How Much of a Small-Sample Heritability Estimate Comes From the Prior?

SNP heritability is the share of variation in a trait, such as height or blood pressure, that can be traced to common genetic variants measured on a genotyping chip. It is estimated by asking whether genetically more similar people are also more similar in the trait. In large cohorts the data settle the answer. Much of the field's interesting science, however, happens in small slices of data: ancestry subgroups, single sexes, age decades, socioeconomic strata, rare diseases. In those slices the data say little.

Bayesian analysis helps here because it returns a full probability distribution over heritability on the interval from 0 to 1, with no impossible negative values. That distribution then depends on prior choices nobody has audited. One is the explicit prior on heritability itself. The other is a hidden one: the assumption about how effect sizes scale with variant frequency, which is built into how genetic similarity between people is computed. This note proposes to measure exactly how far both choices move small-sample heritability conclusions. It also proposes to replace single-prior reports with prior-robust intervals.

The idea in two sentences: for any heritability estimate, compute the full range of answers a defensible family of priors would give, including priors on the frequency-scaling assumption. Then use that range to sort published-style claims into ones the data support and ones the analyst's choices produced.

Core Claim

The prior-sensitivity question for heritability has an unusual structure that existing diagnostics do not exploit. Heritability is one bounded number. Once the genetic similarity matrix is eigendecomposed (an $O(N^3)$ one-time cost), the likelihood at any heritability value costs $O(N)$, as Quantifying the uncertainty in heritability showed. So sensitivity can be computed exactly on a grid for any prior, with no importance sampling and no linearisation.

Prior sensitivity matters only when $N$ is small, and that is exactly when the eigendecomposition is cheap. The supposed tension between biobank scale and Bayesian sensitivity analysis is therefore mostly illusory for the explicit prior. The hard part is the implicit architecture prior. Changing it changes the similarity matrix, so it needs a new decomposition.

We claim three things. First, across a reasonable prior class $\Gamma$, the bounds

$$\Big[\inf_{\pi \in \Gamma} P_\pi(h^2 > c \mid y),\ \sup_{\pi \in \Gamma} P_\pi(h^2 > c \mid y)\Big]$$

are informative and cheap to compute. Here $P_\pi(h^2 > c \mid y)$ is the posterior probability, under prior $\pi$ and phenotype vector $y$, that heritability exceeds a threshold $c$. Second, for subgroups of a few hundred to a few thousand people, these bounds are wide enough to overturn some conclusions that a single-prior posterior or a likelihood interval would report as settled. Third, the frequency-scaling exponent $\alpha$ belongs in $\Gamma$. Its effect on single-component heritability, and especially on differences between strata, can rival the effect of the explicit prior.

Relation to Prior Work

Two papers motivate the proposal by exposing the gap. Quantifying the uncertainty in heritability computed exact grid posteriors for heritability in the ARIC cohort. Posteriors in 4,000 Caucasian women matched the frequentist bell curve. For QT interval in 160 African American women the posterior was nearly flat, and for height in that group it piled up at zero. That paper used a single flat prior on $h^2$ and an inverse-Gamma prior on total variance and never varied them. When the posterior is nearly flat, the reported shape is mostly prior.

Assumptions about frequency-dependent architectures of complex traits bias measures of functional enrichment showed that LD score regression silently fixes $\alpha = 1$ in $\sigma_j^2 \propto [p_j(1-p_j)]^{1-\alpha}$. Here $\sigma_j^2$ is the effect variance of variant $j$ and $p_j$ its allele frequency. The empirical average across 47 traits is about 0.38, and the default biases enrichment estimates.

That paper also partly contradicts us: it found total heritability stable across $\alpha$. We take this head on. The stability came from frequency-bin annotations in stratified LD score regression absorbing the misspecification. A single-similarity-matrix mixed model has no such bins. Whether stability carries over there, particularly for contrasts between strata, is an open empirical question we test rather than assume.

Two lines of work supply method. Detecting and diagnosing prior and likelihood sensitivity with power-scaling perturbs the prior or likelihood by raising it to a power and measures the shift with a cumulative Jensen–Shannon distance, using a threshold of 0.05. We reuse its distance and its separation of "prior dominates" from "prior-data conflict". We compute both exactly on the grid rather than by importance sampling. We also add location shifts, which that paper notes power-scaling cannot produce.

Evaluating Sensitivity to the Stick-Breaking Prior in Bayesian Nonparametrics searches for worst-case prior perturbations through local derivatives and warns that linearisation fails for large shifts. Our one- and two-dimensional problem lets us find worst cases globally instead. The high-dimensional warnings in The Bayesian Analysis of Complex, High-Dimensional Models: Can It Be CODA? are a reason the "uninformative" flat prior on $h^2$ should not get a privileged status.

SumVg and the spike-and-slab model of Bayesian Variable Selection in Searching for Additive and Dominant Effects in Genome-Wide Data are adjacent but address different questions: summary statistics and sparse architectures. We leave both out of scope, except as a sanity check in simulation.

Research Program

The core object is an exact posterior over heritability and $\alpha$ for individual-level data. For each $\alpha$ on a small grid (for example 0, 0.25, 0.38, 0.5, 0.75, 1), we build the frequency-weighted similarity matrix and eigendecompose it once. We then evaluate the marginal likelihood of $y$ on a fine heritability grid, after analytically integrating the variance scale as in Furlotte et al.

This gives a likelihood surface $L(h^2, \alpha)$. Every prior in $\Gamma$ then yields its posterior by reweighting, so we can compute the infimum and supremum of any posterior summary exactly.

The prior class is specified before any real data are touched. It contains three families. The first is Beta priors on $h^2$ with mean anywhere in [0.05, 0.6] and prior sample size at most 4. The second varies the inverse-Gamma hyperparameters on total variance over an order of magnitude. The third covers priors on $\alpha$ ranging from a point mass at 1 (the LD score regression default) to diffuse over [0, 1].

We also include one data-driven member. It is an empirical-Bayes prior fitted to the distribution of large-sample heritabilities across the 551 traits of Phenome-wide heritability analysis of the UK Biobank.

Each analysis then reports three things. The first is the robust interval for $P(h^2 > c)$, and for differences between strata, $P(h^2_A > h^2_B)$. The second is an exact informativeness score: the Jensen–Shannon distance between prior and posterior. The third is a classification of the claim as "data-determined", "prior-dominated" or "prior-data conflict", using the power-scaling logic.

For strata larger than roughly 20,000 people, where decomposition becomes costly, we use the moment-matching estimator from the UK Biobank study. We exploit the fact that sensitivity is negligible at that size, and we verify this on subsamples rather than assuming it.

Baselines

The naive reference point is frequentist restricted maximum likelihood with a Wald interval from the inverted Hessian, as in GCTA-style analyses like Mapping the Genetic Variation of Regional Brain Volumes as Explained by All Common SNPs from the ADNI Study. It is what most papers report. It runs on the same similarity matrix ($\alpha = 1$ and $\alpha = 0.38$) and the same covariates. Furlotte et al. already show it fails at the boundary, so beating it on coverage is necessary but not impressive.

The strongest competitor is a boundary-respecting likelihood interval: the restricted likelihood profiled over heritability, truncated to [0, 1], with the interval derived from the likelihood ratio. It uses no prior, so it is the natural answer to "why bring in priors at all", and it already handles the pile-up at zero. We compute it from exactly the same $L(h^2, \alpha)$ surface, so compute and data are identical by construction.

The single-prior Bayesian posterior of Furlotte et al. (flat on $h^2$) is the third baseline. For the 160-person subgroups, Furlotte et al. report only qualitative findings: a "nearly flat" posterior for QT interval and a peak at zero for height. No reported coverage or interval-width numbers for any of these methods were found on Lacuna, so all quantitative comparisons are new measurements.

The key ablation removes $\alpha$ from the prior class, fixing it at 1 and then at 0.38. This isolates how much of the robust interval's width comes from the architecture prior rather than the explicit prior.

Beating the baselines means two things in simulation. Our intervals must have nominal or better coverage for both levels and stratum contrasts when the generating $\alpha$ and heritability are drawn from realistic ranges. They must also have lower decision error on stratum comparisons, meaning fewer confident claims of a difference when none exists. The likelihood interval at fixed $\alpha$ must fall short on at least one of these. If it does not, the conclusion is that priors add nothing here, and we will say so.

Why not just…?

Why not just use the profile-likelihood interval, which already respects the [0, 1] boundary? This is the most serious threat. For the level of heritability at a correctly specified $\alpha$, it may perform as well as any Bayesian interval. Our bet is that it fails in two places. One is when $\alpha$ is misspecified, since profiling over $\alpha$ in small samples gives near-flat profiles that the interval ignores. The other is in stratum contrasts, where two weak likelihoods are compared. The head-to-head coverage experiment below settles this directly. If the profile interval matches us everywhere, the deliverable shrinks to the audit itself.

Why not just run priorsense on a Stan mixed model? It would work in principle, but it requires a converged MCMC run per stratum. It also explores only the power-scaling direction, which Detecting and diagnosing prior and likelihood sensitivity with power-scaling acknowledges cannot shift the prior's location. Our grid computes the same distance exactly and covers location shifts and $\alpha$. We will run priorsense on the ARIC subgroups as a cross-check that both agree where they overlap.

Why not just adopt one informative empirical-Bayes prior from the phenome-wide distribution? That may be the best single prior, and it is a member of our class. However, it encodes a European, large-cohort architecture and may be most wrong in exactly the non-European small strata where it would carry the most weight. We test it as its own arm. If it alone is calibrated across simulated ancestries and strata, robust intervals are unnecessary and we would recommend it instead.

Proposed Experiments

The first experiment is simulation with real genotypes. We take 1000 Genomes Phase III individuals (about 2,500, across ancestries) and subsets of 160 to 2,000. We simulate phenotypes with heritability in {0, 0.05, 0.2, 0.5} and generating $\alpha$ in {0, 0.38, 1}. We also simulate pairs of strata with equal or unequal heritability.

Metrics are coverage and width of 90% intervals, calibration of $P(h^2 > 0.1)$, and false and true claim rates for stratum differences. We compare Wald, profile likelihood, the flat-prior posterior, the empirical-Bayes prior and the robust bounds, all on the same surface.

The decisive check is whether, at $N \le 1{,}000$ with $\alpha$ misspecified, the profile interval's coverage for contrasts falls measurably below nominal while the robust interval holds.

The second experiment is a reanalysis of ARIC, where access permits. We use the five traits and the three strata of Furlotte et al. The question is which conclusions in the 892- and 160-person groups change sign or certainty across $\Gamma$. The prediction is that QT interval at $N = 160$ is classified as prior-dominated and Caucasian women at $N = 4{,}000$ as data-determined. If the second prediction fails, the prior class is too wide.

The third experiment is a stratified UK Biobank audit. Height (reported $h^2_{\text{SNP}} \approx 0.685$ in the phenome-wide study) serves as a large-$N$ anchor where sensitivity should vanish. We then subsample to stratum sizes typical of age-by-sex-by-deprivation cells. We ask which subsampled contrasts remain data-determined across $\Gamma$. These include the decline of 3.5% per decade in heritability of physical measures, and the near-zero heritability of lifestyle traits.

A clean result would be a sample-size threshold, per trait class, below which heritability contrasts should not be reported without robust bounds. That rule is concrete and checkable for applied geneticists.