Description
A consumer is a person or group of people who are the final users of products and or services generated within a social system. A consumer may be a person or group, such as a household. The concept of a consumer may vary significantly by context, although a common definition is an individual who buys products or services for personal use and not for manufacture or resale.
ABSTRACT
Title of Document: DYNAMIC CONSUMER DECISION
MAKING PROCESS IN
E-COMMERCE
Wei Shi
PhD Candidate in Marketing, 2011
Directed By: Dr. Michel Wedel
PepsiCo Professor of Consumer Science
Marketing Department,
Robert H. Smith School of Business
3303 Van Munching Hall
This dissertation studies the dynamic decision making process in E-commerce. In the first
essay, we use eye tracking to investigate how consumers make information acquisition decisions
on attribute-by-product matrices in online choice environment such as comparison websites.
Hierarchical Hidden Markov Model is used to describe this process. The model consists of three
connected hierarchical layers: a lower layer that describes the eye movements, a middle layer
that identifies product- and attribute-based information acquisition modes, and an upper layer
that flexibly captures switching between these modes over time. Findings of a controlled
experiment show that low-level properties of the eye and the visual brain play an important role
in dynamic information acquisition. Consumer switch frequently between two acquisition modes,
and higher switching frequency increases decision time and reduces easiness of decision making.
These results have implications for web design and online retailing, and may open new
directions for research and theories of online choice.
The second essay investigates how usage experience with different types of decision aids
contributes to the evolution of online shopping behavior over time. In the context of online
grocery stores, we categorize four types of decision aids that are commonly available, namely,
those 1) for nutritional needs, 2) for brand preference, 3) for economic needs, and 4)
personalized shopping lists. We construct a Non-homogeneous Hidden Markov Model of
category purchase incidence and purchase quantity, in which parameters are allowed to vary over
time across hidden states as driven by usage experience with different decision aids. The dataset
was collected during the period when the retailer first launched its web business, which makes it
particularly suited to study the evolution of online purchase behavior. We estimate the model for
the spaghetti sauce and liquid detergent categories. Results indicate that four types of decisions
influence evolution of purchase behavior differently. Findings from this study enrich the
understanding of how purchase behavior may evolve over time in online stores, and provide
valuable insights for online retailers to improvement the design of their store environments.
DYNAMIC CONSUMER DECISION MAKING PROCESS IN E-COMMERCE
By
Wei Shi
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
2011
Advisory Committee:
Dr. Michel Wedel, Chair
Dr. Jie Zhang
Dr. Michael Trusov
Dr. P.K. Kannan
Dr. Ginger Zhe Jin
© Copyright by
Wei Shi
2011
ii
Acknowledgements
I am heartily thankful to my advisor, Dr. Michel Wedel, whose encouragement,
caring, guidance and support enabled me to explore the exciting new subject and
complete this dissertation. The joy and enthusiasm he has for research is contagious
and motivational for me, even during tough times in the Ph.D. pursuit. I would like to
thank Dr. Jie Zhang, who sets an example of a great researcher for her rigor and
passion on research. She has assisted, advised, and supported my research and writing
efforts over the last five years. I am also grateful for Dr. Michael Trusov, who has
generously given his time and expertise to better my work, and helped me to gain a
different perspective to the research problem. Special thanks goes to Dr. P.K. Kannan
and Dr. Ginger Zhe Jin, who graciously agreed to serve on my defense committee. I
am also in debt to Dr. Rik Pieters, whose remarkable insights and valuable
suggestions have contributed greatly to the completion of this dissertation. My
deepest gratitude goes to my parents for their unconditional love and support
throughout my life; this dissertation is simply impossible without them.
iii
Table of Contents
Acknowledgements ....................................................................................................... ii
Table of Contents ......................................................................................................... iii
List of Tables ................................................................................................................ v
List of Figures .............................................................................................................. vi
Chapter I: General Introduction .................................................................................... 1
Chapter II: Information Acquisition during Online Choice: A Model-Based
Exploration .................................................................................................................... 7
2.1. Introduction ........................................................................................................ 8
2.1.1 Modeling Information Acquisition Processes ............................................ 10
2.2 Information Acquisition through Eye movements ............................................ 12
2.2.1 Horizontal Eye Movement Patterns ........................................................... 12
2.2.2 Local Eye Movement Patterns ................................................................... 13
2.2.3 Switching between Information Acquisition Strategies ............................ 13
2.3. Model Formulation .......................................................................................... 15
2.3.1 Motivation .................................................................................................. 15
2.3.2 Specification .............................................................................................. 15
2.3.3 Model Estimation and Testing ................................................................... 18
2.4 Information Acquisition on A Comparison Website ........................................ 19
2.4.1 Experimental Procedure ............................................................................. 19
2.4.2 Eye movement Recording .......................................................................... 20
2.5 Results ............................................................................................................... 20
2.5.1 Model Comparisons ................................................................................... 20
2.5.2 General Estimation Results ........................................................................ 21
2.5.3 Switching between Information Acquisition Strategies ............................ 26
2.5.4 Attention to Products and Attributes.......................................................... 27
2.6. Discussion ........................................................................................................ 29
Chapter III: Usage Experience with Decision Aids and Evolution of Online Purchase
Behavior ...................................................................................................................... 44
3.1. Introduction ...................................................................................................... 45
3.2 Conceptual Development and Literature Review ............................................. 48
3.2.1 Online Decision Aids ................................................................................. 48
3.2.2 Evolution of Online Shopping Behavior.................................................... 50
3.2.3 Usage Experience with Decision Aids and Online Shopping Behavior
Evolution ............................................................................................................. 51
3.2.4 Potential Hidden States of Purchase Behavior ........................................... 53
3.3 Model Formulation ........................................................................................... 54
3.3.1 Type-II Tobit Model of Category Purchase Incidence and Quantity ........ 54
3.3.2 Hidden States and Transition Probabilities ................................................ 56
3.3.3 Prior Distributions and Estimation Method ............................................... 58
3.4 Empirical Analyses ........................................................................................... 59
3.4.1 Data Description ........................................................................................ 59
iv
3.4.2 Operationalization of Key Variables ......................................................... 60
3.4.3 Time-Varying Patterns of Usage Experience with Decision Aids ............. 61
3.4.4 Model Estimation Results .......................................................................... 62
3.4.5 Evolution of Purchase Behavior in the Online Store ................................. 64
3.5 Discussion ......................................................................................................... 68
Chapter IV: Conclusion .............................................................................................. 81
4.1 Summary of the Two Essays............................................................................. 81
4.2 Contribution and Managerial Implications ....................................................... 82
4.3 Future Research ................................................................................................ 86
Appendices .................................................................................................................. 88
Appendix I ............................................................................................................ 88
Appendix IIa ........................................................................................................... 89
Appendix IIb ........................................................................................................... 89
Appendix III ........................................................................................................... 90
Bibliography ............................................................................................................... 91
v
List of Tables
Table 2.1 ……………………………………………………………………... 33
Table 2.2 ……………………………………………………………………... 33
Table 3.1 ……………………………………………………………………... 71
Table 3.2 ……………………………………………………………………... 72
Table 3.3 ……………………………………………………………………... 73
Table 3.4 ……………………………………………………………………... 74
Table 3.5 ……………………………………………………………………... 75
Table 3.6 ……………………………………………………………………... 76
vi
List of Figures
Figure 2.1 ……………………………………………………………………... 34
Figure 2.2 ……………………………………………………………………... 35
Figure 2.3 ……………………………………………………………………... 36
Figure 2.4 ……………………………………………………………………... 37
Figure 3.1 ……………………………………………………………………... 77
Figure 3.2 ……………………………………………………………………... 78
Figure 3.3 ……………………………………………………………………... 79
1
Chapter I: General Introduction
This dissertation is composed of two essays in the discipline of dynamic consumer
decision making in E-commerce. The global E-commerce has experienced robust growth in
recent years, and its revenue is predicted to reach 680 billion dollars by the end of 2011, a 18.9%
increase from 2010
1
. Online shopping websites have become popular channels in which
consumers acquire product information and make purchase decisions. Shopping behavior in
online stores has been shown to be systematically different from that in offline stores (e.g.,
Danaher, Wilson, and Davis 2003, Degeratu, Rangaswamy, and Wu 2000, Zhang and Wedel
2009), and such observed behavioral discrepancy can be partially attributed to the differences in
shopping environments (Zhang and Wedel 2009). The availability and display of product-
attribute information, and the design of interactive decision aids, may all affect consumers‘
information search and evaluation processes, purchase decision making, and the evolution of
shopping behavior in online stores. In this dissertation, we intend to empirically investigate the
impact of the online shopping environment on consumers‘ purchase decision processes, and how
consumers adapt to this increasingly prominent channel. Essay one focuses on the dynamic
information acquisition decisions on comparison shopping websites, and essay two studies the
evolution of purchase decisions in online grocery stores. In both situations, consumers‘ behavior
at one stage may influence their behavior at a later stage. We empirically examine how
consumers dynamically adjust their decision strategies based on the newly acquired information
or experience, in the web-based choice environment. Findings from this dissertation will enrich
the understanding of the dynamic decision making in the information-rich online shopping
1 E-Commerce Report from JP Morgan senior analyst Imran Kahn, http://techcrunch.com/2011/01/03/j-p-morgan-
global-e-commerce-revenue-to-grow-by-19-percent-in-2011-to-680b/
2
environment, and provide valuable insights for online retailers to improve the design of their
store environments.
Dynamic decision making is a field with long tradition in marketing. Constructive
decision making theory proposes that during the decision process, consumers incorporate new
information, develop new standard and construct the decision contingent on task environment
(Bettman and Park1980; Bettman, Luce and Payne 1998). Therefore, decision strategies over
time are interrelated and the decision process is dynamic. Various models have been developed
to study this phenomenon. Time series models, such as autoregressive (AR) and moving
average (MA) models, are used to reveal the over-time impact of marketing variables and make
forecast of purchase behavior (e.g., Dekimpe and Hanssens 2000). The Vector Autoregression
(VAR) model is suitable for studying dynamic interactions (e.g., Dekimpe and Hanssens 1995).
State dependence models are developed to accommodate dynamics in purchase behavior by
including three types of variables: lagged choices such as brand loyalties; lagged marketing
variables such as decaying effects of advertising; and serially correlated error terms in the
random utility function that account for reasons unknown to the researchers (e.g., Erdem 1996;
Guadagni and Little 1983; Heckman 1981, Seetharaman, Ainslie, and Chintagunta 1999). All
these models have made important contributions in studying dynamics of marketing phenomena.
Thanks to recent development in technology, dynamics in the decision making process
has been approached from a new angle: the use of path data (Hui, Bradlow, and Fader 2009a;
2009b). Path data is defined as the ?consumers‘ movement in a spatial configuration?, and
records ?consumers‘ interaction with his environment to achieve his goal? (Hui et al. 2009b).
Grocery shopping paths, eye tracking data, online navigation data, etc., all belong to this
category. The beauty of path data lies in its ability to capture the information search and
3
evaluation processes during decision making, which enables researchers to exploit micro-level
dynamics, and thus offers exciting new research opportunities for this topic.
Researchers have acknowledged the importance of path data in deepening the
understanding of decision processes: Underhill (2004) makes recommendation on shopping
environment design based on the analysis of consumers‘ shopping paths; Hui et al. (2009a)
develop an integrated individual level probability model where consumers‘ entire shopping path
is used to predict store visit and purchase probabilities. Eye-movement paths are another type of
path data that gaining ground in study decision dynamics, especially in the advertising research
(Pieters and Warlop 1999; Rayner 1998; Rosbergen, Pieters, and Wedel 1997). Researchers use
eye trackers to analyze consumers‘ eye fixation paths on advertisements and optimize the design
of them (Pieters and Wedel 2004; Pieters, Wedel and Zhang 2007). Online navigation data offers
great detail on navigation behavior during decision making. It records activities that reflect
consumers‘ preference, provides insights in how consumers search and evaluate information
before purchase, and therefore enables researchers to better interpret and predict online shopping
behavior. Bucklin and Sismeiro (2003) use previous websites visit depth and breadth (numbers
of websites visited) to explain web page browsing strategies. Montgomery and his colleagues
(2004) use sequence of the page viewed to predict the subsequent online navigation path. In this
dissertation, we focus on two types of path data, namely, eye movement paths (essay 1), and
online navigation paths (essay 2), and explore how the information extracted from the path data
help to explain dynamic information acquisition decisions and the evolution of purchase
decisions.
Following up on calls for a flexible evolutionary structure to represent decision dynamics
(Erdem 1996; Hauser and Wisniewski 1982; Heilman , Bowman, and Wright 2000), we provide
4
two approaches for modeling dynamic decision making processes, both fall under the hidden
Markov model framework. Hidden Markov Model (HMM) is a popular probabilistic model for
sequence data analysis. The fundamental premise of HMM is that there several hidden states
governing the observed behavior, and the transitions among these hidden states follow a Markov
process. The probabilistic model effectively extracts diagnostic information from sequence data,
and automatically assigns the observed behavior into latent states, taking uncertainty into
account. The HMM and its extensions are able to flexibly capture the dependence and dynamics
in the decision making process, and has been widely used in a variety of disciplines (e.g.,
Rabiner and Juang 1986; Baldi, Hunkapiller, and Chauvin 1994; Kupiec 1992), including
marketing (Brangule-Vlagsma, Pieters, and Wedel 2002; Du and Kamakura 2006; Montgomery,
et al. 2004; Netzer, Lattin and Srinivasan 2008; Liechty, Pieters and Wedel 2003; Van derLans,
Pieters and Wedel 2008).
Eye movements are fast, adaptive, and partially automatic. The information acquisition
strategies that direct the eyes in searching for information, however, are latent cognitive states. In
addition, the link between eye movement and the unobservable information acquisition strategies
is probabilistic rather than deterministic (Lohse and Johnson 1996; Wedel and Pieters 2000).
Therefore, in essay one, we propose a new Hierarchical Hidden Markov Model (HHMM) for the
analysis of dense eye-movement data. The proposed model extends the standard HMM by
adding another layer of latent states. It is a special case of Hierarchical Hidden Markov Model
(HHMM; Fine, Singer, and Tishby 1998; Heller, The, and Görür 2009), which has been used in
the recognition of human activity (e.g., Kawanaka, Okatani, And Deguchi, 2006), information
extraction, (e.g., Skounakis, Craven, and and Ray 2003), and lexical graphic analysis (e.g.,
Zhang, Yu, Xiong, and Liu, 2003). Three layers in HHMM impose a flexible structure on the
5
information acquisition processes: a lower layer that describes the eye movements, a middle
layer that identifies product-based and attribute-based information acquisition modes, and an
upper layer that captures switching between these modes over time. The model and eye
movement data enable us to diagnose the influence of both conscious strategies and low-level
properties of the eye and the visual brain on dynamic information processes. The paper
empirically investigates the extent of usage and switching of different information acquisition
strategies, how much and what information is processed in these strategies, and the causes and
consequences of strategy switching.
In essay two, we modify the standard HMM by allowing hidden state transition
probabilities to change across time through time-varying covariates, and propose Non-
homogeneous Hidden Markov Model (NHMM) to study how the usage of different decision aids
drives the purchase behavior evolution in online grocery store. Scholars have applied NHMM in
Meteorology and Climatology (Hughes 1993, Hughes and Guttorp 1994a; 1994b); Robertson,
Kirshner, and Smyth 2004; Robertson, Ines, and Hansen 2007), and recently, the model is used
in studying customer relationship dynamics (Netzer et al. 2008). NHMM is able to identify
purchase behavior patterns, which has been a long-standing research interest in the marketing
literature. It is especially suitable for studying drivers behind the behavior evolvement. The
results reveal how consumers switch among behavior states that differ in purchase propensity
and responsiveness to marketing variables, more importantly, how the usage experience with
decision aids for economic needs, non-economic needs, brand preference and personalized
shopping list affects transitions among latent behavior states differently.
6
The organization of this dissertation is as follows: Chapters 2 and 3 discuss the two
essays in depth, respectively; Chapter 5 briefly summarizes each essay, points out the
contributions of this dissertation, and concludes with possible future research avenues.
7
Chapter II: Information Acquisition during Online Choice:
A Model-Based Exploration
Wei Shi
Robert H. Smith School of Business
University of Maryland
Michel Wedel
Robert H. Smith School of Business
University of Maryland
Rik Pieters
Department of Marketing
University of Tilburg
8
2.1. Introduction
Retailers and manufacturers are undergoing competitive pressure to assist consumers in
the large number of choices of products and services that they make online. During the often
short time that it takes to make such a decision, consumers acquire information on the
alternatives presented to them and the way in which they do this affects their decision. This is
recognized by retailers and manufacturers who try to optimize online information displays, in
particular comparison sites that have become popular tools for consumers (for example,
Bizrate.com, Dell.com, and Nextag.com each has over 10 million monthly visitors
2
).
In the academic literature, process tracing methods have been proven important in
helping to understand how consumers process information (Bettman and Jacoby 1976; Bettman
and Kakkar 1977; Johnson, Schulte-Mecklenbeck, and Willemsen 2008; Lohse and Johnson
1996; Payne 1976; Payne, Bettman, and Johnson 1993; Senter and Wedell 1999). Research has
revealed two key information acquisition strategies: attribute-based and product-based (Ball
1997; Bettman, Luce, and Payne 1998; Payne et al. 1993). During attribute-based acquisition,
information is acquired on a single attribute across multiple products, before going to the next
attribute. During product-based acquisition, information is acquired on a single product across
multiple attributes, before going to the next product. Findings on these two strategies, however,
have been obtained with research methods such as information display boards or Mouselab, in
which information becomes available sequentially as consumers manually inspect cells in
attribute-by-product displays. The required motor responses necessarily slow down the decision
process, and render it more controlled and deliberate (Bettman et al. 1998). As a consequence,
2
“15 Most Popular Comparison Shopping Websites|March 2011”, from http://www.ebizmba.com/articles/shopping-
websites.
9
these prior studies could discover mostly high-level and slower cognitive processes that
consumers engage in during decision making.
Comparison sites provide all information on the choice alternatives simultaneously. Then,
decisions can be made much faster and information acquisition is more likely under influence of
both conscious information acquisition strategies and low-level properties of the eye and the
visual brain. Lynch and Srull (1982) call for a methodology that is capable of diagnosing both
?overt and covert?, and ?voluntary and involuntary? aspects of decision processes. Russo (1978),
Russo and Leclerc (1994) and Lohse and Johnson (1996) have argued that in such situations eye
tracking methodology would be ideally suited to provide insights into how consumers acquire
information.
Affordable eye tracking systems are now widely available. Prior eye tracking studies
have often displayed products in a holistic, for instance pictorial, fashion (Lohse and Johnson
1996; Pieters and Warlop 1999; Russo and Rosen 1975; Russo and Leclerc 1994), rather than in
the form of attribute-by product displays. These latter displays reflect online decision contexts,
and increase the potential insights that can be obtained from eye tracking studies. But they also
increase the challenge of providing meaningful descriptions of the massive amounts of data that
the eye tracking studies produce.
To illustrate this, Figure 2.1 shows the eye movements of a single participant while
making a choice, in our study to be described later (participant 36). The decision took only 76
seconds, but involved 140 eye fixations and saccades (excluding re-fixations). Eye fixations are
brief moments that the eye is still and information is extracted (about 2 to 4 times per second).
Saccades are rapid jumps of the eye between fixations to redirect the line of sight to a new
location. During saccades no information acquisition takes place. In the figure, attribute-based
10
and product-based information acquisition strategies are not clearly discernable: the eyes
switched 101 times between making horizontal and vertical movements. Switching seems
extensive and the question is whether this is due to unpredictability in how people move their
eyes, or whether it is a fundamental property of the underlying information acquisition process.
[INSERT FIGURE 2.1 ABOUT HERE]
The present study uses such eye tracking data to investigate information acquisition on
attribute-by-product matrices as encountered in online choice environments. We intend to
provide a description of how people acquire information in these contexts, investigating the
influence of low-level processes, and the extent of usage of and switching between attribute- and
product-based information acquisitions. To deal with challenges posed by the amount of data, we
develop and test a new hierarchical Hidden Markov model of information acquisition.
2.1.1 Modeling I nformation Acquisition Processes
Valid inferences about information acquisition strategies should recognize that they are
fundamentally unobservable. Eye movements reflect them probabilistically rather than
deterministically (Lohse and Johnson 1996; Wedel and Pieters 2000). The information
acquisition strategies that we are interested in are latent cognitive states that direct the eyes in
searching for information. They can only be inferred from the observed eye movement paths.
Whereas describing the patterns of observed eye fixations and saccades (as we did in the earlier
example may) already be insightful, inferring and describing the latent cognitive states is
preferable (Johnson et al. 2008). To be able to identify when during a decision participants
acquire information by-attribute and when by-product, we develop a model that builds on and
extends previous applications of Hidden Markov Models (HMM) to eye movement data
(Liechty, Pieters, and Wedel 2003; Van der Lans, Pieters and Wedel 2008).
11
We extend previous two-layer HMMs by developing a new model that describes
information-acquisition strategies through three hierarchical layers (see Figure 2.2): (1) the lower
layer describes the observed eye movements, (2) the middle layer describes unobserved
(attribute- and product-based) information acquisition strategies, and (3) the upper layer
describes how consumers switch between them. The model captures this moment-to-moment
switching pattern through a number of upper layer states that are a-priori unknown. This allows
the transition probabilities between information acquisition strategies to vary over time, and
overcomes the time-homogeneity of two-level HMMs. The model enables us to identify from
moment-to-moment what the probabilities are that participants process information by-product or
by-attribute.This makes it possible to investigate post-hoc how often participants switch between
attribute-based and product-based information acquisition, and what specific information they
process while doing so. Studying information acquisition processes in this manner is not only of
theoretical interest, but also provides insights into the usability of comparison websites.
[INSERT FIGURE 2.2 ABOUT HERE]
We first provide the theoretical foundation and formulation of our model. Then we
describe the eye movement experiment of decision-making on comparison websites. In the
experiment, we manipulate the information presentation format and provide participants a
comparison website in either a product-row or a product-column format. These are the two main
presentation formats used for comparison sites (e.g., the popular comparison site Bizrate.com
uses a product-row format by default, and Dell.com uses a product-column format). It has been
previously shown that these formats may facilitate specific information acquisition strategies
(Bettman and Kakkar 1977). We then present the results of the model estimation, which reveal
the prevalence of low-level processes, and the extent of switching between the two information
12
acquisition strategies. We conduct an extensive post-hoc analysis to investigate why participants
switch and how much and what information they process in these states. Finally, we discuss
contributions and future research directions.
2.2 Information Acquisition through Eye movements
Prior studies have described high-level cognitive processing during decision-making
(Bettman et al. 1998). Yet, when decisions are fast and more automatic, low-level properties of
the eye and the visual brain may influence the information acquisition process more. Here, we
focus on three aspect of information acquisition: tendencies of horizontal eye movements, of
local eye movements, and of switching between information acquisition strategies. Insight into
these factors will lead to a better understanding of decision-making in information-rich online
choice environments.
2.2.1 Horizontal Eye Movement Patterns
Orderly sequences of eye movements reflect that individuals use systematic information
acquisition strategies. In various visual tasks a horizontal, left-right dominance in consumers' eye
movements has been observed, such as when reading (Rayner 1998) or searching store shelves
(van der Lans et al. 2008). The dominance of horizontal eye movements may be due to the
horizontal layout of many visual displays (Tatler and Vincent 2008). If the information presented
is largely textual (as is typical for comparison websites), this should facilitate such reading-like
patterns even more. But left-to-right eye-movemement tendencies also appear to be independent
of the layout or features of the display and are caused by the more rapid decline in the resolution
of the retina in the vertical than in the horizontal direction (Gilchrist & Harvey 2006).
Information acquisition is thus facilitated when the display is organized consistent with this left-
13
to-right direction (Spalek and Hammad 2005). Therefore, we expect to find a predominant
tendency for left-to-right visual information acquisition, independent of display format.
2.2.2 Local Eye Movement Patterns
The amount of visual detail that can be acquired is highest in the small area around the
exact eye fixation point and rapidly declines towards the visual periphery (Rayner 1998).
Because much of what is present in the periphery is not clearly visible, people need to make eye
saccades when searching for information. It is as if a small attentional "spotlight" moves across
the display, with information being acquired mostly locally within the focus of the spotlight
(Treisman and Gelade 1980). Using a HMM, Liechty et al. (2003) found that information
acquisition from print advertisements indeed took place in a pattern of bursts of multiple
fixations with short saccades on small areas, occasionally separated by one or a few long
saccades to distant areas. A similar pattern is likely to emerge during decision-making on
product-by-attribute displays, where information is mostly textual. Thus, we expect eye
movements – whether attribute-based or product-based – to be mostly confined to contiguous
cells on the information display.
2.2.3 Switching between I nformation Acquisition Strategies
Figure 2.3 presents a hypothetical product-attribute matrix. In attribute-based acquisition
information is gathered on a certain attribute across products, before processing the next
attribute. Solid arrow A expresses this. In product-based acquisition, information is gathered on
attributes of a particular product before evaluating the next product, as dashed arrow B shows.
The two acquisition strategies are comprised of elementary steps (Ball 1997). We propose that
these are reflected in the eye movements.
14
We thus distinguish three elementary eye movements (saccades) between cells of the
comparison matrix. A type-1 saccade is an elementary step in attribute-based information
strategy: the eye jumps between two products for the same attribute (dotted arrow 1 in Figure
2.3). A type-2 saccade is an elementary step in product-based information strategy: the eye
jumps between two attributes for the same product (dotted arrow 2 in Figure 2.3). A type-3
saccade may have several functions: the eye jumps from a particular attribute for one product to
a different attribute for another product. This may reflect a transition between type-1 and 2 eye
movements, an exploratory eye movement (Liechty et al. 2003), or may be a corrective eye
movement (Rayner 1998). An attribute-based information acquisition strategy thus consists of a
sequence of mostly (type-1) attribute-based steps; a product-based information acquisition
strategy is characterized by a sequence of mostly (type 2) product-based steps. Precisely how
many elementary steps or eye movements would constitute such a strategy is not known.
Consumers switch between information acquisition strategies due to a variety of factors,
including task demands (Ball 1997; Pieters and Warlop 1999; Swait and Adamowicz 2001),
experienced accuracy-effort tradeoffs (Bettman et al. 1998) and, importantly, the information
acquired up to that point (Bettman and Park 1980; Russo and Rosen 1975). Correspondingly,
they will switch eye movement directions. Although switching between strategies enables
adaptive decision making, it is also effortful (Gopher, Armony, and Greenshpan 2000; de Jong
2000; Rogers and Monsell 1995) and may reduce the perceived ease of the decision process. To
our knowledge, the extent of strategy switching in natural tasks such as online choice, and its
causes and consequences, have not yet been quantified.
[INSERT FIGURE 2.3 ABOUT HERE]
15
2.3. Model Formulation
2.3.1 Motivation
Our model to identify attribute- and product-based information acquisition is a
Hierarchical Hidden Markov Model (HHMM; Fine, Singer, and Tishby 1998). This model
generalizes Hidden Markov Models (HMM; Rabiner, and Juang 1986), which have been used in
marketing (Du and Kamakura 2006; Montgomery, Srinivasan, and Liechty 2004; Netzer, Lattin,
and Srinivasan 2008), and eye tracking research (Van derLans et al. 2008; Liechty et al. 2003;
Salvucci and Anderson 1998). The link between the observed eye movements and unobservable
information acquisition strategies is probabilistic. For instance, a jump from one attribute to
another in a longer sequence of attribute-based steps has a lower probability to reflect product-
based acquisition, than a similar jump that occurs in a longer sequence of product-based steps.
Therefore, our model needs to capture longer sequences of eye movements. It does this through
three hierarchically related Markov processes, generalizing two-layer HMMs that are constrained
by stationary transition probabilities (Figure 2.2). The lower layer captures (low level) eye
movements; the middle layer of the model captures the attribute-based and product-based
acquisition strategies that drive the eye movements, and the upper layer of the model captures
switching between the two types of acquisition. We intend to use the model as a flexible tool to
describe eye movement patterns and to identify the acquisition strategies that participants use
from moment-to-moment during decision making. We describe and explain those acquisition
strategies in detail, post-hoc.
2.3.2 Specification
We let i = 1,…, I denote participants and t = 1,…, T
fixations. We let ) , (
, , , t i t i t i
p a y =
denote the particular cell in that display defined by attribute (a) and product (p), at eye-fixation t
16
for participant i. In addition,
) , (
1 , 1 , 1 , ÷ ÷ ÷
=
t i t i t i
p a y denotes the cell at the previous eye-fixation t-1.
Figure 2.2 presents the hierarchical relationship among the three layers of the model: the lower
layer describes eye movements y, the middle layer captures information acquisition states
S
1
=1,..., N, and the upper layer states S
2
=1,...,M describe switching between these acquisition
states. Correspondingly, the model is represented by three sets of transition probabilities: the
lower layer transition probabilities ) | (
1 , ,
1
,
÷ t i t i
S
y y P
t i
, the intermediate layer transition probabilities
) | (
1
1 ,
1
,
2
÷
H
t i t i
S
S S , and the upper layer transition probabilities ) | (
2
1 ,
2
, ÷
O
t i t i
S S . These are described
next.
We formulate the middle layer hidden states S
1
=1 and S
1
=2 such that they represent,
respectively, attribute-based and product-based acquisition strategies (Figure 2.2). We assume
) | ( ) | ( ) | (
1 , , 1 , , 1 , , ÷ ÷ ÷
· =
t i t i t i t i t i t i
p p P a a P y y P , given the unobserved state, with ) | (
1 , , ÷ t i t i
y y P the
saccade probability from cell
1 , ÷ t i
y to cell
t i
y
,
,
) | (
1 , , ÷ t i t i
a a P the saccade probability from attribute
1 , ÷ t i
a to attribute
t i
a
,
, and ) | (
1 , , ÷ t i t i
p p P the saccade probability from product
1 , ÷ t i
p to
product
t i
p
,
. Each of the two middle layer states (attribute-based acquisition and product-based
acquisition, S
1
) has a different set of associated saccade probabilities ) | (
1 , , ÷ t i t i
a a P
and ) | (
1 , , ÷ t i t i
p p P .
For state 1 of the middle layer (S
1
=1) we impose a structure on the saccade probabilities
between products that allows for eye movements that are mostly, but not necessarily, consistent
with strict attribute-based acquisition. We assume that at any time the eyes have a probability p
of continuing to move from one product to another along the same attribute, with
probability ) | (
1 , , ÷ t i t i
p p P . These are eye movements consistent with an attribute-based acquisition
strategy. But, with a small probability (1-p) the participant may also make elementary eye
17
movements that are inconsistent with that strategy (types 2 and 3). In the product-based
acquisition state (S
1
=2) the eyes have a probability q of moving from one attribute to the other
along the same product, with probability ) | (
1 , , ÷ t i t i
a a P . This is an eye movement consistent with
product-based acquisition. The eyes, however, also have a small probability of (1-q) of making
inconsistent moves (types 1 and 3). This formulation renders the model conservative in its
identification of switching between the underlying acquisition strategies, because it allows eye
movements to deviate from strict attribute-based and product-based acquisition strategies. These
assumptions yield the following set of equations for the lower-level saccade probabilities:
1 , , 1 , , 1 , ,
2
1 , , 1 , , 1 , ,
2
1 , , 1 , , 1 , ,
1
1 , , 1 , , 1 , ,
1
) | ( ) | (
) 1 ( ) | ( ) | (
) | ( ) | (
) 1 ( ) | ( ) | (
1
1
1
1
÷ ÷ ÷
=
÷ ÷ ÷
=
÷ ÷ ÷
=
÷ ÷ ÷
=
= · =
= ÷ · =
= · =
= ÷ · =
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
p iff p q a a P y y P
p iff p q a a P y y P
a iff a p p p P y y P
a iff a p p p P y y P
(1)
Because we are interested in the extent to which attribute-based and product-based
information acquisition is local (on contiguous cells), or global (on noncontiguous cells) (Liechty
et al. 2003), we re-parameterize the attribute saccade probabilities ) | (
1 , , ÷ t i t i
a a P , and the product
saccade probabilities ) | (
1 , , ÷ t i t i
p p P , as probabilities on spatially contiguous versus spatially
noncontiguous attributes and products.
The switching between the attribute-based and product-based states of the middle layer
(S
1
=1,2) follow a Markov process that is governed by the upper-layer hidden states (S
2
) (see
Figure 2.2). These upper layer states can be conceived of as regimes that govern the middle layer
switching patterns. Given an upper layer state (S
2
), the transitions among middle layer hidden
states (S
1
) follow a Markov process with a transition probability matrix ) | (
1
1 ,
1
,
2
÷
H
t i t i
S
S S . That is,
the upper-layer state influences how participants switch between the attribute-based and product-
18
based (middle layer) states. The transitions between the upper-layer hidden states (S
2
) also
follow a Markov process described by the transition probability matrix ) | (
2
1 ,
2
, ÷
O
t i t i
S S . While the
number of states of the middle layer is dictated to equal two by theory (attribute-based and
product-based acquisition), the number of states of the upper layer is a-priori unknown.
We can now write the saccade probabilities between cells of the comparison matrix,
1 , ,
|
÷ t i t i
y y , conditional on the states of the two hidden layers that participant i is in at fixation t,
1
,t i
S and
2
,t i
S , as:
) | ( ) | ( ) | ( ) , ; | (
1 , ,
1
1 ,
1
,
1
2
1 ,
2
,
1
2
,
1
, 1 , ,
1
1 ,
2
,
2
1 ,
÷ ÷
=
÷
=
÷ ¿ ¿
÷ ÷
H O =
t i t i t i t i
N
S
S
t i t i
M
S
t i t i t i t i
y y P S S S S S S y y P
t i
t i
t i
(2)
with ) | (
1 , ,
1
,
÷ t i t i
S
y y P
t i
given by (1). If Ocollects all parameters, and ) (
1 ,
1
,
i
S
y P
t i
is the initial fixation
probability, the likelihood function of participant i is:
[ ¿ ¿ ¿ [ ¿ ¿
= = =
÷ ÷
= =
÷
= =
÷
H O = O
T
t
M
S
N
S
t i t i
S
t i t i
N
S
S
T
t
t i t i
N
S
M
S
i
S
i
T i T i
t i
t i
t i
t i t i
i
y y P S S S S y P y L
2 1 1
1 , ,
1
1 ,
1
,
1 2
2
1 ,
2
,
1 1
1 ,
2
,
1
,
1
,
1
1 ,
2
,
1
,
2
,
1
1 ,
] ) | ( ) | ( ) | ( ... [ ... ) ( ) | (
(3)
A key output of the model estimation is the marginal posterior probability that participant
i is in information acquisition state
1
,t i
S at fixation t: ) , | (
,
1
,
O
t i t i
y S P , as well as the posterior
probabilities of being in the upper level states
2
,t i
S : ) , | (
,
2
,
O
t i t i
y S P . Those probabilities allow us to
describe the information acquisition strategies that participants use from moment-to-moment.
2.3.3 Model Estimation and Testing
We specify uninformative conjugate priors on all model parameters. We use MCMC to
estimate the model, which is programmed in R. Single move sampling schemes are used to
sample middle layer hidden states S
1
, upper layer hidden states S
2
, and the parameters of the
transition matrices (Robert, Celeux, and Diebolt 1993). Several tests on simulated data reveal
19
accurate recovery of the parameters
3
. In all analyses, we apply the model using 25,000 draws and
discard the first 5,000 draw to burn-in. The chains are stabilized after 5,000 draws. We take one
in ten target draws and report the mean and standard deviation of the resulting 2,000 draws to
summarize the posterior distributions of the parameters.
We compare our model with realistic alternatives described later, and test for the number
of upper layer states of our model using the log-marginal density (LMD). We compute the LMD
using Chib‘s (1995) method, which provides unbiased and stable estimates:
) | ( ) ( ) | ( ) (
* * *
Y In In Y f In Y LMD O ÷ O + O = t t (4)
The calculation of ) (
*
O t In , the log-prior density, and ) | (
*
Y In O t , the log posterior density,
both computed at the ordinate of the posterior density ?*
(for example, the posterior mean) is
straightforward. We compute the high-dimensional sum in the likelihood (3) using Scott‘s (2002)
likelihood recursion method.
2.4 Information Acquisition on A Comparison Website
2.4.1 Experimental Procedure
We chose the Dell website (www.Dell.com) as context for our study. It provides the
option of comparing various personal computers that are relevant for our participants. We
manipulate the display format by transposing rows and columns in a two-group design. The
original comparison matrix that we singled-out from the website contains twelve attributes:
?Picture‘, ?Price‘, ?Processor‘, ?Operating System‘, ?Memory‘, ?Keyboard Mouse‘, ?Monitor‘,
?Hard Drive‘, ?Optical Drive‘, ?Wireless‘, ?Office Software‘ and ?Warranty‘; and four desktop
models: ?Inspiron 864‘, ?Inspiron 819‘, ?Inspiron 758‘ and ?Inspiron 689‘. The comparison
3
30 participants each with 100 observations, transition matrix with 3 columns or 3 rows. All true values fall within the
95% Highest Posterior Density region of the parameters in question; details available upon request.
20
matrix thus lists twelve attributes in the rows, and four desktop products in the columns, plus one
column that contains attribute labels. In the second condition, the comparison matrix is
transposed, so that the attributes are in columns and products in rows. We label the two
conditions as the Product-Column (PC) and the Product-Row (PR) condition, respectively. 108
undergraduate students (55 male) for which choice of a personal computer is relevant
participated. Participants were randomly assigned to either the PC or the PR condition. They read
the instructions that asked them to make a choice for a desktop computer. After having made
their final choice, they were asked to indicate how easy it was to collection information and
decide among the different computers on a 5-point response scale.
2.4.2 Eye movement Recording
Tobii 1750 infrared eye tracking equipment was used (www.Tobii.com). It leaves
participants free to move their heads; cameras in the rim of a LCD-computer monitor (1,280 x
768 pixel resolutions) track the position of the eye and head. Measurements are taken with a
frequency of 35Hz and a precision better than 0.5 degree of visual angle. Instructions and stimuli
were presented on the monitor and participants continued to a next page by pressing the space
bar. Saccades between cells of the comparison website (a total of 18,172 observations) are the
unit-of-analysis; re-fixations were excluded.
2.5 Results
2.5.1 Model Comparisons
We compare the proposed model with four alternatives. Alternative model 1 is a HMM
without the upper layer but else is the same as our proposed model. If this model would be the
best, the probabilities of switching between attribute-based and product-based information
21
acquisition (level 2) would be constant over time. It would imply that a single stage rather than a
two- or multi-stage decision process (Bettman et al. 1998) would direct information acquisition
in the present context. Alternative model 2 is a two-layer HMM without the (level 2) constraints
of attribute-based and product-based acquisition processes. If this model would be the best, it
would imply that eye movements do not reflect the two information acquisition strategies.
Finally, we test for 2, 3 and 4 upper-layer hidden states for our model. We perform all these tests
for both PC and PR formats.
Our model with 2, 3 or 4 upper layer hidden states outperforms both alternative HMM
models, as revealed by the higher LMDs in Table 2.1. This holds for both conditions (PC and
PR). These results provide support for the hierarchical structure of the decision process: a three-
layer model explains the decision process better than a two-layer model. As shown in Table 2.1,
the value of the LMD levels off after three upper-level states. In addition, the parameter
estimates of two hidden states in the 4-state model are quite similar. Thus, we choose the model
with three upper-layer hidden states as our final model.
[INSERT TABLE 2.1 ABOUT HERE]
2.5.2 General Estimation Results
Appendices I and II provide the parameter estimates of the model. Table 2.2 presents, for
the PC and PR conditions, the probabilities that saccades are consistent with attribute-based, or
with product-based acquisition. For the PC presentation format, in the attribute-based state
participants are 14.0 times more likely (calculated as p/(1-p)) to move their eyes within the same
attribute than making other eye movements. For the PR format this ratio drops to 8.1. For the PC
format, in the product-based state participants are 1.7 times more likely (calculated as q/(1-q)) to
make an eye movement within the same product than making other eye movements. For the PR
22
format this ratio increases to 2.8. Two conclusions are apparent. First, the eyes are much more
consistent in moving across products for the same attribute, regardless of whether the products
are presented in the rows or columns. Second, consistency is substantially higher for either
acquisition strategies if it is presented row-wise. This corroborates previous findings on the
dominance of horizontal eye movements in very different research areas (Gilchrist and Harvey
2006, Van der Lans et al. 2008, Tatler and Vincent 2008).
Table 2.2 also shows for the PC and PR conditions the probabilities that participants
make contiguous and non-contiguous eye movements. For the PC format, eye movements
between contiguous attributes are 10.1 times more likely than between non-contiguous ones.
This ratio is 8.2 for the PR format. For contiguous/non-contiguous products that ratio is 2.6 for
the PC, and 1.8 for the PR format. Participants are much more likely to compare information
from neighboring attributes than from neighboring products, regardless of the orientation of the
display. And they are even more likely to do so when attributes are presented in the rows (PC
format). The PR format induces less contiguous information acquisition than the PC format. This
dominance of information acquisition on contiguous cells of the matrix is in line with prior
accounts of the dominance of local eye movement patterns in print advertisements. Yet, it seems
not to have been reported in previous research on decision making on product-attribute displays
where the information content rather than placement is expected to exert stronger impact in
comparison and evaluation. These results provide evidence for mostly local information
acquisition interspersed by short periods of redirecting the attentional spotlight to other
potentially informative regions, with some influence of the presentation format.
[INSERT TABLE 2.2 HERE]
23
We describe the attribute-based and product-based states of the middle layer in more
detail. First, when being in the attribute-based state, the average number of products inspected
within an attribute is a little over two (2.28 in the PC and 2.07 in the PR condition), while one to
two attributes are inspected (1.55 in the PC and 1.38 in the PR condition)
4
. Second, when being
in the product-based state, the average number of attributes inspected within a product is about
three (3.11 in the PC and 3.20 in the PR condition), while on average about a single product is
inspected (1.04 in both conditions). Thus, participants inspect about three attributes for a single
product, or compare two products on one to two attributes, before switching to another strategy
of processing. This is largely independent of the orientation of the display.
Figure 4 plots the cumulative percentages of attributes and products inspected across the
(normalized) decision time. In both formats, participants do not examine all products until about
halfway through the decision process. The increase in the number of attributes inspected with
increasing decision time is almost linear, with around half of the attributes (51% in the PC and
46% in the PR condition) inspected halfway through. About twenty percent of the attributes
(25% in the PR condition and 17% in the PC condition) are never inspected. Both the number of
attributes and the number of products inspected are lower in the PR than in the PC condition.
This selectivity in information acquisition may be due to information overload that limits the
uptake of available information, and is influenced by the presentation format.
[INSERT FIGURE 4 ABOUT HERE]
Appendix I provides estimation results for the middle and upper layers of the model. The
upper layer states in the model govern the prevalence of the middle layer states and the switching
between them. Although it is not always possible or advisable to interpret these upper level states
4
The average number of products inspected is somewhat lower at the beginning and end of the decision process (about
1.7 on average).
24
and one may merely view them as a flexible approximation to the time-structure in the data, they
are clearly differentiated in the present application. That is, upper-layer state S
2
=1 induces state-
dependence with participants having a large probability of sticking to whatever their current
acquisition strategy is (a ?stay? regime). State S
2
=2 induces switching to and staying with
attribute-based acquisition (a ?switch to and stay with attribute? regime). State S
2
=3 induces
switching to and staying with product-based acquisition (a ?switch to and stay with product?
regime?). This interpretation holds for both formats.
Figure 5a presents the (posterior) probabilities of the three upper layer states (?regimes?)
across (normalized) time for all participants; Figure 5b graphs the probabilities of the two middle
layer states (?acquisition strategies?). Counter to theories of two-stage decision making that
postulate a long period of attribute-based acquisition in the first stage followed by a long-period
of product-based acquisition in the second stage, Figure 5a shows that in both formats,
?switching to and staying with attribute-based acquisition? (S
2
=2) dominates in the middle part
of the decision process, and that ?switching to and staying with product-based acquisition? (S
2
=
3) dominates in the early and later parts of the process. In addition, upper-layer state S
2
= 1
(?stay?) is more prevalent in the beginning and the end, which indicates that participants are then
less likely to switch between strategies. Saccade lengths (calculated as the Euclidean distance
between the corresponding cells of the comparison matrix) are much larger for state 1 (S
2
=1)
(35.4) than that for states 2 (S
2
=2) and 3 (S
2
=3) (12.2 and 7.8 respectively). This suggests that,
independent of the format, participants tend to adopt a more global product-based information
acquisition in the beginning and end of the decision process. The prevalence and stickiness of
product-based strategies towards the end of the decision process in particular, is consistent with a
25
pre-decisional gaze bias that reflects a cascading preference formation and decision justification
(Pieters and Warlop 1999; Shimojo, Simion, Shimojo, and Scheier 2003).
As for the format impact, Figure 5a shows that for the PR format, the prevalence of the
upper-layer state 3, which induces ?switching to and staying with product-based acquisition?, is
much higher than that for the PC format. In the latter format, upper-layer state 2, inducing
?switching to and staying with attribute-based acquisition?, dominates. Thus, the PR format
indeed induces more product-based acquisition (Bettman and Kakkar 1977). This is confirmed in
Figure 5b, which shows the prevalence of the two middle layer states over time. The figure
shows, however, that this dominance of product-based acquisition in the PR format manifests
itself only in the middle of the decision process. For the PC format, product-based and attribute-
based acquisitions are equally prevalent in the mid-range of the decision. The dominance of
product-based acquisition at the beginning and end of the decision process is unaffected by the
information display format.
A multinomial logit model with the total (posterior mean) durations of different
acquisition states as predictors (pseudo-R
2
= .821), shows that more attribute-based acquisition
significantly increases the probability of choosing product 1 over 4 (b = 6.32, p = .003), while it
reduces that of the other products (b = -7.01 for product 2, p = .001, -6.33 for product 3, p =
.063). Product 1 (Inspiron 864) is superior on most non-price attributes. Thus, participants in the
PC condition seem more likely to choose the dominating product than those in PR condition
(choice probabilities: PC: .26 vs. PR: .20). They also seem less likely to choose the second
product (choice probabilities: PC: .08 vs. PR: .13), which is inferior on most non-price attributes.
[INSERT FIGURE 2.5 ABOUT HERE]
26
2.5.3 Switching between I nformation Acquisition Strategies
To illustrate the switching between information acquisition states, Figure 6 plots their
probabilities over time for one illustrative participant in the PC and one in the PR condition. The
Figure shows that switching between the two (middle-layer) information acquisition strategies is
very frequent, but for the PR format even more so. On average, participants switch respectively
.61 and .83 times per second for the PC and PR conditions. If we were to count single-step
transitions in the eye movements directly, we obtain even higher switching frequencies: 1.27
(PC) and 1.37 (PR) switches per second. These latter observed frequencies are much higher,
because eye movements are equated with information strategies, rather than being treated as
noisy indicators of them. But, even though our model provides much more conservative
estimates, switching between the two information acquisition strategies is so frequent that it is
unlikely that these reflect conscious "strategies", and we will call them "modes" instead.
[INSERT FIGURE 2.6 ABOUT HERE]
Participants switch between the upper layer states .84 times per second for the PC
condition and .88 times for the PR condition. Thus, while the presentation format affects the
switching between the states of the middle-layer (modes of information processing), it does not
affect switching between the upper-layer states. These states might be associated with higher
cognitive processes (Salvucci and Anderson 1998), which may be less susceptible to bottom-up
influences such as the presentation format.
This frequent switching between acquisition modes significantly increases decision time
(regression coefficient b = 1.53 for middle layer switching, p = .06; b = 2.30 for upper layer
switching, p < .001). Further, switching significantly reduces the experienced ease-of-processing
(average of post-choice evaluation items, M = 2.78; b = -.33 for middle-layer switching, p =
27
.002). Therefore, participants perceive the decision process to be easier in the PC condition (less
switching) than that in the PR condition (M = 3.02 and M = 2.64 respectively; p = .059).
To investigate determinants of switching between states, we estimate logit regressions of
the participant-specific switching indicators on the percentages of products (P), respectively
attributes (A) inspected up to that time point. In the middle-layer, switching from product-based
to attribute-based acquisition is predicted by the percentage of attributes already inspected (PC
condition: b= .29; p=.08; PR condition: b= .38, p=.07). Switching from attribute-based to
product-based acquisition, on the other hand, is predicted by the percentage of products already
inspected (PC condition: b= .48, p=.03; PR condition: b=.96, p<.001). This corroborates that
participants switch to a specific mode of information acquisition as more information has been
acquired in the other mode (Russo and Rosen 1975; Bettman and Park 1980).
2.5.4 Attention to Products and Attributes
Next we show what specific product and attribute information consumers acquire over
time when they are in respectively the attribute-based and product-based acquisition states. For
that purpose we select three participants representative for the PC and the PR conditions, by K-
means clustering participants based on their fixations on products and attributes, and selecting
the most central participant for each cluster. The cumulative numbers of fixations on products in
the attribute-based state and on attributes in the product-based state are shown in Figures 7 (PC
condition) and 8 (PR condition), respectively. The number and pattern of fixations that the
products and attributes receive shows considerable heterogeneity between participants, but four
key insights can be obtained.
First, in many cases, some products seem to be taken out of consideration and receive no
additional gaze after a certain time point: the cumulative product-fixations in the attribute-based
28
state level off, even more so for the PR than for the PC format. This is in line with theories of
hybrid decision-making (Bettman 1979). However, contrary to those theories, in several cases
the eliminated alternative is reconsidered later in the decision process. Take product 1 (Inspiron
864) in the left bottom of figure 7c as an example. The product receives constant attention at
first, then cumulative fixations level off while the other two or three products are compared. At a
later stage of decision-making (after a wait of about 25 fixations), however, this product is being
considered again, as reflected by increasing fixations. It might be that product 1 is being ?put on
hold? while other alternatives are compared, or, it is possible that this product is taken out of
consideration at first and then is being re-considered later. Either way, it seems that the
comparison set changes over the entire time course of decision making. Indeed, the scope of the
product comparison set shrinks and expands (between 1 and 4, SD = 1.10). Thus, in contrast with
hybrid decision-making theories that postulate that the comparison set is constant after a certain
time point, the plots show that it changes over time.
Second, in the product-based acquisition state, attributes enter the decision process
sequentially. Many are considered only once. The temporal sequence of selection of attributes
may reflect their relative importance to the choice goal at hand (Bettman et al. 1998, Wedell and
Senter 1997). It seems that for each of the representative respondents a fair number of attributes
is not considered at all, which is even more so in the PR condition.
Third, regularly attributes and products that have been rejected earlier are briefly
revisited in the final stage of the decision process. This appears to reflect a final verification of
the alternative to which the participant has committed (Russo and LeClerc 1994; Russo and
Rosen 1975). A few attributes appear to be exclusively used in this final stage of the decision.
29
Fourth, in the majority of the cases there is a very clear pre-decisional acceleration of
gaze on the chosen alternative towards the end of the decision. This gaze cascade is predictive of
the alternative that is chosen (Pieters and Warlop 1999; Glaholt and Reingold 2009; Shimojo et
al. 2003).
[INSERT FIGURES 2.7 AND 2.8 ABOUT HERE]
2.6. Discussion
Decision theorists have argued that eye tracking provides insights into fast and partly
automatic information acquisition processes during decision making. In recent years affordable
and easy to use eye tracking equipment has become widely available. With that, the challenge
has become to describe the large volumes of data that eye tracking studies produce in a
meaningful manner. For this purpose, the proposed Hierarchical Hidden Markov Model
(HHMM) can help. In our study we apply the HHMM to eye tracking data to describe moment-
to-moment information acquisition on product-by-attribute matrices.
This revealed that participants switch frequently between product-based and attribute-
based acquisition modes: 50 to 60 times during the 67 seconds that a decision took, on average.
The amount of information already collected on attributes induces switching away from the by-
product mode, and the amount of information on products induces switching away from the by-
attribute mode. This high degree of switching during real-life decision making has not been
documented previously, and its causes are not fully understood. Participants limit their attention
to about three attributes for a single product, or two products for one to two attributes. This
might be due to the restricted rate at which information can be consolidated (Marois and Ivanoff
2005), and limits on the amount of information that can be stored into visual short term memory
at a particular point in time (Cowan 2001, Luck and Vogel 1997). Instead of sequentially
30
adopting attribute-based acquisition to eliminate alternatives followed by product-based
acquisition, participants appear to sample ?parcels? of attribute and product information
(Stewart, Chater, and Brown 2006). They constantly adjust their acquisition mode and the
comparison set of products to consolidate information into memory and make the next
information acquisition decision. Investigating the causes and effects of information parceling
and acquisition mode switching may be an avenue for future research.
The revealed high frequency of switching cast doubts on whether "by-attribute" and "by-
product" information acquisition strategies are conscious and deliberate strategies. Rather,
presentation format, but also low-level eye movement tendencies that could not be incorporated
in previous accounts of decision making, appear to play an important role. Low-level, automatic
and unconscious processes are principal in situations of information overload and time pressure
(Lee and Lee 2004; Pieters and Warlop 1999), which occur when facing data-rich web-based
choice environments. First, there is a predominant left-to-right tendency in the eye movements
that makes it appear as if there is a "by-product" or "by-attribute" strategy of processing,
depending on the orientation of the matrix display. Second, information acquisition is
predominantly local, confined to neighboring cells on the display, because of limited visual
detail beyond the immediate eye fixation point. Third, the end of the decision process reveals a
pre-decisional gaze bias that reflects a cascaded process of preference formation and is predictive
of the final choice. In our study the presentation format impacts the way that information is
acquired, because the format changes but the low-level eye movement tendencies (left-right,
local) do not. If eye movement patterns stay the same when the information display changes,
then this causes different information to be extracted and different decisions to be made. This
makes information acquisition and decision making seem adaptive, while in fact they are not.
31
Instead of purposely using attribute-based and product-based ?strategies?, participants
tend to use a strategy of local information sampling, and tend to make horizontal movements.
Information placed contiguously is likely to exert a strong impact on decisions. Such tendencies
provide opportunities for managers to strategically place product-attribute information to
facilitate comparisons. For example, one could increase the attractiveness of a specific product
by placing it next to a product that it dominates overall, or on specific attributes (Huber, Payne,
and Puto 1982). Comparison website providers can also adopt display formats in a proactive
fashion to stimulate consumers‘ use of specific information acquisition mode favorable to their
goals. Managers need to decide how to balance format, sales objective, and switching cost. For
example, we find that that the Product-Column format facilitates attributed-based acquisition and
increases the chance of the dominant product being chosen. Product-Row format leads to more
evenly distributed choice probabilities. However, the Product-Row format favors by-product
information acquisition, which results in more local eye movements, less products and attributes
being inspected, and increased switching between modes of acquisition that makes choice more
difficult. Some comparison websites (e.g., Nextag.com) now allow consumers to firstly sort
alternatives by certain attribute, such as price, ratings, etc., and then present the sorted product
information in Product-Row format (by default). This may be in conflict with desirable attribute-
based information processing for these websites and make decision making more difficult.
Comparison websites, the context chosen for the present research, are a relatively new
shopping environment, yet rapidly increasing in popularity. We demonstrated the effects of the
row-column orientation of the website, as a useful starting point for research in this area. With
the advance of network engineering, comparison websites are now able to provide more and
more extensive and dynamic product-attribute comparisons. Website designers therefore have
32
increasing abilities to improve comparison website usability and may use the results of this study
to help induce information acquisition modes that are congruent with their goals. We hope that
our study stimulates further research interest in this area, which could include the influence of
website design factors such as sorting (arranging the product order such that it reflects relative
importance), grouping (placing similar or related elements close together), trimming (eliminating
unnecessary information), and highlighting (visually accentuating important information through
colors and shading). According to Johnson et al. (2008) decision research can progress more
quickly when it adopts approaches that provide richer descriptions of the underlying processes.
Our proposed Hierarchical Hidden Markov model may provide a starting point for the
development of such models of online choice.
33
Table 2.1 Log Marginal Density (LMD) of Alternative Models
Experimental
Condition
Proposed HHMM Model
with n States in the Upper Layer
Alternative
Model 1:
Constrained
HMM
Alternative
Model 2:
Unconstrained
HMM Two States Three States Four States
Product-column -10392.00 -9066.07 -9058.13 -13696.97 -13132.81
Product-row -10877.76 -9503.57 -9484.53 -13168.99 -12205.66
Table 2.2 Parameter Estimates For Eye movements (lower layer) and Information
Acquisition Strategies (middle layer) (with standard deviations in parentheses)
Acquisition
Strategy
(Middle Layer)
Components
Information Presentation Format
Product-Column Product-Row
Attribute-based
Consistent
.929 (.001) .892 (.001)
Inconsistent .066 (.022) .110 (.011)
Contiguous (products) .444 (.007) .283 (.006)
Noncontiguous (products) .171 (.009) .157 (.007)
Product-based
Consistent .641 (.005) .736 (.002)
Inconsistent .360 (.016) .264 (.008)
Contiguous (attributes) .362 (.000) .338 (.002)
Noncontiguous (attributes) .036 (.001) .041 (.000)
Note: ?Consistent? is the probability of staying in the same attribute or product (p or q),
?Inconsistent? is the probability of moving to another attribute or product (1-p or 1-q).
?Contiguous? is the average of the transition probabilities between block-diagonal sub-matrices
(all elements on the 2×2 diagonal blocks, see Appendix II, shadowed parts), ?Noncontiguous? is
the average transition probabilities between the cells in the off-diagonal blocks (all elements
except those in the diagonal blocks)
34
Figure 2.1 Eye Movements of A Single Participant Making A Choice on the Dell
Comparison Website.
Circles indicate eye fixations (with fixation numbers), and lines indicate saccades between fixations.
35
Figure 2.2 The Three Layers of the Proposed Hierarchical Hidden Markov Model
Note: Triangles represent the observations in the lower layer, with transition probabilities P
1
or P
2
;
squares represent the hidden states in the middle layer (S
1
), with transition probabilities ?
1
and ?
2
; circles
represent the hidden states in the upper layer (S
2
), with transition probability ?.
S
2
=1 S
2
=2
?
1
?
2
S
1
=1 S
1
=2 S
1
=1 S
1
=2
P
1
P
2
P
1
P
2
?
36
Figure 2.3 Product-Attribute Matrix: Information Acquisition Strategies and
Elementary Eye Movements
product 1 product 2 product 3 product 4
attribute 1
attribute 2
attribute 3
attribute 4
attribute 5
Note: Elementary eye movements (dotted arrows): type 1 is saccade between two attributes of same
product, type 2 is saccade between two products within same attribute, and type 3 is saccade from one
attribute of a product to another attribute of another product. Arrow A is attribute-based acquisition and
comprises type 1 and 3 movements. Arrow B is product-based acquisition and comprises type 2 and 3
movements.
A
1
B 2 3
37
Figure 2.4 Cumulative Percentage of Attributes and Products Inspected across
(normalized) Decision Time for Product-Column (PC) and Product-Row (PR) Conditions.
X-axis: normalized decision time; Y-axis: cumulative percentage of attributes and products inspected
38
Figure 5a Aggregate Time Course of the Prevalence of Upper Layer States: S2=1 (long
dash), S
2
=2 (short dash) and S
2
=3 (solid line) for Two Information Presentation Formats
X-axis: normalized decision time; Y-axis: probabilities of three upper layer states
Product-Column Format Product-Row Format
Figure 5b Aggregate Time Course of Attribute-Based (short dash) and Product-Based
(solid line) Information Acquisition for the Two Information Presentation Formats
X-axis: normalized decision time; Y-axis: probabilities of two middle layer states
Product-Column Format Product-Row Format
39
Figure 2.6 Examples of the Time Course of the Three Upper-Layer States (S
2
=1: long
dash, S
2
=2: short dash and S
2
=3: solid line), and of the Two Middle Layer States (attribute-
based acquisition: short dash, and product-based acquisition: solid line), for One
Participant in Each Condition (presentation format)
X-axis: fixation number; Y-axis: probabilities of each state in the middle and upper layers
Product-Column Condition
(Participant 17)
Upper Layer
Middle Layer
Product-Row Condition
(Participant 3)
Upper Layer
Middle Layer
40
Figure 2.7a The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
39, PC condition. Chosen Alternative: 2 (Dell 819), Decision Time: 1 min and 57 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
Figure 7b The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
42, PC condition. Chosen Alternative: 4 (Dell 689), Decision Time: 2 min and 8 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
41
Figure 2.7c The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
19, PC condition. Chosen Alternative: 1 (Dell 864), Decision Time: 1 min and 28 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
42
Figure 2.8a The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
24, PR condition. Chosen Alternative: 3 (Dell 758), Decision Time: 59 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
Figure 2.8b The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
22, PR condition. Chosen Alternative: 2 (Dell 819), Decision Time: 1 min and 34 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
43
Figure 2.8c The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
41, PR condition. Chosen Alternative: 3 (Dell 758), Decision Time: 3 min and 46 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
44
Chapter III: Usage Experience with Decision Aids and Evolution of
Online Purchase Behavior
Wei Shi
Robert H. Smith School of Business
University of Maryland
Jie Zhang
Robert H. Smith School of Business
University of Maryland
Acknowledgments
The authors thank an anonymous online retailer for providing the data used in this study.
This research is supported by the MSI grant #4-1649, as a winner of the ?Shopping Marketing?
research proposal competition.
45
3.1. Introduction
Internet retailing has experienced explosive growth for over a decade. As more shoppers
begin to make purchases online, it is important to understand how they adapt to this increasingly
prominent channel and whether their purchase behavior evolves as they gain more experience
with the new shopping environment. Prior research has examined the impact of the Internet
environment on purchase decision processes (e.g., Bechwati and Xia 2003; Häubl and Trifts
2000; Hollander and Rassuli 1999; Lee and Geistfeld 1998), and the differences between online
and offline purchase behaviors (e.g., Danaher, Wilson, and Davis 2003; Degeratu, Rangaswamy,
and Wu 2000; Zhang and Wedel 2009).What is lacking in the literature is a comprehensive
examination of the evolving patterns of purchase behavior, such as in terms of store loyalty and
price sensitivity, and more importantly, what may drive purchase behavior changes in online
stores.
It is well documented that the store environment can influence a consumer‘s decision
making process (e.g., Park, Iyer, and Smith 1989; Inman, Winer, and Ferraro 2009). A distinct
feature of the Internet store environment is that online stores offer a variety of interactive
decision aids which can facilitate consumers‘ shopping processes. For example, many online
retailers provide decision aids that allow shoppers to sort alternatives or filter them with certain
criteria, to create personalized shopping lists, or to check total basket spending. Studies have
shown that this kind of interactive decision aids can influence consumers‘ information search
processes, purchase outcomes, and satisfaction (e.g., Bechwati and Xia 2003; Häubl and Trifts
2000; Hollander and Rassuli 1999; Lee and Geistfeld 1998). Therefore, one can speculate that, as
online shoppers accumulate more experience with using various decision aids, their purchase
behavior may also change over time as a consequence.
46
The objective of this research is to conduct an empirical investigation on whether and
how usage experience with various decision aids may drive online purchase behavior changes
over time, and what roles different types of decision aids may play in the process. We intend to
address the following managerial questions: 1) Will shoppers exhibit more habitual behavior as
they get accustomed to an online store, or are they more likely to engage in on-the-spot decisions
as they become more experienced and efficient with using interactive decision aids? 2) How will
this affect their tendency to shop from an online store and their price sensitivity? 3) What kind of
decision aids can mitigate price competition? 4) What kind of decision aids may increase
consumers‘ loyalty to an online store? and 5) How can marketers influence consumers‘ behavior
evolvement by designs of decision aids in an online store? Answers to these questions can help
Internet retailers improve the design of their store environments and provide insights for
manufacturers to modify promotion messages adaptively according to consumers‘ evolving
purchase behavior.
Online shopping behavior has been shown to be systematically different from offline
shopping behavior (e.g., Danaher, Wilson, and Davis 2003; Degeratu, Rangaswamy, and Wu
2000; Zhang and Wedel 2009). For example, the observed behavioral discrepancy can be
attributed to two broad sources: differences in intrinsic characteristics between online and offline
consumers, and differences in the shopping environments (Zhang and Wedel 2009). Researchers
have postulated that interactive decision aids available in online stores can train consumers to
shop in certain fashions, which would attribute to purchase behavior differences in the two types
of shopping environments (e.g., Alba et al. 1997; Degeratu et al. 2000; Zhang and Wedel 2009).
Our study will provide an empirical test of this conjecture and shed light on whether decision
47
aids available in online stores indeed lead to purchase behavior changes for the same consumer
over time.
Like the above mentioned previous studies that compare online and offline purchase
behavior, our empirical investigation is carried out in the context of online grocery stores. After
initial struggles and some high profile failures, Internet grocery retailing has shown a resilient
comeback and experienced steady growth in recent years. According to a recent report by the
Nielsen Company, online grocery retailing has grown at more than 20% compound annual rate
since 2003 and attracts 13 million U.S. Internet users by July 2009 (Swedowsky 2009). Findings
from our study will be relevant to a wide range of companies, especially as more traditional
retailers (e.g., Safeway, Albertson, Wal-Mart) and Internet retailers (e.g., Amazon.com) venture
into the online grocery retailing business.
Our data are provided by a leading Internet grocery retailer which was among the very
first to sell groceries online. The dataset was collected during the period when the retailer first
launched its web business, which makes it particularly appealing to study the evolution of online
purchase behavior. Research has shown that consumers‘ in-store decision making processes may
vary by product categories (Inman, Winer, and Ferraro 2009). Our dataset includes detailed
click-stream navigation information, as well as individual purchase history data in multiple
product categories, and thus allows us to examine potential differences in the patterns across
these categories.
We construct a Non-homogeneous Hidden Markov Model (NHMM) of category
purchase incidence and purchase quantity, in which parameters are allowed to vary over time
across hidden states as driven by usage experience with different decision aids. The Hidden
Markov Model is well suited for the purposes of this research. Shopping behavior, including
48
online shopping behavior, has been classified into different states, such as for ?hedonic? and
?utilitarian? motivations (Babin, Darden, and Griffin 1994; Childers et al. 2001; Hirschman and
Holbrook 1982). These studies suggest that the observed online shopping behavior is likely to be
directed by certain latent ?behavior states?, and the store environment may train consumers and
change their behavior states over time. Our NHMM is built to identify these latent states and
examine how usage experience with online decision aids may drive transitions between these
states.
Understanding how consumers‘ purchase behavior evolves over time as their experience
with decision aids accumulates would offer valuable insights for online retailers to improve the
design of their store environment. It could help manufacturers modify their communication
messages (for example, choose to focus on price/promotion-oriented information or to highlight
specific product attributes), based on the purchase behavior revealed in different latent states.
Moreover, findings from our study will suggest ways for online retailers to offer personalized
shopping environment for individual consumers, or to influence their purchase behavior
evolution.
3.2 Conceptual Development and Literature Review
In this section, we present the conceptual development of our study and provide an
overview of the relevant literature.
3.2.1 Online Decision Aids
Online stores make the shopping process easier and more convenient by offering a
variety of decision aids. These decision aids allow consumers to perform a more ?thorough and
exhaustive search? (Hollander and Rassuli 1999; Lee and Geistfeld 1998). They enable online
49
shoppers to make better purchase decisions and doing so with less effort (Häubl and Trifts 2000).
Decision aids are especially popular in online grocery stores. We classify four types of decision
aids that are commonly available in these stores.
1. Decision aids for nutritional needs: Many online grocery stores provide decision aids
to facilitate the shopping process for consumers who have special dietary needs (Swedowsky
2009) or are concerned of nutritional information. These decision aids include sorting functions
(such as ?by calories?, ?by cholesterol?, ?by sugar?, and ?by fat?) and precluding functions (such
as ?Kosher foods?, ?organic food only?) to rank, compare, or filter the products with certain
criteria. For example, www.groceryexpres.com offers 12 functions to fulfill consumers‘ special
dietary needs. www.freshdirect.com has 16 such functions. Growing health concerns among the
public are believed to have contributed to the prevalence of such decision aids.
2. Decision aids for brand preference: In online stores, consumers with specific brand
preference can choose their preferred products by the brand name using functions such as
?sorting by brand/name? or ?search by (brand name)?. This type of decision aids are ubiquitous
in online stores (see www.freshdirect.com, www.coles.com, www.netgrocer.com for just a few
examples). Compared to brick-and-mortar stores, they make the shopping process particularly
efficient for consumers who have strong brand preferences by avoiding effortful navigations
across physical shelves.
3. Decision aids for economic needs: online decision aids such as ?sorting by price?,
?sorting by promotion?, or ?club special first? (see www.freshdirect.com, www.safteway.com,
www.peapod.com for examples) make the shopping process easier for price-sensitive consumers.
Such decision aids facilitate price comparisons and might induce higher price sensitivity (Alba et
al. 1997). In addition, some online stores allow shoppers to check the total spending before
50
submitting an order, which enables budget-conscious consumers to monitor their spending more
effectively during the shopping session.
4. Personalized shopping lists: Some online stores offer consumers the option to create
personal shopping lists or save previous order lists automatically (such as www.peapod.com,
www.freshdirect.com, www.safeway.com, and www.walmart.com). They allow consumers who
are time-constrained or have relatively consistent shopping baskets to complete the shopping
process quickly. Shopping lists serve as a memory aid (Block and Morwitz 1999). Research has
shown that consumers who use shopping lists tend to make less unplanned purchases (Inman et
al. 2009). Therefore, usage experience with shopping lists may train consumers into habitual
shoppers.
These interactive decision aids offer online stores a unique advantage over their brick-
and-mortar counterparts, by making the purchase decision process less effortful, more efficient,
and more suited to individual‘s needs and preferences. The main objective of this study is to
examine whether and how the usage experience with different types of online decision aids may
drive purchase behavior changes overtime. Although the specific context of our study is online
grocery stores, most of the decision aids classified above, with the exception of those for
nutritional needs, apply to other types of retailers and product categories.
3.2.2 Evolution of Online Shopping Behavior
Previous studies suggest that consumers' purchase behavior may change over time in an
online store (e.g., Ansari, Mela, and Neslin 2008; Zhang and Krishna 2007). In the context of
offline stores and unfamiliar new product categories, Heilman and colleagues (2000) show that
consumers‘ purchase behaviors exhibit evolving patterns as their experience with purchasing the
category increases. Given that the Internet shopping environment is distinctively different from
51
traditional shopping channels with many unique features, shoppers new to the channel are likely
to go through learning and adaptation processes, and thus their purchase behavior may also
evolve over time.
Researchers have suggested several factors that may contribute to the evolution of online
shopping behavior over time, most of which are related to the Internet experience. For example,
comfort with the Internet (Mauldin and Arunachalam 2002), perceived ease of usage, and
perceived usefulness of online shopping all exert a positive impact on the purchase intention
from the Internet channel (Hoffman and Novak 1996; Chen, Gillenson, and Sherroll 2002;
Limayem, Khalifa, and Frini 2000; Pavlou 2003). Yet there has been little research that
empirically investigates how usage experience with various decision aids contributes to the
evolution of online purchase behavior. Our study intends to fill this void.
3.2.3 Usage Experience with Decision Aids and Online Shopping Behavior Evolution
Internet experience has been found to be an important determinant for consumers‘
purchase intention at online stores (see Zhou, Dai,
and Zhang 2007 for a review). Consumers
could gain Internet experience through more time spent online or repeated visits, which would
increase their comfort level with the online shopping environment and thus purchase intention
(Mauldin and Arunachalam 2002). The ?comfort? may come from experience with navigating a
website or familiarity with decision aids available. Positive experience with a website may in
turn induce greater exploratory behavior on the site (Hoffman and Novak 1996; Mathwick and
Rigdon 2004). These two processes could reinforce each other and accelerate the learning of
decision aid usage. Therefore, as shoppers' experience with decision aids accumulates, their
propensity to purchase from an online store is likely to increase as well.
In terms of responsiveness to marketing mix variables, the impact of online decision aids
52
is likely to be more nuanced. We focus on price sensitivity in this discussion. Usage experience
with online decision aids may affect shoppers‘ price sensitivities at both the store choice stage
and the in-store purchase decision stage.
Many online stores allow consumers to create personal shopping lists and/or store other
shopping information in their personal accounts. These decision aids could create a ?lock-in?
effect (Smith, Bailey, and Brynjolfsson 2000): a consumer who creates and uses personal
shopping lists in one store may face higher switching costs if s/he decides to shop at other
(online or offline) stores. In other words, price is not the only factor to be evaluated when
determine where to shop (Bakos 2001). We expect that increased usage of personal shopping
lists are likely to enhance consumers' loyalty to the store, induce habitual purchase behavior, and
soften their price sensitivity when it comes to choose the shopping venue. This ?lock-in? effect
may also apply to other types of decision aids, such as those for nutritional needs.
The Information Integration Theory (Anderson 1971, 1981) provides some guidance in
predicting the impact of decision aid usage experience on price sensitivity at the in-store
purchase decision stage. According to the theory, certain attributes, such as brand or price, can
surrogate information on other attributes if the latter have limited availability; yet when
information on other attributes becomes available, weights of existing attributes will be reduced
(Anderson 1971, 1981; Bettman, Capon, and Lutz 1975). In traditional brick-and-mortar stores,
consumers are more likely to focus on price-related information because it is easily available and
highly salient with frequent feature and display advertisements (Degeratu et al. 2000). In
contrast, decisions aids available in online stores allow consumers to more efficiently access and
utilize information on other product attributes. They now can find the product that meets their
needs based on important attributes other than price at a lower search costs (Alba et al. 1997). In
53
other words, for some consumers, the weight of price information is likely to reduce while the
importance of other attribute information is likely to increase (Degeratu et al. 2000; Smith et al.
2000), and thus price sensitivity may decrease over time as a consequence for these consumers.
On the other hand, online decision aids intended for economic needs make it easier for shoppers
to use price related information more efficiently. Therefore, it is also possible that, at least for
some consumers, experience with online decision aids will train them to be more price sensitive
(Alba et al. 1997). We leave it as an empirical question regarding the pattern of price sensitivity
in online stores over time. More importantly, we intend to find out what types of decision aids
may reduce price sensitivity and what types have the opposite effects.
3.2.4 Potential Hidden States of Purchase Behavior
The classification of different purchase behavior has been a long-standing research
interest in the marketing literature. For instance, based on consumers' motivations for shopping,
their purchase behavior can be classified into ?hedonic? and ?utilitarian? states (Babin et al. 1994;
Childers et al. 2001; Hirschman and Holbrook 1982). Hedonic consumers are ?equivalent to
brick-and-mortar window shoppers for whom the shopping experience is for entertainment and
enjoyment?; while utilitarian consumers (or goal-oriented shoppers) normally ?purchase
products in an efficient and timely manner to achieve their goals with minimum irritation?
(Childers et al. 2001, p513). Cheung et al. (2003) classify online shoppers into ?intention?,
?adoption (purchase)?, and ?continuation (repurchase)? types based on their purchase intentions.
These different stages of purchase intentions may well apply to the same consumer over time.
Heilman and colleagues (2000) conjecture that two competing forces --- consumers' desire to
collect information about alternatives and their aversion to trying risky ones --- drive consumers'
purchase behavior evolution among three hidden states when buying new (unfamiliar) product
54
categories.
In this study, we propose a Hidden Markov Model which allows us to empirically
identify latent states of purchase behavior in the data. The classifications in the literature will
provide valuable guidance for us to interpret these latent states.
3.3 Model Formulation
This study investigates how the usage experience with various online decision aids
affects consumers‘ purchase behavior evolution. We construct a Non-homogeneous Hidden
Markov Model (NHMM) of category purchase incidence and purchase quantity, in which
parameters are allowed to vary over time across hidden states as driven by usage experience with
different decision aids.
The basic premise of our model is that consumers‘ purchase decisions at any given time
is driven by the hidden behavior states they are in, where the hidden states differ in terms of the
baseline tendency to purchase from the online store and their price sensitivity. Consumers switch
between these states as their experience with decision aids accumulates over time. We adopt a
Type II Tobit model (e.g., Amemiya 1984) to jointly capture category purchase incidence and
purchase quantity decisions, where parameters in the Tobit model evolve according to a Hidden
Markov Model (HMM) with the transition probabilities assumed to be driven by the usage
experience with various decision aids. Our model belongs to the category of Non-homogeneous
Hidden Markov Model (NHMM) (see Hughes 1993).
3.3.1 Type-I I Tobit Model of Category Purchase I ncidence and Quantity
Purchase Incidence
Let
c
it
U = household i‘s latent utility of purchasing category c from the online store in
55
week t;
c
it
I
= 1 if household i purchases category c from the online store in week t, 0 otherwise.
Without losing generality, we can scale
c
it
U
such that:
.
otherwise , 0
0 U if , 1
¹
´
¦
=
> =
c
it c
it
I (1)
The utility function is specified as:
), , 0 ( ~ ,
2
c
c
it
c
it
c
t
S c
it
N X U o c c | + = (2)
where
c
t
X
is a vector of marketing mix variables for category c in week t.
S
| is a vector of their
coefficients given that household is in hidden state s, including the intercept. The intercept can
be interpreted as a household‘s baseline tendency to purchase category c from the online store in
state c. In order to get clearly-defined interpretations of the hidden states, we choose to focus on
price for the marketing mix component in our empirical analysis, because price sensitivity is a
key aspect of household purchase behavior that we intend to study here. We fix
1
2
=
c
o
for
identification purposes.
Purchase Quantity
Purchase quantity is observed only when a household purchases a category from the
online store. We denote
* c
it
Q
as household i’s latent purchase quantity of category c in week t
(measured in volume units such as ounces), and
c
it
Q
as the household‘s observed purchase
quantity of category c in week t. Then,
¹
´
¦
=
= =
=
otherwise , 0
1 if ,
* c
it
c
it c
it
I Q
Q . (3)
We specify
* c
it
Q
as:
), N(0, ~ ,
2
c
*
o u u |
c
it
c
it
c
t
S c
it
Z Q + = (4)
56
where
c
t
Z
is a vector of marketing mix variables for category c in week t, and
S
| is a vector of
their coefficients given that household is in hidden state s. Like in the purchase incidence
component, we use price as the key marketing mix variable in the empirical analysis in order to
get a clean interpretation of the hidden states.
We take into account the interdependence of purchase incidence and quantity decisions
by assuming that the error terms in Equations (2) and (4),
c
it
?
and
c
it
?
, follow a bivariate Normal
distribution:
|
|
.
|
\
|
|
|
.
|
\
|
|
|
.
|
\
|
1
,
0
0
~
2
c c
c c c
c
it
c
it
N
o µ
o µ o
c
u
. (5)
The likelihood for household i in week t given latent state s can be written as (see Amemiya
1984, page 31):
( )
( )
( )
( ) ( )
) 1 (
1
/ 1
1
ln * , |
2
1
C
it
C
it
c
c c
I
C
t
S
I
c
c
t
S c
it
c
c
t
S c
it c c
t
S S S c
it
X
Z Q Z Q
X Q l
÷
÷
÷ u
|
|
.
|
\
|
|
|
.
|
\
| ÷
|
|
.
|
\
| ÷
+ u = |
o
|
|
o
| µ
| | |
o
o µ
(6)
where ) (· | is the probability density function and u(·) is the cumulative distribution function of
the Standard Normal distribution, respectively.
3.3.2 Hidden States and Transition Probabilities
Our main research question is whether and how purchase behavior, as measured by the
baseline tendency and price sensitivity in purchase incidence and quantity decisions, evolves by
the usage experience with various decision aids. To this end, we model the evolution of the
hidden-state-specific parameter vectors
S
| and
S
| in the Tobit model according to a Markov
transition matrix P
it
, which is specified as functions of usage experience with various decision
aids. In our model, consumers are allowed to switch back and forth between the hidden states.
57
We model the probabilities of switching from hidden state S
t-1
in week t-1 to hidden state
S
t
in week t for household i in a K-state NHMM using the ordered logit formulation (see Netzer
and Srinivasan 2008, page 190):
,
) exp( 1
) exp(
1 ) , | (
......
,
) exp( 1
) exp(
) exp( 1
) exp(
) , | 2 (
,
) exp( 1
) exp(
) , | 1 (
1 1
1 1
1 1
1 1
1 1
1 1
1 1
1 1
, 1
, 1
1
, 1
, 1
, 2
, 2
1
, 1
, 1
1
it s S iK
it s S iK
it t t it
it s S i
it s S i
it s S i
it s S i
it t t it
it s S i
it s S i
it t t it
D
D
D S K S P
D
D
D
D
D S S P
D
D
D S S P
t t
t t
t t
t t
t t
t t
t t
t t
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ +
÷
÷ = =
÷ +
÷
÷
÷ +
÷
= =
÷ +
÷
= =
÷
÷
÷
÷
÷
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
(7)
where K is the number of hidden states. Parameters
1 1 1
, 1 , 2 , 1
,..., ,
÷ ÷ ÷
÷
t t t
s iK s i s i
ì ì ì
are household i‘s state-
specific cut-off points and their values are set in ascending order for a given S
t-1
e {1, 2, .., K}.
it
D is a vector of household i‘s usage experience with a set of decision aids, and
1 ÷ t
s
¸ is a vector
of their coefficients. In the empirical analysis, we categorize and examine six types of decision
aids, including one for nutritional needs, one for brand preference, two for economic needs, and
two for personalized shopping lists. Details of these decision aids and variable computation are
described later.
Note that, in our model, consumer heterogeneity is accounted for by the individual-
specific hidden state transition matrix. It captures individual differences in the behavior
evolution, and automatically accommodates heterogeneity in parameter values in the consumer
response model (i.e., the Tobit model).
The likelihood function for household i across different hidden states is (see Rabiner and
Juang 1986; Rabiner 1989):
) 8 ( ], ) , | ( [ ... ) ,..., , ( ) | (
1 1 1 2
1 1
1 1 2
| , 2 2 1 1 ¿¿ ¿ [ [
= =
÷
= = =
· = = = = = O
K
s
K
s
K
s
T
t
s s c
it
T
t
s s i is
c
iT it
c
i i
c
i i i
s
i
C
i
T
t t
Q l p Q Y Q Y Q Y P Q L | ¢ t
58
Where } , , , {
1 1
,
÷ ÷
= O
t t
s s is
s s s
i
¸ ì | | , and ?
is
is a vector of the initial state probabilities.
3.3.3 Prior Distributions and Estimation Method
The likelihood function in equation (8) is computationally intractable. Thus, we first re-
write the likelihood function as the following (see Macdonald and Zucchini 1997, page 131)
, ' 1 ) , , ( ) | ( ... ) , , ( ) | ( ) , , ( ) (
1 2 2 1 2 1 1 s
2 1
i,1
· O · · O · O · = O
=
÷
= = s s
i
C
T
C
T T T i
s s
i
C C
i
s s
i
C C
i
iT i i
Z X f s s P Z X f s s P Z X f NL t
(9)
where ) , (
s s
i
C
t
C
t
iT
Z X f
=
O
is an K× K diagonal matrix with likelihood ) , | (
,
s s c
it s s
Q l f | | = ;
) | (
1 ÷ t t i
s s P
is an K×K transition probability matrix from t-1 to t; 1’ is a column vector of ones
with length K;
1
is
t
is the stationary distribution of transition matrix
) | (
1 ÷ t t
s s P
, and is calculated
according to equation (7) with covariates set to their mean values in the data.
Following the approach by Netzer et al. (2008), we divide the parameters into two sets:
those that vary across individuals and those that do not. We use a hierarchical Bayes formulation
to estimate the former group of parameters (e.g., Gelfand and Smith 1990; Rossi and Allenby
2003). The prior and hyper-prior are specified as follows:
(1) Parameters that vary across individuals:
u u
u u
u
u
u u o o o
o u ì u
n
i n n
i s is i
I n df df IW
I V V MVN
MVN
t
= A + A E
= =
E =
×
÷
0 0 0 0
0 1 0 0 0
,
; 5 ~ ); , ( ~
) dim( n ; 100 ; 0 ~ ); , ( ~
) , ( ~ }; {
1
(10)
(2): Parameters that do not vary across individuals:
)) ( )' ( exp( ) (
)) dim( ( 50 ; 0 ~
) , ( ~ }; , , {
0
1
0 2
1
1 0
0
0
0
0
¢ ¢
¢ ¢
¢ ¸ ¢ |
¢
¢ ¢ ¢
¢
÷ + ÷ + · +
+ = =
+ = +
÷
×
V P
n I V
V N
n n
s
S S
(11)
{
i
u } and { +} are drawn from Metropolis Hasting algorithm, while {o } and {
u
E } are drawn
from Gibbs sampler.
59
} { , , , , , |
, }, { |
, }, { |
, , , , , , | } {
i it
C
t
C
t
C
it
i
i
it
C
t
C
t
C
it i
D Z X Q
M
M
D Z X Q
u
o u
u o
o u
u
u
u
+
E
E
E +
(12)
We test our NHMM model on simulation data, where there are 40 households, each with
500 observations. The model recovered well on the simulation data (see Appendix III for the
result).
3.4 Empirical Analyses
3.4.1 Data Description
Our data are provided by a leading Internet grocery retailer which was among the very
first to sell groceries online. The dataset was collected during a 62-week period in 1996-1997
when the retailer first launched its web business. Given that this retailer was a pioneer of the
online grocery business, it is very likely that consumers in our data never had prior exposure to
other online grocery stores. This feature makes our dataset particularly suited to study the
evolution of purchase behavior in a new online environment.
The data include click-stream records of detailed navigation processes and purchase
history information of 225 households, as well as pricing information for multiple product
categories. We estimate the proposed model using data of two distinct product categories,
spaghetti sauce and liquid detergent, in order to test the robustness of the results. These
categories are chosen because they differ in terms of hedonicity and purchase frequency, which
have been shown to affect consumers‘ in-store decision making processes (Inman, Winer, and
Ferraro 2009). Only those households that made at least two category purchases during the 62-
week data period are included in the data for a category.
60
3.4.2 Operationalization of Key Variables
As explained previously, we use price as the key marketing mix variable for the Tobit
model of purchase incidence and quantity decisions in the empirical analysis. We first obtain the
weekly actual price of each stock-keeping units (SKU) in a category by combining its regular
price and price discount (if any), and then convert it to a common unit price (e.g., cents per
ounce) across SKUs. The category-level price variable is computed as a weighted average of the
weekly unit prices of all SKUs in the category, where the weights are sales volume shares of the
SKUs in the entire time period.
Decision aid usage information is extracted from the click-stream data. We classify six
types of decision aids that are commonly available in most online stores and fall into one of the
four broad categories described earlier, including one for nutritional needs, one for brand
preference, two for economic needs, and two for personalized shopping lists (see Table 3.1). For
each decision aid, we measure a household‘s usage experience by week t as the cumulative
number of usage counts up to the prior week t-1. This measure avoids possible reverse causality,
that is, purchase behavior may influence the current usage of decision aids. Note that a
household can use the same decision aid multiple times during a single shopping session and the
cumulative experience variables count for each and every time a given decision aid is used by a
household. A main objective of our research is to investigate how usage experience with
different decision aids may drive purchase behavior changes over time. To this end, the
transition probability functions of the proposed Hidden Markov Model include the usage
experience variables of multiple decision aids. Since the usage experience with each decision aid
increases monotonically with time, a more meaningful measure for the model is the relative
usage experience with each decision aid, computed as a percentage of the household‘s total
61
number of usage counts of all decision aids up to week t-1. We also include the total decision aid
usage variable to account for its potential effect on the transition probabilities.
Descriptive statistics of the key variables for each product category are presented in
Table 3.2. Note that the usage experience variables are store-level measures and thus do not vary
across categories. For each decision aid examined, we report its cumulative usage counts and as
a percentage of the total usage counts, for up to week t-1 as well as by the last week in the data
(Week 62). As the table shows, shopping list is the most frequently used decision aid (30.66
times on average in 62 weeks), far out-numbering the usage of the other types of decision aids,
followed by previous order list (5.27 times), sorting by brand name (4.06 times), sorting by
nutrition (1.93 times), sorting by price (1.17 times) and sorting by promotion (.38 times). As
shown by the standard deviations, the experience measures vary substantially across individuals.
Since we need to exclude at least one of the relative usage experience variables from the model
to avoid perfect collinearity, we choose to take out sorting by promotion because it had the
lowest occurrence amongst the six focal decision aids of interest.
[INSERT TABLES 3.1 & 3.2 HERE]
3.4.3 Time-Varying Patterns of Usage Experience with Decision Aids
To inspect how the usage experience may evolve over time, we take the average of each
relative usage experience variable in a week across households and plot them over time (see
Figure 3.1). Note that the relative usage experience measures do not necessarily sum up to 100%
in all weeks because some households had not tried any decision aids in the earlier weeks. The
relative usage experience with personal shopping list and previous order list, which implies
habitual decision processes, increases over time; while the relative usage experience with sorting
by brand name, by price, and by promotion, which are indicative of on-the-spot decision
62
processes, levels off in the later stage of the observation period. This is consistent with previous
literature which suggests that weights of price information will decrease when other product
attributes are available (Anderson 1971and 1981; Bettman, Capon, and Lutz 1975). The time-
varying patterns of the different decision aids indicate that, on average, consumers tend to adopt
more habitual behavior and make fewer on-the-spot purchase decisions as they get more
accustomed to the type of online shopping environment as studied here.
[INSERT FIGURE 3.1 ABOUT HERE]
3.4.4 Model Estimation Results
In Table 3.3, we compare the BIC, DIC, and log-marginal densities of models with
different numbers of hidden states to determine the best number of states for each category (see
Hughes, Guttorp, and Charles 1999; Netzer et al. 2008). These comparisons indicate that a three-
state model fits the data best for both spaghetti sauce and liquid detergent. Estimation results of
the three-state model for the two categories are presented in Tables 3.4 and 3.5, respectively.
[INSERT TABLES 3.3, 3.4, and 3.5 ABOUT HERE]
As shown by Table 3.4, households‘ purchase behavior differ substantially across the
three hidden states for the spaghetti sauce category. By the model construction, the baseline
purchase incidence tendency increases from hidden state 1 (S1) to hidden state 3 (S3). Our
empirical result shows that the same order of baselines also holds for purchase quantity decisions
across the three states, even though they are not constrained to be so. Price has a negative and
?significant? effect on the purchase incidence probability in all three states
5
, but its effect is the
strongest in S2 (-.620), followed by S1 (-.144), while its effect in S3 is much smaller (-.005). The
5
For the ease of exposition, hereafter, we report the posterior means in parentheses and use the term “significant” to
refer to the case where the posterior 95% credible interval does not cover zero.
63
same pattern of price sensitivity also holds for the purchase quantity decision. Thus, S1
represents a low baseline purchase tendency (i.e., store loyalty) and medium price sensitivity
state, S2 is characterized as having medium baseline purchase tendency and the highest price
sensitivity decisions, and S3 exhibits the highest baseline purchase tendency and the lowest price
sensitivity.
The three hidden states of the liquid detergent category show very similar patterns, and
can be interpreted in the same fashion.
The lower panels of Table 3.4 and Table 3.5 report parameter estimates of the variables
that affect the transition probabilities between hidden states. For the spaghetti sauce category,
when consumers are in S1, more usage experience with sorting by brand name significantly
increases the probability of staying in S1 and reduces the probability of switching to the more
store-loyal states S2 and S3 (posterior mean = -2.578), while the total decision aid usage
significantly increases the probability of switching to a more store-loyal state (posterior mean =
1.153). When consumers are in S2, more usage experiences with sorting by brand name and
sorting by price discourage the transition to S3 while encourages the transition to S1 (posterior
mean = -1.071 and -.420, respectively), while more usage experiences with sorting by nutrition
and personal shopping lists have the opposite effects (posterior mean = .410, and 1.950,
respectively). When consumers are in S3, more usage experiences with personal shopping lists
and previous order lists increase the likelihood of staying in S3 and reduce the transition
probabilities from S3 to a less store-loyal state (posterior mean = .946 and .495, respectively).
The effects of the usage experience variables exhibit a similar pattern in the liquid
detergent category. When consumers are in the least store-loyal state (S1), more usage
experience with sorting by brand name discourages switching to the states with higher levels of
64
store loyalty (S2 and S3). It is possible that consumers with strong brand preference (brand
loyalty) are willing to switch stores if the preferred brand is not available in the focal store.
When consumers are in the medium store-loyalty state (S2), usage experiences of sorting by
brand name and sorting by price increase the probability of switching to S1 and reduce the
probability of switching to the most store-loyal state (S3), while usage experiences with personal
shopping lists and previous order lists, as well as the total decision aid usage count, have the
opposite effects. When consumers are in the most store-loyal state (S3), usage experiences of
personal shopping lists and previous order lists and the total decision aid usage count reinforce
the probability of staying in this state and decrease the chance of switching to a less store-loyal
state, while usage experiences with sorting by brand name and by price do not have any
significant effects any more.
An interesting contrast with results of the spaghetti sauce category is that, for the liquid
detergent category, usage experience with sorting by nutrition does not have any effects on the
purchase behavior evolution. This difference between the two categories is expected, and it
attests to our model‘s ability detect the distinct effects (or the lack of which) of usage experience
with different decision aids.
3.4.5 Evolution of Purchase Behavior in the Online Store
Our results indicate that there are different hidden behavior states and consumers switch
among these states over time, reflecting an evolution of purchase behavior when shopping in a
new online store environment. To investigate the time varying patterns of the purchase behavior,
we need to compute the posterior distribution of the three states for each household in each
week, and then explore the relationship between price sensitivity measures and the usage
experience variables.
65
3.4.5.1 State Probability Distribution
To better understand the evolution of purchase behavior, we firstly examine the
probabilities of the states for each household over time. The filtering probability of household i
belonging to state s at week t (Netzer, et al. 2008) can be calculated as:
(13) ), , | ( / ) , , ( ) , | ( ... ) , , ( ) , | ( ) , , (
) ,..., , , ,... , | ( Pr
1 2 2 1 2 1 1 s
2 1 2 1 ,
2 1
i,1
s s c
it
s s
i
C
t
C
t t t i
s s
i
C C
i
s s
i
C C
it i i it i i t i
Q l Z X f s s s P Z X f s s s P Z X f
Q Q Q I I I s S ob
it i i
| | t
=
÷
= =
O · · O · O · =
=
where P
i
(s
t
|s
t-1,
s) is the s
th
column of transition matrix P(s
t
|s
t-1
), ) , | (
s s c
it
Q l | | is the likelihood of
the observed purchase incidence and quantity up to week t. Figure 3.2 presents the average
probabilities of belonging each state in 62 weeks. For the spaghetti sauce category, the average
probability of belonging to S1, which has the lowest store loyalty and medium price sensitivity,
decreases substantially over time (from 67.86% in week 1 to 16.01% in week 62). In contrast, the
average probabilities of belonging to S2 and S3 increase over time (S2: from 19.03% in week 1
to 30.48% in week 62; S3: from 13.11% in week 1 to 53.50% in week 62). On average,
consumers are more likely to be in S1 than in the other two states in the first half of the observed
period (till week 32). The average probability of belonging to S3, the state with the highest store
loyalty and lowest price sensitivity, surpasses that for S2, and S3 becomes the dominant behavior
state from week 32 and on. This pattern also holds for the liquid detergent category: the average
probability of belonging to S1 declines while those for S2 and S3 increase over time, and S3
becomes the dominant behavior state since week 39, for the average consumer.
[INSERT FIGURE 3.2 ABOUT HERE]
3.4.5.2 Effects of Usage Experience with Decision Aids on Price Sensitivity
Since the states are not in ascending or descending order, the direction of the changes of
price sensitivity, and the relationship between the usage experience with decision aids and the
66
price sensitivity are unknown based on the model estimation per se. Therefore, we compute a
posterior price sensitivity measure for each household in each week, and then regress this
measure on the usage experience variables. To account for uncertainty in posterior distributions
of price coefficients, we calculate a weighted posterior price sensitivity measure.
Specifically, we take the last 1,000 draws of the price coefficients, and for each draw,
price coefficients for household i in week t are weighted by its probabilities of belonging to each
corresponding hidden state in that week. Then we average the weighted price coefficients across
the 1000 draws to get the price sensitivity measure for each household across 62 weeks. The
time-varying patterns of price sensitivity measures are plotted in Figure 3.3 and 3.4, for the
spaghetti sauce and liquid detergent categories, respectively.
For the spaghetti sauce category, the average values of the price sensitivity measures for
purchase incidence and purchase quantity firstly increase then decrease over the observed period
(see the upper panels in Figure 3.3). In addition, the variances of these measures increase over
time, indicating that consumers‘ price sensitivity diverges over time, with some becoming less
price-sensitive while other becoming more price-sensitive, as they become more accustomed to
an online store environment (see the lower panels in Figure 3.3). The plots for the liquid
detergent category reveal the same patterns. Prior research shows that a higher degree of
heterogeneity in price sensitivity is conducive to more granular price promotion customization
(Zhang and Wedel 2009). This divergence in price sensitivity over time suggests that online
retailers have a good opportunity to customize their price promotions to cater to individual
consumers‘ needs and preferences, as consumers become more used to shopping online. In the
following, we will examine the impact of different decision aids on price sensitivity through
post-hoc analysis.
67
[INSERT FIGURE 3.3 and 3.4 ABOUT HERE]
Since the hidden states identified by our model are not ordered by their level of price
sensitivity, we cannot directly infer from the parameter estimates how usage experience with
each decision aids affects consumers‘ price sensitivity evolution. To investigate this issue, we
conduct regression analyses where the dependent variables are the price sensitivity measures for
the purchase incidence or the purchase quantity decision, and the explanatory variables are the
usage experience with different decision aids. We use a random-effect model to allow
heterogeneity in price sensitivity across households. The price coefficients are estimated from
the models, and thus the dependent variables are also measured with uncertainty. Since the
posterior distributions of the price coefficients are unknown, we use the simulated maximum
likelihood estimation (MLE) method to estimate the models, where the draws of the price
coefficients are a natural product of the MCMC procedures of the main model estimation. We
use the last 1,000 draws of each MCMC procedure for the simulated MLE.
Table 3.6 present the results of these regression analyses. For both categories, usage
experience with personal shopping lists and previous order lists can mitigate consumers‘ price
sensitivity in purchase incidence decisions, and make them more loyal to the focal retailer. In
contrast, usage experience with sorting by price can train consumers to become more efficient at
using price information in the liquid detergent category and thus more sensitive to it in their
purchase decisions. Usage experience with sorting by brand name also leads to higher price
sensitivity in the spaghetti sauce category. Thus, the more usage experience with such decision
aids, the more responsive consumers become to price changes.
[INSERT TABLE 3.6 ABOUT HERE]
68
3.5 Discussion
In this study, we investigate whether and how the usage experience with different
decision aids drives the evolution of purchase behavior. We empirically identified three latent
states that direct the purchase behavior over time. They are: hidden state 1where consumers
show lowest store loyalty /medium price sensitivity; hidden state 2 that characterized with
medium store loyalty /high price sensitivity; and hidden 3 that exhibits highest store loyalty /low
price sensitivity. Post-hoc analysis shows that the probabilities of staying in hidden state 1
declines over time, while the probabilities of staying in the other two hidden states demonstrate a
reverse trend. In the latter half of the 62-week period, hidden state 3 dominates. Such pattern
implies that as consumers get more accustomed to the online store environment, their baseline
tendency to purchase from the store increases. In addition, their price sensitivity diverges over
time, with some consumers becoming more price sensitive, while others becoming less price
sensitive.
The transitions among these hidden states are driven by the relative usage experience
with different decision aids. For both spaghetti sauce and liquid detergent categories, more
relative usage experience with the decision aids for brand preference (sorting by brand names)
discourages the transitions from lower store-loyal state to higher store-loyal state. More relative
usage experience with decision aids for economic needs (sorting by price) also reduces the
transition probabilities from lower store-loyal state to higher store-loyal state, in addition, post-
hoc analyses show that usage experience with sorting by price in the liquid detergent category
trains consumers to become more responsive to price changes. In contrast, more relative usage
experience with decision aids for personalized shopping list, such as shopping list and previous
order list, as well as sorting by nutritional information for the spaghetti sauce category, increase
69
the transition from medium store-loyal / high price sensitive state (hidden state 2) to high store-
loyal / low price sensitive state (hidden state 3). Post-hoc analysis also confirmed the inverse
relationship between the usage experience with personalized shopping lists and price sensitivity.
Personalized shopping lists thus help build store loyalty, increase purchase propensity, and ease
price competition.
These findings have important implications for the design of online store environments
and communication messages regarding a firm‘s pricing decisions. To encourage transition from
the more price sensitive state to less price sensitive, online retailers should allow consumers to
create shopping lists, make available their previous order lists, provide decision aids for
searching a variety of nutritional information, and encourage their usage among the customers.
The personalized shopping lists will create a lock-in effect and help consumer build store loyalty.
Decision aids aimed at economic needs, on the other hand, are a double-edged sword. A low-
price online retailer could benefit from higher consumer responsiveness to price by enabling and
promoting the usage of such type of decision aids like sorting by price, but usage of these
decision aids would lower consumers‘ loyalty to the store in the long run. In addition, their usage
could negatively affect the long-term business for retailers that adopt a premier pricing strategy
and do not compete on promotions. What decision aids to offer and to emphasize should depend
on an online retailer‘s overall positioning and pricing strategies, and weigh in the trade-offs of
the retailer‘s short-term versus long-term needs.
In terms of future research, it would be interesting to investigate whether there are carry-
over effects of usage experience across product categories, i.e., would the usage experience in
one product category affect the evolution of purchase behavior in other categories. Effects of the
store-level usage experience variables, as shown by our study, strongly suggest such possibility.
70
Given the proliferation of multi-channel retailing, another worthy direction is to study the impact
of usage experience with online decision aids on offline purchase behavior offline, which would
require matching online navigation data and offline purchase history data. In addition, how the
pattern of decision aids usage evolves over time itself is an interesting topic to explore. All these
topics offer exciting venues for gaining deeper understanding of purchase behavior evolution in
the ever-evolving retail environment.
71
Table 3.1 Types of Online Decision Aids Examined
Broad Category Decision Aid Definition
For nutritional needs Sort by Nutrition
Sort by nutritional criteria, including calories,
sodium, fat, Kosher, sugar, carbohydrates,
fiber, and cholesterol
For brand preference
Sort by Brand Name
Sort by brand name
For economic needs
Sort by Price
Sort by price information, including unit price
and item price
Sort by Promotion Sort by promotion status
Personalized
shopping list
Shopping List
Retrieve and use personal shopping list
Previous Order List
Retrieve and use previous order list
72
Table 3.2 Descriptive Statistics
Variable (mean/sd of variables across
weeks and across individuals)
Cumulative usage count up to
week t-1
As percentage of total
decision aid usage up to
week t-1 (%)
Mean S.D. Mean S.D.
Store Level Tool Usage:
--- Sort by Nutrition
--- Sort by Brand Name
--- Sort by Price
--- Sort by Promotion
--- Shopping List
--- Previous Order List
--- Total decision aid usage (six tools)
1.06
2.32
0.75
0.25
16.70
2.71
23.79
4.23
6.36
2.25
0.73
30.3
5.75
38.04
3.96
7.99
2.91
1.42
55.71
12.42
9.85
16.60
8.81
5.65
37.35
21.73
Variable (mean/sd of last week’s
measure for all individuals)
Cumulative usage count in
week 62
As percentage of total
decision aid usage in
week 62 (%)
Mean S.D. Mean S.D.
Store Level Tool Usage:
--- Sort by Nutrition
--- Sort by Brand Name
--- Sort by Price
--- Sort by Promotion
--- Shopping List
--- Previous Order List
--- Total decision aid usage (six tools)
1.93
4.06
1.17
0.38
30.66
5.27
43.47
6.30
8.52
3.03
0.91
47.18
8.92
57.17
4.55
10.29
3.43
1.50
63.67
16.56
9.19
17.74
10.83
5.34
31.36
23.67
Variable Mean S.D.
Category 1: Spaghetti Sauce (N =137 households)
Purchase frequency (times/per year)
Purchase quantity (ounces/occasion)
Regular price (cents/oz.)
Price discount (cents/oz.)
Paid price (cents/oz.)
5.63
34.62
9.03
0.31
8.64
6.15
21.23
0.37
0.25
0.42
Category 2: Liquid Detergent (N =159 households)
Purchase frequency (times/per year)
Purchase quantity (ounces/occasion)
Regular price (cents/oz.)
Price discount (cents/oz.)
Paid price (cents/oz.)
5.53
124.70
7.15
0.48
6.67
6.49
69.25
0.30
0.38
0.45
73
Table 3.3 Model Comparisons
Spaghetti Sauce Liquid detergent
Number of Hidden
States
LMD BIC DIC LMD BIC DIC
K=2 -4,786.7 9,565.0 9,430.9 -5,992.6 12,078.9 11,901.2
K=3 -4,128.4 8,426.9 8,217.3 -5,337.1 10,848.2 10,692.1
K=4 -4,593.1 9,583.3 9,492.5 -5,889.3 12,184.5 12,103.2
Note: LMD = log-marginal density;
BIC = Bayesian Information Criterion;
DIC = Deviance Information Criterion.
74
Table 3.4 Estimation Result for Spaghetti Sauce
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
Variables in the Tobit Model
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Purchase Incidence
Intercept -2.171 -3.055 -1.287 -1.787 -2.148 4.170 -1.080 -1.255 -0.847
Actual price -0.144 -0.149 -0.140 -0.620 -1.204 -0.036 -0.005 -0.007 -0.003
Purchase Quantity
Intercept 0.960 0.302 1.618 2.190 2.081 2.310 2.258 2.193 3.907
Actual price -0.279 -0.605 0.046 -0.926 -1.718 -0.135 0.001 -0.019 0.021
Variables Affecting the
Transition Probabilities
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
(relative usage experience)
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Cut-off point 1
1.248 -0.014 2.509 2.067 1.231 2.904 0.939 0.545 1.133
Cut-off point 2
1.314 1.183 1.444 2.648 -0.875 6.171 1.937 0.301 3.572
Sort by Brand Name
-2.578 -4.024 -1.131 -1.071 -1.459 -0.683 -0.621 -1.536 0.293
Sort by Price
-0.236 -1.070 0.597 -0.420 -0.597 -0.243 -0.008 -0.269 0.253
Sort by Nutrition
-0.838 -1.846 0.171 0.410 0.019 0.802 -0.585 -2.331 1.162
Shopping List
1.182 -0.051 2.414 1.950 0.928 2.972 0.946 0.351 1.540
Previous Order List
-0.460 -1.273 0.353 0.685 -0.019 1.390 0.495 0.230 0.760
Total decision aid usage
1.153 0.647 1.658 2.013 -0.008 4.035 0.489 -0.713 1.691
Note: The bold font indicates that the 95% credible interval does not cover zero.
75
Table 3.5 Estimation Result for Liquid Detergent
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
Variables in the Tobit Model
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Purchase Incidence
Intercept -0.707 -1.195 -0.219 1.228 0.102 1.027 1.384 0.637 2.130
Actual price -0.228 -2.341 1.886 -0.496 0.121 -0.733 -0.100 -0.116 -0.084
Purchase Quantity
Intercept -1.374 -1.645 -1.103 3.102 -1.164 5.368 3.925 3.052 4.798
Actual price -0.859 -2.057 0.340 -1.441 -1.571 -1.311 -0.120 -0.239 -0.001
Variables Affecting the
Transition Probabilities
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
(relative usage experience)
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Cut-off point 1
-1.310 -2.380 -0.241 -0.235 -0.812 0.342 -1.073 -1.851 -0.296
Cut-off point 2
-0.953 -1.027 -0.878 0.701 0.279 1.123 1.102 0.537 1.667
Sort by Brand Name
-0.380 -0.532 -0.229 -0.104 -0.149 -0.059 -0.053 -1.711 1.604
Sort by Price
-0.804 -1.616 0.009 -1.023 -1.627 -0.420 0.157 -1.250 1.563
Sort by Nutrition
-0.016 -0.578 0.546 0.134 -1.603 1.870 0.035 -1.440 1.510
Shopping List
0.142 -0.115 0.398 1.149 0.743 1.556 1.928 0.792 3.065
Previous Order List
0.767 -1.003 2.537 0.987 0.606 1.367 0.586 0.118 1.054
Total decision aid usage
1.466 -0.228 3.160 1.259 0.247 2.270 1.276 0.924 1.628
Note: The bold font indicates that the 95% credible interval does not cover zero.
76
Table 3.6 The Impact of Usage Experience with Decision Aids on Price Sensitivity (Simulated MLE)
Spaghetti Sauce
Purchase Incidence Purchase Quantity
Mean S.E. t-stat P-value Mean S.E. t-stat P-value
Intercept 0.113 0.006 17.523 0.000 0.228 0.008 29.497 0.000
Intercept_var 0.215 2.508 -0.612 0.542 0.361 1.730 -0.589 0.557
Sort by Brand Name -0.022 0.011 -2.066 0.041 -0.005 0.018 -0.301 0.764
Sort by Price -0.015 0.033 -0.463 0.644 -0.011 0.051 -0.213 0.832
Sort by Nutrition 0.019 0.029 0.677 0.499 0.025 0.034 0.739 0.461
Shopping List 0.032 0.008 4.112 0.000 0.027 0.015 1.781 0.077
Previous Order List 0.022 0.011 1.998 0.048 -0.002 0.009 -0.268 0.789
Liquid Detergent
Purchase Incidence Purchase Quantity
Mean S.E. t-stat P-value Mean S.E. t-stat P-value
Intercept
0.214 0.004 58.584 0.000 0.683 0.053 12.980 0.000
Intercept_var
0.218 0.424 3.586 0.000 0.204 0.213 7.443 0.000
Sort by Brand Name
-0.002 0.010 -0.209 0.835 -0.008 0.019 -0.417 0.677
Sort by Price
-0.135 0.068 1.981 0.049 -0.032 0.022 -1.456 0.147
Sort by Nutrition
0.013 0.015 0.873 0.384 0.003 0.003 1.050 0.295
Shopping List
0.058 0.014 4.056 0.000 0.787 0.037 21.327 0.000
Previous Order List
0.032 0.012 2.739 0.007 0.059 0.010 5.727 0.000
77
Figure 3.1
Time-Varying Patterns of Usage Experience with Decision Aids
78
Figure 3.2 Probabilities of Belonging to Each State
Spaghetti Sauce
Liquid Detergent
79
Figure 3.3 Evolution of Price Sensitivity over Time
Spaghetti Sauce
80
Liquid detergent
81
Chapter IV: Conclusion
This dissertation empirically investigates dynamic consumer decision processes in E-
Commerce settings. In this chapter, we will re-examine the findings from Chapter 2 and 3, points
out the contributions of this dissertation, and conclude with an exploration of possible future
directions for this stream of research.
4.1 Summary of the Two Essays
Essay one uses eye tracking data to investigate information acquisition processes on
attribute-by-product matrices as encountered in online comparison websites. The study identifies
two information acquisition modes (attribute-based or product-based), and documents the extent
of usage and switching of these acquisition modes through Hierarchical Hidden Markov Model
(HHMM). Post-hoc analyses focus on the specific information processed in these modes, the
possible causes and consequences of switching, and the dynamics in the comparison set during
decision making. We find that the low-level properties of the eye and the visual brain (horizontal
and local eye movement tendency) stimulate congruent information processing mode and
contiguous information acquisition, reflecting a stronger influence on information acquisition
processes than previously documented in the traditional process-tracing studies. In addition, we
find frequent switching between acquisition modes when consumers face information-rich
comparison websites. The extensive switching might be explained from the previously obtained
information, and limited visual short term memory. Higher switching frequency significantly
increases decision time, and reduces the experienced ease-of-processing.
Essay two conducts an empirical investigation on whether and how usage experience
with various types of decision aids affects the evolution of online purchase behavior. By
82
adopting Non-homogeneous Hidden Markov Model (NHMM), we empirically identified three
latent behavior states that govern the evolution of online purchase behavior and the drivers
behind the evolvement of behavior. As consumers get more accustomed to the online store
environment, their baseline tendency to purchase from the store increases, and their price
sensitivity diverges over time. More relative usage experience with shopping lists, previous order
list, and sorting by nutrition (in Spaghetti sauce category), encourages the transition from
medium store-loyal/high price sensitivity state to high store-loyal /low price sensitivity state; yet
more relative usage experience with sorting by brand name and sorting by price reduces the
probability of switching from lower store-loyal state to higher store-loyal state. Post-hoc analysis
reveals that more usage experience with sorting by brand name in the spaghetti sauce category
and sorting by price in the liquid detergent category increase price sensitivity, while personalized
shopping lists are important decision aids that help retailers build store loyalty and soften the
price competition.
4.2 Contribution and Managerial Implications
This dissertation has made the following contributions:
Firstly, we take advantage of recent development in technology and investigate the
dynamic decision making process with eye tracking data and online navigation data. These path
data opens up the possibilities to exploit micro-level dynamics during decision making. In the
first study, we rely on eye tracking equipment to accurately document moment to moment
information search activities when consumers making their purchase on comparison websites.
The eye tracking equipment is able to capture fast, adaptive, partially unconscious information
acquisition behavior during real-life decision making. This is different from traditional process
tracing methods such as information display boards. Prior process tracing methods require motor
83
responses, which renders the information acquisition process more controlled and deliberate. As
a consequence, these studies could discover mostly high-level and slower cognitive processes
that consumers engage in during decision making. Thanks to the eye movement data, our study is
able to offer a sneak peek into both ?overt and covert?, and ?voluntary and involuntary? aspects
of decision process (Lynch and Srull 1982). Low-level properties of the eye and the visual brain
that are underexploited in previous accounts of decision process, appear to play an important role
when decision-making is fast and more automatic. Insight into these covert and involuntary
information tendencies will lead to a better understanding of decision-making in information-rich
online choice environments.
In the second study, we collect online navigation data and purchase history data to
investigate purchase behavior evolution. What is lacking in the literature is a comprehensive,
longitudinal examination of the evolving patterns of purchase behavior, and more importantly,
the drivers behind purchase behavior changes in online stores. Our study fills this void by
extracting the information of decision aids usage from online navigation data and using it to
explain the evolution of purchase behavior. Navigation data provides valuable information of
consumers‘ online activities. The usage experience with decision aids extracted from navigation
data reveals the information search and evaluation process before purchase occurs. It impacts the
weights of different attributes including price, the extent of brand and store loyalty, the
importance of personalized shopping environment, etc. The usage experience data collected over
time reflects learning and adaptation processes that consumers go through when shop in a new
channel. This information is helpful in explaining the evolution of purchase propensity and
responsiveness to marketing mix, yet difficult to obtain in offline settings.
84
Secondly, we adopt a flexible evolutionary structure to model decision dynamics. Eye
movement data is dense, stochastic, and highly dynamic. It reflects the latent information
acquisition modes probabilistically, rather than deterministically. The HHMM in essay one deals
with large volumes of eye movement data, unobserved information modes, and a probabilistic
link between these two through structural restrictions. The model effectively copes with the
challenge to extract diagnostic information from the massive amount of eye movement data. The
three hierarchical layers in HHMM represent longer range dependencies in eye-movements, and
quantitatively capture prevalence and flexible switching pattern of information acquisition
modes.
The classification of different purchase behaviors has been a long-standing research
interest in the marketing literature. In essay two, we propose that the observed online shopping
behavior is likely to be directed by certain latent ?behavior states?, and these hidden states differ
in terms of the baseline tendency to purchase from the online store and their price sensitivity.
NHMM automatically identifies three latent behavior states, and allows flexible transitions
among these behavior states. In addition, the non-homogeneous structure in NHMM enables us
to explore the drivers behind the evolutions of these hidden states, and reveals how the store
environment trains consumers and changes their behavior states over time.
Managerially, our results have implications for the design of online shopping websites
which are gaining popularity recently. Retailers and manufacturers try to optimize online
information displays to affect consumers‘ information acquisition processes and their purchase
decision. The first study shows that low-level, automatic and unconscious processes are principal
in situations of information overload and time pressure, which occur when facing data-rich web-
based choice environments. Comparison website developers could display formats in a proactive
85
fashion to stimulate consumers‘ use of specific information acquisition mode favorable to their
goals. Dominance of local information acquisition requires careful allocation of attributes /
products. By strategically placing the contiguous information, retailers may be able to alternate
the attractiveness of certain attribute. A delicate balance among format, sales objective, and
switching cost is also needed. For example, we find that that a Product-Column format invites
more attributed-based acquisition and increases the chance of the dominant product being
chosen; while a Product-Row format leads to more evenly distributed choice probabilities.
However, the latter format leads to more local eye movements and switching between acquisition
modes, smaller numbers of products and attributes being inspected, and reduced ease of decision
making. In addition, some comparison websites that enable sorting by certain attribute may need
to reconsider their default format, which is the Product-Row format. This format may be in
conflict with desirable attribute-based information processing for these websites.
The second study reveals that the more usage experience with personalized shopping list,
and sorting by nutrition (in the spaghetti sauce category) leads to the transition from medium
store-loyal/high price sensitivity behavior state to high store-loyal/low price sensitivity behavior
state; while more usage experience with decision aids for economic needs and brand preference
discourages the transition from lower store-loyal states to higher store-loyal states. If online
retailers could encourage consumers to create and store personal shopping lists, make available
their previous order lists, and encourage their usage among the customers, it will facilitate the
transition from the more price sensitive state to the less price sensitive state, create a lock-in
effect, build store loyalty, and train consumers into habitual shoppers. Retailers should be
careful with the offering of decision aids aimed at economic needs, however. A low-price-driven
online retailer could benefit from higher consumer responsiveness to price by enabling and
86
promoting the usage of such type of decision aids, yet the long-term business may be harmed.
Our study also provides some insights for online retailers to offer personalized shopping
environment. Information of previous usage of decision aids and purchase behavior collected in
the navigation data enables retailers to infer the distribution of different latent behavior states for
individual customers. Retailers could customize the marketing mix variables based on the
purchase propensity and price sensitivity of each behavior states; and encourage the usages of
certain decision aids to influence customer‘s purchase evolution path.
4.3 Future Research
We hope this dissertation will stimulate further research interest in the dynamic decision
making process in the online shopping environment. The first study raises several intriguing
questions, for instance, how to effectively reduce frequent switching between different
information acquisition modes? How do other design factors, such as sorting, grouping,
trimming, and highlighting, affect the dynamic information acquisition processes and post-
decision evaluations? Based on information acquisition patterns, can comparisons websites
automatically adapt display format during decision making and increase the conversion rate?
The second study also offer several exciting research opportunities, for instance, what
drives the evolution of decision aids usage itself over time? How does the usage experience with
decision aids online affect the purchase behavior offline? Will the usage experience with one
product category affect the evolution of purchase behavior in the other category (?spill-over?
effect)? All these future research questions will help us enrich our understandings of dynamic
decision making in E-commerce.
Besides these two essays, we conduct another study of dynamic decision making on a
special type of websites: search engines. Specifically, we collect eye movement data on a major
87
search engine website that includes organic, right sponsored and top sponsored sections, over a
variety of search tasks, and investigate section and snippet inspection decisions that online
shoppers engaged in when search for product or service information on search engine results
page. We explore the impact of two types of stimuli on information processes: one is high level
stimuli, i.e., textual information; the other is low level stimuli, such as snippet location, density,
and section intrinsic property. The integration of newly acquired textual information with prior
knowledge, and the lag effect of low level stimuli, create a context that constantly changes as
decision progress. Therefore, sequential inspection decisions are dynamic and interrelated. We
apply a computational cognitive model with static utilities and dynamic utilities generated from
two types of stimuli to capture this preference formation process. Our results show that
sponsored sections create a strong ?stickiness? effect: once consumers enter these sections, they
are likely to inspect more than one snippet within the same section instead of leaving
immediately. In addition, the impact of snippet location on inspection probabilities varies across
organic and sponsored sections. In terms of high level stimuli, descriptive information, such as
product attribute or quality, decreases consumers‘ ?stickiness? with the current section, while
transactional information, such as price and promotion, has an opposite effect. The study furthers
the understanding of the effect of different types of stimuli, especially the dynamic influence
from snippet content, on the information seeking behavior from a consumer‘s viewpoint. It offers
insights in designing snippet‘s textual information on a search engine result page with respect to
the location and content of its competitors. Implications for the platform (search engines) are
also quite broad – from textual result selection, to textual content pairing, to optimal blending.
88
Appendices
Appendix I Transition Probabilities between Middle and Upper Layer States for the Two
Information Presentation Formats (with standard deviations in parentheses)
Product Column Format
Product Row Format
Middle Layer Transition Matrices
S
1
=AB S
1
=PB
S
1
=AB S
1
=PB
S
1
=AB 0.769 (0.021) 0.231 (0.021)
S
2
=1
S
1
=AB 0.746 (0.047) 0.231 (0.047)
S
1
=PB 0.096 (0.038) 0.904 (0.038) S
1
=PB 0.047 (0.011) 0.953 (0.011)
S
1
=AB S
1
=PB S
1
=AB S
1
=PB
S
1
=AB 0.586 (0.029) 0.414 (0.029)
S
2
=2
S
1
=AB 0.719 (0.036) 0.281 (0.036)
S
1
=PB 0.456 (0.071) 0.544 (0.071) S
1
=PB 0.552 (0.087) 0.448 (0.087)
S
1
=AB S
1
=PB
S
1
=AB S
1
=PB
S
1
=AB 0.520(0.056) 0.480 (0.056)
S
2
=3
S
1
=AB 0.323 (0.019) 0.677 (0.019)
S
1
=PB 0.451(0.013) 0.549 (0.013) S
1
=PB 0.334 (0.038) 0.666 (0.038)
Upper Layer Transition Matrices
S
2
=1 S
2
=2 S
2
=3 S
2
=1 S
2
=2 S
2
=3
0.461 (0.099) 0.244 (0.012) 0.295 (0.090) S
2
=1 0.549 (0.047) 0.161 (0.088) 0.290 (0.070)
0.225 (0.067) 0.400 (0.045) 0.375 (0.034) S
2
=2 0.204 (0.099) 0.493 (0.054) 0.303 (0.101)
0.209 (0.111) 0.446 (0.091) 0.345 (0.089) S
2
=3 0.187 (0.087) 0.496 (0.026) 0.317 (0.095)
Note: AB: attribute-based acquisition; PB: product-based acquisition
89
Appendix IIa Estimated Attribute Transition Matrices
product-column condition : attribute transition probabilities product-row condition: attribute transition probabilities
picture price processor
operating
sys
memory
keyboard
mouse
monitor
hard
drive
optical
drive
wireless
office
software
warranty picture price processor
operating
sys
memory
keyboard
mouse
monitor
hard
drive
optical
drive
wireless
office
software
warranty
0.016 0.420 0.271 0.055 0.055 0.038 0.024 0.040 0.015 0.024 0.025 0.018 picture 0.018 0.584 0.175 0.048 0.033 0.041 0.030 0.031 0.011 0.010 0.012 0.008
0.424 0.004 0.370 0.068 0.024 0.019 0.023 0.007 0.026 0.017 0.007 0.011 price 0.421 0.005 0.325 0.057 0.046 0.022 0.040 0.022 0.031 0.012 0.012 0.007
0.205 0.304 0.007 0.349 0.041 0.032 0.020 0.015 0.009 0.004 0.005 0.008 processor 0.121 0.402 0.004 0.307 0.050 0.029 0.010 0.030 0.018 0.010 0.007 0.010
0.090 0.065 0.311 0.005 0.399 0.058 0.016 0.020 0.010 0.007 0.007 0.013
operating
sys
0.087 0.103 0.253 0.005 0.419 0.065 0.018 0.016 0.007 0.013 0.011 0.004
0.071 0.060 0.096 0.276 0.005 0.391 0.031 0.026 0.007 0.006 0.016 0.015 memory 0.074 0.128 0.149 0.173 0.006 0.322 0.076 0.040 0.009 0.009 0.011 0.004
0.094 0.022 0.036 0.042 0.301 0.004 0.395 0.063 0.014 0.013 0.013 0.003
keyboard
mouse
0.095 0.061 0.068 0.060 0.210 0.005 0.411 0.048 0.008 0.008 0.012 0.015
0.100 0.029 0.020 0.019 0.044 0.293 0.006 0.418 0.021 0.016 0.019 0.015 monitor 0.094 0.085 0.071 0.032 0.071 0.178 0.005 0.325 0.075 0.027 0.029 0.008
0.076 0.016 0.029 0.015 0.016 0.023 0.333 0.013 0.376 0.085 0.008 0.008
hard
drive
0.038 0.083 0.067 0.017 0.022 0.049 0.313 0.007 0.336 0.018 0.031 0.020
0.023 0.019 0.014 0.016 0.011 0.033 0.050 0.348 0.013 0.385 0.075 0.014
optical
drive
0.049 0.033 0.042 0.015 0.013 0.016 0.070 0.202 0.007 0.430 0.075 0.047
0.066 0.024 0.027 0.030 0.024 0.022 0.020 0.038 0.273 0.011 0.385 0.080 wireless 0.025 0.026 0.039 0.010 0.011 0.023 0.018 0.075 0.342 0.007 0.355 0.069
0.073 0.018 0.019 0.006 0.028 0.028 0.025 0.006 0.032 0.292 0.009 0.465
office
software
0.012 0.031 0.041 0.029 0.025 0.006 0.026 0.015 0.072 0.245 0.007 0.492
0.050 0.039 0.073 0.056 0.024 0.041 0.029 0.037 0.068 0.105 0.463 0.015 warranty 0.028 0.085 0.101 0.028 0.030 0.048 0.050 0.077 0.058 0.100 0.381 0.014
Appendix IIb Estimated Product Transition Matrices
Product-column condition:
product transition probabilities
product-row condition:
product transition probabilities
Label Product 1 Product 2 Product 3 Product 4 Label Product 1 Product 2 Product 3 Product 4
0.005 0.767 0.155 0.028 0.045 Label 0.004 0.747 0.159 0.079 0.010
0.186 0.008 0.521 0.193 0.092 Product 1 0.250 0.010 0.508 0.149 0.084
0.037 0.410 0.004 0.418 0.130 Product 2 0.037 0.339 0.006 0.499 0.120
0.007 0.157 0.308 0.008 0.520 Product 3 0.039 0.176 0.308 0.010 0.466
0.042 0.168 0.287 0.486 0.017 Product 4 0.053 0.190 0.224 0.509 0.024
90
Appendix III Model Estimation Test on Synthetic Data
Variables Starting Value
Estimated Value
(posterior mean)
Real Value
PI_0_S1 0.1 0.22875228 0.2
PI_MKT1_S1 0.1 0.61499769 0.5
PI_MKT2_S1 0.1 -0.45321166 -0.3
PI_0_S2 0.1 0.39564543 0.3
PI_MKT1_S2 0.1 0.47681012 0.4
PI_MKT2_S2 0.1 -0.38364839 -0.2
PI_0_S3 0.1 0.54535731 0.5
PI_MKT1_S3 0.1 0.80136447 0.8
PI_MKT2_S3 0.1 -0.61044463 -0.6
PQ_0_S1 0.1 0.66825696 0.5
PQ_MKT1_S1 0.1 0.31771122 0.7
PQ_MKT2_S1 0.1 -0.91325732 -0.8
PQ_0_S2 0.1 0.99265948 0.9
PQ_MKT1_S2 0.1 1.31265599 1.2
PQ_MKT2_S2 0.1 -0.95266036 -0.9
PQ_0_S3 0.1 0.58022575 0.8
PQ_MKT1_S3 0.1 1.15305817 0.9
PQ_MKT2_S3 0.1 -1.42524140 -1.2
TL1_S1 0.1 0.68878361 0.8
TL2_S1 0.1 -0.31458116 -0.5
TL1_S2 0.1 -0.81875761 -0.9
TL2_S2 0.1 0.32136565 0.5
TL1_S3 0.1 0.87714547 0.7
TL2_S3 0.1 0.25091591 0.2
Gamma_S1 0.1 0.84548087 0.9
Gamma_S2 0.1 1.36810117 1.5
Gamma_S3 0.1 0.90530073 1.2
Rho_S1 0.1 0.55360011 0.5
Rho_S2 0.1 0.78094405 0.7
Rho_S3 0.1 0.80169427 0.9
Sigma[1] 0.1 0.26430815 0.2
Sigma[2] 0.2 0.49261638 0.4
Sigma[3] 0.1 0.50207708 0.4
Sigma[4] 0.2 0.69167942 0.7
Sigma[5] 0.1 0.60207708 0.8
Sigma[6] 0.2 1.04969463 1.0
V_theta[1,1] 1.0 0.98644001 1.0
V_theta[2,2] 1.0 1.05860775 1.0
V_theta[3,3] 1.0 1.02430815 1.0
V_theta[4,4] 1.0 0.89679048 1.0
V_theta[5,5] 1.0 0.95966898 1.0
V_theta[6,6] 1.0 0.91105910 1.0
91
Bibliography
Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard Lutz, Alan
Sawyer, and Stacy Wood (1997), ?Interactive Home Shopping: Consumer,
Retailer, And Manufacturer Incentives to Participate in Electronic
Marketplaces,? Journal of Marketing, 61 (3), 38-53.
Amemiya, Tomohiro (1984), ?Tobit Models: A Survey,? Journal of Econometrics, 24
(1-2), 3-61.
Anderson, Norman H. (1971), ?Integration Theory and Attitude Change,?
Psychological Review, 78(3), 171-206.
Anderson, Norman H. (1981), Foundations of Information Integration Theory. New
York: Academic Press.
Ansari, Asim, Carl Mela, and Scott A. Neslin (2008), ?Customer Channel Migration?,
Journal of Marketing Research, 45(1), 60-76.
Babin, Barry J., William R. Darden, and Mitch Griffin (1994), ?Work and/or Fun:
Measuring Hedonic and Utilitarian Shopping Value,? Journal of Consumer
Research, 20(4), 644-656.
Bakos, Yannis (2001), ?The Emerging Landscape for Retail E-Commerce,? Journal
of Economic Perspectives, 15(1), 69-80.
Ball, Christopher (1997), ?A Comparison of Single-Step and Multi-Step Transition
Analyses Of Multiattribute Decision Strategies,? Organizational Behavior Human
Decision Processes, 69 (3), 195-204.
Baldi, Pierre, Yves Chauvint, Tim Hunkapiller, and Marcella A. Mcclure (1994),
?Hidden Markov Models of Biological Primary Sequence Information?,
Proceedings of the National Academy of Science of the United States of America,
91, 1059-1063
Bechwati, Nada Nasr, and Lan Xia (2003), ?Do Computers Sweat? The Impact of
Perceived Effort of Online Decision Aids on Consumers' Satisfaction with The
Decision Process,? Journal of Consumer Psychology, 13( ½), 139-148.
Bettman, James. R., and Jacob Jacoby (1976), ?Patterns Of Processing In Consumer
Information Acquisition,? Advances in Consumer Research, 3(4), 315-320.
———, Noel Capon, and Richard J. Lutz (1975), ?Cognitive Algebra in Multi-
Attribute Attitude Models,? Journal of Marketing Research, 12 (2), 151-64.
——— (1979), An Information Processing Theory of Consumer Behavior. Addison-
Wesley. Reading MA.
92
———, Pradeep Kakkar (1977), ?Effects of Information Presentation Format on
Consumer Information Acquisition Strategies,? Journal of Consumer Research,
3(4), 233-240.
———, Mary F. Luce, John W. Payne (1998), ?Constructive Consumer Choice
Process,? Journal of Consumer Research, 25(3), 187-217.
———, C.Whan. Park (1980), ?Effects of prior knowledge and experience and phase
of the choice process on consumer decision processes: A protocol analysis,?
Journal of Consumer Research, 7(3), 234-248.
Block, Lauren B. and Vicki G. Morwitz (1999), ?Shopping Lists as an External
Memory Aid for Grocery Shopping: Influences on List Writing and List
Fulfillment,? Journal of Consumer Psychology, 8 (4), 343-75.
Brangule-Vlagsma, Kristine, Rik G. M. Pieters
,
and Michel Wedel (2002), ?The
Dynamics of Value Segments: Modeling Framework And Empirical
Illustration?, International Journal of Research In Marketing, 19(3),267-285
Bucklin Randolph E. ,and Catarina Sismeiro (2003), ?A Model of Web Site Browsing
Behavior Estimated on Clickstream Data?, Journal of Marketing Research,
40(3), 249-267
Chen, Lei-da, Mark L. Gillenson, and Daniel L. Sherrell (2002), ?Enticing Online
Consumers: An Extended Technology Acceptance Perspective,? Information
and Management, 39 (8), 705-719.
Cheung, Christy M. K., Lei Zhu, Timothy Kwong, Gloria W.W. Chan, and Moez
Limayem (2003), ?Online Consumer Behavior: A Review and Agenda for
Future Research,? Proceedings of the 16th Bled eCommerce Conference, 194-
218.
Chib, Siddhartha (1995), ?Marginal Likelihood from the Gibbs Output,” Journal of
the American Statistical Association, 90(432), 1313-1321.
Childers, Terry L., Christopher L. Carr, Joann Peck, and Stephen Carson (2001),
?Hedonic and Utilitarian Motivations for Online Retail Shopping Behavior,?
Journal of Retailing, 77 (4), 511-535.
Cowan, Nelson (2001) ?The Magical Number 4 in Short-Term Memory: A
Reconsideration of Mental Storage Capacity,? Behavior and Brain Science, 24,
87–185.
Danaher, Peter J., Isaac W. Wilson, and Robert A. Davis (2003), ?A Comparison of
Online and Offline Consumer Brand Loyalty,? Marketing Science, 22 (4), 461–
76.
93
De Jong, Ritske (2000), ?An Intention-Activation Account Of Residual Switch
Costs,? Control of Cognitive Processes: Attention and Performance eds. S.
Monsell and J. Driver. Vol. XVIII 357–376. Cambridge, MA: MIT Press.
Degeratu, Alexandru M., Arvind Rangaswamy, and Jianan Wu (2000), ?Consumer
Choice Behavior in Online and Traditional Supermarkets: The Effects of Brand
Name, Price, And Other Search Attributes,? International Journal of Research
in Marketing, 17 (1), 55-78.
Dekimpe Marnik G., and Dominique M. Hanssens (2000), ?The Persistence of
Marketing Effects On Sales,? Marketing Science, 14 (1), 1-21.
Dekimpe Marnik G., and Dominique M. Hanssens (2000), ?Time-Series Models In
Marketing: Past, Present and Future,? International Journal of Research in
Marketing, 17(2-3), 183-193
Du, Rex Y., and Wagner A. Kamakura (2006), ?Household Life Cycles and
Lifestyles In The United States,? Journal of Marketing Research, 43(1), 121-
132.
Erdem, Tulin. (1996), ?A Dynamic Analysis of Market Structure Based On Panel
Data,? Marketing Science, 15(4), 359-378.
Fine, Shai, Yoram. Singer, and Naftali Tishby ( 1998), ?The Hierarchical Hidden
Markov Model: Analysis and Applications,? Machine Learning, 32, 41-62.
Gelfand, Alan. E., and Adrian F. M. Smith (1990), ?Sampling based approaches to
calculating marginal densities,? Journal of American Statistical Association,,
85, 398-409.
Gilchrist, Iain.V., and Monika Harvey (2006), ?Evidence for A Systematic
Component Within Scan Paths In Visual Search. Visual Cognition, 14
(4/5/6/7/8), 704-715.
Glaholt, Mackenzie G., and Eyal M. Reingold (2009), ?Stimulus Exposure and Gaze
Bias: A Further Test Of The Gaze Cascade Model,? Attention, Perception &
Psychophysics, 71, 445-450.
Gopher, Daniel, Lilach Armony and Yaakov Greenshpan (2000), ?Switching Tasks
And Attention Policies,? Experimental Psychology, 129 (3), 308-339.
Guadagni, Peter M., and John D.C. Little (1983),?A Logit Model of Brand Choice
Calibrated on Panel Data,? Marketing Science, 2(3), 203-238
Häubl, Gerald and Valerie Trifts (2000), ?Consumer Decision Making in Online
Shopping Environments: The Effects of Interactive Decision Aids,? Marketing
Science, 19 (1), 4–21.
94
Hauser John R., and Kenneth J. Wisniewski (1982), ?Dynamic Analysis of Consumer
Response to Marketing Strategies?, Management Science, 28(5), 455-486.
Heckman, James J. (1981), ?Heterogeneity and State Dependence,? in Studies in
Labor Markets, S. Rosen (Eds.) University of Chicago Press, Chicago 91-139.
Heilman, Carrie M., Douglas Bowman, and Gordon P. Wright (2000), ?The
Evolution of Brand Preferences and Choice Behavior of Consumers New to A
Market,? Journal of Marketing Research, 37 (2), 139-155.
Heller, Katherine A., Yee Whye The, and Dilan Görür (2009), ? Infinite Hierarchical
Hidden Markov Models,? Proceedings of the 12th International Conference on
Artificial Intelligence and Statistics (AISTATS), 5, Clearwater, Florida.
Hirschman, Elizabeth C. and Morris B. Holbrook (1982), ?Hedonic Consumption:
Emerging Concepts, Methods and Propositions,? Journal of Marketing, 46 (3),
92-101.
Hoffman, Donna L. and Thomas P. Novak (1996), ?Marketing in Hypermedia
Computer-Mediated Environments: Conceptual Foundations,? Journal of
Marketing, 60 (3), 50-68.
Hollander, Stanley C. and Kathleen M. Rassuli (1999), ?Shopping with Other
People‘s Money: The Marketing Management Implications of Surrogate-
Mediated Consumer Decision Making,? Journal of Marketing, 63 (2), 102-118.
Huber, Joel, John W. Payne, and Christopher Puto (1982), ?Adding Asymmetrically
Dominated Alternatives: Violations of Regularity and The Similarity
Hypothesis,? Journal of Consumer Research, 9(1), 90-98.
Hui Sam K., Eric T. Bradlow, and Peter S. Fader (2009a), ?Testing Behavioral
Hypotheses Using An Integrated Model Of Grocery Store Shopping Path And
Purchase Behavior,? Journal of Consumer Research, 36, 478-493.
Hui Sam K., Eric T. Bradlow, and Peter S. Fader (2009b), ?Path Data in Marketing:
An Integrative Framework and Prospectus for Model Building,? Marketing
Science, 28(2), 320-335.
Hughes, James P. (1993), ?A Class of Stochastic Models for Relating Synoptic
Atmospheric Patterns to Local Hydrologic Phenomena,? PhD dissertation.
University of Washington. Seattle.
Hughes, James P., and Peter Guttorp (1994a), ?A Class of Stochastic Models for
Relating Synoptic Atmospheric Patterns to Regional Hydrologic Phenomena,?
Water Resource Research, 30(5), 1535–1546.
95
Hughes, James P., and Peter Guttorp (1994b), ?Incorporating Spatial Dependence and
Atmospheric Data In A Model Of Precipitation,? Journal of Applied
Meteorology, 33(12),1503–1515.
Hughes, James P., Peter Guttorp, and Stephen P. Charles (1999), ?A Non-
homogeneous Hidden Markov Model for Precipitation Occurrence,? Journal of
the Royal Statistical Society: Series C (Applied Statistics), 48(1), 15-30.
Inman, J. Jeffrey, Russell S. Winer, and Rosellina Ferraro (2009), ?The Interplay
among Category Characteristics, Customer Characteristics, and Customer
Activities on In-Store Decision Making,? Journal of Marketing, 73 (5), 19-29.
Johnson, Eric J., Michael Schulte-Mecklenbeck, and Martijn C. Willemsen (2008)
?Process Models Deserve Process Data: Comment On Brandstatter. Gigerenzer,
And Hertwig,? Psychological Review, 115(1), 263-273.
Kawanaka, Daiki, Takayuki Okatani, And Koichiro Deguchi , (2006), ?HHMM
Based Recognition of Human Activity?, IEICE TRANSACTIONS on
Information and Systems, E89-D (7), 2180-2185
Kupiec, J Julian (1992), ?Robust Part of Speech Tagging Using A Hidden Markov
Model,? Computer Speech & Language, 6, 225-242.
Lee, Byung-Kwan, and Wei-Na Lee (2004), ?The Effect of Information Overload on
Consumer Choice Quality In An Online Environment,? Psychology and
Marketing, 21(3), 159–183.
Lee, Jinkook and Loren V. Geistfeld (1998), ?Enhancing Consumer Choice: Are We
Making Appropriate Recommendations? ? Journal of Consumer Affairs, 32(2),
227-251.
Liechty, John, Rik Pieters, and Michel Wedel (2003), ?Global and Local Covert
Visual Attention: Evidence from A Bayesian Hidden Markov Model,?
Psychometrika, 68(4), 519-541.
Limayem, Moez, Mohamed Khalifa, and Anissa Frini (2000), ?What Makes
Consumers Buy from Internet? A Longitudinal Study of Online Shopping,?
IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and
Humans, 30 (4), 421-432.
Lohse, Gerald, and Eric Johnson (1996), ?A Comparison of Two Process Tracing
Methods for Choice Tasks,? Proceedings of the 29th Annual Hawaii
International Conference on System Sciences, 86-97.
Luck, Steven J., and Edward K. Vogel (1997), ?The Capacity of Visual Working
Memory For Features And Conjunctions,? Nature, 390, 279–281.
96
Lynch, John G., and Thomas K. Srull (1982), ?Memory and Attentional Factors in
Consumer Choice: Concepts and Research Methods,? Journal of Consumer
Research, 9 (1), 18-37.
MacDonald, Iain L., and Walter Zucchini (1997), Hidden Markov and Other Models
for Discrete-Valued Time Series. Chapman and Hall, London
Marois, Rene´, and Jason Ivanoff (2005), ?Capacity Limits of Information Processing
In The Brain,? Trends in Cognitive Sciences, 9(6), 296-305.
Mathwick, Charla and Edward Rigdon (2004), ?Play, Flow, and the Online Search
Experience,? Journal of Consumer Research, 31 (2), 324-332.
Mauldin, Elaine and Variram Arunachalam (2002), ?An Experimental Examination
of Alternative Forms of Web Assurance for Business-To-Consumer E-
Commerce,? Journal of Information Systems, 16 (1), 33-55.
Montgomery, Alan L., Shibo Li, Kannan Srinivasan, and John C. Liechty (2004),
?Modeling Online Browsing And Path Analysis Using Clickstream Data,?
Marketing Science, 23(4), 579-595.
Nag, R., K. Wong, F. Fallside (1986), ?Script Recognition Using Hidden Markov
Models,? Acoustics Speech and Signal Process, 2071-2074.
Netzer, Oded, James Lattin, and V. Srinivasan (2008), ?A Hidden Markov Model of
Customer Relationship Dynamics,? Marketing Science, 27, 185-204.
Park, C. Whan, Easwer S. Iyer, and Daniel C. Smith (1989), ?The Effects of
Situational Factors on In-Store Grocery Shopping Behavior: The Role of Store
Environment and Time Available for Shopping,? Journal of Consumer
Research, 15 (4), 422-33.
Pavlou, Paul A. (2003), ?Consumer Acceptance of Electronic Commerce: Integrating
Trust and Risk with the Technology Acceptance Model,? International Journal
of Electronic Commerce, 7 (3), 101-134.
Payne, John W. (1976), ?Task Complexity And Contingent Processing In Decision
Making: An Information Search And Protocol Analysis,? Organizational
Behavior Human Decision Processes, 16(2), 366-387.
———, James R. Bettman, and Eric J. Johnson (1993), The Adaptive Decision
Maker, Cambridge University Press, Cambridge, UK.
Pieters, Rik G.M., and Luk Warlop (1999), ?Visual Attention during Brand Choice:
The Impact Of Time Pressure And Task Motivation,? International Journal of
Research in Marketing, 16, 1-16.
97
Pieters, Rik G.M., and Michel Wedel (2004), ?Attention Capture and Transfer in
Advertising: Brand, Pictorial, and Text-Size Effects,? Journal of Marketing,
68(2), 36-50.
Pieters, Rik G.M., Michel Wedel, and Jie Zhang (2007), ?Optimal Feature
Advertising Design Under Competitive Clutter,? Management Science, 53(11),
1815-1828.
Rabiner, Lawrence R., and B. H. Juang (1986), ?An Introduction to Hidden Markov
Models,? IEEE ASSP Magazine, 3(1), 4-16.
Rabiner, Lawrence R. (1989), ?A Tutorial on Hidden Markov Models and Selected
Applications in Speech Recognition?, Proceedings of the IEEE, 77(2), 257-286.
Rayner, Keith, (1998), ?Eye Movement in Reading and Information Processing: 20
Years of Research,? Psychological bulletin, 24(3), 372-422.
Robert, Christian P., Gilles Celeux, and Jean Diebolt (1993), ?Bayesian Estimation
Of Hidden Markov Chains: A Stochastic Implementation,? Statistics and
Probability Letters, 16, 77-83.
Robertson, Andrew W., Sergey Kirshner, and Padhraic Smyth (2004), ?Downscaling
Of Daily Rainfall Occurrence Over Northeast Brazil Using A Hidden Markov
Model,? Journal of Climate, 17(22) 4407-4424.
Robertson, Andrew W., Amor V. M. Ines, and James W. Hansen (2007), ?
Downscaling Of Seasonal Precipitation For Crop Simulation,? Journal of
Applied Meteorology and Climatology, 46, 677-693.
Rogers, Robert D., and Stephen Monsell (1995), ?The Cost of A Predictable Switch
between Simple Cognitive Tasks,? Journal of Experimental Psychology,
General, 124, 207–231.
Rosbergen, Edward , Rik Pieters,and Michel Wedel (1997), ?Visual Attention to
Advertising: A Segment ?Level Analysis?, Journal of Consumer Research,
24(3), 305-314.
Rossi, Peter E., and Greg. M. Allenby (2003), ?Bayesian Statistics and Marketing,?
Marketing Science, 22(3), 304–328.
Russo, J. Edward, and Larry D. Rosen (1975), ?An Eye Fixation Analysis of
Multialternative Choice?, Memory and Cognition, 3, 267-276.
——— (1978), ?Adaptation of Cognitive Processes to the Eye Movement System,?
In Eye Movements and the Higher Psych. Functions. eds. John W. Senders et al.
Hillsdale. NJ: Erlbaum,89-109.
98
——— , and France Leclerc (1994), ?An Eye-Fixation Analysis of Choice Process
For Consumer Nondurables,? Journal of Consumer Research, 21(2), 274-290.
Salvucci, Dario D., and John R. Anderson (1998), ?Tracing Eye Movement Protocols
With Cognitive Process Models,? In Proceedings of the Twentieth Annual
Conference of the Cognitive Science Society, Hillsdale. NJ:Erlbaum, 923-928.
Scott, Steven L (2002), ?Bayesian Methods For Hidden Markov Models: Recursive
Computing In The 21st Century,? Journal of American Statistic Association,
97(457), 337-351.
Seetharaman, P. B., Andrew Ainslie, and Pradeep K. Chintagunta (1999),
?Investigating Household State Dependence Effects Across Categories,?
Journal of Marketing Research, 36(4), 488-500
Senter, Stuart M., and Douglas H. Wedell (1999), ?Information Presentation
Constraints and the Adaptive Decision Maker Hypothesis,? Journal of
Experimental Psychology, 23(2), 428-446.
Shimojo, Shinsuke, Claudiu Simion, Eiko Shimojo, and Christian Scheier (2003),
?Gaze Bias Both Reflects and Influences Preference,? Nature Neuroscience,
6(12), 1317-1322.
Skounakis, Marios, Mark Craven, and Soumya Ray (2003), "Hierarchical Hidden
Markov Models for Information Extraction", Proceedings of the 18th
International Joint Conference on Artificial Intelligence, 427-433.
Smith, Michael D., Joseph Bailey, and Erik Brynjolfsson (2000), Understanding
Digital Markets. E. Brynjolfsson, B. Kahin, eds. MIT Press, Cambridge, MA.
Spalek, Thomas M, and Sherief Hammad (2005), ?The Left-To-Right Bias in
Inhibition of Return Is Due to the Direction of Reading,? Psychological Science,
16(1), 15-18.
Stewart, Stewart, Nick Chater, and Gordon D.A. Brown (2006), ?Decision by
Sampling,? Cognitive Psychology, 53, 1-26.
Swait, Joffre, and Wiktor Adamowicz (2001), ?The Influence of Task Complexity on
Consumer Choice: A Latent Class Model Of Decision Strategy Switching?,
Journal of Consumer Research, 28(1), 135-148.
Swedowsky, Maya (2009), ?Online Grocery Shopping: Ripe Timing for Resurgence,?
The Nielsen Company. (www.nielsen.com, accessed on April 20, 2010)
Tatler, Benjamin W., and Benjamin T. Vincent (2008), ?Systematic Tendencies In
Scene Viewing,? Journal of Eye Movement Research, 2(2), 1-18.
99
Treisman, Anne M., and Garry Gelade (1980), ?A Feature-Integration Theory of
Attention,? Cognitive Psychology, 12 (1), 97–136.
Underhill, Paco (2004), Call of the Mall: The Geography of Shopping. Simon &
Schuster, New York
Van der Lans, Ralf, Rik Pieters, and Michel Wedel (2008), ?Eye Movement Analysis
Of Search Effectiveness,? Journal of American Statistical Association,
103(482), 452-461.
Wedel, Michel, and Rik Pieters (2000), ?Eye Fixation on Advertisements and
Memory for Brands: A Model and Findings,? Marketing Science, 19(4), 297-
312.
Wedell, Douglas H., and Stuart Senter (1997), ?Looking and Weighting in Judgment
and Choice,? Organizational Behavior and Human Decision Processes, 70(1),
41-64.
Zhang, Hua-Ping, Hong-Kui Yu, De-Yi Xiong and Qun Liu (2003), HHMM-based
Chinese lexical analyzer ICTCLAS, Proceedings of the second SIGHAN
workshop on Chinese language processing, 17, 184-187
Zhang, Jie and Aradhna Krishna (2007), ?Brand-Level Effects of Stockkeeping Unit
Reductions,? Journal of Marketing Research, 44 (4), 545-559.
Zhang, Jie and Michel Wedel (2009), ?The Effectiveness of Customized Promotions
in Online and Offline Stores,? Journal of Marketing Research, 46 (2), 190–206.
Zhou, Lina, Liwei Dai,
and Dongsong Zhang (2007), ?Online Shopping Acceptance
Model – A Critical Survey of Consumer Factors in Online Shopping,? Journal
of Electronic Commerce Research, 8 (1), 41-62.
doc_327947935.pdf
A consumer is a person or group of people who are the final users of products and or services generated within a social system. A consumer may be a person or group, such as a household. The concept of a consumer may vary significantly by context, although a common definition is an individual who buys products or services for personal use and not for manufacture or resale.
ABSTRACT
Title of Document: DYNAMIC CONSUMER DECISION
MAKING PROCESS IN
E-COMMERCE
Wei Shi
PhD Candidate in Marketing, 2011
Directed By: Dr. Michel Wedel
PepsiCo Professor of Consumer Science
Marketing Department,
Robert H. Smith School of Business
3303 Van Munching Hall
This dissertation studies the dynamic decision making process in E-commerce. In the first
essay, we use eye tracking to investigate how consumers make information acquisition decisions
on attribute-by-product matrices in online choice environment such as comparison websites.
Hierarchical Hidden Markov Model is used to describe this process. The model consists of three
connected hierarchical layers: a lower layer that describes the eye movements, a middle layer
that identifies product- and attribute-based information acquisition modes, and an upper layer
that flexibly captures switching between these modes over time. Findings of a controlled
experiment show that low-level properties of the eye and the visual brain play an important role
in dynamic information acquisition. Consumer switch frequently between two acquisition modes,
and higher switching frequency increases decision time and reduces easiness of decision making.
These results have implications for web design and online retailing, and may open new
directions for research and theories of online choice.
The second essay investigates how usage experience with different types of decision aids
contributes to the evolution of online shopping behavior over time. In the context of online
grocery stores, we categorize four types of decision aids that are commonly available, namely,
those 1) for nutritional needs, 2) for brand preference, 3) for economic needs, and 4)
personalized shopping lists. We construct a Non-homogeneous Hidden Markov Model of
category purchase incidence and purchase quantity, in which parameters are allowed to vary over
time across hidden states as driven by usage experience with different decision aids. The dataset
was collected during the period when the retailer first launched its web business, which makes it
particularly suited to study the evolution of online purchase behavior. We estimate the model for
the spaghetti sauce and liquid detergent categories. Results indicate that four types of decisions
influence evolution of purchase behavior differently. Findings from this study enrich the
understanding of how purchase behavior may evolve over time in online stores, and provide
valuable insights for online retailers to improvement the design of their store environments.
DYNAMIC CONSUMER DECISION MAKING PROCESS IN E-COMMERCE
By
Wei Shi
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
2011
Advisory Committee:
Dr. Michel Wedel, Chair
Dr. Jie Zhang
Dr. Michael Trusov
Dr. P.K. Kannan
Dr. Ginger Zhe Jin
© Copyright by
Wei Shi
2011
ii
Acknowledgements
I am heartily thankful to my advisor, Dr. Michel Wedel, whose encouragement,
caring, guidance and support enabled me to explore the exciting new subject and
complete this dissertation. The joy and enthusiasm he has for research is contagious
and motivational for me, even during tough times in the Ph.D. pursuit. I would like to
thank Dr. Jie Zhang, who sets an example of a great researcher for her rigor and
passion on research. She has assisted, advised, and supported my research and writing
efforts over the last five years. I am also grateful for Dr. Michael Trusov, who has
generously given his time and expertise to better my work, and helped me to gain a
different perspective to the research problem. Special thanks goes to Dr. P.K. Kannan
and Dr. Ginger Zhe Jin, who graciously agreed to serve on my defense committee. I
am also in debt to Dr. Rik Pieters, whose remarkable insights and valuable
suggestions have contributed greatly to the completion of this dissertation. My
deepest gratitude goes to my parents for their unconditional love and support
throughout my life; this dissertation is simply impossible without them.
iii
Table of Contents
Acknowledgements ....................................................................................................... ii
Table of Contents ......................................................................................................... iii
List of Tables ................................................................................................................ v
List of Figures .............................................................................................................. vi
Chapter I: General Introduction .................................................................................... 1
Chapter II: Information Acquisition during Online Choice: A Model-Based
Exploration .................................................................................................................... 7
2.1. Introduction ........................................................................................................ 8
2.1.1 Modeling Information Acquisition Processes ............................................ 10
2.2 Information Acquisition through Eye movements ............................................ 12
2.2.1 Horizontal Eye Movement Patterns ........................................................... 12
2.2.2 Local Eye Movement Patterns ................................................................... 13
2.2.3 Switching between Information Acquisition Strategies ............................ 13
2.3. Model Formulation .......................................................................................... 15
2.3.1 Motivation .................................................................................................. 15
2.3.2 Specification .............................................................................................. 15
2.3.3 Model Estimation and Testing ................................................................... 18
2.4 Information Acquisition on A Comparison Website ........................................ 19
2.4.1 Experimental Procedure ............................................................................. 19
2.4.2 Eye movement Recording .......................................................................... 20
2.5 Results ............................................................................................................... 20
2.5.1 Model Comparisons ................................................................................... 20
2.5.2 General Estimation Results ........................................................................ 21
2.5.3 Switching between Information Acquisition Strategies ............................ 26
2.5.4 Attention to Products and Attributes.......................................................... 27
2.6. Discussion ........................................................................................................ 29
Chapter III: Usage Experience with Decision Aids and Evolution of Online Purchase
Behavior ...................................................................................................................... 44
3.1. Introduction ...................................................................................................... 45
3.2 Conceptual Development and Literature Review ............................................. 48
3.2.1 Online Decision Aids ................................................................................. 48
3.2.2 Evolution of Online Shopping Behavior.................................................... 50
3.2.3 Usage Experience with Decision Aids and Online Shopping Behavior
Evolution ............................................................................................................. 51
3.2.4 Potential Hidden States of Purchase Behavior ........................................... 53
3.3 Model Formulation ........................................................................................... 54
3.3.1 Type-II Tobit Model of Category Purchase Incidence and Quantity ........ 54
3.3.2 Hidden States and Transition Probabilities ................................................ 56
3.3.3 Prior Distributions and Estimation Method ............................................... 58
3.4 Empirical Analyses ........................................................................................... 59
3.4.1 Data Description ........................................................................................ 59
iv
3.4.2 Operationalization of Key Variables ......................................................... 60
3.4.3 Time-Varying Patterns of Usage Experience with Decision Aids ............. 61
3.4.4 Model Estimation Results .......................................................................... 62
3.4.5 Evolution of Purchase Behavior in the Online Store ................................. 64
3.5 Discussion ......................................................................................................... 68
Chapter IV: Conclusion .............................................................................................. 81
4.1 Summary of the Two Essays............................................................................. 81
4.2 Contribution and Managerial Implications ....................................................... 82
4.3 Future Research ................................................................................................ 86
Appendices .................................................................................................................. 88
Appendix I ............................................................................................................ 88
Appendix IIa ........................................................................................................... 89
Appendix IIb ........................................................................................................... 89
Appendix III ........................................................................................................... 90
Bibliography ............................................................................................................... 91
v
List of Tables
Table 2.1 ……………………………………………………………………... 33
Table 2.2 ……………………………………………………………………... 33
Table 3.1 ……………………………………………………………………... 71
Table 3.2 ……………………………………………………………………... 72
Table 3.3 ……………………………………………………………………... 73
Table 3.4 ……………………………………………………………………... 74
Table 3.5 ……………………………………………………………………... 75
Table 3.6 ……………………………………………………………………... 76
vi
List of Figures
Figure 2.1 ……………………………………………………………………... 34
Figure 2.2 ……………………………………………………………………... 35
Figure 2.3 ……………………………………………………………………... 36
Figure 2.4 ……………………………………………………………………... 37
Figure 3.1 ……………………………………………………………………... 77
Figure 3.2 ……………………………………………………………………... 78
Figure 3.3 ……………………………………………………………………... 79
1
Chapter I: General Introduction
This dissertation is composed of two essays in the discipline of dynamic consumer
decision making in E-commerce. The global E-commerce has experienced robust growth in
recent years, and its revenue is predicted to reach 680 billion dollars by the end of 2011, a 18.9%
increase from 2010
1
. Online shopping websites have become popular channels in which
consumers acquire product information and make purchase decisions. Shopping behavior in
online stores has been shown to be systematically different from that in offline stores (e.g.,
Danaher, Wilson, and Davis 2003, Degeratu, Rangaswamy, and Wu 2000, Zhang and Wedel
2009), and such observed behavioral discrepancy can be partially attributed to the differences in
shopping environments (Zhang and Wedel 2009). The availability and display of product-
attribute information, and the design of interactive decision aids, may all affect consumers‘
information search and evaluation processes, purchase decision making, and the evolution of
shopping behavior in online stores. In this dissertation, we intend to empirically investigate the
impact of the online shopping environment on consumers‘ purchase decision processes, and how
consumers adapt to this increasingly prominent channel. Essay one focuses on the dynamic
information acquisition decisions on comparison shopping websites, and essay two studies the
evolution of purchase decisions in online grocery stores. In both situations, consumers‘ behavior
at one stage may influence their behavior at a later stage. We empirically examine how
consumers dynamically adjust their decision strategies based on the newly acquired information
or experience, in the web-based choice environment. Findings from this dissertation will enrich
the understanding of the dynamic decision making in the information-rich online shopping
1 E-Commerce Report from JP Morgan senior analyst Imran Kahn, http://techcrunch.com/2011/01/03/j-p-morgan-
global-e-commerce-revenue-to-grow-by-19-percent-in-2011-to-680b/
2
environment, and provide valuable insights for online retailers to improve the design of their
store environments.
Dynamic decision making is a field with long tradition in marketing. Constructive
decision making theory proposes that during the decision process, consumers incorporate new
information, develop new standard and construct the decision contingent on task environment
(Bettman and Park1980; Bettman, Luce and Payne 1998). Therefore, decision strategies over
time are interrelated and the decision process is dynamic. Various models have been developed
to study this phenomenon. Time series models, such as autoregressive (AR) and moving
average (MA) models, are used to reveal the over-time impact of marketing variables and make
forecast of purchase behavior (e.g., Dekimpe and Hanssens 2000). The Vector Autoregression
(VAR) model is suitable for studying dynamic interactions (e.g., Dekimpe and Hanssens 1995).
State dependence models are developed to accommodate dynamics in purchase behavior by
including three types of variables: lagged choices such as brand loyalties; lagged marketing
variables such as decaying effects of advertising; and serially correlated error terms in the
random utility function that account for reasons unknown to the researchers (e.g., Erdem 1996;
Guadagni and Little 1983; Heckman 1981, Seetharaman, Ainslie, and Chintagunta 1999). All
these models have made important contributions in studying dynamics of marketing phenomena.
Thanks to recent development in technology, dynamics in the decision making process
has been approached from a new angle: the use of path data (Hui, Bradlow, and Fader 2009a;
2009b). Path data is defined as the ?consumers‘ movement in a spatial configuration?, and
records ?consumers‘ interaction with his environment to achieve his goal? (Hui et al. 2009b).
Grocery shopping paths, eye tracking data, online navigation data, etc., all belong to this
category. The beauty of path data lies in its ability to capture the information search and
3
evaluation processes during decision making, which enables researchers to exploit micro-level
dynamics, and thus offers exciting new research opportunities for this topic.
Researchers have acknowledged the importance of path data in deepening the
understanding of decision processes: Underhill (2004) makes recommendation on shopping
environment design based on the analysis of consumers‘ shopping paths; Hui et al. (2009a)
develop an integrated individual level probability model where consumers‘ entire shopping path
is used to predict store visit and purchase probabilities. Eye-movement paths are another type of
path data that gaining ground in study decision dynamics, especially in the advertising research
(Pieters and Warlop 1999; Rayner 1998; Rosbergen, Pieters, and Wedel 1997). Researchers use
eye trackers to analyze consumers‘ eye fixation paths on advertisements and optimize the design
of them (Pieters and Wedel 2004; Pieters, Wedel and Zhang 2007). Online navigation data offers
great detail on navigation behavior during decision making. It records activities that reflect
consumers‘ preference, provides insights in how consumers search and evaluate information
before purchase, and therefore enables researchers to better interpret and predict online shopping
behavior. Bucklin and Sismeiro (2003) use previous websites visit depth and breadth (numbers
of websites visited) to explain web page browsing strategies. Montgomery and his colleagues
(2004) use sequence of the page viewed to predict the subsequent online navigation path. In this
dissertation, we focus on two types of path data, namely, eye movement paths (essay 1), and
online navigation paths (essay 2), and explore how the information extracted from the path data
help to explain dynamic information acquisition decisions and the evolution of purchase
decisions.
Following up on calls for a flexible evolutionary structure to represent decision dynamics
(Erdem 1996; Hauser and Wisniewski 1982; Heilman , Bowman, and Wright 2000), we provide
4
two approaches for modeling dynamic decision making processes, both fall under the hidden
Markov model framework. Hidden Markov Model (HMM) is a popular probabilistic model for
sequence data analysis. The fundamental premise of HMM is that there several hidden states
governing the observed behavior, and the transitions among these hidden states follow a Markov
process. The probabilistic model effectively extracts diagnostic information from sequence data,
and automatically assigns the observed behavior into latent states, taking uncertainty into
account. The HMM and its extensions are able to flexibly capture the dependence and dynamics
in the decision making process, and has been widely used in a variety of disciplines (e.g.,
Rabiner and Juang 1986; Baldi, Hunkapiller, and Chauvin 1994; Kupiec 1992), including
marketing (Brangule-Vlagsma, Pieters, and Wedel 2002; Du and Kamakura 2006; Montgomery,
et al. 2004; Netzer, Lattin and Srinivasan 2008; Liechty, Pieters and Wedel 2003; Van derLans,
Pieters and Wedel 2008).
Eye movements are fast, adaptive, and partially automatic. The information acquisition
strategies that direct the eyes in searching for information, however, are latent cognitive states. In
addition, the link between eye movement and the unobservable information acquisition strategies
is probabilistic rather than deterministic (Lohse and Johnson 1996; Wedel and Pieters 2000).
Therefore, in essay one, we propose a new Hierarchical Hidden Markov Model (HHMM) for the
analysis of dense eye-movement data. The proposed model extends the standard HMM by
adding another layer of latent states. It is a special case of Hierarchical Hidden Markov Model
(HHMM; Fine, Singer, and Tishby 1998; Heller, The, and Görür 2009), which has been used in
the recognition of human activity (e.g., Kawanaka, Okatani, And Deguchi, 2006), information
extraction, (e.g., Skounakis, Craven, and and Ray 2003), and lexical graphic analysis (e.g.,
Zhang, Yu, Xiong, and Liu, 2003). Three layers in HHMM impose a flexible structure on the
5
information acquisition processes: a lower layer that describes the eye movements, a middle
layer that identifies product-based and attribute-based information acquisition modes, and an
upper layer that captures switching between these modes over time. The model and eye
movement data enable us to diagnose the influence of both conscious strategies and low-level
properties of the eye and the visual brain on dynamic information processes. The paper
empirically investigates the extent of usage and switching of different information acquisition
strategies, how much and what information is processed in these strategies, and the causes and
consequences of strategy switching.
In essay two, we modify the standard HMM by allowing hidden state transition
probabilities to change across time through time-varying covariates, and propose Non-
homogeneous Hidden Markov Model (NHMM) to study how the usage of different decision aids
drives the purchase behavior evolution in online grocery store. Scholars have applied NHMM in
Meteorology and Climatology (Hughes 1993, Hughes and Guttorp 1994a; 1994b); Robertson,
Kirshner, and Smyth 2004; Robertson, Ines, and Hansen 2007), and recently, the model is used
in studying customer relationship dynamics (Netzer et al. 2008). NHMM is able to identify
purchase behavior patterns, which has been a long-standing research interest in the marketing
literature. It is especially suitable for studying drivers behind the behavior evolvement. The
results reveal how consumers switch among behavior states that differ in purchase propensity
and responsiveness to marketing variables, more importantly, how the usage experience with
decision aids for economic needs, non-economic needs, brand preference and personalized
shopping list affects transitions among latent behavior states differently.
6
The organization of this dissertation is as follows: Chapters 2 and 3 discuss the two
essays in depth, respectively; Chapter 5 briefly summarizes each essay, points out the
contributions of this dissertation, and concludes with possible future research avenues.
7
Chapter II: Information Acquisition during Online Choice:
A Model-Based Exploration
Wei Shi
Robert H. Smith School of Business
University of Maryland
Michel Wedel
Robert H. Smith School of Business
University of Maryland
Rik Pieters
Department of Marketing
University of Tilburg
8
2.1. Introduction
Retailers and manufacturers are undergoing competitive pressure to assist consumers in
the large number of choices of products and services that they make online. During the often
short time that it takes to make such a decision, consumers acquire information on the
alternatives presented to them and the way in which they do this affects their decision. This is
recognized by retailers and manufacturers who try to optimize online information displays, in
particular comparison sites that have become popular tools for consumers (for example,
Bizrate.com, Dell.com, and Nextag.com each has over 10 million monthly visitors
2
).
In the academic literature, process tracing methods have been proven important in
helping to understand how consumers process information (Bettman and Jacoby 1976; Bettman
and Kakkar 1977; Johnson, Schulte-Mecklenbeck, and Willemsen 2008; Lohse and Johnson
1996; Payne 1976; Payne, Bettman, and Johnson 1993; Senter and Wedell 1999). Research has
revealed two key information acquisition strategies: attribute-based and product-based (Ball
1997; Bettman, Luce, and Payne 1998; Payne et al. 1993). During attribute-based acquisition,
information is acquired on a single attribute across multiple products, before going to the next
attribute. During product-based acquisition, information is acquired on a single product across
multiple attributes, before going to the next product. Findings on these two strategies, however,
have been obtained with research methods such as information display boards or Mouselab, in
which information becomes available sequentially as consumers manually inspect cells in
attribute-by-product displays. The required motor responses necessarily slow down the decision
process, and render it more controlled and deliberate (Bettman et al. 1998). As a consequence,
2
“15 Most Popular Comparison Shopping Websites|March 2011”, from http://www.ebizmba.com/articles/shopping-
websites.
9
these prior studies could discover mostly high-level and slower cognitive processes that
consumers engage in during decision making.
Comparison sites provide all information on the choice alternatives simultaneously. Then,
decisions can be made much faster and information acquisition is more likely under influence of
both conscious information acquisition strategies and low-level properties of the eye and the
visual brain. Lynch and Srull (1982) call for a methodology that is capable of diagnosing both
?overt and covert?, and ?voluntary and involuntary? aspects of decision processes. Russo (1978),
Russo and Leclerc (1994) and Lohse and Johnson (1996) have argued that in such situations eye
tracking methodology would be ideally suited to provide insights into how consumers acquire
information.
Affordable eye tracking systems are now widely available. Prior eye tracking studies
have often displayed products in a holistic, for instance pictorial, fashion (Lohse and Johnson
1996; Pieters and Warlop 1999; Russo and Rosen 1975; Russo and Leclerc 1994), rather than in
the form of attribute-by product displays. These latter displays reflect online decision contexts,
and increase the potential insights that can be obtained from eye tracking studies. But they also
increase the challenge of providing meaningful descriptions of the massive amounts of data that
the eye tracking studies produce.
To illustrate this, Figure 2.1 shows the eye movements of a single participant while
making a choice, in our study to be described later (participant 36). The decision took only 76
seconds, but involved 140 eye fixations and saccades (excluding re-fixations). Eye fixations are
brief moments that the eye is still and information is extracted (about 2 to 4 times per second).
Saccades are rapid jumps of the eye between fixations to redirect the line of sight to a new
location. During saccades no information acquisition takes place. In the figure, attribute-based
10
and product-based information acquisition strategies are not clearly discernable: the eyes
switched 101 times between making horizontal and vertical movements. Switching seems
extensive and the question is whether this is due to unpredictability in how people move their
eyes, or whether it is a fundamental property of the underlying information acquisition process.
[INSERT FIGURE 2.1 ABOUT HERE]
The present study uses such eye tracking data to investigate information acquisition on
attribute-by-product matrices as encountered in online choice environments. We intend to
provide a description of how people acquire information in these contexts, investigating the
influence of low-level processes, and the extent of usage of and switching between attribute- and
product-based information acquisitions. To deal with challenges posed by the amount of data, we
develop and test a new hierarchical Hidden Markov model of information acquisition.
2.1.1 Modeling I nformation Acquisition Processes
Valid inferences about information acquisition strategies should recognize that they are
fundamentally unobservable. Eye movements reflect them probabilistically rather than
deterministically (Lohse and Johnson 1996; Wedel and Pieters 2000). The information
acquisition strategies that we are interested in are latent cognitive states that direct the eyes in
searching for information. They can only be inferred from the observed eye movement paths.
Whereas describing the patterns of observed eye fixations and saccades (as we did in the earlier
example may) already be insightful, inferring and describing the latent cognitive states is
preferable (Johnson et al. 2008). To be able to identify when during a decision participants
acquire information by-attribute and when by-product, we develop a model that builds on and
extends previous applications of Hidden Markov Models (HMM) to eye movement data
(Liechty, Pieters, and Wedel 2003; Van der Lans, Pieters and Wedel 2008).
11
We extend previous two-layer HMMs by developing a new model that describes
information-acquisition strategies through three hierarchical layers (see Figure 2.2): (1) the lower
layer describes the observed eye movements, (2) the middle layer describes unobserved
(attribute- and product-based) information acquisition strategies, and (3) the upper layer
describes how consumers switch between them. The model captures this moment-to-moment
switching pattern through a number of upper layer states that are a-priori unknown. This allows
the transition probabilities between information acquisition strategies to vary over time, and
overcomes the time-homogeneity of two-level HMMs. The model enables us to identify from
moment-to-moment what the probabilities are that participants process information by-product or
by-attribute.This makes it possible to investigate post-hoc how often participants switch between
attribute-based and product-based information acquisition, and what specific information they
process while doing so. Studying information acquisition processes in this manner is not only of
theoretical interest, but also provides insights into the usability of comparison websites.
[INSERT FIGURE 2.2 ABOUT HERE]
We first provide the theoretical foundation and formulation of our model. Then we
describe the eye movement experiment of decision-making on comparison websites. In the
experiment, we manipulate the information presentation format and provide participants a
comparison website in either a product-row or a product-column format. These are the two main
presentation formats used for comparison sites (e.g., the popular comparison site Bizrate.com
uses a product-row format by default, and Dell.com uses a product-column format). It has been
previously shown that these formats may facilitate specific information acquisition strategies
(Bettman and Kakkar 1977). We then present the results of the model estimation, which reveal
the prevalence of low-level processes, and the extent of switching between the two information
12
acquisition strategies. We conduct an extensive post-hoc analysis to investigate why participants
switch and how much and what information they process in these states. Finally, we discuss
contributions and future research directions.
2.2 Information Acquisition through Eye movements
Prior studies have described high-level cognitive processing during decision-making
(Bettman et al. 1998). Yet, when decisions are fast and more automatic, low-level properties of
the eye and the visual brain may influence the information acquisition process more. Here, we
focus on three aspect of information acquisition: tendencies of horizontal eye movements, of
local eye movements, and of switching between information acquisition strategies. Insight into
these factors will lead to a better understanding of decision-making in information-rich online
choice environments.
2.2.1 Horizontal Eye Movement Patterns
Orderly sequences of eye movements reflect that individuals use systematic information
acquisition strategies. In various visual tasks a horizontal, left-right dominance in consumers' eye
movements has been observed, such as when reading (Rayner 1998) or searching store shelves
(van der Lans et al. 2008). The dominance of horizontal eye movements may be due to the
horizontal layout of many visual displays (Tatler and Vincent 2008). If the information presented
is largely textual (as is typical for comparison websites), this should facilitate such reading-like
patterns even more. But left-to-right eye-movemement tendencies also appear to be independent
of the layout or features of the display and are caused by the more rapid decline in the resolution
of the retina in the vertical than in the horizontal direction (Gilchrist & Harvey 2006).
Information acquisition is thus facilitated when the display is organized consistent with this left-
13
to-right direction (Spalek and Hammad 2005). Therefore, we expect to find a predominant
tendency for left-to-right visual information acquisition, independent of display format.
2.2.2 Local Eye Movement Patterns
The amount of visual detail that can be acquired is highest in the small area around the
exact eye fixation point and rapidly declines towards the visual periphery (Rayner 1998).
Because much of what is present in the periphery is not clearly visible, people need to make eye
saccades when searching for information. It is as if a small attentional "spotlight" moves across
the display, with information being acquired mostly locally within the focus of the spotlight
(Treisman and Gelade 1980). Using a HMM, Liechty et al. (2003) found that information
acquisition from print advertisements indeed took place in a pattern of bursts of multiple
fixations with short saccades on small areas, occasionally separated by one or a few long
saccades to distant areas. A similar pattern is likely to emerge during decision-making on
product-by-attribute displays, where information is mostly textual. Thus, we expect eye
movements – whether attribute-based or product-based – to be mostly confined to contiguous
cells on the information display.
2.2.3 Switching between I nformation Acquisition Strategies
Figure 2.3 presents a hypothetical product-attribute matrix. In attribute-based acquisition
information is gathered on a certain attribute across products, before processing the next
attribute. Solid arrow A expresses this. In product-based acquisition, information is gathered on
attributes of a particular product before evaluating the next product, as dashed arrow B shows.
The two acquisition strategies are comprised of elementary steps (Ball 1997). We propose that
these are reflected in the eye movements.
14
We thus distinguish three elementary eye movements (saccades) between cells of the
comparison matrix. A type-1 saccade is an elementary step in attribute-based information
strategy: the eye jumps between two products for the same attribute (dotted arrow 1 in Figure
2.3). A type-2 saccade is an elementary step in product-based information strategy: the eye
jumps between two attributes for the same product (dotted arrow 2 in Figure 2.3). A type-3
saccade may have several functions: the eye jumps from a particular attribute for one product to
a different attribute for another product. This may reflect a transition between type-1 and 2 eye
movements, an exploratory eye movement (Liechty et al. 2003), or may be a corrective eye
movement (Rayner 1998). An attribute-based information acquisition strategy thus consists of a
sequence of mostly (type-1) attribute-based steps; a product-based information acquisition
strategy is characterized by a sequence of mostly (type 2) product-based steps. Precisely how
many elementary steps or eye movements would constitute such a strategy is not known.
Consumers switch between information acquisition strategies due to a variety of factors,
including task demands (Ball 1997; Pieters and Warlop 1999; Swait and Adamowicz 2001),
experienced accuracy-effort tradeoffs (Bettman et al. 1998) and, importantly, the information
acquired up to that point (Bettman and Park 1980; Russo and Rosen 1975). Correspondingly,
they will switch eye movement directions. Although switching between strategies enables
adaptive decision making, it is also effortful (Gopher, Armony, and Greenshpan 2000; de Jong
2000; Rogers and Monsell 1995) and may reduce the perceived ease of the decision process. To
our knowledge, the extent of strategy switching in natural tasks such as online choice, and its
causes and consequences, have not yet been quantified.
[INSERT FIGURE 2.3 ABOUT HERE]
15
2.3. Model Formulation
2.3.1 Motivation
Our model to identify attribute- and product-based information acquisition is a
Hierarchical Hidden Markov Model (HHMM; Fine, Singer, and Tishby 1998). This model
generalizes Hidden Markov Models (HMM; Rabiner, and Juang 1986), which have been used in
marketing (Du and Kamakura 2006; Montgomery, Srinivasan, and Liechty 2004; Netzer, Lattin,
and Srinivasan 2008), and eye tracking research (Van derLans et al. 2008; Liechty et al. 2003;
Salvucci and Anderson 1998). The link between the observed eye movements and unobservable
information acquisition strategies is probabilistic. For instance, a jump from one attribute to
another in a longer sequence of attribute-based steps has a lower probability to reflect product-
based acquisition, than a similar jump that occurs in a longer sequence of product-based steps.
Therefore, our model needs to capture longer sequences of eye movements. It does this through
three hierarchically related Markov processes, generalizing two-layer HMMs that are constrained
by stationary transition probabilities (Figure 2.2). The lower layer captures (low level) eye
movements; the middle layer of the model captures the attribute-based and product-based
acquisition strategies that drive the eye movements, and the upper layer of the model captures
switching between the two types of acquisition. We intend to use the model as a flexible tool to
describe eye movement patterns and to identify the acquisition strategies that participants use
from moment-to-moment during decision making. We describe and explain those acquisition
strategies in detail, post-hoc.
2.3.2 Specification
We let i = 1,…, I denote participants and t = 1,…, T
fixations. We let ) , (
, , , t i t i t i
p a y =
denote the particular cell in that display defined by attribute (a) and product (p), at eye-fixation t
16
for participant i. In addition,
) , (
1 , 1 , 1 , ÷ ÷ ÷
=
t i t i t i
p a y denotes the cell at the previous eye-fixation t-1.
Figure 2.2 presents the hierarchical relationship among the three layers of the model: the lower
layer describes eye movements y, the middle layer captures information acquisition states
S
1
=1,..., N, and the upper layer states S
2
=1,...,M describe switching between these acquisition
states. Correspondingly, the model is represented by three sets of transition probabilities: the
lower layer transition probabilities ) | (
1 , ,
1
,
÷ t i t i
S
y y P
t i
, the intermediate layer transition probabilities
) | (
1
1 ,
1
,
2
÷
H
t i t i
S
S S , and the upper layer transition probabilities ) | (
2
1 ,
2
, ÷
O
t i t i
S S . These are described
next.
We formulate the middle layer hidden states S
1
=1 and S
1
=2 such that they represent,
respectively, attribute-based and product-based acquisition strategies (Figure 2.2). We assume
) | ( ) | ( ) | (
1 , , 1 , , 1 , , ÷ ÷ ÷
· =
t i t i t i t i t i t i
p p P a a P y y P , given the unobserved state, with ) | (
1 , , ÷ t i t i
y y P the
saccade probability from cell
1 , ÷ t i
y to cell
t i
y
,
,
) | (
1 , , ÷ t i t i
a a P the saccade probability from attribute
1 , ÷ t i
a to attribute
t i
a
,
, and ) | (
1 , , ÷ t i t i
p p P the saccade probability from product
1 , ÷ t i
p to
product
t i
p
,
. Each of the two middle layer states (attribute-based acquisition and product-based
acquisition, S
1
) has a different set of associated saccade probabilities ) | (
1 , , ÷ t i t i
a a P
and ) | (
1 , , ÷ t i t i
p p P .
For state 1 of the middle layer (S
1
=1) we impose a structure on the saccade probabilities
between products that allows for eye movements that are mostly, but not necessarily, consistent
with strict attribute-based acquisition. We assume that at any time the eyes have a probability p
of continuing to move from one product to another along the same attribute, with
probability ) | (
1 , , ÷ t i t i
p p P . These are eye movements consistent with an attribute-based acquisition
strategy. But, with a small probability (1-p) the participant may also make elementary eye
17
movements that are inconsistent with that strategy (types 2 and 3). In the product-based
acquisition state (S
1
=2) the eyes have a probability q of moving from one attribute to the other
along the same product, with probability ) | (
1 , , ÷ t i t i
a a P . This is an eye movement consistent with
product-based acquisition. The eyes, however, also have a small probability of (1-q) of making
inconsistent moves (types 1 and 3). This formulation renders the model conservative in its
identification of switching between the underlying acquisition strategies, because it allows eye
movements to deviate from strict attribute-based and product-based acquisition strategies. These
assumptions yield the following set of equations for the lower-level saccade probabilities:
1 , , 1 , , 1 , ,
2
1 , , 1 , , 1 , ,
2
1 , , 1 , , 1 , ,
1
1 , , 1 , , 1 , ,
1
) | ( ) | (
) 1 ( ) | ( ) | (
) | ( ) | (
) 1 ( ) | ( ) | (
1
1
1
1
÷ ÷ ÷
=
÷ ÷ ÷
=
÷ ÷ ÷
=
÷ ÷ ÷
=
= · =
= ÷ · =
= · =
= ÷ · =
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
t i t i t i t i t i t i
S
p iff p q a a P y y P
p iff p q a a P y y P
a iff a p p p P y y P
a iff a p p p P y y P
(1)
Because we are interested in the extent to which attribute-based and product-based
information acquisition is local (on contiguous cells), or global (on noncontiguous cells) (Liechty
et al. 2003), we re-parameterize the attribute saccade probabilities ) | (
1 , , ÷ t i t i
a a P , and the product
saccade probabilities ) | (
1 , , ÷ t i t i
p p P , as probabilities on spatially contiguous versus spatially
noncontiguous attributes and products.
The switching between the attribute-based and product-based states of the middle layer
(S
1
=1,2) follow a Markov process that is governed by the upper-layer hidden states (S
2
) (see
Figure 2.2). These upper layer states can be conceived of as regimes that govern the middle layer
switching patterns. Given an upper layer state (S
2
), the transitions among middle layer hidden
states (S
1
) follow a Markov process with a transition probability matrix ) | (
1
1 ,
1
,
2
÷
H
t i t i
S
S S . That is,
the upper-layer state influences how participants switch between the attribute-based and product-
18
based (middle layer) states. The transitions between the upper-layer hidden states (S
2
) also
follow a Markov process described by the transition probability matrix ) | (
2
1 ,
2
, ÷
O
t i t i
S S . While the
number of states of the middle layer is dictated to equal two by theory (attribute-based and
product-based acquisition), the number of states of the upper layer is a-priori unknown.
We can now write the saccade probabilities between cells of the comparison matrix,
1 , ,
|
÷ t i t i
y y , conditional on the states of the two hidden layers that participant i is in at fixation t,
1
,t i
S and
2
,t i
S , as:
) | ( ) | ( ) | ( ) , ; | (
1 , ,
1
1 ,
1
,
1
2
1 ,
2
,
1
2
,
1
, 1 , ,
1
1 ,
2
,
2
1 ,
÷ ÷
=
÷
=
÷ ¿ ¿
÷ ÷
H O =
t i t i t i t i
N
S
S
t i t i
M
S
t i t i t i t i
y y P S S S S S S y y P
t i
t i
t i
(2)
with ) | (
1 , ,
1
,
÷ t i t i
S
y y P
t i
given by (1). If Ocollects all parameters, and ) (
1 ,
1
,
i
S
y P
t i
is the initial fixation
probability, the likelihood function of participant i is:
[ ¿ ¿ ¿ [ ¿ ¿
= = =
÷ ÷
= =
÷
= =
÷
H O = O
T
t
M
S
N
S
t i t i
S
t i t i
N
S
S
T
t
t i t i
N
S
M
S
i
S
i
T i T i
t i
t i
t i
t i t i
i
y y P S S S S y P y L
2 1 1
1 , ,
1
1 ,
1
,
1 2
2
1 ,
2
,
1 1
1 ,
2
,
1
,
1
,
1
1 ,
2
,
1
,
2
,
1
1 ,
] ) | ( ) | ( ) | ( ... [ ... ) ( ) | (
(3)
A key output of the model estimation is the marginal posterior probability that participant
i is in information acquisition state
1
,t i
S at fixation t: ) , | (
,
1
,
O
t i t i
y S P , as well as the posterior
probabilities of being in the upper level states
2
,t i
S : ) , | (
,
2
,
O
t i t i
y S P . Those probabilities allow us to
describe the information acquisition strategies that participants use from moment-to-moment.
2.3.3 Model Estimation and Testing
We specify uninformative conjugate priors on all model parameters. We use MCMC to
estimate the model, which is programmed in R. Single move sampling schemes are used to
sample middle layer hidden states S
1
, upper layer hidden states S
2
, and the parameters of the
transition matrices (Robert, Celeux, and Diebolt 1993). Several tests on simulated data reveal
19
accurate recovery of the parameters
3
. In all analyses, we apply the model using 25,000 draws and
discard the first 5,000 draw to burn-in. The chains are stabilized after 5,000 draws. We take one
in ten target draws and report the mean and standard deviation of the resulting 2,000 draws to
summarize the posterior distributions of the parameters.
We compare our model with realistic alternatives described later, and test for the number
of upper layer states of our model using the log-marginal density (LMD). We compute the LMD
using Chib‘s (1995) method, which provides unbiased and stable estimates:
) | ( ) ( ) | ( ) (
* * *
Y In In Y f In Y LMD O ÷ O + O = t t (4)
The calculation of ) (
*
O t In , the log-prior density, and ) | (
*
Y In O t , the log posterior density,
both computed at the ordinate of the posterior density ?*
(for example, the posterior mean) is
straightforward. We compute the high-dimensional sum in the likelihood (3) using Scott‘s (2002)
likelihood recursion method.
2.4 Information Acquisition on A Comparison Website
2.4.1 Experimental Procedure
We chose the Dell website (www.Dell.com) as context for our study. It provides the
option of comparing various personal computers that are relevant for our participants. We
manipulate the display format by transposing rows and columns in a two-group design. The
original comparison matrix that we singled-out from the website contains twelve attributes:
?Picture‘, ?Price‘, ?Processor‘, ?Operating System‘, ?Memory‘, ?Keyboard Mouse‘, ?Monitor‘,
?Hard Drive‘, ?Optical Drive‘, ?Wireless‘, ?Office Software‘ and ?Warranty‘; and four desktop
models: ?Inspiron 864‘, ?Inspiron 819‘, ?Inspiron 758‘ and ?Inspiron 689‘. The comparison
3
30 participants each with 100 observations, transition matrix with 3 columns or 3 rows. All true values fall within the
95% Highest Posterior Density region of the parameters in question; details available upon request.
20
matrix thus lists twelve attributes in the rows, and four desktop products in the columns, plus one
column that contains attribute labels. In the second condition, the comparison matrix is
transposed, so that the attributes are in columns and products in rows. We label the two
conditions as the Product-Column (PC) and the Product-Row (PR) condition, respectively. 108
undergraduate students (55 male) for which choice of a personal computer is relevant
participated. Participants were randomly assigned to either the PC or the PR condition. They read
the instructions that asked them to make a choice for a desktop computer. After having made
their final choice, they were asked to indicate how easy it was to collection information and
decide among the different computers on a 5-point response scale.
2.4.2 Eye movement Recording
Tobii 1750 infrared eye tracking equipment was used (www.Tobii.com). It leaves
participants free to move their heads; cameras in the rim of a LCD-computer monitor (1,280 x
768 pixel resolutions) track the position of the eye and head. Measurements are taken with a
frequency of 35Hz and a precision better than 0.5 degree of visual angle. Instructions and stimuli
were presented on the monitor and participants continued to a next page by pressing the space
bar. Saccades between cells of the comparison website (a total of 18,172 observations) are the
unit-of-analysis; re-fixations were excluded.
2.5 Results
2.5.1 Model Comparisons
We compare the proposed model with four alternatives. Alternative model 1 is a HMM
without the upper layer but else is the same as our proposed model. If this model would be the
best, the probabilities of switching between attribute-based and product-based information
21
acquisition (level 2) would be constant over time. It would imply that a single stage rather than a
two- or multi-stage decision process (Bettman et al. 1998) would direct information acquisition
in the present context. Alternative model 2 is a two-layer HMM without the (level 2) constraints
of attribute-based and product-based acquisition processes. If this model would be the best, it
would imply that eye movements do not reflect the two information acquisition strategies.
Finally, we test for 2, 3 and 4 upper-layer hidden states for our model. We perform all these tests
for both PC and PR formats.
Our model with 2, 3 or 4 upper layer hidden states outperforms both alternative HMM
models, as revealed by the higher LMDs in Table 2.1. This holds for both conditions (PC and
PR). These results provide support for the hierarchical structure of the decision process: a three-
layer model explains the decision process better than a two-layer model. As shown in Table 2.1,
the value of the LMD levels off after three upper-level states. In addition, the parameter
estimates of two hidden states in the 4-state model are quite similar. Thus, we choose the model
with three upper-layer hidden states as our final model.
[INSERT TABLE 2.1 ABOUT HERE]
2.5.2 General Estimation Results
Appendices I and II provide the parameter estimates of the model. Table 2.2 presents, for
the PC and PR conditions, the probabilities that saccades are consistent with attribute-based, or
with product-based acquisition. For the PC presentation format, in the attribute-based state
participants are 14.0 times more likely (calculated as p/(1-p)) to move their eyes within the same
attribute than making other eye movements. For the PR format this ratio drops to 8.1. For the PC
format, in the product-based state participants are 1.7 times more likely (calculated as q/(1-q)) to
make an eye movement within the same product than making other eye movements. For the PR
22
format this ratio increases to 2.8. Two conclusions are apparent. First, the eyes are much more
consistent in moving across products for the same attribute, regardless of whether the products
are presented in the rows or columns. Second, consistency is substantially higher for either
acquisition strategies if it is presented row-wise. This corroborates previous findings on the
dominance of horizontal eye movements in very different research areas (Gilchrist and Harvey
2006, Van der Lans et al. 2008, Tatler and Vincent 2008).
Table 2.2 also shows for the PC and PR conditions the probabilities that participants
make contiguous and non-contiguous eye movements. For the PC format, eye movements
between contiguous attributes are 10.1 times more likely than between non-contiguous ones.
This ratio is 8.2 for the PR format. For contiguous/non-contiguous products that ratio is 2.6 for
the PC, and 1.8 for the PR format. Participants are much more likely to compare information
from neighboring attributes than from neighboring products, regardless of the orientation of the
display. And they are even more likely to do so when attributes are presented in the rows (PC
format). The PR format induces less contiguous information acquisition than the PC format. This
dominance of information acquisition on contiguous cells of the matrix is in line with prior
accounts of the dominance of local eye movement patterns in print advertisements. Yet, it seems
not to have been reported in previous research on decision making on product-attribute displays
where the information content rather than placement is expected to exert stronger impact in
comparison and evaluation. These results provide evidence for mostly local information
acquisition interspersed by short periods of redirecting the attentional spotlight to other
potentially informative regions, with some influence of the presentation format.
[INSERT TABLE 2.2 HERE]
23
We describe the attribute-based and product-based states of the middle layer in more
detail. First, when being in the attribute-based state, the average number of products inspected
within an attribute is a little over two (2.28 in the PC and 2.07 in the PR condition), while one to
two attributes are inspected (1.55 in the PC and 1.38 in the PR condition)
4
. Second, when being
in the product-based state, the average number of attributes inspected within a product is about
three (3.11 in the PC and 3.20 in the PR condition), while on average about a single product is
inspected (1.04 in both conditions). Thus, participants inspect about three attributes for a single
product, or compare two products on one to two attributes, before switching to another strategy
of processing. This is largely independent of the orientation of the display.
Figure 4 plots the cumulative percentages of attributes and products inspected across the
(normalized) decision time. In both formats, participants do not examine all products until about
halfway through the decision process. The increase in the number of attributes inspected with
increasing decision time is almost linear, with around half of the attributes (51% in the PC and
46% in the PR condition) inspected halfway through. About twenty percent of the attributes
(25% in the PR condition and 17% in the PC condition) are never inspected. Both the number of
attributes and the number of products inspected are lower in the PR than in the PC condition.
This selectivity in information acquisition may be due to information overload that limits the
uptake of available information, and is influenced by the presentation format.
[INSERT FIGURE 4 ABOUT HERE]
Appendix I provides estimation results for the middle and upper layers of the model. The
upper layer states in the model govern the prevalence of the middle layer states and the switching
between them. Although it is not always possible or advisable to interpret these upper level states
4
The average number of products inspected is somewhat lower at the beginning and end of the decision process (about
1.7 on average).
24
and one may merely view them as a flexible approximation to the time-structure in the data, they
are clearly differentiated in the present application. That is, upper-layer state S
2
=1 induces state-
dependence with participants having a large probability of sticking to whatever their current
acquisition strategy is (a ?stay? regime). State S
2
=2 induces switching to and staying with
attribute-based acquisition (a ?switch to and stay with attribute? regime). State S
2
=3 induces
switching to and staying with product-based acquisition (a ?switch to and stay with product?
regime?). This interpretation holds for both formats.
Figure 5a presents the (posterior) probabilities of the three upper layer states (?regimes?)
across (normalized) time for all participants; Figure 5b graphs the probabilities of the two middle
layer states (?acquisition strategies?). Counter to theories of two-stage decision making that
postulate a long period of attribute-based acquisition in the first stage followed by a long-period
of product-based acquisition in the second stage, Figure 5a shows that in both formats,
?switching to and staying with attribute-based acquisition? (S
2
=2) dominates in the middle part
of the decision process, and that ?switching to and staying with product-based acquisition? (S
2
=
3) dominates in the early and later parts of the process. In addition, upper-layer state S
2
= 1
(?stay?) is more prevalent in the beginning and the end, which indicates that participants are then
less likely to switch between strategies. Saccade lengths (calculated as the Euclidean distance
between the corresponding cells of the comparison matrix) are much larger for state 1 (S
2
=1)
(35.4) than that for states 2 (S
2
=2) and 3 (S
2
=3) (12.2 and 7.8 respectively). This suggests that,
independent of the format, participants tend to adopt a more global product-based information
acquisition in the beginning and end of the decision process. The prevalence and stickiness of
product-based strategies towards the end of the decision process in particular, is consistent with a
25
pre-decisional gaze bias that reflects a cascading preference formation and decision justification
(Pieters and Warlop 1999; Shimojo, Simion, Shimojo, and Scheier 2003).
As for the format impact, Figure 5a shows that for the PR format, the prevalence of the
upper-layer state 3, which induces ?switching to and staying with product-based acquisition?, is
much higher than that for the PC format. In the latter format, upper-layer state 2, inducing
?switching to and staying with attribute-based acquisition?, dominates. Thus, the PR format
indeed induces more product-based acquisition (Bettman and Kakkar 1977). This is confirmed in
Figure 5b, which shows the prevalence of the two middle layer states over time. The figure
shows, however, that this dominance of product-based acquisition in the PR format manifests
itself only in the middle of the decision process. For the PC format, product-based and attribute-
based acquisitions are equally prevalent in the mid-range of the decision. The dominance of
product-based acquisition at the beginning and end of the decision process is unaffected by the
information display format.
A multinomial logit model with the total (posterior mean) durations of different
acquisition states as predictors (pseudo-R
2
= .821), shows that more attribute-based acquisition
significantly increases the probability of choosing product 1 over 4 (b = 6.32, p = .003), while it
reduces that of the other products (b = -7.01 for product 2, p = .001, -6.33 for product 3, p =
.063). Product 1 (Inspiron 864) is superior on most non-price attributes. Thus, participants in the
PC condition seem more likely to choose the dominating product than those in PR condition
(choice probabilities: PC: .26 vs. PR: .20). They also seem less likely to choose the second
product (choice probabilities: PC: .08 vs. PR: .13), which is inferior on most non-price attributes.
[INSERT FIGURE 2.5 ABOUT HERE]
26
2.5.3 Switching between I nformation Acquisition Strategies
To illustrate the switching between information acquisition states, Figure 6 plots their
probabilities over time for one illustrative participant in the PC and one in the PR condition. The
Figure shows that switching between the two (middle-layer) information acquisition strategies is
very frequent, but for the PR format even more so. On average, participants switch respectively
.61 and .83 times per second for the PC and PR conditions. If we were to count single-step
transitions in the eye movements directly, we obtain even higher switching frequencies: 1.27
(PC) and 1.37 (PR) switches per second. These latter observed frequencies are much higher,
because eye movements are equated with information strategies, rather than being treated as
noisy indicators of them. But, even though our model provides much more conservative
estimates, switching between the two information acquisition strategies is so frequent that it is
unlikely that these reflect conscious "strategies", and we will call them "modes" instead.
[INSERT FIGURE 2.6 ABOUT HERE]
Participants switch between the upper layer states .84 times per second for the PC
condition and .88 times for the PR condition. Thus, while the presentation format affects the
switching between the states of the middle-layer (modes of information processing), it does not
affect switching between the upper-layer states. These states might be associated with higher
cognitive processes (Salvucci and Anderson 1998), which may be less susceptible to bottom-up
influences such as the presentation format.
This frequent switching between acquisition modes significantly increases decision time
(regression coefficient b = 1.53 for middle layer switching, p = .06; b = 2.30 for upper layer
switching, p < .001). Further, switching significantly reduces the experienced ease-of-processing
(average of post-choice evaluation items, M = 2.78; b = -.33 for middle-layer switching, p =
27
.002). Therefore, participants perceive the decision process to be easier in the PC condition (less
switching) than that in the PR condition (M = 3.02 and M = 2.64 respectively; p = .059).
To investigate determinants of switching between states, we estimate logit regressions of
the participant-specific switching indicators on the percentages of products (P), respectively
attributes (A) inspected up to that time point. In the middle-layer, switching from product-based
to attribute-based acquisition is predicted by the percentage of attributes already inspected (PC
condition: b= .29; p=.08; PR condition: b= .38, p=.07). Switching from attribute-based to
product-based acquisition, on the other hand, is predicted by the percentage of products already
inspected (PC condition: b= .48, p=.03; PR condition: b=.96, p<.001). This corroborates that
participants switch to a specific mode of information acquisition as more information has been
acquired in the other mode (Russo and Rosen 1975; Bettman and Park 1980).
2.5.4 Attention to Products and Attributes
Next we show what specific product and attribute information consumers acquire over
time when they are in respectively the attribute-based and product-based acquisition states. For
that purpose we select three participants representative for the PC and the PR conditions, by K-
means clustering participants based on their fixations on products and attributes, and selecting
the most central participant for each cluster. The cumulative numbers of fixations on products in
the attribute-based state and on attributes in the product-based state are shown in Figures 7 (PC
condition) and 8 (PR condition), respectively. The number and pattern of fixations that the
products and attributes receive shows considerable heterogeneity between participants, but four
key insights can be obtained.
First, in many cases, some products seem to be taken out of consideration and receive no
additional gaze after a certain time point: the cumulative product-fixations in the attribute-based
28
state level off, even more so for the PR than for the PC format. This is in line with theories of
hybrid decision-making (Bettman 1979). However, contrary to those theories, in several cases
the eliminated alternative is reconsidered later in the decision process. Take product 1 (Inspiron
864) in the left bottom of figure 7c as an example. The product receives constant attention at
first, then cumulative fixations level off while the other two or three products are compared. At a
later stage of decision-making (after a wait of about 25 fixations), however, this product is being
considered again, as reflected by increasing fixations. It might be that product 1 is being ?put on
hold? while other alternatives are compared, or, it is possible that this product is taken out of
consideration at first and then is being re-considered later. Either way, it seems that the
comparison set changes over the entire time course of decision making. Indeed, the scope of the
product comparison set shrinks and expands (between 1 and 4, SD = 1.10). Thus, in contrast with
hybrid decision-making theories that postulate that the comparison set is constant after a certain
time point, the plots show that it changes over time.
Second, in the product-based acquisition state, attributes enter the decision process
sequentially. Many are considered only once. The temporal sequence of selection of attributes
may reflect their relative importance to the choice goal at hand (Bettman et al. 1998, Wedell and
Senter 1997). It seems that for each of the representative respondents a fair number of attributes
is not considered at all, which is even more so in the PR condition.
Third, regularly attributes and products that have been rejected earlier are briefly
revisited in the final stage of the decision process. This appears to reflect a final verification of
the alternative to which the participant has committed (Russo and LeClerc 1994; Russo and
Rosen 1975). A few attributes appear to be exclusively used in this final stage of the decision.
29
Fourth, in the majority of the cases there is a very clear pre-decisional acceleration of
gaze on the chosen alternative towards the end of the decision. This gaze cascade is predictive of
the alternative that is chosen (Pieters and Warlop 1999; Glaholt and Reingold 2009; Shimojo et
al. 2003).
[INSERT FIGURES 2.7 AND 2.8 ABOUT HERE]
2.6. Discussion
Decision theorists have argued that eye tracking provides insights into fast and partly
automatic information acquisition processes during decision making. In recent years affordable
and easy to use eye tracking equipment has become widely available. With that, the challenge
has become to describe the large volumes of data that eye tracking studies produce in a
meaningful manner. For this purpose, the proposed Hierarchical Hidden Markov Model
(HHMM) can help. In our study we apply the HHMM to eye tracking data to describe moment-
to-moment information acquisition on product-by-attribute matrices.
This revealed that participants switch frequently between product-based and attribute-
based acquisition modes: 50 to 60 times during the 67 seconds that a decision took, on average.
The amount of information already collected on attributes induces switching away from the by-
product mode, and the amount of information on products induces switching away from the by-
attribute mode. This high degree of switching during real-life decision making has not been
documented previously, and its causes are not fully understood. Participants limit their attention
to about three attributes for a single product, or two products for one to two attributes. This
might be due to the restricted rate at which information can be consolidated (Marois and Ivanoff
2005), and limits on the amount of information that can be stored into visual short term memory
at a particular point in time (Cowan 2001, Luck and Vogel 1997). Instead of sequentially
30
adopting attribute-based acquisition to eliminate alternatives followed by product-based
acquisition, participants appear to sample ?parcels? of attribute and product information
(Stewart, Chater, and Brown 2006). They constantly adjust their acquisition mode and the
comparison set of products to consolidate information into memory and make the next
information acquisition decision. Investigating the causes and effects of information parceling
and acquisition mode switching may be an avenue for future research.
The revealed high frequency of switching cast doubts on whether "by-attribute" and "by-
product" information acquisition strategies are conscious and deliberate strategies. Rather,
presentation format, but also low-level eye movement tendencies that could not be incorporated
in previous accounts of decision making, appear to play an important role. Low-level, automatic
and unconscious processes are principal in situations of information overload and time pressure
(Lee and Lee 2004; Pieters and Warlop 1999), which occur when facing data-rich web-based
choice environments. First, there is a predominant left-to-right tendency in the eye movements
that makes it appear as if there is a "by-product" or "by-attribute" strategy of processing,
depending on the orientation of the matrix display. Second, information acquisition is
predominantly local, confined to neighboring cells on the display, because of limited visual
detail beyond the immediate eye fixation point. Third, the end of the decision process reveals a
pre-decisional gaze bias that reflects a cascaded process of preference formation and is predictive
of the final choice. In our study the presentation format impacts the way that information is
acquired, because the format changes but the low-level eye movement tendencies (left-right,
local) do not. If eye movement patterns stay the same when the information display changes,
then this causes different information to be extracted and different decisions to be made. This
makes information acquisition and decision making seem adaptive, while in fact they are not.
31
Instead of purposely using attribute-based and product-based ?strategies?, participants
tend to use a strategy of local information sampling, and tend to make horizontal movements.
Information placed contiguously is likely to exert a strong impact on decisions. Such tendencies
provide opportunities for managers to strategically place product-attribute information to
facilitate comparisons. For example, one could increase the attractiveness of a specific product
by placing it next to a product that it dominates overall, or on specific attributes (Huber, Payne,
and Puto 1982). Comparison website providers can also adopt display formats in a proactive
fashion to stimulate consumers‘ use of specific information acquisition mode favorable to their
goals. Managers need to decide how to balance format, sales objective, and switching cost. For
example, we find that that the Product-Column format facilitates attributed-based acquisition and
increases the chance of the dominant product being chosen. Product-Row format leads to more
evenly distributed choice probabilities. However, the Product-Row format favors by-product
information acquisition, which results in more local eye movements, less products and attributes
being inspected, and increased switching between modes of acquisition that makes choice more
difficult. Some comparison websites (e.g., Nextag.com) now allow consumers to firstly sort
alternatives by certain attribute, such as price, ratings, etc., and then present the sorted product
information in Product-Row format (by default). This may be in conflict with desirable attribute-
based information processing for these websites and make decision making more difficult.
Comparison websites, the context chosen for the present research, are a relatively new
shopping environment, yet rapidly increasing in popularity. We demonstrated the effects of the
row-column orientation of the website, as a useful starting point for research in this area. With
the advance of network engineering, comparison websites are now able to provide more and
more extensive and dynamic product-attribute comparisons. Website designers therefore have
32
increasing abilities to improve comparison website usability and may use the results of this study
to help induce information acquisition modes that are congruent with their goals. We hope that
our study stimulates further research interest in this area, which could include the influence of
website design factors such as sorting (arranging the product order such that it reflects relative
importance), grouping (placing similar or related elements close together), trimming (eliminating
unnecessary information), and highlighting (visually accentuating important information through
colors and shading). According to Johnson et al. (2008) decision research can progress more
quickly when it adopts approaches that provide richer descriptions of the underlying processes.
Our proposed Hierarchical Hidden Markov model may provide a starting point for the
development of such models of online choice.
33
Table 2.1 Log Marginal Density (LMD) of Alternative Models
Experimental
Condition
Proposed HHMM Model
with n States in the Upper Layer
Alternative
Model 1:
Constrained
HMM
Alternative
Model 2:
Unconstrained
HMM Two States Three States Four States
Product-column -10392.00 -9066.07 -9058.13 -13696.97 -13132.81
Product-row -10877.76 -9503.57 -9484.53 -13168.99 -12205.66
Table 2.2 Parameter Estimates For Eye movements (lower layer) and Information
Acquisition Strategies (middle layer) (with standard deviations in parentheses)
Acquisition
Strategy
(Middle Layer)
Components
Information Presentation Format
Product-Column Product-Row
Attribute-based
Consistent
.929 (.001) .892 (.001)
Inconsistent .066 (.022) .110 (.011)
Contiguous (products) .444 (.007) .283 (.006)
Noncontiguous (products) .171 (.009) .157 (.007)
Product-based
Consistent .641 (.005) .736 (.002)
Inconsistent .360 (.016) .264 (.008)
Contiguous (attributes) .362 (.000) .338 (.002)
Noncontiguous (attributes) .036 (.001) .041 (.000)
Note: ?Consistent? is the probability of staying in the same attribute or product (p or q),
?Inconsistent? is the probability of moving to another attribute or product (1-p or 1-q).
?Contiguous? is the average of the transition probabilities between block-diagonal sub-matrices
(all elements on the 2×2 diagonal blocks, see Appendix II, shadowed parts), ?Noncontiguous? is
the average transition probabilities between the cells in the off-diagonal blocks (all elements
except those in the diagonal blocks)
34
Figure 2.1 Eye Movements of A Single Participant Making A Choice on the Dell
Comparison Website.
Circles indicate eye fixations (with fixation numbers), and lines indicate saccades between fixations.
35
Figure 2.2 The Three Layers of the Proposed Hierarchical Hidden Markov Model
Note: Triangles represent the observations in the lower layer, with transition probabilities P
1
or P
2
;
squares represent the hidden states in the middle layer (S
1
), with transition probabilities ?
1
and ?
2
; circles
represent the hidden states in the upper layer (S
2
), with transition probability ?.
S
2
=1 S
2
=2
?
1
?
2
S
1
=1 S
1
=2 S
1
=1 S
1
=2
P
1
P
2
P
1
P
2
?
36
Figure 2.3 Product-Attribute Matrix: Information Acquisition Strategies and
Elementary Eye Movements
product 1 product 2 product 3 product 4
attribute 1
attribute 2
attribute 3
attribute 4
attribute 5
Note: Elementary eye movements (dotted arrows): type 1 is saccade between two attributes of same
product, type 2 is saccade between two products within same attribute, and type 3 is saccade from one
attribute of a product to another attribute of another product. Arrow A is attribute-based acquisition and
comprises type 1 and 3 movements. Arrow B is product-based acquisition and comprises type 2 and 3
movements.
A
1
B 2 3
37
Figure 2.4 Cumulative Percentage of Attributes and Products Inspected across
(normalized) Decision Time for Product-Column (PC) and Product-Row (PR) Conditions.
X-axis: normalized decision time; Y-axis: cumulative percentage of attributes and products inspected
38
Figure 5a Aggregate Time Course of the Prevalence of Upper Layer States: S2=1 (long
dash), S
2
=2 (short dash) and S
2
=3 (solid line) for Two Information Presentation Formats
X-axis: normalized decision time; Y-axis: probabilities of three upper layer states
Product-Column Format Product-Row Format
Figure 5b Aggregate Time Course of Attribute-Based (short dash) and Product-Based
(solid line) Information Acquisition for the Two Information Presentation Formats
X-axis: normalized decision time; Y-axis: probabilities of two middle layer states
Product-Column Format Product-Row Format
39
Figure 2.6 Examples of the Time Course of the Three Upper-Layer States (S
2
=1: long
dash, S
2
=2: short dash and S
2
=3: solid line), and of the Two Middle Layer States (attribute-
based acquisition: short dash, and product-based acquisition: solid line), for One
Participant in Each Condition (presentation format)
X-axis: fixation number; Y-axis: probabilities of each state in the middle and upper layers
Product-Column Condition
(Participant 17)
Upper Layer
Middle Layer
Product-Row Condition
(Participant 3)
Upper Layer
Middle Layer
40
Figure 2.7a The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
39, PC condition. Chosen Alternative: 2 (Dell 819), Decision Time: 1 min and 57 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
Figure 7b The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
42, PC condition. Chosen Alternative: 4 (Dell 689), Decision Time: 2 min and 8 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
41
Figure 2.7c The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
19, PC condition. Chosen Alternative: 1 (Dell 864), Decision Time: 1 min and 28 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
42
Figure 2.8a The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
24, PR condition. Chosen Alternative: 3 (Dell 758), Decision Time: 59 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
Figure 2.8b The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
22, PR condition. Chosen Alternative: 2 (Dell 819), Decision Time: 1 min and 34 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
43
Figure 2.8c The Cumulative Number of Fixations on the Attributes While in the
Product-Based State and on the Products While In the Attribute-Based State; Participant
41, PR condition. Chosen Alternative: 3 (Dell 758), Decision Time: 3 min and 46 sec.
X-axis: fixation number; Y-axis: cumulative number of fixations
44
Chapter III: Usage Experience with Decision Aids and Evolution of
Online Purchase Behavior
Wei Shi
Robert H. Smith School of Business
University of Maryland
Jie Zhang
Robert H. Smith School of Business
University of Maryland
Acknowledgments
The authors thank an anonymous online retailer for providing the data used in this study.
This research is supported by the MSI grant #4-1649, as a winner of the ?Shopping Marketing?
research proposal competition.
45
3.1. Introduction
Internet retailing has experienced explosive growth for over a decade. As more shoppers
begin to make purchases online, it is important to understand how they adapt to this increasingly
prominent channel and whether their purchase behavior evolves as they gain more experience
with the new shopping environment. Prior research has examined the impact of the Internet
environment on purchase decision processes (e.g., Bechwati and Xia 2003; Häubl and Trifts
2000; Hollander and Rassuli 1999; Lee and Geistfeld 1998), and the differences between online
and offline purchase behaviors (e.g., Danaher, Wilson, and Davis 2003; Degeratu, Rangaswamy,
and Wu 2000; Zhang and Wedel 2009).What is lacking in the literature is a comprehensive
examination of the evolving patterns of purchase behavior, such as in terms of store loyalty and
price sensitivity, and more importantly, what may drive purchase behavior changes in online
stores.
It is well documented that the store environment can influence a consumer‘s decision
making process (e.g., Park, Iyer, and Smith 1989; Inman, Winer, and Ferraro 2009). A distinct
feature of the Internet store environment is that online stores offer a variety of interactive
decision aids which can facilitate consumers‘ shopping processes. For example, many online
retailers provide decision aids that allow shoppers to sort alternatives or filter them with certain
criteria, to create personalized shopping lists, or to check total basket spending. Studies have
shown that this kind of interactive decision aids can influence consumers‘ information search
processes, purchase outcomes, and satisfaction (e.g., Bechwati and Xia 2003; Häubl and Trifts
2000; Hollander and Rassuli 1999; Lee and Geistfeld 1998). Therefore, one can speculate that, as
online shoppers accumulate more experience with using various decision aids, their purchase
behavior may also change over time as a consequence.
46
The objective of this research is to conduct an empirical investigation on whether and
how usage experience with various decision aids may drive online purchase behavior changes
over time, and what roles different types of decision aids may play in the process. We intend to
address the following managerial questions: 1) Will shoppers exhibit more habitual behavior as
they get accustomed to an online store, or are they more likely to engage in on-the-spot decisions
as they become more experienced and efficient with using interactive decision aids? 2) How will
this affect their tendency to shop from an online store and their price sensitivity? 3) What kind of
decision aids can mitigate price competition? 4) What kind of decision aids may increase
consumers‘ loyalty to an online store? and 5) How can marketers influence consumers‘ behavior
evolvement by designs of decision aids in an online store? Answers to these questions can help
Internet retailers improve the design of their store environments and provide insights for
manufacturers to modify promotion messages adaptively according to consumers‘ evolving
purchase behavior.
Online shopping behavior has been shown to be systematically different from offline
shopping behavior (e.g., Danaher, Wilson, and Davis 2003; Degeratu, Rangaswamy, and Wu
2000; Zhang and Wedel 2009). For example, the observed behavioral discrepancy can be
attributed to two broad sources: differences in intrinsic characteristics between online and offline
consumers, and differences in the shopping environments (Zhang and Wedel 2009). Researchers
have postulated that interactive decision aids available in online stores can train consumers to
shop in certain fashions, which would attribute to purchase behavior differences in the two types
of shopping environments (e.g., Alba et al. 1997; Degeratu et al. 2000; Zhang and Wedel 2009).
Our study will provide an empirical test of this conjecture and shed light on whether decision
47
aids available in online stores indeed lead to purchase behavior changes for the same consumer
over time.
Like the above mentioned previous studies that compare online and offline purchase
behavior, our empirical investigation is carried out in the context of online grocery stores. After
initial struggles and some high profile failures, Internet grocery retailing has shown a resilient
comeback and experienced steady growth in recent years. According to a recent report by the
Nielsen Company, online grocery retailing has grown at more than 20% compound annual rate
since 2003 and attracts 13 million U.S. Internet users by July 2009 (Swedowsky 2009). Findings
from our study will be relevant to a wide range of companies, especially as more traditional
retailers (e.g., Safeway, Albertson, Wal-Mart) and Internet retailers (e.g., Amazon.com) venture
into the online grocery retailing business.
Our data are provided by a leading Internet grocery retailer which was among the very
first to sell groceries online. The dataset was collected during the period when the retailer first
launched its web business, which makes it particularly appealing to study the evolution of online
purchase behavior. Research has shown that consumers‘ in-store decision making processes may
vary by product categories (Inman, Winer, and Ferraro 2009). Our dataset includes detailed
click-stream navigation information, as well as individual purchase history data in multiple
product categories, and thus allows us to examine potential differences in the patterns across
these categories.
We construct a Non-homogeneous Hidden Markov Model (NHMM) of category
purchase incidence and purchase quantity, in which parameters are allowed to vary over time
across hidden states as driven by usage experience with different decision aids. The Hidden
Markov Model is well suited for the purposes of this research. Shopping behavior, including
48
online shopping behavior, has been classified into different states, such as for ?hedonic? and
?utilitarian? motivations (Babin, Darden, and Griffin 1994; Childers et al. 2001; Hirschman and
Holbrook 1982). These studies suggest that the observed online shopping behavior is likely to be
directed by certain latent ?behavior states?, and the store environment may train consumers and
change their behavior states over time. Our NHMM is built to identify these latent states and
examine how usage experience with online decision aids may drive transitions between these
states.
Understanding how consumers‘ purchase behavior evolves over time as their experience
with decision aids accumulates would offer valuable insights for online retailers to improve the
design of their store environment. It could help manufacturers modify their communication
messages (for example, choose to focus on price/promotion-oriented information or to highlight
specific product attributes), based on the purchase behavior revealed in different latent states.
Moreover, findings from our study will suggest ways for online retailers to offer personalized
shopping environment for individual consumers, or to influence their purchase behavior
evolution.
3.2 Conceptual Development and Literature Review
In this section, we present the conceptual development of our study and provide an
overview of the relevant literature.
3.2.1 Online Decision Aids
Online stores make the shopping process easier and more convenient by offering a
variety of decision aids. These decision aids allow consumers to perform a more ?thorough and
exhaustive search? (Hollander and Rassuli 1999; Lee and Geistfeld 1998). They enable online
49
shoppers to make better purchase decisions and doing so with less effort (Häubl and Trifts 2000).
Decision aids are especially popular in online grocery stores. We classify four types of decision
aids that are commonly available in these stores.
1. Decision aids for nutritional needs: Many online grocery stores provide decision aids
to facilitate the shopping process for consumers who have special dietary needs (Swedowsky
2009) or are concerned of nutritional information. These decision aids include sorting functions
(such as ?by calories?, ?by cholesterol?, ?by sugar?, and ?by fat?) and precluding functions (such
as ?Kosher foods?, ?organic food only?) to rank, compare, or filter the products with certain
criteria. For example, www.groceryexpres.com offers 12 functions to fulfill consumers‘ special
dietary needs. www.freshdirect.com has 16 such functions. Growing health concerns among the
public are believed to have contributed to the prevalence of such decision aids.
2. Decision aids for brand preference: In online stores, consumers with specific brand
preference can choose their preferred products by the brand name using functions such as
?sorting by brand/name? or ?search by (brand name)?. This type of decision aids are ubiquitous
in online stores (see www.freshdirect.com, www.coles.com, www.netgrocer.com for just a few
examples). Compared to brick-and-mortar stores, they make the shopping process particularly
efficient for consumers who have strong brand preferences by avoiding effortful navigations
across physical shelves.
3. Decision aids for economic needs: online decision aids such as ?sorting by price?,
?sorting by promotion?, or ?club special first? (see www.freshdirect.com, www.safteway.com,
www.peapod.com for examples) make the shopping process easier for price-sensitive consumers.
Such decision aids facilitate price comparisons and might induce higher price sensitivity (Alba et
al. 1997). In addition, some online stores allow shoppers to check the total spending before
50
submitting an order, which enables budget-conscious consumers to monitor their spending more
effectively during the shopping session.
4. Personalized shopping lists: Some online stores offer consumers the option to create
personal shopping lists or save previous order lists automatically (such as www.peapod.com,
www.freshdirect.com, www.safeway.com, and www.walmart.com). They allow consumers who
are time-constrained or have relatively consistent shopping baskets to complete the shopping
process quickly. Shopping lists serve as a memory aid (Block and Morwitz 1999). Research has
shown that consumers who use shopping lists tend to make less unplanned purchases (Inman et
al. 2009). Therefore, usage experience with shopping lists may train consumers into habitual
shoppers.
These interactive decision aids offer online stores a unique advantage over their brick-
and-mortar counterparts, by making the purchase decision process less effortful, more efficient,
and more suited to individual‘s needs and preferences. The main objective of this study is to
examine whether and how the usage experience with different types of online decision aids may
drive purchase behavior changes overtime. Although the specific context of our study is online
grocery stores, most of the decision aids classified above, with the exception of those for
nutritional needs, apply to other types of retailers and product categories.
3.2.2 Evolution of Online Shopping Behavior
Previous studies suggest that consumers' purchase behavior may change over time in an
online store (e.g., Ansari, Mela, and Neslin 2008; Zhang and Krishna 2007). In the context of
offline stores and unfamiliar new product categories, Heilman and colleagues (2000) show that
consumers‘ purchase behaviors exhibit evolving patterns as their experience with purchasing the
category increases. Given that the Internet shopping environment is distinctively different from
51
traditional shopping channels with many unique features, shoppers new to the channel are likely
to go through learning and adaptation processes, and thus their purchase behavior may also
evolve over time.
Researchers have suggested several factors that may contribute to the evolution of online
shopping behavior over time, most of which are related to the Internet experience. For example,
comfort with the Internet (Mauldin and Arunachalam 2002), perceived ease of usage, and
perceived usefulness of online shopping all exert a positive impact on the purchase intention
from the Internet channel (Hoffman and Novak 1996; Chen, Gillenson, and Sherroll 2002;
Limayem, Khalifa, and Frini 2000; Pavlou 2003). Yet there has been little research that
empirically investigates how usage experience with various decision aids contributes to the
evolution of online purchase behavior. Our study intends to fill this void.
3.2.3 Usage Experience with Decision Aids and Online Shopping Behavior Evolution
Internet experience has been found to be an important determinant for consumers‘
purchase intention at online stores (see Zhou, Dai,
and Zhang 2007 for a review). Consumers
could gain Internet experience through more time spent online or repeated visits, which would
increase their comfort level with the online shopping environment and thus purchase intention
(Mauldin and Arunachalam 2002). The ?comfort? may come from experience with navigating a
website or familiarity with decision aids available. Positive experience with a website may in
turn induce greater exploratory behavior on the site (Hoffman and Novak 1996; Mathwick and
Rigdon 2004). These two processes could reinforce each other and accelerate the learning of
decision aid usage. Therefore, as shoppers' experience with decision aids accumulates, their
propensity to purchase from an online store is likely to increase as well.
In terms of responsiveness to marketing mix variables, the impact of online decision aids
52
is likely to be more nuanced. We focus on price sensitivity in this discussion. Usage experience
with online decision aids may affect shoppers‘ price sensitivities at both the store choice stage
and the in-store purchase decision stage.
Many online stores allow consumers to create personal shopping lists and/or store other
shopping information in their personal accounts. These decision aids could create a ?lock-in?
effect (Smith, Bailey, and Brynjolfsson 2000): a consumer who creates and uses personal
shopping lists in one store may face higher switching costs if s/he decides to shop at other
(online or offline) stores. In other words, price is not the only factor to be evaluated when
determine where to shop (Bakos 2001). We expect that increased usage of personal shopping
lists are likely to enhance consumers' loyalty to the store, induce habitual purchase behavior, and
soften their price sensitivity when it comes to choose the shopping venue. This ?lock-in? effect
may also apply to other types of decision aids, such as those for nutritional needs.
The Information Integration Theory (Anderson 1971, 1981) provides some guidance in
predicting the impact of decision aid usage experience on price sensitivity at the in-store
purchase decision stage. According to the theory, certain attributes, such as brand or price, can
surrogate information on other attributes if the latter have limited availability; yet when
information on other attributes becomes available, weights of existing attributes will be reduced
(Anderson 1971, 1981; Bettman, Capon, and Lutz 1975). In traditional brick-and-mortar stores,
consumers are more likely to focus on price-related information because it is easily available and
highly salient with frequent feature and display advertisements (Degeratu et al. 2000). In
contrast, decisions aids available in online stores allow consumers to more efficiently access and
utilize information on other product attributes. They now can find the product that meets their
needs based on important attributes other than price at a lower search costs (Alba et al. 1997). In
53
other words, for some consumers, the weight of price information is likely to reduce while the
importance of other attribute information is likely to increase (Degeratu et al. 2000; Smith et al.
2000), and thus price sensitivity may decrease over time as a consequence for these consumers.
On the other hand, online decision aids intended for economic needs make it easier for shoppers
to use price related information more efficiently. Therefore, it is also possible that, at least for
some consumers, experience with online decision aids will train them to be more price sensitive
(Alba et al. 1997). We leave it as an empirical question regarding the pattern of price sensitivity
in online stores over time. More importantly, we intend to find out what types of decision aids
may reduce price sensitivity and what types have the opposite effects.
3.2.4 Potential Hidden States of Purchase Behavior
The classification of different purchase behavior has been a long-standing research
interest in the marketing literature. For instance, based on consumers' motivations for shopping,
their purchase behavior can be classified into ?hedonic? and ?utilitarian? states (Babin et al. 1994;
Childers et al. 2001; Hirschman and Holbrook 1982). Hedonic consumers are ?equivalent to
brick-and-mortar window shoppers for whom the shopping experience is for entertainment and
enjoyment?; while utilitarian consumers (or goal-oriented shoppers) normally ?purchase
products in an efficient and timely manner to achieve their goals with minimum irritation?
(Childers et al. 2001, p513). Cheung et al. (2003) classify online shoppers into ?intention?,
?adoption (purchase)?, and ?continuation (repurchase)? types based on their purchase intentions.
These different stages of purchase intentions may well apply to the same consumer over time.
Heilman and colleagues (2000) conjecture that two competing forces --- consumers' desire to
collect information about alternatives and their aversion to trying risky ones --- drive consumers'
purchase behavior evolution among three hidden states when buying new (unfamiliar) product
54
categories.
In this study, we propose a Hidden Markov Model which allows us to empirically
identify latent states of purchase behavior in the data. The classifications in the literature will
provide valuable guidance for us to interpret these latent states.
3.3 Model Formulation
This study investigates how the usage experience with various online decision aids
affects consumers‘ purchase behavior evolution. We construct a Non-homogeneous Hidden
Markov Model (NHMM) of category purchase incidence and purchase quantity, in which
parameters are allowed to vary over time across hidden states as driven by usage experience with
different decision aids.
The basic premise of our model is that consumers‘ purchase decisions at any given time
is driven by the hidden behavior states they are in, where the hidden states differ in terms of the
baseline tendency to purchase from the online store and their price sensitivity. Consumers switch
between these states as their experience with decision aids accumulates over time. We adopt a
Type II Tobit model (e.g., Amemiya 1984) to jointly capture category purchase incidence and
purchase quantity decisions, where parameters in the Tobit model evolve according to a Hidden
Markov Model (HMM) with the transition probabilities assumed to be driven by the usage
experience with various decision aids. Our model belongs to the category of Non-homogeneous
Hidden Markov Model (NHMM) (see Hughes 1993).
3.3.1 Type-I I Tobit Model of Category Purchase I ncidence and Quantity
Purchase Incidence
Let
c
it
U = household i‘s latent utility of purchasing category c from the online store in
55
week t;
c
it
I
= 1 if household i purchases category c from the online store in week t, 0 otherwise.
Without losing generality, we can scale
c
it
U
such that:
.
otherwise , 0
0 U if , 1
¹
´
¦
=
> =
c
it c
it
I (1)
The utility function is specified as:
), , 0 ( ~ ,
2
c
c
it
c
it
c
t
S c
it
N X U o c c | + = (2)
where
c
t
X
is a vector of marketing mix variables for category c in week t.
S
| is a vector of their
coefficients given that household is in hidden state s, including the intercept. The intercept can
be interpreted as a household‘s baseline tendency to purchase category c from the online store in
state c. In order to get clearly-defined interpretations of the hidden states, we choose to focus on
price for the marketing mix component in our empirical analysis, because price sensitivity is a
key aspect of household purchase behavior that we intend to study here. We fix
1
2
=
c
o
for
identification purposes.
Purchase Quantity
Purchase quantity is observed only when a household purchases a category from the
online store. We denote
* c
it
Q
as household i’s latent purchase quantity of category c in week t
(measured in volume units such as ounces), and
c
it
Q
as the household‘s observed purchase
quantity of category c in week t. Then,
¹
´
¦
=
= =
=
otherwise , 0
1 if ,
* c
it
c
it c
it
I Q
Q . (3)
We specify
* c
it
Q
as:
), N(0, ~ ,
2
c
*
o u u |
c
it
c
it
c
t
S c
it
Z Q + = (4)
56
where
c
t
Z
is a vector of marketing mix variables for category c in week t, and
S
| is a vector of
their coefficients given that household is in hidden state s. Like in the purchase incidence
component, we use price as the key marketing mix variable in the empirical analysis in order to
get a clean interpretation of the hidden states.
We take into account the interdependence of purchase incidence and quantity decisions
by assuming that the error terms in Equations (2) and (4),
c
it
?
and
c
it
?
, follow a bivariate Normal
distribution:
|
|
.
|
\
|
|
|
.
|
\
|
|
|
.
|
\
|
1
,
0
0
~
2
c c
c c c
c
it
c
it
N
o µ
o µ o
c
u
. (5)
The likelihood for household i in week t given latent state s can be written as (see Amemiya
1984, page 31):
( )
( )
( )
( ) ( )
) 1 (
1
/ 1
1
ln * , |
2
1
C
it
C
it
c
c c
I
C
t
S
I
c
c
t
S c
it
c
c
t
S c
it c c
t
S S S c
it
X
Z Q Z Q
X Q l
÷
÷
÷ u
|
|
.
|
\
|
|
|
.
|
\
| ÷
|
|
.
|
\
| ÷
+ u = |
o
|
|
o
| µ
| | |
o
o µ
(6)
where ) (· | is the probability density function and u(·) is the cumulative distribution function of
the Standard Normal distribution, respectively.
3.3.2 Hidden States and Transition Probabilities
Our main research question is whether and how purchase behavior, as measured by the
baseline tendency and price sensitivity in purchase incidence and quantity decisions, evolves by
the usage experience with various decision aids. To this end, we model the evolution of the
hidden-state-specific parameter vectors
S
| and
S
| in the Tobit model according to a Markov
transition matrix P
it
, which is specified as functions of usage experience with various decision
aids. In our model, consumers are allowed to switch back and forth between the hidden states.
57
We model the probabilities of switching from hidden state S
t-1
in week t-1 to hidden state
S
t
in week t for household i in a K-state NHMM using the ordered logit formulation (see Netzer
and Srinivasan 2008, page 190):
,
) exp( 1
) exp(
1 ) , | (
......
,
) exp( 1
) exp(
) exp( 1
) exp(
) , | 2 (
,
) exp( 1
) exp(
) , | 1 (
1 1
1 1
1 1
1 1
1 1
1 1
1 1
1 1
, 1
, 1
1
, 1
, 1
, 2
, 2
1
, 1
, 1
1
it s S iK
it s S iK
it t t it
it s S i
it s S i
it s S i
it s S i
it t t it
it s S i
it s S i
it t t it
D
D
D S K S P
D
D
D
D
D S S P
D
D
D S S P
t t
t t
t t
t t
t t
t t
t t
t t
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ ÷
÷ +
÷
÷ = =
÷ +
÷
÷
÷ +
÷
= =
÷ +
÷
= =
÷
÷
÷
÷
÷
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
¸ ì
(7)
where K is the number of hidden states. Parameters
1 1 1
, 1 , 2 , 1
,..., ,
÷ ÷ ÷
÷
t t t
s iK s i s i
ì ì ì
are household i‘s state-
specific cut-off points and their values are set in ascending order for a given S
t-1
e {1, 2, .., K}.
it
D is a vector of household i‘s usage experience with a set of decision aids, and
1 ÷ t
s
¸ is a vector
of their coefficients. In the empirical analysis, we categorize and examine six types of decision
aids, including one for nutritional needs, one for brand preference, two for economic needs, and
two for personalized shopping lists. Details of these decision aids and variable computation are
described later.
Note that, in our model, consumer heterogeneity is accounted for by the individual-
specific hidden state transition matrix. It captures individual differences in the behavior
evolution, and automatically accommodates heterogeneity in parameter values in the consumer
response model (i.e., the Tobit model).
The likelihood function for household i across different hidden states is (see Rabiner and
Juang 1986; Rabiner 1989):
) 8 ( ], ) , | ( [ ... ) ,..., , ( ) | (
1 1 1 2
1 1
1 1 2
| , 2 2 1 1 ¿¿ ¿ [ [
= =
÷
= = =
· = = = = = O
K
s
K
s
K
s
T
t
s s c
it
T
t
s s i is
c
iT it
c
i i
c
i i i
s
i
C
i
T
t t
Q l p Q Y Q Y Q Y P Q L | ¢ t
58
Where } , , , {
1 1
,
÷ ÷
= O
t t
s s is
s s s
i
¸ ì | | , and ?
is
is a vector of the initial state probabilities.
3.3.3 Prior Distributions and Estimation Method
The likelihood function in equation (8) is computationally intractable. Thus, we first re-
write the likelihood function as the following (see Macdonald and Zucchini 1997, page 131)
, ' 1 ) , , ( ) | ( ... ) , , ( ) | ( ) , , ( ) (
1 2 2 1 2 1 1 s
2 1
i,1
· O · · O · O · = O
=
÷
= = s s
i
C
T
C
T T T i
s s
i
C C
i
s s
i
C C
i
iT i i
Z X f s s P Z X f s s P Z X f NL t
(9)
where ) , (
s s
i
C
t
C
t
iT
Z X f
=
O
is an K× K diagonal matrix with likelihood ) , | (
,
s s c
it s s
Q l f | | = ;
) | (
1 ÷ t t i
s s P
is an K×K transition probability matrix from t-1 to t; 1’ is a column vector of ones
with length K;
1
is
t
is the stationary distribution of transition matrix
) | (
1 ÷ t t
s s P
, and is calculated
according to equation (7) with covariates set to their mean values in the data.
Following the approach by Netzer et al. (2008), we divide the parameters into two sets:
those that vary across individuals and those that do not. We use a hierarchical Bayes formulation
to estimate the former group of parameters (e.g., Gelfand and Smith 1990; Rossi and Allenby
2003). The prior and hyper-prior are specified as follows:
(1) Parameters that vary across individuals:
u u
u u
u
u
u u o o o
o u ì u
n
i n n
i s is i
I n df df IW
I V V MVN
MVN
t
= A + A E
= =
E =
×
÷
0 0 0 0
0 1 0 0 0
,
; 5 ~ ); , ( ~
) dim( n ; 100 ; 0 ~ ); , ( ~
) , ( ~ }; {
1
(10)
(2): Parameters that do not vary across individuals:
)) ( )' ( exp( ) (
)) dim( ( 50 ; 0 ~
) , ( ~ }; , , {
0
1
0 2
1
1 0
0
0
0
0
¢ ¢
¢ ¢
¢ ¸ ¢ |
¢
¢ ¢ ¢
¢
÷ + ÷ + · +
+ = =
+ = +
÷
×
V P
n I V
V N
n n
s
S S
(11)
{
i
u } and { +} are drawn from Metropolis Hasting algorithm, while {o } and {
u
E } are drawn
from Gibbs sampler.
59
} { , , , , , |
, }, { |
, }, { |
, , , , , , | } {
i it
C
t
C
t
C
it
i
i
it
C
t
C
t
C
it i
D Z X Q
M
M
D Z X Q
u
o u
u o
o u
u
u
u
+
E
E
E +
(12)
We test our NHMM model on simulation data, where there are 40 households, each with
500 observations. The model recovered well on the simulation data (see Appendix III for the
result).
3.4 Empirical Analyses
3.4.1 Data Description
Our data are provided by a leading Internet grocery retailer which was among the very
first to sell groceries online. The dataset was collected during a 62-week period in 1996-1997
when the retailer first launched its web business. Given that this retailer was a pioneer of the
online grocery business, it is very likely that consumers in our data never had prior exposure to
other online grocery stores. This feature makes our dataset particularly suited to study the
evolution of purchase behavior in a new online environment.
The data include click-stream records of detailed navigation processes and purchase
history information of 225 households, as well as pricing information for multiple product
categories. We estimate the proposed model using data of two distinct product categories,
spaghetti sauce and liquid detergent, in order to test the robustness of the results. These
categories are chosen because they differ in terms of hedonicity and purchase frequency, which
have been shown to affect consumers‘ in-store decision making processes (Inman, Winer, and
Ferraro 2009). Only those households that made at least two category purchases during the 62-
week data period are included in the data for a category.
60
3.4.2 Operationalization of Key Variables
As explained previously, we use price as the key marketing mix variable for the Tobit
model of purchase incidence and quantity decisions in the empirical analysis. We first obtain the
weekly actual price of each stock-keeping units (SKU) in a category by combining its regular
price and price discount (if any), and then convert it to a common unit price (e.g., cents per
ounce) across SKUs. The category-level price variable is computed as a weighted average of the
weekly unit prices of all SKUs in the category, where the weights are sales volume shares of the
SKUs in the entire time period.
Decision aid usage information is extracted from the click-stream data. We classify six
types of decision aids that are commonly available in most online stores and fall into one of the
four broad categories described earlier, including one for nutritional needs, one for brand
preference, two for economic needs, and two for personalized shopping lists (see Table 3.1). For
each decision aid, we measure a household‘s usage experience by week t as the cumulative
number of usage counts up to the prior week t-1. This measure avoids possible reverse causality,
that is, purchase behavior may influence the current usage of decision aids. Note that a
household can use the same decision aid multiple times during a single shopping session and the
cumulative experience variables count for each and every time a given decision aid is used by a
household. A main objective of our research is to investigate how usage experience with
different decision aids may drive purchase behavior changes over time. To this end, the
transition probability functions of the proposed Hidden Markov Model include the usage
experience variables of multiple decision aids. Since the usage experience with each decision aid
increases monotonically with time, a more meaningful measure for the model is the relative
usage experience with each decision aid, computed as a percentage of the household‘s total
61
number of usage counts of all decision aids up to week t-1. We also include the total decision aid
usage variable to account for its potential effect on the transition probabilities.
Descriptive statistics of the key variables for each product category are presented in
Table 3.2. Note that the usage experience variables are store-level measures and thus do not vary
across categories. For each decision aid examined, we report its cumulative usage counts and as
a percentage of the total usage counts, for up to week t-1 as well as by the last week in the data
(Week 62). As the table shows, shopping list is the most frequently used decision aid (30.66
times on average in 62 weeks), far out-numbering the usage of the other types of decision aids,
followed by previous order list (5.27 times), sorting by brand name (4.06 times), sorting by
nutrition (1.93 times), sorting by price (1.17 times) and sorting by promotion (.38 times). As
shown by the standard deviations, the experience measures vary substantially across individuals.
Since we need to exclude at least one of the relative usage experience variables from the model
to avoid perfect collinearity, we choose to take out sorting by promotion because it had the
lowest occurrence amongst the six focal decision aids of interest.
[INSERT TABLES 3.1 & 3.2 HERE]
3.4.3 Time-Varying Patterns of Usage Experience with Decision Aids
To inspect how the usage experience may evolve over time, we take the average of each
relative usage experience variable in a week across households and plot them over time (see
Figure 3.1). Note that the relative usage experience measures do not necessarily sum up to 100%
in all weeks because some households had not tried any decision aids in the earlier weeks. The
relative usage experience with personal shopping list and previous order list, which implies
habitual decision processes, increases over time; while the relative usage experience with sorting
by brand name, by price, and by promotion, which are indicative of on-the-spot decision
62
processes, levels off in the later stage of the observation period. This is consistent with previous
literature which suggests that weights of price information will decrease when other product
attributes are available (Anderson 1971and 1981; Bettman, Capon, and Lutz 1975). The time-
varying patterns of the different decision aids indicate that, on average, consumers tend to adopt
more habitual behavior and make fewer on-the-spot purchase decisions as they get more
accustomed to the type of online shopping environment as studied here.
[INSERT FIGURE 3.1 ABOUT HERE]
3.4.4 Model Estimation Results
In Table 3.3, we compare the BIC, DIC, and log-marginal densities of models with
different numbers of hidden states to determine the best number of states for each category (see
Hughes, Guttorp, and Charles 1999; Netzer et al. 2008). These comparisons indicate that a three-
state model fits the data best for both spaghetti sauce and liquid detergent. Estimation results of
the three-state model for the two categories are presented in Tables 3.4 and 3.5, respectively.
[INSERT TABLES 3.3, 3.4, and 3.5 ABOUT HERE]
As shown by Table 3.4, households‘ purchase behavior differ substantially across the
three hidden states for the spaghetti sauce category. By the model construction, the baseline
purchase incidence tendency increases from hidden state 1 (S1) to hidden state 3 (S3). Our
empirical result shows that the same order of baselines also holds for purchase quantity decisions
across the three states, even though they are not constrained to be so. Price has a negative and
?significant? effect on the purchase incidence probability in all three states
5
, but its effect is the
strongest in S2 (-.620), followed by S1 (-.144), while its effect in S3 is much smaller (-.005). The
5
For the ease of exposition, hereafter, we report the posterior means in parentheses and use the term “significant” to
refer to the case where the posterior 95% credible interval does not cover zero.
63
same pattern of price sensitivity also holds for the purchase quantity decision. Thus, S1
represents a low baseline purchase tendency (i.e., store loyalty) and medium price sensitivity
state, S2 is characterized as having medium baseline purchase tendency and the highest price
sensitivity decisions, and S3 exhibits the highest baseline purchase tendency and the lowest price
sensitivity.
The three hidden states of the liquid detergent category show very similar patterns, and
can be interpreted in the same fashion.
The lower panels of Table 3.4 and Table 3.5 report parameter estimates of the variables
that affect the transition probabilities between hidden states. For the spaghetti sauce category,
when consumers are in S1, more usage experience with sorting by brand name significantly
increases the probability of staying in S1 and reduces the probability of switching to the more
store-loyal states S2 and S3 (posterior mean = -2.578), while the total decision aid usage
significantly increases the probability of switching to a more store-loyal state (posterior mean =
1.153). When consumers are in S2, more usage experiences with sorting by brand name and
sorting by price discourage the transition to S3 while encourages the transition to S1 (posterior
mean = -1.071 and -.420, respectively), while more usage experiences with sorting by nutrition
and personal shopping lists have the opposite effects (posterior mean = .410, and 1.950,
respectively). When consumers are in S3, more usage experiences with personal shopping lists
and previous order lists increase the likelihood of staying in S3 and reduce the transition
probabilities from S3 to a less store-loyal state (posterior mean = .946 and .495, respectively).
The effects of the usage experience variables exhibit a similar pattern in the liquid
detergent category. When consumers are in the least store-loyal state (S1), more usage
experience with sorting by brand name discourages switching to the states with higher levels of
64
store loyalty (S2 and S3). It is possible that consumers with strong brand preference (brand
loyalty) are willing to switch stores if the preferred brand is not available in the focal store.
When consumers are in the medium store-loyalty state (S2), usage experiences of sorting by
brand name and sorting by price increase the probability of switching to S1 and reduce the
probability of switching to the most store-loyal state (S3), while usage experiences with personal
shopping lists and previous order lists, as well as the total decision aid usage count, have the
opposite effects. When consumers are in the most store-loyal state (S3), usage experiences of
personal shopping lists and previous order lists and the total decision aid usage count reinforce
the probability of staying in this state and decrease the chance of switching to a less store-loyal
state, while usage experiences with sorting by brand name and by price do not have any
significant effects any more.
An interesting contrast with results of the spaghetti sauce category is that, for the liquid
detergent category, usage experience with sorting by nutrition does not have any effects on the
purchase behavior evolution. This difference between the two categories is expected, and it
attests to our model‘s ability detect the distinct effects (or the lack of which) of usage experience
with different decision aids.
3.4.5 Evolution of Purchase Behavior in the Online Store
Our results indicate that there are different hidden behavior states and consumers switch
among these states over time, reflecting an evolution of purchase behavior when shopping in a
new online store environment. To investigate the time varying patterns of the purchase behavior,
we need to compute the posterior distribution of the three states for each household in each
week, and then explore the relationship between price sensitivity measures and the usage
experience variables.
65
3.4.5.1 State Probability Distribution
To better understand the evolution of purchase behavior, we firstly examine the
probabilities of the states for each household over time. The filtering probability of household i
belonging to state s at week t (Netzer, et al. 2008) can be calculated as:
(13) ), , | ( / ) , , ( ) , | ( ... ) , , ( ) , | ( ) , , (
) ,..., , , ,... , | ( Pr
1 2 2 1 2 1 1 s
2 1 2 1 ,
2 1
i,1
s s c
it
s s
i
C
t
C
t t t i
s s
i
C C
i
s s
i
C C
it i i it i i t i
Q l Z X f s s s P Z X f s s s P Z X f
Q Q Q I I I s S ob
it i i
| | t
=
÷
= =
O · · O · O · =
=
where P
i
(s
t
|s
t-1,
s) is the s
th
column of transition matrix P(s
t
|s
t-1
), ) , | (
s s c
it
Q l | | is the likelihood of
the observed purchase incidence and quantity up to week t. Figure 3.2 presents the average
probabilities of belonging each state in 62 weeks. For the spaghetti sauce category, the average
probability of belonging to S1, which has the lowest store loyalty and medium price sensitivity,
decreases substantially over time (from 67.86% in week 1 to 16.01% in week 62). In contrast, the
average probabilities of belonging to S2 and S3 increase over time (S2: from 19.03% in week 1
to 30.48% in week 62; S3: from 13.11% in week 1 to 53.50% in week 62). On average,
consumers are more likely to be in S1 than in the other two states in the first half of the observed
period (till week 32). The average probability of belonging to S3, the state with the highest store
loyalty and lowest price sensitivity, surpasses that for S2, and S3 becomes the dominant behavior
state from week 32 and on. This pattern also holds for the liquid detergent category: the average
probability of belonging to S1 declines while those for S2 and S3 increase over time, and S3
becomes the dominant behavior state since week 39, for the average consumer.
[INSERT FIGURE 3.2 ABOUT HERE]
3.4.5.2 Effects of Usage Experience with Decision Aids on Price Sensitivity
Since the states are not in ascending or descending order, the direction of the changes of
price sensitivity, and the relationship between the usage experience with decision aids and the
66
price sensitivity are unknown based on the model estimation per se. Therefore, we compute a
posterior price sensitivity measure for each household in each week, and then regress this
measure on the usage experience variables. To account for uncertainty in posterior distributions
of price coefficients, we calculate a weighted posterior price sensitivity measure.
Specifically, we take the last 1,000 draws of the price coefficients, and for each draw,
price coefficients for household i in week t are weighted by its probabilities of belonging to each
corresponding hidden state in that week. Then we average the weighted price coefficients across
the 1000 draws to get the price sensitivity measure for each household across 62 weeks. The
time-varying patterns of price sensitivity measures are plotted in Figure 3.3 and 3.4, for the
spaghetti sauce and liquid detergent categories, respectively.
For the spaghetti sauce category, the average values of the price sensitivity measures for
purchase incidence and purchase quantity firstly increase then decrease over the observed period
(see the upper panels in Figure 3.3). In addition, the variances of these measures increase over
time, indicating that consumers‘ price sensitivity diverges over time, with some becoming less
price-sensitive while other becoming more price-sensitive, as they become more accustomed to
an online store environment (see the lower panels in Figure 3.3). The plots for the liquid
detergent category reveal the same patterns. Prior research shows that a higher degree of
heterogeneity in price sensitivity is conducive to more granular price promotion customization
(Zhang and Wedel 2009). This divergence in price sensitivity over time suggests that online
retailers have a good opportunity to customize their price promotions to cater to individual
consumers‘ needs and preferences, as consumers become more used to shopping online. In the
following, we will examine the impact of different decision aids on price sensitivity through
post-hoc analysis.
67
[INSERT FIGURE 3.3 and 3.4 ABOUT HERE]
Since the hidden states identified by our model are not ordered by their level of price
sensitivity, we cannot directly infer from the parameter estimates how usage experience with
each decision aids affects consumers‘ price sensitivity evolution. To investigate this issue, we
conduct regression analyses where the dependent variables are the price sensitivity measures for
the purchase incidence or the purchase quantity decision, and the explanatory variables are the
usage experience with different decision aids. We use a random-effect model to allow
heterogeneity in price sensitivity across households. The price coefficients are estimated from
the models, and thus the dependent variables are also measured with uncertainty. Since the
posterior distributions of the price coefficients are unknown, we use the simulated maximum
likelihood estimation (MLE) method to estimate the models, where the draws of the price
coefficients are a natural product of the MCMC procedures of the main model estimation. We
use the last 1,000 draws of each MCMC procedure for the simulated MLE.
Table 3.6 present the results of these regression analyses. For both categories, usage
experience with personal shopping lists and previous order lists can mitigate consumers‘ price
sensitivity in purchase incidence decisions, and make them more loyal to the focal retailer. In
contrast, usage experience with sorting by price can train consumers to become more efficient at
using price information in the liquid detergent category and thus more sensitive to it in their
purchase decisions. Usage experience with sorting by brand name also leads to higher price
sensitivity in the spaghetti sauce category. Thus, the more usage experience with such decision
aids, the more responsive consumers become to price changes.
[INSERT TABLE 3.6 ABOUT HERE]
68
3.5 Discussion
In this study, we investigate whether and how the usage experience with different
decision aids drives the evolution of purchase behavior. We empirically identified three latent
states that direct the purchase behavior over time. They are: hidden state 1where consumers
show lowest store loyalty /medium price sensitivity; hidden state 2 that characterized with
medium store loyalty /high price sensitivity; and hidden 3 that exhibits highest store loyalty /low
price sensitivity. Post-hoc analysis shows that the probabilities of staying in hidden state 1
declines over time, while the probabilities of staying in the other two hidden states demonstrate a
reverse trend. In the latter half of the 62-week period, hidden state 3 dominates. Such pattern
implies that as consumers get more accustomed to the online store environment, their baseline
tendency to purchase from the store increases. In addition, their price sensitivity diverges over
time, with some consumers becoming more price sensitive, while others becoming less price
sensitive.
The transitions among these hidden states are driven by the relative usage experience
with different decision aids. For both spaghetti sauce and liquid detergent categories, more
relative usage experience with the decision aids for brand preference (sorting by brand names)
discourages the transitions from lower store-loyal state to higher store-loyal state. More relative
usage experience with decision aids for economic needs (sorting by price) also reduces the
transition probabilities from lower store-loyal state to higher store-loyal state, in addition, post-
hoc analyses show that usage experience with sorting by price in the liquid detergent category
trains consumers to become more responsive to price changes. In contrast, more relative usage
experience with decision aids for personalized shopping list, such as shopping list and previous
order list, as well as sorting by nutritional information for the spaghetti sauce category, increase
69
the transition from medium store-loyal / high price sensitive state (hidden state 2) to high store-
loyal / low price sensitive state (hidden state 3). Post-hoc analysis also confirmed the inverse
relationship between the usage experience with personalized shopping lists and price sensitivity.
Personalized shopping lists thus help build store loyalty, increase purchase propensity, and ease
price competition.
These findings have important implications for the design of online store environments
and communication messages regarding a firm‘s pricing decisions. To encourage transition from
the more price sensitive state to less price sensitive, online retailers should allow consumers to
create shopping lists, make available their previous order lists, provide decision aids for
searching a variety of nutritional information, and encourage their usage among the customers.
The personalized shopping lists will create a lock-in effect and help consumer build store loyalty.
Decision aids aimed at economic needs, on the other hand, are a double-edged sword. A low-
price online retailer could benefit from higher consumer responsiveness to price by enabling and
promoting the usage of such type of decision aids like sorting by price, but usage of these
decision aids would lower consumers‘ loyalty to the store in the long run. In addition, their usage
could negatively affect the long-term business for retailers that adopt a premier pricing strategy
and do not compete on promotions. What decision aids to offer and to emphasize should depend
on an online retailer‘s overall positioning and pricing strategies, and weigh in the trade-offs of
the retailer‘s short-term versus long-term needs.
In terms of future research, it would be interesting to investigate whether there are carry-
over effects of usage experience across product categories, i.e., would the usage experience in
one product category affect the evolution of purchase behavior in other categories. Effects of the
store-level usage experience variables, as shown by our study, strongly suggest such possibility.
70
Given the proliferation of multi-channel retailing, another worthy direction is to study the impact
of usage experience with online decision aids on offline purchase behavior offline, which would
require matching online navigation data and offline purchase history data. In addition, how the
pattern of decision aids usage evolves over time itself is an interesting topic to explore. All these
topics offer exciting venues for gaining deeper understanding of purchase behavior evolution in
the ever-evolving retail environment.
71
Table 3.1 Types of Online Decision Aids Examined
Broad Category Decision Aid Definition
For nutritional needs Sort by Nutrition
Sort by nutritional criteria, including calories,
sodium, fat, Kosher, sugar, carbohydrates,
fiber, and cholesterol
For brand preference
Sort by Brand Name
Sort by brand name
For economic needs
Sort by Price
Sort by price information, including unit price
and item price
Sort by Promotion Sort by promotion status
Personalized
shopping list
Shopping List
Retrieve and use personal shopping list
Previous Order List
Retrieve and use previous order list
72
Table 3.2 Descriptive Statistics
Variable (mean/sd of variables across
weeks and across individuals)
Cumulative usage count up to
week t-1
As percentage of total
decision aid usage up to
week t-1 (%)
Mean S.D. Mean S.D.
Store Level Tool Usage:
--- Sort by Nutrition
--- Sort by Brand Name
--- Sort by Price
--- Sort by Promotion
--- Shopping List
--- Previous Order List
--- Total decision aid usage (six tools)
1.06
2.32
0.75
0.25
16.70
2.71
23.79
4.23
6.36
2.25
0.73
30.3
5.75
38.04
3.96
7.99
2.91
1.42
55.71
12.42
9.85
16.60
8.81
5.65
37.35
21.73
Variable (mean/sd of last week’s
measure for all individuals)
Cumulative usage count in
week 62
As percentage of total
decision aid usage in
week 62 (%)
Mean S.D. Mean S.D.
Store Level Tool Usage:
--- Sort by Nutrition
--- Sort by Brand Name
--- Sort by Price
--- Sort by Promotion
--- Shopping List
--- Previous Order List
--- Total decision aid usage (six tools)
1.93
4.06
1.17
0.38
30.66
5.27
43.47
6.30
8.52
3.03
0.91
47.18
8.92
57.17
4.55
10.29
3.43
1.50
63.67
16.56
9.19
17.74
10.83
5.34
31.36
23.67
Variable Mean S.D.
Category 1: Spaghetti Sauce (N =137 households)
Purchase frequency (times/per year)
Purchase quantity (ounces/occasion)
Regular price (cents/oz.)
Price discount (cents/oz.)
Paid price (cents/oz.)
5.63
34.62
9.03
0.31
8.64
6.15
21.23
0.37
0.25
0.42
Category 2: Liquid Detergent (N =159 households)
Purchase frequency (times/per year)
Purchase quantity (ounces/occasion)
Regular price (cents/oz.)
Price discount (cents/oz.)
Paid price (cents/oz.)
5.53
124.70
7.15
0.48
6.67
6.49
69.25
0.30
0.38
0.45
73
Table 3.3 Model Comparisons
Spaghetti Sauce Liquid detergent
Number of Hidden
States
LMD BIC DIC LMD BIC DIC
K=2 -4,786.7 9,565.0 9,430.9 -5,992.6 12,078.9 11,901.2
K=3 -4,128.4 8,426.9 8,217.3 -5,337.1 10,848.2 10,692.1
K=4 -4,593.1 9,583.3 9,492.5 -5,889.3 12,184.5 12,103.2
Note: LMD = log-marginal density;
BIC = Bayesian Information Criterion;
DIC = Deviance Information Criterion.
74
Table 3.4 Estimation Result for Spaghetti Sauce
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
Variables in the Tobit Model
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Purchase Incidence
Intercept -2.171 -3.055 -1.287 -1.787 -2.148 4.170 -1.080 -1.255 -0.847
Actual price -0.144 -0.149 -0.140 -0.620 -1.204 -0.036 -0.005 -0.007 -0.003
Purchase Quantity
Intercept 0.960 0.302 1.618 2.190 2.081 2.310 2.258 2.193 3.907
Actual price -0.279 -0.605 0.046 -0.926 -1.718 -0.135 0.001 -0.019 0.021
Variables Affecting the
Transition Probabilities
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
(relative usage experience)
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Cut-off point 1
1.248 -0.014 2.509 2.067 1.231 2.904 0.939 0.545 1.133
Cut-off point 2
1.314 1.183 1.444 2.648 -0.875 6.171 1.937 0.301 3.572
Sort by Brand Name
-2.578 -4.024 -1.131 -1.071 -1.459 -0.683 -0.621 -1.536 0.293
Sort by Price
-0.236 -1.070 0.597 -0.420 -0.597 -0.243 -0.008 -0.269 0.253
Sort by Nutrition
-0.838 -1.846 0.171 0.410 0.019 0.802 -0.585 -2.331 1.162
Shopping List
1.182 -0.051 2.414 1.950 0.928 2.972 0.946 0.351 1.540
Previous Order List
-0.460 -1.273 0.353 0.685 -0.019 1.390 0.495 0.230 0.760
Total decision aid usage
1.153 0.647 1.658 2.013 -0.008 4.035 0.489 -0.713 1.691
Note: The bold font indicates that the 95% credible interval does not cover zero.
75
Table 3.5 Estimation Result for Liquid Detergent
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
Variables in the Tobit Model
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Purchase Incidence
Intercept -0.707 -1.195 -0.219 1.228 0.102 1.027 1.384 0.637 2.130
Actual price -0.228 -2.341 1.886 -0.496 0.121 -0.733 -0.100 -0.116 -0.084
Purchase Quantity
Intercept -1.374 -1.645 -1.103 3.102 -1.164 5.368 3.925 3.052 4.798
Actual price -0.859 -2.057 0.340 -1.441 -1.571 -1.311 -0.120 -0.239 -0.001
Variables Affecting the
Transition Probabilities
Hidden State 1 (S1) Hidden State 2 (S2) Hidden State 3 (S3)
(relative usage experience)
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Posterior
mean
2.50% 97.50%
Cut-off point 1
-1.310 -2.380 -0.241 -0.235 -0.812 0.342 -1.073 -1.851 -0.296
Cut-off point 2
-0.953 -1.027 -0.878 0.701 0.279 1.123 1.102 0.537 1.667
Sort by Brand Name
-0.380 -0.532 -0.229 -0.104 -0.149 -0.059 -0.053 -1.711 1.604
Sort by Price
-0.804 -1.616 0.009 -1.023 -1.627 -0.420 0.157 -1.250 1.563
Sort by Nutrition
-0.016 -0.578 0.546 0.134 -1.603 1.870 0.035 -1.440 1.510
Shopping List
0.142 -0.115 0.398 1.149 0.743 1.556 1.928 0.792 3.065
Previous Order List
0.767 -1.003 2.537 0.987 0.606 1.367 0.586 0.118 1.054
Total decision aid usage
1.466 -0.228 3.160 1.259 0.247 2.270 1.276 0.924 1.628
Note: The bold font indicates that the 95% credible interval does not cover zero.
76
Table 3.6 The Impact of Usage Experience with Decision Aids on Price Sensitivity (Simulated MLE)
Spaghetti Sauce
Purchase Incidence Purchase Quantity
Mean S.E. t-stat P-value Mean S.E. t-stat P-value
Intercept 0.113 0.006 17.523 0.000 0.228 0.008 29.497 0.000
Intercept_var 0.215 2.508 -0.612 0.542 0.361 1.730 -0.589 0.557
Sort by Brand Name -0.022 0.011 -2.066 0.041 -0.005 0.018 -0.301 0.764
Sort by Price -0.015 0.033 -0.463 0.644 -0.011 0.051 -0.213 0.832
Sort by Nutrition 0.019 0.029 0.677 0.499 0.025 0.034 0.739 0.461
Shopping List 0.032 0.008 4.112 0.000 0.027 0.015 1.781 0.077
Previous Order List 0.022 0.011 1.998 0.048 -0.002 0.009 -0.268 0.789
Liquid Detergent
Purchase Incidence Purchase Quantity
Mean S.E. t-stat P-value Mean S.E. t-stat P-value
Intercept
0.214 0.004 58.584 0.000 0.683 0.053 12.980 0.000
Intercept_var
0.218 0.424 3.586 0.000 0.204 0.213 7.443 0.000
Sort by Brand Name
-0.002 0.010 -0.209 0.835 -0.008 0.019 -0.417 0.677
Sort by Price
-0.135 0.068 1.981 0.049 -0.032 0.022 -1.456 0.147
Sort by Nutrition
0.013 0.015 0.873 0.384 0.003 0.003 1.050 0.295
Shopping List
0.058 0.014 4.056 0.000 0.787 0.037 21.327 0.000
Previous Order List
0.032 0.012 2.739 0.007 0.059 0.010 5.727 0.000
77
Figure 3.1
Time-Varying Patterns of Usage Experience with Decision Aids
78
Figure 3.2 Probabilities of Belonging to Each State
Spaghetti Sauce
Liquid Detergent
79
Figure 3.3 Evolution of Price Sensitivity over Time
Spaghetti Sauce
80
Liquid detergent
81
Chapter IV: Conclusion
This dissertation empirically investigates dynamic consumer decision processes in E-
Commerce settings. In this chapter, we will re-examine the findings from Chapter 2 and 3, points
out the contributions of this dissertation, and conclude with an exploration of possible future
directions for this stream of research.
4.1 Summary of the Two Essays
Essay one uses eye tracking data to investigate information acquisition processes on
attribute-by-product matrices as encountered in online comparison websites. The study identifies
two information acquisition modes (attribute-based or product-based), and documents the extent
of usage and switching of these acquisition modes through Hierarchical Hidden Markov Model
(HHMM). Post-hoc analyses focus on the specific information processed in these modes, the
possible causes and consequences of switching, and the dynamics in the comparison set during
decision making. We find that the low-level properties of the eye and the visual brain (horizontal
and local eye movement tendency) stimulate congruent information processing mode and
contiguous information acquisition, reflecting a stronger influence on information acquisition
processes than previously documented in the traditional process-tracing studies. In addition, we
find frequent switching between acquisition modes when consumers face information-rich
comparison websites. The extensive switching might be explained from the previously obtained
information, and limited visual short term memory. Higher switching frequency significantly
increases decision time, and reduces the experienced ease-of-processing.
Essay two conducts an empirical investigation on whether and how usage experience
with various types of decision aids affects the evolution of online purchase behavior. By
82
adopting Non-homogeneous Hidden Markov Model (NHMM), we empirically identified three
latent behavior states that govern the evolution of online purchase behavior and the drivers
behind the evolvement of behavior. As consumers get more accustomed to the online store
environment, their baseline tendency to purchase from the store increases, and their price
sensitivity diverges over time. More relative usage experience with shopping lists, previous order
list, and sorting by nutrition (in Spaghetti sauce category), encourages the transition from
medium store-loyal/high price sensitivity state to high store-loyal /low price sensitivity state; yet
more relative usage experience with sorting by brand name and sorting by price reduces the
probability of switching from lower store-loyal state to higher store-loyal state. Post-hoc analysis
reveals that more usage experience with sorting by brand name in the spaghetti sauce category
and sorting by price in the liquid detergent category increase price sensitivity, while personalized
shopping lists are important decision aids that help retailers build store loyalty and soften the
price competition.
4.2 Contribution and Managerial Implications
This dissertation has made the following contributions:
Firstly, we take advantage of recent development in technology and investigate the
dynamic decision making process with eye tracking data and online navigation data. These path
data opens up the possibilities to exploit micro-level dynamics during decision making. In the
first study, we rely on eye tracking equipment to accurately document moment to moment
information search activities when consumers making their purchase on comparison websites.
The eye tracking equipment is able to capture fast, adaptive, partially unconscious information
acquisition behavior during real-life decision making. This is different from traditional process
tracing methods such as information display boards. Prior process tracing methods require motor
83
responses, which renders the information acquisition process more controlled and deliberate. As
a consequence, these studies could discover mostly high-level and slower cognitive processes
that consumers engage in during decision making. Thanks to the eye movement data, our study is
able to offer a sneak peek into both ?overt and covert?, and ?voluntary and involuntary? aspects
of decision process (Lynch and Srull 1982). Low-level properties of the eye and the visual brain
that are underexploited in previous accounts of decision process, appear to play an important role
when decision-making is fast and more automatic. Insight into these covert and involuntary
information tendencies will lead to a better understanding of decision-making in information-rich
online choice environments.
In the second study, we collect online navigation data and purchase history data to
investigate purchase behavior evolution. What is lacking in the literature is a comprehensive,
longitudinal examination of the evolving patterns of purchase behavior, and more importantly,
the drivers behind purchase behavior changes in online stores. Our study fills this void by
extracting the information of decision aids usage from online navigation data and using it to
explain the evolution of purchase behavior. Navigation data provides valuable information of
consumers‘ online activities. The usage experience with decision aids extracted from navigation
data reveals the information search and evaluation process before purchase occurs. It impacts the
weights of different attributes including price, the extent of brand and store loyalty, the
importance of personalized shopping environment, etc. The usage experience data collected over
time reflects learning and adaptation processes that consumers go through when shop in a new
channel. This information is helpful in explaining the evolution of purchase propensity and
responsiveness to marketing mix, yet difficult to obtain in offline settings.
84
Secondly, we adopt a flexible evolutionary structure to model decision dynamics. Eye
movement data is dense, stochastic, and highly dynamic. It reflects the latent information
acquisition modes probabilistically, rather than deterministically. The HHMM in essay one deals
with large volumes of eye movement data, unobserved information modes, and a probabilistic
link between these two through structural restrictions. The model effectively copes with the
challenge to extract diagnostic information from the massive amount of eye movement data. The
three hierarchical layers in HHMM represent longer range dependencies in eye-movements, and
quantitatively capture prevalence and flexible switching pattern of information acquisition
modes.
The classification of different purchase behaviors has been a long-standing research
interest in the marketing literature. In essay two, we propose that the observed online shopping
behavior is likely to be directed by certain latent ?behavior states?, and these hidden states differ
in terms of the baseline tendency to purchase from the online store and their price sensitivity.
NHMM automatically identifies three latent behavior states, and allows flexible transitions
among these behavior states. In addition, the non-homogeneous structure in NHMM enables us
to explore the drivers behind the evolutions of these hidden states, and reveals how the store
environment trains consumers and changes their behavior states over time.
Managerially, our results have implications for the design of online shopping websites
which are gaining popularity recently. Retailers and manufacturers try to optimize online
information displays to affect consumers‘ information acquisition processes and their purchase
decision. The first study shows that low-level, automatic and unconscious processes are principal
in situations of information overload and time pressure, which occur when facing data-rich web-
based choice environments. Comparison website developers could display formats in a proactive
85
fashion to stimulate consumers‘ use of specific information acquisition mode favorable to their
goals. Dominance of local information acquisition requires careful allocation of attributes /
products. By strategically placing the contiguous information, retailers may be able to alternate
the attractiveness of certain attribute. A delicate balance among format, sales objective, and
switching cost is also needed. For example, we find that that a Product-Column format invites
more attributed-based acquisition and increases the chance of the dominant product being
chosen; while a Product-Row format leads to more evenly distributed choice probabilities.
However, the latter format leads to more local eye movements and switching between acquisition
modes, smaller numbers of products and attributes being inspected, and reduced ease of decision
making. In addition, some comparison websites that enable sorting by certain attribute may need
to reconsider their default format, which is the Product-Row format. This format may be in
conflict with desirable attribute-based information processing for these websites.
The second study reveals that the more usage experience with personalized shopping list,
and sorting by nutrition (in the spaghetti sauce category) leads to the transition from medium
store-loyal/high price sensitivity behavior state to high store-loyal/low price sensitivity behavior
state; while more usage experience with decision aids for economic needs and brand preference
discourages the transition from lower store-loyal states to higher store-loyal states. If online
retailers could encourage consumers to create and store personal shopping lists, make available
their previous order lists, and encourage their usage among the customers, it will facilitate the
transition from the more price sensitive state to the less price sensitive state, create a lock-in
effect, build store loyalty, and train consumers into habitual shoppers. Retailers should be
careful with the offering of decision aids aimed at economic needs, however. A low-price-driven
online retailer could benefit from higher consumer responsiveness to price by enabling and
86
promoting the usage of such type of decision aids, yet the long-term business may be harmed.
Our study also provides some insights for online retailers to offer personalized shopping
environment. Information of previous usage of decision aids and purchase behavior collected in
the navigation data enables retailers to infer the distribution of different latent behavior states for
individual customers. Retailers could customize the marketing mix variables based on the
purchase propensity and price sensitivity of each behavior states; and encourage the usages of
certain decision aids to influence customer‘s purchase evolution path.
4.3 Future Research
We hope this dissertation will stimulate further research interest in the dynamic decision
making process in the online shopping environment. The first study raises several intriguing
questions, for instance, how to effectively reduce frequent switching between different
information acquisition modes? How do other design factors, such as sorting, grouping,
trimming, and highlighting, affect the dynamic information acquisition processes and post-
decision evaluations? Based on information acquisition patterns, can comparisons websites
automatically adapt display format during decision making and increase the conversion rate?
The second study also offer several exciting research opportunities, for instance, what
drives the evolution of decision aids usage itself over time? How does the usage experience with
decision aids online affect the purchase behavior offline? Will the usage experience with one
product category affect the evolution of purchase behavior in the other category (?spill-over?
effect)? All these future research questions will help us enrich our understandings of dynamic
decision making in E-commerce.
Besides these two essays, we conduct another study of dynamic decision making on a
special type of websites: search engines. Specifically, we collect eye movement data on a major
87
search engine website that includes organic, right sponsored and top sponsored sections, over a
variety of search tasks, and investigate section and snippet inspection decisions that online
shoppers engaged in when search for product or service information on search engine results
page. We explore the impact of two types of stimuli on information processes: one is high level
stimuli, i.e., textual information; the other is low level stimuli, such as snippet location, density,
and section intrinsic property. The integration of newly acquired textual information with prior
knowledge, and the lag effect of low level stimuli, create a context that constantly changes as
decision progress. Therefore, sequential inspection decisions are dynamic and interrelated. We
apply a computational cognitive model with static utilities and dynamic utilities generated from
two types of stimuli to capture this preference formation process. Our results show that
sponsored sections create a strong ?stickiness? effect: once consumers enter these sections, they
are likely to inspect more than one snippet within the same section instead of leaving
immediately. In addition, the impact of snippet location on inspection probabilities varies across
organic and sponsored sections. In terms of high level stimuli, descriptive information, such as
product attribute or quality, decreases consumers‘ ?stickiness? with the current section, while
transactional information, such as price and promotion, has an opposite effect. The study furthers
the understanding of the effect of different types of stimuli, especially the dynamic influence
from snippet content, on the information seeking behavior from a consumer‘s viewpoint. It offers
insights in designing snippet‘s textual information on a search engine result page with respect to
the location and content of its competitors. Implications for the platform (search engines) are
also quite broad – from textual result selection, to textual content pairing, to optimal blending.
88
Appendices
Appendix I Transition Probabilities between Middle and Upper Layer States for the Two
Information Presentation Formats (with standard deviations in parentheses)
Product Column Format
Product Row Format
Middle Layer Transition Matrices
S
1
=AB S
1
=PB
S
1
=AB S
1
=PB
S
1
=AB 0.769 (0.021) 0.231 (0.021)
S
2
=1
S
1
=AB 0.746 (0.047) 0.231 (0.047)
S
1
=PB 0.096 (0.038) 0.904 (0.038) S
1
=PB 0.047 (0.011) 0.953 (0.011)
S
1
=AB S
1
=PB S
1
=AB S
1
=PB
S
1
=AB 0.586 (0.029) 0.414 (0.029)
S
2
=2
S
1
=AB 0.719 (0.036) 0.281 (0.036)
S
1
=PB 0.456 (0.071) 0.544 (0.071) S
1
=PB 0.552 (0.087) 0.448 (0.087)
S
1
=AB S
1
=PB
S
1
=AB S
1
=PB
S
1
=AB 0.520(0.056) 0.480 (0.056)
S
2
=3
S
1
=AB 0.323 (0.019) 0.677 (0.019)
S
1
=PB 0.451(0.013) 0.549 (0.013) S
1
=PB 0.334 (0.038) 0.666 (0.038)
Upper Layer Transition Matrices
S
2
=1 S
2
=2 S
2
=3 S
2
=1 S
2
=2 S
2
=3
0.461 (0.099) 0.244 (0.012) 0.295 (0.090) S
2
=1 0.549 (0.047) 0.161 (0.088) 0.290 (0.070)
0.225 (0.067) 0.400 (0.045) 0.375 (0.034) S
2
=2 0.204 (0.099) 0.493 (0.054) 0.303 (0.101)
0.209 (0.111) 0.446 (0.091) 0.345 (0.089) S
2
=3 0.187 (0.087) 0.496 (0.026) 0.317 (0.095)
Note: AB: attribute-based acquisition; PB: product-based acquisition
89
Appendix IIa Estimated Attribute Transition Matrices
product-column condition : attribute transition probabilities product-row condition: attribute transition probabilities
picture price processor
operating
sys
memory
keyboard
mouse
monitor
hard
drive
optical
drive
wireless
office
software
warranty picture price processor
operating
sys
memory
keyboard
mouse
monitor
hard
drive
optical
drive
wireless
office
software
warranty
0.016 0.420 0.271 0.055 0.055 0.038 0.024 0.040 0.015 0.024 0.025 0.018 picture 0.018 0.584 0.175 0.048 0.033 0.041 0.030 0.031 0.011 0.010 0.012 0.008
0.424 0.004 0.370 0.068 0.024 0.019 0.023 0.007 0.026 0.017 0.007 0.011 price 0.421 0.005 0.325 0.057 0.046 0.022 0.040 0.022 0.031 0.012 0.012 0.007
0.205 0.304 0.007 0.349 0.041 0.032 0.020 0.015 0.009 0.004 0.005 0.008 processor 0.121 0.402 0.004 0.307 0.050 0.029 0.010 0.030 0.018 0.010 0.007 0.010
0.090 0.065 0.311 0.005 0.399 0.058 0.016 0.020 0.010 0.007 0.007 0.013
operating
sys
0.087 0.103 0.253 0.005 0.419 0.065 0.018 0.016 0.007 0.013 0.011 0.004
0.071 0.060 0.096 0.276 0.005 0.391 0.031 0.026 0.007 0.006 0.016 0.015 memory 0.074 0.128 0.149 0.173 0.006 0.322 0.076 0.040 0.009 0.009 0.011 0.004
0.094 0.022 0.036 0.042 0.301 0.004 0.395 0.063 0.014 0.013 0.013 0.003
keyboard
mouse
0.095 0.061 0.068 0.060 0.210 0.005 0.411 0.048 0.008 0.008 0.012 0.015
0.100 0.029 0.020 0.019 0.044 0.293 0.006 0.418 0.021 0.016 0.019 0.015 monitor 0.094 0.085 0.071 0.032 0.071 0.178 0.005 0.325 0.075 0.027 0.029 0.008
0.076 0.016 0.029 0.015 0.016 0.023 0.333 0.013 0.376 0.085 0.008 0.008
hard
drive
0.038 0.083 0.067 0.017 0.022 0.049 0.313 0.007 0.336 0.018 0.031 0.020
0.023 0.019 0.014 0.016 0.011 0.033 0.050 0.348 0.013 0.385 0.075 0.014
optical
drive
0.049 0.033 0.042 0.015 0.013 0.016 0.070 0.202 0.007 0.430 0.075 0.047
0.066 0.024 0.027 0.030 0.024 0.022 0.020 0.038 0.273 0.011 0.385 0.080 wireless 0.025 0.026 0.039 0.010 0.011 0.023 0.018 0.075 0.342 0.007 0.355 0.069
0.073 0.018 0.019 0.006 0.028 0.028 0.025 0.006 0.032 0.292 0.009 0.465
office
software
0.012 0.031 0.041 0.029 0.025 0.006 0.026 0.015 0.072 0.245 0.007 0.492
0.050 0.039 0.073 0.056 0.024 0.041 0.029 0.037 0.068 0.105 0.463 0.015 warranty 0.028 0.085 0.101 0.028 0.030 0.048 0.050 0.077 0.058 0.100 0.381 0.014
Appendix IIb Estimated Product Transition Matrices
Product-column condition:
product transition probabilities
product-row condition:
product transition probabilities
Label Product 1 Product 2 Product 3 Product 4 Label Product 1 Product 2 Product 3 Product 4
0.005 0.767 0.155 0.028 0.045 Label 0.004 0.747 0.159 0.079 0.010
0.186 0.008 0.521 0.193 0.092 Product 1 0.250 0.010 0.508 0.149 0.084
0.037 0.410 0.004 0.418 0.130 Product 2 0.037 0.339 0.006 0.499 0.120
0.007 0.157 0.308 0.008 0.520 Product 3 0.039 0.176 0.308 0.010 0.466
0.042 0.168 0.287 0.486 0.017 Product 4 0.053 0.190 0.224 0.509 0.024
90
Appendix III Model Estimation Test on Synthetic Data
Variables Starting Value
Estimated Value
(posterior mean)
Real Value
PI_0_S1 0.1 0.22875228 0.2
PI_MKT1_S1 0.1 0.61499769 0.5
PI_MKT2_S1 0.1 -0.45321166 -0.3
PI_0_S2 0.1 0.39564543 0.3
PI_MKT1_S2 0.1 0.47681012 0.4
PI_MKT2_S2 0.1 -0.38364839 -0.2
PI_0_S3 0.1 0.54535731 0.5
PI_MKT1_S3 0.1 0.80136447 0.8
PI_MKT2_S3 0.1 -0.61044463 -0.6
PQ_0_S1 0.1 0.66825696 0.5
PQ_MKT1_S1 0.1 0.31771122 0.7
PQ_MKT2_S1 0.1 -0.91325732 -0.8
PQ_0_S2 0.1 0.99265948 0.9
PQ_MKT1_S2 0.1 1.31265599 1.2
PQ_MKT2_S2 0.1 -0.95266036 -0.9
PQ_0_S3 0.1 0.58022575 0.8
PQ_MKT1_S3 0.1 1.15305817 0.9
PQ_MKT2_S3 0.1 -1.42524140 -1.2
TL1_S1 0.1 0.68878361 0.8
TL2_S1 0.1 -0.31458116 -0.5
TL1_S2 0.1 -0.81875761 -0.9
TL2_S2 0.1 0.32136565 0.5
TL1_S3 0.1 0.87714547 0.7
TL2_S3 0.1 0.25091591 0.2
Gamma_S1 0.1 0.84548087 0.9
Gamma_S2 0.1 1.36810117 1.5
Gamma_S3 0.1 0.90530073 1.2
Rho_S1 0.1 0.55360011 0.5
Rho_S2 0.1 0.78094405 0.7
Rho_S3 0.1 0.80169427 0.9
Sigma[1] 0.1 0.26430815 0.2
Sigma[2] 0.2 0.49261638 0.4
Sigma[3] 0.1 0.50207708 0.4
Sigma[4] 0.2 0.69167942 0.7
Sigma[5] 0.1 0.60207708 0.8
Sigma[6] 0.2 1.04969463 1.0
V_theta[1,1] 1.0 0.98644001 1.0
V_theta[2,2] 1.0 1.05860775 1.0
V_theta[3,3] 1.0 1.02430815 1.0
V_theta[4,4] 1.0 0.89679048 1.0
V_theta[5,5] 1.0 0.95966898 1.0
V_theta[6,6] 1.0 0.91105910 1.0
91
Bibliography
Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard Lutz, Alan
Sawyer, and Stacy Wood (1997), ?Interactive Home Shopping: Consumer,
Retailer, And Manufacturer Incentives to Participate in Electronic
Marketplaces,? Journal of Marketing, 61 (3), 38-53.
Amemiya, Tomohiro (1984), ?Tobit Models: A Survey,? Journal of Econometrics, 24
(1-2), 3-61.
Anderson, Norman H. (1971), ?Integration Theory and Attitude Change,?
Psychological Review, 78(3), 171-206.
Anderson, Norman H. (1981), Foundations of Information Integration Theory. New
York: Academic Press.
Ansari, Asim, Carl Mela, and Scott A. Neslin (2008), ?Customer Channel Migration?,
Journal of Marketing Research, 45(1), 60-76.
Babin, Barry J., William R. Darden, and Mitch Griffin (1994), ?Work and/or Fun:
Measuring Hedonic and Utilitarian Shopping Value,? Journal of Consumer
Research, 20(4), 644-656.
Bakos, Yannis (2001), ?The Emerging Landscape for Retail E-Commerce,? Journal
of Economic Perspectives, 15(1), 69-80.
Ball, Christopher (1997), ?A Comparison of Single-Step and Multi-Step Transition
Analyses Of Multiattribute Decision Strategies,? Organizational Behavior Human
Decision Processes, 69 (3), 195-204.
Baldi, Pierre, Yves Chauvint, Tim Hunkapiller, and Marcella A. Mcclure (1994),
?Hidden Markov Models of Biological Primary Sequence Information?,
Proceedings of the National Academy of Science of the United States of America,
91, 1059-1063
Bechwati, Nada Nasr, and Lan Xia (2003), ?Do Computers Sweat? The Impact of
Perceived Effort of Online Decision Aids on Consumers' Satisfaction with The
Decision Process,? Journal of Consumer Psychology, 13( ½), 139-148.
Bettman, James. R., and Jacob Jacoby (1976), ?Patterns Of Processing In Consumer
Information Acquisition,? Advances in Consumer Research, 3(4), 315-320.
———, Noel Capon, and Richard J. Lutz (1975), ?Cognitive Algebra in Multi-
Attribute Attitude Models,? Journal of Marketing Research, 12 (2), 151-64.
——— (1979), An Information Processing Theory of Consumer Behavior. Addison-
Wesley. Reading MA.
92
———, Pradeep Kakkar (1977), ?Effects of Information Presentation Format on
Consumer Information Acquisition Strategies,? Journal of Consumer Research,
3(4), 233-240.
———, Mary F. Luce, John W. Payne (1998), ?Constructive Consumer Choice
Process,? Journal of Consumer Research, 25(3), 187-217.
———, C.Whan. Park (1980), ?Effects of prior knowledge and experience and phase
of the choice process on consumer decision processes: A protocol analysis,?
Journal of Consumer Research, 7(3), 234-248.
Block, Lauren B. and Vicki G. Morwitz (1999), ?Shopping Lists as an External
Memory Aid for Grocery Shopping: Influences on List Writing and List
Fulfillment,? Journal of Consumer Psychology, 8 (4), 343-75.
Brangule-Vlagsma, Kristine, Rik G. M. Pieters
,
and Michel Wedel (2002), ?The
Dynamics of Value Segments: Modeling Framework And Empirical
Illustration?, International Journal of Research In Marketing, 19(3),267-285
Bucklin Randolph E. ,and Catarina Sismeiro (2003), ?A Model of Web Site Browsing
Behavior Estimated on Clickstream Data?, Journal of Marketing Research,
40(3), 249-267
Chen, Lei-da, Mark L. Gillenson, and Daniel L. Sherrell (2002), ?Enticing Online
Consumers: An Extended Technology Acceptance Perspective,? Information
and Management, 39 (8), 705-719.
Cheung, Christy M. K., Lei Zhu, Timothy Kwong, Gloria W.W. Chan, and Moez
Limayem (2003), ?Online Consumer Behavior: A Review and Agenda for
Future Research,? Proceedings of the 16th Bled eCommerce Conference, 194-
218.
Chib, Siddhartha (1995), ?Marginal Likelihood from the Gibbs Output,” Journal of
the American Statistical Association, 90(432), 1313-1321.
Childers, Terry L., Christopher L. Carr, Joann Peck, and Stephen Carson (2001),
?Hedonic and Utilitarian Motivations for Online Retail Shopping Behavior,?
Journal of Retailing, 77 (4), 511-535.
Cowan, Nelson (2001) ?The Magical Number 4 in Short-Term Memory: A
Reconsideration of Mental Storage Capacity,? Behavior and Brain Science, 24,
87–185.
Danaher, Peter J., Isaac W. Wilson, and Robert A. Davis (2003), ?A Comparison of
Online and Offline Consumer Brand Loyalty,? Marketing Science, 22 (4), 461–
76.
93
De Jong, Ritske (2000), ?An Intention-Activation Account Of Residual Switch
Costs,? Control of Cognitive Processes: Attention and Performance eds. S.
Monsell and J. Driver. Vol. XVIII 357–376. Cambridge, MA: MIT Press.
Degeratu, Alexandru M., Arvind Rangaswamy, and Jianan Wu (2000), ?Consumer
Choice Behavior in Online and Traditional Supermarkets: The Effects of Brand
Name, Price, And Other Search Attributes,? International Journal of Research
in Marketing, 17 (1), 55-78.
Dekimpe Marnik G., and Dominique M. Hanssens (2000), ?The Persistence of
Marketing Effects On Sales,? Marketing Science, 14 (1), 1-21.
Dekimpe Marnik G., and Dominique M. Hanssens (2000), ?Time-Series Models In
Marketing: Past, Present and Future,? International Journal of Research in
Marketing, 17(2-3), 183-193
Du, Rex Y., and Wagner A. Kamakura (2006), ?Household Life Cycles and
Lifestyles In The United States,? Journal of Marketing Research, 43(1), 121-
132.
Erdem, Tulin. (1996), ?A Dynamic Analysis of Market Structure Based On Panel
Data,? Marketing Science, 15(4), 359-378.
Fine, Shai, Yoram. Singer, and Naftali Tishby ( 1998), ?The Hierarchical Hidden
Markov Model: Analysis and Applications,? Machine Learning, 32, 41-62.
Gelfand, Alan. E., and Adrian F. M. Smith (1990), ?Sampling based approaches to
calculating marginal densities,? Journal of American Statistical Association,,
85, 398-409.
Gilchrist, Iain.V., and Monika Harvey (2006), ?Evidence for A Systematic
Component Within Scan Paths In Visual Search. Visual Cognition, 14
(4/5/6/7/8), 704-715.
Glaholt, Mackenzie G., and Eyal M. Reingold (2009), ?Stimulus Exposure and Gaze
Bias: A Further Test Of The Gaze Cascade Model,? Attention, Perception &
Psychophysics, 71, 445-450.
Gopher, Daniel, Lilach Armony and Yaakov Greenshpan (2000), ?Switching Tasks
And Attention Policies,? Experimental Psychology, 129 (3), 308-339.
Guadagni, Peter M., and John D.C. Little (1983),?A Logit Model of Brand Choice
Calibrated on Panel Data,? Marketing Science, 2(3), 203-238
Häubl, Gerald and Valerie Trifts (2000), ?Consumer Decision Making in Online
Shopping Environments: The Effects of Interactive Decision Aids,? Marketing
Science, 19 (1), 4–21.
94
Hauser John R., and Kenneth J. Wisniewski (1982), ?Dynamic Analysis of Consumer
Response to Marketing Strategies?, Management Science, 28(5), 455-486.
Heckman, James J. (1981), ?Heterogeneity and State Dependence,? in Studies in
Labor Markets, S. Rosen (Eds.) University of Chicago Press, Chicago 91-139.
Heilman, Carrie M., Douglas Bowman, and Gordon P. Wright (2000), ?The
Evolution of Brand Preferences and Choice Behavior of Consumers New to A
Market,? Journal of Marketing Research, 37 (2), 139-155.
Heller, Katherine A., Yee Whye The, and Dilan Görür (2009), ? Infinite Hierarchical
Hidden Markov Models,? Proceedings of the 12th International Conference on
Artificial Intelligence and Statistics (AISTATS), 5, Clearwater, Florida.
Hirschman, Elizabeth C. and Morris B. Holbrook (1982), ?Hedonic Consumption:
Emerging Concepts, Methods and Propositions,? Journal of Marketing, 46 (3),
92-101.
Hoffman, Donna L. and Thomas P. Novak (1996), ?Marketing in Hypermedia
Computer-Mediated Environments: Conceptual Foundations,? Journal of
Marketing, 60 (3), 50-68.
Hollander, Stanley C. and Kathleen M. Rassuli (1999), ?Shopping with Other
People‘s Money: The Marketing Management Implications of Surrogate-
Mediated Consumer Decision Making,? Journal of Marketing, 63 (2), 102-118.
Huber, Joel, John W. Payne, and Christopher Puto (1982), ?Adding Asymmetrically
Dominated Alternatives: Violations of Regularity and The Similarity
Hypothesis,? Journal of Consumer Research, 9(1), 90-98.
Hui Sam K., Eric T. Bradlow, and Peter S. Fader (2009a), ?Testing Behavioral
Hypotheses Using An Integrated Model Of Grocery Store Shopping Path And
Purchase Behavior,? Journal of Consumer Research, 36, 478-493.
Hui Sam K., Eric T. Bradlow, and Peter S. Fader (2009b), ?Path Data in Marketing:
An Integrative Framework and Prospectus for Model Building,? Marketing
Science, 28(2), 320-335.
Hughes, James P. (1993), ?A Class of Stochastic Models for Relating Synoptic
Atmospheric Patterns to Local Hydrologic Phenomena,? PhD dissertation.
University of Washington. Seattle.
Hughes, James P., and Peter Guttorp (1994a), ?A Class of Stochastic Models for
Relating Synoptic Atmospheric Patterns to Regional Hydrologic Phenomena,?
Water Resource Research, 30(5), 1535–1546.
95
Hughes, James P., and Peter Guttorp (1994b), ?Incorporating Spatial Dependence and
Atmospheric Data In A Model Of Precipitation,? Journal of Applied
Meteorology, 33(12),1503–1515.
Hughes, James P., Peter Guttorp, and Stephen P. Charles (1999), ?A Non-
homogeneous Hidden Markov Model for Precipitation Occurrence,? Journal of
the Royal Statistical Society: Series C (Applied Statistics), 48(1), 15-30.
Inman, J. Jeffrey, Russell S. Winer, and Rosellina Ferraro (2009), ?The Interplay
among Category Characteristics, Customer Characteristics, and Customer
Activities on In-Store Decision Making,? Journal of Marketing, 73 (5), 19-29.
Johnson, Eric J., Michael Schulte-Mecklenbeck, and Martijn C. Willemsen (2008)
?Process Models Deserve Process Data: Comment On Brandstatter. Gigerenzer,
And Hertwig,? Psychological Review, 115(1), 263-273.
Kawanaka, Daiki, Takayuki Okatani, And Koichiro Deguchi , (2006), ?HHMM
Based Recognition of Human Activity?, IEICE TRANSACTIONS on
Information and Systems, E89-D (7), 2180-2185
Kupiec, J Julian (1992), ?Robust Part of Speech Tagging Using A Hidden Markov
Model,? Computer Speech & Language, 6, 225-242.
Lee, Byung-Kwan, and Wei-Na Lee (2004), ?The Effect of Information Overload on
Consumer Choice Quality In An Online Environment,? Psychology and
Marketing, 21(3), 159–183.
Lee, Jinkook and Loren V. Geistfeld (1998), ?Enhancing Consumer Choice: Are We
Making Appropriate Recommendations? ? Journal of Consumer Affairs, 32(2),
227-251.
Liechty, John, Rik Pieters, and Michel Wedel (2003), ?Global and Local Covert
Visual Attention: Evidence from A Bayesian Hidden Markov Model,?
Psychometrika, 68(4), 519-541.
Limayem, Moez, Mohamed Khalifa, and Anissa Frini (2000), ?What Makes
Consumers Buy from Internet? A Longitudinal Study of Online Shopping,?
IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and
Humans, 30 (4), 421-432.
Lohse, Gerald, and Eric Johnson (1996), ?A Comparison of Two Process Tracing
Methods for Choice Tasks,? Proceedings of the 29th Annual Hawaii
International Conference on System Sciences, 86-97.
Luck, Steven J., and Edward K. Vogel (1997), ?The Capacity of Visual Working
Memory For Features And Conjunctions,? Nature, 390, 279–281.
96
Lynch, John G., and Thomas K. Srull (1982), ?Memory and Attentional Factors in
Consumer Choice: Concepts and Research Methods,? Journal of Consumer
Research, 9 (1), 18-37.
MacDonald, Iain L., and Walter Zucchini (1997), Hidden Markov and Other Models
for Discrete-Valued Time Series. Chapman and Hall, London
Marois, Rene´, and Jason Ivanoff (2005), ?Capacity Limits of Information Processing
In The Brain,? Trends in Cognitive Sciences, 9(6), 296-305.
Mathwick, Charla and Edward Rigdon (2004), ?Play, Flow, and the Online Search
Experience,? Journal of Consumer Research, 31 (2), 324-332.
Mauldin, Elaine and Variram Arunachalam (2002), ?An Experimental Examination
of Alternative Forms of Web Assurance for Business-To-Consumer E-
Commerce,? Journal of Information Systems, 16 (1), 33-55.
Montgomery, Alan L., Shibo Li, Kannan Srinivasan, and John C. Liechty (2004),
?Modeling Online Browsing And Path Analysis Using Clickstream Data,?
Marketing Science, 23(4), 579-595.
Nag, R., K. Wong, F. Fallside (1986), ?Script Recognition Using Hidden Markov
Models,? Acoustics Speech and Signal Process, 2071-2074.
Netzer, Oded, James Lattin, and V. Srinivasan (2008), ?A Hidden Markov Model of
Customer Relationship Dynamics,? Marketing Science, 27, 185-204.
Park, C. Whan, Easwer S. Iyer, and Daniel C. Smith (1989), ?The Effects of
Situational Factors on In-Store Grocery Shopping Behavior: The Role of Store
Environment and Time Available for Shopping,? Journal of Consumer
Research, 15 (4), 422-33.
Pavlou, Paul A. (2003), ?Consumer Acceptance of Electronic Commerce: Integrating
Trust and Risk with the Technology Acceptance Model,? International Journal
of Electronic Commerce, 7 (3), 101-134.
Payne, John W. (1976), ?Task Complexity And Contingent Processing In Decision
Making: An Information Search And Protocol Analysis,? Organizational
Behavior Human Decision Processes, 16(2), 366-387.
———, James R. Bettman, and Eric J. Johnson (1993), The Adaptive Decision
Maker, Cambridge University Press, Cambridge, UK.
Pieters, Rik G.M., and Luk Warlop (1999), ?Visual Attention during Brand Choice:
The Impact Of Time Pressure And Task Motivation,? International Journal of
Research in Marketing, 16, 1-16.
97
Pieters, Rik G.M., and Michel Wedel (2004), ?Attention Capture and Transfer in
Advertising: Brand, Pictorial, and Text-Size Effects,? Journal of Marketing,
68(2), 36-50.
Pieters, Rik G.M., Michel Wedel, and Jie Zhang (2007), ?Optimal Feature
Advertising Design Under Competitive Clutter,? Management Science, 53(11),
1815-1828.
Rabiner, Lawrence R., and B. H. Juang (1986), ?An Introduction to Hidden Markov
Models,? IEEE ASSP Magazine, 3(1), 4-16.
Rabiner, Lawrence R. (1989), ?A Tutorial on Hidden Markov Models and Selected
Applications in Speech Recognition?, Proceedings of the IEEE, 77(2), 257-286.
Rayner, Keith, (1998), ?Eye Movement in Reading and Information Processing: 20
Years of Research,? Psychological bulletin, 24(3), 372-422.
Robert, Christian P., Gilles Celeux, and Jean Diebolt (1993), ?Bayesian Estimation
Of Hidden Markov Chains: A Stochastic Implementation,? Statistics and
Probability Letters, 16, 77-83.
Robertson, Andrew W., Sergey Kirshner, and Padhraic Smyth (2004), ?Downscaling
Of Daily Rainfall Occurrence Over Northeast Brazil Using A Hidden Markov
Model,? Journal of Climate, 17(22) 4407-4424.
Robertson, Andrew W., Amor V. M. Ines, and James W. Hansen (2007), ?
Downscaling Of Seasonal Precipitation For Crop Simulation,? Journal of
Applied Meteorology and Climatology, 46, 677-693.
Rogers, Robert D., and Stephen Monsell (1995), ?The Cost of A Predictable Switch
between Simple Cognitive Tasks,? Journal of Experimental Psychology,
General, 124, 207–231.
Rosbergen, Edward , Rik Pieters,and Michel Wedel (1997), ?Visual Attention to
Advertising: A Segment ?Level Analysis?, Journal of Consumer Research,
24(3), 305-314.
Rossi, Peter E., and Greg. M. Allenby (2003), ?Bayesian Statistics and Marketing,?
Marketing Science, 22(3), 304–328.
Russo, J. Edward, and Larry D. Rosen (1975), ?An Eye Fixation Analysis of
Multialternative Choice?, Memory and Cognition, 3, 267-276.
——— (1978), ?Adaptation of Cognitive Processes to the Eye Movement System,?
In Eye Movements and the Higher Psych. Functions. eds. John W. Senders et al.
Hillsdale. NJ: Erlbaum,89-109.
98
——— , and France Leclerc (1994), ?An Eye-Fixation Analysis of Choice Process
For Consumer Nondurables,? Journal of Consumer Research, 21(2), 274-290.
Salvucci, Dario D., and John R. Anderson (1998), ?Tracing Eye Movement Protocols
With Cognitive Process Models,? In Proceedings of the Twentieth Annual
Conference of the Cognitive Science Society, Hillsdale. NJ:Erlbaum, 923-928.
Scott, Steven L (2002), ?Bayesian Methods For Hidden Markov Models: Recursive
Computing In The 21st Century,? Journal of American Statistic Association,
97(457), 337-351.
Seetharaman, P. B., Andrew Ainslie, and Pradeep K. Chintagunta (1999),
?Investigating Household State Dependence Effects Across Categories,?
Journal of Marketing Research, 36(4), 488-500
Senter, Stuart M., and Douglas H. Wedell (1999), ?Information Presentation
Constraints and the Adaptive Decision Maker Hypothesis,? Journal of
Experimental Psychology, 23(2), 428-446.
Shimojo, Shinsuke, Claudiu Simion, Eiko Shimojo, and Christian Scheier (2003),
?Gaze Bias Both Reflects and Influences Preference,? Nature Neuroscience,
6(12), 1317-1322.
Skounakis, Marios, Mark Craven, and Soumya Ray (2003), "Hierarchical Hidden
Markov Models for Information Extraction", Proceedings of the 18th
International Joint Conference on Artificial Intelligence, 427-433.
Smith, Michael D., Joseph Bailey, and Erik Brynjolfsson (2000), Understanding
Digital Markets. E. Brynjolfsson, B. Kahin, eds. MIT Press, Cambridge, MA.
Spalek, Thomas M, and Sherief Hammad (2005), ?The Left-To-Right Bias in
Inhibition of Return Is Due to the Direction of Reading,? Psychological Science,
16(1), 15-18.
Stewart, Stewart, Nick Chater, and Gordon D.A. Brown (2006), ?Decision by
Sampling,? Cognitive Psychology, 53, 1-26.
Swait, Joffre, and Wiktor Adamowicz (2001), ?The Influence of Task Complexity on
Consumer Choice: A Latent Class Model Of Decision Strategy Switching?,
Journal of Consumer Research, 28(1), 135-148.
Swedowsky, Maya (2009), ?Online Grocery Shopping: Ripe Timing for Resurgence,?
The Nielsen Company. (www.nielsen.com, accessed on April 20, 2010)
Tatler, Benjamin W., and Benjamin T. Vincent (2008), ?Systematic Tendencies In
Scene Viewing,? Journal of Eye Movement Research, 2(2), 1-18.
99
Treisman, Anne M., and Garry Gelade (1980), ?A Feature-Integration Theory of
Attention,? Cognitive Psychology, 12 (1), 97–136.
Underhill, Paco (2004), Call of the Mall: The Geography of Shopping. Simon &
Schuster, New York
Van der Lans, Ralf, Rik Pieters, and Michel Wedel (2008), ?Eye Movement Analysis
Of Search Effectiveness,? Journal of American Statistical Association,
103(482), 452-461.
Wedel, Michel, and Rik Pieters (2000), ?Eye Fixation on Advertisements and
Memory for Brands: A Model and Findings,? Marketing Science, 19(4), 297-
312.
Wedell, Douglas H., and Stuart Senter (1997), ?Looking and Weighting in Judgment
and Choice,? Organizational Behavior and Human Decision Processes, 70(1),
41-64.
Zhang, Hua-Ping, Hong-Kui Yu, De-Yi Xiong and Qun Liu (2003), HHMM-based
Chinese lexical analyzer ICTCLAS, Proceedings of the second SIGHAN
workshop on Chinese language processing, 17, 184-187
Zhang, Jie and Aradhna Krishna (2007), ?Brand-Level Effects of Stockkeeping Unit
Reductions,? Journal of Marketing Research, 44 (4), 545-559.
Zhang, Jie and Michel Wedel (2009), ?The Effectiveness of Customized Promotions
in Online and Offline Stores,? Journal of Marketing Research, 46 (2), 190–206.
Zhou, Lina, Liwei Dai,
and Dongsong Zhang (2007), ?Online Shopping Acceptance
Model – A Critical Survey of Consumer Factors in Online Shopping,? Journal
of Electronic Commerce Research, 8 (1), 41-62.
doc_327947935.pdf