I am trying to estimate a threshold regression model (as per Hansen, 1999; and Hansen, 2000) for panel data using the xthreg command in Stata 14. My data is a firm level survey panel data from 2002-2013, where firms exit the panel quite often. Therefore, using the command as it is (meant only for balanced panel data) not only shrinks down my sample considerably, but I would expect it to introduce some kind of selection bias. This may be a very specific question, but I was wondering if anyone has had the same issue and/or knows a way to deal with adjusting the command to deal with unbalanced data.
What I have tried so far:
a. I can modify the ado-file to not produce an error message when the data is unbalanced and the estimation goes through and even produces estimates. However, just because estimates are produced, I am not sure that they are reliable, since I did not do anything to address the reason why the command was not suitable for unbalanced panels in the first place.
b. I looked for literature that uses the same estimation procedure as the xthreg command on unbalanced panel to see how they had dealt with the issue. The suggestion there was that the issue is with the bootstrap method used to estimate threshold significance and that a cluster bootstrap should be used instead of a wild bootstrap. My issue here is that:
1. I am not very familiar with bootstrapping methods, so I'm not entirely sure what the difference between these two methods is.
2. I cannot change the bootstrap method used in the xthreg command, because the author has not made the source code for the estimation function available.
I am looking for suggestions on how to deal with the unbalanced nature of the data (other than balancing it), or alternatively perhaps someone can point me toward any literature pointing out the issue of using unbalanced panels for estimation of threshold regressions (the article that the command xthreg is following (Hansen, 1999) simply mentions that "It is unknown if the results extend to balanced panels"). Finally, if there is a way to change the bootstrap method without having access to the source code, that would also be an option.
Thank you in advance,
Dea
References:
Wang, Q. (2015). Fixed-effect panel threshold model using Stata. Stata Journal, 15(1), 121–134.
Hansen, B. E. (1999). Threshold e!ects in non-dynamic panels: Estimation, testing, and inference. Journal of Econometrics, 24.
Hansen, B. E. (2000). Sample Splitting and Threshold Estimation. Econometrica, 68(3), 575–603. https://doi.org/10.1111/1468-0262.00124
What I have tried so far:
a. I can modify the ado-file to not produce an error message when the data is unbalanced and the estimation goes through and even produces estimates. However, just because estimates are produced, I am not sure that they are reliable, since I did not do anything to address the reason why the command was not suitable for unbalanced panels in the first place.
b. I looked for literature that uses the same estimation procedure as the xthreg command on unbalanced panel to see how they had dealt with the issue. The suggestion there was that the issue is with the bootstrap method used to estimate threshold significance and that a cluster bootstrap should be used instead of a wild bootstrap. My issue here is that:
1. I am not very familiar with bootstrapping methods, so I'm not entirely sure what the difference between these two methods is.
2. I cannot change the bootstrap method used in the xthreg command, because the author has not made the source code for the estimation function available.
I am looking for suggestions on how to deal with the unbalanced nature of the data (other than balancing it), or alternatively perhaps someone can point me toward any literature pointing out the issue of using unbalanced panels for estimation of threshold regressions (the article that the command xthreg is following (Hansen, 1999) simply mentions that "It is unknown if the results extend to balanced panels"). Finally, if there is a way to change the bootstrap method without having access to the source code, that would also be an option.
Thank you in advance,
Dea
References:
Wang, Q. (2015). Fixed-effect panel threshold model using Stata. Stata Journal, 15(1), 121–134.
Hansen, B. E. (1999). Threshold e!ects in non-dynamic panels: Estimation, testing, and inference. Journal of Econometrics, 24.
Hansen, B. E. (2000). Sample Splitting and Threshold Estimation. Econometrica, 68(3), 575–603. https://doi.org/10.1111/1468-0262.00124
Comment