Addressing Causality and Homogeneity Assumptions in Exposure-Response Analyses.

Area of Science:

• Pharmacometrics
• Biostatistics
• Drug Development

Background:

• Exposure-response (PKPD) analyses are crucial for drug development decisions.
• Current PKPD models often lack assessment of causality and homogeneity, potentially leading to biased results.
• Instrumental variable (IV) analysis offers a method to establish causal effects.

Purpose of the Study:

• To adapt and evaluate IV models (predictor substitution and control function) for repeated-measures analyses.
• To compare these IV models against standard PKPD models and a novel partitioned effects (PE) model.
• To assess model performance under various confounding scenarios, including shared latent variables, unmeasured active metabolites, and reversed causality.

Main Methods:

• Adaptation of predictor substitution (PS) and control function (CF) IV models for repeated-measures data.
• Comparison with PKPD models (with and without PK-PD random effect correlations) and the partitioned effects (PE) model.
• Simulation of six scenarios: no confounding, three types of confounding, unmeasured active metabolite, and reversed causality.

Main Results:

• Standard PKPD models yielded biased parameter estimates and inappropriate dosing recommendations in most simulated scenarios.
• PS and CF models provided adequate estimates in a majority of scenarios.
• The novel PE model consistently provided adequate estimates across all simulated scenarios, including complex confounding situations.

Conclusions:

• The PE model offers a robust approach for causal exposure-response modeling, outperforming traditional PKPD methods under confounding.
• The PE model supports adequate individualization of dosing based on concentration or response.
• Adapted IV methods, particularly the PE model, are valuable for reliable drug development decision-making with both single and repeated measures.