Bayesian Sensitivity Analysis for Causality
Also known as: Bayesian sensitivity analysis, Bayesian bias analysis, probabilistic sensitivity analysis for confounding, Bayesian unmeasured confounding analysis
Bayesian sensitivity analysis for causality quantifies how much an unmeasured confounder would need to influence both treatment assignment and outcome to overturn a causal conclusion. Rather than testing a single worst-case scenario, it places prior distributions over the strength of hidden confounding, propagates uncertainty through a full Bayesian model, and reports a posterior distribution for the causal effect that honestly reflects what is and is not identified from observed data.
Key highlights
- Propagates all sources of uncertainty — sampling and confounding — into a single posterior distribution for the causal effect.
- Allows domain knowledge about unmeasured confounders to be formally incorporated via informative priors, making the analysis more realistic than classical worst-case approaches.
- Produces calibrated credible intervals that honestly widen when confounding is plausible, supporting transparent communication of fragility.
- Flexible enough to accommodate complex outcome and treatment models, including survival outcomes, count data, and propensity-score-based designs.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Bayesian sensitivity analysis when you have an observational causal estimate — from difference-in-differences, propensity score matching, or regression — and want to quantify fragility to unmeasured confounding in a probabilistic framework rather than via a single worst-case scenario. It is especially valuable when domain expertise can inform meaningful priors on confounding parameters, and when communicating uncertainty to policymakers or clinicians who need full distributional summaries rather than point estimates. It is not appropriate when the data are from a well-randomised experiment with known assignment mechanism, or when no credible prior information on confounding is available and a simple Rosenbaum-style tipping-point analysis would suffice.
Strengths & limitations
- Propagates all sources of uncertainty — sampling and confounding — into a single posterior distribution for the causal effect.
- Allows domain knowledge about unmeasured confounders to be formally incorporated via informative priors, making the analysis more realistic than classical worst-case approaches.
- Produces calibrated credible intervals that honestly widen when confounding is plausible, supporting transparent communication of fragility.
- Flexible enough to accommodate complex outcome and treatment models, including survival outcomes, count data, and propensity-score-based designs.
- Requires specification of priors on unidentified confounding parameters; results can be sensitive to these choices and critics may question their validity.
- Computationally demanding: MCMC sampling for models with latent variables is slower than classical sensitivity analysis and may require expert tuning.
- Partial identification means the data alone cannot verify whether the priors are correct, demanding careful prior justification and reporting.
- Less familiar to reviewers in many applied disciplines, potentially reducing reproducibility and uptake compared with standard frequentist sensitivity analyses.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does this differ from Rosenbaum's classical sensitivity analysis?
Rosenbaum's approach asks how large a hidden bias parameter Gamma would need to be to invalidate the study, reporting a tipping point. Bayesian sensitivity analysis instead places a prior distribution over the confounding parameters and integrates them out, yielding a full posterior for the causal effect. The Bayesian approach quantifies uncertainty continuously rather than identifying a single threshold.
What priors should I use for confounding parameters?
Ideally, priors should reflect domain knowledge — for example, the plausible prevalence of an unmeasured risk factor in treated and untreated groups, and its known association with the outcome. When evidence is sparse, a range of priors should be explored and the sensitivity of the posterior to those choices should be reported explicitly.
Does a robust posterior mean the study is unconfounded?
No. A robust posterior means the causal conclusion holds across the range of confounding scenarios covered by your prior. If the prior does not include sufficiently strong confounding, robustness is an artifact of prior choice, not evidence of identification. Always justify why the prior range is substantively plausible.
Can I apply this method after propensity score matching?
Yes. McCandless et al. (2007) demonstrated Bayesian sensitivity analysis within a propensity-score framework. After matching on observed covariates, unmeasured confounders are modelled as latent variables with prior distributions, and the posterior for the treatment effect reflects residual confounding uncertainty.
What software can I use?
Stan and BUGS/JAGS are the most common platforms, allowing users to specify the extended outcome and treatment models with latent confounders. The R package 'sensitivitymv' covers classical analysis; dedicated Bayesian packages are found in the 'BayesSens' and 'EValue' ecosystem, alongside bespoke Stan code published in methodological papers.
Sources
- 1.McCandless, L. C., Gustafson, P., & Austin, P. C. (2007). Bayesian propensity score analysis for observational data. Statistics in Medicine, 26(8), 1704-1718.
- 2.Gustafson, P. (2015). Bayesian Inference for Partially Identified Models: Exploring the Limits of Limited Data. CRC Press / Chapman & Hall.ISBN 9781439869390
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Sensitivity Analysis for Causality. ScholarGate. https://scholargate.app/causal-inference/bayesian-sensitivity-analysis-for-causality