How Many More Samples Does a Simple Genetic Score Need?

Two fields have asked the same question about crude rules applied at large scale. In market design, a seller can run an elaborate optimal auction or just post one price. Posted Pricing and Competition in Large Markets showed that as the number of buyers grows, the price of simplicity depends only on the upper tail of buyers' valuations. It also asked how many extra buyers the posted price needs to match the optimal auction (its "competition complexity"). For exponential-like tails the answer is a constant of about 1.781.

Statistical genetics faces the same choice. Its main tool is the genome-wide association study (GWAS), which tests millions of DNA variants one at a time for association with a trait such as height or disease risk. A polygenic score adds up a person's variants, weighted by estimated effects, to predict that trait. The simple version keeps only variants whose association statistic clears a significance cutoff and sums them ("pruning and thresholding"). The elaborate version fits a Bayesian model that shrinks every variant's effect toward zero and keeps them all. The gap between the two is the practical face of missing heritability: much of the genetic signal lives in variants too weak to clear any cutoff.

This note proposes to measure and predict a genetic analogue of competition complexity. The target is the sample-size complexity of a trait: the factor by which the GWAS sample must grow before the simple threshold score matches what the Bayesian score achieves today. We also propose to predict that factor from the tail of the trait's effect-size distribution, estimated from summary statistics alone. The practical payoff is a per-trait answer to a funding question. Should a consortium recruit more participants, or adopt a more elaborate method? The answer would also extend to sample sizes nobody has yet collected.

Core Claim

Write the association statistic of variant $j$ as

$$z_j = \sqrt{N}\,\beta_j + \varepsilon_j,$$

where $N$ is the GWAS sample size, $\beta_j$ is the variant's true standardized effect, and $\varepsilon_j$ is approximately standard normal noise, correlated across nearby variants because neighboring variants are inherited together (linkage disequilibrium, LD). A threshold score keeps variants with $|z_j| > t_M$. Here $t_M$ grows roughly like $\sqrt{2\log M}$ in the number of tested variants $M$, which is where the Gaussian maximum sits. Define

$$\kappa(N) = \frac{N'}{N} \quad\text{such that}\quad R^2_{\text{thresh}}(N') = R^2_{\text{Bayes}}(N),$$

where $R^2$ is out-of-sample predictive variance explained.

Our claim is that $\kappa$ is governed by which extreme-value class the distribution of $|\beta_j|$ falls into, as in Correa et al.'s market result. Extreme-value theory sorts distributions into three families by how their largest values behave. The Fréchet family has polynomial, heavy tails. The Gumbel family has exponential-like tails. The Weibull family has bounded support.

For heavy-tailed, oligogenic architectures (a few large effects plus many tiny ones), we expect $\kappa$ to stay close to a small constant. The large effects clear the cutoff early and carry most of the variance, just as the rare top buyer carries most of the welfare in a Fréchet market. For Gumbel-like, near-infinitesimal architectures, we conjecture that $\kappa$ grows with $\log M$. The cutoff rises with panel density while no individual effect stands out, so simple rules pay a penalty that keeps increasing with scale. That conjecture is the thing we will derive and test.

The conjecture also gives a candidate mechanism for an unexplained observation in Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank by establishment of a large-scale PRS computation platform PRS-hub. That benchmark found that 6.6 million variants often predicted worse than a curated 1.1 million. Under our claim, denser panels should hurt thresholding most on near-infinitesimal traits and barely matter on oligogenic ones.

Research Program

The work has three parts.

The first is theory. We will derive $\kappa(N)$ in closed form, or as tight bounds, for independent variants under each of the three tail families. The derivation adapts the order-statistic arguments of Correa et al. to the additive-noise selection problem above. The Bayesian comparison score is the posterior-mean score under the true prior, which is the best any shrinkage method can do. For this part we keep the problem at the LD-pruned level that SumVg: Total heritability explained by all variants in genome-wide association studies based on summary statistics with standard error estimates uses. SumVg de-noises summary $z$-statistics with an empirical Bayes (Brown–Efron) formula to estimate total genetic variance on pruned, nearly independent variants.

SumVg plays two roles here. It supplies the de-noising machinery we reuse. It also motivates the tail framing. SumVg's main finding is that LD score regression, which assumes every variant has a tiny Gaussian effect, returns zero or negative heritability on immune traits whose architecture is not like that, while SumVg recovers stable 10–20% estimates.

The second part is an estimator. From one set of summary statistics we estimate the tail index of $|\beta_j|$. We will apply a Hill-type estimator to SumVg-de-noised $z$-scores above a data-driven quantile, with a parametric bootstrap for uncertainty. We use that bootstrap because SumVg showed that delete-one jackknife fails for sums over many non-smooth tests, while its fdr-aware bootstraps (fdrboot1 and fdrboot2) had the lowest error. We then plug the tail estimate into the theory to predict $\kappa$ at the current $N$ and at future sample sizes.

The third part addresses LD. Wisdom and Foolishness of Noisy Matching Markets studies selection by cutoff under noise. It shows that such selection recovers true quality when the noise is light-tailed and collapses toward a lottery when the noise is long-tailed. We use this dichotomy as the lens for LD. When a few large effects exist, the signal they leak onto correlated neighbors is heavy-tailed. The effective noise at non-causal variants then inherits a heavy tail, so thresholding selects neighbors of true variants rather than the true variants themselves.

Peng and Garg treat a one-dimensional quality measured by many independent evaluators. We have one evaluation of each of millions of correlated variants, so their theorem does not transfer directly. We will therefore test the lens empirically. The test asks whether the fraction of threshold-selected variants that are causal (known in simulation) falls with panel density at the rate the leakage-tail estimate predicts.

Baselines

The naive reference point is $\kappa = 1$: assume that method choice does not matter and that doubling samples helps both scores equally. Any predictor must beat this on error in the log of $\kappa$ across traits.

The strongest analytic competitor plugs heritability and an effective number of causal variants into the standard expected-accuracy formula for polygenic prediction. That formula relates accuracy to sample size, heritability and the number of effective loci, and it implicitly assumes Gaussian effects. Heritability comes from LD score regression, with SumVg substituted where LD score regression fails. We solve the formula for $\kappa$. Because this plug-in uses only the second moment of the effect distribution, it doubles as the ablation that isolates our key ingredient. It sees the same summary statistics as our estimator and ignores the tail. Beating it on held-out traits is the evidence that tail information matters.

The strongest data-driven competitor fits a mixture prior over effect sizes with SBayesR or SDPR. We then simulate threshold and posterior-mean scores under that fitted prior to read off $\kappa$. Both methods use the same summary statistics and LD panel as our estimator, and both get the same number of prior-grid settings.

For the prediction scores themselves, LDpred2 is the Bayesian reference. PRS-hub found it the best single method on 14 of 19 diseases and 9 of 17 quantitative traits. Optimizing and benchmarking polygenic risk scores with GWAS summary statistics reports that the PUMA-CUBS ensemble exceeds LDpred2 by 6.54% average relative $R^2$, so we also compute $\kappa$ against PUMA-CUBS as a ceiling. No reported result was found on Lacuna for sample-size complexity of threshold scores against any Bayesian score. The ground truth must therefore be measured here.

Why not just…?

Why not just subsample the GWAS and measure $\kappa$ directly? We do this, and it is the ground truth. It has one limit that matters. Subsampling, whether with individual-level data or with PUMA-CUBS's summary-level virtual partitions, can only move down from the observed $N$. The funding question concerns sample sizes not yet collected. The decisive test is therefore extrapolation. We fit on subsamples up to 100,000 and predict $\kappa$ at the full UK Biobank size, then compare against direct measurement. If the subsampled curve extrapolated as well by simple log-linear fitting, the theory would add nothing, and we will report that.

Why not just fit a mixture prior and simulate? This is the real threat. A well-fitted SBayesR or SDPR prior encodes the tail implicitly. Our bet is narrower: an explicit tail index estimated from the top statistics is more stable than mixture components fitted to all of them, especially on small studies like the roughly 8,000-person immune GWAS. If the mixture plug-in matches our estimator across the 36 UK Biobank traits and the immune traits, then the contribution reduces to the quantity $\kappa$ and its theory, not the estimator. We will say so and ship $\kappa$ with the mixture estimator.

Why not just use LD score regression heritability and polygenicity? SumVg documents that LD score regression gives zero or negative heritability on non-infinitesimal traits. Those are exactly the traits where we predict $\kappa$ departs most from the Gaussian answer. Our ablation (the second-moment plug-in) directly measures how much is lost by ignoring the tail.

Proposed Experiments

In simulation, we extend SumVg's design (100,000 variants, $N$ from 5,000 to 200,000, heritability of 0.101, 0.191 and 0.295). We draw effects from Pareto (Fréchet), Gaussian (Gumbel) and bounded-uniform (Weibull) families and add LD blocks from a reference panel. We vary $M$ from 100,000 to several million to test whether $\kappa$ grows with $\log M$ under Gaussian effects and stays flat under Pareto effects. In the LD-block simulations we also track the fraction of threshold-selected variants that are truly causal as panel density increases.

On real data, we use the PRS-hub pipeline on the 36 UK Biobank traits. We rerun GWAS on nested subsamples, and for each trait we measure the $N'$ at which the threshold score reaches LDpred2's $R^2$ at $N$. We then compare the measured $\kappa$ with predictions from our estimator, the mixture plug-in, the second-moment plug-in and $\kappa = 1$. We score each predictor by rank correlation across traits and by extrapolation error at full sample size.

We also run two ordering checks. First, oligogenic traits such as rheumatoid arthritis and Alzheimer's disease, where PUMA-CUBS reports its largest gains, should show the smallest $\kappa$. Second, the 1.1-million versus 6.6-million-variant gap should be largest for traits whose estimated tail is Gumbel-like. Finally, the SumVg immune traits test whether tail estimation still works at small $N$.

The proposal fails if predicted and measured $\kappa$ do not correlate across traits, or if the second-moment plug-in predicts $\kappa$ as well as the tail-based estimator.