RE: Mixture model with logistic regression
Use of mixture model may not be suitable here, if the underlying distribution
of eta's for the different subgroups is not normally distributed. Based on the
description, it looks like you have 3 degenerate eta distributions(ETA ~0 for
no AE; ETA ~1 for 25% AE and ETA ~10 for 7% with all AE), which violates the
normality assumption.
Upon quick search, I came across this article, where they have described almost
a similar situation as yours. They have used a two part mixture distribution
to take care of the large proportions of the subjects with no AE. Hope it is
helpful.
http://www.ncbi.nlm.nih.gov/pubmed/14977163
Kowalski, Kenneth G., Lynn McFadyen, Matthew M. Hutmacher, Bill
Frame, and Raymond Miller. "A Two-Part Mixture Model for Longitudinal Adverse
Event Severity Data." Journal of Pharmacokinetics and Pharmacodynamics 30, no.
5 (October 2003): 315-36.
Thanks.
Mathangi Gopalakrishnan, MS, PhD
Research Assistant Professor
Center for Translational Medicine (CTM)
School of Pharmacy, UMB
Ph: 410-706-7842
http://www.ctm.umaryland.edu/
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of Mark Sale
Sent: Friday, February 19, 2016 5:30 PM
To: [email protected]
Subject: [NMusers] Mixture model with logistic regression
Has anyone every tried to use a mixture model with logistic regression? I have
data on a AE in several hundred patients, measured multiple times (10-20 times
per patient). Examining the data it is clear that, independent of drug
concentration, there is very wide distribution of this AE, 68% of the patients
never have the AE, 25% have it about 20% of the time and the rest have it
pretty much continuously, regardless of drug concentration. (in ordinary
logistic regression, just glm in R, there is also a nice concentration effect
on the AE in addition). Running the usual logistic model, not surprisingly, I
get a really big ETA on the intercept, with 68% of the people having ETA small
negative, 25% ETA ~ 1 and 7% ETA ~ 10. No covariates seem particularly
predictive of the post hoc ETA. I thought I could use a mixture model, with 3
modes, but it refused to do that, giving me essentially 0% in the 2nd and 3rd
distribution, still with the really large OMEGA for the intercept. Even when I
FIX the OMEGA to a reasonable number, I still get essentially no one in the 2nd
and 3rd distribution. I tried fixing the fraction in the 2nd and 3rd
distribution (and OMEGA), and it still gave me a very small difference in the
intercept for the 2nd and 3rd populations.
Is there an issue with using mixture models with logistic regression? I'm just
using FOCE, Laplacian, without interaction, and LIKE.
Any ideas?
Mark
Mark Sale M.D.
Vice President, Modeling and Simulation
Nuventra, Inc. (tm)
2525 Meridian Parkway, Suite 280
Research Triangle Park, NC 27713
Office (919)-973-0383
[email protected]<[email protected]>
http://www.nuventra.com