Randomization-based Z-estimation for evaluating average and individual treatment effects.

The article systematically investigates model-based analyses of treatment effects in randomized experiments within the randomization-based (finite population) inference framework, which relies solely on treatment assignment randomization without distributional assumptions on outcomes or covariates. It establishes asymptotic theory for |$Z$|-estimation, proposes conservative covariance estimation, and compares three strategies for average treatment effect estimation: model-based, model-imputed, and model-assisted, highlighting that the model-assisted approach consistently estimates average effects regardless of model misspecification and can achieve optimal precision via nonlinear least squares. Additionally, the paper introduces a framework for directly modeling individual treatment effects using exponential dispersion family models, addressing identifiability and inference challenges under randomization. These contributions provide robust methods for causal inference in randomized experiments without relying on classical superpopulation assumptions.