RE: Simulating different populations

From: "Kowalski, Ken" Ken.Kowalski@pfizer.com Subject: RE: [NMusers] Simulating different populations Date: Wed, 10 Aug 2005 13:12:09 -0400 Liping, Just to clarify. I indicated that simulating different sets of population estimates based on the var-cov matrix from NONMEM and then assuming multivariate normal and possibly inverse-Wishart (for the variance parameters in Omega and/or Sigma) distributions for the parameter estimates is NOT computationally intensive. However, it is laborious to use these population estimates and feed them back in to NONMEM to perform simulations that take into account this parameter uncertainty because it requires custom coding. Based on Marc's comments it sounds like he and others are working on developing utilities that will automate this process in NONMEM which would make it a lot easier to do this as a general practice. If on the other hand we do not wish to make a parametric assumption using the var-cov matrix and multivariate Normal/Inverse-Wishart distribution we could perform parametric or nonparametric bootstrap simulations and re-fit the model in NONMEM for each of say 1000 bootstrap datasets to obtain 1000 estimates of the population parameters from the posterior distribution and use these in subsequent simulations to account for parameter uncertainty. This approach IS computationally intensive because of having to re-fit the model in NONMEM. My feeling is that it is better to routinely do something to take into account parameter uncertainty using the var-cov matrix and parametric assumptions than to doing nothing to take into account parameter uncertainty. Hopefully, individuals working on utilities to automate this parametric approach will be able to share them with us. Ken