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

1. Problem Statement

In observational studies, unmeasured confounding remains a fundamental epistemological barrier to causal inference. Traditional sensitivity analysis methods, such as the Cornfield condition and the modern E-value, address this by calculating a threshold: the minimum strength of association an unmeasured confounder must have with both exposure and outcome to fully explain away (attenuate to null) an observed effect.

The Core Limitation: While these methods quantify how strong a confounder must be, they fail to answer the more critical inferential question: How plausible is it that such a confounder actually exists?

• Current practice relies on informal, context-dependent reasoning to judge whether an E-value is "large enough" to be credible.
• Existing probabilistic approaches (e.g., Probabilistic Bias Analysis, Bayesian hierarchical models) often require detailed individual-level data and complex modeling assumptions, limiting their applicability to summary-level statistics commonly reported in literature.

2. Methodology

The paper proposes a fully Bayesian reformulation of Cornfield-type sensitivity analysis that treats the strength of unmeasured confounding as a random variable rather than a fixed unknown.

A. Conceptual Framework

• Generative Model: The observed association ($RR_{obs}$) is modeled as the product of the true causal effect ($RR_{true}$) and a multiplicative confounding distortion factor ($\Gamma \geq 1$).
  $$RR_{obs} = RR_{true} \times \Gamma$$
• Log-Linear Transformation: To facilitate additive modeling, the analysis is conducted on the logarithmic scale:
  $$\theta_{obs} = \theta_{true} + \log(\Gamma) + \epsilon$$
  Where $\theta_{obs}$ is the observed log-effect, $\theta_{true}$ is the true log-effect, and $\epsilon$ represents sampling error (assumed Normal).

B. Prior Distributions

Since the model is not identifiable from data alone (true effect and bias cannot be separated without priors), the framework specifies:

1. True Causal Effect ($\theta_{true}$): Modeled as $N(0, \sigma_\tau^2)$, reflecting a prior belief that large effects are possible but not overwhelmingly likely.
2. Confounding Strength ($\log \Gamma$): Modeled as a Half-Normal distribution ($N(0, \sigma_c^2)$ truncated at 0).
   • This ensures $\Gamma \geq 1$.
   • The scale parameter $\sigma_c$ encodes prior beliefs about the plausibility of strong confounding (e.g., smaller $\sigma_c$ implies skepticism about extreme confounding).

C. Inferential Shift

Instead of reporting the threshold $\Gamma^*$ (the E-value) as a static number, the framework calculates the posterior probability that the confounding strength exceeds this threshold:
$$P(\Gamma \geq \Gamma^* \mid \theta_{obs})$$
This transforms the E-value from a deterministic boundary condition into a probabilistic statement about evidential vulnerability.

D. Data Source

The methodology was applied to a dataset of 11 exposure-outcome associations from four domains (smoking, back pain, Alzheimer's, environmental health) derived from a recent study by Xiang et al. (2026). The analysis utilized summary-level data (effect sizes and standard errors) to ensure broad applicability.

3. Key Contributions

1. Probabilistic Reinterpretation of the E-value: The paper redefines the E-value not as a final verdict on robustness, but as a boundary condition within a continuous probabilistic landscape. It quantifies the likelihood that confounding exceeds this boundary.
2. Unified Generative Model: It provides a simple, analytically tractable Bayesian model that links observed summary statistics to true causal effects and bias parameters without requiring individual-level data.
3. Decision-Relevant Metric: It introduces $P(\Gamma \geq \Gamma^*)$ as a direct measure of "evidential vulnerability," allowing researchers to make explicit probabilistic statements about the credibility of causal claims.
4. Bridging Methodological Gaps: It connects the simplicity of the E-value (threshold-based) with the rigor of Bayesian inference (probabilistic), offering a middle ground between heuristic reasoning and complex hierarchical modeling.

4. Results

The application of the framework to empirical case studies yielded the following findings:

• Inverse Relationship: A clear monotonic relationship exists between the magnitude of the E-value and the posterior probability of sufficient confounding.
  • High E-values (e.g., Environmental Health, $E=4.25$): Resulted in extremely low probabilities ($P \approx 0.004$) that a confounder strong enough to explain the effect exists. These associations are deemed highly robust.
  • Low E-values (e.g., Alzheimer's, $E \approx 1.3$): Resulted in high probabilities ($P \approx 0.5$), indicating that confounding of sufficient strength is plausible and the association is vulnerable.
• Granularity: The framework revealed heterogeneity within single studies. For example, within the smoking study, strong associations showed low vulnerability, while weaker contrasts (maternal/household smoking) showed moderate vulnerability ($P \approx 0.16–0.20$), a nuance lost in binary "robust/not robust" classifications.
• Prior Sensitivity:
  • Large E-values: Posterior probabilities remained stable across a wide range of prior specifications ($\sigma_c$), suggesting strong observed effects are robust to assumptions about confounding.
  • Small E-values: Results were highly sensitive to prior choices, reflecting the intrinsic uncertainty of weak effects where prior beliefs about confounding mechanisms heavily influence the conclusion.

5. Significance and Implications

• From Diagnostic to Inference: The work shifts sensitivity analysis from a descriptive diagnostic tool (checking if a threshold could be met) to a coherent inferential procedure (assessing the probability that it is met).
• Explicit Assumptions: By requiring explicit priors on confounding strength, the method forces researchers to articulate their domain knowledge regarding the plausibility of hidden biases, making the reasoning process transparent.
• Practical Applicability: Unlike complex Bayesian bias models requiring raw data, this approach works directly with summary statistics (E-values, effect sizes, SEs), making it immediately applicable to meta-analyses and systematic reviews.
• Philosophical Alignment: It aligns causal inference in observational studies with Bayesian principles, treating probability as a measure of rational belief under uncertainty rather than a long-run frequency, thereby providing a more nuanced framework for scientific judgment.

In conclusion, Costa's framework offers a rigorous, probabilistic alternative to the binary logic of traditional sensitivity analysis, enabling researchers to quantify the plausibility of confounding rather than just its theoretical possibility.