The Validity and Power of Extreme Sampling Schemes for Mediation Analysis

• Activity Number: 188
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Robert Makowsky*+ and Mark Beasley and Gary L. Gadbury and Jeffrey M. Albert and David B. Allison
• Companies: University of Alabama at Birmingham and University of Alabama at Birmingham and Kansas State University and Case Western Reserve University and University of Alabama at Birmingham
• Keywords: Mediation ; Sampling ; Missing data

Abstract

In some cases, the costs of measuring the putative mediator (Z) or the outcome (Y) can be large. Extreme sampling procedures may reduce study costs by increasing power per subject measured on the more expensive variable. Therefore, we explored whether sampling schemes will produce reliable parameter estimates and increase power for fixed study cost or, complementarily, decrease cost for fixed study power. Three sampling strategies were considered; 1) sampling from above and below the 75th and 25th percentiles, respectively, of Z ('quartile-sampling'), 2) quartile-sampling from Z conditional on the study arm, and 3) removal of half of the specimens based on ordering of Z. Simulation results showed that all sampling schemes produced biased correlation coefficients when analyses were based on specimens with complete data, leading to inflated Type 1 Error rates. When maximum likelihood was used to handle missing data, bias in point estimates was eliminated. For all sampling approaches, an increase in the variance of most correlation estimates was observed, and this may account for Type 1 Error rate inflation in some situations, despite lack of bias in the correlation coefficients.