Jakob Ribbing and E. Niclas Jonsson
Division of Pharmacokinetics and Drug Therapy, Uppsala University
Identification and quantification of covariate relationships is an important part of population pharmacokinetic/pharmacodynamic modelling and is usually based on data from a single study. However, the individuals in a clinical study are merely a sample from, and may not faithfully reflect the underlying patient population. Thus, basing covariate selection on data from a single clinical trial may lead to false covariates being included in the model (type-I error), true covariates being omitted (type-II error) and to estimated coefficients being biased due to data-driven selection of covariates (selection bias). Overall this will have negative consequences on the predictive performance of the model. (1)
In this simulation study we investigate to what extent different approaches to covariate identification lead to lower type-I and II errors, less biased estimates of the covariate coefficients and better overall predictive performance.
The approaches examined include:
1. analysing the most recent dataset independently of any previous dataset(s)
2. merging all available datasets into one
3. estimating the model hypothesized from a previous dataset on the most
recent.
Six covariates were sampled with replacement from an empirical distribution containing 1492 patients (2). For each replicate, three PK datasets were simulated using a one-compartment model with first-order absorption and a multivariate-linear-covariate model on typical value of clearance (TVCL). The first and second dataset contained 200 subjects in total. The third dataset consisted of 1000 subjects and was used only to assess the predictive performance on individual TVCL. The first and second datasets were analysed according to the different approaches and the results were compared.
The initial results show that the data-driven approaches all suffer from selection bias in situations where the power of identifying the correct covariate(s) is low. The approach of estimating on the second dataset the model hypothesized from the first dataset is non-data driven and thus produces estimates without selection bias. In situations where selection bias has substantial impact on the predictive performance this approach is considerably better than not using the first dataset at all. However, in all situations examined the approach of merging the two datasets, although producing somewhat biased estimates, performed the best in terms of predictive performance.
References:
[1] Power, Selection Bias and Predictive Performance of the Population Pharmacokinetic Covariate Model. In Press (Journal of Pharmacokinetics and Pharmacodynamics, Vol. 31, No. 2, Apr. 2004)
[2] Population pharmacokinetics of clomethiazole and its effect on the natural course of sedation in acute stroke patients. Br J Clin Pharmacol, Vol. 56, No. 2, Aug 2003
Reference: PAGE 13 (2004) Abstr 492 [www.page-meeting.org/?abstract=492]
Poster: poster