The problem of missing heritability is simple to state and stubborn to close. Studies of twins and families say that a trait such as adult height is roughly 80 percent heritable, yet for over a decade the genetic variants that genome-wide association studies (GWAS) could actually find accounted for only a sliver of that figure. Either the DNA studies are missing something, or the family studies are counting something that is not really genetic, or both. Which one it is matters well beyond genetics journals. It decides how far polygenic scores can ever go in predicting disease, how much of the gap between people is open to environmental change, and whether a half-century of twin research measured what it claimed to measure.

This article lays out what has been measured, the main explanations with the evidence for and against each, and a small simulation built for this article that shows which explanation leaves which fingerprint. The answer has moved a lot since 2022. For height it now looks largely settled. For many other traits, especially behavioral ones, it does not.

Why people keep asking where the heritability went

The question has escaped the specialist literature. In June 2025 the essayist Scott Alexander published a long post titled “Missing Heritability: Much More Than You Wanted To Know” (Alexander 2025), framed around educational attainment: twin studies suggest something like 40 percent heritability, while identified variants explained much less. It drew a follow-up post of reader comments. Five months later the statistical geneticist Sasha Gusev answered on his own newsletter with a post titled “The missing heritability question is now (mostly) answered” (Gusev 2025), arguing that twin studies roughly double the true narrow-sense figure. Both are blogs, not peer-reviewed work, and they are cited here only to show that the dispute is live and public.

The underlying worry is easy to sympathize with. If a parent reads that intelligence or body weight is “about half genetic” and then learns that a DNA test explains a few percent, one of those numbers is misleading. Commercial reports run into this gap, as our piece on how accurate consumer DNA tests are shows from the customer side.

Four numbers that are easy to confuse

Much of the confusion comes from treating four different quantities as one. Each answers a different question, and they are expected to differ even when nothing is wrong.

  • Twin or pedigree heritability. Estimated from how much more alike identical twins are than fraternal twins, or from resemblance across many kinds of relatives. It needs no genotypes at all. Falconer’s classic formula doubles the gap between identical and fraternal twin correlations.
  • Whole-genome (WGS) heritability. The share of variance in a trait that can be attributed to all sequenced variants, common and rare, among people who are not close relatives.
  • SNP heritability. The same idea, restricted to the common variants on a genotyping array (or imputed from one). Rare variants are mostly invisible here.
  • Variance explained by GWAS hits. How much of the trait the variants passing the strict genome-wide significance threshold (p < 5 × 10−8) account for. This is limited by sample size, not by biology.

“Missing heritability” originally meant the gap between the first and last of these. The modern version of the question is mostly about the gap between the first and the second: once every variant is counted, does DNA reach the twin figure?

What is established

Several facts are no longer seriously disputed. The 2009 review that named the problem (Manolio et al. 2009) listed three candidate explanations: many undiscovered variants of small effect, rare or structural variation that arrays could not see, and interactions among genes or between genes and environment that are hard to detect.

Twin estimates are large and consistent. A meta-analysis of fifty years of twin research covering 17,804 traits and about 14.6 million partly dependent twin pairs reported an average of 49 percent, and found that about 69 percent of traits fit a purely additive model (Polderman et al. 2015). In the authors’ words:

“Estimates of heritability cluster strongly within functional domains, and across all traits the reported heritability is 49%.”

Polderman et al., 2015, Nature Genetics

The small-effect explanation turned out to be largely right for common variants. The GIANT consortium’s height study pooled up to 5.4 million people, found 12,111 quasi-independent significant variants, and reported that together they explain close to all of the common-variant SNP heritability in people of European ancestry (Yengo et al. 2022). Out of sample those variants predicted about 40 percent of height variance in European-ancestry groups and roughly 10 to 20 percent in other groups. So the first gap, between hits and SNP heritability, closes when samples get big enough. It was a power problem.

The second gap is where the action is. Twin estimates for height sit around 80 percent, while array-based SNP heritability rose only from about 45 percent in 2010 to 56 percent after imputation, figures summarized by Young (2019). Sequencing, rather than arrays, was the obvious next test.

The competing explanations

Hypothesis 1: rare variants that arrays cannot see

If many variants with a population frequency below 1 percent affect a trait, arrays will miss them, and natural selection makes this plausible: variants that harm fitness are kept rare. The case for this view is now strong for height. Using whole-genome sequences from 25,465 unrelated people, Wainschtein et al. (2022) estimated heritability at 0.68 (standard error 0.10) for height and 0.30 (standard error 0.10) for body mass index, with low-frequency, protein-altering variants enriched for heritability. The follow-up in roughly 347,630 UK Biobank genomes (Wainschtein et al. 2026, online November 2025) reported a WGS heritability of 0.709 for height and summarized the result across 34 traits this way:

“On average across phenotypes, we find that WGS captures approximately 88% of the pedigree-based narrow sense heritability”

Wainschtein et al., 2026, Nature

By that study’s accounting, rare variants supply about 20 percent of pedigree-based heritability and common variants about 68 percent. The argument against leaning on this too hard is that the comparison is to pedigree estimates within UK Biobank, not to classical twin figures, and that rare-variant estimates are statistically fragile. Gusev (2025) notes that another analysis of BMI from the same data reached different rare-variant numbers. Sequencing also struggles in repetitive regions and with structural changes, a limitation discussed in our explainer on what long-read sequencing finds that short reads miss.

Hypothesis 2: the twin denominator is too big

The alternative is that nothing is missing from the DNA and the family estimates are inflated. Two mechanisms are usually named. Identical twins may share more of their environment than fraternal twins (a breach of the equal-environments assumption), and non-additive genetic effects can inflate Falconer-style estimates of the additive share. Zuk et al. (2012) made the second point formally, arguing that

“a substantial portion of missing heritability could arise from overestimation of the denominator”

Zuk et al., 2012, PNAS

They called the result “phantom heritability” and showed with a model of interacting pathways that, for Crohn’s disease, up to 80 percent of the then-missing heritability could be an artifact of assuming no epistasis. Evidence against a large role for this: the Polderman meta-analysis found most traits consistent with additivity, and the 2026 UK Biobank analysis reported that an additive model fit its data well. Evidence for it comes mainly from within-family designs. Relatedness disequilibrium regression in Iceland, which uses only genetic sharing within families, put height at 55 percent (s.e. 4.4), and sibling regression at 68 percent, both below twin figures, as Young (2019) reports.

Hypothesis 3: SNP heritability itself is biased

A third family of explanations targets the DNA-based side from the opposite direction. Population stratification, assortative mating and indirect effects of relatives’ genes (a parent’s genotype shaping the child’s environment) can all leak into estimates drawn from unrelated people. Young (2019) argued that for educational attainment such indirect effects inflate GREML estimates, and warned that

“more sequence data does not mean methodological problems go away”

Young, 2019, PLoS Genetics

This explanation does not close the gap so much as complicate it. It means DNA-based estimates for social outcomes may be too high, not too low, which would make the remaining distance to twin estimates even larger.

Hypothesis 4: within-family estimates are the real benchmark

A recent preprint (not yet peer reviewed) uses up to 547,126 full sibling pairs from 23andMe across four ancestry groups (Yengo et al. 2025, preprint). It reports within-family heritability of 0.70 for height and 0.48 for BMI, and about 0.86 for height after adjusting for assortative mating. The authors conclude that within-family estimates sit much closer to pedigree and twin estimates than to GWAS SNP heritability. For BMI that leaves a real gap: roughly 0.5 within families against roughly 0.3 from population-based DNA methods. Gusev’s reading of other within-family work points the other way for many behavioral traits, toward averages near 30 percent. The two camps are not yet comparing the same traits with the same methods.

A small simulation: which explanation leaves which fingerprint

To make the competing stories concrete, I wrote a small simulation for this article in Python (NumPy and SciPy, fixed seeds). It does not estimate anything about real humans. It builds four artificial worlds where the truth is known and asks what each of the four measurements would report in each world.

Each world has 4,000 common variants (allele frequency 5 to 50 percent) and 4,000 rare ones (0.05 to 1 percent), unlinked to one another, with 1,000 causal variants of each kind. Effects are drawn per standardized genotype, so rare causal variants have large per-allele effects. Every run then does three things. A GWAS on 20,000 unrelated people records how much true variance the significant hits carry. Haseman–Elston regression on a separate 4,000 people estimates SNP heritability (common variants only) and WGS heritability (all variants). Finally, 5,000 identical and 5,000 fraternal twin pairs, bred from simulated parents by random allele transmission and sharing 10 percent of variance in common environment, give a Falconer estimate. Each scenario was run three times.

# core of the simulation (condensed; full script ~70 lines)
p  = np.r_[rng.uniform(.05,.5,MC), rng.uniform(.0005,.01,MR)]  # allele freqs
sd = np.sqrt(2*p*(1-p))
b[ic] = rng.normal(0, np.sqrt(a_c/KC), KC)        # common causal effects
b[ir] = rng.normal(0, np.sqrt(a_r/KR), KR)        # rare causal effects
def gval(G):                                      # additive + AxA epistasis
    Z = (G - 2*p) / sd
    return Z@b + (Z[:,pr[:,0]]*Z[:,pr[:,1]])@g_aa, Z

# 1) GWAS on 20k unrelated: true variance carried by p<5e-8 hits
r = (Z.T@y)/len(y)/Z.std(0); t = r*np.sqrt((n-2)/(1-r**2))
hit = 2*stats.t.sf(abs(t), n-2) < 5e-8
gwas_true = (b[hit]**2).sum()

# 2) Haseman-Elston on 4k unrelated: SNP (common) vs WGS (all)
def he(cols):
    A = Z[:,cols]@Z[:,cols].T/len(cols)
    x, yy = A[iu], np.outer(y,y)[iu]
    return np.cov(x,yy)[0,1]/x.var()
h_snp, h_wgs = he(range(MC)), he(range(MC+MR))

# 3) twins bred from 4 parental haplotypes
def child(P):
    m = rng.integers(0,2,(2,n,M)).astype(bool)
    return np.where(m[0],P[0],P[1]) + np.where(m[1],P[2],P[3])
e_sh = rng.normal(0, np.sqrt(.10 + (eea if mz else 0)), n)  # shared env
twin = 2*(r_mz - r_dz)                                      # Falconer

Results, as mean ± standard deviation over three runs:

Scenario (true additive h²)TwinWGSSNP (common)GWAS hits
A. Additive, common variants only (0.50)0.49 ± 0.010.52 ± 0.020.53 ± 0.030.20 ± 0.01
B. Rare variants carry 0.15 of 0.500.49 ± 0.020.50 ± 0.020.35 ± 0.010.09 ± 0.01
C. Additive 0.35 plus epistasis 0.15 (0.35)0.59 ± 0.090.37 ± 0.060.36 ± 0.040.09 ± 0.01
D. Additive 0.35, identical twins share 0.075 extra environment (0.35)0.49 ± 0.020.36 ± 0.050.36 ± 0.050.08 ± 0.02

The pattern is the point. In all four worlds the GWAS hits fall far short, because 20,000 people is too few to detect most small effects; that gap says nothing about which explanation is true. The informative comparison is between the three other columns. When rare variants matter (B), twin and WGS agree and only the array-based SNP figure falls short, so sequencing closes the gap. When the twin estimate is inflated by epistasis (C) or by unequal environments (D), WGS and SNP heritability agree with each other and with the true additive value, and the twin figure stands alone above them. Sequencing more genomes would never close that gap, because there is nothing left in the DNA to find.

Epistasis also made the twin estimate noisy: the expected Falconer value here is 0.35 + 1.5 × 0.15 = 0.575, near the observed mean, but single runs ranged from 0.51 to 0.68. With only 4,000 people, Haseman–Elston estimates varied by up to about 0.05 between runs.

What the simulation does not show matters as much. Its variants are unlinked, so it ignores linkage disequilibrium and imperfect tagging. It has no population structure, no assortative mating and no indirect genetic effects, all of which bias real estimates. The effect-size model, in which rarer variants have larger per-allele effects, is an assumption. It demonstrates a diagnostic logic, not a conclusion about any human trait.

Reading the real data through that lens

Apply the fingerprint test to the published numbers. For height, sequenced-genome heritability of about 0.71 (Wainschtein et al. 2026) matches within-family sibling estimates of about 0.70 (Yengo et al. 2025, preprint), and the GWAS hits have caught up with common-variant heritability (Yengo et al. 2022). That resembles scenario B with a modest twin excess on top, which assortative mating adjustments may partly explain. For BMI, a population-based DNA figure near 0.3 against within-family estimates near 0.5 does not match either clean scenario. That gap points to something the simulation lacks, possibly downward bias in DNA methods or non-additive variance in sibling designs, the two options the preprint’s authors themselves raise.

For educational attainment and cognitive traits the picture is murkier still, because indirect effects and stratification push population estimates up while within-family estimates come out low. Here the open question is less “where are the genes?” than “which number should count as heritability at all?” Readers who want the evolutionary background on why harmful variants stay rare may find our piece on selection, mutation and drift useful.

The author’s assessment

What follows is my reading of the evidence, not a settled finding. The missing heritability problem has split into two problems. The first, “why don’t GWAS hits explain SNP heritability,” is solved: it was sample size. The second, “does DNA reach the family estimate,” is close to solved for anthropometric traits like height, where rare variants carry a real but modest share and the twin excess is small. For behavioral and social traits I think the more likely story is that the twin denominator is too large, through some mix of unequal environments, assortative mating and non-additive effects, rather than that a large reservoir of undiscovered rare variants exists. That judgment rests on the consistency of low within-family estimates for those traits, and it could be wrong.

Some specific predictions would settle it, and each could fail:

  • If rare variants explain the remaining gap for a trait, then WGS heritability estimated in families (within-sibship GREML on sequence data) should match twin estimates for that trait. If it matches the lower population figure, the twin figure is inflated.
  • If epistasis drives the excess, the ratio of identical-to-fraternal twin correlations should exceed 2 for the affected traits, and the excess should shrink in designs that compare half-siblings, full siblings and cousins jointly.
  • If unequal environments drive it, identical twins raised apart (or misclassified twins who believed they were fraternal) should show lower resemblance than identical twins raised together, by roughly the size of the gap.
  • If WGS heritability for BMI stays near 0.3 as samples grow past a million genomes with family structure included, while sibling estimates stay near 0.5, the residual is not in variants above 0.01 percent frequency and must be sought in structural variation, non-additivity or bias in sibling designs.

A fair test would run all four measurements on the same people, ideally a large sequenced cohort with twins and siblings, using one phenotype definition. That study does not yet exist at the scale needed for behavioral traits.

Common questions

Does missing heritability mean twin studies are wrong?

Not in general. For height the DNA-based estimates now come close to family estimates, which supports the twin method for that trait. For some behavioral traits the twin numbers look higher than what within-family DNA methods find, so they may overstate the additive genetic share.

Is heritability the same as being “genetically determined”?

No. Heritability is a population statistic describing how much of the variation within a particular group, in a particular environment, tracks genetic differences. It says nothing about how fixed a trait is in one person, and it can change when environments change. Gene expression is also shaped by context, as our overview of epigenetics explains.

Will polygenic scores ever predict as well as twin studies suggest?

At best they can approach WGS heritability, and only in ancestries well represented in the training data. For height the European-ancestry ceiling looks close to 0.7, while the 2022 height map predicted only 10 to 20 percent of variance in other ancestry groups. For traits where twin estimates are inflated, the realistic ceiling is lower than the twin figure.

Why do rare variants matter if each one is so rare?

Individually each affects few people, but there are millions of them, and selection tends to keep variants with larger harmful effects at low frequency. Summed across the genome they can carry a meaningful share of variance, about a fifth of pedigree heritability on average in the largest sequencing study so far.

Further reading

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

References

  1. Manolio, T. A., Collins, F. S., Cox, N. J., et al. (2009). Finding the missing heritability of complex diseases. Nature, 461(7265), 747–753. https://doi.org/10.1038/nature08494
  2. Zuk, O., Hechter, E., Sunyaev, S. R., & Lander, E. S. (2012). The mystery of missing heritability: Genetic interactions create phantom heritability. Proceedings of the National Academy of Sciences, 109(4), 1193–1198. https://doi.org/10.1073/pnas.1119675109 (open copy: DSpace@MIT)
  3. Polderman, T. J. C., Benyamin, B., de Leeuw, C. A., et al. (2015). Meta-analysis of the heritability of human traits based on fifty years of twin studies. Nature Genetics, 47, 702–709. https://doi.org/10.1038/ng.3285
  4. Young, A. I. (2019). Solving the missing heritability problem. PLoS Genetics, 15(6), e1008222. https://doi.org/10.1371/journal.pgen.1008222
  5. Wainschtein, P., Jain, D., Zheng, Z., et al. (2022). Assessing the contribution of rare variants to complex trait heritability from whole-genome sequence data. Nature Genetics, 54(3), 263–273. https://doi.org/10.1038/s41588-021-00997-7
  6. Yengo, L., Vedantam, S., Marouli, E., et al. (2022). A saturated map of common genetic variants associated with human height. Nature, 610, 704–712. https://doi.org/10.1038/s41586-022-05275-y
  7. Wainschtein, P., Zhang, Y., Schwartzentruber, J., et al. (2026). Estimation and mapping of the missing heritability of human phenotypes. Nature, 649, 1219–1227 (published online November 2025). https://doi.org/10.1038/s41586-025-09720-6
  8. Yengo, L., Liang, Y., Wang, X., et al. (2025). Within-family heritability estimates for behavioural and disease phenotypes from 500,000 sibling pairs of diverse ancestries. medRxiv (preprint, not peer reviewed). https://doi.org/10.1101/2025.09.17.25336022
  9. Gusev, S. (2025, November 21). The missing heritability question is now (mostly) answered. The Infinitesimal (blog). https://theinfinitesimal.substack.com/p/the-missing-heritability-question
  10. Alexander, S. (2025, June 26). Missing heritability: Much more than you wanted to know. Astral Codex Ten (blog). https://www.astralcodexten.com/p/missing-heritability-much-more-than