search

osorensen / hdme

R-package containing penalized regression methods for High-Dimensional Measurement Error problems (errors-in-variables)
GNU General Public License v3.0
8 stars 3 forks source link

Weighted regression #25

Closed davidaknowles closed 2 years ago

davidaknowles commented 5 years ago

How hard would it be hard to be able to weight data points differently, i.e. different known noise variance for each observation?

Thanks!

davidaknowles commented 5 years ago

I think for corrected_lasso_gaussian at least I can just do Q <- t(W) %*% diag(weights) %*% W - sigmaUU b <- t(W) %*% diag(weights) %*% y right? (where sum(weights)==1).

osorensen commented 5 years ago

Dear @davidaknowles, thanks for this very interesting question! Sorry for my very late reply. Apparently I had turned off notifications from this repository, so did not see this issue.

I will think about it, and get back to you as soon as I can.

osorensen commented 5 years ago

Hello, and sorry again for being so late.

I think what you do makes perfect sense when some data points have a higher reliability than others.

This seems like an interesting and useful feature, and I imagine there must be some references to this in the statistical literature? I will add weight option to a future update of the package - perhaps I will be able to do this already in April.

osorensen commented 2 years ago

This could be an interesting improvement, e.g., for data with correlated residuals, but it requires more methodological development. Closing for now.