Gabor Grothendieck forwarded the following article by Derek Lowe from Medical Progress Today:

FDA Shows Interest in 18th Century Presbyterian Minister
Bayesian statistics may help improve drug development

Not many ideas of 18th-century Presbyterian ministers attract the interest of the pharmaceutical industry. But the works of Rev. Thomas Bayes have improved greatly with age. . . . For decades, no one heard very much about Bayesian statistics at all. One reason for this was they’re much more computationally demanding, which was a real handicap until fairly recently. . . But things are changing. . . . Bayesian and standard “frequentist” statistics are in many ways mirror images of each other, and there are mistakes to be made each way. . . . Bayesian statistics, though, don’t address the likelihood that your observed results might have come out by random chance, but rather give you a likelihood of whether your initial hypothesis is true. . . . As far as I know, no pharma company has yet taken a fully Bayesian clinical package to the FDA for a drug approval. There have been a few dose-finding trials in the cancer area, and Pfizer’s research arm in England ran a large trial of a novel stroke therapy under Bayesian protocols. . .

It’s an interesting article in that it’s presenting an pharmaceutical-industry perspective of something that I usually think of from an academic direction. I suspect that there’s more Bayesian stuff going on in pharmaceuticals than people realize, though: I’d be curious what Amy Racine-Poon or Don Berry or Don Rubin could add to this article. For example, Bayesian data analysis has been used in toxicology for awhile.

Also, as I discussed in response to another recent news article about Bayesian inference, I think the differences between Bayesian and other statistical methods can be overstated. In particular, so-called frequentist methods still require “esitmates” and “predictions” which can (and often are) obtained using Bayesian inference. Also, I don’t think it’s quite right to say that Bayesian methods “give you a likelihood of whether your initial hypothesis is true.” It would be more accurate to say that they allow you to express your unceratainty probabilistically, for example, giving some distribution of the effectiveness of a new drug, as compared to an existing treatment. The idea of a point “hypothesis” is, I think, a holdover from classical statistics that is a hindrance, not a help, in Bayesian inference. Finally, the article has a bit of discussion about where the prior comes from. In many many examples, the prior distribution is estimated from data using hierarchical modeling. There’s not any need in this framework to specify a numerical prior distribution in the way described in the article.

In conclusion, Lowe’s article gives an interesting look at Bayesian inference from another perspective, and also reveals that some of the recent (i.e., since 1980) developments in Bayesian data analysis still have not trickled through to the practitioners. I think that hierarchical modeling is much more powerful than the traditional “hypothesis, prior distribution, posterior distribution” approach to Bayesian statistics.