Increasing Power for Observational Studies of Aberrant Response: An Adaptive Approach

Abstract In many observational studies, the interest is in effect of treatment on bad, aberrant outcomes rather than average outcome. For such settings, traditional approach to define a dichotomous outcome indicating aberration from continuous score and use Mantel–Haenszel test with matched data. example, studies determinants poor child growth World Health Organization’s definition stunting being height-for-age z-score ≤ − 2. The may lose power because it discards potentially useful information about severity aberration. We develop an adaptive that makes this asymptotically dominates approach. our two parts. First, we rank prove novel design sensitivity formula enabling its asymptotic comparison under various settings. Second, new, general approach, two-stage programming method, adaptively combine test. apply study teenage pregnancy stunting.