STA702
Duke University
Regression Model (Sampling model)
Semi-Conjugate Prior Independent Normal Gamma
Regression Model (Sampling model)
Conjugate Normal-Gamma Model: factor joint prior
Normal-Gamma distribution indexed by 4 hyperparameters
Note Prior Covariance for
Likelihood:
Quadratic in Exponential
Posterior density (up to normalizing contants)
Marginal
Conditional Normal for
No need for Gibbs sampling!
Hyperparameters:
Inner product induced by prior precision
Then
Note - true for prior or posterior given
Marginal density
Let
The marginal posterior distribution of
Any linear combination
use for individual
Suppose
What is the predictive distribution of
Use the representation that
need to specify Normal prior mean
need to specify Gamma shape (
hard in higher dimensions!
default choices?
Jeffreys prior is invariant to model parameterization of
log likelihood expressed as function of sufficient statistics
Jeffreys’ Prior (don’t use!)
Use Independent Jeffreys Prior
Formal Posterior Distribution
Bayesian Credible Sets
conditional on
the model in vector form
What if we transform the mean
obtain the posterior for
since
plus the posterior of
Exercise for the Energetic Student
With some linear algebra, show that this is true for a normal prior if
Popular choice is to take
Full conditional
one parameter
if
Conjugate so we could skip Gibbs sampling and sample directly from gamma and then conditional normal!
If
Ridge regression protects against the explosion of variances and ill-conditioning with the conjugate priors:
Posterior for
Posterior mean (or mode) given
related to penalized maximum likelihood estimation
Choice of
Bayes Regression and choice of
Be sure that you can derive the full conditional posteriors for