Making the most of sparse clinical data by using a predictive-model-based analysis, illustrated with a stavudine pharmacokinetic study.

A small-scale clinical investigation was done to quantify the penetration of stavudine (D4T) into cerebrospinal fluid (CSF). A model-based analysis estimates the steady-state ratio of AUCs of CSF and plasma concentrations (R(AUC)) to be 0.270, and the mean residence time of drug in the CSF to be 7.04 h. The analysis illustrates the advantages of a causal (scientific, predictive) model-based approach to analysis over a noncausal (empirical, descriptive) approach when the data, as here, demonstrate certain problematic features commonly encountered in clinical data, namely (i) few subjects, (ii) sparse sampling, (iii) repeated measures, (iv) imbalance, and (v) individual design variation. These features generally require special attention in data analysis. The causal-model-based analysis deals with features (i) and (ii), both of which reduce efficiency, by combining data from different studies and adding subject-matter prior information. It deals with features (iii)--(v), all of which prevent 'averaging' individual data points directly, first, by adjusting in the model for interindividual data differences due to design differences, secondly, by explicitly differentiating between interpatient, interoccasion, and measurement error variation, and lastly, by defining a scientifically meaningful estimand (R(AUC)) that is independent of design.