The causal influence of cognitive ability

Last Updated on July 18, 2022

In a previous post, I cited data showing that cognitive ability is significantly correlated with various important outcomes, such as academic achievement, occupational performance, socioeconomic status, anti-social behavior, and health. However, that data only establishes that there is a statistical association between cognitive ability and these outcomes. The data does not establish that cognitive ability has a causal influence on any of these outcomes.

Now, one possible explanation of these associations is that cognitive ability is at least partially causally responsible. But another possible explanation is that these associations are completely confounded by some third variable and cognitive ability plays no causal role (e.g., one might posit that cognitive ability is correlated with academic achievement not because cognitive ability is causal, but because both factors are the result of, say, family background). For the purposes of this post, I will refer to these two possible explanations as the Causal Hypothesis and the Confounding Hypothesis.

  • Causal Hypothesis: on this hypothesis, the association between cognitive ability and the outcomes listed above is party explained by the fact that cognitive ability has a causal influence on the outcomes.
  • Confounding Hypothesis: on this hypothesis, the association between cognitive ability and the outcomes is entirely explained by other variables that confound the association; cognitive ability has no causal influence on the outcomes in question.

Note that I define the Causal Hypothesis as denoting that the causal influence of cognitive ability only party explains the association between cognitive ability and the relevant outcomes. This does not quantify the magnitude of the effect of cognitive ability’s causal influence. In fact, I believe the causal influence of cognitive ability significantly explains the associations with cognitive ability, often being the primary explanation of the associations. However, this will have to be demonstrated in specific studies and the precise estimate of the influence of cognitive ability will vary from outcome to outcome. So I’ll define Causal Hypothesis as general as possible. More precise estimates of the causal influence of cognitive ability will have to be reported in the individual studies.

Before reviewing the evidence for the Causal Hypothesis, I will mention some considerations that should shift our priors in favor of the Causal Hypothesis. Next, I review some of the primary methods used by the studies in this post that attempt to account for confounding. Finally, I review studies that use these methods to provide evidence in favor of the Causal Hypothesis by ruling out confounding explanations. The studies provide evidence that cognitive ability has a causal influence on academic achievement, occupational performance, socioeconomic success, and anti-social behavior. In the end, I believe the totality of evidence suggests that the Causal Hypothesis is undeniable: cognitive ability has a large causal influence on important life outcomes.

Showing causation

Most people are familiar with the phrase “correlation does not imply causation.” This phrase is true because, for any two correlated variables X and Y, there are three possible explanations of the correlation (assuming the association isn’t pure randomness):

  1. X causes Y.
  2. Y causes X.
  3. Z causes X and Z causes Y.

Note that these explanations are not mutually exclusive. Because the association between X and Y can be explained without X having a causal influence (if explanation 2 or 3 is true), we cannot reasonably infer that X is causal merely from the fact that X is correlated with Y. In order to show that X has a causal influence on Y, there are more conditions that one must demonstrate before the causal inference can be made. One must show the following three conditions (page 146):

  1. There is an empirical association between X and Y.
  2. X occurs before Y.
  3. The association between X and Y is not spurious.

These requirements roughly correspond to the three possible explanations I mentioned earlier. If one fulfills requirement 2 (showing that X occurs before Y), then that rules out possible explanation 2 (that Y causes X). And if one fulfills requirements 3 (showing that the X-Y association is not spurious), then that rules out possible explanation 3 (that Z causes X and Z causes Y). Thus, if one fulfills requirements 2 and 3, then that’s sufficient evidence that explanation 1 is true, i.e. that X causes Y.

The first two requirements are fairly easy to demonstrate. My previous post fulfilled these requirements by showing that cognitive ability measured at one time is strongly associated with various outcomes later in a person’s life, even when the outcomes were assessed decades after the cognitive ability assessment. The more difficult requirement to fulfill is the third condition. To show that the correlation between two variables X and Y are not spurious, one must show that the correlation is not the result of confounding. That is, one must show that the association between X and Y is not the result of a third variable (called a “confounder” variable) Z which causes both X and Y. This is what I plan to show in this post. I spend more time discussing methods to address confounding in a separate post.

Our priors should favor the Causal Hypothesis


Before reviewing the evidence to compare the likelihood of the Causal Hypothesis and the Confounding Hypothesis, we should determine what our priors ought to be with respect to these hypotheses. By that, I mean we should determine which hypothesis we ought to lend more credence prior to reviewing the direct evidence. Priors are important because we do not (and should not) form judgments about hypotheses solely on the basis of the direct evidence that test the hypotheses.

That said, I think there are a number of considerations that should shift our priors towards the Causal Hypothesis. These considerations are as follows. Firstly, there is expert consensus that cognitive ability has a causal influence on outcomes. Secondly, cognitive ability is a fairly stable trait once people reach school age. Lastly, a substantial portion of the variance for cognitive ability is explained by genetic differences and a small proportion is explained by shared environmental factors. These considerations diminish the likelihood that the predictive validity of cognitive ability is the result of some of the most commonly proposed confounders (e.g., family background, school quality, etc.). I will explain this in the rest of the section.

Expert consensus

The expert consensus is that cognitive ability has a causal influence on many of the important life outcomes that I mentioned earlier. I will cite studies that surveyed the opinions of experts about research on cognitive ability to demonstrate this. These are all of the fairly recent surveys of expert opinion on cognitive ability that I’m aware of.

Gottfredson (1997) [archived] was a very brief 3-page statement that outlines conclusions regarded as mainstream by over 50 experts in intelligence and allied fields. Some of the conclusions they reached regarding the causal influence of cognitive ability were as follows (page 14):

  • A high IQ is an advantage in life because virtually all activities require some reasoning and decision-making. Conversely, a low IQ is often a disadvantage, especially in disorganized environments. Of course, a high IQ no more guarantees success than a low IQ guarantees failure in life. There are many exceptions, but the odds for success in our society greatly favor individuals with higher IQs.
  • The practical advantages of having a higher IQ increase as life settings become more complex (novel, ambiguous, changing, unpredictable, or multifaceted). For example, a high IQ is generally necessary to perform well in highly complex or fluid jobs (the professions, management); it is a considerable advantage in moderately complex jobs (crafts, clerical and police work); but it provides less advantage in settings that require only routine decision making or simple problem solving (unskilled work)
  • Differences in intelligence certainly are not the only factor affecting performance in education, training, and highly complex jobs (no one claims they are), but intelligence is often the most important. When individuals have already been selected for high (or low) intelligence and so do not differ as much in IQ, as in graduate school (or special education), other influences on performance loom larger in comparison.

Reeve and Charles (2008) [archived] examined the opinions of 30 experts in the science of mental abilities about their views on cognitive ability and cognitive ability testing. The study found a consensus among experts that general cognitive ability “is measured reasonably well by standardized tests”, that general cognitive ability “enhances performance in all domains of work”, that general cognitive ability “is the most important individual difference variable”, and even that general cognitive ability is “the most important trait determinant of job and training performance” (Table 1). Participants in the survey were selected from individuals on the editorial board of the journal Intelligence, from all registered members of the International Society of Intelligence Researchers, and from persons who had published three or more articles in Intelligence over the last 3 years (page 683). Experts were selected from this group by filtering down to “only individuals with a doctorate degree, and having at least five career publications on the topic of intelligence or testing” (page 683). This study was a replication of Murphy, Cronin, and Tam (2003) [doi], which found largely similar results.

Rindermann, Becker, and Coyle (2020) [doi] surveyed the opinions of over 100 experts in the field of intelligence about a variety of questions. One of the questions in the survey was “to what degree is the average socioeconomic status (SES) in Western societies determined by his or her IQ?” They survey found that “Experts believed 45% of SES variance was explained by intelligence and 55% by non-IQ factors (Table 3). 51% of experts believed that the contribution of intelligence (to SES) was below 50%, 38% above 50%, and 12% had a 50–50 opinion.” That is, experts believe that roughly half of the variance in socioeconomic status in Western societies is due to intelligence.

Stability

An important point to note about cognitive ability is its reliability or stability across an individual’s lifetime. Neisser et al. (1996) [archived] report that “Intelligence test scores are fairly stable during development” (page 81). They note that an individual’s age 17-18 IQ correlates at r=0.86 with their age 5-7 IQ, and correlates at r=0.96 with their age 11-13 IQ. Thus, we can predict with fairly high accuracy a person’s IQ at adulthood once we know their IQ at childhood. Similar points were made by Gottfredson (1997) [archived] who states that “intelligence is highly stable beginning in childhood” (page 87). Similar points are made by Sternberg et al. (2001) [archived] who makes two observations on IQ correlations between ages from age 3 to age 12: “First, the best predictor of IQ in a given year is the IQ from the previous year. Second, the predictive power of IQ in every subsequent year increases with the child’s age” (page 15).

The stability of cognitive ability has also been verified in many recent studies. For example, in a literature review on the stability of intelligence over time, Schneider (2014) notes that there is “broad agreement that the stability of cognitive ability varies as a function of the age of the sample but is rather high from school age onwards”. Also, consider the findings of Yu et al. (2018), who studied the stability of IQ scores in the Fullerton Longitudinal Study from infancy through adolescence. Researchers found that IQ measured at age 6 correlated significantly with IQ measured at age 8 (r = .79), age 12 (r = .72), and age 17 (r = .67) (Table 2). The following graph shows the correlations of IQ scores measured at different ages:

The stability increases significantly over a person’s life span. For example, Deary et al. (2004) [archived] investigated the stability of mental ability using data from The Scottish Mental Surveys, which collected IQ scores for almost every Scottish person born in 1921 and 1936 and attending school on June 1, 1932 and June 4, 1947. Participants completed the Moray House Test at ages 11 and 80 in order to assess their general mental ability. Researchers found that scores at age 11 correlated shockingly well with scores at age 80. The age-11 scores and age-80 scores correlated at about r=63 or r=0.66 depending on the cohort (Figure 2). After correcting for range restrictions, the estimated correlations rose to r=0.73. For reference, a recent study reported that, from childhood (ages 5 to 11 years) to middle adulthood (age 45 years), “IQ was as stable across age as height” (Richmond-Rakerd et al. 2021 [archived], page 4) with a correlation coefficient of r = .77.

The stability of cognitive ability is relevant because it reduces likelihood that certain variables confound the association between cognitive ability certain outcomes. For example, it is highly unlikely that, say, college grades may confound the association between cognitive ability and occupational performance because one’s rank-order cognitive ability does not change much when one enters college. Also, cognitive ability is fairly stable by the time children reach schooling, so it’s unlikely that, say, treatment by teachers confounds the association between cognitive ability and academic achievement. However, the stability of cognitive ability does not provide evidence against the possibility that early social factors may confound the associations with cognitive ability, since cognitive ability is not very stable in early childhood.

Sources of variance

There is scientific consensus that intelligence is substantially heritable. Heritability is formally defined “as a ratio of variances, specifically as the proportion of total variance in a population for a particular measurement, taken at a particular time or age, that is attributable to variation in additive genetic or total genetic values” (Visscher et al. 2008, page 255). The current estimates of heritability range somewhere between 40% and 80% for adults in developed countries (Nisbett et al. 2012, page 132; Plomin and Deary 2015Plomin et al. 2016). A common finding has been what is called the “Wilson Effect”, which is that the heritability of intelligence increases from childhood to adulthood (Neisser et al. 1996, page 85; Bouchard 2013Plomin et al. 2016). For example, Plomin and Deary (2015) [archived] report that “for intelligence, heritability increases linearly, from (approximately) 20% in infancy to 40% in adolescence, and to 60% in adulthood. Some evidence suggests that heritability might increase to as much as 80% in later adulthood but then decline to about 60% after age 80.”

In addition to the Wilson Effect, another common finding is that, as children age, the proportion explained by non-shared environmental factors remains constant, and the proportion explained by shared environmental factors declines dramatically (Haworth et al. 2010, Figure 1; Bouchard 2013, Figure 2; Briley and Tucker-Drob 2013, Figure 2). The result is often that the shared environment explains the smallest proportion of the variance of intelligence for adults in developed countries. For example, a meta-analysis of twin studies by Haworth et al. (2010) [archived] reports that, by age 17, 66% of the variance of cognitive ability is explained by genes, 19% is explained by unshared environment, and 16% is explained by shared environment (Table 3).

Some studies of older individuals have even found that shared environment have zero influence on intelligence differences by the time individuals reach mid to late adulthood (Neisser et al. 1996, page 85; Bouchard 2013, Figure 2). This is similar to many other psychological traits, for which the environmental sources of variance are mostly the result of differences in the non-shared environment (or within-family differences). For example, Plomin et al. (2016) [archived] report that “although environmental effects have a major impact” for psychological traits, “the salient environmental influences do not make siblings growing up in the same family similar.”

The heritability of cognitive ability is relevant because it reduces the likelihood that certain variables confound the associations with cognitive ability. For example, the shared environment explains very little of the variance in cognitive ability, especially during late adulthood. The “shared environment” includes environmental factors shared by siblings, such as e.g. parental income, parental marital status, neighborhood, etc. The “unshared environment” includes environmental factors not shared by siblings, such as e.g. peer groups, teacher interactions, etc. Now, because the shared environment explains only a minuscule proportion of variance for cognitive ability (particularly in late adulthood), it’s rather unlikely that shared environmental factors confound the correlation between cognitive ability and associated outcomes. This is because the substantial correlation between cognitive ability and associated outcome X (where X = academic achievement, occupational performance, etc.) must be the result of substantial covariance between cognitive ability and X. So in order for a candidate confounder C to substantially confound the correlation between cognitive ability and outcome X, C must explain a substantial proportion of the variance for both cognitive ability and outcome X. Therefore, because the shared environment explains only a miniscule proportion of the variance for cognitive ability, it cannot substantially confound the correlation between cognitive ability and outcome X.

Methods to reduce confounding


Before reviewing the data, I will explain some of the methods used to show that cognitive ability has a causal (rather than merely statistical) association with the relevant outcomes. The gold standard for addressing confounding and demonstrating causation in science are randomized control trials. However, randomized control trials are fairly rare in social science for technical and ethical reasons. Instead, social science tends to rely on other methods to make causal inference (e.g., “natural experiments”). Many of these methods are mentioned in a paper on causal inference by Antonakis et al. (2010). In this section, I’ll review three methods that can be used address possible confounding in studies on the effects of cognitive ability.

Statistical controls

The most common method to address confounding (in cognitive ability research) involves statistically controlling (or “adjusting” or “conditioning”) for certain confounders. There are a number of methods to statistically control for confounders. Two common methods involve regression analysis and stratification, both of which are mentioned in two articles about addressing confounding (Normand et al. 2005McDermott and Miller 2008). Normand et al. (2005) [archived] briefly describe regression analyses as follows:

Regression uses the data to estimate how confounders are related to the outcome and produces an adjusted estimate of the intervention effect. It is the most commonly used method for reducing confounding in cohort studies. The outcome of interest is the dependent variable, and the measures of baseline characteristics (such as age and sex) and the intervention are independent variables. The choice of method of regression analysis (linear, logistic, proportional hazards, etc) is dictated by the type of dependent variable. For example, if the outcome is binary (such as occurrence of hip fracture), a logistic regression model would be appropriate; in contrast, if the outcome is time to an event (such as time to hip fracture) a proportional hazards model is appropriate.

Regression analyses estimate the association of each independent variable with the dependent variable after adjusting for the effects of all the other variables. Because the estimated association between the intervention and outcome variables adjusts for the effects of all the measured baseline characteristics, the resulting estimate is called the adjusted effect. For example, regression could be used to control for differences in age and sex between two groups and to estimate the intervention effect adjusted for age and sex differences.

Like all methods to reduce confounding, the article notes that regression analyses rest on a number of assumptions for the results to be valid. The second method described is stratification. Stratification is described as follows:

Stratification is a process in which the sample is divided into subgroups or strata on the basis of characteristics that are believed to confound the analysis. The effects of the intervention are then measured within each subgroup. The goal of stratification is to create subgroups that are more balanced in terms of confounders. If age and sex were confounders, then strata based on age and sex could be used to control for confounding. The intervention effect is calculated by working out the difference in average outcomes between the intervention and comparison groups within each stratum. It is important to determine whether the relation between the intervention and outcome differs across strata. If the effect estimates are the same across strata, a summary estimate can be calculated by pooling the individual estimates.5 However, substantial differences in estimates across strata suggest effect modification, and a summary estimate should not be calculated.

One limitation with statistical controls is that we only have access to a finite amount of variables that we can control in a regression or stratification analysis. We can never be certain that we’ve controlled for all confounding variable. As McDermott and Miller (2008) notes, even “when we’ve controlled for a long list of confounds, we can never be certain that the association between variables is causal in nature. There’s always the possibility of other confounds that we haven’t considered” (page 140). While statistical controls can never prove that an association is not confounded by third variables, these methods are still useful because they can reasonably raise our confidence that an association is causal by showing that the association is robust against controls for a growing number of plausible candidate confounders. Our confidence be reasonably raised even more as the findings from statistical controls are corroborated with other methods (see below).

Sibling analysis

Instead of addressing confounding by statistically adjusting for possible confounding variables, a better method to address confounding involves analyzing sibling studies. These studies report the association between cognitive ability and socioeconomic outcomes within pairs of siblings. That is, they estimate the degree to which siblings with higher cognitive ability outperform siblings with lower cognitive ability. A sibling analysis implicitly controls for a wide range of family factors, because siblings typically have the same parents (and thus the same parental income, education, parenting practices, etc.), live in the same home, live in the same neighborhood, go to similar schools, etc. While this method is not perfect, this is a good design to minimize possible confounding, especially the confounding effects of family background.

For example, Hegelund et al. (2019) [archived] explain the usefulness of sibling designs to address possible confounders of the association between cognitive ability and socioeconomic achievement (page 100):

In the sibling comparison design, siblings with different intelligence test scores are compared whereby all factors shared by the siblings are matched out including both genetic and family environmental factors (Strully & Mishra, 2009). Thus, all familial factors shared by the siblings will be matched out by design, including factors that are unknown, unmeasured or measured with considerable measurement error. This is not to suggest that the observed associations in the sibling comparison design will be completely unconfounded, since this design only matches out factors that are shared by siblings. However, it might be a new and useful step towards understanding how much of the associations between intelligence and educational and occupational achievement can be attributed to familial factors operating in childhood because this study design includes all siblings in a sibship and therefore can be based on large representative samples of the general population.

Genetic analysis

Examining the genetic correlation between cognitive ability and outcomes is another method that can be used to rule out confounding explanations of the association. According to Plomin and Deary 2015 [archived], the genetic correlation between two traits can be thought about as “the probability that genes associated with one trait are also associated with the other trait.” In a 2017 textbook by Robert Plomin and colleagues, Behavior Genetics, the authors go into more detail on the mathematical techniques to estimate the genetic correlation between two traits (page 364):

A genetic correlation of 1.0 would imply that all additive genetic influences on trait X also impact on trait Y. A shared environmental correlation of 0 would imply that the environmental influences that make twins more similar on measure X are independent of the environmental influences that make twins more similar on measure Y. The phenotypic correlation between X and Y can therefore be dissected into genetic and environmental constituents. A high genetic correlation implies that if a gene were found for one trait, there is a reasonable chance that this gene would also influence the second trait.

The authors also go on to explain why the genetic correlation is relevant to my purposes in this post (page 366):

Genetic, shared environmental, and nonshared environmental correlations are independent of univariate heritabilities. That is, two traits might both have low heritabilities but a high genetic correlation. This would mean that, although there are probably only a few genes of modest effect that influence both these traits, whichever gene influences one trait is very likely to influence the other trait also. In this way, the analysis of these three etiological correlations can begin to tell us not just whether two traits are correlated but also why they are correlated.

A related concept that is also relevant is bivariate heritability. The bivariate heritability of two traits is “the proportion of the phenotypic correlation that is due to genetic factors” (Malanchini et al. 2020 [archived]).

These concepts are relevant because if the bivariate heritability between two traits is high, it indicates that the association between the traits is due to shared genetic factors. But if the association between the two traits is due to shared genetic factors, then the association between the two traits is not due to environmental factors. This automatically rules out environmental confounding as a possible explanation of the association between the two traits (at least, for the part of the association that is due to the shared genetic factors). Therefore, showing large bivariate heritability estimates between cognitive ability and outcomes is solid evidence that the association between cognitive ability and outcomes is not the result of environmental confounding.

While this data does rule out environmental confounding as an explanation of the association between cognitive ability and outcomes, it does not rule out all possible confounding explanations. In particular, a high bivariate heritability between two traits is compatible with a genetic confounding as an explanation of the association, without either trait being causal for the other trait. For example, when considering possible explanations of the intelligence-mortality association, Deary et al. (2021) [archived] note a number of possible explanations. Two of the possibilities involve positing that cognitive ability has a causal influence (i.e. the effect of cognitive is mediated through either healthy behaviors or a tendency to seek safer environments). But another possibility that they consider does not involve a causal influence for cognitive ability. They consider the idea that intelligence may be a marker of “general bodily system integrity”. In other words, intelligence test scores might be associated with general “brain health” which is also associated with bodily health more generally. It’s possible for these associations to obtain without any causal influence of cognitive ability, i.e. general bodily system integrity might be responsible for both cognitive ability and health (and other outcomes, such as socioeconomic status, self-control, etc.).

That being said, while genetic confounding cannot be ruled out by the finding of high bivariate heritability, this finding is still useful because it rules out environmental confounding which is the most commonly alleged source of confounding for the predictive validity of cognitive ability.

Converging lines of incremental evidence

Now, none of these methods should be understood as proving that cognitive ability is causal. For example, when we find that cognitive ability remains associated with various outcomes after statistically controlling for possible confounders, it’s always possible that there are additional variables that confound the association that we could not control. When we find that siblings with higher cognitive ability outperform siblings with lower cognitive ability, it’s possible that this is the result of a confounding variable that is not shared by siblings (e.g., perhaps peer group differences or differences in parental treatment explain why one of the siblings has a higher cognitive ability and better outcomes). When we find that cognitive ability and various outcomes are influenced by shared genetic factors, perhaps the genetic factors serve as confounders themselves (e.g., see the “general bodily system integrity” hypothesis as mentioned earlier) without cognitive ability being causal.

But this doesn’t mean these methods are useless. On the contrary, these methods provide useful incremental evidence to the conclusion that cognitive ability is in fact causal. As the evidence gradually accumulates, the confidence that we place into the hypothesis that cognitive ability is causal should correspondingly increase. Of course, our confidence can never reach the level of certainty, but this isn’t really a problem. Science doesn’t deal with certainty. Science deals with provisional theories and models that explain the observable phenomena better than alternative theories and models. Theories and models do not become accepted because they are proven. They become accepted because they provide better explanations of the observable phenomena than alternative theories and because they survive rigorous attempts at falsification. I believe that the hypothesis that cognitive ability is causal should likewise be accepted because the accumulated evidence in favor of the hypothesis is sufficiently large and no theoretically plausible alternative hypothesis can explain the data in an equally parsimonious manner.

That being said, I will now review evidence suggesting that cognitive ability has a causal influence on many important outcomes, starting with academic achievement as the first outcome to investigate.

Academic achievement


My previous post documented the large correlations between cognitive ability and academic achievement. One piece of evidence that this association is causal is the fact that the correlation between cognitive ability and academic achievement is typically greater than the correlation between measures of parental SES and academic achievement. For example, recall the results of Richardson et al. (2012) [archived] which reported that the correlation between intelligence and college GPA (ρ = .21) was greater than the correlation between SES and GPA (ρ = .15). If cognitive ability correlated with academic achievement only because it was a proxy for parental SES, then we would not expect it to have a greater correlation with academic achievement than SES does.

The findings for college GPA are a bit misleading because the predictive validity of cognitive ability will be artificially lowered due to range restriction. To avoid problems with range restriction, recall the results of the meta-analysis by Roth et al. (2015) [archived] which estimated the correlation between general mental ability and grades to be ρ = .54. By comparison, a meta-analysis by Harwella et al (2016) found that parental SES has only modest correlations with IQ (r = .27), GPA (r = .14), and non-IQ achievement outcomes (r .20) (Table 2). Given the rather modest predictive validity for parental SES on with IQ and GPA, parental SES obviously cannot explain the rather large correlation between IQ and grades. Now, one might say that this is not a fair comparison because the correlation reported by Roth et al. was corrected for measurement error and range restriction. However, even without correcting for measurement error or range restriction, the correlation between general mental ability and academic achievement (r = .44, page 123), is still greater than that found for SES.

Controlling for family background

A number of studies show that cognitive ability predicts academic achievement after statistically controlling for parental socioeconomic status. For example, Colom et al. (2006) [archived] analyzed 641 from a Brazilian school to examine the relationship between intelligence, academic achievement, and parental SES. The children were between 7 and 14 years of age. Researchers analyzed the test scores of 3 different samples of children. Fluid intelligence was measured using the Progressive Matrices Test in all 3 samples. Crystallized intelligence was measured using the WISC-III Verbal Scale in the third sample. Parental SES was measured using parental income and parental education. Across all three samples, fluid intelligence correlated with academic achievement (correlations ranged from .38 to .63 after controlling for age) to a significantly greater degree than did parental SES factors (correlations ranged from .01 to .25) (page 247). The authors also ran regression models for each sample, finding that intelligence predicted achievement even after controlling for the parental SES factors. They report that the “regression weights from children’s intelligence to their own scholastic achievement range from .69 to .27 across samples for the measure of fluid intelligence” (page 249).

For another study, see Johnson et al. (2007) [archived], who measured a number of factors related to academic achievement in 617 adoptive and biological families. The researchers used data from the Sibling Interaction and Behavior Study (SIBS), which consists of “a community-based sample of pairs of adoptive and biological siblings and their parents living in the Minneapolis-St. Paul area”. Each adoptive family included an adolescent between the ages of 10 and 22 who was adopted within the first 2 years of life and a second adolescent (not biologically related) within five years of the adoptee’s age. A majority of the adoptees were internationally placed (most from Korea). Each biological family included two adolescents within two years of age who were born to both parents. Academic achievement was measured using reports of student grades. Parental SES was measured using an index of social position based on job held, education, and income. Parenting practices were measured using a 42-item questionnaire completed by each sibling. The questionnaire was split into 5 scales: Structure, Parent’s Regard for Child, Child’s Regard for Parent, Parental Involvement, and Conflict. Parental Expectations for Educational Attainment (PEEA) was measured by asking parents their expectations for their child’s expected educational attainment. IQ was measured using abbreviated versions of the WISC-R or the WAIS-R. Engagement in school was measured using a self-report questionnaire on school behaviors. The results of the analysis were as follows:

  • IQ predicted grades for both biological offspring (r = .31, Table 3) and adoptive offspring (r = .43). Before adjusting for SES range restriction, SES only had minor or modest correlations with grades and IQ. For biological offspring, SES had moderate correlations with IQ (r = .16) and grades (r = .17). For adoptive offspring, SES had minor correlations with IQ (r = .02) and grades (r = .07). After correcting for SES range restriction, the correlations for SES increased. For biological offspring, SES had larger correlations with IQ (r = .24) and grades (r = .27). For adoptive offspring, SES had larger correlations with IQ (r = .12) and grades (r = .19). The fact that IQ predicts grades significantly more than SES predicts either IQ or grades suggests that the association between IQ and grades is not the result of confounding with SES.
  • In a regression model of all predictors on school grades, IQ remained significantly associated with grades (standardized regression coefficient of .26) even after controlling for possible confounding variables such as parental SES, gender, parenting practices, parental expectations, engagement, and even sibling grades (Table 4). The fact that IQ remained associated with school grades after controlling for sibling grades is particularly indicative of causation. By controlling for sibling grades, we are implicitly controlling for any family background factors that might influence offspring grades; thus, the fact that the association between IQ and grades persist after controlling for sibling grades suggests that the association is not the result of confounding with family background factors.

One might criticize some of the previous studies by saying that they used narrow measures of parental SES (usually parental income and parental education). One might say these do not capture the full effects of family background. There might be unmeasured aspects of parenting practices, neighborhood quality, school quality, etc. which confound the relationship between cognitive ability and academic achievement.

One way to avoid this objection is to perform a sibling analysis, where one compares siblings with the same household and measures whether the higher-IQ sibling achieves better outcomes than the lower-IQ sibling. Hegelund et al (2019) [archived] performed such an analysis on a sample of 364,193 Danish men with at least one full brother. The authors performed two analyses: “a conventional cohort analysis and a within-sibship analysis in which the association under investigation was analyzed within siblings while keeping familial factors shared by siblings fixed” (page 102). The findings from the conventional cohort analysis were compared to the findings from the within-sibship analysis to determine the degree to which IQ-outcome correlations were the result of “familial factors shared by siblings”. One of the associations studied was between IQ and GPA. The study found that, while familial factors shared by siblings explained “parts of the associations of IQ with GPA”, much of the association still remained after controlling for such factors. For example, the study notes that “Whereas individuals with an IQ of 70 were found to have a 2.9 times lower GPA than individuals with an IQ of 130 in the within-sibship analysis, individuals with an IQ of 70 were found to have a 4.0 times lower GPA than individuals with an IQ of 130 in the conventional cohort analysis” (page 104). In other words, while some of the association between cognitive ability and GPA can be statistically explained by controlling for “familial factors shared by siblings”, much of the association could not be explained by such factors. Here’s a graph of average GPA by IQ score after controlling for a host of possible confounders:

Accepting that cognitive ability has a causal influence on academic achievement does not imply that one must deny that family background has an influence on achievement. In fact, the two factors (cognitive ability and family background) can serve as mutually supporting explanations. If we accept family background factors influence offspring achievement, we need to explain how family background factors has such an influence. That is, we need to explain the mechanism that explains the connection between family background and offspring achievement. Surely, the impact of family background is not direct. That is, if a parent suddenly improves his income and educational attainment, this will not suddenly improve the child’s academic performance. Any plausible explanation of the causal influence of family background should posit that the influence is mediated through some intermediate variable (also called a “mediator”). Cognitive ability seems to be a plausible candidate factor to fulfill this mediating role. In other words, if family background influences academic achievement, this seems to be easily explained by the hypothesis that family background enhances cognitive development, which influences cognitive ability, which influences academic achievement.

The idea that cognitive ability mediates the association between family background and achievement is not only theoretically plausible. There is also empirical data supporting this causal pathway. For example, return to the study posted by Johnson et al. (2007) [archived]. Recall that this study found that IQ remained associated with school grades even after controlling for parental SES, parenting practices, sibling grades, etc. The researchers also tried different causal models to represent different causal pathways involved with explaining school grades. The authors found that the model that best fit the data involved a causal pathway from parental SES to IQ and from IQ to grades. They note:

Figure 3 shows the path analysis model we constructed based on the order of entry of the independent variables in the HLM regressions and the results that indicated that IQ accounted for many of the effects of SES and PEEA and Engagement accounted for many of the effects of Parenting. We specified this model separately for adoptive and biological offspring. The model shown in the figure provided the best fit to the data. All of the paths shown were significant, and there were no other significant paths. Thus, for example, SES contributed significantly to Grades both directly and indirectly via IQ, but it did not contribute significantly via Engagement.

The path analysis model was as follows:

  • “Par” = parenting. “PEEA” = Parental Expectations for Educational Attainment. “Eng” = Engagement in school.

So family background does not seem to confound the association between cognitive ability and achievement. At best, one could say that cognitive ability mediates the family background and achievement association, such that family background influences cognitive ability which in turn influences achievement.

Controlling for personality

Instead of parental SES, one might posit that personality confounds the association between cognitive ability and academic achievement. The problem with this argument is that cognitive ability typically predicts achievement better than measures of personality. For example, a meta-analysis by Poropat (2009) [archived] reported that intelligence correlated with academic performance (ρ = .25) more than any of the measured personality traits, which were conscientiousness (ρ = .22), openness (ρ = 0.12), agreeableness (ρ = .07), emotional stability (ρ = .02), and extraversion (ρ = −.01) (Table 1). As this meta-analysis shows, conscientiousness is more predictive of academic achievement than any other known personality trait, and it seems almost as predictive as cognitive ability (although recent studies show cognitive ability predicts academic performance considerably better than conscientiousness does, e.g. Spinath et al. 2010Cucina et al. 2016). So conscientiousness might seem like the best candidate of a personality trait that explains the association between cognitive ability and academic achievement. However, the big problem with this hypothesis is that the association between cognitive ability and conscientiousness has often been shown to be negative (Moutafi 2004) or at least not very large if it is positive (Murray 2014). This suggests that, while conscientiousness and cognitive ability are both associated with academic achievement, these two associations are likely to be mostly independent.

In fact, we have pretty good evidence that the respective associations of cognitive ability and personality on academic achievement are likely independent. Consider the following studies:

  • Duckworth et al. (2012) [archived] conducted multiple longitudinal studies showing that self-control predicted changes in report card grades better than did IQ, but that IQ predicted changes in standardized achievement test scores better than did self-control. They conclude that their findings suggest that “report card grades reflect dimensions of student competence related more to self-control than to intelligence, whereas standardized achievement tests reflect dimensions of student competence related more to intelligence than to self-control.”
  • Lechner et al. (2017) [archived] reported similar findings, showing that the predictive power of intelligence “is much higher than that of personality in the case of achievement – but much lower in the case of school grades” (page 89). They reported that 58% of the variance in achievement test scores was explained by intelligence, 17% was explained by personality, and 64% was explained by both factors combined. For school grades, much less of the variance could be explained: only 10% of the variance in school grades was explained by intelligence, 18% was explained by personality, and 29% of the variance was explained by both factors combined. Similar results were reported by Borghans et al. (2011) [archived].

See these findings by Lechner, showing that cognitive ability and personality predict different measures of academic achievement to different degrees:

Galla et al. (2019) [archived] reported similar findings. In one of their samples, researchers studied 1,622 subject to examine the relationship between 4-year college graduation and cognitive ability, self-regulation ratings, SAT scores, and GPA in High School. The study found that self-regulation and cognitive ability predicted academic outcomes through unique pathways. For example, consistent with the prior studies, the study found that, after adjusting for demographic characteristics, regression models showed that self-regulation predicted HSGPA better than cognitive ability did (β = .77 vs β = .22), yet cognitive ability predicted SAT scores better than self-regulation did (β = .74 vs β = .19). This led the researchers to conclude that self-regulation ratings “explained more variance in report card grades than in SAT scores, whereas a battery of cognitive ability tests explained more variance in SAT scores than in report card grades.” Further, the study also found that cognitive ability explained the association between SAT scores and graduation whereas self-regulation explained the association between high school GPA and graduation: “the incremental predictive validity of high school grades for college graduation was explained by self-regulation, whereas the incremental predictive validity of SAT scores for college graduation was explained by cognitive ability.”

The data suggests that cognitive ability is a very strong predictor of standardized test scores which cannot be explained by confounding with personality or self-discipline. For school grades, personality and self-discipline are better predictor than cognitive ability, but cognitive ability still has some predictive validity for school grades that cannot be completely explained by confounding with these variables. This is demonstrated by the fact that using both cognitive ability test scores and personality/self-discipline ratings has greater predictive validity than using only one of the predictors.

Genetic analysis

A number of studies have shown large genetic correlations and bivariate heritability estimates between cognitive ability and academic achievement. For example, in a review article of genetics and intelligence differences, Plomin and Deary (2015) [archived] review a number of studies showing large genetic correlations between intelligence and reading and mathematics performance. One such study was conducted by Davis et al. (2009) [archived]. Researchers assessed the general cognitive ability (g), reading performance, mathematics performance, language, and other aspects of academic achievement of subjects from the Twins Early Development Study (TEDS). They investigated 5,434 twin pairs at age 12 who were born in England and Wales in 1994, 1995, and 1996. The authors found that the genetic factors responsible for g overlapped with the genetic factors responsible for academic achievement. For example, the g had large genetic correlations with reading (r = .88), mathematics (r = .86), and language (r = .91) performance (Table 4).

  • In this figure, A = Additive genetic effects; C = Shared (common) environmental effects; E = non-shared environmental effects (see the ACE model as way to decompose the sources of phenotypic variance). Read this as, e.g., the genetic effects for reading and the genetic effects for mathematics correlated at 0.75.

Further, because the heritability of all of these traits was quite large, g had large bivariate heritability estimates with reading (.65), mathematics (.62), and (.53).

In other words, most of the association between cognitive ability and academic achievement was the result of shared genetic factors. The same applies to the association between different measures of academic achievement. The study ended with the following discussion on “generalist genes”:

The high heritability of our latent factors confirms that genetic effects continue to be important in the etiology of cognitive abilities and disabilities into early adolescence. However, going beyond this, the high genetic correlations between the latent factors representing g, reading, mathematics and language confirm that the genetic influences on these domains continue to be largely shared, in spite of accompanying hormonal and brain changes, with genes accounting for most of the phenotypic correlation between these domains. This implies that when genes influencing reading are found, for example, they are very likely to also be associated with mathematics, general cognitive ability and language.

Trzaskowski et al. (2013) [archived] examined the same sample to estimate the genetic correlation using DNA data of genotyped children instead of twin comparisons. Compatible with the previous study, researchers found that g had significant genetic correlations with reading (r = .89), mathematics (r = .74) and language (r = .81) performance (Table 1).

More recent data also shows large genetic correlations between cognitive ability and academic performance. Rimfeld et al. (2015) [archived] investigated the genetic and environmental influences on subjects of the age-16 UK-wide GSCE examination results for 12,632 twins. Using twin methodology and DNA data, researchers found that scores on all subjects were highly heritable at the end of compulsory education. However, more important for my purposes were their findings of the genetic association between GSCE grades and intelligence scores. Intelligence was assessed using verbal ability (Mill Hill Vocabulary tests) and non-verbal ability (Raven’s Progressive Matrices) tests administered online. Consistent with prior studies, the twin analyses showed that intelligence had large genetic correlations with grades in all GSCE subjects, including English (r = .65), mathematics (r = .69), and science (r = .61) (Table 2). The DNA analyses also showed large genetic correlations (some reaching 100%) between grades across different GSCE subjects. No genetic correlations between intelligence and GSCE grades were reported using the DNA analyses.

Reverse causation

Lastly, one might concede that there is causal association between cognitive ability and achievement, but they might argue that the direction of the causation is reversed. That is, academic achievement might be causing cognitive ability, rather than the other way around. While there is evidence that achievement influences cognitive ability, there is also evidence that cognitive ability influences achievement as well. That is, there is a reciprocal causal relationship between cognitive ability and academic achievement. These findings were reported in a meta-analysis by Peng et al. (2012) [archived]. These researchers analyzed hundreds of studies that examined the relationship between fluid intelligence and reading and mathematics performance. Consistent with studies cited earlier, they found that fluid intelligence was associated with reading (r = .38) and mathematics (r = .41) performance. The more important findings for my purposes here are the reported longitudinal associations. The researchers found support for a reciprocal causal association between cognitive ability and academic achievement. That is, early cognitive ability predicts later academic achievement even after controlling for early achievement, and early achievement predicts later cognitive ability even after controlling for early cognitive ability. The findings were as follows (page 204):

  • Across 42 studies involving 920 correlations, initial fluid intelligence predicted later reading performance even after partialing out initial reading performance (r = .17). The majority of studies focused on children before the age of 13. The time interval between testing spanned from .25 to 7 years. The time interval did not affect the relation.
  • Across 30 studies involving 275 correlations, initial fluid intelligence predicted later mathematics performance even after partialing out initial mathematics performance (r = .21). The majority of studies focused on children before the age of 14. The time interval between testing spanned from .5 to 4 years. The time interval did not affect the relation.
  • Across 9 studies involving 110 correlations, initial reading performance predicted later fluid intelligence even after partialing out initial fluid intelligence (r = .21). All studies focused on children before the age of 11. The time interval between testing spanned from 1 to 3 years. The relation became weaker as the time interval was larger.
  • Across 7 studies involving 52 correlations, initial mathematics performance predicted later fluid intelligence even after partialing out initial mathematics performance (r = .24). The majority of studies focused on children before the age of 11. The time interval between testing spanned from 1 to 3 years. The time interval did not affect the relation.

The authors summarized the longitudinal findings as follows (page 204):

[T]he findings, taken together, suggest that Gf significantly predicted later reading/mathematics performance when initial reading/mathematics performance was controlled for. Likewise, reading/mathematics also significantly predicted later Gf when initial Gf was controlled for. These findings are primarily based on data among children from a relatively short time intervals that generally did not affect these longitudinal relations.

Furthermore, recall the studies cited earlier showing that cognitive ability becomes fairly stable by the time children reach formal schooling (Schneider 2014, Yu et al. 2018). Thus, the strong association between cognitive ability and later academic achievement (Deary et al. 2006) cannot be completely explained by the hypothesis that cognitive ability is a mere byproduct of achievement.

Lastly, more evidence that cognitive ability predicts achievement independently of prior achievement comes from examining the relationship between SAT scores and college GPA. Sackett et al. (2012) [archived] investigated this relationship in several samples of students at colleges and universities. One involved a sample of 143,606 students across 110 colleges and universities in 2006. Another sample involved 136,725 students across 41 schools during 1995-1997. All samples showed that SAT scores provide incremental validity in predicting college grades, even after controlling for high school grades and parental SES (SES was measured using parental education and family income). For example, the 2006 data showed that freshman GPA correlated with SAT scores (r = .35), high school GPA (r = .37), and parental SES (r = .13). After entering all predictors into a regression model, freshman GPA still had modest to large associations with SAT (β = .25) and High school GPA (β = .30), but rather small associations with parental SES (β = .07). After correcting for range restriction, these regression coefficients increased slightly, but the broad pattern remained, with High School GPA being the best predictor of college freshman GPA, followed by SAT scores and SES:

Now, this isn’t ideal evidence for the causal influence of cognitive ability because SAT tests are not cognitive ability tests. However, the correlation between SAT tests scores and g scores are sufficiently large that some researchers have concluded that “the SAT is an adequate measure of general intelligence” (Frey et al. 2004, page 376).

In summary, cognitive ability is a robust predictor of academic achievement, which is not explained by confounders such as parental SES, family background, or personality. Particularly impressive is the fact that cognitive ability remains associated with later achievement even after controlling for prior achievement, sibling GPA, and when comparing siblings from the same family. The fact that cognitive ability remains a powerful predictor after controlling for such a robust set of confounders is strong evidence that there is a causal relationship between cognitive ability and academic achievement. That is, the Causal Hypothesis seems to be true for academic achievement.

Occupational performance


As shown in my previous post, there is a very large correlation between cognitive ability and occupational performance, such that cognitive ability is often the best predictor of occupational performance across a number of different settings (e.g., military vs non-military) and different measures of performance (e.g., work sample vs supervisor ratings).

The question now is whether the correlation between cognitive ability and occupational performance reflects as causal influence. Now, it is more difficult to demonstrate that cognitive ability has a causal influence on occupational performance, since there are no longitudinal datasets that include cognitive ability measured at one point and occupational performance measured years later (as far as I know). There is also less data that involves enough variables that allows us to control for possible confounders. The only study that I know of which attempts to control for possible confounders of the association between cognitive ability and job performance is Kuncel et al. (2014). The study found that cognitive ability predicted job performance even after controlling for family SES, consistent with the findings for other outcomes. However, I wouldn’t place too much weight on this study because the sample is small and it uses a limited measure of early family SES (early family SES is measured via self-report on a scale from 1 to 5, where 1 = “poor” and 5 = “wealthy”).

That being said, there is some indirect evidence that the association between cognitive ability and job performance here is causal. The first piece of evidence is the fact that there is strong evidence that cognitive ability has a causal influence on other predicted outcomes, such as academic achievement (as shown above), anti-social behavior, and socioeconomic status (to be shown below). Given that cognitive ability has causal influence on these other outcomes, and given that cognitive ability predicts occupational performance so well, it would be quite surprising if there were no causal influence on occupational performance. There would need to be some explanation as to why occupational performance differs so much from other outcomes predicted by cognitive ability.

Other indirect evidence that cognitive ability has a causal influence on occupational performance stems from the fact that cognitive ability predicts occupational performance so much better than the most plausible confounders. For example, let us return to the meta-analysis reported in Schmidt et al. (1998) [archived]. The meta-analysis found that general mental ability predicted job performance (r = .51) better than plausible confounders such as job knowledge (r = .48), conscientiousness (r = .31), reference checks (r = .26), job experience (r = .18), years of education (r = .10), and interests (r = .10) (Table 1). General mental ability also predicted job training performance (r = .56) better than plausible confounders such as conscientiousness (r = .30), reference checks (r = .23), job experience (r = .01), years of education (r = .20), and interests (r = .18) (Table 2). These findings cannot be explained by the hypothesis that the association between cognitive ability and occupational performance is the result of confounding with plausible covariates such as personality, job experience, education, etc.

Other indirect evidence of the causal influence of cognitive ability is the fact that cognitive ability predicts occupational performance despite explicit attempts to improve the performance of low-ability workers. For example, Gottfredson (1997) [archived] reported that some organizations have attempted to provide low-ability groups with additional training or instruction in order to reach parity with high-ability groups. These attempts have been mostly unsuccessful (page 86):

Additional evidence of the causal importance of g is provided by the many unsuccessful efforts to eliminate or short-circuit its functional link (correlation) with job proficiency. For example, there have been efforts to train the general cognitive skills that g naturally provides and that jobs require-such as general reading comprehension (which is important for using work manuals, interpreting instructions, and the like). Another approach has been to provide extra instruction or experience to very low-aptitude individuals so that they have more time to master job content. Both reflect what might be termed the training hypothesis, which is that, with sufficient instruction, low-aptitude individuals can be trained to perform as well as high-aptitude individuals. The armed services have devoted much research to such efforts, partly because they periodically have had to induct large numbers of very low-aptitude recruits. Even the most optimistic observers (Sticht, 1975; Sticht, Armstrong, Hickey, & Caylor, 1987) have concluded that such training fails to improve general skills and, at most, increases the number of low-aptitude men who perform at minimally acceptable levels, mostly in lower level jobs.

Gottfredson further states that differences in performance between high-ability and low-ability workers persist even as they acquire substantial experience:

Not even lengthy experience (5 years) eliminates differences in overall job performance between more and less bright men (Schmidt et al., 1988). A large study of military cooks, repairmen, supply specialists, and armor crewmen showed that performance may converge on simpler and oft-performed tasks (Vineberg & Taylor, 1972, p. 55-57). However, even that limited convergence took considerable time, reflecting large differences in trainability. It took men in the 10th to 30th percentiles of ability about 12 to 24 months to catch up with the performance levels on those tasks that were exhibited by men above the 30th percentile with no more than 3 months’ experience on the job. These findings from field settings are consistent with Ackerman’s (1987) review of the experimental literature relating skill learning and ability: individual differences in performance do not decrease with practice, and sometimes increase, when tasks are characterized by “predominantly inconsistent or varied information processing requirements.” In short, tasks that are not easily routinized continue to call forth g.

Socioeconomic success


Controlling for family background

In my last past, I cited data showing large associations youth cognitive ability and socioeconomic outcomes at adulthood. Much additional evidence suggests that this association is causal. For example, Herrnstein and Murray (1994) [archived] used the NLSY79 to estimate the predictive power of youth IQ on socioeconomic outcomes after controlling for parental SES. “Parental SES” is measured based on “information about the education, occupations, and income of the parents of NLSY youths” (page 131). The authors found that youth IQ predicted high school graduation, college degree attainment, and adulthood poverty even among participants with average parental SES (see chapters 5-12). Consider the following examples.

The probability of permanently dropping out of high school for whites with average IQ ( ~10%), was far higher than the same probability for whites with IQs one standard deviation below the mean ( ~26%) and for whites with IQs two standard deviations below the mean ( ~65%). See page 149.

The probability of attaining a bachelor’s degree for whites with average IQ ( ~10%) was far lower than the same probability for whites with IQs one standard deviation above the mean ( ~34%) and for whites with IQs two standard deviations above the mean ( ~75%). See page 152.

The probability of being in poverty for whites with IQs two standard deviations below the mean ( ~27%) was much higher than the same probability for whites with IQs one standard deviations below the mean ( ~15%), whites with mean IQs ( ~9%), whites with IQs one standard deviations above the mean ( ~4%), and whites with IQs two standard deviations above the mean ( ~2%). See page 134.

Keep in mind that these were the results among subjects with average parental SES. Similar analyses were done while fixing IQ to average (100 IQ) and varying parental SES. The general finding was that variations in IQ had a larger impact on socioeconomic outcomes than did variations in parental SES. If the predictive power of IQ was solely due to its correlation with parental SES, then we would not expect IQ to predict outcomes better than parental SES.

The results of Herrnstein and Murray were partially corroborated by a number of other studies. For example, Rindermann and Ceci (2018) [doi] found that “children’s cognitive ability is more important than parental income for children’s later income as adults” in the United States (page 21) based on data from the NLSY79. Eid (2018) [archived] performed a similar analysis using a newer data set: the 1997 National Longitudinal Survey of Youth (NLSY97). He specifically investigated the relative predictive power of IQ versus parental SES on adulthood poverty. The results of the study corroborate the findings from Murray and Herrnstein on the relative predictive power of IQ and parental SES, although the predictive power of both are smaller (page 3):

Without making any meaningful changes to HM’s methodology, we reaffirm the hypothesis that IQ is more important than family SES in avoiding poverty, though both of these covariates’ effects are smaller than those found by HM. Running a logistic regression with IQ, SES, and Age in 1997 as independent variables and poverty status in 2007 as the dependent variable, we find the IQ effect to be approximately three times the size of the SES effect.

One criticism of the above studies is that they rely on a narrow measure of parental SES. Some people have objected that the effects of family background are broader than the narrow effects of Herrnstein and Murray’s measure of parental socioeconomic status (parental income, occupational status, and education). Therefore, the objection goes, when estimating the effects of cognitive ability on socioeconomic outcomes, one must control for a broader set of family background variables than parental income, occupation, and education. Studies using broader controls for family background have replicated Herrnstein and Murray’s finding that cognitive ability has large and robust effects on socioeconomic outcomes.

In “A Reanalysis of The Bell Curve: Intelligence, Family Background, and Schooling”, Koreman and Winship (2000) [archived] estimated the effect of AFQT on various socioeconomic outcomes after controlling for a broader set of family background variables. The additional variables included: family arrangement (e.g. two-parent, single-parent, etc.), urbanicity, number of siblings in the home, mother’s age at birth, and a number of other variables (see page 154). They also include controls for age, age at time of assessment, gender, year, and race/ethnicity (Koreman and Winship analyze all subjects in the sample rather than just non-Hispanic whites, unlike Herrnstein and Murray). The results were as follows:

  • Before adding controls for the additional family controls, the authors performed a regression analysis while controlling only for Herrnstein and Murray’s narrow measure of parental SES. Replicating Herrnstein and Murray’s finding, they report large regression coefficients for zAFQT (z-score for AFQT score) on annual earnings ($4,866), poverty (−0.95), years of schooling completed (0.62), high school dropout (−1.82), B.A. degree attainment (1.76) and a number of other outcome variables after controlling for parental SES (Table 7.4, see column 1).
  • Surprisingly, after introducing the broader set of measures for family background into the model, the effect of AFQT was not substantially reduced. For example, the regression coefficient for zAFQT was still substantial for annual earnings ($4,669), poverty (−0.93), years of schooling completed (0.58), high school dropout (−1.76), and B.A. degree attainment (1.72) after controlling for broader family background (Table 7.4, see column 3).

These findings imply that “when family arrangement and the other family socioeconomic background variables are included in the models, the effect of AFQT is virtually unchanged” (page 158).

Belley and Lochner (2007) [archived] estimate the effects of cognitive ability using a more recent cohort – the 1997 National Longitudinal Survey of Youth (NLSY97) (the study also analyzed the NLSY79, but I will report the findings only for the NLSY97). The authors compared the relative effects of cognitive ability and family income by splitting subject into AFQT and parental income quartiles. They found that “income has moderate effects on high school completion for the lowest ability types and small effects for the higher ability quartiles” (page 13). In other words, family income had large effects only for students of low cognitive ability. On the other hand, cognitive ability had strong effects on high school completion at every quartile. For example, among students in the highest income quartile, the high school completion rate for the highest AFQT-quartile students was nearly 100%, whereas the rate for the lowest AFQT-quartile students was a bit under 90% (Figure 1b). Among students in the lowest income quartile, the high school completion rate for the highest AFQT-quartile was again nearly 100%, whereas the rate for the lowest AFQT-quartile students was slightly over 60%.

The authors then compared the relative effects of cognitive ability and family income on educational attainment after controlling for an expansive set of controls, including sex, race, mother’s age at birth, intact family during adolescence, urban/metropolitan area during adolescence, number of siblings under 18, and mother’s education. The authors then compared the effects of being in a certain AFQT quartile or family income quartile while holding all of the aforementioned variables fixed. They found that AFQT had far larger effects on educational attainment than family income:

  • Scoring in the top AFQT quartile was associated with a 21 percentage point increase in probability of completing high school, whereas having a family income in the top quartile was only associated with a 7 percentage point increase in the same probability (see table 3).
  • Scoring in the top AFQT quartile was associated with a 45 point increase in the probability of completing 4+ years of college by age 23, whereas having a family income in the top quartile was only associated with a 10 point increase in the same probability (table 6, for NLSY97 participants). The same pattern was found for college attendance.

One interesting finding regarding the effects of cognitive ability and parental background on economic outcomes is that cognitive ability tends to have a stronger effect later in one’s career whereas parental background seems to have a stronger effect earlier in one’s career. For example, consider Ganzach (2011) [archived] who investigated a sample of high school graduates from the NLSY79. He used two measures of SEB: (a) a narrow index measured as a composite of parental education, family income, and parental occupational status, and (b) an extended index which included a number of other variables, including number of siblings, whether the participant lived in a two-parents home at age 14, a school composite based on the percent of economically disadvantaged students and non-white students, and a number of other variables (see page 124 for the full list). The results showed that “SEB affected wages solely by its effect on entry pay whereas intelligence affected wages primarily by its effect on mobility. The effect of intelligence on entry pay seems to be weaker than the effect of SEB” (page 127). In other words, both intelligence and SEB impacted entry pay, but only intelligence affected the pace of pay increases throughout one’s career. The following figure shows the trajectory of wages at different levels of intelligence and socioeconomic backgrounds:

Judge, Klinger, and Simon (2010) discovered similar findings while investigating the relationship between general mental ability (GMA) and career success over a 28-year period among participants in the NLSY79. General mental ability was measured using the Armed Forces Qualifying Test (AFQT) in 1980. Researchers controlled for age, gender, race, and a SES composite at the onset of the study. Participants were placed into two groups: high-GMA participants (those scoring one standard deviation above the mean) and low-GMA participants (those scoring one standard deviation below the mean). Researchers found that outcome gaps between high-GMA and low-GMA widened dramatically over time. For example, the income gap between high-GMA and low-GMA participants grew about 25-fold from $1,575 in 1979 ($5,191 vs $3,616) to $38,819 in 2006 ($62,301 vs $23,482) (Figure 2a). The occupational prestige gap grew 6-fold from 7.67 points in 1979 (39.54 vs 31.87) to 49.79 points in 2006 (82.47 vs 32.68) (Figure 2b). See figures here:

Similar trends were found regarding human capital accumulation over time: the gap in education, training, and job complexity between high-GMA and low-GMA participants widened significantly over time (Figure 3). Finally, improvements in education, training, and job complexity were more likely to translate into larger improvements in income and occupational prestige for high-GMA participants (Figures 4 and 5).

The above studies in this section all analyzed data from the same source – the National Longitudinal Survey of Youth. Spengler et al. (2018) [archived] used a different dataset, Project Talent, to produced similar findings as above. Project Talent includes a longitudinal sample of over 81,000 participants followed from high school to late adulthood. The dataset contains information about each participant’s parental SES, personality traits, and cognitive ability while they were in high school. It also contains information on the socioeconomic outcomes (educational attainment, occupational prestige, and annual income) of the participants at two points during adulthood, one that is 11 years after the initial sampling and another that is 50 years after the initial sampling. Parental SES was a composite score consisting of home value, family income, parental education, father’s job status, number of books, number of appliances, number of electronics, and whether the child had a private room. For each outcome, the authors developed regression models to analyze the relative predictive validity of each independent variable after holding each of the other variables constant. All regression models included controls for race, sex, age, IQ, and parental SES. Other regression models included controls for other variables while the subjects were in high school, including interest in school, self-perceived reading skills, self-perceived writing skills, and self-perceived responsibility. The regression models predicting each outcome during the 50-year follow-up were as follows:

These results show that IQ predicts educational attainment, occupational prestige, and income 50 years after high school even after controlling for parental SES, self-perceived academic skills, and basic demographic controls. In fact, IQ predicts these socioeconomic outcomes better than every other variable, with the exception of sex and income (also, model A.5 shows that the regression coefficient of race on educational attainment is .92, but this is almost certainly a typo).

An impressive finding from this study was that IQ predicted occupational prestige and income even after controlling for educational attainment, particularly when these outcomes are measured later in life. For outcomes at the 11-year follow-up, about two-thirds of the association between IQ and occupational prestige and income was mediated by educational attainment (Table 9). In other words, IQ predicted occupational prestige and income 11 years later, and two-thirds of this influence was explained by the fact that IQ is associated with educational attainment and educational attainment is associated with higher occupational prestige and income. For outcomes at the 50-year follow-up, educational attainment only mediated about 50% of the influence of IQ on occupational prestige and only about 32% of the influence of IQ on income. In other words, most of the association between youth IQ and income in late adulthood was not explained by educational attainment.

Damian et al. (2015) [archived] used the same dataset to argue that high intelligence may be able to compensate for low parental SES (page 18):

In the case of personality traits, neither the main effects nor the interactive effects were large enough to compensate for low parental SES and this can be seen on the response surface figures…Therefore, personality traits, while important in the prediction of attainment outcomes, did not suffice to make up for low parental socioeconomic status. However, the story was different for intelligence. There, the slope of the line of perfect disagreement was positive for educational attainment and occupational prestige (and flat for income), indicating that as the discrepancy increased (such that IQ was higher), educational attainment and occupational prestige increased. In other words the smartest but least wealthy people were close to getting a college degree, whereas their least smart but wealthiest counterparts were close to getting an associate’s degree. In sum, even though we have evidence that certain personality traits may compensate for background disadvantage (in the absence of intelligence controls), the effects were not large enough to overcome the main effect of SES. The only individual difference that seemed to be able to do that was intelligence. Thus, we would conclude that the American Dream, as manifest through personality, is more myth than fact. On the other hand, the American Dream manifest through intelligence is still alive and well.

The persistent association between early cognitive ability and later socioeconomic success even after controlling for parental SES is a consistent finding. This finding has been replicated in numerous other countries, such as Britain (Bukodi et al. 2013, tables 3-4; Von Stumm 2009), Scotland (Von Stumm et al. 2010Deary et al. 2005), Sweden (Bergman et al. 2014Sorjonen et al. 2012), Denmark (Hegelund 2018), Ireland (O’Connell and Marks 2021), Germany (Becker et al. 2019), and New Zealand (Fergusson et al. 2005). For example, Fergusson et al. (2005) found that “intelligence had a direct relationship to later educational, occupational and related outcomes independently of other childhood characteristics and family environment” (page 856).

Sibling analysis

In this section, I will list studies showing that sibling analyses replicate the findings reported above: early cognitive ability demonstrate a robust association with later socioeconomic outcomes even after controlling for family background.

In “The Predictive Value of IQ”, Sternberg et al. (2001) [archived] reports that higher-IQ brothers tend to achieve higher socioeconomic success as adults than their lower-IQ brothers (page 9):

Jencks (1979) observed that if two brothers who grew up in the same family were compared on their SES as adults, the brother who had the higher IQ in adolescence would tend to have the higher adult social status and income. This path, however, is mediated by amount of education. The higher-IQ brother would be more likely to get more education and, correspondingly, to have a better chance of succeeding socioeconomically.

Murray (1998) [archived] used the NLSY79 to measure the association between cognitive ability and socioeconomic outcomes in 1993 among siblings from the same household (Chapter 4). He limited his sample to pairs of participants who were full biological siblings and who lived in the same home with both biological parents at least until the younger sibling’s 7th year. The sample was further restricted to include only sibling pairs in which one sibling had an AFQT score in the “normal range” (between 25th and 74th percentile, or 90-109 IQ) and the other sibling had an AFQT score outside of this range. For each sibling pair, the sibling with the AFQT score in this “normal range” was considered the “reference sibling”. The other sibling (with an AFQT score outside of this range) was considered the “comparison sibling”. Each comparison sibling was classified as “very bright” (AFQT score in the 90th+ percentile), “bright” (75th-89th percentile), “dull” (10th-24th percentile), or “very dull” (below in the 10th percentile). This resulted in a total sample of 1,074 sibling pairs. The results showed that siblings with higher AFQT scores achieved far higher levels of socioeconomic success than lower-scoring siblings (Tables 4-3, 4-4, and 4-7):

Cognitive Class (percentile range)Percentage of comparison siblings with a B.A. (reference sibling lacks B.A.)Percentage of comparison siblings with a B.A. (reference sibling has B.A.)Difference in Duncan ScoreDifference in Median Earnings
Very Bright siblings (90th+)58.7%91.3%+10.9+$11,500
Bright siblings (75th – 89th)41.8%75.6%+4.1+$4,000
Normal reference group (25th – 74th)(0)(100)(mean = 42.7, SD = 21.5)$22,000
Dull siblings (10th – 24th)1.2%18.2%−10.4−$5,000
Very Dull siblings (10th−)0.6%0%−18.0−$9,750

These results are largely in line with a sibling analysis by Koreman and Winship (2000) [archived], which found (page 165) results that “mostly confirm with sibling analyses Herrnstein and Murray’s finding that the effects of AFQT are substantial and robust” (although they argue that Herrnstein and Murray underestimate the effects of family environment).

Murray (2002) [archived] repeated his sibling analysis with what he calls a “Utopian Sample”. He limited his analysis to siblings in which “all parents were wed when the child was born (zero illegitimacy) and all young children were brought up by both biological parents during their most formative years (zero early divorce)” (page 339) and in which the parents’ income exceeded the 25th percentile. This resulted in 733 sibling pairs that had “virtually no illegitimacy, divorce, and poverty” (page 339). Even within this idealized sample, Murray found large differences in socioeconomic outcomes between siblings with different AFQT scores. For example, the median family income earned by “very bright” siblings ($70,700) was far higher than the median family income earned by “bright” siblings ($60,500), “dull” siblings ($39,400), “very dull” siblings ($23,600), and the reference group siblings who scored in the “normal” (90th-109th percentile) AFQT range ($52,700) (table 2).

The analyses by Murray demonstrate that the association between early cognitive ability and later socioeconomic outcomes is not primarily due to confounding with familial effects. These findings have been replicated in other studies. For example, Bronars and Oettinger (2006) used the NLSY79 to estimate the returns to schooling and cognitive ability by performing a similar sibling analysis. Their results were largely in line with those reported by Murray, which is that the effect of cognitive ability is not due to confounding with family effects (page 28):

…our estimates suggest that a given difference in AFQT residuals causes roughly the same wage gap between siblings as between unrelated individuals, all else equal. This finding suggests that AFQT scores are valuable measures of labor market skill and that the return to AFQT in OLS regressions is not an artifact of unmeasured family attributes.

Sibling analyses in other countries have also found the important effects of early cognitive ability measures. Hegelund et al (2019) [archived] investigated 364,193 Danish men with at least one full brother to examine the relationship between IQ and later socioeconomic success. The authors employed a conventional cohort analysis and a within-sibship analysis to determine the degree to which the association between IQ and outcomes persisted after controlling for “familial factors shared by siblings”. IQ was measured at age 18 and socioeconomic outcomes were measured at age 30. Consistent with prior studies, the authors found that IQ was associated with gross income in their cohort analysis. In the within-sibship analysis, some of this relationship was reduced, but the majority of the relationship remained (see Figure 3). The authors concluded that while “an appreciable part” of the associations between IQ and socioeconomic outcomes at age 13 could be attributed to familial factors shared by siblings, they found that “most of the associations were not attributable to such factors” (page 106). For example, this graph shows that the correlation between IQ and income barely changes regardless of whether one uses the within-sibship or cohort analysis:

These sibling analyses show that early cognitive ability has robust predictive validity on later socioeconomic outcomes even after controlling for family effects. Before finishing the section on sibling analyses, I would like to emphasize one important insight from these analyses. The sibling data from the NLSY79 suggests that, when estimating the independent effect of cognitive ability, controlling for Herrnstein and Murray’s narrow measure of parental SES (parental income, education, and occupation) is about as good as controlling for all family effects via sibling analyses. In the book Income Inequality and IQMurray (2002) [archived] emphasized this point in the chapter titled “How much difference does it make whether the analysis controls for parental SES or compares siblings”. In this chapter, he concludes that it doesn’t make much of a difference (page 24):

We may produce a more direct comparison by taking the entire sample of siblings (3,802 individuals, comprising 2,859 unique sibling pairs) and running two regression equations. In the first, the 3,802 individuals are treated as unrelated subjects, and the independent effect of IQ on earned is computed after entering the parental SES index into the equation. In the second, the sample consists of the 2,859 sibling pairs, the dependent variable is the sibling difference in earned income, and the sole independent variable is the sibling difference in IQ. The former equation produces a coefficient for IQ of $446. The latter produces a coefficient for IQ of $453. The Bell Curve’s method of controlling for SES and the sibling method of controlling for everything in the family background yield interpretations of the independent role of IQ on income that are substantively indistinguishable.

In other words, whether one controls for parental SES or everything in the family background, the independent effect of cognitive ability on income is virtually identical. These findings were also corroborated in “A Reanalysis of The Bell Curve: Intelligence, Family Background, and Schooling” by Koreman and Winship (2000) [archived], which “present results from family fixed-effect (sibling difference) analyses. With a few exceptions, the fixed-effects estimates for AFQT are remarkably similar to the standard OLS and logit estimates” (page 146). For example, among the sample of siblings in the NLSY79, the regression coefficient for zAFQT (z-score for AFQT scores) on annual earnings was similar regardless of whether one controlled for Herrnstein and Murray’s narrow measure of parental SES (coefficient of zAFQT = $5,548) or one controlled for all sibling fixed-effects (coefficient of zAFQT = $5,317) (table 7.2, compared columns 4 and 7).

Genetic analysis

A number of studies have found genetic correlations between cognitive ability and measures of socioeconomic outcomes. For example, Rowe et al. (1998) examined 1,943 full-siblings and 129 half-siblings from the NLSY79 to estimate the association between IQ at youth and education and income at young adulthood. IQ was measured using AFQT scores when participants were between 14 and 23 years old. Years of education and income were assessed when participants were between 28 and 35 years old. Consistent with prior studies, the study reported significant heritability for IQ (0.64), education (0.68), and income (0.42). Also consistent with prior studies, IQ was estimated to be significantly correlated with both education (r = .63) and income (r = .34). Most importantly, the study found that 68% of the IQ-education correlation and 59% of the IQ-income correlation was the result of common genetic factors. In other words, most of the association between cognitive ability and the socioeconomic outcomes studied here were the result of genetic influences. The study concluded with the following:

This study’s findings also bear on the interpretation of the association among IQ, educational attainment, and income. Within a single generation, it is clear that a substantial part of these associations is genetic. The model-predicted correlations, which were close to the observed ones, were 0.63 between IQ and education and 0.34 between IQ and income. Sixty-eight percent of the former correlation and 59% of the latter was genetically mediated; the remainder was mediated by common shared environment. As genes may create variation in SES through g -> e correlations, any interpretation of IQ-SES associations as purely environmental is untenable.

Luciano et al. (2010) studied a sample of 1,983 families (6,086 individuals) from the Generation Scotland study to examine the causes of the associations between cognitive ability and various risk factors for cardiovascular disease. The authors used an extended pedigree design to examine to what extent these associations were the result of genetic or environmental sources. The authors investigate a number of risk factors, but the risk factors relevant for my purpose here are education and income. Cognitive ability was measured by extracting a g factor from several tests of different cognitive abilities (memory, processing speed, executive function, verbal ability, see page 306). Consistent with prior studies, researchers found that g was significantly correlated with education (r = .44) and income (r = .30) (Table 3). Significant heritability estimates were also reported for g (0.43), education (0.58), and income (0.27) (Table 2). The bivariate heritability estimates were 0.73 between g and education and 0.50 between g and income. In other words, genetic factors explained most of the association between cognitive ability and socioeconomic position. The authors concluded with the following (page 311):

The relationship between g and socioeconomic variables was mediated foremost by genes with additional, equal mediation by maternal effects and the unique environment. The genetic and especially maternal effect correlations were strong, indicating a very high degree of similarity in these factors’ influence on the individual measures. Educational attainment showed greater similarity in these effects than average income, which is unsurprising considering the history of psychometric testing within the education system. If the maternal effect for income level reflects common environment, then combined effects of common and unique environment explain half of the covariance between g and income.

The previous studies estimated genetic correlations between cognitive ability and socioeconomic outcomes by analyzing relatives of varying levels of genetic relatedness. These findings have also been reported by studies that estimate heritability and genetic correlations from the DNA of large samples of people. For example, Marioni et al. (2014) [archived] used the same dataset as the previously cited study – Generation Scotland: the Scottish Family Health Study – to analyze the associations between cognitive ability and various social outcomes in a sample of 6,815 unrelated subjects from the Generation Scotland study who underwent genome-wide testing. The DNA analysis found that the genetic correlation between educational attainment and g to be r = .95 (Table 3). In other words, nearly all of the genetic factors associated with g where associated with educational attainment. The authors note that “this suggests that the genetic influences on the two traits overlap substantially.” The bivariate heritability was estimated to be 0.59 (Table 3), again indicating that most of the association between g and education was due to shared genetic factors. The study also reported small genetic correlations between g and the socioeconomic status (using an index called SIMD), but these findings were less interesting because it used socioeconomic status at the regional level rather than the individual level.

Similar findings were reported by Gale and Deary (2016) [archived]. These authors conducted a genome-wide association study (GWAS) to examine possible genetic influences on social deprivation and household income using over 100k participants of UK Biobank. They reported genetic correlations between these measures of socioeconomic success and a large set of phenotypes previously shown to be associated with SES. The genetic correlation between household income and childhood cognitive ability is of particular concern here. The average age of participants during intelligence assessments was 11 years. The SES measurements were taken at a mean of 57 years. Researchers found that the genetic correlation between childhood cognitive ability and adult household income was r = .67, indicating that most of the genes that influenced household income also influenced cognitive ability. There were also genetic correlations between childhood cognitive ability and adult social deprivation (r = .50), but this was measured at the regional level rather than individual level.

In summary, cognitive ability is a robust predictor of socioeconomic outcomes (measured in terms of education, occupation, and income), which is not explained by confounders such as parental SES, family background, or personality. Particularly impressive is the fact that cognitive ability remains associated with later socioeconomic outcomes even after controlling for earlier socioeconomic outcomes (see Spengler et al. 2018) and when comparing siblings from the same family. The fact that cognitive ability remains a powerful predictor after controlling for such a robust set of confounders is strong evidence that there is a causal relationship between cognitive ability and these socioeconomic outcomes. That is, the Causal Hypothesis seems to be true for socioeconomic success.

Anti-social behavior


Controlling for family background

Someone might state that the association between cognitive ability and anti-social behaviors is not due to a causal effect of cognitive ability. Rather, one might argue that this is the result of confounding. For example, perhaps parental socioeconomic environment is the cause of both cognitive ability and anti-social behavior. But there is plenty of evidence that cognitive ability remains (negatively) associated with criminal engagement even after controlling for socioeconomic status and other risk factors for criminality. I’ll review some of that evidence in this section.

For example, recall that the meta-analysis by Ttofihi et al. (2016) [archived] reported that low intelligence predicts criminal offending even after controlling for various risk factors for crime, such as poor child rearing, prevalence of childhood antisocial behavior, marital disturbance, father imprisonment, etc.

Additional studies show that cognitive ability predicts anti-social behavior even after controlling for parental socioeconomic status. For example, Herrnstein and Murray (1994) [archived] used the NLSY79 to estimate the predictive power of youth IQ on a number of measures of anti-social behaviors after controlling for parental SES. As stated earlier, “parental SES” is measured based on “information about the education, occupations, and income of the parents of NLSY youths” (page 131). The authors found that youth IQ predicted adulthood illegitimacy, welfare usage, incarceration, and unemployment even among participants with average parental SES (see chapters 5-12). Consider the following examples:

The percentage of white men who were incarcerated at the time of the interview for white men with average IQs ( ~3%) was far lower than the same percentage for white men with IQs one standard deviation below the mean ( ~6%) and for white men with IQs two standard deviations below the mean ( ~14%). See page 249.

The percentage of white men who were unemployed for at least one month in 1989 for white men with IQs two standard deviations below the mean ( ~14%) was far higher than the same percentage for white men with IQs one standard deviations below the mean ( ~11%), white men with mean IQs ( ~8%), white men with IQs one standard deviations above the mean ( ~6%), and white men with IQs two standard deviations above the mean ( ~4%). See page 164.

The percentage of white women who had an illegitimate first birth for white women with IQs two standard deviations below the mean ( ~33%) was much higher than the same percentage for white women with IQs one standard deviation below the mean ( ~22%), white women with mean IQs ( ~14%), white women with IQs one standard deviations above the mean ( ~8%), and white women with IQs two standard deviations above the mean ( ~4%). See page 183.

The percentage of white women who went on welfare within a year after birth for white women with IQs two standard deviations below the mean ( ~46%) was much higher than the same percentage for white women with IQs one standard deviation below the mean ( ~35%), white women with mean IQs ( ~24%), white women with IQs one standard deviations above the mean ( ~15%), and white women with IQs two standard deviations above the mean ( ~10%). See page 195.

Note that this last graph involved controls for the poverty status and marital status of the mother. Keep in mind that all of these figures were the results among subjects with average parental SES.

These results were partially corroborated by Levine (2011) [archived] who examined the relationship between criminality and IQ using the same dataset (the NLSY79). One benefit of Levine’s analysis is that he has access to more recent data, with incarceration frequencies from 1982 until 2006. Unsurprisingly, he found a strong association between IQ and incarceration: individuals who were incarcerated had significantly lower IQ scores than those who were not (89.6 vs 100.7, page 1235). After controlling for parental SES (a composite of parental education, father’s occupational prestige, and family income), there was still a significant relationship between IQ and incarceration (Table 1).

The results of the study showed that, among individuals with the same parental SES, individuals with lower IQs were far more likely to be incarcerated (see Figure 1). In fact, individuals with the highest SES and lowest IQ were more likely to be incarcerated than individuals with the lowest SES and highest IQ. The study concluded as follows (page 1237):

[T]he current findings are based on a large-scale longitudinal dataset that consists of a large representative sample. Collectively, the results generally suggest that low SES, low IQ and their interaction prospectively predict crime modestly yet fairly consistently. Thus it may be concluded that it is appropriate to integrate both low IQ and SES in explanations of crime.

Studies from other datasets also show that youth cognitive ability predicts criminality later in life after controlling for socioeconomic status. For example, a longitudinal study by Loeber et al. (2012) [archived] used data from the Pittsburgh Youth Study to examine the relationship between IQ at about age 12 and criminal history at age 28 in a sample of 422 males. IQ was measured using the Wechsler Intelligence Scale for Children–Revised (WISC-R) test. Parental socioeconomic status was measured using Hollingshead’s Index, which is a composite measure of social status based on “marital status, retired/employed status, educational attainment, and occupational prestige”. The results were that IQ was significantly associated with arrest probability for any charge, particularly during adolescence. For example, after controlling for parental SES, race, and age, the probability of arrest for 17-year-old males with low IQs (60-65%) was about three times the probability for those with high IQs (20-25%) (Figure 1).

Low-IQ and high-IQ males were those with IQs one standard deviation below and above the mean, respectively.

Unsurprisingly, studies have also found that cognitive ability predicts criminal victimization, even after controlling for parental SES. For example, Boutwell et al. (2017) used two datasets – the NLSY97 and the CNLSY – to explain correlation between intelligence and criminal victimization. The researchers created various regression models that included family income, age, race, sex, and cognitive ability quartiles (measured using ASVAB scores in the NLSY97 and PIAT/PPVT scores in the CNLSY) as independent variables and victimization as the dependent variable. The study found that lower quartiles of cognitive ability were associated with criminal victimization even after controlling for the other variables in the model (race, sex, family income, and age):

  • In the NLSY97, the odds of victimization for participants in the lowest and second-lowest quartiles of ASVAB scores were more likely to report victimization (odds ratios = 1.01 and 1.14, respectively) compared to participants in the highest and second-highest quartiles (odds ratios = .93 and .75, respectively). See Table 2.
  • In the CNLSY, the odds of victimization for participants in the lowest and second-lowest quartiles of PIAT mathematics scores were more likely to report victimization (odds ratios = 1.15 and 1.04, respectively) compared to participants in the highest and second-highest quartiles (odds ratios = .79 and .90, respectively). See Table 3.

For other studies demonstrating the predictive power of cognitive ability on criminal offending after controlling for parental SES, see Moffitt et al. (1981) [archived] and Lynam et al. (1993) [doi].

As stated in the section on socioeconomic success, one criticism of the above studies is that they control for relatively narrow measures of familial environment (e.g., many of the studies only control for parental education, occupation, and income). One might wonder whether the association between cognitive ability and anti-social behavior persists after controlling for broader measures of family background. In fact, there is evidence that the relationship does persist after controlling for such broader measures.

For example, in “A Reanalysis of The Bell Curve: Intelligence, Family Background, and Schooling”, Koreman and Winship (2000) [archived] estimated the effect of AFQT on various socioeconomic outcomes after controlling for a broader set of family background variables: family arrangement (e.g. two-parent, single-parent, etc.), urbanicity, number of siblings in the home, mother’s age at birth, and a number of other variables. (see page 154). They also include controls for age, age at time of assessment, gender, year, and race/ethnicity (Koreman and Winship analyze all subjects in the sample rather than just non-Hispanic whites, unlike Herrnstein and Murray). The results were as follows:

  • Before adding controls for the additional family controls, the authors performed a regression analysis while controlling only for Herrnstein and Murray’s narrow measure of parental SES. Replicating Herrnstein and Murray’s finding, they report large regression coefficients for zAFQT (z-score for AFQT score) on unemployment (−.44), incarceration (−.91), welfare usage measured by AFDC use (−.54), and illegitimacy (−.46) after controlling for parental SES (see Table 7.4, column 1).
  • Surprisingly, after introducing the broader set of measures for family background into the model, the effect of AFQT was not substantially reduced. For example, the regression coefficient for zAFQT was still substantial for unemployment (−.42), incarceration (−.88), welfare usage measured by AFDC use (−.54), and illegitimacy (−.45) after controlling for broader family background (see Table 7.4, see column 3).

These findings imply that “when family arrangement and the other family socioeconomic background variables are included in the models, the effect of AFQT is virtually unchanged” (page 158). In other words, regardless of whether one controls for the narrow measure of parental SES (parental income, education, and occupation) or the broader measure of family background, the effect of cognitive ability on these anti-social behaviors is about the same.

Finally, there is evidence that cognitive ability remains associated with self-control even after controlling for various measures of parental socioeconomic status. For example, Meldrum et al. (2017) [archived] examined the association between intelligence and self-control in a national sample of United States children. Intelligence was measured using the Wechsler Abbreviated Scale of Intelligence (WASI) which was administered to children during the 4th grade. Self-control was assessed by teacher-reports at the 4th, 5th, and 6th grades and by mothers at the 4th, 5th, and 6th grades as well as at age 15. For parental measures of self-control, the authors constructed a regression model with a number of predictors including sex, race, maternal education, nuclear family structure, maternal age at birth, maternal self-control, child intelligence at grade 4, and a number of other predictors. The authors found that maternal low self-control was the best predictor of child self-control (β = −.23, Table 3). The second-best predictors were parental socialization at grade 3 (β = .15) and child intelligence at grade 4 (β = .15), showing that child intelligence is a significant predictor of later self-control even after controlling for a rich set of covariates. Impressively, the authors found that childhood intelligence predicted later self-control even after controlling for prior self-control (page 6):

We re-estimated all of the models presented in Tables 2 and 3 when controlling for prior child self-control assessed during the third grade (teacher-reported and mother-reported self-control, respectively) to guard against the possibility that prior self-control could be driving the observed associations between child intelligence and self-control. Informatively, with the exception of the model predicting maternal reported self-control at fifth grade, all other models indicated that a statistically significant association (p < 0.05) between child intelligence and self-control remained when accounting for prior self-control. Specifically, the standardized effects for the association between child intelligence at fourth grade and teacher-reported self-control were as follows: 0.09 (predicting self-control at fourth grade); 0.08 (predicting self-control at fifth grade); 0.17 (predicting self-control at sixth grade). The standardized effects for the association between child intelligence at fourth grade and mother-reported self-control were as follows: 0.09 (predicting self-control at fourth grade); 0.04 (predicting self-control at fifth grade, non-significant); 0.07 (predicting self-control at sixth grade); 0.09 (predicting self-control at age 15).

These results led the researchers to the following conclusion (page 6):

Using data drawn from a national sample of children in the United States, our findings aligned with previous empirical work revealing a positive and significant association between intelligence and self-control (e.g., Berg et al., 2014; Boisvert et al., 2013; Petkovsek & Boutwell, 2014). Impressively, this association persisted across time and across different raters, despite adjusting for prior self-control, maternal intelligence and self-control, and child executive functioning. What our findings suggest is that, at the phenotypic level, intelligence is associated with a greater ability to regulate one’s impulses, emotions, and behavior, and may further explain why these two traits (intelligence and self-control) in general are so closely related to important life outcomes such as success in primary and secondary education, economic achievement, and avoiding contact with the criminal justice system (Beaver, Schwartz, et al., 2013; Gottfredson, 1997; Moffitt et al., 2011).

Sibling Analysis

Murray (2002) [archived] used the NLSY79 to measure the association between cognitive ability and socioeconomic outcomes among siblings from the same household (Chapter 4). He limited his sample to pairs of participants who were full biological siblings and who lived in the same home with both biological parents at least until the younger sibling’s 7th year. The sample was further restricted to include only sibling pairs in which one sibling had an AFQT score in the “normal range” (between 25th and 74th percentile, or 90-109 IQ) and the other sibling had an AFQT score outside of this range. For each sibling pair, the sibling with the AFQT score in this “normal range” was considered the “reference sibling”. The other sibling (with an AFQT score outside of this range) was considered the “comparison sibling”. Each comparison sibling was classified as “very bright” (AFQT score in the 90th+ percentile), “bright” (75th-89th percentile), “dull” (10th-24th percentile), or “very dull” (below in the 10th percentile). Murray found that “brighter” siblings were far less likely to have children out of wedlock than their “duller” siblings:

The “Utopian subsample” includes the siblings in which “all parents were wed when the child was born (zero illegitimacy) and all young children were brought up by both biological parents during their most formative years (zero early divorce)” (page 339) and in which the parents’ income exceeded the 25th percentile. This resulted in sibling pairs that had “virtually no illegitimacy, divorce, and poverty” (page 339). This data shows that, even among siblings from the same family, children with higher IQs were much less likely to give birth out of wedlock.

These results are largely in line with a sibling analysis by Koreman and Winship (2000) [archived], which found (page 165) results that “mostly confirm with sibling analyses Herrnstein and Murray’s finding that the effects of AFQT are substantial and robust” (although they argue that Herrnstein and Murray underestimate the effects of family environment). For example, the regression coefficient for zAFQT on unemployment, incarceration , and welfare usage after controlling for Herrnstein and Murray’s measure of parental SES (parental education, occupation, income) was −.44, −.91, and −.54, respectively (Table 7.2, see column 1). The same regression coefficient found in the sibling analysis was −.47, −.91, and −.84 (Table 7.2, see column 7), showing that the association between cognitive ability and these measures of anti-social behavior persists regardless of whether one controls for Herrnstein’s and Murray’s narrow measure of parental SES or broader measures of home environment. However, one interesting finding is that the regression coefficient for zAFQT on illegitimacy decreased in magnitude from −.46 to −.10 (Table 7.2), suggesting that much of the association between cognitive ability and illegitimacy may be due to confounding.

Sibling analysis have also found associations between cognitive ability and anti-social behaviors in other countries. For example, Frisell, Pawitan, and Långström (2012) [archived] examined the association between cognitive ability at age 18 and violent offending among all men in Sweden born between 1961–1975. In total, the researchers analyzed 238,390 full-brothers, 17,594 half-brothers raised together, and 25,148 half-brothers raised apart. Data on adolescent cognitive ability was gathered using the Conscript Register (until 2007, conscription at age 18–20 was mandatory for all Swedish men). Information on socioeconomic characteristics (income, single mother, and urbanicity) was retrieved from the 1970 and 1975 national censuses. Finally, data on criminal convictions was gathered using the The Crime Register, which “covers all convictions in lower court from 1973 and onwards”. Violent offenses included “homicide, assault, robbery, threats and violence against an officer, gross violation of a person’s/woman’s integrity, unlawful coercion, unlawful threat, kidnapping, illegal confinement, arson, and intimidation.” Consistent with prior studies, researchers found a strong association between low cognitive ability and violent offending. They note that “men convicted of violent crime had more than a standard deviation lower cognitive ability than those without such convictions.”

After comparing the general association between cognitive ability and violent offending to the within-pair  association, the authors did find that factors shared by brothers partially confounds the association between cognitive ability and violent offending. However, they note that “most of the association remains even after adjusting for such factors.” The authors conclude that “most of the association is not due to confounding by childhood environment”.

In summary, cognitive ability is a robust predictor of anti-social behavior (criminality, illegitimacy, welfare usage, unemployment), which is not explained by confounders such as parental SES or family background. Particularly impressive is the fact that cognitive ability remains associated with later anti-social behaviors even after controlling for earlier measures of anti-social behavior (e.g., cognitive ability predicts later self-control after controlling for prior self-control), when controlling for import risk factors for criminality (Ttofihi et al. 2016), and when comparing siblings from the same family. The fact that cognitive ability remains a powerful predictor after controlling for such a robust set of confounders is strong evidence that there is a causal relationship between cognitive ability and these anti-social outcomes. That is, the Causal Hypothesis seems to be true for anti-social outcomes

Conclusion


The studies above provide overwhelming evidence that the Causal Hypothesis is true: cognitive ability has a large causal influence on important life outcomes even after controlling for plausible confounders such as parental SES, family background, personality, risk factors for crime, etc. Now, many studies also found that the association between cognitive ability and life outcomes is somewhat reduced after controlling for these confounders, indicating that some of the association between cognitive ability and life outcomes is the result of confounding. Nevertheless, the weight of the evidence suggests that cognitive ability involves a stable set of traits that play a strong causal role in determining a person’s future academic achievement, occupational performance, socioeconomic success, and engagement in anti-social behavior.

These findings have significant political implications. These findings suggest that any plan to improve success in these outcomes for subjects with low cognitive ability may need to focus on improving the cognitive ability of the subjects of concern. Furthermore, because of the lifetime stability of cognitive ability, this suggests that interventions may need to target improving the cognitive ability of children at a very young age (perhaps even before birth). Cognitive ability is a significant factor in success for nearly all societal outcomes that we care about. If we ignore this crucial factor, we are unable to develop a working understanding of the causes of success (and failure) of individuals and groups in Western societies, and we may be unable to develop informed solutions to address inequalities in certain outcomes when these inequalities are the results of differences in cognitive ability.

In closing, I would like to note that cognitive ability is not everything. There are plenty of other factors that influence an individual’s success. For example, most of the studies cited above show that parental socioeconomic status is associated with life outcomes even after controlling for cognitive ability (although the effects of parental SES typically are not as large as the effects of cognitive ability). Additionally, there is a growing body of evidence that “non-cognitive” abilities have a tremendous impact on success. For example, many studies show that childhood self-regulation predicts adulthood success even after controlling for IQ and parental SES (Moffitt et al. 2011, pages 2694-2695; Fergusson et al. 2013, Table 3). Moreover, a number of studies suggest that personality and self-regulation may predict grades (but not standardized test scores) better than cognitive ability does (Duckworth et al. 2012, Lechner et al. 2017, Galla et al. 2019). Conscientiousness has also been shown to predict success independently of cognitive ability. For example, conscientiousness has been shown to predict occupational performance rather well (Hurtz and Donovan 2000, table 2; Dudley et al. 2006, Table 4). Also, in college, where there is significant range restriction in cognitive ability, conscientiousness predicts academic achievement about as well as intelligence does (Poropat 2008, Table 2; Richardson et al. 2012, Table 4). Furthermore, a number of meta-analyses have shown that an internal locus of control is significantly correlated with academic performance (Richardson et al. 2012) and job performance (Ng et al. 2006Judge and Bono 2001).

Leave a Reply

Your email address will not be published. Required fields are marked *