The Role of Cognitive Skills in Economic Development
Journal of Economic Literature 2008, 46:3, 607–668 http:www.aeaweb.org/articles.php?doi=10.1257/jel.46.3.607
The Role of Cognitive Skills in Economic Development
Eric A. Hanushek and Ludger Woessmann*
The role of improved schooling, a central part of most development strategies, has become controversial because expansion of school attainment has not guaranteed improved economic conditions. This paper reviews the role of cognitive skills in pro- moting economic well-being, with a particular focus on the role of school quality and quantity. It concludes that there is strong evidence that the cognitive skills of the population—rather than mere school attainment—are powerfully related to indi- vidual earnings, to the distribution of income, and to economic growth. New empiri- cal results show the importance of both minimal and high level skills, the comple- mentarity of skills and the quality of economic institutions, and the robustness of the relationship between skills and growth. International comparisons incorporating expanded data on cognitive skills reveal much larger skill deficits in developing coun- tries than generally derived from just school enrollment and attainment. The mag- nitude of change needed makes clear that closing the economic gap with developed countries will require major structural changes in schooling institutions.
1. Introduction raise the schooling levels of the population. And, indeed, this is exactly the approach of t takes little analysis to see that schooling the Education for All initiative and a central Ilevels differ dramatically between devel- element of the Millennium Development oping and developed countries. Building Goals (see, e.g., UNESCO 2005 and David upon several decades of thought about E. Bloom 2006). human capital—and centuries of general But there are also some nagging uncer- attention to education in the more advanced tainties that exist with this strategy. First, countries—it is natural to believe that a pro- developed and developing countries differ ductive development strategy would be to in a myriad of ways other than schooling
* Hanushek: Stanford University, CESifo, and NBER. Paul Gertler, Manny Jimenez, Ruth Kagia, Beth King, Woessmann: Ifo Institute, University of Munich, and Lant Pritchett, Emiliana Vegas, the editor, and three CESifo. This project developed through conversa- anonymous referees. Jason Grissom provided important tions with Harry Patrinos, who provided useful com- research assistance. Support has come from the World ments and suggestions along the way. We have also Bank, CESifo, the Program on Education Policy and Gov- benefited from comments by Martha Ainsworth, Luis ernance of Harvard University, and the Packard Humani- Benveniste, François Bourguignon, Deon Filmer, ties Institute. 607 608 Journal of Economic Literature, Vol. XLVI (September 2008) levels. Second, a number of countries—both We will show in graphic terms the dif- on their own and with the assistance of oth- ferences in cognitive skills that exist, even ers—have expanded schooling opportunities after allowing for differences in school without seeing any dramatic catch-up with attainment. Most people would, in casual developed countries in terms of economic conversation, acknowledge that a year of well-being. Third, countries that do not schooling in a school in a Brazilian Amazon function well in general might not be more village was not the same as a year of school- able to mount effective education programs ing in a school in Belgium. They would also than they are to pursue other societal goals. agree that families, peers, and others con- Fourth, even when schooling policy is made tribute to education. Yet, the vast major- a focal point, many of the approaches under- ity of research on the economic impact of taken do not seem very effective and do not schools—largely due to expedience—ig- lead to the anticipated student outcomes. In nores both of these issues. The data suggest sum, is it obvious that education is the driv- that the casual conversation based on dis- ing force or merely one of several factors that parities in school attainment may actually are correlated with more fundamental devel- understate the magnitude of differences in opment forces? true education and skills across countries. The objective of this study is to review We think of education as the broad mix of what research says about the role of educa- inputs and processes that lead to individual tion in promoting economic well-being. We knowledge. Yet, because schooling and edu- pay particular attention to the robustness cation are conflated in common usage, we of the relationship between education and will generally refer specifically to cognitive economic outcomes across tests of alterna- skills, the part of education for which we tive specifications and hypotheses about the have good measures. underlying determinants of outcomes. We provide strong evidence that ignoring The discussion also has one distinctive ele- differences in cognitive skills significantly ment. We have come to conclude that cogni- distorts the picture about the relationship tive skills—particularly in assessing policies between education and economic outcomes. related to developing countries—are THE This distortion occurs at three levels. It misses key issue. It is both conventional and conve- important differences between education and nient in policy discussions to concentrate on skills on the one hand and individual earnings such things as years of school attainment or on the other. It misses an important underly- enrollment rates in schools. These things are ing factor determining the interpersonal dis- readily observed and measured. They appear tribution of incomes across societies. And, it in administrative data and they are published very significantly misses the important ele- on a consistent basis in virtually all countries ment of education in economic growth. of the world. And, they are very misleading The plan of this study is straightforward. in the policy debates. Based on a simple conceptual framework, Cognitive skills are related, among other we document the importance of cognitive things, to both the quantity and quality skills in determining individual earnings of schooling. But schooling that does not and, by implication, important aspects of improve cognitive skills, measured here by the income distribution. We then turn to comparable international tests of mathemat- the relationship of education and economic ics, science, and reading, has limited impact growth. Research into the economics of on aggregate economic outcomes and on growth has itself been a growth area, but economic development. much of the research focuses just on school Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 609 attainment with no consideration of cogni- Yet, while recognizing the impact of these tive skill differences that might arise from overall institutions, we find that cognitive variations in school quality or other sources skills play an important role. Furthermore, of learning. We show, in part with new evi- there is ample evidence that a high-quality dence, that the previous estimation is highly school system can lead to improved cognitive biased by concentration on just quantity of skills. And from a public policy perspective, schooling. interventions in the schools are generally The simple answer in the discussion of viewed as both more acceptable and more economic implications of education is that likely to succeed than, say, direct interven- cognitive skills have a strong impact on indi- tions in the family. vidual earnings. More than that, however, Given the evidence on the importance cognitive skills have a strong and robust of cognitive skills for economic outcomes, influence on economic growth. Models that we turn to what can be said about their include direct measures of cognitive skills production in developing countries. Although can account for about three times the varia- information on enrollment and attainment tion in economic growth than models that has been fairly widely available, cognitive include only years of schooling; including the skills information has not. We use newly cognitive skills measures makes the coeffi- developed data on international assessments cient on years of schooling go to zero; and of cognitive skills (also employed in the anal- the estimates of such more inclusive models ysis of growth) to show that the education are far more robust to variations in the over- deficits in developing countries are larger all model specification. In both the earnings than previously appreciated. and the growth areas, we can reject the most Finally, in terms of framing the analysis frequently mentioned alternative explana- here, we motivate much of the discussion by tions of the relationships. a consideration of world development goals To be sure, while there is a policy link to and the impacts of educational quality on schools, none of this says that schools per se economic outcomes in the developing world. are the answer. Even though it is common to Yet the bulk of the evidence on outcomes treat education and schooling synonymously, has clear and direct implications for devel- it is important to distinguish between oped countries. In simplest terms, available knowledge and skills on the one hand and evidence suggests that the economic situa- schooling. This distinction has important tion in, for example, the United States and substantive underpinnings. Cognitive skills Germany is governed by the same basic edu- may be developed in formal schooling, cational forces as that in developing coun- but—as extensively documented—they may tries of South America. Thus, this should also come from the family, the peers, the not be thought of so much as a treatise on culture, and so forth. Moreover, other fac- policies toward Sub-Saharan Africa but as tors obviously have an important impact on an analysis of the fundamental role of cogni- earnings and growth. For example, overall tive skills on the operations of international economic institutions—a well-defined sys- economies. tem of property rights, the openness of the economy, the security of the nation—can be 2. Conceptual Framework viewed almost as preconditions to economic development. And, without them, education We begin with a very simple earnings and skills may not have the desired impact model: individual earnings (y) are a function on economic outcomes. of the labor market skills of the individual 610 Journal of Economic Literature, Vol. XLVI (September 2008)
(H), where these skills are frequently referred ing H or its determinants in equation (2).1 to simply as the worker’s human capital. For We return to such estimation below but simplicity in equation (1), we assume that first discuss an alternative approach built on this is a one-dimensional index, although this direct measures of cognitive skills. is not important for our purposes: Consider test score measures of cognitive skills—that is, standardized assessments of (1) y 5 gH 1 e. mathematics, science, and reading achieve- ment. If these measures, denoted C, com- pletely capture variations in H, equation (1) The stochastic term, e, represents idiosyn- could be directly estimated with C, and the cratic earnings differences and is orthogonal key parameter (g) could be estimated in an to H. unbiased manner. Further, if school quantity, This abstract model has been refined in a S, is added to the estimation equation, it would wide variety of ways, most importantly by con- have no independent influence (in expecta- sidering the underlying behavior of individuals tion). But nobody believes that existing test in terms of their investments in developing data are complete measures of H. Instead, skills (Yoram Ben-Porath 1967, 1970; James J. the achievement test measures, C, are best Heckman 1976; Flavio Cunha et al. 2006). This thought of as error prone measures of H: basic model of earnings determination is cen- tral to most empirical investigations of wages and individual productivity. Before pursuing (3) C 5 H 1 m. this existing research, however, it is useful to understand where the skills might come from. With classical measurement errors, one These skills are affected by a range of factors would expect the estimate of g to be biased including family inputs (F), the quantity and toward zero and, if S and C are positively cor- quality of inputs provided by schools (which related as would be expected, the coefficient we incorporate as the function, Q(S), where on S in the same estimated equation would S is school attainment), individual ability (A), be biased upward even if S has no indepen- and other relevant factors (X) which include dent effect over and above its relationship labor market experience, health, and so forth. with C. This simple model would imply that Along with a stochastic term assumed uncor- the coefficient on C would be a lower bound related with the other determinants of H, we on the impact of human capital on incomes. can write this simply as The more common development, how- ever, begins with using school attainment (2) H 5 lF 1 fQ(S) 1 dA 1 aX 1 n. as a measure of human capital and proceeds
Human capital is nonetheless a latent vari- 1 Translating the abstract idea of skills into quan- tity of schooling was indeed the genius of Jacob Mincer, able. To be useful and verifiable, it is neces- who pioneered both the consideration of the character sary to specify the measurement of H. The of human capital investment and the empirical deter- vast majority of existing theoretical and mination of wages that built on this (Mincer 1970, 1974). This work spawned an extensive literature about empirical work on earnings determination the determination of earnings and the appropriate solves this by taking the quantity of schooling approach to the estimation. From the Mincer develop- of the individual (S) as a direct measure of H ment and its extensions, S is substituted for H in equa- tion (1) and its coefficient is frequently interpreted as and then dealing in one way or another with the rate of return on investment of one year of schooling the complications of not completely measur- when y is measured as the log of earnings (see below). Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 611 to interpret this under a variety of circum- There are nonetheless two complications stances. Propelled by readily available data in the interpretation of such models. The on school attainment within and across first is that schooling is but one of the fac- countries, economists have devoted enor- tors influencing cognitive skills and human mous energy to estimating the returns to capital formation. The second is that most additional years of schooling, and this work formulations of the Mincer model assume has been summarized and interpreted in a that school quality is either constant or can 2 number of different places. The focal point be captured by addition of direct measures has generally been how to get an unbiased of school quality in equation (4). estimate of g (or a simple transformation of The importance of nonschool influences g) under various considerations of other fac- on cognitive skills, particularly from the tors that might influence earnings and skills. family, has been well documented within For example, the ubiquitous Mincer earnings the literature on education production model takes the form functions, which provides the backdrop for the formulation in equation (2).3 If the 2 ( 4 ) y 5 b0 1 b1S 1 b2Z 1 b3Z 1 b4W 1 e, vector of other factors, W, in the earnings model includes the relevant other influ- ences on human capital from equation where Z is labor market experience, W is a (2), the estimate of b would be simply vector of other measured factors affecting 1 fg 4 incomes, and y is labor market earnings, (as long as school quality is constant). typically measured in logarithms. In this, Unfortunately, having rich information according to equation (2), the estimated about other determinants of skills outside of school attainment is seldom if ever avail- return to a year of schooling (b1) is biased through the correlation of S with F, A, and able, because this requires data about fac- any omitted elements of X; examples of such tors contemporaneous with schooling and correlation would be predicted in standard long before observed labor market data optimizing models of the choice of years are available. While a number of ingenious of schooling (e.g., Card 1999, Paul Glewwe approaches have been used, including for 2002). Recognizing this, significant attention example exploiting the common experi- has concentrated on ability bias arising from ences of twins, it is seldom plausible to correlation of school attainment with A and conclude that the other factors in equation from other selection effects having to do with (2) have been adequately controlled, lead- families and ability (see Card 1999). Indeed, ing to the interpretation of b1 as the com- various estimates of earnings models have bined influence of school and correlated added test score measures to the Mincer model in equation (4) explicitly to control for ability differences. 3 See, for example, the discussion in Eric A. Hanushek (1979). 4 While equation (2) highlights measurement error and its sources, t he histor ica l treatment has concentrated a lmost 2 See early discussions in Mincer (1970) and Zvi exclusively on simple misreporting of years of schooling, as Griliches (1977) and more recent reviews in David Card opposed to potential omitted variables bias from neglect- (1999) and Heckman, Lochner, and Todd (2006). In an ing (correlated) components of the true skill differences international context, see George Psacharopoulos (1994), contained in H. See, for example, Orley Ashenfelter and Psacharopoulos and Harry Anthony Patrinos (2004), Alan B. Krueger (1994). In our context, simple survey and Colm Harmon, Hessel Oosterbeek, and Ian Walker errors in S are a relatively small part of the measurement (2003). errors and omissions in specifying human capital. 612 Journal of Economic Literature, Vol. XLVI (September 2008)
5 but omitted other influences. As such, b1 measures of quality, so that the coefficient is a reduced form coefficient that will give on S might be interpreted as the return to biased estimates of the potential impact a year of average quality school.7 A variety from a policy designed to change school of authors have pursued this approach of attainment alone. adding quality measures generally related The second major issue revolves around to inputs such as spending per pupil or variations in school quality, indicated by Q(S) pupil–teacher ratios. Unfortunately, prior in equation (2). Perhaps the simplest formu- work investigating the determinants of skills lation of this is following equation (2) does not support measuring school quality in this manner.8 (5) Q(S) 5 qS, Additional complications of estimation and interpretation arise if human capital includes important elements of noncognitive skills. where time in school, S, is modified by a qual- Noncognitive skills, while seldom precisely ity index, q. As long as the variation in q is defined, include a variety of interpersonal small, this is not a serious additional impedi- dimensions including communications abil- ment to estimation. In an international con- ity, team work skills, acceptance of social text this assumption is clearly unreasonable, norms, and the like. Along such a line, Samuel as will be described below. But, even in terms Bowles and Herbert Gintis and more recently of individual countries (where earnings esti- Heckman and his coauthors have argued mation is virtually always done), there is also that noncognitive skills are very important substantial evidence that school quality varies for earnings differences.9 Their formula- significantly.6 tion essentially begins with the notion that If estimated in logarithmic form, the specification in equation (5) suggests deal- (6) H 5 C 1 N 1 m ing with school quality by adding direct
5 In terms of equation (2), studies using school- schooling than those not educated in such circumstances ing and income differences of twins (see Card 1999) (Hanushek and Lei Zhang 2008). assume that school quality differences are relatively 7 In reality, not much attention has been given to func- unimportant or unsystematic so that quantity of school- tional form, and much of the estimation is done in linear or ing, S, is the central object. Then, if ability, family semilinear form. An exception is Card and Krueger (1992) circumstances, and other factors affecting skills are who develop a multiplicative specification, but their for- relatively constant across twins, differences in school- mulation has also been criticized (Heckman, Anne Layne- ing can be related to differences in earnings to obtain Farrar, and Todd 1996). an unbiased estimate of g. Of course, the key ques- 8 In terms of school inputs in earnings functions, see tion remaining is why S differs across otherwise iden- Julian R. Betts (1996). More generally in terms of equa- tical twins who presumably face identical investment tion (2), see Hanushek (2003) and Ludger Woessmann payoffs. Other instrumental variable approaches have (2007). An alternative approach related to this paper is also been introduced to deal with the endogeneity the inclusion of test score measures C directly in equa- ofschooling, but they frequently will suffer if human tion (4) with the notion of interpreting C as a direct capital evolves from nonschool factors as in equation (2). measure of school quality that purges any school qual- 6 Moreover, changes in school quality over time within ity variations from the coefficient on years of school- individual countries may be important. Typically, Mincer ing. In the spirit of equation (2), this would not be earnings models compare the earnings associated with easily interpretable if cognitive skills, C, in fact cap- school attainment across individuals of different ages, ture the relevant variations in human capital, H. who necessarily obtained their schooling at different 9 See, for example, the early work along these lines that times, and implicitly assume that a year of schooling is the includes Bowles and Gintis (1976) and Bowles, Gintis, and same, regardless of when received. As a simple example, Melissa Osborne (2001). These is extended in a variety of people educated under wartime conditions may receive ways in Cunha et al. (2006) and Heckman, Jora Stixrud, significantly different educational outcomes per year of and Sergio Urzua (2006). Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 613 and then duplicates the general form of this, measures of N are typically not available, equation (2) to describe the underlying and there is little agreement on even what determinants of C and N.10 This suggests the dimensions might be important.12 If, how- following modification of equation (4), which ever, various noncognitive dimensions are is a reduced form equation that combines important, the estimate of b59 is the reduced influences of both cognitive and noncogni- form effect of cognitive skills and correlated tive factors through the channels of C and S: noncognitive skills. Focusing on measures of cognitive skills has 2 (7) y 5 b90 1 b91S 1 b92Z 1 b93Z 1 b94X a number of advantages. First, they capture variations in the knowledge and ability that 1 b95C 1 e9. schools strive to produce with their curri- cula and thus relate the putative outputs of In this formulation, estimation of equa- schooling to labor market success. Second, by tion (7) with the inclusion of C yields an impli- emphasizing total outcomes of education, they cation that the coefficient onS (i.e., b19) would incorporate skills from any source—families, reflect the impact of human capital differ- schools, ability, and so forth as seen in equa- ences that are not captured by C.11 Yet, for the tion (2). Third, by allowing for differences in same reasons discussed previously, b19 would performance among students with differing not be simply fg. It is still biased by other quality of schooling (but possibly the same omitted determinants of N, such as the family. quantity of schooling), they open the inves- It is important to note, nonetheless, that tigation of the importance of different poli- finding a direct effect of schooling on earn- cies designed to affect the quality aspects of ings to be zero (b19 5 0) after conditioning on schools. In that regard, this approach permits cognitive skills is not the same as saying that aligning investigations of the labor market school attainment does not matter. It merely with the extensive work that has also delved says that the impact of school comes entirely into aspects of educational production func- through the impact on cognitive skills, so that tions (see Hanushek 2002). Finally, recent schooling that does not raise cognitive skills policy attention to accountability in schools— is not productive. In general, the impact of along with the acceptance of parents that school attainment is cognitive skills are important outcomes of schools—reinforce giving more attention to (8) 0y/0S 5 b19 1 b59 (0C/0S). test-based measures of cognitive skills. At the same time, the test score measures A more complete approach might be to of cognitive skills, as indicated, also have dis- include direct measures of noncognitive advantages. As described, the tests that are skills, N, into the estimation in equation (7). given are undoubtedly narrower than either While some attempts have been made to do what is taught in schools or what elements are
10 The focus of this work, Cunha et al. (2006), is some- however, on the specific data samples and questions being what different from the discussion here. They are attempt- investigated. ing to obtain a better description of lifetime investment 12 Reviews of these analyses along with new estima- behavior and of the returns to varying kinds of human tion is found in Bowles, Gintis, and Osborne (2001), capital investment. Cunha et al. (2006), and Heckman, Stixrud, and Urzua 11 Heckman and Edward Vytlacil (2001) go fur- (2006). They use survey measures to capture such ther to argue that it is not possible to separate school behaviors as motivation, persistence, time prefer- attainment and achievement because they are so ence, and self control, but the choice of these appears highly correlated. The importance of this depends, to be heavily dependent upon the specific survey data. 614 Journal of Economic Literature, Vol. XLVI (September 2008) important in the labor market, including non- income instead of individual earnings, y. cognitive skills. This narrowness is clearest Much modeling of economic growth high- when considering individual tests of particu- lights labor force skills or the human capi- lar domains of knowledge, such as primary tal of the country (along with other things as school reading. Most of the available tests are described below). given at the school level, frequently at the end Again, the vast majority of growth modeling of lower secondary education, so that they do has simply taken measures of school attain- not directly capture variation in higher edu- ment to characterize skills. Here, however, cation (although they may do so indirectly differences in educational outcomes (either through their predictive power for obtaining total or per year of schooling) become central further education). Additionally, even as tests because the analysis assumes that, say, ten of specific subject matter at the secondary years of schooling means the same in terms school level, the issue of measurement error of skills regardless of the country. Intuitively, in the tests cannot be ignored. The tests may the error in measuring skills by school attain- suffer from a variety of problems related to ment becomes large when comparing coun- the sampling of knowledge in the particular tries. More importantly, the error will be very domain, the reliability of questions, and even systematic, based on the country and, quite the impact of test taking conditions on scores. likely, different aspects of the country. Again, as described above, these concerns The relevant omitted variables and mea- generally imply that the estimated effects of surement errors are just those depicted cognitive skills will be a lower bound on the in equation (2). The aggregate skills of impact of improved skills. individuals in a country will vary with This discussion was designed to sketch family inputs, school quality, ability differ- out the range of possible issues in estimat- ences, and other country specific factors.13 ing human capital earnings functions, but This motivates our work here. International the importance of the various issues is an test scores provide consistent measures of empirical question. The possibility of biased aggregate differences in cognitive skills estimates through correlated omitted vari- across countries. As such, they do not attri- ables is particularly important in some bute all differences in cognitive skills to contexts, such as the endogeneity of school the schools in different countries, although attainment, but is also relevant elsewhere. many policy discussions will hinge on the Acknowledging the range of possible issues importance of schools in producing differ- does not capture the importance of them ences in cognitive skills. in specific empirical analyses. In the subse- The following sections investigate the exist- quent interpretations of various estimation ing evidence about how educational outcomes schemes, we will focus on the importance of relate to economic outcomes. Throughout we the various issues raised in the conceptual emphasize the impacts of variations in mea- model and emphasize evidence that deals sured cognitive skills on economic outcomes. with particular threats to identification. First, we demonstrate that these educational This discussion of individual economic outcomes has relevance for modeling the 13 Krueger and Mikael Lindahl (2001) are motivated growth of nations. Even though developed by differences in the aggregate and individual estimates of in different ways, many relevant aspects of g, and they focus on measurement error in school attain- growth models can be put in the context of ment. That analysis is directly analogous to the work on measurement error in modeling individual earnings, and equations (1) and (2) where the outcome of it does not consider the more significant issues about other interest is aggregate income or growth in sources of differences in skills (equation (2)). Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 615 outcomes produce the most consistent and estimated schooling effect could include both reliable information about education and the impacts of schooling and the fact that development. Second, these outcomes can those continuing in school could earn more in then be related directly to the relevant policy the absence of schooling. For the most part, options facing nations. The standardly used employing alternative estimation approaches measures of school attainment, on the other dealing with the problems of endogeneity hand, provide much less reliable measures of schooling do not lead to large changes in of skill differences, particularly in the cross- the estimates, and many times they suggest country growth context. that the returns are actually larger with the alternative estimation schemes than with 3. Individual Returns to Education and the simpler modeling strategies (e.g., Philip Economic Inequality Oreopoulos 2006). Harmon, Oosterbeek, and Walker (2003) systematically review 3.1 Impacts of School Attainment on the various issues and analytical approaches Individual Incomes dealing with them along with providing a set Most attention to the value of school- of consistent estimates of returns (largely for ing focuses on the economic returns to OECD countries) and conclude that, while differing levels of school attainment for the estimation approaches can have an impact individuals as depicted in equation (4). on the precise value of the rate of return, it This work, following the innovative anal- is clear that there is a strong causal impact of yses of human capital by Mincer (1970, school attainment on earnings. 1974), considers how investing in differ- The basic estimates of Mincer earnings ing amounts of schooling affects individual models are typically interpreted as the pri- earnings. Over the past thirty years, liter- vate returns to schooling. As is well known, ally hundreds of such studies have been the social returns could differ from the pri- conducted around the world, reviewed and vate returns—and could be either above or interpreted by a variety of studies such below the private returns. The most com- as Psacharopoulos (1994), Card (1999), mon argument is that the social returns Harmon, Oosterbeek, and Walker (2003), will exceed the private returns because of Psacharopoulos and Patrinos (2004), and the positive effects of education on crime, Heckman, Lochner, and Todd (2006). health, fertility, improved citizen participa- These basic estimations of rates of return tion, and (as we discuss below) on growth have uniformly shown that more schooling is and productivity of the economy as a whole. associated with higher individual earnings. Recent studies indeed find evidence of The rate of return to schooling across coun- externalities of education in such areas as tries is centered at about 10 percent with reduced crime (Lochner and Enrico Moretti variations in expected ways based largely on 2004), improved health of children (Janet scarcity: returns appear higher for low income Currie and Moretti 2003), and improved countries, for lower levels of schooling, and, civic participation (Thomas S. Dee 2004, frequently, for women (Psacharopoulos and Kevin Milligan, Moretti, and Oreopoulos Patrinos 2004). 2004). The evidence on direct production Much of the academic debate has focused spillovers of education among workers is on whether these simple estimates provide more mixed, with Moretti (2004) and the credible measures of the causal effect of studies cited therein finding favorable evi- schooling. In particular, if more able people dence and Daron Acemoglu and Joshua D. tend also to obtain additional schooling, the Angrist (2001) and Antonio Ciccone and 616 Journal of Economic Literature, Vol. XLVI (September 2008)
Giovanni Peri (2006) finding no evidence But it is not appropriate simply to presume for this kind of spillovers. that any spending on schools is a productive In the Mincer earnings work, a specific investment. It is instead necessary to ascer- version of social rates of return is frequently tain two things: how various investments calculated. The typical calculations do not translate into skills and how those skills take account of possible positive externali- relate to economic returns. This and the fol- ties, but they instead take account of the fact lowing sections provide a summary of what that the social cost of subsidized education is known about the individual returns spe- exceeds the private costs—thus lowering the cifically to cognitive skills in both developed social rate of return relative to the private rate and developing countries. of return (see Psacharopoulos and Patrinos As discussed in section 2, one of the chal- 2004). In addition, the social return could be lenges to understanding the impact of human below the private return also if schooling was capital has been simply knowing how to mea- more of a selection device than of a means sure human capital. Here we focus on how of boosting knowledge and skills of individu- well cognitive achievement measures—stu- als. The empirical analysis of these issues has dents’ performance on standardized tests— been very difficult because the labor market proxy for the relevant labor market skills outcomes of the screening/selection model when assessed against individuals’ perfor- and the productivity/human capital model mance in the labor market and the economy’s are very similar if not identical. Fabian Lange ability to grow. and Robert Topel (2006) review the theory Until fairly recently, little comprehen- and empirical work and conclude that there is sive data have been available to show any little evidence that the social rate of return to relationship between differences in cognitive schooling is below the private rate of return. skills and any related economic outcomes. Thus, although there are many uncertainties The many analyses of school attainment and about precisely how social returns might dif- Mincer earnings functions rely upon read- fer from private returns, there is overall little ily available data from censuses and surveys, reason to believe that the social returns are which find it easy to collect information on less than the private returns, and there are a earnings, school attainment, age, and other variety of reasons to believe that they could demographic information. On the other hand, be noticeably higher. it is difficult to obtain data on cognitive skills along with earnings and the other determi- 3.2 Impacts of Cognitive Skills on Individual nants of wages. Although cognitive test and Incomes—Developed Countries school resource data are increasingly avail- able at the time of schooling, these are seldom The concentration on school attainment in linked to subsequent labor market informa- the academic literature, however, contrasts tion. Such analyses generally require tracking with much of the policy discussion that, individuals over time, a much more difficult even in the poorest areas, involves elements data collection scheme. Such longitudinal of “quality” of schooling. Most countries are data are, however, now becoming available. involved in policy debates about the improve- A variety of researchers are now able to ment of their schools. These debates, often document that the earnings advantages to phrased in terms of such things as teacher higher achievement on standardized tests are salaries or class sizes, rest on a presumption quite substantial. These results are derived that there is a high rate of return to schools from different specific approaches, but the in general and to quality in particular. basic underlying analysis involves estimating Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 617 a standard “Mincer” earnings function and Murnane et al. (2000) provide evidence adding a measure of individual cognitive from the High School and Beyond and the skills such as equation (7). The clearest analy- National Longitudinal Survey of the High ses are found in several references for the School Class of 1972. Their estimates sug- United States (analyzed in Hanushek 2002; gest some variation with males obtaining a see John Hillman Bishop 1989, 1991, June 15 percent increase and females a 10 per- O’Neill 1990, Jeff Grogger and Eric Eide cent increase per standard deviation of test 1995, McKinley L. Blackburn and David performance. Lazear (2003), relying on a Neumark 1993, 1995, Richard J. Murnane, somewhat younger sample from NELS88, John B. Willett, and Frank Levy 1995, Derek provides a single estimate of 12 percent. A. Neal and William R. Johnson 1996, Casey These estimates are also very close to those in B. Mulligan 1999, Murnane et al. 2000, Mulligan (1999), who finds 11 percent for the Joseph G. Altonji and Charles R. Pierret 2001, normalized AFQT score in the NLSY data.15 Murnane et al. 2001, and Edward P. Lazear Note that these returns can be thought of 2003). While these analyses emphasize dif- as how much earnings would increase with ferent aspects of individual earnings, they higher skills each and every year throughout typically find that measured achievement has the persons’ working career. Thus, the pres- a clear impact on earnings after allowing for ent value of the returns to cognitive skills is differences in the quantity of schooling, the large if these estimates can be interpreted as experiences of workers, and other factors that the structural impact, i.e., g in equation (1). might also influence earnings. In other words, These estimates are obtained fairly early in higher quality as measured by tests similar to the work career (mid-twenties to early thir- those currently being used in accountability ties), and analyses of the impact of cognitive systems around the world is closely related to skills across the entire work life are more lim- individual productivity and earnings. ited. Altonji and Pierret (2001) find that the Three recent U.S. studies provide direct impact of achievement on earnings grows with and quite consistent estimates of the impact of experience because the employer has a chance test performance on earnings (Mulligan 1999, to observe the performance of workers. The Murnane et al. 2000, Lazear 2003). These pattern of how returns change with age from studies employ different nationally represen- their analysis is shown in figure 1, where the tative data sets that follow students after they power of school attainment differences to leave school and enter the labor force. When predict differences in earnings is replaced scores are standardized, they suggest that one by cognitive skills as workers are in the labor standard deviation increase in mathematics force longer. The evidence is consistent with performance at the end of high schools trans- employers relying on readily available infor- lates into 12 percent higher annual earnings.14 mation on school attainment when they do
14 Because the units of measurement differ across tests, of measured cognitive skills on earnings by Bowles, Gintis, it is convenient to convert test scores into measures of the and Osborne (2001) finds that the mean estimate is only distribution of achievement across the population. A one- 0.07, leading them to conclude that the small gains are half standard deviation change would move somebody dwarfed by noncognitive factors. We return to this below. from the middle of the distribution (the fiftieth percen- 15 By way of comparison, we noted that estimates of tile) to the sixty-ninth percentile; a one standard deviation the value of an additional year of school attainment are change would move this person to the eighty-fourth per- typically about 10 percent. Of course, any investment centile. Because tests tend to follow a normal distribution, decisions must recognize that school attainment and test the percentile movements are largest at the center of the performance are generally produced together and that distribution. A separate review of the normalized impact costs of changing each must be taken into account. 618 Journal of Economic Literature, Vol. XLVI (September 2008)
14%
1 year of education 1 SD of academic achievement (AFQT) 12%
10%
8%
6% Increase in earnings 4%
2%
0% 0 1 2 3 4 5 6 7 8 9 10 11 12 13 Work experience (years)
Figure 1. Returns to Observed Educational Quantity and Unobserved Academic Achievement over the Work Life
Notes: Based on data from National Longitudinal Survey of Youth (NLSY) and Armed Forces Qualification Test (AFQT). SD 5 standard deviation. Source: Based on Altonji and Pierret (2001). not have other information and switching to of cognitive skills is not restricted just to observations of skills and performance as that younger workers but holds across the experi- information becomes available through job ence spectrum. performance. On the other hand, Hanushek There are reasons to believe that these esti- and Zhang (2008) do not find that this pattern mates provide a lower bound on the impact of holds for a wider set of countries (although higher achievement. First, the labor market it continues to hold for the United States in experiences that are observed begin in the their data). Thus, there is some uncertainty mid-1980s and extend into the mid-1990s, currently about whether cognitive skills have but other evidence suggests that the value of differential effects on economic outcomes skills and of schooling has grown through- over the work-experience profile. out and past that period. Second, extrapo- Altonji and Pierret (2001) observe only lating from recent patterns, future general a limited age range, so that the chang- improvements in productivity are likely to ing returns which they find may well be lead to larger returns to skill. The considered thought of as leveling off after some amount analyses typically compare workers of dif- of labor market experience. Still, Hanushek ferent ages at one point in time to obtain an and Zhang (2008) find that the importance estimate of how earnings will change for any Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 619 individual. If, however, productivity improve- dents who do better in school, either through ments occur in the economy, these will tend grades or scores on standardized achieve- to raise the earnings of individuals over time. ment tests, tend to go farther in school (see, If recent patterns of skill bias in productiv- for example, Dennis J. Dugan 1976, Charles ity improvement continue, the impact of F. Manski and David A. Wise 1983). Notably, improvements in student skills are likely to Murnane et al. (2000) separate the direct rise over the work life instead of being con- returns to measured skill from the indi- stant as portrayed here (cf. Lawrence F. Katz rect returns of more schooling and suggest and Kevin M. Murphy 1992). On the other that perhaps one-third to one-half of the hand, such skill-biased change has not always full return to higher achievement comes been the case, and technology could push from further schooling. Similarly, Steven G. returns in the opposite direction. Third, Rivkin (1995) finds that variations in test cognitive skills measured by test scores are scores capture a considerable proportion prone to considerable measurement error of the systematic variation in high school as described earlier. Even if the tests were completion and in college continuation, so measuring exactly the relevant skill concept, that test score differences can fully explain we know that there are substantial errors black–white differences in schooling. Bishop in each test.16 These errors will in general (1991) and Hanushek, Rivkin, and Lori L. lead to downward biases in the estimated Taylor (1996), in considering the factors that coefficients. influence school attainment, find that indi- A limited number of additional studies are vidual achievement scores are highly corre- available for developed countries outside of lated with continued school attendance. Neal the United States. Steven McIntosh and Anna and Johnson (1996) in part use the impact of Vignoles (2001) study wages in the United achievement differences of blacks and whites Kingdom and find strong returns to both on school attainment to explain racial dif- numeracy and literacy. Because they look at ferences in incomes. Their point estimates discrete levels of skills, it is difficult to com- of the impact of cognitive skills (AFQT) pare the quantitative magnitudes directly to on earnings and school attendance appear the U.S. work. Ross Finnie and Ronald Meng to be roughly comparable to that found in (2002) and David A. Green and W. Craig Murnane et al. (2000). Jere R. Behrman et Riddell (2003) investigate returns to cogni- al. (1998) find strong achievement effects on tive skills in Canada. Both suggest that lit- both continuation into college and quality eracy has a significant return, but Finnie and of college; moreover, the effects are larger Meng (2002) find an insignificant return to when proper account is taken of the various numeracy. This latter finding stands at odds determinants of achievement. Hanushek and with most other analyses that have empha- Richard R. Pace (1995) find that college com- sized numeracy or math skills. pletion is significantly related to higher test Another part of the return to cognitive scores at the end of high school. Note also skills comes through continuation in school. that the effect of improvements in achieve- There is substantial U.S. evidence that stu- ment on school attainment incorporates
16 In most testing situations, both the reliability of the that would occur if an individual took the same test at dif- specific test and the validity of the test are considered. ferent points in time. Validity refers to the correspondence Reliability relates to how well the test measures the specific between the desired concept (skills related to productivity or material—and would include elements of specific question earnings differences) and the specific choice of test domains development and choice along with individual variations (such as mathematical concepts at some specific level). 620 Journal of Economic Literature, Vol. XLVI (September 2008) concerns about drop out rates. Specifically, Still, even in the absence of convincing higher student achievement keeps students identification strategies for cognitive skills, in school longer, which will lead among other we believe that these reduced form estimates things to higher graduation rates at all levels provide an indication of the potential earn- of schooling.17 ings impacts of policies that lead to improved Tamara Knighton and Patrick Bussière student achievement. The consistency of the (2006) find that higher scores at age fif- estimated impacts across different model teen lead to significantly higher rates of specifications, different time periods, and postsecondary schooling of Canadian nine- different samples provides support for the teen-year-olds. This finding is particularly underlying importance of cognitive skills. At interesting for the international comparisons the same time, the estimation of individual that we consider below, because the analysis earnings models with cognitive skills almost follows up on precisely the international test- always indicates that school attainment has an ing that is used in our analysis of economic independent impact.18 By the development in growth (see section 5 below). The OECD section 2, this finding would imply either that tested random samples of fifteen-year-old the impact of schooling through the noncogni- students across participating countries under tive channel or the measurement errors in cog- the Programme for International Student nitive skills was important. To the extent that Assessment (PISA) in 2000, and students tak- school attainment captures these other dimen- ing these tests in Canada were then followed sions, support for the strength of cognitive and surveyed in 2002 and 2004. skills is increased. More generally, the identifi- The argument on the other side is that none cation of causal impacts of cognitive skills is an of the studies of the impact of cognitive skills important topic for fruitful future research. on either earnings or school attainment have provided any convincing analysis of the causal 3.3 Impacts of Cognitive Skills on impact of scores. By this interpretation, the Individual Incomes—Developing estimated impact of cognitive skills should be Countries considered simply as the reduced form coeffi- cient. Unlike the efforts to identify the rate of Questions remain about whether the clear return to school attainment through a variety impacts of cognitive skills in the United of means, this estimation has stopped without States generalize to other countries, par- providing any substantial evidence that the ticularly developing countries. The existing observed variation in cognitive skills is truly literature on returns to cognitive skills in exogenous. The concern is that other influ- developing countries is restricted to a rela- ences on earnings—say, noncognitive skills or tively limited number of mostly African coun- direct influences of families on earnings—are tries: Ghana, Kenya, Morocco, South Africa, omitted from the estimation, leading to stan- and Tanzania, as well as Pakistan. Moreover, dard omitted variables bias. a number of studies actually employ the
17 This work has not, however, investigated how achievement affects the ultimate outcomes of additional 18 A few previous analyses have shown that the impact schooling. For example, if over time lower-achieving stu- of school attainment is zero or reduced substantially after dents tend increasingly to attend further schooling, these inclusion of cognitive skills (e.g., W. Lee Hansen, Burton schools may be forced to offer more remedial courses, and A. Weisbrod, and William J. Scanlon 1970), but the the variation of what students know and can do at the end review by Bowles, Gintis, and Osborne (2001) indicates of school may expand commensurately. that this is uncommon. Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 621 same basic data, albeit with different ana- Nonetheless, the overall summary is that lytical approaches, but come up with some- the available estimates of the impact of cog- what different results.19 Table 1 provides a nitive skills on outcomes suggest strong eco- simple summary of the quantitative esti- nomic returns within developing countries. mates available for developing countries. The substantial magnitude of the typical The summary of the evidence in estimates indicates that educational quality table 1 permits a tentative conclusion that concerns are very real for developing coun- the returns to cognitive skills may be even tries and that this aspect of schools simply larger in developing countries than in cannot be ignored. developed countries. This of course would Evidence also suggests that cognitive be consistent with the range of estimates skills are directly related to school attain- for returns to quantity of schooling (e.g., ment. In Brazil, a country plagued by high Psacharopoulos 1994 and Psacharopoulos rates of grade repetition and ultimate school and Patrinos 2004), which are frequently dropouts, Ralph W. Harbison and Hanushek interpreted as indicating diminishing mar- (1992) show that higher cognitive skills ginal returns to schooling. (These estimated in primary school lead to lower repetition effects reflect the increase in log earnings rates. Further, Hanushek, Victor Lavy, and associated with a one standard deviation Kohtaro Hitomi (2008) find that lower qual- increase in measured tests; for small changes ity schools, measured by lower value-added in test scores, this estimate is approximately to cognitive achievement, lead to higher the proportionate increase in earnings.) dropout rates in Egyptian primary schools. There are some reasons for caution in Thus, as found for developed countries, the interpreting the precise magnitude of esti- full economic impact of higher educational mates. First, the estimates appear to be quality comes in part through greater school quite sensitive to the estimation methodol- attainment. ogy itself. Both within individual studies and This complementarity of school qual- across studies using the same basic data, the ity and attainment also means that actions results are quite sensitive to the techniques that actually improve quality of schools will employed in uncovering the fundamental yield a bonus in terms of meeting goals for parameter for cognitive skills.20 Second, the attainment. Conversely, simply attempt- evidence on variations within developing ing to expand access and attainment, say countries is not entirely clear. For example, through starting a large number of low Jolliffe (1998) finds little impact of skills quality schools, will be self-defeating to the on farm income, while Behrman, David R. extent that there is a direct reaction to the Ross, and Richard Sabot (2008) suggest an low quality that affects the actual attain- equivalence across sectors at least on theo- ment results. retical grounds. 3.4 Evidence from the International Adult Literacy Survey
19 Unlike in much of the work in developed countries, The preceding analyses in developed coun- these studies collected both earnings information and cognitive skills data at the same time and did not rely on tries rely largely on panel data that follow longitudinal data collections. individuals from school into the labor mar- 20 The sensitivity to estimation approach is not always ket. The alternative approach as found in the the case; see, for example, Dean Jolliffe (1998). A critique and interpretation of the alternative approaches within a International Adult Literacy Survey (IALS) number of these studies can be found in Glewwe (2002). is to test a sample of adults and then to relate 622 Journal of Economic Literature, Vol. XLVI (September 2008)
Table 1 Summary of Estimated Returns to a Standard Deviation Increase in Cognitive Skills
Country Study Estimated Effecta Notes Ghana Glewwe (1996) 0.21**20.3** (government) Alternative estimation approaches yield 0.1420.17 (private) some differences; math effects shown generally more important than reading effects, and all hold even with Raven’s test for ability. Ghana Jolliffe (1998) 0.0520.07* Household income related to average math score with relatively small varia- tion by estimation approach; effect is only observed with off-farm income, and on-farm income is not significantly related to cognitive skills. Ghana Vijverberg (1999) ? Income estimates for math and reading with nonfarm self-employment; highly variable estimates (including both posi- tive and negative effects) but effects not generally statistically significant.
Kenya Boissiere, Knight, and 0.19**20.22** Total sample estimates: small varia- Sabot (1985); Knight and tion by primary and secondary school Sabot (1990) leavers.
Morocco Angrist and Lavy (1997) ? Cannot convert to standardized scores because use indexes of performance; French writing skills appear most important for earnings, but results depend on estimation approach. Pakistan Alderman, Behrman, 0.1220.28* Variation by alternative approaches and Ross, and Sabot (1996) by controls for ability and health; larger and more significant without ability and health controls. Pakistan Behrman, Ross, and Sabot 0.25 Estimates of structural model with (2008) combined scores for cognitive skill; sig- nificant effects of combined math and reading scores which are instrumented by school inputs
South Africa Moll (1998) 0.34**20.48** Depending on estimation method, varying impact of computation; comprehension (not shown) generally insignificant.
Tanzania Boissiere, Knight, and 0.0720.13* Total sample estimates: smaller for Sabot (1985); Knight and primary than secondary school leavers. Sabot (1990)
Notes: * Significant at 0.05 level; ** Significant at 0.01 level. a Estimates indicate proportional increase in wages from a one standard deviation increase in measured test scores. Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 623 these measures to labor market experiences.21 As in the prior analyses, both school attain- An advantage of this data collection approach ment and cognitive skills are seen to enter is that it provides information about the labor into the determination of individual incomes. market experiences across a broader range of With the exception of Poland, literacy scores age and labor market experience.22 have a consistent positive impact on earnings Consistent data on basic skills of literacy (figure 2). The (unweighted) average of the and numeracy for a representative sample impact of literacy scores is 0.093, only slightly of the population aged fifteen to sixty-five less than found previously for the U.S. stud- were collected for a sample of countries ies. The United States is noticeably higher between 1994 and 1998. The analysis here than other countries and the previous U.S. combines the different IALS scores on skills studies, perhaps reflecting that these earnings into a single measure of literacy and numer- are obtained across the entire work life.24 The acy (referred to simply as literacy scores). average excluding the United States is still These data permit direct comparisons of 0.08. Again, the similarity to the prior esti- the relative importance of school attain- mates of the return to cognitive skills, coming ment and cognitive skills across countries, from very different sampling schemes in dif- although the bias toward developed econo- ferent economic markets, lends more support mies remains. Hanushek and Zhang (2008) to the significance of cognitive skills as a con- estimate returns to school attainment and sistent measure of human capital. to literacy scores for the thirteen countries The estimated impact of school attainment where continuous measures of individual across the thirteen countries is 0.059 after earnings are available. Their samples include adjusting the Mincer returns for the literacy full-time workers between twenty-six and scores and down from 0.071 when cognitive sixty-five years of age. The dependent vari- skills are not considered. As noted in sec- able is the logarithm of annual earnings tion 2, however, these estimates are difficult from employment, and control variables are to interpret because of the potential role of gender, potential experience and its square, omitted variables in determining unmeasured and an indicator for living in rural area. cognitive and noncognitive skills. Figures 2 and 3 provide the relevant sum- More interestingly, Hanushek and Zhang mary information on the returns to cogni- (2008) can estimate a Mincer earnings func- tive skills, estimated in a model that jointly tions along the lines of equation (4) where the includes school attainment and literacy scores, measured literacy scores are excluded but the and on the returns to school attainment.23 model is augmented by measures of families,
21 This design was subsequently repeated in 2003 with ferent time periods to be equivalent in quality terms. This the Adult Literacy and Lifeskills Survey (ALL), but only procedure involves equating the marginal impact of a year six countries participated and the data were unavailable of schooling on literacy scores across time (after allow- for this study. The work in developing countries is also ing for other influences on literacy scores). All references more likely to rely upon a single cross section of workers to school attainment here refer to their quality-adjusted who are tested contemporaneously. school attainment. 22 This approach does also present some complications, 24 The previous discussion of the analysis by because the individuals of different ages have both dif- Altonji and Pierret (2001) can reconcile the dif- ferent adult learning experiences and different times of ference in quantitative magnitudes of the impact attending school of possibly different quality. Hanushek of cognitive skills on U.S. earnings. Hanushek and Zhang (2008) consider these alternatives, but they do and Zhang (2008) find that the impact of literacy not change the qualitative results about the impact of cog- scores rises from that for the youngest workers, nitive skills that are presented here. consistent with Altonji and Pierret. They do not, how- 23 An element of the analysis in Hanushek and Zhang ever, find support for this statistical discrimina- (2008) is adjusting the years of schooling obtained in dif- tion hypothesis in the remaining twelve countries. 624 journalof Economic Literature, Vol. XLVI (September2008)
0,26 deviation
standard
per
Increase
XLS. Netherlands Switzerland Chile Finland Germany Hungary Italy Denmark Norway Czech Sweden Poland Rep.
Figure 2. Returns to Cognitive Skills, International Adult Literacy Survey
Source: Hanushek and Zhang (2008).
0.12
csoost Returus to quality-adjusted schooling ecscses Returns accounting for family inputs and ability
0.10
0.08
0.06
0.04 0.02
Chile U.S. Poland Germany Ttaly Czech Switzerland Deumark Hungesy Sweden Norway Netherlands Finland Rep.
Figure 3. Impact of Controlling for Nonschool Taputs on the Estimated Returns to School Attainment, international Adult Literacy Survey
Source: Hanushek and Zhang (2008).
Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 625 school quality, and ability. This permits direct an individual’s cognitive skills yield earn- estimation of the impact of school attainment ings changes similar to those estimated in on wages (conditional on the adequate model- the various models? The experiment we have ing of the factors in equation (2)). As seen from in mind is randomly introducing a different figure 3, this adjustment to the estimated level of cognitive skills across a group of indi- returns to schooling is more significant than viduals and then observing subsequent labor just incorporating literacy scores in the model. market outcomes. In the case of estimation From this figure, it is apparent that the aver- of the returns to school attainment, exten- age returns fall significantly (from 0.071 to sive but highly focused research has delved 0.044) while the variation across countries is into these causality questions (Card 1999). As also lessened considerably. These adjustments sketched in section 2, however, more general are also more significant than is typical in the issues such as systematic variations in school literature concerned with the estimation the quality or other sources of educational out- returns to schooling (Card 1999). comes and skills have not been important The literacy tests in IALS are designed in these analyses. These latter issues are the to measure quite basic skills only, and yet motivation for our work and the reason for the differences are strongly associated with our concentration on cognitive skills as a higher earnings.25 These results, from a broad direct measure of human capital. age spectrum across a number of countries, With cognitive skills, no similar literature reinforce the importance of cognitive skills. on causality questions exists, but the potential The sample of countries for the IALS problems are considerably different. Most of unfortunately contains just one developing the cognitive skills tests used in the devel- country—Chile. Nonetheless, it is suggestive oped country studies, with the exception of that the returns both to quantity of school- the IALS-type sampling, are given at a date ing and cognitive skills exceed those found before the labor market experiences—elim- across the countries with the exception of the inating the reverse causation possibility that United States. In the three transition econo- higher income leads individuals to do things mies (Czech Republic, Hungary, and Poland), that raise their test scores.26 Nevertheless, as the returns to school attainment are also near was developed in section 2, this fact alone the top of the sample, but the returns to cog- does not solve all of the interpretative issues. nitive skills are noticeably lower—perhaps Two other factors threaten the interpretation reflecting institutional aspects of their labor of the estimated effects of cognitive skills: markets. measurement error in tests and other corre- lated but omitted influences on earnings. 3.5 Causality As is well known, test scores inherently For policy purposes, we are interested in have errors in measuring the underlying a rather simple question: Would changing cognitive domains that are being tested. A
25 It should be noted, however, that results from dif- simple correlation of 0.73. This consistency suggests that ferent tests, even when described as measuring dif- the available assessments of cognitive skills may cap- ferent parts of the distribution of cognitive skills, tend ture a wider range of knowledge and skills than would to be highly correlated—particularly at the country be suggested by the descriptions of the individual tests. level. Hanushek and Zhang (2008) correlate the literacy 26 The IALS sampling raises the concern that tests test results from IALS (identified as measuring basic change with age, because of continual learning or simple age skills)with the higher order math test results from the depreciation of skills and knowledge. An investigation of this Third International Mathematics and Science Study by Hanushek and Zhang (2008) suggests that these are not (TIMSS) for comparable age groups in 1995 and find a large concerns in the estimation of the earnings functions. 626 Journal of Economic Literature, Vol. XLVI (September 2008) portion of this is simply that individuals will The implicit inclusion of the correlated por- obtain somewhat different scores if tested tion of the different skills does not lead per again with the same test instrument or with se to large problems as a proxy for overall a different instrument designed to measure labor market skills. It is a more significant the same concepts. This type of error is a problem if we subsequently consider, say, characteristic of the measurement technol- school policies that might operate on just ogy and, in these specific applications, is cognitive skills and not also on the noncog- likely to lead to a standard downward bias nitive skills. in the estimated impact of cognitive skills. As is generally the case, the analyses consid- A second issue is the specificity of the indi- ered here cannot rule out a variety of factors vidual tests employed across the underly- that might bias the estimates and jeopardize ing research; that is, most work relies on the interpretation. Nonetheless, we return specific subject matter tests at a given level to one simple finding: Under a wide range of of difficulty. Again, this is likely to be pres- different labor market conditions, modeling ent quite generally in the various empirical approaches, and samples of individuals, cog- applications. Its impact on the estimates nitive skills are directly related to individual is less clear, because it will depend on the earnings and, moreover, there is a consis- structure of specific tests. On the other tency even in the point estimates. hand, because general cognitive skills tests If we take these estimates as indicating for one subject tend to be very highly cor- a causal relationship, the most important related with the scores in other subjects issue for policy purposes is the source of for individuals, this problem may not be any test score differences. It is natural to overly important. Finally, because indi- believe that schools have an influence on vidual skills may change between the time tests, but clearly other factors also enter. of testing and the period of observed earn- The extensive investigations of the determi- ings, additional error may enter, but its nants of achievement differences indicate implications are unclear in the abstract. that parents, peers, neighborhoods, and More important concerns come from other factors enter along with school factors other omitted factors that might indepen- in determining achievement (see Hanushek dently influence earnings but be corre- 2002). Thus, most importantly, it is inap- lated with cognitive skills. One candidate, propriate to interpret test scores as simply explored early by Bowles and Gintis (1976) reflecting school quality or school policy. In and Bowles, Gintis, and Osborne (2001) and particular, if we can find approaches that extended recently by Heckman, Stixrud, increase skills reliably—whether school and Urzua (2006), is noncognitive skills based or not, the available evidence strongly (see also Cunha et al. 2006). The Heckman, indicates that individual earnings and pro- Stixrud, and Urzua (2006) analysis explores ductivity will also increase. the impact of various measures of noncog- The other side of this issue is also impor- nitive skills (indices of perceived self-worth tant, however. Using just quantity of school- and of one’s control over own life) within ing in the earnings analyses assumes that standard earnings determination models formal schooling is the only source of skill and finds that both cognitive and noncogni- development. But, if a variety of other inputs tive skills influence labor market outcomes. such as families or peers is also important It is less clear, however, how omission of in the formation of human capital, simple these factors might bias the estimates of the years of schooling is subject to this addi- impacts of cognitive skills considered here. tional source of omitted variables bias. Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 627
3.6 Income Distribution Again, these studies do not attempt to describe the causal structure, and it would One implication of the impact of cognitive be inappropriate to attribute the variance in skills on individual earnings is that the distri- earnings simply to differences in the quan- bution of those skills in the economy will have tity or quality of schooling. Nonetheless, to a direct effect on the distribution of income. the extent that these contribute to variations Cognitive skills by themselves do not of course in cognitive skills, it is fair to conclude that determine the full distribution, because other policies aimed at improving school quality factors such as labor market institutions, taxes, (and educational outcomes) will affect the and the like enter. But the importance of skills income distribution. is becoming increasingly evident. Very suggestive evidence on the impact of 4. Quantity of Schooling and skills on the income distribution comes from Economic Growth Stephen Nickell (2004). Nickell, using the Given the microeconomic evidence of the IALS data, considers how differences in the productivity-enhancing effects of education distribution of incomes across countries are and skills, it is natural to extend the view to affected by the distribution of skills and by the macroeconomic perspective of long-run institutional factors including unionization economic growth of countries. Our approach and minimum wages. While union coverage is to the education–growth relationship is statistically significant, he concludes that “the the same as that to the education–earn- bulk of the variation in earnings dispersion is ings relationship. We pursue a simple model generated by skill dispersion” (p. C11).27 that aggregate human capital is relevant to The impact of the skill distribution across growth. Our analysis is designed to compare countries is shown dramatically in figure 4, and contrast simple school attainment mea- which is derived from a comparison of the sures, which have been the near universal dispersion of wages and the dispersion of measure of human capital, with direct inter- prose literacy scores (each measured as the national assessments of cognitive skills. This ratio of the ninetieth to the tenth percentile). section introduces the broad literature based The tight pattern around the regression line on school attainment; the next section pro- reflects a simple correlation of 0.85 (which is vides the contrast with the use of cognitive not affected by including the other institu- skills measures. tional factors). From a theoretical viewpoint, there are at Other studies have also concluded that skills least three mechanisms through which edu- have an increasing impact on the distribution cation may affect economic growth. First, of income (e.g., Chinhui Juhn, Murphy, and just as in the micro perspective, education Brooks Pierce 1993). In the United States, increases the human capital inherent in the the distribution of incomes within schooling labor force, which increases labor produc- groups has been rising (Levy and Murnane tivity and thus transitional growth towards 1992), i.e., holding constant schooling attain- a higher equilibrium level of output (as in ment, the income distribution has become augmented neoclassical growth theories, more dispersed in reflection of growing cf. N. Gregory Mankiw, David Romer, and rewards to individual skills. David N. Weil 1992). Second, education may increase the innovative capacity of the economy, and the new knowledge on new 27 Jose De Gregorio and Jong-Wha Lee (2002) find a (somewhat weaker) positive association between inequal- technologies, products and processes pro- ity in years of schooling and income inequality. motes growth (as in theories of endogenous 628 Journal of Economic Literature, Vol. XLVI (September 2008)
4.5
64" $"/
4.0
3.5 *3& 6,
3.0
Earnings inequality "64
48* /&5 2.5 '*/ (&3 #&- %&/ 48& 2.0 /03
1.5 1.3 1.5 1.7 1.9 Test score inequality
Figure 4. Inequality of Test Scores and Earnings
Notes: Measure of inequality is the ratio of ninth decile to the first decile in both cases; test performance refers to prose literacy in the International Adult Literacy Survey. Source: Based on Nickell (2004). growth, cf., e.g., Robert E. Lucas 1988, Edmund Phelps 1966, Jess Benhabib and Paul M. Romer 1990a, Philippe Aghion Mark M. Spiegel 2005). and Peter Howitt 1998). Third, education 4.1 Results of Initial Cross-Country may facilitate the diffusion and transmis- Growth Regressions sion of knowledge needed to understand and process new information and to imple- Just as in the literature on microeco- ment successfully new technologies devised nomic returns to education, the majority of by others, which again promotes economic the macroeconomic literature on economic growth (cf., e.g., Richard R. Nelson and growth that tries to test these predictions Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 629 employs the quantitative measure of years of cross-country differences in the quality of schooling, now averaged across the labor of education and in the strength of family, force.28 Early studies used adult literacy rates health, and other influences is probably the (e.g., Costas Azariadis and Allan Drazen major drawback of such a quantitative mea- 1990, Romer 1990b) or school enrollment sure of schooling, and we come back to this ratios (e.g., Robert J. Barro 1991, Mankiw, issue in great detail below. Romer, and Weil 1992, Ross Levine and The standard method to estimate the David Renelt 1992) as proxies for the human effect of education on economic growth is capital of an economy. (See Woessmann to estimate cross-country growth regressions 2003 for a survey of measurement and speci- where countries’ average annual growth in fication issues from early growth accounting gross domestic product (GDP) per capita over to current cross-country growth regressions.) several decades is expressed as a function of These were followed by attempts to measure measures of schooling and a set of other vari- average years of schooling based on perpet- ables deemed to be important for economic ual inventory methods (cf. Frederic F. Louat, growth. Following the seminal contributions Dean T. Jamison, and Lawrence J. Lau 1991, by Barro (1991, 1997) and Mankiw, Romer, Vikram Nehru, Eric Swanson, and Ashutosh and Weil (1992),30 a vast early literature of Dubey 1995). An important innovation by cross-country growth regressions has tended Barro and Lee (1993, 2001) was the develop- to find a significant positive association ment of internationally comparable data on between quantitative measures of schooling average years of schooling for a large sample and economic growth (for extensive reviews of countries and years, based on a combina- of the literature, see, e.g., Topel 1999, Temple tion of census or survey data on educational 2001, Krueger and Lindahl 2001, Barbara attainment wherever possible and using lit- Sianesi and John Van Reenen 2003). To give eracy and enrollment data to fill gaps in the an idea of the robustness of this association, census data. in the recent extensive robustness analysis by But, as discussed, using average years of Xavier Sala-i-Martin, Gernot Doppelhofer, schooling as the education measure implicitly and Ronald I. Miller (2004) of sixty-seven assumes that a year of schooling delivers the explanatory variables in growth regres- same increase in knowledge and skills regard- sions on a sample of eighty-eight countries, less of the education system. For example, a primary schooling turns out to be the most year of schooling in Papua New Guinea is robust influence factor (after an East Asian assumed to create the same increase in pro- dummy) on growth in GDP per capita in ductive human capital as a year of schooling 1960–96. in Japan. Additionally, this measure assumes A closely related literature weights years of that formal schooling is the primary (sole) schooling by parameters from microeconomic source of education and, again, that varia- Mincer earnings equations (see section 3.1 tions in nonschool factors have a negligible above) to construct a measure of the stock of effect on education outcomes.29 This neglect human capital, which is then used in growth
28 More precisely, the most commonly used measure is 30 Jonathan Temple and Woessmann (2006) show that average years of schooling in the working-age population, the significantly positive effect of education that Mankiw, usually defined as the population aged fifteen years and Romer, and Weil (1992) find does not depend on their over, instead of the actual labor force. often criticized use of an education flow measure based on 29 If the omitted variables are uncorrelated with enrollment rates, but can be replicated when using years school attainment—an unlikely event—no bias would be of schooling as a measure of the level of human capital in introduced. their model. 630 Journal of Economic Literature, Vol. XLVI (September 2008) accounting and development accounting real GDP per capita in 1960–2000 comes exercises. These use a given macroeconomic from the latest update (version 6.1) of the production function together with parameter Penn World Tables by Alan Heston, Robert estimates from other research to calculate the Summers, and Bettina Aten (2002).32 Figure share of cross-country differences in growth 5 plots the average annual rate of growth in or levels of income which can be accounted GDP per capita over the forty-year period for by cross-country differences in education against years of schooling at the beginning (see Peter J. Klenow and Andres Rodriguez- of the period for a sample of ninety-two Clare 1997 and Robert E. Hall and Charles countries. Both growth and education are I. Jones 1999 for examples). expressed conditional on the initial level of Because this literature has been extensively output, to account for the significant condi- reviewed elsewhere (see references above), tional convergence effect.33 we focus just on a few of the most impor- The regression results depicted by figure 5 tant issues that emerge in the literature with imply that each year of schooling is statisti- respect to the influence of human capital. cally significantly associated with a long-run growth rate that is 0.58 percentage points higher. The association is somewhat lower 4.2 More Recent Evidence on the Effects (at 0.32) but still significant when regional of Levels of and Growth in Years of dummies are added to the regression. The Schooling positive association is substantially larger in the sample of non-OECD countries (at 0.56) To frame the discussion of cognitive skills than in the sample of OECD countries (at that follows, we produce our own estimates 0.26). Alternatively, results based on the sam- of common models that incorporate school ples of countries below the median of initial attainment. These estimates use improved output and above the median are in line with school attainment data and extend the obser- the pattern of larger returns to education in vation period for economic growth through developing countries. 2000. However, after controlling for the institu- We use an extended version of the edu- tional differences reflected in the openness of cation data by Daniel Cohen and Marcelo each country and in the security of property Soto (2007), representing the average years rights, the association with school attainment of schooling of the population aged fifteen 31 to sixty-four. As discussed below, one line 32 Jan Hanousek, Dana Hajkova, and Randall K. Filer of investigation has been the impact of mis- (2004) argue that growth rates are better calculated using measurement of the quantity of education the IMF International Finance Statistics. Our results in this paper are not significantly affected by using these on growth; the Cohen and Soto (2007) data alternative growth measures. improve upon the original quantity data 33 Added-variable plots show the association between by Barro and Lee (1993, 2001). Data on two variables after the influences of other control vari- ables are taken out. Thus, both of the two variables are first regressed on the other controls (in this case, initial- GDP). Only the residual of these regressions, which is the 31 Eliot A. Jamison, Jamison, and Hanushek (2007) part of the variation in the two variables which cannot supplement the Cohen and Soto attainment series with be accounted for by the controls, is used in the graph. In imputed data based on the Barro and Lee series in order so doing, the graph makes sure that the depicted asso- to expand the number of countries available for the growth ciation between the two variables is not driven by the analysis. This approach adds eight countries to the growth control variables. The procedure is numerically equiva- analysis. For details of the extension along with the basic lent to including the other controls in a multivariate data on educational attainment, see Jamison, Jamison, regression of the dependent variable (growth) on the and Hanushek (2006). independent variable under consideration in the graph. Hanushek and Woessmann: The Role of Cognitive Skills in Economic Development 631
6
SGP TWN
4 KOR BWA HKG
THA CYP CHN BRB JPN 2 PRT MYS IRL MUS ROM IDN ESP TUN GRC ASU PAK DOM INDITASYEGTMITA AURITR MUS ISR MAGGABYASURGY ISLIND BEL TUR FRA PAN CERORTITO NLD LSO 0 IRN USARAGRAN COL ZWE DUEASUUWE ASU NPL ASUMEH RY Conditional growth DZA UGA ECU GTM CRIPRGIRGSEHSEHSEA PHL ZAFURIRY PER CHE NZARG JAM BIASTAET IT TZA CIV BENOMR HND BOL 2 MGOTU BDI SEN VEN ZMB NGAMDG MOZ coef 0.581, se 0.095, t 6.10 NIC NER 4