Controlling for both time fixed effect and "entity invariant" variables

[Abdan Syakura]

Dear all,

I have a question about the panel data regression model in general. I have firm-level data in a single country. Is it okay to control for both time fixed effect and entity-invariant variables, such as GDP growth and interest rate (which are the same across firms but vary across years)? Ideally, I want my model to be as parsimonious as possible, i.e., without controlling for the entity-invariant variables. However, my regression result without including them is quite bad (none are statistically significant), while the results are significant by including them. I am not sure whether the latter result is valid. Thank you.

Best regards,

[Carlo Lazzaro]

Abdan:
1) -xtreg,fe- requires -i.timevar- among your predictors;
2) if GDP growth and interest rate are time-varying, it may make sense to include them in the right-hand side of your regression equation.
On a more general note, your specification choice should not be driven by the statistical significance, but by the goal of giving a fair and true view of the data generating process tha you're investgating.

[Abdan Syakura]

Thank you, Carlo. Just to confirm: so it's not wrong to include a time-varying yet entity-invariant variables, such as GDP growth rate and interest rate even though I have controlled for the time fixed effect? The model is like this:

 xtreg investment leverage gdpgrowth irate i.year, fe

Where investment and leverage are firm-specific.

I sometimes hear including time-varying yet entity-invariant variable is unnecessary after controlling for the time fixed effect. This is because all entity-invariant effects are already accounted for by the time fixed effects. Thanks for your help!

[Carlo Lazzaro]

Abdan:
it would be great if you could share not only what you typed but also what Stata gave you back. Thanks.

[Jeff Wooldridge]

Actually, you can’t include both. The best you can do is i.year. I bet in your output some of the year dummies were dropped. The coefficients and the statistical significance of the other variables are meaningless. This is why we ask to see output and not just commands.