Refining starting values stuck during multilevel modelling

[Harry Gommans]

Dear all,

I am running a multilevel model for my paper (individuals nested in countries). For the base model and the random intercepts model, everything goes well. Yet, while trying a random slopes model, Stata keeps stuck at "Refining starting values:". I've waited for hours - nothing happens. I checked beforehand manually which variables were suitable to include as having random slopes with an lrtest.

The code is the following:

xtmelogit csdp solecntatt townatt cntryatt euatt cntryeuatt immcont dismoscow expbor nuclweap gender age education vccnteu rela employment religion || country: solecntatt townatt cntryatt cntryeuatt immcont gender age education vccnteu employment, or var

My dependent variable is binary, hence the odds ratios.

Is there something I am doing wrong?

Best,
Harry

[Joseph Coveney]

Your model has 10 random slopes in addition to the random intercept. You can try specifying the laplace option (or its equivalent, intpoints(1)), but even so the algorithm is liable to take weeks, if not months, to complete its iterations.

[Harry Gommans]

Thank you for the response, Joseph. Several of the random slopes are control variables, so I could consider leaving them out as random slopes. Do you have an idea from which amount of random slopes onwards the model becomes too complicated for Stata to complete quickly?

[Joseph Coveney]

Do you have an idea from which amount of random slopes onwards the model becomes too complicated for Stata to complete quickly?

I can't say for certain in your case, but in my experience, with generalised linear mixed models, integration gets bogged down after about two random effects when they are either a random intercept + random slope (and covariance) or cross-classified.

[Harry Gommans]

I see. I included these random slopes because the variables showed significant variance between clusters. I already dropped my control variables as random slopes which leaves me with 5 random slopes, but I run into the same problem. Is there any statistical rule that states which of the two random slopes that show significant variation (out of the 5 main independent variables that I have) should be included?

[Joseph Coveney]

I included these random slopes because the variables showed significant variance between clusters.

I don't follow that reasoning. That is, I don't see how the variation in the ranges of observed values of an explanatory variable between clusters implies the presence of a random effect for slope.

I would include a random slope because I believe that it is an important component in the data generating process and that its presence in the statistical model improves inference. Otherwise, I'd leave it out for the sake of parsimony.

Is there any statistical rule that states which of the two random slopes that show significant variation (out of the 5 main independent variables that I have) should be included?

Sorry, I don't know what you mean by "random slopes that show significant variation".

As far as giving priority to which among several random slope variables to include in a model, I'm not aware of any rule of thumb for triage, but that doesn't mean much. I would likely choose those few that I believe are the most important to inference.