I love a dramatic entrance. Welcome!
I’ll state for the record that in the past I’ve defended Richard Carrier for using a Bayesian framework to evaluate the historicity of Jesus. Consistency forces me to say that in principle, there is nothing wrong with you or the Dales taking the same approach for evaluating the historicity of the Book of Mormon. And it seems to me that unlike the Dales, you actually understand the statistics.
I’ll leave it to others to debate whether the specific probabilities you are plugging into the formulas are appropriate, but I do want to make one comment about something you said here.
In my field (actuarial science) we continuously repeat clichés like, “All models are wrong. Some are useful.” A few weeks ago Kerry Shirts mentioned the Beta distribution, and since this is related to that, I’ll flesh out a third approach that like all models will also be wrong, but hopefully will also be useful in this context.kyzabee wrote: ↑Sat Jul 24, 2021 3:52 amBilly's laid out a lot here, but I'll try to take it a piece at a time.
"Say I have a coin."
I'd quibble pretty strongly with your particular example, as something like a chi-square or Fisher's Exact test would be much more appropriate than just comparing the raw probability of achieving that specific string, and it would avoid the drama of dealing with small probabilities, but Billy's general point is accurate here. This is in principle what I'm trying to do--to estimate the probability of observing a piece of evidence under the two (or sometimes three) competing hypotheses.
It's worth noting here that the more appropriate test would be comparing p = .43 (chisq(1) = .60) with p = .60 (chisq(1) = .27). . .
Billy looked at the likelihood of getting a specific series of test results with p = 0.5 and p = 0.55. You could take that approach one step further and calculate the likelihood of getting those results with every value of p between 0 and 1. The result would be a likelihood function. You could then shake your hand and do some calculus magic to find the value of p that maximizes the likelihood function. That would be the maximum likelihood estimator of p.
Creating and evaluating a likelihood function is somewhat tedious, but it turns out there is a probability distribution that has almost exactly the same shape as the likelihood function. That is the Beta distribution with parameters Beta(18,12) (the numbers of heads and tails that were rolled). From that, we can easily see that we can be 90% certain that the true probability of flipping a head is between 45% and 74%. That confirms what everybody else has seen—this just isn’t enough data to tell us what p is with a high degree of precision.
But anyway, welcome to the board.