A Bayesian Reinterpretation of Cornfield-Type Sensitivity Analysis: From Thresholds to Probabilities

Sensitivity analysis for unmeasured confounding in observational studies is commonly based on threshold quantities, such as the Cornfield condition or the E-value, which quantify how strong a confounder must be to explain away an observed association. However, these approaches do not address a fundamental inferential question: how plausible is it that such a confounder exists? In this work, we propose a Bayesian reformulation of Cornfield-type sensitivity analysis in which the strength of unmeasured confounding is treated as a random variable. Within this framework, the E-value is reinterpreted as a threshold, and the central inferential quantity becomes the posterior probability that confounding exceeds this threshold. This transforms sensitivity analysis from a descriptive diagnostic into a probabilistic assessment of robustness. We develop a simple generative model linking observed effect estimates to true causal effects and confounding bias, and we specify prior distributions reflecting plausible confounding mechanisms. The resulting framework yields posterior measures of evidential vulnerability that are directly interpretable and applicable to summary-level data. Illustrations based on empirical case studies show that the proposed approach preserves the interpretability of the E-value while providing a more nuanced and decision-relevant characterization of robustness. More broadly, the framework aligns sensitivity analysis with Bayesian principles of inference under uncertainty, offering a coherent alternative to purely threshold-based reasoning.

💡 Research Summary

This paper presents a fully Bayesian reinterpretation of Cornfield‑type sensitivity analysis for unmeasured confounding in observational studies. Traditional approaches such as the Cornfield inequality and the more recent E‑value provide a deterministic threshold: the minimum strength of an unmeasured confounder required to completely explain away an observed association. While useful for quantifying how strong a confounder would need to be, these methods do not address the inferential question of how plausible it is that such a confounder actually exists.

The authors propose to treat the strength of unmeasured confounding, denoted Γ (≥ 1), as a random variable and to embed it within a simple generative model linking the observed effect estimate to the true causal effect. On the log‑scale, the observed log‑risk ratio θ̂ is modeled as
θ̂ ∼ Normal(θ₀ + log Γ, s²),
where θ₀ is the true log‑risk ratio and s is the reported standard error. Because the model is not identifiable from the data alone, prior distributions are placed on both θ₀ and log Γ. The causal effect receives a weakly informative Normal(0, σ*²) prior, reflecting the belief that large effects are possible but not overwhelmingly likely. The confounding strength receives a Half‑Normal(0, σ⁺) prior on log Γ, guaranteeing Γ ≥ 1 while allowing the analyst to encode skepticism about strong confounding through the scale σ⁺ (e.g., σ⁺ = 0.5 makes Γ > 2–3 increasingly unlikely).

Within this framework, the E‑value is re‑interpreted as the deterministic threshold Γ* that would reduce the observed effect to the null. The central inferential quantity becomes the posterior probability
P(Γ ≥ Γ* | θ̂),
which directly answers “what is the probability that unmeasured confounding is strong enough to explain away the observed association?” This probability is derived from the joint posterior p(θ₀, Γ | θ̂) obtained via Bayes’ theorem and can be computed using standard MCMC tools (e.g., Stan).

To demonstrate the method, the authors use a publicly available dataset compiled by Xiang et al. (2026), which contains summary‑level effect estimates (risk ratios, odds ratios, hazard ratios) and their standard errors from a variety of epidemiological domains. They select 11 exposure‑outcome pairs spanning smoking, back pain, Alzheimer’s disease, and environmental health. For each case they compute the E‑value, fit the Bayesian model with a moderate prior (σ⁺ = 0.5), and evaluate P(Γ ≥ Γ*). The results show a clear monotonic relationship: larger E‑values correspond to smaller posterior probabilities of sufficient confounding. For example, an environmental health study with E = 4.25 yields P ≈ 0.004, indicating that confounding strong enough to nullify the finding is highly implausible. In contrast, Alzheimer’s disease associations with E‑values around 1.3–1.4 produce P ≈ 0.5, suggesting that confounding of the required magnitude is quite plausible under the chosen prior. Even within a single study, different exposure contrasts produce markedly different probabilities (e.g., smoking intensity yields P ≈ 0.16–0.20), illustrating the method’s ability to provide nuanced, case‑specific assessments rather than a uniform qualitative judgment.

The authors discuss several advantages of their approach: (1) it operates on summary statistics, making it applicable when individual‑level data are unavailable; (2) it makes prior scientific knowledge about plausible confounding explicit, increasing transparency; (3) the posterior probability is an intuitive metric that can be directly incorporated into policy or clinical decision‑making. They also acknowledge limitations: the single‑parameter Γ simplification may not capture complex confounding structures; results are sensitive to the choice of σ⁺, requiring careful prior elicitation or sensitivity analysis; and the model assumes that the reported adjusted estimate is unbiased conditional on measured covariates, which may not hold in all settings.

Future work suggested includes extending the framework to multiple confounders, developing systematic prior‑sensitivity procedures, and integrating the Bayesian sensitivity analysis with causal‑graphical models or hierarchical meta‑analytic structures.

In conclusion, by converting the deterministic Cornfield/E‑value threshold into a probabilistic statement about the plausibility of sufficient unmeasured confounding, the paper provides a coherent Bayesian alternative that enriches the interpretability of sensitivity analyses and aligns them with the broader principles of Bayesian inference under uncertainty.