Studying Continuous, Time-varying, and/or Complex Exposures Using Longitudinal Modified Treatment Policies.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Causal inference methods are crucial for understanding treatment effects in observational studies.
• Existing methods like static and dynamic interventions have limitations with complex time-varying data.
• Novel parameters and generalizations are needed for robust causal effect estimation.

Purpose of the Study:

• To introduce and explain the methodology of longitudinal modified treatment policies (LMTPs).
• To demonstrate the mathematical formalization, identification, and estimation capabilities of LMTPs.
• To illustrate practical applications and provide estimation strategies for LMTPs.

Main Methods:

• Utilizes longitudinal modified treatment policies for causal inference.
• Applies to diverse exposures (binary, multivariate, continuous) and outcomes (survival, binary, continuous).
• Accommodates time-varying treatments, confounders, competing risks, and loss to follow-up.

Main Results:

• LMTPs generalize parameters such as the average treatment effect.
• They allow for the definition of alternative estimands with potentially more satisfiable positivity assumptions than static interventions.
• The tutorial provides examples, including estimating the effect of delayed intubation on COVID-19 mortality.

Conclusions:

• LMTPs offer a flexible and powerful framework for causal inference with complex longitudinal data.
• The methodology extends existing intervention strategies and allows for estimation of novel causal parameters.
• The open-source R package lmtp facilitates the practical application of LMTPs.