Melogit and testing for a random slope - New iterations (not concave) keep coming up

[ChR]

Hello,

I am running a multilevel logit regression analysis in Stata 13.1 using the melogit command to test for a random slope. I use the following syntax:

melogit dummyy x|| Group: x, cov(uns) || Subject:

However, the analysis keeps running and generates non concave iterations. Am I doing something wrong or does this analysis just takes much time to finish?

Thank you for your help!

Regards ChR

[ChR]

Ok, I have found out the problem is in the covariance part; the analysis works without this part. Therefore, I also tried cov(exchangeable), but that does not work either. Can I leave this part out? I thought it should be in there, but I have found a UCLA website that does not use cov for three level mixed effects logistic regression.

[Clyde Schechter]

I'm guessing from your question that you may not fully understand what the cov() option does in your model. When left unspecified, Stata defaults to cov(independent)--which is less likely to have convergence problems because Stata does not need to estimate the covariance between your random intercepts and random effects. Estimating the latter can be somewhat delicate. But the question of whether you can exploit this depends on the meaning of your data.

You need to think about whether you expect your random intercepts at the Group level and the random slopes at the Group level to be sensibly thought of as independent. Bear in mind, too, that this will depend critically on what the meaning of x = 0 in your data is. The same data can and will produce very different random effect covariance matrices depending on whether x = 0 is something in the center of the data as a whole, something in the center of each group's data, something at the boundary of the data, or something outside the data altogether (the last being typically not very useful). So you need to think about what the expected relationships are in relationship to x = 0. If you don't know what I'm talking about here, you might want to do the Bristol University on-line course in multi-level modeling before continuing to work on your model.

Although exchangeable is an admissible specification for cov(), it is computationally no different from unstructured when there is only one level of nesting in the data, and it is, in any case, something of an unusual specification for the random effects covariance. Perhaps you are really thinking about the covariance structure of the residuals instead of the random effects? In that case you want to specify that in the residuals() option, not the cov() option.