We still do not understand family resemblance

Edward Ruscha. Blood Strip. 1973

Why do kids look and act like their parents? One reason is that parents pass down their genes to their kids. We understand genetic transmission very well, so if genes are important, we might expect simple models of genetic transmission to provide good explanations of family resemblance. But parents also pass down their environments, and this “cultural” transmission can mimic genetic relationships, inflating the estimates from simple genetic models. An interesting new study by Eftedal et al. (2025) sought to disentangle these influences using a massive dataset and some cleverly chosen family structures.

The study relied on a registry of every Norwegian student that took a standardized test from 2007-2019, totaling nearly 1 million students and largely free of selection bias. From this data, they extract a variety of different relatedness classes and investigate the patterns of family resemblance in test performance. They fit a “Fisherian” model (based on Ronald Fisher’s seminal 1918 derivations) where phenotypic correlation is a function of additive transmission (h2) and assortative mating (m), such that the correlation for n-th degree relatives can be predicted from these two parameters alone:

\(\rho_n = h^2 \left( \frac{1 + m}{2} \right)^n\)

As has been noted in prior work (and these authors acknowledge¹), the h2 parameter here is not really “heritability” in the direct genetic sense we may expect, because it can also capture components of resemblance that happen environmentally (aka “shared environments”, “cultural” transmission from parents, “dynastic” transmission from relatives). But this study does something very clever to address the identifiability issue by also including individuals who share environments but not genes: relatives by marriage but not by kinship (step siblings/cousins in-law), for whom resemblance is only driven by the assortative mating of their parents (m); and relatives through adoption but not kinship, for whom resemblance is expected to be zero. Then they fit h2 and m to the data using (1) only biological relatives or (2) both biological and spousal relatives.

The big result is that no single Fisherian model provides a good fit to the data². This is summarized in a single figure below, showing the deviation between the observed resemblances and the fit from the two different models (re-generated from Table 1 in the paper). Neither model fits well, implying that the heritability parameters cannot generalize across all of the observed relatedness classes.

This figure captures several additional interesting findings:

The resemblance of twins cannot be reconciled with any model.

The biggest data/model deviation was the monozygotic (MZ) twin correlation, which was significantly higher than expected from the model fit only with biological relatives, and much higher than the model fit with in-laws. By extension, the differences in correlations seen between MZ and DZ twins, the workhorse of the Classic Twin Design, do not generalize to the correlations observed for other biological relationships. This observation that monozygotic twins often exhibit higher than expected correlations continues to be a fundamental mystery, with two plausible explanations: gene-gene/gene-environment interactions inflating their resemblance due to sharing 100% of their genes; or unequal environmental influences (e.g. identical twin mimicking and convergence) biasing their resemblance due to the artificial sharing of environments.

And for a more detailed discussion of gene-gene interaction effects in twin studies and beyond, see this recent post:

Beneath the surface of the sum

The resemblance of adoptees cannot be reconciled with any model.

The other clear outlier was the correlation of adopted relatives, which is expected to be zero under a Fisherian model but is clearly and significantly not zero. Interestingly, using the CTD to estimate heritability under random mating produced a value for the shared environment that fits the adopted siblings fairly well, but does not fit the other biological relatives. On the other hand, accounting for assortative mating could fit most of the biological resemblances but reduced the estimate of the shared environment to zero and thus no longer fit the adopted and in-law resemblances. MZ twins, in-laws, and adopted siblings all exhibit higher correlations than expected, so accounting for one pushes the model away from the other and vice versa³.

The resemblance among adopted relatives appears to be “dynastic”.

For the adopted relatives, which are similar only due to environmental influences, we can ask if their resemblance matches a simple “cultural transmission” model where offspring phenotypes are only directly influenced by the average phenotype of their parents and randomness. To reproduce the adopted sibling correlation of 0.15, one would need a cultural transmission effect of sqrt(0.15) or 0.39. However, extending a cultural transmission effect of 0.39 to the next generation (and adding realistic assortative mating) would, in turn, produce an adopted 1st cousin correlation of ~0.014, much lower than the actual adopted 1st cousin correlation of 0.045. If we flip it around and try to reproduce the adopted 1st cousin correlation, we would need a cultural transmission effect of 0.51. But this would produce a sibling correlation of 0.51^2 = 0.26, which is significantly higher that what was observed for adopted siblings. To the extent that cultural transmission drives the adopted resemblance, it must therefore do so through both direct and indirect means. This aligns with the theory of “dynastic effects” recently observed using genetic data, where offspring phenotypes were as or more significantly correlated with the genetic scores (and presumably phenotypic influences) of extended relatives than the genetic scores of their own parents.

One final methodological quirk is that the adoptees can only have international origins in order to be identified in the registry. Given the unique language/assimilation barriers and pre-adoption environments for non-native adoptions (in addition to the fact that adoption is itself an unusual environmental shock), even these correlations are likely to be underestimates of dynastic effects.

In-law relationships are critical to spotting model misspecification.

A key methodological takeaway from this work is that in-law relationships are critical to disentangling gene-environment masking. When the model was estimated only with biological data, it appeared to provide a generally good fit, with some deviation for twins and a handful of sex-specific relationships. From that analysis alone, one might easily discard the few deviations as outliers and conclude that resemblance is largely explained by a simple genetic model. In fact, this is more or less what happened in the recent work of Clark (2023) PNAS, which used a large biological pedigree reconstructed from rare surnames. It is only when relationships through in-laws are brought in that the genetic model was revealed to be significantly misspecified, with environments apparently mimicking genetic transmission. In-law relationships, of course, only share some of their environment: the component that is captured by assortative mating and the component after marriage. It therefore remains unknown how much lower the “h2” estimate could go if full environmental sharing was modeled.

Family resemblance in molecular data

While the current study looked at correlations for specific relatedness pairs, several prior studies investigated the components of phenotypic resemblance across all pairs of individuals. The way to think about this is to treat the phenotypic similarity (i.e. covariance) as the sum of genotypic similarity, pedigree/kinship similarity, shared environment, and, finally, a random environmental/residual term. The weights that correspond to each of these terms can tell us the relative contribution of different sources to phenotypic resemblance.

Zaitlen et al. (2013) was one of the first studies to show that pedigree relationships could be estimated directly from genetic data simply by setting weakly genetically related pairs to zero in the kinship matrix. This enabled analyses of complex relationships without having to laboriously reconstruct family trees. The authors also pointed out the fundamental limitation we saw above: if the shared environment component is not known, the kinship component will suck up any correlated explanatory factors and treat them as genes. Thus, kinship heritability should not be interpreted as direct genetic variation, but rather as the combination of genetic and environmental sharing⁴. Analyzing 23 representative traits, the results broadly mirrored those of Eftedal et al. Yet again, heritability estimates from twins did not generalize to (genetically inferred) kinship-based estimates and were consistently inflated⁵. To diagnose the source of inflation, the authors turned to relatedness classes and re-estimated heritability using classic quantitative models.

Heritability estimates (and confidence intervals) from different pairs of relationships. Estimates reproduced from Figure 1 of Zaitlen et al. (2013).

Yet again, the estimates varied significantly depending on the relationship that was used for estimation, ranging from 0.20 to 0.35 (averaged across 17 traits). Notably, the lowest kinship heritability estimates were from the cross-generational comparisons — avuncular and grandparental — that are less likely to share environments. The authors concluded that such patterns were not consistent with genetic dominance/epistasis alone and were likely to be driven by the shared environment⁶.

So what’s going on?

The main takeaway is that behavioral/status traits are complicated and we still do not have good unified models of their family resemblance. The high correlation for MZ twins points to interactions. The high correlation of adopted relatives points to direct and dynastic environmental influences. And the fact that a Fisherian model with biological relatives fit fairly well but became significantly misspecified with the inclusion of in-laws points to an environmental influence that, at least in part, mimics genetic relatedness. This latter point (which I’ve alluded to several times) is important, as it suggests that when we see similarities in families and ascribe them to genes we are at least to some extent misinterpreting the causal processes at play. Ironically, the authors are mildly guilty of this themselves: they argue that their model provides a challenge to the claim that “genetic factors have only a very limited role” in educational outcomes. But the reality is that no genetic model actually fit their data and so no clear conclusion about genetic factors can be drawn. Perhaps if we understand interactions the estimated heritability will go back up, or (as Zaitlen et al. hypothesized) if we understand environments the estimated heritability will go even further down.

More broadly, a common argument in favor of heritability estimates from classic twin/pedigree/adoption studies is that they converge on similar results (for example, that was one of the conclusions in the recent Astral Codex Ten review of “missing heritability”⁷). This is usually stated more as intuition than as fact, because such studies are mostly conducted in different cohorts, using different phenotypes, and employing different adjustments for confounders. I’ve previously written about how twin studies do not agree on heritability estimates even internally, driven by the assumptions they make about assortative mating, shared environment, and interactions. Now this study, in a single massive cohort with consistent modeling, shows that the disagreement persists for extended pedigree and adoption data as well.

Over a decade ago, Zaitlen et al. concluded that better models of family-environmental influences were needed to properly interpret pedigree-based heritability estimates⁸. It looks like we are not there yet.

1. “Plausibly, genetic and environmental effects can “mimic” each other when looking only at biological relatives. Genetic effects are correlated most strongly between close relatives, and their correlation drops off at a constant rate as relatives become more distant: environmental effects could very well be correlated between relatives in a similar pattern, causing Fisherian models to interpret them as genetic.” ~ Eftedal et al. (2025)

2. “We investigated the hypothesis that this family resemblance can be fully explained by additive genetic effects and assortative mating, through fitting models based on the work of Fisher. Our conclusion is that these factors indeed appear to be important, but that a complete model would need other sources of family resemblance as well. Environmental effects appear necessary to fully account for correlations between adoptive relatives, between relatives-in-law, and between maternal relatives. Additionally, the high correlations we see between monozygotic twins are suggestive of nonadditive genetic effects and/or gene–environment interplay.” ~ Eftedal et al. (2025)

3. “The predicted gap between rMZ and rDZ would then narrow if genetic effects are substituted for shared-environmental effects. Above, we argued that shared-environmental effects are necessary to appropriately model relatives-in-law and adoptive relatives. The rMZ–rDZ gap is too narrow even in our model with no shared-environmental effects at all, however, so every such step toward making the model fit better with relatives-in-law and adoptive relatives would exacerbate this problem.” ~ Eftedal et al. (2025)

4. “There are two major challenges in comparing [kinship h2] and [genotype h2] to quantify missing heritability. First, there is the potential for inflation of estimates based on closely related individuals such as MZ/DZ twins. It is well known that epistatic interactions can inflate heritability estimates in studies of related individuals. … Other factors that could also lead to inflated estimates of using closely related pairs of individuals include dominance and shared environment.” ~ Zaitlen et al. (2013)

5. “We find, for all of the quantitative phenotypes, that our estimates of are smaller than those from the literature that were based on MZ/DZ twins. Our results indicate that previous estimates were inflated by the impact of epistasis or shared environment.” ~ Zaitlen et al. (2013)

6. “The differences in heritability estimate between classes of relationship are consistent with a shared-environment only effect on phenotypic correlation, but not with a dominance only or epistasis only effect on phenotypic correlation.” ~ Zaitlen et al. (2013)

7. “I think the twin / pedigree / adoption estimates are mostly right. They are strong designs, their assumptions are well-validated, and they all converge on similar results.” ~ Scott Alexander / ACX

8. “A standard way to quantify the contribution of environmental effects is to fit an ACE model. However, a complexity with this approach is that it is unclear which relative classes should be modeled as sharing a common environment. For example, do parent/child pairs have the same environmental sharing as siblings? We believe this merits further investigation, although it is outside the scope of our current work.” ~ Zaitlen et al. (2013)