Using latent outcome trajectory classes in causal inference.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Causal Inference

Background:

• Longitudinal studies reveal critical subgroup differences in treatment response.
• Growth mixture analysis identifies heterogeneous outcome trajectories.
• Integrating latent classes with causal inference remains challenging.

Purpose of the Study:

• To propose a novel 2-step approach for estimating causal treatment effects in the presence of latent trajectory classes.
• To separate the formulation of trajectory strata from the identification of causal effects.
• To provide a framework for understanding differential treatment effects across subpopulations.

Main Methods:

• A 2-step approach is proposed: 1. Stratify individuals into trajectory classes using growth mixture analysis within a reference condition. 2. Estimate causal treatment effects for each trajectory stratum, treating stratum membership as partly known.
• This method aligns with principal stratification principles, enabling sensitivity analyses.
• The approach is demonstrated using longitudinal data on childhood attention deficit.

Main Results:

• The proposed method allows for the estimation of treatment effects tailored to specific outcome trajectories.
• Results indicate how subpopulations with distinct prognoses under one condition alter their prognosis under another treatment.
• The approach facilitates systematic sensitivity analyses for causal effect estimation.

Conclusions:

• The 2-step method effectively integrates growth mixture modeling with causal inference for longitudinal data.
• This approach enhances the understanding of treatment effects across distinct subpopulations defined by outcome trajectories.
• The methodology offers a robust framework for analyzing complex longitudinal data and informing personalized interventions.