Coefficient sign flips after applying Fixed Effects

[Dimitar Chernev]

Good day everybody,

I am currently writing my master thesis and I research whether family owned companies (FFF=1 if family owned, 0 otherwise) have a positive or negative effect on the performance of the company and according to the literature I should expect a positive relationship (reasoning: family wealth is tied to the financial wealth of the company, long term investments, less agency costs, maybe less consumption on the job, better monitoring etc.).

My dependent variable is company performance (approximated by Tobin's Q) and my independent variables are CashFlow, Dividends, TotalAssets, TotalDebt, RoA and dummy variables for the year (2007-2016) and the Industry.

The panel data contains around 25 000 observations for companies from 17 European countries for the period between 2007 and 2016.

Firs I ran a OLS regression and found that the coefficient of FFF is negative and significant. Than I tested the OLS-assumptions. The Breush-Pagan test suggested possible presence of heteroskedasticity, so I conducted the Hausman test for fixed vs. random effects which suggested that I need to use Fixed Effects.

Now when I apply Fixed effects to the same variables from the OLS model (excluding the Industry dummies) I find a positive relationship between Performance and the Family fummy (FFF). Does anybody know why the coefficient sign of FFF changes eventhough the variables I use in both regression are the same?

Kind regards,
Dimitar

[Carlo Lazzaro]

Dimitar:
-first off, you seem to mistake -estat hettest- (that actually checks whether the OLS residual distribution shows evidence of heteroskedasticity) with the -xttest0- (that actually checks whether -xtreg, re- shows shows evidence of random effect);
- -hausman- is (asymptotically) helpful to test -re- fits your data better than -fe- (or is the other way round). Unfortunately, if you detect heteroskedasticity after -xtreg- and you wisely invoke non-default standard errors, -hausman- is no more an option and you should replace it with the community-contributed command -xtoverid- (which, in turn, does not support -fvvarlist- notation);
- I find difficult to comment any further without taking a look at what you typed and what Stata gave you back (as recommended by the FAQ).

[Marcos Almeida]

Just a side note, after Carlo's excellent advice.

Even if the main predictors are the same, there is a difference (in perspective) between OLS and fixed-effect models

You may wish to read this thread.