Chapter 11

The Mincer Equation Thirty Years after Schooling, Experience, and Earnings*

ABSTRACT This paper evaluates the empirical performance of the standard M incer earnings equation thirty years after the publication of Schooling, Experience and Earnings. Over this period, there has been a dramatic expansion in micro data and estimation techniques available to labor economists. How does the M incer equation stand in light of these advances in empirical labor economics? Is it time to revise our benchmark model? On the basis of the existing literature and some new empirical estimates, I conclude that the M incer equation remains an accurate benchmark for estimating wage determination equations provided that it is adjusted by 1) including a quartic function in potential experience instead of just a quadratic, 2) allowing for a quadratic term in years of schooling to capture the growing convexity in the relationship between schooling and wages, and 3) allowing for cohort effects to capture the dramatic growth in returns to schooling among cohorts born after 1950.

*: An early version of this paper was prepared for an invited lecture at the Journes de la microconomie applique at Universit Laval, June 8 2000. I would like to thank Bernard Fortin for inviting me to present this lecture. I would also like to thank David Card, Nicole Fortin, Shoshana Grossbard, James Heckman, and an anonymous referee for useful comments, and SSHRC, FCAR, NIHCD (1 R01 HD39921-01) and the Center for Labor Economics (UC Berkeley) for financial support.

11.1. Introduction Thirty years ago, Jacob M incer published his landmark book Schooling, Experience, and Earnings (Jacob M incer, 1974) that had a profound and lasting influence on empirical work in the field of labor economics.1 On the basis of both theoretical and empirical arguments carefully reviewed in this volumes chapters by Barry Chiswick and Solomon Polachek, M incer modeled the natural logarithm of earnings as a function of years of education and years of potential labor market experience (age minus year of schooling minus six). In the most widely used version of M incers "human capital earnings function", log earnings are modelled as the sum of a linear function of years of education and a quadratic function of years of potential experience:2 (1) log y = log y0 + rS + _1X + _2X2 ,

where y is earnings (y0 is the level of earnings of an individual with no education and no experience), S is years of schooling, and X is years of potential labor market experience. Equation (1) has become the "workhorse" of empirical research on earnings determination. It has been estimated on thousands of data sets for a large number of countries and time periods which clearly makes it one of the most widely used model in empirical economics. What explains the popularity of the Mincer equation? One part of the answer lies in the fact that equation (1) is based on a formal model of investment in human capital (see the chapters by Chiswick and Polachek). Another part of the answer is that the M incer equation provides a parsimonious specification that fits the data remarkably well in most contexts. In this regard, the key contribution of Schooling, Experience and Earnings was the introduction of potential

See Sherwin Rosen (1992) and the other chapters in this volume for a survey of Jacob M incer's major contributions to empirical labor economics. 2 Though the human capital earnings function is typically associated with the name and the work of Jacob M incer, other authors also made important contributions to that line of research. In the words of Robert Willis (1986): "..A major development in the literature, initiated by Gary Becker and Barry Chiswick (1966) and carried to full fruition by M incer (1974), sought to use the theory to restrict the functional form of the earnings function ... This work was carried out with such ingenuity, sophistication, and care by M incer (1974) that the resulting function is often referred to as "the" human capital earnings function..." See also Chiswick (this volume) for an excellent survey of the historical and intellectual origin of the human capital earnings function.

experience as a standard regressor in the earnings regression. It was known prior to M incer's work that earnings grew as a (concave) function of age.3 In his early work, M incer (1958) also pointed out that the resulting "age-earnings profile" was steeper for more educated workers than for less educated workers. In other words, log earnings are not a strictly separable function of education and age. There is no such thing as a single rate of return to education but rather a different rate of return for each age group. By contrast, M incer pointed out in Schooling, Experience and Earnings that the experienceearnings profiles were relatively parallel for different education groups. This point can be readily seen in Figure 11.1 that reproduces Mincer's original charts of earnings as a function of both age and potential experience (using data from the 1960 U.S. Census). Introducing potential experience as opposed to age in the earnings equation is, therefore, a parsimonious way of capturing both the shape of the age-earnings profile and the differential slope of the age-earnings profile across education groups. Another advantage of this additively separable model is that, conditional on years of potential experience, there is a single rate of return to education, r, in the labor market. From this point of view, the M incer equation provides the foundation for the large and growing literature that attempts to estimate the causal effect of education on earnings (see David Card, 1999, for a review). It is quite remarkable that, thirty years after Schooling, Experience and Earnings, most studies still tend to estimate earnings regression that are very closely related to equation (1). Though a list of other regressors are typically added to the basic M incer equation, the three key variables in equation (1) still appear in most empirical estimates of earnings regressions. Furthermore, the logarithmic specification for earnings is almost always used in these models. Unless the M incer earnings equation is really a law of earnings determination, it is difficult to believe that, for most data sets, equation (1) is truly the most parsimonious model of earnings determination that would be obtained through careful econometric specification testing. On the one hand, as long as the M incer earnings regression remains a good approximation for the "true" earnings equation, it is quite valuable to keep estimating the same equation for the sake of comparability across studies. If, on the other hand, the Mincer equation is not, or is no longer, a good approximation, we may be getting a quite inaccurate picture of earnings determination from all these studies that all make the same mistakes.
3

See Herman Miller (1960) and Polachek (this volume) for a more extensive discussion.
2

The goal of this paper is to critically reappraise the standard M incer earnings equation thirty years after the publication of Schooling, Experience and Earnings. Since those days, there has been a phenomenal expansion in the computer power and in the number of micro data sets available for empirical labor economic research. Sophisticated parametric and non-parametric procedures are now available to perform careful specification analyses. How does the M incer equation stand in light of all these advances in empirical labor economics? Is it time to revise our benchmark model? This chapter is a limited attempt at answering these specific empirical questions. I only focus on the issue of the robustness of M incers original specification to complement the chapters by Chiswick and Polachek in this volume. Chiswick provides a detailed derivation of the earnings function, while Polachek discusses the implications and extensions of the earnings function. In this chapter, I only review the results of few recent studies that have implications for the choice of an accurate "benchmark" earnings equation la M incer, and supplement the main findings of these studies with some new empirical results based on the Current Population Survey (CPS) for the years 1979 to 2001. I make no attempt to cover the much larger literature on earnings determination that has been well surveyed elsewhere (see, for example, Solomon Polachek and Stanley Siebert (1993)). The remainder of the paper is divided in five themes. In Section 11.2, I look at whether the natural logarithm is the appropriate transformation for earnings. In the following two sections, I discuss whether education should enter linearly (Section 11.3) and potential experience should enter as a quadratic function (Section 11.4) in a separable earnings equation. The issue of separability between schooling and experience is examined in Section 11.5, while Section 11.6 looks at whether cohort effects should be included in the earnings equation. I conclude in Section 11.7. 11.2. Earnings in logs or not?

As mentioned earlier, the dependent variable in the standard M incer earnings equation is the log, as opposed to the level, of earnings. While logs are typically used in econometric models for reasons of convenience or fit, there is a strong theoretical rationale for using log earnings in a human capital earnings regression. As pointed out by M incer (1958), education should have a multiplicative effect on earnings in a simple model where identical individuals maximize the present value of future
3

income which is equalized for all education levels in equilibrium. The reason is that investments in human capital, like other investments, are only undertaken as long as the rate of return (not the absolute return) on the investment exceeds the discount rate. Log-linearity of earnings as a function of years of schooling is in fact a key empirical implication of the human capital model with identical individuals proposed by M incer (1958). The existing evidence generally supports the log-earnings specification. For example, James Heckman and Polachek (1974) estimate a Box-Cox model and could not reject the log specification. M ore recently, Nicole Fortin and Thomas Lemieux (1998) use a more flexible "rank regression" model in which earnings are specified as a relatively unrestricted monotonic transformation of a M incer-type human capital index (sum of a linear function of education, polynomial function of potential experience, few other regressors and a normally distributed error term). Using large samples from the 1979 and 1991 outgoing rotation group (ORG) supplement of the CPS, they find that the log wage is close to a log-linear function of the human capital index for values of the wage above the minimum wage. Their findings are reproduced for men in Figure 10.2.4 The human capital index appears to have a smaller effect on wages around the minimum wage, which is consistent with minimum wages compressing the wage distribution at the low end of the skill distribution.5 Figure 11.2 shows, nevertheless, that the assumption of log-linearity is very accurate for most of the range of the wage distribution. 11.3. Linear education?

Though M incer (1974) considered several functional forms for the earnings equation, the most
4

Fortin and Lemieux (1998) divide the wage distribution into 200 intervals and estimate an ordered probit model for the probability that wages are in a given interval, conditional on the observables (as captured by education, a quartic in experience, and marital status). Figure 11.2 is simply a plot of average log wages in each interval as a function of the estimated thresholds in the ordered probit. If the true model was log(wage) = xb +e, where e is normal (Mincer model with normal errors), then the function plotted in Figure 11.2 should be linear. 5 The function is flatter around the minimum wage because too many people with different skill levels are bunched up at the minimum wage. This suggests that some workers who would have earned less than the minimum in the absence of a minimum wage retained their jobs and are now paid the minimum wage, though they may be working less hours than in absence of a minimum wage. If all of these workers had lost their jobs (perfect disemployment effects), the function
4

commonly used is equation (1). There are several reasons, however, why a simple linear specification for years of education may be inaccurate. For example, log earnings will be a concave function of years of schooling in a simple human capital investment model in which individuals have different preferences (discount rates) but all face the same concave production function (the return to a year of schooling declines as years of schooling increase). M ore generally, M incer (1997) shows that in a Becker (1975) type model where individuals are heterogenous in their preferences and earnings opportunities, average log earnings may either be a convex or a concave function of years of schooling. 6 Another possibility is that in addition to years of schooling per se, educational credentials also have a direct impact on earnings. In the presence of "credential or sheepskin effects, the return to a year of schooling should be higher between 11 and 12 years of education (high school credential effect) and between 15 and 16 years of schooling (college credential effect) than for other years of schooling. One simple way of testing the linearity of years of schooling is to estimate log earnings as an unrestricted (non-parametric) function of years of schooling and see whether this unrestricted function is approximately linear. Card and Alan Krueger (1992) estimate such an unrestricted earnings function as part of their study of the impact of school quality on earnings.7 Their results for three separate birth cohorts of white men (Figure 2, p. 7) show that log earnings appear to be a linear function of years of schooling for most levels of schooling. One exception is between 15 and 16 years of schooling where, consistent with credential effects, the return to schooling is systematically higher than for other schooling levels.8 Interestingly, however, there are no abnormal returns associated with high would simply be truncated at the minimum wage and log-linearity would be preserved. 6 This model was introduced by Becker in his 1967 Woytinski Lecture which is reproduced in Becker (1975). 7 Card and Krueger estimate regression models for log weekly earnings using data for white men from the 1980 U.S. Census. The explanatory variables included in their regressions are a set of dummy variables for each single year of schooling plus standard controls for potential experience, marital status, and geographical location. 8 An alternative explanation to sheepskin or credential effects is selectivity in college completion and graduation. For example, youths entering college may expect a rate of return greater than or equal to their discount rate. During the college experience, some re-assess their expected rate of return downward and are more likely to drop out. Those doing as expected or better than expected go on to graduate. It is not possible to separate this reassessment hypothesis from the sheepskin hypothesis in Figures 11.3 to 11.5.
5

school completion. The other exception is at very low years of schooling where returns to schooling are relatively low. This phenomenon is particularly clear for the younger (and most educated) cohort born in 1940-49 for which returns below five years of schooling are essentially zero. Card and Krueger (1992) conclude from their analysis that log earnings are an approximately linear function of years of schooling except for the lowest two percentiles (for a cohort) of the schooling distribution. Whether or not log earnings is a linear function remains an open debate. For example, using similar data as Card and Krueger (1992), Heckman et al. (1996) conclude that log earnings are a nonlinear function of schooling because of the large increase in earnings between 15 and 16 years of schooling. Perhaps more importantly, however, M incer (1997) and Olivier Deschnes (2001) show that since 1980, log earnings have become an increasingly convex function of years of schooling. In Figures 11.3 to 11.5, I report some new estimates of effect of schooling on male wages for 1979-81, 1989-91, and 1999-2001, respectively. Unlike the aforementioned studies that use weekly earnings from the Census or the CPS, I use hourly wage from the ORG supplements of the CPS as dependent variable in the regression analysis.9 10 For each time period, I estimate an OLS regression of log wages on an unrestricted set of year dummies for schooling, experience, and calendar year. The estimated coefficients on the year of schooling dummies, along with the estimated confidence intervals (at the 95 percent confidence levels), are reported in Figures 11.311.5. The results for 1979-1981 (Figure 11.3) are qualitatively similar to those of Card and
9

Samples include all men age 16 to 64 with 0 to 40 years of potential experience and nonallocated wages between $1 and $100 (in $1979). Top-coded observations are adjusted by multiplying wages or earnings by 1.4. I pool three years of data for each period to increase sample sizes to 220,737 observations in 1979-1981, 211,759 observations in 1989-1991, and 153,265 observations in 1999-2001. 10 I use log hourly wages as dependent variable since Mincers human capital earnings function is a model for the determination of the hourly price of labor. Traditionally, the model has often been estimated using annual or weekly earnings because hourly wages were generally not available. Mincer (1974) in fact shows that some key results are sensitive to the choice of the earnings measure. For example, he rejects the hypothesis of parallel earnings profiles and linear education when using annual earnings, but fails to reject these restrictions when annual hours are controlled for. Note also that most workers report directly an hourly wage rate in the ORG supplement of the CPS. This limits potential measurement problems linked with defining hourly wages as earnings divided by hours. See Lemieux (2005) for more details on these measurement
6

Krueger (1992) in the sense that they are approximately linear for most values of years of schooling.11 On the one hand, the estimated schooling wage relationship does not systematically depart from a linear specification. When a (weighted) quadratic regression is fit to these data, the quadratic term in schooling is positive but not statistically significant.12 On the other hand, the linear approximation is clearly inaccurate for 15 (overpredicts) or 16 (underpredicts) years of schooling, as in Heckman et al. (1996). M ore generally, a goodness-of-fit test strongly rejects the linear or quadratic specification mostly (but not solely) because of the significant departures from linearity at 15 and 16 years of schooling.13 Figure 11.4 reports the same estimates of the schooling-wage relationship for the 1989-91 period. Consistent with M incer (1997) and Deschnes (2001), Figure 11.4 shows that log wages are now clearly a convex function of years of schooling. When a (weighted) quadratic regression is fit to these data, the quadratic term in schooling is positive and statistically significant (t-statistic of 4.9). The magnitude of the estimated quadratic coefficient, 0.0031, is economically substantial. It implies that the return to a single year of education is three percentage points larger at 18 than 8 years of schooling. However, the quadratic specification still underpredicts wages at 15 year of schooling and overpredicts wages at 16 years of schooling. These departures are nonetheless not as visually important as the strong overall convexity in the schooling-wage relationship. Figure 11.5 shows that by 1999-2001, returns to schooling are an even more convex function of years of schooling than they were in 1989-1991. The estimated coefficient on the quadratic term is now 0.0044 with a t-statistic of 5.9. Note also that because of the changes in the education question introduces in the 1992 CPS, it is no longer possible to identify workers with either 15 or 17 year of schooling. As a result, there is no longer much of a systematic departure from a linear or quadratic specification around 16 years of schooling. This helps clarify the earlier point that the issues. 11 The relatively large confidence bands for levels of education of five years or less in Figure 11.311.5 reflect that fact that only few workers have such a low level of education. Immigrants represent a significant fraction of these low-education workers but cannot be identified in the ORG CPS. 12 The weights used are the inverse of the variance of the estimated coefficients from the microlevel wage regression (coefficients on years of schooling dummies). 13 The chi-square test statistic is 355, which largely exceeds the critical value at the 95 percent confidence level (17 degrees of freedom). The statistics is reduced by almost two thirds when the test is limited to years of schooling other than 15 and 16.
7

growing convexity of the schooling-wage relationship, as opposed to local non-linearities around 16 years of schooling, is now the dominant source of non-linearity in these data. M incer (1997) argues that this growing convexity is readily explained by an increase in the relative demand for skilled labor in a Becker (1975) type human capital investment model with heterogenous workers (heterogenous preferences and earnings capacity). This model yields an hedonic equilibrium where marginal returns to schooling can either be an increasing or a decreasing function of years of schooling. In this model, an abrupt growth in relative demand that is not matched by a corresponding increase in relative supply of schooling increases the marginal return to schooling for more educated workers relative to less-educated workers. This results in a more convex schooling-wage relationship since the marginal return to schooling is just the slope of the schooling wage relationship reported in Figures 11.3 to 11.5. This suggests that the linear approximation may only be accurate in a stable environment where the growth in relative demand in matched by a corresponding growth in relative supply. By contrast, studies such as Lawrence Katz and Kevin Murphy (1992) suggests that the post-1980 period has been unstable in the sense that relative supply did not increase enough to match the growth in relative demand. This has apparently resulted in both an increase in the returns to education (Katz and M urphy, 1992) and in the convexity of the schooling wage relationship (M incer, 1997). As it turns out, I will argue in Section VI that changes in relative supply and demand for different labor categories can also explain other recent failures of the most simple version of the M incer equation (equation (1)). In particular, experience profiles are no longer parallel in recent data (Section 11.5) because of systematic cohort effects (Section 11.6) that are linked to a slowdown in the rate of growth of educational attainment for workers born after 1950. 11.4. Quadratic experience?

As mentioned earlier, the most popular version of the M incer equation includes a quadratic function in years of potential experience to capture the fact that on-the-job training investments decline over time in a standard life-cycle human capital model. M incer (1974) shows that the quadratic profile is implied by a model in which investments decline linearly over time. This choice of specification as
8

the preferred empirical specification was mostly driven by practical considerations. It is indeed much easier to estimate a regression model with experience and experience squared included as regressors than to estimate the more complex non-linear experience profile implied by the assumption of a constant rate of decline in investments which is more consistent with economic theory. M urphy and Finis Welch (1990) examine in detail whether the standard quadratic specification in years of potential experience captures well the empirical experience-earnings profile. Using M arch CPS data from 1964 to 1987, they conclude that a quadratic function is not flexible enough to capture the main features of the experience-earnings profile. The main problem is that the quadratic function understates earnings growth over the first 10 to 15 years of career. By contrast, they find that a quartic function in years of experience captures very well the main features of the empirical experience-earnings profiles. Figures 11.6 to 11.8 confirm the findings of M urphy and Welch using more recent CPS data on hourly wage rates. While M urphy and Welch estimate separate experience-wage profiles for four broad education groups, Figures 11.6 to 11.8 simply report estimates of the experience earnings profiles from a pooled specification with all education groups included in. I first estimate regressions with a full set of education, experience, and year dummies for each of the three periods (1979-1981, 1989-1991, and 1999-2001). All three figures show the 95 percent level confidence bands around the unrestricted estimates of the experience-wage profiles. I then re-estimate two restricted versions of the model where the set of experience dummies is either replaced by a quadratic or a quartic function of experience. Consistent with Murphy and Welch (1990), the figures show that the standard quadratic specific is not flexible enough to capture important features of the data. In particular, it systematically understates the growth in wages for the first 10-15 years of experience. As a result, the quadratic predictions are systematically above the confidence band for workers with 3 or less years of experience, and systematically below the confidence bands for workers with about 4 to 14 years of experience. The quadratic also predicts too much of a decline in wages past 25 years of experience. In fact, the confidence bands show that wages are quite stable between 25 and 40 years of experience. The predicted decline in wages in the last 10-15 years of career is an artifact of the quadratic specification.
9

By contrast, Figures 11.6-11.8 show that the quartic specification rarely falls outside of the confidence bands for the unrestricted model. The only noticeable exception is around 35 years of experience in 1989-91 where there is an unusual (likely spurious) drop in the unrestricted confidence bands. In summary, recent research and the new evidence presented in this paper show that it is important to "fine-tune" the standard M incer equation by adding higher order polynomials in potential experience. Otherwise, the standard M incer equation understates wage growth for younger workers. The quadratic specification also predicts a spurious decline in wages among older workers. 11.5. Additive separability in education and experience?

As mentioned earlier, one main empirical innovation in M incer (1974) was the introduction of a concave function of years of potential experience as a parsimonious way of capturing both the (concave) shape of the age-earnings profile and its interaction with schooling. One may wonder, however, whether the empirical relationship identified for data from more than forty years ago (1960 U.S. Census) still holds in more recent data. Figures 11.9 to 11.11 compare the empirical experience-wage profile for male college and high school graduates to the profiles predicted by a M incer model where these two groups are constrained to have the same (quartic) experience-wage profile. The predicted profiles are obtained by fitting a regression to a pooled sample of men with exactly a high school or college degree for the same three time periods as before: 1979-1981, 1989-1991, and 1999-2001. The regressors used are a quartic function of experience, a dummy for college degree, and year dummies. For the sake of simplicity, I only look at college and high school graduates which are the two largest education groups in these samples. At the bottom of each graph, I plot the confidence intervals around the empirical college-high school wage gaps (the difference between the actual college and high school wages in the graph) and compare them to the gap predicted by the M incer equation. They key empirical prediction of the standard M incer model with additively separable experience and education effects is that the collegehigh school gap should be a constant function of experience.
10

Figure 11.9 shows that in 1979-1981, the empirical (or actual) experience profiles for college and high school graduates look approximately parallel. Indeed, for most of the values of experience, the constant college-high school wage gap predicted by the model (0.323) is within or close to the confidence bands around the empirical wage gaps. A closer examination of the evidence indicates, however, that the empirical college high school wage gap tends to grow as a function of experience. For example, the empirical wage gap is significantly lower than predicted by the model around 3-9 years of experience, but higher than predicted around 27-35 years of experience. This systematic pattern is confirmed by a linear regression of the empirical college-high school wage gap on experience. The estimated coefficient (0.0018) is significantly different from zero (t-statistic of 3.1). This implies that the wage gap is 7-8 percentage point larger for workers with 40 years of experience than for young workers with no experience. The results for 1989-1991 (Figure 11.10) and 1999-2001 (Figure 11.11) are markedly different from those for the earlier period. In both periods, the college-high school wage gap now tends to decline as a function of experience. The estimated coefficients in a simple regression of the wage gap on experience are now -0.0020 (t-statistic of -4.8) and -0.0042 (t-statistic of -5.9) for 1989-1991 and 1999-2001, respectively. The results imply that in 1999-2001, the college-high school wage gap is 15-20 percentage points larger for the youngest workers (0 years of experience) than for the oldest workers (40 years of experience). Figure 11.11 indeed shows that the empirical experience-wage profiles are no longer parallel in more recent data. The experience-wage profile has clearly become steeper for high school than college graduates over the last two decades. In summary, the standard M incer equation in which log earnings is an additively separable (no interaction term) function of schooling and potential experience does not fit the recent data nearly as well as it used to do in the 1960 Census. This finding is confirmed by Card and Lemieux (2001a) and Heckman et al. (2003) who conclude that experience-earning profiles are no longer parallel in more recent data. I now look at potential explanations for this change in the next section. 10.6. Cohort effects?

There has always been an important gap between the theory underlying the human capital earnings function and the data used to estimate it. On the theoretical side, the age-earnings profile pertains to
11

the evolution of the earnings of a given individual (or cohort of individuals) over the life-cycle. By contrast, empirical age-earnings are typically based on a cross-section of individuals at different points in their life-cycle. M incer was well aware of this problem in his early work (M incer, 1958), and conjectured that the cross-sectional age-earnings profiles were probably understating life-cycle earnings growth since, in those days, there was substantial secular growth in average earnings.14 Interestingly, many researchers now believe that the cross-sectional age-earnings profile overstates life-cycle earnings growth. For example, Paul Beaudry and David Green (2000) conclude in their recent study on the earnings of Canadian men that the slope of the cross-sectional ageearnings profile has increased because new cohorts of men are entering the labor market at increasingly low earnings level. This highlights the well known problem of distinguishing between age, cohort, and year effects in earnings growth. Since these three variables are linear combinations of each other, it is difficult to identify separately the effect of these three variables on earnings even when repeated cross sections are available.15 Another way of thinking about the problem is to simply ask whether one "needs" cohort effects to "fit" the empirical relationship among schooling, experience and earnings. On the one hand, M incer (1974) analyzed this issue in detail and correctly concluded that, on the basis of the 1960 Census data, it was not necessary to include cohort effects to fit these data. On the other hand, the standard M incer equation does not fit very well the more recent data for the 1980s and 1990s. Could the addition of cohort effects help improve the fit of the model? Card and Lemieux (2001a) investigate this issue in detail while trying to understand the sources of change in the college-high wage gap for the United States, Canada, and the United Kingdom over the last three decades. In the case of U.S. men, they use the 1960 Census and a large number of March CPS supplements to study the evolution of the college-high school wage gaps by age groups (5-year age groups) between 1959 and 1996. For the sake of comparability with the other figures presented in the paper, I show the same college-high school wage gaps using hourly

14

M incer (1974) examines this issue in more detail and concludes that, apart from understating earnings growth rates by a constant, cross-sectional age-earnings profiles are a good approximation of the life-cycle profiles. 15 Examples of recent studies attempting to separate cohort from age and year effects are Thomas M aCurdy and Thomas M roz (1991) for the United States, Amanda Gosling, Stephen M achin and Costas M eghir (1999) for the United Kingdom, and Beaudry and Green (2000) for Canada.
12

wages from the M arch CPS for the years 1975-76, 1979-81, 1984-86, 1989-91 and 1994-96.16 The wage gaps are reported in Figure 11.12. The figure shows how the age profile of the college-high school wage gap evolved from a positively sloped profile in the mid-1970s, which is consistent with Mincer's model, to an essentially flat profile in the mid-1990s, which is inconsistent with M incer's model.17 Figure 11.12 strongly suggests that the profile became increasingly flatter because cohorts of men born after 1950 experienced increasingly high returns to schooling relative to older workers. Between 1975-75 and 1979-81, the entire profile of the college-high school gap shifted down, with the exception of the youngest age group (age 26-30 in 1980 i.e. born in 1950-54) for whom the gap remained constant. By the mid-1980s the gaps for older workers were back to their levels of the mid-1970s, but the gaps for the two youngest age groups were much higher. M oving to 1989-91, the gaps for the three youngest age groups were substantially higher than those in the mid-1970s, while those for other workers were not too different. Finally, in the mid-1990s, the gaps for the four youngest age groups were well above the levels of the mid-1970s, but the gaps for older workers were still comparable to those 20 years earlier. Except for these four youngest cohorts, the data in Figure 11.12 are roughly consistent with the M incer equation (positively sloped profile). In fact, Card and Lemieux (2001a) show more formally that a simple Mincer model explains well the data from 1959 to 1995 provided that cohort dummies are introduced for the four youngest cohorts. One promising avenue for understanding why the M incer equation does not fit the recent data as well is, therefore, to understand what happened to the post-1950 cohorts. Card and Lemieux (2001a) investigate this issue in detail and conclude that the dramatic increase in the college-high school wage gap for post-1950 cohorts is primarily a consequence of an equally dramatic break in inter-cohort trends in educational achievement. They show that, until
16

Most of the empirical analysis in Card and Lemieux (2001a) is based on the weekly earnings of full-time workers. Starting with the March 1976 CPS (earnings year 1975), however, it is possible to compute an hourly wage rate by dividing annual earnings by annual hours. Card and Lemieux present a brief analysis of these data (Table 5, p. 731) and find results very similar to the main results based on weekly earnings of full-time workers. 17 When log earnings are the same concave function of potential experience for high school and college graduates, the college-high school wage gap must increase with age. To see this, replace the quadratic function by a generic concave function f(.) in equation (1): log y = log y0 + rS + f(age-S-6). The return to education is given by log y/S = r -f '(age-S-6). The return to education grows as a
13

1950, each successive cohort of men was more educated than the preceding cohort. This secular trend abruptly stopped with men born around 1950. As a result, men born in the late 1960s are no more educated than those born 20 years earlier.18 The relative supply of highly educated workers is, therefore, much lower for post-1950 cohorts that what would have been predicted on the basis of the pre-1950 trend. As long as age groups are imperfect substitutes for each others, this relative decline in the relative supply for post-1950 cohorts should result in a higher college-high school wage gap for these cohorts in a standard supply and demand model of the labor market.19 The results of Card and Lemieux (2001a) can be viewed as a complement, rather than a substitute, to the basic human capital earnings model of M incer. Until the mid-1970s, there was a smooth improvement in the level of schooling of each successive cohort of young men entering the labor market. This increase in the relative supply of highly educated labor was more or less enough to offset the increasing relative demand due, for instance, to skill biased technical change. This changed after 1975 with the entry of the post-1950 cohorts that were no more educated, and in some case less educated, than previous cohorts. Relative supply could no longer offset the rising demand for skilled young workers and, as a result, the college-high school wage gap started expanding dramatically for young workers. In summary, recent evidence suggests that the basic M incer human capital earnings model remains a parsimonious and accurate model in a stable environment where educational achievement grows smoothly across cohorts. Such an environment prevailed in the data analyzed by M incer (1974) who correctly concluded that it was not necessary to control for cohort effects in an earnings regression. In a less stable environment, however, major shifts in the relative supply of different age-education groups can induce important changes in the structure of wages that have to be taken into account when estimating a standard M incer equation. This can either be achieved by adding cohort effects to a standard M incer equation, or by explicitly modeling the relative supply and demand for different groups of workers.

function of age since 2log y/Sage = -f ''(age-S-6), which is positive when f(.) is concave (f ''<0). 18 See Card and Lemieux (2001b) for a detailed analysis of inter-cohort trends in educational achievement. 19 Welch (1979) shows that cohort relative supplies have a negative and significant effect on wages, which is consistent with imperfect substitution among age groups in production.
14

11.7. Conclusions Two broad kinds of conclusions can be drawn from this paper. A first conclusion has to do with "fine-tuning" of the simple M incer human capital earnings function. Though equation (1) is a good approximation in many cases, it may overstate or understate the effect of experience and schooling on earnings for some groups. In particular, it tends to understate the effect of experience on the earnings of young workers. This problem is readily fixed by adding higher order polynomials (up to a quartic) of potential experience to the basic model. The model also tends to overstate the effect of skills (either experience or schooling) on earnings at the very low end of the skill distributions, because, for example, of compression effects associated with the minimum wage. A more substantial set of concerns is that the basic M incer human capital earnings function does not appear to fit the data nearly as well in the 1980s and 1990s as it did in the 1960s and 1970s. The first problem is that wages are an increasingly convex function of years of schooling. The second difficulty is that experience-wage profiles are no longer parallel for different education groups. In particular, the college-high school wage is now much larger for less experienced than more experienced workers. Existing research suggests, however, that these departures from the simple M incer model are primarily a consequence of the fact that recent increases in the relative supply of educated labor have not kept up with the growth in relative demand. In particular, the educational achievement of men born in the 1950s and 1960s is no larger than for men born at the end of the 1940s. On balance, the evidence suggests that the Mincer equation remains a useful and accurate benchmark in a stable environment where educational achievement grows smoothly over successive cohorts of workers, as it did in the time period originally studied by M incer (1974). Thirty years later, the M incer equation does not fit the data quite as well as it used to because we are no longer in a stable environment where educational achievement grows smoothly over successive cohorts of workers. In fact, it would have been truly remarkable if the dramatic changes in the wage structure that happened since the late 1970s had had no impact on the quality of the fit of the M incer equation. It will be interesting to see whether the fit of the M incer equation improves in the next decade or so, now that the wage structure appears to be stabilizing again (Card and John DiNardo, 2002).
15

In terms of econometric practice, these findings suggest that it is important to verify the robustness of the standard M incer equation to the inclusion of a quadratic term in years of schooling and cohort effects. These considerations aside, the M incer human capital earnings function remains a parsimonious and relatively accurate way of modelling the relationship between earnings, schooling and experience. Its status as the "workhorse" of empirical labor economic research on earnings determination is well deserved.

16

REFERENCES Beaudry, Paul, and David Green (2000), Cohort Patterns in Earnings and the Skill-Biased Technical Change Hypothesis, Canadian Journal of Economics 33, November, 907-936. Becker, Gary S. (1975), Human capital (2nd Edition), University of Chicago Press. Becker, Gary S., and Barry Chiswick (1966) Education and the Distribution of Earnings, American Economic Review 56, M ay, 358-369. Card, David (1999), The Causal Effect of Education on Earnings, in Orley Ashenfelter and David Card, editors, Handbook of Labor Economics, North Holland: Amsterdam and New York. Card, David, and John DiNardo (2002), Skill-Biased Technological Change and Rising Wage Inequality: Some Problems and Puzzles, Journal of Labor Economics 20, October, 733-783 Card, David and Alan Krueger (1992), Does School Quality M atter? Returns to Education and the Characteristics of Public Schools in the United States, Journal of Political Economy 100, February, 1-40. Card, David, and Thomas Lemieux (2001a), Can Falling Supply Explain the Rising Return to College for Younger M en? A Cohort-Based Analysis, Quarterly Journal of Economics 116, May 2001, 705-46. Card, David, and Thomas Lemieux (2001b), Dropout and Enrollment Trends in the Post War Period: What Went Wrong in the 1970s? in J. Gruber (ed.) An Economic Analysis of Risky Behavior Among Youth, Chicago: University of Chicago Press for NBER, pp. 439-482. Chiswick, Barry (this volume), Jacob Mincer, Experience, and the Distribution of Earnings Deschnes, Olivier (2001), Unobserved Ability, Comparative Advantage and the Rising Return to Education in the United States: A Cohort-Based Approach, Princeton University Industrial Relations Section Working Paper No. 465, August. Fortin, Nicole M ., and Thomas Lemieux (1998), Rank Regressions, Wage Distributions, and the Gender Gap, Journal of Human Resources 33, Summer, 610-643. Gosling, Amanda, Stephen M achin, and Costas M eghir (2000), The Changing Distribution of M ale Wages in the U.K., Review of Economic Studies 67, October, 635-66. Heckman, James J., Anne Layne-Farrar, and Petra E. Todd (1996) Human Capital Pricing Equations with an Application to Estimating the Effect of Schooling Quality on Earnings, Review of Economics and Statistics 78, November, 562-610. Heckman, James J., Lance J. Lochner, Petra E. Todd (2003) Fifty Years of M incer Earnings
17

Regressions, National Bureau of Economic Research Working Paper No. 9732, M ay. Heckman, James J., and Solomon Polachek (1974), Empirical Evidence on the Functional Form of the Earnings-Schooling Relationship, Journal of the American Statistical Association 69, 350-354. Katz, Lawrence, and Kevin Murphy (1992) Changes in Relative Wages, 1963-1987: Supply and Demand Factors, Quarterly Journal of Economics 107, January, 35-78 Lemieux, Thomas (2005), Increasing Residual Wage Inequality: Composition Effects, Noisy Data, or Rising Demand for Skill? University of British Columbia mimeo. M aCurdy, Thomas, and Thomas M roz (1991), Estimating M acroeconomic Effects on Wages from Cohort Specifications, Stanford University mimeo. M iller, Herman P. (1966), Income distribution in the United States, Bureau of the Census, U.S. Department of Commerce: Washington, D.C. M incer, Jacob (1958), Investment in Human Capital and Personal Income Distribution, Journal of Political Economy 66, August, 281-302. M incer, Jacob (1974), Schooling, Experience and Earnings, Columbia University Press: New York. M incer, Jacob (1997), Changes in Wage Inequality, 1970-1990, Research in Labor Economics 16, 1-18. M urphy, Kevin M . and Finis Welch (1990), Empirical Age-Earnings Profiles, Journal of Labor Economics 8, 202-229. Polachek, Solomon W. (this volume), Proving Mincer Right: Mincer's Overtaking Point and the Lifecycle Earnings Distribution Polachek, Solomon W., and W. Stanley Siebert (1993), The Economics of Earnings, Cambridge: Cambridge University Press. Rosen, Sherwin (1992), Distinguished Fellow: M incering Labor Economics, Journal of Economic Perspective 6, Spring, 157-70. Welch, Finis (1979), Effects of Cohort Size on Earnings: The Baby Boom Babies' Financial Bust, Journal of Political Economy 87, S65-S98. Willis, Robert J. (1986) Wage Determinants: A Survey and Reinterpretation of Human Capital Earnings Functions, in Orley Ashenfelter and Richard Layard, editors, Handbook of Labor Economics, North Holland: Amsterdam and New York.

18

Figure 11.1

Source: Jacob Mincer, Schooling, Experience, and Earnings

Figure 11.2: Transformation Model for Log Wages: Men in the 1979 and 1991 CPS

3.5

3.0 1979

Linear fit above the minimum wage: ----

2.5

Log Wage

2.0

1991

1.5 1979 minimum 1991 minimum wage

1.0

0.5

0.0
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5

-3.5

-3.0

Normalized Skill Index (xb+e, e~N(0,1))

Figure 11.3: Return to Single Year of Schooling, 1979-81 CPS

1.4

1.2

1.0

0.8

0.6

Log Wage Differential

0.4

0.2

0.0

-0.2 2 3 4 5 6 7 8 9 10 11 Year of Completed Schooling 12 13 14 15 16 17 18

Figure 11.4: Return to Single Year of Schooling, 1989-91 CPS

1.4

1.2

1.0

0.8

0.6

Log Wage Differential

0.4

0.2

0.0

-0.2 2 3 4 5 6 7 8 9 10 11 Years of Completed Schooling 12 13 14 15 16 17 18

Figure 11.5: Return to Single Year of Schooling, 1999-2001 CPS

1.4

1.2

1.0

0.8

0.6

Log Wage Differential

0.4

0.2

0.0

-0.2 2 3 4 5 6 7 8 9 10 11 Years of Completed Schooling 12 13 14 15 16 17 18

Figure 11.6: Experience Profiles for Men, 1979-81 CPS

1.0 Confidence band on unrestricted profile

0.9

0.8

0.7

0.6

0.5

0.4

Log Wage Differential

0.3

0.2 Quadratic fit: Quartic fit: -----

0.1

0.0

-0.1 5 10 15 20 25 Years of Potential Experience 30 35 40

Figure 11.7: Experience Profiles for Men, 1989-91 CPS

1.0

0.9

Confidence band on unrestricted profile

0.8

0.7

0.6

0.5

0.4

Log Wage Differential

0.3

0.2 Quadratic fit: Quartic fit: ------

0.1

0.0

-0.1 5 10 15 20 25 Years of Potential Experience 30 35 40

Figure 11.8: Experience Profiles for Men, 1999-2001 CPS

1.0

0.9

0.8

Confidence band on unrestricted profile

0.7

0.6

0.5

0.4

Log Wage Differential

0.3

0.2 Quadratic fit: Quartic fit: ------

0.1

0.0

-0.1 5 10 15 20 25 Years of Potential Experience 30 35 40

Figure 11.9: Experience Profiles and Wage Gap for College and High School Graduates, 1979-1981 CPS

2.5

College

2.0

1.5

Actual: Fitted: -----

High School

Log Wage (or Wage Gap)

1.0

Confidence band on wage gap

0.5

Fitted gap 5 10 15 20 25 Years of Potential Experience 30 35 40

0.0

Figure 11.10: Experience profiles and Wage Gap for College and High School Graduates, 1989-1991 CPS

2.5

College

2.0

1.5

Actual: Fitted: ------

High School

Log Wage (or Wage Gap)

1.0

0.5 Confidence band on wage gap 5 10 15 20 25 Years of Potential Experience 30 35 40

Fitted gap

0.0

Figure 11.11: Experience Profiles and Wage Gap for College and High School Graduates, 1999-2001 CPS

2.5

College

2.0

1.5

Actual: Fitted: ----High School

Log Wage (or Wage Gap)

1.0

0.5 Fitted gap Confidence band on wage gap 10 15 20 25 Years of Potential Experience 30 35 40

0.0 5

Figure 11.12: College-High School Wage Gap by Age Group and Year

0.50

0.45

1994-96

0.40

0.35

1989-91

0.30

0.25

1984-86

0.20

1974-76 1979-81

College-high school wage gap

0.15

0.10

0.05

0.00 31-35 36-40 41-45 Age group 46-50 51-55 56-60

26-30