Multiway Clustering of standard errors with probit regression and panel data

[Christina Chara]

Hello,

I am working with a short individual level panel dataset from a single country in which I merged a number of macro level variables that are grouped at different levels. For example, I have the unemployment level by sex and age, the unemployment level at the time of labour market entry and the percentage change in the supply of labour by occupation.

I am using a panel probit model to find the effect of a number of micro and the above mentioned macro variables on the likelihood of a person being in a job matching their skills versus being mismatched.

Here is the code I used without including the macro variables :

xtprobit mismatched age experience gender disabled married, vce(cluster id)

Now, in addition to the above micro level variables, I want to examine the effect of the macro variables on the mismatched dummy. However, I haven't been able to find a solution to the problem of multiway clustering of my standard errors.

I am aware of the cgmreg ado file used in “Robust Inference with multi-way clustering” by Cameron, Gelbach, and Miller (2006, NBER TWP 327), found on: http://gelbach.law.upenn.edu/~gelbach/ado/cgmreg.ado, however I am not clear as to whether this can be used with probit specifications.
Moreover, given that my aggregate variables are grouped by different combinations of variables, for example "age and sex" for the unemployment level; "year of labour market entry" for the unemployment at labour market entry variable and "occupation" for the percentage change in labour supply variable, I am not clear as to how to choose the variables I need to cluster for.

The 2010 presentation by Baum, Nichols, Schaffer titled: Evaluating one-way and two-way cluster-robust covariance matrix estimates (http://www.stata.com/meeting/boston10/boston10_baum.pdf) also discusses this issue, however I am not clear as how to apply the formula provided in slide 27 in Stata.



Any insight will be greatly appreciated.

[sbrazys]

Justin Caskey at UTA has written an .ado for cgmlogit (among others)

https://sites.google.com/site/judsoncaskey/data

[Christina Chara]

Thank you very much for this. Any idea if anything like this has been written for probit models?

[Roberto Liebscher]

A little late but maybe of interest for other users searching the forum: Mitchell Petersen wrote ado files for probit, logit and tobit models with double clustered standard errors. See here: http://www.kellogg.northwestern.edu/...rogramming.htm

[Bene Eil]

I want to use two clusters for my standard errors.

Is Stata handle it like this:
If I have M1 clusters among the first dimension, and M2 among the second, the number of clusters used is min(M1,M2)?

[Roberto Liebscher]

It depends on what you mean with "Stata". If you cluster on the intersection of the two variables (say states x year) you have M1 x M2 clusters. But in most cases this would be wrong. For an excellent review on this see Collin and Millers's Pracitioners Guide (http://cameron.econ.ucdavis.edu/rese...ober152013.pdf).

If you instead mean their procedure ("cgmreg") then the three VCEs forming the "new" double-clustered VCE $$\hat{V}_{2way}=\hat{V}_1+\hat{V}_2-\hat{V}_{1\cap 2}$$, V1 would have M1 clusters, V2 M2 clusters and the last Matrix of the formula would have M1 x M2 clusters.

[Christian Fink]

How do you integrate fixed effects into the probit2 regression (probit2.ado)?