A scoping review of causal methods enabling predictions under hypothetical interventions

Area of Science:

• Causal inference
• Machine learning
• Health informatics

Background:

• Standard prediction models lack causal interpretability, limiting their use in decision-making for hypothetical scenarios.
• Predicting outcomes under interventions requires methods beyond traditional correlational approaches.

Purpose of the Study:

• To identify and analyze methods for developing causal prediction models for risk estimation under hypothetical interventions.
• To outline approaches, assumptions, estimands, and challenges in causal prediction modeling.
• To highlight current limitations and future research directions in the field.

Main Methods:

• Systematic literature review of health domain papers published by December 2019.
• Inclusion of methodologies from statistical and machine learning literature, and applied studies.
• Focus on papers using causal considerations for prediction under hypothetical interventions.

Main Results:

• Identified 13 relevant papers from an initial search of 5034.
• Found methods rooted in both statistical and machine learning fields.
• Marginal structural models and g-estimation are prominent for causal inference from observational data.

Conclusions:

• Two primary methods exist: augmenting observational models with trial data or direct causal estimation from observational data.
• Further development is needed for dynamic treatment regimes and clinical decision support systems.
• Validation techniques for 'causal prediction models' are underdeveloped.