Bayesian Negative Binomial Regression
Also known as: Bayesian NB regression, Bayesian negbin model, Bayesian overdispersed count regression, Bayesian NB-2 model
Bayesian Negative Binomial Regression models non-negative integer count outcomes that exhibit overdispersion — where the variance exceeds the mean — by placing a negative binomial likelihood on the data and specifying prior distributions over the regression coefficients and the dispersion parameter. Posterior inference is typically performed via Markov chain Monte Carlo (MCMC) or variational methods, yielding full posterior distributions rather than point estimates.
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 Negative Binomial Regression when the outcome is a count, overdispersion is present (conditional variance exceeds the conditional mean, detectable via a dispersion test or by comparing Poisson and NB fits), and you want full uncertainty quantification over all parameters including dispersion. It is especially valuable with small-to-moderate samples where likelihood-based asymptotic intervals are unreliable, when incorporating prior information from previous studies, or when you need to propagate uncertainty into downstream decisions. Do not use it when the outcome is truly Poisson-distributed (no overdispersion), when counts are bounded and better suited to a binomial model, or when the computational cost of MCMC is prohibitive and a frequentist NB GLM suffices.
Strengths & limitations
- Handles overdispersed counts directly through a dispersion parameter, outperforming Poisson when variance exceeds the mean.
- Provides full posterior distributions for all parameters, enabling exact finite-sample credible intervals rather than asymptotic approximations.
- Prior distributions enable principled regularisation, improving stability in small samples or with sparse predictors.
- Naturally propagates parameter uncertainty into predictions through the posterior predictive distribution.
- Model comparison via LOO-CV or WAIC is straightforward within the Bayesian framework.
- MCMC sampling is computationally expensive for large datasets or complex hierarchical extensions.
- Prior specification for the dispersion parameter phi requires care; vague priors can lead to poor mixing or unrealistic posteriors.
- Results depend on prior choices, which may be difficult to justify or communicate to non-Bayesian audiences.
- Convergence diagnostics add complexity compared to frequentist maximum likelihood fitting.
- Not suitable when zero-inflation is substantial — a Bayesian zero-inflated negative binomial is then preferable.
Frequently asked
How do I tell whether my count data need a negative binomial rather than Poisson model?
Fit both models and compare them with LOO-CV (in the Bayesian framework) or a likelihood ratio test. Also inspect the ratio of residual deviance to degrees of freedom in the Poisson fit — a ratio well above 1 suggests overdispersion. A posterior predictive check that shows the Poisson model consistently under-predicts the tails is a clear visual signal.
What prior should I use for the dispersion parameter phi?
A Half-Normal(0, 1) or Exponential(1) prior on the reciprocal of phi (sometimes called the shape parameter) is a widely recommended weakly informative default. Avoid Uniform(0, large) priors as they can place implausible mass on extreme values and cause slow MCMC mixing.
Can I add random effects (hierarchical structure) to a Bayesian NB model?
Yes. Bayesian NB regression extends naturally to hierarchical models by placing further priors on group-level coefficients. This is straightforward in brms or Stan and is one of the main advantages of the Bayesian approach over frequentist alternatives.
How do I interpret the regression coefficients?
Exponentiating a posterior coefficient exp(beta_j) gives the incidence rate ratio: the multiplicative change in the expected count for a one-unit increase in predictor j, holding other predictors constant. A credible interval for exp(beta_j) that excludes 1 indicates a credibly non-null effect.
What if many counts are zero?
If zeros are more frequent than a negative binomial distribution predicts, a Bayesian zero-inflated negative binomial (ZINB) model is more appropriate. Check with a posterior predictive plot comparing observed versus predicted zero counts.
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
- Cameron, A. C., & Trivedi, P. K. (2013). Regression Analysis of Count Data (2nd ed.). Cambridge University Press. ISBN: 978-1107667273
How to cite this page
ScholarGate. (2026, June 3). Bayesian Negative Binomial Regression. ScholarGate. https://scholargate.app/en/statistics/bayesian-negative-binomial-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 Poisson RegressionStatistics↔ compare
- Bayesian Zero-inflated modelStatistics↔ compare
- Negative Binomial RegressionEconometrics↔ compare
- Poisson RegressionEconometrics↔ compare
- Zero-inflated modelStatistics↔ compare