Choosing the right modeling technique available in STATA

[Sam McCaw]

Dear Stata Experts:

I am working on a project to predict state policy on a particular issue. The objective is more to predict state policy rather than find causal relationships between x and y.

My policy outcome variable has 5 choices, coded from 0 - 4. I am using 0 as the base outcome. I have 10 years of panel data but many of the variables change very slowly over time.

My initial hunch were to use the following models:

1) multinomial logit - since this is not a panel estimator I would have to limit this to a cross-sectional model with only 50 states which I think would be a problem. I am not sure if I can use the panel data as a "pooled" data without significant problems. Grateful for any advice on this.

Also, I was planning on using state and year dummies if I was able to use panel data.

2) xtologit: I could also use the random effects model. I could potentially argue that the outcomes are ordered but I was hoping to avoid this assumption.

3) gesem, mlogit: I am learning about this program right now and it seems promising. Does this seem like an appropriate option? Also, does gsem with mlogit allow me to predict outcome probabilities?

4) I was also advised to look into mixed methods. Any suggestions?

Grateful for any insight on the different choices I have.

Thanks.

SAM

[Clyde Schechter]

I could potentially argue that the outcomes are ordered but I was hoping to avoid this assumption.

Could you argue it with a straight face? Or is it a far-fetched argument?

If the ordering is natural and sensible, then -xtologit- would be my preferred approach. But if the ordering is just some kind of kludge, I wouldn't do it that way. Instead I'd stick with an -mlogit- link.

There is no -memlogit- command, nor is there an -xtmlogit-. But a random effects -mlogit- can be emulated using -gsem, mlogit- (N.B. not gesem), and this is probably what I would do if there is no convincing ordering of the outcomes.

As an aside, there is an -meologit-, but I'm pretty sure that in 2-level data it just estimates the same model as -xtologit-, and I imagine that -xtologit- is faster and less prone to convergence problems.

[Sam McCaw]

Many many thanks for your comments.

I could definitely argue the ordering with a straight face , as it has been discussed in the literature as well.

I will ultimately run mlogit, xtologit, and gsem, mlogit to compare and for robustness but since I wanted some advice on how to decide on the primary and base model your comments are immensely helpful and it seems like an xtologit would be the first go to. Thanks!

Best,

SAM