What is Posterior Predictive Distribution?
A Posterior Predictive Distribution is the probability distribution of a new or replicated outcome after integrating its conditional likelihood over the posterior distribution of unknown model quantities.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Marginalize unknown quantities on the outcome scale
Draw or integrate plausible parameter values from the posterior, then generate or evaluate the outcome conditional on each value. Averaging p(y*|x*,theta) over those values forms the posterior predictive mixture. A plug-in prediction p(y*|x*,theta_hat) omits parameter uncertainty, while a latent-function distribution can omit observation noise. Preserve these identities in APIs and stored artifacts instead of labeling every interval as confidence.
Use replicated data to criticize the model
The current Stan User's Guide defines posterior predictive checks by simulating replicated data and comparing decision-relevant statistics with observed data. Check location, dispersion, tails, zero counts, dependence, and important slices chosen for the domain. A model can reproduce the mean while missing variance or extremes. Posterior predictive p-value-like summaries are diagnostic and are not generally uniform frequentist p-values.
Keep model checks separate from future performance
PyMC's predictive workflow distinguishes replications at observed predictors from predictions at new inputs. Reusing observed data helps reveal model-data conflicts, but it is not an untouched generalization test. Evaluate future predictions on a time-, entity-, or population-appropriate holdout, report proper scores and coverage, quantify Monte Carlo error, and rerun checks after prior, likelihood, feature, or population changes.
Key Characteristics
- Conditions on observed data and averages over posterior uncertainty
- Produces a distribution over outcomes rather than only model parameters
- Can combine parameter uncertainty with modeled observation variability
- Supports prediction, interval construction, and model criticism
- Depends on prior, likelihood, inference, input, and population assumptions
- Requires held-out evaluation in addition to posterior predictive checks
Common Use Cases
- Generating probability distributions for future counts or continuous outcomes
- Constructing posterior prediction intervals for decision planning
- Checking whether replicated data reproduce tails, dispersion, or dependence
- Comparing a Bayesian model with a plug-in point-estimate predictor
- Propagating parameter uncertainty into downstream expected-cost calculations
Example
Loading code...Frequently Asked Questions
What does a Posterior Predictive Distribution represent?
It represents possible future or replicated outcomes after averaging their likelihood over uncertainty in model quantities under the posterior. It is an outcome distribution conditioned on observed data, not merely a distribution over parameters.
How is a posterior distribution different from a Posterior Predictive Distribution?
The posterior describes unknown parameters or latent quantities after observing data. The posterior predictive pushes that posterior through the likelihood to describe outcomes. This can add observation variability to parameter uncertainty.
What is the difference between a credible interval and a prediction interval?
A credible interval can summarize a posterior parameter or latent function. A posterior prediction interval summarizes a future observed outcome and usually includes both uncertainty about the latent mean and modeled observation noise, so it is often wider.
Does a posterior predictive check prove that a model is correct?
No. It can reveal that the model fails to reproduce selected features, but many wrong models can reproduce the same summaries. Use multiple domain-relevant checks, computational diagnostics, sensitivity analysis, and untouched predictive evaluation.
Can posterior predictive samples be used as independent test data?
No. They are generated from the fitted model and inherit its assumptions. They support model criticism and decision simulation, while real generalization evidence requires observations withheld by an appropriate time, entity, or population split.