search

JuliaStats / Distributions.jl

A Julia package for probability distributions and associated functions.
Other
1.11k stars 415 forks source link

Posterior-predictive distributions #145

Closed nfoti closed 9 years ago

nfoti commented 11 years ago

The interface for dealing with posterior distributions for Bayesian computations should include computing posterior-predictive distributions, that is the distribution of new observations conditioned on previous observations. These have a variety of uses, for example:

  1. Model checking (i.e. posteror-predictive checks).
  2. Computing the probability of a new cluster in a Dirichlet Process mixture model.

I'm filing this issue as a reminder to myself to implement this when I have time, but if anyone else wants to take a stab at it that would be great.

lindahua commented 11 years ago

Posterior-predictive distribution is good to have.

However, I think it is better to explore Bayesian nonparametric models (such as DPMM) in a separate package.

johnmyleswhite commented 11 years ago

Is it better to have specialized functions for this or to have both posterior and predict functions? Is the use case that it's more efficient to construct predictions without constructing the full posterior?

nfoti commented 11 years ago

@lindahua, I totally agree that addressing Bayesian nonparametrics is best done in another package, I was just using DPMMs as an example of where posterior-predictive distributions can arise.

@johnmyleswhite, in a conjugate setting the posterior-predictive can be analytically computed so there is no need to ever construct the full posterior. So as you said it's more efficient to just evaluate (or sample from) the posterior-predictive.

nfoti commented 11 years ago

I've started implementing this on the postpred branch of my fork.

lindahua commented 9 years ago

Should be moved to ConjugatePriors package.