Growing Wealth Gaps in Education

Although it is not the aim of this contribution to assess whether the association between family wealth and children’s education is causal, it is worth nothing that Lovenheim’s evidence on the causal relationship between housing wealth and college entry is an important advance in the literature, especially in the context of continuing debates about the causal influences of family income (see, e.g., Cameron and Taber 2004; Mayer 1997).

The net cost of attending college (i.e., tuition/fees/board minus all financial aid and tax credits) has risen less steeply than sticker prices but still profoundly. In the 25 years between 1990 and 2015, the average net cost of attendance at public four-year colleges rose by 84 % (while the sticker price rose by 111 %); the corresponding increase at private four-year colleges was 39 % (sticker price rose by 78 %) (Ma et al. 2015).

The PSID has been conducted biannually since 1997, so I assess educational attainment at ages 20 and 25 if surveyed in that year but at adjacent ages (older, if available) otherwise. Online Resource 1, section 1, provides an overview of the different measurement years for each birth cohort. It also details how birth cohorts were differently affected by the 1997 PSID sample reduction but shows that the conclusions presented here do not appear to be substantially influenced by it.

To best capture the economic conditions in which the child grows up, I define family wealth as a characteristic of the household in which the child resides at ages 10–14, irrespective of the household’s structure. A different measurement approach would instead link children to the wealth reports of their parents, which, for nonintact families, can provide additional information on the wealth of nonresident parents (an alternative approach that could also be applied in studies focused on family income but typically is not). However, this information is available for only a selective set of cases in which the nonresident parent continues to be interviewed as a PSID respondent. In addition, it is debatable whether and how a nonresident parent’s wealth should be taken into account. Including the wealth of a “truly absent” parent may induce as much measurement error as failing to include the wealth of a nonresident parent with continued parenting commitments (undivided by new parenting commitments to stepchildren). In other work on the intergenerational effects of wealth (Pfeffer and Killewald 2017), we tested the sensitivity of results to these two distinct measurement approaches and concluded that they do not yield substantively different findings.

Stability analyses based on linear probability models are presented in Online Resource 1, section 3.

Based on the Current Population Survey March Supplement, I estimate a college graduation rate for comparable individuals—specifically, individuals who are heads of households and age 25 in survey years 1995 through 2009—of 28 %, compared with 27.2 % in the analytic sample used here.

Here, the lowest group contains those whose parents do not own a home (home value of zero), about 30 % of the sample, while the second lowest group (about 10 % of the sample) consists of children from owned homes valued at most about $64,000 (see Table 4, Appendix 1). The remaining groups are standard quintiles (20 % each).

This comparative assessment could be influenced by differences in the measurement error present in the income and wealth measures. Although separate assessments of the quality of PSID’s income and wealth measures do exist (with generally positive conclusions; see Gouskova and Schoeni 2007; Pfeffer et al. 2016), it is difficult to draw firm conclusions about the relative degree of measurement error in these two variables. However, most researchers would probably be ready to assume more measurement error in wealth than in income, submitting that it may be more difficult to capture (e.g., when held in complex financial products) and more difficult for the respondent to recall and estimate (e.g., paycheck information is recent, but home valuation may not be). If this assumption is correct, the estimated size of the wealth coefficients relative to that of the income coefficients would be underestimated, making for a conservative assessment of the relative importance of wealth.

For an explanation of why statistical significance tests should be based on estimates of discrete change, see Long and Freese (2014:297).

Stability analyses reported in Online Resource 1, section 4, further reinforce the contrast between stagnating college persistence rates among the bottom three quintiles and sharply increasing rates among the top quintile.

Stability analyses based on linear probability models (see Online Resource 1, section 3) reveal only one notable difference: an even more pronounced increase in the growth of college attainment among children from the top wealth quintile; the main conclusion about growing wealth gaps in higher educational attainment based on average marginal effects from logistic regression models, as presented here, thus appears to be conservative.

Note that distributions that include negative values, as is the case for wealth, can produce a Gini coefficient above 1.0.

Additional analyses reported in Online Resource 1, section 5, reveal that the Gini coefficient of nonhousing net worth (net worth excluding home equity) followed a similarly sharp increase, now reaching an astounding level of 0.98 (but see also the previous footnote).

The specific model used here has been calibrated to provide the best empirical fit (discussed further below; see also Online Resource 1, section 6).

Furthermore, in the context of the specific model applied here, I also need to assume that the parameterization of the model remains equally valid—that is, the absolute thresholds chosen for the spline knots that were drawn based on the earlier cohort remain equally useful for the later cohort.

Later ages would introduce endogeneity concerns because college attendance is expected to affect family wealth.