Estimands – What should they be? What can they be?

If you want to find out, that you share quite a lot of confusion regarding the estimand concept with the RSS/PSI experts, I recommend listening to the audio recording of the RSS/PSI journal club on estimands from 2016-07-06 on YouTube [1].
Personally, I think that much of the confusion around the estimand concept (except not being clearly defined by the regulators, and no explicit examples are given) is, that while “estimand” is correctly translated from  Latin gerundive aestimandum as”that what should be estimated” the actual underlying statistical concept is “that what can be estimated”.
The first definition/translation implies that the estimand can be derived from research question, after understanding the research question enough, you arrive at the estimand which should be estimated. The second definition/translation is based on the idea that you cannot observe what you really want to know, ie the true treatment effect from a randomized controlled trial, because of post-randomization effects (treatment switching, missings), which compromise the trial design, you have to use an estimand, ie a coefficient which can be estimated based on the available data, which is closest to what you actually want to know. See also [2].
[2] Clement de Chaisemantin (Fall 2015). Lecture notes – Econometrics 1. University of Warwick.