Potential bug: more clusters than observations in cragg hurdle model (churdle).

[Philipp Haendel]

Dear All,

I'm currently using a cragg hurdle model to analyze some survey data using the churdle command.

The data was conducted in different villages in two countries.
I did an overall cluster robust estimation using a cragg hurdle model and clustering on the village level using vce(cluster village).

To get a better understanding of the country specific situations, I also did separate cluster robust estimations for each country, restricting the observations using the "if-command" (if country).

This is where the bug occurred.

As I use more explanatory variables than available clusters/villages for the country specific sample, I chose to cluster on the individual level. This led to the problem that I now have more clusters than observations and the Wald-Chi2 statistic displays as not significant.

see Screen shot:

Clustering on the village level reduces the problem of an insignificant p-value of the wald-chi2 tests, but still it shows more clusters than it should..

The number of clusters represents the number of clusters in the overall sample, but as I only use a subsample, the number of clusters is not fitting.

I assume this is a bug in the programming of this command that needs to be fixed.

I'm a bit confused on how I should proceed as I'm not sure whether I can trust the results, or rather not do cluster robust estimations using a cragg hurdle model.

Any suggestions on that?

[Enrique Pinzon (StataCorp)]

Dear Philipp,

Could you please send your data and a do-file, or a reproducible example, to [email protected]. We will look into it and fix the problem.

Update: This is a reporting issue. The standard errors are correctly computed. The number in the table header is incorrect.