Compoenents Variance in Random Effects models

[Dave_the_Forecaster]

Hi all,

This should hopefully be fairly simple to answer.

When runnning a panel regression with random effects, Eviews offers a few different methods to calculate the variance of the error components. Swamy-Arora and Wansbeek-Kapteyn (which I've taken to be the Amemiya method) involve first performing a fixed effects regression to obtain an unbiased estimate of the idiosyncratic (within) error variance.

My question is: how does Eviews handle time-invariants during this intermediate FE regression? Are they simply dropped from the model specification - if so, are the degrees of freedom adjusted for this when calculating the idiosyncratic error variance from the SSR?


Cheers,
Dave

[Dave_the_Forecaster]

Any Eviews master willing to help out with this question?

[EViews Glenn]

Sorry. Didn't see the original post.

You can think of the between regression as a regression where we drop redundant regressors, or as simply a projection onto the space orthogonal to the regressors (I prefer the latter interpretation). We use the rank of the between regression as the number of regressors for determining the residual variance. Thus, there is no account for the time-invariant regressors in the df computation. One way of thinking about this is that if someone told you to that you were going to use a particular set of between regressors (the ones not including the dropped regressors) and between dependent variable to estimate the residual variance, how would you do so?

[Dave_the_Forecaster]

Hmm, makes sense. Thanks very much Glenn.