Introduction: In pharmacometric modeling, modellers choose the covariate-parameter scope to consider. For example, all covariates on all structural parameters (a full model), all covariates on only one of the parameters, or different sets of covariates on different parameters. Scope reductions aim to simplify models, reduce runtime, and enhance stability, usually based on the assumption that some parameters are unaffected by certain covariates. However, in practice such assumptions may not hold and may lead to biased estimates of the parameters [1].
The current work considers omission bias and inclusion bias. These biases may occur when a covariate is predictive for two parameters but only included on one of them (omission bias) or when a covariate effect is estimated for a parameter it is not predictive of (inclusion bias). These biases may affect any of the parameters in the model.
Objectives: • To investigate the impact of misspecified fixed effects (FFEM) covariate models on the omission and inclusion bias in the estimated parameters. • To investigate the impact of misspecified full random effects (FREM) [2] covariate models on the inclusion bias in the estimated parameters.
Methods: 100 richly sampled subjects were simulated from a one compartment model with first order absorption for eight different data generating mechanisms, ranging from WT on both CL and V with a correlation between CL and V, via models with WT on only one of the parameters, with and without correlation, to a model without any covariate effect and correlation.
The simulated data was estimated using estimation models including the simulation models and two FREM models with covariates on either CL, V and KA, or only CL and V.
The simulation and re-estimations were repeated 100 times to investigate the impact on omission and inclusion bias.
Results: • The estimates of CL, V and KA were sensitive to both biases. For instance, biases in CL estimates ranged from 7-75% under different FFEM scenarios. • The accuracy of covariate coefficients was significantly impacted (up to a seven-fold increase in bias), especially when the parameters were correlated. • The variance parameters in the FFEM were also sensitive to inclusion and omission bias. • FREM behaves robustly regardless of simulation model and does not exhibit any inclusion bias. • Identical parameter estimates were obtained from equivalent FFEM and FREM.
Conclusions Reducing the covariate scope without considering the underlying data generating mechanisms risks introducing omission and inclusion bias.
FREM is insensitive to inclusion bias and including covariates on a parameter where it has no effect result in that the impact of the covariate is estimated to 0.
References [1] Ivaturi et al “Selection Bias in Pre-Specified Covariate Models,” PAGE 20 (2011) Abstr 2228 [2] Yngman et al “An introduction to the full random effects model,” CPT:PSP 11(2) 2022, doi: https://doi.org/10.1002/psp4.12741
Citations: [1] Ivaturi et al “Selection Bias in Pre-Specified Covariate Models,” PAGE 20 (2011) Abstr 2228 [2] Yngman et al “An introduction to the full random effects model,” CPT:PSP 11(2) 2022, doi: https://doi.org/10.1002/psp4.12741
