Is it possible or reasonable to use the coefficient of variation in a GAM / GLM as a weight to incorporate uncertainty into a response variable?

I've got density estimates that have a CV value and need to model while taking into account the observation error.


If I understand correctly then yes you can. Though whether you can in practice and how you do this will depend on the software implementation you wish to use. For example:

In R's glm() and mgcv::gam() there is a weights argument. ?glm says

Non-NULL weights can be used to indicate that different observations have different dispersions (with the values in weights being inversely proportional to the dispersions); or equivalently, when the elements of weights are positive integers w_i, that each response y_i is the mean of w_i unit-weight observations.

From that description, I think you would need to use weights = 1/CV to get the desired effect.

  • $\begingroup$ Thanks - yes, this is what I ended up interpreting as well. However, I had to normalize the weights (weights = weights/mean(weights) ) to get it to work. Not 100% sure why, but otherwise it just didn't work properly. $\endgroup$ Apr 21 '17 at 9:11
  • $\begingroup$ I've seen that happen too in one case, but not with others data sets. I think what might be happening is the likelihood is being swamped by the weights and that's leading the model to be erroneously estimated with very low error — in the one example where it didn't work well confidence bands on smooth went to almost zero. $\endgroup$ Apr 21 '17 at 15:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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