Demystifying Statistical Inference When Using Machine Learning in Causal Research.

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• Causal inference in public health increasingly uses machine learning (ML).
• Valid statistical inference is crucial for reliable ML applications in research.
• Methodological advancements are needed to address challenges in ML-based causal inference.

Purpose of the Study:

• To comment on recommendations for valid statistical inference using ML in causal research.
• To highlight the importance of the Super Learner library in ensemble methods.
• To demonstrate alternative approaches for achieving valid inference.

Main Methods:

• Review of prominent methodological work in ML for causal inference.
• Simulation studies to evaluate targeted maximum likelihood estimation (TMLE) and Super Learner libraries.
• Comparison of TMLE with and without sample-splitting.

Main Results:

• Targeted maximum likelihood estimation (TMLE) can achieve low bias and valid statistical inference without sample-splitting.
• A Super Learner library excluding tree-based methods but including regression splines was effective.
• The necessity of extremely data-adaptive algorithms and sample-splitting is problem-dependent.

Conclusions:

• The composition of the Super Learner library is critical for valid causal inference.
• Sample-splitting may not always be necessary when using TMLE.
• Further research is needed for practical guidance on selecting ML methods in epidemiology.