What Austin summarized was in the sense of my original post.
But a further comment on detecting within-group correlations. So far I have
also read some other papers on the topic but found Petersens' (see below)
hands on approach very intuitive.
Basically you run the model with and without clustered SEs and check if SEs
in the clustered case are much larger which would indicate that you should
cluster in this dimension.
But I would be happy to learn about other approaches.
(e.g. the -iclassr- routine mentioned by Jason. I had a look at it, but how
exactly (maybe a short example) do you apply the routine?)
- Tom
Quote (p.41) :
Thus in addition to providing a guide to the correct estimation of standard
errors, the techniques in this paper can be used to help researchers
diagnose potential problems with their models. By comparing the different
standard errors, one can quickly observe the presence and magnitude of a
firm and/or a time effect. As we saw in Section VI, when the standard errors
clustered by firm are much larger than the White standard errors (three to
four times larger), this indicates the presence of a firm effect in the data
(Table 7). When the standard errors clustered by time are much larger
than the White standard errors (two to four times larger), this indicates
the presence of a time effect in the data (Table 6). When the standard
errors clustered by firm and time are much larger than the standard errors
clustered by only one dimension, this can indicate the presence of both a
firm and a time effect in the data. Which dependencies are most important
will vary across data sets and thus researchers must consult their data.
Petersen, 2006. Estimating Standard Errors in Finance Panel Data Sets:
Comparing Approaches, Working paper.
http://www.kellogg.northwestern.edu/faculty/petersen/htm/working.htm
-----Urspr�ngliche Nachricht-----
Von: [email protected]
[mailto:[email protected]] Im Auftrag von Austin Nichols
Gesendet: Donnerstag, 18. Januar 2007 21:54
An: [email protected]
Betreff: st: panel re vs. fe model
Jason-
I understood Tom's original post to say that he was including firm and
year fixed effects (firm effects via use of -xtreg- and year effects
via a set of indicator variables), and was clustering on firm to allow
arbitrary serial correlation of errors within panel, but was concerned
about the implications of the cluster() option for a downstream test,
which question Mark answered quickly. I.e. he was not confusing the
cluster-robust estimator of the covariance matrix of the estimates
with the estimator that allows for individual-specific intercepts. I
could be wrong, of course.
The trouble he alludes to in a subsequent post is that once you
consider the possibility of arbitrary serial correlation of errors
within panel, it is a small step to worry about correlation across
panels within a year. As far as I know, a cluster-robust VCE
estimator with multiple levels of clustering is only possible via
-svyset- and the -svy- commands, but I would be happy to be
contradicted.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
