MIMIC fail to converge

[janina steinert]

Hello everybody,

I am trying to fit a MIMIC model measuring poverty in a household sample. The model looks like this: (HUMCAP -> education job) (HOUSINGQUAL -> floor wall dwelling tv refrigerator radio electricity lighting heating cooking washingmachine) (POV->HUMCAP HOUSINGQUAL)

However, when trying to fit this model it iterates forever. I don't understand why bc according to the k(k+1)/2 - rule my model should be identified.

I would be very thankful if someone had any ideas.

Thank you !!

[janina steinert]

cross-posted: http://stats.stackexchange.com/quest...il-to-converge

[Joseph Coveney]

My understanding is that that k(k+1)/2 rule-of-thumb should not be considered dispositive--you can satisfy it and still not have model identification. Regardless, even if the model is identified, it doesn’t mean that maximum likelihood iterations will converge.

You might want to try some of the following:

It seems that most (all?) of your indicator variables are binary. If so, then you could try forming the polyserial/polychoric (or tetrachoric if they’re all binary) correlation matrix and running an exploratory factor analysis on that just to see whether there is a solid-enough factor structure present to warrant pursuing your desired model.

You could try fitting just the Housing Quality submodel and use that model fit’s coefficients as starting values for the corresponding loading factors and factor variance in the larger model.

You might have to impose additional constraints on things just to get convergence in the first place. Once you have that, you can look for where the problem(s) lie by inspection of that initial fit, and then proceed from there.

(By the way, is that really a MIMIC model? It looks more like a two-level confirmatory factor model.)