Directed acyclic graphs and causal thinking in clinical risk prediction modeling

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• Causal inference and prediction modeling have distinct histories in epidemiology.
• Directed Acyclic Graphs (DAGs) are crucial for modeling causal assumptions and variable selection.
• While prediction tools aid causal inference, the use of causal knowledge in prediction is less explored.

Purpose of the Study:

• To assess the benefits of using DAGs in clinical risk prediction.
• To explore how causal structure knowledge improves model transportability across settings.
• To investigate causal knowledge's role in enhancing predictor selection for risk models.

Main Methods:

• Theoretical analysis of causal structures in prediction.
• Simulation-based study to evaluate DAGs' impact.
• Exploration of transportability and predictor selection using causal insights.

Main Results:

• Models with predictors in the causal direction show better transportability.
• The Markov Blanket (parents, children, and parents of children of the outcome) is identified as the optimal predictor set.
• Empirical evidence supports the theoretical findings on predictor optimality.

Conclusions:

• Findings provide a theoretical basis for using causes as predictors in risk models for better transportability.
• Utilizing DAGs to identify Markov Blanket variables offers an efficient predictor selection strategy.
• This approach is valuable when causal structure knowledge is available or learnable.