IS THE MEASURED BLACK-WHITE WAGE GAP AMONG WOMEN TOO SMALL? Derek Neal University of Wisconsin Presented Nov 6, 2000 PRELIMINARY

Transcription

1 IS THE MEASURED BLACK-WHITE WAGE GAP AMONG WOMEN TOO SMALL? Derek Neal University of Wisconsin Presented Nov 6, 2000 PRELIMINARY

2 Over twenty years ago, Butler and Heckman (1977) raised the possibility that racial differences in rates of labor force participation may frustrate attempts to measure racial differences in wage rates. Because the wages of employed workers are not randomly sampled from the distribution of potential wages, researchers cannot make direct inferences concerning racial gaps in potential wages from data on observed wage rates. Since the participation rates of black men have fallen considerably below the rates for white men over the past four decades, there is reason to suspect that wage gap measures based on samples of observed wages may overstate black wage gains relative to whites. Thus, numerous studies attempt to assess the importance of selection bias in measuring racial wage inequality. However, these studies deal exclusively with the data on men s wages and participation decisions. Researchers have not yet examined how selection bias affects measures of racial wage inequality among women. There are at least two reasons that previous studies have not explored the link between selection bias and measured black-white wage gaps among women. First, in direct contrast to the trends observed among men, racial differences in participation rates have declined over time among women. In fact, in the 1990 census, aggregate participation rates among black women were almost identical to those observed among white women. Second, many of the strategies used to correct measured wage gaps among men do not seem appropriate for analyses of data on women s wages. Both Brown (1984) and Johnson, Kitamura, and Neal (2000) use methods that assign relatively low wages to non-participants. This strategy seems less appealing for analyses of women than for analyses of men. Compared to their male counterparts, women with high wage offers are much more likely to take a break from their careers in order to raise children. Below, I demonstrate that, while overall participation rates for black women are similar to those observed for white women, the relationship between labor force participation and family structure differs notably by race. In every census year, the sample of white women who do not participate is dominated by married women who are raising children. However, by 1990, the modal family structure among black women who did not work was single motherhood. If family structures serve as signals of unmeasured traits among women, the observed racial differences in relationships between family structure and participation behavior suggest that current measures of black-white wage gaps among women may be contaminated by selection bias.. This paper documents a variety of estimates of the black-white gap in wage offers using

3 data from both the 1990 census and the National Longitudinal Survey of Youth. I emphasize four findings. First, in both data sets, accounting for selection on observed characteristics implies black-white wage gaps that are somewhat larger than those observed among samples of workers. Second, models that use census data to produce measures of the black-white wage gap among women that account for selection on unmeasured traits consistently indicate that, relative to black women, white women are negatively selected into work. The black-white gaps in log wage offers implied by this selection correction are at least.2 greater than the corresponding gaps in measured wages. Third, with respect to gaps measured on samples of workers, wage data from the NLSY yield higher participation rates and larger black-white log wage gaps than those implied by census data on participants. However, the NLSY gaps remain smaller than those implied by the selection corrected estimates from the PUMS analyses. Finally, detailed data on sources on income from the NLSY indicate that among women who do not work over a five-year period, black women and white women are quite different. The typical white woman in this group is receiving income for a spouse and raising children. The vast majority of black women in this group are single mothers receving public assistance. I estimate race-specific wage equations using Heckman s two-step estimator. My exclusion restrictions involve predicted measures of family structure that have two important properties. First, they are based on group characteristics and not individual choices. Second, they measure relative positions in marriage markets and are not correlated with local labor market conditions. These predicted measures of family structure are correlated with participation decisions because they capture variation in the expected value of time devoted to home production, but they may be excluded from my wage equations because they are orthogonal to any unmeasured component of wages that captures individual traits or local labor market conditions. 1. Participation Patterns Table 1 describes trends in two measures of labor force participation. Both measures are constructed using census data from the PUMS. The in labor force column captures whether or 2

4 not respondents were working or seeking employment at the time they completed the census form. The full-time, full-year category measures the fraction of women who usually worked at least 35 hours per week and also worked at least 48 weeks in calendar year In 1990, participation rates among black women were quite similar to those of white women. However, this was not true in previous years. In fact, the overall participation rates among black women in both 1960 and 1970 were more than 10 percentage points higher than corresponding rates for white women. In recent analyses of wage data for men, researchers have devoted considerable attention to the fact that participation rates among black men have fallen notably since Chandra (2000) reports that the participation rate in 1990 among black men age was only 83.5, compared to a corresponding rate of 93.7 among white men. Chandra and others argue that measured black-white wage gaps for men understate the degree of blackwhite inequality because a significant fraction of black men who face low wage offers do not participate and therefore do not report wages. Chandra also asserts that selection bias should not be a major concern in measuring black-white wage gaps among women because participation rates for women differ little by race in Tables 2a and 2b present more detailed data on participation patterns for women. These data indicate that, among women, the similarity of overall participation rates by race masks important racial differences in the composition of the labor force. In every year, among white women age who are not in the labor force, 80 percent or more are married, and over 60 percent are married with children. Further, the corresponding fractions of married women in the labor force are always smaller. However, this is not true among black women. By 1980, married women made up 50 percent of the black female labor force but only 48 percent of black female non-participants. In 1990, married women accounted for 42 percent of the black female labor force and only 36 percent of black female non-participants. Also, among black women, the fraction of non-participants who were single mothers grew each decade, and by 1990, the modal non-participant in the black sample was a single mother. This trend in part reflects that fact that single motherhood among black women increased dramatically from 1960 to However, the 1 See footnote 1 page

5 fraction of participants who were single mothers did not rise as rapidly over this period. Many theories of marriage markets imply that adults who are not married possess less human capital than adults who are married. 2 Further, Neal (2000) outlines a model where it is not only true that married women have higher endowments than single women, but it is also true that single women with children possess smaller endowments than single women who are childless. Because white non-participants have high marriage rates relative to white participants and because black non-participants have high rates of single motherhood relative to black participants, the relative endowments of black non-participants may be lower than the relative endowments of white non-participants. Thus, even though the overall participation rates for black and white women are quite similar by 1990, measured racial wage gaps among women may still be contaminated by selection bias. 2. Empirical Method Here, I describe an empirical method for assessing how selection bias affects the measurement of black-white wage gaps among women. Consider the following model of wages and labor force participation. W i X i β e i s i X i θ Z i γ v i where s i is the latent value of participating in the labor market for individual i. W i is the market wage offer for individual i. The econometrician observes W i if s i is positive. If s i 0, W i is missing. X i is a vector of human capital characteristics the influence market wages. Z i measures 2 See Becker (1981) and Lam (1988) for examples. 4

6 factors that influence the value of participating, either through their influence on the value of nonmarket time or their influence on pecuniary costs associated with market work. The statistical properties of this model are well understood. Heckman s (1979) method provides a consistent estimator for β under the assumption that e i and v i are drawn from a bivariate normal distribution. While the assumption of joint normality provides identification, even in the absence of explicit exclusion restrictions, Z i, it is desirable to have a set of characteristics Z i that do help explain participation decisions but have no direct influence on wage offers. Creating or discovering such measures is quite difficult. Measures of family structure provide natural places to start, but marriage decisions are influenced by the surplus available in potential matches, and the total surplus available in a given match is a function of both partners skill endowments. Since skills determine market wages and shadow prices of time in home production, individual marital status may be correlated with e i. Further, since market wages help determine the cost of time spent with children, individual family size may also be correlated with e i. I propose an estimator that employs predicted family structures as determinants of the expected value of time in home production. The method I propose cannot be generalized to achieve identification in a non-parametric setting because predicted family structures do not vary among individuals i and j if X i = X j. However, the method does not achieve identification through an arbitrary choice of functional form. I create the variables in Z i by fleshing out the implications of a specific type of assortative mating in marriage markets. Assume that marriage markets exhibit assortative mating on education and consider the following algorithm. Use census data from the PUMS to construct marriage markets for women of a particular age and race who live in a specific state. Given a market with M men and F women, rank the men and women in each market by education level. If M < F, add F-M men with no education to the bottom of the men s ranking. Then, for women at each education level, compute the average education level of men who share the same rank. 3 This average provides a 3 For example, consider a market defined within the PUMS that contains 600 women. Given a ranking of these women by education level, consider all women with exactly 16 years of schooling. If these women are ranked in positions 50 to 120, compute the average level of education among men ranked from 50 to 120, and denote this average as the expected education match for women in this market with exactly 16 years of schooling. 5

7 measure of expected match quality for women of a particular schooling level in a specific marriage market under the assumption that courting opportunities reflect assortative mating on education. Using this measure of expected match quality as well as the rank of each women s education group in her marriage market, I estimate multinomial logit models in order to construct predicted rates of marriage with children, marriage without children, single motherhood, and single without children. I estimate 4 separate models on samples drawn from the 1990 census. I estimate models separately for black and white women, and I divide the samples further into samples of women ages 25 to 39 and women ages 40 to 55. The regressors in the model include not only rank in the marriage market and the expected education of a woman s potential match but also indicator variables for state, age, and education level as well as interaction terms between education level and the expected education of the potential match, an interaction between marriage market rank and a five year average of the state-specific maximum AFDC benefit, and another interaction between rank and a five year average of food stamps benefits. 4 Based on the coefficient estimates from these models, I create predicted probabilities of each family structure that are specific to women who share the same race, birth cohort, education level, and state of residence. These predicted probabilities as well as interaction terms between these probabilities and several other variables make up Z i, my set of exclusion restrictions. The following summarizes the specification of the two equation model presented above. X i includes indicators for state of residence and education plus a quartic in potential experience. Z i includes estimated probabilities of three family structures: (i) married with children, married without children, and single with children. In addition, there are interaction terms between the probabilities of structures involving marriage and the expected education of the potential match. There is also an interaction between the probability of being a single 4 I thank Michael Keane and Kenneth Wolpin for data on AFDC and Food Stamp program from 1967 to I restrict the samples to states that contain at least 100 women of each race in each seven year age range that defines a marriage market. 6

8 mother and the sum of welfare and food stamp benefits available in a given state and year. Finally, there are interactions between age and the predicted probabilities of structures involving children. As I note above, I estimate separate sets of equations by race and age group. The Z i variables give predicted family structures based on the distribution of male and female education endowments in each marriage market and the position of particular women in these markets. Because the variables involve expected family structures, each element of Z i should be correlated with the expected value of time spent in home production. I include interaction terms between age and the probabilities associated with structures involving children because the age of women will be correlated with the age of their children. The age of children should influence the value of home time. The interaction between the probability of single motherhood and aid levels captures that fact that, among single mothers, the level of welfare benefits represents an opportunity cost associated with leaving home production and entering market work. Finally, the wealth or total surplus in marriage will be greater in markets where potential spouses are better educated, and this wealth effect may affect the value of time at home. Therefore, I include interactions between the probabilities of the two marriage structures and the expected education of potential mates. 3. Results from the PUMS I begin by describing results from the first-stage of the Heckman two-step estimator. Table 3 provides coefficients and standard errors from the first-stage participation probits. The table presents estimated coefficients for the variables that are excluded from the wage equation. I estimate the model using several different definitions of participation and wages. Table 3 contains results from two specific models, but the patterns observed here are common in the other specifications. Several patterns are noteworthy. To begin, the table includes, for each model, the estimate of the correlation coefficient between the errors in the wage offer and participation equations. Note that, in the analyses of black women, one can never reject the null hypothesis of no selection on unmeasured traits. 7

9 However, this is not true for the models involving white women. In all four models presented here, there is statistically significant evidence that white women are negatively selected into work. The estimated coefficients on the variables that are excluded from the second stage are often statistically significant, and the estimated coefficients for whites follow a pattern that might be expected given standard models of labor supply and household production. The most striking difference between the results for blacks and whites involves relationships between participation rates and the expected probability of being married with children. Among black women, participation rates rise with the likelihood of being married with children. Tables 4a and 4b give mean log wages and predicted mean log wages by race. These results come from four different versions of the two-step model. One model involves weekly wages and defines participation as full-time, full-year work. The other models use constructed hourly wages and three different participation definitions. In all four cases, the following patterns emerge. i) The gap in mean log wages based on samples of participants in less than.1 ii) If one assumes that there is no problem of selection on unmeasured traits, but corrects the estimated gaps for selection on observed traits by constructing predicted wages for non-workers, the implied black-white gap in mean log wages increases by.03 to.04. In every instance, correcting for selection on observed education, potential experience, and state of residence implies an increase of over 40% in the implied black-white wage gap. iii) The black-white wage gaps implied by the corrections of selection on unmeasured traits are much larger. While the predicted averages for black women are barely affected by these corrections, the implied adjustments in the white sample are quite large. iv) Given that one cannot reject the null of no selection on unmeasured traits among black women, it is appropriate to construct estimated black-white log wage gaps based on the selection 8

10 adjusted predictions for white women and the OLS full-sample predictions for black women. This procedure yields estimated gaps that range from.31 to.40 in absolute value. In all cases, the effect of the selection correction among white women is to increase predicted wages across the board but dampen the estimated returns to education. Therefore, the correction for selection on umeasured traits implies the greatest increase in black-white gaps among less educated women. 5 It is interesting to note that the selection corrected gaps yields patterns across education classes that more closely resemble the patterns observed among men. Racial wage inequality is most pronounced among the least educated. 4. Results from the NLSY The results presented in Tables 4a-4d suggest that published data on wage rates may grossly understate the level of racial wage inequality among women. To gather additional evidence on the issue, I now turn to another data set, the National Longitudinal Survey of Youth. The NLSY sample is much smaller than the PUMS sample, and thus, the Heckman two-step model employed above is not a viable option. However, the NLSY does provide detailed panel data on work experience and wages. The panel nature of the study creates more complete data on participation and wages in any particular year, and the panel also permits the construction of multi-year wage averages. The use of multi-year averages allows me to create samples with much higher participation rates. Tables 5a and 5b present results based on two samples of workers. The first includes women who were interviewed and whose data did not contain coding errors in The second contains women with valid data in any year between 1988 and Note that the participation rates in the five year samples are over 90 percent for both black and white women. Several results are noteworthy: 5 See Appendix Table 1 for details. 9

11 i) The black-white wage gaps based on the samples of participants are larger in the NLSY than in the 1990 census data. In both the 1990 and the samples, the log gap among participants is.17. ii) In both samples, the implied black-white gap increases in absolute value to.19 if one accounts for selection on observed characteristics by creating predicted wages based on the coefficients from OLS regressions on participants. iii) The bulk of the difference between the measured gaps in NLSY and the measured gaps in the PUMS data reflects the fact that the NLSY records lower average wages for black women. On the surface, the NLSY results may be taken as further evidence that census data on the earnings, hours, and weeks worked of participants yield a black-white wage gap among women that is too small. However, the age range of the NLSY samples is only 26-33, and in contrast to the census samples used to produce Tables 4a-4d, the NLSY data cover all states. Tables 6a and 6b compare results from the NLSY and the PUMS. The PUMS samples are drawn for the age ranges that match the NLSY. Further, the NLSY sample does not include workers who report valid wages but do not report valid hours and weeks information. Without this restriction, the two samples are not comparable because constructing wages in the census data requires valid information on labor supply. However, rate of pay information in the NLSY is much richer and there are many cases in which respondents provide valid wage data without providing valid labor supply data. In Table 6a and 6b, the NLSY participation rates remain above participation rates in the PUMS data. However, they are well below the five-year rates obtained using all available wage data in the NLSY. NLSY wages are higher in these samples. Female workers in the NLSY who report wages but do not report complete hours or weeks records earn below average wages. Nonetheless, the implied black-white wage gaps in Table 6b are quite similar to those reported in Table 5a. The black-white gaps reported in Table 6a come from IPUMS data for the NLSY age 10

12 range. For full-time workers, analyses of weekly wages yield black-white log wages gaps around.13 if one corrects for selection on measured characteristics. This is true not only in Table 6a, but also in Table 6a. Among women with positive weeks and hours, the corresponding gap is slightly less than.12. In Table 4d, the corresponding gap is.08. In sum, the age composition of the NLSY sample does not account for the difference in the PUMS and NLSY results. The NLSY data consistently imply larger gaps between blacks and whites. However, the NLSY gaps implied by corrections for selection on measured traits are never more than.19. Given that the results in Table 5a are based on participation rates of over 90 percent, one might worry that the NLSY results are inconsistent with the large gaps of.3 or more implied by the two-step method described in the analyses related to Tables 4a-4d.. Table 7 above sheds some light on this issue by providing descriptive statistics on sources on income for women who are either represented or not represented in the wage samples employed in Table 5a. The contrast between black women who do and do not participate is striking. Ten percent of black women in the NLSY sample did not work between 1988 and Of these women, 73 percent received SSI, AFDC, and/or Food Stamp payments in each year between 1988 and Among black women who worked at least once during this period, the corresponding number is.14. Further, black women who did not work over this period averaged 4.13 years of aid receipt. Among black women who did work over this period, the corresponding average is only The pattern is reversed with respect to receipt of spousal income. Among women who did not work from , less than 3 percent were receiving spousal income in each year. The corresponding number is 14 percent among those who did work. However, among white women who do not work during the period, 47 percent recieved spousal income in each year. The corresponding fraction for participants is only.38. Among white women, the modal non-participant is receiving income from a spouse. Among black women, the modal non-participant is receiving public assistance. Although receipt of public assistance in more common among non-participants in both the black and white samples, the overall patterns suggest that non-particpation in the white sample is associated with higher rates of spousal support while non-participation in the black sample is associated with receiving public 11

13 assistance. The differences are large, and if one assumes that family structures and sources of income provide information about expected human capital endowments, the measured racial wage gaps of.19 in the NLSY may significantly understate the true extent of racial wage inequality. Conclusion These analyses have not addressed the sources of black-white wage inequality among women. They provide no direct evidence concerning the part of racial wage gaps that should be attributed to discrimination. However, the results do indicate that standard methods of measuring racial wage inequality may understate the level of inequality among women. The results presented here suggest that overall black-white wage gaps among women may be similar in magnitude to the large gaps observed among men. Further work remains. 12

14 References: Altonji, Joseph and Blank, Rebecca. Race and Gender in the Labor Market in Orley Ashenfelter and David Card, eds., Handbook of Labor Economics. Vol. 3, Amsterdam: North Holland, Brown, Charles. Black-White Earnings Ratios since the Civil Rights Acts of 1964: The Importance of Labor Market Dropouts." Quarterly Journal of Economics, February 1984, 91(1), pp Butler, Richard and Heckman, James J. The Government s Impact on the Labor Market Status of Black Americans: A Critical Review, in L. Hausman, et al, eds. Equal Rights and Industrial Relations. Madison: Industrial Relations Research Assocation, Chandra, Amitabh. Labor Market Dropouts and the Racial Wage Gap: ," American Economic Review. May, pp Heckman, James J. "The Impact of Government." in S. Shulman and W. Darity, eds., The Question of Discrimination. Middletown, CT: Wesleyan University Press, 1989, pp Neal, Derek and Johnson, William. The Role of Premarket Factors in Black-White Wage Differences. Journal of Political Economy, October 1996, 104(5), pp Smith, James and Welch, Finis. Black Economic Progress after Myrdal. Journal of Economic Literature. June 1989, 27(2), pp

15 14

16 Table 1: Female Labor Force Participation by Race, Black White In Labor Force Worked Full-Time In Labor Force Worked Full-Time Note: Sample includes women between 25 and 55 years of age. In Labor Force refers to status at the time of census interview. Worked Full-time refers to persons who worked at least 35 hours in the survey week and worked at least 48 weeks in the previous calendar year. 15

17 Table 2a: Family Structure for Female Black Labor Force Participants and Non- Participants, In Labor Force Not in Labor Force With Child Without Child With Child Without Child Married Single Married Single Married Single Married Single Full-Time, Full-Year Not Full-Time, Full-Year With Child Without Child With Child Without Child Married Single Married Single Married Single Married Single Note: See Table 1 for sample definitions Table 2b: Family Structure for Female White Labor Force Participants and Non- Participants, In Labor Force Not in Labor Force With Child Without Child With Child Without Child Married Single Married Single Married Single Married Single Full-Time, Full-Year Not Full-Time, Full-Year With Child Without Child With Child Without Child Married Single Married Single Married Single Married Single

18 Table 3: 1990 Hourly and Weekly Wage Selection Equation Coefficients By Age, Race and Level of Labor Supply Hourly Wages: Hours>=1, Weeks>=1 Weekly Wages: Full-time, Full-year Blacks Whites Blacks Whites Age Age Age Age Age Age Age Age Pr(Marr. w/ child) 5.13 (1.55) 1.96 (2.87) (0.61) (0.84) 2.93 (1.44) (2.52) (0.53) (0.79) Pr(Marr. no child (1.97) 0.37 (0.99) 0.37 (0.80) (0.52) (1.97) (0.97) 0.98 (0.73) (0.50) Pr(Single w/ child) 0.66 (1.75) (3.17) (1.64) 0.43 (2.57) (1.64) (2.83) (1.55) 1.29 (2.47) Pr(Marr. w/ child)* Age (0.049) (0.062) (0.021) (0.016) (0.045) (0.056) (0.018) (0.014) Pr(Single w/ child)* Age (0.051) (0.066) 0.41 (0.05) (0.054) (0.048) 0.20 (0.06) 0.55 (0.04) (0.052) Pr(Marr. w/ child)* (Exp. Spouse Educ.) (0.030) (0.031) (0.009) (0.010) (0.030) (0.030) (0.009) (0.010) Pr(Marr. no child)* (Exp. Spouse Educ.) (0.117) (0.062) (0.038) (0.014) 0.31 (0.11) (0.061) (0.035) (0.014) Pr(Single w/ child)* (Aid Value) (0.0005) (0.001) (0.0004) (0.0007) (0.0004) (0.0008) (0.0004) (0.0007) From 2 nd stage: implied rho [t-statistic] [0.31] [-.41] [-3.26] [-3.21] [-0.13] [-0.19] [-3.08] [-1.68] Note: Standard errors in parentheses except for implied rho. 17

19 Table 4a: Actual and Predicted 1990 Wages by Education Group Full-time, Full-year Workers Weekly Wages Black White Working Full-time, Full-year Full Sample Working Full-time, Full Sample Unadjusted Adjusted Full-year Unadjusted Adjusted No High School Some High School High School Graduate Some College College Graduate All Education Levels Hourly Wages Black White Working Full-time, Full-year Full Sample Working Full-time, Full Sample Unadjusted Adjusted Full-year Unadjusted Adjusted No High School Some High School High School Graduate Some College College Graduate All Education Levels

20 Table 4b: Actual and Predicted 1990 Hourly Wages by Education Group Non-Full-time, Full-year Workers Part-time, Part-Year Workers Black White Hours>=15, Weeks>=26 Full Sample Hours>=15, Full Sample Weeks>=26 Unadjusted Adjusted Unadjusted Adjusted No High School Some High School High School Graduate Some College College Graduate All Education Levels Workers with Positive Hours and Weeks Black White Hours>=1, Weeks>=1 Full Sample Hours>=1, Full Sample Weeks>=1 Unadjusted Adjusted Unadjusted Adjusted No High School Some High School High School Graduate Some College College Graduate All Education Levels

21 Table 5a: Actual and Predicted 1990 Log Wages NLSY Avg. Log Wage: Participants Blacks Predicted Avg. Log Wage: Full Sample Participation Rate Avg. Log Wage: Participants Whites Predicted Avg. Log Wage: Full Sample Participation Rate One-year Five-year Average Table 5b: Wage Regression Coefficients NLSY Blacks Whites 1990 Log Wage 5-year Avg. Log Wage 1990 Log Wage 5-year Avg. Log Wage 1-7 Yrs Experience (0.024) (0.019) (0.016) (0.014) 8-12 Yrs Experience (0.010) (0.008) (0.009) (0.007) 12+ Yrs Experience (0.016) (0.012) (0.014) (0.010) Education (0.009) (0.007) (0.007) (0.005) Constant (0.239) (0.187) (0.168) (0.136) R N 1,011 1,254 1,872 2,307 Note: Standard errors in parentheses except for implied rho. 20

22 Table 6: Average and Predicted 1990 Wages Census vs. NLSY Ages 26 to 33 Table 6a: Census Black White Avg. Wage: Participants Avg. Wage: Full Sample Participation Rate Avg. Wage: Participants Avg. Wage: Full Sample Participation Rate 35+ Hours, 48+ Weeks 15+ Hours, 26+ Weeks 1+ Hours, 1+ Weeks Table 6b: NLSY Black White Avg. Wage: Participants Avg. Wage: Full Sample Participation Rate Avg. Wage: Participants Avg. Wage: Full Sample Participation Rate 35+ Hours, 48+ Weeks 15+ Hours, 26+ Weeks 1+ Hours, 1+ Weeks

23 Table 7: Sources of Income and Education NLSY, Years of Govt. Aid Years of Spousal Income Fraction with 5 Years of Aid Fraction with 5 Years of Spousal Income Education Blacks 1990 Participants Non-Participants Participants Non-Participants Whites 1990 Participants Non-Participants Participants Non-Participants

24 Appendix Table 1: Returns to Education 1990 Hourly Wages Blacks Whites Age Age Age Age Education Unadj. Adjusted Unadj. Adjusted Unadj. Adjusted Unadj. Adjusted 7 years (0.094) (0.255) (0.069) (0.168) (0.028) (0.076) (0.030) (0.105) 9 years (0.087) (0.226) (0.071) (0.176) (0.030) (0.078) (0.032) (0.117) 10 years (0.078) (0.196) (0.067) (0.179) (0.028) (0.074) (0.030) (0.126) 11 years (0.074) (0.191) (0.064) (0.182) (0.026) (0.071) (0.029) (0.129) 12 years (0.073) (0.202) (0.063) (0.216) (0.024) (0.076) (0.027) (0.155) 13 years (0.073) (0.215) (0.063) (0.246) (0.024) (0.080) (0.027) (0.176) 14 years (0.073) (0.222) (0.065) (0.245) (0.025) (0.081) (0.028) (0.172) 16 years (0.073) (0.231) (0.065) (0.236) (0.024) (0.081) (0.028) (0.145) 17 years (0.075) (0.226) (0.067) (0.199) 1.10 (0.025) (0.079) (0.028) (0.144) 18 years (0.084) (0.232) (0.085) (0.265) 1.21 (0.027) 1.05 (0.085) 1.01 (0.033) (0.184) 20 years 1.05 (0.126) 1.06 (0.267) 1.08 (0.098) 1.08 (0.257) 1.12 (0.037) (0.095) 1.09 (0.037) (0.207) Note: Standard errors in parentheses. 23