From: Meredith Goldwasser
Subject: [NMusers] post hoc pk parameter associations with clinical covariates/response
Date: Fri, 19 Aug 2005 15:02:03 -0400

Dear NONMEM users,
I am a new NONMEM user and also new to the area of non-linear mixed
effects modeling of pk data. In the clinical pk literature, I've seen
predicted (post hoc) estimates of pk parameters for each subject
generated from a population pk model and then used in standard
association tests of covariates or models of response, e.g. using
t-tests, Wilcoxon or Kruskal-Wallis tests to compare clearance between
groups of subjects with or without a particular polymorphism, or using
logistic regression to model toxicity on drug clearance.  Is this
two-step approach statistically appropriate? For instance, a standard
assumption of these tests and models is independence between subjects,
but it would appear that these predicted pk parameters are not
independent.  I've read some of the discussion of simultaneous versus
sequential estimation in pk/pd analysis of Karlsson, Zhang, and
others, but I'm not sure if this applies to the situation of
associating pk estimates with a clinical or pharmacogenetic endpoint.
Is a single endpoint like toxicity (dichotomous variable: yes or no)
or survival outcome (time to event variable) considered to be PD data,
as it's usually modeled in NONMEM with an Emax model?

Finally, if the joint likelihood of the pk and outcome data are
modeled simultaneously in the population pk model, does NONMEM provide
frequentist-like inference measures, like a p-value of association
between pk and the outcome variable, based on the posterior

Any guidance or reference to relevant articles would be most appreciated.