Twin studies with unmet assumptions are biased towards genetic heritability

Abstract

SummaryFor a century [1,2], studies of monozygotic and dizygotic twins have yielded estimates of trait heritability. The clever logic behind them is that while both types of twins share environments, their genetic overlap is different. Hence, larger trait correlations between monozygotic compared to dizygotic twins indicate heritability (nature), whereas similar correlations indicate shared environmental influences (nurture), and low correlations indicate shaping through non-shared environments (external influences and measurement error). While many have written on the assumptions that both types of twins share equal environments [3–5], and that parental genetics and environment are independent [6,7]; fewer have put their data where their mouth is. Here, the impacts of unmet assumptions were investigated using a generative mixture model of twin phenotypes. The results indicated that violations of the equal environments assumption yielded large overestimations of heritability and underestimations of shared environmental influences. On the other hand, when parental genetics shaped twins’ shared environments, only minor non-linear biases against heritability emerged. Finally, realistic levels of measurement error uniformly depressed estimates for genetic and shared environmental factors. In sum, twin studies are particularly susceptible to overestimation of genetic and non-shared environmental influences. This bias could explain why some traits, such as attitudes towards property taxes [8], show suspiciously high heritability without a biologically plausible mechanism. Particularly in the context of traits with convincing mechanisms of cultural transmission [9–11] and complex gene-environment interactions [6], researchers should not allow biases in twin studies to overestimate heritability.

Variation in subcortical anatomy: relating interspecies differences, heritability, and brain-behavior relationships

“…Firstly, heritability estimation, like any statistical model, relies on a simplification of reality that is subject to certain limitations. For instance, previous work has shown that gene-environment interactions lead to overestimation of heritability as they are often confounded for additive genetic effects (Dalmaijer, 2020). Secondly, parental passive genetic influence results in the marginal under-estimation of shared environmental effects and measurement error increases the estimation of non-shared environmental effects (Dalmaijer, 2020).…”

“…For instance, previous work has shown that gene-environment interactions lead to overestimation of heritability as they are often confounded for additive genetic effects (Dalmaijer, 2020). Secondly, parental passive genetic influence results in the marginal under-estimation of shared environmental effects and measurement error increases the estimation of non-shared environmental effects (Dalmaijer, 2020). Thirdly, although the purpose of our comparative neuroanatomy work was to try to better understand human brain evolution, it should be acknowledged that pairwise contrasts with only chimpanzees cannot fully take allometric scaling into account, nor can chimpanzee brain structure be assumed to perfectly represent the ancestral condition from which human neuroanatomy evolved (Staes et al ., 2019).…”

Tracing the development and lifespan change of population-level structural asymmetry in the cerebral cortex

“…Second, while GAMMs are considered an optimal modelling technique for longitudinal lifespan data and are robust to non-normal age distributions 73 , relative underrepresentation of the mid-adulthood age-range may drive trajectory inflection points around this age 11 , suggesting caution is warranted regarding interpreting mid-life inflection points as reflecting real change. Third, though the differing heritability methods applied enabled replication for SA, twin studies are prone to overestimating heritability due to unmet assumptions 79,81 , whereas SNP-based methods may not capture all phenotype-relevant genetic variance, and have their own assumptions 136,137 . Indeed, we found twin-based estimates were often substantially higher even where only nominally significant, agreeing with recent calls for caution when interpreting twin-based heritability estimates 81 .…”

“…Third, though the differing heritability methods applied enabled replication for SA, twin studies are prone to overestimating heritability due to unmet assumptions 79,81 , whereas SNP-based methods may not capture all phenotype-relevant genetic variance, and have their own assumptions 136,137 . Indeed, we found twin-based estimates were often substantially higher even where only nominally significant, agreeing with recent calls for caution when interpreting twin-based heritability estimates 81 . Fourth, we imposed a necessary cluster size limit for overlapping asymmetry effects across samples, and thus more focal asymmetries may also be informative in relation to the factors tested here (see 29 ).…”

“…Data were reformatted such that rows represented family-wise observations. As is standard, we set A to be 1 for MZ twins assumed to share 100% of their segregating genes (but see 79 ), 0.5 for DZ twins and siblings that share 50% on average, and shared environment was assumed equal for twins and non-twin siblings (but see 80,81 ). Prior to estimating heritability, for each phenotype we regressed out the effects of age and sex via linear models and computed z-scores.…”