www.cmm.bristol.ac.uk/forum

Dear all,

I am running a multiple membership multiple classification model with individuals nested in schools, areas and (one or more) friendship networks (cf., Tranmer et al. 2014). My outcome is a count variable. To account for overdispersion, I have added a pseudo-level as suggested elsewhere on this forum (http://www.cmm.bristol.ac.uk/forum/view ... ?f=3&t=774). My code for a model without predictors looks as follows:

sort school area egonet individual individual
quietly runmlwin count cons, ///
level5(school:cons) ///
level4(area:cons) ///
level3(egonet:cons) ///
level2(individual:cons) level1(individual:) ///
discrete(distribution(poisson) link(log)) rigls nopause

runmlwin count cons, ///
level5(school:cons) ///
level4(area:cons) ///
level3(egonet:cons, mmids(egonet-egonet7) mmweights(weight1-weight7)) ///
level2(individual:cons) level1(individual:) ///
discrete(distribution(poisson) link(log)) mcmc(cc burnin(10000) chain(100000)) initsprevious

The model output contains, in addition to an estimate for the intercept (cons), a random effect parameter for school, area, and egonet, and an individual-level random effect that’s supposed to take care of the overdispersion, like this:

Level 5: school - var(cons) - estimate
Level 4: area - var(cons) - estimate
Level 3: egonet - var(cons) - estimate
Level 2: individual - var(cons) - estimate

Given this output, I want to compare the amount of variance at different levels – individual, school, area, egonet – and how much these can be accounted for by predictors to be added later, but I am struggling to understand how to do this for this particular model, given the Poisson distribution, the inclusion of a pseudo-level, and the use of a multiple membership structure for the egonets. So that is my question. Help would be much appreciated!

Many thanks,
Bram van Leeuwen


Tranmer, M., Steel, D., & Browne, W. J. (2014). Multiple‐membership multiple‐classification models for social network and group dependences. Journal of the Royal Statistical Society: Series A (Statistics in Society), 177(2), 439-455.

Dear Bram,

I think you would be best posting this on the JISC multilevel email list as it more a general question about interpreting changes in the absolute and relative magnitude of variance parameters as you add covariates rather than a runmlwin specific question. Needless to say, as soon as you move from continuous to discrete response models these issues are no longer straightforward.

Best wishes

George

Dear George,

Thank you for the advice. I'll post the question at JISCMail.

Best,
Bram