Re: st: Testing joint significance of fixed effects in presence of heteroskedasticity and auto-correlation

[Christian Wagener <wagener.christian@googlemail.com>]

Thanks again for your help, we really appreciate it. Just two more
questions on your comment:

1. The use of "robust bw(4)" is - to our understanding - recommended
for large T, small N samples. Since we have small T and large N, is it
okay to replace "robust bw(4)" by "vce(cluster cluster_id)" in your
example? This has the side-effect mentioned earlier, that dfr drops to
the number of clusters-1.

2. In addition, we wonder why you suggest using "xtivreg2" instead of
"xtreg" in this context? To our knowledge, "xtivreg2" without
specification of an instrumental variable and "xtreg" differ with
regard to the conservatism in the adjustment of df - is that the
reason?

Thanks again,

Thomas and Christian

2010/6/10 Christopher Baum <kit.baum@bc.edu>:
> <>
> On Jun 10, 2010, at 2:33 AM, Christian wrote:
>
>> Just to be sure, we would be testing the joint significance of fixed
>> effects in the presence of heteroskedasticity and auto-correlation by
>> using the following code:
>>
>> xtreg y x, fe vce(cluster cluster_id)
>> scalar rss1 = e(rss)
>> scalar dfr = e(df_r)
>> scalar dfa = e(df_a)
>> regress y x, vce(cluster cluster_id)
>> scalar rss2 = e(rss)
>> scalar fstat = ((rss2-rss1)/dfa)/(rss1/dfr)
>> di "Resid SS with dummies " rss1
>> di "Resid SS without dummies " rss2
>> di "F statistic with " dfa " and " dfr " d.f. = " fstat
>>
>> This example is exactly the one from you we quoted in our initial
>> posting, we just replaced areg with xtreg and robust with vce(cluster
>> cluster_id) to account for auto-correlation. Would this be a viable
>> way or are we overlooking something important?
>>
>> When using vce(cluster cluster_id), we noticed that the rss stay the
>> same as in the case without vce(cluster cluster_id). Only the df_r
>> drop down to the number of clusters - 1 and equals the df_a. The same
>> happens when using areg instead of the xtreg-command.
>>
>> Is this drop of df_r a problem for the test of fixed-effect
>> significance we are looking for?
>
> I don't think so. The cluster-robust estimator does allow for arbitrary autocorrelation within clusters, but I think if you want to test in the context of a HAC estimator, you should use one. E.g., with webuse grunfeld and Schaffer's -xtivreg2- from SSC,
>
> xtivreg2 invest mvalue, fe robust bw(4)
> scalar rss1 = e(rss)
> scalar dfr = e(Fdf2)
> scalar dfa = e(df_a)
> ivreg2 invest mvalue, robust bw(4)
> scalar rss2 = e(rss)
> scalar fstat = ((rss2-rss1)/dfa)/(rss1/dfr)
> di "Resid SS with dummies " rss1
> di "Resid SS without dummies " rss2
> di "F statistic with " dfa " and " dfr " d.f. = " fstat
>
>
> Kit Baum   |   Boston College Economics & 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/