Dear all,
I'm fitting a beta regression for a proportion and although I like a good
model for all the data, I'm interested the most in a good power prediction
for the interval (0, 0.05). But the data in that interval is only 3% of the
total and the model is not good predicting for that interval. So, I did an
undersampling, I mean I did two same size groups one over 0.05 and one under
0.05 (like undersampling for a logistic regression and the sample has a good
size) and the model works very well. But, now I realize that I only have
papers that talk about undersampling and oversampling for logistic
regression or binary regression but not for continuous response. Do you know
any work in this problem? Undersampling for continuous response?. Now I
have a good model but I have no theory support!.
I'll really appreciate your help.
Angelica Neisa
Research Statistician
http://www.ehealthinformation.ca/
|