Panel Data Analysis - Hausman test fails

[Joca Almeida]

Hello,

I'm writing my master thesis and I want to test the impact of crowdfunding in renewable energy production. FOr that, I built a database with 27 countries, with the time span of 6 years, totalizing 162 observations.

When I try to to perform the Hausman test to check which specification is preferred, I get this message: "model fitted on the data fails to meet the asymptotic assumptions of the Hausman test."

Here are the results I get from Stata:

I'm really lost here. I don't know what kind of specification I should use.

Hope you can help me. Thanks in advance!

[Maxence Morlet]

The Hausman test makes very unrealistic assumptions, that will almost never be met (e.g. homoscedasticity). Try the Mundlak (1978) test: https://blog.stata.com/2015/10/29/fi...dlak-approach/

Other than that, I know at uni we all learn to do things by the books and test whether the random effects assumption is valid.

In practice though, without even conducting these tests, economists tend to just execute two way fixed effects (TWFE) models by default, and cluster standard errors by the x variable in xtset.

This is for causality purposes; the assumption made by the random effects model is simply implausible, therefore it will be extremely difficult to convince a reader that you've obtained causal estimates using an RE model...

For a mindblowingly amazing review of TWFE estimators, and pooled OLS, see https://papers.ssrn.com/sol3/papers....act_id=3906345.

Hope this helps!

[Joca Almeida]

Exactly, I was performing the test just as uni taught me, but I'll check the Mundlak test.

Regarding my insignificant variables, do you have any suggestion on how to fix it? Do you think it is because my sample is not big enough?

[Maxence Morlet]

Your sample is very small indeed. Also, you cannot cluster your standard errors (27 groups is not enough, you'd need at least 50) as this will result in bias, so only use robust standard errors.

I would recommend getting a larger sample if it's possible.

I see you're running a log-log model, you could try varying that (e.g. see what happens with a log-linear model for instance).

Otherwise, there's not much you can do when you have insignificant regressors.

I strongly advise that you do not remove the insignificant regressors from your model however, as this will affect how other coefficients are calculated, and may introduce bias.