On Tuesday, July 1, 2003, at 02:33 AM, Vince wrote:
Mark Schaffer <[email protected]> is estimating a model with an
indicator
(dummy) variable that is 1 in only a single observation and 0
everywhere else
and he wants an explanation for some things he notices about the
variance-covariance matrix,
Thanks to Vince Wiggins for his cogent explanation of what is going on
in this case (where the VCV becomes less than full rank). He and I
corresponded offlist about the results I had posted
to Statalist using RATS, in which I found it strange that the robust
standard errors did not agree, even in the case where the dummy (sorry,
indicator) is excluded. We agreed that the difference in the latter
case (e.g. regress weight length, robust) was a degrees-of-freedom
issue. If one scales RATS' robust covariance matrix to divide by (N-k)
rather than N, the two programs agree exactly.
With the dummy included, they do not; but inverting RATS' covariance
matrix leads to a (slightly different) matrix of rank 2. Since RATS
indicates that a generalized inverse is being used, and there is more
than one way to generate a g-inverse, close in this case may be good
enough (as it is in the case of horseshoes and hand grenades).
Kit
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
