Thanks for the comment, Doug. For the first scenario, luckily our baseline outcomes are balanced between treatment and control if we exclude the lost treatment individuals from this analysis. Hence, balance on baseline outcomes is not the issue, rather the only reason to consider either option 2 or 3 is to deal with mechanical bias in attrition analysis. Specifically, if we regress "baseline individual present at endline" on a treatment dummy, ceteris paribus, we will see a negative coefficient because the lost endline data consists entirely of individuals in the treatment group. I think a bigger problem would have been if we lost balance on baseline outcomes as well, and perhaps a future blogpost can elaborate on strategies to deal with that.