Currently at the end of an estimate for prior variance in each SER, we compare the loglik of the estimate vs that from setting prior to zero. We will use the zero variance estimate instead unless the estimated (non-zero) variance achieves an obviously better likelihood: that is, in our current code, 1.1 times the likelihood from zero variance. See this line where check_null_tol on the log scale was set to 0.1 . This allows us to be a bit more stringent about detecting additional effects.
There is a potential issue with it: since we allow for certain tolerance like this, the new prior estimate might no longer maximizing the ELBO. As a result ELBO might decrease.
Currently at the end of an estimate for prior variance in each SER, we compare the loglik of the estimate vs that from setting prior to zero. We will use the zero variance estimate instead unless the estimated (non-zero) variance achieves an obviously better likelihood: that is, in our current code, 1.1 times the likelihood from zero variance. See this line where
check_null_tol
on the log scale was set to 0.1 . This allows us to be a bit more stringent about detecting additional effects.There is a potential issue with it: since we allow for certain tolerance like this, the new prior estimate might no longer maximizing the ELBO. As a result ELBO might decrease.