Going several directions at once in tweaking the warning signals code to make sure I have clear comparisons that are not resulting from any limitations in the numerical algorithms. Performing a variety of error checks, etc:
Comparing fixed models vs comparing models estimated from the data
Ensure that const model doesn’t attempt to estimate m (doesn’t impact likelihood anyway
Ensure only positive values of sigma are returned (trying use trigger in setLTC and setLSN to return infinite variance if sigma is negative)
tau distributions compare totally separate null and test models, rather than ones both estimated from the same dataset.
Today’s simulations [flickr-gallery mode=“search” tags=“stochpop” min_upload_date=“2011-02-02 00:00:00” max_upload_date=“2011-02-03 23:59:59”]