Randomized lifestyle trials are beginning to ask whether people with high genetic risk benefit more or less from intervention. In the Lung Health Study, each standard-deviation increase in a genetic score for smoking initiation cut the intervention's quit success by about 2.5 percentage points. By year five, treated participants with high genetic risk smoked no more than those with low risk, a pattern the authors called an "equalizer" (Basu et al.). That analysis, like almost every such analysis, used only European-ancestry participants. The obvious next step is to repeat it in African, East Asian, South Asian or admixed trial populations. This note argues that the naive version of that step will give misleading answers.
Polygenic scores are sums of many genetic variants weighted by their estimated effects on a trait. They track true genetic liability much less well outside Europe. A treatment-by-score interaction estimated with a noisier score shrinks toward zero. So a finding that "genetic risk moderates the intervention in Europeans but not in African Americans" may reflect nothing more than score quality. We propose to estimate each ancestry's interaction per unit of true genetic liability, not per unit of the score. The correction uses ancestry-specific estimates of how well the score tracks liability. It rests on a simple observation: when the intervention is randomized, the size of this attenuation can be written down exactly. We then test the correction on a case where the right answer is known.
Core Claim
Within ancestry group $k$, suppose we fit $Y = \alpha + \tau T + \beta P + \gamma_k\, T P + \epsilon$. Here $Y$ is the trial outcome, $T$ is randomized treatment, $P$ is the polygenic score standardized within that ancestry, and $\gamma_k$ is the observed treatment-by-score interaction. Let $\gamma_k^*$ be the interaction per standard deviation of true genetic liability $G$. Randomization makes $T$ independent of both $P$ and $G$. The expected outcome given $T$ and $P$ therefore depends on $G$ only through its best prediction from the score, $E[G \mid P]$. If that prediction is linear, then
$$\gamma_k = \rho_k\, \gamma_k^*, \qquad \rho_k = \mathrm{corr}(P, G \mid k) = \sqrt{R^2_k / h^2_k},$$
where $R^2_k$ is the variance in the trait the score explains in ancestry $k$ and $h^2_k$ is the trait's heritability captured by common variants in that ancestry.
Observational gene-by-environment studies cannot use this identity, because there the environment is correlated with genotype. It is valid in randomized trials. The ratio $\rho_{\text{EUR}}/\rho_k$ predicts how much smaller the observed interaction should be in group $k$ even if the biology is identical. Only a difference in $\gamma_k^*$ is evidence that the intervention's genetic moderation truly differs across ancestries.
The idea extends one from PIGEON. PIGEON is a summary-statistics method that estimates the "oracle" gene-by-environment interaction an error-free score would show, and it reports estimates 20% to 55% larger than naive score-based ones. PIGEON's method assumes a single European reference for how nearby variants are inherited together, so it cannot say how the oracle differs across ancestries. It also does not give trial-level estimates by ancestry. We supply both, in the setting where randomization makes the correction cleanest.
Why the Existing Tools Do Not Settle This
Two lines of work motivate the problem. The Lung Health Study analyses (Basu et al., PIGEON) show that genetic moderation of a behavioral intervention is real and clinically meaningful. Both are restricted to Europeans. The multidomain trials that matter most for aging, the Cui et al. cognitive cohort and the protocols for SINGER and SENSS, run in Asian and European settings. None of them models genotype.
Two further lines supply methods but leave the key quantity untouched. Vargas et al. calibrate scores against continuous West African ancestry ("genetic similarity to Yoruba"). This fixes shifts in the score's mean and spread and reclassifies 33.6% of the top type 2 diabetes decile. Calibration also raised the height score's correlation with measured height from 0.24 to 0.32. Even so, the calibrated score still tracks liability less well than in Europeans, so interactions remain attenuated. DisPred learns an ancestry-invariant disease representation that predicts Alzheimer's disease better in African, admixed and South Asian groups. Better prediction raises $\rho_k$ but does not remove the need to know it.
Two papers constrain the design. Gorfine et al. prove that odds-ratio interactions with binary outcomes are biased by non-collapsibility: an odds ratio averaged over a population differs from the odds ratios within its subgroups, even with no confounding. We therefore work on the risk-difference scale, using linear probability models as Basu et al. did. Randomization also means we need no genetic instruments for the exposure, which sidesteps the instrument-validity problems of Mendelian randomization. Mendelian randomization uses genetic variants as natural experiments to estimate causal effects of an exposure.
Research Program
The work has three parts, each testing a different weakness of the correction.
The first part is estimating $\rho_k$ reliably. We estimate $R^2_k$ directly within each ancestry. The hard part is $h^2_k$, which is poorly estimated in small non-European samples. We report three versions of the correction: one assuming equal heritability across ancestries, one using ancestry-specific estimates, and bounds over a plausible heritability range. All uncertainty is propagated by bootstrap.
The second part replaces the single global $\rho_k$ with a function of continuous ancestry, $\rho(a)$. This matters in admixed individuals, whose genomes are mosaics of segments inherited from different ancestral populations. We estimate $\rho(a)$ along the Yoruba-similarity gradient in REGARDS, which directly addresses the global-versus-local ancestry gap Vargas et al. leave open.
The third part checks the linearity assumption. We compare the linear correction with a nonparametric estimate of $E[G \mid P]$, using the trait itself as a noisy stand-in for $G$ in held-out data.
Baselines
The naive reference point is the Basu et al. specification applied to each ancestry. That is a linear probability or least-squares model with treatment, uncalibrated score, their interaction, and 20 genetic principal components (summary axes of genetic variation). The same model is also fit pooled across ancestries with ancestry covariates. It is the analysis most groups would run first. Vargas et al.'s "covariate trap" result suggests that the pooled version produces unstable coefficients.
The strongest existing competitor is the Vargas calibration, followed by a stratified interaction model with no reliability correction. It is the best published fix for ancestry distortion of scores, and it already improves correlation (0.24 to 0.32 for height). No reported result was found on Lacuna for calibrated scores used as treatment-effect moderators, so there is no published number to beat. Our correction must beat this baseline on the convergence criterion defined below.
A second strong competitor replaces the score with DisPred's ancestry-invariant representation, retrained on the same genotypes. It raises $\rho_k$ through better prediction rather than by correcting for it. We run it both with and without our correction.
The PIGEON oracle estimate is also run, using European reference data, on the European subsets where it applies. It anchors the European $\gamma^*$.
The key ablation applies calibration alone, correction alone, and both, with global versus continuous-ancestry $\rho$. Every method uses the same genotypes, the same score weights (from PRS-CS-auto, a Bayesian score-building method used by Basu et al.), the same covariates and the same bootstrap budget.
"Beating" a baseline has a precise meaning. On a target where the true interaction is shared across ancestries, the corrected estimates should show significantly less between-ancestry spread than the baseline, measured by a heterogeneity test. They should also remain unbiased relative to the European oracle.
Why not just…?
Why not just adjust for genetic principal components or self-reported ethnicity? Principal components remove confounding by population structure. They do nothing about how well the score tracks liability, which is what attenuates the interaction. Vargas et al. further show that entering ancestry as a covariate next to an uncalibrated score creates spurious ancestry effects, their "statistical ghosts." The naive baseline above tests this directly.
Why not just calibrate the scores and stop there? This is the real threat. If calibrated scores already give interaction estimates that agree across ancestries, our correction is unnecessary. We will report that outcome. The theory predicts they will not agree: calibration standardizes the score's scale, but the remaining gap in $R^2_k$ still shrinks $\gamma_k$ in proportion to $\rho_k$. The sex experiment below decides between the two.
Why not just run the trial stratified by ethnicity, without genetics? Stratification estimates average benefit by group. That is valuable, and we report it. It cannot say whether benefit is concentrated among the genetically high-risk within each group, and that is the targeting question the Lung Health Study raised. Hu et al. found that benefit in a 24,396-person workplace hypertension trial depended strongly on education. This shows that moderators matter, and that they may be correlated with the genetic score. We therefore include score-by-education terms to keep the genetic moderator from standing in for education.
Why not use Mendelian randomization to study lifestyle effects across ancestries? With a randomized intervention, Mendelian randomization is unnecessary for identifying the treatment effect. It would also bring in the non-collapsibility and instrument problems documented by Gorfine et al.
Proposed Experiments
The decisive check uses an "environment" that is effectively randomized at conception and available in every ancestry: sex. PIGEON's atlas of genetic effects that differ by sex across 530 UK Biobank traits supplies the European reference. Many such differences, such as the stronger effect of anorexia genetic risk on body-mass index in females, plausibly share biology across ancestries.
Within each UK Biobank ancestry group, we estimate score-by-sex interactions for a panel of these traits with each method. If attenuation drives the ancestry differences, uncorrected interactions should shrink roughly in proportion to $\sqrt{R^2_k}$. Corrected interactions should converge. If the uncorrected interactions do not scale with $R^2_k$, the mechanism is wrong and we stop.
The second experiment moves to an admixture gradient. In REGARDS we estimate how interactions between the score and physical activity or diet, on type 2 diabetes risk, vary with Yoruba similarity. We test whether $\rho(a)$ explains the gradient. These exposures are observational, so this experiment tests only the measurement model, not causal moderation.
The third step applies the validated correction to real trials. In the Lung Health Study, we check whether the corrected European interaction matches the PIGEON oracle direction and magnitude. Basu et al. report 2.5 percentage points per SD for cessation as the uncorrected reference. We also design a genotyping add-on to a multi-ethnic multidomain trial such as SINGER, whose Singapore population includes Chinese, Malay and Indian participants. The pre-registered hypothesis is whether the equalizer pattern for cognition and frailty scores holds per unit of true liability in each group.
The main risk is that the heritability estimates for non-European groups are too uncertain for the bounds to exclude zero difference. The sex experiment will reveal this early, before any trial genotyping is funded.
References
- Understanding the Role of Genetic Heterogeneity in Smoking Interventions: Experimental Evidence from the Lung Health Study
- Reimagining Gene-Environment Interaction Analysis for Human Complex Traits
- To Evaluate The Effectiveness Of Non-Pharmacological Multifactorial Interventions On Cognitive Impairment In Geriatric Patients
- The SINgapore GERiatric Intervention Study to Reduce Cognitive Decline and Physical Frailty (SINGER): Study Design and Protocol
- The SENSS (Stress, Exercise, Nutrition, Sleep, Self-management) study: study protocol for a randomized controlled trial to evaluate an integrated, personalized and stepped care lifestyle approach for people with Parkinson’s disease
- Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults
- Improving genetic risk prediction across diverse population by disentangling ancestry representations
- Unveiling challenges in Mendelian randomization for gene–environment interaction
- Workplace-based primary prevention intervention reduces incidence of hypertension: a post hoc analysis of cluster randomized controlled study