I want to employ the "local projections method" (LPM) by Oscar Jorda (AER, 2005) to panel data, in a similar way as Gourio, Messer and Siemer in their paper "Firm "Entry and Macroeconomic Dynamics: A State-Level Analysis" (AER, 2016), also freely available as a Washington Fed working paper.
I know that LPM has some advantages over VARs and one of them is that it is more robust to various types of misspecifications. But I'm curious how is it possible that Gourio et al (2016) use a mixture of stationary and non-stationary variables. (I have their replication files and I performed panel unit root tests to make sure, they didn't bother). Also, other papers that employ LPM don't seem to be concerned with the stationarity issue so probably there is a good and obvious reason.
My concern is due to the fact that LPM relies on a simple OLS estimator (only with two-way clustered standard errors).
Can someone please explain it shortly?
This may be due to the fact that if I(1) variables cointegrate, there is no problem with including stationary variables in system. The crucial question concerns cointegration. What is the result of the Kao or Pedroni test for cointegration?
czesc Wojtek ;-)
I didn't run cointegration tests - there are quite a few models which they estimate. The point is that they don't seem to care about stationarity at all. Neither in the paper, nor in their codes I see any track of performing stationarity tests, let alone cointegration tests. I attach the paper, it's very short.
Thanks!
Dear Pawel,
I am not sure about the reliability of your results in this case. Mixing stationary and nom-stationary variables may generate unfavorable bias in your model. Please pay attention to this issue.
Regards,
Ehsan
Distinguish between the two questions of (I) is the model logically coherent and (ii) are the statistical methods reliable.
(I) if the LHS variable is I(1) then a mix of I(1) and I(0) on the RHS is logically coherent. If the LHS is I(0) then it is illogical to have I(1) variables on the RHS unless they cointegrate.
(ii) I imagine that the consequences (for inference) of non-stationarity in the time dimension become less problematic as the x-section becomes wider relative to the sample time period. (This is a guess so I would be grateful for comment on this point.)
- Badges
- Science method
- Similar topics
- Methodology
- Research Methodology
- Qualitative Research