Test equations in FE/RE tests

[beginner202]

Hi there,

I'm having big problems with my model specification. The explanatory variables' coefficients look correct when I DON'T specify fixed effects, however according to the tests that EViews provides, the fixed effects are significant and thus fixed effects model should be employed. Which would be fine, but the fixed effects model throws my coefficients out completely, for example telling me that investment has a negative effect on econ growth! This is just the background to my question...

When the tests are run on likelihood ratio of redundant cross-section and period fixed effects, it produces 3 test equations. What I want to know, is what the coefficients in these tests represent? The final test equation "Cross-section and period fixed effects test equation" produces coefficients on the explanatory variables that look correct! But without understanding what these coefficients mean, I don't know where to go from here.

Can anyone point me in the direction of where to find out? I have read the EViews 7 Help guide but am still none the wiser.

Thanks,

Katie

[EViews Glenn]

It sounds from your description though your estimated model has both fixed cross-section and period effects.

Redundant testing using this equation performs three tests:

1) drop the cross-section effects and test the unrestricted model with both effects against one with period effects only
2) drop the period effects and test the unrestricted model with both effects against one with cross-section effects only
3) drop both the cross-section and period effects and test the unrestricted model with both effects against one with no effects

The test equations are the restricted specifications with dropped effects (1- period only, 2 - cross-section only, 3 - none).

So if I'm understanding what you wrote correctly, the test equation "Cross-section and period fixed effects test equation" is one where there are no effects.

Note that the following models that are not tested in this framework: one where you test the model with cross-effects against one with no effects, and one where you test the model with period effects against one with no effects. For those you'll have to estimate the one-way effects model and then do the redundant tests...

As to what is happening with your coefficient estimates, how many observations do you have? Have you thought about employing random effects instead of fixed?

[beginner202]

Thanks for your swift and helpful reply EViews Glenn.

I have a panel consisting of 43 developing countries with 242 observations spanning the 5 year periods between 1960 and 2005. The observations are rather restricted as I am using inequality data which is notoriously scarce.

I have tried the random effects model, but the Hausman test indicates that the results are significantly different from the fixed effects model (p=0.0000) and therefore the fixed model is more appropriate.. assuming I have interpreted that statistic correctly!

I have also tried running the one-way fixed effects model for both cross-section and period, and cross section fixed effects are far more significant than period.. But when I run the likelihood test on the two-way fixed effects model, the test results seem to indicate that both are significant.

This is the second sample that I have used to test this data, as initially I thought I may have had some obscurities within the sample so I went and collected more data to test it again, but unfortunately the results are still not as they should be. Even when I run a very basic growth regression without the inequality variable, when the fixed effects model is employed the coefficients are very strange.

Any light you can shed on this issue will be very much appreciated!

[EViews Glenn]

Are the period effects themselves significant? The Period F and Period Chi-square. (Not the joint test of the fixed/period)...

Beyond that I'm afraid you're at a point where it's between your and your data. There may be other specification issues that are leading to your unexpected result.

[beginner202]

Period F: p=0.0032
Period Chi-square: p=0.0021

So period fixed effects are quite significant but not as much so as cross-section. When I use random effects the coefficients look correct, but as I said before apparently the random effects model is not appropriate here.

I'm coming round to the fact that there is a problem in the data as you said, but I just wish I knew what that was! Having collected it twice and experimented with different indicators, all to no avail, I am starting to think I should give up!

[EViews Glenn]

When I say "data", I really mean data and model. Misspecification is certainly a possibility.