Bayesian Ex Post Facto Design — Bayesian Retrospective Causal Research
Bayesian Ex Post Facto Research Design · Also known as: Bayesian causal-comparative design, Bayesian after-the-fact design, Bayesian observational causal design, Bayesian retrospective causal study
Bayesian ex post facto design investigates possible causal relationships among variables that have already occurred, without researcher manipulation of those variables, and quantifies uncertainty about those relationships using Bayesian statistical inference. The researcher selects groups that differ on an outcome or a presumed cause after the fact, then uses prior knowledge and observed data together — via Bayes' theorem — to estimate credible effect sizes, group differences, or predictors.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian ex post facto design when: (1) random assignment is impossible or unethical — the independent variable is a pre-existing characteristic such as gender, socioeconomic status, diagnosis, or educational background; (2) you want to incorporate prior research quantitatively rather than ignore it; (3) small or moderate samples make frequentist power inadequate and Bayesian shrinkage or informative priors can stabilise estimates; and (4) credible intervals and probability statements about effect sizes are more useful to stakeholders than p-values. Do NOT use this design when: a randomised controlled experiment is feasible and the effect of interest can be isolated — the ex post facto element means causal claims remain tentative regardless of statistical framework; or when prior knowledge is genuinely absent and the Bayesian machinery adds complexity without interpretive gain.
Strengths & limitations
- Enables principled incorporation of prior research through formal prior distributions, accumulating knowledge across studies.
- Produces posterior distributions over effect sizes rather than binary reject/fail-to-reject decisions, giving richer uncertainty quantification.
- Works well with small to moderate samples when informative priors can offset limited data, unlike purely frequentist approaches.
- Credible intervals have a direct probabilistic interpretation — there is a 95% probability the parameter lies in the interval — which is more intuitive than frequentist confidence intervals.
- Sensitivity analyses over priors make hidden assumptions explicit and testable.
- Causal inference is inherently limited because variables were not randomly assigned; unmeasured confounders remain a threat to validity regardless of the Bayesian framework.
- Prior specification requires justification; poorly chosen or undisclosed priors can silently bias posterior estimates, especially with small samples.
- MCMC-based Bayesian analysis is computationally demanding and requires familiarity with probabilistic programming tools such as Stan, JAGS, or brms.
- Results — credible intervals, posterior distributions — can be unfamiliar to audiences trained only in frequentist statistics, increasing the communication burden.
Frequently asked
Is Bayesian ex post facto design the same as causal-comparative research?
They overlap substantially. Causal-comparative research (a term common in education research) is essentially synonymous with ex post facto design — both study pre-existing group differences without manipulation. The Bayesian label specifies the statistical framework used: instead of t-tests and ANOVA with p-values, the researcher uses prior distributions and posterior distributions. So Bayesian ex post facto design is causal-comparative research analysed with Bayesian methods.
Can I claim causation from a Bayesian ex post facto study?
No. The Bayesian framework improves uncertainty quantification but does not resolve the fundamental limitation of the ex post facto design: because the independent variable was not randomly assigned, unmeasured confounders may explain the observed differences. Strong causal language is not warranted. Instead, language such as 'groups differed' or 'the variable was associated with the outcome' is appropriate, with the posterior probability quantifying the strength of that association.
What software is typically used for Bayesian ex post facto analysis?
The most commonly used tools are Stan (via the RStan or CmdStan interfaces), JAGS (via rjags), and the brms package in R, which provides a high-level formula interface to Stan. Python users can use PyMC. For simpler models with conjugate priors, closed-form posteriors can be computed without MCMC. JASP, a free GUI-based statistical software, also supports Bayesian t-tests and ANOVAs suitable for two-group ex post facto comparisons.
How do I choose priors when I have little prior literature to draw on?
When prior evidence is sparse, weakly informative priors are preferred over completely flat (improper) priors. Weakly informative priors encode broad domain knowledge — for example, that standardised effect sizes in social science rarely exceed 1.5 in absolute value — without strongly pulling estimates toward a specific value. The brms and Stan documentation provides recommended weakly informative defaults. Always report what priors were used and run sensitivity checks with alternative reasonable priors.
How large a sample do I need for Bayesian ex post facto research?
There is no fixed minimum. Bayesian methods are more robust than frequentist approaches with small samples when informative priors are available, because prior information supplements the data. With very small samples (e.g., n < 20 per group) and vague priors, estimates will be highly uncertain — wide credible intervals are the honest result. Posterior predictive checks and prior sensitivity analyses help assess whether the sample is informative enough to update beliefs meaningfully.
Sources
- Kerlinger, F. N. (1973). Foundations of Behavioral Research (2nd ed.). Holt, Rinehart and Winston. link ↗
- Kruschke, J. K. (2015). Doing Bayesian Data Analysis: A Tutorial with R, JAGS, and Stan (2nd ed.). Academic Press. ISBN: 978-0124058880
How to cite this page
ScholarGate. (2026, June 3). Bayesian Ex Post Facto Research Design. ScholarGate. https://scholargate.app/en/research-design/bayesian-ex-post-facto-design
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian InferenceStatistics↔ compare
- Causal-Comparative ResearchResearch Design↔ compare
- Ex Post Facto DesignResearch Design↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Retrospective Cohort StudyEpidemiology↔ compare