3

I am a beginner level stat learner (with graduate training in Applied Math) I have just read a sage book which states the following:

"Dummy variables help address the issue of heteroscedasticity in residual variances given a well specified linear model"

This threw me off a bit for I was always under the impression that the choice of Dummy Variables is a structurally driven decision given a fair knowledge of presence of heterogeneous groups in the DGP where as WLS is almost a ex post adjustment proviso the error terms exhibit changing variance.

Can some one kindly point me to some literature about the same or comment regarding the relative merits of one as opposed to the other?

Ferdi
  • 4,882
  • 7
  • 42
  • 62
schuler
  • 31
  • 2

1 Answers1

1

Dummy vars and WLS (with exchangeable correlation structures) are both methods that can handle additive effects of cluster based samples.

The dummy variable approach is the most appropriate method. The only reason to discourage its use is when the number of clusters is high relative to the sample size.

Weighted least squares specifies a dependence structure and then iteratively estimates the model effects and covariance structure until an optimal result is found. Unlike the above example, this one gives consistent estimates even when there is a relatively large number of clusters. The approach is very similar to fitting a mixed effects model with random intercepts.

You can find great a reference in Longitudinal Data Analysis by Diggle Heagerty Liang and Zeger.

AdamO
  • 52,330
  • 5
  • 104
  • 209