Data Fusion Estimators

The estimators are semiparametrically efficient when nuisance functions are correctly specified. The proposed one-step estimators can use nuisance estimates from parametric, semiparametric, or machine-learning models. With flexible learners, the paper proposes cross-fitted debiased machine learning using K-fold sample splitting. Non-monotone cumulative incidence estimates can be projected onto non-decreasing functions without changing asymptotic behavior. Participant-level data fusion can estimate counterfactual clinical incidence curves for vaccine regimens evaluated only through immunologic outcomes, subject to causal assumptions.