Black Suits, Gowns, & Skin: SAT Scores by Income, Education, & Race

People with highly educated or wealthy parents score higher on the SAT than those from poor, uneducated families. Obvious statistic is obvious, but how important are dollars and degrees compared to race? The College Board, the organization that oversees the SAT, holds tight to its information on the subject, but incomplete leaks have occurred for 1995, 1997, 2003, and 2008. 1995’s top income bracket only started at $70,000, so the wealthiest African-American students that year did not outscore even the poorest white students. As shocking as that fact is, it provides no controls for confounding variables and neglects the currently largest minority, Hispanic Americans. Therefore, I decided to approach the question using multiple linear regression of state data. ("M" in the graph below stands for the math subtest. "V" stands for the critical reading or verbal subtest.)

First, I shall review the important news from this year’s SAT and ACT score reports. I used the ACT-to-SAT conversion equation that I extracted from the conversion table to construct a summary graph of overall ACT-SAT scores for each race and gender. Asian scores continued to rise, despite the College Board’s South Korean crackdown, which was based upon suspicions of widespread cheating. Meanwhile, overall scores fell, and whites, Native Americans, and men declined this year more than any of the previous sixteen available years.

Native Americans now barely score higher than Hispanic Americans. Native-American ACT scores slipped especially fast, and their average score on the optional ACT writing exam now equals those of African Americans. Since many white people have Native-American ancestry, and Canadian and US Native Americans tend to have high amounts of European ancestry, Native-American score trends could reflect changing cultural attitudes about racial identification, but their absolute number of test participants has not changed greatly.

I previously commented on evidence for possible white decline. So far, the evidence is subtle. If future scores demonstrate long-term decline, it could signify the “dumbing down” of education or culture, dysgenics, minority-centered education reforms, or low rates of whites taking test preparation services.

Because Asian score increases have been steady and measured, I believe that these represent genuine progress, even with or perhaps due to the same root causes as the reported cheating scandals among Asians. However, I suspect that Asian progress will eventually level off because most racial score gaps have stayed remarkably constant, and I think nature influences testing potential, especially among those of adequate means.

SAT annual reports provide scores of students grouped by their parents’ income and education levels. These graphs of that data should not surprise anyone.

The fluctuation of income categories tells an interesting story about the past two decades. The number of people with six-figure incomes took off not that long ago.

Despite the fact that levels of (parents’) education have been trending upwards, coming from an educated family increasingly predicts obtaining a higher SAT score. This graph of educational advantage is a Cohen’s d graph with the vertical axis zoomed in. I defined those parents with a bachelor’s or graduate degree as “educated.” (This graph uses the older term for the critical reading subtest, verbal.)

Parents’ education and income show clear links to SAT scores, but so do many other variables. When I mapped state SAT scores, I discovered that Midwestern states achieved higher scores than other states because a small percentage of studious Midwesterners took the SAT, and most other college-bound students in those states only took the ACT. Simple linear regression of state SAT scores with only participation rate as the predictor variable explains 78 percent of variance (P = 10^-273). Usually psychology research conducts this kind of analysis with a sample population, but I cannot access the College Board’s raw data, obviously. I am trying to reverse-engineer an SAT database with 16 years of state data. However, a single data point for an income category could represent tens of thousands of people or as few as three. Plus, the year could potentially influence scores due to inflation or even societal changes in cognitive abilities, the so-called Flynn effect. I can appropriately control for those variables, but controlling for race would require estimation because racial group proportions are not broken down every year within each income bracket and educational degree. Therefore, I turned income and education levels into continuous variables for multiple linear regression.

For linear regression, I turned income into a continuous variable by dividing the number of students whose parents earn six-figure incomes by the number whose parents earn less than $20,000 per year. I compared results for this income “gap” to an income “divide” based on the number whose parents earn more than $60,000 per year divided by those whose parents earn less. I created a variable for education based on the number whose parents achieved at least a bachelor’s degree divided by the number whose parents did not. The racial variable was the number of whites and Asians divided by the number of another race. Simple linear regression for state population size showed that it was a significant predictor of SAT scores (P=4.3 x 10^-47) that explained 22 percent of variance. However, it became insignificant with all further analysis except when the other predictor variables were either participation, year, and income gap or participation, year, education, and income divide with or without race. Year did not produce a significant P value for simple linear regression (P = 0.5) but always did in multiple linear regression of income or education, lowering scores over time like a reverse-Flynn effect. Both income gap and income divide were significant predictor variables in multiple linear regression, but income divide slightly better predicted SAT scores, explaining 86 percent of variance with year and participation variables, compared to 85 percent for income gap with year, size, and participation variables. Multiple linear regression with participation rate, year, income divide, and education explained 90 percent of variance with all variables achieving statistical significance, but the addition of education caused the P value of the income divide to worsen from 10^-83 to 10^-3. When I added race as an additional variable, the income divide was no longer a significant predictor of SAT scores (P = 0.4). Most of the impact of income on SAT scores stems from its ability to predict parents’ education levels. Multiple linear regression with the remaining significant variables (state sample size, participation rate, year, education, and race) explained 92 percent of SAT variance.

Graphs of actual SAT scores for income brackets and education levels show the distinctiveness of children whose parents have graduate degrees. Trendlines are given with R-squared variances for the highest and lowest categories.

The graphs for income or education with race reach the provocative result that race affects scores more among the lower rungs of society. As the data only represents state racial proportions, the results leave room for debate. For instance, the furthest left data in the graphs below represent Washington, DC, which I actually treated as a state. Higher-income families there are probably more likely to be white or Asian-American than are lower-income families. Many racially diverse states like California have industries or attractions that pull in successful, educated whites. Nevertheless, one could use these results to defend upper-class affirmative-action beneficiaries or to call for a new class-based system to benefit the many poor, but intelligent whites and Asian Americans.

Also of note is the fact that these graphs look totally awesome.










Anonymous (1998). Why Family Income Differences Don't Explain the Racial Gap in SAT Scores The Journal of Blacks in Higher Education (20), 6-8 DOI: 10.2307/2999198

Anonymous (2008). Why Family Income Differences Don't Explain the Racial Gap in SAT Scores The Journal of Blacks in Higher Education (62), 10-12

Ezekiel Dixon-Roman, Howard Everson, & John McArdle (2013). Race, Poverty and SAT Scores: Modeling the Influences of Family Income on Black and White High School Students’ SAT Performance Teachers College Record, 115 (4), 1-33

The SAT-ACT Score Map

(Note: Following this post, I shall focus on finishing the MAOA bibliography probably until it is up-to-date and until I can better quantify its data.)

In my last post on college-entrance exams, I left incomplete the task of properly controlling for test participation a state map of combined SAT and ACT scores. I had already explored group average SAT gaps by race and gender and SAT score distributions. Finally, I am posting above what I consider the definitive state map, which is properly controlled for test type and state participation levels.

The map demonstrates my contention that American demographic changes contribute to a North-South educational divide. Detailed mapping of potential academic decline can help inform discussion of policies like “immigration reform,” help extrapolate future global competitiveness of the American workforce, and delineate regional economic fault lines. In the explanation that follows, I compare the effects of state participation in the SAT and ACT and race with regression analysis. Then, I shall review an important study on the relative importance of these scores and what might augment or replace them.

Testing associations publish a standardized table to convert between ACT and SAT scores. The primary table converts between the composite ACT score and the combined SAT mathematics and critical-reading (formerly verbal) score. (I divided the scores in half for purposes of comparison with SAT subtest scores.) A separate table converts between the newer SAT writing score and the score for the optional ACT writing exam. However, I shall neglect the writing scores data at this time, but the previous post maps raw SAT writing scores.

My last attempt at ACT-to-SAT score conversion amounted to a crude estimate that only accounted for the highest and lowest ACT scores with a line drawn between for all others. The tests follow different scales with several possible SAT scores coinciding with almost all ACT scores. Therefore, I created a new formula based on linear regression of the plot of average SAT scores for each ACT increment. Since state ACT averages tend to hover around 21, this graph illuminates how my previous formula unfairly underestimated states that emphasize the ACT over the SAT.


As states increasingly have required the ACT for all high-school graduates, their average scores have declined. Plotting below each state’s yearly ACT and SAT score since 1998 by participation level confirms the association. All associations achieve statistical significance (P=9.54 x 10^-27 for the ACT, P=3.75 x 10^-33 for the weighted SAT-ACT scores) with the SAT, alone, achieving the greatest significance of any tested relationship in this entire effort (P=5.4 x 10^-245) and the largest coefficient of determination (0.769). Midwestern states that strongly emphasize the ACT achieved impressive average SAT scores and seem to have an outsized impact on this finding. The combined SAT-ACT participation rates are out of a possible 200%, which would require all high-school graduates to take the SAT and the ACT.


The comparison of score maps to demographic trends, which I presented with a map of the percentage of SAT examinees who are white or Asian, fits the familiar national racial group mean gaps. Simple linear regression better quantifies the effect of race, and multiple linear regression can tease apart the effect of state participation levels. Asians have the highest scores, but I lumped whites with Asians because Asians are a relatively small group. I would expect other racial gaps to be too confounding to separate Asians from the rest. The distinction I drew might seem arbitrary, but many institutions separate data on Asians and whites from that of “underrepresented minorities.” The graphs of scores by racial proportion appear to show that these are linear associations, all of which are significant (P=3.51 x 10^-53 for the SAT, P=1.33 x 10^-41 for the ACT, and P=1.48 x 10^-40 for the tests combined). Multiple regression shows that all associations remain significant. For the combined SAT-ACT score, participation had a P-value of 3.35 x 10^-22, and race had a P-value of 1.29 x 10^-29. The multiple regression model set score equal to 986 – 0.782 x participation (as a whole number) + 97.5 x the percentage white or Asian.


The residuals from subtracting this model from the raw data fit a Gaussian distribution, as expected. So, I recalculated the SAT-ACT composite scores by adding the residuals back to the model under the assumption of 100% (out of a possible 200%) participation.


While the score map might not appear identical to the map of demographic trends, one can make out a North-South gradient, and state efforts to adopt test requirements seem well-controlled. Further analysis could use ANCOVA for income categories or compare racial gaps within states. Composite writing scores could prove useful, despite the shorter timeframe for the SAT and the optional nature of the ACT written exam. In fact, the states of Texas, Nevada, and Florida might not seem to perform so badly in this map, given their diversity, but their raw SAT writing scores were especially low. Then again, immigration weighing down English writing skills could resolve with acculturation.

Now, I wish to review a 2009 study on the relative relevance of SAT and ACT scores. Schmitt et al compared the predictive value of those scores to high-school grades and twelve “noncognitive predictors”: knowledge, curiosity, adaptability, perseverance, ethics (“not cheating”), career orientation, healthy behaviors, interpersonal skills, “leadership,” community volunteer activities, “artistic and cultural appreciation,” and “appreciation for diversity” (“e.g. by culture, ethnicity, religion, or gender”). As the US Supreme Court revisits the issue of Affirmative Action in college admissions, universities might apply such predictors to lessen the influence of standardized tests. None of the “noncognitive predictors” could predict college grades even half as well as either high-school grades or the SAT/ACT scores, which had correlations of 0.531 and 0.539, respectively. Knowledge came closest, but I think knowledge is cognitive. Career orientation actually was alone in its statistically significant negative association with college grade-point average. The authors offered as their only explanatory hypothesis the poor performance of African Americans “for whom career mobility and a career orientation was a major reason for college attendance.” Indeed, career orientation was the strongest advantage for African Americans, who barely scored higher than whites on “appreciation for diversity.” Their only other advantage was perseverance. High-school grades underestimated African-American college grades, but not as much as SAT/ACT scores overestimated. Adding the “noncognitive” criteria to potential admissions selection would lower college grades, in general, but raise African-American and Hispanic-American admissions at the 15% most exclusive universities, at the expense of white and especially Asian-American applicants. However, African-American college graduation rates would fall eight percentage points at such institutions. SAT/ACT scores were significantly associated with higher college classroom absenteeism and lower “organizational citizenship behavior,” with which “appreciation for diversity” had a significant positive association. In other words, the more intelligent students were less inclined to go to all lectures, promote “the university to outsiders,” defend “it against criticism,” and participate “in student government or other clubs” to make the university “a better place.” The authors did not conclude that those are relatively unintelligent behaviors.










Schmitt N, Keeney J, Oswald FL, Pleskac TJ, Billington AQ, Sinha R, & Zorzie M (2009). Prediction of 4-year college student performance using cognitive and noncognitive predictors and the impact on demographic status of admitted students. The Journal of applied psychology, 94 (6), 1479-97 PMID: 19916657