Non-linear mixed effects modeling as a tool for causal inference

Intercurrent events, which are events occurring after the initiation of treatment, can confound dose-exposure-response relationships even in randomized controlled trials. When the confounding variables are unmeasured, addressing scientific questions (estimands) that are hypothetical in nature — such as about the dose-response relationship had the intercurrent events not occurred — may become a challenging inferential task. Here, we show that non-linear mixed effects (NLME) modeling can be employed for valid causal inference in the presence of unobserved confounders that affect both the treatment received and the outcome of interest. Such situations are e.g. encountered in studies in which intercurrent events related to either the efficacy or tolerability of a drug may lead to deviations from the assigned treatment schedule. The key to utilize NLME modeling as a tool for causal inference in such scenarios is to condition on the individual pharmacokinetics (PK) and/or pharmacodynamics (PD) parameters of the NLME model. This blocks the spurious associations induced by the unmeasured confounders. Specifically, we show that in certain scenarios, when the PK and PD of a drug are well characterized, NLME PKPD models provide an unbiased estimator of hypothetical estimands of the dose-exposure-response relationship. Having illustrated that NLME modeling can be a valid tool for causal inference, we hope that this work encourages utilizing NLME modeling and its well-established strengths, including the integration of longitudinal data and prior knowledge, to address challenging causal questions in clinical development, and encourages pharmacometricians to engage more deeply with questions of causality.