A journal club presentation on Kallus (2021) “More Efficient Policy Learning via Optimal Retargeting” in a Kosorok lab meeting
Kallus (2021) introduced a systematic approach to policy learning with observational data where a lack of overlap is frequently seen. An overlap, here, means that the data at hand (or the population) include all the treatment arms with enough proportions for each of the covariate values. This is a stronger condition than positivity and thus is frequently observed in practice. I present here why lack of overlap matters (instability of policy learning), what the author’s systematic solution is, and how it performs.
Be First to Comment