MCMC estimates when IGLS algorithm fails to converge

Welcome to the forum for runmlwin users. Feel free to post your question about runmlwin here. The Centre for Multilevel Modelling take no responsibility for the accuracy of these posts, we are unable to monitor them closely. Do go ahead and post your question and thank you in advance if you find the time to post any answers!

Go to runmlwin: Running MLwiN from within Stata >> http://www.bristol.ac.uk/cmm/software/runmlwin/
Post Reply
kathrynmbarker
Posts: 9
Joined: Fri Sep 22, 2017 10:27 pm

MCMC estimates when IGLS algorithm fails to converge

Post by kathrynmbarker »

Hello,

I am running a cross-classified binary logit model using the following code:

Code: Select all

sort nhood_w1 scid modmax_groupid aid
xi: quietly runmlwin pregfw2 i.race i.parent_highestedu ///
i.region i.h1co10 cons, level4(nhood_w1:cons) level3(scid: cons) ///
level2(modmax_groupid:cons) level1(aid:) ///
discrete(dist(binomial) link(logit) denom(cons) pql2) nopause
xi: runmlwin pregfw2 i.race i.parent_highestedu ///
i.region i.h1co10 cons, level4(nhood_w1:cons) level3(scid: cons) ///
level2(modmax_groupid:cons) level1(aid:) ///
discrete(dist(binomial) link(logit) denom(cons)) mcmc(cc) initsprevious nopause
Output shows parameter estimates for the MCMC estimation, but under the IGLS output I see, "WARNING: IGLS algorithm failed to converge." How can one interpret the MCMC results given the IGLS algorithm failed to converge? Should one dismiss the MCMC results outright given the initialization values are in question? As a solution, should I use MQL1 instead of PQL2 (though I had chosen PQL2 given it is the most accurate of the four quasi-likelihood methods and my within higher level units are small (mean 2.6 for levels 2 and 4) and the response proportion is (relatively) extreme (prob=0.113)?

Thank you!
Kate
ChrisCharlton
Posts: 1351
Joined: Mon Oct 19, 2009 10:34 am

Re: MCMC estimates when IGLS algorithm fails to converge

Post by ChrisCharlton »

The model being fitted with IGLS will be different from the MCMC version, as IGLS always treats models as hierarchical whereas the MCMC model in this case is specified as cross-classified. This may be part of the reason why the IGLS version does not converge.

I would suggest performing the usual MCMC diagnostics to gauge the stability of the MCMC estimate chains, as well as running the model with a diverse set of starting values to check that the estimates always end up in the same place. You may have to run a longer burnin that usual if the starting estimates are further away from the correct values.
kathrynmbarker
Posts: 9
Joined: Fri Sep 22, 2017 10:27 pm

Re: MCMC estimates when IGLS algorithm fails to converge

Post by kathrynmbarker »

Thanks, Chris! I bump into this issue only after I add a covariate with a large number of missing observations that decreases my sample size to: level-1, n=1135; l-2, n=430; l-3, n=89; l-4, n=457. Before proceeding with the diagnostics that you helpfully suggest, I just wanted to be sure that this issue is not due to the reduced sample size. Thanks again, Kate
ChrisCharlton
Posts: 1351
Joined: Mon Oct 19, 2009 10:34 am

Re: MCMC estimates when IGLS algorithm fails to converge

Post by ChrisCharlton »

Have you tried simulating or imputing the missing observations so that the sample size does not decrease? If not this may be a way of telling if it is the reduction in sample size causing the problem.
Post Reply