[Clyde Schechter]

For this reason it's usually the only accepted choice of estimator in economics, finance or disciplines dealing with observational data.

Well, epidemiology deals primarily with observational data, but the fixed effects estimator is almost never used. One might argue that it should be, but as a matter of practice it is quite uncommon. (Metanalyses are an exception to this generalization, but even there, random effects estimation is more common.) Consistency of estimates is only one of several qualities one can desire from an estimator, and not always the most important. Fixed effects estimators have many limitations and rigidities that make them unsuitable for many purposes.

But I think this advice is misguided in a more fundamental way. The fixed-effects model provides only estimates of within-panel effects. If the research question specifically addresses between-panel effects (as appears to be the case here) then the fixed effects estimator is giving consistent estimates of the wrong parameter. So the inconsistency of an OLS or random effects estimator just has to be accommodated by including as many covariates as you reasonably can and hope that you are left with errors that are uncorrelated or only weakly correlated with the predictors, and then you live with it.

[Morten Gravesen]

Thank you all, it is very helpfull advice.

Philip Gigliotti, I understand the difference in OLS and Fixed effects in that perspective now (within versus between). However, the main variable that I am interested in is the "external weighted average market-to-book ratio (EFWAMB)". The others, including firm size is control variables.

The EFWAMB is estimated for each firm in each year by weighting each market-to-book value by the external finance in any given year. Thus for instance, the EFWAMB for firm A in year 2017 uses the whole panel of market-to-book ratios (14 m-to-b ratios from 2004-2017) in estimating the EFWAMB. For year 2009 it only uses 6 m-to-b ratios (2004-2009).

What I want to test is, if the capital structure of firms today is smaller for firms, that issue equity when their market-to-book values are high.

Thus, I am interested in the differences between firms with large and small EFWAMB ratios in any year. By using fixed effects, I am looking at the variance in the EFWAMB from year to year of firm A and the impact on leverage for that firm. I find it hard to figure it out, but there is just something that tells me, that using fixed effects is not the right choice here, at it measures something wrong.

However, using OLS measures in a correct way I think, but gives me ineffecient estimates because of these fixed unobservable factors,

[Carlo Lazzaro]

Morten:
let's start from square one:
what does the user-written command -xtoverid- tells you if you adopt -re- specification in your panel data regression (I recommend -xtoverid- because -hausman- does not allow non-default standard errors).
As an aside, please also note the -fe- specification (ie, within estimator) gets rid of any observed and unobserves source of heterogeneity related to time-invariant predictors. It does not shelter you from heterogeneity sources related to time-varying predictors.

[Morten Gravesen]

Carlo Lazzaro I ran the test and it gave me something very significan with a p-value very very close to 0. The Hausman test also gives me a very significan result. So I am guessing that these models suggest that I use fixed effects? However, as I am not interested in the within firm effect but the between firm effect (to test if firms with a high EFWAMB has lower leverage), I don't find the fixed effects model useful. Is there any other model, that I can use?

[Clyde Schechter]

You can use -xtreg, be-. This is a pure between-group effects estimator.