Panel Data - Fixed Effect model with low within variation

[Stan Johnson]

Dear all,

I have read several similar posts about this topic. Still, I havn't found a clear answer. Thus, I hope you can help my with this matter.

For my thesis, I'm using panel data (8years) to analyse the relationship between different kind of ownership types (dummy variables) and firm performance.
These different ownership types are not completly time-constant variables. Nevertheless, they are relatively stable and thus, show little within variation (5%-25%).

As I want to consider unobserved heterogeneity at the firm level, I want to conduct a fixed or random effect regression model.
However, according to Hausman test, I should use a fixed effect model.

As the within variation is relatively small, I am concerned that the model would not be appropriate.
In contrast, the random effect model seems to the wrong choice as well as there seems to be a correlation between individual-specific effects and some of the explanatory variables.

While looking for a solution, I found the approach of Mundlak (1978). However, it seems that this approach is not common in my field of research (Impact of ownership structures on corporate decisions).

Would it maybe a solution to conduct both, the random and fixed effect model, while considering their advantages and disadvantages?
It seems that there is no optimal solution.

Thank you very much for your help.

[Alberto Camus]

As you mention, there are problems regarding the use of FE vs. RE that go beyond what the Hausman test can possibly tell you to do. It is quite common practice, especially if you are unsure about what to use, to run the three models and present all of them in your results (OLS vs. FE vs. RE) like in the table below, and then briefly explain in your discussion the problems associated with each. I am not familiar with Mundlak's approach -- could you explain it?

(Taken from Hartwig, 2011)

[Stan Johnson]

Thank you very much for reply.

I do not understand the Mundlak approach 100%. But the basic idea is that you can use a random effect model even if the Hausman test suggest the opposite. Therefore, you relax the assumption of independece between the time-varying characteristics and the error term by including in the model the individual mean of the time-varying covariates.

However, as I did not find a empirical paper which used this approach, I did not analyse this approach in greater detail.

[Carlo Lazzaro]

Stan:
the following references cover Mundlak correction:
https://uk.sagepub.com/en-gb/eur/a-p...ta/book237578: 192-3; 205-9.
http://www.springer.com/us/book/9783642329135: 164
http://econ.msu.edu/faculty/wooldridge/docs/cre1_r4.pdf

[Stan Johnson]

Thanks Carlo.

Would you recommend it although it is not common in empirical papers?
I look forward to hear from you.

[Carlo Lazzaro]

Stan:
I would consider it an option, especially when -hausman- test outcome sounds problematic or barely helpful.
As an aside, I would skim through the literature in your research field and see whether somebody else followed that road when presented with the same research topic.