Focus on Bayesian Methods
Last year, SAS introduced Bayesian capabilities in SAS/STAT software with experimental web downloads for three procedures: GENMOD, LIFEREG, and PHREG. With the recent release of SAS 9.2 Phase 1, these capabilities become production. In addition, SAS 9.2 includes a new experimental procedure, PROC MCMC, which provides a flexible tool for Bayesian modeling.
So what's the big deal about Bayesian methods?
Bayesian analysis has become a standard statistical technique in recent years. The ideas are not new, as the original concept was laid out in the late 1800s, but recent computing advances have made the methodology more plausible. Bayesian methods incorporate existing information, based on expert knowledge, past studies, and so on, into your current data analysis. This 'prior' information takes the form of a prior distribution, and the data likelihood is effectively weighted by the prior distribution as the data analysis results are computed. Obviously, the selection of the prior is very important in this type of analysis.
The main outcome of a Bayesian analysis is a set of probabilities, or the posterior distribution, rather than a point estimate and its standard error. This allows you to address scientific questions of interest more directly. This is due to the fact that, once the model parameters are estimated, it is rather easy to compute the posterior distributions for any functions of the parameters or any quantities of research interests. Some statisticians produce Bayesian analyses simply to operate in the Bayesian framework. They use noninformative priors that produce very similar results to say, maximum likelihood-based analysis.
How is the software download program going?
Users have downloaded over 1800 copies of the "B" procedures, as we call them (the GENMOD, LIFEREG, and PHREG procedures got temporary B prefixes for the 9.1.3 release). Seventy people took the Bayesian tutorial at SAS Global Forum this year and the same number took the tutorial at ENAR (Eastern North American Regional of the International Biometrics Society). The new MCMC procedure attracted lots of interest at our exhibition table at ENAR and we're looking forward to feedback.
I've used the 9.1.3 download procedures. What's different in 9.2?
In SAS 9.2, the capabilities in the GENMOD, LIFEREG, and PHREG procedures are now production. BY-processing is available in all three procedures, and so is the DIC statistic. In addition, the PHREG procedure provides survival function estimation using the BASELINE statement for the Bayesian analysis, and it also provides estimates of the hazard ratio function.
Of course, the biggest difference is the availability of the experimental MCMC procedure. You can fit models by specifying the standard prior distributions and likelihood functions that are provided, or by specifying arbitrary hierarchical models with SAS programming statements, similar to the style in which you specify models in PROC NLMIXED.
How can I learn more about these methods and the capabilities in SAS software?
A good starting place would be the new Introduction to Bayesian Analysis Procedures. This chapter also provides a comprehensive reading list, with suggestions sorted by different levels.
The next chance to take a course is August 6 at the Joint Statistical Meetings, taught by Fang Chen, SAS Institute Inc.