Hannah Li

Randomized controlled trials are the gold standard for evaluating interventions — yet a broad class of trials systematically violates a core assumption of standard experimental design. When interventions are delivered by capacity-constrained resources — teachers, healthcare providers, social workers, and countless others — participants compete for finite service capacity, creating hidden variation in the treatment itself. We call this operational dosage. Using queueing-theoretic methods, we show that in such settings, treatment effects are both capacity- and sample-size-dependent, and that statistical power is non-monotone in sample size: adding participants beyond a critical threshold reduces power, directly contradicting conventional power calculations. We validate these insights through simulations calibrated to a real-world tuberculosis intervention trial in Kenya. Our results reveal a fundamental design lever — capacity selection — that has been largely ignored in experimental practice, and offer a unifying explanation for why promising interventions so often fail to replicate or scale.

Joint work with Justin Boutilier, Jonas Jonasson, and Erez Yoeli.