American Statistical Association New York City Metropolitan Area Chapter Memorial Sloan-Kettering Cancer Center Biostatistics Seminar

To address a study objective to estimate a mean or mean difference of an expensive endpoint Y, one approach employs a two-phase sampling design, wherein inexpensive auxiliary variables W predictive of Y are measured in everyone, Y is measured in a random sample, and the semiparametric efficient doubly robust estimator is applied. If the auxiliaries are correlated with Y, this approach is more efficient than analyzing the subgroup with Y measured. Moreover, efficiency is maximized by specifying the phase-two selection probabilities as optimal functions of the auxiliary variables and measurement costs. While this approach is familiar to survey samplers, it apparently has seldom (if ever) been used in clinical trials, and several results practicable for clinical trials are developed. Simulations are presented to identify settings where the optimal approach significantly improves efficiency compared to simpler approaches.

The results are developed to design a Phase I HIV vaccine trial, where the primary objective is to compare the mean importance-weighted breadth (Y) of the HIV-specific T cell response among randomized vaccine groups. The trial will collect a crude measure of breadth (W) for all subjects, which is highly predictive of Y, and will measure Y in an optimally chosen subset. Depending on the optimality result and the assumptions about measurement costs, we show that the optimal design-estimation approach is expected to confer up to a 20% efficiency gain for this trial compared to the approach that uses the same estimator but simple random sampling. In general, the potential for efficiency gain via optimal sampling increases with the variability of Var(Y|W) in W. The simulations show that accurate estimation of E[Y|W] based on the phase-two data is important for realizing the efficiency gain, which is aided by collecting a large enough phase-two sample and by using a robust fitting method.