[Carlo Lazzaro]

Jon:
not quite.
You should consider the 0/1 levels of the variables you have plugged in the right-hand side of your regression equation if you want top calculate the effect on ROA of different scenario.
Basically, other things being equal, you can have four different scenarios -Family_Firm_Identifier- and -Founder_Identity both =1; -Family_Firm_Identifier- and -Founder_Identity both =0; -Family_Firm_Identifier-=1 and -Founder_Identity both =0; Family_Firm_Identifier-=0 and -Founder_Identity both =1. See -predict-; a good exercise is to calculate the fitted values by hand and then compare your calculation with the corresponding -predict- values.

[Jon Hoefer]

Jon:
not quite.
If the regression model does not pass the -estat ovtest- or -linktest- is misspecified.

Thanks once more. Now with regards to improving, if I (rather randomly) ln terms (including y) and add squared terms I get the link test to become insignificant. However, is there any systematic approach / scientific approach to it, which makes it plausibel to transform the variables in a given manner? Before I only had some transformed by ln due to prior research doing so, too or due to being heavily skewed. But for instance, I wouldn't see a reason as to why I ln ROA (other than it makes linktest insignificant).

[Carlo Lazzaro]

Jon:
quant economists/econometricians often present log-linear regression model.
As reported in any decent econometrics textbook, a good reason for logging the regressand (ROA in you case), provided that on their raw scale data are >0, is that you can report in percentage terms the contribution of each predictor (when adjusted for the other ones) to explain the variation in the regressand.