What exactly is the difference between Bayesian Model Averaging (BMA) and the Stochastic Search Variable Selection (SSVS) prior when we talk about linear regression models?

The SSVS prior is used for variable selection, which is conceptually different from model averaging, right? However, both result in coefficient estimates and posterior inclusion probabilities and are often mentioned together in literature (see for instance the article on BMA linked above).

Can anyone provide more systematic insights regarding the relationship of SSVS and BMA? Where do they differ? Where are similarities to be found? Do they generally give similar results? Is the model space they explore the same?



SSVS samples from the higher dimensional posterior of all parameters and models. You don’t need to sample models to do BMA, though—you can fit each of the many models separately, and then use their marginal likelihoods to compute a quantity you’re interested in. The authors of the SSVS paper explain that estimating all $2^p$ regression models isn’t feasible for a large number of possible predictors $p$, and so sampling is preferable.

  • $\begingroup$ So is SSVS a type of BMA? Or a generalization of BMA? Or a completely different animal? $\endgroup$ – Mr. Zen Nov 25 '18 at 15:34
  • $\begingroup$ @Mr.Zen BMA is pretty vague term and might refer to either model selection or prediction. SSVS is a specific example of how to implement the model selection part, that’s sampling-based, and that’s in the case of regression models. Maybe there are other papers that have extended the methodology to other applications though. $\endgroup$ – Taylor Nov 25 '18 at 20:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.