Now up at #EuroCIM2020: Saskia le Cessie and @GelovenNan with ‘Prediction meets causal inference: the role of treatment in clinical prediction models’

AKA ‘predictimands’ (you wanna hear this one).

Mapping the ICH E9(R1) addendum on estimands to clinical prediction models with treatment initiation following the moment of prediction

1. 'Ignore treatment' strategy, similar to treatment policy estimand, hopefully self-explanatory

2. 'Composite' strategy, where treatment initiation counts as an event (so you model min of time to event and time to treatment). Estimation straightforward. Assumptions about future treatment initiation might be wobbly.

3. 'While untreated' predicts the risk of an event before time t. Estimation using competing risk approaches

4. 'Hypothetical' strategy predicts the risk in a hypothetical world where treatment does not exist. Nice in that it can inform future treatment decisions. Estimation is more difficult and requires various CI assumptions.

. @GelovenNan is now showing an example in practice. Patients with end-stage renal disease starting on dialysis; some receive a transplant; main outcome of interest is death

Are you wondering whether any of this matters? Here are predictions for the four predictimands

We're now having a role play where @GelovenNan plays a doctor explaining predictions to her patient, Saskia le Sessie. This is AMAZING!

Conclusion: the meaning of predictions provided to patients should be well-defined and crystal clear

So the paper is accepted by European Journal of Epi
and if you just can't wait, there is a pre-print up at https://arxiv.org/abs/2004.06998 
(CoI: I live-tweeted this because I'm an author)

Interesting questions:
@RuthHKeogh – What about model validation/assessment when we have defined predictimands?
Ian White – What about model selection? Might you use different covariates for different predictimands?
[I missed the other one]

Here's the hypothetical conversation between doctor and patient that explains the different predictimands. IMO this demonstrates nicely how each predictimand is of interest.

The eagle eyed among you will have noticed that Nan didn't map a principal-stratum predictimand. Yep. We just couldn't see the value.