economic project
G ender D ifferences in R eturns on Education I. Introduction For a society that claim s to value equality in the w orkplace, the gender gap in w ages in A m erica seem s aw fully persistent. This paper investigates the differences in w ages betw een m en and w om en at different levels of education using data from a sub sam ple of the C urrent Population Survey (2012). Such analysis w ill help reveal the nature of the gender gap, and m ay help identify the segm ents in w hich discrim ination in the w orkforce m ay exist. U sing linear regressions, I first confirm the w age gap in the data and that returns to education are positive. N ext, I use interaction variables to illum inate gender differences on returns at the different levels of education (high school, bachelor’s, and m aster’s). O verall, I find that fem ales see higher returns than m en for com pleting high school and college, but not for graduate school. II. D ata The data set consists of 999 observations of w orking individuals betw een the ages of 18 and 54: The average age in the sam ple is 39.11 years old. O n average, individuals m ade $16.92 an hour w ith a standard deviation of $9.80. The average highest grade com pleted, 13.28, show s that m ost graduated high school. 88% of the sam ple have high school diplom as, 24% hold a bachelor's degree, and 7.4% have com pleted at least a m aster’s. A m ajority w as w hite (81.6% ). 10% of the individuals w ere black, 9% w ere other races. 22.7% of the w orkers w ere parttim e. A pproxim ately half of the sam ple w as fem ale. The follow ing histogram show s the distribution of education level: M ost of the data lies on the m ilestone years. The 12, 14, 16, and 16 areas represent high school diplom as, associate's, bachelor’s, and m aster’s degrees. H ow ever there is som e am biguity at the 14th grade level: these observations could be both associate’s degree holders or four year college dropouts. III. Em pirical M ethodology To com pare gender differences in the returns on w ages at different levels of education I run a linear regression on log w ages: The particular variables of interest are B 9, B 10, and B 11. These interaction variables w ill show the additional percentage point increase or decrease in w ages that fem ales accrue at the different levels of education. B ecause the distribution of w ages is skew ed right, I choose to use log w ages, w hich are m ore norm ally distributed and thus m ay increase the goodness of fit. B ased on prior research, I expect to see positive, though dim inishing, returns to age. Thus, one w ould expect B 1 to be positive and B 2 to be negative. Incom e inequality betw een w hites and blacks is w ell established in econom ic literature, so I expect B 3 to be negative. B 4 is also likely negative since m any of the higher paying jobs w ould be full tim e. I expect a negative coefficient on the fem ale variable, m atching m y hypothesis that the w age gap is present in the data. Lastly, the coefficients on the dum m y variables for com pletion of high school, com pletion of a bachelor’s, and com pletion of a m aster’s are expected to be positive because higher education levels allow individuals to access higher w age positions. There are som e potential concerns w ith this m ethodology. First, there is inevitably a sam ple selection problem since w e are only looking at the data of em ployed people. For exam ple, if it w ere the case that being fem ale low ered the probability of being em ployed due to discrim ination, then the sam ple observations w ould only represent the fem ales w ith a relatively high m arginal productivity of labor. Thus, B 5 m ight underestim ate the true m agnitude of the w age gap. A nother possible concern is om itted variable bias. For exam ple, living in a city is likely positively correlated w ith higher w ages. M oreover, it m ay be the case that having a m aster’s degree is correlated w ith a higher probability of living in a city. These positive correlations w ould cause an upw ard bias on B 8. IV . R esults C olum n 1 show s the baseline specification w hich includes all the race dum m y variables and does not include the interaction variable for fem ale and education. A ll the variables are significant at the 5% level except for the race dum m y variables for A m erican Indian, A sian, M ixed, and H ispanic. A ll signs on the significant coefficients are consistent w ith the predictions discussed in the previous section. I conducted an F (4,986) test on the variables for A m erican Indian, A sian, M ixed, and H ispanic and found that none had a statistically significant im pact on w ages, all else constant (p=.18). Thus, I decided to rem ove them from the regression in colum n 2. The adjusted R ^2 for colum n 1 w as .164. In colum n 2, I add the interaction variables for fem ales and education level. The adjusted R ^2 slightly im proved in this specification to .168. The interaction variables are interpreted as follow s: fem ales see an additional return of 5.6% com pared to m en for graduating high school, an additional 24% return com pared to m en for graduating college, and a negative 26% return com pared to m en for a m aster’s degree, all else equal. A ll the interaction variables are significant at the 5% level. The signs on the rest of the coefficients are consistent w ith the initial predictions, except for the coefficient on bachelor’s degrees. The statistical insignificance of the bachelor’s variable in this regression can be explained as a point in the education level w here fem ale w ages catch up to the m ale w ages (i.e. the w age gap closes). In fact, m y regression predicts that at the bachelor’s degree level of education, fem ale w ages average about 8 percentage points higher than m en. H ow ever, the gap reappears for graduate level jobs. A t that point, the m odel predicts that m en see about a 20% higher return than w om en, all else constant. The follow ing graph show s how the w age gap is “pinched” for bachelor’s degree w ages: V . C onclusion In theory, it is not surprising that a w age gap persists at low er levels of education. Jobs that do not require degrees tend to involve m ore m anual labor, thus have positive returns on physical strength. A ccording to m y results, fem ale w ages catch up to m ale w ages w ith a bachelor’s degree, but lag behind m ale w ages at the graduate level. This m ay be evidence of gender discrim ination for senior positions. A ll in all, these results lend insight into the nature of the gender gap: Since the gap closes w ith a bachelor’s degree, there does not seem to be evidence that w om en earn less doing the sam e jobs as m en. It seem s a m ore likely explanation for the the overall w age gap is that a disproportionate am ount of m en get hired for top paying positions. Further investigation could involve using linear probability m odels to test gender differences in the probability of being em ployed in senior executive positions.