Wave-Resolved Ancestry: Using Admixture Timing to Split Genetic from Environmentally Mediated Effects

When two previously isolated populations mix, the descendants inherit chromosomes that are mosaics of ancestral segments. Recombination cuts those segments shorter every generation, so the length of a segment acts as a crude clock: long stretches of, say, European DNA in an African American genome mean that ancestry arrived recently; short stretches mean it arrived long ago. Population genetics has learned to read these clocks well enough to date admixture events — African Americans' main pulse at roughly 5.4 generations ago, Maasai at two waves around 16 and 68 generations ago (Inference of multiple-wave population admixture by modeling decay of linkage disequilibrium with polynomial functions). Yet in genetic association studies, ancestry is almost always treated as a static label: a covariate, a principal component, or a categorical local-ancestry call. Nobody uses the clock itself as a covariate.

This proposal argues the clock is not just demographic trivia but an identification tool. A risk allele carried on an ancestral segment exerts its biological effect regardless of when that segment entered the genome. But effects that travel through environment and culture — language, diet, socioeconomic position, discrimination — accumulate with the number of generations an ancestral background has been embedded in the receiving environment. So if two individuals carry the same amount of a source ancestry but that ancestry arrived at different times, outcome differences that scale with generation count measure the environmentally mediated portion of the ancestry effect, while the timing-independent portion measures the genetic portion. Admixture timing is thus a moderator that scales the indirect path of a mediation model and leaves the direct path untouched — and a moderator that nature has conveniently randomized across individuals and even across chromosomal regions within one individual. The plan is to estimate wave-resolved ancestry features (how much of each source ancestry arrived in which pulse), enter them into a blockwise hierarchical regression, and use the timing-by-dosage interaction to decompose ancestry–trait associations into direct and mediated components.

Core Claim

The core claim is twofold. First, replacing static ancestry covariates with wave-resolved ones — per-indivial estimates of ancestry dosage per admixture pulse, obtained from tract-length mixture distributions — removes stratification confounding that both principal components and static local ancestry miss, because two admixture waves from the same source create distinct LD and haplotype backgrounds that current methods collapse into one. Second, and more importantly, the interaction between ancestry dosage and pulse age identifies the mediated (environmental, vertically transmitted) share of an ancestry effect, something no existing stratification correction or genetic mediation framework can do. Existing mediation tools built on Mendelian randomization assume homogeneous, stratification-free samples and balance pleiotropy by assumption (Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration), while the sensitivity analysis that does confront stratification (Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy) handles only univariable bounds. Neither models admixture history at all.

The quantitative intuition is one regression. Let $D_s$ be an individual's dosage of source ancestry $s$ and $G_s$ the generation-weighted average age of that dosage (the expectation of segment age under the individual's tract-length mixture):

$$Y = \beta_1 D_s + \beta_2 (D_s \cdot G_s) + \gamma^\top \mathbf{C} + \epsilon$$

with $\mathbf{C}$ the usual covariate vector. $\beta_1$ is the timing-invariant (genetic or fixed-environment) component; $\beta_2$ is the component that grows with generations of exposure — the mediated share, in the spirit of mediation analysis where an indirect effect is the product of a path into a mediator and a path out (Quantifying Indirect Pathways with Statistical Mediation Analysis). The form of the interaction comes from assuming cumulative environmental exposure adds linearly over generations since admixture; the linear form is an approximation we will test against alternatives in simulation.

How This Differs from the Closest Existing Work

Four lines of work abut this proposal, and each stops exactly where we start. iMAAPs (Inference of multiple-wave population admixture by modeling decay of linkage disequilibrium with polynomial functions) estimates multi-pulse admixture dates at the population level with careful correction for source-LD confounding (uncorrected ancestral LD fabricates "ghost" admixture signals at 500–1,250 generations), but never touches phenotypes or regression; we port its wave decomposition to individuals and push it downstream into association models. HAUDI (An Efficient Lasso Framework for Admixture-Aware Polygenic Scores) uses local ancestry from FLURE-style inference to give variants ancestry-specific effect sizes, reporting a 31.9% improvement in predictive $R^2$ for white blood cell count in Hispanic/Latino participants of the PAGE study — but its local ancestry is a static, undated categorical label; it cannot distinguish two European pulses of different ages, and it has no mechanism for splitting genetic from culturally transmitted effects. Lawson et al.'s Haplotype Components (From individuals to ancestries: towards attributing trait variation to haplotypes) capture fine-scale structure that principal components miss — 80.19% birthplace recall versus 62.68% for PCs, and four times the variance explained in educational attainment in roughly 460,000 UK Biobank participants — but they are descriptive variance partitions, explicitly silent on when the haplotype structures formed and on any causal decomposition. Finally, the MS study of ancestral risk modification (Ancestral risk modification for multiple sclerosis susceptibility detected across the Major Histocompatibility Complex in a multi-ethnic population) fits exactly the kind of $G \times$ ancestry interactions we care about — the HLA-DQB106:02* risk allele roughly doubles the odds ratio on a European versus African background — but it cannot tell whether such interactions are biological or artifacts of wave-specific LD differences, a tension raised directly by Is population structure in the genetic biobank era irrelevant, a challenge, or an opportunity?. Our framework can adjudicate: a biological interaction should be stable across pulses of different ages; an LD artifact should shift with the wave.

Research Program

The first build is a per-individual wave-resolved ancestry estimator. For each individual and each pair of source populations, we fit a mixture of exponentials to the distribution of ancestral tract lengths, where each component corresponds to one admixture pulse; this extends the population-level polynomial LD-decomposition logic of iMAAPs to the individual level, inheriting its source-LD correction so that background LD in the source populations is not misread as an ancient pulse. The estimator's outputs are, per source $s$: a vector of pulse dosages $(d_{s,1}, d_{s,2}, \dots)$ with associated pulse ages $(t_{s,1}, t_{s,2}, \dots)$, and the scalar summaries $D_s$ and $G_s$. We will validate the estimator on the HapMap/1000 Genomes populations whose pulse structure is already published — ASW at about 5.4 generations, Uyghurs at 33.3, Maasai at two waves near 16 and 68 (Inference of multiple-wave population admixture by modeling decay of linkage disequilibrium with polynomial functions) — by checking that our individual-level estimates average to the known population values and that the between-individual spread is calibrated against simulation truth.

The second build is the hierarchical regression and mediation layer. We enter covariates in blocks and measure incremental variance, the standard blockwise-$\Delta R^2$ logic of hierarchical regression, which exists precisely to isolate what an interaction adds beyond main effects without the multicollinearity of simultaneous entry (Modeling Conditional Effects with Hierarchical Regression): block one is standard covariates and principal components; block two adds static local-ancestry dosages (what current best practice uses); block three adds wave-resolved dosages; block four adds the dosage-by-age interactions. Inference on $\beta_2$ and the block-four $\Delta R^2$ uses non-parametric bootstrapping, not Sobel-style normal approximations, because products and interactions of normally distributed coefficients are skewed and the Sobel test loses power (Quantifying Indirect Pathways with Statistical Mediation Analysis). Where explicit mediators are measured (for example socioeconomic indices in a biobank), we fit the full two-equation mediation system with pulse age moderating the exposure-to-mediator path, so the mediated effect becomes $a(G_s) \cdot b$ and its generation-dependence is estimable directly.

A known constraint from the admixture-timing literature bounds the whole program: LD signals older than roughly 500 generations decay into background noise (Inference of multiple-wave population admixture by modeling decay of linkage disequilibrium with polynomial functions), so the method applies to recent and intermediate admixture (roughly the last 500 generations) and we will state that scope explicitly rather than overclaim.

Proposed Experiments

The first experiment is a simulation study built for identification, not just prediction. We simulate multi-pulse admixed populations under known truth — a direct genetic effect proportional to ancestry dosage, and a mediated effect proportional to dosage times generations of environmental exposure, with the mediator optionally measured — plus a "decoy" condition where waves differ only in LD structure and no biology differs, to test false-positive interactions. Primary metrics: bias and bootstrap coverage of $\beta_1$, $\beta_2$, and the indirect effect $a \cdot b$; false-positive rate of wave-interaction detection in the decoy condition; and recovery of the known split between direct and mediated shares. Success means coverage near nominal and near-zero decoy false positives, at pulse separations realistic for Maasai-like histories (16 versus 68 generations).

The second experiment is real-data prediction on the PAGE cohorts (approximately 17,300 African American and 11,776 Hispanic/Latino individuals, as used by An Efficient Lasso Framework for Admixture-Aware Polygenic Scores), where the strongest existing number is HAUDI's 31.9% improvement in predictive $R^2$ over ancestry-agnostic models for white blood cell count in Hispanic/Latino participants. We run wave-resolved covariates through the same Lasso pipeline on the same individuals and same phenotypes; beating the baseline means a larger improvement in held-out $R^2$ under an identical tuning budget, with the ancestry-agnostic model as the "do nothing" reference point.

The third experiment is the interaction-adjudication test on the ARHMS multiple sclerosis cohorts (2,652 Hispanic and 2,435 African American individuals with local ancestry from RFMix v2; Ancestral risk modification for multiple sclerosis susceptibility detected across the Major Histocompatibility Complex in a multi-ethnic population). We re-fit their $G \times$ ancestry models for HLA-DQB106:02* with wave-resolved covariates. If the doubled odds ratio on European background is stable across pulse ages, it is biological; if it attenuates or shifts once wave-specific LD is modeled, it was a tagging artifact.

The fourth experiment applies the mediation split in UK Biobank, where fine-scale structure is known to survive principal-component adjustment (Apparent latent structure within the UK Biobank sample has implications for epidemiological analysis) and where chromosome painting explained 8% of the variance in educational attainment in ALSPAC versus 0.8% for standard PCA (Is population structure in the genetic biobank era irrelevant, a challenge, or an opportunity?). Using educational attainment and a socioeconomic mediator, we estimate the generation-scaling coefficient $\beta_2$ and compare its magnitude against the sibling-trio evidence from the same paper that standard GWAS estimates are 30–40% inflated by stratification.

Baselines

The naive reference point is the ancestry-agnostic model: a standard GWAS or polygenic score with no ancestry handling beyond principal components, run on exactly the same individuals, phenotypes, and folds. This is the reference against which An Efficient Lasso Framework for Admixture-Aware Polygenic Scores reports its 31.9% $R^2$ improvement for white blood cell count in Hispanic/Latino participants, so it anchors all comparisons to a published number. The strongest competitor for prediction is HAUDI itself — the fused-Lasso re-parameterization using FLURE local ancestry that uniformly outperformed its predecessor GAUDI across 18 clinical phenotypes in PAGE — because static local ancestry is the best existing capture of admixed genomes and any gain from dating the tracts must come on top of it. We will run HAUDI's released pipeline and ours with the same variant sets, the same train/test splits, and the same hyperparameter search budget, and we count ourselves as beating it only if held-out $R^2$ improvements exceed its 31.9% figure on the Hispanic/Latino white-blood-cell trait and hold across the other phenotypes. For fine-scale structure, the competitor is the Haplotype Components of From individuals to ancestries: towards attributing trait variation to haplotypes (80.19% birthplace recall and four times the educational-attainment variance of PCs in UK Biobank); we compare blockwise $\Delta R^2$ once HCs and wave-resolved components are entered in both orders, which directly measures how much timing adds beyond haplotype structure. For mediation, the competitor is MAGIC (Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration), whose estimating equations held 95% confidence-interval coverage where standard two-step MR dropped to 70–80%; we run it on the same summary statistics and ask whether our timing-scaled indirect effects agree where its assumptions hold and diverge where stratification is present. The key ablation is our own model with the clock removed: identical pipeline, but all pulse ages set to a single pooled date, so tract lengths inform only dosage and not timing. Every claimed contribution must disappear under this ablation; if it does not, timing was not the active ingredient.

Why not just…?

Why not just adjust for the top 10–20 principal components? This is standard practice, and the file documents exactly why it falls short: latent geographic structure remains detectable in UK Biobank even after adjusting for dozens of PCs (Apparent latent structure within the UK Biobank sample has implications for epidemiological analysis), PCs reach only 62.68% birthplace recall against 80.19% for haplotype components (From individuals to ancestries: towards attributing trait variation to haplotypes), and PC adjustment can overcorrect when ancestry sits on the causal path, biasing true effects toward zero (Is population structure in the genetic biobank era irrelevant, a challenge, or an opportunity?). More fundamentally, PCs are timing-blind by construction: they cannot represent two pulses of the same ancestry at different dates, and they cannot identify a mediated pathway even in principle.

Why not just use static local ancestry, as HAUDI already does? This is the real threat, and we treat it as the primary baseline rather than arguing around it. For pure prediction, static local ancestry may capture most of the signal, and if our held-out $R^2$ does not exceed HAUDI's 31.9% improvement on the same PAGE trait, the prediction half of the proposal fails honestly. But static local ancestry cannot do the identification half: it collapses waves of different ages into one label, so the dosage-by-age interaction — the entire basis for splitting genetic from environmentally mediated effects — is unidentifiable. The decoy simulation settles this directly: when truth contains only LD differences between waves and no biology, a static-ancestry interaction model has no variable that could distinguish it from a real effect, whereas ours does.

Why not just run two-step Mendelian randomization or MAGIC for the mediation question? Off-the-shelf mediation tools assume homogeneous, stratification-free samples and mean-zero direct SNP effects (Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration), while Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy shows that residual stratification drives false-positive rates of 46% to 80% in large samples even for methods marketed as robust. More basically, MR needs instruments for an exposure; in our question the ancestral dosage is itself the exposure-like quantity, and its mediated pathway runs through vertically transmitted environment, which genetic instruments do not touch. MAGIC remains the right baseline where its assumptions hold, and agreement between the two methods in clean regions is a validity check, not a redundancy.

Decisive Checks

The proposal stands or falls on three outcomes. The simulation must show that $\beta_2$ recovers the true mediated share with near-nominal bootstrap coverage while the decoy condition yields near-zero false positives; if timing cannot separate LD artifacts from biological interactions in simulation, nothing downstream matters. The PAGE prediction test must beat HAUDI's published 31.9% $R^2$ improvement under matched conditions, or the prediction claim must be narrowed to the identification setting. And at least one real interaction — the MS HLA-DQB106:02* case being the cleanest candidate — must be adjudicated one way or the other, with the result reported whichever way it goes, because a framework whose first real-world test is a coin flip is not yet a framework.