Re: st: GLS estimator for xtivreg?

[Christopher Baum <kit.baum@bc.edu>]

On Dec 9, 2011, at 2:33 AM, Philip wrote:

> I have multiple observations per firm and a large number of firms.  Firms vary massively in scale. The model includes one endogenous variable.  I'd like to do xtivreg with a gls correction for the scale differences across firms.  As I understand it, the robust estimator fixes some problems with the standard errors, but leaves the large firms dominating the estimates of the betas.
> 
> It seems like I could either (i) do a pre-estimate using xtreg, xtivreg, or xtivreg2 to predict residuals which I then use to estimate firm standard deviations followed by a xtivreg2 with weights, or (ii) do a separate estimate of the instrument for the endogenous variable followed by xtgls using the instrument allowing for heteroskedasticity by panel.

What variables are differing widely in scale? Usually when working with firm-level panel data we utilise ratios to get around these problems: e.g. rather than looking at capital investment, we consider the I/K ratio (where K is the capital stock), or rather than output, we consider output per worker, or return on assets rather than dollar-value net earnings. That in itself deals with much of the scale-related heteroskedasticity.

I would recommend using ratio variables, and employing Mark Schaffer's -xtivreg2- from SSC to estimate IV models for firm-level panel data. You should also consider Mark's -ivhettest- which can provide consistent tests of heteroskedasticity in an IV framework. Both are described in Baum, Schaffer, Stillman, SJ 2007, but you should download the SSC versions of these routines, as they have been updated since the SJ article.

Kit

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/