The Relative Returns to Graduating from a Historically Black College/University: Propensity Score Matching Estimates from the National Survey of Black Americans

NLS-72 (Jones 1986) limits labor market experience to approximately 10 years, as it follows high school seniors—some of whom eventually attended and/or graduated from college—from 1972–1986. The B&B (Wine et al. 2005) limits the labor market experience of actual college graduates to 11 years with the initial survey in the 1992–1993 academic year, and the most recent follow-up survey in 2003. While the C&B (Bowen and Bok 1998) captures data on college attendees/graduates starting in 1951, it only includes 4 HBCUs.

The specification of the utility function incorporates the idea that as institutions, colleges have economic (e.g, imparting skills that have a return in the labor market) and sociological (e.g, imparting on students particular identities corresponding to social ideals) functions (Akerlof and Kranton 2002). Allowing earnings to be embodied directly in effort recognizes the economic and human capital dimension of college. The sociological dimension of college—particularly its role in shaping social ideals, is captured by allowing identity/self-image to indirectly and multiplicatively affect expected post-college earnings. In this context, the model differs from Akerlof and Kranton (2002) as investments in identity/self-image can have an indirect tangible economic return. This is similar to the approach of Darity et al. (2006). Our theoretical framework treats I as identity and self-image. Another interpretation of I—particularly in a context where it corresponds to an identity of high confidence/self-esteem is that it is a measure of psychological capital. Thus, our model also embodies the effects of psychological capital on earnings as considered by Goldsmith et al. (1997).

Wave 1 of the NSBA was administered to 2,107 individuals, Wave 2 was administered to 951 individuals (including 935 from Wave 1), Wave 3 to 793 individuals (including 779 individuals from wave 2), and Wave 4 to 659 individuals (including 1 individual from Wave1, 28 individuals from Wave 2, and 623 individuals from Wave 4).

The first wave of the NSBA is the only one that reports numeric values of income for individuals, and it is top-coded. The last three waves do not report any individual measures of income, providing only categorical measures of total family/household income.

The Duncan index is measured on a numeric scale from 0 to 96, and the numeric value is proportional to the prestige of an occupation. A two-stage procedure is utilized to generate a Duncan Index. In the first stage, prestige rankings for a few occupations are regressed on a measure of education and earnings for an occupation. In the second stage, estimated parameters from the first stage are used to rank a broader set of occupations, then scaled to formulate an index. For a more detailed description see Duncan (1961) and Duncan et al. (1972).

Matching estimators are becoming popular in applied economics/sociology research—including estimating the returns to attending/gradating from college as in the recent analysis of Brand and Halaby (2006).

See Augurzky and Kluve (2007) for a consideration of propensity score distance matching.

These HBCUs include Xavier, Morehouse, Spelman, and Howard.

For the cross product term, at least one of the pre-treatment characteristics is non-binary.

We were unable to estimate a propensity score for graduating from an elite HBCU, as there was simply not enough observations on these HBCUs to get meaningful Probit parameter estimates. As such, in our treatment effect specifications of elite HBCUs, we use the estimated propensity score for graduating from HBCUs in general.

For estimating the treatment effect of HBCU graduation, our matching estimators are based on four matches with replacement, as there is evidence that matching parameter estimates are robust when selecting between one and four matches with replacement (Imbens 2004).

For estimating the treatment effect of graduating from an elite HBCU, our matching estimators are based on one match, as in each wave, there was no more than one respondent who graduated from an elite HBCU. Identification of treatment effects with matching estimators requires that the number of matches equal the minimum of the number of treated and controls in the sample (Abadie et al. 2001).

See Rilling et al. (2008) for an application of this procedure.

For a critique on the empirical performance of propensity score based estimators, see Wilde and Hollister (2007).