It is straightforward to find many more examples of this trend at all levels of the educational hierarchy.<\/p>\n\n\n\n
There are typically several motivations given for these changes:<\/p>\n\n\n\n
- \n
- Intellectual ability may not be a coherent or quantifiable concept in the first place<\/li>\n\n\n\n
- Standardized exams may not capture the full range of students’ intellectual abilities<\/li>\n\n\n\n
- Socioeconomic adversity or racial discrimination may lead to certain subpopulations having depressed standardized exam scores relative to their intrinsic ability<\/li>\n\n\n\n
- Families and cultures that place high value on exam preparation may lead to certain subpopulations having overly high standardized exam scores relative to their intrinsic ability<\/li>\n<\/ul>\n\n\n\n
Typically, the disadvantaged groups (exam scores underestimate ability) are thought to be poor, Black, Hispanic, or Native American students, and the advantaged groups (exam scores overestimate ability) are thought to be Asian-Americans.<\/p>\n\n\n\n
These motivations are wrong.<\/p>\n\n\n\n
However, it is important to be clear about why they are wrong. They are not wrong because I disagree with them on an ideological basis; rather, they are wrong because they conflict with the existing scientific literature on cognitive ability, the heritability of intelligence, and standardized testing. They are empirically wrong<\/em> and run against decades’ worth of detailed studies spanning the fields of psychology, sociology, education, and modern genetics.<\/p>\n\n\n\n
In this post, I outline a clear, step-by-step argument that lays out a strong case for the pro-standardized testing viewpoint. I establish the following points:<\/p>\n\n\n\n
- \n
- Intelligence is measured by a single factor, g<\/em><\/li>\n\n\n\n
- The majority of differences in intelligence are genetic<\/li>\n\n\n\n
- Common objections to heritability estimates are invalid<\/li>\n\n\n\n
- Standardized test scores are good measures of intelligence<\/li>\n\n\n\n
- Test scores are unaffected by parental income or education<\/li>\n\n\n\n
- Test preparation has a minimal effect on test scores<\/li>\n\n\n\n
- Asian culture does not explain Asian-American exam outperformance<\/li>\n<\/ol>\n\n\n\n
In sum, I supply a succinct argument that rebuts all of the typical arguments in favor of eliminating standardized exams. Upon a comprehensive review of the literature, I find that cognitive ability is a coherent, innate, and heritable trait; that standardized exams are good measures of cognitive ability, and are largely unaffected by parental income or education; finally, that exam scores are not meaningfully affected by student preparation, motivation, educational quality, or parental pressure, and certainly not to a degree necessary to explain group differences in performance.<\/p>\n\n\n\n
The conclusion is that standardized exams are an accurate and effective measurement of innate cognitive ability. Their usage does not introduce or exacerbate preexisting societal biases or privileges, and their elimination is tantamount to an explicit policy of discrimination against higher-performing groups in favor of lower-performing groups. That is to say: anti-Asian discrimination.<\/p>\n\n\n\n
I would like to preemptively thank Cr\u00e9mieux<\/a> for a number of useful Twitter threads which I extensively consulted while compiling this post. Much of the hard labor is his; I am but a mere synthesizer of preexisting content. The same gratitude is, naturally, extended to all others whose work I have cited in this post.<\/em><\/p>\n\n\n\n
Intelligence is measured by a single factor, g<\/em><\/strong><\/p>\n\n\n\n
First, it’s important to understand how cognitive ability is structured. As early as 1927, it was observed by Spearman (in The Abilities of Man<\/em>) that all tests of cognitive abilities were highly positively correlated with each other. For example:<\/p>\n\n\n
\n![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-4-1024x783.png\")
<\/figure><\/div>\n\n\nThe common factor which explains the majority of the variance in the results for any given test is termed g<\/em>, or the g<\/em> factor, denoting “general intelligence.” The validity of this construct has been demonstrated many times over the course of the past century. For example, Johnson et al.<\/em> (2008)<\/a>, working to replicate results from prior work in 2004, continue to find support for “the existence of a unitary higher-level general intelligence construct whose measurement is not dependent on the specific abilities assessed.” The validity of modeling cognitive ability as derived from a unitary g<\/em> factor has also been confirmed in modern cohort studies such as Panizzon et al.<\/em> (2014)<\/a>. Attempts to construct models of human cognitive ability either produce multi-factor models which fail to have the same predictive ability as g<\/em> (such as Gardner’s “multiple intelligences,” which have long fallen out of favor) or, ultimately, recapitulate a hierarchical model where a unitary factor, analogous to g<\/em>, stands at the top of the hierarchy. For a more detailed review of the background literature on g<\/em>, the reader is advised to consult Stuart Ritchie’s Intelligence: All That Matters<\/a>.<\/p>\n\n\n\n
It is worth noting that the existence of g<\/em> is not obvious a priori<\/em>. For athletics, for instance, there is no intuitively apparent “a<\/em> factor” which explains the majority of the variation in all<\/em> domains of athleticism. While many sports do end up benefiting from the same traits, in certain cases, different types of athletic ability may be anticorrelated:<\/em> for instance, the specific body composition and training required to be an elite runner will typically disadvantage someone in shotput or bodybuilding. However, when it comes to cognitive ability, no analogous tradeoffs are known.<\/p>\n\n\n\n
We will later discuss the relationship between g<\/em> and exam scores. Insofar as the two are correlated, a direct causal relationship between intelligence and life outcomes (income, wealth, physical health, overall life satisfaction) follows as a direct consequence. However, the reader should note that there is a rich literature demonstrating the relevance of the abstract construct of g<\/em> to everyday life, ranging from classic work by Gottfredson (1997)<\/a> to modern replications such as those by Zisman and Ganzach (2022)<\/a>. Differences in g<\/em> are not merely statistical artifacts of purely academic interest; they reflect real differences between individuals with substantial predictive value for numerous aspects of adult life.<\/p>\n\n\n\n
Popular critiques of g<\/em>, or of the science of cognitive ability in general, surface with some regularity. The criticisms are typically flawed, typified by a lack of theoretical understanding as well as a deep ignorance of the empirical literature. While the purpose of this report is not to supply a history of the scientific debate on g<\/em>; nevertheless, because they are cited with such great frequency, they are worth mentioning in passing:<\/p>\n\n\n\n
- \n
- Gould (1981)<\/a> formulated a number of wide-ranging criticisms of classical research on cognitive ability. A number of central empirical claims made by Gould regarding a classic lead-shot study by Morton were demonstrated in Lewis et al.<\/em> (2011)<\/a> to themselves be “poorly supported or falsified.” Similarly, a reanalysis by Warne et al.<\/em> (2018)<\/a> concluded that Gould’s negative analysis of the Army Beta intelligence test “mischaracterized” critical aspects of the data under consideration. Finally, many of the broader aspects of Gould’s work (for example, his examination of the historical literature) were refuted by Bartholomew (2014)<\/a>.<\/li>\n\n\n\n
- Shalizi (2007)<\/a> (archive)<\/a> advances a number of superficially technical criticisms. For example, he claims that the Perron\u2013Frobenius theorem<\/a> guarantees the extraction of a general factor from any test battery; however, there is no guarantee, before looking at the actual correlations, that all test scores are positively correlated with each other, nor any guarantee that the first latent factor explains as much of the shared variance as does g<\/em> with cognitive batteries. Shalizi also ignores the long history of failed attempts to find assessments of cognitive ability not largely explained by g<\/em>. Comprehensives responses to Shalizi were given by by Dalliard (2013a)<\/a> (archive)<\/a> and Dalliard (2013b)<\/a> (archive)<\/a>.<\/li>\n\n\n\n
- Taleb (2019)<\/a> (archive)<\/a> supplies a very discursive and imprecise critique. The claims made are too broad-ranging and vague to properly address in this article but have been thoroughly re-analyzed by a number of public rebuttals (1<\/a>, 2<\/a>).<\/li>\n<\/ul>\n\n\n\n
Beyond their historical significance in the public discourse surrounding cognitive ability and g<\/em>, responses to poorly formulated critiques of g<\/em> themselves serve as excellent primers to the field of intelligence research, as they thoroughly address a number of intuitive or oft-cited misconceptions.<\/p>\n\n\n\n
The majority of differences in intelligence are genetic<\/strong><\/p>\n\n\n\n
So far, we have said nothing about the origin of cognitive ability. The existence of g<\/em> is compatible with fully-hereditarian models, where cognitive ability is fixed at birth, and with fully-environmental models, where all humans begin with the same “blank slate” at birth and develop differing levels of cognitive abilities due to their parental, etc. environments, and, of course, with anything in between these two extremes.<\/p>\n\n\n\n
The place where traits fall on this spectrum is represented by heritability,<\/em> defined as the amount of variance in the trait in question explained by genetic variation between individuals. The traditional gold-standard for measuring heritability ever since the advent of modern genetics has been the twin study,<\/a> which uses genetic similarities between twins reared together or by different families to cleanly separate out genetic vs. environmental influences on traits. A review conducted by Bouchard (2013)<\/a> summarized a number of twin studies to show that the heritability of g<\/em> increases over age to a terminal value of 0.8 or above in adulthood:<\/p>\n\n\n
\n![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-8-1024x697.png\")
<\/figure><\/div>\n\n\nThis is quite a significant finding. First, it suggests that in adulthood, cognitive ability is at least 80% predictable from a given person’s genetic makeup. (Importantly, we should note that this value is actually an underestimate<\/em> of the true heritability, due to noise in measurement of g<\/em> as well as the presence of assortative mating, which biases twin-study estimates of heritability downward.) Second, it shows that while cognitive abilities are more environmentally dependent in childhood, the environmental influences tend to “wash out” as people age and regress toward their “genetic ability,” roughly speaking.<\/p>\n\n\n\n
Results from other analytic methods arrive at the same results:<\/p>\n\n\n\n
- \n
- For example, a more sophisticated twin study conducted by Vinkhuysen et al.<\/em> (2012)<\/a>, which uses detailed family background to adjust for assortative mating, finds a total genetic heritability of 82% for intelligence.<\/li>\n\n\n\n
- For those who object to twin studies in general, Schwabe, Janss, and van den Berg (2017)<\/a> uses nearly one million data points from the same Dutch birth cohort with detailed pedigree information to estimate a heritability of 85% for educational achievement.<\/li>\n<\/ul>\n\n\n\n
While there exist some twin studies which yield lower heritability estimates, their results are often indicative of measurement errors or other systematic flaws. For example, Willoughby et al.<\/em> (2021)<\/a> produce a genetic heritability of 40% for ICAR-16 cognitive scores; however, the effect of parental environment is also estimated at 1% (which is much lower than almost all other empirical findings, as well as being highly implausible), suggesting that a great deal of “unmeasured heritability” is being inaccurately bundled into the 50% contribution of non-shared environment. (One might even note that studies which are obvious underestimates<\/em> such as Willoughby et al.<\/em> (2021) are actually useful as very tight lower bounds on the true values!) A magisterial overview of the history and broader landscape of twin studies of intelligence and its heritability is given in Lee (2010)<\/a> (itself a critical review of Nisbet<\/a>t (2009)<\/a>, which cherry-picks twin studies in order to inaccurately conclude that the true heritability of g<\/em> is low).<\/p>\n\n\n\n
Overall, multiple decades’ worth of studies, including both twin studies and cross-sectional analyses, convincingly demonstrate that well over 80% of individual variation in cognitive ability (g<\/em>) is accounted for by genetic variation, at least in modern, first-world, developed countries. The reader should note that this is a truly remarkable<\/em> result: it means that an individual’s genome, which is fixed at birth, can be used to predict, with relatively high accuracy, a factor (g<\/em>) which has found to affect life outcomes as far-ranging as income or even health and longevity! For instance, Wolfram (2023)<\/a> demonstrates that occupations are highly stratified with respect to intelligence:<\/p>\n\n\n
\n![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-18-1024x729.png\")
<\/figure><\/div>\n\n\nIn conjunction with the high heritability of g<\/em> estimated by multiple studies, we may conclude that one can largely determine, at the moment of conception, the professions most suitable for a given child. (There will still be a great deal of variability in specific preferences and abilities; however, the narrowing of the space of possibilities is nevertheless very dramatic.) Of course, this report is not intended to be a review of the total predictive abilities of g<\/em>, so we will not spend too much time dwelling on the literature relating measurements of g<\/em> to life outcomes; we merely bring up the existence of such work so that the reader may gain some appreciation for the practical significance of the heritability of human intelligence and, if so they so desire, explore such tangents at their leisure.<\/p>\n\n\n\n
Common objections to heritability estimates are invalid<\/strong><\/p>\n\n\n\n
(The first-time reader is advised to skip this section if they are relatively unfamiliar with the topic. It is placed here for narrative consistency, but is slightly more nitpicky in nature.)<\/em><\/p>\n\n\n\n
Three common objections exist to the above heritability estimates. The first is that shared prenatal environment inflates twin-study estimates of heritability:<\/p>\n\n\n
\n![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-9-1024x437.png\")
<\/figure><\/div>\n\n\nThis rejoinder is of course simply invalid for estimates of heritability which are not derived from twin studies, such as the estimate from Schwabe, Janss, and van den Berg (2017)<\/a>. However, shared prenatal environment in general<\/a> is also not known to have any strong effect on cognitive ability. For example, a comprehensive review by van Beijsterveldt et al.<\/em> (2016)<\/a> finds that “the influence on the MZ twin correlation of the intrauterine prenatal environment, as measured by sharing a chorion type, is small and limited to a few phenotypes [implying] that the assumption of equal prenatal environment of mono- and DC MZ twins, which characterizes the classical twin design, is largely tenable.” Indeed, if womb effects increased similarity between mother and child, we would expect this effect to show up in studies of adoptees; however, per Loehlin et al.<\/em> (2021)<\/a>, adoptees’ measured IQs (in the Texas and Colorado Adoption Projects) were correlated in similar degrees to both maternal and paternal IQ at ages 6 and 16. Overall, the claim that shared prenatal environment leads to overestimated heritability is a testable empirical claim,<\/em> and the results of such tests show conclusively that it is a false claim.<\/em><\/p>\n\n\n\n
Second, there is the question of “missing heritability.” For example, GCTA estimates of the heritability of intelligence seem to find much lower heritability values compared to those reported above. Similarly, polygenic scores (PGS) constructed to predict either intelligence or educational attainment only manage to explain a very small fraction of the variance in such traits.<\/p>\n\n\n
\n![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-10-1024x418.png\")
<\/figure><\/div>\n\n\nThe explanation, of course, is that both of these estimates (GCTA and PGS) are really weak lower bounds<\/em> for heritability. (While PGS estimates can be cleanly interpreted as lower bounds, GCTA estimates are subject to a variety of biases themselves, and are not necessarily straightforwardly understandable as strict lower bounds.) The first only estimates the variability attributable to common variants, while the latter relies on the explicit identification of genetic variants which predict cognitive ability.<\/p>\n\n\n\n
We can reason about the applicability of GCTA estimates through analogy to other human traits. For example, Yang et al.<\/em> (2011)<\/a> first provided a GCTA estimate of the heritability of human height at ~0.5. A decade later, Yengo et al.<\/em> (2022)<\/a> supplied, again, an estimate of ~0.5 heritability based on common variants. At the same time, however, whole genome<\/em> estimates of the heritability of human height are able to recover the same values (~0.7) given by pedigree-based heritability, such as in Wainschtein et al.<\/em> (2019)<\/a>. In the specific case of height, estimates of heritability based on common variants, while stable over the course of years, were clearly underestimates of the actual heritability, which includes contributions from a large number of rare variants. Similarly, with cognitive ability, the “heritability gap” is likely to be due to the contribution of rare variants not assessed by the genotyping arrays used for GCTA-based estimates. (Some of the difference may also result from the usage of primarily European data for imputation, which does not generalize perfectly to non-European populations.)<\/p>\n\n\n\n
Gusev further claims that studies such as Howe et al.<\/em> (2022)<\/a> or Young et al.<\/em> (2022)<\/a> demonstrate that estimates of the heritability of g<\/em> (either via traditional GCTA or via the construction of polygenic scores) are in fact upper<\/em> bounds on the heritability of g<\/em>. This argument is prima facie<\/em> absurd: even accepting these studies’ findings that traditional heritability estimates for g<\/em> are overestimated due to assortative mating and shared environment, the presence of an upward bias does not turn an estimate into an upper bound!<\/em> The first of the two papers, for example, estimates a much lower heritability for cognitive ability (n<\/em> = 27,638) than for height (n<\/em> = 149,174), and also finds that the heritability of cognitive ability is “deflated” by a greater degree than height after indirect genetic effects are taken into account. However, especially<\/em> in light of the drastic difference in sample size, as well as the inherently higher measurement error in simple tests of cognition compared to measurements of height, this simply does not logically imply that the heritability estimates for g<\/em> are upper bounds. It merely implies, exactly as it claims to find<\/em>, that the estimates are potentially biased upward due to biometric interactions; to conflate that with the unadjusted estimates forming an upper bound is, pure and simple, an error in logical reasoning. The identification of such interactions is also fully consistent with the results of twin studies; it would be surprising if they were not<\/em> found within cross-sectional genomic data, and their existence does not change the fact that as cohort data improves, estimates of heritability continue to march upward.<\/p>\n\n\n\n
As for PGS estimates of heritability, the genetic architecture of intelligence, and other complex human traits, is typically such that thousands, if not tens of thousands, of genetic variants affect the final trait value, as described in Hsu (2014).<\/a> As data quality improves, the PGS-derived heritability is also expected to increase, much as it has done for height and a number of other complex phenotypic traits throughout the 2010s; moreover, PGS scores for both cognitive ability and educational attainment continue to improve with the collection of greater quantities of biobank data, with little sign of a “plateau” indicating that we have reached anything close to the best possible PGS. Even setting aside the lack of sufficient data to train a high-quality polygenic model for g<\/em>, it is purely nonsensical<\/em> to claim, ever,<\/em> that a polygenic score forms anything but a lower bound on the true heritability of the trait it is predicting; it follows from the definition of what a polygenic score is in the first place<\/em> that if you can explicitly cognitive ability with a PGS to some degree d<\/em>, the true heritability must<\/em> be greater than d<\/em>. Overall, vague allusions to GCTA\/PGS estimates of heritability, and especially to the deflation of heritability estimates as a rationale for interpreting them upper (rather than just lower) bounds, are logically incoherent at best and disingenuous at worst.<\/p>\n\n\n\n
The final objection is that heritability is depressed by adverse environmental circumstances. For example, if all members of a population are severely malnourished and mistreated, it is likely that they will all have extremely low cognitive abilities in adulthood regardless of genetic makeup. The contention is therefore that if we were to examine less privileged groups, which face adverse circumstances in childhood, we would find lower heritability values. Therefore, the values of >80% arrived at by twin studies in predominantly white subpopulations of developed nations may effectively overstate the heritability of intelligence in other groups.<\/p>\n\n\n\n
More generally, this is known as the Scarr\u2013Rowe hypothesis<\/a>: if true, groups which face some sort of adversity in early life (poverty, discrimination, and so on) should result in lower heritability for cognitive ability. Alternatively, if we observe that heritability differs across groups, that ought to constitute evidence that the group with lower heritability faced greater overall adversity in early childhood. The extent to which measured heritability actually differs across different subsections of the population is, of course, an empirically answerable question.<\/p>\n\n\n\n
While this hypothesis is likely true to some nonzero extent in principle<\/em>, we should note that even among a sub-Saharan population mired in extreme poverty, Hur and Bates (2019)<\/a> were able to estimate a heritability of ~35% for intelligence. It is difficult to imagine that any reasonable subpopulation of a first-world country faces circumstances as adverse as those of poor sub-Saharan teenagers; correspondingly, 35% ought to be taken as a fairly hard lower bound for how far environmental deprivation can reduce the heritability of intelligence. (However, one may wish to take these results with a grain of salt, given that no replication studies in comparable populations presently exist.) In de Rooij et al.<\/em> (2010)<\/a>, the authors found that even a 5-month-long famine at the end of World War II was barely associated with reduced cognitive performance in later life, and only in one of four tests of cognitive ability they tried (nonsignificant results for the others). Finally, Fuerst (2014)<\/a> (later systematically confirmed by Pesta et al.<\/em> (2020)<\/a> and Pesta et al.<\/em> (2023)<\/a>) found that heritability of cognitive ability is essentially the same across different ethnic subgroups in the United States. Since socioeconomic status and (presumed) exposure to racial discrimination are highly variable across different American ethnicities, this result suggests that neither of those two factors significantly affects the heritability of cognitive ability, at least within the milieu of modern Western society.<\/p>\n\n\n\n
One particular study, Turkheimer et al.<\/em> (2003)<\/a>, is often cited in support of a lower heritability of g<\/em> among low-income substrates of American society. However, this is a problematic piece of evidence. First, heritability estimates were only provided for 7-year-olds; because we know that environmental effects asymptotically “wash away” with age, this is highly unreflective of heritability estimates for late adolescence and early adulthood, which is what we are most interested in. Second, larger follow-up studies on older cohorts such as Rask-Andersen et al.<\/em> (2021)<\/a> often fail to reproduce a clear association between lower socioeconomic status and lower heritability of g<\/em>. The most likely explanation is that Turkheimer et al.<\/em> (2003) is either purely artefactual, a reflection of transiently depressed heritability in early childhood, or a combination of the two. (The reader may also be interested to know that a re-analysis of the same cohort data used by Turkheimer et al.<\/em>, namely by Beaver et al.<\/em> (2013)<\/a>, found no heritability differences by race, suggesting that if one accepted wholesale the results of Turkheimer et al.<\/em>, one ought to also consequently accept the conclusion that even among economically disprivileged families, where one might expect to find the most severe levels of racial discrimination and prejudice, racial ancestry does not appear to have any discernible differential impact upon the cognitive abilities of different ethnic subgroups.)<\/p>\n\n\n\n
Overall, we find that objections to a high heritability estimate of g<\/em> are, generally, lacking in substance. In some cases, they are simply false claims; in other cases, they seem to reflect a fundamental misunderstanding of the methods under question and the results cited. The one trait that these criticisms do share in common, however, is that they all have the flavor of being “isolated demands for rigor.”<\/a> Seldom are claims for the heritability of other complex human traits, such as height, disease risk, or psychological attributes questioned to such depth; at any given point in time, given sufficient effort and motivation,<\/em> it would likely have been possible to identify both theoretical and empirical objections to the heritability of height. Yet it is now well accepted that height is well over 80% heritable in a modern Western country with adequate nutrition\u2014an estimate which has steadily ratcheted upward over the course of the 21st century.<\/p>\n\n\n\n
One final note: critics of heritability estimates of g<\/em> often claim that “hereditarians” have erred in their predictions (of, say, the eventual potential to use polygenic scores to estimate g<\/em> to a precision equal to its broad-sense heritability). Such claims are then used to support the notion that cognitive ability is too complex or confounded of a trait, and that further research into “hereditarian” directions are therefore fruitless. This argument is self-contradictory: knowing that we have an order of magnitude less data for cognitive ability than we do for other complex traits like height or cardiovascular disease risk, and that existing measurements of g<\/em> in cohort data are often noisy and range-bound, the fact that we have failed to generate high-quality genetic predictors of g<\/em> (or educational attainment and so on) are evidence that we should collect more, and better, datasets, not<\/em> evidence that we should stop<\/em> collecting and analyzing such data, or that doing so is in any sense fruitless or doomed to failure.<\/p>\n\n\n\n
Standardized test scores are good measures of intelligence<\/strong><\/p>\n\n\n\n
We understand now that intelligence is summarized by a single factor, g<\/em>, which is highly heritable (most variation is causally explained on the genetic level). However, we have yet to bring in standardized testing into the discussion. Are SAT scores actually associated with g<\/em>?<\/p>\n\n\n\n
The answer is unequivocally yes<\/em>. Frey and Detterman (2004)<\/a> found that measures of g<\/em> derived from the Armed Services Vocational Aptitude Battery (ASVAB) were correlated with SAT score at r<\/em> = 0.86 among samples in the National Longitudinal Survey of Youth (NLSY) 1979. A further reanalysis by Cr\u00e9mieux<\/a> not only reproduced the same result using the NLSY ’79 dataset but in fact revealed an even stronger and cleaner relationship in NLSY ’97. (The changes are likely attributable to less range-restriction on the ’97 ASVAB and the increasing commonality of taking the SAT, meaning that the population of SAT test-takers is less selected for high-g<\/em> bias).<\/p>\n\n\n\n
![\"\"](\"https:\/\/milkyeggs.com\/wp-content\/uploads\/2023\/05\/image-7-1024x397.png\")
<\/figure>\n\n\n\nA brief aside: Because cognitive ability is predominantly measurable through a single factor, known as g<\/em>, it is a priori<\/em> reasonable to use a small number of standardized exam scores (such as SAT Critical Reading and SAT Mathematics subscores) to quantify a given student’s cognitive abilities. Were this not the case, then it might be reasonable to suggest that we need a more holistic approach, where we measure a large cornucopia of different abilities. However, because all cognitive tests seem to be essentially downstream of individual variation in g<\/em>, it suffices to have one or two metrics of intellectual ability. Additionally, because practical constraints in examination length mean that there is a tradeoff between the “quality” of the given subtest (precision, length, range, versus the number of subtests given, the existence of g<\/em> means that the information-maximizing choice is set at a fairly low number of in-depth subtests.<\/p>\n\n\n\n
The correspondence between academic achievement and cognitive ability has also been explicitly tested numerous times. Notably, Kaufman et al.<\/em> (2012)<\/a> studies the latent factors extractable from batteries of academic and cognitive tests, finding that so-called “academic g<\/em>” and “cognitive g<\/em>” correlate at r<\/em> = 0.83. (In actuality, there is only one g<\/em>, cognitive g<\/em>, so the terminology here is a slight abuse of notation.) A number of other cohorts have also been studied, although it is important to note that when the test subjects are too young, the heritability of their cognitive abilities is still relatively low (environmental effects have yet to “wash away” and genetic potential has yet to “manifest”), likely introducing a great degree of noise into the calculated correlations. That is precisely one advantage of the analysis performed by Kaufman et al.<\/em>, who explicitly calculate the variation of the academic\/cognitive g<\/em> correlation over time (and find that it increases, in fact, to a value >0.9).<\/p>\n\n\n\n
Test scores are unaffected by parental income or education<\/strong><\/p>\n\n\n\n
The reader might object that just because g<\/em> is correlated with SAT scores, the possibility remains that both are simply downstream of socioeconomic privilege (as reflected by parental income, wealth, or educational status). It is well understood that, cross-sectionally, higher socioeconomic status (SES) is strongly correlated with higher test scores on both standardized exams like the SAT and explicit IQ tests. Of course, just looking cross-sectionally, it is not obvious that which way the causation lies, i.e.,<\/em> whether children do better because their parents are rich, thereby leading to a favorable educational environment with tutors and leisure time, or because their parents are smart first and rich second, leading to both higher incomes and genetic transmission of intelligence.<\/p>\n\n\n\n
Given the tight correlation of r<\/em> = 0.85 or higher between g<\/em> and SAT scores, as well as the known heritability of cognitive ability at >80%, it seems relatively implausible for differences in SAT scores to be largely causally downstream of differences in parental SES. However, to remove all doubt, we may consult the results of a number of adoption studies, all of which show that an adopted child’s score on a test of intelligence is related to the educational attainment or cognitive ability of their biological parents<\/em> but not<\/em> of their adoptive parents:<\/p>\n\n\n\n
- \n
- Skodak and Skeels (1949).<\/a> Adoptee IQ is correlated >0.4 with biological parents’ education levels but <0.1 with adoptive parents’ education levels. (further discussion)<\/a><\/li>\n\n\n\n
- McGue et al.<\/em> (2007).<\/a> Intelligence, as well as delinquency and drug use, are not significantly associated with adoptive parents’ socioeconomic status. However, all three outcomes are significantly correlated with parents’ SES for biological children. (further discussion)<\/a><\/li>\n\n\n\n
- Loehlin et al.<\/em> (2021).<\/a> Among both the Texas and Colorado Adoption Projects, adoptee child IQ was significantly associated with their biological mother’s IQ at both ages 6 and 16, but was not associated with adoptive family SES at either age. (further discussion)<\/a><\/li>\n\n\n\n
- Odenstad (2008).<\/a> In a Swedish cohort, higher parental education was significantly associated with the IQ of biological children, but completely unrelated with the IQ of international adoptees. (further discussion)<\/a><\/li>\n<\/ul>\n\n\n\n
These results suggest, in fact, that effectively none<\/em> of the individual variation in IQ or SAT scores is causally downstream of either family environment or parental income and wealth. Otherwise, we would see some relationship between the cognitive performance of adoptees and the educational levels or incomes of their adoptive parents. The fact that we find no such relationship across multiple studies and different cohorts is therefore quite striking.<\/p>\n\n\n\n
Finally, if socioeconomic status was causally upstream of test scores in a manner that was independent of parental intelligence, we would expect to see these differences manifest in the predictive validity<\/em> of exam scores. For example, if SAT scores were purely determined by SES, we would expect that controlling for SES would attenuate the correlation between SAT scores and university grades to zero. However:<\/p>\n\n\n\n
- \n
- Sackett et al.<\/em> (2009)<\/a>, however, show that this is not true; in fact, controlling for SES only reduces the correlation between SAT scores and first-year grades from r<\/em> = 0.47 to r<\/em> = 0.44, a very modest effect size.<\/li>\n\n\n\n
- Work by Higdem et al.<\/em> (2016)<\/a> further demonstrated that even after subdividing by ethnic group or sex, \u201cthe test\u2013grade relationship is only slightly diminished when controlling for SES [and] SES is a substantially less powerful predictor of academic performance than both SAT and HSGPA.” Overall, the near-complete lack of SES to provide incremental predictive power for first-year university performance on top of SAT scores is a clear refutation of the idea that SES is a significant contributor to students’ performance on standardized exams, both in aggregate as well as within any specific racial subgroup or sex.<\/li>\n\n\n\n
- An extraordinarily large (
- Work by Higdem et al.<\/em> (2016)<\/a> further demonstrated that even after subdividing by ethnic group or sex, \u201cthe test\u2013grade relationship is only slightly diminished when controlling for SES [and] SES is a substantially less powerful predictor of academic performance than both SAT and HSGPA.” Overall, the near-complete lack of SES to provide incremental predictive power for first-year university performance on top of SAT scores is a clear refutation of the idea that SES is a significant contributor to students’ performance on standardized exams, both in aggregate as well as within any specific racial subgroup or sex.<\/li>\n\n\n\n
- McGue et al.<\/em> (2007).<\/a> Intelligence, as well as delinquency and drug use, are not significantly associated with adoptive parents’ socioeconomic status. However, all three outcomes are significantly correlated with parents’ SES for biological children. (further discussion)<\/a><\/li>\n\n\n\n
- For those who object to twin studies in general, Schwabe, Janss, and van den Berg (2017)<\/a> uses nearly one million data points from the same Dutch birth cohort with detailed pedigree information to estimate a heritability of 85% for educational achievement.<\/li>\n<\/ul>\n\n\n\n
- Shalizi (2007)<\/a> (archive)<\/a> advances a number of superficially technical criticisms. For example, he claims that the Perron\u2013Frobenius theorem<\/a> guarantees the extraction of a general factor from any test battery; however, there is no guarantee, before looking at the actual correlations, that all test scores are positively correlated with each other, nor any guarantee that the first latent factor explains as much of the shared variance as does g<\/em> with cognitive batteries. Shalizi also ignores the long history of failed attempts to find assessments of cognitive ability not largely explained by g<\/em>. Comprehensives responses to Shalizi were given by by Dalliard (2013a)<\/a> (archive)<\/a> and Dalliard (2013b)<\/a> (archive)<\/a>.<\/li>\n\n\n\n
- Intelligence is measured by a single factor, g<\/em><\/li>\n\n\n\n