A plug-in estimator of the non-iterative conditional expectation g-formula using long short-term memory models

Key points

• In practical applications, parametric models are commonly used to estimate the joint distribution of confounders, treatments, and outcomes over time.1
• As an alternative to parametric models, long short-term memory (LSTM) models, designed to capture both short-term patterns and long-range dependencies in sequential data, may better accommodate complex temporal dependencies among time-varying confounders, treatments, and outcomes.1
• Here, we provide proof of principle that, under a complex data generation mechanism that resembles data from a cohort of persons with HIV and sufficiently large sample sizes, an LSTM-based NICE g-formula plug-in estimator results in less bias than parametric estimators while achieving near-nominal coverage for 95% confidence intervals.1
• This work is intended to spur the development of a theoretical framework that encompasses deep learning plug-in estimators for causal inference based on high-dimensional longitudinal data.1

Sources (1)

• [1]A plug-in estimator of the non-iterative conditional expectation g-formula using long short-term memory models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 10:30 PM

  The non-iterative conditional expectation (NICE) plug-in estimator of the g-formula can be used to estimate the causal effects of sustained treatment strategies, even in the presence of treatment-confounder feedback, under the assumptions of exchangeability, positivity, and consistency.

  In practical applications, parametric models are commonly used to estimate the joint distribution of confounders, treatments, and outcomes over time.

Extractive summary: sentences quoted from the sources.