Kermit Daniel

The Wharton School of the University of Pennsylvania Philadelphia, PA 19104ñ6372

Dan Black

Department of Economics University of Kentucky Lexington, KY 40506ñ0034

Jeffrey Smith

Department of Economics University of Western Ontario London, Ontario N6A 5C2

Kermit Daniel received support under the Education Research and Development Center program, agreement number R117Q00011-91, CFDA 84.117Q, as administered by the Office of Educational Research and Improvement, U. S. Department of Education. The findings and opinions expressed in this paper do not necessarily reflect the position or policies of the Office of Educational Research and Improvement or the U.S. Department of Education. Dan Black completed some of this work while visiting the Department of Economics at the Australian National University; he thanks them for their generous support. Linda Bornyasz, Dan Polsky, and Ann Rickert provided excellent research assistance. The authors wish to thank Kathryn Neckerman and participants in the Labor Economics workshop at Columbia University and in the Public Policy seminar at Wharton for helpful comments.

Using the rich data from the National Longitudinal Survey of Youth, we show that several dimensions of college quality have positive impacts on young womenís wages. We find evidence of ability sorting, but controlling for ability, women who attend higher quality colleges earn higher wages. Women receive smaller gains from college quality than do men; black women receive greater gains from college quality than do white women. Controlling for quality, women who attend private colleges earn more than those who attend public colleges, and women earn lower wages, the higher the proportion of their collegeís students were women. A significant part of the return to college quality appears to arise from a greater likelihood of working in high-wage occupations and industries.

The college education of children represents a major investment. With tuition costs at elite colleges now in excess of $20,000 annually, parents and children must make a major financial commitment if their children are to attend such a school. For all the importance of this decision, little is known about the decision process and the subsequent impacts this decision has on students. Moreover, most of the research in this area has focused exclusively on the impact of college quality on male earnings.

In this paper, we begin to redress this imbalance by examining the impact of college quality on the wages of young women. Only Morgan and Duncan (1979), Sweetman (1994), and Behrman et al. (1995) consider the impact of college quality on the wages of women. Their results are mixed: Morgan and Duncan (1979) find little evidence that college quality increases the wages of women, while Sweetman (1994) and Behrman et al. (1995) find evidence that college quality does increase the wages of women. Controlling for ability, we find evidence of a modest, positive effect of college quality on the wages of young women. This positive effect is substantially larger for black women than for white women, and appears to operate in large part through the effect of college quality on the occupational distribution of women. We also find that women who attend colleges with a higher fraction of female students tend to earn lower wages than those who attend colleges with relatively many male students.

College selection involves an array of factors, ranging from familiesí financial situation, to studentsí intellectual abilities, to the perceived benefits of attending various colleges. These and other factors influence both studentsí choice of college and collegesí choice of students. An economist looking at the consequences of these decisions can hardly ignore the non-random nature of this process and the potential for selection bias, especially ability bias. Ability bias may have been a problem for Morgan and Duncan (1979), who used the 1974 Panel Study of Income Dynamics. While the PSID contains rich data on individual and family background characteristics, it has only limited information on individualsí intellectual abilities. Behrman et al. (1995) estimate college quality effects among female twins born in Minnesota between 1936 and 1955. While twins provide an excellent opportunity to evaluate the role of unobserved endowment differences, they are an unusual segment of the population. The educational experiences of twins, especially those who attend different colleges, may be atypical.

In contrast to these earlier studies, we use unusually rich data from the National Longitudinal Survey of Youth (NLSY) to estimate quality effects from detailed wage regressions that control for previous labor market experiences, family and other background characteristics, and characteristics of respondentsí high schools. The data also include a very good ability measure: individual scores on the Armed Services Vocational Aptitude Battery (ASVAB). The NLS High School Class of 1972 (NLSHS72) data used by Sweetman (1994) also include information about individualsí intellectual abilities, high school, and family background, but the NLSY contains much better controls for the labor market experiences of workers. In addition, our data cover the most recent cohorts of young women currently available. This may be important because the pattern of results from research on high school quality suggests that the link between school quality and wages has weakened over time at the secondary school level (Betts 1994).

College quality may affect wages directly, but may also act indirectly through labor force participation, the student's likelihood of attending graduate or professional school, where she lives, her industry of employment, or her marital status and the attributes of her spouse. If college quality affects any of these outcomes, then controlling for them will bias the estimated returns to college quality. Ideally, we would like to estimate a structural model that accounted for the interrelationship among these variables. Our goal for this paper, however, is more modest: we seek to establish the reduced-form relationship between the quality of college education received and female wages. We begin by reporting baseline estimates obtained by regressing log wages on a constant and individual college characteristics, with and without controls for ability. These estimates are unbiased only if all factors correlated with wages are orthogonal to college quality.

While our measures of college quality are often highly significant when we enter them individually in a wage equation, when we enter them collectively most are statistically insignificant. This reflects a high degree of correlation among the measures. We exploit this collinearity to construct a general college quality index, which we substitute for the individual college characteristics in our wage regressions. We then explore the robustness of our basic finding that women who attend higher quality colleges have higher wages, and in the process determine the relationship between wages and several college characteristics after controlling for college quality. We conclude with an empirical assessment of the plausibility of our quality effect estimates.

Our main findings are easily summarized. First, we find evidence of a positive return to college quality for young women, but this evidence is less robust than that reported for men. Without college major and industry controls 12 of the 14 quality indicators have positive coefficients statistically significant at 5 percent, but once we control for college major and industry of employment, only 4 of our 14 quality indicators produce statistically significant positive effects on the wages of women. In contrast, Daniel, Black, and Smith [DBS] (1995) report a robust relationship between many measures of college quality and young menís wages.

Second, we find that the return to college quality appears to be larger among black women than among white women. This is consistent with what Loury and Garman (1995) and DBS (1995) found for men.

Third, we find some evidence that conditional on college quality and womenís abilities, attending a private college increases wages for women. Behrman et al. (1995) also find that women who attend private colleges earn more. In contrast, Sweetman (1994) finds no evidence that private colleges increase the wages of female students, and DBS (1995) and Solmon (1975) find no evidence among men of a wage premium for attending a private college once one controls for college characteristics.

In addition, we find that white women are more likely to work if they attended higher quality colleges, but accounting for the participation decision when estimating wages has no effect on the estimated impact of college quality on wages. Even after controlling for college quality, college major, industry of employment, family background, and high school quality, women who attended colleges with a higher fraction of female students earn lower wages. Finally, we find that womenís college quality is unrelated to husbandsí earnings.

DATA

Our data come from three sources. Our primary source is the National Longitudinal Survey of Youth (NLSY), a panel data set based on annual surveys of a sample of men and women who were 14ñ21 years old on January 1, 1979. Respondents were first interviewed in 1979 and have been re-interviewed each year since then. Of the five subsamples that comprise the NLSY, we use only the representative cross-section and the minority oversamples. After deleting observations with missing or inconsistent data, we are left with a sample of about 3,000 women.

The NLSY provides unusually detailed information about respondentsí employment histories, including detailed information about wages, total labor market experience, and tenure with each employer. It also provides the identity of colleges respondents attended, thereby allowing us to match college characteristics to respondents.

Our sources for college characteristics are the Department of Educationís Integrated Post-secondary Education Data System (IPEDS) for 1990 and the US News and World Reportís Directory of Colleges and Universities (1991). The former source provides most of our information about the physical and financial characteristics of colleges; the latter provides most of our information about student body composition. We only included information for four-year colleges; roughly one half of the women in our sample with some type of post-secondary education attended a four-year college. The variables are described more fully in the Appendix.

Whenever necessary, we redefine the raw college characteristics so that larger values correspond to obvious notions of quality. For example, we transform the acceptance rate reported in the raw data into a rejection rateóreflecting the assumption that more selective colleges are of higher qualityóbut leave size untransformed because larger colleges may realize gains from specialization. Of course, each variable is an imperfect indicator of quality. For example, observed rejections are not the correct measure of selectivity when applying is costly at the margin. One can raise similar concerns about the other measures, but there are no compelling reasons to believe the characteristics as redefined are inversely related to quality. Larger values should correspond to higher quality.

ESTIMATION RESULTS FOR COLLEGE QUALITY

In this section we report estimates of the effect of college quality on wages. Except where noted, the dependent variable in each regression is the natural log of the individualís real wage in 1987, and the estimated coefficients are from OLS regressions on the set of working respondents not enrolled in school. Using Koenkerís (1981) robust version of Breusch and Paganís (1979) test, we easily reject the null hypothesis of homoskedastic errors in our baseline quality index regression, described below in the section ìCollege Quality Index in Wage Regressions.î Therefore, we estimate covariance matrices using Whiteís (1980) heteroskedasticity-consistent estimator.

We estimate standard Mincer wage equations of the form:

(1)

where c is a constant, Q is a vector of college characteristics (often a single characteristic), A is our set of ability controls, and X is a standard set of wage regressors. The Z_J, , are vectors of background characteristics, where Z_H describes the early home environment, Z_P describes the respondentís parents, and Z_HS describes the respondentís high school. The residual captures the effect of all omitted influences. The regressors are explained in more detail later.

Baseline Quality Regressions

Table 1 reports the coefficients on college characteristics in wage regressions that exclude X and Z_J. These are the correct, possibly reduced-form, measures of the impact of college quality only if the college quality indicators are uncorrelated with all the omitted wage determinants or if quality determines those omitted variables for which the correlation is not zero. Neither situation is likely, but these estimates are a convenient way to present the correlations in the raw data.

Each college characteristic appears in two wage regressions. The first includes a single college characteristic, the post-secondary school indicator, and a constant; the second adds controls for ability (A). If the college characteristic was missing for a respondent, it was set to zero; all regressions include a dummy variable indicating when the college characteristic was missing. The ability controls are based on the Armed Services Vocational Aptitude Battery (ASVAB). The ability controls are the first two principal components of respondentsí age-adjusted scores on the 10 exams that comprise the ASVAB, and these two variables squared. We describe the ASVAB and our age adjustment procedure in the Appendix.

These estimates reveal evidence of ability sorting; for almost all characteristics the point estimates are smaller when ability controls are present. Without ability controls, all 13 of the characteristics with quality interpretations have positive coefficients that are statistically significant at 10 percent or better. All but the SAT interquartile ratio (the ratio of the 25th to 75th percentile SAT scores)óa measure of the dispersion of student abilitiesóare significant at the 5 percent level. None of the characteristics alone explain more than 8 percent of the variance in the log wage.

When ability controls are added to the regressions the point estimates decline, but with two exceptionsósize and the SAT interquartile ratioóthe estimates remain statistically significant. High ability women attend better colleges, but by almost any of our measures, conditional on ability, women who attended higher quality colleges earn higher wages.

College Characteristics in Wage Regressions

This section reports estimates obtained from wage regressions that include X and Z_J. At a minimum, each regression includes a constant, quartics in four variablesóage, tenure, pre-college-graduation labor market experience, and post-graduation experienceócontrols for race, region, urban residence, any post-secondary school completed, years of school completed, years of post-secondary school completed, having a 4-year degree, and the ability controls described earlier. These variables are summarized in the Appendix. We separated pre- and post- graduation experience because college quality might be correlated with forfeiting occupation- or industry-specific human capital acquired from pre-graduation work.

In each of the tables we report results from specifications that include only the basic set of regressors described above, as well as those that add college major and/or industry and union status indicators. If better colleges produce their effect by channeling students into more remunerative majors or into jobs in high wage industries, then the basic specification provides the correct reduced form estimates.

Table 2 reports coefficients for the 13 college characteristics with quality interpretations and for percent students female. All of the point estimates for the 13 quality indicators are positive, except those for the faculty-student ratio (which are never statistically significant), but the statistical significance of the estimates is not robust across specifications. Several fail to achieve standard levels of statistical significance when controls for college major and industry are present.

Because the aggregate SAT measuresó25th and 75th percentile SAT, and mean SATómight proxy for individual scores, regressions that include aggregate SAT measures also include respondentsí combined verbal and math SAT scores (see Table 2, page 25, lower panels, and accompanying note). Whether or not individual SAT scores are present makes essentially no difference.

Table 2 also reports results from regressions that include the percent of students who are female. Some have argued that students, especially female students, who attend colleges with their own gender will learn more (for example, see Tidball 1989). We find no evidence of this for women. When percent female enters linearly, the point estimates are negative and not statistically significant. When percent female is captured by quartile dummy variables, the negative effect is large and statistically significant for the top two quartiles. The estimated negative effect for women in the top quartile is about the same as that for men, as reported in DBS (1995).

In Table 3 we explore the possibility that among women, tuition, size, spending per student, or the faculty/student ratio are proxying for whether the college is public or private. We find evidence of a private college premium. When we control for spending per student, size, or the faculty/student ratio, we find that women who attended private colleges earn wages about 6 to 8 percent higher than otherwise comparable women who attended public colleges. We found no evidence of a positive private college effect for men (see DBS 1995). When the private indicator is present, size, spending per student, and the faculty/student ratio all fail to achieve statistical significance at the 10 percent level in any specification.

College Quality Index in Wage Regressions

In our sample, characteristics are strongly correlated across colleges, a feature noted by others as well (Solmon 1975; Morgan and Duncan 1979; James et al. 1988). This high correlation makes precise estimation difficult. For this reason, and to simplify the interpretation of our empirical results, we constructed an overall ìquality indexî equal to the first principal component of a subset of college characteristics.

Because many colleges fail to report one or more characteristics, we were generally able to calculate a quality index for a larger number colleges, the smaller was the set of characteristics considered. Balancing this incentive to keep the set of characteristics small was the desire to use as much information as possible about each college, and the empirical fact that if we used only two measures, the rank order correlations between the indices produced were relatively small, in the range of .5 to .7. Once at least three or four variables were included in the sets, it did not matter much which variables we chose to include; all indices produced were highly correlated.

The quality index is the first principal component of six variables: spending per student, the faculty/student ratio, the rejection rate, average SAT of first year students, and the percent of first year students who were in the top 10 and top 25 percent of their high school classes. The principal components analysis for the six college characteristics is summarized in the Appendix. There is one dominant principal component, which explains 63 percent of the variance of the six variables. The Appendix includes a list of the top 50 colleges as ranked by this quality index. Our quality indicator ranks schools consistent with a priori notions of college quality. We report evidence later that after controlling for other college characteristics, tuition is positively related to this quality index.

As in the previous subsection, each table reports results from specifications that differ by whether we control for the college major of the respondent and the industry in which she works. In addition, we include background controls that fall into three categories: those describing respondentsí home environments, parents, and high schools. Such characteristics have been found to be correlated with schooling in other contexts, and are plausibly correlated with omitted asset or other variables independently affecting both college choice and human capital. Some of them may be correlated with cognitive or other abilities that affect both quality choice and later wages. The background variables are described in the Appendix.

Table 4 reports results from regressions that exclude X and Z_J. These regressions are identical to those reported in Table 1, except that our quality index now replaces the 14 individual college characteristics. We find a positive relationship between wages and quality, whether or not we control for ability. In each specification, the coefficient on the quality index is almost three times as large for black women as for white women, and this difference is statistically significant at 5 or 6 percent.

Table 5 reports results from wage regressions that include all of the X and Z_J controls. These regressions are identical to those described earlier, but include the college quality index in place of individual college characteristics. The table reports results for two separate quality measures: the raw quality index and four dummy variables indicating the quality index quintile into which each college falls (the lowest quintile is the omitted category). The estimates for the quality index are significant at 5 percent in the basic regressions, and when college major controls are added, but are smaller and not significant at 10 percent when both major and industry controls are present. The quality quintile estimates provide weaker evidence of a quality effect, although the estimates for the highest quintile are roughly consistent with the quality effects implied by the linear estimates. The point estimates are generally larger for the higher quintiles, but only the highest achieves 10 percent statistical significance, and it does so only when college major and industry controls are absent. For both of the measures, the point estimates for women are about half the size of comparable estimates for men reported in DBS (1995).

It is possible that college quality is related to later labor force participation. In Table 6 we explore this possibility. The table compares OLS quality effect estimates to maximum likelihood quality effect estimates from joint estimation of participation and wages (see Heckman 1976, 1979). Estimating wages jointly with participation makes no difference to the quality estimates, but among white women, participation is more likely, the higher is the quality of college attended. Among black women, the estimated effect on participation is positive, but not statistically significant. One cannot reject the hypothesis that white women receive no benefit from attending higher quality colleges. The estimated effect for black women is somewhat larger than that for white men (see DBS 1995).

The next table reports estimates from regressions that include family and high school background variables. Table 7 reports results from adding background variables to the basic regressions in Table 5. Adding background controls has little effect on the estimates, although the point estimates are generally a little smaller when all background controls are present. Adding these background variables to regressions that include college major and industry controls in addition to college major controls also has little effect, as does using the most detailed college major controls (See Appendix).

Table 8 reports evidence on the relationship, after controlling for college quality, between wages and the percent of a collegeís students who are black. Some have made arguments suggesting this should be a quality measure for black students, but we find little evidence of this. We divided colleges into four categories: less than 5 percent black students, between 5 to 7 percent black, 8 to 17 percent black, and more than 17 percent black. We report results for the whole sample as well as subsamples of white and black women. For the sample of all women, those attending colleges with between 5 and 7 percent black students earned higher wages than otherwise similar women attending colleges in any other category. This is true whether or not we control for family and high school background characteristics. This pattern, however, is not evident for white women considered separately. There is strong evidence that black women who attended colleges with between 5 and 7 percent black students earn more than those who attended colleges with less than 5 percent black students, and the point estimates suggest they earn more than otherwise similar women who attended college with more than 8 percent black students. DBS (1995) report stronger evidence that racial diversity among students may raise later wages for both black and white men.

The estimates in Table 4 and Table 8 show that black women receive a greater return to quality than do white women. Loury and Garman (1995) and DBS (1995) find similar results for men in the NLS High School Class of 1972 and in the NLSY, respectively.

Table 9 reports our estimates of the effect of tuition, size, and percent female, controlling for quality and for whether or not the college is private. Controlling for quality, we find little evidence that womenís wages are positively related to tuition. The estimated effect is positive, but not statistically significant. Once we control for quality, size appears to have no effect, but there is a private college premium that is statistically significant except when both college major and industry are present.

Women who attend a college with a relatively high proportion of female students have lower wages. When percent female enters linearly, the estimates are negative, but not statistically significant. When percent female is represented by quartile dummies, however, the negative relationship is statistically significant. These results are similar to those reported earlier that did not control for college quality.

Difference Estimates

We now briefly explore how college quality affects wages. Specifically, we consider the possibility that a high quality college education complements or substitutes for on-the-job human capital accumulation. So far, we have assumed this is not the case; the estimates reported here suggest that this assumption is justified.

All of the estimates we have reported so far are from single period regressions of the form

(2)

where Z represents all non-time varying controls except quality, Q, and X_t denotes all time-varying controls. In differences, this becomes

(3)

We can write our alternative hypothesis, that high quality education affects the ability to accumulate human capital, so that it includes equation (2) as a special case:

(4)

where t is tenure or post-college experience. For someone employed in both periods used to calculate the difference, this becomes

(5)

Equation (5) nests equation (3), so there are two ways to test whether equation (2) or (4) is correct. We can estimate equation (4) directly or we can estimate equation (5). In either case, under the alternative hypothesis the data should reject . If we have specified the functional form correctly and there is no relevant unobserved heterogeneity, then either test will reveal an interaction between quality and human capital accumulation. Equation (5) is equivalent to a fixed effects estimator, with the fixed effects removed through differencing. Thus, if the functional form is correct only up to fixed effects, then equation (5), but not equation (4), will produce an unbiased estimate of , and the two strategies might produce different estimates.

Whether we estimate Equation (4) or (5), we can cannot reject the null hypothesis represented by Equation (2). The top and bottom panels of Table 10 report estimates of Equations (4) and (5), respectively. There is little evidence that college quality is related to the speed with which women later accumulate human capital. The estimates are positive, but none are statistically significant Because we earlier found evidence of a private college effect, we added a private school indicator to the difference equation in the bottom panel. Women who attended private colleges had slower wage growth. Apparently, women who attend private colleges have flatter wage profiles. We repeated the analysis in Table 10 with our ability controls present to allow for the possibility that higher ability workers might accumulate human capital at different rates than lower ability workers. Including these controls had no effect on the qualitative results.

Are the Estimates Plausible?

If we are willing to make some simplifying assumptions we can answer the question: based on our estimates do wages increase enough to pay for the added financial investment required to attend a good college? Given that the additional assumptions are reasonable, we can use the answer to gauge the believability of our quality effect estimates. If the market for college quality is in equilibrium, then at the margin the price of quality should just offset its return. We have estimates of the average return. As most students are inframarginal, this average return should exceed the price of quality.

We compare our direct estimates of the wage effect produced by college quality to estimates of the wage effect that would just offset the additional cost of quality (the ìbreak-evenî wage effect). The direct financial benefit from additional college quality is the present value of the additional earnings produced by the increment in quality. The cost of quality is the present value of the additional tuition a student must pay each year to attend a higher quality college.

Denote the post-graduation wage at time t by w(t;q), where q is college quality. Assume students live forever, can borrow at their discount rate r, and take c years to complete college. Ignore any effort or other non-pecuniary costs or benefits to students of attending a higher quality college, and denote the difference in post-graduation wage associated with a difference in quality of by . Assume is constant over time and therefore equal to the increment produced in the first wage received after college, where b is the log wage regression estimate and w(c) is the first wage received after college. Denote this constant, persistent wage effect by . To the extent that college quality increases wage growth understates the value of quality; if quality effects diminish over time, then overstates its value. Discounting reduces the importance of either of these sources of bias.

The dollar price of quality is DT per year in added tuition. Additional quality is worth buying if

(6)

To evaluate this expression, we need an estimate of the price of quality, . We estimate this price by regressing tuition on the quality index and controls not used to construct the quality index that we hope are correlated with amenities and/or the costs of providing a college education. We performed the regressions on the set of all colleges for which we had information. The controls are quartics in size, percent black, and percent female, a public college indicator, and interaction terms between this indicator and all other variables in the regression. We experimented with several functional forms, each differing in the way quality entered the regression. Based on R²s and inspection of residuals we included higher-order quality terms up to the fourth power. This functional form fits the data quite well.

Table 11 reports estimates from this admittedly ad hoc hedonic price regression; Figure 1 graphs the implied tuition-quality relationship for private colleges. Table 12 uses these price estimates to derive the break-even wage effect for private colleges under different assumptions about the interest rate and time to complete a college degree. It also reports the wage effects implied by some of our earlier quality estimates for three different increments in college quality: moving from a college at the 25th percentile of the quality distribution to one at the median, moving from the median to the 75th percentile, and moving from the median to the 95th percentile. The wage effects are those implied for a woman earning $23,529, the beginning salary offered Bachelorís degree recipients in Business in 1990 (Census 1994; Table 289).

The calculations show that our price and quality effect estimates are roughly consistent with one another. At an interest rate of 10 percent, all of the implied effects are somewhat smaller than the break-even effects. At an interest rate of 6 percent, the quality effects implied by our estimates are about the same size as the break-even effects. The implied quality effects are generally a little larger than the break-even effects when college major, industry, and background controls are absent; the implied quality effects are a little smaller that the break-even effects when all the controls are present. Assuming our price estimates are approximately correct, these figures suggest our quality effect estimates imply plausible wage effects. Our estimates of the quality effect on wages are small, but are in line with what it costs to buy quality.

Do Higher Quality Colleges Provide Higher Quality Husbands?

In DBS (1995) we report estimates for men that generally exceeded the break-even values; our estimates for women equal or fall a little short of the break-even values. This suggests that women may be receiving some benefit from college quality other than human capital. Especially given that men receive a larger quality premium, it is possible that women find higher quality husbands at better colleges. In order to see whether this is so, we regressed husbandsí wages on controls for womenís age, educational achievement, ability, region of current residence, and her collegeís quality. Table 13 reports the results. For white women, the estimates of the quality effect are positive and slightly larger than those for her own wage, but they are not statistically significant. The point estimates are smaller when womenís family background and high school characteristics are present. For black women, the estimates are very small, sometimes negative, and are not statistically significant. The results are similar if the samples are restricted to those who attended some post-secondary school (not shown). Because black men receive a larger quality premium that do white men, the weaker evidence of a quality effect on husbandsí wages for black women is additional evidence that women are not meeting higher-earning spouses at better colleges. For white women, size, percent female, and percent black all increase their husbandsí wages. Among black women, size and percent black have little effect, but the proportion of students who are women appears to raise husbandsí wages, although only when we control for college major is the effect statistically significant.

In Table 14 we explore whether a womanís college quality is related to the probability she is married. It is not. For both black and white women, the estimated effect of college quality is small and not statistically significant. For white women, however, other characteristics of their colleges affect the probability they are married. They are more likely to be married if they attended a larger college, and less likely to be married, the higher proportion of students who were black. The estimated effect of the proportion of students female is positive, but not statistically significant at customary levels. There are no statistically significant relationships for black women.

SUMMARY AND CONCLUSION

We find evidence of a small positive return to college quality for young women. The gain from college quality is substantially larger for black women than for white women. White women are more likely to work if they attended higher quality colleges, but accounting for the participation decision has no effect on the estimated effect of quality.

Holding constant college quality and other wage determinants, including ability, women who attend private colleges earn more than women who attend public colleges. Our most surprising finding is that women who attend schools with higher fractions of female students earn less. This pattern holds even after controlling for college quality, college major, industry of employment, family background, and high school quality.

The evidence of a positive return to college quality is less robust for young women than that for young men in DBS (1995). For young men, the relationship between college quality and earnings was positive and statistically significant in every specification examined. Comparing similar specifications, the estimated coefficient on the college quality index was generally about half as large for young women as for young men. Moreover, after controlling for college major and industry of employment, the estimated quality effect for women is substantially reduced; this was not true for the men in DBS (1995). This suggests that for women, a major advantage of attending a higher quality college may be an increased probability of winding up in a high-wage occupation or industry. We find little or no evidence that attending higher quality colleges results in women marrying higher quality husbands.

(N=2,984) (1) (2) (3) (4) (5) (6) (7) (8)

Ability Controls YES YES YES YES

College Characteristics:

Tuition ($10,000s) .227** .190**

(.042) (.042)

Spending per Student .056** .038*

($10,000s) (.021) (.021)

Faculty / Student Ratio 1.40** 1.21**

(.609) (.604)

Size (10,000s) .032** .022

(.014) (.014)

R² (adjusted) .08 .11 .07 .11 .07 .11 .07 .11

(N=2,984) (1) (2) (3) (4) (5) (6) (7) (8)

Ability Controls YES YES YES YES

College Characteristics:

Rejection Rate .277** .277**

(decimal) (.089) (.089)

1^ST Year Retention Rate .600** .430**

(decimal) (.109) (.109)

Graduation Rate .420** .294**

(decimal) (.089) (.090)

HS Top 10% .406** .304**

(decimal) (.098) (.102)

R² (adjusted) .07 .11 .08 .08 .08 .11 .07 .11

(N=2,984) (1) (2) (3) (4) (5) (6) (7) (8)

Ability Controls YES YES YES YES

College Characteristics:

Average SAT .064** .047**

(100s) (.011) (.011)

75th Percentile SAT .049** .039**

(100s) (.014) (.014)

25th Percentile SAT .054** .042**

(100s) (.015) (.015)

SAT Interquartile Ratio .786* .629

(25th / 75th) (.428) (.405)

R² (adjusted) .08 .11 .08 .11 .08 .11 .07 .11

(N=2,984) (1) (2) (3) (4)

Ability Controls YES YES

College Characteristics:

Faculty PhDs .385** .298**

(decimal) (.089) (.089)

Proportion Female -.114 -.042

(decimal) (.135) (.128)

R² (adjusted) .08 .11 .07 .11

Note: ** indicates significance at 5 percent; * indicates significance at 10 percent. Standard errors are estimated by Whiteís (1980) heteroskedasticity-consistent method. The dependent variable is the natural log of real wage for the year ending at the 1987 interview. ìAbility Controlsî are the first two principal components of respondentsí age-adjusted ASVAB scores, and these two variables squared. All regressions include a constant, a variable indicating whether the respondent has any post-secondary schooling, and a variable indicating when the college characteristic was unavailable.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8) (9)

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Tuition .112** .092** .073*

($10,000s) (.039) (.040) (.039)

Spending per Student .031* .026 .013

($10,000s) (.018) (.018) (.017)

Faculty/Student Ratio -1.41 -1.27 -1.28

(1.70) (1.76) (1.69)

R² (adjusted) .29 .29 .34 .29 .29 .34 .29 .29 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8) (9)

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Size .009 .011 .009

(10,000s) (.013) (.013) (.013)

Rejection Rate .170** .169** .136

(decimal) (.084) (.085) (.085)

1st Yr Retention Rate .196** .172* .132

(decimal) (.095) (.094) (.090)

R² (adjusted) .29 .29 .34 .29 .29 .34 .29 .29 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8) (9)

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Graduation Rate .179** .154* .136*

(decimal) (.082) (.082) (.081)

HS Top 10% .209** .197** .152*

(decimal) (.093) (.094) (.091)

SAT Interquartile .137 .087 .066

Ratio (25th / 75th) (.372) (.372) (.358)

R² (adjusted) .29 .29 .34 .29 .29 .34 .29 .29 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

College Major Controls NO NO YES YES NO NO YES YES

Industry Controls NO NO YES YES NO NO YES YES

Respondentís SAT NO YES NO YES NO YES NO YES

Average SAT .027** .027** .023** .023**

(100s) (.010) (.010) (.010) (.010)

75th Percentile SAT .011 .012 .005 .005

(100s) (.012) (.013) (.012) (.012)

R² (adjusted) .29 .29 .34 .34 .29 .29 .34 .34

(N=2,831) (1) (2) (3) (4)

College Major Controls NO NO YES YES

Industry Controls NO NO YES YES

Respondentís SAT NO YES NO YES

25th Percentile SAT .012 .015 .008 .008

(100s) (.014) (.014) (.013) (.014)

R² (adjusted) .29 .29 .34 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8) (9)

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Faculty Ph.Ds .078 .072 .050

(decimal) (.081) (.080) (.078)

Proportion Female -.135 -.133 -.114

(decimal) (.113) (.113) (.120)

2nd Quartile % Female -.062* -.056 -.046

(.032) (.037) (.037)

3rd Quartile % Female -.078* -.072* -.067*

(.042) (.041) (.040)

Top Quartile % Female -.100** -.094** -.085**

(.039) (.039) (.038)

R² (adjusted) .29 .29 .34 .29 .29 .34 .29 .29 .34

Note: ** indicates significance at 5 percent; * indicates significance at 10 percent. The covariance matrix is estimated by Whiteís (1980) heteroskedasticity-consistent method. The dependent variable is the natural log of real wage for the year ending at the 1987 interview. All regressions include a constant, quartics in four variablesóage, tenure, pre-college-graduation labor market experience, and post-graduation experienceócontrols for race, geographical region, urban residence, any post-secondary school completed, years of school completed, years of post-secondary school completed, receipt of a BA degree, and the ability controls described in the text and Appendix. ìCollege Major Controlsî are the 24 basic college majors described in the Appendix; ìIndustry Controlsî are the 15 industry indicators described in the Appendix; union status is included whenever industry controls are. ìRespondentís SATî is their combined math and verbal score from 1981 and is only available for students in college that year. Missing values were set to zero, and a dummy variable was included to indicate missing values.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

College Major Controls NO YES NO YES NO YES NO YES

Industry Controls NO NO YES YES NO NO YES YES

Tuition .131* .110 .106 .086

($10,000s) (.071) (.070) (.070) (.071)

Private -.017 -.016 -.012 -.011

(.050) (.050) (.050) (.050)

JOINT SIG. PÑVALUE: .02 .07 .06 .18

Spending per Student .019 .016 .007 .004

($10,000s) (.019) (.018) (.018) (.029)

Private .061** .051* .056* .046

(.029) (.029) (.029) (.029)

R² (adjusted) .29 .29 .34 .34 .29 .29 .34 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

College Major Controls NO YES NO YES NO YES NO YES

Industry Controls NO NO YES YES NO NO YES YES

Size .025 .024 .021 .019

(10,000s) (.031) (.015) (.014) (.014)

Private .086** .072** .071** .057**

(.031) (.031) (.030) (.030)

Faculty/Student Ratio -1.46 -1.33 -1.34 -1.33

(1.73) (1.79) (1.64) (1.71)

Private .075** .060** .065** .051*

(.030) (.030) (.029) (.029)

R² (adjusted) .29 .29 .33 .34 .29 .29 .33 .34

See note to Table 2.

ALL WOMEN WHITE WOMEN BLACK WOMEN

(N=2,984;1618;886) (1) (2) (3) (4) (5) (6)

Quality Index .044** .034** .028** .024* .071** .073**

(.010) (.010) (.013) (.013) (.019) (.019)

PÑVALUE OF BLACK/WHITE DIFFERENCE: .06 .05

(TWO TAIL)

Any Post-secondary School .199** .151** .231** .164** .197** .174**

(.018) (.018) (.026) (.027) (.030) (.030)

Ability Controls NO YES NO YES NO YES

R² (adjusted) .07 .11 .08 .11 .07 .08

Note: The dependent variable is log real wage. Ability controls are the first two principal components of respondentsí age-adjusted ASVAB scores, and these two variables squared. The regressions include a constant in addition to the variables shown. The standard deviation of the ìQuality Indexî is 1.95.

(N=2,831) (1) (2) (3) (4) (5) (6)

College Major Controls NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES

Quality Index .020** .019** .014

(.009) (.009) (.009)

Quality Quintiles

2nd -.005 -.010 -.018

(.057) (.057) (.053)

3rd .028 .011 .018

(.057) (.057) (.055)

4th .020 .030 .021

(.062) (.062) (.061)

highest .097* .080 .050

(.058) (.058) (.056)

R² (adjusted) .29 .29 .34 .29 .29 .34

Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The standard deviation of the ìQuality Indexî is 1.95. ìQuality Quintilesî are based on the Quality Index. Substituting the first principal component of spending per student and the student/faculty ratio for the Quality Index increases the size and standard error of the quality estimate slightly, but otherwise has little effect.

ALL WOMEN WHITE WOMEN BLACK WOMEN

(1) (2) (3) (4) (5) (6)

LOG WAGE:

Quality Index .020** .020** .017 .015 .037** .037**

(.009) (.009) (.011) (.011) (.017) (.017)

Any Post-secondary .060 .058 .020 .010 .089 .086

(.056) (.056) (.080) (.080) (.087) (.085)

Years Post-secondary .044** .044** .043** .043** .064** .063**

(.009) (.009) (.012) (.012) (.017) (.017)

4-Year Degree .087** .086** .046 .045 .085 .089

(.032) (.031) (.042) (.042) (.059) (.058)

PARTICIPATION:

Quality Index . 097* .252** .034

(.053) (.089) (.128)

Any Post-secondary .435 .895* 1.29

(.309) (.471) (.881)

Years Post-secondary .062 .040 .096

(.053) (.070) (.125)

4-Year Degree .016 -.057 -.338

(.186) (.256) (.431)

-.107 -.231* -.212

n 2,831 3,413 1,556 1,832 831 1,024

log likelihood -1,945 -1,099 -429

Note: ** indicates significance at 5 percent; * indicates significance at 10 percent. The regressors are described in the note to Table 2. The participation equation includes all the regressors in the wage equation, indicators for married, single divorced, and single cohabiting, four indicator variables for the presence of children 2 through 5 years old, an indicator for the presence of children 6 through 13 years old, interactions between the married indicator and the presence of children under 6 years old, and interactions between this indicator and the divorced indicator. Adding family and high school background controls does not change the qualitative results. For white women, adding background controls reduces the quality coefficient in the wage equation to .013 and raises it in the participation equation to .435. For black women, the quality coefficient in the wage equation is unchanged, but rises to .593 (still not significant at the 10 percent level) in the participation equation.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Home YES YES YES YES

Parents YES YES YES YES

High School YES YES YES YES

Quality Index .020** .019** .019** .017*

(.009) (.009) (.009) (.009)

Quality Quintile

2nd -.008 -.001 -.002 .003

(.057) (.056) (.057) (.056)

3rd .025 .026 .026 .028

(.057) (.057) (.057) (.057)

4th .021 .019 .018 .020

(.062) (.063) (.062) (.063)

highest .092 .096* .088 .085

(.006) (.058) (.058) (.058)

R² (adjusted) .29 .29 .29 .29 .29 .29 .29 .29

Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The ìBasic Regressorsî do not include college major or industry controls. Quality coefficient estimates from regressions including these controls are shown in the Appendix. Substituting the first principal component of spending per student and the student/faculty ratio for the quality measure increases the size and standard error of the quality estimate slightly, but otherwise has little effect.

(N=2,831; 1,619; 760) (1) (2) (3) (4) (5) (6) (7) (8) (9)

Background Controls NO NO NO NO NO NO NO NO NO

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Quality Index .019** .017* .012 .016 .015 .012 .034** .026* .014

(.009) (.009) (.009) (.012) (.012) (.012) (.015) (.016) (.015)

Percent Black:

5%ñ7% .078* .080** .076* .050 .052 .042 .165* .196** .244**

(.040) (.040) (.039) (.052) (.052) (.050) (.098) (.099) (.088)

8%ñ17% -.006 -.006 -.005 .037 .029 .027 .029 .026 .072

(.039) (.038) (.037) (.048) (.046) (.046) (.099) (.099) (.086)

greater than 17% .012 .023 -.006 .046 .048 .074 .125 .133 .130

(.045) (.045) (.044) (.092) (.100) (.105) (.103) (.103) (.089)

R² (adjusted) .29 .29 .34 .26 .27 .31 .32 .32 .38

(N=2,831; 1,619; 760) (1) (2) (3) (4) (5) (6) (7) (8) (9)

Background Controls YES YES YES YES YES YES YES YES YES

College Major Controls NO YES YES NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES NO NO YES

Quality Index .016 .014 .009 .015 .014 .010 .035** .027* .016

(.009) (.010) (.009) (.012) (.012) (.012) (.016) (.016) (.015)

Percent Black:

5%ñ7% .075* .076* .072* .045 .045 .033 .174* .208** .239**

(.040) (.040) (.039) (.051) (.052) (.049) (.097) (.101) (.089)

8%ñ17% -.012 -.012 -.013 .015 .012 .006 .021 .023 .067

(.038) (.038) (.038) (.047) (.046) (.046) (.099) (.101) (.086)

greater than 17% .018 .021 -.010 .035 .038 .059 .135 .146 .145

(.045) (.045) (.044) (.093) (.101) (.106) (.103) (.105) (.090)

R² (adjusted) .29 .30 .35 .27 .28 .33 .30 .30 .37

See the notes to Table 2. The results are nearly identical when a private college control is included. The omitted first category includes 45 percent of the white women and 11 percent (30 women) of black women whose colleges report percent black. Thirty-four percent of white women in the sample, and 24 percent of black women, attended colleges with 5 to 7 percent black students. Seventeen percent of white women, and 22 percent of black women, attended colleges with between 8 and 17 percent black students. Four percent of white women (23 women), and 43 percent of black women attended colleges with greater than 17 percent black students. The only differences between black and white women significant at 5% are those for the coefficients on 5ñ7 percent students black when both major and industry controls are present.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

College Major Controls NO YES NO YES NO YES NO YES

Industry Controls NO NO YES YES NO NO YES YES

Quality Index .008 .009 .006 .007 .012 .012 .010 .009

(.011) (.011) (.011) (.011) (.010) (.010) (.010) (.010)

Tuition .109 .083 .090 .067

($10,000s) (.086) (.086) (.085) (.085)

Private -.008 -.005 -.006 -.004 .072** .059* .060* .048

(.055) (.054) (.062) (.053) (.035) (.034) (.033) (.033)

Size .021 .012 .017 .016

(10,000s) (.016) (.010) (.016) (.015)

R² (adjusted) .29 .29 .33 .34 .29 .29 .34 .34

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Background Controls NO NO YES YES NO NO YES YES

College Major Controls NO YES NO YES NO YES NO YES

Industry Controls NO YES NO YES NO YES NO YES

Quality Index .019** .014 .016* .010 .014 .009 .010 .005

(.009) (.009) (.009) (.009) (.010) (.010) (.010) (.010)

Proportion Female -.079 -.077 -.010 -.091

(decimal) (.111) (.110) (.113) (.112)

Proportion Female

2nd quartile -.053 -.043 -.065 -.051

(.040) (.040) (.040) (.039)

3rd quartile -.075* -.074* -.080** -.076*

(.043) (.041) (.043) (.041)

highest quartile -.083** -.074* -.090** -.077*

(.042) (.040) (.042) (.040)

R² (adjusted) .29 .34 .29 .34 .30 .35 .29 .34

See the note to Table 2. The estimates in the top panel do not change appreciably when background controls are added.

Equation 4

Dependent Variable: Log Wage

(N=2,262) (1) (2) (3) (4) (5) (6) (7) (8)

Background Controls NO NO NO NO YES YES YES YES

College Major Controls NO NO YES YES NO NO YES YES

Industry Controls NO NO YES YES NO NO YES YES

Quality Index .012 .009 .007 .007 .012 .009 .006 .007

(.010) (.019) (.010) (.018) (.010) (.019) (.010) (.018)

Quality Experience .050 -.013 .047 -.018

(1,000s) (.275) (.262) (.275) (.261)

R² (adjusted) .29 .29 .34 .36 .29 .29 .33 .33

Equation 5

Dependent Variable: (lnW₈₉-lnW₈₇)

(N=2,262) (1) (2) (3) (4)

Ability Controls NO NO YES YES

(Quality Experience) .047 .097 .027 .078

(1,000s) (.089) (.092) (.090) (.094)

Private -.060* -.062*

(.035) (.035)

R² (adjusted) .05 .05 .09 .09

Note: The dependent variable is natural log of real wage; the regressions are described in the note to Table 2. For all regressions the sample is restricted to those who worked and had valid data in both 1987 and 1989. ìExperienceî in the interaction term is post-college experience.

Dependent Variable: Tuition
Regressors

(N=1,289)

Quality Index -1,774**

(422.4)

Quality² 858.4**

(128.9)

Quality³ -91.5**

(14.9)

Quality⁴ 2.94**

(.569)

public -5,012**

(1,497)

public Quality 429.8

(429.3)

public Quality² -56.8

(244)

public Quality³ -55.3

(43.6)

public Quality⁴ 5.35**

(2.38)

R² (adjusted) .80



Note: The dependent variable is tuition (in-state tuition for public schools). The regressions are estimated using the full set of four-year colleges for which we have information. In addition to the variables shown, they also include a constant, quartics in size, proportion female, and percent black, and interactions between each of these variables and the public school indicator. The quality index is normalized so that its minimum value is zero. Adding the quality variables increases R² by about 0.15.

Dw (break-even) Dw (quality effects)

(1) (2) (3)

Major & Industry Controls NO YES

Background Controls NO YES

Graduate in 4 years, r=.06:

25th %ile to Median $220 $400 $240

Median to 75th %ile $390 $520 $310

Median to 95th %ile $1,450 $2,250 $1,320

Graduate in 4 years, r=.10:

25th %ile to Median $390 $400 $240

Median to 75th %ile $700 $520 $310

Median to 95th %ile $2,570 $2,250 $1,320

Graduate in 5 years, r=.06:

25th %ile to Median $290 $400 $240

Median to 75th %ile $510 $520 $310

Median to 95th %ile $1,870 $2,250 $1,320

Graduate in 5 years, r=.10:

25th %ile to Median $520 $400 $240

Median to 75th %ile $910 $520 $310

Median to 95th %ile $3,370 $2,250 $1,320

Note: See Equation 6 on page 15. The break-even point in the first column is based on the estimated cost of quality at private colleges reported in Table 11. The quality effects are from estimates in Table 5 and Table A10. They are evaluated at $23,529, the average beginning salary offered Bachelorís degree recipients in Business in 1990 (Census 1994; Table 289). Quality effects () are linear in starting salary, so evaluation at the average beginning salary for graduates in the Humanities ($23,213), Social Sciences ($21,627), or Computer Science ($29,804) does not substantially change the result. The 25th, 50th, 75th and 95th percentiles of normalized quality are 2.08, 2.95, 4.05, and 7.42, respectively.

Dependent Variable: Log Husbandís Wage

(N=855) (1) (2) (3) (4) (5)

Background Controls NO YES NO YES YES

College Major Controls NO NO NO NO YES

Quality Index .032 .019 .030 .018 .024

(.029) (.029) (.029) (.030) (.033)

Size .116** .106** .113**

(10,000s) (.034) (.035) (.036)

Proportion Students Female 1.03** .973** 1.10**

(.421) (.438) (.382)

Proportion Students Black 1.81** 1.44** 1.14*

(.660) (.691) (.676)

R² (adjusted) .07 .07 .09 .09 .09

(N=228) (1) (2) (3) (4) (5)

Background Controls NO YES NO YES YES

College Major Controls NO NO NO NO YES

Quality Index .014 -.023 .018 -.010 -.015

(.040) (.042) (.045) (.048) (.059)

Size .034 .026 .033

(10,000s) (.103) (.089) (.110)

Proportion Students Female .818 1.31** 1.07*

(.508) (.554) (.589)

Proportion Students Black .060 .112 .123

(.247) (.243) (.265)

R² (adjusted) .11 .20 .09 .19 .17

Note: ** indicates significance at 5 percent; * indicates significance at 10 percent. All regressions include quartics in the womanís age, ability controls, years of school completed, any post-secondary school, years of post-secondary school completed, an indicator for having a 4-year degree, and regional indicator variables. The results are similar if the samples are restricted to those who attended some post-secondary school.

Dependent Variable: Married in 1987

(N=1,896) dP/dx dP/dx dP/dx

Background Controls YES YES YES

College Major Controls NO NO YES

Quality Index -.018 -.006 -.021 -.008 -.018 -.006

(.034) (.035) (.034)

Size .106* .039 .098 * .036

(10,000s) (.055) (.056)

Proportion Students Female .770 .280 .786 .284

(decimal) (.523) (.539)

Proportion Students Black -1.82** -.663 -1.80** -.653

(decimal) (.894) (.864)

Log Likelihood -1,186 -1,181 -1,173

(N=1,115) dP/dx dP/dx dP/dx

Background Controls YES YES YES

College Major Controls NO NO YES

Quality Index -.029 -.011 -.025 -.009 -.001 -.000

(.061) (.064) (.068)

Size .026 .010 .025 .009

(10,000s) (.105) (.109)

Proportion Students Female .957 .351 1.06 .384

(decimal) (.871) (.932)

Proportion Students Black -.286 -.105 -.313 -.113

(decimal) (.270) (.277)

Log Likelihood -699 -697 -687

Note: ** indicates significance at 5 percent; * indicates significance at 10 percent. The columns labeled ìdP/dxî report the mean numerical derivatives implied by the probit coefficient estimates. All regressions include quartics in the womanís age, ability controls, years of school completed, any post-secondary school, years of post-secondary school completed, an indicator for having a 4-year degree, regional indicator variables, a catholic religion indicator, and family and high school background variables.

Behrman, Jere, Mark Rosenzweig, and Paul Taubman. 1995. ìIndividual Endowments, College Choice and Wages: Estimates Using Data on Female Twins.î University of Pennsylvania. Mimeo.

Betts, Julian. 1994. ìIs There a Link Between School Inputs and Earnings? Fresh Scrutiny of an Old Literature.î in Gary Burtless, ed., Does Money Matter? The Link Between Schools, Student Achievement, and Adult Success. Washington, DC: The Brookings Institution. Forthcoming.

Breusch, T. S., and A. R. Pagan. 1979. ìA Simple Test for Heteroscedasticity and Random Coefficient Variation.î Econometrica 47(5): 1287ñ1294.

Cawley, John, James Heckman, Dimitry Hindanov, Karen Conneely, and Edward Vytlacil. 1995. "Measuring the Effects of Cognitive Ability." University of Chicago. Mimeo.

Daniel, Kermit, Dan Black, and Jeffrey Smith. 1995. "College Quality and the Wages of Young Men." University of Pennsylvania. Mimeo.

Heckman, James J. 1976. ìThe Common Structure of Statistical Models of Truncation, Sample Selection, and Limited Dependent Variables and a Simple Estimator for Such Models.î The Annals of Economic and Social Measurement 5: 475ñ492.

_________. 1979. ìSample Selection Bias as a Specification Error.î Econometrica 47: 153ñ161.

James, Estelle, Nabeel Alsalam, Joseph Conaty, and Duc-Le To. 1989. ìCollege Quality and Future Earnings: Where Should You Send Your Child to College?î American Economic Review 79(2): 247ñ252.

Koenker, Roger. 1981. ìA Note on Studentizing a Test for Heteroscedasticity.î Journal of Econometrics 17: 107ñ112.

Loury, Linda, and David Garman. 1995. ìCollege Selectivity and Earnings,î Journal of Labor Economics 13(2): 289ñ308.

Morgan, James, and Greg Duncan. 1979. ìCollege Quality and Earnings.î Research in Human Capital and Development 1: 103ñ121.

Reed, Ritchie, and Herman Miller. 1970. ìSome Determinants of Variation in Earnings for College Men.î Journal of Human Resources 5(2): 177ñ90.

Solmon, Lewis. 1975. ìThe Definition of College Quality.î Explorations in Economic Research 2: 537ñ87.

Solmon, Lewis, and Paul Wachtel. 1975. ìThe Effects on Income of Type of College Attended.î Sociology of Education 48(winter): 75ñ90.

Sweetman, Olive. 1994. ìCollege Choice: Does it Make a Difference?î University of Newcastle-upon-Tyne. Mimeo.

Tidball, M. Elizabeth. 1989. ìWomenís Colleges: Exceptional Conditions, Not Exceptional Talent, Produce High Achievers,î in C. Pearson, D. Shavlik, and J. Touchton, eds., Educating the Majority: Women Challenge Tradition in Higher Education. New York: MacMillan.

U.S. Bureau of the Census. 1994. Statistical Abstract of the United States: 1994 (114th edition.) Washington, DC.

U.S. News and World Report. 1991. 1992 College Guide, ìDirectory of Colleges and Universities,î September.

Wales, Terence. 1973. ìThe Effect of College Quality on Earnings: Results from the NBER-Thorndike Data.î Journal of Human Resources 8(summer): 306ñ315.

Wachtel, Paul. 1976. ìThe Effect on Earnings of School and College Investment Expenditures.î Review of Economics and Statistics 58: 326ñ331.

The Wall Street Journal. 1995. ìCheat Sheets,î April 5, 225(66): A1.

Weisbrod, Burton, and Peter Karpoff. 1968. ìMonetary Returns to College Education, Student Ability, and College Quality.î Review of Economics and Statistics 50(4): 491ñ7.

White, Halbert. 1980. ìA Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.î Econometrica 48(4): 817ñ838.

Table A1 describes the basic set of regressors included in the log wage regressions. Table A2 reports the results of the principal components analysis used to create the ability controls. Our ability controls were created in two steps. First, we created age-adjusted ASVAB scores by regressing each of the ten ASVAB scores for each individual on age dummy variables and an indicator of whether or not the respondent had completed high school when she took the ASVAB. The residuals from these regressions are the age-adjusted scores. These are the data for the principal components analysis. The first two principal components of the age-adjusted scores are the ability controls used in the wage regressions throughout the paper.

Table A3 describes the family and other background variables included as controls in some of the regressions. Table A4 describes the college characteristics we use as indicators of quality. Table A5 reports descriptive statistics for the college characteristics. Table A6 lists and describes our two sets of college major controls. Table A7 summarizes the principal components analysis that produced our quality index; Table A8 lists the top 50 colleges and universities as ranked by this index. Tables A9 and A10 show that adding background controls to the regressions reported in the text has little effect on the quality estimates. Table A11 shows that using the most disaggregated college major controls has little effect on the quality estimates. Table A12 shows that adding the second principal component of the quality measures has little effect. Table A13 reports coefficient estimates for many of the individual background controls in regressions that control for college quality. Table A14 and Table A15 report the distributions of white and black women across schools, defined by the percent of students who are black.

We tested the exclusion restrictions in the joint estimation reported in Table 6. We did so by including subsets of the excluded variables in the wage regression (the probit always had the full set of excluded variables) and testing the restriction that their coefficients are jointly zero. We cannot reject the exclusion restrictions for most subsets of variables at usual levels of significance, although whether or not we reject the exclusion restrictions is somewhat sensitive to how the variables are grouped for testing. Across all specifications created by including subsets of the excluded variables the quality estimates are extremely stable; the coefficients are always within +/- .002 of those we report.

log wage Log of average real wage (1982 dollars) on all jobs held during the year 1987 or 1989, as indicated

west, south, northeast dummy variables indicating region of residence at the interview 1987 or 1989 interview, as indicated

smsa dummy variable indicating that respondent lived in an SMSA at 1987 or 1989 interview, as indicated

union indicates whether any job held during the year was covered by a collective bargaining agreement

age respondent's age at the 1987 interview

experience total months the respondent has been employed since age 16

tenure total months the respondent has worked for the current employer

BA degree dummy variable indicating the respondent has a 4-year college degree as of 1987 interview

highest grade completed highest grade or year of school the respondent completed as of the 1987 interview

black dummy variable indicating the respondent is black

hispanic dummy variable indicating the respondent is hispanic

(black & hispanic are mutually exclusive)

ability The first and second principal components of the 10 age-adjusted Armed Services Vocational Aptitude Battery scores, and these two principal components squared. The ASVAB was administered in 1980. The age adjustment procedure and principal components analysis are described in this Appendix.

Note: The ASVAB scores are adjusted for age by regressing each test score on age dummy variables and a variable indicating whether the respondent had completed high school when the ASVAB was administered. Principal components analysis is performed on the residuals from these regressions.

Home: magazine ìWhen you were about 14 years old, did you or anyone else living with you get magazines regularly?î

Home: newspaper ìWhen you were about 14 years old, did you or anyone else living with you get a newspaper regularly?î

Home: library card ìWhen you were about 14 years old, did you or anyone else living with you have a library card?î

Parents: mom education Highest grade or year of school completed by respondentís mother.

Parents: mom living Was the respondentís mother living at the 1979 interview (when respondents were between 14 and 22 years old)?

Parents: mom age At the 1987 interview.

Parents: dad education Highest grade or year of school completed by respondentís father.

Parents: dad living Was the respondentís father living at the 1979 interview?

Parents: dad age At the 1987 interview.

Parents: living together Indicator for whether the respondentís mother and father lived in the same household at the 1979 interview.

Parents: mom occupation Occupation of job held longest by mother or stepmother in 1978, represented by dummy variables for each Census 2-digit occupation.

Parents: dad occupation Occupation of job held longest by father or stepfather in 1978, represented by dummy variables for each Census 2-digit occupation.

HS: Size Asked of respondentsí high schools: ìAs of 10/1/79 [or nearest date] what was [your] total enrollment?î

HS: books Asked of respondentsí high schools: ìWhat is the approximate number of catalogued volumes in the school library (enter 0 if your school has no library).î [in 1979]

HS: teacher salary Asked of respondentsí high schools: ìWhat is the first step on an annual salary contract schedule for a beginning certified teacher with a bachelorís degree?î [in 1979]

HS: disadvantaged Asked of respondentsí high schools: ìWhat percentage of the students in [the respondentís high school] are classified as disadvantaged according to ESEA [or other] guidelines?î [in 1979]

Whenever a school quality measure appears, missing values are set to zero and a dummy variable indicating whether the quality measure is missing is included.

Agriculture (24)

agriculture

Architecture (11)

architecture

interior design

misc. architecture

Area Studies (1)

area studies

Biology (54)

biochemistry

biology

microbiology

pre-med

zoology

misc. biology

Business (282)

accounting

banking and finance

business administration

economics

international business

institutional management

marketing

misc. business

Communications (45)

advertising

communications

journalism

radio & TV

misc. Communications

Computer Science (41)

computer science

misc. computer science

Education (150)

education

Engineering (104)

aerospace engineering

chemical engineering

civil engineering

electrical engineering

engineering technologies

general engineering

industrial engineering

mechanical engineering

misc. engineering

Fine Arts (58)

art

commercial arts

drama

design

fine arts

music performing

music liberal arts

studio arts

misc. fine arts

Foreign Languages (14)

french

spanish

misc. foreign language

Health Professions (159)

nursing

misc. health professions

Home Economics (16)

home economics

nutrition

misc. home economics

Interdisciplinary Studies (66)

general studies

liberal arts, & other

interdisciplinary studies

misc. other

Law (11)

law

pre-law

misc. law

Letters (25)

english

misc. letters

Mathematics (9)

applied mathematics

mathematics

misc. mathematics

Military (1)

military science

Office Occupations (53)

data processing

secretarial studies

Physical science (31)

chemistry

geology

physics

misc. physical sciences

Psychology (35)

psychology

misc. psychology

Public Service (46)

law enforcement

social work

misc. public service

Social Science (66)

anthropology

history

political Science

sociology

misc. social science

Theology (12)

theology

This table lists the two aggregations of college majors used. Respondents for whom there is no college major information have all of the indicator variables set to zero. Regressions that include college major controls also include a dummy variable indicating whether college major information was available for the respondent.

Detailed controls: The detailed measure recognizes 82 distinct college majors, listed above.

Basic controls: The 82 ìdetailedî college majors were aggregated into 24 areas. These are listed in boldface, with the number of respondents reporting a major in each of these categories given in parentheses.

standard deviation of quality index: 1.94

Note: The quality index is the first principal component, constructed by multiplying each term in the 1st eigenvector by the corresponding variable. The variables were normalized for the principal components analysis.

Quality

Index

1.	California Institute of Technology	11.20
2.	Johns Hopkins University (MD)	8.43
3.	Yale University (CT)	8.34
4.	Stanford University (CA)	8.27
5.	Massachusetts Institute of Technology	7.95
6.	Harvard University (MA)	7.41
7.	Princeton University (NJ)	7.20
8.	Columbia University (NY)	7.03
9.	Dartmouth College (NH)	7.01
10.	Washington University (MO)	6.99
11.	United States Air Force Academy (CO)	6.89
12.	Duke University (NC)	6.87
13.	Wake Forest University (NC)	6.55
14.	University of Chicago (IL)	6.43
15.	United States Military Academy (NY)	6.21
16.	University of Pennsylvania	5.94
17.	Cornell University (NY)	5.90
18.	Amherst College (MA)	5.88
19.	Rice University (TX)	5.75
20.	Pomona College (CA)	5.65
21.	United States Coast Guard Academy (CT)	5.57
22.	Williams College (MA)	5.56
23.	Cleveland Institute of Music (OH)	5.56
24.	Case Western Reserve University (OH)	5.55
25.	Harvey Mudd College (CA)	5.53
26.	Swarthmore College (PA)	5.51
27.	Northwestern University (IL)	5.45
28.	Bowdoin College (ME)	5.34
29.	Georgetown University (DC)	5.09
30.	Vanderbilt University (TN)	5.03
31.	Brown University (RI)	4.99
32.	Tufts University (MA)	4.67
33.	Wesleyan University (CT)	4.63
34.	University of California at Berkeley	4.62
35.	Wellesley College (MA)	4.59
36.	University of California at San Diego	4.50
37.	Haverford College (PA)	4.49
38.	University of Notre Dame (IN)	4.44
39.	University of Virginia	4.41
40.	University of Rochester (NY)	4.40
41.	Carnegie Mellon University (PA)	4.18
42.	Claremont McKenna College (CA)	4.15
43.	Carleton College (MN)	4.14
44.	Univ. of North Carolina at Chapel Hill	4.14
45.	Middlebury College (VT)	4.12
46.	Bryn Mawr College (PA)	4.07
47.	Emory University (GA)	4.02
48.	Vassar College (NY)	3.88
49.	Washington and Lee University (VA)	3.85
50.	New York University	3.80

Note: This list represents approximately the upper 9 percent of colleges for which we were able to construct our quality index (50 colleges represent about 4 percent of 4-year colleges). Because there appears to be a tendency for better colleges to report more data, ìupper 9 percentî probably understates the relative quality of the colleges on this list compared to their position in the distribution of quality if we could calculate our quality index for all colleges.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Home YES YES YES YES

Parents YES YES YES YES

High School YES YES YES YES

Quality Index .018** .017* .017* .017*

(.009) (.009) (.009) (.009)

Quality Quintile

2nd -.013 -.003 -.006 -.001

(.057) (.055) (.056) (.055)

3rd .008 .012 .009 .014

(.057) (.058) (.057) (.057)

4th .030 .031 .027 .031

(.062) (.063) (.062) (.063)

highest .074 .079 .073 .070

(.058) (.057) (.058) (.058)

R² (adjusted) .30 .29 .30 .29 .29 .29 .29 .30








Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The college major controls are described in the text and the Appendix. Substituting the first principal component of spending per student and the student/faculty ratio for the quality measure increases the size and standard error of the quality estimate slightly, but otherwise has little effect.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Home YES YES YES YES

Parents YES YES YES YES

High School YES YES YES YES

Quality Index .014 .012 .013 .012

(.009) (.009) (.009) (.009)

Quality Quintile

2nd -.021 -.009 -.015 -.009

(.053) (.052) (.053) (.052)

3rd .014 .019 .017 .020

(.055) (.054) (.054) (.054)

4th .021 .022 .018 .022

(.061) (.062) (.060) (.061)

highest .044 .048 .043 .038

(.056) (.056) (.056) (.056)

R² (adjusted) .34 .34 .34 .34 .34 .34 .34 .34

Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The college major controls are described in the text and the Appendix. The industry controls include union status. Substituting the first principal component of spending per student and the student/faculty ratio for the quality measure increases the size and standard error of the quality estimate slightly, but otherwise has little effect.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Home YES YES YES YES

Parents YES YES YES YES

High School YES YES YES YES

Quality Index .015* .014 .015 .013

(.009) (.009) (.009) (.009)

Quality Quintile

2nd -.009 .000 -.003 .003

(.059) (.058) (.059) (.058)

3rd .031 .032 .032 .038

(.059) (.059) (.059) (.059)

4th .061 .061 .056 .061

(.063) (.064) (.063) (.064)

highest .065 .072 .064 .064

(.058) (.057) (.058) (.057)

R² (adjusted) .30 .30 .30 .30 .30 .30 .30 .30









Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. There are 81 ìDetailed College Major Controls,î as described in the Appendix. Substituting the first principal component of spending per student and the student/faculty ratio for the quality measure increases the size and standard error of the quality estimate slightly, but otherwise has little effect.

(N=2,831) (1) (2) (3) (4) (5) (6)

College Major Controls NO YES YES NO YES YES

Industry Controls NO NO YES NO NO YES

1st Principal Component .020** .019** .015*

(Quality Index) (.009) (.009) (.009)

2nd Principal Component .004 -.001 -.008 .011 .018 .019

(.017) (.017) (.017) (.018) (.027) (.026)

1st Principal Component Quintiles

2nd .012 .005 .000

(.055) (.054) (.052)

3rd .047 .027 .036

(.055) (.055) (.053)

4th .041 .046 .038

(.060) (.061) (.059)

highest .116** .095* .068

(.056) (.056) (.055)

R² (adjusted) .29 .29 .34 .29 .29 .34

Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The first principal component is the ìquality indexî used throughout the paper.

(N=2,831) (1) (2) (3) (4) (5) (6) (7) (8)

Detailed College Major NO NO NO NO YES YES YES YES

Industry NO NO NO NO YES YES YES YES

QUALITY INDEX .020** .019** .019** .017* .014 .012 .13 .011

(.009) (.009) (.009) (.009) (.090) (.009) (.009) (.009)

HOME (JOINT SIG.) (1%) (4%) (1%) (4%)

magazine .037** .036* .031* .030*

(.018) (.018) (.017) (.018)

newspaper .015 .008 .022 .016

(.019) (.019) (.018) (.019)

library card .038 ** .032* .038** .032*

(.018) (.018) (.018) (.018)

PARENTS (JOINT SIG.) (1%) (1%) (1%) (4%)

mom education .047 .035 .043 .031

(10s) (.036) (.036) (.035) (.035)

mom living .028 .029 -.001 .003

(.061) (.060) (.060) (.059)

dad education .012 -.004 .020 .004

(10s) (.031) (.031) (.030) (.030)

dad living .000 -.001 -.000 .002

(.035) (.035) (.034) (.034)

living together -.041 -.049 -.053 -.060

(.053) (.053) (.050) (.050)

mom occupation YES YES YES YES

dad occupation YES YES YES YES

HIGH SCHOOL (JOINT SIG.) (1%) (2%) (1%) (2%)

size .310** .298** .335** .270**

(10,000s) (.118) (.121) (.137) (.118)

books .000 .000 -.008 -.003

(10,000s) (.008) (.008) (.012) (.008)

disadvantaged -.065 -.057 -.017 -.056

(decimal) (.042) (.043) (.041) (.042)

teacher salary .070 .078 .193** .107**

($10,000s) (.074) (.074) (.077) (.070)

R² (adjusted) .29 .29 .29 .29 .34 .34 .34 .34
Note: The dependent variable is log real wage; the regressions are described in the note to Table 2. The 81 detailed college major controls are described in the text and the Appendix. The ìHome,î ìParents,î and ìHigh Schoolî variables are described in the Appendix. The covariance matrix is estimated using Whiteís (1980) heteroskedasticity-consistent method.

Collegeís

Percent Cumulative

Students Number of Percent of

Black White Women White Women

Collegeís

Percent Cumulative

Students Number of Percent of

Black Black Women Black Women