Following on a previous thread regarding the Hausman test (here) in which the reference unfortunately disappeared @briatte , I am faced with some additional Hausman related questions. While the previous thread alluded to the fact that the taking of the absolute value might be a problem in the Hausman execution, I am wondering whether the problems I am having are caused by this indeed.
Depending on the random.method
specification I use, the results can be dramatically different, which makes me wonder which ones to trust (maybe none).
In the following 'vit' represents a formula with on the right side only variables that vary both over time and over individuals, while full represents the same variables + Year fixed effect dummies + some variables that differ for every year but are the same for every firm (firm fixed effects).
Here are some results
phtest(vit,data = p.data, effect="twoways", random.method="amemiya")
Hausman Test
data: vit
chisq = 13.0489, df = 30, p-value = 0.9969
alternative hypothesis: one model is inconsistent
Versus
phtest(vit,data = p.data, effect="twoways", random.method="walhus")
Hausman Test
data: vit
chisq = 76.3063, df = 30, p-value = 6.617e-06
alternative hypothesis: one model is inconsistent
While it is of course possible to get different results with different tests, the radical difference here makes me wonder...
A similar situation occurs with the LM tests:
plmtest(full, data=p.data, effect="twoways",type="kw")
Lagrange Multiplier Test - two-ways effects (King and Wu)
data: full
normal = -2.3819, df = 2, p-value = 0.01722
alternative hypothesis: significant effects
versus
plmtest(full, data=p.data, effect="twoways",type="ghm")
Lagrange Multiplier Test - two-ways effects (Gourieroux, Holly and Monfort)
data: full
chisq = 0.0025, df = 2, p-value = 0.9987
alternative hypothesis: significant effects
Can anyone make a suggestion on what to do?
Thanks,
Simon