I sometimes have to vectorise the Huber weights from a robust regression and use them in a lm. Recently I've had to do something similar for a logistic model but I'm slightly worried because I don't get very similar results

ROB <- glmrob(Y ~Volume+Rate, family=binomial("logit"), data=vaso)
glm(Y ~Volume+Rate,data=vaso,family=binomial("logit"))
glm(Y ~Volume+Rate,data=vaso,weights=ROB$w.r,family=binomial("logit"))

The coefficients from the weighted glm are more similar to the robust regression than the unweighted glm, but is there a way to make them the same? I can get the same results with a robust (rlm) and weighted lm but this doesn't seem to be the case with glm. I haven't looked at the glm robust regression in detail so what I'm asking may be impossible...

Thanks for your help


1 Answer 1


That’s no surprise. Because the binomial distribution is skewed (except pi equals 0.5), a consistency correction is need in each iteration step when a robust psi function is applied. Hence, the robust estimated is not equivalent to a weighted maximum likelihood estimator in glm. Although for the Gaussian case, you obtain the same estimated values as the robust method using a weighted least-square method (using the “correct” weights), the results will differ for confidence intervals and hypothesis tests. The robust result will consider that the robust method may downweight some “good” observations.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.