking for different occupations.
Thus, I develop an instrumental variable to measure the penetration of internet for
teleworking across occupation and city size using the work schedule supplement to the
May 2001 CPS, from which we know whether a respondent ever worked at home and
what equipment they used when they were working at home. I calculate the percentage of
employees for each of 270 (45 x 6) occupation-by-MSA-size combinations who ever
worked at home and used the internet (hereafter referred to as internet penetration). The
higher the value, the more likely a person in the occupation-by-city-size cell is to work
from home and possibly telecommute. The advantage of exploiting the variation in the
interaction of occupation and city size is that the effects of unmeasured occupation and
urban structure attributes on commuting behavior can be purged by the introduction of
occupation and city fixed effects in the model. To ensure measurement accuracy, the
occupation-by-city-size cells with fewer than 50 observations are not used in the baseline
analysis. This results in 179 cells covering 37 occupations and 6 MSA sizes. The cell-size
weighted mean (standard deviation) of internet penetration is 0.088 (0.113). In the
sensitivity analysis, I lower the cell selection criterion to 30 observations.
Figure 2-2 shows that internet penetration varies substantially across both
occupations and city sizes. In general, white-collar workers such as professionals,
teachers, and sales representatives have higher average internet penetration as well as
larger variation across city sizes than blue-collar workers such as mechanics and
22
repairmen, or transportation and production workers. College teachers and lawyers and
judges have the highest teleworking internet penetration (0.5 or above), which seems
reasonable since these two occupations are information intensive as well as flexible in
where work is performed. Sales in finance, business and non-retail commodities have
much higher percentages of internet-using teleworkers than retail sales probably because
the latter require personal presence and more face-to-face interaction with customers.
It is plausible to assume that the instrumental variable is not systematically
correlated with other unobservables that affect commuting behavior conditional on the
occupation and city fixed effects. However, to address the concern about this assumption,
I also construct a set of variables at the occupation-by-city level with the PUMS and test
how robust the instrumental variable estimates are to including these variables. More
detail about the occupation-by-city variables and the test is presented in the next section.
2.5.2 The Two-Sample Instrumental Variables (TSIV) Method
Traditionally, instrumental variables estimation is performed when the outcome
variable, the potentially endogenous variable of interest and the instrumental variable
exist in one data set. In addition, a large sample is generally needed for IV estimation to
produce sufficient statistical power. In our case, the instrumental variable discussed
above is measured for two-digit occupation by city size. It cannot be assigned to the
NHTS samplebecause the NHTS contains little information about respondents’ jobs. (A
five-category variable is used to describe occupation as opposed to 45 two-digit
occupations in the CPS.) To the best of my knowledge, there is no other (large) data set
23
that contains information oncommuting, telecommuting, occupation and MSA of
residence.
8
Therefore, a traditional instrumental variable method is infeasible here.
Angrist and Krueger (1992) developed a two-sample instrumental variables (TSIV)
technique that allows one to apply IV estimation to a joint sample with two data sets, one
of which has the outcome and the instrumental variable and the other the endogenous
explanatory variable and the instrument. The work schedule supplement to the May 2001
CPS collected information about respondents’ working at home, occupation and MSA.
The 2000 5% PUMS collected information about journey-to-work as well as occupation
and MSA. Moreover, they were both intended to represent the US population within the
same period and contain many of the same questions. Thus, they constitute a suitable case
for the TSIV method to work.
Formally, suppose the model of interest is
, y X| c = +
where y and c are 1 n× vectors and X is an n k × matrix of regressors, some of
which are correlated with c . An n l × (l k > ) matrix Z is needed to consistently estimate
| , where Z is not correlated with c and ( ) lim / 0
n
p Z X n
÷·
' = . Angrist and Krueger point
out that in the case when only X and Z (but not y ) are observed in one data set and
only y and Z (but not X ) are observed in the other, | can still be consistently
estimated when certain assumptions, which will be discussed in detail in the next
8
The NLSY79, an alternative data set, provides only commuting time in the early 1990s for fewer than
10,000 workers. The information about telecommuting in the NLSY79 is limited to hours worked at home.
24
subsection, hold for the two samples. Many researchers have since used the two-sample
approach (e.g., Currie and Yelowitz 2000, Dee and Evans 2003) to circumvent the data
constraint. In practice, a two-stage least squares procedure is usually adopted to produce
the following estimator
( )
1
2 2 2 2
TSIV
X X X y |
÷
' ' =
? ? ? ?
,
9
where ( )
1
2 2 1 1 1 1
X Z Z Z Z X
÷
' =
?
,
1
X and
1
Z are from the first sample, and
2
y and
2
Z are
from the second.
Now suppose equation (2-1) describes the structural model of commute length (or
mode):
ikc ikc ikc k c ikc
y a W B T u ì µ v = + + + + + (2-1)
where
ikc
y is the commute time or travel mode of individual i living in MSA c with
occupation k ,
ikc
W is a vector of individual specific exogenous variables,
k
µ and
c
v are
occupation and MSA fixed effects, and
ikc
u is idiosyncratic disturbance. The potentially
endogenous variable,
ikc
T , is an indicator for telecommuting. The parameter of interest,
ì , measures the causal impact of telecommuting on commute length or travel mode.
The first stage in calculating the TSIV estimate of ì is to estimate a model of
telecommuting adoption as described by equation (2-2),
1 1 1 1 1 1 1 1 1
,
ikc ikc ks k c ikc
T a W B Z u ì µ v = + + + + +
1
1, , i n = ? (2-2)
9
Inoue and Solon (2005) called this estimator the two-sample two-stage least squares (TS2SLS) estimator
and showed that it is different from the TSIV estimator originally proposed by Angrist and Krueger. They
proved that the TS2SLS estimator is asymptotically more efficient than the TSIV estimator. That being said,
I continue to label the estimator TSIV to distinguish it from the one-sample IV approach.
25
where the subscript 1 denotes the CPS sample,
1
n is the sample size in the CPS, and
ks
Z
is the instrumental variable measured at occupation by MSA size level (s ). The
parameters estimates are applied to the second sample, i.e. the PUMS sample, to predict
telecommuting status,
2ikc
T
?
. In the second stage, the TSIV estimate of ì is generated by
regressing the outcome variables in the PUMS,
2ikc
y , on the predicted telecommuting
status,
2ikc
T
?
and other covariates. In an exactly identified case such as ours, we can
alternatively fit a reduced-form equation, i.e. equation (2-3), using the PUMS sample,
2 2 2 2 2 2 2 2 2
,
ikc ikc ks k c ikc
y a W B Z u ì µ v = + + + + +
2
1, , i n = ? (2-3)
where subscript 2 denotes the PUMS sample and
2
n is the sample size of the PUMS. The
TSIV estimate is just the ratio between the reduced-form and first-stage coefficients
before
ks
Z , i.e.
1
2
ˆ
ˆ
ˆ
ì
ì
ì =
TSIV
.
Standard errors of the TSIV estimator can be computed using a linear Taylor series
approximation assuming zero covariance between the first-stage and reduced-form
estimators. That is
|
|
.
|
\
|
+ =
2
2
2
2
2
1
2
1
2
1
2
2 2
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ì
o
ì
o
ì
ì
o
TSIV
(2-4)
where
1
ˆ o and
2
ˆ o are estimated standard errors of
1
ˆ
ì and
2
ˆ
ì , respectively.
10
10
Note that with this approximation formula the t-statistic of the TSIV estimates is the following function
of the t-statistics of the first-stage and reduced-form estimates:
2 2
2 1 2
2 2
1 2
TSIV
t t
t
t t
=
+
. When
1
t , the first-stage t-
statistic outweighs
2
t ,
TSIV
t approaches
2
t .
26
2.5.3 CPS and PUMS Samples
In addition to assumptions underlying the traditional IV model, the TSIV approach
imposes some conditions on the joint sample. The key one is that the two data sets must
represent the same population. It is plausible to argue that these conditions hold for the
samples constructed from the CPS and the PUMS. The CPS is administered by the
Bureau of the Census for the Bureau of Labor Statistics. The former is also in charge of
implementing the decennial census of the US, from which the PUMS was created. Both
the CPS and the PUMScollected a rich set of information from US households on
individuals’ demographic characteristics, labor force experience, household attributes and
economic status. In addition to the similarity in content, the phrasing of questions and
coding of potential responses are similar across the CPS and the PUMS. While the CPS
and PUMS are both intended to be representative of the US population, the PUMS
includes institutionalized individuals, who are excluded from the CPS. I remove these
observations from the PUMS in constructing the joint sample. Moreover, every variable
in the sample is ensured to have the same support across the two sources. For instance,
only workers who are 16 years old or above, live in an MSA and are not self-employed
on the main job are retained in my data. MSAs that appear in just one data set are
removed. The final data includes 234 common MSAs.
Nevertheless, potential mismatches between the CPS and the PUMS might exist due
to the differences in sampling design, response rates and survey times. The CPS selects
households by primary sampling units (PSUs) based on the 1990 Census while the PUMS
27
draws households with sampling rates varying with the housing density of census blocks
or tracts.
11
Second, the Census spent tremendous effort to induce people to fill out the
survey forms, which led to higher response rates in the Census than the CPS. Finally, the
CPS data were collected in May 2001 roughly one year after the 2000 Census was
conducted. A visual comparison of the weighted means
12
of the CPS and the PUMS
samples does not suggest significant differences for most variables between the samples.
However, t-tests reject the mean equality for several variables across the two samples.
13
Table 2-3 presents descriptive statistics for telecommuters and non-telecommuters in
the CPS sample. Telecommuters are 3 – 4 years older than commuters on average, and
disproportionately white and better educated. They are more likely to be married and live
in smaller households, with higher annual incomes. In terms of job types, telecommuters
are concentrated in occupations such as executives, administrators, managers, math and
computer scientists, teachers of all levels, lawyers and judges, and sales representatives in
finance and business services—workers who are generally in the upper levels of the job
hierarchy. White-collar workers in the service sector and blue-collar workers have fewer
11
See http://usa.ipums.org/usa/chapter2/chapter2.shtml for a detailed explanation.
12
Sample weights contained in the CPS and PUMS are applied in calculating summary statistics and
estimation to adjust the over-sampling in each survey.
13
In the notation of section 2.5.2, the condition on the joint sample can be formally written as
( ) ( )
1 2
1 1 1 2 2 2
lim / lim /
ZX
n n
p Z X n p Z X n
÷· ÷·
' ' = = E
.
It requires that the first and second moments of explanatory variables including the instrumental variable of
the two samples converge to the same matrix. It can be tested for the variables that are observed in both
samples. A t-test on the means just examines the first moment and is likely to reject the null given large
sample size. With increasing applications of the TSIV method, formal ways to test the assumption and to
evaluate the potential bias resulting from mismatch of the two samples are probably desirable.
28
opportunities to telecommute. Finally, a higher proportion of telecommuters than non-
telecommuters live in large MSAs with populations over one million. As in Table 2-1,
Table 2-3 showsthat telecommuters and non-telecommuters differ in many observed
ways. It is therefore likely that they also differ in unobserved variables that are correlated
with commuting behavior.
In the PUMS, commute length is recordedin minutes and measures how long it
usually took the respondent to get from home to work during the past week. White
(1988a) argues that time is a better measure of commuting costs than distance because
time is the scarce resource that people economize. Moreover, Table2-2 shows that the
same set of variables explains more variation in commute time than commute distance,
This suggests that the noise associated with distance is larger than with time: A commuter
can estimate commuting time more accurately than distance. The translation of the
impact of telecommuting on commute time into an impact on commute distance is
considered below.
In the PUMS, 1.7 percent of the sample or over 60,000 respondents have travel times
that are top-coded at 99 minutes. Exploiting the properties of the Pareto distribution, I
replace the top-coded values with an estimate of the conditional expectation for top-
coded values. The procedure is described in Appendix B, which suggests a range for the
imputed values of 120 to 165 minutes. I use the lower bound, 120 minutes, in the
benchmark analysis. Since the likelihood of being top-coded is positively correlated with
telecommuting adoption, using the lower bound value works against finding a positive
29
effect of telecommuting. I check whether different imputed values affect the results in the
sensitivity analysis. The weighted average commute times in the PUMS are 24.2, 24.7
and 27.5 minutes for married women, single women and men, respectively. These figures
are slightly higher than in the NHTS sample, which are 22.1, 22.9 and 25.2, respectively.
In the PUMS, a higher share (93.3%) of married women drives to work than single
women (85.1%) and men (89.8%). Similar patterns are observed in the NHTS sample, for
which the shares are 94.1%, 87.7%, and 91.7%, respectively.
2.6 The First-Stage Estimates
In the first stage, I estimate a linear probability model of telecommuting adoption
(equation (2-2)). The dependent variable is a binary indicator equal to one if the worker is
telecommuting. In the baseline model, the explanatory variables include age, age squared,
gender, race, educational achievement, number of household members, presence of
children 5 years of age or younger, children between 6 and 15 years of age, spouse (for
the male sample only), annual household income, and the occupation-by-city-size internet
penetration measure. Industry and job class variables are not included in the model
because they are individual choices that are likely to be correlated with home location,
work location or commute length. Neither the wage, housing price, or travel mode and
time is used as an explanatory variable. All of these are chosen simultaneously with
commute length and, therefore, are endogenous. Fixed effects for MSA-of-residence and
2-digit occupation category are controlled for, assuming people do not sort into a city and
2-digit occupation based on their preferences for commute length or telecommuting.
30
The model is estimated for married women, single women and men separately using
individual weights provided by the CPS. Results are reported in Table2-4. In general, the
estimates reflect the differences between telecommuters and commuters in Table2-3.
People who are older, white, possess a college or advanced degree, have children and
come from affluent households are more likely to telecommute. Less obvious from the
descriptive statistics is that black employees have a higher probability of working at
home than other groups, although a lower probability than whites. Being married does
not seem to play a role in the telecommuting decisions for male employees. All else
equal, telecommuting is significantly more popular among professionals and sales
representatives in finance and business services, but less popular among engineers and
supervisors. Surprisingly, blue-collar workers are not less likely to telecommute than
white collar workers, conditional on demographic and economic covariates. This may be
because people with less education are offered more telecommuting opportunities when
working in blue-collar jobs than in white-collar jobs.
Several variables have differing influences on telecommuting adoption across the
samples. Race plays an important role in telecommuting for married women but not for
single women. In contrast, household size and income are more important for the latter
than for the former. The likelihood of telecommuting increases with age at a decreasing
rate for women workers. This pattern is much weaker and statistically insignificant for
men. A male employee with a graduate degree has a substantially larger propensity to
telecommute than one with a college degree, but this is not the case for a female
31
employee. Men tend to work at home if there are older children but not younger children
in the household. The reverse is true for married women – suggesting that married
women may use telecommuting as a way to combine work and childcare.
The coefficients on the instrumental variables are of paramount importance and vary
substantially across samples. In the case of married women, a 10 percentage point
increase in occupation/MSA internet penetration causes the probability of telecommuting
to rise by 5.4 percentage points once 2-digit occupation and MSA fixed effects are
controlled for. This effect is statistically significant at the1% level. On the contrary, the
estimate for single women is smaller (0.143) and statistically insignificant, suggesting the
instrumental variable has little explanatory power for single women employees. For male
employees, a 10 percentage point increase in internet penetration increases the probability
of telecommuting by 2.9 percentage points, an effect that is significant at the 5% level.
One critical assumption underlying the IV approach is that teleworking technology
penetration is not correlated with any unobservable that influences commute length or
mode. There might be concerns that the instrumental variable is correlated with
occupation-specific local labor market conditions. For instance, the urban economics
literature hypothesizes that individuals are forward looking when they choose home
location and commute length. They take into account labor market dynamics and
potential moving costs. Specifically, Crane (1996) predicts a shorter commute for persons
with lower probability of changing jobs within the local labor market. Likewise, van
32
Ommeren et al. (1997) argue that commuting distance is decreasing in the arrival rates of
job offers and increasing in moving costs.
One way to deal with this concern is to control in the model for occupation-by-city
attributes. Lacking clear theory informing what those attributes should be, I construct a
rich set of covariates using the PUMS data. I calculate the fraction of employees within
each 2-digit occupation and MSA combination who are: male, white, black, have a high
school degree, some college experience, a college degree, an advanced degree (omitting
high school dropouts), in the transportation and communication industries, in trade, in
finance, in services, in public administration(omitting the manufacturing and
construction industries), working for private for profit employers, and working for private
non-profit employers (omitting government). I also compute the labor market share,
median hourly wage, and difference between the 75
th
percentile wage and the 25
th
percentile wage of each occupation by MSA. Finally, using the CPS sample, I calculate
the fraction of employees for each 2-digit occupation and MSA size combination who
have flexible work hours.
Even columns in Table2-4report estimation results for the model with inclusion of
these occupation-city specific covariates. The coefficients of demographic and household
variables do not change much, although some occupation fixed effects vary. This
suggests that the constructed covariates pick up part of the variation in telecommuting
explained by occupation. The coefficient on internet penetration declines slightly to 0.48
for married women while statistical significance is maintained at the 1% level. The
33
coefficient for men is unchanged up to two decimal places. These results indicate that the
instrumental variable is likely to be orthogonal to the local labor market conditions
described by those covariates.
2.7 Reduced-Form and TSIV Estimates
2.7.1 Reduced-Form Estimates
Equation (2-3) is estimated only for married women and men since the instrumental
variable is not statistically significant for single women. The exogenous explanatory
variables are the same as in the first-stage except that they are from the PUMS sample.
Results are reported in Table 2-5. In the baseline model, commute length increases with
age at a decreasing rate for both women and men. Race makes a substantial difference in
commute length, which may reflect residential segregation and employment separation.
Black male workers on average spend 2 more minutes on the road than white and other
workers and black females travel 4 minutes longer than white females. Regardless of
gender, college graduates and those from high-income households live farther from their
workplace than employees without a college degree and workers from low income
households. Married men travel 1 minute longer to work than single men. When there are
younger children in the household, both married women and men travel longer to work,
while the presence of older children has the opposite, but smaller, effect for women.
Commute time increases with the number of household members for men and decreases
for married women. Overall, the results are consistent with those from the NHTS and
largely agree with those in White (1986). Commute length varies significantly across jobs
34
even conditioning on factors like age, race, and education. One possible reason is the
variation in geographic concentrations of different occupations. For example, school
teachers have short commutes because schools are scattered throughout a city.
The instrumental variable shows large positive impacts on married women's
commute lengths but not on men's commute lengths. In the baseline model without
controlling for occupation-MSA covariates, i.e., Columns 1 and 3, a 10 percentage point
increase in the proportion of employees of each 2-digit occupation and MSA size
combination who ever use internet when working at home leads to 0.60 minute longer
commuting trip for married women. The estimate is statistically significant. In contrast,
the coefficient estimate of the internet penetration for male workers is 0.13 minutes and
statistical insignificant.
When the occupation-by-city covariates are controlled for in the model, few changes
occur in the coefficients of the demographic and household variables. However, a number
of occupation fixed effects vary dramatically. This suggests the importance of
heterogeneity in local markets for different occupations in determining commute length.
The coefficients of the occupation-MSA covariates imply that conditional on individual
characteristics, commute length increases if the person works in an occupation that has
more human capital, is concentrated in finance and services industries, is more
represented in the private for profit sector and has a larger labor market share. The last
result seems to be consistent with Crane’s theory that a person values commuting
distance less if more potential employers are available.
35
The effect of internet penetration on commuting length declines slightly and retains
statistical significance for married female workers. Now, a 10 percentage point increases
in internet penetration lead to an additional 0.46 minutes in commute time for married
women. The estimate for male workers is less than 0.2 minutes and statistically
insignificant. The results, consistent with those without occupation-by-city covariates,
suggest that the instrumental variable is unlikely to pick up the occupation-city specific
attributes as confounding factors.
2.7.2 TSIV Estimates of the Effects on Commute Length
First-stage estimates indicate that the internet penetration instrumental variable has
statistically significant and positive impacts on the telecommuting status of married
women and men in the 2001 May CPS. The reduced-form estimates indicate that the
instrumental variable has a substantial positive effect on one-way commute time of
married women but little effect for male employees in the 2000 PUMS. The TSIV
procedure ties these two sets of results together to generate consistent estimates of the
causal effects of telecommuting on commute length.
Table 2-6a presents the TSIV estimates calculated as the ratios of reduced-form
estimates to the first-stage estimates of the instrumental variable. In the exactly identified
case, it yields the same estimates as the two-stage least square estimation in the two
sample case (TS2SLS). The standard errors of the TSIV estimates are computed using
equation (2-4). The TSIV estimates suggest that telecommuting has a substantial positive
impact on married women’s commute lengths. All else equal, working at home at least
36
one day every two weeks, on average, causes a married women employee to commute 9 –
11 minutes longer than if she commutes every day. The impact for male employees is
smaller in magnitude (around 5 minutes) and statistically indistinguishable from zero. In
comparison with OLS estimates, the TSIV estimates yield qualitatively similar results.
However, OLS results underestimate the effect of telecommuting for married women,
which is consistent with the fact that married women usually have short commutes if they
do not telecommute.
2.7.3 Effects of Telecommuting on Commute Mode
OLS analysis of the NHTS data shows that male and married female telecommuters
are less likely to drive to work than non-telecommuters. It is difficult to find a compelling
reason why telecommuting leads people to forego driving to work. The OLS estimates
are susceptible to an omitted variable bias that fails to account for sorting of women who
take public transit to work into telecommuting. Moreover, driving usually is faster than
taking public transit or any other travel mode.
14
If telecommuting does cause a worker to
commute by a mode other than driving, the lengthened commute time might be a result of
choosing a slower travel mode rather than an increase in commute distance. Therefore, it
is important to identify the true effect of telecommuting on commute mode.
I apply the same TSIV procedure to the travel mode variable available in the PUMS
sample. Using the same argument that internet penetration is unlikely to affect travel
14
The average speeds for commuting by driving, by rail, by bus, and by bicycle in the NHTS are 0.53, 0.36,
0.28, and 0.23 miles per minute, respectively.
37
mode choice directly, TSIV produces consistent estimates of the effects of telecommuting
on travel mode choice. Table 2-6b reports both reduced-form and TSIV estimates for
travel mode. In the baseline model, the TSIV estimates are small, positive and without
statistical significance for both married women and men. When the occupation-by-city
covariates are added, the estimate for married women is almost zero while the estimate
for men becomes negative with a large standard error. Overall, the TSIV point estimates
do not support the OLS results that telecommuting reduces a married woman’s
probability of driving to work. The negative OLS estimates could result from the fact that
employees who commute by public transit also prefer to telecommute. However, the
TSIV estimates are not sufficiently precise to let us draw definite conclusions about the
effect.
2.7.4 Sensitivity Analysis
I examine the sensitivity of the above results to different sample restrictions and
alternative imputed values for the top-coded commute times. Tables 2-7a and 2-7b report
the estimates for the commute time and travel mode models, respectively. In Panel A of
each table, the samples are extended to include the occupation-MSA size cells that
contain 30 or more CPS observations, which results in 216 cells covering 38 2-digit
occupations and 6 MSA sizes. In the first stage, the instrumental variable has a smaller
effect for married women while the coefficient for men does not change much as
compared to the case with cells containing over 50 observations. It continues to have a
large, statistically significant reduced-form effect on married women's commute time and
38
little effect on men's commute time. The TSIV estimates show that telecommuting
increases married women's commute time by 13 minutes though they lack enough
statistical power in the case with job-by-city covariates included. The effect of
telecommuting among male employees falls to 3 and 4 minutes, andthe t-statistics are
less than 1. As far as travel mode is concerned, telecommuting shows some positive
effects for both married women and men, but again the estimates are not distinguishable
from zero. These results are highly consistent with the baseline case with cells larger than
50 observations.
Telecommuting is often thought of as a choice for office workers only. Programs and
policies that aim at promoting telecommuting usually target these occupations rather than
the entire working population. Therefore, it may be of interest to examine the effects of
telecommuting on commuting behavior for office workers. One way to define office
workers is to narrow the sample down to the 2-digit occupations coded 1 through 26.
Included in this group are managerial, professional specialty, technical, sales, and
administrative support occupations. 2-digit occupation codes greater than 26, including
service, precision production, craft, repair, farming, forestry and fishing occupations and
operators, fabricators and laborers, are excluded. Panels B of Tables 2-7a and 2-7b
present the estimates for the sample of office workers. The instrumental variable affects
only the telecommuting propensity of married women. Telecommuting is estimated to
lengthen the one-way commute time of married women by 8 - 9 minutes, which is
statistically significant at the 10% level. Again, telecommuting has a positive but
39
statistically insignificant effect on married women’s commute mode, contrary to the OLS
estimates. In sum, estimates with different sample restrictions demonstrate that the effects
of telecommuting on commuting show stability and a certain degree of homogeneity
across occupations. Panel C of Table 2-7a shows that replacing the top-coded commute
time by 165 minutes instead of 120 minutes has no impact on the effects of
telecommuting on commute time.
2.8 Discussion
The TSIV estimates of the effects of telecommuting on commute time for married
women equal 9 to 12 minutes, which are 3 to 4 times the OLS estimates from the NHTS.
The results are plausible in that married women have shorter commutes on average. The
OLS analysis tends to underestimate the effects of telecommuting in this case. The
magnitude of the adjustment in the commute made by married women appear reasonable
given that the average commute time for married women in the PUMS is 24.2 minutes
with a standard deviation of 19. TSIV estimates suggest that telecommuting increases
commute time by about half of a standard deviation.
TSIV estimation could be biased if the internet penetration measured by occupation
crossed with MSA size is correlated with some unobservables that impact individual
commute lengths. The concern may be less serious as the models control for a rich set of
occupation-by-city specific covariates as well as occupation and city fixed effects.
Another potential source of bias is that the teleworking technology penetration is
measured with 2001 CPS data. When internet access expanded rapidly to a wider
40
population and more regions in the early 2000s, the variation across occupation and cities
declined quickly with time. Therefore, the impact of internet penetration on
telecommuting adoption estimated using 2001 data may underestimate the impact in year
2000 when the PUMS were collected, which would result in an overestimation of the
TSIV coefficients.
I am interested in translating the effects of telecommuting on commute time into the
effects on commute distance. I use the NHTS data to estimate a relationship between
commute time and distance for people driving to work. Table 2-8 shows the coefficients
of models that project commute time onto commute distance and distance squared.
15
Commute time is a concave function of distance with an intercept greater than zero,
which suggests a positive fixed cost and an increasing marginal speed. The relationship
between commute time and distance varies by sex, with women having greater concavity.
Using these estimates, we can recover the approximate distance from travel time. For
instance, suppose a woman drove 24 minutes to work before choosing to telecommute.
Applying the projection estimates implies that on average her commute distance was 13
miles. If her one-way commute timeincreases to 33 minutes after telecommuting, the
one-way commute time increases to 20.5 miles, a7.5 miles increase. If she works from
home 2 days a week (the national average for telecommuting women is 2.2 days per
week), the total weekly commutes are 198 minutes or 123 miles, representing 17 percent
15
Higher order polynomials were tried. They produce very bad predictions for distances on the high end.
Moreover, the predictions for the mid-range values do not differ with and without the higher order terms.
41
and 5.5 percent declines relative to the before-telecommuting commute times and
commute miles, respectively.
2.9 Conclusion
Telecommuting has been promoted as a means to deal with congestion and
automobile emissions by researchers and public policy makers. However, there are
concerns that telecommuting workers will make a longer commute in response to the
lower commute frequency. Naïve (OLS) estimates based on the NHTS show that a
married woman commutes 3 minutes or 3 miles longer if she telecommutes. The NHTS
estimates also show that telecommuters except single women are less likely to drive to
work than non-telecommuters. However, these estimates could be biased because
telecommuting is not randomly assigned among workers. Furthermore, theory cannot
predict the direction of the bias.
By applying two-sample instrumental variables technique to the CPS and PUMS
samples, I find that telecommuting causes married women employees’ commuting trips
to increase by 9 to 12 minutes. The effect for male workers is also positive, but smaller
and not precisely estimated. For single women, the instrumental variable does not have
enough power to explain telecommuting choice. In addition, TSIV estimates show a
small, positiveeffect of telecommuting on the probability of commuting by car for
married women. Although lacking statistical power, this does not agree with the negative
relationship between telecommuting and driving to work found in the OLS analysis.
Given the sizable “rebound” effect on one-way commute time found among married
42
women, thetotal commute miles traveled by an average married women worker are
unlikely to decline in proportion to telecommuting frequency.
Unfortunately, the instrumental variable developed in this paper does not have
enough information to let us estimate the effects of telecommuting for men and single
women. This needs to be explored in future research. Moreover, to understand whether
telecommuters adjust their commute distance by changing residential location or
employment location is important for both research and policy purposes and should also
be examined.
43
Figures and Tables for Chapter 2
Figure 2-1. Distributions of Telecommuters and Commuters by Commute Time and
Distance
0
.
1
.
2
.
3
.
4
<=10 11-20 21-30 31-40 41-50 51-60 >60 min
Source: Author's calculation using NHTS 2001.
Distributions of Workers by One-way Commute Time
Commuters
Telecommuters
0
.
1
.
2
.
3
.
4
<=5 5-10 10-15 15-30 30-45 45-60 >60 miles
Source: Author's calculation using NHTS 2001.
Distributions of Workers by One-way Commute Distance
Commuters
Telecommuters
44
Figure 2-2. Internet Penetration by 2-Digit Occupation and MSA size (2001 CPS)
0 .2 .4 .6
Teleworking Techonology Penetration
Farm Workers
Other Handlers
Freight & Stock Handlers
Construction Laborers
Other Transportation
Motor Vehicle Operators
Fabr. Assem. Inspe. Samp.
Machine Oper. & Tenders
Other Precis. Prod. & Craft
Construction Trades
Mechanics & Repairers
Personal Service
Clean & Bldg.Service
Health Service
Food Service
Protective Service
Other Admin. Support,
Mail And Message Distr.
Finan. Rcds. Process.
Secre. Steno. & Typists
Super,, Admin. Support
Sales, Retail & Personal
Sales, Non-retail Comm,
Sales, Finance. & Busi.
Super. & Propr., Sales
Other Technicians
Engineering & Sci. Tech.
Health Tech.
Other Professional
Lawyers & J udges
School Teachers
College/Univ. Teachers
Health Assess. & Treat.
Math & Computer Sci.
Engineers
Management
Other Exe, Admin.
Sources: Work Schedule Supplement to May 2001 CPS
Teleworking Technology Penetration by Occupation and MSA Size
MSA Size:
100-250K
250-500K
500K-1M
1-2.5M
2.5-5M
>5M
Note: Internet penetration is calculated as the weighted percentage of employees who ever work at home
and use the internet within each occupation-by-MSA-size cell using data from the Work Schedule
Supplement to the May 2001 CPS.
45
Table 2-1. Descriptive Statistics of NHTS Sample
Non-telecommuters Telecommuters
Variables Mean Std. Dev. Mean Std. Dev.
Raw N 44556 3174
Age 39.205 12.370 42.250 11.243
Male 0.541 0.498 0.599 0.490
White 0.708 0.455 0.807 0.395
Black 0.123 0.329 0.060 0.238
Asian 0.029 0.168 0.044 0.206
Hispanic 0.110 0.313 0.056 0.231
High School Degree 0.290 0.454 0.111 0.314
Some College 0.303 0.460 0.252 0.434
College Degree 0.216 0.411 0.369 0.483
Graduate Degree 0.115 0.319 0.248 0.432
Spouse 0.608 0.488 0.664 0.472
Child Age 0 – 5 in HH 0.211 0.408 0.225 0.418
Child Age 6 – 15 in HH 0.310 0.463 0.312 0.464
Household Size 3.152 1.441 2.990 1.364
HH Income $40 – 70K 0.322 0.467 0.233 0.423
HH Income $70 – 100K 0.191 0.393 0.256 0.437
HH Income >$100K 0.152 0.359 0.343 0.475
Sales or Services 0.266 0.442 0.236 0.425
Clerical or Administrative
Support 0.136 0.342 0.059 0.236
Manufacturing, Construction,
Maintenance, or Framing 0.180 0.384 0.061 0.239
Professional, Managerial, or
Technical 0.417 0.493 0.644 0.479
Time to Work 23.688 17.889 27.167 22.362
Distance to Work 12.628 12.800 15.238 16.194
Drive to Work 0.917 0.276 0.884 0.321
Note: Sample is constructed from the 2001 NHTS including workers who live in an MSA, have an
outside-home fixed workplace, and have one-way commute distance less than 180 miles, commute time
less than 180 minutes and commute speed less than 1.5 miles per minute and greater than 0.01 miles per
minute. Observations with missing values for any of the listed variables are also dropped. Means and
standard deviations are calculated using the weights from the NHTS.
46
Table 2-2. OLS Estimates of the "Effect" of Telecommuting on Commute Lengths and
Travel Mode, 2001 NHTS
Married
Women
Single
Women
Men
Telecommuting (1) (2) (3)
A. COMMUTE TIME (MINUTES)
Coefficient 2.904** 0.452 1.652
Standard Error (1.225) (1.410) (1.040)
R-sq 0.11 0.12 0.08
B. COMMUTE DISTANCE (MILES)
Coefficient 3.124*** 1.063 1.144
Standard Error (0.968) (1.010) (0.699)
R-sq 0.09 0.09 0.05
C. DRIVE TO WORK
Coefficient -0.043** 0.013 -0.051***
Standard Error (0.019) (0.025) (0.014)
R-squared 0.11 0.17 0.13
#Observations 14176 8939 24615
Note: The sample is the same as in Table 2-1. All models include
age, age squared, race, education, household composition, annual
household income, and job category and MSA fixed effects.
Heteroscedastic-robust standard errors without clustering are in
parentheses. * indicates significant at 10%, ** significant at 5%,
and *** significant at 1%.
47
Table 2-3. Summary Statistics of CPS Sample by Gender and Telecommuting Status
Women Men
Non-
telecommuters Telecommuters
Non-
telecommuters Telecommuters
Variables
Mean Std. Mean Std. Mean Std. Mean Std.
Raw N 13528 594 14481 544
Age 38.678 12.806 41.663 10.829 38.437 12.614 43.020 11.367
White 0.795 0.404 0.886 0.318 0.833 0.373 0.921 0.271
Black 0.148 0.355 0.081 0.273 0.111 0.314 0.045 0.208
High School Degree 0.285 0.452 0.139 0.346 0.283 0.450 0.096 0.295
Some College 0.314 0.464 0.227 0.419 0.272 0.445 0.175 0.381
College Degree 0.204 0.403 0.399 0.490 0.201 0.401 0.416 0.493
Graduate Degree 0.086 0.281 0.202 0.402 0.095 0.293 0.300 0.459
Spouse Present 0.501 0.500 0.642 0.480 0.577 0.494 0.730 0.444
With Child 0 – 5 in HH. 0.194 0.395 0.222 0.416 0.221 0.415 0.196 0.397
With Child 6 – 15 in HH. 0.327 0.469 0.338 0.474 0.314 0.464 0.320 0.467
Household Size 3.076 1.495 3.029 1.444 3.236 1.592 2.924 1.426
Annual Family Income <$40K 0.364 0.481 0.163 0.369 0.320 0.467 0.087 0.282
Annual Family Income $40 – 75K 0.337 0.473 0.310 0.463 0.355 0.479 0.245 0.430
Annual Family Income >$75K 0.299 0.458 0.527 0.500 0.325 0.469 0.669 0.471
2-digit Occupation
01 Public Administrators and
Officials 0.000 0.016 0 0 0.001 0.025 0 0
02 Other Executive,
Administrators, and Managers 0.099 0.299 0.165 0.371 0.120 0.324 0.252 0.435
03 Management Related
Occupations 0.053 0.224 0.077 0.267 0.032 0.175 0.063 0.244
04 Engineers 0.004 0.065 0.003 0.057 0.035 0.184 0.046 0.210
05 Math. and Computer Scientists 0.013 0.111 0.038 0.192 0.025 0.156 0.051 0.221
06 Natural Scientists 0.003 0.058 0.010 0.100 0.005 0.069 0.006 0.077
07 Health Diagnosing Occupations 0.005 0.067 0.004 0.059 0.007 0.082 0.010 0.102
08 Health Assessment and Treating 0.049 0.216 0.026 0.160 0.007 0.083 0 0
09 College and University
Teachers 0.007 0.082 0.045 0.208 0.007 0.085 0.057 0.232
10 Other Teachers 0.065 0.246 0.152 0.359 0.021 0.143 0.039 0.194
11 Lawyers and J udges 0.004 0.066 0.010 0.097 0.007 0.082 0.032 0.177
12 Other Professional Specialty 0.042 0.200 0.097 0.296 0.032 0.177 0.110 0.313
13 Health Technologists and
Technicians 0.025 0.155 0.007 0.084 0.004 0.060 0 0
14 Engineering and Science
Technicians 0.007 0.084 0.003 0.052 0.015 0.123 0.009 0.093
15 Other Technicians 0.010 0.100 0.014 0.116 0.015 0.121 0.027 0.162
16 Sales Supervisors and
Proprietors 0.027 0.162 0.024 0.154 0.033 0.180 0.032 0.176
17 Sales Representatives, Finance
and Business Service 0.018 0.132 0.055 0.229 0.018 0.131 0.089 0.285
18 Sales Representatives,
Commodities except Retail 0.006 0.075 0.017 0.129 0.018 0.131 0.073 0.261
19 Sales Workers, Retail and
Personal Services 0.067 0.249 0.026 0.159 0.036 0.186 0.020 0.139
20 Sales Related Occupations 0.001 0.030 0 0 0.000 0.021 0 0
21Supervisors, Administrative
Support 0.010 0.099 0.001 0.038 0.004 0.064 0.003 0.055
22 Computer Equipment Operators 0.004 0.060 0.001 0.036 0.003 0.054 0 0
23 Secretaries, Stenographers, and 0.042 0.201 0.040 0.197 0.001 0.031 0.003 0.055
48
Typists
24 Financial Records Processing 0.028 0.164 0.030 0.170 0.003 0.056 0.003 0.057
25 Mail and Message Distributing 0.005 0.073 0 0 0.010 0.099 0.002 0.044
26 Other Administrative Support
Occupations, including Clerical 0.153 0.360 0.078 0.269 0.046 0.209 0.009 0.094
27 Private Household Service 0.001 0.025 0.004 0.061 0 0 0 0
28 Protective Service Occupations 0.009 0.093 0.002 0.044 0.031 0.172 0.004 0.066
29 Food Service Occupations 0.058 0.234 0 0 0.044 0.206 0 0
30 Health Service Occupations 0.036 0.186 0.007 0.086 0.004 0.064 0.003 0.056
31 Cleaning and Building Service 0.022 0.145 0 0 0.025 0.156 0 0
32 Personal Service 0.041 0.199 0.046 0.210 0.009 0.093 0.002 0.041
33 Mechanics and Repairs 0.004 0.061 0.003 0.052 0.061 0.240 0.024 0.153
34 Construction Trades 0.002 0.048 0 0 0.071 0.257 0.007 0.082
35Other Precision Production 0.012 0.111 0.004 0.066 0.040 0.195 0.003 0.050
36 Machine Operators and Tenders 0.024 0.153 0.001 0.023 0.039 0.194 0.004 0.065
37 Fabricators, Assemblers,
Inspectors, and Samplers 0.016 0.124 0.007 0.083 0.025 0.155 0.001 0.036
38 Motor Vehicle Operators 0.008 0.089 0 0 0.052 0.222 0 0
39 Other Transportation and
Material Moving 0.001 0.027 0 0 0.018 0.131 0 0
40 Construction Laborer 0.000 0.015 0 0 0.013 0.114 0.005 0.069
41 Freight, Stock and Material
Handlers 0.012 0.107 0.003 0.057 0.034 0.182 0 0
42 Other Handlers, Equipment
Cleaners, and Laborers 0.005 0.067 0 0 0.011 0.102 0.001 0.036
43 Farm Operators and Managers 0.000 0.018 0 0 0.000 0.021 0.004 0.060
44 Farm Related Workers 0.006 0.079 0 0 0.021 0.142 0.003 0.056
45 Forestry and Fishing
Occupations 0.000 0.008 0 0 0.001 0.022 0.003 0.051
MSA w/ Population 100k – 250k 0.089 0.284 0.062 0.241 0.084 0.278 0.076 0.266
MSA w/ Population 250k – 500k 0.140 0.347 0.115 0.319 0.134 0.341 0.094 0.291
MSA w/ Population 500k – 1m 0.166 0.372 0.139 0.346 0.156 0.363 0.111 0.314
MSA w/ Population 1m – 2.5m 0.306 0.461 0.339 0.474 0.316 0.465 0.390 0.488
MSA w/ Population 2.5m – 5m 0.168 0.374 0.195 0.397 0.176 0.380 0.191 0.393
MSA w/ Population 5m+ 0.131 0.338 0.150 0.357 0.134 0.341 0.138 0.346
Note: Sample is constructed from the May 2001 CPS including workers who live in an MSA and are not self-employed
on the main job. Observations with missing values for any of the listed variables are also dropped. Means and standard
deviations are calculated using the weights from the CPS.
49
Table 2-4. First-Stage Estimates of Telecommuting Models, May 2001 CPS
Married Women Single Women Men
(1) (2) (3) (4) (5) (6)
Age 0.004** 0.004** 0.002** 0.002** 0.001 0.001
(0.002) (0.002) (0.001) (0.001) (0.001) (0.001)
Age Squared -0.000* -0.000* -0.000* -0.000* -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
White 0.044*** 0.042*** 0.014 0.013 0.026*** 0.026***
(0.009) (0.008) (0.010) (0.010) (0.006) (0.006)
Black 0.023* 0.020 0.010 0.010 0.019*** 0.019***
(0.013) (0.013) (0.012) (0.012) (0.007) (0.007)
High Scholl Degree -0.013 -0.014 -0.005 -0.005 0.001 0
(0.011) (0.011) (0.004) (0.003) (0.003) (0.003)
Some College -0.008 -0.010 0.001 0.001 -0 -0.001
(0.012) (0.012) (0.005) (0.005) (0.004) (0.004)
College Degree 0.034** 0.033** 0.026*** 0.027*** 0.025*** 0.024***
(0.014) (0.014) (0.008) (0.008) (0.008) (0.008)
Graduate Degree 0.023 0.021 0.009 0.010 0.035*** 0.035***
(0.019) (0.019) (0.015) (0.015) (0.011) (0.011)
Spouse 0.001 0.001
(0.005) (0.005)
With Child0-5 in HH. 0.022*** 0.021** 0.013* 0.013* 0.007 0.007
(0.008) (0.008) (0.007) (0.007) (0.005) (0.005)
With Child 6-15 in HH. 0.010 0.009 0.014* 0.014* 0.013*** 0.012***
(0.009) (0.009) (0.007) (0.007) (0.005) (0.004)
Household Size -0 0 -0.004* -0.004* -0.005*** -0.004***
(0.003) (0.003) (0.002) (0.002) (0.002) (0.001)
HH Income $40 – 75K -0.001 -0 0.008 0.008 0.004 0.003
(0.007) (0.007) (0.006) (0.006) (0.004) (0.004)
HH Income >75K 0.007 0.007 0.037*** 0.038*** 0.030*** 0.029***
(0.009) (0.009) (0.009) (0.009) (0.005) (0.005)
0.027 0.038 0.009 0.018 0.018 0.027 03 Management Related
Occupations (0.023) (0.037) (0.020) (0.032) (0.018) (0.024)
04 Engineers -0.032 -0.019 -0.049*** 0.009 -0.026** -0.059*
(0.038) (0.061) (0.014) (0.043) (0.013) (0.032)
05 Math. and Computer Scientists -0.025 -0.013 0.073** 0.112*** -0.027 -0.019
(0.035) (0.043) (0.033) (0.040) (0.021) (0.025)
0.025 0.083 0.010 0.045 -0.015 -0.016 08 Health Assessment and
Treating (0.032) (0.069) (0.029) (0.045) (0.025) (0.053)
-0.169** -0.143 0.052 0.079 -0.026 0.043 09 College and University
Teachers (0.076) (0.117) (0.136) (0.122) (0.089) (0.097)
10 Other Teachers -0.026 0.089 0.050*** -0.019 -0.019 0.021
(0.019) (0.073) (0.016) (0.078) (0.018) (0.051)
11 Lawyers and J udges -0.150 -0.200 -0.078** 0.071 0.005 0.045
(0.097) (0.144) (0.033) (0.076) (0.072) (0.090)
12 Other Professional Specialty 0.041* 0.039 0.027 0.040 0.054*** 0.074**
(0.023) (0.038) (0.018) (0.028) (0.017) (0.030)
0.067* 0.142* 0.014 0.019 0.007 0.045 13 Health Technologists and
Technicians (0.039) (0.076) (0.027) (0.045) (0.026) (0.050)
0.047 0.096* 0.001 0.001 0.015 0.043 14 Engineering and Science
Technicians (0.034) (0.050) (0.027) (0.041) (0.022) (0.035)
15 Other Technicians 0.051 0.069* -0.010 0.031 0.025 0.017
(0.037) (0.042) (0.026) (0.036) (0.026) (0.029)
50
0.050* 0.156** 0.028 0.003 0.011 -0.017 16 Sales Supervisors and
Proprietors (0.026) (0.069) (0.021) (0.046) (0.019) (0.034)
0.024 -0.015 0.041 0.017 0.090*** 0.148** 17 Sales Representatives, Finance
and Business Service (0.033) (0.059) (0.028) (0.056) (0.032) (0.065)
-0.050 0.048 0.133* 0.118 0.051* -0.010 18 Sales Representatives,
Commodities except Retail (0.054) (0.069) (0.070) (0.075) (0.028) (0.040)
0.082** 0.205** 0.021 -0.040 0.034 0.003 19 Sales Workers, Retail and
Personal Services (0.038) (0.094) (0.029) (0.056) (0.029) (0.047)
-0.013 0.016 0.022 0.032 -0.037* -0.019 21Supervisors, Administrative
Support (0.025) (0.045) (0.040) (0.049) (0.019) (0.028)
0.086** 0.134** 0.028 0.020 0.120 0.150 23 Secretaries, Stenographers, and
Typists (0.038) (0.066) (0.031) (0.044) (0.082) (0.093)
0.096** 0.140** 0.015 0.016 0.041 0.055 24 Financial Records Processing
(0.038) (0.062) (0.028) (0.042) (0.043) (0.049)
0.042 -0.021 0.004 -0.011 0.015 0.062 25 Mail and Message Distributing
(0.042) (0.130) (0.033) (0.077) (0.030) (0.055)
0.067** 0.072 0.018 -0.018 0.014 0.043 26 Other Administrative Support
Occupations, including Clerical (0.033) (0.059) (0.027) (0.041) (0.027) (0.040)
0.044 0.088 0.016 0.061 0.007 0.091** 28 Protective Service Occupations
(0.035) (0.096) (0.036) (0.060) (0.023) (0.041)
0.063 0.189** 0.020 -0.031 0.029 0.018 29 Food Service Occupations
(0.038) (0.087) (0.032) (0.056) (0.030) (0.048)
0.076* 0.165** 0.018 0.013 0.050 0.091 30 Health Service Occupations
(0.039) (0.081) (0.030) (0.052) (0.042) (0.070)
0.057 0.160** 0.014 -0.002 0.021 0.077 31 Cleaning and Building Service
(0.038) (0.070) (0.032) (0.053) (0.029) (0.051)
32 Personal Service 0.103** 0.151** 0.049 0.030 0.026 0.058
(0.041) (0.068) (0.030) (0.043) (0.029) (0.049)
33 Mechanics and Repairs 0.033 0.067 0.064 0.066 0.027 0.047
(0.035) (0.057) (0.066) (0.077) (0.026) (0.044)
34 Construction Trades 0.064 0.136* 0.002 -0.027 0.021 0.035
(0.039) (0.071) (0.030) (0.058) (0.028) (0.043)
35Other Precision Production 0.073* 0.163*** 0.023 -0.006 0.014 0.018
(0.040) (0.060) (0.028) (0.046) (0.027) (0.037)
0.071* 0.181** 0.011 -0.029 0.027 0.041 36 Machine Operators and
Tenders (0.039) (0.072) (0.032) (0.055) (0.030) (0.044)
0.067* 0.173** 0.042 0.009 0.019 0.031 37 Fabricators, Assemblers,
Inspectors, and Samplers (0.037) (0.068) (0.039) (0.056) (0.028) (0.041)
38 Motor Vehicle Operators 0.057 0.037 0.008 0.009 0.020 0.020
(0.039) (0.083) (0.032) (0.057) (0.029) (0.047)
-0.001 0.098 -0.014 -0.019 0.017 0.021 39 Other Transportation and
Material Moving (0.035) (0.074) (0.035) (0.059) (0.029) (0.045)
40 Construction Laborer 0.051 0.154* -0.035 -0.055 0.022 0.048
(0.040) (0.080) (0.033) (0.064) (0.029) (0.046)
0.087 0.140* 0.019 -0.006 0.026 0.033 41 Freight, Stock and Material
Handlers (0.054) (0.082) (0.032) (0.055) (0.030) (0.046)
0.047 0.168** 0.008 -0.032 0.019 0.034 42 Other Handlers, Equipment
Cleaners, and Laborers (0.048) (0.075) (0.033) (0.057) (0.031) (0.048)
44 Farm Related Workers 0.079** 0.174** 0.010 -0.026 0.029 0.066
(0.038) (0.071) (0.032) (0.057) (0.030) (0.046)
Internet Penetration 0.539*** 0.481*** 0.143 0.183 0.288** 0.285**
(0.160) (0.163) (0.127) (0.134) (0.119) (0.115)
Constant -0.194*** -0.446** -0.081** 0.089 -0.072** -0.247**
51
(0.053) (0.181) (0.039) (0.159) (0.032) (0.102)
J ob-by-city Covariates N Y N Y N Y
Observations 6936 6936 6553 6553 13809 13809
R-squared 0.07 0.08 0.07 0.07 0.09 0.09
Note: All models include MSA fixed effects. Occupation-by-city covariates include fractions of employees within
each 2-digit occupation and MSA who are male, white, black, have high school degree, some college, college degree,
advanced degree, work in industries of transportation and communication, trade, finance, services, or public
administration, and work in private profit or private non-profit sectors. Also included are occupation’s local labor
market share, median log of wage, inter-quartile log of wage, and fraction of employees that have flexible work
hours. Robust standard errors are estimated clustering on MSA. * indicates significant at 10%, ** significant at 5%,
and *** significant at 1%.
52
Table 2-5. Reduced-Form Estimates of Commute Time Model, 2000 PUMS
Married Women Men
(1) (2) (3) (4)
Age 0.174*** 0.174*** 0.575*** 0.574***
(0.019) (0.018) (0.017) (0.017)
Age Squared -0.003*** -0.003*** -0.006*** -0.006***
(0) (0) (0) (0)
White -1.532*** -1.498*** -0.018 -0.002
(0.344) (0.337) (0.222) (0.208)
Black 2.322*** 2.329*** 2.048*** 2.034***
(0.374) (0.377) (0.314) (0.326)
High Scholl Degree -0.731*** -0.640*** 0.194*** 0.229***
(0.151) (0.133) (0.071) (0.069)
Some College 0.126 0.181 0.276*** 0.286***
(0.153) (0.134) (0.098) (0.098)
College Degree 0.935*** 0.943*** 0.762*** 0.711***
(0.181) (0.187) (0.231) (0.226)
Graduate Degree 1.628*** 1.603*** -0.182 -0.324
(0.290) (0.261) (0.267) (0.247)
Spouse 1.193*** 1.191***
(0.107) (0.107)
With Child 0-5 in HH. 1.370*** 1.354*** 0.667*** 0.649***
(0.099) (0.098) (0.088) (0.085)
With Child 6–15 in HH. -0.535*** -0.533*** -0.091 -0.103
(0.106) (0.104) (0.092) (0.089)
Household Size -0.455*** -0.452*** 0.301*** 0.313***
(0.043) (0.041) (0.038) (0.036)
HH Income $40 – 75K 0.320*** 0.337*** 0.330*** 0.341***
(0.097) (0.095) (0.079) (0.075)
HH Income >75K 1.278*** 1.256*** 1.118*** 1.099***
(0.151) (0.148) (0.146) (0.142)
1.908*** 2.335*** 1.741*** 1.546*** 03 Management Related
Occupations (0.247) (0.506) (0.239) (0.525)
04 Engineers 1.074*** 3.083*** 1.155*** 2.090***
(0.411) (1.185) (0.273) (0.795)
05 Math. and Computer Scientists 3.007*** 4.387*** 2.604*** 3.340***
(0.333) (0.569) (0.333) (0.650)
08 Health Assessment and Treating -0.434 -0.379 -1.485** -0.359
(0.662) (1.024) (0.603) (1.228)
-1.924** 3.412* -2.793*** 0.201 09 College and University
Teachers (0.862) (1.956) (0.922) (2.106)
10 Other Teachers -7.506*** -3.951*** -5.218*** -0.868
(0.369) (1.271) (0.362) (1.900)
11 Lawyers and J udges -0.911 0.464 -0.947 -3.326
(0.967) (2.855) (0.660) (2.456)
12 Other Professional Specialty -1.984*** 1.559** -2.100*** 0.711
(0.204) (0.637) (0.175) (1.009)
-0.090 2.122** -0.468 1.705 13 Health Technologists and
Technicians (0.640) (1.068) (0.600) (1.605)
14 Engineering and Science
Technicians
2.347*** 6.782*** 0.791* 3.843***
(0.595) (1.043) (0.438) (1.216)
15 Other Technicians 3.615*** 3.371*** 2.792*** 2.707***
53
(0.373) (0.668) (0.446) (0.741)
-1.611*** 3.980*** -1.555*** 2.511** 16 Sales Supervisors and
Proprietors (0.368) (1.082) (0.364) (1.080)
-1.752*** -3.124** -0.679** -4.056** 17 Sales Representatives, Finance
and Business Service (0.273) (1.558) (0.336) (1.799)
2.259*** 8.260*** 2.597*** 7.295*** 18 Sales Representatives,
Commodities except Retail (0.408) (1.080) (0.287) (1.002)
-3.976*** 2.761* -3.205*** 2.712* 19 Sales Workers, Retail and
Personal Services (0.630) (1.419) (0.615) (1.438)
0.143 3.418*** -0.112 2.892*** 21Supervisors, Administrative
Support (0.411) (0.755) (0.447) (0.864)
-0.096 3.904*** -0.016 3.984*** 23 Secretaries, Stenographers, and
Typists (0.542) (0.996) (0.585) (1.324)
24 Financial Records Processing 0.176 4.643*** 0.730 4.785***
(0.586) (1.047) (0.580) (1.025)
25 Mail and Message Distributing -0.399 3.378* -3.019*** 5.746***
(0.800) (1.911) (0.683) (1.757)
-0.095 1.157 -1.027* 0.525 26 Other Administrative Support
Occupations, including Clerical (0.568) (1.025) (0.533) (0.847)
28 Protective Service Occupations 0.034 4.622** -1.556** 2.642
(0.861) (1.885) (0.668) (1.744)
29 Food Service Occupations -4.354*** 2.308 -3.580*** 1.958
(0.713) (1.562) (0.653) (1.517)
30 Health Service Occupations -0.555 3.314*** -1.573** 1.725
(0.666) (1.216) (0.694) (1.730)
31 Cleaning and Building Service 0.610 5.096*** -2.378*** 0.985
(0.623) (1.424) (0.660) (1.754)
32 Personal Service -1.586*** 3.159** -1.509** 3.387**
(0.582) (1.218) (0.598) (1.598)
33 Mechanics and Repairs 2.594*** 6.459*** 0.015 2.849**
(0.601) (1.132) (0.583) (1.401)
34 Construction Trades 4.153*** 10.447*** 5.044*** 9.022***
(0.830) (1.477) (0.650) (1.545)
35Other Precision Production -0.472 5.480*** -0.701 3.599***
(0.595) (1.187) (0.602) (1.345)
36 Machine Operators and Tenders 0.151 6.061*** -1.077* 3.070**
(0.649) (1.417) (0.637) (1.553)
0.558 6.950*** 0.233 4.696*** 37 Fabricators, Assemblers,
Inspectors, and Samplers (0.678) (1.302) (0.637) (1.495)
38 Motor Vehicle Operators -2.504*** 1.576 -1.549** 3.569**
(0.789) (1.422) (0.714) (1.548)
1.275 7.537*** 1.806*** 6.669*** 39 Other Transportation and
Material Moving (1.075) (1.653) (0.665) (1.596)
40 Construction Laborer 5.827** 13.929*** 5.626*** 10.816***
(2.508) (2.834) (0.695) (1.737)
-0.629 5.467*** -1.286** 4.183*** 41 Freight, Stock and Material
Handlers (0.632) (1.462) (0.650) (1.557)
-0.339 6.301*** -0.761 3.934** 42 Other Handlers, Equipment
Cleaners, and Laborers (0.827) (1.652) (0.643) (1.765)
44 Farm Related Workers 0.732 6.945*** -1.091 3.227*
54
(0.968) (1.714) (0.739) (1.679)
Internet Penetration 5.988** 4.625** 1.364 1.599
(2.704) (2.311) (2.660) (2.188)
Constant 24.001*** 5.690 12.845*** -3.926
(1.012) (3.475) (0.735) (3.755)
J ob-by-city Covariates N Y N Y
Observations 832956 832956 1720931 1720931
R-squared 0.08 0.08 0.07 0.07
Note: All models include MSA fixed effects. Robust standard errors are estimated clustering on
MSA. * indicates significant at 10%, ** significant at 5%, and *** significant at 1%.
55
Table 2-6. TSIV Estimates of the Effects of Telecommuting on Commute Time and
Mode
Table 2-6a. Commute Time
Married Women Male
(1) (2) (3) (4)
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 5.988** 4.625** 1.364 1.599
Standard Error 2.704 2.311 2.66 2.188
#of Observations 832956 832956 1720931 1720931
TSIV
Coefficient 11.109* 9.615* 4.736 5.610
Standard Error 6.004 5.805 9.441 8.004
J ob-by-city Covariates N Y N Y
Table 2-6b. Commute Mode
Married Women Male
(1) (2) (3) (4)
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 0.013 0.003 0.006 -0.043
Standard Error 0.047 0.036 0.05 0.037
#of Observations 849904 849904 174329 1743292
TSIV
Coefficient 0.024 0.006 0.021 -0.151
Standard Error 0.087 0.075 0.174 0.143
J ob-by-city Covariates N Y N Y
Note: All models include age, age squared, race, education, household
composition, annual household income, and occupation and MSA fixed effects.
Robust standard errors are estimated clustering on MSA.* indicates significant at
10%, ** significant at 5%, and *** significant at 1%.
56
Table 2-7. Robustness Check of the TSIV Estimates
Table 2-7a. Commute Time
Married Women Male
(1) (2) (3) (4)
A. OCCUPATION – MSA SIZE CELL >=30
First-Stage
Coefficient 0.419** 0.346** 0.302*** 0.325***
Standard Error 0.169 0.174 0.115 0.115
#of Observations 7157 7157 14724 14724
Reduced-Form
Coefficient 5.427*** 4.519*** 1.285 0.909
Standard Error 2.086 1.713 2.212 1.839
#of Observations 864097 864097 1831544 1831544
TSIV
Coefficient 12.952* 13.061 4.255 2.797
Standard Error 7.216 8.225 7.502 5.744
B. OFFICE WORKERS
First-Stage
Coefficient 0.574*** 0.532*** 0.254 0.244
Standard Error 0.169 0.171 0.175 0.164
#of Observations 5477 5477 6974 6974
Reduced-Form
Coefficient 5.369* 4.583* 1.858 0.966
Standard Error 3.051 2.355 2.054 2.137
#of Observations 654502 654502 845161 845161
TSIV
Coefficient 9.354 8.615* 7.315 3.959
Standard Error 5.986 5.221 9.529 9.154
C. TOPCODED COMMUTE TIME REPLACED BY 165 MIN
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 6.088** 4.517* 1.63 1.806
Standard Error 2.792 2.473 2.856 2.5
#of Observations 832956 832956 1720931 1720931
TSIV
Coefficient 11.295* 9.391 5.660 6.337
Standard Error 6.170 6.047 10.189 9.137
J ob-by-city Covariates N Y N Y
57
Table 2-7b. Commute Mode
Married Women Male
(1) (2) (3) (4)
A. OCCUPATION – MSA SIZE CELL >=30
First-Stage
Coefficient 0.419** 0.346** 0.302*** 0.325***
Standard Error 0.169 0.174 0.115 0.115
#of Observations 7157 7157 14724 14724
Reduced-Form
Coefficient 0.032 0.022 0.035 0.023
Standard Error 0.043 0.033 0.047 0.036
#of Observations 881551 881551 1855151 1855151
TSIV
Coefficient 0.076 0.064 0.116 0.071
Standard Error 0.107 0.101 0.162 0.114
B. OFFICE WORKERS
First-Stage
Coefficient 0.574*** 0.532*** 0.254 0.244
Standard Error 0.169 0.171 0.175 0.164
#of Observations 5477 5477 6974 6974
Reduced-Form
Coefficient 0.045 0.035 -0.059 -0.054
Standard Error 0.059 0.04 0.040 0.049
#of Observations 667751 667751 862045 862045
TSIV
Coefficient 0.078 0.066 -0.232 -0.221
Standard Error 0.105 0.078 0.225 0.250
J ob-by-city Covariates N Y N Y
Note: All models include age, age squared, race, education, household composition,
annual household income, and occupation and MSA fixed effects. Robust standard
errors are estimated clustering on MSA.* indicates significant at 10%, ** significant at
5%, and *** significant at 1%.
58
Table 2-8. Projection of Commute Time (Minute) onto Commute Distance (Mile)
All Women Men
(1) (2) (3)
Commute Distance 1.253 1.363 1.224
(0.018) (0.042) (0.023)
Distance Squared -0.002 -0.004 -0.0019
(0.0003) (0.001) (0.0003)
Constant 7.279 6.742 7.375
(0.143) (0.229) (0.203)
#Observations 44,218 21,404 22,814
Adj. R-squared 0.739 0.684 0.765
Note: The sample includes only those who drive to workin the
sample in Table2-1.
59
3 Do People Drive Less on Code Red Days?
3.1 Introduction
By 2007, 347 counties with 141 million residents were designated by EPA as
ground-level ozone nonattainment areas.
16
This means that nearly half of the US
population breathes air with ozone concentration above a harmful level. Besides the
established fact that ozone has adverse effects on the respiratory system, recent studies
(e.g., Bell et al. 2004) also link ozone levels with increases in mortality. Therefore,
bringing the ozone levels into compliance with the EPA standard is a goal of high priority
for public policy.
Ozone is formed when its precursors, oxides of nitrogen (NOx) and volatile organic
compounds (VOCs), react in the atmosphere. Peak ozone levels typically occur on hot,
dry and sunny summer days. Emissions from motor vehicle exhaust, industrial facilities
and electric utilities are the main sources of NOx and VOCs. Dramatic increases in the
number of cars and miles they are driven contribute significantly to the ozone problem in
urban areas, in spite of the fact that individual vehicles are getting cleaner. According to
the EPA (2003), motor vehicles account for 56% and 45% of emissions of NOx and VOC
nationwide, respectively.
A number of metropolitan areas have implemented public information programs that
aim at mitigating the ozone problem by encouraging voluntary driving reductions on high
ozone days. Examples include the Air Quality Action Days (AQAD) program in the
16
See http://www.epa.gov/air/oaqps/greenbk/o8index.html
60
Washington-Baltimore metropolitan area, the Spare the Air (STA) program in the San
Francisco Bay area, and the Ozone Action Days program in Atlanta, to name a few.
Undoubtedly, these programs have low implementation and enforcement costs, in
contrast to mandatory control programs. They also take advantage of the episodic feature
of the ozone problem, a strategy that theoretically promotes economic efficiency (Teller
1967, Krupnick 1988). However, the first question that needs to be addressed is how
effective these programs are in getting cars off the road. People receiving forecast
information may cancel trips due to the concerns about getting unhealthy exposure and/or
the environmental impacts of driving on those days. Nevertheless, for many people,
including most commuters, it is very costly to change travel schedules, if not impossible.
The question is whether the information provided creates enough incentives for the
recipients to take action.
Identifying the impact of the programs on vehicle driving also serves a practical
purpose in air quality regulation. These public information programs all fall into the
category of Voluntary Mobile Source Emission Reduction Programs (VMEPs). Since
1997, the U.S. EPA allows states with non-attainment areas to claim credits up to 3% of
projected emissions reductions for VMEPs when filing State Implementation Plans
(SIPs). To do so requires that mobile emission reductions through voluntary programs be
quantified.
Several studies have looked at how these voluntary information programs impact
travel behavior. Henry and Gordon (2003), MWCOG (2003), and Fox and Sarkar (2002)
61
use individual survey data to examine to what extent the ozone alerts have altered
behavior. They all find that a significant share of respondents reported taking actions
during ozone episodes to help abate pollution. For example, Fox and Sarkar report that in
the Washington-Baltimore area, 7-9 percent of respondents said they drove less on code
red days, days when the ozone levelsare predicted to exceed the EPA standard. A
common issue, however, with the self-reported information is that it may be biased due to
recollection difficulty or other subjective factors. Instead, Cummings and Walker (2000),
Cutter and Neidell (2007) and Welch et al. (2005) directly examine traffic volumes in
Atlanta and the San Francisco Bay area and train ridership in Chicago, respectively.
Cummings and Walker and Welch et al. found either the traffic reductions were too small
to be surely attributed to the program or the ozone advisories increased transit ridership
only in a small part of the Chicago area. To the contrary, Cutter and Neidell found that
STAs reduce total daily traffic by 2.5 to 3.5 percent, with most effects occurring during
and just after the morning peak hours.
The focus of this study is to examine the effectiveness of the AQAD program in the
Baltimore area, an area relying more on automobile driving than the Chicago, San
Francisco and Washington DC metropolitan areas.
17
The program forecasts daily ozone
level one day ahead and uses color codes to indicate expected ozone severity. When the
17
Besides a bus system, Baltimore’s transit system consists of a single-line metro subway and a three-line
light rail, most parts of which overlap. In terms of commuting, the percentage of drivers is higher and the
percentage of rail riders is lower than the national average, based on data from NHTS 2001. See Table A2
for a comparison of the distributions of commuters by commute mode across several cities.
62
ozone level is predicted to reach or exceed the one-hour ozone standard, 125 ppb, a code
red is announced. I use a regression discontinuity (RD) design to see whether traffic
volumes are lower on code red days due to the announcement. The study is closest to
Cutter and Neidell in methodology, and obtains somewhat similar results. The main
finding is that the coderedday announcement reduces inbound traffic volumes during
morning peak hours by 3-5%. Outbound traffic volumes in the evening peak hours fall
correspondingly. In contrast, on code orange days, when ozone levelsare predicted to
exceed 105 ppb but lie below 125 ppb, I do not observe a reduction in vehicle driving.
The rest of the chapter is organized as follows. Section 3.2 documents the detailsof
the AQAD program. Section 3.3 presents a theoretical account of potential behavioral
changes in response to code red days. Sections 3.4 and 3.5 describe the empirical
methods and data used, respectively. Section 3.6 presents the results, and section 3.7
discusses the policy implications of my findings.
3.2 AQAD Program in Baltimore Area
The Baltimore area, with over 2.5 million residents in 2000, consists of five
counties
18
andBaltimore city. It is designated as a nonattainment area by EPA under both
the 1-hour and 8-hour ozone standards. Since the mid-1990s the area has been
implementing the AQAD program jointly with the Washington metropolitan area. Under
the coordination of the Metropolitan Washington Council of Governments (MWCOG), a
18
They are Anne Arundel, Baltimore, Carroll, Harford and Howard. See
http://www.epa.gov/oar/oaqps/greenbk/baltimo.html for a regional map. The designated area is different
from the census MSA definition.
63
daily forecast of the ozone level
19
is conducted for each area by a panel of meteorologists
every day from May 1stto mid-September.
The ozone level is predicted as a quadratic function of a vector of variables including
maximum and average surface temperatures, wind speed, relative humidity, solar zenith
angel, and lagged ozone observations. Note that the predicting variables measure only the
most relevant air and climatological conditions. Because they do not include variables
that forecast vehicle travel demand and electric utility production, the model does not
account for human behavior. The parameters of the function are estimated using
historical observations and remain unchanged for the current year. The model produces
forecasts for each of seven locations in Baltimore area where ozone monitoring stations
are located. The highest one is chosen as the initial forecast for the area.
The expert panel meets on a conference call at 3:00 pm every afternoon to discuss
and make adjustments to the initial forecast. This stage is subjective to the extent that it
relies on the experience of the participants. A personal communication with one of the
panel members indicates that the rational for this subjective procedure is multi-fold. First,
some factors are hard to quantify or are insignificant in model estimation, e.g., the
direction of wind from outside the area, but need to be taken into account. The day of the
week is sometimes taken into account toaddressconcerns about traffic. The panel also
needs to consider different versions of weather forecasts as well as to adjust ozone
forecastsat the lower and upper ends because the model seems to perform better in the
19
The one-hour ozone level was forecast until 2003. Since 2004, eight-hour levels have been forecast..
64
middle range of ozone levels than at the extremes. Although the changes often involve
only a couple of units, they may result in a shift of the air quality category in which the
day falls.
A color code is assigned to the day based on the consensus forecast value. Table3-1
shows the ranges of forecast one-hour ozone concentrations and the corresponding code
colors. When the ozone level is predicted to exceed the EPA standard, i.e. 125 ppb, a red
code is designated and the day is called a code red day. The last column in Table3-1
shows the distribution of summer days across air quality categories. 28 days were
announced as code red days in 2001 through 2003, accounting for 6.8 percent of the
season. Code orange days indicate that the ozone concentration will reach a level
unhealthy for sensitive populations. These days account for 10.6 percent of the period.
The forecast as well as the code color are publicized throughvarious communication
channelsonce they are available. People who subscribe to a mailing list receive email
notification. Local employers who enroll in the clean air partner program receive an
email or fax. Major newspapers and TV and radio stations will report air quality forecasts
together with weather forecasts. News sources will highlight code red days to enhance
visibility for the program. People are urged to take actions to reduce ozone precursors
emissions on high ozone days. On the top of the action list is reducing driving by all
means, includingcarpooling, teleworking, riding transit, and consolidating trips.
65
3.3 Theory
A simple discrete choice model can be used to analyze individual's choice between
driving and its substitutes on code red days. Specifically, we consider staying/working at
home and using public transit as the alternatives an individual may choose. Other travel
options such as carpool and bicycle may be incorporated into the framework easily and
would not affect the main results obtained below.
Suppose an individual chooses to drive (d ), to ride public transit ( p ) or stay/work at
home (h ) in order to maximize her utility
ij ij ij
U V c = +
where i indexes individual, { , , } j d p h e . The utility is the sum of a deterministic part V
and an idiosyncratic part c . Further assume that the deterministic utility is a weighted
linear combination of travel benefits and a variety of travel costs. That is
0 1 2 3 ij ij ij ij ij
V B TC HC EC | | | | = + + +
where B stands for trip benefit ( 0
h
B = ), TC is travel cost, including gas, bus fare, and
time, HC is health cost resulting from exposure to bad air quality, and EC is the
environment cost associated with one's choice. The model assumes that everybody has
common weights, | 's, and
0
0 | > ,
1 2 3
, , 0 | | | < although the benefits/costs of each
choice vary by individuals. In one case, people may differ in the extent that they
internalize the negative impact on air quality for the same amount of driving, but they
value air quality equally. Also, people with existing respiratory problem may have greater
health costs than those without when exposed to the same air pollution.
66
We assume that
ij
c are independently distributed as type-I extreme value. Let
i
y
denote the choice of individual i that maximize the utility, i.e. argmax( , , )
i id ip ih
y U U U = .
The probability of choosing j is
3 3
0 0
Pr( | ) exp / exp
i i k ijk k ijk
k j k
y j x x x | |
= =
(
| | | |
= =
( | |
\ . \ .
¸ ¸
¿ ¿ ¿
where
0
x B
·
= ,
1
x TC
·
= ,
2
x HC
·
= , and
3
x EC
·
= . For all individuals,
( )/ ( )[1 ( )]
j jk j j k
p x x p x p x | c c = ÷ (3-1)
and
( )/ ( ) ( )
j lk j l k
p x x p x p x | c c = ÷ , l k = . (3-2)
We are interested in how probabilities of choosing different alternatives change when
it is a code red day. Equations (3-1) and (3-2) tell us that the probability change for any
alternative depends on the changes in each benefit/cost factors for the option itself and all
others on code red days. To derive further results from the model, we assert the following
changes and relationships
(i) 0
d p h
B B B A = A = A = ,
(ii) 0
d h p
TC TC TC A s = A s A ,
(iii) 0
h d p
HC HC HC = A s A < A ,
(iv) , 0
p h d
EC EC EC A A s s A .
Relationship (i) states that the benefit from making the trip does not change on
code red days for any option; (ii) reflects the assumption that people may expect other
people to forego driving for riding public transit. Thus, traffic is expected to be lighter
67
while transit becomes more crowded and uncomfortable; (iii) indicates that when air
quality gets worse, people walking to and waiting at the bus stop are more exposed to
ozone. Driving in a car may or may not increase risk while staying indoors is always safe;
and (iv) implies that people are altruistic and may gain satisfaction (negative cost) for not
driving on code red days or may feel guilty for driving. These relationships are sensible
and not all are necessary for reaching the theoretical conclusion below.
Taking into account the cost changes on code red days, the change in the
probability of driving (also taking transit and staying home) for an individual is
ambiguous. This is mainly because declining air quality lowers the travel cost and health
cost of driving relative to riding bus, although there may be some utility gain from
reducing emissions. Even if people do not speculate about the improved traffic on code
red days, or in some areas bus fares are waived for riders on high ozone days, which
results in lower travel cost for riding bus, bus ridership may still not go up due to health
concern about the air quality.
The above analysis shows that the voluntary information program does not
provide people incentives necessarily consistent with reducing driving on code red days.
It is important to empirically measure the impact of the program on driving amount.
3.4 Empirical Strategy
The primary question this study attemptsto answer is whether the AQAD program
changes individuals' travel behavior episodically. Do people reduce vehicle trips and/or
miles traveled on code red days? Ideally, we would like to have a random sample of
68
households in the Baltimore area, together withtheir daily VMTs for all summer days.
However, such micro-data do not exist. What is available, instead, is data measuring
traffic volumes during short time intervals on highways in the Baltimore area. (These will
be described in detail in the next section). With these data, we can estimate the following
model to measure changes in traffic on code red days that can be attributed to the AQAD
program.
it t t i it
y CRD X B ¸ u c = + + + (3-3)
where
it
y is (log) number of vehicles passing by traffic monitor i
20
on date t .
t
CRD is
an indicator for day t to be a code red day and the parameter ¸ measures the impact of
code red day announcement on highway traffic volumes. The vector
t
X contains other
time varying factors that may affect vehicle trips, such as contemporaneous and lagged
weather, the forecast 1-hour ozone concentration and observed ozone levels for the
previous day, contemporaneous and lagged gas prices, public holiday dummies and a set
of dummies for year, month, and day of the week. In a specification check, I include
lagged traffic of the same time block on previous days and seven days ago.
i
u represents
a monitor fixed effect to capture the time-invariant traffic characteristics for each
monitor. The variable
it
c is an unobserved idiosyncratic term.
The problem in consistently estimating ¸ is that code red daysare not random. Even
conditional on all those covariates, there still could be some variables missing in the
20
Please note this is different from the ozone monitoring stations mentioned earlier. Coincidentally, the
number of traffic monitors in the sample is also seven.
69
model that are correlated with the code red day announcement and traffic flow. For
instance, forecast weather plays a crucial role in predicting ozone concentration and
determining code color. It is also arguably important in affecting people's travel decision
for the coming day. However, historical forecast weather data is not readily accessible.
Although we control for the observed weather and its lag, they may fail to account
adequately for the forecast weather. If it is the case, a naïve regression estimation would
yield a spurious estimate of ¸ .
However, if we could control for the conditional expectation of the unobservables in
the model, we would still be able to estimate ¸ consistently. That is to estimate the
following model instead of equation (3-3),
( ) |
it t it t t i it
y CRD E CRD X B ¸ ì c u u = + + + + (3-4)
where ( ) |
it t
E CRD c is expectation of
it
c conditional on the code red day indicator, and
( | , , )
it it it t t i
y E y CRD X u u = ÷ . If
t
CRD is the only variable correlated with
it
c , OLS
estimation of equation (3-4) yields a consistent ¸ estimate. In practice, ( ) |
it t
E CRD c is
not observed. However, we know that the code red day announcement is completely
dependent on the forecast ozone concentration, denoted by O. When and only when the
forecast level exceeds a threshold, will it be a code red day. Formally,
*
( ) 1{ }
t t t
CRD f O O O = = >=
where
*
O denotes the threshold equal to 125 ppb. Thus, we could exploit the sharp
regression discontinuity design (e.g. van der Klaauw 2002) to measure the impact of code
red days. Since O, referred to as a running variable in the literature, captures all the
70
information contained in CRD , ( ) | ( | )
it t it t
E CRD E O c c = . Thus, we could estimate
equation (3-5)
( )
it t t t i it
y CRD k O X B ¸ u u = + + + + (3-5)
where ( ) k O is a flexible functional specification for ( | ) E O c . In the literature, ( ) k O
often takes the form of high-order polynomial series.
As noted above, the vector X contains a linear term inthe forecast ozone level. It
is, however, possible that the linear term is insufficient to completely account for the
correlation between CRD and c . Figure3-1 illustrates that the estimates (¸ ' ) obtained
by controlling only for the linear term in the forecast ozone level will underestimate (left
panel) or overestimate (right panel) the true effect when the correlation between CRD
and c is a nonlinear function of O. In the estimation, I include up to a fifth order
polynomial in the ozone forecast.
Two key assumptions must be satisfied in order to apply the regression discontinuity
method (Imbens and Lemieux 2008). First, it is assumed that there is no manipulation of
the running variable. In our case, if the expert panel adjusts the forecast ozone level to
move a day into or out of the code red category based on expected transportation
volumes, concern about the validity of RD strategy might be raised. A communication
from one of the panel members stated that no sophisticated traffic information (e.g.,
forecasted daily traffic volumes) beyond the day of the week was considered in
forecasting ozone concentration. More specifically, it happened occasionally that the
forecast level was adjusted upward for Monday or downward for Friday based on the
71
general impression about traffic patterns on these days. However, this is the only channel
through which traffic is taken into account in code red day classification. In the next
section, it is shown that Mondays and Fridays are not statistically more likely to be (or
not be) a code red day. In addition, all modelsare estimated controlling for day-of-week
dummy. In the robustness check, I exclude Mondays and Fridays from the sample used
for estimation.
The other assumption underlying the RD model is that the unobserved variables that
may affect traffic volumes evolve continuously at the cutoff point, i.e.125 ppb. This
assumption cannot be verified directly. As a specification test presented in the next
section, I check the discontinuity of the control variables, especially weather covariates.
If some variables are found discontinuous at 125 ppb, it casts doubt on the continuity
assumption for the unobservables.
3.5 Data
The Maryland Department of Environment (MDE) has archived the forecast and
observed daily maximum one-hour ozone concentrations for the Baltimore area. I use
data from May through mid-September --- the ozone season when the AQAD program is
in place --- from 2001 to 2003. I focus on these three years because traffic data is
available from 2001 and the color code assignment started to be based on an 8-hour
ozone forecast and the 8-hour standard in 2004.
21
The code color is also available from
21
The 8-hour standard is stricter in the sense that more days are designated as exceedance days. However,
most exceedance days are code orange. It may be interesting to examine how the scheme change affects
people's responses.
72
MDE. Alternatively, it can be determined by applying the rule described in Table3-1. The
latter matches the recorded one perfectly, which confirms the relationship between code
color and forecast ozone. A sharp RD rather than fuzzy RD model is therefore appropriate.
In the early 2000s, the Maryland State High Administration (SHA) started to install
detectors
22
along major roads to monitor and record traffic conditions. The detectors
count the number of vehiclesand volume is reported in five-minute intervals. For the
project, weekday traffic volumes of the years2001 through 2003 were obtained from the
University of Maryland's Center for Advanced Transportation Technology Laboratory
(CATT Lab), which archives data from the Maryland SHA.
As the detector system was establishedshortly before the period we examine, the
performance of detectors and the data transfer network was not ideal. This resulted in
considerable missing data. I restrict the set of detectors to be analyzed to those with less
than 30 percent of 5-minute intervals missing, whichgives seven detectors located on
four major interstate highways in the Baltimore metropolitan area.
23
Table3-2 provides
information about the detectors and the traffic they are monitoring. The routes where
these detectors are located all carry heavy traffic from the surrounding areas into and out
of the Baltimore urban area. These roads rank from 3rd to 8th in terms of 2003 annual
average daily traffic (AADT) in the area. Unfortunately, we do not have data for I-95 and
22
Different from those buried underneath the road surface, this type of detectors is usually mounted on
existing side-of-the-road poles and work with microwave sensor technology. See http://www.rtms-by-
eis.com/rtms_features.html for more information.
23
See Figure3-2 for a map showing major freeways of the region and locations of detectors.
73
I-695, two major routes through and around the Baltimore area, respectively. Five
detectors monitor inbound traffic while two monitor outbound traffic.
I aggregate the 5-minute volumes into four time blocks following the definitions in
the Baltimore Metropolitan Council's travel demand model (BMC 2004). They are
morning peak (6 AM – 10 AM), mid-day (10 AM – 3 PM), evening peak (3 PM – 7 PM),
and other times (7 PM – 6 AM). This aggregation largely overcomes the short-term
fluctuations in traffic flow caused by traffic conditions. More importantly, the time
blocks group together hours with homogenous traffic patterns and separate those with
different patterns. It is therefore more appropriate to study traffic pattern changes at the
time block level than at 5-minute interval or hourly levels.
It is difficult to fill in missing observations on traffic volume. In general, filling in
missing values of the dependent variable in a single-equation regression may lead to
biased estimation (Greene 2003). So time blocks with one or more missing 5-minute
interval are dropped from the regression. This may lead to an efficiency loss but should
have no effect on estimator consistency so long as the time blocks that do not enter the
volume regression are not systematically correlated with the explanatory variable of
interest, i.e. CRD. Table3-3 presents checks on the correlation between dropped time
blocks and code red days. Each column represents a probit model specification with
incremental inclusion of control variables. When CRD is the only explanatory variable
(column (1)), it seems to affect the missing pattern of all times of day except for the
morning peak period. However, sinceCRD occurs on hot, sunny days it may pick up
74
weather impacts on the detector system. When the linear forecast ozone level and a full
set of covariates including weather conditions are included (columns (2) and (3)) in the
model, the effect of CRD becomes smaller and statistically insignificant. Adding
polynomials terms in the forecast ozone level (column (4)) does not change the result at
all. Thus, we conclude that estimating equation (3-5) with only the complete time blocks
should not give us biased estimates due to missing observations.
Daily weather measuresincluding temperature (maximum and minimum), wind
speed, relative humidity (maximum and minimum) and precipitation wereobtained from
the National Climatic Data Center and are observed at the weather station located in the
Baltimore-Washington International Airport. Daily average prices for regular unleaded
gasoline in Baltimore area are provided by the GasBuddy Organization. I use the first
through seventh lags of gasoline price to control for the impact of gas price on travel
demand.
Columns (1) and (2) inTable3-4 presents the means and standard deviations of
control variables for non-code-red days and code red days, respectively. Column (3)
shows differences in means between the two types of days and the associated standard
errors. Generally speaking, a code red day is more likely to occur on hot, dry days with
lower wind speed. The observed ozone level for the day before the forecasted day is
significantly higher for the code red day. However, neither the short-term historical retail
gas price nor the day of week differ significantly, which is consistent with the fact that the
ozone forecasting model is basically a meteorological model rather than a behavioral one.
75
Although adjustments may have been made accounting for the day of the week, it seems
to be a rare unsystematic practice.
One key identification assumption mentioned earlier is that the conditional mean of
the unobservable, i.e. ( | ) E O c , is continuous at
*
O =125 ppb. The evidence in support of
this assumption can be found by testing the continuity of the observed covariates. Figure
3-3 plots the average daily characteristics including temperature (max and min),
precipitation, wind speed, humidity (max and min), retail gas price (lags), ozone
observation (lag) and Monday and Friday dummies, against the forecast of ozone level.
The predicted values from a fifth-order polynomial in the forecast level as well as the 95
percent confidence intervals are also presented. The figures suggest that there is no large,
statistically significant break for these variables when ozone levels change from non-
code-red days to code red days. Columns (4) and (5) of Table3-4 provide quantitative
support for this finding. Although code red days are different from non-code-red days, as
shown in column (3), when the comparison is narrowed between code red days and code
orange days in column (4), the difference diminishes dramatically in magnitude across all
variables and only the max temperature and observed ozone level remain statistically
significant. The higher max temperature and observed ozone level the day before most
likely reflect only the higher forecast ozone level for code red days. Column (5),
equivalent to Figure3-3, reports the estimated coefficient of CRD when a fifth-order
polynomial in forecast ozone level is included in the regression. It indicates that the
difference between code red days and non-code-red days is small and statistically
76
insignificant conditional on the forecast level. These results suggest that the unobserved
characteristics are unlikely to be discontinuous at the CRD cutoff point.
3.6 Results
Table3-5presents the estimates of the effects of CRD on traffic volumes by time of
day. Each model is estimated for a pooled sample as well as two sub-samples separating
inbound detectors from outbound detectors. Although detector fixed effectsaccount for
the unique features of traffic pattern for each location and direction, it may be true that
inbound and outbound traffic respond to CRD in distinct ways. Further, given symmetry
between morning and evening travel, i.e. the morning inbound (outbound) traffic returns
in the evening on the same routes, we should expect to see CRD have similar impacts on
morning inbound (outbound) traffic and evening outbound (inbound) traffic. The sample
is therefore split to explore the heterogeneity in theeffects of CRD on inbound and
outbound traffic. Common covariates across models include weather conditions and their
lags, the observed ozone level for the previous day, lagged retail gas prices, and dummies
for year, month, day of the week, public holidays and detectors.
24
Overall, the models
explain traffic patterns reasonably well, with an
2
R above 0.90 for the full sample and
above 0.80 for inbound and outbound sub-samples.
Columns (1)-(3) report the results of the first specification, in which the model
controls for the ozone forecast in linear form only. For the full sample, CRD decreases
morning peak traffic by 1.7 percent but increases mid-day traffic by 3.3 percent. The
24
Models including lagged traffic on previous days and seven days ago yield no different estimates.
77
average weekday morning peak and mid-day traffic volumes across monitors are about
8500 and 8000, respectively. Applying the estimates suggests that on average 145 or so
vehicle trips from 6 AM to 10 AM were cancelled or moved to other time periods. For
the mid-day hours 10 AM to 3 PM, trips rose by about 264, which could be a result of
trip rescheduling from the morning and/or people switching to driving to avoid ozone
exposure. However, it is not obvious why people would postpone their vehicle travel
closer to noon. When we look at inbound and outbound traffic separately, the CRD has
little effect on inbound traffic except for increasing mid-day volumes by 4 percent. It
lowers outbound traffic by 2.6 percent in the morning and 3.3 percent in the evening.
These results are not consistent with a symmetric traffic pattern between inbound and
outbound routes. As we discussed before, these estimates could be biased if the control
function of the forecast level is not flexible enough.
Columns (4)-(6) are the baseline regression discontinuity modelsusing a fifth order
polynomial in forecast ozone to proxy for ( | ) E O c . Column (4) shows that the CRD
reduces morning traffic by 5 percent for the pooled sample, which is equal to about 425
vehicle trips on average. In contrast to column (2), the CRD does not exhibit a
statistically significant impact on traffic during other periodsof a day. When inbound and
outbound traffic are examined separately (see columns (5) and (6)), the CRD is found to
lower the morning peakinbound traffic by 5 percent. Moreover, this reduction is matched
in the outbound traffic, which declines 2.6 percent in the evening peak and 5 percent in
other hours on code red days. The coefficients of CRD are positive for the inbound
78
sample during mid-day and evening and negative for morning outbound, but neither is
statistically significant. These resultssuggest that the code red day alert indeed reduces
traffic, albeit by a small proportion.
25
Columns (7)-(9) maintain the RD specification and exclude the code green days
from the sample, to test whether the results are driven by observations far away from the
cutoff point.
26
The main finding remains unchanged: morning inbound traffic declines by
5 percent and the evening outbound traffic declines correspondingly. The difference is
that outbound trip reductions are concentrated in the evening peak hours rather than in
other hours. Theseestimates imply that the RD approach is appropriate for measuring the
effect of the AQAD program, which a normal regression fails to capture.
Table3-6 reportsadditional tests of the robustness of the results. The expert panel
occasionally manipulatesthe ozone forecast and/or code color on Mondays and Fridays
to account for traffic patterns, but not on other days of the week. The RD strategy is
plausible if it yields similar estimates with a sample containing only Tuesday through
Thursday. Columns (1)-(3) show that morning traffic is lower by 3 percent for the pooled
sample and lower by 4 percent for the inbound sub-sample on code red days. Outbound
traffic is reduced by 3 percent, thoughthe effect is not statistically significant. Although
25
Table A3 provides full estimation results for models specified in columns (4) through (6) of Table3-5.
26
Table3-1shows that code green days account for 51 percent of sample days and code yellow days
account for 31 percent. The estimates appear to be sensitive to individual observations when code yellow
days are removed. It is more likely because of the dramatic decrease in sample size.
79
the samples diminish in size by two fifths, we still find evidence consistent with the
baseline results
Another test of the findings is to see whether similar reductions occur on other days,
Code orange days mean an air quality alert to the public, although not to the same degree
ason a code red day. Therefore, we expect no or a smaller decline in vehicle trips on
code orange days. The hypothesis is tested in two specifications. I replacethe CRD
dummy with a dummy for code orange and code red days in the first case, and use two
dummies, one for code orange days and one for code red days, in the second case.
Columns (4)-(6) of Table3-6 report the first specification and columns (7)-(9) report the
second. The dummy for code orange and red days does not have a negative impact on
traffic volumes. Instead, it increases the morning inbound traffic slightly. When models
includetwo separate dummies, the code red dummy has significant negative effects on
morning inbound volumeswhile the code orange dummy has statistically insignificant
effects sometimes opposite to code red. Both results suggest that drivers do not respond
to code orange as they do to code red days.
Code red days often occur on consecutive days. The cost of foregoing driving may
rise on the second or third code red day. For instance, it may be easier for a professional
to work from home one day a week than two or three days a week. On the other hand, if
the marginal cost of driving on code red days increases, an individual is more likely to
take some action on the second or third code red day than on the first. I estimate the
following models to see the effect of consecutive code red days. In addition to the code
80
red day dummy, I add to the model a dummy equal to one if it is the second, third or
fourth (the longest string is four).code red day in a row, or a dummy for the third or
fourth code red day. Columns (1)-(3) and (4)-(6) of Table3-7report the results for these
two cases respectively. For the pooled sample, morning traffic is lower by 6 percent on
the first code red day but only 4 percent lower for the second, third or fourth code red day.
When the sample is split between inbound and outbound detectors, however, the effect
loses statistical significance. A dummy for the third or fourth code red days in a row does
not exhibit reinforcing or offsetting effects either.
3.7 Discussion and Conclusion
The findings of this study are similar to those in Cutter and Neidell's (2007), which
examines a similar program in the San Francisco Bay Area. Both results suggest that the
voluntary information programs lead to a small reduction in vehicle trips and the effect
most is concentratedin the morning peak period. The evidence that the reduction occurs
for the morning inbound traffic and evening outbound traffic seems to provide additional
support for the main results.
The timing of the effect suggests that it is commuting trips that are reduced. The
workers who usually drive to work could potentially work at home or switch to other
travel modes such as taking public transportation, walking/biking and carpooling on code
red days. The former is especially likely since, according to the 2001 National Household
Travel Survey, 7.7 percent of workers work at home at least one day every month and 4.7
81
percent at least one day every week in the Washington-Baltimore metropolitan area.
These people should generally have more flexibility to avoid driving on code red days.
Although the program is demonstrated to have some expected impact, the magnitude
seems too small to reduce vehicle emissions dramatically. As the literature suggests
voluntary programs are unlikely to improve air quality sufficiently to bring a region into
compliance status. An innovative approach would be a permit program that restricts
driving on high ozone days unless a permit is bought for each vehicle. The program could
be effective if the permit price is set high enough, which presents a strong disincentive
for many people to drive. Imposing the control on an episodic basis means the program
could be more economically efficient than programs with year-round controls.
One objective of the programs like AQAD is to see how education and persuasion
might alter individuals' behavior in favor of the environment. This study indicates that
these efforts are not made in vain. In addition to limiting driving, the program also asks
people to refuel vehicles after dusk or on another day. It may be worth investigating
whether this is an easier behavioral change for people to make once data are available.
82
Figures and Tables for Chapter 3
Figure 3-1. Illustration of Biased Estimates with Linear Forecast Ozone Level
Note: The graphs show that the estimated CRD effect, ¸ ' , controlling for linear forecast ozone level could
underestimate (left) or overestimate (right) the true effect of CRD, ¸ , when the underlying relationship
between y and forecast ozone level is nonlinear.
0
Forecast ozone level (ppb)
125
¸
¸ '
125
¸
¸ '
Forecast ozone level (ppb)
0
y y
83
Figure 3-2. Map of Baltimore Region Major Freeways and Maryland SHA’s Traffic
Detectors
Source: Created with ArcMap using Census 2000 TIGER/Line®Shapefiles and Bureau of Trasportation
Statistics’ Highway Performance Monitoring System data. Locations of detectors are not accurate but
illustrative.
84
Figure 3-3. Similarity of Covariates around Code Red Day Cutoff Point
5
0
6
0
7
0
8
0
9
0
1
0
0
D
a
i
l
y
m
a
x
i
m
u
m
t
e
m
p
e
r
a
t
u
r
e
40 60 80 100 125140
Forecate ozone level
4
0
5
0
6
0
7
0
8
0
D
a
i
l
y
m
i
n
i
m
u
m
t
e
m
p
e
r
a
t
u
r
e
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
2
0
0
2
5
0
D
a
i
l
y
p
r
e
c
i
p
i
t
a
t
i
o
n
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
2
0
0
A
v
e
r
a
g
e
d
a
i
l
y
w
i
n
d
s
p
e
e
d
40 60 80 100 125140
Forecate ozone level
8
0
8
5
9
0
9
5
1
0
0
1
0
5
M
a
x
i
m
u
m
r
e
l
a
t
i
v
e
h
u
m
i
d
i
t
y
40 60 80 100 125140
Forecate ozone level
2
0
4
0
6
0
8
0
1
0
0
M
i
n
i
m
u
m
r
e
l
a
t
i
v
e
h
u
m
i
d
i
t
y
40 60 80 100 125140
Forecate ozone level
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
1
d
a
y
a
g
o
40 60 80 100 125140
Forecate ozone level
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
3
d
a
y
s
a
g
o
40 60 80 100 125140
Forecate ozone level
1
.
2
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
7
d
a
y
s
a
g
o
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
l
a
g
o
f
o
b
s
e
r
v
e
d
o
z
o
n
e
40 60 80 100 125140
Forecate ozone level
-
.
5
0
.
5
1
M
o
n
d
a
y
40 60 80 100 125140
Forecate ozone level
-
.
5
0
.
5
1
1
.
5
F
r
i
d
a
y
40 60 80 100 125140
Forecate ozone level
Note: The dots represent the average daily characteristics for each forecast ozone level. The continuous line
is the predicted values from a fifth-order polynomial in forecast level with the dashed lines for 95 percent
confidence interval.
85
Table 3-1. One-hour Ozone Level and Code Colors in Baltimore Area
1-hour Ozone
(ppb)
Code Color Health Concern Number of Days
2001-2003
0 – 79 Green Good 214
80 – 104 Yellow Moderate 128
105 – 124 Orange Unhealthy for sensitive groups 44
>125 Red Unhealthy 28
Source: http://www.mwcog.org/environment/air/downloads/calendar_2003.pdf, MWCOG.
86
Table 3-2. Description of Detectors and Traffic in Baltimore Area
Detector Highway AADT 2003 Direction In/Out Bound Missing %
1 I-83 86,293 (4
th
) S In 9.7
2 I-83 86,293 (4
th
) S In 26.3
3 I-83 86,293 (4
th
) N Out 18.5
4 I-795 81,500 (5
th
) S In 17.8
5 I-97 105,008 (3
rd
) N In 17.3
6 I-97 105,008 (3
rd
) S Out 9.4
7 I-70 44,142 (8
th
) E In 10.4
Note: The third column presents annual average daily traffic in 2003. The top two roads missing
here are I-95 with AADT of 169,534 and I-695 with AADT of 167,473.
87
Table 3-3. Correlation between Missing Time Block and Code Red Day
(1) (2) (3) (4)
Morning peak 0.272 0.325 0.219 0.453
(0.249) (0.221) (0.148) (0.350)
0.002 0.003 0.117 0.120
Mid-day 0.445 -0.068 -0.172 0.025
(0.160) (0.157) (0.167) (0.244)
0.006 0.021 0.154 0.155
Evening peak 0.665 -0.064 -0.043 0.137
(0.125) (0.138) (0.145) (0.206)
0.014 0.044 0.179 0.183
Other 0.349 0.300 0.142 0.218
(0.179) (0.174) (0.176) (0.237)
0.004 0.004 0.088 0.089
N 2898 2884 2849 2849
Linear forecast ozone level N Y Y Y
2
nd
to 5
th
order polynomials
of the forecast ozone N N N Y
Control variables N N Y Y
Note: Dependent variable is a binary indicator equal to 1 if the time block is to be
dropped from traffic volume equation. The first row of each time-of-day panel is
probit estimates of the coefficient of CRD, the second row is standard error
clustered on each week, and the third row is the pseudo-R squared. Control
variables include weather variables and their lags, observed ozone levels, lags of
daily gas price, and year, month, day-of-week, holiday, and monitor dummies,.
88
Table 3-4. Summary Statistics and Difference in Selected Covariates Between Code Red
Days and Other Days
Non-CRD CRD
CRD vs.
Non-CRD
Orange
vs. Red Polynomials
(1) (2) (3) (4) (5)
Max temperature 80.903 94.462 13.558 2.873 1.051
8.771 3.313 (0.842) (0.875) (1.935)
Min temperature 60.824 69.538 8.714 2.480 1.630
8.959 4.282 (1.002) (1.347) (2.826)
Precipitation 14.896 1.000 -13.896 3.412 -4.202
39.252 2.966 (2.463) (2.533) (9.279)
Wind speed 62.654 52.654 -10 -5.434 -0.375
25.356 12.103 (2.833) (3.385) (7.139)
Min relative humidity 51.639 40.846 -10.793 -1.624 3.273
16.004 8.698 (1.965) (2.246) (4.974)
Max relativehumidity 94.457 90.808 -3.650 -0.928 3.329
6.794 6.487 (1.338) (1.682) (3.025)
Gas price 1 day ago 1.462 1.449 -0.013 0.021 0.048
0.121 0.116 (0.024) (0.034) (0.063)
Gas price 3 days ago 1.464 1.45 -0.014 0.016 0.041
0.121 0.107 (0.022) (0.032) (0.063)
Gas price 7 days ago 1.465 1.445 -0.020 0.002 0.036
0.118 0.121 (0.025) (0.033) (0.066)
Ozone (lag) 75.526 123.654 48.128 17.595 -7.890
23.293 20.829 (4.327) (5.607) (12.863)
Monday 0.197 0.192 -0.005 -0.102 0.063
0.398 0.402 (0.082) (0.112) (0.245)
Friday 0.208 0.154 -0.054 0.007 0.200
0.407 0.368 (0.076) (0.095) (0.154)
N 269 26 295 60 295
Note: Columns (1) and (2) are means and standard deviations (underneath) for non-code red days and code
red days, respectively. Column (3) is the difference between CRD and non-CRD. Column (4) is the
difference between CRD and code orange days. Column (5) is the estimate of CRD coefficient regressing
each covariate on CRD and a fifth order polynomial in forecast ozone level. In the parentheses are standard
errors. Those standard errors in the column (5) account for within-week clustering.
89
Table 3-5. Impact of Code Red Day Announcement on Traffic Volumes by Time of Day
All Inbound Outbound All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Morning Peak
Coefficient -0.017** -0.012 -0.026** -0.051* -0.051** -0.034 -0.063* -0.052* 0.008
Std. errors (0.007) (0.009) (0.011) (0.030) (0.024) (0.046) (0.032) (0.027) (0.026)
N 1520 1059 461 1520 1059 461 736 517 219
2
R
0.96 0.97 0.96 0.96 0.97 0.96 0.97 0.98 0.97
Mid-day
Coefficient 0.033* 0.039* 0.018 0.026 0.045 -0.006 -0.084* -0.092 -0.02
Std. errors (0.018) (0.020) (0.021) (0.043) (0.057) (0.030) (0.049) (0.056) (0.042)
N 1119 795 324 1119 795 324 485 360 125
2
R
0.93 0.94 0.87 0.93 0.94 0.87 0.96 0.96 0.93
Evening Peak
Coefficient -0.007 0.004 -0.033* 0.03 0.07 -0.026* 0.011 0.068 -0.069**
Std. errors (0.02) (0.026) (0.017) (0.059) (0.093) (0.014) (0.102) (0.164) (0.029)
N 1157 787 370 1157 787 370 461 316 145
2
R
0.90 0.86 0.88 0.90 0.86 0.88 0.91 0.89 0.92
Other
Coefficient 0.006 0.012 -0.013 -0.019 -0.01 -0.050* 0.009 0.018 -0.028
Std. errors (0.012) (0.011) (0.018) (0.019) (0.018) (0.027) (0.027) (0.026) (0.049)
N 1201 839 362 1201 839 362 566 398 168
2
R
0.95 0.96 0.83 0.95 0.96 0.83 0.97 0.98 0.88
Note: Dependent variable is log of traffic volumes. Control variables include weather conditions and their lags, forecast 1-hour ozone concentration,
observed ozone level for the day before, lagged retail gas prices, and dummies for year, month, day of the week, public holiday and monitor. Columns (1)-
(3) control for linear ozone forecast only. Columns (4)-(9) control for a fifth-order polynomial in ozone forecast. Columns (7)-(9) focus on a sub-sample
excluding code green days. Standard errors account for within-week clustering. * indicates significance at 10 percent level while ** at 5 percent level.
90
Table 3-6. Robustness Check
All Inbound Outbound All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Morning Peak
Code orange or both 0.019 0.036* -0.003 0.005 0.025 -0.016
(0.02) (0.021) (0.025) (0.014) (0.018) (0.019)
Code red day -0.028* -0.037* -0.018 -0.049* -0.038** -0.042
(0.015) (0.019) (0.026) (0.026) (0.019) (0.043)
N 903 627 276 1520 1059 461 1520 1059 461
2
R
0.98 0.98 0.99 0.96 0.97 0.96 0.96 0.97 0.96
Mid-day
Code orange or both -0.017 -0.025 -0.003 -0.011 -0.013 -0.006
(0.027) (0.035) (0.028) (0.030) (0.037) (0.033)
Code red day -0.001 0.006 0.002 0.02 0.038 -0.009
(0.059) (0.080) (0.040) (0.047) (0.063) (0.035)
N 669 473 196 1119 795 324 1119 795 324
2
R
0.94 0.94 0.90 0.93 0.94 0.87 0.93 0.94 0.87
Evening Peak
Code orange or both -0.024 -0.053 0.019 -0.018 -0.039 0.012
(0.032) (0.044) (0.016) (0.040) (0.057) (0.017)
Code red day 0.053 0.123 -0.032 0.019 0.047 -0.019
(0.114) (0.188) (0.020) (0.070) (0.110) (0.014)
N 679 459 220 1157 787 370 1157 787 370
2
R
0.92 0.90 0.60 0.90 0.86 0.88 0.90 0.86 0.88
Other
Code orange or both 0.016 0.02 0.021 0.013 0.02 0.007
(0.014) (0.014) (0.022) (0.014) (0.015) (0.025)
Code red day 0.030 0.023 0.029 -0.012 0 -0.046
(0.028) (0.029) (0.046) (0.020) (0.019) (0.032)
N 718 506 212 1201 839 362 1201 839 362
2
R
0.96 0.97 0.64 0.95 0.96 0.83 0.95 0.96 0.83
Note: Dependent variable is log of traffic volumes. Columns (1)-(3) use samples excluding Monday and Friday. Columns (4)-(6) replace the CRD dummy with a
code-red-or-orange dummy. Columns (7)-(9) add a code-orange dummy in addition to the CRD dummy. Standard errors account for within-week clustering. *
indicates significance at 10 percent level while ** at 5 percent level.
91
Table 3-7. Impact of CRDs in Sequence on Traffic Volumes by Time of Day
All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6)
Morning Peak
Code red days -0.057* -0.058** -0.034 -0.054* -0.051** -0.035
(0.030) (0.024) (0.047) (0.028) (0.022) (0.047)
CRDs in Seq. 0.013* 0.015 0.001 0.006 -0.001 0.002
(0.007) (0.010) (0.009) (0.007) (0.010) (0.010)
N 1520 1059 461 1520 1059 461
2
R
0.96 0.97 0.96 0.96 0.97 0.96
Mid-day
Code red days 0.020 0.040 -0.009 0.011 0.029 -0.009
(0.042) (0.056) (0.028) (0.039) (0.053) (0.026)
CRDs in Seq. 0.015 0.015 0.006 0.043 0.049 0.007
(0.035) (0.044) (0.016) (0.029) (0.034) (0.028)
N 1119 795 324 1119 795 324
2
R
0.93 0.94 0.87 0.93 0.94 0.87
Evening Peak
Code red days 0.020 0.056 -0.029 0.045 0.094 -0.033**
(0.047) (0.072) (0.019) (0.062) (0.098) (0.016)
CRDs in Seq. 0.017 0.024 0.005 -0.035 -0.054 0.017
(0.030) (0.047) (0.021) (0.023) (0.033) (0.021)
N 1157 787 370 1157 787 370
2
R
0.90 0.86 0.88 0.90 0.86 0.88
Other
Code red days -0.011 -0.002 -0.043 -0.016 -0.006 -0.045*
(0.023) (0.024) (0.029) (0.022) (0.023) (0.027)
CRDs in Seq. -0.017 -0.021 -0.014 -0.008 -0.011 -0.014
(0.020) (0.019) (0.025) (0.023) (0.025) (0.025)
N 1201 839 362 1201 839 362
2
R
0.95 0.96 0.83 0.95 0.96 0.83
Note: Dependent variable is log of traffic volumes. Columns (1)-(3) has one additional dummy equal
to one if the code red day is the second, third or fourth one in a row. Columns (4)-(6) has a dummy
equal to one if the code red day is the third or fourth one in a row. Standard errors account for within-
week clustering. * indicates significance at 10 percent level while ** at 5 percent level.
92
4 Concluding Comments
4.1 Summary of Results
Chapter 2 calculates the percent of workers who use the Internet when
working at home in a person’s two-digit occupation by city size cell to instrument for
telecommuting choice. After controlling for occupation and city fixed effects, as well
as individual and household characteristics, this variable still predicts men and
married women’s probability to telecommute: a 10 percentage point increase in the
internet penetration causes telecommuting probability to rise by 5 percentage points
for married women and 3 percentage points for men.
Using this variable to instrument for telecommuting choice yields IV
estimates that telecommuting leads a married women’s one-way commute time to
increase by 9 to 12 minutes. The effect on men’s commute length is smaller, at about
5 minutes and statistically indistinguishable from zero. The results are robust for
different specifications and sub-samples. Contrary to the OLS estimates, IV
estimation finds that probability of commuting by driving does not decline due to
telecommuting. For the average married female worker who commutes 24 minutes
one way five days a week, telecommuting lowers weekly total commute time from
240 minutes to 198 minutes if the woman telecommutes two days a week. This means
a 17 percent reduction, less proportional to the reduction in commuting frequency.
Chapter 3 finds that the code red day announcement results in a 4-5 percent
reduction in vehicle commute trips in morning peak hours. However, this estimate is
obtained only when we include in the regression a flexible function of forecast ozone
levels, which is designed to control for the correlation between the code red day
93
indicator and any non-random unobservables. If only the linear term of forecast ozone
level is controlled in the regression, the estimate is as small as 1.7 percent. The
difference highlights the importance of the identification strategy used.
The conclusion from these results is that the two TDM strategies work to
some degree. Telecommuting is not shown to have a rebound effect on men’s
commute length. For married women, the effect seems moderate enough to result in a
net reduction in commute miles. Consistent with findings in northern California,
information about bad air quality could induce a small proportion of people to refrain
from driving. It is more likely that these people will work from home rather than to
switch to another travel mode. The effect, however, is not large enough to cause air
quality improvements.
4.2 Directions for Future Research
Some questions related this dissertation remain unsolved. The instrumental
variable developed in Chapter 2 does not have much power in explaining single
women’s telecommuting choices. Thus, the analysis cannot provide information
about the responses of single females to telecommuting.
Commuting length reflects both residential location choice and work location
choice. Telecommuters who choose longer one-way commute distances could choose
to live farther from work or to work farther from home. It is important to distinguish
the two possibilities from a public policy perspective.
For a two-earner household, housing location is determined jointly by both
husband’s and wife’s employment locations. A change in one person’s commuting
cost might lead to changes in the commute lengths of both people. In this case, the
94
sum of commute lengths of the household would be be the variable of interest. More
research is needed to understand the impact of household members’ telecommuting
on total household commute length.
A natural extension of Chapter 3 is to apply the same technique to similar
programs that have free bus fares. Free bus fares decrease the cost of riding a bus.
However, for people who are used to driving, a larger share of the cost of switching to
transit is the time and inconvenience to get on a bus. Moreover, it is of interest to
know whether such a program passesthe cost-benefit test. The voluntary program
takes an episodic approach to controlling ozone, which is valuable in designing a
pricing control scheme. Since ozone episodes occur only on hot, sunny days, the
government could set a price for driving on those days. Daily automobile travel and
resultant emissions could be managed by choosing a permit price. The cost-
effectiveness of such a program if implemented in the Washington metropolitan area
is being evaluated in an ongoing project.
Finally, economists may not want to give up the idea of managing travel
demand via non-pricing strategies. Many TDM strategies may be effective in various
contexts and even cost-effective if the political costs of pricing strategies are taken
into account. Clearly, economists should participate in the design and evaluation of
the TDM programs.
95
Appendices
Appendix 1 A Monocentric City Model with Commuters and Telecommuters
In a closed city, each household has only one worker and all employment
concentrates in the central business district (CBD). Workers commute to work at the
CBD along a radial network. Commuting costs per mile traveled are e , so a worker
who lives d miles from the CBD spends 2ed on daily commuting. All workers earn
the same income y per day. Household utility is described by a strictly quasi-
concave function ( , ) u c h , where c represents consumption of a composite non-
housing good and h is consumption of housing that could be measured in square feet
of floor space or number of rooms. The price of the composite good is assumed to be
the same across different locations of the city and normalized to 1. The daily rental
price of a unit of housing, denoted p , depends on location.
Initially, suppose all workers are identical. They maximize household utility to
reach a constant level, u . That is
{ , }
max ( , )
c h
u c h u = (A1)
s.t. 2 c ph ed y + + = .
Substitute 2 c y ph ed = ÷ ÷ into Eq. (A1) and notice that equilibrium housing price
and consumption are both functions of distance to the CBD, i.e. d . We have
( ) ( ) ( ) 2 , ( ) u y p d h d ed h d u ÷ ÷ = . (A2)
96
Totally differentiating Eq. (A2) and applying the envelop theorem, we get the well-
known conditions on the market equilibrium rent gradient that,
( ) 2
'( )
( )
p d e
p d
d h d
c
= = ÷
c
, (A3)
and
2
2 2
( ) 2 '( )
"( )
( )
p d eh d
p d
d h d
c
= =
c
. (A4)
Eqs. (A3) and (A4) imply that the housing price declines with commuting distance
and the rent gradient gets flatter as distance increases since '( ) 0 h d > . Plotted on a
plane with distance to the CBD as the x-axis and rent as the y-axis, the rent curve is a
downward-sloping convex function. Intuitively, workers who live in the suburbs with
longer commute are compensated by cheaper and larger homes.
Now, extend the model to including two types of otherwise identical workers:
commuters (c ) and telecommuters (tc ). Because the latter commute less often than
the former, the average daily commuting costs are lower for telecommuters than for
commuters. Therefore, there are separate rent offer curves for the two types of
workers, respectively. They are characterized as
2
( )
( )
i
i
i
e
p d
h d
'
= ÷
where , i c tc = . Assuming that housing is a normal good, then ( ) ( )
c tc
h d h d < .
Together with
c tc
e e > , we have
( ) ( )
c tc
p d p d
' '
> .
97
The rent offer of telecommuters declines slower than that of commuters. Figure A1
illustrates the two rent offer curves and the market rent gradient in equilibrium. The
telecommuters' rent offer curve (CD) is flatter than commuters' rent offer curve (AB)
while the two intersect at a certain distance
o
d d = . Commuters outbid telecommuters
for housing at locations closer to the CBD (
o
d d < ), as segment AO lies above CO,
and vice versa for locations beyond
o
d . The market equilibrium rent gradient is the
upper segments of the two offer curves (AO and OD). This means in equilibrium
commuters occupy the entire ring-shaped region around the CBD from distance 0 to
o
d while telecommuters sort into the surrounding ring from
o
d to
*
d , the city edge
determined by exogenous farmland rent. Thus, telecommuters have longer commutes
than commuters.
98
Appendix 2 Imputation of Top-Coded Commuting Time in the PUMS
First, I estimate a Pareto distribution to approximate the right-hand tail of the
commute time distribution, i.e.
( 1)
( )
a a
f x ab x
÷ +
= , for b x s s ·
where a is the parameter of the distribution, b is a constant from which commuting
time is assumed to follow a Pareto distribution, x is observed individual commuting
time equal to or greater than b . To obtain an estimate for a , I estimate Pr( ) x t > ,
where t is the top-coded value, i.e. 99 in PUMS, by the fraction of people commuting
b or more minutes who are top-coded, and exploit the relationship ( ) Pr( ) /
a
x t b t > = .
Therefore,
( )
t b
t x
a
ln ln
) r( P
ˆ
ln
ˆ
÷
>
= .
Then, the top-coded values are replaced by the estimated conditional expectation
of commuting time,
1
( | ) ( 1) E x x t ta a
÷
> = ÷
? ?
where t is the top-coded value, i.e. 99 in PUMS. Thus, different values for b yield
different estimates of a and the imputing value for top-coded observations.
For instance, let 50 b = , then 378,211 observations have commuting time equal
to or above 50 minutes, 16.4 percent of which are top-coded observations. Thus,
( ) ln(0.164)/ ln(50) ln(99) 2.65 a = ÷ =
?
. The conditional expectation for top-coded
individuals equals 159.1. When b is varied from 40 to 90 in increments of 10, the
conditional expectation estimates vary from 123 to 165 with an average of 150.
99
Figure A1. Bid Rent Curves in a Monocentric City with Telecommuters and
Commuters
p(d)
d
A
B
C
D
O
d* do 0
100
Table A1. CPS and NHTS Sample Construction
May 2001 CPS 2001 NHTS
Original sample 131,997 160,758
15 or older, employed with information on
working at home
50,743 65,697
Not self-employed 45,217
Reasonable commute distance and speed 62,283
MSA residents 32,272 50,810
Final sample without missing values on
any covariates
29,147 47,730
Note: Reasonable commute distance refers to one-way commute time below 180 minutes and
commute distance below 180 miles; reasonable speed refers to speed between 0.01 mile per
minute and 1.5 miles per minute.
101
Table A2. Distributions by Commute Mode across Cities
Commute mode Driving Rail Bus
Nationwide cities with 1 million or more
population
0.878 0.044 0.463
Atlanta, GA 0.964 0.002 0.013
Baltimore, MD 0.883 0.032 0.053
Chicago-Gary-Kenosha, IL-IN-WI 0.819 0.112 0.034
San Francisco-Oakland-San J ose, CA 0.822 0.041 0.071
Washington DC-VA-MD-WV 0.831 0.072 0.064
Source: Author’s calculation using NHTS 2001. Commute mode is defined as transportation mode to
work last week covering most of the distance.
102
Table A3. Full Results of Regression Discontinuity Models (Code red day coefficients correspondto Columns (4)-(6) inTable 3-5)
All Inbound Outbound
Morning Mid-day Evening Other Morning Mid-day Evening Other Morning Mid-day Evening Other
Ozone forecast -0.016 -0.009 0.083 -0.036 -0.032 -0.013 0.151 -0.037 0.020 0.030 -0.032 -0.030
(0.052) (0.058) (0.064) (0.027) (0.049) (0.079) (0.091) (0.027) (0.067) (0.033) (0.037) (0.041)
2
nd
order forecast 0.001 0.000 -0.002 0.001 0.001 0.001 -0.004 0.001 -0.000 -0.001 0.001 0.001
(0.001) (0.002) (0.002) (0.001) (0.001) (0.002) (0.002) (0.001) (0.002) (0.001) (0.001) (0.001)
3
rd
order forecast -0.000 -0.000 0.000 -0.000 -0.000 -0.000 0.000 -0.000 0.000 0.000 -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
4
th
order forecast 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 -0.000 -0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
5
th
order forecast -0.000 -0.000 0.000 -0.000 -0.000 -0.000 0.000 -0.000 0.000 0.000 -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag observed ozone -0.000** 0.000 0.001** 0.000 -0.000* 0.000 0.001* 0.000 -0.000 0.000 0.000 0.000
(0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000)
Avg. wind speed 0.000 -0.000 0.000 -0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Min. humidity 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000
(0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000)
Max. humidity 0.001* 0.000 0.001 0.001 0.001 0.000 0.001 0.000 0.001 -0.000 -0.000 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Max. temperature -0.001 -0.003** -
0.005***
-0.001 -0.001* -
0.005***
-0.007** -0.001 0.000 0.001 -0.000 -0.001
(0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.003) (0.001) (0.001) (0.001) (0.001) (0.001)
Min temperature -0.002** -0.001 -0.000 -0.001 -0.002* -0.001 -0.001 -0.001 -0.001 -0.000 0.000 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Precipitation -
0.000***
0.000 0.000 -0.000 -
0.000***
0.000 0.000 -0.000 -
0.000***
-0.000 -0.000** -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag avg. wind spd. 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 0.000 0.000* -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag min. humidity -0.001** -0.000 -0.001 -0.001 -0.001** -0.001 -0.001 -0.000 -0.000 0.000 -0.000 -0.001
(0.000) (0.000) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.001) (0.000) (0.000) (0.000)
Lag max. humidity 0.001* 0.000 0.000 0.001 0.001* 0.001 0.000 0.001 0.001 0.000 0.001 -0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.000) (0.001)
103
Lag max. temp. -0.000 -0.001 -0.005** -0.001 -0.000 -0.001 -0.006** -0.001 0.000 0.001 -0.002 -0.001
(0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.003) (0.001) (0.001) (0.001) (0.002) (0.002)
Lag min. temp. 0.000 0.002** 0.004*** 0.001 0.001 0.003** 0.006*** 0.001* -0.000 -0.001 -0.000 0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Lag precipitation -0.000 -0.000 0.000 0.000 -0.000** 0.000 0.000 -0.000 -0.000 -0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag gas price -0.078 -0.109 0.108 -0.008 -0.081 -0.155 0.115 -0.004 -0.038 -0.074 0.024 -0.063
(0.052) (0.099) (0.180) (0.086) (0.050) (0.124) (0.257) (0.088) (0.087) (0.109) (0.058) (0.101)
2
nd
lag gas price 0.030 0.081 -0.071 0.160 0.059 0.091 -0.017 0.172 -0.040 0.119 -0.051 0.150
(0.066) (0.108) (0.162) (0.118) (0.060) (0.140) (0.230) (0.116) (0.131) (0.074) (0.046) (0.131)
3
rd
lag gas price -0.012 -0.126 -0.202 -0.087 -0.014 -0.151 -0.373 -0.103 0.103 -0.039 -0.081 0.003
(0.100) (0.122) (0.168) (0.122) (0.082) (0.155) (0.235) (0.123) (0.109) (0.091) (0.097) (0.123)
4
th
lag gas price 0.073 0.109 -0.034 -0.005 -0.036 0.206 0.005 -0.011 0.088 -0.046 -0.007 0.001
(0.116) (0.118) (0.150) (0.116) (0.085) (0.160) (0.213) (0.123) (0.092) (0.121) (0.118) (0.151)
5
th
lag gas price -0.173 -0.194* 0.153 -0.082 -0.172 -0.203 0.212 -0.062 -0.108 -0.235* 0.089 -0.181
(0.128) (0.114) (0.150) (0.116) (0.145) (0.128) (0.238) (0.113) (0.125) (0.135) (0.098) (0.171)
6
th
lag gas price 0.052 0.033 0.251 0.111 0.065 -0.095 0.340 0.099 -0.015 0.220 -0.001 0.189
(0.124) (0.138) (0.204) (0.138) (0.144) (0.156) (0.277) (0.140) (0.100) (0.261) (0.087) (0.195)
7
th
lag gas price 0.081 0.190* -0.108 -0.076 0.152* 0.299** -0.152 -0.082 -0.015 0.019 0.041 -0.087
(0.065) (0.097) (0.140) (0.065) (0.078) (0.126) (0.181) (0.087) (0.074) (0.171) (0.085) (0.089)
Year 2002 0.012 0.020** 0.008 0.019 0.011 0.013 -0.001 0.009 0.017* 0.045*** 0.027** 0.036**
(0.009) (0.009) (0.013) (0.014) (0.012) (0.012) (0.018) (0.014) (0.009) (0.011) (0.013) (0.015)
Year 2003 0.052*** 0.055*** 0.016 0.052*** 0.048*** 0.053*** 0.010 0.046*** 0.055*** 0.066*** 0.039*** 0.060***
(0.007) (0.009) (0.012) (0.011) (0.008) (0.012) (0.019) (0.011) (0.008) (0.008) (0.008) (0.013)
J une 0.025** -0.016 -
0.072***
0.029** 0.014 -0.033* -
0.107***
0.031** 0.034** 0.022 0.001 0.030**
(0.011) (0.016) (0.024) (0.012) (0.011) (0.019) (0.039) (0.012) (0.016) (0.021) (0.008) (0.015)
J uly 0.022* 0.027** -0.008 0.042*** 0.001 0.012 -0.014 0.040*** 0.050*** 0.056*** 0.001 0.049***
(0.012) (0.013) (0.020) (0.013) (0.012) (0.017) (0.029) (0.014) (0.018) (0.020) (0.011) (0.015)
August 0.029** 0.037*** 0.015 0.047*** 0.014 0.024 0.016 0.045*** 0.045** 0.054** 0.019 0.052***
(0.014) (0.014) (0.020) (0.016) (0.015) (0.018) (0.029) (0.015) (0.019) (0.023) (0.012) (0.018)
September 0.004 -0.026** 0.009 -0.031* 0.013 -0.037** 0.012 -0.024 -0.019 -0.017 0.010 -0.041*
(0.010) (0.012) (0.019) (0.018) (0.013) (0.016) (0.030) (0.017) (0.016) (0.013) (0.009) (0.021)
Tuesday 0.019*** -0.007 0.003 0.034*** 0.009* -0.004 0.004 0.024*** 0.033*** -0.013 0.012 0.055***
(0.006) (0.008) (0.015) (0.007) (0.005) (0.011) (0.019) (0.007) (0.010) (0.009) (0.009) (0.012)
104
Wednesday 0.030*** 0.002 0.011 0.081*** 0.015** 0.001 0.004 0.071*** 0.055*** 0.007 0.032*** 0.107***
(0.008) (0.009) (0.016) (0.009) (0.006) (0.010) (0.022) (0.010) (0.014) (0.009) (0.007) (0.011)
Thursday 0.049*** 0.020* 0.035** 0.131*** 0.029*** 0.007 0.029 0.109*** 0.083*** 0.055*** 0.056*** 0.179***
(0.008) (0.011) (0.014) (0.009) (0.006) (0.013) (0.019) (0.009) (0.014) (0.009) (0.007) (0.012)
Friday 0.033*** 0.139*** 0.099*** 0.220*** -0.006 0.116*** 0.113*** 0.184*** 0.111*** 0.201*** 0.075*** 0.295***
(0.008) (0.013) (0.016) (0.012) (0.008) (0.013) (0.022) (0.012) (0.014) (0.014) (0.008) (0.014)
Public holiday -
1.202***
-
0.097***
-
0.375***
-
0.119***
-
1.365***
-0.084** -
0.237***
-0.090** -
0.896***
-
0.120***
-
0.652***
-
0.183***
(0.054) (0.034) (0.042) (0.034) (0.043) (0.037) (0.059) (0.038) (0.080) (0.035) (0.017) (0.030)
Detector 2 1.480*** 0.475*** 0.101*** 0.422*** 1.479*** 0.463*** 0.096** 0.424***
(0.019) (0.023) (0.036) (0.013) (0.020) (0.025) (0.036) (0.013)
Detector 3 0.451*** 0.599*** 0.983*** 0.494*** -
0.801***
0.078*** -
0.019***
(0.014) (0.018) (0.030) (0.011) (0.010) (0.006) (0.005)
Detector 4 1.624*** 0.994*** 0.881*** 0.750*** 1.625*** 0.995*** 0.885*** 0.752***
(0.014) (0.018) (0.033) (0.009) (0.014) (0.018) (0.034) (0.010)
Detector 5 1.427*** 0.965*** 0.859*** 0.633*** 1.427*** 0.968*** 0.867*** 0.635***
(0.013) (0.015) (0.031) (0.012) (0.013) (0.015) (0.032) (0.013)
Detector 6 1.250*** 0.841*** 0.931*** 0.514*** 0.213***
(0.014) (0.019) (0.033) (0.011) (0.008)
Detector 7 1.114*** 0.160*** -0.045 -
0.049***
1.114*** 0.158*** -0.045 -
0.047***
(0.018) (0.017) (0.029) (0.012) (0.018) (0.017) (0.029) (0.012)
Code red days -0.051* 0.026 0.030 -0.019 -0.051** 0.045 0.070 -0.010 -0.034 -0.006 -0.026* -0.050*
(0.030) (0.043) (0.059) (0.019) (0.024) (0.057) (0.093) (0.018) (0.046) (0.030) (0.014) (0.027)
Constant 8.050*** 8.487*** 7.485*** 8.907*** 8.347*** 8.643*** 6.551*** 9.008*** 8.620*** 8.425*** 9.807*** 9.205***
(0.702) (0.804) (0.891) (0.427) (0.675) (1.091) (1.242) (0.418) (0.903) (0.455) (0.553) (0.630)
Observations 1520 1119 1157 1201 1059 795 787 839 461 324 370 362
R-squared 0.96 0.93 0.90 0.95 0.97 0.94 0.86 0.96 0.96 0.87 0.88 0.83
Standard errors in the parenthesis account for within-week clustering.* indicates significance at 10%; ** significant at 5%; *** significant at 1%
105
Bibliography
[1] Angrist, J oshua D., and Alan B. Krueger, 1992, “The Effect of Age at School
Entry on Educational Attainment: An Application of Instrumental Variables
with Moments from Two Samples”, Journal of the American Statistical
Association 87, 328-336.
[2] Bloom, H. S., et al., 2004, “Baltimore Region Travel Demand Model for Base
Year 2000”, Task Report 04-01, Baltimore Metropolitan Council.
[3] Bell, M. L., et al., 2004, “Ozone and Short-term Mortality in 95 US Urban
Communities, 1987-2000”, Journal of the American Medical Association,
292, 2372-2378.
[4] Brueckner, J an K., 2001, “Urban Sprawl: Lessons from Urban Economics”, in
William G. Gale, and J anet Rothenberg Pack, eds.: Brookings-Wharton
Papers on Urban Affairs, Brookings Institution Press.
[5] Crane, Randall, 1996, “The Influence of Uncertain J ob Location on Urban
Form and the J ourney to Work”, Journal of Urban Economics 39, 342-356.
[6] Cutter, W. Bowman and Matthew Neidell, 2007, “Voluntary Information
Programs and Environmental Regulation: Evidence from 'Spare the Air'”,
Working paper.
[7] Cummings, R. G. and M. B. Walker, 2000, “Measuring the Effectiveness of
Voluntary Emission Reduction Programmes”, Applied Economics, 32, 1719-
1726.
[8] Dee, Thomas S., and William N. Evans, 2003, “Teen Drinking and
Educational Attainment: Evidence from Two-Sample Instrumental Variables
Estimates”, Journal of Labor Economics 21, 178-209.
[9] Drucker, J oshua, and Asad J . Khattak, 2000, “Propensity to Work from Home:
Modeling Results from the 1995 Nationwide Personal Transportation
Survey”, Transportation Research Record 1706, 108-117.
[10] Ellen, Ingrid Gould, and Katherine Hempstead, 2002, “Telecommuting and
the Demand for Urban Living: A Preliminary Look at White-Collar Workers”,
Urban Studies 39, 749-66.
[11] FCC, 2002, “Third Report on the Availability of High Speed and Advanced
Telecommunications Capability”, Federal Communications Commission,
Washington D.C.
106
[12] Fox, J . C. and M. Sarkar, 2002, “An Analysis of Air Pollution-Related
Knowledge, Attitudes, and Behaviors across Time: the End of Season and
Episodic Surveys”, MWCOG.
[13] Glaeser, Edward L., and Matthew E. Kahn, 2001, “Decentralized Employment
and the Transformation of the American City”, Brookings-Wharton Papers on
Urban Affairs 1-47.
[14] Greene, W. H., 2003, Econometric Analysis, 5
th
ed. Prentice Hall, New J ersey.
[15] Grubesic, Tony H., and Alan T. Murray, 2004, “Waiting for Broadband: Local
Competition and the Spatial Distribution of Advanced Telecommunication
Services in the United States”, Growth and Change 35, 139-165.
[16] Grubesic, Tony H., and Morton E. O'Kelly, 2002, “Using Points of Presence
to Measure Accessibility to the Commercial Internet”, Professional
Geographer 54, 259-278.
[17] Guimaraes, T., and P. Dallow, 1999, “Empirically testing the benefits,
problems, and success factors for telecommuting programmes”, European
Journal of Information Systems 8, 40-54.
[18] Hahn, J ., P. Todd and W. van der Klaauw, 2001, “Identification and
Estimation of Treatment Effects with a Regression-Discontinuity Design,”
Econometrica, 69(1), 201-09.
[19] Hausman, J erry A., J . Gregory Sidak, and Hal J . Singer, 2001, “Cable
Modems and DSL: Broadband Internet Access for Residential Customers”,
American Economic Review 91, 302-307.
[20] Heckman, J ., R. LaLonde and J . Smith, 1999, “The Economics and
Econometrics of Active Labor Market Programs,” in Orley Ashenfelter and
David Card, eds., Handbook of Labor Economics, Volume 3A, Section 7.4.6.
Amsterdam: North-Holland.
[21] Henderson, Dennis K., and Patricia L. Mokhtarian, 1996, “Impacts of Center-
Based Telecommuting on Travel and Emissions: Analysis of the Puget Sound
Demonstration Project”, Transportation Research: Part D: Transport and
Environment 1, 29-45.
[22] Henry, G. T. and C. S. Gordon, 2003, “Driving Less for Better Air: Impacts of
a Public Information Campaign,” Journal of Policy Analysis and
Management, 22(1), 45-63.
[23] Hensher, D. J ., W. H. Greene, and J . M. Rose, 2005, Applied Choice Analysis:
a Primer, Cambridge University Press.
107
[24] Imbens, Guido W., and J oshua D. Angrist, 1994, “Identification and
Estimation of Local Average Treatment Effects”, Econometrica 62, 467-475.
[25] Imbens Guido W. and Thomas Lemieux, 2008, “Regression Discontinuity
Designs: A Guide to Practice”, Journal of Econometrics, 142(2), 615-635.
[26] Inoue, Atsushi, and Gary Solon, 2005, “Two-Sample Instrumental Variables
Estimators”, NBER Working Paper t0311.
[27] Kain, J ohn F., 1968, “Housing Segregation, Negro Employment, and
Metropolitan Decentralization”, The Quarterly Journal of Economics 82, 175-
197.
[28] Kitamura, Ryuichi, Patricia Mokhtarian, Ram Pendyala, and Konstadinos
Goulias, 1991, “An Evaluation of Telecommuting as a Trip Reduction
Measure”, University of California Transportation Center, CA: University of
California.
[29] Koening, Brett, Dennis Henderson, and Patricia Mokhtarian, 1996, “The
Travel and Emissions Impacts of Telecommuting for the State of California
Telecommuting Pilot Project”, Transportation Research C 4, 13-32.
[30] Krupnick, A. J ., 1988, “Temporal and Spatial Control of Ambient Ozone”,
Discussion Paper QE88-05, Resources for the Future, Washington DC.
[31] Kuzmyak, J . R. and Street Smarts, 2003, “Effectiveness of Free Bus Fares on
Forecast Air Quality Code Red Days”, Interim Report for Northern Virginia
Transportation Commission.
[32] Lee, D. S., E. Moretti, M. Butler, 2004, “Do Voters Affect or Elect Policies?
Evidence from the U.S. House Election”, Quarterly Journal of Economics,
119, 807-859.
[33] Lomax, Tim, and David Schrank, 2005, “2005 Annual Urban Mobility
Report”, Texas Transportation Institute, College Station, Texas.
[34] Lund, J . R., and P. L. Mokhtarian, 1994, “Telecommuting and Residential
Location: Theory and Implications for Commute Travel in Monocentric
Metropolis”, Transportation Research Record 1463, 10-14.
[35] Malecki, Edward J ., 2002, “The Economic Geography of the Internet's
Infrastructure”, Economic Geography 78, 399-424.
[36] Meyer, Peter B., and Anastasiya M. Osborne, 2005, “Proposed Category
System for 1960-2000 Census Occupations”, BLS Working Papers 383.
108
[37] Mokhtarian, P. L., and I. Salomon, 1996, “Modeling the Choice of
Telecommuting: 3. Identifying the Choice Set and Estimating Binary Choice
Models for Technology-Based Alternatives”, Environment and Planning A 28,
1877-94.
[38] Mokhtarian, Patricia L., Gustavo O. Collantes, and Carsten Gertz, 2004,
“Telecommuting, Residential Location, and Commute-Distance Traveled:
Evidence from State of California Employees”, Environment and Planning A
36, 1877-97.
[39] Mokhtarian, Patricia L., Ilan Salomon, and Sangho Choo, 2005, “Measuring
the Measurable: Why can't we Agree on the Number of Telecommuters in the
U.S.?”, Quality and Quantity 39, 423-452.
[40] MWCOG, 2005, Recommended Air Quality Forecast Procedures Baltimore-
Washington Region.
[41] MWCOG, 2003, “2003 Air Pollution Episodic Surveys: Report of Findings
for the Washington”, DC and Baltimore Metropolitan Areas.
[42] Parry, Ian W. H., Margaret Walls, and Winston Harrington, 2007,
“Automobile Externalities and Policies”, Journal of Economic Literature
XLV, 373-399.
[43] Plaut, Pnina O., 2006, “The Intra-Household Choices Regarding Commuting
and Housing”, Transportation Research: Part A: Policy and Practice 40, 561-
571.
[44] Popuri, Y., and C. R. Bhat, 2003, “On Modeling Choice and Frequency of
Home-Based Telecommuting”, Transportation Research Record 1858, 55-60.
[45] Pratt, J oanne H., 2000, “Asking the Right Questions about Telecommuting:
Avoiding Pitfalls in Surveying Home-based Work”, Transportation 27, 99-
116.
[46] Prystash, D. T., 1995, “Corporate attitudes and commitments to
telecommuting”, Professional Communication, IEEE Transactions on 38, 95-
99.
[47] Small, Kenneth A., 1992a, "Trip Scheduling in Urban Transportation
Analysis", American Economic Review Papers and Proceedings, 82(2), 482-
486.
[48] Small, Kenneth A., 1992b, Urban Transportation Economics, Harwood,
Philadelphia.
109
[49] Teller, A., 1967, “Air Pollution Abatement: Economic Rationality and
Reality,” Daedalus, 96, 1082-98.
[50] U.S. Environmental Protection Agency (EPA), 1993, “Automobiles and
Ozone”, 400/F-92-006 (J anuary).
[51] _______, 1997, “Episodic Emission Control Programs”, 420/F-97-022
(December).
[52] _______, 1999, “Smog: Who Does It Hurt?: What You Need To Know About
Ozone and Your Health”, EPA-452/K-99-001 (J uly).
[53] _______, 2003, “Ozone: Good Up High Bad Nearby”, 451/K-03-001 (J une).
[54] _______, 2004, “The Ozone Report: Measuring Progress through 2003”,
454/K-04-001 (April).
[55] van der Klaauw, W., 2002, “Estimating the Effect of Financial Aid Offers on
College Enrollment: A Regression-Discontinuity Approach,” International
Economic Review, 43(4), 1249-87.
[56] van Ommeren, J os, Piet Rietveld, and Peter Nijkamp, 1997, “Commuting: In
Search of J obs and Residences”, Journal of Urban Economics 42, 402-21.
[57] Walls, Margaret A., Elena A. Safirova, and Yi J iang, 2007, “What Drives
Telecommuting? The Relative Impact of Worker Demographics, Employer
Characteristics, and J ob Types”, Transportation Research Record, No. 2010,
111-120.
[58] Walls, Margaret, and Elena Safirova, 2004, “A Review of the Literature on
Telecommuting and Its Implications for Vehicle Travel and Emissions”,
Discussion Papers 04-44, Resources for the Future.
[59] Welch E., X. Gu, and L. Kramer, 2005, “The Effects of Ozone Action Day
Public advisories on Train Ridership in Chicago”, Transportation Research
Park D, 10, 445-458.
[60] White, Michelle J ., 1977, “A Model of Residential Location Choice and
Commuting by Men and Women Workers”, Journal of Regional Science 17,
41-52.
[61] White, Michelle J ., 1986, “Sex Differences in Urban Commuting Patterns”,
American Economic Review 76, 368-72.
[62] White, Michelle J ., 1988, “Location Choice and Commuting Behavior in
110
Cities with Decentralized Employment”, Journal of Urban Economics 24,
129-52.
[63] White, Michelle J ., 1988, “Urban Commuting J ourneys Are Not 'Wasteful'”,
Journal of Political Economy 96, 1097-110.
[64] Wooldridge, J . M., 2001, Econometric Analysis of Cross Section and Panel
Data, MIT press, Cambridge, MA.
doc_751225720.pdf
Urban economics is broadly the economic study of urban areas; as such, it involves using the tools of economics to analyze urban issues such as crime, education, public transit, housing, and local government finance.
ABSTRACT
Title of Dissertation: TWO EMPIRICAL ESSAYS IN
ENVIRONMENTAL AND URBAN
ECONOMICS
Yi J iang
Doctor of Philosophy, 2008
Directed By: ProfessorsMaureen Cropper and William Evans
Externalities associated with automobile use have long been an important
topic in environmental and urban economics. Air pollution and traffic congestion
constitute two main external costs of driving (Parry, Walls and Harrington 2007).
Because pricing approaches such as higher fuel taxes and road pricing are unpopular,
varioustravel demand management (TDM) programs aiming to control vehicle travel
demand through non-pricing approaches have been adopted by government agencies
across the country. These programs provide public information, use persuasion,
subsidize transit riding, and promote carpooling and telecommuting. However,
whether these programs generate incentives for people to reduce driving remains an
open question.
I address this question with respect to two types of TDM strategies:
telecommuting and public information provision. The first essay examines whether
telecommutingopportunities lead employees to have longer commute lengths.
Because telecommuting is often jointly chosen with commuting patterns and no
single dataset contains sufficient information to solve the endogeneity problem, I use
a two-sample instrumental variables technique to estimate the causal impact of
telecommuting on commute length. The data for the project are assembled from the
May 2001 Current Population Survey (CPS) and the 2000 Census 5% Public Use
Micro-data Series (PUMS). The results suggest that telecommuting increases married
female workers’ one-way commute time by 9 – 12 minutes, but the effect on male
workers’ commute lengthis not precisely estimated. Although telecommuting may
still cut down total commute miles, it is less effective than expected, in particular for
married women.
The second essay assesses the effectiveness of the Air Quality Action Days
program in the Baltimore metropolitan area in getting cars off the road on high ozone
days. The program asks people to reduce vehicle trips on code red days when the
ozone level is forecast to exceed the EPA’s standard. I look at traffic volumes on
highways in the Baltimore area, and using aregression discontinuity design, measure
the extent that traffic is lower due to the announcement. I find that the program
generally has little effect except that it reduces morning inbound traffic by 4-5
percent. Evening outbound traffic declines correspondingly.
TWO EMPIRICAL ESSAYS IN ENVIRONMENTAL AND URBAN ECONOMICS
By
Yi J iang
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park, in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2008
Advisory Committee:
Dr. Maureen Cropper, Chair
Dr. William Evans, Co-chair
Dr. Mahlon Straszheim
Dr. J udith Hellerstein
Dr. Anna Alberini
©Copyright by
Yi J iang
2008
ii
Dedication
This work is dedicated to my parents, who have always been the source of my
aspiration, energy and strength, and to my wife Ruoyun (Rebecca), for her constant
support, encouragement and love.
iii
Acknowledgements
I am pleased that I chose environmental economics as my major field and am
extremely fortunate to have Maureen Cropper as my advisor. She has spent
uncounted hours instructing and mentoring me. Beyond this dissertation, Maureen’s
help and support has been complete and unconditional: from obtaining external
funding to recommending me for a Resources for the Futureinternship; from eliciting
information from knowledgeable researchers for my projects to supporting my
participation in academic workshops and conferences. I think the best way to return
her great support is to be a sound environmental economist, which I will strive to be.
I am grateful to Bill Evans for his guidance and employment advice. His
ability to creatively combine empirical methods and data to answer a significant
question will always be an inspiration to me. I also benefited tremendously from his
Applied Microeconomics course, which is one of the best prepared and most
enjoyable classes I have ever taken.
I thank Mahlon Straszheim and J udy Hellerstein for their constructive insights
on the papers herein as well as their service on my dissertation committee. I thank
Anna Alberini for serving on the committee and for collaborating with Maureen and
me on a project on episodic ozone control. I am also indebted to Dave Evans, J onah
Gelbach, Christopher McKelvey, Peter Nelson, J eff Smith, Elena Safirova, Margaret
Walls, and to several seminar participants for helpful suggestions and comments.
Finally, special thanks go to J ennifer Desimone at the Metropolitan
Washington Council of Government, Charles Piety at the Department of Atmospheric
and Oceanic Science, University of Maryland, and Michael Pack and the staff of the
iv
Center for Advanced Transportation Technology, University of Maryland for
providing data and background information for the second essay.
v
Table of Contents
Dedication.....................................................................................................................ii
Acknowledgements......................................................................................................iii
List of Tables............................................................................................................... vi
List of Figures.............................................................................................................vii
1 Introduction........................................................................................................... 1
1.1 Objective of the dissertation......................................................................... 1
1.2 Contribution to the Literature....................................................................... 4
1.3 Plan of the Dissertation................................................................................. 6
2 The Impact of Telecommuting on the J ourney to Work: A Two-Sample
Instrumental Variables Approach................................................................................. 7
2.1 Introduction................................................................................................... 7
2.2 Urban Problems and Telecommuting......................................................... 11
2.3 Theory and Empirical Liteature.................................................................. 14
2.4 The NHTS and an Empirical Baseline........................................................ 16
2.5 Empirical Strategy...................................................................................... 19
2.5.1 Instrumental Variable.............................................................................. 19
2.5.2 The Two-Sample Instrumental Variables (TSIV) Method..................... 22
2.5.3 CPS and PUMS Samples........................................................................ 26
2.6 The First-Stage Estimates........................................................................... 29
2.7 Reduced-Form and TSIV Estimates........................................................... 33
2.7.1 Reduced-Form Estimates........................................................................ 33
2.7.2 TSIV Estimates of the Effects on Commute Length.............................. 35
2.7.3 Effects of Telecommuting on Commute Mode...................................... 36
2.7.4 Sensitivity Analysis................................................................................ 37
2.8 Discussion................................................................................................... 39
2.9 Conclusion.................................................................................................. 41
Figures and Tables for Chapter 2............................................................................ 43
3 Do People Drive Less on Code Red Days?........................................................ 59
3.1 Introduction................................................................................................. 59
3.2 AQAD Program in Baltimore Area............................................................ 62
3.3 Theory......................................................................................................... 65
3.4 Empirical Strategy...................................................................................... 67
3.5 Data............................................................................................................. 71
3.6 Results......................................................................................................... 76
3.7 Discussion and Conclusion......................................................................... 80
Figures and Tables for Chapter 3............................................................................ 82
4 Concluding Comments........................................................................................ 92
4.1 Summary of Results.................................................................................... 92
4.2 Directions for Future Research................................................................... 93
Appendices.................................................................................................................. 95
Appendix 1 A Monocentric City Model with Commuters and Telecommuters..... 95
Appendix 2 Imputation of Top-Coded Commuting Time in the PUMS................ 98
Bibliography............................................................................................................. 105
vi
List of Tables
Table 2-1. Descriptive Statistics of NHTS Sample.................................................... 45
Table 2-2. OLS Estimates of the "Effect" of Telecommuting on Commute Lengths
and Travel Mode, 2001 NHTS.................................................................................... 46
Table 2-3. Summary Statistics of CPS Sample by Gender and Telecommuting Status
..................................................................................................................................... 47
Table 2-4. First-Stage Estimates of Telecommuting Models, May 2001 CPS........... 49
Table 2-5. Reduced-Form Estimates of Commute Time Model, 2000 PUMS........... 52
Table 2-6. TSIV Estimates of the Effects of Telecommuting on Commute Time and
Mode........................................................................................................................... 55
Table 2-7. Robustness Check of the TSIV Estimates................................................. 56
Table 2-8. Projection of Commute Time (Minute) onto Commute Distance (Mile).. 58
Table 3-1. One-hour Ozone Level and Code Colors in Baltimore Area.................... 85
Table 3-2. Description of Detectors and Traffic in Baltimore Area........................... 86
Table 3-3. Correlation between Missing Time Block and Code Red Day................. 87
Table 3-4. Summary Statistics and Difference in Selected Covariates Between Code
Red Days and Other Days........................................................................................... 88
Table 3-5. Impact of Code Red Day Announcement on Traffic Volumes by Time of
Day.............................................................................................................................. 89
Table 3-6. Robustness Check...................................................................................... 90
Table 3-7. Impact of CRDs in Sequence on Traffic Volumes by Time of Day.......... 91
Table A1. CPS and NHTS Sample Construction...................................................... 100
Table A2. Distributions by Commute Mode across Cities....................................... 101
Table A3. Full Results of Regression Discontinuity Models (Code red day
coefficients correspond to Columns (4)-(6) in Table 3-5)........................................ 102
vii
List of Figures
Figure 2-1. Distributions of Telecommuters and Commuters by Commute Time and
Distance....................................................................................................................... 43
Figure 2-2. Internet Penetration by 2-Digit Occupation and MSA size (2001 CPS).. 44
Figure 3-1. Illustration of Biased Estimates with Linear Forecast Ozone Level........ 82
Figure 3-2. Map of Baltimore Region Major Freeways and Maryland SHA’s Traffic
Detectors..................................................................................................................... 83
Figure 3-3. Similarity of Covariates around Code Red Day Cutoff Point.................. 84
Figure A1. Bid Rent Curves in a Monocentric City with Telecommuters and
Commuters.................................................................................................................. 99
1
1 Introduction
Automobile use generates significant negative externalities that can be quite
costly to the society. Some economists estimate that external costs amount to 5 cents per
mile in the form of congestion, 2 cents per mile in air pollution and 3 cents per mile in
accidents (Parry, Walls and Harrington 2007). The externalities associated with driving
are not an easy problem to solve. For instance, road capacity expansion has not proved
effective in mitigating congestion due to latent demand (e.g. Small 1992a), and it
exacerbates the air pollution problem by encouraging more driving. Economic theory
suggests that excess driving can be reduced by adding the extra social cost caused by
vehicle driving to a driver's personal cost calculation. This can be accomplished with
instruments like higher fuel taxes and/or a congestion toll. However, in the US where car
ownership is prevalent and demand for vehicle travel is price inelastic, it is not surprising
that these pricing approaches havelittle political support. Meanwhile, government
agencies such as urban planning boards have been interested in controlling the demand
for vehicle travel directly. They have developed various travel demand management
(TDM) programs that provide public information, use persuasion, subsidize transit riding,
and promote carpooling and telecommuting. However, whether these programs generate
incentives for people to reduce driving remains an open question.
1.1 Objective of the dissertation
This dissertation evaluates two distinct TDM strategies: telecommuting and a
public information program. The focus of Chapter 2 is the impact of telecommuting on
2
total commute miles. Telecommuting refers to working from home instead of traveling to
work at least two days every month. In addition to being a favorable arrangement
between employers and employees, telecommuting is increasingly used as a TDM
strategy in order to mitigate urban congestion and air pollution attributable to commute
trips. However, to what extent telecommuting reduces vehicle miles traveled (VMTs)
depends on how it affects commute VMTs and non-commute VMTs. The impact of
telecommuting on commute VMTs depends on the effect of telecommuting on one-way
commuting length as well as telecommuting frequency. The naïve conclusion that
telecommuting reduces total commute miles in the same proportion as it reduces
commuting frequency could be wrong because telecommuting may actually increase one-
way commuting length.
The goal of Chapter 2 is to empirically measure the impacts of telecommuting on
one-way commute length and the probability of driving to work. One of the difficulties is
that telecommuting is not a random choice. Individuals who telecommute could be
systematically different from non-telecommuters in unobserved ways that are also
correlated with commuting behavior. I employ instrumental variable methodsto obtain
consistent estimates, and develop an instrumental variable that captures variation in
telecommuting opportunity across occupation and city size. Another difficulty is that
there exists no single data set that contains telecommuting, commuting and the
instrumental variable. The problem is tackled by using the two-sample instrumental
variable technique, initially developed in Angrist and Krueger (1992). The two samples
3
come from the May 2001 CPS and 2000 PUMS, both of which are nationally
representative.
Chapter 3 looks at a specific public information program---the Air Quality Action
Days program in the Baltimore metropolitan area---which features a code red day alert
when ground-level ozone is forecast to exceed the EPA’s standard. The program not only
warns the public of high ozone levels but also tries to lower ozone concentrations on
those days bypersuading people not to drive. The program can be viewed as a TDM
initiative that aims to influence vehicle travel episodically for environmental purposes. It
is unclear, however, whether an individual would forego driving on code red dayseven if
he/she internalizes the environmental cost resulting from his/her driving. Driving might
still be the optimal mode choice because a person has a smaller risk of being exposed to
bad air when driving than when walking to transit. This chapter conducts an evaluation of
effectiveness of the program in reducing on-road vehicles. A regression discontinuity
design is employed to overcome potential omitted variable bias, since the code red day is
a discontinuous function of an observed continuous variable, forecast ozone level.
In sum, the dissertation studies two popular TDM strategies from an economic
perspective. Potential behavioral responses are taken into account and state-of-the-art
econometric techniques are used in the analysis to carefully examine the effectiveness of
the strategies in lowering vehicle travel. Both positive and negative findings should be
useful to economists. The former may lead us to re-think TDM programs and consider
4
combining pricing approaches with TDM. A negative finding could strengthen the
argument for a complete pricing strategy.
1.2 Contribution to the Literature
Most empirical studies about telecommuting are designed to identifywhat factors
explain the choice of telecommuting. Relatively few studies examine the impacts of
telecommuting on commute length or total VMTs. The results from previous studies are
mixed, with earlier ones suggesting a reduction in total VMTs and later ones showing
positive effects on one-way commute distance and total commute miles. But most studies
are subject to two critical shortcomings. First, telecommuting is assumed to be exogenous
in explaining commute distance or miles. Second, the datasets used in the analyses are
usually small and not representative.
The main contribution of Chapter 2 is that I tackle the endogeneity issue with an
instrumental variable procedure and use large, nationally representative samples
assembled from the 2001 CPS and 2000 PUMS. The instrumental variable measures the
internet penetration for working from home across different occupation by city size cells,
which should capture inexogenous telecommuting opportunity for individuals. Because
of the data constraints mentioned in the previous section, I apply the two-sample
instrumental variable technique of Angrist and Krueger (1992) to information on
telecommuting from the May 2001 CPS and information on commuting from the 2000
PUMS. As both datasets are national samples, the results have implications for different
regions and working groups.
5
The main findings are that telecommuting leadsmarried female workers’ one-way
commute time to increase by 9 – 12 minutes and has a smaller, positive, but statistically
insignificant effect on men’s commute time. Exploring heterogeneous impacts between
married women, single women and men is an innovation to the literature. It is plausible to
expect that married women are more responsive to lower commuting costs as they are
more often the secondary earner in a two-earner household and are likely to be more
constrained in workplace locations. The results confirm these expectations.
Another contribution is to apply the same method to the probability of driving to
work. The OLS estimates using the 2001 National Household Travel Survey (NHTS)
data show that the propensity of driving to work declines among telecommuters. This
result implies that telecommuting provides an extra bonus by changing commute modes
in favor of transit. However, the two-sample instrumental variables estimates indicate
that telecommuting has a positive but statistically insignificant effect on commute mode
choice. They suggest that the OLS estimates could be misleading without correcting the
endogeneity problem. Overall, the chapter suggests that telecommuting is less effective in
lowering total commute miles than people think it is, in particular for married women.
Chapter 3 contributes to the literature by examining a public information program
in Baltimore, andexploiting the program’s institutional features to identify the traffic
reduction attributable to the program. Similar advisory programs have been implemented
in other urban areas that have ozone problem. Earlier studies either rely on survey
respondents’ stated information or are unable to detect a change in traffic caused by the
6
program. Only the recent study by Cutter and Neidell (2007) uses the same econometric
technique and reaches similar conclusions as my study does, but instead looks at the San
Francisco Bay area. I find that the code red day announcements result in morning traffic
reductions by 3 – 5%. This reduction occurs only for inbound traffic in the morning and
outbound traffic in the evening.
1.3 Plan of the Dissertation
Chapters 2 and 3 are both self-contained essays. Chapter 2 examines the impact of
telecommuting on one-way commuting length and on the probability of driving to work.
Chapter 3 estimates the effect of code red day announcements on traffic volumes in the
Baltimore Metropolitan area. Each chapter starts with an introduction of the research
question, methodology and main findings. More detailed research or institutional
background is provided next, and followed by a simple theoretical model to convey the
key intuitions. Empirical methods as well as the data used are described. The main results
and sensitivity checks are presented in subsequent sections. Each essay ends with a
further discussion of the findings and conclusions. Chapter 4 summarizes the main
findings in both essays and draws some common conclusions. It also discusses the
questions stemming from the study that deserve future research.
7
2 The Impact of Telecommuting on the J ourney to Work: A Two-
Sample Instrumental Variables Approach
2.1 Introduction
Telecommuting reduces both the monetary and psychological costs of commuting.
Employers, by allowing workers to telecommute, can recruit and retain valued employees
and possibly reduce the costs of office space and administrative support. More
importantly, telecommuting is increasingly suggested as a solution to traffic congestion
and air pollution in urban areas. For instance, the Connecticut Department of
Transportation established a statewide initiative "Telecommute Connecticut!" to help
employers within the state set up and run telecommuting programs.
1
In May 2006, the
U.S. Department of Transportation announced its new National Strategy to Reduce
Congestion on America's Transportation Network, which highlights "Four Ts" – tolling,
transit, telecommuting and technology – as an approach to reducing traffic congestion.
From the perspective of reducing congestion and pollution caused by vehicle miles
traveled, a key policy question is what impact telecommuting has on total commute miles
traveled.
2
At first blush it would appear that greater telecommuting should decrease
1
See their website http://www.telecommutect.com for more information about "Telecommute
Connecticut!".
2
The impact of telecommuting on non-commute VMTs is another important topic but is beyond the scope
of the dissertation. In theory, telecommuting could affect non-commute VMTs in multiple ways. For
example, flexible working schedules allow telecommuters to go shopping or run errands more often. If the
individual changes home location in response to telecommuting, her/his demand for non-commute travel is
likely to change as well. Walls and Safirova (2004) review a series of telecommuting papers and find no
8
commute miles. However, since telecommuting decreases the cost of commuting, it is
plausible that telecommuting actually induces workers to work farther from home. For
example, a woman who works at home one day a week reduces her commuting costs by
20% compared to a non-telecommuter. The decline in commuting costs provides an
incentive for the woman to live farther away from her workplace or work farther from
home.
Telecommuting may not achieve its policy objectives if it leads to a longer journey
to work. However, it is not easy to obtain a consistent estimate of the causal impact of
telecommuting on commute length. For research purposes, the ideal situation would be to
randomly assign the opportunity to telecommute to a panel of workers and then examine
how often they telecommute and the length of their commutes before and after the
intervention. However, this type of experiments have never been performed. Because
commute length and the decision to telecommute are jointly determined, estimates of the
impact of telecommuting on travel time may be biased. Yet, the direction of the bias is
unclear. On one hand, workers who have longer commute distances may be more likely
to telecommute. At the same time, people who have a distaste for commuting would, all
else equal, live closer to work as well as welcome a telecommuting opportunity.
In this paper, I examine the impact of telecommuting on total commute miles
traveled while controlling for the endogeneity of the telecommuting decision. Because
study show evidence of significant increase in non-commute travel for telecommuters. However, a common
shortcoming of those studies is that they are based on small samples of workers.
9
information on whether an individual telecommutes and the length of his commute are
not contained in one data set, I utilize the two-sample instrumental variables (TSIV)
technique developed by Angrist and Krueger (1992). The key data sets include the work
schedule supplement to the May 2001 Current Population Survey (CPS) that contains
telecommuting data, and the 5-percent Public Use Micro-data Series (PUMS) of the 2000
Census that contains information about one-way commute time and mode. An instrument
is developed from the CPS sample that measures internet utilization for working at home
for each 2-digit occupation and MSA-size combination. The instrument exploits the fact
that certain occupations and MSA combinations are more open to telecommunication
technology than others. These differences are by and large determined by job
characteristics and internet infrastructure distribution, which, once I control for MSA and
occupation fixed-effects, should be orthogonal to individuals' commutes. I also examine
the effect of telecommuting on travel mode choice using the same method.
It is well documented in the literature that men and women exhibit distinct
commuting patterns (White 1977, 1986), especially with respect to marital status and
family composition. I conjecture that telecommuting might have differential effects on
married women and single women for the following reasons. First, in a dual-earner
household, the woman is more often the secondary earner rather than the primary. She is
more likely than her husband to have a part-time or lower-paying job. Therefore,
commuting costs, which will be reduced by telecommuting, may be more important to
her workplace location than to her husband’s. Second, the husband’s job situation is
10
likely to dominate the residential location of the household and affect the workplace
location of the wife. Married women restrict the geographic ranges of their job search and
often work closer to home than their husbands. People who are in occupations where
telecommuting is an option will consider a larger range of workplace location than those
who are not. Since married women may be more constrained in their job search than
single women, telecommuting may have a larger impact on married women choosing
workplace locations than on single women. Therefore, I estimate each model for men,
married women and single women separately to explore the heterogeneity in the response
across these demographic groups.
3
TSIV estimates demonstrate that telecommuting has a large positive effect on
commute length for married female workers: Married women tend to work farther from
home when they can substitute working at home for commuting. Being able to
telecommute causes married women to increase their one-way commute an additional 9-
12 minutes. This finding is consistent with the fact that married female workers have
short commuteswhen telecommuting is not an option. The effect for male workers is
smaller and statistically insignificant. For an average married women who works from
home two out of five days a week, telecommuting reducestotal commute miles, but not
by 40 percent. My analysis also suggests that telecommuting is unlikely to affect the
probability of a worker driving to work.
3
It could be argued that women with children are more constrained in their choice of workplace location
than women without children, regardless of marital status. When, however, the sample is split between
women with and without children, the instrumental variable does not have enough explanatory power.
11
The rest of the paper is organized as follows: Section 2.2 defines telecommuting and
provides background information about telecommuting and relevant studies. Section 2.3
presents baseline estimates of the "effect" of telecommuting on journey-to-work from
OLS analysis of the 2001 Nationwide Household Transportation Survey (NHTS). Section
2.4 describes the identification strategy and the data. TSIV estimation results appear in
Sections 2.5 and 2.6, with discussion and conclusions following in Sections 2.7 and 2.8.
2.2 Urban Problems and Telecommuting
Traffic congestion is a problem for many urban areas in the US and around the
world. The social costs of having millions of cars stuck in traffic are high. The Texas
Transportation Institute estimates that, in 2003, congestion in the 85 largest urban areas
in the US caused 3.7 billion vehicle-hours of delay, resulting in a cost of $63 billion.
According to Lomax and Schrank (2005), each rush hour traveler pays an annual
congestion tax of $800 to $1,600 in lost time and fuel in the 10 most congested areas of
the US. The costs of congestion extend to the environment as well. Automobile emissions
are an important source of ozone precursors—nitrogen oxides (NOx) and volatile organic
compounds (VOCs). In 2003 more than 100 million people lived in counties that violated
the federal ozone standard (EPA, 2004). This is a serious public health problem since it is
well established that ozone can induce respiratory symptoms, and cause decrements in
lung function and inflammation of the airways (EPA, 2003).
While pricing instruments such as congestion tolls and gasoline taxes are a way to
internalize the external costs of driving, they are unpopular in the US. More attention has
12
therefore been devoted to non-pricing strategies that control the demand for automobile
travel directly. A subset of these strategies, Commute Trip Reduction (CTR) programs,
focuses on commute trips, the largest contributor to rush hour traffic and one of the main
contributors to the total vehicle miles traveled (VMT). These programs, often
implemented through cooperation agreements between government authorities,
employers and individuals, provide persuasion (e.g., Earth Day fairs), incentives (e.g.,
transit subsidies) and/or facilitate carpooling. Telecommuting is one of the most popular
components of these programs (Pollution Probe 2001).
The literature has not settled on a consistent definition of telecommuting. Some
studies include as telecommuters people who take work home and never substitute
working from home for commuting on a work day. I refer to these people as teleworkers.
Some research includes the self-employed who work at home sometimes as
telecommuters. As a result, counts of telecommuters vary dramatically across studies.
Mokhtarian et al. (2005) reviewed a number of papers using various data sets and
concludes that the percentage of telecommuters in the late 1990s ranged from 3% to 20%.
The latter figure includes the home-based self-employed and all teleworkers.
In this paper, I define a telecommuter as an employee who works at home instead of
traveling to a workplace at least one day every two weeks. People who commute every
day even though they sometimes work from home, as well as those who telecommute
infrequently are not counted as telecommuters in my definition. My definition also
excludes the self-employed since they are not the target population of TDM policy.
13
Finally, telecommuting does not require that the individual use information and
communication technology (ICT) when working at home, although technology (ICT)
plays a significant role in enhancing telecommuting opportunities.
The May 2001 CPS supplemental survey collected information about work schedules
and working at home from 51,000 working adults from approximately 47,000
households. The final CPS sample in this analysis consists of 29,147workers who lived
in an MSA and were not self-employed in their main jobs.
4
Among them 1,138 were
telecommuters, accounting for 4 percent of the sample. This figure falls at the low end of
the range identified in Mokhtarian et al. (2005).
Many studies of telecommuting have examined who telecommutes or why people
telecommute.
5
For instance, Drucker and Khattak (2000) found in the 1995 Nationwide
Personal Transportation Survey sample that ceteris paribus, males, older people, those
with more education, those with higher incomes, parents of young children, those in rural
areas and those with inferior access to transit are more likely to telecommute. They also
found that one-way commute distance negatively impacts the propensity to telecommute.
Popuri and Bhat (2003) and Walls et al. (2007) analyzed large data sets from New York
and Southern California, respectively. They confirmed the role of the aforementioned
demographic characteristics in determining telecommuting status. In addition, they found
4
Table A1 provides information on sample construction for both the May 2001 CPS sample and the
2001 NHTS sample.
5
The literature contains various definitions of telecommuting. For a comprehensive review, see Walls and
Safirova (2004).
14
that job types and employer characteristics such as employer size and industry have
significant power in explaining telecommuting adoption. However, some variables such
as home location and job tenure may be affected by telecommuting status as well. Using
them directly as explanatory variables yields biased model estimates in these studies.
2.3 Theory and Empirical Liteature
The question of interest here is what effects telecommuting has on workers' journey-
to-work, and, in particular, on commute length. A monocentric-city framework as
described in Brueckner (2001) can be utilized to convey some simple intuition about the
likely impact of telecommuting on commute length. Suppose two types of workers,
commuters and telecommuters, live in a city where all employment is concentrated in the
central business district (CBD). Telecommuters travel to the CBD for work only part of
the week while commuters go five days a week. Because telecommuters have lower
commuting costs than commuters, all else equal, they bid less for homes close to the
CBD and more for homes in suburban areas than commuters. In equilibrium, commuters
live close to the CBD and telecommuters sort into the surrounding region with longer
commutes (see Appendix A for a formal exposition).
The monocentric model, though simple and stylized, predicts that telecommuting
results in a longer commute distance due to a reduction in the marginal cost of
commuting. In a more realistic model that features cities with multiple employment
centers (Glaeser and Kahn 2001), the result may not be so straightforward. In a
polycentric city, employers who are located farther from regions where potential
15
qualified employees live may use telecommuting as a tool in the recruitment (e.g.,
Prystash 1995, Guimaraes and Dallow 1999). This would attract individuals who would
choose to work near their homes if they had to commute everyday. This seems
particularly likely for married women who are more often the secondary earner of the
family and, on average, have shorter commutes than their husbands. Thus, telecommuters
could have longer commutes than non-telecommuters in a polycentric city if they choose
an employer located farther from their home who offerstelecommuting.
The preceding discussion suggests that the impact of telecommuting on commute
length is an empirical question. The difficulty of testing the hypothesis that
telecommuting increases one-way commute distance lies in that telecommuting choice is
unlikely to be exogenous to commuting preference and/or behavior. If original longer
commute encourages an individual to work from home when allowed, a regression of
commute length on telecommuting status will overestimate the effect of telecommuting.
On the contrary, telecommuters could be those who feel more pressures from traffic.
They would have shorter commutes in the absence of telecommuting opportunities. This
unobserved selection will lead to a downward base in the regression estimates. The
existing literature has started to notice the policy significance of the question, but has not
addressed it satisfactorily.
Earlier studies (e.g., Kitamura et al. 1991; Koening et al. 1996; Henderson and
Mokhtarian 1996) found that telecommuting led to a large reduction in total VMTs.
These studies all treat the decision to telecommute as exogenous. Among recent studies,
16
Mokhtarian et al. (2004) analyzedretrospective data from a survey of 218 California state
government employees regarding their telecommuting and commuting behavior over a
ten-year period, from 1988 to 1998. The authors found that telecommuters had higher
one-way commuting lengths than non-telecommuters. Again, assuming telecommuting is
an exogenous choice, the study was unableto tell whether longer commuting distances
encouraged telecommuting or telecommuting facilitated residential relocation farther
from work. Ellen and Hempstead (2002) examinedthe correlation between
telecommuting and city size using the work schedule supplement to the May 1997 CPS.
Their results showed that telecommuters were more likely to live in large, high-density
metropolitan areas. As the authors acknowledge, these results fail to shed light on a
causal relationship: telecommuting opportunities were more likely to appear in
information-intensive service businesses, which tend to concentrate in large, dense
metropolitan areas.
2.4 The NHTS and an Empirical Baseline
The NHTS is a survey of the daily and long-distance travel behavior of the American
public conducted periodically by the Federal Highway Administration (FHWA) since
1969. In the 2001 NHTS, 69,817 households wereinterviewed. The survey collected
detailed information about travel of all sorts including the journey to work. A
shortcoming of the NHTS data is that it does not have much information about a
respondent's job, so that it is difficult to instrument for telecommuting as I do below. I
instead use the NHTS to generate a conditional correlation between telecommuting and
17
commute length, which sets a baseline for comparison with the two-sample instrumental
variables estimates I obtain from the combined CPS and PUMS samples.
The sample constructed from the 2001 NHTS includes individuals who lived in a
Metropolitan Statistical Areas (MSA) and had a job at the time of the survey.
Unfortunately, the NHTS did not ask whether the individual was self-employed. The
problem is mitigated by excluding those who always work at home or have no fixed
workplace. A small portion of respondents with outlier values for commute length or
speed are also removed from the sample.
6
The final sample contains 47,730 individuals
from 33,326 households. I treat as telecommuters those who substitute working from
home for traveling to their usual workplace once every month or more. In this case,
telecommuters constitute of 7.1 percent of the sample. This figure is higher than in the
2001 CPS because the self-employed who work in a fixed place outside the home some
days and at home other days are counted as telecommuters.
7
Table 2-1 reports means and standard deviations of key variables for telecommuters
and non-telecommuters in the NHTS sample. It is clear that the two groups of workers
differ considerably in demographic and socioeconomic characteristics. Telecommuters
6
As there is no way to identify whether outliers are due to misreporting, I employ conservative thresholds
on commute length and speed in sample selection. Individuals reporting one-way commute time greater
than 180 minutes, commute distance longer than 180 miles, or speed lower than 0.01 or greater than 1.5
miles per minute are removed from the sample, which results in 189 exclusions.
7
Due to data constraint, the minimum frequency requirement (one day every month) in the NHTS
definition is lower than that (one day every two weeks) of the CPS. Counting the self-employed, the
percentage of telecommuters in the CPS goes up to 6.2. The difference in the definitions may explain part
of the remaining gap.
18
are, on average, more likely to be male, white, older, better educated, more likely to be
married, have young children and have higher household incomes compared to non-
telecommuters. Telecommuters are more concentrated in professional, managerial, or
technical occupations than non-telecommuters. In terms of commuting patterns, an
average telecommuter spends about 3.5 more minutes and travels an additional 2.6 miles
for a one-way trip to his workplace than an average non-telecommuter. Figures 2-1graph
the distributions of commuters and telecommuters across groups defined by commuting
distance or commuting time. The proportions of telecommuters that fall in groups with
longer commutes are higher than the proportions of commuters in those groups. The last
row of Table 2-1 shows that driving is the main travel mode for 92 percent of commuters.
The proportion of workers commuting by car is 3 percentage points lower among
telecommuters.
A naïve approach to examining how telecommuting impacts the journey to work is to
estimate a single-equation regression model with a commuting variable (i.e. length or
travel mode) as the dependent variable and telecommuting status, together with other
relevant variables, as the explanatory variables. Table 2-2 reports the OLS coefficient
estimates of the telecommuting dummy. Telecommuting has a large positive effect on
both commute time and commute distance for married women. A married female
telecommuter is estimated to travel 3 minutes or 3 miles longer to work than a married
female commuter, ceteris paribus. The estimates for single women and men are smaller
and statistically indistinguishable from zero. In terms of travel mode, telecommuters,
19
except for single women, are less likely by 4 – 5 percentage points—to drive to work
than the average commuter. However, none of these results should be interpreted as the
causal effects of telecommuting as it is likely that people choose telecommuting based on
how far and by which means they commute. The confounding factors would cause OLS
estimates to be biased and the direction of the bias is unclear. To obtain consistent
estimates of the impacts of telecommuting on the journey-to-work, we need to instrument
for telecommuting choice.
The coefficient estimates for other variables indicate that commute lengths as well as
probability of driving to work increase with age (at a decreasing rate), education, and
household income across different population groups. Black workers commute longer
than white, Hispanic and Asian workers, as is documented in the spatial mismatch
literature (e.g., Kain 1968). Married men commute longer than single men. The variables
have qualitatively the same effects on the probability of driving to work.
2.5 Empirical Strategy
2.5.1 Instrumental Variable
The opportunities for teleworking and telecommuting vary substantially from job to
job because of the variation in the relativeproductivity of working from home to working
on-site, which is generally determined by the need for face-to-face communication with
colleagues and customers, as well as the need for team-work. Theapplication of
telecommunication technology during teleworking could alter the substitutability of
teleworking for face-to-face contact. For some jobs, internet technology maintains or
20
even increases the productivity of employees working from home, while for others, it
appears less helpful. The employees in the former case are likely to have more options for
teleworking and telecommuting. While a variable measuring the occupational technology
penetration for teleworking may explain individual’s telecommuting choice, some
unobserved occupational characteristics that affect commute length might be correlated
with that variable. For instance, a high school teacher uses the internet less often when
she works at home than a college professor does. Furthermore, there are more high
schools geographically scattered in a city than colleges. An instrumental variable that
shows that a high school teacher has fewer telecommuting opportunities may also capture
the difference in the geographical distributions of the two jobs if the latter is not well
measured or controlled for in the model.
In the early 2000s internet services, and in particular the broadband capacity, were
not evenly distributed across the country. Some studies show that internet infrastructure
investment or city accessibility to the internet was biased toward larger metropolitan
areas and a group of midsized urban areas (e.g. Malecki, 2002; Grubesic and O'Kelly,
2002). Consequently, the competitiveness of the broadband market varied considerably
across regions. The Federal Communications Commissions (2002) shows that 40.5% of
zip codes had none or one broadband line, in contrast to 27.6% of zip codes with four or
more high-speed lines by J une 2001. The number of broadband providers increased with
population density (Grubesic and Murray, 2004), and rural and smaller metropolitan
areas failed to attract significant levels of competition. The spatial variation in the
21
internet and broadband markets could have led to spatial differences in technology
options for teleworking for different occupations.
Thus, I develop an instrumental variable to measure the penetration of internet for
teleworking across occupation and city size using the work schedule supplement to the
May 2001 CPS, from which we know whether a respondent ever worked at home and
what equipment they used when they were working at home. I calculate the percentage of
employees for each of 270 (45 x 6) occupation-by-MSA-size combinations who ever
worked at home and used the internet (hereafter referred to as internet penetration). The
higher the value, the more likely a person in the occupation-by-city-size cell is to work
from home and possibly telecommute. The advantage of exploiting the variation in the
interaction of occupation and city size is that the effects of unmeasured occupation and
urban structure attributes on commuting behavior can be purged by the introduction of
occupation and city fixed effects in the model. To ensure measurement accuracy, the
occupation-by-city-size cells with fewer than 50 observations are not used in the baseline
analysis. This results in 179 cells covering 37 occupations and 6 MSA sizes. The cell-size
weighted mean (standard deviation) of internet penetration is 0.088 (0.113). In the
sensitivity analysis, I lower the cell selection criterion to 30 observations.
Figure 2-2 shows that internet penetration varies substantially across both
occupations and city sizes. In general, white-collar workers such as professionals,
teachers, and sales representatives have higher average internet penetration as well as
larger variation across city sizes than blue-collar workers such as mechanics and
22
repairmen, or transportation and production workers. College teachers and lawyers and
judges have the highest teleworking internet penetration (0.5 or above), which seems
reasonable since these two occupations are information intensive as well as flexible in
where work is performed. Sales in finance, business and non-retail commodities have
much higher percentages of internet-using teleworkers than retail sales probably because
the latter require personal presence and more face-to-face interaction with customers.
It is plausible to assume that the instrumental variable is not systematically
correlated with other unobservables that affect commuting behavior conditional on the
occupation and city fixed effects. However, to address the concern about this assumption,
I also construct a set of variables at the occupation-by-city level with the PUMS and test
how robust the instrumental variable estimates are to including these variables. More
detail about the occupation-by-city variables and the test is presented in the next section.
2.5.2 The Two-Sample Instrumental Variables (TSIV) Method
Traditionally, instrumental variables estimation is performed when the outcome
variable, the potentially endogenous variable of interest and the instrumental variable
exist in one data set. In addition, a large sample is generally needed for IV estimation to
produce sufficient statistical power. In our case, the instrumental variable discussed
above is measured for two-digit occupation by city size. It cannot be assigned to the
NHTS samplebecause the NHTS contains little information about respondents’ jobs. (A
five-category variable is used to describe occupation as opposed to 45 two-digit
occupations in the CPS.) To the best of my knowledge, there is no other (large) data set
23
that contains information oncommuting, telecommuting, occupation and MSA of
residence.
8
Therefore, a traditional instrumental variable method is infeasible here.
Angrist and Krueger (1992) developed a two-sample instrumental variables (TSIV)
technique that allows one to apply IV estimation to a joint sample with two data sets, one
of which has the outcome and the instrumental variable and the other the endogenous
explanatory variable and the instrument. The work schedule supplement to the May 2001
CPS collected information about respondents’ working at home, occupation and MSA.
The 2000 5% PUMS collected information about journey-to-work as well as occupation
and MSA. Moreover, they were both intended to represent the US population within the
same period and contain many of the same questions. Thus, they constitute a suitable case
for the TSIV method to work.
Formally, suppose the model of interest is
, y X| c = +
where y and c are 1 n× vectors and X is an n k × matrix of regressors, some of
which are correlated with c . An n l × (l k > ) matrix Z is needed to consistently estimate
| , where Z is not correlated with c and ( ) lim / 0
n
p Z X n
÷·
' = . Angrist and Krueger point
out that in the case when only X and Z (but not y ) are observed in one data set and
only y and Z (but not X ) are observed in the other, | can still be consistently
estimated when certain assumptions, which will be discussed in detail in the next
8
The NLSY79, an alternative data set, provides only commuting time in the early 1990s for fewer than
10,000 workers. The information about telecommuting in the NLSY79 is limited to hours worked at home.
24
subsection, hold for the two samples. Many researchers have since used the two-sample
approach (e.g., Currie and Yelowitz 2000, Dee and Evans 2003) to circumvent the data
constraint. In practice, a two-stage least squares procedure is usually adopted to produce
the following estimator
( )
1
2 2 2 2
TSIV
X X X y |
÷
' ' =
? ? ? ?
,
9
where ( )
1
2 2 1 1 1 1
X Z Z Z Z X
÷
' =
?
,
1
X and
1
Z are from the first sample, and
2
y and
2
Z are
from the second.
Now suppose equation (2-1) describes the structural model of commute length (or
mode):
ikc ikc ikc k c ikc
y a W B T u ì µ v = + + + + + (2-1)
where
ikc
y is the commute time or travel mode of individual i living in MSA c with
occupation k ,
ikc
W is a vector of individual specific exogenous variables,
k
µ and
c
v are
occupation and MSA fixed effects, and
ikc
u is idiosyncratic disturbance. The potentially
endogenous variable,
ikc
T , is an indicator for telecommuting. The parameter of interest,
ì , measures the causal impact of telecommuting on commute length or travel mode.
The first stage in calculating the TSIV estimate of ì is to estimate a model of
telecommuting adoption as described by equation (2-2),
1 1 1 1 1 1 1 1 1
,
ikc ikc ks k c ikc
T a W B Z u ì µ v = + + + + +
1
1, , i n = ? (2-2)
9
Inoue and Solon (2005) called this estimator the two-sample two-stage least squares (TS2SLS) estimator
and showed that it is different from the TSIV estimator originally proposed by Angrist and Krueger. They
proved that the TS2SLS estimator is asymptotically more efficient than the TSIV estimator. That being said,
I continue to label the estimator TSIV to distinguish it from the one-sample IV approach.
25
where the subscript 1 denotes the CPS sample,
1
n is the sample size in the CPS, and
ks
Z
is the instrumental variable measured at occupation by MSA size level (s ). The
parameters estimates are applied to the second sample, i.e. the PUMS sample, to predict
telecommuting status,
2ikc
T
?
. In the second stage, the TSIV estimate of ì is generated by
regressing the outcome variables in the PUMS,
2ikc
y , on the predicted telecommuting
status,
2ikc
T
?
and other covariates. In an exactly identified case such as ours, we can
alternatively fit a reduced-form equation, i.e. equation (2-3), using the PUMS sample,
2 2 2 2 2 2 2 2 2
,
ikc ikc ks k c ikc
y a W B Z u ì µ v = + + + + +
2
1, , i n = ? (2-3)
where subscript 2 denotes the PUMS sample and
2
n is the sample size of the PUMS. The
TSIV estimate is just the ratio between the reduced-form and first-stage coefficients
before
ks
Z , i.e.
1
2
ˆ
ˆ
ˆ
ì
ì
ì =
TSIV
.
Standard errors of the TSIV estimator can be computed using a linear Taylor series
approximation assuming zero covariance between the first-stage and reduced-form
estimators. That is
|
|
.
|
\
|
+ =
2
2
2
2
2
1
2
1
2
1
2
2 2
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ì
o
ì
o
ì
ì
o
TSIV
(2-4)
where
1
ˆ o and
2
ˆ o are estimated standard errors of
1
ˆ
ì and
2
ˆ
ì , respectively.
10
10
Note that with this approximation formula the t-statistic of the TSIV estimates is the following function
of the t-statistics of the first-stage and reduced-form estimates:
2 2
2 1 2
2 2
1 2
TSIV
t t
t
t t
=
+
. When
1
t , the first-stage t-
statistic outweighs
2
t ,
TSIV
t approaches
2
t .
26
2.5.3 CPS and PUMS Samples
In addition to assumptions underlying the traditional IV model, the TSIV approach
imposes some conditions on the joint sample. The key one is that the two data sets must
represent the same population. It is plausible to argue that these conditions hold for the
samples constructed from the CPS and the PUMS. The CPS is administered by the
Bureau of the Census for the Bureau of Labor Statistics. The former is also in charge of
implementing the decennial census of the US, from which the PUMS was created. Both
the CPS and the PUMScollected a rich set of information from US households on
individuals’ demographic characteristics, labor force experience, household attributes and
economic status. In addition to the similarity in content, the phrasing of questions and
coding of potential responses are similar across the CPS and the PUMS. While the CPS
and PUMS are both intended to be representative of the US population, the PUMS
includes institutionalized individuals, who are excluded from the CPS. I remove these
observations from the PUMS in constructing the joint sample. Moreover, every variable
in the sample is ensured to have the same support across the two sources. For instance,
only workers who are 16 years old or above, live in an MSA and are not self-employed
on the main job are retained in my data. MSAs that appear in just one data set are
removed. The final data includes 234 common MSAs.
Nevertheless, potential mismatches between the CPS and the PUMS might exist due
to the differences in sampling design, response rates and survey times. The CPS selects
households by primary sampling units (PSUs) based on the 1990 Census while the PUMS
27
draws households with sampling rates varying with the housing density of census blocks
or tracts.
11
Second, the Census spent tremendous effort to induce people to fill out the
survey forms, which led to higher response rates in the Census than the CPS. Finally, the
CPS data were collected in May 2001 roughly one year after the 2000 Census was
conducted. A visual comparison of the weighted means
12
of the CPS and the PUMS
samples does not suggest significant differences for most variables between the samples.
However, t-tests reject the mean equality for several variables across the two samples.
13
Table 2-3 presents descriptive statistics for telecommuters and non-telecommuters in
the CPS sample. Telecommuters are 3 – 4 years older than commuters on average, and
disproportionately white and better educated. They are more likely to be married and live
in smaller households, with higher annual incomes. In terms of job types, telecommuters
are concentrated in occupations such as executives, administrators, managers, math and
computer scientists, teachers of all levels, lawyers and judges, and sales representatives in
finance and business services—workers who are generally in the upper levels of the job
hierarchy. White-collar workers in the service sector and blue-collar workers have fewer
11
See http://usa.ipums.org/usa/chapter2/chapter2.shtml for a detailed explanation.
12
Sample weights contained in the CPS and PUMS are applied in calculating summary statistics and
estimation to adjust the over-sampling in each survey.
13
In the notation of section 2.5.2, the condition on the joint sample can be formally written as
( ) ( )
1 2
1 1 1 2 2 2
lim / lim /
ZX
n n
p Z X n p Z X n
÷· ÷·
' ' = = E
.
It requires that the first and second moments of explanatory variables including the instrumental variable of
the two samples converge to the same matrix. It can be tested for the variables that are observed in both
samples. A t-test on the means just examines the first moment and is likely to reject the null given large
sample size. With increasing applications of the TSIV method, formal ways to test the assumption and to
evaluate the potential bias resulting from mismatch of the two samples are probably desirable.
28
opportunities to telecommute. Finally, a higher proportion of telecommuters than non-
telecommuters live in large MSAs with populations over one million. As in Table 2-1,
Table 2-3 showsthat telecommuters and non-telecommuters differ in many observed
ways. It is therefore likely that they also differ in unobserved variables that are correlated
with commuting behavior.
In the PUMS, commute length is recordedin minutes and measures how long it
usually took the respondent to get from home to work during the past week. White
(1988a) argues that time is a better measure of commuting costs than distance because
time is the scarce resource that people economize. Moreover, Table2-2 shows that the
same set of variables explains more variation in commute time than commute distance,
This suggests that the noise associated with distance is larger than with time: A commuter
can estimate commuting time more accurately than distance. The translation of the
impact of telecommuting on commute time into an impact on commute distance is
considered below.
In the PUMS, 1.7 percent of the sample or over 60,000 respondents have travel times
that are top-coded at 99 minutes. Exploiting the properties of the Pareto distribution, I
replace the top-coded values with an estimate of the conditional expectation for top-
coded values. The procedure is described in Appendix B, which suggests a range for the
imputed values of 120 to 165 minutes. I use the lower bound, 120 minutes, in the
benchmark analysis. Since the likelihood of being top-coded is positively correlated with
telecommuting adoption, using the lower bound value works against finding a positive
29
effect of telecommuting. I check whether different imputed values affect the results in the
sensitivity analysis. The weighted average commute times in the PUMS are 24.2, 24.7
and 27.5 minutes for married women, single women and men, respectively. These figures
are slightly higher than in the NHTS sample, which are 22.1, 22.9 and 25.2, respectively.
In the PUMS, a higher share (93.3%) of married women drives to work than single
women (85.1%) and men (89.8%). Similar patterns are observed in the NHTS sample, for
which the shares are 94.1%, 87.7%, and 91.7%, respectively.
2.6 The First-Stage Estimates
In the first stage, I estimate a linear probability model of telecommuting adoption
(equation (2-2)). The dependent variable is a binary indicator equal to one if the worker is
telecommuting. In the baseline model, the explanatory variables include age, age squared,
gender, race, educational achievement, number of household members, presence of
children 5 years of age or younger, children between 6 and 15 years of age, spouse (for
the male sample only), annual household income, and the occupation-by-city-size internet
penetration measure. Industry and job class variables are not included in the model
because they are individual choices that are likely to be correlated with home location,
work location or commute length. Neither the wage, housing price, or travel mode and
time is used as an explanatory variable. All of these are chosen simultaneously with
commute length and, therefore, are endogenous. Fixed effects for MSA-of-residence and
2-digit occupation category are controlled for, assuming people do not sort into a city and
2-digit occupation based on their preferences for commute length or telecommuting.
30
The model is estimated for married women, single women and men separately using
individual weights provided by the CPS. Results are reported in Table2-4. In general, the
estimates reflect the differences between telecommuters and commuters in Table2-3.
People who are older, white, possess a college or advanced degree, have children and
come from affluent households are more likely to telecommute. Less obvious from the
descriptive statistics is that black employees have a higher probability of working at
home than other groups, although a lower probability than whites. Being married does
not seem to play a role in the telecommuting decisions for male employees. All else
equal, telecommuting is significantly more popular among professionals and sales
representatives in finance and business services, but less popular among engineers and
supervisors. Surprisingly, blue-collar workers are not less likely to telecommute than
white collar workers, conditional on demographic and economic covariates. This may be
because people with less education are offered more telecommuting opportunities when
working in blue-collar jobs than in white-collar jobs.
Several variables have differing influences on telecommuting adoption across the
samples. Race plays an important role in telecommuting for married women but not for
single women. In contrast, household size and income are more important for the latter
than for the former. The likelihood of telecommuting increases with age at a decreasing
rate for women workers. This pattern is much weaker and statistically insignificant for
men. A male employee with a graduate degree has a substantially larger propensity to
telecommute than one with a college degree, but this is not the case for a female
31
employee. Men tend to work at home if there are older children but not younger children
in the household. The reverse is true for married women – suggesting that married
women may use telecommuting as a way to combine work and childcare.
The coefficients on the instrumental variables are of paramount importance and vary
substantially across samples. In the case of married women, a 10 percentage point
increase in occupation/MSA internet penetration causes the probability of telecommuting
to rise by 5.4 percentage points once 2-digit occupation and MSA fixed effects are
controlled for. This effect is statistically significant at the1% level. On the contrary, the
estimate for single women is smaller (0.143) and statistically insignificant, suggesting the
instrumental variable has little explanatory power for single women employees. For male
employees, a 10 percentage point increase in internet penetration increases the probability
of telecommuting by 2.9 percentage points, an effect that is significant at the 5% level.
One critical assumption underlying the IV approach is that teleworking technology
penetration is not correlated with any unobservable that influences commute length or
mode. There might be concerns that the instrumental variable is correlated with
occupation-specific local labor market conditions. For instance, the urban economics
literature hypothesizes that individuals are forward looking when they choose home
location and commute length. They take into account labor market dynamics and
potential moving costs. Specifically, Crane (1996) predicts a shorter commute for persons
with lower probability of changing jobs within the local labor market. Likewise, van
32
Ommeren et al. (1997) argue that commuting distance is decreasing in the arrival rates of
job offers and increasing in moving costs.
One way to deal with this concern is to control in the model for occupation-by-city
attributes. Lacking clear theory informing what those attributes should be, I construct a
rich set of covariates using the PUMS data. I calculate the fraction of employees within
each 2-digit occupation and MSA combination who are: male, white, black, have a high
school degree, some college experience, a college degree, an advanced degree (omitting
high school dropouts), in the transportation and communication industries, in trade, in
finance, in services, in public administration(omitting the manufacturing and
construction industries), working for private for profit employers, and working for private
non-profit employers (omitting government). I also compute the labor market share,
median hourly wage, and difference between the 75
th
percentile wage and the 25
th
percentile wage of each occupation by MSA. Finally, using the CPS sample, I calculate
the fraction of employees for each 2-digit occupation and MSA size combination who
have flexible work hours.
Even columns in Table2-4report estimation results for the model with inclusion of
these occupation-city specific covariates. The coefficients of demographic and household
variables do not change much, although some occupation fixed effects vary. This
suggests that the constructed covariates pick up part of the variation in telecommuting
explained by occupation. The coefficient on internet penetration declines slightly to 0.48
for married women while statistical significance is maintained at the 1% level. The
33
coefficient for men is unchanged up to two decimal places. These results indicate that the
instrumental variable is likely to be orthogonal to the local labor market conditions
described by those covariates.
2.7 Reduced-Form and TSIV Estimates
2.7.1 Reduced-Form Estimates
Equation (2-3) is estimated only for married women and men since the instrumental
variable is not statistically significant for single women. The exogenous explanatory
variables are the same as in the first-stage except that they are from the PUMS sample.
Results are reported in Table 2-5. In the baseline model, commute length increases with
age at a decreasing rate for both women and men. Race makes a substantial difference in
commute length, which may reflect residential segregation and employment separation.
Black male workers on average spend 2 more minutes on the road than white and other
workers and black females travel 4 minutes longer than white females. Regardless of
gender, college graduates and those from high-income households live farther from their
workplace than employees without a college degree and workers from low income
households. Married men travel 1 minute longer to work than single men. When there are
younger children in the household, both married women and men travel longer to work,
while the presence of older children has the opposite, but smaller, effect for women.
Commute time increases with the number of household members for men and decreases
for married women. Overall, the results are consistent with those from the NHTS and
largely agree with those in White (1986). Commute length varies significantly across jobs
34
even conditioning on factors like age, race, and education. One possible reason is the
variation in geographic concentrations of different occupations. For example, school
teachers have short commutes because schools are scattered throughout a city.
The instrumental variable shows large positive impacts on married women's
commute lengths but not on men's commute lengths. In the baseline model without
controlling for occupation-MSA covariates, i.e., Columns 1 and 3, a 10 percentage point
increase in the proportion of employees of each 2-digit occupation and MSA size
combination who ever use internet when working at home leads to 0.60 minute longer
commuting trip for married women. The estimate is statistically significant. In contrast,
the coefficient estimate of the internet penetration for male workers is 0.13 minutes and
statistical insignificant.
When the occupation-by-city covariates are controlled for in the model, few changes
occur in the coefficients of the demographic and household variables. However, a number
of occupation fixed effects vary dramatically. This suggests the importance of
heterogeneity in local markets for different occupations in determining commute length.
The coefficients of the occupation-MSA covariates imply that conditional on individual
characteristics, commute length increases if the person works in an occupation that has
more human capital, is concentrated in finance and services industries, is more
represented in the private for profit sector and has a larger labor market share. The last
result seems to be consistent with Crane’s theory that a person values commuting
distance less if more potential employers are available.
35
The effect of internet penetration on commuting length declines slightly and retains
statistical significance for married female workers. Now, a 10 percentage point increases
in internet penetration lead to an additional 0.46 minutes in commute time for married
women. The estimate for male workers is less than 0.2 minutes and statistically
insignificant. The results, consistent with those without occupation-by-city covariates,
suggest that the instrumental variable is unlikely to pick up the occupation-city specific
attributes as confounding factors.
2.7.2 TSIV Estimates of the Effects on Commute Length
First-stage estimates indicate that the internet penetration instrumental variable has
statistically significant and positive impacts on the telecommuting status of married
women and men in the 2001 May CPS. The reduced-form estimates indicate that the
instrumental variable has a substantial positive effect on one-way commute time of
married women but little effect for male employees in the 2000 PUMS. The TSIV
procedure ties these two sets of results together to generate consistent estimates of the
causal effects of telecommuting on commute length.
Table 2-6a presents the TSIV estimates calculated as the ratios of reduced-form
estimates to the first-stage estimates of the instrumental variable. In the exactly identified
case, it yields the same estimates as the two-stage least square estimation in the two
sample case (TS2SLS). The standard errors of the TSIV estimates are computed using
equation (2-4). The TSIV estimates suggest that telecommuting has a substantial positive
impact on married women’s commute lengths. All else equal, working at home at least
36
one day every two weeks, on average, causes a married women employee to commute 9 –
11 minutes longer than if she commutes every day. The impact for male employees is
smaller in magnitude (around 5 minutes) and statistically indistinguishable from zero. In
comparison with OLS estimates, the TSIV estimates yield qualitatively similar results.
However, OLS results underestimate the effect of telecommuting for married women,
which is consistent with the fact that married women usually have short commutes if they
do not telecommute.
2.7.3 Effects of Telecommuting on Commute Mode
OLS analysis of the NHTS data shows that male and married female telecommuters
are less likely to drive to work than non-telecommuters. It is difficult to find a compelling
reason why telecommuting leads people to forego driving to work. The OLS estimates
are susceptible to an omitted variable bias that fails to account for sorting of women who
take public transit to work into telecommuting. Moreover, driving usually is faster than
taking public transit or any other travel mode.
14
If telecommuting does cause a worker to
commute by a mode other than driving, the lengthened commute time might be a result of
choosing a slower travel mode rather than an increase in commute distance. Therefore, it
is important to identify the true effect of telecommuting on commute mode.
I apply the same TSIV procedure to the travel mode variable available in the PUMS
sample. Using the same argument that internet penetration is unlikely to affect travel
14
The average speeds for commuting by driving, by rail, by bus, and by bicycle in the NHTS are 0.53, 0.36,
0.28, and 0.23 miles per minute, respectively.
37
mode choice directly, TSIV produces consistent estimates of the effects of telecommuting
on travel mode choice. Table 2-6b reports both reduced-form and TSIV estimates for
travel mode. In the baseline model, the TSIV estimates are small, positive and without
statistical significance for both married women and men. When the occupation-by-city
covariates are added, the estimate for married women is almost zero while the estimate
for men becomes negative with a large standard error. Overall, the TSIV point estimates
do not support the OLS results that telecommuting reduces a married woman’s
probability of driving to work. The negative OLS estimates could result from the fact that
employees who commute by public transit also prefer to telecommute. However, the
TSIV estimates are not sufficiently precise to let us draw definite conclusions about the
effect.
2.7.4 Sensitivity Analysis
I examine the sensitivity of the above results to different sample restrictions and
alternative imputed values for the top-coded commute times. Tables 2-7a and 2-7b report
the estimates for the commute time and travel mode models, respectively. In Panel A of
each table, the samples are extended to include the occupation-MSA size cells that
contain 30 or more CPS observations, which results in 216 cells covering 38 2-digit
occupations and 6 MSA sizes. In the first stage, the instrumental variable has a smaller
effect for married women while the coefficient for men does not change much as
compared to the case with cells containing over 50 observations. It continues to have a
large, statistically significant reduced-form effect on married women's commute time and
38
little effect on men's commute time. The TSIV estimates show that telecommuting
increases married women's commute time by 13 minutes though they lack enough
statistical power in the case with job-by-city covariates included. The effect of
telecommuting among male employees falls to 3 and 4 minutes, andthe t-statistics are
less than 1. As far as travel mode is concerned, telecommuting shows some positive
effects for both married women and men, but again the estimates are not distinguishable
from zero. These results are highly consistent with the baseline case with cells larger than
50 observations.
Telecommuting is often thought of as a choice for office workers only. Programs and
policies that aim at promoting telecommuting usually target these occupations rather than
the entire working population. Therefore, it may be of interest to examine the effects of
telecommuting on commuting behavior for office workers. One way to define office
workers is to narrow the sample down to the 2-digit occupations coded 1 through 26.
Included in this group are managerial, professional specialty, technical, sales, and
administrative support occupations. 2-digit occupation codes greater than 26, including
service, precision production, craft, repair, farming, forestry and fishing occupations and
operators, fabricators and laborers, are excluded. Panels B of Tables 2-7a and 2-7b
present the estimates for the sample of office workers. The instrumental variable affects
only the telecommuting propensity of married women. Telecommuting is estimated to
lengthen the one-way commute time of married women by 8 - 9 minutes, which is
statistically significant at the 10% level. Again, telecommuting has a positive but
39
statistically insignificant effect on married women’s commute mode, contrary to the OLS
estimates. In sum, estimates with different sample restrictions demonstrate that the effects
of telecommuting on commuting show stability and a certain degree of homogeneity
across occupations. Panel C of Table 2-7a shows that replacing the top-coded commute
time by 165 minutes instead of 120 minutes has no impact on the effects of
telecommuting on commute time.
2.8 Discussion
The TSIV estimates of the effects of telecommuting on commute time for married
women equal 9 to 12 minutes, which are 3 to 4 times the OLS estimates from the NHTS.
The results are plausible in that married women have shorter commutes on average. The
OLS analysis tends to underestimate the effects of telecommuting in this case. The
magnitude of the adjustment in the commute made by married women appear reasonable
given that the average commute time for married women in the PUMS is 24.2 minutes
with a standard deviation of 19. TSIV estimates suggest that telecommuting increases
commute time by about half of a standard deviation.
TSIV estimation could be biased if the internet penetration measured by occupation
crossed with MSA size is correlated with some unobservables that impact individual
commute lengths. The concern may be less serious as the models control for a rich set of
occupation-by-city specific covariates as well as occupation and city fixed effects.
Another potential source of bias is that the teleworking technology penetration is
measured with 2001 CPS data. When internet access expanded rapidly to a wider
40
population and more regions in the early 2000s, the variation across occupation and cities
declined quickly with time. Therefore, the impact of internet penetration on
telecommuting adoption estimated using 2001 data may underestimate the impact in year
2000 when the PUMS were collected, which would result in an overestimation of the
TSIV coefficients.
I am interested in translating the effects of telecommuting on commute time into the
effects on commute distance. I use the NHTS data to estimate a relationship between
commute time and distance for people driving to work. Table 2-8 shows the coefficients
of models that project commute time onto commute distance and distance squared.
15
Commute time is a concave function of distance with an intercept greater than zero,
which suggests a positive fixed cost and an increasing marginal speed. The relationship
between commute time and distance varies by sex, with women having greater concavity.
Using these estimates, we can recover the approximate distance from travel time. For
instance, suppose a woman drove 24 minutes to work before choosing to telecommute.
Applying the projection estimates implies that on average her commute distance was 13
miles. If her one-way commute timeincreases to 33 minutes after telecommuting, the
one-way commute time increases to 20.5 miles, a7.5 miles increase. If she works from
home 2 days a week (the national average for telecommuting women is 2.2 days per
week), the total weekly commutes are 198 minutes or 123 miles, representing 17 percent
15
Higher order polynomials were tried. They produce very bad predictions for distances on the high end.
Moreover, the predictions for the mid-range values do not differ with and without the higher order terms.
41
and 5.5 percent declines relative to the before-telecommuting commute times and
commute miles, respectively.
2.9 Conclusion
Telecommuting has been promoted as a means to deal with congestion and
automobile emissions by researchers and public policy makers. However, there are
concerns that telecommuting workers will make a longer commute in response to the
lower commute frequency. Naïve (OLS) estimates based on the NHTS show that a
married woman commutes 3 minutes or 3 miles longer if she telecommutes. The NHTS
estimates also show that telecommuters except single women are less likely to drive to
work than non-telecommuters. However, these estimates could be biased because
telecommuting is not randomly assigned among workers. Furthermore, theory cannot
predict the direction of the bias.
By applying two-sample instrumental variables technique to the CPS and PUMS
samples, I find that telecommuting causes married women employees’ commuting trips
to increase by 9 to 12 minutes. The effect for male workers is also positive, but smaller
and not precisely estimated. For single women, the instrumental variable does not have
enough power to explain telecommuting choice. In addition, TSIV estimates show a
small, positiveeffect of telecommuting on the probability of commuting by car for
married women. Although lacking statistical power, this does not agree with the negative
relationship between telecommuting and driving to work found in the OLS analysis.
Given the sizable “rebound” effect on one-way commute time found among married
42
women, thetotal commute miles traveled by an average married women worker are
unlikely to decline in proportion to telecommuting frequency.
Unfortunately, the instrumental variable developed in this paper does not have
enough information to let us estimate the effects of telecommuting for men and single
women. This needs to be explored in future research. Moreover, to understand whether
telecommuters adjust their commute distance by changing residential location or
employment location is important for both research and policy purposes and should also
be examined.
43
Figures and Tables for Chapter 2
Figure 2-1. Distributions of Telecommuters and Commuters by Commute Time and
Distance
0
.
1
.
2
.
3
.
4
<=10 11-20 21-30 31-40 41-50 51-60 >60 min
Source: Author's calculation using NHTS 2001.
Distributions of Workers by One-way Commute Time
Commuters
Telecommuters
0
.
1
.
2
.
3
.
4
<=5 5-10 10-15 15-30 30-45 45-60 >60 miles
Source: Author's calculation using NHTS 2001.
Distributions of Workers by One-way Commute Distance
Commuters
Telecommuters
44
Figure 2-2. Internet Penetration by 2-Digit Occupation and MSA size (2001 CPS)
0 .2 .4 .6
Teleworking Techonology Penetration
Farm Workers
Other Handlers
Freight & Stock Handlers
Construction Laborers
Other Transportation
Motor Vehicle Operators
Fabr. Assem. Inspe. Samp.
Machine Oper. & Tenders
Other Precis. Prod. & Craft
Construction Trades
Mechanics & Repairers
Personal Service
Clean & Bldg.Service
Health Service
Food Service
Protective Service
Other Admin. Support,
Mail And Message Distr.
Finan. Rcds. Process.
Secre. Steno. & Typists
Super,, Admin. Support
Sales, Retail & Personal
Sales, Non-retail Comm,
Sales, Finance. & Busi.
Super. & Propr., Sales
Other Technicians
Engineering & Sci. Tech.
Health Tech.
Other Professional
Lawyers & J udges
School Teachers
College/Univ. Teachers
Health Assess. & Treat.
Math & Computer Sci.
Engineers
Management
Other Exe, Admin.
Sources: Work Schedule Supplement to May 2001 CPS
Teleworking Technology Penetration by Occupation and MSA Size
MSA Size:
100-250K
250-500K
500K-1M
1-2.5M
2.5-5M
>5M
Note: Internet penetration is calculated as the weighted percentage of employees who ever work at home
and use the internet within each occupation-by-MSA-size cell using data from the Work Schedule
Supplement to the May 2001 CPS.
45
Table 2-1. Descriptive Statistics of NHTS Sample
Non-telecommuters Telecommuters
Variables Mean Std. Dev. Mean Std. Dev.
Raw N 44556 3174
Age 39.205 12.370 42.250 11.243
Male 0.541 0.498 0.599 0.490
White 0.708 0.455 0.807 0.395
Black 0.123 0.329 0.060 0.238
Asian 0.029 0.168 0.044 0.206
Hispanic 0.110 0.313 0.056 0.231
High School Degree 0.290 0.454 0.111 0.314
Some College 0.303 0.460 0.252 0.434
College Degree 0.216 0.411 0.369 0.483
Graduate Degree 0.115 0.319 0.248 0.432
Spouse 0.608 0.488 0.664 0.472
Child Age 0 – 5 in HH 0.211 0.408 0.225 0.418
Child Age 6 – 15 in HH 0.310 0.463 0.312 0.464
Household Size 3.152 1.441 2.990 1.364
HH Income $40 – 70K 0.322 0.467 0.233 0.423
HH Income $70 – 100K 0.191 0.393 0.256 0.437
HH Income >$100K 0.152 0.359 0.343 0.475
Sales or Services 0.266 0.442 0.236 0.425
Clerical or Administrative
Support 0.136 0.342 0.059 0.236
Manufacturing, Construction,
Maintenance, or Framing 0.180 0.384 0.061 0.239
Professional, Managerial, or
Technical 0.417 0.493 0.644 0.479
Time to Work 23.688 17.889 27.167 22.362
Distance to Work 12.628 12.800 15.238 16.194
Drive to Work 0.917 0.276 0.884 0.321
Note: Sample is constructed from the 2001 NHTS including workers who live in an MSA, have an
outside-home fixed workplace, and have one-way commute distance less than 180 miles, commute time
less than 180 minutes and commute speed less than 1.5 miles per minute and greater than 0.01 miles per
minute. Observations with missing values for any of the listed variables are also dropped. Means and
standard deviations are calculated using the weights from the NHTS.
46
Table 2-2. OLS Estimates of the "Effect" of Telecommuting on Commute Lengths and
Travel Mode, 2001 NHTS
Married
Women
Single
Women
Men
Telecommuting (1) (2) (3)
A. COMMUTE TIME (MINUTES)
Coefficient 2.904** 0.452 1.652
Standard Error (1.225) (1.410) (1.040)
R-sq 0.11 0.12 0.08
B. COMMUTE DISTANCE (MILES)
Coefficient 3.124*** 1.063 1.144
Standard Error (0.968) (1.010) (0.699)
R-sq 0.09 0.09 0.05
C. DRIVE TO WORK
Coefficient -0.043** 0.013 -0.051***
Standard Error (0.019) (0.025) (0.014)
R-squared 0.11 0.17 0.13
#Observations 14176 8939 24615
Note: The sample is the same as in Table 2-1. All models include
age, age squared, race, education, household composition, annual
household income, and job category and MSA fixed effects.
Heteroscedastic-robust standard errors without clustering are in
parentheses. * indicates significant at 10%, ** significant at 5%,
and *** significant at 1%.
47
Table 2-3. Summary Statistics of CPS Sample by Gender and Telecommuting Status
Women Men
Non-
telecommuters Telecommuters
Non-
telecommuters Telecommuters
Variables
Mean Std. Mean Std. Mean Std. Mean Std.
Raw N 13528 594 14481 544
Age 38.678 12.806 41.663 10.829 38.437 12.614 43.020 11.367
White 0.795 0.404 0.886 0.318 0.833 0.373 0.921 0.271
Black 0.148 0.355 0.081 0.273 0.111 0.314 0.045 0.208
High School Degree 0.285 0.452 0.139 0.346 0.283 0.450 0.096 0.295
Some College 0.314 0.464 0.227 0.419 0.272 0.445 0.175 0.381
College Degree 0.204 0.403 0.399 0.490 0.201 0.401 0.416 0.493
Graduate Degree 0.086 0.281 0.202 0.402 0.095 0.293 0.300 0.459
Spouse Present 0.501 0.500 0.642 0.480 0.577 0.494 0.730 0.444
With Child 0 – 5 in HH. 0.194 0.395 0.222 0.416 0.221 0.415 0.196 0.397
With Child 6 – 15 in HH. 0.327 0.469 0.338 0.474 0.314 0.464 0.320 0.467
Household Size 3.076 1.495 3.029 1.444 3.236 1.592 2.924 1.426
Annual Family Income <$40K 0.364 0.481 0.163 0.369 0.320 0.467 0.087 0.282
Annual Family Income $40 – 75K 0.337 0.473 0.310 0.463 0.355 0.479 0.245 0.430
Annual Family Income >$75K 0.299 0.458 0.527 0.500 0.325 0.469 0.669 0.471
2-digit Occupation
01 Public Administrators and
Officials 0.000 0.016 0 0 0.001 0.025 0 0
02 Other Executive,
Administrators, and Managers 0.099 0.299 0.165 0.371 0.120 0.324 0.252 0.435
03 Management Related
Occupations 0.053 0.224 0.077 0.267 0.032 0.175 0.063 0.244
04 Engineers 0.004 0.065 0.003 0.057 0.035 0.184 0.046 0.210
05 Math. and Computer Scientists 0.013 0.111 0.038 0.192 0.025 0.156 0.051 0.221
06 Natural Scientists 0.003 0.058 0.010 0.100 0.005 0.069 0.006 0.077
07 Health Diagnosing Occupations 0.005 0.067 0.004 0.059 0.007 0.082 0.010 0.102
08 Health Assessment and Treating 0.049 0.216 0.026 0.160 0.007 0.083 0 0
09 College and University
Teachers 0.007 0.082 0.045 0.208 0.007 0.085 0.057 0.232
10 Other Teachers 0.065 0.246 0.152 0.359 0.021 0.143 0.039 0.194
11 Lawyers and J udges 0.004 0.066 0.010 0.097 0.007 0.082 0.032 0.177
12 Other Professional Specialty 0.042 0.200 0.097 0.296 0.032 0.177 0.110 0.313
13 Health Technologists and
Technicians 0.025 0.155 0.007 0.084 0.004 0.060 0 0
14 Engineering and Science
Technicians 0.007 0.084 0.003 0.052 0.015 0.123 0.009 0.093
15 Other Technicians 0.010 0.100 0.014 0.116 0.015 0.121 0.027 0.162
16 Sales Supervisors and
Proprietors 0.027 0.162 0.024 0.154 0.033 0.180 0.032 0.176
17 Sales Representatives, Finance
and Business Service 0.018 0.132 0.055 0.229 0.018 0.131 0.089 0.285
18 Sales Representatives,
Commodities except Retail 0.006 0.075 0.017 0.129 0.018 0.131 0.073 0.261
19 Sales Workers, Retail and
Personal Services 0.067 0.249 0.026 0.159 0.036 0.186 0.020 0.139
20 Sales Related Occupations 0.001 0.030 0 0 0.000 0.021 0 0
21Supervisors, Administrative
Support 0.010 0.099 0.001 0.038 0.004 0.064 0.003 0.055
22 Computer Equipment Operators 0.004 0.060 0.001 0.036 0.003 0.054 0 0
23 Secretaries, Stenographers, and 0.042 0.201 0.040 0.197 0.001 0.031 0.003 0.055
48
Typists
24 Financial Records Processing 0.028 0.164 0.030 0.170 0.003 0.056 0.003 0.057
25 Mail and Message Distributing 0.005 0.073 0 0 0.010 0.099 0.002 0.044
26 Other Administrative Support
Occupations, including Clerical 0.153 0.360 0.078 0.269 0.046 0.209 0.009 0.094
27 Private Household Service 0.001 0.025 0.004 0.061 0 0 0 0
28 Protective Service Occupations 0.009 0.093 0.002 0.044 0.031 0.172 0.004 0.066
29 Food Service Occupations 0.058 0.234 0 0 0.044 0.206 0 0
30 Health Service Occupations 0.036 0.186 0.007 0.086 0.004 0.064 0.003 0.056
31 Cleaning and Building Service 0.022 0.145 0 0 0.025 0.156 0 0
32 Personal Service 0.041 0.199 0.046 0.210 0.009 0.093 0.002 0.041
33 Mechanics and Repairs 0.004 0.061 0.003 0.052 0.061 0.240 0.024 0.153
34 Construction Trades 0.002 0.048 0 0 0.071 0.257 0.007 0.082
35Other Precision Production 0.012 0.111 0.004 0.066 0.040 0.195 0.003 0.050
36 Machine Operators and Tenders 0.024 0.153 0.001 0.023 0.039 0.194 0.004 0.065
37 Fabricators, Assemblers,
Inspectors, and Samplers 0.016 0.124 0.007 0.083 0.025 0.155 0.001 0.036
38 Motor Vehicle Operators 0.008 0.089 0 0 0.052 0.222 0 0
39 Other Transportation and
Material Moving 0.001 0.027 0 0 0.018 0.131 0 0
40 Construction Laborer 0.000 0.015 0 0 0.013 0.114 0.005 0.069
41 Freight, Stock and Material
Handlers 0.012 0.107 0.003 0.057 0.034 0.182 0 0
42 Other Handlers, Equipment
Cleaners, and Laborers 0.005 0.067 0 0 0.011 0.102 0.001 0.036
43 Farm Operators and Managers 0.000 0.018 0 0 0.000 0.021 0.004 0.060
44 Farm Related Workers 0.006 0.079 0 0 0.021 0.142 0.003 0.056
45 Forestry and Fishing
Occupations 0.000 0.008 0 0 0.001 0.022 0.003 0.051
MSA w/ Population 100k – 250k 0.089 0.284 0.062 0.241 0.084 0.278 0.076 0.266
MSA w/ Population 250k – 500k 0.140 0.347 0.115 0.319 0.134 0.341 0.094 0.291
MSA w/ Population 500k – 1m 0.166 0.372 0.139 0.346 0.156 0.363 0.111 0.314
MSA w/ Population 1m – 2.5m 0.306 0.461 0.339 0.474 0.316 0.465 0.390 0.488
MSA w/ Population 2.5m – 5m 0.168 0.374 0.195 0.397 0.176 0.380 0.191 0.393
MSA w/ Population 5m+ 0.131 0.338 0.150 0.357 0.134 0.341 0.138 0.346
Note: Sample is constructed from the May 2001 CPS including workers who live in an MSA and are not self-employed
on the main job. Observations with missing values for any of the listed variables are also dropped. Means and standard
deviations are calculated using the weights from the CPS.
49
Table 2-4. First-Stage Estimates of Telecommuting Models, May 2001 CPS
Married Women Single Women Men
(1) (2) (3) (4) (5) (6)
Age 0.004** 0.004** 0.002** 0.002** 0.001 0.001
(0.002) (0.002) (0.001) (0.001) (0.001) (0.001)
Age Squared -0.000* -0.000* -0.000* -0.000* -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
White 0.044*** 0.042*** 0.014 0.013 0.026*** 0.026***
(0.009) (0.008) (0.010) (0.010) (0.006) (0.006)
Black 0.023* 0.020 0.010 0.010 0.019*** 0.019***
(0.013) (0.013) (0.012) (0.012) (0.007) (0.007)
High Scholl Degree -0.013 -0.014 -0.005 -0.005 0.001 0
(0.011) (0.011) (0.004) (0.003) (0.003) (0.003)
Some College -0.008 -0.010 0.001 0.001 -0 -0.001
(0.012) (0.012) (0.005) (0.005) (0.004) (0.004)
College Degree 0.034** 0.033** 0.026*** 0.027*** 0.025*** 0.024***
(0.014) (0.014) (0.008) (0.008) (0.008) (0.008)
Graduate Degree 0.023 0.021 0.009 0.010 0.035*** 0.035***
(0.019) (0.019) (0.015) (0.015) (0.011) (0.011)
Spouse 0.001 0.001
(0.005) (0.005)
With Child0-5 in HH. 0.022*** 0.021** 0.013* 0.013* 0.007 0.007
(0.008) (0.008) (0.007) (0.007) (0.005) (0.005)
With Child 6-15 in HH. 0.010 0.009 0.014* 0.014* 0.013*** 0.012***
(0.009) (0.009) (0.007) (0.007) (0.005) (0.004)
Household Size -0 0 -0.004* -0.004* -0.005*** -0.004***
(0.003) (0.003) (0.002) (0.002) (0.002) (0.001)
HH Income $40 – 75K -0.001 -0 0.008 0.008 0.004 0.003
(0.007) (0.007) (0.006) (0.006) (0.004) (0.004)
HH Income >75K 0.007 0.007 0.037*** 0.038*** 0.030*** 0.029***
(0.009) (0.009) (0.009) (0.009) (0.005) (0.005)
0.027 0.038 0.009 0.018 0.018 0.027 03 Management Related
Occupations (0.023) (0.037) (0.020) (0.032) (0.018) (0.024)
04 Engineers -0.032 -0.019 -0.049*** 0.009 -0.026** -0.059*
(0.038) (0.061) (0.014) (0.043) (0.013) (0.032)
05 Math. and Computer Scientists -0.025 -0.013 0.073** 0.112*** -0.027 -0.019
(0.035) (0.043) (0.033) (0.040) (0.021) (0.025)
0.025 0.083 0.010 0.045 -0.015 -0.016 08 Health Assessment and
Treating (0.032) (0.069) (0.029) (0.045) (0.025) (0.053)
-0.169** -0.143 0.052 0.079 -0.026 0.043 09 College and University
Teachers (0.076) (0.117) (0.136) (0.122) (0.089) (0.097)
10 Other Teachers -0.026 0.089 0.050*** -0.019 -0.019 0.021
(0.019) (0.073) (0.016) (0.078) (0.018) (0.051)
11 Lawyers and J udges -0.150 -0.200 -0.078** 0.071 0.005 0.045
(0.097) (0.144) (0.033) (0.076) (0.072) (0.090)
12 Other Professional Specialty 0.041* 0.039 0.027 0.040 0.054*** 0.074**
(0.023) (0.038) (0.018) (0.028) (0.017) (0.030)
0.067* 0.142* 0.014 0.019 0.007 0.045 13 Health Technologists and
Technicians (0.039) (0.076) (0.027) (0.045) (0.026) (0.050)
0.047 0.096* 0.001 0.001 0.015 0.043 14 Engineering and Science
Technicians (0.034) (0.050) (0.027) (0.041) (0.022) (0.035)
15 Other Technicians 0.051 0.069* -0.010 0.031 0.025 0.017
(0.037) (0.042) (0.026) (0.036) (0.026) (0.029)
50
0.050* 0.156** 0.028 0.003 0.011 -0.017 16 Sales Supervisors and
Proprietors (0.026) (0.069) (0.021) (0.046) (0.019) (0.034)
0.024 -0.015 0.041 0.017 0.090*** 0.148** 17 Sales Representatives, Finance
and Business Service (0.033) (0.059) (0.028) (0.056) (0.032) (0.065)
-0.050 0.048 0.133* 0.118 0.051* -0.010 18 Sales Representatives,
Commodities except Retail (0.054) (0.069) (0.070) (0.075) (0.028) (0.040)
0.082** 0.205** 0.021 -0.040 0.034 0.003 19 Sales Workers, Retail and
Personal Services (0.038) (0.094) (0.029) (0.056) (0.029) (0.047)
-0.013 0.016 0.022 0.032 -0.037* -0.019 21Supervisors, Administrative
Support (0.025) (0.045) (0.040) (0.049) (0.019) (0.028)
0.086** 0.134** 0.028 0.020 0.120 0.150 23 Secretaries, Stenographers, and
Typists (0.038) (0.066) (0.031) (0.044) (0.082) (0.093)
0.096** 0.140** 0.015 0.016 0.041 0.055 24 Financial Records Processing
(0.038) (0.062) (0.028) (0.042) (0.043) (0.049)
0.042 -0.021 0.004 -0.011 0.015 0.062 25 Mail and Message Distributing
(0.042) (0.130) (0.033) (0.077) (0.030) (0.055)
0.067** 0.072 0.018 -0.018 0.014 0.043 26 Other Administrative Support
Occupations, including Clerical (0.033) (0.059) (0.027) (0.041) (0.027) (0.040)
0.044 0.088 0.016 0.061 0.007 0.091** 28 Protective Service Occupations
(0.035) (0.096) (0.036) (0.060) (0.023) (0.041)
0.063 0.189** 0.020 -0.031 0.029 0.018 29 Food Service Occupations
(0.038) (0.087) (0.032) (0.056) (0.030) (0.048)
0.076* 0.165** 0.018 0.013 0.050 0.091 30 Health Service Occupations
(0.039) (0.081) (0.030) (0.052) (0.042) (0.070)
0.057 0.160** 0.014 -0.002 0.021 0.077 31 Cleaning and Building Service
(0.038) (0.070) (0.032) (0.053) (0.029) (0.051)
32 Personal Service 0.103** 0.151** 0.049 0.030 0.026 0.058
(0.041) (0.068) (0.030) (0.043) (0.029) (0.049)
33 Mechanics and Repairs 0.033 0.067 0.064 0.066 0.027 0.047
(0.035) (0.057) (0.066) (0.077) (0.026) (0.044)
34 Construction Trades 0.064 0.136* 0.002 -0.027 0.021 0.035
(0.039) (0.071) (0.030) (0.058) (0.028) (0.043)
35Other Precision Production 0.073* 0.163*** 0.023 -0.006 0.014 0.018
(0.040) (0.060) (0.028) (0.046) (0.027) (0.037)
0.071* 0.181** 0.011 -0.029 0.027 0.041 36 Machine Operators and
Tenders (0.039) (0.072) (0.032) (0.055) (0.030) (0.044)
0.067* 0.173** 0.042 0.009 0.019 0.031 37 Fabricators, Assemblers,
Inspectors, and Samplers (0.037) (0.068) (0.039) (0.056) (0.028) (0.041)
38 Motor Vehicle Operators 0.057 0.037 0.008 0.009 0.020 0.020
(0.039) (0.083) (0.032) (0.057) (0.029) (0.047)
-0.001 0.098 -0.014 -0.019 0.017 0.021 39 Other Transportation and
Material Moving (0.035) (0.074) (0.035) (0.059) (0.029) (0.045)
40 Construction Laborer 0.051 0.154* -0.035 -0.055 0.022 0.048
(0.040) (0.080) (0.033) (0.064) (0.029) (0.046)
0.087 0.140* 0.019 -0.006 0.026 0.033 41 Freight, Stock and Material
Handlers (0.054) (0.082) (0.032) (0.055) (0.030) (0.046)
0.047 0.168** 0.008 -0.032 0.019 0.034 42 Other Handlers, Equipment
Cleaners, and Laborers (0.048) (0.075) (0.033) (0.057) (0.031) (0.048)
44 Farm Related Workers 0.079** 0.174** 0.010 -0.026 0.029 0.066
(0.038) (0.071) (0.032) (0.057) (0.030) (0.046)
Internet Penetration 0.539*** 0.481*** 0.143 0.183 0.288** 0.285**
(0.160) (0.163) (0.127) (0.134) (0.119) (0.115)
Constant -0.194*** -0.446** -0.081** 0.089 -0.072** -0.247**
51
(0.053) (0.181) (0.039) (0.159) (0.032) (0.102)
J ob-by-city Covariates N Y N Y N Y
Observations 6936 6936 6553 6553 13809 13809
R-squared 0.07 0.08 0.07 0.07 0.09 0.09
Note: All models include MSA fixed effects. Occupation-by-city covariates include fractions of employees within
each 2-digit occupation and MSA who are male, white, black, have high school degree, some college, college degree,
advanced degree, work in industries of transportation and communication, trade, finance, services, or public
administration, and work in private profit or private non-profit sectors. Also included are occupation’s local labor
market share, median log of wage, inter-quartile log of wage, and fraction of employees that have flexible work
hours. Robust standard errors are estimated clustering on MSA. * indicates significant at 10%, ** significant at 5%,
and *** significant at 1%.
52
Table 2-5. Reduced-Form Estimates of Commute Time Model, 2000 PUMS
Married Women Men
(1) (2) (3) (4)
Age 0.174*** 0.174*** 0.575*** 0.574***
(0.019) (0.018) (0.017) (0.017)
Age Squared -0.003*** -0.003*** -0.006*** -0.006***
(0) (0) (0) (0)
White -1.532*** -1.498*** -0.018 -0.002
(0.344) (0.337) (0.222) (0.208)
Black 2.322*** 2.329*** 2.048*** 2.034***
(0.374) (0.377) (0.314) (0.326)
High Scholl Degree -0.731*** -0.640*** 0.194*** 0.229***
(0.151) (0.133) (0.071) (0.069)
Some College 0.126 0.181 0.276*** 0.286***
(0.153) (0.134) (0.098) (0.098)
College Degree 0.935*** 0.943*** 0.762*** 0.711***
(0.181) (0.187) (0.231) (0.226)
Graduate Degree 1.628*** 1.603*** -0.182 -0.324
(0.290) (0.261) (0.267) (0.247)
Spouse 1.193*** 1.191***
(0.107) (0.107)
With Child 0-5 in HH. 1.370*** 1.354*** 0.667*** 0.649***
(0.099) (0.098) (0.088) (0.085)
With Child 6–15 in HH. -0.535*** -0.533*** -0.091 -0.103
(0.106) (0.104) (0.092) (0.089)
Household Size -0.455*** -0.452*** 0.301*** 0.313***
(0.043) (0.041) (0.038) (0.036)
HH Income $40 – 75K 0.320*** 0.337*** 0.330*** 0.341***
(0.097) (0.095) (0.079) (0.075)
HH Income >75K 1.278*** 1.256*** 1.118*** 1.099***
(0.151) (0.148) (0.146) (0.142)
1.908*** 2.335*** 1.741*** 1.546*** 03 Management Related
Occupations (0.247) (0.506) (0.239) (0.525)
04 Engineers 1.074*** 3.083*** 1.155*** 2.090***
(0.411) (1.185) (0.273) (0.795)
05 Math. and Computer Scientists 3.007*** 4.387*** 2.604*** 3.340***
(0.333) (0.569) (0.333) (0.650)
08 Health Assessment and Treating -0.434 -0.379 -1.485** -0.359
(0.662) (1.024) (0.603) (1.228)
-1.924** 3.412* -2.793*** 0.201 09 College and University
Teachers (0.862) (1.956) (0.922) (2.106)
10 Other Teachers -7.506*** -3.951*** -5.218*** -0.868
(0.369) (1.271) (0.362) (1.900)
11 Lawyers and J udges -0.911 0.464 -0.947 -3.326
(0.967) (2.855) (0.660) (2.456)
12 Other Professional Specialty -1.984*** 1.559** -2.100*** 0.711
(0.204) (0.637) (0.175) (1.009)
-0.090 2.122** -0.468 1.705 13 Health Technologists and
Technicians (0.640) (1.068) (0.600) (1.605)
14 Engineering and Science
Technicians
2.347*** 6.782*** 0.791* 3.843***
(0.595) (1.043) (0.438) (1.216)
15 Other Technicians 3.615*** 3.371*** 2.792*** 2.707***
53
(0.373) (0.668) (0.446) (0.741)
-1.611*** 3.980*** -1.555*** 2.511** 16 Sales Supervisors and
Proprietors (0.368) (1.082) (0.364) (1.080)
-1.752*** -3.124** -0.679** -4.056** 17 Sales Representatives, Finance
and Business Service (0.273) (1.558) (0.336) (1.799)
2.259*** 8.260*** 2.597*** 7.295*** 18 Sales Representatives,
Commodities except Retail (0.408) (1.080) (0.287) (1.002)
-3.976*** 2.761* -3.205*** 2.712* 19 Sales Workers, Retail and
Personal Services (0.630) (1.419) (0.615) (1.438)
0.143 3.418*** -0.112 2.892*** 21Supervisors, Administrative
Support (0.411) (0.755) (0.447) (0.864)
-0.096 3.904*** -0.016 3.984*** 23 Secretaries, Stenographers, and
Typists (0.542) (0.996) (0.585) (1.324)
24 Financial Records Processing 0.176 4.643*** 0.730 4.785***
(0.586) (1.047) (0.580) (1.025)
25 Mail and Message Distributing -0.399 3.378* -3.019*** 5.746***
(0.800) (1.911) (0.683) (1.757)
-0.095 1.157 -1.027* 0.525 26 Other Administrative Support
Occupations, including Clerical (0.568) (1.025) (0.533) (0.847)
28 Protective Service Occupations 0.034 4.622** -1.556** 2.642
(0.861) (1.885) (0.668) (1.744)
29 Food Service Occupations -4.354*** 2.308 -3.580*** 1.958
(0.713) (1.562) (0.653) (1.517)
30 Health Service Occupations -0.555 3.314*** -1.573** 1.725
(0.666) (1.216) (0.694) (1.730)
31 Cleaning and Building Service 0.610 5.096*** -2.378*** 0.985
(0.623) (1.424) (0.660) (1.754)
32 Personal Service -1.586*** 3.159** -1.509** 3.387**
(0.582) (1.218) (0.598) (1.598)
33 Mechanics and Repairs 2.594*** 6.459*** 0.015 2.849**
(0.601) (1.132) (0.583) (1.401)
34 Construction Trades 4.153*** 10.447*** 5.044*** 9.022***
(0.830) (1.477) (0.650) (1.545)
35Other Precision Production -0.472 5.480*** -0.701 3.599***
(0.595) (1.187) (0.602) (1.345)
36 Machine Operators and Tenders 0.151 6.061*** -1.077* 3.070**
(0.649) (1.417) (0.637) (1.553)
0.558 6.950*** 0.233 4.696*** 37 Fabricators, Assemblers,
Inspectors, and Samplers (0.678) (1.302) (0.637) (1.495)
38 Motor Vehicle Operators -2.504*** 1.576 -1.549** 3.569**
(0.789) (1.422) (0.714) (1.548)
1.275 7.537*** 1.806*** 6.669*** 39 Other Transportation and
Material Moving (1.075) (1.653) (0.665) (1.596)
40 Construction Laborer 5.827** 13.929*** 5.626*** 10.816***
(2.508) (2.834) (0.695) (1.737)
-0.629 5.467*** -1.286** 4.183*** 41 Freight, Stock and Material
Handlers (0.632) (1.462) (0.650) (1.557)
-0.339 6.301*** -0.761 3.934** 42 Other Handlers, Equipment
Cleaners, and Laborers (0.827) (1.652) (0.643) (1.765)
44 Farm Related Workers 0.732 6.945*** -1.091 3.227*
54
(0.968) (1.714) (0.739) (1.679)
Internet Penetration 5.988** 4.625** 1.364 1.599
(2.704) (2.311) (2.660) (2.188)
Constant 24.001*** 5.690 12.845*** -3.926
(1.012) (3.475) (0.735) (3.755)
J ob-by-city Covariates N Y N Y
Observations 832956 832956 1720931 1720931
R-squared 0.08 0.08 0.07 0.07
Note: All models include MSA fixed effects. Robust standard errors are estimated clustering on
MSA. * indicates significant at 10%, ** significant at 5%, and *** significant at 1%.
55
Table 2-6. TSIV Estimates of the Effects of Telecommuting on Commute Time and
Mode
Table 2-6a. Commute Time
Married Women Male
(1) (2) (3) (4)
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 5.988** 4.625** 1.364 1.599
Standard Error 2.704 2.311 2.66 2.188
#of Observations 832956 832956 1720931 1720931
TSIV
Coefficient 11.109* 9.615* 4.736 5.610
Standard Error 6.004 5.805 9.441 8.004
J ob-by-city Covariates N Y N Y
Table 2-6b. Commute Mode
Married Women Male
(1) (2) (3) (4)
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 0.013 0.003 0.006 -0.043
Standard Error 0.047 0.036 0.05 0.037
#of Observations 849904 849904 174329 1743292
TSIV
Coefficient 0.024 0.006 0.021 -0.151
Standard Error 0.087 0.075 0.174 0.143
J ob-by-city Covariates N Y N Y
Note: All models include age, age squared, race, education, household
composition, annual household income, and occupation and MSA fixed effects.
Robust standard errors are estimated clustering on MSA.* indicates significant at
10%, ** significant at 5%, and *** significant at 1%.
56
Table 2-7. Robustness Check of the TSIV Estimates
Table 2-7a. Commute Time
Married Women Male
(1) (2) (3) (4)
A. OCCUPATION – MSA SIZE CELL >=30
First-Stage
Coefficient 0.419** 0.346** 0.302*** 0.325***
Standard Error 0.169 0.174 0.115 0.115
#of Observations 7157 7157 14724 14724
Reduced-Form
Coefficient 5.427*** 4.519*** 1.285 0.909
Standard Error 2.086 1.713 2.212 1.839
#of Observations 864097 864097 1831544 1831544
TSIV
Coefficient 12.952* 13.061 4.255 2.797
Standard Error 7.216 8.225 7.502 5.744
B. OFFICE WORKERS
First-Stage
Coefficient 0.574*** 0.532*** 0.254 0.244
Standard Error 0.169 0.171 0.175 0.164
#of Observations 5477 5477 6974 6974
Reduced-Form
Coefficient 5.369* 4.583* 1.858 0.966
Standard Error 3.051 2.355 2.054 2.137
#of Observations 654502 654502 845161 845161
TSIV
Coefficient 9.354 8.615* 7.315 3.959
Standard Error 5.986 5.221 9.529 9.154
C. TOPCODED COMMUTE TIME REPLACED BY 165 MIN
First-Stage
Coefficient 0.539*** 0.481*** 0.288** 0.285**
Standard Error 0.16 0.163 0.119 0.115
#of Observations 6936 6936 13809 13809
Reduced-Form
Coefficient 6.088** 4.517* 1.63 1.806
Standard Error 2.792 2.473 2.856 2.5
#of Observations 832956 832956 1720931 1720931
TSIV
Coefficient 11.295* 9.391 5.660 6.337
Standard Error 6.170 6.047 10.189 9.137
J ob-by-city Covariates N Y N Y
57
Table 2-7b. Commute Mode
Married Women Male
(1) (2) (3) (4)
A. OCCUPATION – MSA SIZE CELL >=30
First-Stage
Coefficient 0.419** 0.346** 0.302*** 0.325***
Standard Error 0.169 0.174 0.115 0.115
#of Observations 7157 7157 14724 14724
Reduced-Form
Coefficient 0.032 0.022 0.035 0.023
Standard Error 0.043 0.033 0.047 0.036
#of Observations 881551 881551 1855151 1855151
TSIV
Coefficient 0.076 0.064 0.116 0.071
Standard Error 0.107 0.101 0.162 0.114
B. OFFICE WORKERS
First-Stage
Coefficient 0.574*** 0.532*** 0.254 0.244
Standard Error 0.169 0.171 0.175 0.164
#of Observations 5477 5477 6974 6974
Reduced-Form
Coefficient 0.045 0.035 -0.059 -0.054
Standard Error 0.059 0.04 0.040 0.049
#of Observations 667751 667751 862045 862045
TSIV
Coefficient 0.078 0.066 -0.232 -0.221
Standard Error 0.105 0.078 0.225 0.250
J ob-by-city Covariates N Y N Y
Note: All models include age, age squared, race, education, household composition,
annual household income, and occupation and MSA fixed effects. Robust standard
errors are estimated clustering on MSA.* indicates significant at 10%, ** significant at
5%, and *** significant at 1%.
58
Table 2-8. Projection of Commute Time (Minute) onto Commute Distance (Mile)
All Women Men
(1) (2) (3)
Commute Distance 1.253 1.363 1.224
(0.018) (0.042) (0.023)
Distance Squared -0.002 -0.004 -0.0019
(0.0003) (0.001) (0.0003)
Constant 7.279 6.742 7.375
(0.143) (0.229) (0.203)
#Observations 44,218 21,404 22,814
Adj. R-squared 0.739 0.684 0.765
Note: The sample includes only those who drive to workin the
sample in Table2-1.
59
3 Do People Drive Less on Code Red Days?
3.1 Introduction
By 2007, 347 counties with 141 million residents were designated by EPA as
ground-level ozone nonattainment areas.
16
This means that nearly half of the US
population breathes air with ozone concentration above a harmful level. Besides the
established fact that ozone has adverse effects on the respiratory system, recent studies
(e.g., Bell et al. 2004) also link ozone levels with increases in mortality. Therefore,
bringing the ozone levels into compliance with the EPA standard is a goal of high priority
for public policy.
Ozone is formed when its precursors, oxides of nitrogen (NOx) and volatile organic
compounds (VOCs), react in the atmosphere. Peak ozone levels typically occur on hot,
dry and sunny summer days. Emissions from motor vehicle exhaust, industrial facilities
and electric utilities are the main sources of NOx and VOCs. Dramatic increases in the
number of cars and miles they are driven contribute significantly to the ozone problem in
urban areas, in spite of the fact that individual vehicles are getting cleaner. According to
the EPA (2003), motor vehicles account for 56% and 45% of emissions of NOx and VOC
nationwide, respectively.
A number of metropolitan areas have implemented public information programs that
aim at mitigating the ozone problem by encouraging voluntary driving reductions on high
ozone days. Examples include the Air Quality Action Days (AQAD) program in the
16
See http://www.epa.gov/air/oaqps/greenbk/o8index.html
60
Washington-Baltimore metropolitan area, the Spare the Air (STA) program in the San
Francisco Bay area, and the Ozone Action Days program in Atlanta, to name a few.
Undoubtedly, these programs have low implementation and enforcement costs, in
contrast to mandatory control programs. They also take advantage of the episodic feature
of the ozone problem, a strategy that theoretically promotes economic efficiency (Teller
1967, Krupnick 1988). However, the first question that needs to be addressed is how
effective these programs are in getting cars off the road. People receiving forecast
information may cancel trips due to the concerns about getting unhealthy exposure and/or
the environmental impacts of driving on those days. Nevertheless, for many people,
including most commuters, it is very costly to change travel schedules, if not impossible.
The question is whether the information provided creates enough incentives for the
recipients to take action.
Identifying the impact of the programs on vehicle driving also serves a practical
purpose in air quality regulation. These public information programs all fall into the
category of Voluntary Mobile Source Emission Reduction Programs (VMEPs). Since
1997, the U.S. EPA allows states with non-attainment areas to claim credits up to 3% of
projected emissions reductions for VMEPs when filing State Implementation Plans
(SIPs). To do so requires that mobile emission reductions through voluntary programs be
quantified.
Several studies have looked at how these voluntary information programs impact
travel behavior. Henry and Gordon (2003), MWCOG (2003), and Fox and Sarkar (2002)
61
use individual survey data to examine to what extent the ozone alerts have altered
behavior. They all find that a significant share of respondents reported taking actions
during ozone episodes to help abate pollution. For example, Fox and Sarkar report that in
the Washington-Baltimore area, 7-9 percent of respondents said they drove less on code
red days, days when the ozone levelsare predicted to exceed the EPA standard. A
common issue, however, with the self-reported information is that it may be biased due to
recollection difficulty or other subjective factors. Instead, Cummings and Walker (2000),
Cutter and Neidell (2007) and Welch et al. (2005) directly examine traffic volumes in
Atlanta and the San Francisco Bay area and train ridership in Chicago, respectively.
Cummings and Walker and Welch et al. found either the traffic reductions were too small
to be surely attributed to the program or the ozone advisories increased transit ridership
only in a small part of the Chicago area. To the contrary, Cutter and Neidell found that
STAs reduce total daily traffic by 2.5 to 3.5 percent, with most effects occurring during
and just after the morning peak hours.
The focus of this study is to examine the effectiveness of the AQAD program in the
Baltimore area, an area relying more on automobile driving than the Chicago, San
Francisco and Washington DC metropolitan areas.
17
The program forecasts daily ozone
level one day ahead and uses color codes to indicate expected ozone severity. When the
17
Besides a bus system, Baltimore’s transit system consists of a single-line metro subway and a three-line
light rail, most parts of which overlap. In terms of commuting, the percentage of drivers is higher and the
percentage of rail riders is lower than the national average, based on data from NHTS 2001. See Table A2
for a comparison of the distributions of commuters by commute mode across several cities.
62
ozone level is predicted to reach or exceed the one-hour ozone standard, 125 ppb, a code
red is announced. I use a regression discontinuity (RD) design to see whether traffic
volumes are lower on code red days due to the announcement. The study is closest to
Cutter and Neidell in methodology, and obtains somewhat similar results. The main
finding is that the coderedday announcement reduces inbound traffic volumes during
morning peak hours by 3-5%. Outbound traffic volumes in the evening peak hours fall
correspondingly. In contrast, on code orange days, when ozone levelsare predicted to
exceed 105 ppb but lie below 125 ppb, I do not observe a reduction in vehicle driving.
The rest of the chapter is organized as follows. Section 3.2 documents the detailsof
the AQAD program. Section 3.3 presents a theoretical account of potential behavioral
changes in response to code red days. Sections 3.4 and 3.5 describe the empirical
methods and data used, respectively. Section 3.6 presents the results, and section 3.7
discusses the policy implications of my findings.
3.2 AQAD Program in Baltimore Area
The Baltimore area, with over 2.5 million residents in 2000, consists of five
counties
18
andBaltimore city. It is designated as a nonattainment area by EPA under both
the 1-hour and 8-hour ozone standards. Since the mid-1990s the area has been
implementing the AQAD program jointly with the Washington metropolitan area. Under
the coordination of the Metropolitan Washington Council of Governments (MWCOG), a
18
They are Anne Arundel, Baltimore, Carroll, Harford and Howard. See
http://www.epa.gov/oar/oaqps/greenbk/baltimo.html for a regional map. The designated area is different
from the census MSA definition.
63
daily forecast of the ozone level
19
is conducted for each area by a panel of meteorologists
every day from May 1stto mid-September.
The ozone level is predicted as a quadratic function of a vector of variables including
maximum and average surface temperatures, wind speed, relative humidity, solar zenith
angel, and lagged ozone observations. Note that the predicting variables measure only the
most relevant air and climatological conditions. Because they do not include variables
that forecast vehicle travel demand and electric utility production, the model does not
account for human behavior. The parameters of the function are estimated using
historical observations and remain unchanged for the current year. The model produces
forecasts for each of seven locations in Baltimore area where ozone monitoring stations
are located. The highest one is chosen as the initial forecast for the area.
The expert panel meets on a conference call at 3:00 pm every afternoon to discuss
and make adjustments to the initial forecast. This stage is subjective to the extent that it
relies on the experience of the participants. A personal communication with one of the
panel members indicates that the rational for this subjective procedure is multi-fold. First,
some factors are hard to quantify or are insignificant in model estimation, e.g., the
direction of wind from outside the area, but need to be taken into account. The day of the
week is sometimes taken into account toaddressconcerns about traffic. The panel also
needs to consider different versions of weather forecasts as well as to adjust ozone
forecastsat the lower and upper ends because the model seems to perform better in the
19
The one-hour ozone level was forecast until 2003. Since 2004, eight-hour levels have been forecast..
64
middle range of ozone levels than at the extremes. Although the changes often involve
only a couple of units, they may result in a shift of the air quality category in which the
day falls.
A color code is assigned to the day based on the consensus forecast value. Table3-1
shows the ranges of forecast one-hour ozone concentrations and the corresponding code
colors. When the ozone level is predicted to exceed the EPA standard, i.e. 125 ppb, a red
code is designated and the day is called a code red day. The last column in Table3-1
shows the distribution of summer days across air quality categories. 28 days were
announced as code red days in 2001 through 2003, accounting for 6.8 percent of the
season. Code orange days indicate that the ozone concentration will reach a level
unhealthy for sensitive populations. These days account for 10.6 percent of the period.
The forecast as well as the code color are publicized throughvarious communication
channelsonce they are available. People who subscribe to a mailing list receive email
notification. Local employers who enroll in the clean air partner program receive an
email or fax. Major newspapers and TV and radio stations will report air quality forecasts
together with weather forecasts. News sources will highlight code red days to enhance
visibility for the program. People are urged to take actions to reduce ozone precursors
emissions on high ozone days. On the top of the action list is reducing driving by all
means, includingcarpooling, teleworking, riding transit, and consolidating trips.
65
3.3 Theory
A simple discrete choice model can be used to analyze individual's choice between
driving and its substitutes on code red days. Specifically, we consider staying/working at
home and using public transit as the alternatives an individual may choose. Other travel
options such as carpool and bicycle may be incorporated into the framework easily and
would not affect the main results obtained below.
Suppose an individual chooses to drive (d ), to ride public transit ( p ) or stay/work at
home (h ) in order to maximize her utility
ij ij ij
U V c = +
where i indexes individual, { , , } j d p h e . The utility is the sum of a deterministic part V
and an idiosyncratic part c . Further assume that the deterministic utility is a weighted
linear combination of travel benefits and a variety of travel costs. That is
0 1 2 3 ij ij ij ij ij
V B TC HC EC | | | | = + + +
where B stands for trip benefit ( 0
h
B = ), TC is travel cost, including gas, bus fare, and
time, HC is health cost resulting from exposure to bad air quality, and EC is the
environment cost associated with one's choice. The model assumes that everybody has
common weights, | 's, and
0
0 | > ,
1 2 3
, , 0 | | | < although the benefits/costs of each
choice vary by individuals. In one case, people may differ in the extent that they
internalize the negative impact on air quality for the same amount of driving, but they
value air quality equally. Also, people with existing respiratory problem may have greater
health costs than those without when exposed to the same air pollution.
66
We assume that
ij
c are independently distributed as type-I extreme value. Let
i
y
denote the choice of individual i that maximize the utility, i.e. argmax( , , )
i id ip ih
y U U U = .
The probability of choosing j is
3 3
0 0
Pr( | ) exp / exp
i i k ijk k ijk
k j k
y j x x x | |
= =
(
| | | |
= =
( | |
\ . \ .
¸ ¸
¿ ¿ ¿
where
0
x B
·
= ,
1
x TC
·
= ,
2
x HC
·
= , and
3
x EC
·
= . For all individuals,
( )/ ( )[1 ( )]
j jk j j k
p x x p x p x | c c = ÷ (3-1)
and
( )/ ( ) ( )
j lk j l k
p x x p x p x | c c = ÷ , l k = . (3-2)
We are interested in how probabilities of choosing different alternatives change when
it is a code red day. Equations (3-1) and (3-2) tell us that the probability change for any
alternative depends on the changes in each benefit/cost factors for the option itself and all
others on code red days. To derive further results from the model, we assert the following
changes and relationships
(i) 0
d p h
B B B A = A = A = ,
(ii) 0
d h p
TC TC TC A s = A s A ,
(iii) 0
h d p
HC HC HC = A s A < A ,
(iv) , 0
p h d
EC EC EC A A s s A .
Relationship (i) states that the benefit from making the trip does not change on
code red days for any option; (ii) reflects the assumption that people may expect other
people to forego driving for riding public transit. Thus, traffic is expected to be lighter
67
while transit becomes more crowded and uncomfortable; (iii) indicates that when air
quality gets worse, people walking to and waiting at the bus stop are more exposed to
ozone. Driving in a car may or may not increase risk while staying indoors is always safe;
and (iv) implies that people are altruistic and may gain satisfaction (negative cost) for not
driving on code red days or may feel guilty for driving. These relationships are sensible
and not all are necessary for reaching the theoretical conclusion below.
Taking into account the cost changes on code red days, the change in the
probability of driving (also taking transit and staying home) for an individual is
ambiguous. This is mainly because declining air quality lowers the travel cost and health
cost of driving relative to riding bus, although there may be some utility gain from
reducing emissions. Even if people do not speculate about the improved traffic on code
red days, or in some areas bus fares are waived for riders on high ozone days, which
results in lower travel cost for riding bus, bus ridership may still not go up due to health
concern about the air quality.
The above analysis shows that the voluntary information program does not
provide people incentives necessarily consistent with reducing driving on code red days.
It is important to empirically measure the impact of the program on driving amount.
3.4 Empirical Strategy
The primary question this study attemptsto answer is whether the AQAD program
changes individuals' travel behavior episodically. Do people reduce vehicle trips and/or
miles traveled on code red days? Ideally, we would like to have a random sample of
68
households in the Baltimore area, together withtheir daily VMTs for all summer days.
However, such micro-data do not exist. What is available, instead, is data measuring
traffic volumes during short time intervals on highways in the Baltimore area. (These will
be described in detail in the next section). With these data, we can estimate the following
model to measure changes in traffic on code red days that can be attributed to the AQAD
program.
it t t i it
y CRD X B ¸ u c = + + + (3-3)
where
it
y is (log) number of vehicles passing by traffic monitor i
20
on date t .
t
CRD is
an indicator for day t to be a code red day and the parameter ¸ measures the impact of
code red day announcement on highway traffic volumes. The vector
t
X contains other
time varying factors that may affect vehicle trips, such as contemporaneous and lagged
weather, the forecast 1-hour ozone concentration and observed ozone levels for the
previous day, contemporaneous and lagged gas prices, public holiday dummies and a set
of dummies for year, month, and day of the week. In a specification check, I include
lagged traffic of the same time block on previous days and seven days ago.
i
u represents
a monitor fixed effect to capture the time-invariant traffic characteristics for each
monitor. The variable
it
c is an unobserved idiosyncratic term.
The problem in consistently estimating ¸ is that code red daysare not random. Even
conditional on all those covariates, there still could be some variables missing in the
20
Please note this is different from the ozone monitoring stations mentioned earlier. Coincidentally, the
number of traffic monitors in the sample is also seven.
69
model that are correlated with the code red day announcement and traffic flow. For
instance, forecast weather plays a crucial role in predicting ozone concentration and
determining code color. It is also arguably important in affecting people's travel decision
for the coming day. However, historical forecast weather data is not readily accessible.
Although we control for the observed weather and its lag, they may fail to account
adequately for the forecast weather. If it is the case, a naïve regression estimation would
yield a spurious estimate of ¸ .
However, if we could control for the conditional expectation of the unobservables in
the model, we would still be able to estimate ¸ consistently. That is to estimate the
following model instead of equation (3-3),
( ) |
it t it t t i it
y CRD E CRD X B ¸ ì c u u = + + + + (3-4)
where ( ) |
it t
E CRD c is expectation of
it
c conditional on the code red day indicator, and
( | , , )
it it it t t i
y E y CRD X u u = ÷ . If
t
CRD is the only variable correlated with
it
c , OLS
estimation of equation (3-4) yields a consistent ¸ estimate. In practice, ( ) |
it t
E CRD c is
not observed. However, we know that the code red day announcement is completely
dependent on the forecast ozone concentration, denoted by O. When and only when the
forecast level exceeds a threshold, will it be a code red day. Formally,
*
( ) 1{ }
t t t
CRD f O O O = = >=
where
*
O denotes the threshold equal to 125 ppb. Thus, we could exploit the sharp
regression discontinuity design (e.g. van der Klaauw 2002) to measure the impact of code
red days. Since O, referred to as a running variable in the literature, captures all the
70
information contained in CRD , ( ) | ( | )
it t it t
E CRD E O c c = . Thus, we could estimate
equation (3-5)
( )
it t t t i it
y CRD k O X B ¸ u u = + + + + (3-5)
where ( ) k O is a flexible functional specification for ( | ) E O c . In the literature, ( ) k O
often takes the form of high-order polynomial series.
As noted above, the vector X contains a linear term inthe forecast ozone level. It
is, however, possible that the linear term is insufficient to completely account for the
correlation between CRD and c . Figure3-1 illustrates that the estimates (¸ ' ) obtained
by controlling only for the linear term in the forecast ozone level will underestimate (left
panel) or overestimate (right panel) the true effect when the correlation between CRD
and c is a nonlinear function of O. In the estimation, I include up to a fifth order
polynomial in the ozone forecast.
Two key assumptions must be satisfied in order to apply the regression discontinuity
method (Imbens and Lemieux 2008). First, it is assumed that there is no manipulation of
the running variable. In our case, if the expert panel adjusts the forecast ozone level to
move a day into or out of the code red category based on expected transportation
volumes, concern about the validity of RD strategy might be raised. A communication
from one of the panel members stated that no sophisticated traffic information (e.g.,
forecasted daily traffic volumes) beyond the day of the week was considered in
forecasting ozone concentration. More specifically, it happened occasionally that the
forecast level was adjusted upward for Monday or downward for Friday based on the
71
general impression about traffic patterns on these days. However, this is the only channel
through which traffic is taken into account in code red day classification. In the next
section, it is shown that Mondays and Fridays are not statistically more likely to be (or
not be) a code red day. In addition, all modelsare estimated controlling for day-of-week
dummy. In the robustness check, I exclude Mondays and Fridays from the sample used
for estimation.
The other assumption underlying the RD model is that the unobserved variables that
may affect traffic volumes evolve continuously at the cutoff point, i.e.125 ppb. This
assumption cannot be verified directly. As a specification test presented in the next
section, I check the discontinuity of the control variables, especially weather covariates.
If some variables are found discontinuous at 125 ppb, it casts doubt on the continuity
assumption for the unobservables.
3.5 Data
The Maryland Department of Environment (MDE) has archived the forecast and
observed daily maximum one-hour ozone concentrations for the Baltimore area. I use
data from May through mid-September --- the ozone season when the AQAD program is
in place --- from 2001 to 2003. I focus on these three years because traffic data is
available from 2001 and the color code assignment started to be based on an 8-hour
ozone forecast and the 8-hour standard in 2004.
21
The code color is also available from
21
The 8-hour standard is stricter in the sense that more days are designated as exceedance days. However,
most exceedance days are code orange. It may be interesting to examine how the scheme change affects
people's responses.
72
MDE. Alternatively, it can be determined by applying the rule described in Table3-1. The
latter matches the recorded one perfectly, which confirms the relationship between code
color and forecast ozone. A sharp RD rather than fuzzy RD model is therefore appropriate.
In the early 2000s, the Maryland State High Administration (SHA) started to install
detectors
22
along major roads to monitor and record traffic conditions. The detectors
count the number of vehiclesand volume is reported in five-minute intervals. For the
project, weekday traffic volumes of the years2001 through 2003 were obtained from the
University of Maryland's Center for Advanced Transportation Technology Laboratory
(CATT Lab), which archives data from the Maryland SHA.
As the detector system was establishedshortly before the period we examine, the
performance of detectors and the data transfer network was not ideal. This resulted in
considerable missing data. I restrict the set of detectors to be analyzed to those with less
than 30 percent of 5-minute intervals missing, whichgives seven detectors located on
four major interstate highways in the Baltimore metropolitan area.
23
Table3-2 provides
information about the detectors and the traffic they are monitoring. The routes where
these detectors are located all carry heavy traffic from the surrounding areas into and out
of the Baltimore urban area. These roads rank from 3rd to 8th in terms of 2003 annual
average daily traffic (AADT) in the area. Unfortunately, we do not have data for I-95 and
22
Different from those buried underneath the road surface, this type of detectors is usually mounted on
existing side-of-the-road poles and work with microwave sensor technology. See http://www.rtms-by-
eis.com/rtms_features.html for more information.
23
See Figure3-2 for a map showing major freeways of the region and locations of detectors.
73
I-695, two major routes through and around the Baltimore area, respectively. Five
detectors monitor inbound traffic while two monitor outbound traffic.
I aggregate the 5-minute volumes into four time blocks following the definitions in
the Baltimore Metropolitan Council's travel demand model (BMC 2004). They are
morning peak (6 AM – 10 AM), mid-day (10 AM – 3 PM), evening peak (3 PM – 7 PM),
and other times (7 PM – 6 AM). This aggregation largely overcomes the short-term
fluctuations in traffic flow caused by traffic conditions. More importantly, the time
blocks group together hours with homogenous traffic patterns and separate those with
different patterns. It is therefore more appropriate to study traffic pattern changes at the
time block level than at 5-minute interval or hourly levels.
It is difficult to fill in missing observations on traffic volume. In general, filling in
missing values of the dependent variable in a single-equation regression may lead to
biased estimation (Greene 2003). So time blocks with one or more missing 5-minute
interval are dropped from the regression. This may lead to an efficiency loss but should
have no effect on estimator consistency so long as the time blocks that do not enter the
volume regression are not systematically correlated with the explanatory variable of
interest, i.e. CRD. Table3-3 presents checks on the correlation between dropped time
blocks and code red days. Each column represents a probit model specification with
incremental inclusion of control variables. When CRD is the only explanatory variable
(column (1)), it seems to affect the missing pattern of all times of day except for the
morning peak period. However, sinceCRD occurs on hot, sunny days it may pick up
74
weather impacts on the detector system. When the linear forecast ozone level and a full
set of covariates including weather conditions are included (columns (2) and (3)) in the
model, the effect of CRD becomes smaller and statistically insignificant. Adding
polynomials terms in the forecast ozone level (column (4)) does not change the result at
all. Thus, we conclude that estimating equation (3-5) with only the complete time blocks
should not give us biased estimates due to missing observations.
Daily weather measuresincluding temperature (maximum and minimum), wind
speed, relative humidity (maximum and minimum) and precipitation wereobtained from
the National Climatic Data Center and are observed at the weather station located in the
Baltimore-Washington International Airport. Daily average prices for regular unleaded
gasoline in Baltimore area are provided by the GasBuddy Organization. I use the first
through seventh lags of gasoline price to control for the impact of gas price on travel
demand.
Columns (1) and (2) inTable3-4 presents the means and standard deviations of
control variables for non-code-red days and code red days, respectively. Column (3)
shows differences in means between the two types of days and the associated standard
errors. Generally speaking, a code red day is more likely to occur on hot, dry days with
lower wind speed. The observed ozone level for the day before the forecasted day is
significantly higher for the code red day. However, neither the short-term historical retail
gas price nor the day of week differ significantly, which is consistent with the fact that the
ozone forecasting model is basically a meteorological model rather than a behavioral one.
75
Although adjustments may have been made accounting for the day of the week, it seems
to be a rare unsystematic practice.
One key identification assumption mentioned earlier is that the conditional mean of
the unobservable, i.e. ( | ) E O c , is continuous at
*
O =125 ppb. The evidence in support of
this assumption can be found by testing the continuity of the observed covariates. Figure
3-3 plots the average daily characteristics including temperature (max and min),
precipitation, wind speed, humidity (max and min), retail gas price (lags), ozone
observation (lag) and Monday and Friday dummies, against the forecast of ozone level.
The predicted values from a fifth-order polynomial in the forecast level as well as the 95
percent confidence intervals are also presented. The figures suggest that there is no large,
statistically significant break for these variables when ozone levels change from non-
code-red days to code red days. Columns (4) and (5) of Table3-4 provide quantitative
support for this finding. Although code red days are different from non-code-red days, as
shown in column (3), when the comparison is narrowed between code red days and code
orange days in column (4), the difference diminishes dramatically in magnitude across all
variables and only the max temperature and observed ozone level remain statistically
significant. The higher max temperature and observed ozone level the day before most
likely reflect only the higher forecast ozone level for code red days. Column (5),
equivalent to Figure3-3, reports the estimated coefficient of CRD when a fifth-order
polynomial in forecast ozone level is included in the regression. It indicates that the
difference between code red days and non-code-red days is small and statistically
76
insignificant conditional on the forecast level. These results suggest that the unobserved
characteristics are unlikely to be discontinuous at the CRD cutoff point.
3.6 Results
Table3-5presents the estimates of the effects of CRD on traffic volumes by time of
day. Each model is estimated for a pooled sample as well as two sub-samples separating
inbound detectors from outbound detectors. Although detector fixed effectsaccount for
the unique features of traffic pattern for each location and direction, it may be true that
inbound and outbound traffic respond to CRD in distinct ways. Further, given symmetry
between morning and evening travel, i.e. the morning inbound (outbound) traffic returns
in the evening on the same routes, we should expect to see CRD have similar impacts on
morning inbound (outbound) traffic and evening outbound (inbound) traffic. The sample
is therefore split to explore the heterogeneity in theeffects of CRD on inbound and
outbound traffic. Common covariates across models include weather conditions and their
lags, the observed ozone level for the previous day, lagged retail gas prices, and dummies
for year, month, day of the week, public holidays and detectors.
24
Overall, the models
explain traffic patterns reasonably well, with an
2
R above 0.90 for the full sample and
above 0.80 for inbound and outbound sub-samples.
Columns (1)-(3) report the results of the first specification, in which the model
controls for the ozone forecast in linear form only. For the full sample, CRD decreases
morning peak traffic by 1.7 percent but increases mid-day traffic by 3.3 percent. The
24
Models including lagged traffic on previous days and seven days ago yield no different estimates.
77
average weekday morning peak and mid-day traffic volumes across monitors are about
8500 and 8000, respectively. Applying the estimates suggests that on average 145 or so
vehicle trips from 6 AM to 10 AM were cancelled or moved to other time periods. For
the mid-day hours 10 AM to 3 PM, trips rose by about 264, which could be a result of
trip rescheduling from the morning and/or people switching to driving to avoid ozone
exposure. However, it is not obvious why people would postpone their vehicle travel
closer to noon. When we look at inbound and outbound traffic separately, the CRD has
little effect on inbound traffic except for increasing mid-day volumes by 4 percent. It
lowers outbound traffic by 2.6 percent in the morning and 3.3 percent in the evening.
These results are not consistent with a symmetric traffic pattern between inbound and
outbound routes. As we discussed before, these estimates could be biased if the control
function of the forecast level is not flexible enough.
Columns (4)-(6) are the baseline regression discontinuity modelsusing a fifth order
polynomial in forecast ozone to proxy for ( | ) E O c . Column (4) shows that the CRD
reduces morning traffic by 5 percent for the pooled sample, which is equal to about 425
vehicle trips on average. In contrast to column (2), the CRD does not exhibit a
statistically significant impact on traffic during other periodsof a day. When inbound and
outbound traffic are examined separately (see columns (5) and (6)), the CRD is found to
lower the morning peakinbound traffic by 5 percent. Moreover, this reduction is matched
in the outbound traffic, which declines 2.6 percent in the evening peak and 5 percent in
other hours on code red days. The coefficients of CRD are positive for the inbound
78
sample during mid-day and evening and negative for morning outbound, but neither is
statistically significant. These resultssuggest that the code red day alert indeed reduces
traffic, albeit by a small proportion.
25
Columns (7)-(9) maintain the RD specification and exclude the code green days
from the sample, to test whether the results are driven by observations far away from the
cutoff point.
26
The main finding remains unchanged: morning inbound traffic declines by
5 percent and the evening outbound traffic declines correspondingly. The difference is
that outbound trip reductions are concentrated in the evening peak hours rather than in
other hours. Theseestimates imply that the RD approach is appropriate for measuring the
effect of the AQAD program, which a normal regression fails to capture.
Table3-6 reportsadditional tests of the robustness of the results. The expert panel
occasionally manipulatesthe ozone forecast and/or code color on Mondays and Fridays
to account for traffic patterns, but not on other days of the week. The RD strategy is
plausible if it yields similar estimates with a sample containing only Tuesday through
Thursday. Columns (1)-(3) show that morning traffic is lower by 3 percent for the pooled
sample and lower by 4 percent for the inbound sub-sample on code red days. Outbound
traffic is reduced by 3 percent, thoughthe effect is not statistically significant. Although
25
Table A3 provides full estimation results for models specified in columns (4) through (6) of Table3-5.
26
Table3-1shows that code green days account for 51 percent of sample days and code yellow days
account for 31 percent. The estimates appear to be sensitive to individual observations when code yellow
days are removed. It is more likely because of the dramatic decrease in sample size.
79
the samples diminish in size by two fifths, we still find evidence consistent with the
baseline results
Another test of the findings is to see whether similar reductions occur on other days,
Code orange days mean an air quality alert to the public, although not to the same degree
ason a code red day. Therefore, we expect no or a smaller decline in vehicle trips on
code orange days. The hypothesis is tested in two specifications. I replacethe CRD
dummy with a dummy for code orange and code red days in the first case, and use two
dummies, one for code orange days and one for code red days, in the second case.
Columns (4)-(6) of Table3-6 report the first specification and columns (7)-(9) report the
second. The dummy for code orange and red days does not have a negative impact on
traffic volumes. Instead, it increases the morning inbound traffic slightly. When models
includetwo separate dummies, the code red dummy has significant negative effects on
morning inbound volumeswhile the code orange dummy has statistically insignificant
effects sometimes opposite to code red. Both results suggest that drivers do not respond
to code orange as they do to code red days.
Code red days often occur on consecutive days. The cost of foregoing driving may
rise on the second or third code red day. For instance, it may be easier for a professional
to work from home one day a week than two or three days a week. On the other hand, if
the marginal cost of driving on code red days increases, an individual is more likely to
take some action on the second or third code red day than on the first. I estimate the
following models to see the effect of consecutive code red days. In addition to the code
80
red day dummy, I add to the model a dummy equal to one if it is the second, third or
fourth (the longest string is four).code red day in a row, or a dummy for the third or
fourth code red day. Columns (1)-(3) and (4)-(6) of Table3-7report the results for these
two cases respectively. For the pooled sample, morning traffic is lower by 6 percent on
the first code red day but only 4 percent lower for the second, third or fourth code red day.
When the sample is split between inbound and outbound detectors, however, the effect
loses statistical significance. A dummy for the third or fourth code red days in a row does
not exhibit reinforcing or offsetting effects either.
3.7 Discussion and Conclusion
The findings of this study are similar to those in Cutter and Neidell's (2007), which
examines a similar program in the San Francisco Bay Area. Both results suggest that the
voluntary information programs lead to a small reduction in vehicle trips and the effect
most is concentratedin the morning peak period. The evidence that the reduction occurs
for the morning inbound traffic and evening outbound traffic seems to provide additional
support for the main results.
The timing of the effect suggests that it is commuting trips that are reduced. The
workers who usually drive to work could potentially work at home or switch to other
travel modes such as taking public transportation, walking/biking and carpooling on code
red days. The former is especially likely since, according to the 2001 National Household
Travel Survey, 7.7 percent of workers work at home at least one day every month and 4.7
81
percent at least one day every week in the Washington-Baltimore metropolitan area.
These people should generally have more flexibility to avoid driving on code red days.
Although the program is demonstrated to have some expected impact, the magnitude
seems too small to reduce vehicle emissions dramatically. As the literature suggests
voluntary programs are unlikely to improve air quality sufficiently to bring a region into
compliance status. An innovative approach would be a permit program that restricts
driving on high ozone days unless a permit is bought for each vehicle. The program could
be effective if the permit price is set high enough, which presents a strong disincentive
for many people to drive. Imposing the control on an episodic basis means the program
could be more economically efficient than programs with year-round controls.
One objective of the programs like AQAD is to see how education and persuasion
might alter individuals' behavior in favor of the environment. This study indicates that
these efforts are not made in vain. In addition to limiting driving, the program also asks
people to refuel vehicles after dusk or on another day. It may be worth investigating
whether this is an easier behavioral change for people to make once data are available.
82
Figures and Tables for Chapter 3
Figure 3-1. Illustration of Biased Estimates with Linear Forecast Ozone Level
Note: The graphs show that the estimated CRD effect, ¸ ' , controlling for linear forecast ozone level could
underestimate (left) or overestimate (right) the true effect of CRD, ¸ , when the underlying relationship
between y and forecast ozone level is nonlinear.
0
Forecast ozone level (ppb)
125
¸
¸ '
125
¸
¸ '
Forecast ozone level (ppb)
0
y y
83
Figure 3-2. Map of Baltimore Region Major Freeways and Maryland SHA’s Traffic
Detectors
Source: Created with ArcMap using Census 2000 TIGER/Line®Shapefiles and Bureau of Trasportation
Statistics’ Highway Performance Monitoring System data. Locations of detectors are not accurate but
illustrative.
84
Figure 3-3. Similarity of Covariates around Code Red Day Cutoff Point
5
0
6
0
7
0
8
0
9
0
1
0
0
D
a
i
l
y
m
a
x
i
m
u
m
t
e
m
p
e
r
a
t
u
r
e
40 60 80 100 125140
Forecate ozone level
4
0
5
0
6
0
7
0
8
0
D
a
i
l
y
m
i
n
i
m
u
m
t
e
m
p
e
r
a
t
u
r
e
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
2
0
0
2
5
0
D
a
i
l
y
p
r
e
c
i
p
i
t
a
t
i
o
n
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
2
0
0
A
v
e
r
a
g
e
d
a
i
l
y
w
i
n
d
s
p
e
e
d
40 60 80 100 125140
Forecate ozone level
8
0
8
5
9
0
9
5
1
0
0
1
0
5
M
a
x
i
m
u
m
r
e
l
a
t
i
v
e
h
u
m
i
d
i
t
y
40 60 80 100 125140
Forecate ozone level
2
0
4
0
6
0
8
0
1
0
0
M
i
n
i
m
u
m
r
e
l
a
t
i
v
e
h
u
m
i
d
i
t
y
40 60 80 100 125140
Forecate ozone level
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
1
d
a
y
a
g
o
40 60 80 100 125140
Forecate ozone level
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
3
d
a
y
s
a
g
o
40 60 80 100 125140
Forecate ozone level
1
.
2
1
.
3
1
.
4
1
.
5
1
.
6
1
.
7
R
e
t
a
i
l
g
a
s
p
r
i
c
e
7
d
a
y
s
a
g
o
40 60 80 100 125140
Forecate ozone level
0
5
0
1
0
0
1
5
0
l
a
g
o
f
o
b
s
e
r
v
e
d
o
z
o
n
e
40 60 80 100 125140
Forecate ozone level
-
.
5
0
.
5
1
M
o
n
d
a
y
40 60 80 100 125140
Forecate ozone level
-
.
5
0
.
5
1
1
.
5
F
r
i
d
a
y
40 60 80 100 125140
Forecate ozone level
Note: The dots represent the average daily characteristics for each forecast ozone level. The continuous line
is the predicted values from a fifth-order polynomial in forecast level with the dashed lines for 95 percent
confidence interval.
85
Table 3-1. One-hour Ozone Level and Code Colors in Baltimore Area
1-hour Ozone
(ppb)
Code Color Health Concern Number of Days
2001-2003
0 – 79 Green Good 214
80 – 104 Yellow Moderate 128
105 – 124 Orange Unhealthy for sensitive groups 44
>125 Red Unhealthy 28
Source: http://www.mwcog.org/environment/air/downloads/calendar_2003.pdf, MWCOG.
86
Table 3-2. Description of Detectors and Traffic in Baltimore Area
Detector Highway AADT 2003 Direction In/Out Bound Missing %
1 I-83 86,293 (4
th
) S In 9.7
2 I-83 86,293 (4
th
) S In 26.3
3 I-83 86,293 (4
th
) N Out 18.5
4 I-795 81,500 (5
th
) S In 17.8
5 I-97 105,008 (3
rd
) N In 17.3
6 I-97 105,008 (3
rd
) S Out 9.4
7 I-70 44,142 (8
th
) E In 10.4
Note: The third column presents annual average daily traffic in 2003. The top two roads missing
here are I-95 with AADT of 169,534 and I-695 with AADT of 167,473.
87
Table 3-3. Correlation between Missing Time Block and Code Red Day
(1) (2) (3) (4)
Morning peak 0.272 0.325 0.219 0.453
(0.249) (0.221) (0.148) (0.350)
0.002 0.003 0.117 0.120
Mid-day 0.445 -0.068 -0.172 0.025
(0.160) (0.157) (0.167) (0.244)
0.006 0.021 0.154 0.155
Evening peak 0.665 -0.064 -0.043 0.137
(0.125) (0.138) (0.145) (0.206)
0.014 0.044 0.179 0.183
Other 0.349 0.300 0.142 0.218
(0.179) (0.174) (0.176) (0.237)
0.004 0.004 0.088 0.089
N 2898 2884 2849 2849
Linear forecast ozone level N Y Y Y
2
nd
to 5
th
order polynomials
of the forecast ozone N N N Y
Control variables N N Y Y
Note: Dependent variable is a binary indicator equal to 1 if the time block is to be
dropped from traffic volume equation. The first row of each time-of-day panel is
probit estimates of the coefficient of CRD, the second row is standard error
clustered on each week, and the third row is the pseudo-R squared. Control
variables include weather variables and their lags, observed ozone levels, lags of
daily gas price, and year, month, day-of-week, holiday, and monitor dummies,.
88
Table 3-4. Summary Statistics and Difference in Selected Covariates Between Code Red
Days and Other Days
Non-CRD CRD
CRD vs.
Non-CRD
Orange
vs. Red Polynomials
(1) (2) (3) (4) (5)
Max temperature 80.903 94.462 13.558 2.873 1.051
8.771 3.313 (0.842) (0.875) (1.935)
Min temperature 60.824 69.538 8.714 2.480 1.630
8.959 4.282 (1.002) (1.347) (2.826)
Precipitation 14.896 1.000 -13.896 3.412 -4.202
39.252 2.966 (2.463) (2.533) (9.279)
Wind speed 62.654 52.654 -10 -5.434 -0.375
25.356 12.103 (2.833) (3.385) (7.139)
Min relative humidity 51.639 40.846 -10.793 -1.624 3.273
16.004 8.698 (1.965) (2.246) (4.974)
Max relativehumidity 94.457 90.808 -3.650 -0.928 3.329
6.794 6.487 (1.338) (1.682) (3.025)
Gas price 1 day ago 1.462 1.449 -0.013 0.021 0.048
0.121 0.116 (0.024) (0.034) (0.063)
Gas price 3 days ago 1.464 1.45 -0.014 0.016 0.041
0.121 0.107 (0.022) (0.032) (0.063)
Gas price 7 days ago 1.465 1.445 -0.020 0.002 0.036
0.118 0.121 (0.025) (0.033) (0.066)
Ozone (lag) 75.526 123.654 48.128 17.595 -7.890
23.293 20.829 (4.327) (5.607) (12.863)
Monday 0.197 0.192 -0.005 -0.102 0.063
0.398 0.402 (0.082) (0.112) (0.245)
Friday 0.208 0.154 -0.054 0.007 0.200
0.407 0.368 (0.076) (0.095) (0.154)
N 269 26 295 60 295
Note: Columns (1) and (2) are means and standard deviations (underneath) for non-code red days and code
red days, respectively. Column (3) is the difference between CRD and non-CRD. Column (4) is the
difference between CRD and code orange days. Column (5) is the estimate of CRD coefficient regressing
each covariate on CRD and a fifth order polynomial in forecast ozone level. In the parentheses are standard
errors. Those standard errors in the column (5) account for within-week clustering.
89
Table 3-5. Impact of Code Red Day Announcement on Traffic Volumes by Time of Day
All Inbound Outbound All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Morning Peak
Coefficient -0.017** -0.012 -0.026** -0.051* -0.051** -0.034 -0.063* -0.052* 0.008
Std. errors (0.007) (0.009) (0.011) (0.030) (0.024) (0.046) (0.032) (0.027) (0.026)
N 1520 1059 461 1520 1059 461 736 517 219
2
R
0.96 0.97 0.96 0.96 0.97 0.96 0.97 0.98 0.97
Mid-day
Coefficient 0.033* 0.039* 0.018 0.026 0.045 -0.006 -0.084* -0.092 -0.02
Std. errors (0.018) (0.020) (0.021) (0.043) (0.057) (0.030) (0.049) (0.056) (0.042)
N 1119 795 324 1119 795 324 485 360 125
2
R
0.93 0.94 0.87 0.93 0.94 0.87 0.96 0.96 0.93
Evening Peak
Coefficient -0.007 0.004 -0.033* 0.03 0.07 -0.026* 0.011 0.068 -0.069**
Std. errors (0.02) (0.026) (0.017) (0.059) (0.093) (0.014) (0.102) (0.164) (0.029)
N 1157 787 370 1157 787 370 461 316 145
2
R
0.90 0.86 0.88 0.90 0.86 0.88 0.91 0.89 0.92
Other
Coefficient 0.006 0.012 -0.013 -0.019 -0.01 -0.050* 0.009 0.018 -0.028
Std. errors (0.012) (0.011) (0.018) (0.019) (0.018) (0.027) (0.027) (0.026) (0.049)
N 1201 839 362 1201 839 362 566 398 168
2
R
0.95 0.96 0.83 0.95 0.96 0.83 0.97 0.98 0.88
Note: Dependent variable is log of traffic volumes. Control variables include weather conditions and their lags, forecast 1-hour ozone concentration,
observed ozone level for the day before, lagged retail gas prices, and dummies for year, month, day of the week, public holiday and monitor. Columns (1)-
(3) control for linear ozone forecast only. Columns (4)-(9) control for a fifth-order polynomial in ozone forecast. Columns (7)-(9) focus on a sub-sample
excluding code green days. Standard errors account for within-week clustering. * indicates significance at 10 percent level while ** at 5 percent level.
90
Table 3-6. Robustness Check
All Inbound Outbound All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Morning Peak
Code orange or both 0.019 0.036* -0.003 0.005 0.025 -0.016
(0.02) (0.021) (0.025) (0.014) (0.018) (0.019)
Code red day -0.028* -0.037* -0.018 -0.049* -0.038** -0.042
(0.015) (0.019) (0.026) (0.026) (0.019) (0.043)
N 903 627 276 1520 1059 461 1520 1059 461
2
R
0.98 0.98 0.99 0.96 0.97 0.96 0.96 0.97 0.96
Mid-day
Code orange or both -0.017 -0.025 -0.003 -0.011 -0.013 -0.006
(0.027) (0.035) (0.028) (0.030) (0.037) (0.033)
Code red day -0.001 0.006 0.002 0.02 0.038 -0.009
(0.059) (0.080) (0.040) (0.047) (0.063) (0.035)
N 669 473 196 1119 795 324 1119 795 324
2
R
0.94 0.94 0.90 0.93 0.94 0.87 0.93 0.94 0.87
Evening Peak
Code orange or both -0.024 -0.053 0.019 -0.018 -0.039 0.012
(0.032) (0.044) (0.016) (0.040) (0.057) (0.017)
Code red day 0.053 0.123 -0.032 0.019 0.047 -0.019
(0.114) (0.188) (0.020) (0.070) (0.110) (0.014)
N 679 459 220 1157 787 370 1157 787 370
2
R
0.92 0.90 0.60 0.90 0.86 0.88 0.90 0.86 0.88
Other
Code orange or both 0.016 0.02 0.021 0.013 0.02 0.007
(0.014) (0.014) (0.022) (0.014) (0.015) (0.025)
Code red day 0.030 0.023 0.029 -0.012 0 -0.046
(0.028) (0.029) (0.046) (0.020) (0.019) (0.032)
N 718 506 212 1201 839 362 1201 839 362
2
R
0.96 0.97 0.64 0.95 0.96 0.83 0.95 0.96 0.83
Note: Dependent variable is log of traffic volumes. Columns (1)-(3) use samples excluding Monday and Friday. Columns (4)-(6) replace the CRD dummy with a
code-red-or-orange dummy. Columns (7)-(9) add a code-orange dummy in addition to the CRD dummy. Standard errors account for within-week clustering. *
indicates significance at 10 percent level while ** at 5 percent level.
91
Table 3-7. Impact of CRDs in Sequence on Traffic Volumes by Time of Day
All Inbound Outbound All Inbound Outbound
(1) (2) (3) (4) (5) (6)
Morning Peak
Code red days -0.057* -0.058** -0.034 -0.054* -0.051** -0.035
(0.030) (0.024) (0.047) (0.028) (0.022) (0.047)
CRDs in Seq. 0.013* 0.015 0.001 0.006 -0.001 0.002
(0.007) (0.010) (0.009) (0.007) (0.010) (0.010)
N 1520 1059 461 1520 1059 461
2
R
0.96 0.97 0.96 0.96 0.97 0.96
Mid-day
Code red days 0.020 0.040 -0.009 0.011 0.029 -0.009
(0.042) (0.056) (0.028) (0.039) (0.053) (0.026)
CRDs in Seq. 0.015 0.015 0.006 0.043 0.049 0.007
(0.035) (0.044) (0.016) (0.029) (0.034) (0.028)
N 1119 795 324 1119 795 324
2
R
0.93 0.94 0.87 0.93 0.94 0.87
Evening Peak
Code red days 0.020 0.056 -0.029 0.045 0.094 -0.033**
(0.047) (0.072) (0.019) (0.062) (0.098) (0.016)
CRDs in Seq. 0.017 0.024 0.005 -0.035 -0.054 0.017
(0.030) (0.047) (0.021) (0.023) (0.033) (0.021)
N 1157 787 370 1157 787 370
2
R
0.90 0.86 0.88 0.90 0.86 0.88
Other
Code red days -0.011 -0.002 -0.043 -0.016 -0.006 -0.045*
(0.023) (0.024) (0.029) (0.022) (0.023) (0.027)
CRDs in Seq. -0.017 -0.021 -0.014 -0.008 -0.011 -0.014
(0.020) (0.019) (0.025) (0.023) (0.025) (0.025)
N 1201 839 362 1201 839 362
2
R
0.95 0.96 0.83 0.95 0.96 0.83
Note: Dependent variable is log of traffic volumes. Columns (1)-(3) has one additional dummy equal
to one if the code red day is the second, third or fourth one in a row. Columns (4)-(6) has a dummy
equal to one if the code red day is the third or fourth one in a row. Standard errors account for within-
week clustering. * indicates significance at 10 percent level while ** at 5 percent level.
92
4 Concluding Comments
4.1 Summary of Results
Chapter 2 calculates the percent of workers who use the Internet when
working at home in a person’s two-digit occupation by city size cell to instrument for
telecommuting choice. After controlling for occupation and city fixed effects, as well
as individual and household characteristics, this variable still predicts men and
married women’s probability to telecommute: a 10 percentage point increase in the
internet penetration causes telecommuting probability to rise by 5 percentage points
for married women and 3 percentage points for men.
Using this variable to instrument for telecommuting choice yields IV
estimates that telecommuting leads a married women’s one-way commute time to
increase by 9 to 12 minutes. The effect on men’s commute length is smaller, at about
5 minutes and statistically indistinguishable from zero. The results are robust for
different specifications and sub-samples. Contrary to the OLS estimates, IV
estimation finds that probability of commuting by driving does not decline due to
telecommuting. For the average married female worker who commutes 24 minutes
one way five days a week, telecommuting lowers weekly total commute time from
240 minutes to 198 minutes if the woman telecommutes two days a week. This means
a 17 percent reduction, less proportional to the reduction in commuting frequency.
Chapter 3 finds that the code red day announcement results in a 4-5 percent
reduction in vehicle commute trips in morning peak hours. However, this estimate is
obtained only when we include in the regression a flexible function of forecast ozone
levels, which is designed to control for the correlation between the code red day
93
indicator and any non-random unobservables. If only the linear term of forecast ozone
level is controlled in the regression, the estimate is as small as 1.7 percent. The
difference highlights the importance of the identification strategy used.
The conclusion from these results is that the two TDM strategies work to
some degree. Telecommuting is not shown to have a rebound effect on men’s
commute length. For married women, the effect seems moderate enough to result in a
net reduction in commute miles. Consistent with findings in northern California,
information about bad air quality could induce a small proportion of people to refrain
from driving. It is more likely that these people will work from home rather than to
switch to another travel mode. The effect, however, is not large enough to cause air
quality improvements.
4.2 Directions for Future Research
Some questions related this dissertation remain unsolved. The instrumental
variable developed in Chapter 2 does not have much power in explaining single
women’s telecommuting choices. Thus, the analysis cannot provide information
about the responses of single females to telecommuting.
Commuting length reflects both residential location choice and work location
choice. Telecommuters who choose longer one-way commute distances could choose
to live farther from work or to work farther from home. It is important to distinguish
the two possibilities from a public policy perspective.
For a two-earner household, housing location is determined jointly by both
husband’s and wife’s employment locations. A change in one person’s commuting
cost might lead to changes in the commute lengths of both people. In this case, the
94
sum of commute lengths of the household would be be the variable of interest. More
research is needed to understand the impact of household members’ telecommuting
on total household commute length.
A natural extension of Chapter 3 is to apply the same technique to similar
programs that have free bus fares. Free bus fares decrease the cost of riding a bus.
However, for people who are used to driving, a larger share of the cost of switching to
transit is the time and inconvenience to get on a bus. Moreover, it is of interest to
know whether such a program passesthe cost-benefit test. The voluntary program
takes an episodic approach to controlling ozone, which is valuable in designing a
pricing control scheme. Since ozone episodes occur only on hot, sunny days, the
government could set a price for driving on those days. Daily automobile travel and
resultant emissions could be managed by choosing a permit price. The cost-
effectiveness of such a program if implemented in the Washington metropolitan area
is being evaluated in an ongoing project.
Finally, economists may not want to give up the idea of managing travel
demand via non-pricing strategies. Many TDM strategies may be effective in various
contexts and even cost-effective if the political costs of pricing strategies are taken
into account. Clearly, economists should participate in the design and evaluation of
the TDM programs.
95
Appendices
Appendix 1 A Monocentric City Model with Commuters and Telecommuters
In a closed city, each household has only one worker and all employment
concentrates in the central business district (CBD). Workers commute to work at the
CBD along a radial network. Commuting costs per mile traveled are e , so a worker
who lives d miles from the CBD spends 2ed on daily commuting. All workers earn
the same income y per day. Household utility is described by a strictly quasi-
concave function ( , ) u c h , where c represents consumption of a composite non-
housing good and h is consumption of housing that could be measured in square feet
of floor space or number of rooms. The price of the composite good is assumed to be
the same across different locations of the city and normalized to 1. The daily rental
price of a unit of housing, denoted p , depends on location.
Initially, suppose all workers are identical. They maximize household utility to
reach a constant level, u . That is
{ , }
max ( , )
c h
u c h u = (A1)
s.t. 2 c ph ed y + + = .
Substitute 2 c y ph ed = ÷ ÷ into Eq. (A1) and notice that equilibrium housing price
and consumption are both functions of distance to the CBD, i.e. d . We have
( ) ( ) ( ) 2 , ( ) u y p d h d ed h d u ÷ ÷ = . (A2)
96
Totally differentiating Eq. (A2) and applying the envelop theorem, we get the well-
known conditions on the market equilibrium rent gradient that,
( ) 2
'( )
( )
p d e
p d
d h d
c
= = ÷
c
, (A3)
and
2
2 2
( ) 2 '( )
"( )
( )
p d eh d
p d
d h d
c
= =
c
. (A4)
Eqs. (A3) and (A4) imply that the housing price declines with commuting distance
and the rent gradient gets flatter as distance increases since '( ) 0 h d > . Plotted on a
plane with distance to the CBD as the x-axis and rent as the y-axis, the rent curve is a
downward-sloping convex function. Intuitively, workers who live in the suburbs with
longer commute are compensated by cheaper and larger homes.
Now, extend the model to including two types of otherwise identical workers:
commuters (c ) and telecommuters (tc ). Because the latter commute less often than
the former, the average daily commuting costs are lower for telecommuters than for
commuters. Therefore, there are separate rent offer curves for the two types of
workers, respectively. They are characterized as
2
( )
( )
i
i
i
e
p d
h d
'
= ÷
where , i c tc = . Assuming that housing is a normal good, then ( ) ( )
c tc
h d h d < .
Together with
c tc
e e > , we have
( ) ( )
c tc
p d p d
' '
> .
97
The rent offer of telecommuters declines slower than that of commuters. Figure A1
illustrates the two rent offer curves and the market rent gradient in equilibrium. The
telecommuters' rent offer curve (CD) is flatter than commuters' rent offer curve (AB)
while the two intersect at a certain distance
o
d d = . Commuters outbid telecommuters
for housing at locations closer to the CBD (
o
d d < ), as segment AO lies above CO,
and vice versa for locations beyond
o
d . The market equilibrium rent gradient is the
upper segments of the two offer curves (AO and OD). This means in equilibrium
commuters occupy the entire ring-shaped region around the CBD from distance 0 to
o
d while telecommuters sort into the surrounding ring from
o
d to
*
d , the city edge
determined by exogenous farmland rent. Thus, telecommuters have longer commutes
than commuters.
98
Appendix 2 Imputation of Top-Coded Commuting Time in the PUMS
First, I estimate a Pareto distribution to approximate the right-hand tail of the
commute time distribution, i.e.
( 1)
( )
a a
f x ab x
÷ +
= , for b x s s ·
where a is the parameter of the distribution, b is a constant from which commuting
time is assumed to follow a Pareto distribution, x is observed individual commuting
time equal to or greater than b . To obtain an estimate for a , I estimate Pr( ) x t > ,
where t is the top-coded value, i.e. 99 in PUMS, by the fraction of people commuting
b or more minutes who are top-coded, and exploit the relationship ( ) Pr( ) /
a
x t b t > = .
Therefore,
( )
t b
t x
a
ln ln
) r( P
ˆ
ln
ˆ
÷
>
= .
Then, the top-coded values are replaced by the estimated conditional expectation
of commuting time,
1
( | ) ( 1) E x x t ta a
÷
> = ÷
? ?
where t is the top-coded value, i.e. 99 in PUMS. Thus, different values for b yield
different estimates of a and the imputing value for top-coded observations.
For instance, let 50 b = , then 378,211 observations have commuting time equal
to or above 50 minutes, 16.4 percent of which are top-coded observations. Thus,
( ) ln(0.164)/ ln(50) ln(99) 2.65 a = ÷ =
?
. The conditional expectation for top-coded
individuals equals 159.1. When b is varied from 40 to 90 in increments of 10, the
conditional expectation estimates vary from 123 to 165 with an average of 150.
99
Figure A1. Bid Rent Curves in a Monocentric City with Telecommuters and
Commuters
p(d)
d
A
B
C
D
O
d* do 0
100
Table A1. CPS and NHTS Sample Construction
May 2001 CPS 2001 NHTS
Original sample 131,997 160,758
15 or older, employed with information on
working at home
50,743 65,697
Not self-employed 45,217
Reasonable commute distance and speed 62,283
MSA residents 32,272 50,810
Final sample without missing values on
any covariates
29,147 47,730
Note: Reasonable commute distance refers to one-way commute time below 180 minutes and
commute distance below 180 miles; reasonable speed refers to speed between 0.01 mile per
minute and 1.5 miles per minute.
101
Table A2. Distributions by Commute Mode across Cities
Commute mode Driving Rail Bus
Nationwide cities with 1 million or more
population
0.878 0.044 0.463
Atlanta, GA 0.964 0.002 0.013
Baltimore, MD 0.883 0.032 0.053
Chicago-Gary-Kenosha, IL-IN-WI 0.819 0.112 0.034
San Francisco-Oakland-San J ose, CA 0.822 0.041 0.071
Washington DC-VA-MD-WV 0.831 0.072 0.064
Source: Author’s calculation using NHTS 2001. Commute mode is defined as transportation mode to
work last week covering most of the distance.
102
Table A3. Full Results of Regression Discontinuity Models (Code red day coefficients correspondto Columns (4)-(6) inTable 3-5)
All Inbound Outbound
Morning Mid-day Evening Other Morning Mid-day Evening Other Morning Mid-day Evening Other
Ozone forecast -0.016 -0.009 0.083 -0.036 -0.032 -0.013 0.151 -0.037 0.020 0.030 -0.032 -0.030
(0.052) (0.058) (0.064) (0.027) (0.049) (0.079) (0.091) (0.027) (0.067) (0.033) (0.037) (0.041)
2
nd
order forecast 0.001 0.000 -0.002 0.001 0.001 0.001 -0.004 0.001 -0.000 -0.001 0.001 0.001
(0.001) (0.002) (0.002) (0.001) (0.001) (0.002) (0.002) (0.001) (0.002) (0.001) (0.001) (0.001)
3
rd
order forecast -0.000 -0.000 0.000 -0.000 -0.000 -0.000 0.000 -0.000 0.000 0.000 -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
4
th
order forecast 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 -0.000 -0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
5
th
order forecast -0.000 -0.000 0.000 -0.000 -0.000 -0.000 0.000 -0.000 0.000 0.000 -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag observed ozone -0.000** 0.000 0.001** 0.000 -0.000* 0.000 0.001* 0.000 -0.000 0.000 0.000 0.000
(0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000)
Avg. wind speed 0.000 -0.000 0.000 -0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Min. humidity 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000
(0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000)
Max. humidity 0.001* 0.000 0.001 0.001 0.001 0.000 0.001 0.000 0.001 -0.000 -0.000 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Max. temperature -0.001 -0.003** -
0.005***
-0.001 -0.001* -
0.005***
-0.007** -0.001 0.000 0.001 -0.000 -0.001
(0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.003) (0.001) (0.001) (0.001) (0.001) (0.001)
Min temperature -0.002** -0.001 -0.000 -0.001 -0.002* -0.001 -0.001 -0.001 -0.001 -0.000 0.000 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Precipitation -
0.000***
0.000 0.000 -0.000 -
0.000***
0.000 0.000 -0.000 -
0.000***
-0.000 -0.000** -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag avg. wind spd. 0.000 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 0.000 0.000* -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag min. humidity -0.001** -0.000 -0.001 -0.001 -0.001** -0.001 -0.001 -0.000 -0.000 0.000 -0.000 -0.001
(0.000) (0.000) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.001) (0.000) (0.000) (0.000)
Lag max. humidity 0.001* 0.000 0.000 0.001 0.001* 0.001 0.000 0.001 0.001 0.000 0.001 -0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.000) (0.001)
103
Lag max. temp. -0.000 -0.001 -0.005** -0.001 -0.000 -0.001 -0.006** -0.001 0.000 0.001 -0.002 -0.001
(0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.003) (0.001) (0.001) (0.001) (0.002) (0.002)
Lag min. temp. 0.000 0.002** 0.004*** 0.001 0.001 0.003** 0.006*** 0.001* -0.000 -0.001 -0.000 0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)
Lag precipitation -0.000 -0.000 0.000 0.000 -0.000** 0.000 0.000 -0.000 -0.000 -0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Lag gas price -0.078 -0.109 0.108 -0.008 -0.081 -0.155 0.115 -0.004 -0.038 -0.074 0.024 -0.063
(0.052) (0.099) (0.180) (0.086) (0.050) (0.124) (0.257) (0.088) (0.087) (0.109) (0.058) (0.101)
2
nd
lag gas price 0.030 0.081 -0.071 0.160 0.059 0.091 -0.017 0.172 -0.040 0.119 -0.051 0.150
(0.066) (0.108) (0.162) (0.118) (0.060) (0.140) (0.230) (0.116) (0.131) (0.074) (0.046) (0.131)
3
rd
lag gas price -0.012 -0.126 -0.202 -0.087 -0.014 -0.151 -0.373 -0.103 0.103 -0.039 -0.081 0.003
(0.100) (0.122) (0.168) (0.122) (0.082) (0.155) (0.235) (0.123) (0.109) (0.091) (0.097) (0.123)
4
th
lag gas price 0.073 0.109 -0.034 -0.005 -0.036 0.206 0.005 -0.011 0.088 -0.046 -0.007 0.001
(0.116) (0.118) (0.150) (0.116) (0.085) (0.160) (0.213) (0.123) (0.092) (0.121) (0.118) (0.151)
5
th
lag gas price -0.173 -0.194* 0.153 -0.082 -0.172 -0.203 0.212 -0.062 -0.108 -0.235* 0.089 -0.181
(0.128) (0.114) (0.150) (0.116) (0.145) (0.128) (0.238) (0.113) (0.125) (0.135) (0.098) (0.171)
6
th
lag gas price 0.052 0.033 0.251 0.111 0.065 -0.095 0.340 0.099 -0.015 0.220 -0.001 0.189
(0.124) (0.138) (0.204) (0.138) (0.144) (0.156) (0.277) (0.140) (0.100) (0.261) (0.087) (0.195)
7
th
lag gas price 0.081 0.190* -0.108 -0.076 0.152* 0.299** -0.152 -0.082 -0.015 0.019 0.041 -0.087
(0.065) (0.097) (0.140) (0.065) (0.078) (0.126) (0.181) (0.087) (0.074) (0.171) (0.085) (0.089)
Year 2002 0.012 0.020** 0.008 0.019 0.011 0.013 -0.001 0.009 0.017* 0.045*** 0.027** 0.036**
(0.009) (0.009) (0.013) (0.014) (0.012) (0.012) (0.018) (0.014) (0.009) (0.011) (0.013) (0.015)
Year 2003 0.052*** 0.055*** 0.016 0.052*** 0.048*** 0.053*** 0.010 0.046*** 0.055*** 0.066*** 0.039*** 0.060***
(0.007) (0.009) (0.012) (0.011) (0.008) (0.012) (0.019) (0.011) (0.008) (0.008) (0.008) (0.013)
J une 0.025** -0.016 -
0.072***
0.029** 0.014 -0.033* -
0.107***
0.031** 0.034** 0.022 0.001 0.030**
(0.011) (0.016) (0.024) (0.012) (0.011) (0.019) (0.039) (0.012) (0.016) (0.021) (0.008) (0.015)
J uly 0.022* 0.027** -0.008 0.042*** 0.001 0.012 -0.014 0.040*** 0.050*** 0.056*** 0.001 0.049***
(0.012) (0.013) (0.020) (0.013) (0.012) (0.017) (0.029) (0.014) (0.018) (0.020) (0.011) (0.015)
August 0.029** 0.037*** 0.015 0.047*** 0.014 0.024 0.016 0.045*** 0.045** 0.054** 0.019 0.052***
(0.014) (0.014) (0.020) (0.016) (0.015) (0.018) (0.029) (0.015) (0.019) (0.023) (0.012) (0.018)
September 0.004 -0.026** 0.009 -0.031* 0.013 -0.037** 0.012 -0.024 -0.019 -0.017 0.010 -0.041*
(0.010) (0.012) (0.019) (0.018) (0.013) (0.016) (0.030) (0.017) (0.016) (0.013) (0.009) (0.021)
Tuesday 0.019*** -0.007 0.003 0.034*** 0.009* -0.004 0.004 0.024*** 0.033*** -0.013 0.012 0.055***
(0.006) (0.008) (0.015) (0.007) (0.005) (0.011) (0.019) (0.007) (0.010) (0.009) (0.009) (0.012)
104
Wednesday 0.030*** 0.002 0.011 0.081*** 0.015** 0.001 0.004 0.071*** 0.055*** 0.007 0.032*** 0.107***
(0.008) (0.009) (0.016) (0.009) (0.006) (0.010) (0.022) (0.010) (0.014) (0.009) (0.007) (0.011)
Thursday 0.049*** 0.020* 0.035** 0.131*** 0.029*** 0.007 0.029 0.109*** 0.083*** 0.055*** 0.056*** 0.179***
(0.008) (0.011) (0.014) (0.009) (0.006) (0.013) (0.019) (0.009) (0.014) (0.009) (0.007) (0.012)
Friday 0.033*** 0.139*** 0.099*** 0.220*** -0.006 0.116*** 0.113*** 0.184*** 0.111*** 0.201*** 0.075*** 0.295***
(0.008) (0.013) (0.016) (0.012) (0.008) (0.013) (0.022) (0.012) (0.014) (0.014) (0.008) (0.014)
Public holiday -
1.202***
-
0.097***
-
0.375***
-
0.119***
-
1.365***
-0.084** -
0.237***
-0.090** -
0.896***
-
0.120***
-
0.652***
-
0.183***
(0.054) (0.034) (0.042) (0.034) (0.043) (0.037) (0.059) (0.038) (0.080) (0.035) (0.017) (0.030)
Detector 2 1.480*** 0.475*** 0.101*** 0.422*** 1.479*** 0.463*** 0.096** 0.424***
(0.019) (0.023) (0.036) (0.013) (0.020) (0.025) (0.036) (0.013)
Detector 3 0.451*** 0.599*** 0.983*** 0.494*** -
0.801***
0.078*** -
0.019***
(0.014) (0.018) (0.030) (0.011) (0.010) (0.006) (0.005)
Detector 4 1.624*** 0.994*** 0.881*** 0.750*** 1.625*** 0.995*** 0.885*** 0.752***
(0.014) (0.018) (0.033) (0.009) (0.014) (0.018) (0.034) (0.010)
Detector 5 1.427*** 0.965*** 0.859*** 0.633*** 1.427*** 0.968*** 0.867*** 0.635***
(0.013) (0.015) (0.031) (0.012) (0.013) (0.015) (0.032) (0.013)
Detector 6 1.250*** 0.841*** 0.931*** 0.514*** 0.213***
(0.014) (0.019) (0.033) (0.011) (0.008)
Detector 7 1.114*** 0.160*** -0.045 -
0.049***
1.114*** 0.158*** -0.045 -
0.047***
(0.018) (0.017) (0.029) (0.012) (0.018) (0.017) (0.029) (0.012)
Code red days -0.051* 0.026 0.030 -0.019 -0.051** 0.045 0.070 -0.010 -0.034 -0.006 -0.026* -0.050*
(0.030) (0.043) (0.059) (0.019) (0.024) (0.057) (0.093) (0.018) (0.046) (0.030) (0.014) (0.027)
Constant 8.050*** 8.487*** 7.485*** 8.907*** 8.347*** 8.643*** 6.551*** 9.008*** 8.620*** 8.425*** 9.807*** 9.205***
(0.702) (0.804) (0.891) (0.427) (0.675) (1.091) (1.242) (0.418) (0.903) (0.455) (0.553) (0.630)
Observations 1520 1119 1157 1201 1059 795 787 839 461 324 370 362
R-squared 0.96 0.93 0.90 0.95 0.97 0.94 0.86 0.96 0.96 0.87 0.88 0.83
Standard errors in the parenthesis account for within-week clustering.* indicates significance at 10%; ** significant at 5%; *** significant at 1%
105
Bibliography
[1] Angrist, J oshua D., and Alan B. Krueger, 1992, “The Effect of Age at School
Entry on Educational Attainment: An Application of Instrumental Variables
with Moments from Two Samples”, Journal of the American Statistical
Association 87, 328-336.
[2] Bloom, H. S., et al., 2004, “Baltimore Region Travel Demand Model for Base
Year 2000”, Task Report 04-01, Baltimore Metropolitan Council.
[3] Bell, M. L., et al., 2004, “Ozone and Short-term Mortality in 95 US Urban
Communities, 1987-2000”, Journal of the American Medical Association,
292, 2372-2378.
[4] Brueckner, J an K., 2001, “Urban Sprawl: Lessons from Urban Economics”, in
William G. Gale, and J anet Rothenberg Pack, eds.: Brookings-Wharton
Papers on Urban Affairs, Brookings Institution Press.
[5] Crane, Randall, 1996, “The Influence of Uncertain J ob Location on Urban
Form and the J ourney to Work”, Journal of Urban Economics 39, 342-356.
[6] Cutter, W. Bowman and Matthew Neidell, 2007, “Voluntary Information
Programs and Environmental Regulation: Evidence from 'Spare the Air'”,
Working paper.
[7] Cummings, R. G. and M. B. Walker, 2000, “Measuring the Effectiveness of
Voluntary Emission Reduction Programmes”, Applied Economics, 32, 1719-
1726.
[8] Dee, Thomas S., and William N. Evans, 2003, “Teen Drinking and
Educational Attainment: Evidence from Two-Sample Instrumental Variables
Estimates”, Journal of Labor Economics 21, 178-209.
[9] Drucker, J oshua, and Asad J . Khattak, 2000, “Propensity to Work from Home:
Modeling Results from the 1995 Nationwide Personal Transportation
Survey”, Transportation Research Record 1706, 108-117.
[10] Ellen, Ingrid Gould, and Katherine Hempstead, 2002, “Telecommuting and
the Demand for Urban Living: A Preliminary Look at White-Collar Workers”,
Urban Studies 39, 749-66.
[11] FCC, 2002, “Third Report on the Availability of High Speed and Advanced
Telecommunications Capability”, Federal Communications Commission,
Washington D.C.
106
[12] Fox, J . C. and M. Sarkar, 2002, “An Analysis of Air Pollution-Related
Knowledge, Attitudes, and Behaviors across Time: the End of Season and
Episodic Surveys”, MWCOG.
[13] Glaeser, Edward L., and Matthew E. Kahn, 2001, “Decentralized Employment
and the Transformation of the American City”, Brookings-Wharton Papers on
Urban Affairs 1-47.
[14] Greene, W. H., 2003, Econometric Analysis, 5
th
ed. Prentice Hall, New J ersey.
[15] Grubesic, Tony H., and Alan T. Murray, 2004, “Waiting for Broadband: Local
Competition and the Spatial Distribution of Advanced Telecommunication
Services in the United States”, Growth and Change 35, 139-165.
[16] Grubesic, Tony H., and Morton E. O'Kelly, 2002, “Using Points of Presence
to Measure Accessibility to the Commercial Internet”, Professional
Geographer 54, 259-278.
[17] Guimaraes, T., and P. Dallow, 1999, “Empirically testing the benefits,
problems, and success factors for telecommuting programmes”, European
Journal of Information Systems 8, 40-54.
[18] Hahn, J ., P. Todd and W. van der Klaauw, 2001, “Identification and
Estimation of Treatment Effects with a Regression-Discontinuity Design,”
Econometrica, 69(1), 201-09.
[19] Hausman, J erry A., J . Gregory Sidak, and Hal J . Singer, 2001, “Cable
Modems and DSL: Broadband Internet Access for Residential Customers”,
American Economic Review 91, 302-307.
[20] Heckman, J ., R. LaLonde and J . Smith, 1999, “The Economics and
Econometrics of Active Labor Market Programs,” in Orley Ashenfelter and
David Card, eds., Handbook of Labor Economics, Volume 3A, Section 7.4.6.
Amsterdam: North-Holland.
[21] Henderson, Dennis K., and Patricia L. Mokhtarian, 1996, “Impacts of Center-
Based Telecommuting on Travel and Emissions: Analysis of the Puget Sound
Demonstration Project”, Transportation Research: Part D: Transport and
Environment 1, 29-45.
[22] Henry, G. T. and C. S. Gordon, 2003, “Driving Less for Better Air: Impacts of
a Public Information Campaign,” Journal of Policy Analysis and
Management, 22(1), 45-63.
[23] Hensher, D. J ., W. H. Greene, and J . M. Rose, 2005, Applied Choice Analysis:
a Primer, Cambridge University Press.
107
[24] Imbens, Guido W., and J oshua D. Angrist, 1994, “Identification and
Estimation of Local Average Treatment Effects”, Econometrica 62, 467-475.
[25] Imbens Guido W. and Thomas Lemieux, 2008, “Regression Discontinuity
Designs: A Guide to Practice”, Journal of Econometrics, 142(2), 615-635.
[26] Inoue, Atsushi, and Gary Solon, 2005, “Two-Sample Instrumental Variables
Estimators”, NBER Working Paper t0311.
[27] Kain, J ohn F., 1968, “Housing Segregation, Negro Employment, and
Metropolitan Decentralization”, The Quarterly Journal of Economics 82, 175-
197.
[28] Kitamura, Ryuichi, Patricia Mokhtarian, Ram Pendyala, and Konstadinos
Goulias, 1991, “An Evaluation of Telecommuting as a Trip Reduction
Measure”, University of California Transportation Center, CA: University of
California.
[29] Koening, Brett, Dennis Henderson, and Patricia Mokhtarian, 1996, “The
Travel and Emissions Impacts of Telecommuting for the State of California
Telecommuting Pilot Project”, Transportation Research C 4, 13-32.
[30] Krupnick, A. J ., 1988, “Temporal and Spatial Control of Ambient Ozone”,
Discussion Paper QE88-05, Resources for the Future, Washington DC.
[31] Kuzmyak, J . R. and Street Smarts, 2003, “Effectiveness of Free Bus Fares on
Forecast Air Quality Code Red Days”, Interim Report for Northern Virginia
Transportation Commission.
[32] Lee, D. S., E. Moretti, M. Butler, 2004, “Do Voters Affect or Elect Policies?
Evidence from the U.S. House Election”, Quarterly Journal of Economics,
119, 807-859.
[33] Lomax, Tim, and David Schrank, 2005, “2005 Annual Urban Mobility
Report”, Texas Transportation Institute, College Station, Texas.
[34] Lund, J . R., and P. L. Mokhtarian, 1994, “Telecommuting and Residential
Location: Theory and Implications for Commute Travel in Monocentric
Metropolis”, Transportation Research Record 1463, 10-14.
[35] Malecki, Edward J ., 2002, “The Economic Geography of the Internet's
Infrastructure”, Economic Geography 78, 399-424.
[36] Meyer, Peter B., and Anastasiya M. Osborne, 2005, “Proposed Category
System for 1960-2000 Census Occupations”, BLS Working Papers 383.
108
[37] Mokhtarian, P. L., and I. Salomon, 1996, “Modeling the Choice of
Telecommuting: 3. Identifying the Choice Set and Estimating Binary Choice
Models for Technology-Based Alternatives”, Environment and Planning A 28,
1877-94.
[38] Mokhtarian, Patricia L., Gustavo O. Collantes, and Carsten Gertz, 2004,
“Telecommuting, Residential Location, and Commute-Distance Traveled:
Evidence from State of California Employees”, Environment and Planning A
36, 1877-97.
[39] Mokhtarian, Patricia L., Ilan Salomon, and Sangho Choo, 2005, “Measuring
the Measurable: Why can't we Agree on the Number of Telecommuters in the
U.S.?”, Quality and Quantity 39, 423-452.
[40] MWCOG, 2005, Recommended Air Quality Forecast Procedures Baltimore-
Washington Region.
[41] MWCOG, 2003, “2003 Air Pollution Episodic Surveys: Report of Findings
for the Washington”, DC and Baltimore Metropolitan Areas.
[42] Parry, Ian W. H., Margaret Walls, and Winston Harrington, 2007,
“Automobile Externalities and Policies”, Journal of Economic Literature
XLV, 373-399.
[43] Plaut, Pnina O., 2006, “The Intra-Household Choices Regarding Commuting
and Housing”, Transportation Research: Part A: Policy and Practice 40, 561-
571.
[44] Popuri, Y., and C. R. Bhat, 2003, “On Modeling Choice and Frequency of
Home-Based Telecommuting”, Transportation Research Record 1858, 55-60.
[45] Pratt, J oanne H., 2000, “Asking the Right Questions about Telecommuting:
Avoiding Pitfalls in Surveying Home-based Work”, Transportation 27, 99-
116.
[46] Prystash, D. T., 1995, “Corporate attitudes and commitments to
telecommuting”, Professional Communication, IEEE Transactions on 38, 95-
99.
[47] Small, Kenneth A., 1992a, "Trip Scheduling in Urban Transportation
Analysis", American Economic Review Papers and Proceedings, 82(2), 482-
486.
[48] Small, Kenneth A., 1992b, Urban Transportation Economics, Harwood,
Philadelphia.
109
[49] Teller, A., 1967, “Air Pollution Abatement: Economic Rationality and
Reality,” Daedalus, 96, 1082-98.
[50] U.S. Environmental Protection Agency (EPA), 1993, “Automobiles and
Ozone”, 400/F-92-006 (J anuary).
[51] _______, 1997, “Episodic Emission Control Programs”, 420/F-97-022
(December).
[52] _______, 1999, “Smog: Who Does It Hurt?: What You Need To Know About
Ozone and Your Health”, EPA-452/K-99-001 (J uly).
[53] _______, 2003, “Ozone: Good Up High Bad Nearby”, 451/K-03-001 (J une).
[54] _______, 2004, “The Ozone Report: Measuring Progress through 2003”,
454/K-04-001 (April).
[55] van der Klaauw, W., 2002, “Estimating the Effect of Financial Aid Offers on
College Enrollment: A Regression-Discontinuity Approach,” International
Economic Review, 43(4), 1249-87.
[56] van Ommeren, J os, Piet Rietveld, and Peter Nijkamp, 1997, “Commuting: In
Search of J obs and Residences”, Journal of Urban Economics 42, 402-21.
[57] Walls, Margaret A., Elena A. Safirova, and Yi J iang, 2007, “What Drives
Telecommuting? The Relative Impact of Worker Demographics, Employer
Characteristics, and J ob Types”, Transportation Research Record, No. 2010,
111-120.
[58] Walls, Margaret, and Elena Safirova, 2004, “A Review of the Literature on
Telecommuting and Its Implications for Vehicle Travel and Emissions”,
Discussion Papers 04-44, Resources for the Future.
[59] Welch E., X. Gu, and L. Kramer, 2005, “The Effects of Ozone Action Day
Public advisories on Train Ridership in Chicago”, Transportation Research
Park D, 10, 445-458.
[60] White, Michelle J ., 1977, “A Model of Residential Location Choice and
Commuting by Men and Women Workers”, Journal of Regional Science 17,
41-52.
[61] White, Michelle J ., 1986, “Sex Differences in Urban Commuting Patterns”,
American Economic Review 76, 368-72.
[62] White, Michelle J ., 1988, “Location Choice and Commuting Behavior in
110
Cities with Decentralized Employment”, Journal of Urban Economics 24,
129-52.
[63] White, Michelle J ., 1988, “Urban Commuting J ourneys Are Not 'Wasteful'”,
Journal of Political Economy 96, 1097-110.
[64] Wooldridge, J . M., 2001, Econometric Analysis of Cross Section and Panel
Data, MIT press, Cambridge, MA.
doc_751225720.pdf