RE: RE: RE: st: Robust Standard Errors in Paneldatasets

[Amy Dunbar <Amy.Dunbar@business.uconn.edu>]

Thank you so much, Kit.  I am learning so much from statalist.  I appreciate the time everyone takes to help out those of us who struggle with comprehending the language of econometrics.
Amy Dunbar

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Christopher F Baum
Sent: Tuesday, October 26, 2010 3:14 PM
To: statalist@hsphsun2.harvard.edu
Subject: re: RE: RE: st: Robust Standard Errors in Paneldatasets

<>	
Amy wrote 

Petersen wrote (p. 458):

One way empirical finance researchers can address two sources of correlation is to parametrically estimate one of the dimensions (e.g., by including dummy variables). Since many panel data sets have more firms than years, a common approach is to include dummy variables for each time period (to absorb the time
effect) and then cluster by firm (Lamont and Polk, 2001; Anderson and Reeb, 2004; Gross and Souleles, 2004; Sapienza, 2004; and Faulkender and Petersen, 2006). If the time effect is fixed (e.g., Equation (15)), the time dummies completely remove the correlation between observations in the same time period.
In this case, there is only a firm effect left in the data. As seen in Section 1, OLS and Fama-MacBeth standard errors are biased in this case, while standard errors clustered by firm are unbiased (results available from the author).

Stas pointed out that clustered SEs are never unbiased. Point well taken. I would not suggest they are.

The above statement by Petersen makes no sense to me. Compare the contemporaneous correlations ("correlation between observations in the same time period") for a one-way FE, two-way FE and the strange creature he suggests: one-way FE with time dummies, clustered by firm.* In all three cases the contemporaneous correlations are quite similar. The contemporaneous correlation cannot be removed by including time dummies.

This should not be surprising, for correlation within a firm's observations is not removed by including firm dummies (fixed effects); you need to use the cluster VCE to allow for such correlations (and as Stock/Watson have shown, that is mandatory unless you assume iid errors). So why would you expect that allowing intercepts to shift for each time period would have the effect of removing correlations across firms for each time period? I'm sorry, but that quote from a published article is just plain wrong.


* do http://fmwww.bc.edu/cfb/stata/twowayFE.do          to see this demonstrated.

Kit

Kit Baum   |   Boston College Economics and DIW Berlin   |   http://ideas.repec.org/e/pba1.html
An Introduction to Stata Programming   |   http://www.stata-press.com/books/isp.html
An Introduction to Modern Econometrics Using Stata   |   http://www.stata-press.com/books/imeus.html


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/