Schooling substantially improves intelligence, but neither lessens nor widens the impacts of socioeconomics and genetics

Schooling substantially improves intelligence, but neither lessens nor widens the impacts of socioeconomics and genetics. Nicholas Judd, Bruno Sauce & Torkel Klingberg. npj Science of Learning volume 7, Article number: 33, Dec 15 2022. https://www.nature.com/articles/s41539-022-00148-5

Abstract: Schooling, socioeconomic status (SES), and genetics all impact intelligence. However, it is unclear to what extent their contributions are unique and if they interact. Here we used a multi-trait polygenic score for cognition (cogPGS) with a quasi-experimental regression discontinuity design to isolate how months of schooling relate to intelligence in 6567 children (aged 9–11). We found large, independent effects of schooling (β ~ 0.15), cogPGS (β ~ 0.10), and SES (β ~ 0.20) on working memory, crystallized (cIQ), and fluid intelligence (fIQ). Notably, two years of schooling had a larger effect on intelligence than the lifetime consequences, since birth, of SES or cogPGS-based inequalities. However, schooling showed no interaction with cogPGS or SES for the three intelligence domains tested. While schooling had strong main effects on intelligence, it did not lessen, nor widen the impact of these preexisting SES or genetic factors.

Discussion

Schooling showed substantial and independent effects for each intelligence domain tested: cIQ, fIQ, and WM. In line with previous research, we found the raw effect of schooling on cIQ to be larger than for fIQ, though this difference was not significant13. This was also true for the relative influence of schooling compared to the effect of chronological age, as that ratio for cIQ was almost double that of fIQ (1.1 vs 0.54), showing almost near equal influences per year of schooling and age. Surprisingly, WM had the highest ratio (2.2), with the effect of schooling being more than double that of age. While this is in agreement with a previous study in younger children21, it should be interpreted with caution as the WM measure comprised of only one task59.

As expected, SES and cogPGS were highly correlated with each other, highlighting the need to isolate the independent effects of each. Both had large, independent effects on cIQ, fIQ, and WM. In a follow-up analysis, we estimated the contribution of each SES component separately. Notably, all SES components were significant for each intelligence domain, with the effects from parental education and income being similar in size while neighborhood quality was roughly half.

A child’s SES is not independent of their cogPGS31, which makes it difficult to support causal inferences of these factors as well as interpret the interaction between them (i.e., the endogeneity problem)60. This gene-environment dependence can cause spurious gene-environment interactions41,61. Our sibling analysis sheds light on this issue. We found the within-family effect of cogPGS to be roughly half of the between-family effect for cIQ, in line with the literature51,52. This indicates the presence of passive genotype-environment correlations—whereby parents create family environments consistent with their genotypes, which in turn facilitate the development of their children’s intelligence. Since we only had data from 392 families, a lack of statistical power is most likely the reason for our null findings for fIQ and WM. While our cogPGS estimate in the full sample should be interpreted with caution, previous research has shown SES to be the major source of these between-family effects51,62.

Predominant theories of GE-interplay imply a positive sign—genetically endowed cognition influences one’s proximal environment, and that environment, in turn, influences one’s cognition in continuous, reciprocal interactions, such as the multiplier theory34, the transactional model35 and the bioecological model36. In line with this, a meta-analysis on twin research found the heritability of intelligence to increase with higher SES in the United States63. But this effect is far from consistent. The same meta-analysis did not replicate this finding on data collected outside the United States. Furthermore, a large twin study found negative and null results for heritability by SES interaction regarding mathematics and reading in Florida64. Two studies using a similar polygenic marker to ours did not find evidence of GE-interplay for EA65,66, yet a study with 130,000 adults in the UK found a very small negative interaction of SES with neighborhood quality for fIQ and EA40. Crucially, the standardized effect size of this interaction was (β < 0.02 SD)—translating to less than a third IQ point throughout one’s entire life—in turn having no practical effect for the individual.

We did not find any significant interaction between schooling and SES or cogPGS. One strength of our design is that schooling is, in principle, independent from cogPGS and SES. We are not aware of any other research looking at GE-interplay with schooling. A recent study did find low PGS children in high-SES schools to continue with mathematics much longer than genetically similar children in low-SES schools39. However, we see their result as more relevant to inform a cogPGS-SES effect rather than the gene-by-schooling interaction.

We expected our interaction terms to either compensate or accelerate preexisting differences. Schooling, for example, could increase (i.e., Mathew’s effect) economic/genetic inequality or lessen these differences between children (i.e., catch-up effect). The Coleman Report, a seminal study with more than half a million students and over 3000 schools in 1966, controversially concluded schools did not contribute to widening achievement gaps between children29. Conversely, there is some evidence that schooling might lessen socioeconomic disparities between children (i.e., a catch-up effect) for cognitive skills67.

Our study indicates schooling to not be a major driving force for either increasing or decreasing differences due to SES or cogPGS. Yet, we emphasize caution in interpreting these null effects as our range of schooling (3rd–5th grade) was limited, and Bayesian analysis showed that an effect of less than 0.02 SD could not be ruled out (Fig. 3). Since any interaction effect with schooling could accumulate—that is, continue to increase each year—a very small (<0.05 SD) effect size could be of practical relevance54,55. For example, an interaction with schooling as small as 0.02 SD could accumulate over five years to 0.1 SD or roughly 30% of the largest SES main effect (i.e., cIQ = 0.29). This is, of course, a simplistic scenario assuming no counteracting mechanisms, yet it illustrates how very small effect sizes can become consequential68,69. In contrast, cogPGS–SES’ interaction is a lifelong effect and does not have the potential to accumulate in the same way. However, our sample had a slightly lower SES (Cohen’s d = −0.23) than the average for the United States, therefore, we cannot rule out an interaction at the lower tail of the SES distribution.

One limitation of this study is that the 1.1 million individuals used to estimate the cogPGS are heavily biased towards those of European descent and from higher SES areas30,70. This means our results regarding genetics should only be generalized to white populations. Furthermore, GWAS methods cannot detect certain types and sizes of GE interactions since they are intended to detect additive effects71. Another consideration is how to interpret findings with multi-trait GWAS’s—in our case cognitive ability, mathematics, and educational attainment—since one of the supplementary phenotypes could be driving the results72. The extent of this issue depends on (1) the relative sample size differences between the GWAS included and (2) the genetic correlation of these traits. In our case, there are sample size differences between educational attainment and cognitive ability, yet the very high (r ~ 0.75) genetic correlation between these traits most likely mitigates this issue73. Lastly, a strength of our study is that we controlled for ancestry-based genetic PCs in the full model, rather than just correcting cogPGS. While this means SES’s relationship with cognition is controlled for population stratification, it also brought the limitation that we had to exclude subjects without DNA, resulting in the average level of SES increasing.

We found that schooling causes relatively large increases in children’s intelligence. The two years of schooling (3rd to 5th grade) caused a larger difference in intelligence than either SES or cogPGS. However, schooling did not change the rank order of individuals’ intelligence. This was shown by the lack of significant two-way interactions between Schooling, SES, and cogPGS, although our power to detect potentially meaningful small effects for schooling was limited. Intriguingly, we did not find any interaction between SES and cogPGS, this means that children’s genetic differences do not matter more, or less, for intelligence dependent upon their SES background.