www.cmm.bristol.ac.uk/forum

Hello,
I am running a two-level (multilevel) multinomial (3 categories for the response) logistic regression model using runmlwin in STATA. I tried the models for both second order penalised quasi-likelihood linearization and MCMC which ran correctly without the survey weights as below:

quietly runmlwin outcome cons female, ///
level2(v001x: cons) ///
level1(caseid_1:) ///
discrete(distribution(multinomial) link(mlogit) denominator(cons) ///
basecategory(0)) ///
nopause rrr

/*Second order penalised quasi-likelihood linearization */
runmlwin outcome cons female, ///
level2(v001x: cons) ///
level1(caseid_1:) ///
discrete(distribution(multinomial) link(mlogit) denominator(cons) pql2 ///
basecategory(0)) ///
initsprevious nopause rrr

/* MCMC */
runmlwin outcome cons female, ///
level2(v001x: cons) ///
level1(caseid_1:) ///
discrete(distribution(multinomial) link(mlogit) denominator(cons) ///
basecategory(0)) ///
mcmc(burnin(1000) chain(20000) thinning(2) orthogonal) initsprevious nopause rrr


The challenge: I wish to incorporate survey weights into the Second order penalised quasi-likelihood linearization since this is not possible for MCMC but I got an error saying "Weights are only valid for univariate models estimated using (R)IGLS" after running the code below:

/*Second order penalised quasi-likelihood linearization with survey weight (wt_w=level 2 weights and wt_c=level 1 weights*/
runmlwin outcome cons female, ///
level2(v001x: cons,weightvar(wt_w)) ///
level1(caseid_1:,weightvar(wt_c)) ///
discrete(distribution(multinomial) link(mlogit) denominator(cons) pql2 ///
basecategory(0)) ///
initsprevious nopause rrr

I am using STATA 14.2 and MLwiN v3.05.

Any help will be highly appreciated.

Unfortunately, as indicated by the message, you can only specify weights for models that are both fitted with (R)IGLS and have a univariate response via runmlwin.

Thank you Chris for confirming this.
Do you have some interest of incorporating this functionality in the near future?
Cheers

I want to include survey weights in the Second order penalized quasi-likelihood linearization because MCMC does not allow it skibidi toilet but I get an error that says "Weights are only valid for univariate models estimated using (R)IGLS."