(W-011) Data Gaps, Model Mishaps: Quantifying the Impact of Missing Pharmacometrics Data on Pharmacodynamic Projections

Hillary Husband, Ph.D. – Research Scientist, M&S, Metrum Research Group; Megan Cala Pane, Ph.D. – Research Scientist, M&S, Metrum Research Group; Curtis Johnston, Ph.D. – Group Leader, M&S, Metrum Research Group

Objectives: To determine the bias and precision of pharmacodynamic (PD) estimates and predictions when different levels of pharmacokinetic (PK) data are missing. As the percentage of PK data missing completely at random (MCAR) was increased, the impact on the bias and precision of PD estimates and predictions (measured as percentage relative mean squared error [%RRMSE] calculated across simulation replicates) were assessed.

Methods:
1. PK data were generated with a two-compartment model with first-order absorption, and covariates (sample from an NHANES database) of weight (WT), estimated glomerular filtration rate (eGFR), age, and albumin. Variability was introduced by applying a 33%, coefficient of variation (CV) around CL/F, and a proportional residual variance of 20% CV. 500 replicates of 800 subjects were simulated (200 per dose group of 0, 50, 150, and 200 mg).
2. Six PK datasets were generated, each with a different percentage between 0 and 25% of missing individual subjects’ PK data. Missingness was determined randomly (MCAR). All patients assigned to have missing PK data were subsequently assigned typical exposure metrics conditional upon their covariate values.
3. PD time-courses were simulated, using an indirect response model with a proportional stimulatory effect on kin. EC50 was calculated as approximately the median AUC for the simulated 150 mg dose group. Intersubject variability was included on kout and baseline (CV% of 31% and 45%, respectively). A proportional error model contributed 12% CV residual variability to the simulations.
4. PD parameters were estimated and a landmark PD endpoint (12-week change from baseline [CFB]) was determined from each of the 6 datasets.
5. Relative bias and %RRMSE were calculated for each of the six datasets across dose groups and simulated replicates.

Results: For fixed parameter estimates, the magnitude of bias was greatest for kout but there was little difference between groups (median of 5.68E-3 and 0.60E-3 for 0% and 25% missing PK, respectively). There also was very little difference between levels of missingness in %RRMSE (median of 7.03E-5 and 6.81E-5 for 0% and 25% missing PK, respectively). For the PD endpoint of 12-week CFB, although there was a slight trend towards an increase in mean bias with level of missingness, the magnitude of bias was very small (0.0841 and 0.130 for 0% and 25% missing PK, respectively).

Conclusions: When predicting PD outcomes, imputing population-level PK estimates for subjects missing PK is an acceptable practice under conditions with reasonably large dosing groups, as even a relatively large proportion of subjects’ missing PK does not make a marked difference in estimated PD parameters or simulated responses. Additional characterization of the operating characteristics of the impact of imputing typical exposures for PD analyses is warranted.

Citations: None