This is really nice and I can see what happened intuitively. Can anyone provide an explanation of why truncating the likelihood leads to a better model? If there are no data points outside the truncation bounds, then shouldn’t the log-likelihood of the existing points be the same whether it’s truncated or not?