Clustered Standard Errors

For technical questions regarding estimation of single equations, systems, VARs, Factor analysis and State Space Models in EViews. General econometric questions and advice should go in the Econometric Discussions forum.

Moderators: EViews Gareth, EViews Moderator

mikeguima
Posts: 19
Joined: Wed Mar 14, 2018 8:07 am

Re: Clustered Standard Errors

Postby mikeguima » Fri Apr 27, 2018 10:00 am

Thank you Glenn!

I then noticed from reading other reply of yours here in the forum that in fact the cross section estimators address clustering by time and period estimators address clustering by cross section, as you say.

I have few time periods (12) and plenty of firms (over 2600), but I tested clustering by firm and year and SEs were higher when clustering by time (and pretty much unchanged when clustering by firm), which led me to conclude that perhaps (and following Petersen in "Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches") clustering by time would be best.

Therefore, I'm considering using the cross section estimators you have under the "coef covariance method", which I reckon are robust to heterosckedasticity and cross section dependence (from what I gathered from the manual). I was trying to decide between white cross section and cross section SUR (PSCE), but couldn't quite undertsnad the practical from the manual. Would both be suitable in my case? What would be the differences?

Thank you very much again, Glenn!

EViews Glenn
EViews Developer
Posts: 2671
Joined: Wed Oct 15, 2008 9:17 am

Re: Clustered Standard Errors

Postby EViews Glenn » Fri Apr 27, 2018 5:14 pm

Bear in mind that the large sample properties of clustering-by-cross-section rely on lots of cross-sections, and the opposite for cluster-by-period.

So your performance in allowing for between period correlation (clustering by cross-section) will be much better than in the other dimension. Intuitively, you are computing a moment and if you don't have many observations to compute the moment, then the covariance estimator will not work well.

The PCSE estimators compute a residual moment involving only the residuals and not the data (as you would do if you were doing GLS), then substitute this estimate in the general (non GLS) covariance formula. They allow for a restricted form of correlation and so are "less robust" in that sense, but _may_ have lesser data requirements since the correlation is restricted.

While we don't offer advice on econometrics, I will point out that just because a result is different, does not mean that it is better. You are best off first thinking about what sorts of correlations you would expect to find in your data and why.


Return to “Estimation”

Who is online

Users browsing this forum: No registered users and 15 guests