Inference for treatment-specific survival curves using machine learning.

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning in Healthcare

Background:

• Randomized trials are ideal for causal inference but often unavailable.
• Observational data requires advanced methods to estimate treatment-specific survival curves.
• Accurate estimation of survival curves is crucial for time-to-event outcomes.

Purpose of the Study:

• To propose a novel doubly-robust estimator for treatment-specific survival curves using observational data.
• To incorporate data-adaptive methods, like machine learning, for improved estimation.
• To establish theoretical guarantees for the estimator's consistency and asymptotic linearity.

Main Methods:

• Developed a cross-fitted, doubly-robust estimator.
• Utilized data-adaptive estimators (e.g., machine learning) for conditional survival functions.
• Proposed an ensemble learner to combine multiple survival estimators.
• Methods accommodate discrete, continuous, or mixed-time events.

Main Results:

• The proposed estimator is consistent and asymptotically linear under specified conditions.
• Theoretical properties hold both pointwise and uniformly over time.
• Numerical studies and a real-world application demonstrated practical performance.

Conclusions:

• The novel estimator provides a robust approach for causal inference on survival outcomes from observational data.
• The methods are flexible, handling various event time data structures.
• This work advances the statistical toolkit for analyzing time-to-event data in the absence of randomized trials.