Einstein Circle Lecture | Nan van Geloven

Einstein Circle Lecture

“(The Lack of) Causal Inference in Prediction Models”

A key aim of prediction models is to provide individualized risk estimates that guide treatment choices. Most prediction models, however, are derived from observational data that include some individuals who have already received the treatment that the model is meant to inform. The causal effects of these earlier treatments blur the interpretation of regular prediction models, making them unfit for decision support.

An alternative approach that recently gained traction is to estimate risks under specific (hypothetical) treatment options, referred to as ‘prediction under interventions’. This approach requires tailored causal inference techniques. In this talk, van Geloven will discuss methods for developing and evaluating models for prediction under interventions, as well as a recent application of these methods at the neonatal intensive care unit.