Re: st: Re: firthlogit

["Joseph Coveney" <jcoveney@bigplanet.com>]

Nyasha Tirivayi wrote:

Thanks for your reply. Are you suggesting that it would be best to
remove the problematic variable from the PSM model, since balance
would still be difficult to achieve even if firthlogit is successful?

--------------------------------------------------------------------------------

No, I wasn't suggesting that.

You'd essentially be relegating the problematic variable to ignorable status 
(as in "unobservable" or "unobserved variable") and appealing to sensitivity 
analysis for salvation.

The option you were considering at first is to leave the problematic 
variable(s) in the propensity score model and match on the basis of predictions 
from -firthlogit-.  Even if you could find matches for the affected 
observations, are you sure that predictions from -firthlogit- deserve to be 
treated as propensity scores in the same manner as predictions from -logit-?

An alternative, perhaps, would be to omit the affected cases so that you
no longer have separation, and then restrict the domain of inference to the
subpopulation represented by the remaining observations.  This might not be 
very appealing from a research-goal prospective . . .

Propensity scores have been around for what, now, a quarter of a century?
I'm guessing that yours isn't the first time separation has cropped up.  The 
literature on propensity score seems to be unsettled in many areas;
nevertheless, it ought to have some sort of discussion of the problem and 
some recommendations or guidance by the experts, however conflicting.

Joseph Coveney


*
*   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/