Updating probabilities & characterizing CAR: when conditioning succeeds and when it fails.

P. Grünwald
CWI and Universiteit Leiden

Abstract:

How should we update a probability distribution when given new information? The standard answer is: "conditioning", but this is not always correct: as shown by examples such as the three prisoners problem, the Monty Hall (three cars, quizmaster) problem and the two-children puzzle, at least when applied in the common "naive" form, conditioning often gives the wrong answers. We give a detailed explanation of this phenomenon and show that a criterion known as CAR (`coarsening at random') in the statistical literature characterizes when "naive" conditioning works. We provide two characterizations of CAR. First we show that in many situations, CAR essentially *cannot* hold, so that naive conditioning must give the wrong answer. This is problematic since people apply naive conditioning all the time. Second, we precisely characterize CAR by providing a randomized algorithm that can simulate all and only CAR distributions. This solves a tricky open problem first posed by Gill, van der Laan and Robins (1997).

The talk is based on joint work with R. Gill (Leiden) and J. Halpern (Cornell).

back to the list of talks