In uplift modeling with an X-learner metalearner (Künzel et al. 2019), predictions from the two first-stage models are used in training the two second-stage models. Question: What datasets/splits should these predictions be run on exactly? E.g. should the treatment-group first-stage model predict the entire training set from the control-group first-stage model, and vice-versa?
Context: Say we have e.g. 20 weeks of data, hold out the last 4 weeks as a test set, and use cross-validation on the first 16 weeks when training the first-stage models. (Please feel free to comment if you see any issues with this approach.) Our ultimate outcome and our treatment are both binary.
Thanks all!
