Mendelian randomization (MR) uses inherited genetic variants as natural experiments to ask whether a biomarker, such as body mass index (BMI) or a blood metabolite, causes a disease. It rests on the idea that genes are assigned at conception and so are shielded from lifestyle, income and environment. In biobanks with hundreds of thousands of people, that shield leaks. Subtle ancestry structure, assortative mating and batch effects reintroduce environmental confounding that is tiny in variance terms but large relative to the very small standard errors such samples produce. Bayesian MR methods were built to handle a different problem, "pleiotropy," where a variant affects the disease through some route other than the biomarker. These methods put priors on which variants are valid instruments. We argue that they also carry a hidden prior that is never stated: they assign zero probability to residual confounding shared across instruments. In large samples the data cannot identify that shared bias, so this unstated prior, not the likelihood, decides the answer.
We propose to make that prior explicit and to calibrate it against something observable. The proposed model adds a residual-confounding term to a Bayesian MR model with a sparse pleiotropy prior. The prior on the strength of that term is set as a multiple of how much the measured genetic principal components explain. We then report a prior robustness value: the smallest calibration multiple at which the posterior stops supporting a causal effect. Put plainly, this is a Bayesian MR analysis that states how much unmeasured ancestry-like confounding it is willing to believe in, ties that belief to the measured ancestry covariates, and reports when the conclusion breaks. Unlike existing frequentist sensitivity tools, it also shows how that judgment interacts with the analyst's choice of pleiotropy prior.
Core Claim
In biobank-scale MR, residual environmental confounding that acts coherently across many instruments creates a flat ridge in the likelihood. Along this ridge, the causal effect and a shared bias term trade off almost one for one. On the Geometry of Bayesian Graphical Models with Hidden Variables shows algebraically that hidden variables in graphical models produce exactly such flat ridges and boundary maxima. On these ridges the prior acts as the tie-breaker. That paper motivates our view but only treats small discrete networks. The flat ridge we target is continuous and multi-instrument.
Current Bayesian MR methods implicitly fix the shared bias at zero. Their posteriors are therefore narrow for a reason that has nothing to do with the data. Our first prediction is that adding a confounding term whose prior scale is benchmarked to the observed principal components will restore honest uncertainty, so posterior credible intervals should achieve close to nominal coverage under simulated residual stratification. Our second prediction is that it should flag fragile findings that pleiotropy priors alone miss.
The second, more specific claim concerns interaction. The pleiotropy prior and the confounding prior are not separable. A broad slab, meaning the continuous part of a spike-and-slab prior that allows invalid variants, can absorb part of a coherent bias as "many invalid instruments." A narrow slab forces the same bias into the causal estimate. A sensitivity analysis that perturbs only one of the two will misjudge robustness. A purely frequentist robustness bound, which has no pleiotropy prior at all, cannot represent this interaction.
Research Program
The model starts from the individual-level linear instrumental-variable setup used by MR-SENSEMAKR. An exposure $X$ depends on genotypes $G$ through effects $\gamma$. The outcome $Y$ depends on $X$ through the causal effect $\beta$, on $G$ directly through pleiotropic effects $\alpha$, and on an unmeasured residual confounder $U$ that remains after adjusting for the top principal components $P$:
$$Y = \beta X + G\alpha + \lambda U + P\eta + \varepsilon .$$
Here $\lambda$ is the confounder's effect on the outcome, $\eta$ the principal-component effects, and $\varepsilon$ noise. $U$ is also allowed to correlate with $G$, which is how stratification breaks instrument independence.
Following Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy, we parameterize $U$ by two partial $R^2$ values: its association with the instruments and its association with the outcome. That paper supplies the parameterization, the omitted-variable-bias algebra and the idea of benchmarking against principal components. It stops at deterministic contours and robustness values, with no posterior and no pleiotropy model.
We place on both partial $R^2$ values a prior of the form $k \cdot R^2_{\text{PC}}$, where $R^2_{\text{PC}}$ is what the top 20 principal components explain. The multiple $k$ receives a log-normal prior whose scale is the single interpretable hyperparameter, read as "how many times stronger than measured ancestry could the unmeasured remainder plausibly be." On $\alpha$ we use the spike-and-slab prior of MASSIVE: Tractable and Robust Bayesian Learning of Many-Dimensional Instrumental Variable Models. In that prior, a point mass at zero means a valid instrument and a broad slab means a pleiotropic one, explored by model averaging with Laplace-approximated evidence. We reuse its machinery and add the confounding block, which enters the evidence computation through an inflation of the residual covariance.
For each analysis we report the posterior for $\beta$. We also report the prior robustness value $k^*$, the smallest benchmark multiple at which the posterior probability of the reported effect sign drops below 0.975. Finally, we report a joint sensitivity surface over $k$, the prior validity probability $\pi$, and the slab variance. We compute these by importance-reweighting one set of posterior samples drawn under a deliberately broad reference prior, and fall back to reruns where the effective sample size collapses.
For screening-scale use on summary statistics, as in the 130-metabolite by 93-trait matrix, we build a reduced version. There, each variant's ancestry loading from a reference panel serves as a measured stratification score, and the confounding term becomes a shared coefficient on that score with the same benchmarked prior.
The evidence-based-prior idea comes from Bayesian Sensitivity of Causal Inference Estimators under Evidence-Based Priors. That paper replaces worst-case sensitivity with expected sensitivity under informative priors on assumption violations. It shows worst-case values drift to uninformative bounds as covariate dimension grows. It was built for treatment-control unconfoundedness and uses rejection Monte Carlo. We move the idea to instrumental variables, anchor the prior to measured principal components rather than to loose elicitation, and replace rejection sampling with reweighting.
The Bayesian Analysis of Complex, High-Dimensional Models: Can It Be CODA? acts as a constraint. It shows that "uninformative" priors in high dimensions can behave badly, and that good frequentist behavior requires deliberately constructed priors. This is why we calibrate $k$ against an observable rather than defaulting to a flat prior on $R^2$.
Baselines
The naive reference point is the inverse-variance-weighted ratio estimate with its standard confidence interval. This estimate ignores both pleiotropy and residual confounding, and it is the result most published two-sample MR studies report. We run it on exactly the same instruments and covariate adjustment. Beating it means correct false-positive rates where it fails. No reported result for this estimator on these datasets was found on Lacuna, so we compute it ourselves.
The strongest frequentist competitor is MR-SENSEMAKR, because it answers the closest question: how strong residual confounding must be to overturn a result. On the 291,274-person UK Biobank subset, it reports that BMI on diastolic blood pressure would require a confounder 14 times stronger than the top 20 principal components to explain away. It reports that BMI on Townsend deprivation could be explained away by a confounder half as strong as those components. In UK Biobank-like simulations with slight pleiotropy, it reports false-positive rates of 46% to 80% for MR-Egger and MR-Mix. We run it on identical data and principal components. A fair comparison requires our $k^*$ to rank findings in broadly the same order on cases with no pleiotropy, and to differ only where the pleiotropy prior matters. Beating it means one of two things. Either the posterior intervals have better calibrated coverage, or we find cases where accounting for pleiotropy uncertainty changes a robustness verdict, and simulation ground truth shows our verdict is the correct one.
The strongest Bayesian competitors are MASSIVE and BWMR. Bayesian weighted Mendelian randomization for causal inference based on summary statistics down-weights outlying variants with continuous weights fitted by variational expectation-maximization and corrects the variational variance. It reports a stable Type I error with 50% to 80% invalid instruments, without a numeric rate in our file, and 46 significant metabolite-to-trait effects. MASSIVE reports BMI raising psoriasis risk by about 9% per unit. Both are run with their default hyperpriors, using the BWMR code, and also with a matched hyperparameter tuning budget. Beating them means lower false-positive rates and nominal coverage under simulated residual stratification, at equal computation.
Two ablations isolate the key ingredient. The first keeps the confounding term but gives $k$ an uncalibrated flat prior on $R^2$, which tests whether benchmarking to principal components is what helps. The second keeps the benchmarked confounding prior but drops the spike-and-slab, using a single Gaussian pleiotropy term. This tests the interaction claim.
Why not just…?
Why not just run MR-SENSEMAKR next to MASSIVE and report both? This is the real threat. In data with little pleiotropy the two analyses should agree, and our method would add mostly a probabilistic summary. The proposal earns its place only if the interaction claim holds. A broad versus narrow slab must change how much coherent bias is absorbed, so that reporting the two analyses separately mislabels some findings as robust or fragile. The simulation arm with known pleiotropy and stratification settles this directly. If side-by-side reporting matches our joint analysis in ranking and coverage, we will say so and recommend the simpler practice.
Why not just rerun the Bayesian method over a grid of hyperparameters? For three hyperparameters on one exposure-outcome pair, a grid is perfectly adequate, and we use it as the gold standard against which reweighting is validated. A grid does not supply the missing ingredient, though: no grid over $\pi$ and slab variance places any prior mass on shared residual confounding. A grid also becomes costly over the roughly twelve thousand pairs of the metabolite matrix. Speed is a convenience; the contribution is the calibrated confounding prior.
Why not just use a heavier-tailed pleiotropy prior such as a horseshoe? Heavy tails help when a few variants are wildly invalid. Residual stratification is different: it shifts many variants in the same direction in proportion to their ancestry differentiation, which looks like a causal effect rather than like outliers. We test this by including a horseshoe variant of the pleiotropy prior in the stratification simulations. We expect its false-positive rate to stay close to MASSIVE's. If it instead matches our method, the confounding term is unnecessary and we will report that.
Proposed Experiments
The decisive experiment is a simulation built on UK Biobank genotype structure, mirroring the MR-SENSEMAKR design. It injects residual confounding with strength $k_{\text{true}}$ from 0.1 to 20 times the top-20 principal components. It crosses this with 0%, 20% and 50% pleiotropic instruments and with a true null or a modest true $\beta$. Across replicates we measure false-positive rate at the null, coverage of 95% intervals, and interval width. The headline check is that our method keeps false-positive rates near nominal whenever $k_{\text{true}}$ lies within its prior's bulk, while MASSIVE and BWMR do not. A second check is whether the joint sensitivity surface correctly predicts the slab variances at which the competitors fail.
On real data, the UK Biobank pairs BMI on diastolic blood pressure and BMI on Townsend deprivation act as a positive control and a fragility control. A credible method must report a high $k^*$ for the first and a low $k^*$ for the second, consistent with the 14-times and half-strength benchmarks. We will treat a reversal as a failure. BMI on psoriasis tests whether calibrated confounding uncertainty leaves MASSIVE's roughly 9% per unit estimate intact. For the summary-data variant, we re-screen the BWMR metabolite matrix and report how many of the 46 reported effects survive at a prespecified prior scale. We will check the survivors against the Alzheimer's result, where BWMR found no metabolite with a causal effect, as a sanity check that the method does not invent effects. The pooled UK Biobank and GECCO colorectal cancer data serve as a stress test with a binary outcome. Using a probit link, we ask whether the confounding prior interacts with the non-collapsibility problems documented in Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction, where the MR log-odds ratio was about 0.23 per 5 BMI units against 0.14 observationally.
Individual-level UK Biobank access is required for the headline arm. If access is delayed, the summary-data variant with reference-panel ancestry loadings is the fallback, and we will state its weaker identification.
References
- On the Geometry of Bayesian Graphical Models with Hidden Variables
- Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy
- MASSIVE: Tractable and Robust Bayesian Learning of Many-Dimensional Instrumental Variable Models
- Bayesian Sensitivity of Causal Inference Estimators under Evidence-Based Priors
- The Bayesian Analysis of Complex, High-Dimensional Models: Can It Be CODA?
- Bayesian weighted Mendelian randomization for causal inference based on summary statistics
- Unveiling challenges in Mendelian randomization for gene–environment interaction