Hi I’m using pymc on a hierarchical binomial model. The logit of the probability is a Gaussian with a certain std.dev. that I am trying to recover (called ‘si’ in the code’). While the MC samples look fine, find_MAP() does not seem to find the right values when data is limited (N=100 in the code).…

This is a known phenomenon, and basically the reason why MAP estimation has fallen out of favor. You can somewhat mitigate it by using a zero-avoiding prior on the sigma of the group-level effect (e.g. si = pm.Gamma('si', alpha=2, beta=0)). This is the advice of the STAN prior FAQ. In general, if y…

find_MAP struggles with hierarchical model

Questions version agnostic

cosine February 19, 2025, 11:20am 3

thank you for the quick reply.
(This nice community is sufficient reason to do MCMC:)

2 Likes

Topic		Replies	Views
Best practices for MAP estimation Questions	3	1247	September 6, 2019
Unconsistent logp, Evaluating the results of find_MAP on Custom Likelihood Questions	3	610	August 28, 2018
Evaluating log likelihood using estimates from find_MAP results Questions	3	1360	January 25, 2020
A little change to the sample code of MAP, however Questions	10	1136	April 1, 2018
MAP not working with shape>1 and VI on a pooled model Questions	11	836	November 25, 2017

find_MAP struggles with hierarchical model

Related topics