mlogit with pooled data (not mlogit for panel model)

[Sam McCaw]

I have panel data for 40 individuals and their characteristics for 10 years.

The DV is an outcome variables with 4 choices (which some argue are ordered and others argue are not ordered).

I am already using an xtologit model under the "ordered outcome" assumption. However, I would also like to run a mlogit with the pooled data (not a panel model using gsem).

My questions are:

1) Can I use pooled data for individuals across time to run an mlogit model with individual dummies (fixed effects)? When I tried to run this, my model did not converge.

2) What are big concerns if I ran a simple mlogit regression with pooled data and are there ways to correct for it? Would clustering around stateid resolve any of these issues?

3) For later iterations of the model I will explore and learn to apply the gsem, mlogit as well but for the current time constraint I have I was wondering if there was a way to get around it.


Thanks.

SAM

[Clyde Schechter]

1. No, the use of individual indicators ("dummies") as a way to emulate fixed-effects modeling only works for linear models. It is not valid for multinomial logistic regression.

2. The results may be quite misleading as they fail to account for heterogeneity among individuals and intra-individual correlation, The use of clustered vce is better than nothing, but it does not fully correct for this problem.

3. The only thing I can think of that might be a suitable "quick and dirty" approach that you can do without taking time to learn something better would be to reduce it to ordinary logistic regression, by combining sufficiently similar outcome categories to reduce it to 2 levels instead of 4. And then you could run -xtlogit-.

[Sam McCaw]

Many thanks, Clyde, for your helpful comments. Not surprised that my models kept failing to converge. I'll add the ordinary logit model instead.

Best,

SAM

[Clyde Schechter]

You're welcome.

Just to be clear, the non-convergence is a different matter. The likelihood function of -mlogit- models is often "ill-behaved" and finding the maximum can be challenging. That is, even when the structure of the data make -mlogit- the appropriate model, difficulties with convergence are not uncommon.

In your case, the data aren't really suitable for a (non-panel) -mlogit- analysis. But the non-convergence is not necessarily a reflection of that. Even if the calculations had converged, the results would have been incorrect.

[Sam McCaw]

Many thanks again for the clarification. This is very helpful.

[William Kratochwill]

However, if I have individual data like Employment cross-section microdata, and I want to run mlogit and include State dummies? Or in the case I have two years' cross-sections, can I include State and year dummies? How much wrong can I be?