Causal use of prediction models

A friend and I had a discussion around the following scenario and I’m interested in some outsider opinions.

Suppose you build a prediction model or models that allow you to predict that a patient with characteristics [X] under treatment course Y would have a survival of 3 years, and a patient with those same characteristics [X] under treatment course Z would have a survival of 5 years.

  1. Are you using your model “in a causal way” if you switch all patients with characteristics [X] to treatment course Z?
  2. Is this a reasonable thing to do, given that the prediction models were developed following best practices for development of a prediction model?

If you have developed the models following best practices, could you instead reformulate your assessment of the models in terms of the probability of treatment course Y being better (longer life) than treatment course Z in a Bayesian framework? This take on the model might not answer your underlying question about whether you’re using the model in the way it was designed. But using a Bayesian model, you might be able to use the information gained from prediction to update your personal prior on any causal effect (or was that statement just jibberish?)