Efficient Two-Phase Sampling Designs for Predicting a Binary Outcome Variable

• Activity Number: 153
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): Jinbo Chen*
• Companies: University of Pennsylvania Perelman School of Medicine

Abstract

Two-phase design can be used as an efficient sampling strategy for assessing the value of a variable in predicting disease risk. When the first phase consists of a frequency-matched case-control sample, we develop efficient sampling designs for selecting phase II subjects. To this end, we considered three sampling strategies: 1) the second phase is a simple random sample that is stratified on case-control status only, not on the matching variables; 2) The second phase is a random sample that is stratified on both case-control status and matching variables; 3) The second phase is stratified on case-control status and phase-I variables that were not used in matching. Using a logistic regression model to describe the association between case-control status, matching variables, and phase-I and phase II variables, we considered inverse probability weighting methods and pseudo-likelihood methods for estimating odds ratio association parameters under each study design. We compared the efficiency of the three study designs for estimating statistical measures for quantifying the predictive accuracy of the subsequently developed risk prediction model