At the risk of self-promotion, in https://www.ssc.wisc.edu/~b... I strongly recommend jackknife/HC3 standard errors, as HC1/HC2 have potentially unbounded bias and size. The same recommendation is made independently in the forthcoming JoE review paper by MacKinnon, Nielsen, and Webb https://www.sciencedirect.c..., based on extensive simulation evidence. Both of these papers focus on the broader setting of clustered observations, but this is not an essential difference.
↧