I asked a related question before here on the difference between GEE method with exchangeable varcov structure v. Robust standard errors known as Huber White method in group randomized trials. As Macro pointed out Freedman in his 2006 paper The American Statistician called On The So-Called “Huber Sandwich Estimator” and “Robust Standard Errors explains that with a little modification of the huber white method, we can get valid inference by both solving the issue with Heteroscedasticity and Correlation among clusters.

I have a study with 34 schools randomly assigned to treament/control. The ICC in the study is 0.04 (although is small but is statistically significant, more info here) with the design effect of 17.7. The ICC indicates some degree of similarities among students within schools (i.e. clusters). Schools have different number of students varying from 150 to 800.

My question is if you wanted to analyze the data, which method you would choose to get valid inference? Huber-White or GEE with exchangeable varcov matrix? I'm in favor of GEE because of two reasons:

1) we have unbalanced clusters 2) Our estimates with GEE are more efficient.

If you choose GEE, could you explain why and what GEE can do that HuberWhite cannot do and the other way around?

I appreciate your help.

  • $\begingroup$ Why is an exchangeable covariance matrix structure appropriate for your data? $\endgroup$ – Michael R. Chernick Aug 10 '12 at 14:59
  • $\begingroup$ Hi @MichaelChernick, because I have indication that there is correlation between students within schools (=clusters). Also, it seems reasonable to assume that any two students in a same school have the same correlation (no specific pattern). We have unbalanced cluster sizes and this is what taken care of with GEE. These are the reasons that I think are appropriate. My question is actually finding the best answer to your question since my reasons above are not theoretical reasons. $\endgroup$ – Sam Aug 10 '12 at 15:07
  • $\begingroup$ Thanks. I don't know the answer to your question. $\endgroup$ – Michael R. Chernick Aug 10 '12 at 15:11
  • 1
    $\begingroup$ @MichaelChernick, I think of huber-white as a robust method with minimum possible assumptions. I think by adding more relevant information to the model we can gain more efficiency and I think GEE with exchangeable varcov can do that. But on the other hand, H-W method is robust in terms of model misspecification whereas GEE is not and that might blow up everything. I think my challenge is to find enough evidence to choose between efficiency and robustness. In my study we have many latent variables and there is a high chance of model misspecification and this might be a reason against GEE. $\endgroup$ – Sam Aug 10 '12 at 15:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.