Panel of firms - FE/RE model specification

[Piotr Szafranski]

Dear Experts,

I want to evaluate the effect of economic policy uncertainty on investment in various business sectors. I have an unbalanced panel of 1500 firms from 11 sectors for the 1990-2019 period. To conduct the analysis I perform regressions in each of the sectors. I would like to seek your help regarding the model specification.

Dependent variable: CAPEX/Total Assets
Main independent variable: economic policy uncertainty
Control variables: Firm-level control variables
Macroeconomic control variables

Considering the highly heterogeneous nature of the panel data I try to control for the investment opportunity and capacity, but still there might remain unobservable effects that vary across firms but are constant over time. Is it a valid approach to simply conduct a Hausman test to establish whether to estimate with RE or FE or maybe I should consider other aspects?

[Piotr Szafranski]

EDIT: The Hausman test indicates FE and RE specifications depending on the sector.

[Carlo Lazzaro]

Piotr:
I would not peform separate regression for each Sector; I would add -i.sector- as a predictor in the right-hand side of the regression equation, instead.
After running -xtreg-, I would check -hausman- outcome and decide about -fe- or -re- specification accordingly.

[Piotr Szafranski]

Dear Carlo,
thank you very much for your answer. I believe my data is heterogeneous on both firm and sectoral (please, see the individual regressions attached) level. In the fixed effects model firm and industry effects are perfectly collinear, so they cannot be included together. In the estimation of the impact of policy uncertainty on capital expenditures of firms in certain sectors why is it preferred to perform a regression for entire panel with i.sector variable?

[Carlo Lazzaro]

Piotr:
if -i.sector- is, at it seems, a relevant predictor, running separate regressions may lose relevant piece of information.
That said, if you have more than one fixed effect to be investigated, you may want to take a look at the community-contributed programme -reghdfe-.

[Phil Bromiley]

What is your panel control? Is it the sector or the firm? If it is the firm, then a sector panel control is unneeded - firms almost never change sector. But, following Carlo's line of thinking, you could use sector to allow different parameters on the other variables for each sector [e.g., xtreg y c.x1 c.x2 i.sector#(c.x1 c.x2) ]