Panel data analysis: Random effects model - controlling for year and industry possible?

[Guest]

Hello everyone!

I am new to the forum and have been going through a lot of past post, but I just haven't found the right answers to the questions, I am facing as part of my master thesis project. I will be very grateful for any help and hints!

I have a data set with roughly 1200 observations across 99 companies and 13 years, so panel data. My two main IV's are "task-related diversity" and "relations-oriented diversity" (on the TMT) on my DV "short-term orientation". After intensive research and consultation with my supervisor, we came to the conclusion that a random, or fixed effects model (determined by the Hausmann test) would be the most suitable analysis. Results of the Hausman test for my data indicated random effects to be the right model. In addition, our supervisor suggested to use the command vce(robust) to account for heteroskedasticity and auto correlation.

So far so good, if I don't want to control for industry differences by including the industry dummies (11 industries -1, so 10 of them), everything works "fine" (insignificant but seem reasonable). However, for both some IV's (ROA or number of employees), as well as for the DV (short-termism), there may be differences across industries (certainly the means differ, I checked that with an ANOVA), so I want to control for that. So I tried to options:

1) Including industry dummies in the xtreg re vce(robust): This results in getting no test statistic (Wald Chi2), which may be because industry for a certain company doesn't change over time. So this isn't really an option.

2) My supervisor suggested to industry-adjust the performance IV's, which would be mainly ROA and then don't include any dummies. This does not seem very logical to me, because then I still don't account for industry differences on my DV, nor any other variables in the model. In addition, using the variable ROAt-3industryadjusted turns out highly insignificant, while ROAt-3 not adjusted is highly significant.

What would be possible options for me to stick to OLS random effects model (if possible as it seems the right choice) while controlling for year and industry clusters?

Please check my attached screenshots, I think this will make things a lot clearer. I am very lost and would highly appreciate any help! I hope I have explained things sufficiently.


1) Random effects model including industry dummies; no test statistics available (missing WaldChi2)

2) Random effects model with industry dummies but without command vce(robust)

3) Random effects mode without industry dummies, instead ROA industry-adjusted (as suggested by my supervisor):

[Guest]

UPDATE and potential solution: I have now eliminated 4 specific companies that were the only ones in their industry, respectively. And the problem seems to be revolved (see picture). Although its unfortunate that I have to drop 4 companies and have less industries representative in my analysis, this could be an option. Do you agree?

[Carlo Lazzaro]

Guest:
I would go Random effects model with industry dummies but without the -vce(robust)- option, as the defauls standard errors do not seem that different from the robustified ones (please note that, unlike other commands, such as -regression-, all over the -xt-realm the option -robust- and -cluster- work the same).

[Guest]

Grazie mille Carlo! I agree with you and will use the RE model. I also run various tests, including Hausman, and xtoverid with and without robust, and all indicate RE.
In addition, I have been wondering if it makes sense to include i.FiscalYear, as it might have an effect on Short-termism across year (financial crisis etc.), that is possible in RE and the model does NOT automatically account for differences across year - does it? Thanks a lot!

[Carlo Lazzaro]

Guest:
you can try adding -i.FiscalYear- and see what happens.