Dirichlet multinomial likelihood
Craig44 opened this issue · 8 comments
There is a branch that has implemented 90% the dirichlet multinomial likelihood.https://github.com/NIWAFisheriesModelling/CASAL2/tree/DirichletMultinomial
There are unit-tests
There is a usermanual description
However, when estimating the theta parameter it was running to bounds so this needs furthur investigation
Added TestModel "LIM_tc_Dm", the "LIM_tc" model with the Dirichlet multinomial likelihood used for the composition observations.
I reviewed and updated most comments in recent Pull Request.
However the comment made here #405 (comment) would benefit from additional looking. The SSB has doubled vs the multinomial test case. Is this reasonable? @amart
Touche we may be comparing a weighted estimation (Dirichlet multinomial) with an unweighted estimation (multinomial with no stage two weightings). If you are happy with this you can close and merge.
For TestModel LIN_tc, on CentOS 7.9 with Casal2 compiled with devtoolset-11, the MPD B0 estimate is 308529. The composition data have sample sizes ranging from 13 to 120. Changing all of the sample sizes to 1/10 of their current value for the 5 sets of composition observations results in the MPD B0 estimate of 396892. Using the Dirichlet multinomial likelihood results in effective sample sizes of less than 5, most of which are less than 2, and the MPD B0 estimate of 688682.
Merged pull request #405 on 2022-09-08.