Response-dependent two-phase sampling designs for biomarker studies

Abstract

Two-phase sampling designs are developed and investigated for use in the context of a rheumatology study where interest lies in the association between a biomarker with an expensive assay and disease progression. We derive optimal phase-II stratum-specific sampling probabilities for analyses from parametric maximum likelihood (ML), mean score (MS), inverse probability weighted (IPW), and augmented IPW (AIPW) estimating equations. The easy-to-implement optimally efficient design for the MS estimator is found to be asymptotically optimal for the IPW and AIPW estimators we consider, and is shown to result in efficiency gains over balanced and simple random sampling even when analyses are likelihood-based. We further demonstrate the robustness of this optimal design and show that it results in very efficient estimation even when the model or parameters used in its derivation are misspecified.