Last Updated on March 18, 2023
In this post, I will review studies comparing the predictive validity of cognitive ability and parental socioeconomic status (SES) on academic achievement. I’ll start with some important statistical background that is necessary to understand the studies cited in this post. After those preliminaries, I start by reviewing meta-analyses reporting the correlations between parental SES and achievement and between cognitive ability and academic achievement. Next, I review individual studies investigating the relative predictive validity of ability and parental SES on grades. The final section will review individual studies investigating the relative predictive validity of ability and parental SES on test scores. A clear picture emerges from each study cited in this post: cognitive ability is a far better predictor of academic achievement than is parental SES.
Before reviewing the main studies for this post, some important statistical background is necessary. To compare the predictive validity of cognitive ability and parental SES, I will rely heavily on correlation coefficients and standardized regression coefficients. Therefore, in order to understand this post, it is essential that one understand these statistical concepts at least at a high level. I will briefly outline these concepts in this section (as well as clarify the measure of parental SES). If you don’t have at least a rough understanding of these concepts, then you probably shouldn’t even bother with the rest of the post.
At a high level, correlation coefficients quantify the statistical association between two variables. A correlation coefficient can range from -1 to +1. A correlation coefficient of 1 between X and Y indicates that all values for Y can be predicted perfectly from a linear relationship with the values for X (and vice-versa). A correlation coefficient of 0 indicates no statistical association whatsoever.
For someone unfamiliar with differential psychology, it may be unclear whether a given correlation coefficient in this field is large, small, or medium. I recommend such readers to check out this post where I provide empirical data on the distributions of correlation coefficients within the field and I also list the correlation coefficients for many commonly understood variables. Following the standards mentioned in that post, one can treat low, medium, and large correlations as correlation coefficients in the ranges of |r| < .15, .15 < |r| < .30, and |r| > .30, respectively.
Now, the purpose of this post is primarily statistical rather than causal, i.e. I’m concerned with the predictive power of various factors (i.e. cognitive ability and parental SES) on academic achievement. However, because I’m concerned with comparing the validities of these different predictors, it is problematic to rely solely on correlation coefficients. This is because the predictive validity of one predictor may be due to confounding with the other predictor. This is particularly a problem for my purposes because parental SES and offspring cognitive ability are correlated.
For example, imagine that parental SES and achievement correlate at r = 0.30 whereas cognitive ability and achievement correlate at r = 0.40. This may give the impression that parental SES is almost as important as cognitive ability in predicting achievement. However, it could be the case that parental SES correlates with achievement only because parental SES also correlates with cognitive ability, which has an effect on achievement. It is perfectly possible that parental SES has a minor association with achievement when holding cognitive ability fixed, whereas cognitive ability maintains a strong association with achievement even when holding parental SES fixed. In this hypothetical scenario, we would want to say that cognitive ability is a much more important predictor than parental SES, even though comparing correlation coefficients would give a different impression.
To avoid this problem, we cannot rely solely on correlation coefficients. We also want to quantify the independent associations of cognitive ability and parental SES on academic achievement. By this, I mean the associations of each predictor on academic achievement while holding the other predictor fixed. This is where regression coefficients come in.
To quantify the independent associations of parental SES vs cognitive ability on academic achievement, I will reference studies that model academic achievement using regression analysis. In a regression, the dependent variable (in this case, academic achievement) is modeled as a function of independent variable(s) simultaneously (in this case, parental SES, cognitive ability, and potentially other variables, depending on the study). In a regression analysis, each independent variable is assigned a regression coefficient which quantifies the statistical association between that independent variable and the dependent variable, while holding all other independent variables fixed.
One problem is that regression coefficients are often unstandardized. An unstandardized regression coefficient indicates the change in the dependent variable associated with a one-unit increase in the independent variable. This makes comparisons of coefficients meaningless because different coefficients will have different units. For example, let’s say we have a regression model with SAT score as the dependent variable with the following 2 independent variables: parental education (in years of schooling) and parental income (in dollars). Let’s say the unstandardized regression coefficient for both independent variables is 5. This would indicate that an additional year of parental education is associated with an additional 5 SAT points, and an additional dollar of parental income is also associated with 5 SAT points.
In this highly unrealistic hypothetical, it is obvious that parental income has a far greater association with SAT scores (even though both independent variables have an unstandardized coefficient of 5), because there is far more variation in dollars of parental income than there is in years of parental education. In other words, a one dollar difference in parental income is far more common than a one year difference in parental education. So the typical variation in parental income is associated with a far greater change in SAT scores than the typical variation in parental education (because a one dollar change in parental income is associated with the same increase in SAT scores as a one year change in parental education). Therefore, even though it is obvious that parental income has a far greater association with SAT scores than does parental education in this hypothetical, this difference is not apparent from comparing unstandardized coefficients.
It is clear that comparing unstandardized regression coefficients is meaningless. But we also cannot rely on comparing correlation coefficients for the reasons given earlier. Therefore, to compare the independent effects of different predictors, I will primarily rely on comparing standardized regression coefficients. Standardized regression coefficients quantify associations not in terms of the units of either set of variables. Instead, it quantifies associations in terms of standard deviations. Standardized regression coefficients must fall between -1 and +1, like correlation coefficients. More specifically, a standardized regression coefficient indicates the change in one-standard-deviation of the dependent variable associated with a one-standard-deviation increase in the independent variable.
For example, a standardized coefficient of 0.3 indicates that a 1 standard deviation change in the independent variable is associated with a 0.3 standard deviation increase in the dependent variable. This avoids the problem with unstandardized regression coefficients mentioned earlier. For example, return to the example with parental income and parental education from above. The key problem with comparing unstandardized coefficients for parental income (in dollars) and parental education (in years) was that there is far more variation in the units for former than the latter. But this is not a problem when we use standard deviations instead of the original units. Whereas a one dollar difference in parental income is far more common than a one year difference in parental education, a one standard deviation difference in parental income is just as common as a one standard deviation difference in parental education.
the position of an individual or group on the socioeconomic scale, which is determined by a combination of social and economic factors such as income, amount and kind of education, type and prestige of occupation, place of residence, and—in some societies or parts of society—ethnic origin or religious background.
In 2007, the APA published “Report of the APA Task Force on Socioeconomic Status” stating that “At the individual level, most research on the effects of social stratification has used educational attainment, income (personal or household), and/or occupation as indicators of SES” (page 9). A number of other studies have agreed that this tripartite model is the traditional measure of SES (Adler and Newman 2002, Bradley and Corwyn 2002, Winkleby et al. 1992). SES has also been measured using these three components in a number of meta-analyses studying the relationship between SES and a number of social outcomes, including future SES (Strenze 2007, page 406), self-esteem (Twenge and Campbell 2002), well-being (Pinquart and Sörensen 2000), and countless others.
It seems fair to say that SES is traditionally defined as a combination of education, income, and occupation. In fact, the following meta-analyses agree that these three factors are traditionally used to measure parental SES:
- In a meta-analysis on the relationship between SES and academic achievement, Sirin (2005) [archived] reports that “there seems to be an agreement on Duncan, Featherman, and Duncan’s (1972) definition of the tripartite nature of SES that incorporates parental income, parental education, and parental occupation as the three main indicators of SES” (page 418).
- In another meta-analysis on the relationship between peer socioeconomic status on student achievement, Ewijka and Sleegers (2010) [archived] also states that “There also seems to be agreement on a three-componential view of SES which states that SES can be indicated by either parental education, parental occupation, or parental income” (page 138).
- In another meta-analysis on the relationship between socioeconomic status and college admissions tests and academic performance, Sackett et al. (2009) reports “There is no uniform agreement on measurement of this construct, although most studies have focused on some combination of three measures: parental education, parental income, and parental occupational status” (page 3).
In this post, almost all of the studies use some variation of this tripartite measure of SES, i.e. they use some combination of parental income, parental occupational status, and parental education. Most of the studies cited below measure parental SES using at least parental income and parental education. Some studies also include parental occupational status as well. One might complain that this measure of parental SES is too narrow (e.g., it doesn’t include wealth). I address this objection in the conclusion of the post.
Before analyzing any individual studies, it is important to review the meta-analytic literature to get a handle on the commonly reported associations between parental SES and cognitive on academic achievement. I will begin by reviewing meta-analyses for parental SES. Then I will consider meta-analyses for cognitive ability.
Meta-analyses reveal that the correlation between parental SES and offspring achievement is fairly modest. A recent meta-analysis by Harwell et al. (2016) reported a correlation of just r = 0.22 between parental SES and offspring achievement from approximately 300 effect sizes across over 100 studies. The observed correlations were also much lower for older children. The correlation between parental SES and achievement was r = 0.33 for students in kindergarten, r = 0.23 for students in elementary school, and r = 0.16 for students in middle and high school (Table 2).
The authors note the following as the most important finding of the meta-analysis:
The most important finding was that the SES-achievement relationship assuming a random-effect model was relatively weak (M = .22), which was similar to White’s (1982) average ES using unweighted correlations (M = .24). Thus our results generally confirm White’s conclusion that the average SES-achievement correlation is weak. The average ES for a fixed-effect model was even smaller (M = .09) as a result of a small number of studies in our sample with quite large sample sizes and quite small SES-achievement correlations. For example, Coleman et al. (1966) used a national stratified sample of approximately 300,000 students and produced values corresponding to the first, second, and third quartiles of the distribution of unweighted Pearson correlations between SES and achievement of .06, .13, and .21, respectively. These findings undermine empirical rationales for including SES measures in analyses of achievement data with the goal of increasing statistical power, controlling for the effects of SES, and enhancing causality arguments.
There were not large differences in correlation based on the measure of parental SES. The correlation between achievement and parental SES were all modest, whether parental SES was measured using parental income (r = 0.26), parental education (r = 0.17), parental occupation (r = 0.21), or a composite (r = 0.23).
Importantly, parental SES had slightly larger correlations with offspring IQ (r = 0.27) than with GPA (r = 0.14). These findings were also reported in Table 2:
For older meta-analyses on the correlation between parental SES and achievement, see White (1982) and Sirin (2005). See Kim et al. (2021) for a meta-analysis in developing countries. See Korous et al. (2021) for a review of meta-analyses.
For a fairly new meta-analysis investigating this question, see Selvitopu and Kaya (2021). These authors meta-analyzed 62 samples from 48 studies that were published between 2010 and 2019. In line with the results from Harwell, they found an average correlation between parental SES and achievement of r = 0.25 (see page 5). The results of the moderator analysis were presented in Table 3.
- Note: the effect sizes are Pearson correlation coefficients (r) (see page 4).
Again, the results here are mostly in line with the results presented by Harwell. The correlation between parental SES and achievement in the United States is about the same as in the overall dataset (r = 0.23). Also, the correlation is the weakest when achievement is measured using GPA (r = 0.17).
von Stumm et al. (2022) [archived] examined the association between parental SES and primary school performance across 16 British cohorts and over 90k children from 1921 to 2011. The authors found that the average correlation between parental SES and school performance was r = 0.28, in line with the findings reported from earlier meta-analyses. The authors also reported that the magnitude of the correlation was incredibly stable throughout all cohorts. They note there was no “systematic increasing or decreasing trends in the strength of the association could be observed across the past 95 years”. The following figure shows trends in the correlation between parental SES and school performance over time. The second graph is limited to cohorts with over 1,000 participants born between 1946 and 2011:
The average correlation between parental SES and school performance in the analysis with > 1,000 participants was fairly similar to the analysis with all cohorts (r = 0.30 vs r = 0.28).
In contrast to parental SES, meta-analyses demonstrate that the correlation between school grades and offspring intelligence is rather large. A meta-analysis by Roth et al. (2015) [archived] reported a correlation of r = 0.54 between Intelligence and grades after correcting for sampling error, measurement error, and range restriction.
Even without these corrections, the reported correlation between intelligence and grades (r = 0.44) was much greater than that reported between parental SES and grades in the other meta-analyses (r = 0.14 or r = 0.17). The findings from this meta-analysis led the authors to the following conclusions:
The results of our study clearly show that intelligence has substantial influence on school grades and thus can be regarded as one of the most (if not the most) influential variables in this context. Although intelligence turned out to be a significant predictor on all moderator levels, we were able to identify some scenarios in which even higher validities can be obtained. First of all, the population correlation was highest for tests relying on both verbal and nonverbal materials, indicating that a broad measure of intelligence or g respectively is the best predictor of school grades. Furthermore, the importance of intelligence increases throughout grade levels. This leads us to the conclusion that intelligence has special importance in educational contexts which deal with content that is more complex and thus can be mastered fully only with an appropriate cognitive ability level.
Cognitive ability is also a much better predictor of standardized achievement test scores. For example, a meta-analysis by Zaboski et al. (2018) [archived] found that g was significantly correlated with academic achievement (average r^2 = 0.54, so r = 0.73). Assessments of cognitive ability and academic achievement were done mostly using the Woodcock-Johnson Tests of Cognitive Abilities and Woodcock-Johnson Tests of Achievement, respectively. The authors limited assessments of achievement to tests measuring one of 4 achievement areas: basic reading, reading comprehension, basic math skills, or math reasoning. The average number of studies for each achievement area was about K = 4.9 and the average sample size was N = 7,079. The average correlation between g and achievement ranged from r = 0.72 to r = 0.76. See Table 3 to summarize the findings:
Similar results were found in a recent meta-analysis by Kriegbaum et al. (2018) on the importance of intelligence and motivation as predictors of school achievement. Across 74 studies and over 80,000 subjects, the authors reported an average correlation of r = 0.44 between intelligence and school achievement. After correcting for measurement error and range restriction, this correlation rose to over r = 0.60. School achievement was measured as a combination of grades and achievement tests.
Conclusion: These meta-analyses show that intelligence is a far better predictor of academic achievement than parental SES. To summarize the above findings on grades and achievement test scores:
- Grades: parental SES correlates with high school grades at about r ~ 0.15. By contrast, cognitive ability correlates with grades at about r ~ 0.50, slightly higher or lower depending on the controls added. Thus, the meta-analytic data suggests that the correlation between cognitive ability and grades is about 3 to 4 times the correlation between parental SES and grades.
- Achievement test scores: Parental SES correlates with achievement test scores at somewhere between r = 0.20 and r = 0.30. By contrast, cognitive ability correlates with achievement test scores at about r = 0.73. Thus, the meta-analytic data suggests that the correlation between cognitive ability and achievement test scores is about 2 to 3 times that of the correlation between parental SES and achievement test scores.
Individual studies: grades
The meta-analytic data shows that the correlation between cognitive ability and achievement in the literature is far greater than the same correlation between parental SES and achievement. While these findings are highly important, it is also important to review various individual studies to truly understand the relative predictive validity of cognitive ability and parental SES. There are a few reasons for this.
First, reviewing individual studies allows one to dive into the details of a few high quality studies, which is arguably equally as important as understanding meta-analytical findings. Second, analyzing individual studies allows us to ensure that the effect sizes for parental SES and cognitive ability are derived from the same sample. Third, another benefit is that this allows us to examine the degree to which parental SES and ability predict achievement while holding fixed the other predictor (which is important, because parental SES and cognitive ability are correlated). I will begin by considering individual studies that report zero-order correlations with the variables of question (i.e. correlations without controlling for any other variables).
Zero order correlations
For example, Sackett et al. (2009) [archived] reported on the relationship between parental SES, SAT scores, high school grades, and college grades for college freshmen from three entering cohorts (1995, 1996, and 1997) in 41 colleges and universities. For now, I’ll focus on the findings reported on parental SES and high school grades. The 41 colleges were geographically diverse, included small and large schools, public and private institutions, and a broad range of selectivity on SAT scores. The mean freshman class size was 1,369 students, so the total sample was well over 50k students. Parental SES was measured based on student-reported parental education and family income. The observed correlation between parental SES and high school grades was r = 0.19 after correcting for range restriction (Table 8).
Similar findings were reported by Westrick et al. (2015). These authors analyzed the impact of parental SES, ACT scores, and high school grades on college grades. For now, I’ll focus on the reported correlations between parental SES and high school grades. The authors analyzed background factors for all students who took the ACT from 1999 to 2006 (page 28), resulting in a sample size exceeding 6 million. Parental SES was measured using self-reported measures of parental income. The authors reported that parental SES and high school grades correlated at r = 0.20 (Table 3), which is almost exactly the same as the findings of the Sackett study.
For more examples of the association between cognitive ability and grades, see Cucina et al. (2016). These authors conducted two studies on this association for high school students. Cognitive ability in both studies was measured based on the general factor of intelligence (g) extracted from a battery of tests of different mental abilities.
- The first study uses the base-year dataset from Project TALENT, which contains data on over 300,000 high school students conducted in the 1960s. The correlation between g and standardized GPA varied between r = 0.37 and r = 0.40 (Table 3), depending on whether g included measures of crystalized intelligence or not. These correlations rose to r = 0.40 and r = 0.43 after correcting for predictor and criterion unreliability (see ρTS from Table 3).
- The second study uses data from the 1997 cohort of the National Longitudinal Study of Youth (NLSY97), a longitudinal survey following representative samples of thousands of participants in their youth starting in the 1997. The correlation between g and GPA was r = 0.44 (Table 6). This correlation rose to r = 0.47 after correcting for predictor and criterion unreliability (see ρTS from Table 6).
Note that these correlations were not adjusted for range restriction, so are likely underestimates. That said, these findings are in line with the average uncorrected correlation (r = 0.44) between intelligence and high school grades as reported in the meta-analysis by Roth et al. (2015). Recall that this correlation rose to r = 0.54 after controlling for sampling error, measurement error, and range restriction in the meta-analysis.
Given the fact that both parental SES and cognitive ability are correlated with grades, and the fact that both of these variables are correlated with each other, it is reasonable to assume that much of the association between parental SES and grades will be due to confounding with cognitive ability (and vice-versa, swapping parental SES and cognitive ability). Therefore, it will be useful to consider studies that attempt to estimate the independent associations of parental SES and cognitive ability with grades, i.e. the associations between each predictor variable and grades that are independent of the other predictor variable. I will consider such studies in this section.
The first study to consider that performs this analysis comes from Zisman and Ganzach (2022) [archived]. These authors examined six databases to compare the predictive validity of intelligence vs personality in predicting important life outcomes. While their primary purpose was to compare intelligence and personality, 3 of the datasets allowed comparisons between intelligence and SES as predictors of academic achievement, so they can be included for my purposes here. The 3 datasets are as follows:
- The NLSY79 and NLSY97 are longitudinal surveys following representative samples of thousands of participants starting in their youth with recurring interviews extending until middle-age. The NLSY79 followed a sample of over 10,000 participants since 1979, when participants were aged 14-22 years old. The NLSY97 followed a slightly smaller sample since 1997, when participants were 12-17 years old. Cognitive ability was measured using the Armed Forces Qualifying Test (AFQT) scores. Socioeconomic background (SEB) was measured as a composite of both parents’ income and educational attainment.
- The Wisconsin Longitudinal Study (WLS). The WLS contains data on over 10,000 individuals aged 17-20 when the study started in 1957. These participants accounted for one-third of all seniors in Wisconsin high school. Cognitive ability was measured using the Henmon-Nelson Test of Mental Abilities. Socioeconomic background (SEB) was measured using Duncan’s socioeconomic-index (SEI), a measure of occupational status.
Academic achievement was measured by high school GPA on a 4-point grading scale in all 3 datasets. All 3 datasets showed intelligence to be a much greater predictor of GPA than family SEB. This is shown when comparing pairwise correlation coefficients of GPA with intelligence vs with SEB:
- In the NLSY79, intelligence correlated with GPA (r = 0.42) much more than did SEB (r = 0.21) (Table A.1).
- In the NLSY97, intelligence correlated with GPA (r = 0.52) much more than did SEB (r = 0.33) (Table A.2).
- In the WLS, intelligence correlated with GPA (r = .50) much more than did SEB (r = 0.12) (Table A.5).
The average correlations across all datasets are presented here:
Notice that the average correlation between GPA and intelligence (r = 0.48) is over twice that of the average correlation between GPA and socioeconomic background (r = 0.22), in line with the previous results. The disparities become even greater when comparing the standardized coefficients of intelligence and SEB when they are considered simultaneously in the same model. Consider the following results:
- In the regression model containing both intelligence and SEB in the NLSY79, the standardized coefficient for intelligence (β = 0.48) was far greater than the coefficient for SEB (β = -0.02) (Model 3.1 of Table D.1).
- In the regression model containing both intelligence and SEB in the NLSY97, the standardized coefficient for intelligence (β = 0.47) was far greater than the coefficient for SEB (β = 0.13) (Model 3.1 of Table D.2).
- In the regression model containing both intelligence and SEB in the WLS, the standardized coefficient for intelligence (β = 0.49) was far greater than the coefficient for SEB (β = 0.001) (Model 3.1 of Table D.5).
Similar findings were reported in a recent study by Marks (2022) [archived]. The author also analyzed the predictive validity of cognitive ability and family background on various outcomes (e.g., GPA, test scores, wages, educational attainment, etc.) in the NLSY79 and NLSY97. For the purposes of this section of the post, I will focus on GPA (I will mention test scores in later sections). Parental SES was measured using family income and parental education. However, the author also performed supplementary analysis with “extended models” that also incorporated parental occupation and wealth.
Now, recall that these two datasets – the NLSY79 and NLSY97 – were both analyzed by Zisman and Ganzach. This study differs from the Zisman study in two important ways. Firstly, Marks measures cognitive ability by extracting general ability g from 10 ASVAB subtests. By contrast, Zisman and Ganzach measure cognitive ability simply as the summed score from 4 subtests of the ASVAB (i.e., the AFQT). Secondly, the sample size analyzed by Marks was greater than that analyzed by Zisman and Ganzach, probably because Zisman and Ganzach included a larger number of variables in their models (e.g., personality traits), which was likely only available for fewer participants.
Now, let’s start by noting the correlations between the relevant variables of interest:
- Note: the figures below the diagonal are for the NLSY79 whereas the figures above are for the NLSY97.
First, compare how these correlations relate to the meta-analyses cited earlier. The association between cognitive ability and grades is in line with the previously cited meta-analytically derived figures; the correlations of r = 0.60 (NLSY79) and r = 0.44 (NLSY97) are consistent with the findings from Roth et al. (2015), which found an IQ-grades correlation of r = 0.68 for studies published before 1983 and r = 0.47 for studies published afterwards (Table 1). The associations between parental SES and grades are actually slightly greater than that reported in the meta-analyses cited above; the SES-grades correlations varied between r = 0.11 and r = 0.29 depending on the measure used, whereas Harwell et al. (2016) reported an average correlation of just r = 0.14 between parental SES and grades.
The following table shows the independent associations between each predictor and grades when they are simultaneously entered in the same model:
There are a few points to explain in order to read this table. Each regression model represents a different combination of independent variables. For example, model 1 measures the impact of parental education and family income on grades. The “R square” value indicates the percentage of variance in the dependent variables (in this case, grades) that is (statistically) explained by the independent variables in the models. The values under “Est” indicate the unstandardized coefficients for each independent variable. The values under β indicate the standardized coefficients for each independent variable.
Now, before comparing the relative predictive validities of cognitive ability and parental SES, it is worth noting that neither of the measures of parental SES has a consistent statistically significant association with grades after controlling for cognitive ability. See model 3 in the NLSY79 and NLSY97. This model shows that, when controlling for cognitive ability, parental education is not significantly associated with high school grades in the NLSY79, whereas family income has no significant association with grades in the NLSY97.
That being said, there are a few different ways to compare the relative predictive validity of cognitive ability and parental SES:
- Compare the standardized coefficients of each independent variable in model 3. Both the NLSY79 and NLSY97 indicate that a one standard deviation change in cognitive ability (holding fixed the other variables) is associated with a far greater change in grades than a one standard deviation change in either parental education or family income (holding fixed the other variables). For example, in the NLSY79, the standardized coefficient of cognitive ability (β = 0.58) is nearly 10 times the coefficient for family income (β = 0.06). In the NLSY97, the standardized coefficient for cognitive ability (β = 0.41) is nearly 6 times the coefficient for parental education (β = 0.07).
- Compare the R^2 of models 1 and 2. Start with the NLSY79. The model with both parental education and family income (model 1) explains only 12% of the variance in school grades, whereas the model with just cognitive ability explains 37% of the variance (model 2), showing the much greater predictive validity of cognitive ability. Similar findings hold true for the NLSY97.
- Compare the R^2 of models 2 and 3. Both models explain 37% of the variance, which means that including parental education and family income provides almost no incremental validity over just cognitive ability in predicting grades. Similar findings hold true for the NLSY97.
All methods converge on the same conclusion: cognitive ability is far more important than parental SES for predicting grades.
Now, one might object that the measure of parental SES was too limited. To account for this objection, the author created extended models of both models 1 and 3 to account for a broader measure of parental SES. In the NLSY79, the extended model adds parental occupational status (based on the Duncan socioeconomic index of occupations), average family income from 1979 to 1986 (instead of a snapshot of family income), and average family wealth for 1979 and 1980. In the NLSY97, the extended model adds family wealth in 1997. The extended models did not provide large incremental predictive validity over the standard models. That is, incorporating variables such as parental occupational status or family wealth did not make the models substantially better able to predict academic achievement (see Tables A1 and A2 and also commentary throughout the study).
For example, in the NLSY79, the percent of variance explained in extended model 1 was just 2 percentage points higher than standard model 1 (14% vs 12%, see page 8). Thus, adding parental occupation, average family income, and parental wealth only provides marginal incremental predictive validity over just parental education and a snapshot of family income; and all of these variables combined still explain a much smaller percent of variance than just cognitive ability.
The basic findings for grades were summarized as follows:
Table 4 presents the estimates from the analyses of grades at school. Model 1 shows that parents’ education and family income poorly explain grades: 12% of the variance for high school grades in the NLSY79, and 7% for grade 8 and high school grades in the NLSY97.
In contrast, ability accounts for 37% of the variance for high school grades in the NLSY79 and about 20% in the NLSY97. For the NLSY79, the variance explained by the extended model 1 is 2 percentage points higher than that for model 1, 14%. Since grades are measured differently in the two studies, it is not possible to compare the metric coefficients across the two cohorts. Comparison of the standardized g coefficients indicates that g is a weaker predictor for high school grades in the NLSY97 (β = 0.41) than the NLSY79 (β = 0.58). The 95% confidence intervals do not overlap, so the difference is statistically significant
Model 3 includes all three variables. Compared to model 2, the variance explained is unchanged indicating the parents’ education and family income provide no additional explanatory power beyond that from g. Furthermore, the effects of g decline only marginally. The addition of parents’ education and family income decreased the coefficient for g by 4% in the NLSY79 and by 5% for grade 8 grades and 7% for high school grades in the NLSY97.
Similar findings were also reported by Guez et al. (2018) [archived]. These authors analyzed predictors of the achievement gap in France in a longitudinal study of over 20k French middle school students. Academic achievement was measured using a battery of five tests measuring school-related skills. The tests assessed phonics skills, grammar, mathematics, reading comprehension, and academic knowledge (page 105-106). Non-verbal IQ was measured using Chartier’s Reasoning Test on Playing Cards which was inspired from Raven’s progressive matrices and designed to assess non-verbal logical reasoning skills. Other predictors included self-efficacy, household income, parental education, cultural objects in the house, and a number of other variables. The results showed that “non-verbal IQ was by far the most predictive factor of achievement in grades 6 and 9” in the sample (page 111). The standardized coefficients of each predictor in a regression model are as follows:
It should be noted that this study likely underestimates the influence of cognitive ability for a number reasons:
- Cognitive ability is measured with just a non-verbal test in the spirit of Raven’s progressive matrices. Cognitive ability tends to have larger predictive validity when extracting g from a test that measures a broader range of cognitive abilities (e.g., spatial reasoning, verbal reasoning, logical reasoning, etc.).
- The study measures the impact of IQ after controlling for self-efficacy, i.e. one’s belief in their own ability. The problem here is that self-efficacy is influenced by actual cognitive ability, i.e. individuals who are high in cognitive ability are more likely to believe that they have high ability. Thus, by controlling for self-efficacy, one is controlling for a possible mediator between cognitive ability and grades, which will underestimate the true effect (as I have explained here). The same potentially applies to “Priority education school”. If a student’s cognitive ability has an influence on whether they attend a priority education school, then controlling for priority education school will underestimate the effect of cognitive ability.
Despite these limitations, the study still found a much larger effect of IQ compared to other variables. In particular, the standardized coefficient for non-verbal IQ was about 2 times that of parental education and 7 to 8 times that of household income. The findings of the study led the authors to conclude that IQ was “by far” the most predictive factor for academic achievement, although IQ was a poor predictor for progression in achievement between grades 6 and 9:
The aim of this study was to assess the contributions of various environmental and individual factors to the IQ-achievement gap – i.e., the part of achievement that is not predicted by IQ – during middle school. Our study confirmed that cognitive ability is an important predictor of academic achievement, even when taking into account a wide range of contextual and individual factors. Indeed, non-verbal IQ was by far the most predictive factor of achievement in grades 6 and 9, and not including it induced biased estimates for other independent variables. However, it had a small effect on the variations in achievement during middle school compared to other factors. These results thus confirm the undeniable and widely reported role of intelligence (here non-verbal intelligence) in explaining academic performance at a given time, but show that it has a marginal role in explaining progression.
Similar findings were reported by Becker et al. (2019) who examined how parental SES, childhood intelligence, and gender influence socioeconomic outcomes in German children. Unfortunately, the study did not include parental income in the measure of parental SES (it only included parental occupational status and educational attainment). That said, childhood intelligence correlated with GPA in general education far more than did parental SES (r = 0.37 vs r = 0.22, Table 4). When parental SES, childhood intelligence, and gender were entered into a path model to predict GPA, education, occupational status, and income, the path from childhood intelligence to GPA in general education was over 7 times the path from parental SES to GPA (0.38 vs 0.05, Figure 1). In fact, the path from parental SES to GPA wasn’t even statistically significant.
Conclusion: the findings from these studies are consistent with the previously cited studies. The standardized regression coefficients of cognitive ability on grades hovered in the β = 0.40 to β = 0.50 range. By contrast, the coefficients for parental SES on grades hovered in the β = 0.0 to β = 0.10 range in the U.S. (with many of the estimates being statistically insignificant) and about β ~ 0.20 in France.
Individual studies: test scores
As with grades, it is important to review individual studies that compare the predictive validity of parental SES and cognitive ability on achievement test scores. I will begin by considering the zero-order correlations reported by individual studies involving these variables.
Zero order correlations
First, recall the meta-analyses cited earlier on the association between parental SES and academic performance. In particular, Selvitopu and Kaya (2021) reported that the correlation between parental SES and various achievement tests (e.g., national achievement tests, international achievement tests, etc.) between around r = 0.2 and r = 0.3.
Now, consider some individual studies that analyze the association between parental SES and specifically SAT/ACT test scores. For example, Sackett et al. (2009) reported correlations between SAT scores and parental SES for college freshmen from three entering cohorts (1995, 1996, and 1997) in 41 colleges and universities. The 41 colleges were geographically diverse, included small and large schools, public and private institutions, and a broad range of selectivity on SAT scores. The mean freshman class size was 1,369 students, so the total sample was well over 50k students. Parental SES was measured based on student-reported parental education and family income. The observed correlation between parental SES and SAT scores was r = 0.22 within a given institution (page 4). The correlation rose to r = 0.42 when considering the entire test-taking population (due to range restriction within an institution).
Westrick et al. (2015) reported similar findings for ACT scores. These authors reported correlations between ACT scores and parental SES among all students who took the ACT from 1999 to 2006 (page 28). Parental SES was measured using self-reported measures of parental income. The total sample size exceeded 6 million students (Table 3). The reported correlation between parental income and ACT scores was r = 0.34.
To appreciate the magnitude of the IQ-SAT/ACT correlations, note that these correlations are on par with, and sometimes greater than, the correlations between different subtests of the SAT and ACT.
In the second study, researchers correlated ACT scores with Raven’s Advanced Progressive Matrices (Raven’s APM) scores among a sample of 149 college students. They found a correlation of r = .61 between Raven’s APM and Composite ACT score (page 157), which increased to r = .75 after correction for range restriction (page 158).
These large correlations led the authors to conclude that “the ACT is an acceptable measure of general intelligence” (page 158). They ended with the following discussion:
The analyses presented above demonstrate a significant relationship between measures of cognitive ability and ACT scores. Based upon correlations with conventional intelligence tests and the first factor of the ASVAB, it appears that ACT is a measure of general intelligence. Indeed, based on the correlations among the tests in Study 1, the ACT is indistinguishable from other tests that are identified as intelligence tests. In addition, the ACT shows a high correlation with the SAT, itself considered to be a measure of intelligence (Frey & Detterman, 2004). The jackknife analysis confirms the stability of these results.
Similar findings were reported for the SAT by Frey and Detterman (2004) [archived] who performed a similar analysis on the same dataset but for SAT scores. These authors also found large correlations between SAT tests and conventional intelligence tests. Again, the correlations ranged between r = 0.53 and r = 0.83, with all correlations exceeding r = 0.65 except for one (Table 2). These authors found similarly large correlations between SAT scores and g extracted from ASVAB scores (r = 0.72 or r = 0.86 after correction, depending on the sample). The authors also found a correlation of r = 0.48 between Raven’s APM scores and SAT scores, which increased to r = 0.72 after correction for range restriction. These findings led the authors to conclude that “the SAT is an adequate measure of general intelligence” (page 377).
One might object that the ASVAB is not technically an IQ test as it was not constructed to measure general cognitive ability. While this is true, other studies have shown that measures of g extracted from different test batteries correlated almost perfectly with one another (Johnson et al. 2004, Johnson et al. 2008). For example, Johnson et al. (2004) found that when g is extracted from three different test batteries from a sample of over 400 adults, the extracted g factors correlate almost perfectly (r > 0.98). The authors note that this suggests the “existence of a higher-level g factor and suggests that its measurement is not dependent on the use of specific mental ability tasks”.
Perhaps more impressively, Kaufman et al. (2012) also analyzed the association between g factors extracted from different test batteries. In particular, they analyzed g factor scores extracted from achievement tests (in reading, mathematics, and writing) and g factor scores extracted from cognitive ability tests. Despite the distinct content of the tests, the authors observed a mean correlation coefficient of r = 0.83 for the g extracted from the different tests. Furthermore, this correlation gradually increased with age. See the following table:
Thus, g extracted from cognitive ability tests correlates with g extracted from achievement tests extremely highly. Again, the average correlation is r = 0.83, which is about the degree to which different cognitive ability tests correlate with themselves. Furthermore, this correlation rises to about r = 0.86 to r = 0.88 once subjects reach late adolescence. This shows that g extracted from any test battery can be used as a reliable measure of cognitive ability, so long as the test battery measures a sufficiently broad range of abilities.
These findings were replicated in a recent study by Coyle (2015). Like the previous studies, this author also analyzed the relationship between g (extracted from ASVAB tests) and standardized test scores. However, unlike the previous studies, he used a more recent dataset – The NLSY97 – and he analyzed the relationship between g with SAT, ACT, and PSAT scores. Nearly 2,000 participants were analyzed in this dataset. The following table indicates the correlation between g and the composite scores of each test as well as subtest scores.
As you can see, the correlation between g and test scores range from r = 0.66 to r = 0.76, with larger correlations for the composite measures (r = 0.73 to r = 0.76). In fact, the author notes “g predicted composite scores of the three tests (Mr = .74) better than subtest scores of the tests (Mr = .68)” (page 18). These correlations are consistent with the correlations by Koenig et al. (2008) and Frey and Detterman (2004). Thus, it seems that data from both the NLSY79 and NLSY97 indicate that SAT and ACT scores are acceptable measures of general cognitive ability.
Longitudinal studies also show that IQ is a powerful predictor of performance on achievement tests conducted years into the future. For example, McCoach et al. (2017) [archived] examined the stability of cognitive ability and academic achievement tests for a cohort of children from elementary school through high school. The authors relied on data from the Fullerton Longitudinal Study, a program that was launched in 1979 which followed 130 children from infancy into adulthood with repeated assessments of cognitive ability and academic achievement.
Academic achievement was measured between ages 7 and 17 years based on mathematics and reading achievement from the Woodcock Johnson Psycho-Educational Batteries. The participants were given different assessments for cognitive ability based on their age: they completed the Bayley Scales of Infant Development (BSID) when children were 24 months or younger, the McCarthy Scales of Children’s Abilities (MSCA) between 30 to 42 months of age, the Wechsler Intelligence Scales for Children-Revised between the ages of 6 and 12 years, the Wechsler Intelligence Scale for Children-Third Edition (WISC-III) at ages 12 and 15 years, and finally the Wechsler Adult Intelligence Scale-Revised (WAIS-R) at 17 years of age.
Consistent with prior evidence, the authors found large correlations between scores on intelligence tests and scores on achievement tests. Furthermore, intelligence tests were significant predictors of future achievement tests. The following figures show the correlations between IQ test scores and achievement test scores conducted at different ages.
- Note: I added the red lines as guides to represent a correlation of r = 0.7.
As you can see, there are substantial associations between academic achievement and cognitive ability. After about age 6, at any given point in time, a subject’s IQ correlates with their mathematics achievement at about r = 0.7 to r = 0.8 when both measures are conducted at the same time. When IQ is measured years before mathematics achievement, the associations are still substantial. For example, IQ measured at age 7 correlates with mathematics achievement between ages 7 and 17 at between r = 0.6 and r = 0.7. Similar patterns hold for reading achievement, although the absolute correlations are about 0.1 points lower.
The finding that early cognitive ability significantly predicts later test scores has also been replicated in large sample studies in the United Kingdom. For example, Deary et al. (2006) [archived] examined a 5-year prospective longitudinal survey of a representative sample of over 70,000 children in England. These researchers measured the relationship between the general factor of intelligence (g) measured at age 11 and GCSE test points at age 16. Researchers found that the correlation between g measured at age 11 and GCSE test points at age 16 was r = 0.69. The largest correlation was found between g and mathematics (r = 0.77).
Conclusion: the studies summarized in this section imply that achievement test scores correlate substantially with cognitive ability. The correlations typically ranged between r = 0.7 and r = 0.8, which is in line with the meta-analyses cited earlier. Indeed, a number of studies report that achievement tests just are tests of general cognitive ability. By contrast, the correlation between parental SES and test scores varied between r = 0.3 to r = 0.4, which is also in line with the meta-analyses cited earlier.
Given the fact that both parental SES and cognitive ability are correlated with test scores, and the fact that both of these variables are correlated with each other, it is reasonable to assume that much of the association between parental SES and test scores will be due to confounding with cognitive ability (and vice-versa, swapping parental SES and cognitive ability). Therefore, it will be useful to consider studies that attempt to estimate the independent associations of parental SES and cognitive ability with test scores. I will consider such studies in this section.
I will start with Marks (2022) [archived] which was cited earlier. Remember that this study analyzed the predictive validity of cognitive ability and parental SES on various outcomes (e.g., GPA, test scores, wages, educational attainment, etc.) in the NLSY79 and NLSY97. I already reported the results for GPA above. Now let us consider the predictive validity of parental SES and cognitive ability on test scores, specifically SAT and ACT scores. First, consider again the zero-order correlations between the main variables of interest:
Again, the correlations here are in line with the figures reported earlier. Parental income and parental education correlate with SAT/ACT scores at somewhere between r = 0.20 and r = 0.40, which is expected given the studies reported earlier. Also, general cognitive ability correlates with SAT/ACT scores at somewhere between r = 0.80 and r = 0.87, which is also in line with the findings reported earlier, although these figures are on the higher end of the estimates reported earlier.
Now, consider the independent effects of cognitive ability versus parental SES scores when they are simultaneously entered into models to predict test scores. Let us begin with SAT scores:
The first thing to note here is that family income is not statistically significantly associated with SAT scores when controlling for both cognitive ability and parental education in both datasets (see model 3). In fact, there is a negative association between family income and SAT scores in both datasets (though this could just be noise).
Now, as I did with grades earlier in this post, there are a few ways to compare the independent associations between cognitive ability vs parental SES on SAT scores:
- Compare the standardized coefficients when they are entered into the same model. In model 3, the standardized coefficient of cognitive ability is about 10 times the coefficient for parental education in both the NLSY79 (β = 0.86 vs β = 0.09) and the NLSY97 (β = 0.84 vs β = 0.08). The coefficient for family income is negative.
- Compare the R^2 of models 1 and 2. The model with both parental education and family income (model 1) explains only 18% of the variance in SAT scores, whereas the model with just cognitive ability explains 75 to 81% of the variance.
- Now, compare the R^2 of 2 and 3. The R^2 for model 3 is only 1 percentage point greater than the R^2 for model 2 in both of the datasets. This suggests that parental education and family income provide almost no incremental validity over just cognitive ability in predicting SAT scores.
As I mentioned in the section on grades, one might object that the measure of parental SES was too limited. The “extended models” with broader measures of parental SES (described earlier) did not provide large incremental predictive validity over the standard models. That is, incorporating variables such as parental occupational status or family wealth did not make the models substantially better able to predict academic achievement (see Tables A1 and A2 and also commentary throughout the study).
The findings for ACT scores are basically the same as the findings for SAT scores, so I won’t cover them in too much detail. I’ll post the table here for the viewer’s sake:
As with the SAT scores, cognitive ability is a much better predictor of ACT scores than is parental SES.
- Academic achievement was measured from a variety of measures. Firstly, participants reported scores on a national standardized test of various subjects conducted when participants were aged 15-16 years old. Secondly, in wave 3, when participants were 17-years-old, they conducted tests of verbal, vocabulary, and numerical attainment.
- Cognitive ability was measured at age 13 based on the Drumcondra Reasoning Test (DRT), which assesses numerical, verbal and overall reasoning ability. The verbal subtest is based on synonyms, classifications, analogies and antonyms. Numerical ability is assessed by examining operations with numbers, relationships with numbers, sequential ordering and numerical abstractions.
- Parental SES was measured using a number of different measures. Household income was measured as average disposable income adjusted for household size from waves 1 and 2. Social class was based on the occupation of the adults, which was categorized in one of 4 groups: Professionals (9.4%), Managerial and technical (30.5%), white-collar (19.7%), and Manual and other (40.3%). Parental education was measured as the average of the sum of the number of full-time years of education of both parents.
For each measure of academic achievement, three regression models were used to predict that measure of achievement. Model 1 (M1) included gender and measures of parental SES (e.g., social class, parental education, and household income, etc.). M2 included gender and cognitive ability. M3 included parental SES, gender, and cognitive ability.
The results show that childhood cognitive ability is a far better predictor of future academic performance than childhood SES. Perhaps the simplest way to see this is to compare the R^2 of different models for each measure of academic performance. For example, consider national examination scores. Only 16.3% of the variance in national examination scores can be explained by gender and the parental SES measures (M1). By contrast, 32.7% of variance can be explained by gender and cognitive ability (M2), showing that cognitive ability explains far more variance than parental SES. Furthermore, when all predictors are included together, 36.8% of the variance is explained (M3). Thus, introducing parental SES into the model with gender and cognitive ability only marginally increases the percentage of variance explained (from 32.7% to 36.8%), suggesting that adding parental SES only contributes marginal incremental validity over just cognitive ability and gender.
Similar findings can be observed for the other measures of academic achievement. Perhaps the most dramatic finding can be observed by analyzing the results for the numeracy measures. The results show that adding parental SES to the model that already includes cognitive ability and gender only raises the percentage of variance explained from 38.5% to 38.9%. Thus, parental SES barely has any incremental predictive validity for numeracy after accounting for gender and cognitive ability.
These findings led the authors to conclude the following:
The analysis focused on the effects of SES measured by social class, household income and parental education vis-à-vis cognitive ability on a range of adolescent outcomes in Ireland. Cognitive ability measured at age 13 had strong associations with educational, cognitive, life difficulties, and relationship outcomes. On the other hand, SES factors–family social class, household income, and parents’ educational attainment–had much weaker effects with outcomes often considered strongly linked to SES.
Similar findings were reported by Marks (2021) [archived] in a study of students in Australia. The author analyzed the relative predictive validity of childhood cognitive ability and parental SES on academic achievement in a longitudinal sample of Australian children. The participants were studied across six waves, starting when children were aged 4 or 5 with another wave every 2 years. Parental SES was measured by averaging a composite of income, parental education, and parental occupational status over waves 2 through 6. Childhood cognitive ability was measured based on Who Am I? (WAI) at wave 1, the Peabody Picture Vocabulary Test (PPVT) from waves 1 and 3, and a Matrix Reasoning test from waves 2 to 4. The WAI consisted of “11 pages on which children were to write their names, copy shapes, and write words and numbers.” For academic achievement, I will focus on numeracy and reading achievement tests taken in the final wave when children were aged 14 to 15 years of age (year 9). First, let us note the zero order correlations among these variables:
The correlations between childhood cognitive ability and year 9 achievement were around r = 0.52 to r = 0.54 which is slightly lower than the average correlation between ability and test scores reported in earlier meta-analyses. The correlations between the parental SES composite and year 9 achievement were about r = 0.40, in line with the earlier meta-analyses. When both parental SES and cognitive ability are entered into regression models for year 9 achievement, we see that cognitive ability has a far greater association with achievement than does parental SES:
The findings show that prior achievement (year 7 achievement) is obviously the best predictor of year 9 achievement (Model 3). However, when only cognitive ability and parental SES are included in the model (model 2), the standard coefficient for cognitive ability is over twice that of parental SES.
O’Connell and Marks (2022) [archived] analyzed the influences of academic achievement among a nationally representative sample of 8,303 children from the United Kingdom. The participants were analyzed in seven waves, at 9 months, 3 years, 5 years, 7 years, 11 years, 14 years, and 17 years of age. Academic achievement was measured based on self-reported grades on the GCSE, typically taken when participants were 16 years old. Cognitive ability was measured as a single score extracted from a battery of different tests of cognitive abilities taken between waves 2 and 6. Parental SES was measured using a variety of measures: household income adjusted for family size, parental education, and parental occupational status. First, note the zero-order correlations among all of these variables:
The correlation between child cognitive score and GCSE exam score was r = 0.47, which is somewhat lower than the typical correlations between cognitive ability and achievement test scores that were reported in meta-analyses above. By contrast, the correlation between the parental SES measures and GCSE exam score varied between r = 0.20 and r = 0.30, which is in line with meta-analytic data reported earlier. To test the independent effects of these variables (among others, including maternal cognitive ability and personality), a number of regression models were fitted to the data. There were 5 different models considered: a personality model, an SES model, a cognitive model, a psychological model, and a final model containing all variables from previous models. The results are displayed as follows:
There are a few ways to compare the independent predictive validity of cognitive ability and parental SES. First, note that the R^2 for the cognitive model is far greater than that of the SES model. Unfortunately, the cognitive model also includes the effect of maternal cognitive test scores, so the “cognitive” model doesn’t solely test the effect of childhood cognitive ability. Also, the regression coefficients listed are unstandardized, so we cannot meaningfully compare coefficients of different variables. However, we can easily standardize the coefficients because we know the standard deviation of both the independent variables and dependent variables from Table 2 above. The standard deviation for child cognitive score, maternal cognitive score, and GCSE exam scores were 1.0, 4.4, and 22.5 points, respectively. Thus, the standardized regression coefficients for child cognitive score and maternal cognitive score in model 3 would be (10.99 * 1.0 / 22.5) = 0.49 and (0.69 * 4.4 / 22.25) = 0.13, respectively. Thus, the higher explanatory power of the cognitive model can be almost entirely explained by childhood cognitive ability.
Furthermore, these models show that introducing parental SES to a model that already includes child personality, maternal cognitive test scores, and child cognitive test scores has only a marginal effect on the explanatory power (compare models 4 and 5). In particular, the R^2 for the “psychological” model is 28.5% compared to just 30.1% for the full model.
These results led the authors to conclude that “Cognitive ability and conscientiousness are more important than SES for educational attainment”. The paper ended with the following discussion:
The key result was a multivariate analysis of performance of a representative sample of sixteen-year-olds in a UK state exam comparing several models. The optimal model for educational attainment includes cognitive ability and conscientiousness. SES measures have much weaker effects.
The result demonstrates the inadequacy of the dominant SES-achievement paradigm, as social class and income have much weaker effects than cognitive ability and conscientiousness. Given their small effects, it does not make sense to pursue improvements in educational attainment by using SES as a lever…
The findings from this study were corroborated in a replication attempt by Bittmann (2022) on a study of German students, although this study found much larger effects of personality. Similar findings were also reported in a recent study of academic achievement among Swedish students by Boman (2022). I don’t go into much detail for this study because the author used a somewhat underdeveloped measure of parental SES (just parental education). However, for the sake of completeness, I’ll note that they find the same relative predictive validities, with cognitive ability being on top of parental SES. See the author’s summary of the results (page 13):
The results from the longitudinal analysis echo that of for example Gustafsson (2007), who stresses that earlier achievement (e.g., Grade 6) is the strongest predictor of later achievement (e.g., Grade 8), as well as Guez et al. (2018) who examined similar relations in the contemporary French context. Nonetheless, the other major predictors are still significant. The overarching results confirm the findings of, for example, Guez et al. (2018), Hartman et al. (2010), Laidra et al. (2007), Roth et al. (2015), and Vazsonyi et al. (2022). The rank order of predictor variables is that cognitive ability is a stronger predictor than non-cognitive abilities such as self-efficacy or academic self-concept, followed by SES.
Other studies comparing the effects of parental SES and cognitive ability on academic achievement include Colom and Mendoza (2007) in Brazil, Mendoza et al. (2021) in Latin America, and Marks and O’Connell (2021) in the United States.
Conclusion: the findings from these studies are consistent with the conclusion from the previous sections. Not all studies reported the standardized coefficients for cognitive ability and parental SES on test scores, so these cannot be used to summarize the findings from all studies. However, when standardized coefficients were reported, the effects of cognitive ability were typically several times that of parental SES. When comparing regression models that include just cognitive ability versus those that include just parental SES, the models with just cognitive ability typically explained several times as much of the variance in test scores as did the models with just parental SES. Finally, when comparing models with cognitive ability to models with cognitive ability and parental SES, the addition of parental SES only provided marginal incremental validity on the prediction of test scores. In fact, several models show that parental SES no longer demonstrated a statistically significant relationship with test scores after including cognitive ability in the model.
The studies summarized in this post converge on one undeniable finding: cognitive ability is a far better predictor of academic achievement than is parental SES. The disparity becomes even greater when considering the independent effects of both predictors, i.e. the effects of each predictor while holding the other predictor fixed (e.g., with regression analyses). The disparities are particularly large when academic achievement is measured with achievement test scores rather than grades. In fact, cognitive ability predicts standardized achievement test scores so well that some researchers have deemed prominent tests like the SAT and ACT to be adequate tests of general cognitive ability.
This post focused on statistical associations, so no interesting causal claims can be straightforwardly derived from any of this data. However, given some reasonable assumptions, I believe a number of plausible causal inferences can be made.
- I’ve argued elsewhere that much of the association between cognitive ability and academic outcomes is in fact due to a causal relationship (e.g. we know that much of the association between cognitive ability and academic performance is driven by genes rather than some kind of environmental confound). Therefore, given the findings in this post, it is reasonable to infer that cognitive ability has a much greater causal effect on academic outcomes than does parental SES.
- If parental SES does have a causal effect on academic outcomes, then much of this effect needs to be mediated by cognitive ability. That is, the causal path from parental SES to academic outcomes must involve a path from parental SES to cognitive ability and a path from cognitive ability to the academic outcome, because the association between parental SES and academic achievement diminishes greatly after controlling for cognitive ability. However, there are independent reasons to believe that parental SES does not have a large impact on cognitive ability (but this would need to be explained in a later post).
That said, there are a couple of points I would like to address before ending this post.
Poor measures of parental SES?
Someone might object that the measure of parental SES used in this post is too narrow. Most of the studies relied on some combination of parental income, education, and occupation (I tried to focus on studies that included at least income and one other measure). Perhaps, if other measures were included, such as wealth, then the effect of parental SES would be even larger. There are several points to make in response. First, one study, Marks (2022), did incorporate wealth and did not have a substantially larger effect on academic achievement.
Second, any measure of parental SES will fall under the “shared environment”, which plays a relatively minor role (compared to genetic differences) in explaining academic achievement (the “shared environment” is discussed briefly here). For example, meta-analyses have reported that the shared environment accounts for only about 10% to 16% of variation in academic achievement for older children (de Zeeuw et al. 2015, Little et al. 2016). That amounts to a correlation coefficient of about r = 0.32 to r = 0.40, which is in line with the correlations between parental SES and achievement reported in the studies above. Therefore, because the effect of the shared environment represents the upper bound of the effect of parental SES, and because the effect of the shared environment is about the same as the effects of parental SES reported here, a broader measure of parental SES likely would not have had a much larger association on academic achievement.
Much of the association reported here between parental SES and achievement is not in fact causal, but is instead due to confounding. For example, we know that much of the association between parental SES and academic achievement must be due to genetic confounding, because high-SES parents are more likely to have high intelligence and because individuals with high intelligence have higher intelligence largely due to genetics. In fact, there is evidence that about half of the correlation between family SES and test scores can be explained by genes. Now, the studies cited above in this post would have implicitly controlled for some of this genetic confounding because they controlled for childhood cognitive ability. However, this only controls for the genetic confounding that is mediated through offspring cognitive ability. We know that academic achievement is also driven by many traits other than cognitive ability, traits which are also heavily influenced by genes (e.g., personality, self-control, etc.).
Furthermore, there may also be confounding other than genetic confounding. For example, as already noted, high-SES parents tend to be smarter than average. If there are environmental benefits to having smarter parents (independent of socioeconomic background), then part of the association between parental SES and academic achievement will be due to environmental confounding with parental cognitive ability. There is some evidence that such confounding actually exists. For example, Marks and O’Connell (2021) find that the effects of parental SES decrease by 50-60% after controlling for a measure of maternal cognitive ability. Similar points apply to any parental traits that may cause both high SES and high achieving students (e.g., parental intelligence, parental parenting styles, parental personality, etc.).