Clinical risk prediction models enable predictions of a person’s risk of an outcome (e.g. mortality) given their observed characteristics. It is often of interest to use risk predictions to inform whether a person should initiate a particular treatment. However, when standard clinical prediction models are developed in a population in which patients follow a mix of treatment strategies, they are unsuitable for informing treatment decisions. Counterfactual risk predictions aim to address this problem – they are estimates of what a person’s risk would be if they were to follow a particular treatment strategy, given their individual characteristics that are also predictive of the outcome.
Causal inference methods typically focus on estimating population average treatment effects. In this talk I will discuss how causal inference methods can be used for individual counterfactual risk prediction using longitudinal observational data on treatment use, patient characteristics and a time-to-event outcome. An essential step in development and reporting of prediction models is to validate their performance. I will discuss the challenges of this, and describe some new methods for assessing the predictive performance of counterfactual risk prediction.
In a motivating example, we are interested in counterfactual risk predictions for mortality in patients awaiting a liver transplant under the strategies of receiving or not receiving a transplant. I will illustrate the methods using data from the US Scientific Registry of Transplant Patients.