Unknown structural break in Panel data

[NENG ZHAP]

Dear Experts,

"fresh on board".

I have an unbalanced panel data of all NYSE, AMEX and NASDAQ firms from 1998 to 2018. After doing the F-test, Hausman test, the results suggest me to use Fixed Effects.

However, most of my main explanatory variables get the wrong sign (I expect a + sign, but it is a - sign) and are not statistically significant; whereas the coefficients from using the Pooled OLS and Random Effects have the correct sign and are statistically significant.

My questions are:

1. what should I do next in ur opinion?
2. I expect to have a sub-period analysis. (maybe the above issue is due to a structural break?) I have checked many previous topics about testing unstructured breaks in panel data, but could not find a direct answer. How can I test an unknown structural break in panel data?
3. BTW, I got negative chi-square in some of the Hausman tests.

Hope to hear from you soon.

Thanks in advance.

Neng

[NENG ZHAP]

BTW, I have controlled for time and industry dummies in my Pooled OLS and Random Effects, while Time dummy only for Fixed Effects.

[Phil Bromiley]

See my comments in:
https://www.statalist.org/forums/forum/general-stata-discussion/general/1550130-heteroskedasticity-or-other-problem-with-regression

While there are some automated ways to identify structural breaks, I don't know that they have been programmed for panel data. Furthermore, with your large sample, it would be problematic to run the regression with all the dummies for each firm which is what would be a precursor to most of the structural break analyses.

However, you only have a 20 year time series. You can run the model 19 times with 19 different breaks and look at whatever criteria is normally applied to identify structural break. Often, it is easiest to do this sort of thing by including a dummy for the break, and interacting it with the right hand side variables. This then gives you an automatic test whether the parameters differ.

[NENG ZHAP]

@Phil Bromiley Thanks for your help.

I could check the structural break manually for 20 times, however, I have 7 different measures for my dependent variable and I am not sure if I need to check each of them for 20 times (maybe they don't have the same break?).

Or probably it will be better for my situation to find some reference for a structural break as you said there doesn't have any programmed code for an unknown structural break test in panel data in STATA.