Description
A paradox is an argument that produces an inconsistency, typically within logic or common sense.
Human capital is the key to the IT productivity paradox
Gudmundur Gunnarsson Erik Mellander Eleni Savvidou
WORKING PAPER 2004:13
The Institute for Labour Market Policy Evaluation (IFAU) is a research institute under the Swedish Ministry of Industry, Employment and Communications, situated in Uppsala. IFAU’s objective is to promote, support and carry out: evaluations of the effects of labour market policies, studies of the functioning of the labour market and evaluations of the labour market effects of measures within the educational system. Besides research, IFAU also works on: spreading knowledge about the activities of the institute through publications, seminars, courses, workshops and conferences; creating a library of Swedish evaluational studies; influencing the collection of data and making data easily available to researchers all over the country. IFAU also provides funding for research projects within its areas of interest. There are two fixed dates for applications every year: April 1 and November 1. Since the researchers at IFAU are mainly economists, researchers from other disciplines are encouraged to apply for funding. IFAU is run by a Director-General. The authority has a traditional board, consisting of a chairman, the Director-General and eight other members. The tasks of the board are, among other things, to make decisions about external grants and give its views on the activities at IFAU. A reference group including representatives for employers and employees as well as the ministries and authorities concerned is also connected to the institute. Postal address: P.O. Box 513, 751 20 Uppsala Visiting address: Kyrkogårdsgatan 6, Uppsala Phone: +46 18 471 70 70 Fax: +46 18 471 70 71 [email protected] www.ifau.se
Papers published in the Working Paper Series should, according to the IFAU policy, have been discussed at seminars held at IFAU and at least one other academic forum, and have been read by one external and one internal referee. They need not, however, have undergone the standard scrutiny for publication in a scientific journal. The purpose of the Working Paper Series is to provide a factual basis for public policy and the public policy discussion.
ISSN 1651-1166
Human capital is the key to the IT productivity paradox?
Gudmundur Gunnarsson† Erik Mellander‡ Eleni Savvidou§ October 4, 2004
Abstract Unlike previous analyses, we consider (i) possible externalities in the use of IT and ii) IT and human capital interactions. Examining, hypothetically, the statistical consequences of erroneously disregarding (i) and (ii) we shed light on the small or negative growth e?ects found in early studies of the e?ects of IT on productivity growth, as well as the positive impacts reported more recently. Our empirical analysis uses a 14-industry panel for Swedish manufacturing 1986-95. We ?nd that human capital developments made the average e?ect of IT essentially zero in 1986 and steadily increasing thereafter, and, also, generated large di?erences in growth e?ects across industries. JEL codes : O33, L23, L60 Keywords : IT Productivity Paradox, Applied Econometrics
? We thank Per-Anders Edin, Karolina Ekholm, Nils Gottfries, Bertil Holmlund, Matthew Lindquist, Mikael Lundholm, Eva Mörk, Sten Nyberg, Hans Wijkander and two anonymous referees for constructive comments Anders Hintze and Michael Wolf at Statistics Sweden generously helped us with data. Financial support from the Swedish Council for Work Life Research, the Swedish Transport & Communications Research Board and the Swedish Agency for Innovation Systems is gratefully acknowledged. † Dept. of Business Administration and Information Systems, Mälardalen University, P.O. Box 883, SE — 721 23 Västerås, Sweden; email: [email protected] ‡ Institute for Labour Market Policy Evaluation (IFAU), P.O. Box 513, SE-751 20 Uppsala, Sweden; email: [email protected] § Dept. of Economics, Uppsala University, P.O. Box 513, SE-751 20 Uppsala, Sweden; email: [email protected]
IFAU–Human capital is the key to the IT productivity paradox
1
Contents
1 Introduction 2 Literature review: attempts to explain the paradox 3 A stylized model 4 Data and empirical speci?cation 4.1 The growth rate in total factor productivity 4.2 Speci?cation of the explanatory variables . 4.3 Measures of IT equipment and IT use . . . 4.4 The human capital data . . . . . . . . . . . 4.5 Control variables . . . . . . . . . . . . . . . 3 5 9 17 18 20 22 25 27
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
5 Results 29 5.1 Testing the implications of the stylized model . . . . . . . . 30 5.2 Econometric issues . . . . . . . . . . . . . . . . . . . . . . . 32 5.3 Multivariate speci?cations of human capital . . . . . . . . . 38 6 Summary and conclusions A Computation of omputer capital 48 57
2
IFAU–Human capital is the key to the IT productivity paradox
1
Introduction
The IT productivity paradox was formulated in response to the fact that the massive investments in information technology (IT) that started around 1980 did not seem to have any positive e?ects on productivity growth. In the words of Nobel laureate Robert Solow: ”You can see the computer age everywhere but in the productivity statistics.” [Solow (1987)] In recent years, the original focus on computers has been broadened to include also communication devices: the concept of IT has been extended to ICT, information and communication technology. In this paper, we account for the development of communications equipment. We have kept the term IT, however. In empirical studies, the IT productivity paradox has been veri?ed in analyses based on early (pre—1990) data for the U.S. and Canada. Mostly, the results show either very small or insigni?cant e?ects of IT on productivity growth; see for instance Harris & Katz (1991) and Parsons, Gotlieb, & Denny (1993). Indeed, some studies have reported signi?cantly negative e?ects; cf. Loveman (1988) and Berndt & Morrison (1995). Some of the explanations suggested for these counter-intuitive results are: the time required for IT investments to yield productivity increases has been underestimated, the magnitude of the investments have been overestimated and measurement problems on both the input side and the output side have concealed the productivity e?ects. However, a couple of more recent studies, using data extending to the end of the 1990’s, have found productivity—increasing e?ects of IT. Oliner & Sichel (2000) argue that the reason why there were no e?ects earlier is that, in the U.S., IT investments did not really take o? until 1995. When they did, the e?ects were substantial, however: Oliner & Sichel claim that
IFAU–Human capital is the key to the IT productivity paradox 3
IT accounted for about two—thirds of the acceleration in the labor productivity between the ?rst and second halves of the 1990’s. Bresnahan, Brynjolfsson, & Hitt (2002), while focusing primarily on skillbiased technical change rather than productivity, make an important contribution towards the resolution of the IT productivity paradox by extending the idea of capital-skill complementarity hypothesis discussed by Griliches (1969) and Lucas (1990). Bresnahan et al. (op.cit.) argue that too much attention has been paid to IT investments and too little attention has been paid to work organization and human capital structure. Accounting for both IT and human capital, they ?nd that the balance between the two is crucial. Firms with high levels of both IT and human capital are found to be the most productive. More interesting: ?rms with low levels of both IT and human capital are shown to be more productive than ?rms that are high on IT and low on human capital, or vice versa. The framework we suggest in this paper is similar to the Bresnahan et al. (op.cit.) approach in the sense that we, too, conjecture that human capital is a key element in the explanation of the IT productivity paradox. However, we extend the analysis by incorporating a phenomenon often discussed in the context of endogenous growth theory, namely knowledge spillovers. While it seems very natural to consider knowledge spillovers in an evaluation of the productivity e?ects of IT, these have barely been discussed in earlier studies. The next section contains a review of some attempts to explain the IT productivity paradox. In Section 3 we develop a simple stylized growth model. By means of this model we discriminate between some of the suggested explanations for the IT productivity and, second, propose a way to account for knowledge spillovers.
4 IFAU–Human capital is the key to the IT productivity paradox
Our empirical analysis is based on data for 14 industries in the Swedish manufacturing sector observed annually during the period 1986—95. It appears that in the Swedish manufacturing sector the productivity-enhancing e?ects of IT started to show already in the ?rst half of the 1990s, i.e. a couple of years earlier than, e.g., in the U.S. Otherwise, the developments in Sweden seems to have been qualitatively similar to that in several other countries. Our data are described in Section 4 and the results are provided in Section 5. Section 6 contains a summary of our results and our conclusions.
2
Literature review: attempts to explain the paradox
For brevity, we here only provide a very condensed and selective list of some the explanations suggested for the IT productivity paradox.1 1. Investments in IT became massive only towards the end of the 1990s. Thus, early analyses were unable to capture positive growth e?ects from IT simply because, at the time, these investments were still comparatively small. Studies using later data should be able to discern positive growth e?ects. This view is supported by the study by Oliner & Sichel (2000). However, this explanation says nothing about the signi?cant negative e?ects of IT on productivity estimated by, e.g., Loveman (1988) and Berndt & Morrison (1995). 2. It takes time before the productivity-enhancing e?ects of a new technology can be realized. This point has perhaps been most convinc1 For a more extensive discussion see, e.g., Triplett (1999). Also, for the view that there is essentially no paradox to explain, because the importance of the introduction of IT has been vastly exaggerated, compared to the signi?cance of other technological developments like the adoption of electricity, see Gordon (2000).
IFAU–Human capital is the key to the IT productivity paradox
5
ingly made by David (1990). From an empirical point of view, this explanation is similar to the previous one. An important di?erence, however, is that this explanation can account for (initial) negative e?ects of IT on productivity, provided that the di?usion of IT use is associated with learning costs that decrease over time, as a function of the increasing number of users. This explanation also points to the importance of (positive) externalities. More wide-spread knowledge about (how to exploit) IT will speed up the rate of di?usion. The resulting increase in people with access to IT will raise the bene?ts accruing to individual users, which will further accelerate di?usion. The importance of this spiralling e?ect has been especially notable in the 1990’s, with the rapidly expanding use of email and the Internet. 3. No account has been taken of the complementarity between IT and skilled workers. Although the capital-skill complementarity hypothesis was put forward already by Griliches (1969), the connection between IT and human capital has almost invariably been disregarded in assessments of the productivity e?ects of IT.2 Presumably, this is primarily due to lack of data. However, by matching two di?erent data sets Bresnahan, Brynjolfsson, & Hitt (2002) have overcome this problem. Splitting their data into four categories according to whether ?rms are ”high” or ”low” on IT and human capital, they ?nd high levels of productivity in ?rms that are either high on both IT and human capital or low in both of these dimensions. Relatively lower levels of productivity are found in ?rms that are high in one
2 However, complementarity between IT and skilled workers has been documented in several studies of labor demand and skill-biased technical change. Two seminal contributions are Berman, Bound, & Griliches (1994) and Autor, Katz, & Kreuger (1998). For a study using Swedish data, see Mellander (1999).
6
IFAU–Human capital is the key to the IT productivity paradox
of the two dimensions and low in the other.3 Using a di?erent approach, Kaiser (2003) also ?nds strong evidence for complementarity between expenditures on IT capital and outlays for IT personnel. 4. IT is a general purpose technology (GPT), the e?cient implementation of which requires changes in work practices and skill upgrading. This explanation contains elements of explanations 2 and 3. The idea is that the introduction of GPTs like IT will initially lead to a slowdown in productivity, as it takes time to implement and learn to use the GPT e?ciently. In particular, assuming skilled labor to have a learning advantage over unskilled labor, the theory holds that skill premia will rise, inducing an increased supply of skills. When the increased supply comes about and the work organization is properly adapted to the GPT, productivity starts increasing again. The notion of GPTs was introduced by Bresnahan & Trajtenberg (1995) and the relation between GPTs and productivity growth is discussed in, e.g., Helpman & Trajtenberg (1998), and Greenwood & Yorukoglu (1997). 5. Mismeasurement of outputs. According to this explanation, the use of information technology has increased the quality of existing products and services and created new goods, neither of which are (fully) captured in the o?cial statistics. This has led to a downward bias in the estimated growth e?ects; see, e.g., Brynjolfsson (1993) and Dean (1999). Nevertheless, it is essential to point out, like Lee & Barua (1999) do, that e?ciency related gains in the production of
A related approach is taken by Siegel (1997), who considers the possibility that the investments in IT may induce enhanced e?ciency of labor which, in turn, positively a?ects productivity growth. He ?nds some, although not unambiguous, support for this hypothesis.
3
IFAU–Human capital is the key to the IT productivity paradox
7
the ”old” goods should still be accounted for by conventional output measures. That is to say, while mismeasurement of output certainly is part of the puzzle it cannot resolve it entirely. 6. Mismeasurement of inputs. On the input side the issue of mismeasurement is less clear-cut than on the output side. On the one hand, it can be argued that early (U.S.) measures of IT were overstated because they included equipment that one would not ordinarily associate with IT like, e.g., typewriters and accounting machinery.4 On the other hand, the often noted di?culties to adjust for quality increases in IT price indexes implies a tendency to underestimate the volumes of IT investments.5 And the presence of positive externalities in the use of IT, cf. the second point above, points in the same direction. Failure to account for these externalities will, again, bias measures of IT inputs downwards. 7. Overinvestments in IT, in the latter half of the 1980s. This explanation has been suggested by Morrison (1997), based on the ?nding that in U.S. manufacturing industries estimated bene?t—cost ratios (Tobin’s q ) for IT capital dropped signi?cantly below 1 by the mid 1980’s. It is natural to interpret the term ”overinvestment” in a relative sense here, i.e. that IT investments were too large compared
These were included in Bureau of Economic Analysis category ”O?ce Computing and Accounting Machinery; cf Berndt & Morrison (1995). After 1982 this category was replaced by ”Information Processing and Related Equipment”, see Lee & Barua (1999). 5 For a hedonic approach to the estimation of price indexes for computers, see Berndt, Griliches & Rappaport (1995) and Berndt & Rappaport (2001). Observing that IT involves non—computer equipment, too, Lee & Barua (1999) have turned upside down the argument about how quality adjustment a?ects the measured volumes of IT. In their examination of the study by Loveman (1988), they argue that by applying a computer price index to all types of IT Loveman overestimated the volumes of IT investments, as computer prices have fallen faster than the prices of other IT products. While this criticism is probably foremost valid with respect to early de?nitions of IT that involved many items whose IT character could be questioned, the argument is supported by Jorgenson’s (2001) study of relative prices for di?erent kinds of IT equipment in the US since the late 1940s.
4
8
IFAU–Human capital is the key to the IT productivity paradox
to outlays on other factors of production, notably human capital; cf. points 3 and 4. There are thus rather diverse results on the connection between IT and growth, and the explanations for these ?ndings are quite diverse, too.
3
A stylized model
We here consider a stylized version of the model that we use in our empirical analysis. Our discussion serves two purposes. The ?rst is to reconcile the di?erent results of the earlier studies and to discriminate between some of the explanations that have been suggested for the IT productivity paradox. The second purpose is to consider how knowledge spillovers and capital-skill complementarity might a?ect productivity growth. Our stylized model captures four features: i) measurement error in the IT input variable(s), ii) mismeasurement of output, iii) positive externalities in the use of IT, and iv) the connection between IT and human capital. The analysis is consistent with both a neoclassical growth theory framework and with endogenous growth models. We can thus here disregard the fact that these two theoretical frameworks have di?erent implications for the empirical analysis, notably with respect to how IT and human capital are operationalized.6 Regarding feature i., it was noted in Section 2 that the IT measurement error can be both negative and positive. A simple speci?cation allowing for this is ITt? = ITt + wt
6
(1)
The empirical speci?cation of the model will be discussed in Section 4.2. 9
IFAU–Human capital is the key to the IT productivity paradox
where ITt? is the observed mesure of IT in period t, ITt the true measure and wt a random error, such that E (wt ) = 0, V ar (wt ) = ? 2 w, Cov (ITt , wt ) = 0. (2)
With respect to feature ii., non-recorded quality improvements in output should introduce a downward bias in measures of productivity growth (cf. point 5 in Section 2). Like the mismeasurement of IT, the mismeasurement of output is likely to vary over time, cf. Basu et al. (2003). We therefore specify the di?erence between the ?rm’s true rate of TFP growth, gt , and
? , as a random variable with positive expectation, ? , the observed rate, gt 0
according to
? = ? 0 + ut , gt ? gt
? 0 > 0,
(3)
and E (ut ) = 0,
Feature iii. can be modeled by assuming that the productivity e?ects from IT at the ?rm and industry level are a?ected by the use of IT in the aggregate economy; see the last paragraph of point 2, Section 2. Assuming that there is an index of the Total Use of IT in the Swedish Economy, T U IT E , we posit that T U IT E has the e?ect of scaling up the IT input. Using an increasing function, ? , and allowing for a delayed impact on the rate of growth we arrive at the following direct e?ect of IT on gt : ? 1t ·ITt?1 ; ? 1t = ? (T U IT Et?1 ) and ? 0 > 0. (5) The scaling e?ect can thus be expressed in terms of a time-varying parameter, ? 1t . Note that we do not assume that this parameter is positive, a priori.
10 IFAU–Human capital is the key to the IT productivity paradox
¡ ¢ 2 E u2 t = ?u ,
Cov (ut , wt ) = 0.
(4)
The motivation for (5) is that, by de?nition, an externality is an e?ect which is not accounted for by individual ?rms and, hence, shows up in TFP growth. In a neoclassical context, this would mean that the capital rental price of IT would overstate the real cost of IT capital.7 In an endogenous growth context, as in, e.g., Barrro and Sala—i—Martin (1999) it is natural to relate to a learning—by—investing mechanism; as successively more ?rms invest in IT, the knowledge about the properties of the new technology increases and becomes more widespread. With respect to feature iv., our analysis will be based on the maintained hypothesis that information technology and human capital are complements, in accordance with, e.g., Bresnahan et al. (2002) and Kaiser (2003). We model the complementarity by means of an interaction variable, taken to a?ect gt positively. Allowing, again, for a delayed impact we get an indirect e?ect of IT on gt : ? 2 · (IT × HC )t?1 ; ? 2 > 0. (6)
Ordinarily, interaction e?ects should be captured already in the measure of productivity growth.8 In the context of externalities in the use of IT and/or measurement error in the IT input, the interaction e?ect may not be properly accounted for, however. There may be knowledge spillovers arising through networks: employees working with computers form networks (via the Internet) with colleagues in other ?rms, networks which facilitate the transfer of knowledge.9
7 Siegel (1997) tries to capture IT externalities within a neoclassical framework. However, instead of considering the total use of IT in the economy he uses a measure of the IT investments made by the industry’s suppliers. 8 We are assuming here that the TFP growth measure corresponds to a ?exible representation of the technology, implying that it allows for interactions between inputs; see Section 4.1 9 One might wonder why we allow for both ?rst- and second-order e?ects of IT on productivity growth but only for a second-order e?ect of HC. The reason is that the
IFAU–Human capital is the key to the IT productivity paradox
11
Taking the total e?ect of IT on gt to be the sum of the direct e?ect (5) and the indirect e?ect (6) and using (3) we obtain the following equation:
? = ?? o + ? 1t ITt?1 + ? 2 (IT × HC )t?1 ? ut . gt
(7)
By (7), the e?ect of "true" IT on the observed rate of TFP growth equals
? ?gt = ? 1t + ? 2 HCt?1 . ?ITt?1
(8)
Note that although the e?ect of IT on productivity growth is increasing in human capital, the total e?ect can be negative, provided that ? 1t is negative and su?ciently large in magnitude. Before proceeding to analyse the implications of our simple model, a word of caution is in order. A causal interpretation, from IT and HC to
? , is justi?ed only if the one year lag on IT and HC makes it possible gt
to treat these variables as predetermined. This, in turn, hinges upon the absence of serial correlation in the data. This is an empirical matter that we consider in Section 5.2 Using the framework given by equations (1) — (8) we now discuss three issues that have arisen in connection with earlier studies: I. Can the negative e?ects of IT on productivity growth found in studies based on pre—1990 data be explained by measurement error in the IT variable as argued by Lee & Barua (1999), or are the results indicative of a truly negative return to early IT investments, as argued by Morrison (1997)? II. Why is it that models similar to the one just outlined yield positive
features i — iv above, do not involve mismeasurement in human capital and also not externalities in human capital per se. The externalities that we consider are associated with IT, either through IT investments or through the use of IT. However, from an empirical point of view there might nevertheless be a place for a ?rst-order e?ect of HC in the model. This point is discussed in Section 5.2.
returns when applied to later data?
12
IFAU–Human capital is the key to the IT productivity paradox
III. If complementarity between IT and skilled labor is allowed for, like in Bresnahan et al. (2002), what will happen to the estimated direct e?ect?
? is simply regressed on IT ? , using data for the Assume, ?rst, that gt t?1
pre—1990 period and post—1990 period, respectively. This implies that the measurement error in IT is ignored, that the variable (IT × HC )t?1 is omitted, and that no account is taken of the fact that ? 1t is a time—varying coe?cient. For illustrative purposes we will here assume that the function ? is a step function, taking on the values ? 1,pre-90 during the pre-1990 period ? 1,post-90 in the post-1990 period. To derive the probability limit of the OLS estimate of ? 1t under this conditions, we apply a result stated in Lam & Schoeni (1993).10 This yields ³ ´ b b plim ? 1,K = ? 1,K ? ? 1,K · ? + ? 2 ? (1 ? ?) , K = pre-90, post-90 (9)
where the IT measurement error is accounted for by the parameter ?, de?ned as ?? V ar (w) , V ar (IT ? ) (10)
and b ? is the coe?cient from a hypothetical regression of IT × HC on IT : Cov (IT × HC, IT ) b , ?= V ar (IT ) b ? > 0.
(11)
From (9) it can be seen that the bias in the estimate of ? 1,K has two components. The ?rst, ?? 1,K · ?, is the measurement error bias (MEB). The second component, due to omission of the variable IT × HC , is the omitted variable bias (OVB). While the OVB is invariably positive, given
In a returns to schooling context, Lam & Schoeni (op.cit.) consider how the estimated e?ect on earnings from another year of schooling is a?ected when data on ”ability” are lacking and there is measurement error in the schooling variable.
10
IFAU–Human capital is the key to the IT productivity paradox
13
the assumptions ? 2 > 0 and b ? > 0, the sign of the MEB is determined by the sign of the true parameter ? 1,K . If ? 1,K is positive the MEB will be negative, and if ? 1,K is negative, the MEB will be positive. Equation (9) can be used to derive bounds on the probability limit of b . These bounds are given in Table 1, for various the OLS estimate ? 1,K assumptions about the true parameter and the magnitude of the omitted variable bias. We can now consider issue I. As can be seen in Table 1, the estimated e?ect of IT on productivity growth can be negative only if the corresponding true e?ect is negative. In this case, c), the true e?ect is ´ ³ b negative and smaller than the lower bound of plim ? 1,pre-90 ; this is so because the omitted variable bias, ? b ? , is positive. Furthermore, this con2
clusion is una?ected by measurement error in the IT variable. The upper ´ ³ b bound of plim ? 1,pre-90 is equal to zero, irrespective of whether there is measurement error or not. Our analysis thus supports Morrison’s (1997) suggestion of overinvestment in IT during the latter part of the 1980’s, as overinvestment would, eventually, result in a negative e?ect of IT on productivity. And, as our conclusion is invariant to measurement error in the IT variable, we reject the claim in Lee & Barua (1999) that measurement errors were behind estimated negative e?ects of IT on productivity
growth.11
Actually, Lee & Barua state that ”.... the negative contribution of IT .... is attributable primarily to the choices of the IT de?ator and modeling technique.” However, they do not provide any assessment making it possible to disentangle the impacts of these two factors.
11
14
IFAU–Human capital is the key to the IT productivity paradox
Table 1: Ranges for the probability limit of the OLS estimator of ? 1,K ,for di?erent signs of the true e?ect and di?erent magnitudes of the omitted variable bias ´ ³ b ? ? +? b a) ? >0 =? 0?plim ? ?
1,K 1,K 1,K 2
b)
c)
? 1,K ¯? 1,K ¯
=? =?
Note: The index K denotes either pre-90 or post-90
? 1,K 1990
? 1,post-90 > ? 1,pre-90
(12)
It should be noted that (12) is not su?cient to determine the sign of ? 1,post-90 . If ? 1,pre-90 < 0 then ? 1,post-90 may be negative, too. Unfortub nately, the sign of the estimate ? 1,post-90 is no help here. In Table 1, we ´ ³ b see that plim ? 1,post-90 > 0 is consistent with both ? 1,post-90 > 0 and ? 1,post-90 < 0; cf cases a) and b), respectively. However, we can discriminclude a vector of proxy variables for the omitted variable, i.e. IT × HC . This will a?ect the estimate of ? 1,post-90 di?erently depending on the sign of the true parameter ? 1,post-90 . To show this, denote vector of proxy variIFAU–Human capital is the key to the IT productivity paradox 15
inate between the two cases by expanding the simple OLS regression to
b ables by P, and the corresponding estimate of ? 1,K by ? (1,K )·P . Then ´ ³ b = ? 1,K ? ? 1,K 1?R2 plim ? (1,K )·P
?
IT ? ×HC,P
(13)
2 2 ? where RIT ? ×HC,P denotes the R obtained when IT × HC is regressed
+ ? 2b ? (1 ? ?) · ? (IT ? , IT ? × HC, P)
on P, and ? (·) is a function that under fairly general conditions satis?es 0 < ? (·) < 1.12 Comparing (9) and (13) we note that ³ ´ ³ ´ b b ? 1,K > 0 =? plim ? ? < plim (1,K )·P 1,K . (14)
The implication (14) is due to the fact that the inclusion of proxy variables a?ects the measurement error bias (MEB) and the omitted variable bias (OVB) in the same direction when ? 1,K > 0. With respect to the MEB, ´ ³ 2 ?]0, 1[ implies that including proxies makes the fact that 1 ? RIT ? ×HC,P OVB, while positive, becomes smaller, too, because 0 < ? (·) < 1. On the other hand, if ? 1,K < 0 the e?ect of the proxy variables is ambiguous, the ambiguity being due to the fact that in this case the MEB and the OVB change in di?erent directions. Thus, by studying the e?ects of including proxy variables we should be able to infer the sign of the true parameter ? 1,post-90 . If ? 1,post-90 is indeed positive, then the estimate of ? 1,post-90 should be positive when human capital variables are excluded from the regression and this positive estimate should decrease towards zero when proxy variables for human capital are included.
Like (9), this equation draws on Lam & Schoeni (1993). They provide a similar expression to assess the e?ect on the estimated return to schooling when a proxy variable for the missing ability measure is included in the regression.
12
the MEB larger in magnitude, i.e. smaller because of the minus sign. The
16
IFAU–Human capital is the key to the IT productivity paradox
The analysis also provides the answer to issue III. It shows that the answer depends on the sign of the true direct e?ect. If the true direct e?ect is positive, allowing for indirect e?ects will decrease the estimated direct e?ect, cf.(14). If, on the other hand, the true direct e?ect is negative, allowing for indirect e?ects will have an ambiguous impact on the estimated direct e?ect.
4
Data and empirical speci?cation
Our empirical analysis covers 14 industries in the Swedish manufacturing sector, observed annually over the period 1986—95. The industry codes are given in Table 2. To indicate the relative size of the industries we also show their shares in manufacturing employment in the middle of the observation period. The data are Table 2: The industries considered and their shares in total manufacturing employment in 1991.
Industry code Industry Employment share 1991, % 9.4 3.0 8.5 14.7 7.9 3.3 4.0 11.5 13.5 8.1 12.3 2.2 0.8 0.8 100.0
3100 Food, Beverages and Tobacco 3200 Textile, Apparel & Leather 3300 Saw Mills and Wood Products 3400 Pulp, Paper and Printing & Publishing 3500 Chemical, Plastic Products. and Petroleum 3600 Non-Metallic Mineral Products 3700 Basic Metals 3810 Metal Products 3820 Machinery & Equipment, not elsewhere classi?ed 3830 Electrical Machinery, not elswhere classi?ed 3840 Transport Equipment, except Shipyards 3850 Instruments, Photographic & Optical Devices 3860 Shipyards 3900 Other Manufacturing 3000 Total Manufacturing Note: The classi?cation system used here is very close to the ISIC codes.
from the o?cial statistics produced by Statistics Sweden; from the National Accounts, the Employment Register, the Labor Force Surveys, varIFAU–Human capital is the key to the IT productivity paradox 17
ious Investment Surveys and the Trade Statistics. The cross-sectional dimension of the data set has been determined by the most detailed break—down of IT investments provided in the Investment Surveys. In the time series dimension, the starting point is given by the ?rst year of the Employment Register. The end point is the result of a change in the industrial classi?cation system, making it impossible to extend the time series beyond 1995.
4.1
The growth rate in total factor productivity
The yearly TFP growth rates have been computed by means of a Törnqvist index. This index corresponds to the translog production function and allows for interactions among inputs like, e.g., complementarity between IT and human capital.13 Suppressing industry indexes and denoting the volume of gross output by Y and the volume of input i by Xi , the TFP growth rate g , is de?ned as
gt ? ? ln T F Pt = ? ln Yt ? ? ln Xt
t = 1986, ...., 1995
(15)
where ? is the di?erence operator, de?ned such that ? ln Zt ? ln Zt ? ln Zt?1 . The growth in aggregate input, Xt , is given by: ? ln Xt =
8 X i=1
wi,t ? ln Xi,t ,.
(16)
where the weights wit are de?ned in terms of average cost shares according to wi,t
13
1 = 2
µ
Cf. Jorgenson et al. (1973) and Caves et al (1982).
P Pi,t Xi,t X Pn i,t?1 i,t?1 + Pn k=1 Pk,t?1 Xk,t?1 k=1 Pk,t Xk,t
¶
,
(17)
18
IFAU–Human capital is the key to the IT productivity paradox
and Pi is price of input i. We consider the following eight inputs, which will be discussed below, KC = Stock of computer equipment capital, KM = Stock of non-computer equipment capital, KS = Stock of structure capital, L1 = # of full-time employees with elementary school (less than 9 years), L2 = # of full-time employees with 9 year compulsory school, L3 = # of full-time employees with upper secondary school, L4 = # of full-time employees with tertiary and postgraduate education, IG = Intermediate goods. Figure 1 shows how the industry-weighted average of TFP growth has evolved over time. While the period 1986—90 showed low but stable
growth, the growth rates during 1991—95 were much higher and also more volatile. Also, Figure 2 shows that the variation around the average is smaller in 1991—95 than in 1986—90. Thus, the higher average growth in the ?rst half of the 1990s is not merely the result of high growth rates in some large industries.As noted in the introduction, the turning point apparently occurred quite early in Sweden. For instance, Stiroh (2002) estimates that the breakpoint in U.S. manufacturing was passed in 1993.
IFAU–Human capital is the key to the IT productivity paradox
19
Figure 1: Weighted averages of TFP growth rates in Swedish manufacturing 1986-1995. Industry weights equal to employment shares
0.04 0.03
0.02
0.01
0
86 87 88 89 90 91 92 93 94 95
-0.01
-0.02
Figure 2: The industry variation around the weighted average. All observations lie within the bounds given by the dashed lines.
0.15 0.1
0.05
0
86 87 88 89 90 91 92 93 94 95
-0.05
-0.1
-0.15
It can be argued, of course, that the increase in TFP growth in the latter half of the period is not only due to IT developments, but also to business cycle changes. We thus control for the business cycle in the empirical analysis, cf. Section 4.5.
4.2
Speci?cation of the explanatory variables
We consider three alternative speci?cations of the explanatory variables.
20 IFAU–Human capital is the key to the IT productivity paradox
The ?rst, due to neoclassical growth theory as originally formulated by Solow (1956), implies that the explanatory variables should be speci?ed in terms of growth rates. In a neoclassical context, the primary reason for explaining variations in TFP growth by means input growth rates is presence of input measurement error. While less natural, externalities can also be used as a motivation.14 The second framework is endogenous growth theory, which predicts that the levels of (some) inputs determine the rate of productivity growth. Endogenous growth theory explicitly deals with the rôle of externalities in explaining growth; see, e.g., Barro & Sala-i-Martin (1999). There are also endogenous growth models where growth is increased by devoting resources to R&D [Romer (1990) and Aghion & Howitt (1992)].Since resources devoted to R&D are essentially resources devoted to sophisticated capital equipment (IT) and highly educated workers, these models provide a motivation for the current study. Another argument can be derived from the literature on GPTs: successful implementation of a new GPT and the generation of skills needed to operate it e?ciently is a cumulative process. As such, it should be better captured by the developments of stocks (of IT and human capital) than by yearly ?ows, i.e. growth rates. The third framework is due to Jones’ (1995, 1999) critique of endogenous growth models. Jones (1995) argues that the claim that the level of R&D should determine the rate of growth is inconsistent with empirical data. He notes, however, that a simple way to avoid that increases in the levels of inputs can increase growth without limit is to substitute input proportions for input levels. For instance, if resources devoted to R&D can be approximated by "research labor" then, instead of having the number
14 A study framed in the neoclassical tradition which considers both measurement errors and externalities is Siegel (1997).
IFAU–Human capital is the key to the IT productivity paradox
21
of research workers determining the rate of growth, one could have the share of research workers in total employment. As there are no clear theoretical arguments for preferring one of these speci?cations in favor of the others, we have estimated models according to each one of them. Our general conclusions can be formulated as follows. Similar to the experience of Benhabib & Spiegel (1994), the neoclassical speci?cation with explanatory variables in growth rates yielded largely insigni?cant results. The level speci?cation of the original endogenous growth models to a larger extent resulted in signi?cant estimates but these were often implausible with respect to sign. The input proportions speci?cation yielded the best results in terms of signi?cance, signs and goodness-of-?t. We thus focus on this alternative.15
4.3
Measures of IT equipment and IT use
As our measure of IT, we use the share of computers in the total capital stock, KC /K . The computer capital stock has been constructed by means of data on computer investments collected through investment surveys conducted by Statistics Sweden. The computer investments cover investments made both for o?ce use and for use in the production process, e.g., CNC (computer numerically controlled) equipment and CAD / CAM — systems.16 For the manufacturing sector as a whole, computer investments for use in the production process were 3—4 times as large as those for o?ce use, during the period that we study. By means of the computer investments data we have broken down
However, results corresponding to the rates and levels speci?cations are avaiable on request. 16 The de?nition of IT investments employed here di?ers from de?nitions used in some recent U.S. studies. For example, Gordon (2000), Jorgenson & Stiroh (2000), and Oliner & Sichel (2000) de?ne IT investments as investments in hardware, software, and telecommunications.
15
22
IFAU–Human capital is the key to the IT productivity paradox
the industry-speci?c stocks of equipment capital provided in the National Accounts into computer capital stocks, KC , and stocks of non-computer equipment, KM . Details on the computation are provided in the Appendix. Table 3: Capital stock shares in Swedish manufacturing
Industry 3100 3200 3300 3400 3500 3600 3700 3810 3820 3830 3840 3850 3860 3900 3000 Computers 1985 1990 1994 2.8 5.5 7.8 3.5 6.6 6.9 3.0 17.2 12.6 9.2 13.8 14.1 4.0 7.0 12.1 2.0 6.1 6.7 2.2 9.9 10.8 8.8 18.0 15.6 13.4 17.8 21.0 16.1 16.2 32.7 19.7 21.0 36.2 23.6 15.7 21.0 1.9 3.1 7.2 2.1 5.0 6.5 7.9 13.4 17.3 Equipment 1985 1990 1994 48.6 48.8 48.7 60.7 56.4 49.0 47.1 33.2 39.1 56.0 54.2 53.4 61.4 60.4 55.5 50.8 50.5 49.9 56.6 50.6 51.8 44.8 41.0 44.1 33.5 42.0 40.5 41.7 48.5 32.2 30.0 36.0 25.2 39.7 56.4 49.5 42.3 34.9 30.2 37.6 38.9 35.2 49.2 47.8 44.9 Structures 1985 1990 1994 48.6 45.7 43.5 35.9 37.0 44.1 49.9 49.6 48.3 34.8 32.0 32.5 34.6 32.6 32.4 47.2 43.4 43.4 41.2 39.4 37.3 46.5 41.0 40.3 53.1 40.1 38.5 42.2 35.3 35.1 50.4 43.0 38.6 36.7 27.9 29.5 55.8 62.0 62.5 60.4 56.2 58.3 42.9 38.9 37.8
Table 3 shows the shares of computers, non-computer equipment and structures in the capital stock, for the beginning, middle and end of the period.17 In Table 3, we see that, for the manufacturing sector as a whole, the computer share in the capital stock more than doubled over the period 1985-94, from 7.9 percent to 17.3 percent. This is especially remarkable in view of the fact that computer capital depreciates much faster than other types of capital; we have assumed the rate of depreciation for computer capital to be 1/3. Table 3 also shows that in relative terms the largest increases in the computer shares took place between 1985 and 1990, rather than between 1990 and 1994. It can also be seen that there is a lot of variation across industries. This is important because the relatively short
17
The capital stocks for year t are de?ned as January 1.
IFAU–Human capital is the key to the IT productivity paradox
23
period covered by our data makes cross-sectional variation crucial in our empirical analysis. Figure 3: Index of total use of IT in Sweden, 1984=100
1000.0 900.0 800.0 700.0 600.0 500.0 400.0 300.0 200.0 100.0 0.0 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
To model the externalities associated with IT, we use an index of the Total Use of IT in the Swedish Economy, T UIT E , cf. (5). This index includes both computers & peripherals, and communication equipment. It is de?ned as
N N N + IM PIT,t ? EXPIT,t T U IT Et = P RODIT,t
(18)
N , IM P N , and EXP N denoting volumes of production, imP RODIT,t IT,t IT,t
ports and exports of IT at the national level. Figure 3 shows the evolution of T U IT E. It can be seen that the use of IT has increased extremely rapidly, especially from 1992 and onwards; between 1992 and 1995 the increase was threefold. Both KC /K and T U IT E are included in the regressions we with a one year lag, again to avoid endogeneity problems.
24 IFAU–Human capital is the key to the IT productivity paradox
4.4
The human capital data
The human capital variables have been constructed by means of the Swedish Employment Register and the Labor Force Surveys. The Employment Register contains employee information on industry, level of education and ?elds-of-study, age, sex, and immigrant status, and yearly earnings. The Labor Force Surveys provide data on work hours per week, by industry and sex, enabling an approximate conversion of number of employees into full-time equivalents.18 . Just like the use of capital, employment of labor is endogenously determined. In the empirical analysis, the human capital variables are thus also lagged one year, relative to productivity growth. Accordingly, the cross-classi?cations of labor for 1985, 1990 and 1994 in Table 4 are to be related to productivity growth rates in 1986, 1991 and 1995, respectively. The four cells in the upper left corner of the three sub-tables in Table 4 are identically zero, because the cross-classi?cation by ?elds-of-study is possible only for labor with at least upper secondary school. For the latter, quite detailed ?eld-of-study information is available, however. The labels ”engineering” and ”business administration” are used for brevity only; both encompass several sub?elds. The table shows that the human capital in the Swedish manufacturing sector changed dramatically during the period that we are studying. For instance, in 1985 almost half of the workers (49 percent) had no more than 9 years of schooling. In 1994, the share was 1/3. And, at the other end of the distribution, the share of workers with tertiary education almost doubled, from 9 to 16 percent. There is also considerable cross-section
18 The approximate nature of the conversion is due to the fact that the Labor Force Survey does not contain work hours by level of education.
IFAU–Human capital is the key to the IT productivity paradox
25
variation; in the empirical analysis we employ cross-classi?cations like Table 4 that di?er both by to industry and year. Table 4: Employment shares in Swedish manufacturing, by level of education and ?elds—of—study, 1985, 1990 and 1994. 1985:
Level of education < 9 years 9 years Upper secondary Tertiary P Level of education < 9 years 9 years Upper secondary Tertiary P Level of education < 9 years 9 years Upper secondary Tertiary P Engineering 0 0 0.25 0,06 0.31 Field-of-study Business administration 0 0 0.08 0,02 0.10
”other” 0.30 0.19 0.09 0.01 0.59
0.30 0.19 0.42 0.09 1
P
1990:
Field-of-study Engineering Business administration 0 0 0 0 0.29 0.09 0.08 0.03 0.37 0.12 Field-of-study Business administration 0 0 0.09 0.04 0.13
”other” 0.22 0.17 0.10 0.02 0.51
0.22 0.17 0.48 0.13 1
P
1994:
Engineering 0 0 0.31 0.10 0.41
”other” 0.18 0.16 0.11 0.02 0.47
0.18 0.16 0.51 0.16 1
P
In addition to levels of education and ?elds-of-study we also account for the workers’ age. The age structure can matter in two di?erent ways. On the hand, an education’s ”IT content” is higher the more recently the education was obtained, i.e. the younger the worker. This would point to a negative relation between age and productivity growth. On the other hand, older workers have accumulated more work experience than younger workers. If skills acquired in the workplace are more important for produc26 IFAU–Human capital is the key to the IT productivity paradox
tivity than computer skills acquired in school, then the relation between age and productivity growth should be positive instead. To empirically assess which of these two opposing forces that dominate the other we use the following variable # 16-29 year olds . # [(16-29) + (50-74)] year olds (19)
The idea underlying this variable is to capture e?ects of relative changes in tails of the age distribution; all employees in our data belong to the age interval 16-74 years.19 It should be noted that the ratio (19) can change even if the total number of 16-29 year olds plus the number of 50-74 year olds doesn’t change. Thus, e.g., substituting a given number of older workers with an equal number of younger worker will increase the ratio.20
4.5
Control variables
To account for cyclical variations in TFP growth, we have used a business cycle indicator, BCI , for the Swedish manufacturing sector, cf Figure 4. The indicator together data on orders, stocks of ?nished goods, and expected production.21
19 In terms of years, the right tail is longer than the left tail. However, the number of people working beyond the retirement age of 65 is very small. Hence, for practical purposes the tails can be considered to be equally long. 20 The fact that we model age structure e?ects by means of (19) should not be taken to mean that we deny the importance of changes in the share of 30-49 year olds for productivity growth; as shown by Malmberg (1994) workers aged 40-49 have made substantial positive contributions to growth in Sweden (along with 50-64 year olds) and Feyrer (2002) obtains similar results for a data set covering 108 di?erent countries. However, unlike these authors we are not primarily interested in the direct link between age demographics and productivity, but on e?ects working via interactions between workers of di?erent ages and IT. It is then natural to focus on the age categories that di?er the most in this respect, i.e. the youngest and the oldest workers. 21 The indicator has been constructed by the Swedish Institute for Economic Analysis (Konjunkturinstitutet).
IFAU–Human capital is the key to the IT productivity paradox
27
Comparing Figure 4 and Figure 1, we see that the BCI captures the turning points in TFP growth quite well. However, the BCI cannot explain the relative magnitudes of growth at di?erent points in time. In particular, it does not capture that TFP growth was much higher during 1991—95 than during 1986—90.22 Figure 4: The business cycle indicator (BCI ) for the Swedish manufacturing sector 1986-1995
20 10
0 1986 -10 1987 1988 1989 1990 1991 1992 1993 1994 1995
-20
-30
-40
To take into account that computer investments partly depend on other capital investments, we include the share of non-computer equipment in total capital, KM /K .23 As KC /K + KM /K + KS /K = 1 by de?nition, including KM /K together with KC /K means that we fully control for the industries’ capital structures. Finally, we include the shares of females and immigrants among the employees.Gender might be important for two reasons. Weinberg (2000) argues that computers create job openings for women by replacing physically demanding blue-collar jobs by jobs that require computer knowledge. Second, Lindbeck & Snower (2000) point out that modern work organizations are increasingly characterized by multi-tasking. If women are better
22 We do not want to use time dummies to control for the time variation that is common to all industries. Using time dummies amounts to eliminating the general time pro?le of the endogenous variable, i.e. the pro?le given in Figure 1. But that time pro?le is part of what we want to explain; one thing we want to test is whether our simple model can capture the change in the TFP growth pattern that occurred between the end of the 1980s and the beginning of the 1990s. 23 In this respect we follow earlier studies; see, e.g., Berndt and Morrison (1995).
-50
28
IFAU–Human capital is the key to the IT productivity paradox
suited to multi-tasking than men, as is often claimed, this should favor ?rms with a large female labor share. Regarding immigrants the direction of causality is more ambiguous. On the one hand, it can be conjectured that the increased international communication brought about by IT could be facilitated by a work-force comprising employees with di?erent cultural backgrounds. On the other hand, imperfect knowledge of the host country language might have an adverse e?ect on productivity.
5
Results
In the ?rst part of this section we test the empirical implications of the stylized model in Section 3, on our Swedish data. In the next subsection we consider various econometric issues. To focus on methodological aspects, the analysis is conducted within a modeling framework entailing a univariate representation of human capital. Based on our results in this section we decide upon a basic formulation of the model and an appropriate estimation method. In the last subsection we extend the basic model through multivariate speci?cations of human capital.24 Before discussing the results we will brie?y comment upon three features that are common to all the regressions. First, the estimations are based on weighted least squares (WLS), where the di?erent industries are weighted by their shares in manufacturing employment. Methodologically we thus follow, e.g., Berman, Bound, & Griliches (1994) and Kahn & Lim (1998). The motivation for the WLS procedure can be found in the latter paper: it is reasonable to assume the
While not ideal, this sequential approach is necessary due to the fact that our data set is rather small. Considering the issues of model formulation, estimation methods, and multivariate speci?cations of human capital simultaneously, we would simply run out of degrees of freedom.
24
IFAU–Human capital is the key to the IT productivity paradox
29
data for small industries to be noisier than the data for large industries. This assumption can be modeled by assuming that the standard errors of the (unweighted) residuals are inversely proportional to the square of employment. Weighting industries by employment shares will then make the residuals homoscedastic. Second, the following control variables are always included in the regressions: the (contemporaneous) business cycle indicator, BCI , the (lagged) share of non-computer equipment capital in the total capital stock, KM /K , and the shares of females and immigrants among the employees. Third, we do not explicitly account for possible measurement error in the IT variable, because we lack information on this issue.
5.1
Testing the implications of the stylized model
The ?rst point made in Section 3 was that the negative e?ects of IT on productivity growth reported in studies using early (pre—1990) data are not mere statistical artefacts. To see what can be said of the Swedish manufacturing sector in this respect, we estimate the following equation for the ?rst half of our study period: 1986-90:
? = ? 0.036 + controls ? 0.004 · ght (1.71) (0.08) KC K h,t?1 ,
R2 = 0.18 (20)
where absolute values of t—statistics are in parentheses.25 The e?ect of IT, i.e. the coe?cient of (KC /K )h,t?1 is negative. The theoretical analysis tells us that, although the estimate is insigni?cant, this indicates that IT had a negative impact on growth in Sweden, too, during the latter part of
25 To save space, we do not report the coe?cients for the control variables here, as they are of no interest with respect to theoretical implications that we consider.
30
IFAU–Human capital is the key to the IT productivity paradox
the 1980s. The intercept is negative as expected (although insigni?cant). According to the theoretical analysis, this means that the observed rate of
? , underestimates the true rate, g , by, on average, productivity growth, ght ht
3.6 percent; cf. (3). The second point made in Section 3 was that if the e?ect of IT on productivity growth turned positive in the 1990’s then we would expect, ?rst, a positive estimate of the impact of IT when ignoring human capital variables and, second, that this positive estimate should decrease after inclusion of human capital variables. The following regression shows that the ?rst condition is satis?ed: 1991-95: ght = ? 0.072 + controls + 0.204 ·
(1.83) (2.80) KC K h,t?1 ,
R2 = 0.51 (21)
The coe?cient for (KC /K )h,t?1 is now positive, and strongly signi?cant. It can also be noted that the intercept is still negative, as expected, and that it has increased in magnitude. This, too, is in line with expectations: one e?ect of the positive impact of IT will be quality improvements in output; to the extent that these are not captured in the data output growth and, hence, productivity growth will be (further) underestimated. To check the second condition we include the share of workers with tertiary education as a crude proxy for skilled labor. Interacting it with KC /K we obtain: 1991-95: ght = ? 0.067 + controls + 0.184 ·
(1.18) (1.02) KC K KC K h,t?1
+ 0.066 ·
(0.12)
³
# T ertiary # Employees
×
´
h,t?1
,
R2 = 0.52 (22)
The inclusion of the interaction variable decreases the estimated direct
IFAU–Human capital is the key to the IT productivity paradox 31
e?ect of IT from 0.204 to 0.184, i.e. the second condition is satis?ed, too. To summarize: these very simplistic regressions based on our stylized model point to a (small) negative e?ect on TFP in Swedish manufacturing during the second half of the 1980s and a positive e?ect after 1990. That is to say, they indicate a development qualitatively similar to the one experienced in the US, but with the turning point occurring somewhat earlier.
5.2
Econometric issues
In this section we will consider the following four issues: (1) the modeling of the time-varying e?ects of IT; cf. (5),(2) the potential presence of ?rst-order e?ects of human capital on TFP growth, in addition to the second-order interaction e?ect given by (6), (3) industry ?xed e?ects, and (4) serial correlation. Our starting point is the last speci?cation of the previous subsection, i.e. (22). We here estimate that model for the entire period of study, 1986-95, cf column I of Table 5.26 It can be seen that in contrast to the results obtained for the 1991-95 period the point estimate of the direct e?ect of IT on TFP growth is negative. Thus, when the impact is not allowed to vary over time, the positive e?ect during 1991-95 reported in (22) is dominated by a negative impact during 1986-90.27
In this section we also report the estimates obtained for the control variables. This is veri?ed when we apply the speci?cation used in (22) to data for 1986-90. This yields an estimate of the direct e?ect of IT that is equal to ?0.314 and signi?cant at the 1 % level.
27
26
32
IFAU–Human capital is the key to the IT productivity paradox
Table 5: Alternative model speci?cations, given univariate measure of human capital Dependent variable : ght
Intercept I -0.0239 (0.976) 0.0002 (2.046) 0.0880 (2.133) 0.0104 (0.325) -0.3010 (2.334) -0.0875 (1.009) 0.0002 (1.430) 0.0002 (1.441) 0.0349 (0.363) 1.0826 (3.276) No No 0.34 0.4961 (1.957) No No 0.34 0.3248 (0.606) No No 0.34 1.3426 (1.973) Yesa No 0.44 0.4225 (1.798) No Yesb 0.39 -0.0001 (0.483) 0.0002 (1.948) II -0.0515 (2.720) 0.0003 (2.569) 0.1179 (3.181) 0.0010 (0.313) -0.1868 (1.327) III -0.0471 (2.313) 0.0003 (2.526) 0.1072 (2.259) 0.0121 (0.371) -0.1972 (1.369) IV 0.0974 (0.894) 0.0002 (1.094) 0.1478 (1.998) -0.3340 (1.394) -0.5453 (1.581) V -0.0545 (3.207) 0.0003 (2.697) 0.1245 (3.644) 0.0078 (0.267) -0.1586 (1.225)
Control variables : BCIt ³ ´
KM K
³ ³
h,t?1
# F emales # Employees h,t?1 # Immigrants # Employees h,t?1 KC K
´
Direct ³ ´ e?ect of IT :
h,t?1
´
h i C T U IT E × ( K ) h K
#Tertiary #Employees h,t?1
t?1
Direct e?ect´of human capital : ³
h,t?1
IT interaction: ³ and human capital ´ #Tertiary. Kc #Employees × K
Industry dummies Correction for AR(1) residuals
R2
aThe reference industry is 3100 = Food, Beverages and Tobacco. b Iterative Parks (1967) procedure, second-round estimates.
Having thus established the need for a time-varying e?ect, we turn to the ?rst issue, the speci?cation of an explicit form for the function ? (T U IT E )t?1 . We have chosen to approximate ? by a linear function since our data only cover ten years, making it di?cult to precise estimate
IFAU–Human capital is the key to the IT productivity paradox 33
higher order approximations: ? (T U IT E )t?1 = ? · T U IT Et?1 ; ? > 0, (23)
where ? is a parameter and T U IT E the index described in Section 4.3.28 The e?ect of incorporating (??) can be assessed by comparing columns I and II in Table 5. It is clear that all the parameter estimates are a?ected. In particular, the point estimate of the direct e?ect of IT changes from ?0.0875 to 0.0002. And while the indirect e?ect decreases, the two changes do not cancel each other out; the partial derivative of gh,t with respect to (KC /K )h.t?1 [cf. (8)] increases in magnitude. As the time-varying speci?cation is in line with our theoretical model and does have an impact, we will stick to it in the following. The next issue concerns the possibility of direct, ?rst-order, e?ects of human capital on gh,t . While our theoretical analysis does not imply that human capital should have a direct e?ect on growth — cf. footnote 10 — there might still be empirical grounds for such a direct e?ect. To assess this possibility we compare columns II and III in Table 5, which di?er only by the inclusion of the human capital variable in column III. It can be seen that the direct e?ect of human capital is small and very imprecisely estimated. With respect to the other estimates, the only one a?ected is the coe?cient measuring the indirect, interaction, e?ect. That coe?cient becomes smaller and insigni?cant. Taken together, it appears that the
28
A disadvantage with the linear form is that it cannot allow the e?ect of IT on
TFP growth to change sign over time. As a result, the partial derivative (8) cannot be negative, under the assumptions made in Section 3. However, when we turn to a multivariate speci?cation of human capital, in Section 5.3, there is no reason to restrict all the IT and human capital interaction e?ects to be positive. The partial derivative of TFP growth with respect to IT might then change sign over time. It will be seen that this does indeed happen in our estimations.
34
IFAU–Human capital is the key to the IT productivity paradox
inclusion of a direct human capital e?ect has the clear disadvantage of creating multicollinearity problems but no discernible empirical advantage. Henceforth, we will therefore not consider direct e?ects of human capital. The third issue, allowing for industry ?xed e?ects amounts, in this context, to allow for cross-industry di?erences in the expected mismeasurement in output, cf. (3). While desirable, this generalization is quite costly in terms of degrees of freedom. Comparing columns II and IV in Table 5, we see that allowing for industry ?xed e?ects results in the estimate of the direct e?ect of IT becoming less signi?cant, both economically and statistically, while the economic signi?cance of the indirect e?ect is substantially increased. The ?xed e?ects themselves take on implausible values, however. For industry 3100 = Food, Beverages and Tobacco, which is the reference industry, the ?xed e?ect is given by the intercept. While insigni?cant, the estimate of the intercept says that the mismeasurement in output in industry 3100 is such that, on average, the (true) rate of productivity growth is overestimated by 9.7 percent. For the other industries, the ?xed e?ects are given by deviations from the reference level of 9.7 percent, determined by means of estimated coe?cients on industry dummies. These coe?cients imply that the estimated ?xed e?ects are positive for all the other industries as well.29 As we ?nd it really hard to believe that IT has resulted in TFP growth being overestimated in every industry we will disregard industry-speci?c ?xed e?ects from now on. The issue of serial correlation, ?nally, is important because the interpretation of the lagged explanatory variables as predetermined is valid
29 The coe?cients, which should be added to the intercept, are, by industry, 3200: 0.0693, 3300: -0.0684, 3400: -0.0526?? , 3500: -0.0262, 3600: -0.0901? , 3700: -0.0766, 3810: -0.0600, 3820: -0.0879, 3830: -0.0211, 3840: -0.0838, 3850: -0.0841? , 3860: 0.0934, 3900: 0.0193, where * and ** denote signi?cantly di?erent from zero at the 10 and 5 % level, respectively.
IFAU–Human capital is the key to the IT productivity paradox
35
only if the regression residuals ful?ll the assumption of being random disturbances and, hence, not correlated over time. As our panel only covers a ten-year period, formal tests for autocorrelation will, unfortunately, have very low power. Nevertheless, it is possible to estimate the parameters of a simple autoregressive structure. To this end we apply an iterated version of the procedure suggested by Parks (1967) to correct for ?rst-order autocorrelation in a multiple-equation context. The assumed autocorrelation structure is given by: uh,t = ?h uh,t?1 + eh,t , |?| < 1 (24)
where the eh,t are white noise disturbances. Note that the autocorrelation parameter, ?, is allowed to vary across industries. We apply this structure to the model given by column II in Table 5. The ?rst-round estimates of the ?h are obtained by application of (24) to the estimated residuals of the column II speci?cation. All 14 estimates ful?ll the requirement that |?| < 1. As judged from the t-statistics, only one estimate is signifantly di?erent from zero, at the 10 % level. Still, the ?rst-round estimates, denoted by b ?1h , are used to estimate the model: where
? (b yh,t ?1h ) ? (b ?1h ) yh,t
? (b ?1h ) = x? ?1h ) + u? ?1h yh,t h,t (? , b h,t b
1
(25)
= (1 ? b ?1h ) 2 yh,t ,
1 2
for t = 1986 for t = 1987, ..., 1995 (26) for t = 1986 for t = 1987, ..., 1995
the 1986 variables being constructed according to the Prais-Winsten transformation. The resulting ?-estimates were qualitatively similar to the ones
36 IFAU–Human capital is the key to the IT productivity paradox
x? ?1h ) = xh,t ? b ?1h · xh,t?1 , h,t (? , b
?1h ) = (1 ? b ?1h ) x? x? h,t (? , b h,t (? ) ,
= yh,t ? b ?1h · yh,t?1 ,
in column II of Table 5 with small di?erences in magnitude and signi?cance. By means of the u? ?1h ), second-round estimates b ?2h were obtained. h,t (b
Two of these estimates were signi?cantly di?erent from zero at the 10 % level, thus indicating no improvement with respect to autocorrelation, as compared to the original speci?cation (where only one of the estimated autocorrelation parameters was signi?cantly di?erent from zero at the 10 % level). The estimate of the vector ? obtained from the regression model transformed by means of the b ?2h was extremely close to the original ?
estimate; compare columns V and II in Table 5. From the table it can be seen that the t-statistics are very close, too. But again, there was no discernable improvement with respect to the residuals; of the b ?3h estimates
one was signi?cant, at the 5 % level. Upon further iterations, the initial pattern was repeated: the estimates of the structural parameters shifted back and forth between one alternative similar to the original column II speci?cation and one alternative extremely close to this speci?cation. In no case was there any improvement with respect to the serial correlation of the residuals, as compared to the column II speci?cation. Thus, there is no strong indication that the residuals of the model in column II of Table 5 are autocorrelated and application of a standard procedure to correct for possible autocorrelation has no e?ect on the parameter estimates and and seems to make the residuals less well-behaved. Based on the results of this section we conclude that, in line with the theoretical arguments in Section 3, it seems important to allow the e?ects of IT to vary over time. We do not ?nd that our modeling framework needs to be extended to account for the other three issues that we have considered — potential ?rst-order e?ects of human capital on TFP growth,
IFAU–Human capital is the key to the IT productivity paradox 37
industry ?xed e?ects, and serial correlation. Using speci?cation II in Table 5 as our starting point we now proceed to consider more detailed, multivariate speci?cations of human capital.
5.3
Multivariate speci?cations of human capital
Apart from indicating the need for relative measures (cf. Section 4.2) theory does not provide any guidance regarding the implementation of a more detailed speci?cation of human capital. We have constructed variables such that the model can tell the e?ects of marginal changes in the educational structure. The e?ect that we are interested in is given by the partial derivative of total factor productivity growth with respect to this measure: X ?ght b = ?i · Xi ? (KC /K )h,t?1
i=1 m
(27)
ciated human capital variable. The variance of this partial derivative is equal to " #
where b ?j denotes an estimated coe?cient and Xj represents an assom X i=1 m m X ´ ³ ´ ³ X b · V ar ?i + 2 Xi Xj Cov b ? ib ?j i=1 j>i
?ght V ar ? (KC /K )h,t?1
=
Xi2
(28)
As the variance computation is a bit complicated we will, to begin with, merely consider the individual terms in (27), implying that we only have to consider the corresponding t — ratios.
38
IFAU–Human capital is the key to the IT productivity paradox
Table 6: Growth regressions allowing for externalities in the use of IT Dependent variable : ght I II III
Intercept -0.0273 (1.132) 0.0002 (2.082) 0.0586 (1.426) 0.0201 (0.592) 0.0159 (0.110) -0.2440 (0.952) 0.0002 (1.902) 0.0545 (1.168) -0.0021 (0.051) 0.0440 (0.205) -0.0225 (1.226) 0.0002 (2.000) 0.0547 (1.753)
Control variables : BCIt ³ ´
KM K
³ ³
h,t?1
# F emales # Employees h,t?1 # Immigrants # Employees h,t?1
´
Direct e?ect of IT : h i C T U IT E × ( K ) h K
´
t?1
0.00006 (0.505) 0.6460 (2.061)
0.00001 (0.096)
Direct e?ect of human capital ´: ³ #Tertiary KC #(Upper sec.+Tertiary ) × ( K ) h,t?1 h i #Tertiary Engineers Kc × K #(Upper sec.+Tertiary Engineers) h h h h
#Tertiary Business adm. #(Upper sec.+Tertiary Bus. adm.) #Tertiary ”Other” #(Upper sec.+Tertiary ”Other”) #Upper sec. #(9 years+Upper sec.)
×
Kc K h,t?1
×
×
Kc K h,t?1
i
Kc K h,t?1
i
i
h,t?1
0.8497 (3.413) -0.8646 (2.039) 0.9498 (1.198) 0.4240 (1.812) -0.8122 (3.464) 0.403 0.9104 (2.996) -1.2996 (3.393) 0.437
0.8779 (5.289) -0.8324 (2.383) 0.8779 (5.289) 0.8779 (5.289) -1.2593 (5.877) 0.437
#16-29 year olds #(16-29+50-74 year olds)
×
Kc K h,t?1
i
R2
Table 6 reports the results of three di?erent speci?cations. In column I we have allowed for the possiblity that, in addition to tertiary educated
IFAU–Human capital is the key to the IT productivity paradox 39
workers, employees with upper secondary education also belong to the ?rm’s skilled workers. The number of employees with tertiary education has been related to the number of employees with upper secondary or tertiary education. Similarly, the number of upper secondary educated workers has been normalized by the number of workers with 9 years of education or upper secondary education. We also use the variable (19) to account for the age structure aspect of human capital. Clearly, accounting for upper secondary education and the age structure are important extensions. The corresponding parameter estimates are strongly signi?cant. Interestingly, the indirect e?ect of IT associated with the age structure is negative. This implies that the negative e?ect of lost work experience caused by old workers retiring outweighs the positive e?ect of the entry of young workers with high ”IT content” in their basic education. Comparing column I of Table 6 with column II of Table 5 we see that the more detailed modeling of human capital renders the estimated direct e?ect of IT smaller and that among the control variables only the business cycle indicator stays signi?cant. The next step is to disaggregate the measures of human capital even further, by ?elds of study; cf. column II of Table 6. We ?nd considerable di?erences across ?elds. In particular, while there is a positive indirect e?ect of IT associated with the relation between engineers with university education and engineers with upper seconday education there is a negative indirect e?ect connected with the corresponding categories in business administration. While the this di?erence is somewhat counter-intuitive, there are results in the literature that point in this direction. For example, Murphy et al. (1991) claim that while "entrepreneurs" a?ect growth pos40 IFAU–Human capital is the key to the IT productivity paradox
itively "rent-seekers" are harmful to growth. Proxying entrepreneurs and rent-seekers with engineers and lawyers, respectively, they ?nd empirical support for their claim. As our category Business administrators includes lawyers, this ?nding is relevant for our results. Further, Mellander and Skedinger (1999) show that in the mid 1990s wage premia for university education were much higher among business administrators than engineers in seven European countries, including Sweden, in spite of an engineering degree requiring more years of study. A possible interpretation is that the university wage premium for business administrators is ”too high”, relative to their contribution to productivity. The see if the regression model in column II can be expressed in a more parsimonious way, we test the following composite hypothesis: h ³ ´ i C (i) The coe?cients of T U IT E × K K are zero. (ii) The coe?cients of h equal . h h i #Females , # Employees × h
# Immigrants # Employees h,t?1
h t?1
h,t?1
and
i
# Tertiary Engineers # (Upper sec. + Tertiary Engineers )
# Tertiary ”Other” # (Upper sec. + Tertiary ”Other” )
×
KC K h,t?1 , h i # Upper sec. KC K h,t?1 and # (9 years + Upper sec.)
i
×
KC K h,t?1 are
i
With respect to hypothesis ii) it should be emphasized that equality among the coe?cients does not imply that the associated indirect e?ects of IT on productivity growth are equal. If the coe?cients are equal, the corresponding indirect e?ects will be determined by the relative magnitudes of the human capital variables. Among these, the ratio is invariably the largest. As indicated by the fact that there is no di?erence between the R2 s in columns II and III, the composite hypothesis cannot be rejected at any
IFAU–Human capital is the key to the IT productivity paradox 41
# Upper sec. # (9 years + Upper sec.)
standard level of signi?cance. We thus end up with a model containing only six parameters, which explains 44 percent of the variation in total factor productivity growth across industries and over time! What, then, are the relative magnitudes of the indirect e?ects in our preferred speci?cation, i.e. column III in Table 6? For the manufacturing sector as a whole this question can be answered by means (5.8) and Table 4. The largest positive indirect e?ect is the one associated with the ratio
# Upper sec. # (9 years + Upper sec.) ;
for a marginal increase in the share of com-
puters in total capital the e?ect varies between 0.60 percentage points in 1986 and 0.67 percentage points in 1995. The largest negative indirect e?ect, which is the one channeled through the age structure, i.e. the ratio
# 16 — 29 year olds # (16 - 29 + 50 - 74 year olds) ,
decreases in magnitude over time, from -0.68
percentage points in 1986 to 0.60 percentage points in 1995.30 The next to largest positive indirect e?ect is associated with the relation between university educated engineers and engineers with upper secondary education, the ratio
# Tertiary Engineers # (Upper sec. + Tertiary Engineers ) ;
the indirect
e?ect increases from 0.17 percentage points in 1986 to 0.21 percentage points in 1995. This e?ect is however o?set by the negative indirect e?ect connected to business administrators, which decreases from -0.17 percentage points in 1986 to -0.26 percentage points in 1995. Finally, a positive indirect e?ect stemming from the relation between employees with "other" university and upper secondary education, respectively, makes upp the balance: this positive e?ect increases from 0.09 percentage points in 1986 to 0.14 percentage points in 1995. While these results for the entire manufacturing sector provide a gen30 To save space, the age structure data have not been provided in Section 4.4. However, for the years 1985 and 1994 the age structure ratio is equal to 0.536 and 0.479, respectively, re?ecting a declining in?ow of young people and ageing of the incumbents.
42
IFAU–Human capital is the key to the IT productivity paradox
eral feeling for the time pro?le of the e?ect of IT on total factor productivity growth, an important feature of the model is that it allows the e?ect of marginal increases in computers’ share to vary over time and by industries. This is illustrated in Figures 5a—c, which are based on computations using speci?cation III in Table 6. The diagrams show the distributions of the partial derivatives (5.8) across industries at three points in time, 1986, 1991 and 1995. The estimates’ precision have been computed according to (5.9).The estimates can be interpreted as answering the following question: If the share of computers in total capital increases by 1 percent, what is the resulting change in the rate of growth in total factor productivity, in percentage points? The bars indicate the e?ects for individual industries. The solid line is a weighted average e?ect, where the industries are weighted by their employment shares. Looking at the development over time, we see that the marginal e?ects of computer investments have increased steadily over time. The weighted average e?ect rises from about 0.01 percentage point in 1986 to 0.05 in 1991, ending up at 0.17 percentage points in 1995. These average changes have been caused by upward shifts in the entire distributions of e?ects across industries. For instance, while only two industries record e?ects above
1 10
of a percentage point in 1986, e?ects of this magnitude are found
in six industries in 1991 and in 11 in 1995. In the latter year, the point estimates are 0.25 or higher in ?ve industries, indicating that a 1 percent increase in computers’ share in total capital increases the rate of TFP growth by
1 4
of a percentage point or more.
IFAU–Human capital is the key to the IT productivity paradox
43
Figure 5: Distributions over industries of the e?ects of a marginal increase in computers´ share of capital on TFP growth; regression III in Table 6, evaluated in 1986, 1991 and 1995.
a:1986
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth. Weighted mean value
-0.10 -0.15
b:1991
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth Weighted mean value
-0.10 -0.15
c:1995
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth Weighted mean value
-0.10 -0.15
Note: Stars indicate signi?cance level: ”*” denoting 10 percent, ”**” 5 percent and ”***” 1 percent.
Among the three years covered by Figure 5a—c, the largest variation across industries is found in 1986. In that year the spread is 0.46 percentage points, the range being given by a negative e?ect of ?0.12 percentage points in 3840 = Transportation and a positive e?ect of 0.34 percentage
44 IFAU–Human capital is the key to the IT productivity paradox
31 0 38 0 40 * 33 * 00 38 ** 10 * 34 ** 00 * 36 ** 00 ** 38 * 30 * 38 ** 20 * 39 ** 00 ** 37 * 00 * 32 ** 00 * 35 ** 00 * 38 ** 50 ** 38 * 60 ** *
Industries
-0.05
31 00 38 40 36 00 38 10 33 00 34 00 38 2 39 0 00 38 ** 30 * 35 ** 00 ** 37 * 00 * 38 ** 60 * 32 ** 00 ** 38 * 50 ** *
Industries
-0.05
0* * 38 10 31 00 33 00 38 20 34 00 36 00 37 00 39 00 * 38 5 32 0 00 ** 38 30 35 * 00 38 ** 60 ** *
Industries
-0.05
38 4
points in 3860 = Shipyards.31 In 1991 and 1995 the spread is considerably smaller — about 0.30 percentage points in both years. Moreover, in 1995 the e?ects are positive in all industries. There are thus two ?ndings pointing to a fundamental di?erence between the beginning and the end of the period that we study: compared to 1986 the variation across industries is smaller in 1995 and the estimated e?ects are con?ned entirely to the positive domain, unlike 1986 when about a third were negative. In line with our basic hypothesis of the importance of human capital, a comparison of Figure 5 and Table 3 shows that the industries that had the largest increases in the shares of computers in total capital were not in general the industries that had the largest growth-enhancing e?ects of IT. For instance, the industries 3300 = Saw Mills and Wood Products and 3700 = Basic Metals increased the relative size of their computer capital stock dramatically between 1985 and 1990; cf Table 3. These investments did not result in top-ranking marginal e?ects of IT in either 1991 or 1995, however; see Figure 5. Conversely, industry 3850 = Instruments, Photographic & Optical Devices experienced very large IT-induced growth e?ects in 1991 and 1995. In this industry the share of computers decreased between 1985 and 1990 — cf Table 3. Instead, the share of skilled workers increased strongly in this industry.32 Finally, a notable result is that, compared to the U.S., we ?nd positive impacts of IT on growth in a broader spectrum of industries. According to
31 The shipyards rank very high in 1991 and 1995, too. Since the Swedish shipyards have undergone major structural changes since the mid 70’s and have been facing severe problems with low and, sometimes, negative pro?ts this industry could be seen as a potential outlier. To check this, we reestimated the model given by column III in Table 6, leaving out the shipyards. The parameters changes were entirely negligible, however. The reason is the WLS estimation procedure where the industries are weighted by employment; the shipyards account for less than 1 percent of manufacturing employment, during the period studied. 32 The latter fact cannot be inferred from the paper but can be seen when the Table 4 is broken down by industry.
IFAU–Human capital is the key to the IT productivity paradox
45
Gordon (2000), in the U.S. the e?ects of computer investments were essentialy zero outside the IT-producing industries and the industries producing durable manufacturing goods. In the Swedish manufacturing sector, these industries roughly correspond to: 3810, 3820, 3830, 3840, 3850, and 3860; see Table 2. From Figure 5 it can be seen that while we ?nd large marginal e?ects in some of these industries, notably in 3850 = Instruments and 3860 = Shipyards, we also see examples of negative or very small effects as in, e.g., in 3810 = Metals and 3840 = Transportation. On the other hand, there are several industries outside this group recording large positive e?ects like 3200 = Textiles and 3500 = Chemicals.33 Table 7: Statistics for non-nested tests of the presence of Kc /K in growth equation; critical value at 1% signi?cance level ±2.57 Model I II Ho : include Kc /K -0.329 -0.686 Ha : exclude Kc /K Ho : exclude Kc /K 3.193 4.112 Ha : include Kc /K
Note: i) the model speci?cations refer to the columns in Table 6 ii) ”include Kc /K” refers to the regressions in Table 6 while ”exclude Kc /K” means setting Kc /K=1 in those regressions iii) the test statistic is asymptotically normally distributed.
However, while our results certainly seem to indicate that the human capital variables are essential, one might wonder about the importance of the computer capital share, Kc /K . Is this variable really essential, too, or can the human capital variables do the job by themselves? This is an important question because our interpration of human capital being the key to the IT productivity paradox relies on the assumption that it
33 Using more recent U.S. data than Gordon (op.cit) and dummy variable techniques, Stiroh (2002) ?nds indications of substantial e?ects of IT after 1995 not only in industries producing IT and durable goods, but also in IT-intensive industries, de?ned as industries having above median shares of computers in total capital. He does not link these ?ndings to human capital structures, however.
46
IFAU–Human capital is the key to the IT productivity paradox
is the interaction between Kc /K and human capital that matters. To check if this is the case it is necessary to conduct a non-nested test of whether Kc /K should be included in the growth equations or not. To this end we use the J test proposed by Davidson and MacKinnon (1981). The results of applying this test to the speci?cations I and II in Table 6 are given in Table 7. Note that the results concern the testing of two hypotheses. An intrinsic feature of a non-nested test is that there is no natural null hypothesis. Being a speci?cation test, the non-nested test merely investigates how two alternative models ?t the data. In the ?rst row of Table 7 we provide the test statistics for the case when the speci?cations in Table 6 constitute the null hypotheses. The alternative, Ha , corresponds to when Kc /K = 1 in the regressions. In none of the tests can the null be rejected at any standard level of signi?cance. In the second row, the roles of the null hypothesis and the alternative hypothesis have been reversed. The null is very clearly rejected in favor of the alternative. These results provide strong evidence for the model speci?cations in Table 6 and reject the alternative speci?cations where KC /K = 1. Put di?erently, the outcomes give convincing support for the notion that it is the interaction between IT capital and human capital that drives our results. This conclusion is further strengthened by the fact that it is quite unusual that non-nested tests yield results as clear as in this case; often the tests produce inconsistent results (reject both of the null hypotheses) or inconclusive results (reject neither).34
34 The reason why we have not performed the test on speci?cation III in Table 6 is that the Davidson-MacKinnon test cannot be applied to models incorporating linear constraints. Pesaran and Hall (1998) discuss non-nested tests allowing for general linear restrictions. However, given the very clear outcomes of the tests reported in Table 7 and the fact that, statistically, the speci?cations II and III in Table 6 are very close we have not taken the trouble to formulate such a general test.
IFAU–Human capital is the key to the IT productivity paradox
47
6
Summary and conclusions
Our principal conclusion from this study is that human capital is the key to the IT productivity paradox. We substantiate this general conclusion with both theoretical and empirical results. Our theoretical analysis investigates the consequences of erroneously disregarding human capital aspects in assessments of the e?ects of IT on productivity growth. Speci?cally, we consider a model where IT a?ects growth both directly and indirectly, through complementarity with human capital, and analyze what happens to the estimate of the direct e?ect when the indirect e?ect is omitted. Regarding the negative e?ects of IT on growth reported in several studies using early (pre—1990) U.S. data, our conclusion is that these results are likely to indicate a truly negative e?ect, as suggested by Morrison (1997), rather than be a consequence of measurement error, as argued by, e.g., Lee and Barua (1999). The positive relation between IT and productivity growth found in studies based on more recent data is in our theoretical analysis attributed to positive external e?ects in the use of IT. These external e?ects are assumed to be increasing in the total use of IT, implying that as more and more IT capital is accumulated, the growth e?ects change from negative to positive. In the empirical analysis, we ?rst con?rm that the predictions generated in the theoretical analysis are valid for our data on the Swedish manufacturing sector. We then proceed to include successively more information about interactions between IT and human capital. As shown by the theoretical analysis, accounting for indirect e?ects of IT in this way reduces the estimated direct e?ect. Eventually, the direct e?ect ?nally
48 IFAU–Human capital is the key to the IT productivity paradox
vanishes altogether. We end up with a model that is very parsimonious in terms of parameters but, nevertheless, explains well over 40 percent of the variation in total factor productivity growth across industries and over time. In this model, all the interaction variables between IT and human capital are highly signi?cant. In general, the maintained hypothesis of complementarity between IT and skilled workers is con?rmed. The largest indirect e?ects of IT on growth are associated with workers having upper secondary education, relative to workers with only 9 years of education. Disaggregating by ?elds of study, we ?nd the next to largest e?ect to be associated with the relation between university educated engineers compared to engineers with upper secondary education. An exception to the complementarity relation between IT and skilled labor concerns workers within the ?eld of business administration and law. For these, the relation between university educated and workers with upper secondary education gives rise to a negative indirect impact on productivity growth. In the spirit of Murphy et al. (1991), we interpret the negative estimate as indicating rent-seeking behavior among business administrators and lawyers. Regarding the connection between human capital and the age structure we ?nd that replacing workers aged 50 or older by workers below 30 has a negative impact on productivity growth rates. This indicates that, during the period studied, the advantage of many of the younger workers of having become acquainted with IT during their school years did not outweigh the work experience acquired by the older workers. This negative indirect e?ect is quite large but decreasing, due to a declining in?ow of young
IFAU–Human capital is the key to the IT productivity paradox 49
people to the manufacturing sector. For the manufacturing sector as a whole, the model predicts that in the beginning of the period, in 1986, a 1 percent increase in the share of computers in total capital increased productivity growth by 0.01 percentage points only, i.e. an entirely negligible e?ect. In the middle of the period, in 1991, this average e?ect had grown to 0.05 percentage points, while at the end of the period, in 1995, it was up to 0.17 percentage points. The variation in e?ects across industries decreases over time. Moreover, while the e?ects of IT on growth are negative in several industries in 1986, the e?ects are positive in all industries in 1995. In ?ve of them the estimated e?ect was 0.25 or higher, saying that a 1 percent increase in computers’ capital share increased productivity growth by at least percentage point. To check that our results are not driven solely by human capital developments but by complementarity between IT and human capital, we perform non-tested tests for the presence of the IT variable in the growth equations. These tests provide very strong support for the complementarity hypothesis. In line with our basic hypothesis, we ?nd that the industries were the (relative) increases in computer capital have been particularly large are not necessarily the industries that show the largest marginal e?ects of IT on productivity growth. With respect to di?erences in e?ects across industries, we also relate our ?ndings to the claim in Gordon (2000) that IT has increased productivity growth only in a small number of U.S. industries. We show that, unlike in the U.S., the Swedish IT development has had positive e?ects outside the sectors producing IT and durable manufacturing goods. We
50 IFAU–Human capital is the key to the IT productivity paradox
1 4
of a
?nd strongly positive e?ects also in, e.g., the chemical industry and, even more interesting, in the textile industry. Regarding policy considerations, one conclusions is immediate: measures to promote increased use of IT should be followed up by measures promoting skill upgrading, especially from elementary to upper secondary education. Another implication is that measures aimed at facilitating early retirement among older workers, in order to make more room for young labor market entrants, can be (strongly) harmful for growth. It should be remembered, however, that our study is based on data ending quite a few years back. Our results on the age structure might have changed during recent years. Investigating whether this is the case is an important task for future research. Also, it should be noted that our ?ndings concern only the manufacturing sector and cannot be extended to the service sector or the economy as a whole. While analyses of the service and the entire economy lie beyond the scope of the present paper because of data limitations, we believe that such analyses are important tasks for future research.
IFAU–Human capital is the key to the IT productivity paradox
51
References
[1] Aghion, P. & P. Howitt (1992) ”A Model of Growth through Creative Destruction”, Econometrica, Vol. 60, pp. 323—351. [2] Autor, D.H., L.F. Katz, & A.B. Kreuger (1998) ”Computing Inequality: Have Computers Changed the Labor Market?”, Quarterly Journal of Economics, Vol. CXIII, pp. 1169-1213. [3] Barro, R.J. & X. Sala—i—Martin (1999) Economic Growth, MIT Press, Cambridge, Massachusetts, U.S.A. [4] Basu, S., J.G. Fernald, N. Oulton, & S. Srinivasan (2003) ”The Case of the Missing Productivity Growth: Or, Does Information Technology Explain Why Productivity Accelerated in the United States but not the United Kingdom?” NBER Macroeconomics Annual 2003 [5] Benhabib, J. & M.M. Spiegel (1994) ”The Role of Human Capital in Economic Development: Evidence from Aggregate Cross-Country Data”, Journal of Monetary Economics, Vol. 34, pp. 143-173. [6] Berman, E., J. Bound, & Z. Griliches (1994) ”Changes in the Demand for Skilled Labor within U.S. Manufacturing: Evidence from the Annual Survey of Manufactures”, Quarterly Journal of Economics, Vol. CIX, pp. 367—398. [7] Berndt, E.R., Z. Griliches & N.J. Rappaport (1995) ”Econometric Estimates of Price Index for Personal Computers”, Journal of Econometrics, Vol. 68, pp. 243-268. [8] Berndt, E.R. & C.J. Morrison (1995) ”High—tech capital formation and Economic Performance in U.S. manufacturing industries: An exploratory analysis”, Journal of Econometrics, Vol. 65, pp. 9—43. [9] Berndt, E.R. & N.J. Rappaport (2001) "Price and Quality of Desktop and Mobile Personal Computers: A Quarter-Century Historical Overview", American Economic Review, Papers and Proceedings, Vol. 91, pp. 268—273. [10] Bresnahan, T.F., E. Brynjolfsson, & L.M. Hitt (2002) ”Information Technology, Workplace Organization, and the Demand for Skilled
52 IFAU–Human capital is the key to the IT productivity paradox
Labor: Firm—Level Evidence”, Quarterly Journal of Economics, Vol. CXVII, pp. 339—376. [11] Bresnahan, T.F. & M. Trajtenberg (1995) ”General Purpose Technologies: ’Engines of Growth’ ?”, Journal of Econometrics, Vol. 65, pp. 83-108. [12] Brynjolfsson, E. (1993) ”Information Technology and the Productivity Paradox: Review and Assessment”, Communications of the ACM, Vol. 35, pp. 66—77. [13] Caves, D.W., L.R. Christensen, & W. E. Diewert (1982) ”The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity”, Econometrica, Vol. 50, pp. 1393—1414. [14] David, P. A. (1990) ”The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, American Economic Review, Papers and Proceedings, Vol. 80, pp. 355—361. [15] Davidson, R. & J. MacKinnon (1981) ”Several Tests for Model Speci?cation in the Presence of Alternative Hypotheses”, Econometrica, Vol. 49, pp. 781-793. [16] Dean, E.R. (1999) ”The Accuracy of the BLS Productivity Measures”, Monthly Labor Review, pp. 24-34. [17] Feyrer, J.D. (2002) ”Demographics and Productivity”, mimeo, Dept. of Economics, Dartmouth College. [18] Gordon, R.J. (2000) ”Does the ’New Economy’ Measure up to the Great Inventions of the Past?”, Journal of Economic Perspectives, Vol. 14, pp. 49-74. [19] Greenwood, J. & M. Yorukoglu (1997) ”1974”, Carniege-Rochester Series on Public Policy, Vol. 46, pp. 49-95. [20] Griliches, Z. (1969) ”Capital-Skill Complementarity”, Review of Economics and Statistics, Vol. LI, pp. 465—468. [21] Harris, S. E.& J. L. Katz (1991) ”Organizational Performance and Information Technology Investment Intensity in the Insurance Industry”, Organizational Science, Vol. 2, pp. 263—296.
IFAU–Human capital is the key to the IT productivity paradox 53
[22] Helpman, E. & M. Trajtenberg (1998) ”A Time to Sow and a Time to Reap: Growth Based on General Purpose Technologies”, in E. Helpman (ed.): General Purpose Technologies, MIT Press. [23] Jones, C.I. (1995) ”R&D-based Models of Economic Growth” Journal of Political Economy, Vol. 103, pp. 759—784. [24] Jones, C.I. (1999) ”Growth: With or Without Scale E?ects” American Economic Review, Papers and Proceedings, Vol. 89, pp. 139—144. [25] Jorgenson, D.W. (2001) ”Information Technology and the US Economy”, American Economic Review, Vol. 91, pp. 1—32. [26] Jorgenson, D.W., L.R. Christensen, & L.J. Lau (1973) Transcendental Logarithmic Production Frontiers”, Review of Economics and Statistics, Vol. 55, pp. 28—45. [27] Jorgenson, D.W. & K.J. Stiroh (2000) ”Raising the Speed Limit: U.S. Economic Growth in the Information Age”, Brookings Papers on Economic Activity, pp. 125—211. [28] Kahn, J.A. & J.—S. Lim (1998) ”Skilled Labor—Augmenting Technical Progress in U.S. Manufacturing”, Quarterly Journal of Economics, Vol. CXIII, pp. 1281—1308. [29] Kaiser, U.(2003) ”Strategic Complementarities Between Di?erent Types of ICT-Expenditures”, ZEW Discussion Paper No 03-46. [30] Lam, D. & R. F. Schoeni (1993) ”E?ects of Family Background on Earnings and Returns to Schooling: Evidence from Brazil”, Journal of Political Economy, Vol. 101, pp. 710—740. [31] Lee, B. & A. Barua (1999) ”An Integrated Assessment of Productivity and E?ciency Impacts of Information Technology Investments: Old Data, New Analysis and Evidence”, Journal of Productivity Analysis, Vol. 12, pp. 21—43. [32] Lindbeck, A. & D.J. Snower (2000) ”Multi-Task Learning and the Reorganization of Work: From Tayloristic to Holistic Organization”, Journal of Labor Economics, Vol. 18, pp. 353—376.
54 IFAU–Human capital is the key to the IT productivity paradox
[33] Loveman, G.W. (1988) ”An Assessment of the Productivity Impact of Information Technologies”, in T.J. Allen and M.S.S. Morton (eds): Information Technology and the Corporation of the 1990s: Research Studies, Cambridge, MA: MIT Press. [34] Lucas, R. (1990) ”Why Doesn´t Capital Flow from Rich to Poor Countries?”, American Economic Review, Vol. 80, pp. 92-96. [35] Malmberg, B. (1994) ”Age Structure E?ects on Economic Growth — Swedish Evidence”, Scandinavian Economic History Review, Vol. XLII, pp. 279-295. [36] Mellander, E. (1999) ”The Multi—Dimensional Nature of Technical Change and Skill—Biased Technical Change”, IUI Working Paper No 518. [37] Mellander, E. & P. Skedinger (1999) ”Corporate Job Ladders in Europe: Wage Premia for University vs High School—Level Jobs”, Swedish Economic Policy Review, Vol. 6, pp. 449—487. [38] Morrison, C.J. (1997) “Assessing the Productivity of Information Technology Equipment in U.S.Manufacturing Industries”, Review of Economics and Statistics, Vol. LXXIX, No. 3, pp. 471—481. [39] Murphy, K.M., A. Schleifer, & R. W. Vishny (1991) ”The Allocation of Talent: Implications for Growth”, Quarterly Journal of Economics, Vol. CVI, pp. 503-530. [40] Parks, R.W. (1967) ”E?cient Estimation of a System of Regression Equations when Disturbances are Both Serially and Contemporaneously Correlated”, Journal of the American Statistical Association, Vol. 62, pp. 500-509. [41] Oliner, S.D. & D. E. Sichel (2000) ”The Resurgence of Growth in the Late 1990s: Is Information Technology the Story?”, Journal of Economic Perspectives, Vol. 14, pp. 3-22. [42] Parsons, D., C.C. Gotlieb, & M. Denny (1993) ”Productivity and Computers in Canadian Banking”,Journal of Productivity Analysis, Vol. 4 pp. 95—113.
IFAU–Human capital is the key to the IT productivity paradox 55
[43] Pesaran, M. H. & A.D. Hall (1998) ”Tests of Non-Nested Linear Regression Models Subject to Linear Restrictions”, Economics Letters, Vol. 27, pp. 341-348. [44] Romer, P.M. (1990) ”Endogenous Technical Change”, Journal of Political Economy, Vol. 98, pp. S71-S102. [45] Siegel, D. (1997) ”The Impact of Computers on Manufacturing Productivity Growth: A Multiple—Indicators, Multiple—Cause Approach”, Review of Economics and Statistics, Vol. LXXIX, No. 1, pp. 69—78. [46] Solow, R.L. (1956) ”A Contribution to the Theory of Growth”, Quarterly Journal of Economics, Vol. LXX, pp. 65—94. [47] Solow, R.L. (1987) ”We’d Better Watch Out”, New York Times Book Review, July 12, p. 36. [48] Stiroh, K.J. (2002) ”Information Technology and the US Productivity Revival: What Do the Industry Data Say?”, American Economic Review, Vol. 92, pp. 1559-1576. [49] Triplett, J.E. (1999) ”The Solow Productivity Paradox: What Do Computers Do to Productivity?”, Canadian Journal of Economics, Vol. 32, pp. 309—334. [50] Weinberg, B.A. (2000) ”Computer Use and the Demand for Female Workers”, Industrial and Labor Relations Review, Vol. 53, pp. 290— 308.
56
IFAU–Human capital is the key to the IT productivity paradox
A
Computation of omputer capital
The Swedish National Accounts (SNA) provides data on capital stocks of equipment and structures (buildings) by 2- or 3-digit industries. In this section we show how the equipment capital stock can be decomposed into two parts, one computer capital stock and one stock för non-computer equipment. To this end, we ?rst have to to consider the computation of the SNA capital stocks and and the corresponding capital rental prices. To simplify the notation, we here suppress industry indices and denote the equipment stocks by KE,t and the stocks of structures by KB,t .The stocks are de?ned such that the period t stock denotes the stock as of January 1, year t. The perpetual inventory method used in the SNA to compute the stocks implies that they can be closely approximated by the following accumulation formula ¢ ¡ i = E, B . (29) Ki,t = 1 ? ? i Ki,t?1 + Ii,t?1 , The capital rental prices for equipment and structure capital are constructed according to " !# ¡ ¢e à ¡ ¢e PIi ,t|t?1 ? PIi ,t?1 PIi ,t|t?1 ? PKi ,t = PIi ,t?1 rt?1 + ? i (30) PIi ,t?1 PIi ,t?1
where PKi ,t denotes the rental price for type i captal at the beginning of period t, PIi ,t?1 is the gross investment price index for type i capital and is a long-term interest rate measured at the very end of period t ? 1, rt?1 ¡ ¢e period t ? 1, and PIi ,t|t?1 is the expected value of the investment price index for type i capital in period t, given information this index ¡ about¢e up to (and including) period t ? 1. The di?erence PIi ,t|t?1 ? PIi ,t?1 measures the expected windfall pro?t (loss) that accrues to the owner of the capital asset through an increase (decrease) in the renewal cost.35 Like the ? i , the PIi are obtained from the SNA. The interest rate r is measured by means of the nominal ¢e Swedish long-term industrial ¡ rate on bonds. The expectional variable PIi ,t|t?1 is implemented by means of
35 The rental price formula (30) corresponds to the one given by equation (B4) in Jorgenson & Stiroh (2000). The only di?erence being that Jorgenson and Stiroh (op.cit.) assume perfect foresight with respect to the investment price index, thus substituting ¢e ¡ PIi ,t for PIi ,t|t?1 .
IFAU–Human capital is the key to the IT productivity paradox
57
a univariate Kalman ?lter.36 The rental prices are normalized to unity in a base-year to — here set to 1991 — yielding: eK ,t = PKi ,t . (31) P i PKi ,to To preserve the property that price × quantity = cost, the quantity of capital is normalized accordingly, i.e. e i,t = PK ,to Ki,t K i (32)
eK ,t K e i,t = PK ,t Ki,t . such that P i i To obtain the computer capital stock, we split the equipment stock KE into KEC and KEM where subindex C denotes Computers and subindex M stands for machines (that are not computers). In analogy with (29): KEC ,t = (1 ? ? EC ) KEC ,t?1 + IEC ,t?1 (33)
To make (33) operational, we have to decide on a value for ? EC and on an initial value for KEC . We have set ? EC = 1 3 . One motivation is that in the SNA depreciation rates for equipment (including computers) varies between 0.16 and 0.21. As computer capital depreciates much faster than other types of equipment ? EC should considerably larger than 0.21, making 1 3 a rather reasonable number. It is also close to the depreciation rate of 0.315 (from the Bureau of Economic Analysis) employed by Jorgenson & Stiroh (op.cit.). The initial value for KEC is obtained by extrapolating gross investments, IEC , backwards. To this end, we have assumed that investments during the period 1980-1994 can be approximated by the arithmetic average of the 1985 and 1986 gross investments. For the computation of the TFP growth rate according to Section 5.1, we also need a capital rental price for computer capital. The computation of this rental price is very similar to (30). For the gross investment price index PIEC ,t we use an import price for computers and peripherals, normalized to unity in 1991. Unfortunately, this index can only be computed for 1984-1995. During this period the index shows a continous decrease
36 This ?lter amounts to modeling the price index by means of a transition equation and a measurement equation. The former models the "true" investment price index as a random walk, incorporating a drift in the form of a deterministic quadratic time trend. The measurement equation models the observed price index as the sum of the "true" index and a random error.
58
IFAU–Human capital is the key to the IT productivity paradox
in the price of computers and peripherals, at an increasing rate. Between 1984 and 1985 the rate of decrease was very small, only 0.1 %, while between 1994 and 1995 the index fell by 14.3 %. The arithmetic mean of the rates of price decreases over the period was around 6.5 %.37 As our time series on PIEC ,t is so short we cannot model the expected investment price index by means of a Kalman ?lter. Instead we have simply ?tted a linear trend to the log-di?erences of the index, to estimate the average rate of decrease in the yearly price reductions, i.e. the discrete analogue of the second order derivative. We obtain an estimate of -1.24 percent annually, implying that for computer capital the last term within brackets in (30).is equal to zero in 1985 and the falls cumulatively by -1.24 each year, to reach -12.4 percent in 1995. Given the stock of computer capital and the computer capital rental price we can consistently solve for the expenditures on (non-computer) machinery equipment. Denoting these expenditures by VKEM ,t we get ³ ´ eK ,t K eK ,t K eK ,t K e K ,t = P e E,t ? P e E ,t VKEM ,t ? P EM EM E EC C (34)
because rental expenditures on computers and non-computer machinery have to add up to total rental expenditures on equipment capital. e K ,t . To solve for P eK ,t , eK ,t and K The ?nal step is determine P EM EM EM we ?rst assume that the rental price of equipment capital can be approxeK ,t : eK ,t and P imated by a translog aggregate of P E E eK ,t = ? ln P E
1 2
+
eK ,t (St?1 + St ) · ? ln P EC
1 2
where St =
eK ,t [(1 ? St?1 ) + (1 ? St )] · ? ln P EM eK P EC eK ,t K e E ,t P EC C e ,t KE ,t + VK
C
(35)
.
(36)
EM ,t
37 This may seem like a rather small rate of price decrease. It is smaller than similar estimates for the US but the di?erence is not as large as one might think. For comparison, Jorgenson and Stiroh (2000) report an average rate of decrease in the price of computer investments equal to 12.8 percent over the period 1985-1995. For communications investment they ?nd a much smaller rate of decrease, namely 0.6 percent over the same period. Thus, the decline in prices di?ers substanntially between di?erent types of computerrelated equipment. In our case, it might be that the prices of peripherals have fallen not fallan as fast as the prices of computers. Unfortunately, we cannot check this conjecture, as there is no separate price index for computers.
IFAU–Human capital is the key to the IT productivity paradox
59
eK ,t , we obtain Solving for ? ln P EM eK ,t = ? ln P EM ?
1
1 (St?1 +St ) 2
1 [(1?St?1 )+(1?St )] 2
1 [(1?St?1 )+(1?St )] 2
eK ,t · ? ln P E
(37)
eK ,t but not its The equation (37) determines the rate of change in P EM eK ,t , level. However, the level is determined by the normalization that P EM eK ,t and P eK ,t , should be equal to unity in the base-year. just like P E EC Thus, eK ,to ? 1.0 . (38) P EM e K ,t according to eK ,t we can ?nally solve for K Given P EM EM e K ,t = K EM VKEM ,t , eK ,t P
EM
eK ,t . · ? ln P EC
(39)
which constitutes the ?nal step in the break-down of the equipment capital stock into computer capital and (non-computer) machinery capital.
60
IFAU–Human capital is the key to the IT productivity paradox
Publication series published by the Institute for Labour Market Policy Evaluation (IFAU) – latest issues
Rapporter/Reports 2004:1 Björklund Anders, Per-Anders Edin, Peter Fredriksson & Alan Krueger “Education, equality, and efficiency – An analysis of Swedish school reforms during the 1990s” Lindell Mats “Erfarenheter av utbildningsreformen Kvalificerad yrkesutbildning: ett arbetsmarknadsperspektiv” Eriksson Stefan & Jonas Lagerström ”Väljer företag bort arbetslösa jobbsökande? Forslund Anders, Daniela Fröberg & Linus Lindqvist ”The Swedish activity guarantee” Franzén Elsie C & Lennart Johansson “Föreställningar om praktik som åtgärd för invandrares integration och socialisation i arbetslivet” Lindqvist Linus ”Deltagare och arbetsgivare i friårsförsöket” Larsson Laura ”Samspel mellan arbetslöshets- och sjukförsäkringen” Ericson Thomas ”Personalutbildning: en teoretisk och empirisk översikt” Calmfors Lars & Katarina Richardson ”Marknadskrafterna och lönebildningen i landsting och regioner”
2004:2 2004:3 2004:4 2004:5 2004:6 2004:7 2004:8 2004:9
2004:10 Dahlberg Matz & Eva Mörk ”Kommunanställda byråkraters dubbla roll” 2004:11 Mellander Erik, Gudmundur Gunnarsson & Eleni Savvidou ”Effekter av IT i svensk industri” Working Papers 2004:1 Frölich Markus, Michael Lechner & Heidi Steiger “Statistically assisted programme selection – International experiences and potential benefits for Switzerland” Eriksson Stefan & Jonas Lagerström “Competition between employed and unemployed job applicants: Swedish evidence” Forslund Anders & Thomas Lindh “Decentralisation of bargaining and manufacturing employment: Sweden 1970–96” Kolm Ann-Sofie & Birthe Larsen “Does tax evasion affect unemployment and educational choice? Schröder Lena “The role of youth programmes in the transition from school to work”
2004:2 2004:3 2004:4 2004:5
2004:6 2004:7 2004:8 2004:9
Nilsson Anna “Income inequality and crime: The case of Sweden” Larsson Laura & Oskar Nordström Skans “Early indication of program performance: The case of a Swedish temporary employment program” Larsson Laura “Harmonizing unemployment and sickness insurance: Why (not)?” Cantoni Eva & Xavier de Luna “Non-parametric adjustment for covariates when estimating a treatment effect”
2004:10 Johansson Per & Mårten Palme “Moral hazard and sickness insurance: Empirical evidence from a sickness insurance reform in Sweden” 2004:11 Dahlberg Matz & Eva Mörk “Public employment and the double role of bureaucrats” 2004:12 van den Berg Gerard J, Maarten Lindeboom & Peter J Dolton “Survey nonresponse and unemployment duration” 2004:13 Gunnarsson Gudmundur, Erik Mellander & Eleni Savvidou “Human capital is the key to the IT productivity paradox” Dissertation Series 2002:1 2002:2 2002:3 2002:4 2003:1 2003:2 Larsson Laura “Evaluating social programs: active labor market policies and social insurance” Nordström Skans Oskar “Labour market effects of working time reductions and demographic changes” Sianesi Barbara “Essays on the evaluation of social programmes and educational qualifications” Eriksson Stefan “The persistence of unemployment: Does competition between employed and unemployed job applicants matter?” Andersson Fredrik “Causes and labor market consequences of producer heterogeneity” Ekström Erika “Essays on inequality and education”
doc_665828832.pdf
A paradox is an argument that produces an inconsistency, typically within logic or common sense.
Human capital is the key to the IT productivity paradox
Gudmundur Gunnarsson Erik Mellander Eleni Savvidou
WORKING PAPER 2004:13
The Institute for Labour Market Policy Evaluation (IFAU) is a research institute under the Swedish Ministry of Industry, Employment and Communications, situated in Uppsala. IFAU’s objective is to promote, support and carry out: evaluations of the effects of labour market policies, studies of the functioning of the labour market and evaluations of the labour market effects of measures within the educational system. Besides research, IFAU also works on: spreading knowledge about the activities of the institute through publications, seminars, courses, workshops and conferences; creating a library of Swedish evaluational studies; influencing the collection of data and making data easily available to researchers all over the country. IFAU also provides funding for research projects within its areas of interest. There are two fixed dates for applications every year: April 1 and November 1. Since the researchers at IFAU are mainly economists, researchers from other disciplines are encouraged to apply for funding. IFAU is run by a Director-General. The authority has a traditional board, consisting of a chairman, the Director-General and eight other members. The tasks of the board are, among other things, to make decisions about external grants and give its views on the activities at IFAU. A reference group including representatives for employers and employees as well as the ministries and authorities concerned is also connected to the institute. Postal address: P.O. Box 513, 751 20 Uppsala Visiting address: Kyrkogårdsgatan 6, Uppsala Phone: +46 18 471 70 70 Fax: +46 18 471 70 71 [email protected] www.ifau.se
Papers published in the Working Paper Series should, according to the IFAU policy, have been discussed at seminars held at IFAU and at least one other academic forum, and have been read by one external and one internal referee. They need not, however, have undergone the standard scrutiny for publication in a scientific journal. The purpose of the Working Paper Series is to provide a factual basis for public policy and the public policy discussion.
ISSN 1651-1166
Human capital is the key to the IT productivity paradox?
Gudmundur Gunnarsson† Erik Mellander‡ Eleni Savvidou§ October 4, 2004
Abstract Unlike previous analyses, we consider (i) possible externalities in the use of IT and ii) IT and human capital interactions. Examining, hypothetically, the statistical consequences of erroneously disregarding (i) and (ii) we shed light on the small or negative growth e?ects found in early studies of the e?ects of IT on productivity growth, as well as the positive impacts reported more recently. Our empirical analysis uses a 14-industry panel for Swedish manufacturing 1986-95. We ?nd that human capital developments made the average e?ect of IT essentially zero in 1986 and steadily increasing thereafter, and, also, generated large di?erences in growth e?ects across industries. JEL codes : O33, L23, L60 Keywords : IT Productivity Paradox, Applied Econometrics
? We thank Per-Anders Edin, Karolina Ekholm, Nils Gottfries, Bertil Holmlund, Matthew Lindquist, Mikael Lundholm, Eva Mörk, Sten Nyberg, Hans Wijkander and two anonymous referees for constructive comments Anders Hintze and Michael Wolf at Statistics Sweden generously helped us with data. Financial support from the Swedish Council for Work Life Research, the Swedish Transport & Communications Research Board and the Swedish Agency for Innovation Systems is gratefully acknowledged. † Dept. of Business Administration and Information Systems, Mälardalen University, P.O. Box 883, SE — 721 23 Västerås, Sweden; email: [email protected] ‡ Institute for Labour Market Policy Evaluation (IFAU), P.O. Box 513, SE-751 20 Uppsala, Sweden; email: [email protected] § Dept. of Economics, Uppsala University, P.O. Box 513, SE-751 20 Uppsala, Sweden; email: [email protected]
IFAU–Human capital is the key to the IT productivity paradox
1
Contents
1 Introduction 2 Literature review: attempts to explain the paradox 3 A stylized model 4 Data and empirical speci?cation 4.1 The growth rate in total factor productivity 4.2 Speci?cation of the explanatory variables . 4.3 Measures of IT equipment and IT use . . . 4.4 The human capital data . . . . . . . . . . . 4.5 Control variables . . . . . . . . . . . . . . . 3 5 9 17 18 20 22 25 27
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
5 Results 29 5.1 Testing the implications of the stylized model . . . . . . . . 30 5.2 Econometric issues . . . . . . . . . . . . . . . . . . . . . . . 32 5.3 Multivariate speci?cations of human capital . . . . . . . . . 38 6 Summary and conclusions A Computation of omputer capital 48 57
2
IFAU–Human capital is the key to the IT productivity paradox
1
Introduction
The IT productivity paradox was formulated in response to the fact that the massive investments in information technology (IT) that started around 1980 did not seem to have any positive e?ects on productivity growth. In the words of Nobel laureate Robert Solow: ”You can see the computer age everywhere but in the productivity statistics.” [Solow (1987)] In recent years, the original focus on computers has been broadened to include also communication devices: the concept of IT has been extended to ICT, information and communication technology. In this paper, we account for the development of communications equipment. We have kept the term IT, however. In empirical studies, the IT productivity paradox has been veri?ed in analyses based on early (pre—1990) data for the U.S. and Canada. Mostly, the results show either very small or insigni?cant e?ects of IT on productivity growth; see for instance Harris & Katz (1991) and Parsons, Gotlieb, & Denny (1993). Indeed, some studies have reported signi?cantly negative e?ects; cf. Loveman (1988) and Berndt & Morrison (1995). Some of the explanations suggested for these counter-intuitive results are: the time required for IT investments to yield productivity increases has been underestimated, the magnitude of the investments have been overestimated and measurement problems on both the input side and the output side have concealed the productivity e?ects. However, a couple of more recent studies, using data extending to the end of the 1990’s, have found productivity—increasing e?ects of IT. Oliner & Sichel (2000) argue that the reason why there were no e?ects earlier is that, in the U.S., IT investments did not really take o? until 1995. When they did, the e?ects were substantial, however: Oliner & Sichel claim that
IFAU–Human capital is the key to the IT productivity paradox 3
IT accounted for about two—thirds of the acceleration in the labor productivity between the ?rst and second halves of the 1990’s. Bresnahan, Brynjolfsson, & Hitt (2002), while focusing primarily on skillbiased technical change rather than productivity, make an important contribution towards the resolution of the IT productivity paradox by extending the idea of capital-skill complementarity hypothesis discussed by Griliches (1969) and Lucas (1990). Bresnahan et al. (op.cit.) argue that too much attention has been paid to IT investments and too little attention has been paid to work organization and human capital structure. Accounting for both IT and human capital, they ?nd that the balance between the two is crucial. Firms with high levels of both IT and human capital are found to be the most productive. More interesting: ?rms with low levels of both IT and human capital are shown to be more productive than ?rms that are high on IT and low on human capital, or vice versa. The framework we suggest in this paper is similar to the Bresnahan et al. (op.cit.) approach in the sense that we, too, conjecture that human capital is a key element in the explanation of the IT productivity paradox. However, we extend the analysis by incorporating a phenomenon often discussed in the context of endogenous growth theory, namely knowledge spillovers. While it seems very natural to consider knowledge spillovers in an evaluation of the productivity e?ects of IT, these have barely been discussed in earlier studies. The next section contains a review of some attempts to explain the IT productivity paradox. In Section 3 we develop a simple stylized growth model. By means of this model we discriminate between some of the suggested explanations for the IT productivity and, second, propose a way to account for knowledge spillovers.
4 IFAU–Human capital is the key to the IT productivity paradox
Our empirical analysis is based on data for 14 industries in the Swedish manufacturing sector observed annually during the period 1986—95. It appears that in the Swedish manufacturing sector the productivity-enhancing e?ects of IT started to show already in the ?rst half of the 1990s, i.e. a couple of years earlier than, e.g., in the U.S. Otherwise, the developments in Sweden seems to have been qualitatively similar to that in several other countries. Our data are described in Section 4 and the results are provided in Section 5. Section 6 contains a summary of our results and our conclusions.
2
Literature review: attempts to explain the paradox
For brevity, we here only provide a very condensed and selective list of some the explanations suggested for the IT productivity paradox.1 1. Investments in IT became massive only towards the end of the 1990s. Thus, early analyses were unable to capture positive growth e?ects from IT simply because, at the time, these investments were still comparatively small. Studies using later data should be able to discern positive growth e?ects. This view is supported by the study by Oliner & Sichel (2000). However, this explanation says nothing about the signi?cant negative e?ects of IT on productivity estimated by, e.g., Loveman (1988) and Berndt & Morrison (1995). 2. It takes time before the productivity-enhancing e?ects of a new technology can be realized. This point has perhaps been most convinc1 For a more extensive discussion see, e.g., Triplett (1999). Also, for the view that there is essentially no paradox to explain, because the importance of the introduction of IT has been vastly exaggerated, compared to the signi?cance of other technological developments like the adoption of electricity, see Gordon (2000).
IFAU–Human capital is the key to the IT productivity paradox
5
ingly made by David (1990). From an empirical point of view, this explanation is similar to the previous one. An important di?erence, however, is that this explanation can account for (initial) negative e?ects of IT on productivity, provided that the di?usion of IT use is associated with learning costs that decrease over time, as a function of the increasing number of users. This explanation also points to the importance of (positive) externalities. More wide-spread knowledge about (how to exploit) IT will speed up the rate of di?usion. The resulting increase in people with access to IT will raise the bene?ts accruing to individual users, which will further accelerate di?usion. The importance of this spiralling e?ect has been especially notable in the 1990’s, with the rapidly expanding use of email and the Internet. 3. No account has been taken of the complementarity between IT and skilled workers. Although the capital-skill complementarity hypothesis was put forward already by Griliches (1969), the connection between IT and human capital has almost invariably been disregarded in assessments of the productivity e?ects of IT.2 Presumably, this is primarily due to lack of data. However, by matching two di?erent data sets Bresnahan, Brynjolfsson, & Hitt (2002) have overcome this problem. Splitting their data into four categories according to whether ?rms are ”high” or ”low” on IT and human capital, they ?nd high levels of productivity in ?rms that are either high on both IT and human capital or low in both of these dimensions. Relatively lower levels of productivity are found in ?rms that are high in one
2 However, complementarity between IT and skilled workers has been documented in several studies of labor demand and skill-biased technical change. Two seminal contributions are Berman, Bound, & Griliches (1994) and Autor, Katz, & Kreuger (1998). For a study using Swedish data, see Mellander (1999).
6
IFAU–Human capital is the key to the IT productivity paradox
of the two dimensions and low in the other.3 Using a di?erent approach, Kaiser (2003) also ?nds strong evidence for complementarity between expenditures on IT capital and outlays for IT personnel. 4. IT is a general purpose technology (GPT), the e?cient implementation of which requires changes in work practices and skill upgrading. This explanation contains elements of explanations 2 and 3. The idea is that the introduction of GPTs like IT will initially lead to a slowdown in productivity, as it takes time to implement and learn to use the GPT e?ciently. In particular, assuming skilled labor to have a learning advantage over unskilled labor, the theory holds that skill premia will rise, inducing an increased supply of skills. When the increased supply comes about and the work organization is properly adapted to the GPT, productivity starts increasing again. The notion of GPTs was introduced by Bresnahan & Trajtenberg (1995) and the relation between GPTs and productivity growth is discussed in, e.g., Helpman & Trajtenberg (1998), and Greenwood & Yorukoglu (1997). 5. Mismeasurement of outputs. According to this explanation, the use of information technology has increased the quality of existing products and services and created new goods, neither of which are (fully) captured in the o?cial statistics. This has led to a downward bias in the estimated growth e?ects; see, e.g., Brynjolfsson (1993) and Dean (1999). Nevertheless, it is essential to point out, like Lee & Barua (1999) do, that e?ciency related gains in the production of
A related approach is taken by Siegel (1997), who considers the possibility that the investments in IT may induce enhanced e?ciency of labor which, in turn, positively a?ects productivity growth. He ?nds some, although not unambiguous, support for this hypothesis.
3
IFAU–Human capital is the key to the IT productivity paradox
7
the ”old” goods should still be accounted for by conventional output measures. That is to say, while mismeasurement of output certainly is part of the puzzle it cannot resolve it entirely. 6. Mismeasurement of inputs. On the input side the issue of mismeasurement is less clear-cut than on the output side. On the one hand, it can be argued that early (U.S.) measures of IT were overstated because they included equipment that one would not ordinarily associate with IT like, e.g., typewriters and accounting machinery.4 On the other hand, the often noted di?culties to adjust for quality increases in IT price indexes implies a tendency to underestimate the volumes of IT investments.5 And the presence of positive externalities in the use of IT, cf. the second point above, points in the same direction. Failure to account for these externalities will, again, bias measures of IT inputs downwards. 7. Overinvestments in IT, in the latter half of the 1980s. This explanation has been suggested by Morrison (1997), based on the ?nding that in U.S. manufacturing industries estimated bene?t—cost ratios (Tobin’s q ) for IT capital dropped signi?cantly below 1 by the mid 1980’s. It is natural to interpret the term ”overinvestment” in a relative sense here, i.e. that IT investments were too large compared
These were included in Bureau of Economic Analysis category ”O?ce Computing and Accounting Machinery; cf Berndt & Morrison (1995). After 1982 this category was replaced by ”Information Processing and Related Equipment”, see Lee & Barua (1999). 5 For a hedonic approach to the estimation of price indexes for computers, see Berndt, Griliches & Rappaport (1995) and Berndt & Rappaport (2001). Observing that IT involves non—computer equipment, too, Lee & Barua (1999) have turned upside down the argument about how quality adjustment a?ects the measured volumes of IT. In their examination of the study by Loveman (1988), they argue that by applying a computer price index to all types of IT Loveman overestimated the volumes of IT investments, as computer prices have fallen faster than the prices of other IT products. While this criticism is probably foremost valid with respect to early de?nitions of IT that involved many items whose IT character could be questioned, the argument is supported by Jorgenson’s (2001) study of relative prices for di?erent kinds of IT equipment in the US since the late 1940s.
4
8
IFAU–Human capital is the key to the IT productivity paradox
to outlays on other factors of production, notably human capital; cf. points 3 and 4. There are thus rather diverse results on the connection between IT and growth, and the explanations for these ?ndings are quite diverse, too.
3
A stylized model
We here consider a stylized version of the model that we use in our empirical analysis. Our discussion serves two purposes. The ?rst is to reconcile the di?erent results of the earlier studies and to discriminate between some of the explanations that have been suggested for the IT productivity paradox. The second purpose is to consider how knowledge spillovers and capital-skill complementarity might a?ect productivity growth. Our stylized model captures four features: i) measurement error in the IT input variable(s), ii) mismeasurement of output, iii) positive externalities in the use of IT, and iv) the connection between IT and human capital. The analysis is consistent with both a neoclassical growth theory framework and with endogenous growth models. We can thus here disregard the fact that these two theoretical frameworks have di?erent implications for the empirical analysis, notably with respect to how IT and human capital are operationalized.6 Regarding feature i., it was noted in Section 2 that the IT measurement error can be both negative and positive. A simple speci?cation allowing for this is ITt? = ITt + wt
6
(1)
The empirical speci?cation of the model will be discussed in Section 4.2. 9
IFAU–Human capital is the key to the IT productivity paradox
where ITt? is the observed mesure of IT in period t, ITt the true measure and wt a random error, such that E (wt ) = 0, V ar (wt ) = ? 2 w, Cov (ITt , wt ) = 0. (2)
With respect to feature ii., non-recorded quality improvements in output should introduce a downward bias in measures of productivity growth (cf. point 5 in Section 2). Like the mismeasurement of IT, the mismeasurement of output is likely to vary over time, cf. Basu et al. (2003). We therefore specify the di?erence between the ?rm’s true rate of TFP growth, gt , and
? , as a random variable with positive expectation, ? , the observed rate, gt 0
according to
? = ? 0 + ut , gt ? gt
? 0 > 0,
(3)
and E (ut ) = 0,
Feature iii. can be modeled by assuming that the productivity e?ects from IT at the ?rm and industry level are a?ected by the use of IT in the aggregate economy; see the last paragraph of point 2, Section 2. Assuming that there is an index of the Total Use of IT in the Swedish Economy, T U IT E , we posit that T U IT E has the e?ect of scaling up the IT input. Using an increasing function, ? , and allowing for a delayed impact on the rate of growth we arrive at the following direct e?ect of IT on gt : ? 1t ·ITt?1 ; ? 1t = ? (T U IT Et?1 ) and ? 0 > 0. (5) The scaling e?ect can thus be expressed in terms of a time-varying parameter, ? 1t . Note that we do not assume that this parameter is positive, a priori.
10 IFAU–Human capital is the key to the IT productivity paradox
¡ ¢ 2 E u2 t = ?u ,
Cov (ut , wt ) = 0.
(4)
The motivation for (5) is that, by de?nition, an externality is an e?ect which is not accounted for by individual ?rms and, hence, shows up in TFP growth. In a neoclassical context, this would mean that the capital rental price of IT would overstate the real cost of IT capital.7 In an endogenous growth context, as in, e.g., Barrro and Sala—i—Martin (1999) it is natural to relate to a learning—by—investing mechanism; as successively more ?rms invest in IT, the knowledge about the properties of the new technology increases and becomes more widespread. With respect to feature iv., our analysis will be based on the maintained hypothesis that information technology and human capital are complements, in accordance with, e.g., Bresnahan et al. (2002) and Kaiser (2003). We model the complementarity by means of an interaction variable, taken to a?ect gt positively. Allowing, again, for a delayed impact we get an indirect e?ect of IT on gt : ? 2 · (IT × HC )t?1 ; ? 2 > 0. (6)
Ordinarily, interaction e?ects should be captured already in the measure of productivity growth.8 In the context of externalities in the use of IT and/or measurement error in the IT input, the interaction e?ect may not be properly accounted for, however. There may be knowledge spillovers arising through networks: employees working with computers form networks (via the Internet) with colleagues in other ?rms, networks which facilitate the transfer of knowledge.9
7 Siegel (1997) tries to capture IT externalities within a neoclassical framework. However, instead of considering the total use of IT in the economy he uses a measure of the IT investments made by the industry’s suppliers. 8 We are assuming here that the TFP growth measure corresponds to a ?exible representation of the technology, implying that it allows for interactions between inputs; see Section 4.1 9 One might wonder why we allow for both ?rst- and second-order e?ects of IT on productivity growth but only for a second-order e?ect of HC. The reason is that the
IFAU–Human capital is the key to the IT productivity paradox
11
Taking the total e?ect of IT on gt to be the sum of the direct e?ect (5) and the indirect e?ect (6) and using (3) we obtain the following equation:
? = ?? o + ? 1t ITt?1 + ? 2 (IT × HC )t?1 ? ut . gt
(7)
By (7), the e?ect of "true" IT on the observed rate of TFP growth equals
? ?gt = ? 1t + ? 2 HCt?1 . ?ITt?1
(8)
Note that although the e?ect of IT on productivity growth is increasing in human capital, the total e?ect can be negative, provided that ? 1t is negative and su?ciently large in magnitude. Before proceeding to analyse the implications of our simple model, a word of caution is in order. A causal interpretation, from IT and HC to
? , is justi?ed only if the one year lag on IT and HC makes it possible gt
to treat these variables as predetermined. This, in turn, hinges upon the absence of serial correlation in the data. This is an empirical matter that we consider in Section 5.2 Using the framework given by equations (1) — (8) we now discuss three issues that have arisen in connection with earlier studies: I. Can the negative e?ects of IT on productivity growth found in studies based on pre—1990 data be explained by measurement error in the IT variable as argued by Lee & Barua (1999), or are the results indicative of a truly negative return to early IT investments, as argued by Morrison (1997)? II. Why is it that models similar to the one just outlined yield positive
features i — iv above, do not involve mismeasurement in human capital and also not externalities in human capital per se. The externalities that we consider are associated with IT, either through IT investments or through the use of IT. However, from an empirical point of view there might nevertheless be a place for a ?rst-order e?ect of HC in the model. This point is discussed in Section 5.2.
returns when applied to later data?
12
IFAU–Human capital is the key to the IT productivity paradox
III. If complementarity between IT and skilled labor is allowed for, like in Bresnahan et al. (2002), what will happen to the estimated direct e?ect?
? is simply regressed on IT ? , using data for the Assume, ?rst, that gt t?1
pre—1990 period and post—1990 period, respectively. This implies that the measurement error in IT is ignored, that the variable (IT × HC )t?1 is omitted, and that no account is taken of the fact that ? 1t is a time—varying coe?cient. For illustrative purposes we will here assume that the function ? is a step function, taking on the values ? 1,pre-90 during the pre-1990 period ? 1,post-90 in the post-1990 period. To derive the probability limit of the OLS estimate of ? 1t under this conditions, we apply a result stated in Lam & Schoeni (1993).10 This yields ³ ´ b b plim ? 1,K = ? 1,K ? ? 1,K · ? + ? 2 ? (1 ? ?) , K = pre-90, post-90 (9)
where the IT measurement error is accounted for by the parameter ?, de?ned as ?? V ar (w) , V ar (IT ? ) (10)
and b ? is the coe?cient from a hypothetical regression of IT × HC on IT : Cov (IT × HC, IT ) b , ?= V ar (IT ) b ? > 0.
(11)
From (9) it can be seen that the bias in the estimate of ? 1,K has two components. The ?rst, ?? 1,K · ?, is the measurement error bias (MEB). The second component, due to omission of the variable IT × HC , is the omitted variable bias (OVB). While the OVB is invariably positive, given
In a returns to schooling context, Lam & Schoeni (op.cit.) consider how the estimated e?ect on earnings from another year of schooling is a?ected when data on ”ability” are lacking and there is measurement error in the schooling variable.
10
IFAU–Human capital is the key to the IT productivity paradox
13
the assumptions ? 2 > 0 and b ? > 0, the sign of the MEB is determined by the sign of the true parameter ? 1,K . If ? 1,K is positive the MEB will be negative, and if ? 1,K is negative, the MEB will be positive. Equation (9) can be used to derive bounds on the probability limit of b . These bounds are given in Table 1, for various the OLS estimate ? 1,K assumptions about the true parameter and the magnitude of the omitted variable bias. We can now consider issue I. As can be seen in Table 1, the estimated e?ect of IT on productivity growth can be negative only if the corresponding true e?ect is negative. In this case, c), the true e?ect is ´ ³ b negative and smaller than the lower bound of plim ? 1,pre-90 ; this is so because the omitted variable bias, ? b ? , is positive. Furthermore, this con2
clusion is una?ected by measurement error in the IT variable. The upper ´ ³ b bound of plim ? 1,pre-90 is equal to zero, irrespective of whether there is measurement error or not. Our analysis thus supports Morrison’s (1997) suggestion of overinvestment in IT during the latter part of the 1980’s, as overinvestment would, eventually, result in a negative e?ect of IT on productivity. And, as our conclusion is invariant to measurement error in the IT variable, we reject the claim in Lee & Barua (1999) that measurement errors were behind estimated negative e?ects of IT on productivity
growth.11
Actually, Lee & Barua state that ”.... the negative contribution of IT .... is attributable primarily to the choices of the IT de?ator and modeling technique.” However, they do not provide any assessment making it possible to disentangle the impacts of these two factors.
11
14
IFAU–Human capital is the key to the IT productivity paradox
Table 1: Ranges for the probability limit of the OLS estimator of ? 1,K ,for di?erent signs of the true e?ect and di?erent magnitudes of the omitted variable bias ´ ³ b ? ? +? b a) ? >0 =? 0?plim ? ?
1,K 1,K 1,K 2
b)
c)
? 1,K ¯? 1,K ¯
=? =?
Note: The index K denotes either pre-90 or post-90
? 1,K 1990
? 1,post-90 > ? 1,pre-90
(12)
It should be noted that (12) is not su?cient to determine the sign of ? 1,post-90 . If ? 1,pre-90 < 0 then ? 1,post-90 may be negative, too. Unfortub nately, the sign of the estimate ? 1,post-90 is no help here. In Table 1, we ´ ³ b see that plim ? 1,post-90 > 0 is consistent with both ? 1,post-90 > 0 and ? 1,post-90 < 0; cf cases a) and b), respectively. However, we can discriminclude a vector of proxy variables for the omitted variable, i.e. IT × HC . This will a?ect the estimate of ? 1,post-90 di?erently depending on the sign of the true parameter ? 1,post-90 . To show this, denote vector of proxy variIFAU–Human capital is the key to the IT productivity paradox 15
inate between the two cases by expanding the simple OLS regression to
b ables by P, and the corresponding estimate of ? 1,K by ? (1,K )·P . Then ´ ³ b = ? 1,K ? ? 1,K 1?R2 plim ? (1,K )·P
?
IT ? ×HC,P
(13)
2 2 ? where RIT ? ×HC,P denotes the R obtained when IT × HC is regressed
+ ? 2b ? (1 ? ?) · ? (IT ? , IT ? × HC, P)
on P, and ? (·) is a function that under fairly general conditions satis?es 0 < ? (·) < 1.12 Comparing (9) and (13) we note that ³ ´ ³ ´ b b ? 1,K > 0 =? plim ? ? < plim (1,K )·P 1,K . (14)
The implication (14) is due to the fact that the inclusion of proxy variables a?ects the measurement error bias (MEB) and the omitted variable bias (OVB) in the same direction when ? 1,K > 0. With respect to the MEB, ´ ³ 2 ?]0, 1[ implies that including proxies makes the fact that 1 ? RIT ? ×HC,P OVB, while positive, becomes smaller, too, because 0 < ? (·) < 1. On the other hand, if ? 1,K < 0 the e?ect of the proxy variables is ambiguous, the ambiguity being due to the fact that in this case the MEB and the OVB change in di?erent directions. Thus, by studying the e?ects of including proxy variables we should be able to infer the sign of the true parameter ? 1,post-90 . If ? 1,post-90 is indeed positive, then the estimate of ? 1,post-90 should be positive when human capital variables are excluded from the regression and this positive estimate should decrease towards zero when proxy variables for human capital are included.
Like (9), this equation draws on Lam & Schoeni (1993). They provide a similar expression to assess the e?ect on the estimated return to schooling when a proxy variable for the missing ability measure is included in the regression.
12
the MEB larger in magnitude, i.e. smaller because of the minus sign. The
16
IFAU–Human capital is the key to the IT productivity paradox
The analysis also provides the answer to issue III. It shows that the answer depends on the sign of the true direct e?ect. If the true direct e?ect is positive, allowing for indirect e?ects will decrease the estimated direct e?ect, cf.(14). If, on the other hand, the true direct e?ect is negative, allowing for indirect e?ects will have an ambiguous impact on the estimated direct e?ect.
4
Data and empirical speci?cation
Our empirical analysis covers 14 industries in the Swedish manufacturing sector, observed annually over the period 1986—95. The industry codes are given in Table 2. To indicate the relative size of the industries we also show their shares in manufacturing employment in the middle of the observation period. The data are Table 2: The industries considered and their shares in total manufacturing employment in 1991.
Industry code Industry Employment share 1991, % 9.4 3.0 8.5 14.7 7.9 3.3 4.0 11.5 13.5 8.1 12.3 2.2 0.8 0.8 100.0
3100 Food, Beverages and Tobacco 3200 Textile, Apparel & Leather 3300 Saw Mills and Wood Products 3400 Pulp, Paper and Printing & Publishing 3500 Chemical, Plastic Products. and Petroleum 3600 Non-Metallic Mineral Products 3700 Basic Metals 3810 Metal Products 3820 Machinery & Equipment, not elsewhere classi?ed 3830 Electrical Machinery, not elswhere classi?ed 3840 Transport Equipment, except Shipyards 3850 Instruments, Photographic & Optical Devices 3860 Shipyards 3900 Other Manufacturing 3000 Total Manufacturing Note: The classi?cation system used here is very close to the ISIC codes.
from the o?cial statistics produced by Statistics Sweden; from the National Accounts, the Employment Register, the Labor Force Surveys, varIFAU–Human capital is the key to the IT productivity paradox 17
ious Investment Surveys and the Trade Statistics. The cross-sectional dimension of the data set has been determined by the most detailed break—down of IT investments provided in the Investment Surveys. In the time series dimension, the starting point is given by the ?rst year of the Employment Register. The end point is the result of a change in the industrial classi?cation system, making it impossible to extend the time series beyond 1995.
4.1
The growth rate in total factor productivity
The yearly TFP growth rates have been computed by means of a Törnqvist index. This index corresponds to the translog production function and allows for interactions among inputs like, e.g., complementarity between IT and human capital.13 Suppressing industry indexes and denoting the volume of gross output by Y and the volume of input i by Xi , the TFP growth rate g , is de?ned as
gt ? ? ln T F Pt = ? ln Yt ? ? ln Xt
t = 1986, ...., 1995
(15)
where ? is the di?erence operator, de?ned such that ? ln Zt ? ln Zt ? ln Zt?1 . The growth in aggregate input, Xt , is given by: ? ln Xt =
8 X i=1
wi,t ? ln Xi,t ,.
(16)
where the weights wit are de?ned in terms of average cost shares according to wi,t
13
1 = 2
µ
Cf. Jorgenson et al. (1973) and Caves et al (1982).
P Pi,t Xi,t X Pn i,t?1 i,t?1 + Pn k=1 Pk,t?1 Xk,t?1 k=1 Pk,t Xk,t
¶
,
(17)
18
IFAU–Human capital is the key to the IT productivity paradox
and Pi is price of input i. We consider the following eight inputs, which will be discussed below, KC = Stock of computer equipment capital, KM = Stock of non-computer equipment capital, KS = Stock of structure capital, L1 = # of full-time employees with elementary school (less than 9 years), L2 = # of full-time employees with 9 year compulsory school, L3 = # of full-time employees with upper secondary school, L4 = # of full-time employees with tertiary and postgraduate education, IG = Intermediate goods. Figure 1 shows how the industry-weighted average of TFP growth has evolved over time. While the period 1986—90 showed low but stable
growth, the growth rates during 1991—95 were much higher and also more volatile. Also, Figure 2 shows that the variation around the average is smaller in 1991—95 than in 1986—90. Thus, the higher average growth in the ?rst half of the 1990s is not merely the result of high growth rates in some large industries.As noted in the introduction, the turning point apparently occurred quite early in Sweden. For instance, Stiroh (2002) estimates that the breakpoint in U.S. manufacturing was passed in 1993.
IFAU–Human capital is the key to the IT productivity paradox
19
Figure 1: Weighted averages of TFP growth rates in Swedish manufacturing 1986-1995. Industry weights equal to employment shares
0.04 0.03
0.02
0.01
0
86 87 88 89 90 91 92 93 94 95
-0.01
-0.02
Figure 2: The industry variation around the weighted average. All observations lie within the bounds given by the dashed lines.
0.15 0.1
0.05
0
86 87 88 89 90 91 92 93 94 95
-0.05
-0.1
-0.15
It can be argued, of course, that the increase in TFP growth in the latter half of the period is not only due to IT developments, but also to business cycle changes. We thus control for the business cycle in the empirical analysis, cf. Section 4.5.
4.2
Speci?cation of the explanatory variables
We consider three alternative speci?cations of the explanatory variables.
20 IFAU–Human capital is the key to the IT productivity paradox
The ?rst, due to neoclassical growth theory as originally formulated by Solow (1956), implies that the explanatory variables should be speci?ed in terms of growth rates. In a neoclassical context, the primary reason for explaining variations in TFP growth by means input growth rates is presence of input measurement error. While less natural, externalities can also be used as a motivation.14 The second framework is endogenous growth theory, which predicts that the levels of (some) inputs determine the rate of productivity growth. Endogenous growth theory explicitly deals with the rôle of externalities in explaining growth; see, e.g., Barro & Sala-i-Martin (1999). There are also endogenous growth models where growth is increased by devoting resources to R&D [Romer (1990) and Aghion & Howitt (1992)].Since resources devoted to R&D are essentially resources devoted to sophisticated capital equipment (IT) and highly educated workers, these models provide a motivation for the current study. Another argument can be derived from the literature on GPTs: successful implementation of a new GPT and the generation of skills needed to operate it e?ciently is a cumulative process. As such, it should be better captured by the developments of stocks (of IT and human capital) than by yearly ?ows, i.e. growth rates. The third framework is due to Jones’ (1995, 1999) critique of endogenous growth models. Jones (1995) argues that the claim that the level of R&D should determine the rate of growth is inconsistent with empirical data. He notes, however, that a simple way to avoid that increases in the levels of inputs can increase growth without limit is to substitute input proportions for input levels. For instance, if resources devoted to R&D can be approximated by "research labor" then, instead of having the number
14 A study framed in the neoclassical tradition which considers both measurement errors and externalities is Siegel (1997).
IFAU–Human capital is the key to the IT productivity paradox
21
of research workers determining the rate of growth, one could have the share of research workers in total employment. As there are no clear theoretical arguments for preferring one of these speci?cations in favor of the others, we have estimated models according to each one of them. Our general conclusions can be formulated as follows. Similar to the experience of Benhabib & Spiegel (1994), the neoclassical speci?cation with explanatory variables in growth rates yielded largely insigni?cant results. The level speci?cation of the original endogenous growth models to a larger extent resulted in signi?cant estimates but these were often implausible with respect to sign. The input proportions speci?cation yielded the best results in terms of signi?cance, signs and goodness-of-?t. We thus focus on this alternative.15
4.3
Measures of IT equipment and IT use
As our measure of IT, we use the share of computers in the total capital stock, KC /K . The computer capital stock has been constructed by means of data on computer investments collected through investment surveys conducted by Statistics Sweden. The computer investments cover investments made both for o?ce use and for use in the production process, e.g., CNC (computer numerically controlled) equipment and CAD / CAM — systems.16 For the manufacturing sector as a whole, computer investments for use in the production process were 3—4 times as large as those for o?ce use, during the period that we study. By means of the computer investments data we have broken down
However, results corresponding to the rates and levels speci?cations are avaiable on request. 16 The de?nition of IT investments employed here di?ers from de?nitions used in some recent U.S. studies. For example, Gordon (2000), Jorgenson & Stiroh (2000), and Oliner & Sichel (2000) de?ne IT investments as investments in hardware, software, and telecommunications.
15
22
IFAU–Human capital is the key to the IT productivity paradox
the industry-speci?c stocks of equipment capital provided in the National Accounts into computer capital stocks, KC , and stocks of non-computer equipment, KM . Details on the computation are provided in the Appendix. Table 3: Capital stock shares in Swedish manufacturing
Industry 3100 3200 3300 3400 3500 3600 3700 3810 3820 3830 3840 3850 3860 3900 3000 Computers 1985 1990 1994 2.8 5.5 7.8 3.5 6.6 6.9 3.0 17.2 12.6 9.2 13.8 14.1 4.0 7.0 12.1 2.0 6.1 6.7 2.2 9.9 10.8 8.8 18.0 15.6 13.4 17.8 21.0 16.1 16.2 32.7 19.7 21.0 36.2 23.6 15.7 21.0 1.9 3.1 7.2 2.1 5.0 6.5 7.9 13.4 17.3 Equipment 1985 1990 1994 48.6 48.8 48.7 60.7 56.4 49.0 47.1 33.2 39.1 56.0 54.2 53.4 61.4 60.4 55.5 50.8 50.5 49.9 56.6 50.6 51.8 44.8 41.0 44.1 33.5 42.0 40.5 41.7 48.5 32.2 30.0 36.0 25.2 39.7 56.4 49.5 42.3 34.9 30.2 37.6 38.9 35.2 49.2 47.8 44.9 Structures 1985 1990 1994 48.6 45.7 43.5 35.9 37.0 44.1 49.9 49.6 48.3 34.8 32.0 32.5 34.6 32.6 32.4 47.2 43.4 43.4 41.2 39.4 37.3 46.5 41.0 40.3 53.1 40.1 38.5 42.2 35.3 35.1 50.4 43.0 38.6 36.7 27.9 29.5 55.8 62.0 62.5 60.4 56.2 58.3 42.9 38.9 37.8
Table 3 shows the shares of computers, non-computer equipment and structures in the capital stock, for the beginning, middle and end of the period.17 In Table 3, we see that, for the manufacturing sector as a whole, the computer share in the capital stock more than doubled over the period 1985-94, from 7.9 percent to 17.3 percent. This is especially remarkable in view of the fact that computer capital depreciates much faster than other types of capital; we have assumed the rate of depreciation for computer capital to be 1/3. Table 3 also shows that in relative terms the largest increases in the computer shares took place between 1985 and 1990, rather than between 1990 and 1994. It can also be seen that there is a lot of variation across industries. This is important because the relatively short
17
The capital stocks for year t are de?ned as January 1.
IFAU–Human capital is the key to the IT productivity paradox
23
period covered by our data makes cross-sectional variation crucial in our empirical analysis. Figure 3: Index of total use of IT in Sweden, 1984=100
1000.0 900.0 800.0 700.0 600.0 500.0 400.0 300.0 200.0 100.0 0.0 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
To model the externalities associated with IT, we use an index of the Total Use of IT in the Swedish Economy, T UIT E , cf. (5). This index includes both computers & peripherals, and communication equipment. It is de?ned as
N N N + IM PIT,t ? EXPIT,t T U IT Et = P RODIT,t
(18)
N , IM P N , and EXP N denoting volumes of production, imP RODIT,t IT,t IT,t
ports and exports of IT at the national level. Figure 3 shows the evolution of T U IT E. It can be seen that the use of IT has increased extremely rapidly, especially from 1992 and onwards; between 1992 and 1995 the increase was threefold. Both KC /K and T U IT E are included in the regressions we with a one year lag, again to avoid endogeneity problems.
24 IFAU–Human capital is the key to the IT productivity paradox
4.4
The human capital data
The human capital variables have been constructed by means of the Swedish Employment Register and the Labor Force Surveys. The Employment Register contains employee information on industry, level of education and ?elds-of-study, age, sex, and immigrant status, and yearly earnings. The Labor Force Surveys provide data on work hours per week, by industry and sex, enabling an approximate conversion of number of employees into full-time equivalents.18 . Just like the use of capital, employment of labor is endogenously determined. In the empirical analysis, the human capital variables are thus also lagged one year, relative to productivity growth. Accordingly, the cross-classi?cations of labor for 1985, 1990 and 1994 in Table 4 are to be related to productivity growth rates in 1986, 1991 and 1995, respectively. The four cells in the upper left corner of the three sub-tables in Table 4 are identically zero, because the cross-classi?cation by ?elds-of-study is possible only for labor with at least upper secondary school. For the latter, quite detailed ?eld-of-study information is available, however. The labels ”engineering” and ”business administration” are used for brevity only; both encompass several sub?elds. The table shows that the human capital in the Swedish manufacturing sector changed dramatically during the period that we are studying. For instance, in 1985 almost half of the workers (49 percent) had no more than 9 years of schooling. In 1994, the share was 1/3. And, at the other end of the distribution, the share of workers with tertiary education almost doubled, from 9 to 16 percent. There is also considerable cross-section
18 The approximate nature of the conversion is due to the fact that the Labor Force Survey does not contain work hours by level of education.
IFAU–Human capital is the key to the IT productivity paradox
25
variation; in the empirical analysis we employ cross-classi?cations like Table 4 that di?er both by to industry and year. Table 4: Employment shares in Swedish manufacturing, by level of education and ?elds—of—study, 1985, 1990 and 1994. 1985:
Level of education < 9 years 9 years Upper secondary Tertiary P Level of education < 9 years 9 years Upper secondary Tertiary P Level of education < 9 years 9 years Upper secondary Tertiary P Engineering 0 0 0.25 0,06 0.31 Field-of-study Business administration 0 0 0.08 0,02 0.10
”other” 0.30 0.19 0.09 0.01 0.59
0.30 0.19 0.42 0.09 1
P
1990:
Field-of-study Engineering Business administration 0 0 0 0 0.29 0.09 0.08 0.03 0.37 0.12 Field-of-study Business administration 0 0 0.09 0.04 0.13
”other” 0.22 0.17 0.10 0.02 0.51
0.22 0.17 0.48 0.13 1
P
1994:
Engineering 0 0 0.31 0.10 0.41
”other” 0.18 0.16 0.11 0.02 0.47
0.18 0.16 0.51 0.16 1
P
In addition to levels of education and ?elds-of-study we also account for the workers’ age. The age structure can matter in two di?erent ways. On the hand, an education’s ”IT content” is higher the more recently the education was obtained, i.e. the younger the worker. This would point to a negative relation between age and productivity growth. On the other hand, older workers have accumulated more work experience than younger workers. If skills acquired in the workplace are more important for produc26 IFAU–Human capital is the key to the IT productivity paradox
tivity than computer skills acquired in school, then the relation between age and productivity growth should be positive instead. To empirically assess which of these two opposing forces that dominate the other we use the following variable # 16-29 year olds . # [(16-29) + (50-74)] year olds (19)
The idea underlying this variable is to capture e?ects of relative changes in tails of the age distribution; all employees in our data belong to the age interval 16-74 years.19 It should be noted that the ratio (19) can change even if the total number of 16-29 year olds plus the number of 50-74 year olds doesn’t change. Thus, e.g., substituting a given number of older workers with an equal number of younger worker will increase the ratio.20
4.5
Control variables
To account for cyclical variations in TFP growth, we have used a business cycle indicator, BCI , for the Swedish manufacturing sector, cf Figure 4. The indicator together data on orders, stocks of ?nished goods, and expected production.21
19 In terms of years, the right tail is longer than the left tail. However, the number of people working beyond the retirement age of 65 is very small. Hence, for practical purposes the tails can be considered to be equally long. 20 The fact that we model age structure e?ects by means of (19) should not be taken to mean that we deny the importance of changes in the share of 30-49 year olds for productivity growth; as shown by Malmberg (1994) workers aged 40-49 have made substantial positive contributions to growth in Sweden (along with 50-64 year olds) and Feyrer (2002) obtains similar results for a data set covering 108 di?erent countries. However, unlike these authors we are not primarily interested in the direct link between age demographics and productivity, but on e?ects working via interactions between workers of di?erent ages and IT. It is then natural to focus on the age categories that di?er the most in this respect, i.e. the youngest and the oldest workers. 21 The indicator has been constructed by the Swedish Institute for Economic Analysis (Konjunkturinstitutet).
IFAU–Human capital is the key to the IT productivity paradox
27
Comparing Figure 4 and Figure 1, we see that the BCI captures the turning points in TFP growth quite well. However, the BCI cannot explain the relative magnitudes of growth at di?erent points in time. In particular, it does not capture that TFP growth was much higher during 1991—95 than during 1986—90.22 Figure 4: The business cycle indicator (BCI ) for the Swedish manufacturing sector 1986-1995
20 10
0 1986 -10 1987 1988 1989 1990 1991 1992 1993 1994 1995
-20
-30
-40
To take into account that computer investments partly depend on other capital investments, we include the share of non-computer equipment in total capital, KM /K .23 As KC /K + KM /K + KS /K = 1 by de?nition, including KM /K together with KC /K means that we fully control for the industries’ capital structures. Finally, we include the shares of females and immigrants among the employees.Gender might be important for two reasons. Weinberg (2000) argues that computers create job openings for women by replacing physically demanding blue-collar jobs by jobs that require computer knowledge. Second, Lindbeck & Snower (2000) point out that modern work organizations are increasingly characterized by multi-tasking. If women are better
22 We do not want to use time dummies to control for the time variation that is common to all industries. Using time dummies amounts to eliminating the general time pro?le of the endogenous variable, i.e. the pro?le given in Figure 1. But that time pro?le is part of what we want to explain; one thing we want to test is whether our simple model can capture the change in the TFP growth pattern that occurred between the end of the 1980s and the beginning of the 1990s. 23 In this respect we follow earlier studies; see, e.g., Berndt and Morrison (1995).
-50
28
IFAU–Human capital is the key to the IT productivity paradox
suited to multi-tasking than men, as is often claimed, this should favor ?rms with a large female labor share. Regarding immigrants the direction of causality is more ambiguous. On the one hand, it can be conjectured that the increased international communication brought about by IT could be facilitated by a work-force comprising employees with di?erent cultural backgrounds. On the other hand, imperfect knowledge of the host country language might have an adverse e?ect on productivity.
5
Results
In the ?rst part of this section we test the empirical implications of the stylized model in Section 3, on our Swedish data. In the next subsection we consider various econometric issues. To focus on methodological aspects, the analysis is conducted within a modeling framework entailing a univariate representation of human capital. Based on our results in this section we decide upon a basic formulation of the model and an appropriate estimation method. In the last subsection we extend the basic model through multivariate speci?cations of human capital.24 Before discussing the results we will brie?y comment upon three features that are common to all the regressions. First, the estimations are based on weighted least squares (WLS), where the di?erent industries are weighted by their shares in manufacturing employment. Methodologically we thus follow, e.g., Berman, Bound, & Griliches (1994) and Kahn & Lim (1998). The motivation for the WLS procedure can be found in the latter paper: it is reasonable to assume the
While not ideal, this sequential approach is necessary due to the fact that our data set is rather small. Considering the issues of model formulation, estimation methods, and multivariate speci?cations of human capital simultaneously, we would simply run out of degrees of freedom.
24
IFAU–Human capital is the key to the IT productivity paradox
29
data for small industries to be noisier than the data for large industries. This assumption can be modeled by assuming that the standard errors of the (unweighted) residuals are inversely proportional to the square of employment. Weighting industries by employment shares will then make the residuals homoscedastic. Second, the following control variables are always included in the regressions: the (contemporaneous) business cycle indicator, BCI , the (lagged) share of non-computer equipment capital in the total capital stock, KM /K , and the shares of females and immigrants among the employees. Third, we do not explicitly account for possible measurement error in the IT variable, because we lack information on this issue.
5.1
Testing the implications of the stylized model
The ?rst point made in Section 3 was that the negative e?ects of IT on productivity growth reported in studies using early (pre—1990) data are not mere statistical artefacts. To see what can be said of the Swedish manufacturing sector in this respect, we estimate the following equation for the ?rst half of our study period: 1986-90:
? = ? 0.036 + controls ? 0.004 · ght (1.71) (0.08) KC K h,t?1 ,
R2 = 0.18 (20)
where absolute values of t—statistics are in parentheses.25 The e?ect of IT, i.e. the coe?cient of (KC /K )h,t?1 is negative. The theoretical analysis tells us that, although the estimate is insigni?cant, this indicates that IT had a negative impact on growth in Sweden, too, during the latter part of
25 To save space, we do not report the coe?cients for the control variables here, as they are of no interest with respect to theoretical implications that we consider.
30
IFAU–Human capital is the key to the IT productivity paradox
the 1980s. The intercept is negative as expected (although insigni?cant). According to the theoretical analysis, this means that the observed rate of
? , underestimates the true rate, g , by, on average, productivity growth, ght ht
3.6 percent; cf. (3). The second point made in Section 3 was that if the e?ect of IT on productivity growth turned positive in the 1990’s then we would expect, ?rst, a positive estimate of the impact of IT when ignoring human capital variables and, second, that this positive estimate should decrease after inclusion of human capital variables. The following regression shows that the ?rst condition is satis?ed: 1991-95: ght = ? 0.072 + controls + 0.204 ·
(1.83) (2.80) KC K h,t?1 ,
R2 = 0.51 (21)
The coe?cient for (KC /K )h,t?1 is now positive, and strongly signi?cant. It can also be noted that the intercept is still negative, as expected, and that it has increased in magnitude. This, too, is in line with expectations: one e?ect of the positive impact of IT will be quality improvements in output; to the extent that these are not captured in the data output growth and, hence, productivity growth will be (further) underestimated. To check the second condition we include the share of workers with tertiary education as a crude proxy for skilled labor. Interacting it with KC /K we obtain: 1991-95: ght = ? 0.067 + controls + 0.184 ·
(1.18) (1.02) KC K KC K h,t?1
+ 0.066 ·
(0.12)
³
# T ertiary # Employees
×
´
h,t?1
,
R2 = 0.52 (22)
The inclusion of the interaction variable decreases the estimated direct
IFAU–Human capital is the key to the IT productivity paradox 31
e?ect of IT from 0.204 to 0.184, i.e. the second condition is satis?ed, too. To summarize: these very simplistic regressions based on our stylized model point to a (small) negative e?ect on TFP in Swedish manufacturing during the second half of the 1980s and a positive e?ect after 1990. That is to say, they indicate a development qualitatively similar to the one experienced in the US, but with the turning point occurring somewhat earlier.
5.2
Econometric issues
In this section we will consider the following four issues: (1) the modeling of the time-varying e?ects of IT; cf. (5),(2) the potential presence of ?rst-order e?ects of human capital on TFP growth, in addition to the second-order interaction e?ect given by (6), (3) industry ?xed e?ects, and (4) serial correlation. Our starting point is the last speci?cation of the previous subsection, i.e. (22). We here estimate that model for the entire period of study, 1986-95, cf column I of Table 5.26 It can be seen that in contrast to the results obtained for the 1991-95 period the point estimate of the direct e?ect of IT on TFP growth is negative. Thus, when the impact is not allowed to vary over time, the positive e?ect during 1991-95 reported in (22) is dominated by a negative impact during 1986-90.27
In this section we also report the estimates obtained for the control variables. This is veri?ed when we apply the speci?cation used in (22) to data for 1986-90. This yields an estimate of the direct e?ect of IT that is equal to ?0.314 and signi?cant at the 1 % level.
27
26
32
IFAU–Human capital is the key to the IT productivity paradox
Table 5: Alternative model speci?cations, given univariate measure of human capital Dependent variable : ght
Intercept I -0.0239 (0.976) 0.0002 (2.046) 0.0880 (2.133) 0.0104 (0.325) -0.3010 (2.334) -0.0875 (1.009) 0.0002 (1.430) 0.0002 (1.441) 0.0349 (0.363) 1.0826 (3.276) No No 0.34 0.4961 (1.957) No No 0.34 0.3248 (0.606) No No 0.34 1.3426 (1.973) Yesa No 0.44 0.4225 (1.798) No Yesb 0.39 -0.0001 (0.483) 0.0002 (1.948) II -0.0515 (2.720) 0.0003 (2.569) 0.1179 (3.181) 0.0010 (0.313) -0.1868 (1.327) III -0.0471 (2.313) 0.0003 (2.526) 0.1072 (2.259) 0.0121 (0.371) -0.1972 (1.369) IV 0.0974 (0.894) 0.0002 (1.094) 0.1478 (1.998) -0.3340 (1.394) -0.5453 (1.581) V -0.0545 (3.207) 0.0003 (2.697) 0.1245 (3.644) 0.0078 (0.267) -0.1586 (1.225)
Control variables : BCIt ³ ´
KM K
³ ³
h,t?1
# F emales # Employees h,t?1 # Immigrants # Employees h,t?1 KC K
´
Direct ³ ´ e?ect of IT :
h,t?1
´
h i C T U IT E × ( K ) h K
#Tertiary #Employees h,t?1
t?1
Direct e?ect´of human capital : ³
h,t?1
IT interaction: ³ and human capital ´ #Tertiary. Kc #Employees × K
Industry dummies Correction for AR(1) residuals
R2
aThe reference industry is 3100 = Food, Beverages and Tobacco. b Iterative Parks (1967) procedure, second-round estimates.
Having thus established the need for a time-varying e?ect, we turn to the ?rst issue, the speci?cation of an explicit form for the function ? (T U IT E )t?1 . We have chosen to approximate ? by a linear function since our data only cover ten years, making it di?cult to precise estimate
IFAU–Human capital is the key to the IT productivity paradox 33
higher order approximations: ? (T U IT E )t?1 = ? · T U IT Et?1 ; ? > 0, (23)
where ? is a parameter and T U IT E the index described in Section 4.3.28 The e?ect of incorporating (??) can be assessed by comparing columns I and II in Table 5. It is clear that all the parameter estimates are a?ected. In particular, the point estimate of the direct e?ect of IT changes from ?0.0875 to 0.0002. And while the indirect e?ect decreases, the two changes do not cancel each other out; the partial derivative of gh,t with respect to (KC /K )h.t?1 [cf. (8)] increases in magnitude. As the time-varying speci?cation is in line with our theoretical model and does have an impact, we will stick to it in the following. The next issue concerns the possibility of direct, ?rst-order, e?ects of human capital on gh,t . While our theoretical analysis does not imply that human capital should have a direct e?ect on growth — cf. footnote 10 — there might still be empirical grounds for such a direct e?ect. To assess this possibility we compare columns II and III in Table 5, which di?er only by the inclusion of the human capital variable in column III. It can be seen that the direct e?ect of human capital is small and very imprecisely estimated. With respect to the other estimates, the only one a?ected is the coe?cient measuring the indirect, interaction, e?ect. That coe?cient becomes smaller and insigni?cant. Taken together, it appears that the
28
A disadvantage with the linear form is that it cannot allow the e?ect of IT on
TFP growth to change sign over time. As a result, the partial derivative (8) cannot be negative, under the assumptions made in Section 3. However, when we turn to a multivariate speci?cation of human capital, in Section 5.3, there is no reason to restrict all the IT and human capital interaction e?ects to be positive. The partial derivative of TFP growth with respect to IT might then change sign over time. It will be seen that this does indeed happen in our estimations.
34
IFAU–Human capital is the key to the IT productivity paradox
inclusion of a direct human capital e?ect has the clear disadvantage of creating multicollinearity problems but no discernible empirical advantage. Henceforth, we will therefore not consider direct e?ects of human capital. The third issue, allowing for industry ?xed e?ects amounts, in this context, to allow for cross-industry di?erences in the expected mismeasurement in output, cf. (3). While desirable, this generalization is quite costly in terms of degrees of freedom. Comparing columns II and IV in Table 5, we see that allowing for industry ?xed e?ects results in the estimate of the direct e?ect of IT becoming less signi?cant, both economically and statistically, while the economic signi?cance of the indirect e?ect is substantially increased. The ?xed e?ects themselves take on implausible values, however. For industry 3100 = Food, Beverages and Tobacco, which is the reference industry, the ?xed e?ect is given by the intercept. While insigni?cant, the estimate of the intercept says that the mismeasurement in output in industry 3100 is such that, on average, the (true) rate of productivity growth is overestimated by 9.7 percent. For the other industries, the ?xed e?ects are given by deviations from the reference level of 9.7 percent, determined by means of estimated coe?cients on industry dummies. These coe?cients imply that the estimated ?xed e?ects are positive for all the other industries as well.29 As we ?nd it really hard to believe that IT has resulted in TFP growth being overestimated in every industry we will disregard industry-speci?c ?xed e?ects from now on. The issue of serial correlation, ?nally, is important because the interpretation of the lagged explanatory variables as predetermined is valid
29 The coe?cients, which should be added to the intercept, are, by industry, 3200: 0.0693, 3300: -0.0684, 3400: -0.0526?? , 3500: -0.0262, 3600: -0.0901? , 3700: -0.0766, 3810: -0.0600, 3820: -0.0879, 3830: -0.0211, 3840: -0.0838, 3850: -0.0841? , 3860: 0.0934, 3900: 0.0193, where * and ** denote signi?cantly di?erent from zero at the 10 and 5 % level, respectively.
IFAU–Human capital is the key to the IT productivity paradox
35
only if the regression residuals ful?ll the assumption of being random disturbances and, hence, not correlated over time. As our panel only covers a ten-year period, formal tests for autocorrelation will, unfortunately, have very low power. Nevertheless, it is possible to estimate the parameters of a simple autoregressive structure. To this end we apply an iterated version of the procedure suggested by Parks (1967) to correct for ?rst-order autocorrelation in a multiple-equation context. The assumed autocorrelation structure is given by: uh,t = ?h uh,t?1 + eh,t , |?| < 1 (24)
where the eh,t are white noise disturbances. Note that the autocorrelation parameter, ?, is allowed to vary across industries. We apply this structure to the model given by column II in Table 5. The ?rst-round estimates of the ?h are obtained by application of (24) to the estimated residuals of the column II speci?cation. All 14 estimates ful?ll the requirement that |?| < 1. As judged from the t-statistics, only one estimate is signifantly di?erent from zero, at the 10 % level. Still, the ?rst-round estimates, denoted by b ?1h , are used to estimate the model: where
? (b yh,t ?1h ) ? (b ?1h ) yh,t
? (b ?1h ) = x? ?1h ) + u? ?1h yh,t h,t (? , b h,t b
1
(25)
= (1 ? b ?1h ) 2 yh,t ,
1 2
for t = 1986 for t = 1987, ..., 1995 (26) for t = 1986 for t = 1987, ..., 1995
the 1986 variables being constructed according to the Prais-Winsten transformation. The resulting ?-estimates were qualitatively similar to the ones
36 IFAU–Human capital is the key to the IT productivity paradox
x? ?1h ) = xh,t ? b ?1h · xh,t?1 , h,t (? , b
?1h ) = (1 ? b ?1h ) x? x? h,t (? , b h,t (? ) ,
= yh,t ? b ?1h · yh,t?1 ,
in column II of Table 5 with small di?erences in magnitude and signi?cance. By means of the u? ?1h ), second-round estimates b ?2h were obtained. h,t (b
Two of these estimates were signi?cantly di?erent from zero at the 10 % level, thus indicating no improvement with respect to autocorrelation, as compared to the original speci?cation (where only one of the estimated autocorrelation parameters was signi?cantly di?erent from zero at the 10 % level). The estimate of the vector ? obtained from the regression model transformed by means of the b ?2h was extremely close to the original ?
estimate; compare columns V and II in Table 5. From the table it can be seen that the t-statistics are very close, too. But again, there was no discernable improvement with respect to the residuals; of the b ?3h estimates
one was signi?cant, at the 5 % level. Upon further iterations, the initial pattern was repeated: the estimates of the structural parameters shifted back and forth between one alternative similar to the original column II speci?cation and one alternative extremely close to this speci?cation. In no case was there any improvement with respect to the serial correlation of the residuals, as compared to the column II speci?cation. Thus, there is no strong indication that the residuals of the model in column II of Table 5 are autocorrelated and application of a standard procedure to correct for possible autocorrelation has no e?ect on the parameter estimates and and seems to make the residuals less well-behaved. Based on the results of this section we conclude that, in line with the theoretical arguments in Section 3, it seems important to allow the e?ects of IT to vary over time. We do not ?nd that our modeling framework needs to be extended to account for the other three issues that we have considered — potential ?rst-order e?ects of human capital on TFP growth,
IFAU–Human capital is the key to the IT productivity paradox 37
industry ?xed e?ects, and serial correlation. Using speci?cation II in Table 5 as our starting point we now proceed to consider more detailed, multivariate speci?cations of human capital.
5.3
Multivariate speci?cations of human capital
Apart from indicating the need for relative measures (cf. Section 4.2) theory does not provide any guidance regarding the implementation of a more detailed speci?cation of human capital. We have constructed variables such that the model can tell the e?ects of marginal changes in the educational structure. The e?ect that we are interested in is given by the partial derivative of total factor productivity growth with respect to this measure: X ?ght b = ?i · Xi ? (KC /K )h,t?1
i=1 m
(27)
ciated human capital variable. The variance of this partial derivative is equal to " #
where b ?j denotes an estimated coe?cient and Xj represents an assom X i=1 m m X ´ ³ ´ ³ X b · V ar ?i + 2 Xi Xj Cov b ? ib ?j i=1 j>i
?ght V ar ? (KC /K )h,t?1
=
Xi2
(28)
As the variance computation is a bit complicated we will, to begin with, merely consider the individual terms in (27), implying that we only have to consider the corresponding t — ratios.
38
IFAU–Human capital is the key to the IT productivity paradox
Table 6: Growth regressions allowing for externalities in the use of IT Dependent variable : ght I II III
Intercept -0.0273 (1.132) 0.0002 (2.082) 0.0586 (1.426) 0.0201 (0.592) 0.0159 (0.110) -0.2440 (0.952) 0.0002 (1.902) 0.0545 (1.168) -0.0021 (0.051) 0.0440 (0.205) -0.0225 (1.226) 0.0002 (2.000) 0.0547 (1.753)
Control variables : BCIt ³ ´
KM K
³ ³
h,t?1
# F emales # Employees h,t?1 # Immigrants # Employees h,t?1
´
Direct e?ect of IT : h i C T U IT E × ( K ) h K
´
t?1
0.00006 (0.505) 0.6460 (2.061)
0.00001 (0.096)
Direct e?ect of human capital ´: ³ #Tertiary KC #(Upper sec.+Tertiary ) × ( K ) h,t?1 h i #Tertiary Engineers Kc × K #(Upper sec.+Tertiary Engineers) h h h h
#Tertiary Business adm. #(Upper sec.+Tertiary Bus. adm.) #Tertiary ”Other” #(Upper sec.+Tertiary ”Other”) #Upper sec. #(9 years+Upper sec.)
×
Kc K h,t?1
×
×
Kc K h,t?1
i
Kc K h,t?1
i
i
h,t?1
0.8497 (3.413) -0.8646 (2.039) 0.9498 (1.198) 0.4240 (1.812) -0.8122 (3.464) 0.403 0.9104 (2.996) -1.2996 (3.393) 0.437
0.8779 (5.289) -0.8324 (2.383) 0.8779 (5.289) 0.8779 (5.289) -1.2593 (5.877) 0.437
#16-29 year olds #(16-29+50-74 year olds)
×
Kc K h,t?1
i
R2
Table 6 reports the results of three di?erent speci?cations. In column I we have allowed for the possiblity that, in addition to tertiary educated
IFAU–Human capital is the key to the IT productivity paradox 39
workers, employees with upper secondary education also belong to the ?rm’s skilled workers. The number of employees with tertiary education has been related to the number of employees with upper secondary or tertiary education. Similarly, the number of upper secondary educated workers has been normalized by the number of workers with 9 years of education or upper secondary education. We also use the variable (19) to account for the age structure aspect of human capital. Clearly, accounting for upper secondary education and the age structure are important extensions. The corresponding parameter estimates are strongly signi?cant. Interestingly, the indirect e?ect of IT associated with the age structure is negative. This implies that the negative e?ect of lost work experience caused by old workers retiring outweighs the positive e?ect of the entry of young workers with high ”IT content” in their basic education. Comparing column I of Table 6 with column II of Table 5 we see that the more detailed modeling of human capital renders the estimated direct e?ect of IT smaller and that among the control variables only the business cycle indicator stays signi?cant. The next step is to disaggregate the measures of human capital even further, by ?elds of study; cf. column II of Table 6. We ?nd considerable di?erences across ?elds. In particular, while there is a positive indirect e?ect of IT associated with the relation between engineers with university education and engineers with upper seconday education there is a negative indirect e?ect connected with the corresponding categories in business administration. While the this di?erence is somewhat counter-intuitive, there are results in the literature that point in this direction. For example, Murphy et al. (1991) claim that while "entrepreneurs" a?ect growth pos40 IFAU–Human capital is the key to the IT productivity paradox
itively "rent-seekers" are harmful to growth. Proxying entrepreneurs and rent-seekers with engineers and lawyers, respectively, they ?nd empirical support for their claim. As our category Business administrators includes lawyers, this ?nding is relevant for our results. Further, Mellander and Skedinger (1999) show that in the mid 1990s wage premia for university education were much higher among business administrators than engineers in seven European countries, including Sweden, in spite of an engineering degree requiring more years of study. A possible interpretation is that the university wage premium for business administrators is ”too high”, relative to their contribution to productivity. The see if the regression model in column II can be expressed in a more parsimonious way, we test the following composite hypothesis: h ³ ´ i C (i) The coe?cients of T U IT E × K K are zero. (ii) The coe?cients of h equal . h h i #Females , # Employees × h
# Immigrants # Employees h,t?1
h t?1
h,t?1
and
i
# Tertiary Engineers # (Upper sec. + Tertiary Engineers )
# Tertiary ”Other” # (Upper sec. + Tertiary ”Other” )
×
KC K h,t?1 , h i # Upper sec. KC K h,t?1 and # (9 years + Upper sec.)
i
×
KC K h,t?1 are
i
With respect to hypothesis ii) it should be emphasized that equality among the coe?cients does not imply that the associated indirect e?ects of IT on productivity growth are equal. If the coe?cients are equal, the corresponding indirect e?ects will be determined by the relative magnitudes of the human capital variables. Among these, the ratio is invariably the largest. As indicated by the fact that there is no di?erence between the R2 s in columns II and III, the composite hypothesis cannot be rejected at any
IFAU–Human capital is the key to the IT productivity paradox 41
# Upper sec. # (9 years + Upper sec.)
standard level of signi?cance. We thus end up with a model containing only six parameters, which explains 44 percent of the variation in total factor productivity growth across industries and over time! What, then, are the relative magnitudes of the indirect e?ects in our preferred speci?cation, i.e. column III in Table 6? For the manufacturing sector as a whole this question can be answered by means (5.8) and Table 4. The largest positive indirect e?ect is the one associated with the ratio
# Upper sec. # (9 years + Upper sec.) ;
for a marginal increase in the share of com-
puters in total capital the e?ect varies between 0.60 percentage points in 1986 and 0.67 percentage points in 1995. The largest negative indirect e?ect, which is the one channeled through the age structure, i.e. the ratio
# 16 — 29 year olds # (16 - 29 + 50 - 74 year olds) ,
decreases in magnitude over time, from -0.68
percentage points in 1986 to 0.60 percentage points in 1995.30 The next to largest positive indirect e?ect is associated with the relation between university educated engineers and engineers with upper secondary education, the ratio
# Tertiary Engineers # (Upper sec. + Tertiary Engineers ) ;
the indirect
e?ect increases from 0.17 percentage points in 1986 to 0.21 percentage points in 1995. This e?ect is however o?set by the negative indirect e?ect connected to business administrators, which decreases from -0.17 percentage points in 1986 to -0.26 percentage points in 1995. Finally, a positive indirect e?ect stemming from the relation between employees with "other" university and upper secondary education, respectively, makes upp the balance: this positive e?ect increases from 0.09 percentage points in 1986 to 0.14 percentage points in 1995. While these results for the entire manufacturing sector provide a gen30 To save space, the age structure data have not been provided in Section 4.4. However, for the years 1985 and 1994 the age structure ratio is equal to 0.536 and 0.479, respectively, re?ecting a declining in?ow of young people and ageing of the incumbents.
42
IFAU–Human capital is the key to the IT productivity paradox
eral feeling for the time pro?le of the e?ect of IT on total factor productivity growth, an important feature of the model is that it allows the e?ect of marginal increases in computers’ share to vary over time and by industries. This is illustrated in Figures 5a—c, which are based on computations using speci?cation III in Table 6. The diagrams show the distributions of the partial derivatives (5.8) across industries at three points in time, 1986, 1991 and 1995. The estimates’ precision have been computed according to (5.9).The estimates can be interpreted as answering the following question: If the share of computers in total capital increases by 1 percent, what is the resulting change in the rate of growth in total factor productivity, in percentage points? The bars indicate the e?ects for individual industries. The solid line is a weighted average e?ect, where the industries are weighted by their employment shares. Looking at the development over time, we see that the marginal e?ects of computer investments have increased steadily over time. The weighted average e?ect rises from about 0.01 percentage point in 1986 to 0.05 in 1991, ending up at 0.17 percentage points in 1995. These average changes have been caused by upward shifts in the entire distributions of e?ects across industries. For instance, while only two industries record e?ects above
1 10
of a percentage point in 1986, e?ects of this magnitude are found
in six industries in 1991 and in 11 in 1995. In the latter year, the point estimates are 0.25 or higher in ?ve industries, indicating that a 1 percent increase in computers’ share in total capital increases the rate of TFP growth by
1 4
of a percentage point or more.
IFAU–Human capital is the key to the IT productivity paradox
43
Figure 5: Distributions over industries of the e?ects of a marginal increase in computers´ share of capital on TFP growth; regression III in Table 6, evaluated in 1986, 1991 and 1995.
a:1986
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth. Weighted mean value
-0.10 -0.15
b:1991
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth Weighted mean value
-0.10 -0.15
c:1995
0.35 0.30 0.25 Percentage points 0.20 0.15 0.10 0.05 0.00
Effect of a marginal increase in computers´ share of capital on TFP growth Weighted mean value
-0.10 -0.15
Note: Stars indicate signi?cance level: ”*” denoting 10 percent, ”**” 5 percent and ”***” 1 percent.
Among the three years covered by Figure 5a—c, the largest variation across industries is found in 1986. In that year the spread is 0.46 percentage points, the range being given by a negative e?ect of ?0.12 percentage points in 3840 = Transportation and a positive e?ect of 0.34 percentage
44 IFAU–Human capital is the key to the IT productivity paradox
31 0 38 0 40 * 33 * 00 38 ** 10 * 34 ** 00 * 36 ** 00 ** 38 * 30 * 38 ** 20 * 39 ** 00 ** 37 * 00 * 32 ** 00 * 35 ** 00 * 38 ** 50 ** 38 * 60 ** *
Industries
-0.05
31 00 38 40 36 00 38 10 33 00 34 00 38 2 39 0 00 38 ** 30 * 35 ** 00 ** 37 * 00 * 38 ** 60 * 32 ** 00 ** 38 * 50 ** *
Industries
-0.05
0* * 38 10 31 00 33 00 38 20 34 00 36 00 37 00 39 00 * 38 5 32 0 00 ** 38 30 35 * 00 38 ** 60 ** *
Industries
-0.05
38 4
points in 3860 = Shipyards.31 In 1991 and 1995 the spread is considerably smaller — about 0.30 percentage points in both years. Moreover, in 1995 the e?ects are positive in all industries. There are thus two ?ndings pointing to a fundamental di?erence between the beginning and the end of the period that we study: compared to 1986 the variation across industries is smaller in 1995 and the estimated e?ects are con?ned entirely to the positive domain, unlike 1986 when about a third were negative. In line with our basic hypothesis of the importance of human capital, a comparison of Figure 5 and Table 3 shows that the industries that had the largest increases in the shares of computers in total capital were not in general the industries that had the largest growth-enhancing e?ects of IT. For instance, the industries 3300 = Saw Mills and Wood Products and 3700 = Basic Metals increased the relative size of their computer capital stock dramatically between 1985 and 1990; cf Table 3. These investments did not result in top-ranking marginal e?ects of IT in either 1991 or 1995, however; see Figure 5. Conversely, industry 3850 = Instruments, Photographic & Optical Devices experienced very large IT-induced growth e?ects in 1991 and 1995. In this industry the share of computers decreased between 1985 and 1990 — cf Table 3. Instead, the share of skilled workers increased strongly in this industry.32 Finally, a notable result is that, compared to the U.S., we ?nd positive impacts of IT on growth in a broader spectrum of industries. According to
31 The shipyards rank very high in 1991 and 1995, too. Since the Swedish shipyards have undergone major structural changes since the mid 70’s and have been facing severe problems with low and, sometimes, negative pro?ts this industry could be seen as a potential outlier. To check this, we reestimated the model given by column III in Table 6, leaving out the shipyards. The parameters changes were entirely negligible, however. The reason is the WLS estimation procedure where the industries are weighted by employment; the shipyards account for less than 1 percent of manufacturing employment, during the period studied. 32 The latter fact cannot be inferred from the paper but can be seen when the Table 4 is broken down by industry.
IFAU–Human capital is the key to the IT productivity paradox
45
Gordon (2000), in the U.S. the e?ects of computer investments were essentialy zero outside the IT-producing industries and the industries producing durable manufacturing goods. In the Swedish manufacturing sector, these industries roughly correspond to: 3810, 3820, 3830, 3840, 3850, and 3860; see Table 2. From Figure 5 it can be seen that while we ?nd large marginal e?ects in some of these industries, notably in 3850 = Instruments and 3860 = Shipyards, we also see examples of negative or very small effects as in, e.g., in 3810 = Metals and 3840 = Transportation. On the other hand, there are several industries outside this group recording large positive e?ects like 3200 = Textiles and 3500 = Chemicals.33 Table 7: Statistics for non-nested tests of the presence of Kc /K in growth equation; critical value at 1% signi?cance level ±2.57 Model I II Ho : include Kc /K -0.329 -0.686 Ha : exclude Kc /K Ho : exclude Kc /K 3.193 4.112 Ha : include Kc /K
Note: i) the model speci?cations refer to the columns in Table 6 ii) ”include Kc /K” refers to the regressions in Table 6 while ”exclude Kc /K” means setting Kc /K=1 in those regressions iii) the test statistic is asymptotically normally distributed.
However, while our results certainly seem to indicate that the human capital variables are essential, one might wonder about the importance of the computer capital share, Kc /K . Is this variable really essential, too, or can the human capital variables do the job by themselves? This is an important question because our interpration of human capital being the key to the IT productivity paradox relies on the assumption that it
33 Using more recent U.S. data than Gordon (op.cit) and dummy variable techniques, Stiroh (2002) ?nds indications of substantial e?ects of IT after 1995 not only in industries producing IT and durable goods, but also in IT-intensive industries, de?ned as industries having above median shares of computers in total capital. He does not link these ?ndings to human capital structures, however.
46
IFAU–Human capital is the key to the IT productivity paradox
is the interaction between Kc /K and human capital that matters. To check if this is the case it is necessary to conduct a non-nested test of whether Kc /K should be included in the growth equations or not. To this end we use the J test proposed by Davidson and MacKinnon (1981). The results of applying this test to the speci?cations I and II in Table 6 are given in Table 7. Note that the results concern the testing of two hypotheses. An intrinsic feature of a non-nested test is that there is no natural null hypothesis. Being a speci?cation test, the non-nested test merely investigates how two alternative models ?t the data. In the ?rst row of Table 7 we provide the test statistics for the case when the speci?cations in Table 6 constitute the null hypotheses. The alternative, Ha , corresponds to when Kc /K = 1 in the regressions. In none of the tests can the null be rejected at any standard level of signi?cance. In the second row, the roles of the null hypothesis and the alternative hypothesis have been reversed. The null is very clearly rejected in favor of the alternative. These results provide strong evidence for the model speci?cations in Table 6 and reject the alternative speci?cations where KC /K = 1. Put di?erently, the outcomes give convincing support for the notion that it is the interaction between IT capital and human capital that drives our results. This conclusion is further strengthened by the fact that it is quite unusual that non-nested tests yield results as clear as in this case; often the tests produce inconsistent results (reject both of the null hypotheses) or inconclusive results (reject neither).34
34 The reason why we have not performed the test on speci?cation III in Table 6 is that the Davidson-MacKinnon test cannot be applied to models incorporating linear constraints. Pesaran and Hall (1998) discuss non-nested tests allowing for general linear restrictions. However, given the very clear outcomes of the tests reported in Table 7 and the fact that, statistically, the speci?cations II and III in Table 6 are very close we have not taken the trouble to formulate such a general test.
IFAU–Human capital is the key to the IT productivity paradox
47
6
Summary and conclusions
Our principal conclusion from this study is that human capital is the key to the IT productivity paradox. We substantiate this general conclusion with both theoretical and empirical results. Our theoretical analysis investigates the consequences of erroneously disregarding human capital aspects in assessments of the e?ects of IT on productivity growth. Speci?cally, we consider a model where IT a?ects growth both directly and indirectly, through complementarity with human capital, and analyze what happens to the estimate of the direct e?ect when the indirect e?ect is omitted. Regarding the negative e?ects of IT on growth reported in several studies using early (pre—1990) U.S. data, our conclusion is that these results are likely to indicate a truly negative e?ect, as suggested by Morrison (1997), rather than be a consequence of measurement error, as argued by, e.g., Lee and Barua (1999). The positive relation between IT and productivity growth found in studies based on more recent data is in our theoretical analysis attributed to positive external e?ects in the use of IT. These external e?ects are assumed to be increasing in the total use of IT, implying that as more and more IT capital is accumulated, the growth e?ects change from negative to positive. In the empirical analysis, we ?rst con?rm that the predictions generated in the theoretical analysis are valid for our data on the Swedish manufacturing sector. We then proceed to include successively more information about interactions between IT and human capital. As shown by the theoretical analysis, accounting for indirect e?ects of IT in this way reduces the estimated direct e?ect. Eventually, the direct e?ect ?nally
48 IFAU–Human capital is the key to the IT productivity paradox
vanishes altogether. We end up with a model that is very parsimonious in terms of parameters but, nevertheless, explains well over 40 percent of the variation in total factor productivity growth across industries and over time. In this model, all the interaction variables between IT and human capital are highly signi?cant. In general, the maintained hypothesis of complementarity between IT and skilled workers is con?rmed. The largest indirect e?ects of IT on growth are associated with workers having upper secondary education, relative to workers with only 9 years of education. Disaggregating by ?elds of study, we ?nd the next to largest e?ect to be associated with the relation between university educated engineers compared to engineers with upper secondary education. An exception to the complementarity relation between IT and skilled labor concerns workers within the ?eld of business administration and law. For these, the relation between university educated and workers with upper secondary education gives rise to a negative indirect impact on productivity growth. In the spirit of Murphy et al. (1991), we interpret the negative estimate as indicating rent-seeking behavior among business administrators and lawyers. Regarding the connection between human capital and the age structure we ?nd that replacing workers aged 50 or older by workers below 30 has a negative impact on productivity growth rates. This indicates that, during the period studied, the advantage of many of the younger workers of having become acquainted with IT during their school years did not outweigh the work experience acquired by the older workers. This negative indirect e?ect is quite large but decreasing, due to a declining in?ow of young
IFAU–Human capital is the key to the IT productivity paradox 49
people to the manufacturing sector. For the manufacturing sector as a whole, the model predicts that in the beginning of the period, in 1986, a 1 percent increase in the share of computers in total capital increased productivity growth by 0.01 percentage points only, i.e. an entirely negligible e?ect. In the middle of the period, in 1991, this average e?ect had grown to 0.05 percentage points, while at the end of the period, in 1995, it was up to 0.17 percentage points. The variation in e?ects across industries decreases over time. Moreover, while the e?ects of IT on growth are negative in several industries in 1986, the e?ects are positive in all industries in 1995. In ?ve of them the estimated e?ect was 0.25 or higher, saying that a 1 percent increase in computers’ capital share increased productivity growth by at least percentage point. To check that our results are not driven solely by human capital developments but by complementarity between IT and human capital, we perform non-tested tests for the presence of the IT variable in the growth equations. These tests provide very strong support for the complementarity hypothesis. In line with our basic hypothesis, we ?nd that the industries were the (relative) increases in computer capital have been particularly large are not necessarily the industries that show the largest marginal e?ects of IT on productivity growth. With respect to di?erences in e?ects across industries, we also relate our ?ndings to the claim in Gordon (2000) that IT has increased productivity growth only in a small number of U.S. industries. We show that, unlike in the U.S., the Swedish IT development has had positive e?ects outside the sectors producing IT and durable manufacturing goods. We
50 IFAU–Human capital is the key to the IT productivity paradox
1 4
of a
?nd strongly positive e?ects also in, e.g., the chemical industry and, even more interesting, in the textile industry. Regarding policy considerations, one conclusions is immediate: measures to promote increased use of IT should be followed up by measures promoting skill upgrading, especially from elementary to upper secondary education. Another implication is that measures aimed at facilitating early retirement among older workers, in order to make more room for young labor market entrants, can be (strongly) harmful for growth. It should be remembered, however, that our study is based on data ending quite a few years back. Our results on the age structure might have changed during recent years. Investigating whether this is the case is an important task for future research. Also, it should be noted that our ?ndings concern only the manufacturing sector and cannot be extended to the service sector or the economy as a whole. While analyses of the service and the entire economy lie beyond the scope of the present paper because of data limitations, we believe that such analyses are important tasks for future research.
IFAU–Human capital is the key to the IT productivity paradox
51
References
[1] Aghion, P. & P. Howitt (1992) ”A Model of Growth through Creative Destruction”, Econometrica, Vol. 60, pp. 323—351. [2] Autor, D.H., L.F. Katz, & A.B. Kreuger (1998) ”Computing Inequality: Have Computers Changed the Labor Market?”, Quarterly Journal of Economics, Vol. CXIII, pp. 1169-1213. [3] Barro, R.J. & X. Sala—i—Martin (1999) Economic Growth, MIT Press, Cambridge, Massachusetts, U.S.A. [4] Basu, S., J.G. Fernald, N. Oulton, & S. Srinivasan (2003) ”The Case of the Missing Productivity Growth: Or, Does Information Technology Explain Why Productivity Accelerated in the United States but not the United Kingdom?” NBER Macroeconomics Annual 2003 [5] Benhabib, J. & M.M. Spiegel (1994) ”The Role of Human Capital in Economic Development: Evidence from Aggregate Cross-Country Data”, Journal of Monetary Economics, Vol. 34, pp. 143-173. [6] Berman, E., J. Bound, & Z. Griliches (1994) ”Changes in the Demand for Skilled Labor within U.S. Manufacturing: Evidence from the Annual Survey of Manufactures”, Quarterly Journal of Economics, Vol. CIX, pp. 367—398. [7] Berndt, E.R., Z. Griliches & N.J. Rappaport (1995) ”Econometric Estimates of Price Index for Personal Computers”, Journal of Econometrics, Vol. 68, pp. 243-268. [8] Berndt, E.R. & C.J. Morrison (1995) ”High—tech capital formation and Economic Performance in U.S. manufacturing industries: An exploratory analysis”, Journal of Econometrics, Vol. 65, pp. 9—43. [9] Berndt, E.R. & N.J. Rappaport (2001) "Price and Quality of Desktop and Mobile Personal Computers: A Quarter-Century Historical Overview", American Economic Review, Papers and Proceedings, Vol. 91, pp. 268—273. [10] Bresnahan, T.F., E. Brynjolfsson, & L.M. Hitt (2002) ”Information Technology, Workplace Organization, and the Demand for Skilled
52 IFAU–Human capital is the key to the IT productivity paradox
Labor: Firm—Level Evidence”, Quarterly Journal of Economics, Vol. CXVII, pp. 339—376. [11] Bresnahan, T.F. & M. Trajtenberg (1995) ”General Purpose Technologies: ’Engines of Growth’ ?”, Journal of Econometrics, Vol. 65, pp. 83-108. [12] Brynjolfsson, E. (1993) ”Information Technology and the Productivity Paradox: Review and Assessment”, Communications of the ACM, Vol. 35, pp. 66—77. [13] Caves, D.W., L.R. Christensen, & W. E. Diewert (1982) ”The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity”, Econometrica, Vol. 50, pp. 1393—1414. [14] David, P. A. (1990) ”The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, American Economic Review, Papers and Proceedings, Vol. 80, pp. 355—361. [15] Davidson, R. & J. MacKinnon (1981) ”Several Tests for Model Speci?cation in the Presence of Alternative Hypotheses”, Econometrica, Vol. 49, pp. 781-793. [16] Dean, E.R. (1999) ”The Accuracy of the BLS Productivity Measures”, Monthly Labor Review, pp. 24-34. [17] Feyrer, J.D. (2002) ”Demographics and Productivity”, mimeo, Dept. of Economics, Dartmouth College. [18] Gordon, R.J. (2000) ”Does the ’New Economy’ Measure up to the Great Inventions of the Past?”, Journal of Economic Perspectives, Vol. 14, pp. 49-74. [19] Greenwood, J. & M. Yorukoglu (1997) ”1974”, Carniege-Rochester Series on Public Policy, Vol. 46, pp. 49-95. [20] Griliches, Z. (1969) ”Capital-Skill Complementarity”, Review of Economics and Statistics, Vol. LI, pp. 465—468. [21] Harris, S. E.& J. L. Katz (1991) ”Organizational Performance and Information Technology Investment Intensity in the Insurance Industry”, Organizational Science, Vol. 2, pp. 263—296.
IFAU–Human capital is the key to the IT productivity paradox 53
[22] Helpman, E. & M. Trajtenberg (1998) ”A Time to Sow and a Time to Reap: Growth Based on General Purpose Technologies”, in E. Helpman (ed.): General Purpose Technologies, MIT Press. [23] Jones, C.I. (1995) ”R&D-based Models of Economic Growth” Journal of Political Economy, Vol. 103, pp. 759—784. [24] Jones, C.I. (1999) ”Growth: With or Without Scale E?ects” American Economic Review, Papers and Proceedings, Vol. 89, pp. 139—144. [25] Jorgenson, D.W. (2001) ”Information Technology and the US Economy”, American Economic Review, Vol. 91, pp. 1—32. [26] Jorgenson, D.W., L.R. Christensen, & L.J. Lau (1973) Transcendental Logarithmic Production Frontiers”, Review of Economics and Statistics, Vol. 55, pp. 28—45. [27] Jorgenson, D.W. & K.J. Stiroh (2000) ”Raising the Speed Limit: U.S. Economic Growth in the Information Age”, Brookings Papers on Economic Activity, pp. 125—211. [28] Kahn, J.A. & J.—S. Lim (1998) ”Skilled Labor—Augmenting Technical Progress in U.S. Manufacturing”, Quarterly Journal of Economics, Vol. CXIII, pp. 1281—1308. [29] Kaiser, U.(2003) ”Strategic Complementarities Between Di?erent Types of ICT-Expenditures”, ZEW Discussion Paper No 03-46. [30] Lam, D. & R. F. Schoeni (1993) ”E?ects of Family Background on Earnings and Returns to Schooling: Evidence from Brazil”, Journal of Political Economy, Vol. 101, pp. 710—740. [31] Lee, B. & A. Barua (1999) ”An Integrated Assessment of Productivity and E?ciency Impacts of Information Technology Investments: Old Data, New Analysis and Evidence”, Journal of Productivity Analysis, Vol. 12, pp. 21—43. [32] Lindbeck, A. & D.J. Snower (2000) ”Multi-Task Learning and the Reorganization of Work: From Tayloristic to Holistic Organization”, Journal of Labor Economics, Vol. 18, pp. 353—376.
54 IFAU–Human capital is the key to the IT productivity paradox
[33] Loveman, G.W. (1988) ”An Assessment of the Productivity Impact of Information Technologies”, in T.J. Allen and M.S.S. Morton (eds): Information Technology and the Corporation of the 1990s: Research Studies, Cambridge, MA: MIT Press. [34] Lucas, R. (1990) ”Why Doesn´t Capital Flow from Rich to Poor Countries?”, American Economic Review, Vol. 80, pp. 92-96. [35] Malmberg, B. (1994) ”Age Structure E?ects on Economic Growth — Swedish Evidence”, Scandinavian Economic History Review, Vol. XLII, pp. 279-295. [36] Mellander, E. (1999) ”The Multi—Dimensional Nature of Technical Change and Skill—Biased Technical Change”, IUI Working Paper No 518. [37] Mellander, E. & P. Skedinger (1999) ”Corporate Job Ladders in Europe: Wage Premia for University vs High School—Level Jobs”, Swedish Economic Policy Review, Vol. 6, pp. 449—487. [38] Morrison, C.J. (1997) “Assessing the Productivity of Information Technology Equipment in U.S.Manufacturing Industries”, Review of Economics and Statistics, Vol. LXXIX, No. 3, pp. 471—481. [39] Murphy, K.M., A. Schleifer, & R. W. Vishny (1991) ”The Allocation of Talent: Implications for Growth”, Quarterly Journal of Economics, Vol. CVI, pp. 503-530. [40] Parks, R.W. (1967) ”E?cient Estimation of a System of Regression Equations when Disturbances are Both Serially and Contemporaneously Correlated”, Journal of the American Statistical Association, Vol. 62, pp. 500-509. [41] Oliner, S.D. & D. E. Sichel (2000) ”The Resurgence of Growth in the Late 1990s: Is Information Technology the Story?”, Journal of Economic Perspectives, Vol. 14, pp. 3-22. [42] Parsons, D., C.C. Gotlieb, & M. Denny (1993) ”Productivity and Computers in Canadian Banking”,Journal of Productivity Analysis, Vol. 4 pp. 95—113.
IFAU–Human capital is the key to the IT productivity paradox 55
[43] Pesaran, M. H. & A.D. Hall (1998) ”Tests of Non-Nested Linear Regression Models Subject to Linear Restrictions”, Economics Letters, Vol. 27, pp. 341-348. [44] Romer, P.M. (1990) ”Endogenous Technical Change”, Journal of Political Economy, Vol. 98, pp. S71-S102. [45] Siegel, D. (1997) ”The Impact of Computers on Manufacturing Productivity Growth: A Multiple—Indicators, Multiple—Cause Approach”, Review of Economics and Statistics, Vol. LXXIX, No. 1, pp. 69—78. [46] Solow, R.L. (1956) ”A Contribution to the Theory of Growth”, Quarterly Journal of Economics, Vol. LXX, pp. 65—94. [47] Solow, R.L. (1987) ”We’d Better Watch Out”, New York Times Book Review, July 12, p. 36. [48] Stiroh, K.J. (2002) ”Information Technology and the US Productivity Revival: What Do the Industry Data Say?”, American Economic Review, Vol. 92, pp. 1559-1576. [49] Triplett, J.E. (1999) ”The Solow Productivity Paradox: What Do Computers Do to Productivity?”, Canadian Journal of Economics, Vol. 32, pp. 309—334. [50] Weinberg, B.A. (2000) ”Computer Use and the Demand for Female Workers”, Industrial and Labor Relations Review, Vol. 53, pp. 290— 308.
56
IFAU–Human capital is the key to the IT productivity paradox
A
Computation of omputer capital
The Swedish National Accounts (SNA) provides data on capital stocks of equipment and structures (buildings) by 2- or 3-digit industries. In this section we show how the equipment capital stock can be decomposed into two parts, one computer capital stock and one stock för non-computer equipment. To this end, we ?rst have to to consider the computation of the SNA capital stocks and and the corresponding capital rental prices. To simplify the notation, we here suppress industry indices and denote the equipment stocks by KE,t and the stocks of structures by KB,t .The stocks are de?ned such that the period t stock denotes the stock as of January 1, year t. The perpetual inventory method used in the SNA to compute the stocks implies that they can be closely approximated by the following accumulation formula ¢ ¡ i = E, B . (29) Ki,t = 1 ? ? i Ki,t?1 + Ii,t?1 , The capital rental prices for equipment and structure capital are constructed according to " !# ¡ ¢e à ¡ ¢e PIi ,t|t?1 ? PIi ,t?1 PIi ,t|t?1 ? PKi ,t = PIi ,t?1 rt?1 + ? i (30) PIi ,t?1 PIi ,t?1
where PKi ,t denotes the rental price for type i captal at the beginning of period t, PIi ,t?1 is the gross investment price index for type i capital and is a long-term interest rate measured at the very end of period t ? 1, rt?1 ¡ ¢e period t ? 1, and PIi ,t|t?1 is the expected value of the investment price index for type i capital in period t, given information this index ¡ about¢e up to (and including) period t ? 1. The di?erence PIi ,t|t?1 ? PIi ,t?1 measures the expected windfall pro?t (loss) that accrues to the owner of the capital asset through an increase (decrease) in the renewal cost.35 Like the ? i , the PIi are obtained from the SNA. The interest rate r is measured by means of the nominal ¢e Swedish long-term industrial ¡ rate on bonds. The expectional variable PIi ,t|t?1 is implemented by means of
35 The rental price formula (30) corresponds to the one given by equation (B4) in Jorgenson & Stiroh (2000). The only di?erence being that Jorgenson and Stiroh (op.cit.) assume perfect foresight with respect to the investment price index, thus substituting ¢e ¡ PIi ,t for PIi ,t|t?1 .
IFAU–Human capital is the key to the IT productivity paradox
57
a univariate Kalman ?lter.36 The rental prices are normalized to unity in a base-year to — here set to 1991 — yielding: eK ,t = PKi ,t . (31) P i PKi ,to To preserve the property that price × quantity = cost, the quantity of capital is normalized accordingly, i.e. e i,t = PK ,to Ki,t K i (32)
eK ,t K e i,t = PK ,t Ki,t . such that P i i To obtain the computer capital stock, we split the equipment stock KE into KEC and KEM where subindex C denotes Computers and subindex M stands for machines (that are not computers). In analogy with (29): KEC ,t = (1 ? ? EC ) KEC ,t?1 + IEC ,t?1 (33)
To make (33) operational, we have to decide on a value for ? EC and on an initial value for KEC . We have set ? EC = 1 3 . One motivation is that in the SNA depreciation rates for equipment (including computers) varies between 0.16 and 0.21. As computer capital depreciates much faster than other types of equipment ? EC should considerably larger than 0.21, making 1 3 a rather reasonable number. It is also close to the depreciation rate of 0.315 (from the Bureau of Economic Analysis) employed by Jorgenson & Stiroh (op.cit.). The initial value for KEC is obtained by extrapolating gross investments, IEC , backwards. To this end, we have assumed that investments during the period 1980-1994 can be approximated by the arithmetic average of the 1985 and 1986 gross investments. For the computation of the TFP growth rate according to Section 5.1, we also need a capital rental price for computer capital. The computation of this rental price is very similar to (30). For the gross investment price index PIEC ,t we use an import price for computers and peripherals, normalized to unity in 1991. Unfortunately, this index can only be computed for 1984-1995. During this period the index shows a continous decrease
36 This ?lter amounts to modeling the price index by means of a transition equation and a measurement equation. The former models the "true" investment price index as a random walk, incorporating a drift in the form of a deterministic quadratic time trend. The measurement equation models the observed price index as the sum of the "true" index and a random error.
58
IFAU–Human capital is the key to the IT productivity paradox
in the price of computers and peripherals, at an increasing rate. Between 1984 and 1985 the rate of decrease was very small, only 0.1 %, while between 1994 and 1995 the index fell by 14.3 %. The arithmetic mean of the rates of price decreases over the period was around 6.5 %.37 As our time series on PIEC ,t is so short we cannot model the expected investment price index by means of a Kalman ?lter. Instead we have simply ?tted a linear trend to the log-di?erences of the index, to estimate the average rate of decrease in the yearly price reductions, i.e. the discrete analogue of the second order derivative. We obtain an estimate of -1.24 percent annually, implying that for computer capital the last term within brackets in (30).is equal to zero in 1985 and the falls cumulatively by -1.24 each year, to reach -12.4 percent in 1995. Given the stock of computer capital and the computer capital rental price we can consistently solve for the expenditures on (non-computer) machinery equipment. Denoting these expenditures by VKEM ,t we get ³ ´ eK ,t K eK ,t K eK ,t K e K ,t = P e E,t ? P e E ,t VKEM ,t ? P EM EM E EC C (34)
because rental expenditures on computers and non-computer machinery have to add up to total rental expenditures on equipment capital. e K ,t . To solve for P eK ,t , eK ,t and K The ?nal step is determine P EM EM EM we ?rst assume that the rental price of equipment capital can be approxeK ,t : eK ,t and P imated by a translog aggregate of P E E eK ,t = ? ln P E
1 2
+
eK ,t (St?1 + St ) · ? ln P EC
1 2
where St =
eK ,t [(1 ? St?1 ) + (1 ? St )] · ? ln P EM eK P EC eK ,t K e E ,t P EC C e ,t KE ,t + VK
C
(35)
.
(36)
EM ,t
37 This may seem like a rather small rate of price decrease. It is smaller than similar estimates for the US but the di?erence is not as large as one might think. For comparison, Jorgenson and Stiroh (2000) report an average rate of decrease in the price of computer investments equal to 12.8 percent over the period 1985-1995. For communications investment they ?nd a much smaller rate of decrease, namely 0.6 percent over the same period. Thus, the decline in prices di?ers substanntially between di?erent types of computerrelated equipment. In our case, it might be that the prices of peripherals have fallen not fallan as fast as the prices of computers. Unfortunately, we cannot check this conjecture, as there is no separate price index for computers.
IFAU–Human capital is the key to the IT productivity paradox
59
eK ,t , we obtain Solving for ? ln P EM eK ,t = ? ln P EM ?
1
1 (St?1 +St ) 2
1 [(1?St?1 )+(1?St )] 2
1 [(1?St?1 )+(1?St )] 2
eK ,t · ? ln P E
(37)
eK ,t but not its The equation (37) determines the rate of change in P EM eK ,t , level. However, the level is determined by the normalization that P EM eK ,t and P eK ,t , should be equal to unity in the base-year. just like P E EC Thus, eK ,to ? 1.0 . (38) P EM e K ,t according to eK ,t we can ?nally solve for K Given P EM EM e K ,t = K EM VKEM ,t , eK ,t P
EM
eK ,t . · ? ln P EC
(39)
which constitutes the ?nal step in the break-down of the equipment capital stock into computer capital and (non-computer) machinery capital.
60
IFAU–Human capital is the key to the IT productivity paradox
Publication series published by the Institute for Labour Market Policy Evaluation (IFAU) – latest issues
Rapporter/Reports 2004:1 Björklund Anders, Per-Anders Edin, Peter Fredriksson & Alan Krueger “Education, equality, and efficiency – An analysis of Swedish school reforms during the 1990s” Lindell Mats “Erfarenheter av utbildningsreformen Kvalificerad yrkesutbildning: ett arbetsmarknadsperspektiv” Eriksson Stefan & Jonas Lagerström ”Väljer företag bort arbetslösa jobbsökande? Forslund Anders, Daniela Fröberg & Linus Lindqvist ”The Swedish activity guarantee” Franzén Elsie C & Lennart Johansson “Föreställningar om praktik som åtgärd för invandrares integration och socialisation i arbetslivet” Lindqvist Linus ”Deltagare och arbetsgivare i friårsförsöket” Larsson Laura ”Samspel mellan arbetslöshets- och sjukförsäkringen” Ericson Thomas ”Personalutbildning: en teoretisk och empirisk översikt” Calmfors Lars & Katarina Richardson ”Marknadskrafterna och lönebildningen i landsting och regioner”
2004:2 2004:3 2004:4 2004:5 2004:6 2004:7 2004:8 2004:9
2004:10 Dahlberg Matz & Eva Mörk ”Kommunanställda byråkraters dubbla roll” 2004:11 Mellander Erik, Gudmundur Gunnarsson & Eleni Savvidou ”Effekter av IT i svensk industri” Working Papers 2004:1 Frölich Markus, Michael Lechner & Heidi Steiger “Statistically assisted programme selection – International experiences and potential benefits for Switzerland” Eriksson Stefan & Jonas Lagerström “Competition between employed and unemployed job applicants: Swedish evidence” Forslund Anders & Thomas Lindh “Decentralisation of bargaining and manufacturing employment: Sweden 1970–96” Kolm Ann-Sofie & Birthe Larsen “Does tax evasion affect unemployment and educational choice? Schröder Lena “The role of youth programmes in the transition from school to work”
2004:2 2004:3 2004:4 2004:5
2004:6 2004:7 2004:8 2004:9
Nilsson Anna “Income inequality and crime: The case of Sweden” Larsson Laura & Oskar Nordström Skans “Early indication of program performance: The case of a Swedish temporary employment program” Larsson Laura “Harmonizing unemployment and sickness insurance: Why (not)?” Cantoni Eva & Xavier de Luna “Non-parametric adjustment for covariates when estimating a treatment effect”
2004:10 Johansson Per & Mårten Palme “Moral hazard and sickness insurance: Empirical evidence from a sickness insurance reform in Sweden” 2004:11 Dahlberg Matz & Eva Mörk “Public employment and the double role of bureaucrats” 2004:12 van den Berg Gerard J, Maarten Lindeboom & Peter J Dolton “Survey nonresponse and unemployment duration” 2004:13 Gunnarsson Gudmundur, Erik Mellander & Eleni Savvidou “Human capital is the key to the IT productivity paradox” Dissertation Series 2002:1 2002:2 2002:3 2002:4 2003:1 2003:2 Larsson Laura “Evaluating social programs: active labor market policies and social insurance” Nordström Skans Oskar “Labour market effects of working time reductions and demographic changes” Sianesi Barbara “Essays on the evaluation of social programmes and educational qualifications” Eriksson Stefan “The persistence of unemployment: Does competition between employed and unemployed job applicants matter?” Andersson Fredrik “Causes and labor market consequences of producer heterogeneity” Ekström Erika “Essays on inequality and education”
doc_665828832.pdf