Thank you @f2harrell - I think that this helps me to get my head around this partially. I’m still pondering if individual treatment effects (which requires some form of prediction because we only see each participant under one condition and not the other) should be thought of as an out of sample prediction problem where we should only use data that was available at the time of randomization, or more akin to imputation which I think of as an within sample prediction problem (albeit with purposefully added noise) were we use any data that we have to “fill in the blanks”.