Counterfactual prediction modeling is different from prediction modeling since it involves the evaluation of hypothetical scenarios where input variables can be intervened upon (i.e., changed) to modify the outcome. For instance, changing a behavior to decrease risk of disease, or changing a drug to improve disease outcome.
If data and experiments were randomized with respect to an intervention variable, prediction and counterfactual prediction would be basically the same. However, many data sources used to develop prediction models are not randomized (e.g., electronic health records) and can contain different biases (e.g., confounding, selection). Therefore, prediction models are not assured to be usable in counterfactual mode to assess hypothetical interventions. Yet, this important difference in capabilities of prediction models is sometimes overlooked.
The aim of this collection is to encourage methodological and applied studies concerning the development, validation, and application of counterfactual prediction models in healthcare.
