In one of R packages for advanced survival analysis, the frailtypack, the output contains standard errors calculated in two ways, named: H (using the inverse of Hessian) and HIH (H-1 * I * H-1; where I is the Fisher Information). The values are quite close. I am wondering what the two are telling me in general (I suppose this is an universal method, not bound solely to this package and method) and which one should be reported by default (and when)?

  • $\begingroup$ The second appears to be the sandwich estimator, which is a heteroskedasticity-consistent estimator of the standard errors. See en.wikipedia.org/wiki/… for a description in the context of linear regression. It's intended to correct for the effects of heteroskedasticity on the estimated standard errors. $\endgroup$
    – jbowman
    Feb 25, 2020 at 21:49
  • $\begingroup$ It does a lot of sense. I saw this previously in the context of the GEE model, where it reported both "naive" and "robust" SEs. The structure (like a sandwich wrapping the Fisher information) resembles that. I wasn't sure, however, if this is just a coincidence... Thank you for your answer! $\endgroup$
    – Paolinetta
    Feb 25, 2020 at 21:56


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.