not only defended but also applied (rev’d)

Following a very positive and encouraging review by The American Statistician of our paper with Andrew Gelman on Feller’s misrepresentation of Bayesian statistics in the otherwise superb Introduction to Probability Theory , we have submited a revised version, now posted on arXiv. Hopefully, we will be able to publish this historic-philosophical note in The American Statistician, and maybe even get a discussion paper on the issue of misconceptions on Bayesian analysis.

the birthday problem [X'idated]

The birthday problem (i.e. looking at the distribution of the birthdates in a group of n persons, assuming [wrongly] a uniform distribution of the calendar dates of those birthdates) is always a source of puzzlement [for me]! For instance, here is a recent post on Cross Validated:

I have 360 friends on facebook, and, as expected, the distribution of their birthdays is not uniform at all. I have one day with that has 9 friends with the same birthday. So, given that some days are more likely for a birthday, I’m assuming the number of 23 is an upperbound.

The figure 9 sounded unlikely, so I ran the following computation:

extreme=rep(0,360)
for (t in 1:10^5){
  i=max(diff((1:360)[!duplicated(sort(sample(1:365,360,rep=TRUE)))]))
  extreme[i]=extreme[i]+1
  }
extreme=extreme/10^5
barplot(extreme,xlim=c(0,30),names=1:360)

whose output shown on the above graph. (Actually, I must confess I first forgot the sort in the code, which led me to then believe that 9 was one of the most likely values and post it on Cross Validated! The error was eventually picked by one administrator. I should know better than trust my own R code!) According to this simulation, observing 9 or more people having the same birthdate has an approximate probability of 0.00032… Indeed, fairly unlikely!

Incidentally, this question led me to uncover how to print the above on this webpage. And to learn from the X’idated moderator whuber the use of tabulate. Which avoids the above loop:

> system.time(test(10^5)) #my code above
user  system elapsed
26.230   0.028  26.411
> system.time(table(replicate(10^5, max(tabulate(sample(1:365,360,rep=TRUE))))))
user  system elapsed
5.708   0.044   5.762

[weak] information paradox

While (still!) looking at questions on Cross Validated on Saturday morning, just before going out for a chilly run in the park, I noticed an interesting question about a light bulb problem. Once you get the story out of the way, it boils down to the fact that, when comparing two binomial probabilities, p₁ and p₂, based on a Bernoulli sample of size n, and when selecting the MAP probability, having either n=2k-1 or n=2k observations lead to the same (frequentist) probability of making the right choice. The details are provided in my answers here and there. It is a rather simple combinatoric proof, once you have the starting identity [W. Feller, An Introduction to Probability Theory and Its Applications, vol. 1, 1968, [II.8], eqn (8.6)]

but I wonder if there exists a more statistical explanation to this weak information paradox…

“Not only defended but also applied”: The perceived absurdity of Bayesian inference

After a first unsuccessful attempt at publishing a note on the great Willliam Feller’s dismissive attitude towards Bayesian statistics, in An Introduction to Probability Theory and Its Applications, and more broadly about misconceptions on Bayesianism, jointly with Andrew Gelman, last year, we have rewritten some of it and resubmitted to The American Statistician. It has also been re-arXived. Here is the abstract:

Abstract. The missionary zeal of many Bayesians has been matched, in the other direction, by a view among some theoreticians that Bayesian methods are absurd—not merely misguided but obviously wrong in principle. We consider several examples, beginning with Feller’s classic text on probability theory and continuing with more recent cases such as the perceived Bayesian nature of the so-called doomsday argument. We analyze in this note the intellectual background behind various misconceptions about Bayesian statistics, without aiming at a complete historical coverage of the reasons for this dismissal.

“Not only defended but also applied”

On page 124 of his superb Introduction to Probability Theory book (volume 1), William Feller has this strange remark about Bayesian inference:

“Unfortunately Bayes’ rule has been somewhat discredited by metaphysical applications of the type described above. In routine practice, this kind of argument can be dangerous. A quality control engineer is concerned with one particular machine and not with an innite population of machines from which one was chosen at random. He has been advised to use Bayes’ rule on the grounds that it is logically acceptable and corresponds to our way of thinking. Plato used this type of argument to prove the existence of Atlantis, and philosophers used it to prove the absurdity of Newton’s mechanics. In our case it overlooks the circumstance that the engineer desires success and that he will do better by estimating and minimizing the sources of various types of errors in predicting and guessing.The modern method of statistical tests and estimation is less intuitive but more realistic. It may be not only defended but also applied.”

When we were discussing about this great book, Andrew Gelman pointed out to me this strong dismissal of Bayesian techniques (note that I had overlooked so far) and, given that it is still quoted as an argument against a Bayesian approach to inference, we ended up writing [well, mostly Andrew!] a short note on the motivations and implications of this remark, now published on arXiv. One of the points is that Feller’s sentence has the interesting feature that it is actually the opposite of the usual demarcation: typically it is the Bayesian who makes the claim for inference in a particular instance and the frequentist who restricts claims to infinite populations of replications. Another point is the naïve faith in the classical Neyman-Pearson theory to solve practical problems in statistics.

Actually, Persi Diaconis took a (deeper) look at Feller’s stance as well, as mentioned in this review of Jaynes’s Probability Theory. Using Amazon Look Inside tool,  I spotted Feller being mentioned more than 30 times in Jaynes’s book, one of the best quotes being “The date was 1956 when the author met Willy Feller“! More to the point, Jaynes identifies Feller’s dismissal of the “old wrong ways” (volume 2, p.76), which is to be opposed to the “modern method” above. (Persi Diaconis and Susan Holmes also wrote a nice piece entitled “A Bayesian peek into Feller volume 1″ that does not relate directly to this issue.) In a loosely related point, Persi’s warning that he sees “a strong trend against measure theory in modern statistics departments: [he] had to fight to keep the measure theory requirement in Stanford’s statistics graduate program“, to which I completely subscribe, should be heard more widely…