Mendelian randomization (MR) is a technique that uses inherited genetic variants as stand-ins for randomized treatment assignments: because genes are shuffled at conception, a variant associated with an exposure can be used to estimate whether the exposure causes a disease, free of most confounding. Its known weakness across ancestries is genomic — a "tagging" variant that reliably marks a causal variant in Europeans may not mark it in Africans because the surrounding DNA blocks differ. This proposal argues that for infectious and immune-related traits there is a second, entirely non-genomic mechanism that has never been characterized: people's infection risk depends on whom they contact, contacts are strongly ancestry-correlated, and contacts' genotypes affect their susceptibility. Under these conditions, a genetic instrument influences your outcome through your neighbors' biology — a pathway that looks exactly like pleiotropy (the instrument acting through routes other than the exposure) but that no ancestry-aware genomic correction can remove.
The two research literatures that motivate this have never touched. Epidemiological work on relaxing the homogeneous mixing assumption — the classical fiction that everyone contacts everyone equally — shows that structured contact networks change epidemic outcomes drastically: ignoring slum populations in Delhi influenza models underestimated attack rates by 10% to 50% and missed the epidemic peak by weeks Effect of modelling slum populations on influenza spread in Delhi, and stay-at-home interventions were overestimated because physical isolation is impossible in dense settlements Disparities in spread and control of influenza in slums of Delhi: findings from an agent-based modelling study. Separately, work on multi-profile networks proves that highly susceptible sub-populations act as "epidemiological anchors" that drag otherwise resistant groups into sustained transmission Virus Propagation in Multiple Profile Networks. Meanwhile, the ancestry-bias literature documents that instruments and diagnostic criteria discovered in European cohorts fail qualitatively elsewhere — European-derived testing criteria for monogenic diabetes misclassify patients of other ancestries rather than merely shifting thresholds Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing, and large-scale MR tools like GSMR are locked to European reference panels Causal associations between risk factors and common diseases inferred from GWAS summary data. The anchor result supplies precisely the interference mechanism that MR's assumptions forbid, yet no study has connected them.
Core Claim
When contact networks are ancestry-homophilous and susceptibility is partly genetic, the MR estimand acquires a network-mediated term that is indistinguishable from horizontal pleiotropy and is invisible to every correction that operates on one individual's genome at a time. Formally, for individual $i$ with instrument $G_i$ and contact weights $w_{ij}$, the infection-related outcome obeys
$$Y_i = \beta\, G_i + \theta \sum_j w_{ij} G_j + \varepsilon_i,$$
where $\beta$ is the direct genetic effect MR targets and $\theta \sum_j w_{ij} G_j$ is the anchor effect: your neighbors' instruments shift their susceptibility and hence your exposure to infection. Under homophily, $\sum_j w_{ij} G_j$ correlates with both $G_i$ and ancestry, so $\hat\beta$ converges to $\beta + \theta \cdot f(h, C, I_t)$, where $h$ is a homophily parameter, $C$ the between-group contact structure, and $I_t$ the time-varying infected fraction inside a stratified susceptible-infectious-removed model with per-stratum force of infection $\lambda_k = \beta_{\text{tx}} \sum_l C_{kl} (I_l/N_l)^{\alpha}$. The crucial, testable property is that the offset depends on epidemic state $I_t$: it is not a fixed confounder, so principal components and ancestry covariates cannot absorb it. We will derive this offset analytically, quantify it in simulation, and build a correction.
Research Program
The first component is theory. We extend the stratified susceptible-infectious-removed model above to $K$ ancestry strata, each with its own tagging-variant–causal-variant correlation (mimicking the shorter linkage-disequilibrium blocks of African-ancestry genomes), and derive the probability limit of the two-sample MR estimator. This yields a decomposition of "ancestry bias" into two mechanistically distinct parts: weak-instrument bias from correlation decay, and interference bias from the anchor term. The decomposition is the conceptual deliverable — it says these two biases have different signatures and demand different fixes.
The second component is a simulation platform. We build an agent-based model in the tradition of the Delhi slum studies Disparities in spread and control of influenza in slums of Delhi: findings from an agent-based modelling study and the interaction-based computational epidemiology program An Interaction-Based Approach to Computational Epidemiology, with the multi-profile anchor dynamics of Virus Propagation in Multiple Profile Networks made genetic: each agent carries a genotype whose effect on susceptibility is known by construction. Homophily $h$ and instrument-tag correlation per stratum are dialable, giving a factorial design that isolates each bias source.
The third component is the estimator. The naive fix — a per-stratum exponent $\alpha_k$ in a discrete-time time-series SIR (TSIR) model — is ruled out by prior evidence: the negative binomial likelihood in TSIR is valid only for linear birth processes, and with $\alpha \neq 1$ it produced the highest root-mean-squared error and 95% credible interval coverage of only about 49% on a 629-trajectory synthetic ensemble Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models. Instead we develop a stratified version of that paper's continuous-time Bayesian birth-death model (BayesSIR), whose exact transition kernels are computed with the multiBD matrix-exponential machinery the same paper introduces. Full stratification explodes the state space, so we use a composite likelihood: exact within-stratum birth-death kernels with cross-stratum forces treated as time-varying external inputs. We validate this approximation's coverage explicitly, since it is the method's known soft spot. The genetic stage then estimates $\beta$ on residuals from the fitted transmission model, with instruments screened per stratum using stratum-matched reference panels — addressing what GSMR cannot, given its European-panel lock-in Causal associations between risk factors and common diseases inferred from GWAS summary data.
Proposed Experiments
Experiment one reproduces the existing benchmarks to certify the machinery. We regenerate the 629-trajectory stochastic SIR ensemble (ground truth $\alpha = 0.8$, $\beta = 0.0045$, $\gamma = 1$) from Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models and confirm that our unstratified implementation matches its reported behavior: BayesSIR recovers the true parameters, while TSIR with relaxed mixing shows the reported roughly 49% credible-interval coverage. We then replicate the London measles analysis (biweekly incidence, 1944–1951) and check we reproduce the paper's finding that BayesSIR gives strong evidence for $\alpha < 1$ where TSIR's intervals often include 1.0.
Experiment two is the bias decomposition. Holding instrument strength fixed across strata and sweeping homophily $h$ from 0 to 1, we run GSMR-style two-sample MR on the pooled simulated summary statistics; the causal-effect bias should rise monotonically with $h$ even though the instruments are perfectly valid in every stratum. Symmetrically, sweeping tag correlation at $h = 0$ isolates pure weak-instrument bias. If the two curves produce distinguishable signatures (scale with epidemic intensity versus with tag-variant correlation), the decomposition claim is confirmed; if they do not, that itself is a decisive negative result worth reporting.
Experiment three evaluates the estimator on the stratified ensemble: parameter recovery (RMSE), 95% credible interval coverage for $\beta$ and $\beta_{\text{tx}}$, and attack-rate and time-to-peak errors in a Delhi-style two-stratum scenario, where the reference failure magnitude is the reported 10%–50% attack-rate underestimation of homogeneous models Effect of modelling slum populations on influenza spread in Delhi. We state honestly that no real multi-ancestry infection GWAS with contact data exists in the reviewed corpus; the IL-18 asthma–ulcerative colitis mediation study Causal relationship between asthma and ulcerative colitis and the mediating role of interleukin-18: a bidirectional Mendelian study and mediation analysis and the FGF23 aortic-stenosis analysis Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis are homogeneous-cohort MR studies whose cross-ancestry transferability is untested, so real-data validation is a stated follow-up, not a claim.
Baselines
The "do nothing" reference is pooled two-sample MR in the GSMR style Causal associations between risk factors and common diseases inferred from GWAS summary data, run on mixed-ancestry summary statistics with a European reference panel; no reported numeric performance for GSMR appears in the reviewed corpus, so its role is as the standard practice we measure bias against. Beating it means causal-effect bias reduced toward zero and near-nominal interval coverage when homophily is present, at matched sample size and tuning budget.
The strongest existing competitor is BayesSIR, the continuous-time birth-death model that recovered the true parameters ($\alpha = 0.8$, $\beta = 0.0045$, $\gamma = 1$) on the 629-trajectory ensemble where TSIR failed Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models. We run it two ways — pooled across strata, and separately per stratum with no coupling — under identical priors, MCMC budgets, and data. It is the strongest competitor because it is the only statistically principled non-homogeneous model in the corpus; beating it means correct recovery of both transmission parameters and $\beta$ in the stratified setting where the pooled version conflates strata and the per-stratum version ignores the anchor coupling. We also run the stratified TSIR with per-stratum $\alpha_k$, and a principal-component ancestry adjustment of standard MR, both as weaker comparators; and an ablation of our own estimator with the cross-stratum contact matrix set to the homogeneous value, which isolates whether the coupling term is the ingredient doing the work.
Why not just…?
Why not just add ancestry principal components or ethnicity covariates to the MR regression? This is the natural first move, and the corpus gives two reasons it fails. Clinically, European-derived criteria misclassify rather than shift — the monogenic diabetes study found qualitative failure of transferred thresholds Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing. Statistically, our theory says the interference offset depends on the epidemic state $I_t$, so it is not a fixed coefficient a covariate can absorb; experiment two tests this directly by showing bias persists and grows with $h$ even with perfect ancestry labels. If PC adjustment did remove the bias in simulation, our core claim would be falsified and we would report that.
Why not just fit stratified TSIR models with one mixing exponent per sub-population? The reviewed evidence rules this out: the negative binomial likelihood underlying TSIR is derived for linear birth processes, and with a non-homogeneous exponent the model was systematically biased, with 95% credible interval coverage of only about 49% and the highest RMSE on the synthetic ensemble Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models. Stacking several flawed strata compounds the non-linearity. We still run it as a baseline so the comparison is empirical, not just asserted.
Why not just use multivariable MR with network position or contact frequency as an additional exposure? Multivariable MR adjusts for exposures of the index individual, but the anchor mechanism means a neighbor's genotype affects your outcome through their infection status — an indirect effect that cannot be blocked by conditioning on your own covariates, as the anchor dynamics result demonstrates formally Virus Propagation in Multiple Profile Networks. Worse, contact structure is partly chosen, so conditioning on it induces selection on a collider with genotype. We include an MVMR-style adjustment in experiment two; if it matched our estimator's bias reduction, the simpler method would win and we would say so.
Why not just collect better-powered non-European GWAS? That fixes the weak-instrument half of the bias but not the interference half — the two mechanisms are orthogonal in our decomposition, and the $h$-sweep at fixed instrument strength shows bias from interference alone. Diverse biobanks are necessary but, on this account, insufficient for infectious traits.
Decisive Checks
The proposal lives or dies on three numbers. First, the $h$-sweep must show monotone MR bias at fixed, perfect instrument strength — the signature of network interference. Second, our stratified estimator must achieve near-nominal 95% coverage where stratified TSIR sits near the reported ~49% and pooled BayesSIR is biased by stratum conflation. Third, the diagonal-contact ablation must collapse back to pooled-model performance, proving the coupling term is what matters. Each check has a pre-registered failure mode, and any of them failing would meaningfully revise how the field thinks about transferring genetic causal estimates for infection-related traits across ancestries.
References
- Effect of modelling slum populations on influenza spread in Delhi
- Disparities in spread and control of influenza in slums of Delhi: findings from an agent-based modelling study
- Virus Propagation in Multiple Profile Networks
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- An Interaction-Based Approach to Computational Epidemiology
- Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models
- Causal relationship between asthma and ulcerative colitis and the mediating role of interleukin-18: a bidirectional Mendelian study and mediation analysis
- Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis