Debiasing the estimate of treatment effect on the treated with time-varying counfounders

TL;DR

This paper introduces a debiased method for estimating treatment effects on the treated in time-varying settings, improving accuracy over previous models using large health databases and real-world ICU data.

Contribution

It generalizes Gran's additive intensity regression model to account for repeated outcomes and multiple covariates, providing a more accurate ATT estimation method.

Findings

Debiased estimator outperforms Gran's original method in simulations.

Method effectively applied to ICU data to assess vasopressor effects.

Improved treatment effect estimates in complex time-dependent health data.

Abstract

With the increased availability of large health databases comes the opportunity of evaluating treatment effect on new data sources.Through these databases time-dependent outcomes can be analysed as events that can be measured using counting processes. Estimating average treatment effect on the treated (ATT) requires modelling of time-varying covariate and time-dependent treatment and outcome. Gran et al. proposed an easy-to-implement method based on additive intensity regression to estimate ATT. We introduce a debiased estimate of the ATT based on a generalization of the Gran's model for a potentially repeated outcome and in the presence of multiple time-dependent covariates and baseline covariates. Simulation analyses show that our corrected estimator outperforms Gran's uncorrected estimator. Our method is applied to intensive care real-life data from MIMIC-III databases to estimate vasoppressors effect on patients with sepsis.