Hausman test

[Ting1]

Hello,

I use 2sls to estimate coefficients for 2 equations. After running 2sls with random effect, I did hausman test to see whether random effect is appropriate for each equation. However, based on the result, it shows that it's proper to use fixed effect in equation 1, while random effect in equation 2. Is it normal to have different effect in one model?

THanks for reply!
Also, I attached 2 pictures showing the result of Hausman test. Please see the words in red box. Could you explain what do they mean? Do they mean that I cannot judge the effect based on p-value<.05?

[EViews Glenn]

The second result is telling you that the estimator of the random effects gives you results which indicate that there are no random effects (the variance is zero). The first result indicates that the Hausman test is obtaining an invalid variance, probably because of small sample issues. It's hard to say, but my first guess is that your samples are so small that the estimators don't have enough data to determine much of anything.

[Ting1]

The second result is telling you that the estimator of the random effects gives you results which indicate that there are no random effects (the variance is zero). The first result indicates that the Hausman test is obtaining an invalid variance, probably because of small sample issues. It's hard to say, but my first guess is that your samples are so small that the estimators don't have enough data to determine much of anything.

Hello, Glenn,

Thank you very much for your reply.
I don't think my sample is small. I am using pool object and it's an unbalanced panel data with 1015 companies in 5 years, a total 4071 observations.
I have another question about the difference of 2sls in pool object and 2sls in system object. I run both of them and also 3sls in system object. Not mention the 3sls result, I got totally different 2sls results. Why?

I guess it may be because of the data format? In pool object, year is not a dummy variable. It follows the identifiers as part of the name, like company1-2006. However, in system object, I use year dummy variables since I don't know how to create a similar data set as the one in pool object. I open my Excel file in Eviews with year dummy variables and create a system object to run 3sls and 2sls.

If this is the reason, could you tell me how to create a similar data set as in pool object? If not, what is the reason for the difference?

Thank you for your input.
Ting

[EViews Glenn]

It is small in the T dimension which is all you have to identify the fixed effects. The asymptotics of the Hausman test require T to go to infinity.

As to the differences, I'm afraid that I don't understand what you are comparing...

[tomka]

It follows the identifiers as part of the name, like company1-2006.

I don't know if I understand you correctly, but this identifier might actually cause your problem. If you're using the identifier including the year as the cross section ID for a panel structure, E-Views will see each year's reading of a company as an individual company and see the whole sample as a single period. That could explain why the sample would show no variance and why you cannot calculate fixed effects.

To me it appears that you have to define an identifier which only includes the company and use this one as cross section ID. Then define an indicator which only shows the year and use this as the date series indicator.