Estimating the causal effects of sustained treatment strategies is hard when time-varying confounders are affected by prior treatment. The non-iterative conditional expectation (NICE) plug-in estimator of the g-formula handles this under standard causal assumptions, but practical implementations typically rely on parametric models that may miss complex temporal dependencies.
The authors propose using long short-term memory (LSTM) networks instead. LSTMs are designed to capture both short-term patterns and long-range dependencies in sequential data, making them a natural fit for longitudinal confounders, treatments, and outcomes. However, the statistical properties of an LSTM-based plug-in estimator had not been studied.
In simulations resembling a cohort of people with HIV, the LSTM-based estimator produced less bias than parametric estimators and achieved near-nominal 95% confidence interval coverage when sample sizes were large. The authors frame this as proof of principle, not a full theoretical treatment, and call for future work to develop a rigorous framework for deep learning plug-in estimators in high-dimensional longitudinal causal inference.