Use of vce(cluster) legitimate?

[Ellen Sterk]

Hi all,

I am analysing a Discrete Choice Experiment with the cmxtmixlogit command. I had three different types of construction clients in my sample, which is why I thought to use "vce(cluster type of client)". However, now I came across entries that said the number of clusters should be "sufficient" and three did not seem to meet this criterion. So should I just use "vce(robust)" instead? How can I control for the different groups I have?
Note: the experiment was unlabelled, so entering type of client as a case variable is meaningless.

Thank you!

[George Ford]

3 too few. But could try it then boottest after.

[Carlo Lazzaro]

Ellen:
welcome to this forum.
With three clusters only, -vce(cluster clusterid)- standard errors are surely misleading.
In addition, -vce(robust)- won't help either, as quoting -cmxtmixlogit- entry, Stata .pdf manual:

Specifying vce(robust) is equivalent to specifying
vce(cluster panelvar), where panelvar is the variable that identifies the panels.

[Ellen Sterk]

Thank you, George and Carlo, for your answers! That already helps a lot.
@Carlo: the variable that identifies my panel is actually the individuals (> 1000), they all made 8 choices. Then -vce(robust)- would be ok, right?

[Carlo Lazzaro]

Yes, but -vce(cluster clusterid)- would work out fine, too.

[Miguel Henry]

Kezdi, Gabor. 2004. “Robust Standard Error Estimation in Fixed-Effects Panel Models.” Hungarian Statistical ReviewSpecial(9): 96-116 shows that 50 clusters (with roughly equal cluster sizes) are often close enough to infinity for accurate inference.

[Jeff Wooldridge]

I'll just add to the helpful remarks: You need a good reason to cluster on a variable. I don't see how clustering on type of client makes sense. Why not cluster on, say, years of schooling? Race? Experience in the workforce?

You have a large number of individuals and, as best I can tell, the assignment of the scenarios was done at the individual level. If so, cluster at the individual level and no more.

[Ellen Sterk]

These are very helpful, thank you all!