How can the clustered robust standard errors be smaller than the model based ones?

Question

I ran a GEE model and I used it to check the difference between the empirical standard errors and the model-based one. For almost the variables, the empirical standard error was greater than the OLS standard errors. Except for ONE variable that has its model-based standard error bigger than the empirical standard error.

This contradicts what I've been reading this whole time with " It is anticipated that the Standard Errors of the robust model are more inflated than the model-based standard errors". Now I'm just confused.

My dependent variable is binary.

Answer 1

• If I understand you correctly, you are essentially asking why robust standard errors under clustering design are greater than those computed by the OLS? If that is indeed what you are asking, then typically OLS standard errors are calculated based on the assumption that Simple Random Sampling (SRS) was used. In practice, it is very common that clustered standard errors are greater than those produced by OLS under SRS. In contrast, if stratification was used (rather than clustering), standard errors tend to be smaller compared to the OLS.

• This Chapter says the following: "The greater the correlation among units within a group (that is, the bigger Intraclass correlation is) the greater the impact on the standard error." Please see page 12/20 of this document for more details.