Re: st: Fixed effects and standard errors and two-way clustered SE

[Austin Nichols <austinnichols@gmail.com>]

startistiker <diudas1@gmail.com> :
I would be inclined to use SEs clustered by firm; 14 years is not a
large number for these purposes, but 52 is probably large enough.  Of
course, you should simulate the performance of these SEs using your
data if you want a better answer.


On Wed, Sep 8, 2010 at 4:24 AM, startistiker <diudas1@gmail.com> wrote:
> I have a question regarding standard errors in a two-way fixed effects
> covariance model.
>
> I'm using panel data (firm (52 obs.), year (14 obs.)) which is unbalanced
> (some of the firms have missing firm-years) in a two-way fixed effects model
> (firm and year dummies).
>
> Eq. (1): Y = b1*X1 + b2*X2 + b3*FE(firm)+b4*FE(year)+e
> (intercept is suppressed)
>
> All variables are scaled, the first (b1) is a scaled intercept (1/scale). I
> want to test the effect of a treatment by adding a dummy (treatment=1,
> otherwise 0) for the years a firm is affected by the treatment:
>
> Eq. (2): Y = b1*X1 + b2*X2 + b3*FE(firm)+b4*FE(year)+b5*treatment+e
> (intercept is suppressed)
>
> The whole sample consists of 220 treatment years and 400 non-treatment
> years. Every firm has at least on treatment and one non-treatment year.
>
> Question: Now I am not sure about the correct kind of standard errors I
> should use in my t-statistics?
>
> So far I tried the standard t-test, Huber/White standard errors
> (,cluster(mark) command) and two-way clustered standard errors (Cameron,
> Gelbach, and Miller; cgmreg in Stata).
>
> Question: Which one is recommandable and why (in context of the two-way FE
> model)?
> Results are, as expected, much weaker by using the latter.
>
> Question: To you have general recommendations what to test in this kind of
> approach or to what i have to pay special attention?
>
> Tests I have also done so far:
> - Hausman between FE and RE -> FE recommended (Chi2 11.88)
> - F-Test of year dummies -> inclusion recommended (F-Test 2.99)
> - Wooldridge test for autocorrelation -> AR exists (F-Test 11.590)
> - Breusch-Pagan for heteroskedasticity -> Het exists (F-Test 77.93)
>
> Question: Regarding the Hausman test: should I run it with the treatment
> dummies (Eq. 2) or without (Eq. 1)? At the moment the results for the
> Hausman test above are with usual intercept and without the scaled intercept
> of Eq. 1, is that ok? Can anybody explain to me, why I need a scaled
> intercept? (used it because the original model used it, but more or less
> without explaining).
>
*
*   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/