Panel data analysis with timevariant and time invariant variables

[Sabine Anna]

Hello everyone,

I am working on my master thesis about family firms in Germany, it is my first empirical work and right now I am struggeling with the analysis method I could use.
My aim is to see wheter family firms are competitive towards non family firms. I have unbalanced panel data for 10 years (measured yearly) for 90 firms.
Return on asset is my dependen variable. I chose a dummy for family firms (1=yes it is a family firm , 0= no it is not), capital expenditure, debt to equity, r&d intensity, beta and tobin's q as independent variables. After doing the Hausmantest I knew that I had to use the fixed estimation but in this method I can not include my dummy for family firms it is time invariant but this is the effect which I definitely need to control for.
I looked at some multilevel regressions but those didn't seem right. Now I did two regressions one for only family firms (roa if fb==1) and one for only non family firms (roa if fb==0) and now i want to compare those results (see picture below) . Still this approach seems not statistically correct. What do you think can I do it this way? Or can you think of any better idea how to address my issue? I am very thankful for every idea or comment!

Kindest regards and thanks in advance
Sabine

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - include Stata code in code delimiters, readable Stata output, and sample data using dataex. Avoid pictures.

You don't want to use firm fixed effects because the variable of interest is a stable firm characteristic. You may also want to think about your control variables. To the extent that family firms do things differently, their difference in outcome might come through that variables you are using as controls. Then your dummy for family firm would misrepresent the effect. What you may really have is a mediated model where family influences these firm behaviors and family and these firm behaviors influence performance. I'd also be careful about including a performance measure like Tobin's Q on the rhs in a model that attempts to explain performance.

For a master's thesis, I'd propose a panel estimate using random effects for firms. There are fancier ways to attack the problem, but they are somewhat more complicated.