[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: IVREG2 resulting in underidentification and CONDIVREG in unbounded intervals
I need to instrument 7 possibly endogenous variables at a time, 5 of them binary, 2 others count variables. I use 17 excluded instruments, mostly cross-products and squares of the exogenous variables included in the main regression equation.
The IVREG2 results are unreliable because of poor instruments. What is the alternative? I tried the following:
Condivreg regressions for each endogenous variable separately: this resulted in unbounded confidence sets. However, if I eliminate instruments from the list, the confidence intervals become bounded. For the elimination, I used the following selection rule: - drop the instrument if it was not significant at the 10% level in the first stage of the IVREG2 procedure. I applied this rule for each endogenous variable, so that I work with different instrument sets for each variable.
This strategy seems very unconventional to me. I wonder whether it makes sense. Also, when instrumenting for each endogenous variable separately, I wonder if I should include the other six possibly endogenous variables in the main regression, i.e. as included exogenous variables. The alternative is to leave them out completely, possible resulting in omitted variable bias.
I am looking forward to receiving some suggestions,
* For searches and help try: