When geneticists find that a causal effect estimated in Europeans shrinks or flips in East Asian or African cohorts, they almost always blame the genome: different linkage disequilibrium (LD) patterns — the statistical correlations between nearby genetic variants — different allele frequencies, different pleiotropy. When plant physiologists find that a genotype looks heat-tolerant in one region and heat-sensitive in another, they know to first blame the atmosphere: heating air changes its vapor pressure deficit (VPD), the evaporative drying power of the air, so a "temperature" experiment is silently also a drought experiment. These two literatures have never spoken to each other. Yet in human population genetics, ancestry and climate are confounded by geography — the same gradient that shaped LD structure also determined where each cohort's participants live and what thermal and humidity conditions they experience. The question this proposal attacks is simple and, on the current evidence, unanswerable: when a thermal or metabolic causal effect "fails to port" across ancestries, how much of that failure is genetic architecture and how much is unmeasured coupling between ambient heat and atmospheric moisture in the environments the cohorts actually occupy?
The proposal builds a causal inference pipeline in which the environmental exposure entering a gene–environment interaction is, by construction, thermally decoupled from humidity, and applies it across ancestrally diverse cohorts. The motivating tension comes from two sides. On the physiology side, the Hot Serial Cereal experiment showed that warming a wheat canopy by 1.3 °C day / 2.7 °C night under constant relative humidity produces clean temperature response curves, and the AmbControl chamber work showed that even the best semi-open chambers leave relative humidity and VPD co-varying with the temperature treatment — meaning that most field and observational "heat" signals are heat-plus-drying signals. On the genetics side, Gorfine et al. built two-stage instrumental-variable estimators for gene–environment (G×E) interactions but never touched climatic exposures or ancestry-specific LD, and X-Wing showed that cross-ancestry portability is local (for mean corpuscular volume, roughly 1% of the genome carries over 60% of the European–East Asian genetic covariance) but treats the environment as background noise.
Core Claim
The core claim is that cross-ancestry heterogeneity in causal estimates of thermal-environment effects on human physiology is decomposable, and that current methods misattribute part of it. Specifically: an estimator that (a) replaces the raw ambient thermal exposure with two separate exposures — sensible heat $T$ and evaporative demand $D$ (VPD) — matched to each cohort's recruitment geography, and (b) applies two-stage residual inclusion (2SRI) with ancestry-specific instruments, will recover ancestry-invariant causal G×$T$ effects where naive single-exposure estimators report spurious ancestry-specific effects. The misattribution mechanism is concrete: because ancestry tracks geography, and geography tracks the coupling between $T$ and $D$ (humid-hot versus dry-hot climates), an estimator that ingests a coupled thermal index absorbs the $D$-effect into its "genetic non-portability" term. The estimator's outcome model is:
$$\text{logit}\,P(Y{=}1)=\beta_0+\beta_G G+\beta_T\hat T+\beta_D\hat D+\beta_{GT}G\hat T+\beta_{GD}G\hat D+\gamma \hat R$$
where $\hat T$ and $\hat D$ are fitted values from first-stage regressions of the ambient thermal index and VPD index on genetic instruments, $\hat R$ is the first-stage residual, and $G$ is a polygenic score built with ancestry-specific weights. The coefficients of interest are $\beta_{GT}$ (genetically moderated response to heat itself) and $\beta_{GD}$ (genetically moderated response to evaporative demand). Under the hypothesis, $\beta_{GT}$ estimates agree across ancestries while $\beta_{GD}$ carries the geography-correlated signal that naive pipelines load onto ancestry.
Research Program
Three connected workstreams. First, a semi-synthetic benchmark, built in the spirit of SpaCE — which showed that standard causal models overfit smooth spatial confounding in 124 heat- and pollution-exposure datasets — but extended with a genetic layer: simulated genomes with ancestry-specific LD decay, real gridded $T$ and humidity fields as the confounding substrate, and ground-truth G×$T$ effects that are exactly ancestry-invariant. The benchmark's job is to measure how often each estimator falsely reports ancestry-specific heat effects when none exist. Second, an empirical re-analysis of the critical-care temperature result of Soltani et al., which pooled 118,721 ICU patients from MIMIC-IV and eICU into one aggregate GAM (a flexible spline regression) and found the mortality nadir at 36.8 °C in MIMIC-IV and 37.0 °C in eICU. We stratify that analysis by recorded ethnicity and test whether the nadir and the mortality slope of the "temperature gap" (daily temperature variability) are invariant; following the monogenic-diabetes lesson of Jones et al. — where White-derived diagnostic cutoffs misclassified minority patients — we expect some drift, and the stratified GAM quantifies it. Third, a two-sample G×E MR application across European (UK Biobank) and East Asian (BioBank Japan / China Kadoorie Biobank) cohorts, using adiposity instruments (validated by GSMR, where a 1-SD BMI increase caused a 3.29-fold type-2 diabetes risk increase in up to 405,072 participants) as the thermoregulation-relevant genetic axis, with decoupled $T$ and $D$ exposures matched to recruitment regions, Steiger directionality tests, and HEIDI-outlier filtering as in Zhu et al..
Baselines
The "do nothing" reference point is the aggregate-population GAM of Soltani et al., which reports the pooled mortality nadir of 36.8 °C (MIMIC-IV) and 37.0 °C (eICU), an adjusted odds ratio of 1.08 to 1.18 for spending 10% more time below 36 °C, and 1.09 for 10% more time above 38 °C. It is the strongest competitor for the ICU question because it is the only published quantitative anchor for the temperature–mortality curve; we re-run it identically (same spline basis, same acuity adjustments) but per ancestry stratum, on identical data and compute. Beating it means showing that per-ancestry nadirs or temperature-gap slopes differ beyond bootstrap confidence intervals, and that a model conditioning on the decoupled $T$/$D$ exposures explains part of that difference. The strongest methodological competitor for the causal leg is Gorfine et al.'s 2SRI G×E estimator, which in 44,500 colorectal cancer cases and 52,235 controls recovered an MR log-odds ratio of 0.23 per 5-unit BMI increase versus 0.14 for naive observational regression, and which their simulations showed controls Type-I error far better than 2SPS. We run their exact 2SRI with a single coupled ambient-temperature exposure and European-derived instruments applied unchanged to all cohorts — the fairest version of "the best existing method, used as people actually use it." Beating it means lower false-positive ancestry-heterogeneity rates in the benchmark at matched sample sizes, and empirically, cross-ancestry agreement in $\beta_{GT}$ that the coupled version does not achieve. For the prediction-oriented sanity check, the baseline is X-Wing local portability, which delivered relative predictive $R^2$ gains of 14.1% to 119.1% across non-European cohorts (median 35.2% in East Asians across 31 traits). Finally, the key ablation runs our full estimator twice, once with the decoupled $(T, D)$ pair and once with the raw coupled thermal index, everything else identical; the difference in spurious heterogeneity isolates exactly the humidity-decoupling ingredient and nothing else.
Why not just…?
Why not just adjust for humidity and ethnicity as covariates in an observational regression? For the ICU stratification, we in fact do exactly that — observational adjustment is all the data supports, and we are honest that this leg cannot establish causation. But the broader answer is quantitative: observational regression estimated the BMI–colorectal-cancer log-odds ratio at 0.14 where instrumental variables gave 0.23 (Gorfine et al.), and observational associations between psychiatric disorders and glaucoma dissolved under two-sample MR — depression on glaucoma: OR 1.15, p = 0.20 in Europeans and OR 1.38, p = 0.30 in East Asians (Zhang et al.). Covariate adjustment also does not fix spatial overfitting, which SpaCE showed makes OLS and XGBoost look excellent on validation while failing at causal recovery. So we treat covariate adjustment as the descriptive layer, not the causal one.
Why not just use standard 2SPS? Because Gorfine et al. proved across thousands of simulations that 2SPS is structurally biased for binary outcomes due to the non-collapsibility of the odds ratio (the odds ratio shifts when you condition on covariates, even absent confounding), with poor Type-I error control, while 2SRI is consistently "less wrong." We adopt 2SRI wholesale; the proposal's novelty is the decoupled exposure pair and ancestry-specific instrument weighting, not the two-stage machinery.
Why not just do a European GWAS and adjust principal components? Because portability is local, not global: X-Wing showed that for some traits about 1% of the genome carries most of the cross-population covariance, so global PC correction cannot rescue transfer, and DisPred showed that linear models mistake European-specific variants for disease variants in admixed cohorts, with explicit ancestry/disease latent disentanglement yielding higher AUC in African and admixed groups. Our instrument weighting follows the local-portability logic rather than global adjustment.
Why not just skip the humidity decoupling and use raw ambient temperature? This is the closest real threat, because a reviewer could argue the VPD coupling is a second-order detail. We take it head-on rather than argue: the AmbControl work — the most careful chamber system in the file, holding setpoints within ±0.3 °C across +1.5 to −4 °C offsets — still observed RH and VPD co-varying with the temperature treatment inside its semi-open chambers (Sha Zhang et al.), and 30 AgMIP crop models diverged exactly where canopy transpiration and moisture demand had to be separated (Martre et al.). The benchmark's decisive experiment makes the coupled-vs-decoupled comparison directly: if spurious ancestry heterogeneity is the same under both, the reviewer is right and the decoupling earns nothing; our hypothesis is that it differs substantially. We pre-commit to reporting that result either way.
Proposed Experiments and Decisive Checks
The decisive checks, in order of informativeness. First, the simulation benchmark: generate data with exactly ancestry-invariant G×$T$ effects and geography-correlated $T$–$D$ coupling; the decoupled 2SRI must hold false-positive ancestry-heterogeneity rates near nominal (say, within 1.5× of 0.05) while the coupled 2SRI and 2SPS inflate them — if the coupled estimator is already near-nominal, the core claim fails and the proposal collapses to a negative result, which we would report. Second, the stratified ICU re-analysis on MIMIC-IV (43,537 patients) and eICU (75,184): report per-ancestry mortality nadirs with bootstrap intervals against the pooled 36.8/37.0 °C benchmark, plus per-ancestry temperature-gap slopes. Third, the two-sample application: per-ancestry $\beta_{GT}$ and $\beta_{GD}$ estimates with Steiger filtering, where the falsifiable prediction is $\beta_{GT}$ agreement across UKB and BBJ/CKB paired with $\beta_{GD}$ disagreement that tracks regional climate, echoing the cross-ancestry invariance logic of Zhang et al.. Two limitations must be stated plainly: the file lists no genome-wide genotyping for eICU or MIMIC-IV and no cross-ancestry GWAS of core body temperature itself — no reported result found on Lacuna for either — so the ICU leg is stratified-observational and the MR leg uses adiposity instruments, whose thermoregulatory relevance (insulation, metabolic rate) is indirect. Both gaps define exactly what a positive result would and would not license: it would license the claim that environmental decoupling changes cross-ancestry causal conclusions, not that the 37 °C optimum is itself genetically causal.
References
- The Hot Serial Cereal Experiment for modeling wheat response to temperature: field experiments and AgMIP-Wheat multi-model simulations
- Manipulating natural heatwaves in AmbControl chambers: bridging controlled and field conditions for warming studies in plant–soil systems
- Unveiling challenges in Mendelian randomization for gene–environment interaction
- Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics
- SpaCE: The Spatial Confounding Environment
- Association of body temperature and mortality in critically ill patients: an observational study using two large databases
- 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
- Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
- Improving genetic risk prediction across diverse population by disentangling ancestry representations