When someone in a household gets infected, who else in that home gets sick is usually attributed to physics: how many people share a bedroom, how long they spend indoors, how poor the family is. Yet the same outcome data routinely show disparities that track ancestry, and epidemiology has no way to tell whether elevated secondary attack rates in admixed or minority households reflect the domestic environment or the hosts' genetic susceptibility. The two literatures that could answer this have never been joined. Household transmission studies measure secondary attack rates meticulously but collect no host genomics — the Comvida 2 study tracked 2,024 people across 593 households in a Rio de Janeiro favela and found within-household infection risk of 11.4% (95% credible interval 5.7–17.2%) versus 74.2% (70.3–77.8%) from the community, concentrated in homes with more than four persons per bedroom, but sequenced nothing SARS-CoV-2 transmission in a highly vulnerable population of Brazil: results from a household longitudinal study. Genetic-ancestry studies do the opposite: multi-omic profiling of 1,049 COVID-19 patients at Stanford used local ancestry inference to find severity loci on chromosomes 14 and 21 tied to African and Oceanian ancestry, and the HLA-B*07:02 allele raising severe-disease risk 2.7-fold, but the cohort was hospital-based individuals with no household structure, so no transmission could be observed Deconvoluting complex correlates of COVID-19 severity with a multi-omic pandemic tracking strategy.
The gap matters beyond attribution. Ioannidis et al. themselves showed that during Northern California's second wave, cases shifted toward people with more than 10% Indigenous American ancestry with no accompanying shift in viral clades — ancestry acting as a proxy for social exposure and domestic crowding rather than biology. If ancestry-associated infection disparities are environmental, interventions should target housing; if they are immunogenetic, risk stratification should target hosts. Getting this wrong in either direction misallocates public health resources, and current methods cannot adjudicate it because no study has jointly modeled local genetic ancestry and domestic physical structure in one causal framework.
Core Claim
The proposal claims that a within-household discordance design — comparing infection outcomes among contacts in the same household who share the crowding, income, and community exposure but differ in their inherited local ancestry tracts — cleanly separates host-genetic susceptibility from household-environmental confounding, and that this separation can be certified with formal sensitivity bounds rather than asserted. In an admixed population, siblings and co-residents inherit different mosaics of ancestral chromosome segments through ordinary Mendelian segregation, so there is genuine within-household variation in local ancestry at any given locus even when the family's overall ancestry mix is fixed. Conditioning on the household removes every household-level confounder at once; the residual question is only whether unmeasured person-level confounders (age, prior immunity) could mimic a genetic effect, and that residual question has a quantitative answer via the Robustness Value machinery of MR-SENSEMAKR Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy.
Research Program
Three components are built and analyzed together.
The first is a conditional chain-binomial transmission model. Following the Comvida 2 design's Bayesian chain binomial structure fit in Stan, each contact $i$ in household $h$ develops infection within the follow-up window with probability $1 - e^{-\lambda_h s_{ih}}$, where $\lambda_h$ is the household's force of infection from the index case and community importation combined, and $s_{ih} = e^{\beta_g g_{ih} + \beta_a a_{ih}}$ is that contact's relative susceptibility, a function of local ancestry dosage $a_{ih}$ at candidate loci (from RFMix-style admixture mapping, as deployed by Ioannidis et al.) and genotype indicators $g_{ih}$ for immunogenetic variants such as HLA alleles. Conditioning on the observed number of infected contacts $k_h$ out of $n_h$, the shared household hazard $\lambda_h$ cancels from the likelihood, leaving only the relative susceptibilities — the same cancellation that makes conditional logistic regression robust to shared confounders. Every household-level variable — persons per bedroom, income, transit use, neighborhood viral load — drops out by construction rather than by adjustment. Local ancestry inference at biobank scale is now cheap enough for this to be practical, using RFMix for tract-level mapping and the neural autoencoder clustering of Neural ADMIXTURE, which runs orders of magnitude faster than classical matrix factorization and scales to large cohorts in hours rather than weeks Neural ADMIXTURE: rapid population clustering with autoencoders.
The second component is an estimator for gene-by-environment interaction on the binary infection endpoint that avoids the trap Gorfine et al. documented: standard two-stage instrumental methods (Two-Stage Predictor Substitution and Two-Stage Residual Inclusion) are biased for binary outcomes because the odds ratio is non-collapsible — it does not return to its true value even when confounding is perfectly controlled, and both estimators showed persistent bias and type-I error inflation in their simulations Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction. We therefore place the interaction on a log-risk (Poisson-style) scale, where risk ratios are collapsible, and estimate $\beta_g$ jointly with a household-density interaction term inside the conditional likelihood. This is the one place the proposal departs from existing software rather than reusing it.
The third component is certification. For every claimed genetic or local-ancestry effect on susceptibility, we report a Robustness Value $RV_\alpha$: the minimum strength, on a variance-explained scale, of an unmeasured person-level confounder correlated with ancestry that would shrink the effect to a fraction $\alpha$ of its estimate. Following Cinelli et al., this bound is benchmarked against observed principal components rather than chosen arbitrarily, so the reader sees whether the effect survives confounders weaker or stronger than the genetic structure already in the data. This responds directly to their "Large-N Trap" demonstration: in the UK Biobank ($n = 291{,}274$), residual confounding explaining a mere 0.01% of variance produced $p < 0.05$ false positives, and traditional robust MR methods showed false-positive rates of 46% to 80% under slight pleiotropy. Any infectious-disease genetics finding on modern cohorts that does not survive this scrutiny is not publishable, and none of the transmission literature has applied it.
Proposed Experiments
The primary experiment extends the Comvida 2 design: add low-pass whole-genome sequencing (the ~2.6x coverage Ioannidis et al. showed sufficient for imputation-assisted local ancestry) or dense SNP genotyping to the existing 2,024 individuals across 593 households, re-run the 14- and 28-day follow-up testing with RT-PCR and serology, and estimate within-household susceptibility effects with the conditional model. The design's reported parameters calibrate everything: within-household risk of 11.4%, extra-household risk of 74.2%, density and income gradients as the environmental mixture, so simulations can be planted with known effect sizes (including a 2.7-fold susceptibility allele modeled on the Stanford HLA-B*07:02 severity effect) and known environmental confounding drawn to match the real household covariate structure. Because 593 households may underpower within-household genetic contrasts — siblings share much of their genome, so the effective sample for local-ancestry discordance is smaller than the headcount — a pre-registered power simulation on the real pedigree and admixture structure determines whether pooling multiple admixed sites (Brazil plus a comparable US cohort) is required before any data collection begins.
A second experiment tests generalizability directly, the population-transfer question at the heart of the ancestry-bias direction. The ORCHARDS cohort in southcentral Wisconsin measured secondary infection rates of 12.2% (95% CI 8.1–17.4%) for human metapneumovirus and 19.2% (13.8–25.7%) for seasonal coronavirus across 128 families, with transmission linked to persons per bedroom for coronavirus ($p = .048$) but to index age for HMPV ($p = .004$) — notably, severity symptoms predicted nothing, showing the endpoint is susceptibility, not severity Household transmission of human metapneumovirus and seasonal coronavirus. We re-fit our susceptibility model across cohorts of different ancestry composition and test whether effect estimates from one population predict attack rates in another. Given that diagnostic criteria calibrated in White populations failed outright in non-European groups for monogenic diabetes Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing, we expect transfer to fail for ancestry-specific loci and to hold for universal exposures like density — quantifying exactly which component of transmission risk is portable is itself the deliverable.
A third simulation suite stresses the estimator: null simulations with ancestry-correlated unmeasured confounding calibrated to Cinelli et al.'s Large-N scenarios, measuring false-positive rates against the 46–80% inflation they documented for traditional robust MR under slight pleiotropy; implanted-effect simulations measuring bias and credible-interval coverage of our log-risk estimator versus logistic 2SPS and 2SRI, where Gorfine et al.'s finding of persistent non-collapsibility bias is the expected failure mode; and saturation simulations where crowding is so extreme that exposure overwhelms any genetic differences, testing the honest limit of the design.
Baselines
The naive reference point is what most household studies would do with genomics bolted on: a logistic regression of secondary infection on self-reported race/ethnicity and persons-per-bedroom, with the top 10–20 principal components as covariates. It is the reference because it is the field's default, and because Cinelli et al. showed precisely why it fails in large samples — PC adjustment did not protect the BMI–Townsend Deprivation Index association, which was overturned by a confounder only half as strong as the observed PCs, while BMI on diastolic blood pressure required a confounder 14 times stronger than the first 20 PCs to invalidate. A baseline that cannot distinguish robust from fragile associations cannot answer the attribution question. It will be run on the identical simulated and real datasets with the same genotype inputs.
The strongest existing competitor is the Comvida 2 Bayesian chain binomial transmission model in Stan, extended with a standard genome-wide association layer — that is, the environmental transmission machinery of the strongest household study combined with the standard genetic adjustment of the GWAS tradition, fit between households rather than within them. This is the real rival because its environmental half already explains the transmission landscape well (it produced the 11.4% / 74.2% decomposition), and its genetic half is what any genotyping team would add by default. It will be fit to the same data with the same MCMC compute budget and the same tuning iterations as our model, differing only in the absence of within-household conditioning and local-ancestry discordance. Beating it means three things: recovering implanted genetic susceptibility effects with bias under 10% of the true log-risk ratio where the between-household baseline absorbs the crowding–ancestry correlation into its genetic coefficients; a false-positive rate at or below 5% in the ancestry-confounded null simulations where we expect the logistic-PC baseline to inflate toward the 46–80% range Cinelli et al. reported for traditional methods; and calibration of predicted attack rates across density strata at least as good as the chain-binomial environmental model, since the genetic layer must not degrade the environmental fit.
The key ablation isolates each ingredient separately: the same conditional likelihood run with self-reported ethnicity replacing RFMix local ancestry, and the same genetics run between households instead of within. If the within-household, local-ancestry version does not beat both ablations on the bias and false-positive criteria, the design's specific ingredients — not just genotyping or just conditioning — are not earning their place, and the simpler ablation wins.
Why not just…?
Why not just adjust for the top 10–20 principal components in a standard GWAS? Because PC adjustment provably does not survive the Large-N Trap. Cinelli et al. demonstrated in the UK Biobank ($n = 291{,}274$) that a confounder explaining 0.01% of variance generates significant false positives, that the BMI–Townsend association — an environmentally embedded outcome much like infection risk — was explained away by a confounder half as strong as the PCs already controlled for, and that traditional robust MR methods reached 46–80% false-positive rates under slight violations Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy. PCs capture between-population structure, but crowding and ancestry are correlated within the residual variation PCs leave behind; the within-household conditioning removes that entire axis rather than adjusting for its leading components. Our null simulations run this baseline head-to-head and will report its false-positive rate directly.
Why not accept that crowding explains everything and skip the genetics? This is the most serious objection, and the evidence is genuinely mixed. In the ORCHARDS cohort of the same 128 families, household density predicted seasonal coronavirus transmission ($p = .048$) but not HMPV, where index age drove spread ($p = .004$) — exposure metrics did not explain both viruses in the same homes, and symptom severity predicted neither, so physical proximity is demonstrably not the whole story Household transmission of human metapneumovirus and seasonal coronavirus. On the genetic side, HLA-B07:02 raised severe COVID-19 risk 2.7-fold with HLA-C15:02 protective. But the objection could still be right for transmission specifically, especially at extreme saturation: Comvida 2 found within-household risk maximized above four persons per bedroom, where exposure may indeed swamp susceptibility differences. We treat this as an empirical question with a pre-registered answer format — the saturation simulations map effect attenuation as a function of density, and if the genetic coefficient goes to zero above a density threshold while remaining detectable below it, the crowding-only view wins in exactly the strata where it wins, and the study says so.
Why not just use standard two-stage instrumental variable methods for the gene–environment interaction? Because Gorfine et al. proved they break on binary endpoints: non-collapsibility of the odds ratio biases both 2SPS and 2SRI for binary outcomes even under perfect confounding control, with 2SRI merely "less wrong," and the authors conclude the problem is unresolved Unveiling Challenges in Mendelian randomization for Gene-Environment Interaction. Our log-risk formulation exists precisely because of this result, and our simulation suite runs both estimators as internal comparisons — if either matches our estimator's bias and coverage, we will report that the simpler standard tool suffices and drop the new estimator.
Why not just use self-reported ethnicity from standard epidemiological surveys instead of sequencing? Self-report conflates genetic ancestry with socioeconomic exposure at exactly the resolution this question needs. Ioannidis et al. found that self-reported categories obscured what admixture mapping revealed — ancestry-specific loci on chromosomes 14 and 21 plus the shift of cases toward people with over 10% Indigenous American ancestry without any viral-clade shift — and in the Comvida 2 population, self-reported distancing behavior showed no protective association, indicating the survey channel itself is compromised by reporting bias Deconvoluting complex correlates of COVID-19 severity with a multi-omic pandemic tracking strategy. The monogenic diabetes experience — White-calibrated diagnostic criteria failing in non-European groups — is the cautionary precedent for coarse categories. Still, this is why the self-report ablation exists: if it performs as well as local ancestry in the blinded analyses, the cheap version wins and the expensive one is unjustified.
Decisive Checks
The study succeeds if three checks pass on the real admixed household data: within-household local-ancestry or genotype effects on susceptibility are estimated with credible intervals that exclude zero and survive Robustness Values benchmarked against observed PCs, where the between-household logistic-PC baseline attributes the same signal to genetics under ancestry-correlated crowding; the environmental coefficients from the conditional model reproduce the Comvida 2 density and income gradients (within-household risk 11.4%, extra-household 74.2%) without degradation; and the null simulations hold false positives at nominal rates while logistic 2SPS/2SRI show the non-collapsibility bias Gorfine et al. documented. The study fails informatively if the power simulation shows the admixture discordance among co-residents is too thin in available cohorts — in which case the deliverable becomes the design and power map for the multi-site cohort that can answer the question, plus the demonstration that every method currently in the literature would confound genes with crowding if applied naively.
References
- SARS-CoV-2 transmission in a highly vulnerable population of Brazil: results from a household longitudinal study
- Deconvoluting complex correlates of COVID-19 severity with a multi-omic pandemic tracking strategy
- Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy
- Neural ADMIXTURE: rapid population clustering with autoencoders
- Unveiling challenges in Mendelian randomization for gene–environment interaction
- Household transmission of human metapneumovirus and seasonal coronavirus
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing