I estimate an IV model with one explanatory variable that is endogenous (X1), several other explanatory variables (X2), and one instrument (Z) using xtivreg2 (Model 1).
In some specifications (Model 2), I also include the interactions between the variable X1 and one of the variables of interest (X2) (e.g. time dummy variable / continuous variable) – X1*X2. Thus I also add the corresponding interaction of X2 with the instrument – Z*X2 to the xtivreg2 specification.
Under Model 1 (the model with X1 and its instrument only), the Cragg-Donald and Kleibergen-Paap F statistics indicate that the weak identification test is passed (F>20). The Kleibergen-Paap test is my key reference since I use clustered standard errors.
Yet, under the model with interactions between X1 and X2, Model 2, the Kleibergen-Paap F stats is in a range between 0.5-3.0 in most specifications.
1) Can I still report and use the results under Model 2 given that Model 1 is passing the weak identification test? Model 2’s F stats estimate are likely driven due to the inclusion of interactions between X1 and X2 but the coefficient estimates that I receive are statistically significant and make sense. However, I am not sure whether this is a common approach in the IV analysis with such interaction terms, or not.
2) On the other extreme, in some cases I also get Cragg-Donald & Kleibergen-Paap F statistics that exceed values of 500-1000. Should there be any concerns if the F stats under both tests are that high?
I couldn’t find much information in the Stata forum or in the Wooldridge textbook, so I hope you could help.
In some specifications (Model 2), I also include the interactions between the variable X1 and one of the variables of interest (X2) (e.g. time dummy variable / continuous variable) – X1*X2. Thus I also add the corresponding interaction of X2 with the instrument – Z*X2 to the xtivreg2 specification.
Under Model 1 (the model with X1 and its instrument only), the Cragg-Donald and Kleibergen-Paap F statistics indicate that the weak identification test is passed (F>20). The Kleibergen-Paap test is my key reference since I use clustered standard errors.
Yet, under the model with interactions between X1 and X2, Model 2, the Kleibergen-Paap F stats is in a range between 0.5-3.0 in most specifications.
1) Can I still report and use the results under Model 2 given that Model 1 is passing the weak identification test? Model 2’s F stats estimate are likely driven due to the inclusion of interactions between X1 and X2 but the coefficient estimates that I receive are statistically significant and make sense. However, I am not sure whether this is a common approach in the IV analysis with such interaction terms, or not.
2) On the other extreme, in some cases I also get Cragg-Donald & Kleibergen-Paap F statistics that exceed values of 500-1000. Should there be any concerns if the F stats under both tests are that high?
I couldn’t find much information in the Stata forum or in the Wooldridge textbook, so I hope you could help.
Comment