1

If a cluster trial had homoskedastic data, but the regression model used by the authors used 'robust standard errors' intended for use with heteroskedastic data (or vice-versa), would the implications be clinically significant with regards to their conclusions?

user356816
  • 543
  • If the method is 'robust' then it is presumably little affected by mis-specification of the skedasticity. – Michael Lew Nov 28 '23 at 23:24
  • 1
    With a clustered trial, you would typically use cluster-robust (aka clustered) standard errors that allow for different variances in the errors across observations and correlation of the errors within clusters but not across clusters. Robust errors generally allow only for heteroskedasticity but not autocorrelation. Can you clarify which ones the authors are using here? – dimitriy Nov 28 '23 at 23:58

1 Answers1

1

I think it would depend on exactly which estimator they used. If, as you say, there was no problem with heteroscedasticity, then perhaps there was a problem with autocorrelation. The Huber-White estimator is typically used to correct only for heteroscedasticity, while the Newey-West estimator is a heteroscedasticity and autocorrelation consistent (HAC) estimator. As to

would the implications be clinically significant with regards to their conclusions?

I don't think you can conclude anything like that. Robust standard errors are very often larger than those obtained with ordinary least squares, but there can be cases where they are biased downwards, as discussed in the following thread:

Can robust standard errors be less than those from normal OLS?

Robert Long
  • 60,630