Vijay Ivaturi, Mathangi Gopalakrishnan, Serge Guzy
Center for Translational Medicine, University of Maryland Baltimore School of Pharmacy
Objectives:The aim was to investigate and compare the performance with respect to bias and precision of the QRPEM and LAPLACE methods available in Phoenix NLME for discrete data and repeated time to event data.
Methods: Discrete data models with “mu-modeled” parameters, representing different types, binary, count, ordered categorical (OC) and a repeated time-to-event (RTTE) model were used to simulate 200 datasets that were subsequently re-estimated using Phoenix NLME (v1.4 beta). Each model was simulated with different starting values for the random effects representing low, medium and high between subject variability on the fixed effect parameter estimates, resulting in ten different scenarios. All datasets in each scenarios were analyzed starting with initial estimates set to the true simulation values of that scenario. Root mean squared error (RMSE), relative bias (RB) and relative estimation error (REE) in estimates were evaluated for each parameter across scenarios.
Results:QRPEM performed equal or better as compared to LAPLACE across all scenarios investigated. The absolute RB was less than 5 % for fixed effects and in the range of 10-20 % for random effects parameters. The RMSE was less than 5 % across all different models and scenarios. On an average the runtimes for LAPLACE were quicker than QRPEM.
Conclusions:We present preliminary results evaluating the new QRPEM algorithm in Phoenix NLME for discrete data models. As all parameters were “mu-modeled”, this evaluation tested QRPEM specifically as opposed to using Sampling Importance Resampling algorithm (SIR). QRPEM performs equal or better than LAPLACE in most scenarios, with the bias on random effects higher than fixed effects. Further evaluation of QRPEM and comparison to similar algorithms in other software is currently in progress which will allow a complete evaluation with respect to discrete data models.
Reference: PAGE 23 (2014) Abstr 3226 [www.page-meeting.org/?abstract=3226]
Poster: Methodology - Estimation Methods