Bayesian Poisson Regression
Also known as: Bayesian log-linear count model, Bayesian GLM Poisson, Poisson regression with priors, Bayesian count regression
Bayesian Poisson regression models non-negative integer count outcomes using a Poisson likelihood with a log link, placing prior distributions on the regression coefficients. Posterior inference — combining prior beliefs with the data likelihood — produces full probability distributions over the coefficients rather than single-point estimates, enabling coherent uncertainty quantification and incorporation of domain knowledge.
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 Poisson regression when the outcome is a non-negative integer count (events per time period, cases per region, occurrences per subject) and you want full posterior uncertainty rather than a point estimate plus asymptotic confidence interval. It is especially valuable when samples are small or sparse (where frequentist maximum-likelihood estimates can be unstable), when informative prior knowledge from prior studies or theory is available, or when model comparison via information criteria such as WAIC or LOO-CV is needed. Avoid it — or prefer the Negative Binomial variant — when the count data are overdispersed (variance substantially exceeds the mean), as the Poisson assumption of equidispersion will inflate posterior confidence and produce poor predictive checks. Also avoid when the outcome is continuous or when computational cost of MCMC is prohibitive for a very large dataset and a simpler GLM suffices.
Strengths & limitations
- Produces full posterior distributions over coefficients, enabling genuine probabilistic statements about parameter values and predictions.
- Naturally regularises estimates through priors, reducing overfitting in small or high-dimensional data settings.
- Straightforward incorporation of external information (prior studies, expert elicitation) through informative priors.
- Rate ratios (exp(beta)) are directly interpretable as multiplicative effects on the expected count.
- Principled model comparison using WAIC and leave-one-out cross-validation (LOO-CV).
- Posterior predictive checks provide a transparent, visual tool for diagnosing model misfit.
- Computationally intensive: MCMC sampling can be slow for large datasets or complex hierarchical structures.
- Requires careful prior specification; poorly chosen priors can dominate the likelihood when sample sizes are small.
- The Poisson equidispersion assumption (mean = variance) is often violated in real count data; overdispersion produces unreliable posterior inference.
- Results and convergence diagnostics (R-hat, effective sample size) require more statistical expertise to interpret than a standard GLM output.
- Sensitive to misspecification of the likelihood family if the generating process deviates substantially from Poisson.
Frequently asked
What prior should I use for the regression coefficients?
A common weakly-informative choice is Normal(0, 2.5) on standardised predictors (recommended by Gelman et al. and the rstanarm defaults). This keeps coefficients within plausible ranges without overwhelming the likelihood. For rate-ratio interpretability, check that exp(beta) stays within a reasonable range given the domain.
How do I tell if my data are overdispersed?
After fitting the model, run a posterior predictive check: simulate replicated datasets from the posterior and compare the variance-to-mean ratio of the simulated counts to the observed data. If simulated variance is systematically lower than observed variance, the Poisson assumption is too restrictive and Bayesian Negative Binomial regression is a better choice.
When should I add an offset term?
Add an offset when the exposure or population at risk differs across observations — for example, if one region has 10 times the population of another, the raw count is not comparable without adjusting for population size. The offset log(exposure_i) is added to the linear predictor with a coefficient fixed at 1.
How does Bayesian Poisson regression compare to frequentist Poisson GLM?
Both share the same likelihood and log link. The Bayesian version adds priors and returns full posterior distributions rather than point estimates and asymptotic confidence intervals. Bayesian credible intervals have a direct probability interpretation, whereas frequentist confidence intervals do not. In large samples the two agree closely; differences are most pronounced in small samples or with informative priors.
Can I use this model for zero-inflated counts?
Standard Bayesian Poisson regression does not separately model an excess of zeros. If your data have more zeros than a Poisson distribution predicts, use a Bayesian zero-inflated Poisson or hurdle model instead, which explicitly accounts for a structural zero-generating process.
Sources
- Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd ed.). CRC Press. ISBN: 978-1439840955
- McCullagh, P., & Nelder, J. A. (1989). Generalized Linear Models (2nd ed.). Chapman and Hall. ISBN: 978-0412317606
How to cite this page
ScholarGate. (2026, June 3). Bayesian Poisson Regression. ScholarGate. https://scholargate.app/en/statistics/bayesian-poisson-regression
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 Generalized Linear ModelStatistics↔ compare
- Bayesian Multiple linear regressionStatistics↔ compare
- Bayesian Negative Binomial RegressionStatistics↔ compare
- Negative Binomial RegressionEconometrics↔ compare
- Poisson RegressionEconometrics↔ compare
- Zero-inflated modelStatistics↔ compare