Continuing my recent, seemingly interminable, series of too-technical posts on probability theory… To understand this one you’ll need to remember Bayes’ Theorem, and the resulting need for a Bayesian statistician to come up with an appropriate prior distribution to describe her state of knowledge in the absence of the experimental data she is considering, updated to the posterior distribution after considering that data. I should perhaps follow the guide of blogging-hero Paul Krugman and explicitly label posts like this as “wonkish”.
(If instead you’d prefer something a little more tutorial, I can recommend the excellent recent post from my colleague Ted Bunn, discussing hypothesis testing, stopping rules, and cheating at coin flips.)
Deborah Mayo has begun her own series of posts discussing some of the articles in a recent special volume of the excellently-named journal, “Rationality, Markets and Morals” on the topic Statistical Science and Philosophy of Science.
She has started with a discussion Stephen Senn’s “You May Believe You are a Bayesian But You Are Probably Wrong”: she excerpts the article here and then gives her own deconstruction in the sequel.
Senn’s article begins with a survey of the different philosophical schools of statistics: not just frequentist versus Bayesian (for which he also uses the somewhat old-fashioned names of “direct” versus “inverse” probability), but also how the practitioners choose to apply the probabilities that they calculate: either directly in terms of inferences about the world versus using those probabilities to make decisions in order to give a further meaning to the probability.
Having cleaved the statistical world in four, Senn makes a clever rhetorical move. In a wonderfully multilevelled backhanded compliment, he writes
If any one of the four systems had a claim to our attention then I find de Finetti’s subjective Bayes theory extremely beautiful and seductive (even though I must confess to also having some perhaps irrational dislike of it). The only problem with it is that it seems impossible to apply.
He discusses why it is essentially impossible to perform completely coherent ground-up analyses within the Bayesian formalism:
This difficulty is usually described as being the difficulty of assigning subjective probabilities but, in fact, it is not just difficult because it is subjective: it is difficult because it is very hard to be sufficiently imaginative and because life is short.
And, later on:
The … test is that whereas the arrival of new data will, of course, require you to update your prior distribution to being a posterior distribution, no conceivable possible constellation of results can cause you to wish to change your prior distribution. If it does, you had the wrong prior distribution and this prior distribution would therefore have been wrong even for cases that did not leave you wishing to change it. This means, for example, that model checking is not allowed.
I think that these criticisms mis-state the practice of Bayesian statistics, at least by the scientists I know (mostly cosmologists and astronomers). We do not treat statistics as a grand system of inference (or decision) starting from single, primitive state of knowledge which we use to reason all the way through to new theoretical paradigms. The caricature of Bayesianism starts with a wide open space of possible theories, and we add data, narrowing our beliefs to accord with our data, using the resulting posterior as the prior for the next set of data to come across our desk.
Rather, most of us take a vaguely Jaynesian view, after the cranky Edwin Jaynes, as espoused in his forty years of papers and his polemical book Probability Theory: The Logic of Science — all probabilities are conditional upon information (although he would likely have been much more hard-core). Contra Senn’s suggestions, the individual doesn’t need to continually adjust her subjective probabilities until she achieves an overall coherence in her views. She just needs to present (or summarise in a talk or paper) a coherent set of probabilities based on given background information (perhaps even more than one set). As long as she carefully states the background information (and the resulting prior), the posterior is a completely coherent inference from it.
In this view, probability doesn’t tell us how to do science, just analyse data in the presence of known hypotheses. We are under no obligation to pursue a grand plan, listing all possible hypotheses from the outset. Indeed we are free to do ‘exploratory data analysis’ using (even) not-at-all-Bayesian techniques to help suggest new hypotheses. This is a point of view espoused most forcefully by Andrew Gelman (author of another paper in the special volume of RMM).
Of course this does not solve all formal or philosophical problems with the Bayesian paradigm. In particular, as I’ve discussed a few times recently, it doesn’t solve what seems to me the most knotty problem of hypothesis testing in the presence of what one would like to be ‘wide open’ prior information.