st: Re: Suggestions On Event Study Implementation Using Sureg

[Kit Baum <baum@bc.edu>]

<>
(1) I do not understand why the dummy is being dropped. Even if it was  
a singleton dummy it would work (useless, but it would compute). There  
must be something idiosyncratic about the way you've set up the data.  
I would look at, say, 3 firms and see if it happens there too with  
your original setup of a dummy hitting for two days only. In an  
example like this there is no problem:

webuse grunfeld,clear
reshape wide invest mvalue kstock, i(year) j(company)
g onedum = (year>1953)
forv i=1/10 {
	local eqn "`eqn' (invest`i' mvalue`i' kstock`i' onedum)"
}
sureg `eqn', noh

(2) -sureg- cannot estimate a model in which there are more equations  
than time periods, or in your case more firms than days. As long as  
your timespan is long enough, perhaps not a problem.

(3) -sureg- would seem to be an attractive way to look at how an event  
impacted returns, allowing for he return generating process to differ  
across firms. As implemented in Stata, though, -sureg- is not a very  
attractive estimator, as unlike most other Stata commands, it does not  
allow you to specify the covariance structure (as both -sureg- and - 
reg3-, on which it depends, are Version 6 commands, that's not too  
surprising: something about which I have grumbled on numerous  
occasions). There is no reason why -sureg- (or -reg3-) cannot offer  
the same array of covariance structures as do other least-squares- 
based commands, but the fact of the matter is that it does not. So  
robustifying the estimates is somewhat problematic. If you could, for  
instance, justify constraining the regressors' coefficients across  
equations, you could use -xtivreg2- with firm-level clustering, allow  
for autocorrelation, etc. That would entail calculating a single  
coefficient for the dummy, rather than firm-specific coefficients as  
you are trying to do.


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


On Nov 7, 2008, at 02:33 , Tom wrote:

> Sorry I missed your reply until now.  Actually, the code below places
> a 1 in days 1381 and 1382 and even if I extend the event period to
> twenty days it still drops the dummy variable in the results.  I found
> a couple papers that recommended this methodology (seemingly unrelated
> regression and a dummy for the event days)  for regulatory type
> events.  Further, for similar firms (all of my firms are financial)
> with identical event windows this method (using -sureg-) allows one to
> reflect the time series and cross-sectional relationship of the errors
> in the model but still provide differing loadings on the independent
> variables.  The "standard" method leaves me with either measuring the
> standard deviation of the abnormal returns for the event period across
> very few observations if I do it firm by firm or else having to
> combine all abnormal returns of the firms in the measurement of an
> event abnormal return standard deviation.  It is also unclear how to
> adjust the errors unless I use something like -pcse-.  Most of the
> papers utilizing the standard event study approach I reviewed were
> able to rely on some kind of randomness in the individual events (such
> as earnings or dividend announcements) across firms as opposed to one
> event for all firms as well as diversity in the firm type.
>
> 1.  Is there any reason you would expect the dummy to be dropped for a
> longer event window from a -sureg- design viewpoint as opposed to a
> statistical measurement viewpoint?
>
> 2.  I don't understand your comment that -sureg- prevents me from
> examining any number of firms
>
> 3.  Do you have any suggestions or are you familiar with any standard
> event study approaches that have similar characteristics to my problem
> (short event window, same window for multiple firms, all firms of very
> similar error structure - heteroskedastic, strong cross sectional and
> time series correlation, etc.)
>
> 4.  Would you simply use firm specific abnormal return standard
> deviations or combine into one abnormal return standard deviation
> across all firms?

*
*   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/