Bayesian analysis with Stata 14

Keen Stata users will be aware that  Stata 14 was released this week. I got a email flyer advertising the new release and the lead item was the fact that Bayesian analysis is now part of official Stata.

I have been campaigning for wider use of Bayesian methods and the fact that Stata now offers this facility will mark a huge step forward. I am delighted but also a little puzzled.

When I was writing the book on Bayesian analysis with Stata, I naturally asked about plans for incorporating Bayesian analysis into the official release of Stata and I was told that it was a long-term goal but that nothing was currently underway. Then last year, at the UK Stata Users Group meeting in London, I gave a talk based on my book. William Gould, the presedent of StataCorp and the driving force behind Stata, was in the audience and he repeated to me that nothing immediate was planned. Perhaps, this was just sensible commerical confidentiality, or maybe my talk extolling the benefits of Bayesian analysis had more impact than I realised, because to my surprise, Stata now has its first official commands for Bayesian analysis.

StataCorp has very kindly given me a complimentary copy of Stata 14, so over the coming weeks I will have the opportunity to evaluate the new commands and I’ll report my findings in this blog.

My immediate reaction is that the new facilities are heavily influenced by the first few chapters of my book, although of course, they have more of the Stata house style in their presentation and no doubt the coding is to a much higher standard. The proof of their long-term success will lie in the speed of the algorithms and I have not yet had the opportunity to test that.

As we have seen many times, both in the book and then in this blog, programming Bayesian analyses in Stata is easy but it can be very slow and for even moderately sized models we either need Mata or we need to export the problem to WinBUGS or some other purpose written program. I assume that StataCorp will have followed the advice in Chapter 12 of my book and used Mata for the standard models, but much of the advantage of Bayesian analysis lies in its flexibility and the ease with which one can adapt models to better reflect one’s own problem; coding in Stata, this might not be practical.

It would be unfair to judge StataCorp on this first attempt at Bayesian analysis, but now that they have started, they will find themselves under constant pressure to make their code better and that can only be to the good. Perhaps, we can look forward to the day when a WinBUGS-like program is written in Mata and integrated fully into Stata.