Influence analysis explores heterogeneity in database before data processing by a parametric population method
N. Frances (1), L. Claret (2), F. Schaedeli Stark (3), R. Bruno (2), A. Iliadis (1)
(1) Dpt of Pharmacokinetics, UMR-MD3, University of Méditerranée, Marseilles, France; (2) Pharsight Corp., Mountain View, USA; (3) Hoffman-La Roche, Basel, Switzerland
Objectives: How to proceed when heterogeneous population data are intended to be analyzed by means of a parametric population method like NONMEM. By means of an influence analysis, examine whether covariates are in the origin of this heterogeneity.
Methods: We previously developed a model describing tumor size dynamics in metastatic breast cancer treated by two drugs used in combination [1]. 222 patients received Docetaxel plus Capecitabine and have been monitored up to 50 weeks. The model includes K-PD components for the two drugs and resistance mechanism. Population analyses were performed using NONMEM v6 and simulation analyses, using Matlab v8a. Population analysis of these data results in high values of shrinkage for post hoc estimates. Therefore, Bayesian estimation procedure turns unreliable; the model can not be used to predict evolution disease or to individualize treatment. We attempted to partition the data in homogeneous groups which can be analyzed by a parametric method. Using the same model and the same database, three approaches controlling heterogeneity were applied: the live-one-out approach, the mixture model within NONMEM [2] and the non-parametric approach implemented within NONMEM [3].
Results: The live-one-out method shares the database in two groups. This result was confirmed by the mixture model. The non-parametric approach is under investigation. Administration and observation protocols did not explain the partition of data in two groups. No other covariates were available to explain this partition.
Conclusion: Influence analysis was applied to explore heterogeneity in the recorded database. It would be of interest to find covariates powerful to share patients in homogeneous groups before data processing. These covariates could be biologic markers (characteristics of tumor) or pharmacokinetic parameters (characteristics of drugs or individuals).
References:
[1]. N. Frances, L. Claret, F. Schaedeli Stark, R. Bruno, A. Iliadis. Modeling of longitudinal tumor size data in clinical oncology studies of drugs in combination. in PAGE. 2008. Marseille.
[2]. B. Frame, R. Miller, R. L. Lalonde, Evaluation of Mixture Modeling with Count Data using NONMEM. Journal of Pharmacokinetics and Pharmacodynamics, 2003. 30: p. 167-183.
[3]. A. Lemenuel-Diot, C.Laveille, N. Frey, R. Jochemsen, A. Mallet, Mixture Modeling for the Detection of Subpopulations in a Pharmacokinetic/Pharmacodynamic analysis. Journal of Pharmacokinetics and Pharmacodynamics, 2006. 34: p. 157-181.
