Open aagnello opened 9 years ago
We might consider the mean log likelihood as a better approximation to the Bayesian evidence. It differs by the information gain between prior and posterior - I'm not sure how much of the "Occam factor" is contained within this information term, but if both Nebula and Lens are capturing similar amounts of information, the mean log L might be useful in approximating the class probabilities.
@aagnello You also had an idea for some sort of weighted BIC, which in the spirit of finding engineering solutions might be thrown into a score comparison.
BIC just counts parameters: when we go to the trouble of including all teh information in the OM10 lens parameter prior, we're going to want to use a better approximation of the evidence. I suggest we look into the mean log likelihood over the posterior samples. This is a blunt tool, but hopefully sharp enough for our purposes. Note the need for samples, though.