Hi Anderson,
Thanks for your reply. I think my case belongs to b) with df1=1 and df2~=600. Inf values were generated on few observations among 60k observations were fine, so I liked to correct the code instead of my inputs, to just make them skip the errors and generate results.
I agree with you in general, but I want to check if the numerical accuracy was cut in this part of palm_gcdf function:
gcdf(ig) = 1 - gcdf(ig);
This cuts values to 1 when gcdf() is lower than around 1e-16, and the cut value feeded into erfcinv in palm_gtoz.m generates Inf. Maybe it needs to use erfinv(x) instad of erfcinv(1-x) in this case?
Thanks,
Dongjin
|