Bayesian Probit Model
Bayesian Probit Regression Model · Also known as: Bayesian probit regression, probit model with data augmentation, Gibbs sampling probit, Albert-Chib probit
The Bayesian Probit model is a binary regression method that models the probability of a binary outcome using the normal CDF (probit link) within a Bayesian framework. It assigns prior distributions to regression coefficients and updates them with observed data, yielding a full posterior distribution rather than a single point estimate. The Albert-Chib data-augmentation algorithm makes posterior sampling computationally efficient via Gibbs sampling.
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 the Bayesian Probit model when the outcome is binary and you want full uncertainty quantification via posterior distributions rather than point estimates alone — especially valuable in small samples where likelihood-based MLE is unstable, when informative prior knowledge is available, or when you need predictions with calibrated uncertainty bands. It is also preferred over Bayesian logistic regression in some neuroscience and psychometric applications where the probit link better matches the underlying latent-variable theory. Do not use it if you need fast, deterministic computation on very large datasets (the MCMC sampler is slower than MLE), if the outcome has more than two ordered categories (use Bayesian ordinal probit), or if computational resources are limited and a frequentist probit gives adequate coverage.
Strengths & limitations
- Provides a full posterior distribution over coefficients, yielding credible intervals and predictive distributions rather than just point estimates.
- The Albert-Chib data-augmentation Gibbs sampler is exact and computationally efficient, avoiding numerical integration.
- Informative priors stabilise estimates in small samples or with sparse events, reducing the separation/complete-separation problem common in standard probit MLE.
- Naturally propagates parameter uncertainty into predictions, producing well-calibrated probability forecasts.
- Extends cleanly to hierarchical and multilevel specifications via hyperpriors on beta.
- MCMC sampling is slower than MLE for large datasets; convergence must be monitored (Rhat, trace plots).
- Results depend on prior choice; poorly specified priors can distort inference, particularly in small samples.
- The probit link produces coefficients on the latent-variable scale, which are less directly interpretable as odds ratios compared to logistic regression.
- Requires more statistical expertise to implement and diagnose than standard frequentist probit.
Frequently asked
What is the difference between Bayesian probit and Bayesian logistic regression?
Both model binary outcomes but use different link functions: probit uses the normal CDF (Phi), logistic uses the logistic function. The probit link has a natural latent-variable interpretation (a hidden Gaussian score), which is theoretically motivated in psychometrics and signal detection. In practice the two models produce very similar predicted probabilities; logistic coefficients can be converted to odds ratios, which probit coefficients cannot.
How do I choose the prior for beta?
A common default is a weakly informative normal prior, such as N(0, 2.5) on each standardised coefficient, following recommendations in the Stan and rstanarm literature. This shrinks extreme estimates without strongly dominating the data. Use informative priors only when you have genuine prior knowledge and document the sensitivity of results to prior choice.
How many MCMC iterations do I need?
There is no universal answer, but a common starting point is 2000 warm-up and 2000 sampling iterations across 4 chains. Monitor convergence with Rhat (should be below 1.01) and effective sample size (ESS should exceed 400 per parameter). Increase iterations if chains have not converged.
Can I use Bayesian probit with complete separation?
Yes — this is one of its main advantages over MLE probit. When a predictor perfectly separates the two outcome classes, MLE coefficients diverge to infinity, but a weakly informative prior keeps the posterior concentrated at a finite, interpretable value.
Is Bayesian probit suitable for large datasets?
It can be, but MCMC becomes slow when n is very large. Variational Bayes approximations or Laplace approximations (as in INLA or rstanarm's optimisation mode) can provide fast approximate posteriors. Alternatively, modern samplers such as Stan's NUTS handle moderately large datasets efficiently.
Sources
- Albert, J. H., & Chib, S. (1993). Bayesian analysis of binary and polychotomous response data. Journal of the American Statistical Association, 88(422), 669-679. DOI: 10.1080/01621459.1993.10476321 ↗
- 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
How to cite this page
ScholarGate. (2026, June 3). Bayesian Probit Regression Model. ScholarGate. https://scholargate.app/en/statistics/bayesian-probit-model
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 Logistic RegressionBayesian↔ compare
- Bayesian Multinomial Logistic RegressionStatistics↔ compare
- Bayesian Ordinal Logistic RegressionStatistics↔ compare
- Logistic RegressionResearch Statistics↔ compare
- Probit ModelEconometrics↔ compare