Two-phase biomarker studies for disease progression with multiple registries

We consider the design and analysis of two-phase studies of the association between an expensive biomarker and disease progression when phase I data are obtained by pooling registries having different outcome-dependent recruitment schemes. We utilize two analysis methods, namely maximum-likelihood and inverse probability weighting (IPW), to handle missing covariates arising from a two-phase design. In the likelihood framework, we derive a class of residual-dependent designs for phase II sub-sampling from an observed data likelihood accounting for the phase I sampling plans used by the different registries. In the IPW approach, we derive and evaluate optimal stratified designs that approximate Neyman allocation. Simulation studies and an application to a motivating example demonstrate the finite sample improvements from the proposed designs over simple random sampling and standard stratified sampling schemes.