OLS with Fixed Effects for Sparse Panel data

[Richard Bui]

Hi Statalist,

I am running a simple OLS on a time-series panel data in which each firm-year is unique. However, my dataset is quite sparse with many firms having only a few observations.
Given that, I've been contemplated whether to incorporate FEs into my model (specifically firm FE).
My concern is, as the within-firm variation is low (due to sparse data) compared to the between-firm variation (also, the key predictor is an event that rarely occurs in any given firm), adding it can result in a loss of statistical significance.

As a first-year research student, I am still quite new to modelling.
Any advice/insights on this is very much appreciated. Thank you.

[Daniel Feenberg]

You are right to be concerned, significance may fall dramatically. However, the only way to find out is to give FE a try. Note that there is no law that says every possible fixed effect has to be estimated. And even if it is, does the new esimate of the coefficient of interest overlap with the OLS estimate? If so, the FE hasn't really contradicted the OLS estimate even if it is insignificant. Perhaps you can argue that the inter-firm effect is a legitimate source of variation.

[Carlo Lazzaro]

Richard:
did you run the -hausman- test (or the community-contributed module -xtoverid-, if you imposed non-default standard errors)?

[Richard Bui]

Richard:
did you run the -hausman- test (or the community-contributed module -xtoverid-, if you imposed non-default standard errors)?

Thanks for the suggestion, Carlo. If my understanding is correct, you wanted me to check whether FE or RE is more appropriate, right?
Unfortunately, in my area of research, RE isn't typically preferred, so I am leaning toward FE regardless.

[Richard Bui]

You are right to be concerned, significance may fall dramatically. However, the only way to find out is to give FE a try. Note that there is no law that says every possible fixed effect has to be estimated. And even if it is, does the new esimate of the coefficient of interest overlap with the OLS estimate? If so, the FE hasn't really contradicted the OLS estimate even if it is insignificant. Perhaps you can argue that the inter-firm effect is a legitimate source of variation.

Thank you, Daniel. Sadly the results turned completely sideway when I added firm FE, but not year and/or industry FE.
I understand there is no law that we have to estimate every possible FEs, yet it's quite common practice to include especially firm FE in my area of research.
Do you think my sparse data (many firms having only a few observations) is a valid reason to argue against the inclusion of firm FE then?

[Carlo Lazzaro]

Richard:
exactly.
I'd go with two-way -fe- (-panelid- and -year-; the latter can be estimated adding -i.year- in the right-hand side of your -xtreg,fe- equation).
That said, if -re- is the way to go, your -fe- estimator remains consistent but inefficient.
As a sidelight, please share what you typed and what Stata gave you back to increase the chance of getting (more) helpful replies (as per FAQ). Thanks.