When to use robust standard errors and "normal" standard errors?

[Pea]

Hi!

I have been consistently looking at robust standard errors and t-ratios when interpreting my results (I'm conducting a choice experiment and I analyse the results through multinominal and mixed logit models in Apollo). Now I have become in doubt as to whether I should have focused on the "normal" standard errors and t-ratios instead. Shouldn't the (simulated) maximum likelihood estimator automatically correct for heteroscedasticity? I guess the reason I've focused on robust errors and t-ratios is this quote from the manual "All that will happen by using apollo_panelProd is that the calculation of the robust standard errors recognises that the choices come from the same individual". Maybe I have falsely interpreted this quote as recommending robust standard errors. When should robust standard errors and "normal" standard errors be used, respectively?

Best wishes
Pea

[stephanehess]

Pea

apologies for the slow reply, David and I have been away.

I always use robust standard errors. They make far fewer assumptions than classical ones.

Stephane