Dear all,
I am currently working on my master's thesis where I estimate a FE model which looks the following:
y_it = a_0 + a_i + x_it*beta1+ crisis*(d_0 + d_i + x_it*beta2) + epsilon_it,
where "crisis" is a dummy that has value one with the beginning of a macroeconomic crisis. "a_i" and "d_i" are constant fixed effects and "x_it" explanatory variables. The model allows to investigate changing coefficients "beta1" to "beta2" when the crisis occurs.
When I wanted to perform tests like xtserial or do multiway clustering with xtivreg2 that was not possible because factor variables are not allowed in these cases. My idea to approximate the results is to drop the crisis interactions in my model and estimate
y_it = a_0 + a_i + x_it*beta1 + epsilon_it
once for the crisis and the pre-crisis period. Though coefficients are the same as in the complete model (robust) standard errors slightly differ. I then get test results that in both periods serial correlation is present and would use these seperate results to conclude that I have to account for serial correlation in the original model.
I have two questions on that:
1) Why do standard errors differ between the two approaches while coefficients do not? Is it plausible or something wrong with my model?
2) Is it a legit approach to deal with the fact that factor-variables are not allowed in the commands? As a professor reading the thesis: Would you consider this idea to be unscientific and not precise enough or to be a suitable idea?
Any help is very much appreciated.
Best,
Chris
I am currently working on my master's thesis where I estimate a FE model which looks the following:
y_it = a_0 + a_i + x_it*beta1+ crisis*(d_0 + d_i + x_it*beta2) + epsilon_it,
where "crisis" is a dummy that has value one with the beginning of a macroeconomic crisis. "a_i" and "d_i" are constant fixed effects and "x_it" explanatory variables. The model allows to investigate changing coefficients "beta1" to "beta2" when the crisis occurs.
When I wanted to perform tests like xtserial or do multiway clustering with xtivreg2 that was not possible because factor variables are not allowed in these cases. My idea to approximate the results is to drop the crisis interactions in my model and estimate
y_it = a_0 + a_i + x_it*beta1 + epsilon_it
once for the crisis and the pre-crisis period. Though coefficients are the same as in the complete model (robust) standard errors slightly differ. I then get test results that in both periods serial correlation is present and would use these seperate results to conclude that I have to account for serial correlation in the original model.
I have two questions on that:
1) Why do standard errors differ between the two approaches while coefficients do not? Is it plausible or something wrong with my model?
2) Is it a legit approach to deal with the fact that factor-variables are not allowed in the commands? As a professor reading the thesis: Would you consider this idea to be unscientific and not precise enough or to be a suitable idea?
Any help is very much appreciated.
Best,
Chris