Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
Pharmacometrics modeling in combination with G-formula to adjust for time-varying confounder for time-to-event analysis
Authors: Siyan Xu, Harriet Longley, Thomas Dumortier, Yu-Yun Ho,
Presenting Author: Siyan Xu*
Treatment crossover (XO) from control to experimental in oncology complicates estimation of treatment effect of overall survival (OS). A hypothetical estimand is: the probability of survival in control arm under a counterfactual scenario where XO is disallowed. Naïve censoring at XO time yields biased estimates when XO is driven by post randomization variables that also affect survival.
In prostate cancer trials, prostate specific antigen (PSA) influences both XO and OS, making PSA a time varying confounder of the relationship between time to XO and time to death. Assuming PSA is the sole confounder, conditioning on PSA can render time to XO independent of time to death, enabling unbiased estimation of the survival even when censoring at XO time.
We estimate hypothetical estimand using g-formula, with a pharmacometrics modelling framework and parametric time-to-event analysis. Trial data is simulated from an “assumed ground truth” kinetic-pharmacodynamic model where PSA is the only time-varying confounder that impacts both XO and death hazards. G-formula based survival estimates were compared to the known “ground truth” survival curve.
Simulations show that, under correct model assumptions and no unmeasured confounders, the g-formula approach accurately estimates the hypothetical estimand. This finding highlights the value of integrating pharmacometric modeling with causal inference to improve the estimation of treatment effect when time-varying confounders are present.