More on wobbling of priors

In two recent posts (here and here), I made arguments based on the idea that wobbliness in priors translates to wobbliness in posteriors. The posts while mathematically correct neglect an epistemologically important fact: a wobble in a prior may be offset be a countervailing wobble in a Bayes’ factor, resulting in a steady posterior.

Here is an example of this phenomenon. Either a fair coin or a two-headed coin was tossed by Carl. Alice thinks Carl is a normally pretty honest guy, and so she thinks it’s 90% likely that a fair coin was tossed. Bob thinks Carl is tricky, and so he thinks there is only a 50% chance that Carl tossed the fair coin. So:

• Alice’s prior for heads is (0.9)(0.5)+(0.1)(1.0) = 0.55

• Carl’s prior for heads is (0.5)(0.5)+(0.5)(1.0) = 0.75.

But now Carl picks up the coin, mixes up which side was at the top, and both Alice and Bob have a look at it. It sure looks to them like there is a head on one side of it. As a result, they both come to believe that the coin is very, very likely to be fair, and when they update their credences on their observation of the coin, they both come to have credence 0.5 that the coin landed heads.

But a difference in priors should translate to a corresponding difference in posteriors given the same evidence, since the force of evidence is just the addition of the logarithm of the Bayes’ factor to the logarithm of the prior odds ratio. How could they both have had such very different priors for heads, and yet a very similar posterior, given the same evidence?

The answer is this. If the only relevant difference between Alice’s and Carl’s beliefs were their priors for heads, then indeed they couldn’t get the same evidence and both end up very close to 0.5. But their Bayes’ factors also differ.

• For Alice: P(looks fair | heads)≈0.82; P(looks fair | tails)≈1; Bayes’ factor for heads vs. tails ≈0.82

• For Bob: P(looks fair | heads)≈0.33; P(looks fair | tails)≈1; Bayes’ factor for heads vs. tails ≈0.33.

Thus, for Alice, that the coin looks fair is pretty weak evidence against heads, lowering her credence from 0.55 to around 0.5, while for Bob, that the coin looks fair is moderate evidence against heads, lowing his credence from 0.75 to around 0.5. Both end up at roughly the same point.

Thus, we cannot assume that a difference with respect to a proposition in the priors translates to a corresponding difference in the posteriors. For there may also be a corresponding difference in the Bayes’ factors.

I don’t know if the puzzling phenomena in my two posts can be explained away in this way. But I don’t know that they can’t.