Hausman test following xtivreg

[Ian Bowen]

I am wondering what the best approach is to conduct the Hausman test for RE vs FE following xtivreg. I have checked the help for xtoverid, but I was not able to find an answer. I also tried the hausman command after storing the estimates from xtivreg models with random effects and fixed effects. However, the results are rather bizarre; I am getting Prob>chi2 = 1.

[Jeff Wooldridge]

A simple test is covered in my 2010 MIT Press book in the balanced case and here

Joshi and Wooldridge

in the general unbalanced case. Simply compute the time averages of all exogenous variables and include them as extra regressors. Then estimate by REIV and test the time averages for joint significance -- using vce(cluster id) for a robust test.

[Ian Bowen]

Excellent. Thank you very much.

However, what about for simple IV models without exogenous regressors, either with or without year fixed effects? In that case, there would be no variables for which to compute the time averages.

[Jeff Wooldridge]

Oh, then you just use the usual control function test based on pooled OLS. Generally, with endogenous w, exogenous covariates xj, and IVs zh:

reg w x1 x2 ... xk z1 ... zm i.year
predict vhat, resid
reg y w vhat x1 ... xk i.year, vce(cluster id)

[Ian Bowen]

Great. And is the null hypothesis beta_vhat = 0, which, if rejected, implies that fixed effects are preferred?

[Jeff Wooldridge]

Hmmmm. I thought you didn't want fixed effects. What you're testing with the basic CF test is whether you need to use instrumental variables or not. The test of pooled OLS versus FE is different and doesn't involve IVs. I think you need to show your model and be very clear about what you want to test. What are the two estimators you wish to compare?