www.cmm.bristol.ac.uk/forum

Hi Dear

I had run multinomial multilevel model using runmlwin command as follow. My model has 3 levels:

xi: runmlwin Outcome Sex cons , level3( District: cons) level2( Accident :cons) level 1(Driver:) discrete(dist (multinomial) link(mlogit) denom (cons) basecategory(1)) mlwinsettings(optimat) nopause maxiterations(35)

in stata outputs all amounts for var (cons_2), cov(cons_2,cons_3) and var(cons_3) are 0.

How can I solve problem?

Regards

Dear Ali,

This doesn't sound like a runmlwin problem per se. That is, I would expect you to run into similar problems if you used other software. Having said this, you might as well try all the usual things like building up your model gradually using the parameter estimates of simpler models as starting values for more complex models. See previous posts on this issue.

However your problem sounds more to do with your choice of model given the characteristics of your data. What I mean by this is that perhaps you don't have enough level three units to fit a three level model. Or perhaps three is very little true variation at level three and so that is why the variance covariance matrix gets stuck at zero. Alternatively, perhaps one of your multinomial categories is very rare or while not far on average, is rare or non existent in many of your level three units. You need to carefully explore your data to try to work out why the model you have specified is struggling to fit.

Sorry not to be of more help.

George

Hi Dear George

Many thanks, I also think that the problem is with my data or model, but there is a question. When I tried to run the model with two levels (accident and driver), there is no problem and estimates for random effect part (accident) appear in stata outputs. Adding "district" to the model as third level, all estimated amounts for second level (accident) will be zero, but not for higher level (district). Is it true that i forget about third level and run model with just two levels?

Best Wishes

Dear Ali,

Again this is a general multilevel modelling question not a runmlwin specific one. That is, I expect you would run into similar situation using other software. If you are unable to get sensible looking results for your desired three-level model then first you want to work out why this is and if, armed with this knowledge, you are still unable to get the model working, then yes, you will need to simplify your model. One possible simplification would be reduce the model down to two-levels as you suggest by leaving out either the third or second level, whichever makes more sense substantively.

Sorry not to be of more help

Best wishes

George