##  [Posterior Predictive Distribution](/posterior-predictive-distribution-0) 

 Definition

The probability distribution of future or unobserved observables obtained by averaging the model likelihood for new data over the posterior distribution of parameters conditioned on observed data.

 

 

 

 

 

 





## Principle

Principle

Propagate parameter uncertainty into predictions by integrating the likelihood for new observables against the posterior measure, yielding predictive probabilities that account for both data and parameter uncertainty.

 

 

 

 

 





## Demonstration

Demonstration

In a count model with Poisson likelihood p(y|θ) = Poisson(θ) and a conjugate posterior for θ from observed counts, the posterior predictive for a new count y* is obtained by integrating Poisson(y*|θ) against the posterior density of θ, giving a closed-form mixture distribution (e.g., negative binomial in common conjugate cases).

 

 

 

 

## Misapplication

Misapplication

Using the posterior predictive distribution as the sole measure of model adequacy without checking predictive calibration or using it to compare models across different likelihood specifications without accounting for overfitting.

 

 

 

 

 





## Consequence

Consequence

Provides a coherent tool for forecasting, uncertainty quantification for new observations, and principled model checking via posterior predictive checks and predictive intervals.

 

 

 

 

## Reversal

Reversal

The prior predictive distribution integrates the likelihood against the prior rather than the posterior and is used to assess implications of prior assumptions before seeing data.

 

 

 

 

 





## Boundary

Boundary

Requires a specified likelihood model and a (computable) posterior distribution; ill-defined when the posterior is improper, non-normalizable, or when the likelihood for new data cannot be evaluated or integrated over the posterior.

 

 

 

 

 





## Semantic Tension

Semantic Tension

Tension between posterior predictive distribution and cross-validation predictive estimates: the former conditions on the entire data set and averages over parameters, while the latter estimates out-of-sample performance by withholding data and may better diagnose overfitting.

 

 

 

 

 





## Synthesis

Synthesis

The posterior predictive distribution is the predictive law for future observables produced by averaging the model's sampling distribution over posterior uncertainty in parameters, yielding forecasts and diagnostics that incorporate data-driven parameter learning.