Log likelihood with binary dummy? (using OS X)

[kylem133]

Hello,

I'm using runmlwin on Mac OS X, using the method outlined here
https://www.cmm.bristol.ac.uk/forum/vie ... f=3&t=1073 , (which, by the way I am eternally grateful for!)

Now, I don't know if my question is specific to OSX users, or to all users of runmlwin, but I hope someone can help out.

I'm basically estimating a model with a binary dependent variable, of the following form (for example)

Code: Select all

runmlwin depvar cons, level3(level3_id: cons) level2(level2_id: cons) level1(level1_id:) discrete(distribution(binomial) link(logit)denom(cons) pql2) 

At this stage, all I'm attempting to do is decide on appropriate multilevel structure for my model (i.e. how many levels, and which levels to include). The usual way I would do this when using, say, xtmelogit in stata is to compute an LR test either with the command lrtest or manually, using the Log-Likelihood statistics. However I would prefer to use runmlwin over xtmelogit as the computation times with xtmelogit are frustratingly slow for 3 or 4 level models.

However the runmlwin output, having specified the binary option, does not display the Log Likelihood. At the same time, I get an error above each output stating "Warning: getversion plugin could not be loaded". Is this error the reason why there is no Log-likelihood displayed (and if so is this fixable?), or does runmlwin not specify the log-likelihood for binary response models (and if so, is there a way to obtain this?)

Furthermore, when attempting to use the lrtest following the runmlwin commands (having stored 2 sets of estimates), I get the error "model1 does not contain scalar e(ll)". Might there be an option to store the 'scaler e(ll)', allowing for the use of 'lrtest'?

Apologies if my questions are rudimentary in nature and I thank you in advance for any assistance you can provide!

Kyle

[ChrisCharlton]

As discrete responses via (R)IGLS are estimated via quasi-likelihood it isn't possible to provide an accurate log-likelihood value (see http://www.bristol.ac.uk/cmm/software/s ... equawindow). As this isn't available using a Wald test is the recommended way to examine the model fit. As the results can sometimes be biased (see http://www.bristol.ac.uk/cmm/software/s ... entresults) we also recommend fitting at least your final models via MCMC, and these can be compared via the DIC statistic.

The Warning: getversion plugin could not be loaded message that you are getting can be ignored. This plugin is used when calling -runmlwin- on Windows machines so that the MLwiN version used can be displayed in the model results, and to prevent the command being run on versions that are too old. This plugin does not currently have an equivalent for the version used on Mac or Linux machines.

[kylem133]

Hello,

Apologies for the delay in replying but thanks for the advice!

Kyle