Double Robust Efficient Estimators of Longitudinal Treatment Effects: Comparative Performance in Simulations and a...

Area of Science:

• Causal inference methods in longitudinal studies
• Statistical modeling for time-series data
• Biostatistics and epidemiology

Background:

• Estimating intervention effects in longitudinal studies is crucial for understanding treatment impacts.
• Existing sophisticated estimators lack direct comparative research, hindering method selection.
• Longitudinal data, particularly from HIV cohorts, presents unique challenges for causal effect estimation.

Purpose of the Study:

• To directly compare the performance of six distinct causal inference estimators in a longitudinal setting.
• To evaluate estimators using both simulated data and real-world human immunodeficiency virus (HIV) cohort data.
• To assess the impact of positivity violations on estimator accuracy and bias.

Main Methods:

• Evaluated six estimators: iterated conditional expectation, inverse propensity weighting (IPW), augmented IPW, and three double robust (DR) variants, including a targeted minimum loss-based estimator (TMLE).
• Nuisance parameters were estimated using pooled data and data-adaptive machine learning algorithms.
• Simulations featured increasing positivity violations across six time points; applied analysis used HIV cohort data.

Main Results:

• Double robust estimators showed minimal bias when at least one nuisance parameter model was correct; performance degraded under full misspecification.
• Weighted estimators generally outperformed covariate-based estimators.
• All estimators' bias and mean squared error worsened with increased positivity violations; covariate-based DR estimators were particularly susceptible.

Conclusions:

• The targeted minimum loss-based estimator (TMLE) demonstrated efficiency, parameter space adherence, and robust performance.
• The pooled and weighted TMLE is recommended for longitudinal causal effect estimation, especially when positivity violations are a concern.
• Inverse propensity weighting (IPW) estimator showed marked deviation in applied analyses as positivity violations increased.