|From||"Zezza, Alberto (ESAE)" <Alberto.Zezza@fao.org>|
|Subject||st: Overidentification test in clogit|
|Date||Wed, 18 May 2005 18:19:05 +0200|
I use Stata 8 for Windows.
I am trying to estimate a model that explains the choice of selling on a specific market (in my case on the farm, in the local market, on a distant wholesale market) on a number of (a) farmer, (b) market, and (c) transaction characteristics.
Data are by transaction, some households sell their harvest in more that one transactions so I have 120 households and 280 transactions approximately.
One of my explanatory variable in the conditional logit is quantities sold which is endogenous to the market choice (I decide where and how much to sell simultaneously). I therefore want to instrument the quantity and have estimated a predicted value with a first stage OLS regression. My questions:
1) How can I test that the instruments are valid, that is that the exogenous variables included in the OLS are indeed exogenous? In a normal 2sls this is trivial, but in this type of mixed ‘clogit’ (see Long & Freese, p. 243) data are transformed so that the number of observations is now different from the first stage OLS model (observations are multiplied by the number of alternatives).
2) Is the fact that both my regressions are clustered (to control for household fixed effects) a problem?
3) Commands like overid use a Sargan test that is based on R2. How does the fact that clogit yields a (variety of) pseudo-R2 (with all their problems) affect the possibility of meaningfully performing a test along those lines?
4) Related to that, is it possible to save residuals in clogit?