The question of missing heritability has long haunted behavioral geneticists. After the failure of the candidate gene era (Arango 2017; Border et. al 2019; Chabris et. al 2012; Charney 2012; Houlihan et. al 2009), the search for the missing heritability has been ongoing (Kim et. al 2017). Given that the contribution of common variants to the variation in most traits has just about maxed out (following better ascertainment and phenotyping), people have been searching for the remaining sources of the variance (Bourrat 2017; Koch 2014; Slatkin 2009; Sverdlov 2016; van IJzendoorn et. al 2011; Zuk et. al 2012). My humble task here today is to suggest that the ‘missing heritability’ may not be so… ‘missing’, but rather non-existent.
The contribution of common variants has an upper bound. Lee et. al (2018) estimated a SNP of about 12-13% , while Holland et. al (2019) employed a new method to infer that again, about 12-13% of the variance in IQ and educational attainment is associated with common genetic variants with MAF less than 0.002 (0.2%). Evans et. al (2018) got similar results, estimating an of fluid intelligence using the least biased method (LDSC) .
It also seems rare non-synonymous exonic variants don’t contribute to intelligence differences either  (Marioni et. al 2014). Spain et. al (2016) also reports that there are no rare alleles / exonic variants associated with extremely high intelligence or in the normal range  (and it seems that extremely high intelligence is genetically continuous with intelligence in the normal range (Zabenah 2016)). Known intellectual disability genes have also been found to not associate with intelligence in the normal range (Franic et. al 2015; Hill et. al 2016). Genes do also not materialize in the human leukocyte antigen region (Zabenah et. al 2017). De novo mutations are also unlikely to contribute to variation in intelligence in the normal range (Arslan et. al 2014; D’Onofrio et. al 2014, eAppendix 3, figure 2; McGrath et. al 2013; Myrskylä et. al 2013) . Copy number variants are also unlikely to play a role in intelligence (Abdellaoui et. al 2015; MacLeod et. al 2012; McRae et. al 2013).
There have not very many large-scale studies on mitochondrial DNA since the candidate gene QTL days (Moises et. al 1998; Petrill et. al 1997), though the only study I could find found no associations (Byrne et. al 2009).
This leaves one major category for additive variance left: rare variants . I have computed the share of variance explained by rare variants if the estimates for height are correct in this thread, indicating only 4.3-5% of the variance can be expected to be explained by the population stratification biased estimates of rare variant contribution .
|Category||Tenable||(Dis)confirmation||Contribution So Far||Extrapolated Contribution|
|Common variants||L||Lee et. al 2018||1-2%||13-14%|
|Common exonic variants||L||Savage et. al (2018);||“1.7%” ||~1.1%|
|Rare exonic variants||U||Marioni et. al 2014; Spain et. al 2016||0%||0-1%|
|Overall||L||Shadrin et. al (2019)||0%||~1.7%|
|Rare exonic variants||N||Marioni et. al 2014; Spain et. al (2016)||0%||0-1%|
|Other rare variants||U||Akiyama et. al 2019||0%||5-6%|
|De novo mutations||N||Arslan et. al 2014; D’Onofrio et. al 2014; McGrath et. al 2013; Myrskylä et. al 2013||0%||0%|
|Copy number variants||N||Abdellaoui et. al 2015; Bagshaw et. al 2013; Kirkpatrick et. al 2014; MacLeod et. al 2012; McRae et. al 2013||0%||0-2%|
|Other structural variants||?|
|Mitochondrial DNA||U||Byrne et. al 2009||0%||0-1%|
|Overall Heritability||N/A||See Young et. al (2018)||~2%||19.7-26.7%|
Given the facts about existing known genomic contribution to IQ phenotypes, and the population genetic facts about the genome, it is unlikely that the rest of the missing heritability will ever be recovered. The statistical issues involved in modeling genetic contribution (Kumar et. al 2015), and increasing sample sizes (Calude & Longo 2017), it seems very likely that the missing heritability was never missing, just misattributed.
 We should note two things. The first is that their estimates are definitely upwardly biased by population stratification (Dandine-Roulland et. al 2016). Secondly, the associations with intelligence are also likely to not be the true causal variants, but merely in linkage disequilibrium with the actual causal variants (Gong et. al 2019).
 However, they used a method of binning SNPs into MAF strata that is known to bias estimates (Hou et. al 2019).
 It also seems likely that if rare variants are going to significantly contribute to intelligence, they will be found relatively quickly. Simulations have demonstrated that the power for genetic associations with rare variants decays rapidly as the proportion of heritability explained by rare variants decreases (Bandyopadhyay et. al 2017). As such, the longer we go without finding rare variants, we should update our priors to smaller and smaller rare variant contribution to heritability. There are other reasons why rare variants are a priori unlikely, like the partitioning of heritability into strata (Gazal et. al 2018).
 Teng et. al (2018) reports some nominal associations, but they disappear with the use of different cognitive outcomes and different age cohorts, indicating a likely spurious correlation. Additionally, Ganna et. al (2016) has reported some associations, but I have yet to look deeply into the methodology.
 See Gisbon (2012) for a summary of arguments/for against rare variants. We should note that in the case of intelligence, there is likely selection for intelligence in the long term (Srinivasan et. al 2018), so #1 is irrelevant (as is #3’s claim also about diseases), and #4, as we observed above, is not the case for intelligence (and the other claims about CNVs often do not replicate [Hehir-Kwa et. al 2013]). That leaves #2 (which has limited relevance to IQ, though it should be noted that the fitness effects of SNPs is still controversial) and #5 (which is merely consistent with rare variants, and not positive evidence for rare variant contribution).
 Calculated from the supplementary materials of Savage et. al (2018). Note that per Savage et. al (2019), only 1.37% of associated variants are exonic, consistent with Smith (2019)‘s estimates. However, since Savage et. al (2018) did not include proper controls for population stratification or cryptic relatedness, I have downwardly adjusted the estimate in accordance with the Shardin et. al (2019) estimates.
 Note that the categories above are not mutually exclusive, so simply adding up the final column with overcount.