The effect of banning affirmative action on human capital accumulation prior to college entry

Abstract

This paper examines how banning affirmative action in university admissions affects both overall academic achievement and the racial gap in academic achievement prior to college entry. Focusing on college-bound high school students, we use a difference-in-difference methodology to analyze the impact of the end of race-based affirmative action at the University of California in 1998 on both the overall level of SAT scores and high school GPA, and the racial gap in SAT scores and high school GPA. We find little evidence of either a decline in academic achievement or a widening of the racial gap in academic achievement after the ban.

JEL codes

I21; I24

1 Introduction

Universities in the United States are increasingly limited in their ability to practice race-based affirmative action. In the last two decades, public universities in a growing number of states have stopped practicing race-based affirmative action due to various court rulings, voter initiatives, and administrative decisions¹. In addition, the United States Supreme Court’s 2013 decision in Fisher v. Texas makes it more difficult for universities to justify using race as a factor in admissions, and its 2014 ruling in Schuette v. Coalition leaves the door open for additional states to implement statewide bans on affirmative action. The implications of eliminating race-based affirmative action in college admissions are far-reaching and have been the subject of considerable legal, political, and scholarly debate. In this paper, we focus on the potential effects of banning affirmative action on academic achievement prior to college entry.

Economic theory suggests that eliminating affirmative action could have important implications for human capital accumulation. There are a number of channels through which this effect could operate. First, the removal of racial preferences directly affects admissions probabilities, which in turn affects the return to studying prior to college application. Second, banning affirmative action could lead underrepresented minorities to feel institutionally discouraged from attending college, and they could respond by putting forth less effort in preparing for postsecondary education. Finally, if colleges react to bans on affirmative action by changing their admissions process more broadly, then this too could lead students to shift the focus of their efforts in high school.

While economic theory clearly predicts that ending affirmative action could affect human capital investment, it does not yield definitive predictions about whether the overall level of human capital investment or the racial gap in human capital investment will rise of fall². Thus, the effect of banning affirmative action on human capital investment is largely an empirical question.

In an attempt to answer this question, this paper explores the impact of Proposition 209, which prohibited public universities in California from practicing race-based affirmative action, on both the SAT scores and high school GPA of college-bound high school students. In particular, using data from the College Board, we examine how these measures of academic achievement changed in California relative to other states (a difference-in-difference estimate), and how the racial gap in these measures changed in California relative to other states (a triple-difference estimate). We find little evidence that the ban had a negative impact on either overall academic achievement or the racial gap in academic achievement. This finding is consistent with previous research documenting a limited behavioral response to California’s ban on affirmative action in terms of both college application behavior and college enrollment behavior (see, for example, Antonovics and Backes (2013) and Antonovics and Sander (2013)).

Our paper proceeds as follows. Section 2 discusses the related literature, and Section 3 presents an overview California’s affirmative action ban. Section 4 discusses our empirical strategy and gives a short overview of our data. Section 5 presents our results, and Section 6 concludes.

2 Related literature

The theoretical literature on affirmative action in college admissions emphasizes the distinction between “color-sighted affirmative action”, wherein colleges use explicit racial preferences in admissions, and “color-blind affirmative action”, wherein colleges implicitly favor minorities by using admissions rules that favor students who possess characteristics that are positively correlated with being a minority (see, for example, Fryer et al. (2008), Ray and Sethi (2010) and Fryer and Loury (2013))³. Both forms of affirmative action stand in contrast to laissez-faire admission regimes in which race is not considered either explicitly or implicitly. In the case of California, it seems clear that Prop 209 shifted most University of California (UC) schools from color-sighted to color-blind affirmative action. That is, the UC administration openly acknowledged that diversity remained a high priority even after Prop 209, and Antonovics and Backes (2014) provide evidence that, after Prop 209, UC schools changed their admissions process to implicitly favor minorities.

In their seminal paper, Coate and Loury (1993) show that affirmative action in the labor market has an ambiguous theoretical effect on human capital investment. In the context of college admissions, Fryer et al. (2008) establish that moving from color-sighted to color-blind affirmative action alters students’ incentives to invest in human capital, but their model does not yield definitive predictions about whether investment will go up or down. For example, if the move from color-sighted to color-blind affirmative action lowers minorities’ admission rate (as was the case for California), then this could dampen minorities’ incentives to invest in human capital since there is no point in investing if you have no chance of being admitted. On the other hand, reducing minorities’ chances of admission could increase human capital investment since it may become more important to distinguish yourself as the field becomes more competitive. Finally, Hickman (2012) and Hickman (2013) also model the link between different admissions regimes and human capital investment, but do not consider color-blind affirmative action, which we feel is essential to understanding the effects of Prop 209⁴.

A substantial number of papers examine the effect of affirmative action on academic achievement (for a recent survey, see Arcidiacono and Lovenheim (2014)). Of these, only a handful use data from before and after a ban on affirmative action, with most focussing on academic achievement after college entry (see, for example, Arcidiacono et al. (2014), Backes (2012), Cortes (2010) and Hinrichs (2012)). To our knowledge, only two previous studies have directly examined the effect of a ban on affirmative action on academic achievement prior to college entry, Furstenberg (2010) and Caldwell (2010). Both find evidence that banning affirmative action increases the racial test score gap. We discuss these papers in greater detail when we present our results in Section 5.

3 Background on California’s ban on affirmative action

The first threat to affirmative action in California was in July 1995, when the Board of Regents of the University of California passed a resolution (SP-1), which stipulated that the UC would discontinue considering race in admissions by the beginning of 1997. The implementation of SP-1, however, was delayed. Then, in November 1996, California voters approved Prop 209, which banned the use of racial preferences at public universities⁵. Prop 209 underwent various legal challenges until the Supreme Court denied further appeals in November 1997. Thus, the incoming class of 1998 was the first to be admitted under the statewide ban on affirmative action. Table 1 presents a timeline of the events leading up to the ban.

Table 1 California’s proposition 209 timeline

Although the prohibition against affirmative action applied to all public universities in California, in practice it only affected UC schools because at the time most Cal State schools were not very selective, admitting the vast majority of applicants, regardless of race. Given that one in six Californian high school graduates apply to at least one UC campus, however, it is reasonable to think that major policy changes at the UC could affect the human capital investment for the state as a whole⁶. In addition, because it was so controversial, Proposition 209 received a tremendous amount of coverage in the media. For example, between July 1996 and July 2001, over 900 articles in the Los Angeles Times contained the phrase “Proposition 209” (roughly one every other day). Thus, it seems likely that high school students were aware of Prop 209.

Prop 209 also had an enormous impact on the minorities at the more selective UC schools⁷. For example, the fraction of URM applicants admitted to Berkeley plummeted from 52 percent to 25 percent between the three-year period immediately before and the three-year period immediately after the ban. In contrast, the admission rates of non-minority applicants to Berkeley only fell from 32 percent to 28 percent over that same time period.

Evidence that students’ human capital investment decisions respond to the incentives created by college admissions policies is evident in Cortes and Friedson (2014) and Cullen et al. (2013), both of which find evidence of students moving to lower quality high schools after Texas introduced its top ten percent plan, which guaranteed admission to any public university in Texas for students who graduated in the top ten percent of their high school class.

Nonetheless, isolating the effect of the end of race-based affirmative action at the UC is complicated by a number of concomitant policy changes. First, in an effort to minimize the effects of Prop 209 on minority enrollment, the UC system substantially increased minority outreach efforts⁸. It is unclear, however, how much of an effect the new outreach programs had on the applicant pool in the years immediately following Prop 209 since many of these programs took years to fully develop and were long-term in nature. To the extent that increased outreach had an effect on academic achievement, our estimates represent the net effect of ending race-based affirmative action and the accompanying change in outreach.

In addition to the increase in outreach, there were two major policy changes in California around the same time as Prop 209. First, the 1999 passage of the Public Schools Accountability Act brought about substantial changes to the public K-12 schools in California. In particular, beginning in 2000, schools were eligible for rewards if all ethnic subgroups within schools either scored above a certain threshold or met targets for test score growth. In addition, schools with low test scores could opt into an intervention program designed for schools not meeting growth targets. Second, in 2001, the UC implemented Eligibility in the Local Context (ELC), guaranteeing any student in the top four percent of his or her high school class admission to at least one UC school (conditional on completing specified coursework). The new policy was designed to attract students from high schools that did not typically send many students to the UC, giving the UC a way to potentially increase minority enrollment. Since both of these policy changes are likely to have affected human capital accumulation, we perform robustness checks by removing observations from 2000 and later from our analysis and find no substantial impact on our main results.

Finally, we note that the long-run effects of Proposition 209 on SAT scores and high school GPA are likely to be larger than the short-run effects. In the short run, any change in students’ SAT scores or high school GPA is likely to primarily reflect a change in student effort. In the long run, however, changes in students’ SAT scores and high school GPA additionally may reflect changes in accumulated knowledge and skill. Since our data only allow us to examine students graduating from high school in the four-year period immediately following the implementation of Prop 209 (1998–2001), our results are likely to reflect the intermediate-run effects Prop 209.

4 Estimation and data

Antonovics and Backes (2014) provide evidence that Prop 209 led to widespread shifts in the entire admission process at many UC schools. In particular, the more selective UC schools appear to have decreased the weight placed on SAT scores and increased the weight given to high school GPA and family background characteristics in order to boost minority admission rates after Prop 209 went into effect. This is important because the apparent reduced emphasis on SAT scores and increased emphasis on high school GPA might have led to a shift away from human capital investments that increase SAT scores towards those that increase high school GPA. For this reason, we separately examine the effect of Prop 209 on these two measures of academic achievement.

In addition, we measure the effects of the policy change in two ways. First, we explore the reaction of Californians relative to the rest of the country. Second, we document how the gap between whites and underrepresented minorities (URMs) changed in California relative to the rest of the country⁹. Each of the two measures is important. Since Californians of all races were affected by the changes in the admissions policies at UC schools, they may be thought of as one treated group. Comparing Californians to those in other states reveals the extent to which Prop 209 changed academic achievement for all Californians. On the other hand, affirmative action policies are generally thought of as a way to address the gap between white and minority students. Viewed in this way, it is natural to ask how the removal of explicit racial preferences affected the racial gap in academic achievement.

4.1 Empirical strategy

To examine the effect of Prop 209 on academic achievement, we estimate the following:

$$\begin{array}{ll}{\text{Outcome}}_{\mathit{\text{ist}}}& ={\beta }_0{\text{CA}}_s+{\beta }_1{\text{Post}}_{\mathit{\text{st}}}+{\beta }_2{\text{URM}}_i\\ +{\beta }_{01}{\text{CA}}_s{\text{Post}}_{\mathit{\text{st}}}+{\beta }_{02}{\text{CA}}_s{\text{URM}}_i+{\beta }_{12}{\text{Post}}_{\mathit{\text{st}}}{\text{URM}}_i\\ +{\beta }_{012}{\text{CA}}_s{\text{Post}}_{\mathit{\text{st}}}{\text{URM}}_i\\ +\beta X_i+{\epsilon }_{\mathit{\text{ist}}},\end{array}$$

(1)

where Outcome _{i s t} is the outcome (SAT score or high school GPA) for student i in state s in year t. Post _{s t} is an indicator for whether the affirmative action ban was in place in state s in year t, CA _s is an indicator for whether the student resides in California, and URM _i is an indicator for an individual’s race. Finally, X _i includes controls for parental income, education, gender, whether English is the student’s first language, citizenship status, and a constant term. Following the discussion at the beginning of this section, there are two coefficients of interest. First, to the extent that Californians had a common response to Prop 209, it would be captured by β₀₁, which represents the change in the dependent variable for white Californians relative to the rest of the country. Second, β₀₁₂ represents the change in the minority-white test score gap in California relative to the rest of the country.

There are two important choices that must be made in estimating Equation (1). The first is which cohorts to include. As discussed above, there was an extended period of uncertainty surrounding the end of race-based affirmative action at the UC. Table 1 shows the timing of events leading up to the implementation of Prop 209 along with the timing of when various graduation cohorts are likely to have taken the SAT. As the table indicates, the first major threat to affirmative action at the UC came in July 1995 when the Regents of California passed SP-1, which committed the UC system to an eventual ban on the use of racial preference in admissions. It was not until over two years later, however, in November 1997, when Prop 209 officially became law that the UC stopped the use of explicit racial preferences. Thus, students who took the SAT between July 1995 and November 1997 (those in the 1996–1998 graduation cohorts) did so during a time of considerable uncertainty about the future of race-based affirmative action. As a result, we experiment with two different pre and post period definitions: (i) 1994–1996 vs 1998–2001 and (ii) 1994–1995 vs 1999–2001. The first choice of sample years drops only the 1997 cohort, which took the SAT during the election cycle in which Prop 209 was on the ballot. The second choice of sample years takes a more conservative approach by dropping any cohort that could have taken the SAT during a period of uncertainty about Prop 209. However, it leaves only two years in the pre period and three in the post period.

The second important choice is whether to include an interaction term between the post period and demographic characteristics. By not including such an interaction, time-varying changes in the relationship between demographic characteristics and the outcome variable would be picked up by the coefficient on Post*URM, due to the correlation between race and demographic characteristics. However, if the effects of the policy operated in part through changes in the relationship between demographic characteristics and the outcome variable (due to, for example, changes in the weights placed on different student characteristics in determining admissions), the researcher may be interested in omitting the interaction between the post period and demographics to allow the changes to load onto the URM*Post coefficient. We present results from both specifications.

In practice, we also expand our estimating equation to include the full set of interactions for blacks, Hispanics, and Asians, with whites as the excluded group. We also use year fixed effects and state fixed effects, and we include state-specific linear time trends. We drop observations from Florida, Louisiana, Mississippi, Texas, and Washington, which were affected by their own affirmative action policy changes during our sample period. We also drop North Dakota and Wyoming due to small sample sizes of minority SAT test takers¹⁰. Finally, we normalize test scores and GPA to be mean zero with standard deviation one.

Asians, who constitute a large portion of the college-going population in California, are not considered URMs for the purposes of admissions to the UC. We estimate effects on Asians separately but generally do not focus on their results because blacks and Hispanics were the intended beneficiaries of affirmative action policy at the UC and because the outcomes for Asians and whites are generally similar¹¹.

4.2 Estimation of standard errors with limited treated units

A growing number of papers have documented the inadequacy of typical methods of obtaining standard errors when the number of treated units is small (see Moulton (1990), Wooldridge: Cluster-sample methods in applied econometrics: an extended analysis, unpublished, Donald and Lang (2007), Abadie et al. (2010), and Buchmueller et al. (2011)). To illustrate the problem, consider SAT scores as the outcome variable and suppose we are interested in β₀₁, which measures the change in test scores for Californians relative to the rest of the country. Since SAT scores naturally fluctuate from year to year within a state even in the absence of a policy change, it is important to distinguish these fluctuations from the true policy effect. This is done in the typical DD setting by assuming that these state-specific, year-to-year fluctuations average out to a mean of zero over a large number of treatment and control states. In our case, there is only one treated unit, so there is little reason to believe this assumption holds.

We follow an established method of dealing with the problem of only one treated state by using the remaining untreated states to conduct a permutation test in which we construct the empirical distribution of β₀₁ by estimating ${\widehat{\beta }}_{01}$ for each of the control states, treating each control state as the treated state (see Abadie et al. (2010) and Buchmueller et al. (2011))¹². Thus, the control states are used to estimate the variability of ${\widehat{\beta }}_{01}$ that is driven by year-to-year variation in test scores. The null hypothesis – that Prop 209 had no effect on academic achievement in California – is rejected when ${\widehat{\beta }}_{01}^{\text{California}}$ is large relative to the estimated empirical distribution of ${\widehat{\beta }}_{01}^j$ for the control states (j indexes the control states). This procedure tests whether the change in test scores in California is large relative to the naturally occurring variation in test scores observed in other states. We also estimate ${\widehat{\beta }}_{012}^{\text{California}}$ and ${\widehat{\beta }}_{012}^j$ for each of the states in our data to measure whether the change in the minority-white gap in California was extreme relative to the change in states that did not ban affirmative action in the same time period. This correction reveals substantial within-state year-to-year variability in SAT scores.

Of course, one solution to the above problem would be to include more treated states. However, most states that banned affirmative action have done so too recently to be contained in our data¹³. Two states that banned affirmative action during the period for which we have data are Texas and Florida; however, each introduced top x % plans in which the top x % of students within a high school were guaranteed admission to an in-state public university¹⁴. Louisiana and Mississippi were both affected by the court ruling that ended affirmative action in Texas, but both states were under desegregation orders, so they may not have been under pressure to comply with the ruling¹⁵. In addition, neither state’s public universities are as selective as the institutions typically affected by bans on racial preferences (see Blume and Long (2014)). Of the affirmative action banning states, the sole remainder is Washington, which voted to ban affirmative action in 1999. In principle, Washington could be included with our analysis; however, using two policy change states instead of one would still be insufficient for classical estimation of standard errors¹⁶. Finally, there are several advantages to using California. First, it is a large state with a significant URM population. In addition, it has been well-established that the more selective UC schools practiced significant race-based affirmative action prior to Prop 209, and finally, the measurement of how admissions rules changed at the UC after Prop 209 has been documented by Antonovics and Backes (2014).

4.3 Data

Our College Board data consist of SAT test takers who are expected to graduate from high school between 1994 and 2001¹⁷. The sample consists of all black and Hispanic test takers nationwide, all Californian test takers, and a 25 percent random sample of the rest of the country. The College Board includes a range of descriptive variables that are generated when students fill out the Student Descriptive Questionnaire before taking the exam. These include race, gender, parental characteristics, college aspirations, high school GPA, and many other variables. Our data also identify each students’ high school graduation cohort, and we assume that students in a given graduation cohort took the SAT between the spring and fall of the preceding year. In addition, to the extent that students took the SAT multiple times, our data only reveal students’ last SAT score. Vigdor and Clotfelter (2003) show that students’ SAT scores tend to increase when they retake the SAT. Thus, if Prop 209 lowered students’ incentive to retake the SAT (perhaps because SAT scores became less important in determining admission), then this is likely to show up as a decline in average SAT scores.

One advantage of using the College Board sample is that nearly all SAT takers are interested in going to college, so they should be the ones most readily affected by the affirmative action ban. On the other hand, a potential problem with using the College Board is that a student’s decision about whether to take the SAT could be affected by Prop 209, leading to possible sample selection bias. Dickson (2006) finds that removal of affirmative action in Texas led to a decline in the percentage of minority high school graduates who take either the ACT or SAT of roughly 3–4 percent. If Prop 209 led to a similar drop in the fraction of URMs who took the SAT in California, then to the extent that these students were differentially likely to score poorly on the SAT, this could lead us to understate the true drop in the relative performance of URMs. We discuss this issue in a later section below.

Basic summary statistics for our College Board sample are displayed in Table 2. Californians are quite similar to the rest of the country in terms of SAT scores and high school GPA, but, as might be expected, a smaller fraction are U.S. citizens and a larger fraction speak English as a second language. In both California and the rest of the U.S., blacks and Hispanics tend to score lower on the SAT, have lower GPAs, and have parents with lower levels of education.

Table 2 College board summary statistics

5 Results

5.1 Overview of trends in test scores

As a first pass at gauging the effects of Prop 209 on SAT scores and high school GPA, we plot normalized (mean zero and standard deviation one) average SAT scores by race and year in Figure 1. Panels (a) and (b) show normalized SAT math and verbal scores for Californians and the rest of the U.S. Although whites tend to score higher than URMs, the gap appears to be roughly stable over time. Panel (c) shows normalized high school GPA. The patterns in the figure underscore the importance of controlling for state-specific time trends: there was a gradual rise in GPA over time that began before the implementation of the preference ban.

Figure 1

College Board outcome variables. Each panel shows the mean of the normalized outcome (mean zero and standard deviation one) by race and year. The U.S. panel excludes California, Florida, Louisiana, Mississippi, Texas, Washington, North Dakota, and Wyoming. (a) SAT Math Scores (b) SAT Verbal Scores (c) High School GPA.

5.2 Regression results

Panel 1 of Table 3 presents results when including all but the 1997 cohort. The first three columns display our results when we include controls for demographic characteristics but not for the interaction between demographic characteristics and the post-Prop 209 indicator. In terms of the effect of Prop 209 on the overall level of academic achievement, for California relative to the rest of the country, we find that after Prop 209 there was a 0.003 standard deviation increase in SAT math scores and a 0.020 standard deviation increase in SAT verbal scores. We also find a 0.019 standard deviation relative increase in high school GPA. While the direction of these point estimates indicates an overall increase in academic achievement in California relative to the rest of the country, we note that the magnitudes are small and, on balance, suggest that Prop 209 had no meaningful impact on the overall level of academic achievement.

Table 3 College board SAT and GPA DDD estimates

In terms of the racial gap in academic achievement, for California relative to the rest of the country, we find a reduction in both the black-white SAT gap and the black-white GPA gap. In addition, while we find an increase in Hispanic-white SAT gap (with the estimated 0.026 standard deviation increase in the Hispanic-white SAT math gap, for example, representing about 3.5 SAT points, or about 4 percent of the Hispanic-white SAT math gap in California), we also find a reduction in the Hispanic-white high school GPA gap. Thus, our findings do not generally support the notion that banning affirmative action widens racial gaps in academic achievement.

Columns (4)-(6) present results when allowing the coefficients on the demographic controls to vary in the post period. Results are generally similar but smaller in magnitude since part of the effect that was previously captured by the URM*Post coefficient is now contained in the Demographics*Post coefficients.

Panel 2 further restricts the sample by dropping each of the 1996 through 1998 cohorts from the analysis. Many coefficients become larger in magnitude, especially when comparing Californians to the rest of the country. For example, the estimated performance gain of Californians on the SAT verbal section increases from.02 standard deviations to.046 standard deviations when dropping the additional years. Again, the estimates become somewhat smaller in magnitude when including Post*Demographic interactions as shown in columns (4)-(6).

The standard errors in Table 3 are generated by clustering at the state level. However, as discussed above, the presence of one policy change state may result in standard errors that are misleadingly small. As a result, we conduct permutation tests in which we generate a coefficient ‘estimate’ for each of the 44 states in our data using the specification which resulted in the largest point estimates – columns (1)-(3) of panel 2. Figure 2 plots histograms of these placebo coefficient estimates, with the red line indicating the coefficient estimate for California. In addition, the confidence intervals obtained by the permutation tests are shown in brackets in each of the regression tables. The point estimates for California are generally extremely small relative to the other states. However, given the wide range of coefficient estimates from other states, we would be unable to detect even a large change in California.

Figure 2

Placebo Test Coefficients. Distribution of DDD coefficient estimates for each of the 44 states in the sample, with California represented by red line. See text. (a) CA*Post (b) CA*Post*Black (c) CA*Post*Hispanic.

In seeking to understand the imprecision of our estimates, we plot average SAT math scores by state and year for the first eight states alphabetically¹⁸. The plot is displayed in Figure 3, with California represented by the dashed line¹⁹. A naive look at the graphic would suggest a very small increase in the Californian SAT math scores in the post period. However, two factors prevent being able to make a definitive causal statement about the change in Californian performance due to Prop 209. First, the small increase in SAT math scores was part of a general upward trend in both California and the rest of the country (see Panel (a) of Figure 1). Second, the other states shown in Figure 3 generally have substantial year-to-year variability, which does not show up in the “U.S.” panel of Figure 1a since all the states are averaged together. Thus, considerable year-to-year variation within each state makes it difficult to make definitive statements about the causal effect of the policy change despite the very large sample sizes in the College Board.

Figure 3

Mean of normalized SAT math scores. Mean zero and standard deviation one.

Despite the inherent problem with performing a difference-in-difference estimation with one policy change state, we believe that the small point estimates are noteworthy in themselves. As discussed above, Prop 209 led to substantial changes in URM admissions rates and in the UC admissions process. However, despite these changes, we find that the performance of Californians as a whole and of black students relative to white students was relatively stable. Even Hispanics, who appear to have performed worse on the SAT in the post period, saw an increase in their high school GPAs. Thus, for neither Californians as a whole nor for any minority group within California do we find consistent evidence of lower academic achievement in response to Prop 209.It is possible that our small estimated effect sizes are the result of including many students who would be unaffected by Prop 209 because they are unlikely to attend the top-tier UC campuses where the effects of Prop 209 were strongest. In our sample, only one in six Californians applies to a UC campus, and even fewer to a selective campus where the effects of Prop 209 were strongest. In an effort to isolate a sample of students likely to be most affected by the policy change, we predict which students would have a high likelihood of sending an SAT score to either Berkeley or UCLA. Specifically, for cohorts in the pre-Prop 209 period, we regress the likelihood of score-sending to at least one of Berkeley or UCLA on demographic characteristics that plausibly were not affected by Prop 209, which consist of parental education and income, race, gender, whether English was the first language spoken, and citizenship status. We then use the coefficients to predict score-sending for the entire sample, both in the pre and post periods. The density of predicted Berkeley or UCLA score-sending is shown in Figure 4, with the vertical lines denoting cutoff points between quartiles of score-sending likelihood. We keep the top quartile of predicted score-senders to generate a sample of students likely to be most affected by Prop 209.

Figure 4

Density of predicted score-sending. Predicted score-sending obtained by regressing the likelihood of score-sending for Californian pre-period cohorts to at least one of Berkeley or UCLA on demographic characteristics consisting of parental education and income, race, gender, first language spoken, and citizenship status. Details in text. Each vertical line represents the boundary of a quartile.

We then run the same regressions as previously on this selected sample. Results for these likely Berkeley or UCLA score-senders are shown in Table 4. Coefficient estimates continue to be small, with the possible exception of an estimated reduction in the Hispanic-white GPA gap²⁰.

Table 4 College board SAT and GPA DDD estimates for students in top quartile by likelihood of score-sending to Berkeley or UCLA

5.3 Robustness to selection of SAT test takers

As mentioned earlier, a potential concern about using a sample of SAT test takers is that if Prop 209 led to a drop in the fraction of URMs who took the SAT in California, then to the extent that these students were disproportionately drawn from the low end of the SAT distribution (conditional on observable characteristics), this would lead us to understate the true drop in the relative performance of URMs.

To explore this issue, we examine the trend in the number of SAT takers from California as a fraction of the number of California public high school graduates from 1995–2001 using data from the 1997–2003 Digest of Education Statistics²¹. Results are shown in Figure 5. There is some evidence that black students were relatively less likely to take the SAT after Prop 209, especially in 1998. Relative to white students, black students were two percentage points less likely to take the SAT in the post period (authors’ calculation), which is equivalent to about four percent of black SAT test takers. On the other hand, relative to white students, there was no change in the share of Hispanic students who took the SAT.As mentioned above, if it were the case that the fall in the fraction of black SAT takers was driven by blacks at the bottom end of the SAT score distribution, our estimates would understate the drop in the relative SAT scores of blacks. In order to gauge the magnitude of this potential bias, we conduct a simulation in which we artificially increase the number of black SAT test takers in the post period in California by identifying the bottom one percent of Californian post-period black SAT scorers and replicating each four times (to “recover” the estimated 4 percent drop suggested by Figure 5).

Figure 5

Share of SAT Takers. Fraction of high school graduates by race in California who took the SAT.

Results with these “extra” low-achieving Californian post-period black students are shown in Table 5. Not surprisingly, we now find a drop in the relative SAT scores of blacks after Prop 209. Overall, the results for blacks now mirror those for Hispanics, with an increase in the minority-white SAT gap and decrease in the minority-white GPA gap. This pattern is interesting as it suggests that blacks and Hispanics may have responded to the new post-Prop 209 admissions rules at the UC, which placed less weight on SAT scores and more weight on high school GPA. Importantly, however, we do not find evidence of a uniform drop in relative minority achievement.

Table 5 Sensitivity of estimates to selection by black students

As a second way of addressing potential selection into taking the SAT, we note that high-achieving students likely to apply to Berkeley and UCLA are unlikely to have reduced the rate at which they took the SAT after Prop 209. As shown in Table 4, when we restrict our sample to students with a high probability of applying to Berkeley or UCLA, we find no evidence on an increase in either the minority-white SAT gap or the minority-white GPA gap.

5.4 Why our findings differ from previous work

Standing in contrast to the results presented here, Furstenberg (2010) finds a statistically significant widening of the black-white SAT gap in California following Prop 209. However, our analysis improves upon his in several ways. First, his College Board sample only includes the 1996–2000 cohorts; his paper compares the 1996–1997 cohorts to the 1998–2000 cohorts. However, interpreting results from 1996–1998 is difficult since, as discussed above, the UC Regents first announced their intention to end their use of racial preferences in July 1995. Thus, it is possible that students began responding to the policy change long before 1998, his first post-policy change year. Second, Furstenberg’s data consist of a 30-percent sample of SAT takers, while we have obtained a more comprehensive dataset containing all Californian test takers, all black and Hispanic test takers nationwide, and a 25-percent sample of the remaining non-Californian whites. All together, our sample contain 51 percent of test takers during the 1994 to 2001 period. As a result, our estimates are obtained from a much larger sample.

An additional previous empirical study of the impact of affirmative action bans on human capital accumulation, Caldwell (2010), examines PIAT math test scores using the Children of the NLSY 79 (CNLSY). Caldwell finds that Prop 209 increased the racial test score gap. We obtained the same restricted version of the CNLSY used by Caldwell and successfully replicated his findings. However, there are several issues with his analysis.

First, the CNLSY was not designed to yield a representative sample at the state level. From the NLS FAQ²²,

The National Longitudinal Surveys are designed to represent specific birth cohorts at the national level. The surveys cannot provide representative estimates for States [...] NLS data files with geographic variables are available on a restricted basis for authorized researchers to use, but the permitted uses do not include producing estimates for States.

Thus, it may be difficult to generalize the findings obtained from the CNLSY because it is not a randomly selected sample of Californians. While the College Board data used in this paper is also not representative of all Californian high school students, it does include all Californian SAT test takers and these are the most likely to respond to Prop 209.

Compounding the non-representativeness of the CNLSY sample are the relatively small sample sizes in the panel. For example, one of Caldwell’s strongest results – suggesting that the PIAT math scores of 13 and 14 year olds fell considerably for blacks in California relative to blacks in the rest of the country – is estimated from only 62 Californian blacks, with 17 of these from the period after Prop 209. This is important because, as noted in Wooldridge: Cluster-sample methods in applied econometrics: an extended analysis, unpublished, small sample sizes exacerbate the problems inherent to estimation with only one policy change state²³.

Finally, neither Furstenberg nor Caldwell takes into account the limited number of treated states in evaluating the statistical significance of their results.

6 Conclusion

Much of the popular debate surrounding affirmative action in higher education focuses on how it affects the allocation of students to universities, taking the achievement of high school graduates as fixed. However, the disparities in educational preparation which drive racial differences in enrollment at selective colleges arise early in the education process and are formed well before college admissions come into play. As affirmative action was originally conceived to mitigate these gaps in racial achievement, it is natural to ask whether and how the removal of racial preferences affects these gaps. In addition, a number of scholars have pointed out that since policies such as Prop 209 give colleges and universities an incentive to place a greater weight on non-academic factors in determining admissions, they could lower student quality by weakening all students’ incentives to invest in their academic qualifications prior to college entry.

Our results provide little support for the idea that the end of raced-based affirmative action at the UC lowered either overall academic achievement or the academic achievement of minorities relative to whites. This finding is consistent with several previous studies that have shown a minimal response to Prop 209 in terms of both application behavior and enrollment behavior.

Endnotes

¹ Establishing the list of states in which race-based affirmative action has been prohibited is complicated by ambiguities in case law, but arguably includes Alabama, Arizona, California, Florida, Georgia, Louisiana, Michigan, Mississippi, Nebraska, New Hampshire, Oklahoma, Texas and Washington. See Blume and Long (2014) for a nice discussion of the policy environment surrounding affirmative action.

² We discuss the theoretical literature in greater detail below.

³ In this paper, we use the terms “color-sighted affirmative action” and “race-based affirmative action” interchangeably.

⁴ Hickman’s color-blind admission policy (in which universities do not consider race at all) differs from color-blind affirmative action (in which colleges implicitly favor minorities).

⁵ Searching the LexisNexis article database gives the first mention of Prop 209 in July 1996.

⁶ Calculation made using UC application data (http://www.ucop.edu/news/archives/2001/applications_2001/table1.pdf) and California high school graduation data (http://www.cpec.ca.gov/completereports/2003reports/03-09/Display1.PDF).

⁷ The UC uses the term underrepresented minority to refer to blacks, Hispanics and Native Americans.

⁸ For example, “In an attempt to improve minority access to UC without the help of affirmative action, the university’s investment in kindergarten-through-12th-grade outreach has rocketed from about $60 million in 1995 to $180 million last year and a planned $250 million this year” (2000, January 21). UC Regents Urged to Step Up Minority Outreach at Schools. The San Francisco Chronicle.

⁹ At the UC, URMs include Hispanics, blacks and Native Americans.

¹⁰ In principle, these two states could be included in our base specification estimated by Equation 1. However, including these states in the permutation tests we discuss below is problematic because the limited number of minority students cause wildly variable estimates within those states.

¹¹ Adding Asians to the excluded group in the DDD regressions gives similar results.

¹² When estimating β₀₁ for states other than California, California is treated as a control state.

¹³ These states include Alabama (2002), Arizona (2010), Georgia (2002), Michigan (2006), Nebraska (2008), and Oklahoma (2012).

¹⁴ In Texas, students could attend a university of their choice while in Florida, admission was guaranteed to at least one public university.

¹⁵ For example: Healy, Patrick (1998, April 24). Affirmative Action Survives at Colleges in Some States Covered by Hopped Ruling. The Chronicle of Higher Education. Retrieved from http://www.chronicle.com.

¹⁶ The results for the black-white and Hispanic-white gaps presented in this paper are similar when including Texas and Washington with the treatment group.

¹⁷ In April 1995, the College Board recentered the SAT score scales to reestablish a mean score of about 500. To ensure consistency over time, we use College Board-provided recentered scores for all years.

¹⁸ The other control states are similar to those shown here but are not displayed in the figure due to lack of clarity when too many states are plotted at once.

¹⁹ The relative ranking of the states shown in Figure 3 is largely driven by SAT participation rates, with low participation rates generally corresponding to high average scores. For example, Alabama has relatively high average SAT scores despite its poor performance on most standardized tests, such as the NAEP, because only about 10% of high school graduates in Alabama take the SAT.

²⁰ We also investigated children of parents who attended at least some college. Unsurprisingly, these results are similar to the full sample since about 80% of the sample has at least one parent who attended at least some college.

²¹ Published by the National Center for Education Statistics, Institute of Education Sciences, U.S. Department of Education. Washington, DC. Available at http://nces.ed.gov/programs/digest/; Accessed: 05/18/2012.

²²http://www.bls.gov/nls/nlsfaqs.htm#anch14.

²³ The authors also obtained restricted-use data from the eighth grade math sample of the National Assessment of Educational Progress (NAEP), which has the advantage of being designed to be representative at the state level. However, the NAEP suffers from two serious drawbacks. First, background information is relatively sparse: for example, parental education but not income is available. Second, smaller sample sizes result in estimates that are considerably less precise than those from the College Board data.