Open tykim6 opened 1 year ago
Most of our estimators fit linear treatment effect models, in which case the treatment effect θ(X) (where X is the set of features) will indicate how strongly each treatment effects the outcome, so (for costless treatments) you'd optimally set each treatment with a positive coefficient as high as possible and each treatment will a negative coefficient as low as possible. In practice, you would want to apply domain knowledge to determine what range of treatments your existing data covers and how far you'd be comfortable extrapolating from that, since treatment effects are unlikely to really be linear for arbitrarily high values in general settings.
If you want to include non-linear effects, then you can featurize your treatment array to include higher-order terms (and our next release will make this much easier; for now you'd need to do it by hand), and you could also incorporate treatment costs by hand in devising your own strategy for optimizing treatment.
Is there a method for predicting optimal treatments based on an unseen dataset of covariates? Is there a valid way to choose an optimal treatment for an individual record?
If it makes a difference, I am using multiple treatments (3)