On Mediation Analysis with Terminal Measures Data

• Activity Number: 245
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Roshan N. Serasinghe*+ and Gary L. Gadbury and Jeffrey M. Albert and Mark Beasley and Brad A. Rikke and Thomas E. Johnson and David B. Allison
• Companies: Kansas State University and Kansas State University and Case Western Reserve University and University of Alabama at Birmingham and University of Colorado at Boulder and University of Colorado at Boulder and University of Alabama at Birmingham
• Keywords: Causality ; Indirect effect ; potential outcomes

Abstract

The total causal effect of a treatment on an outcome variable, Y, may include the direct effect of a treatment on Y and an indirect effect through another variable(s), say Z. This indirect effect is called a mediating effect (ME), and the analysis of such data is mediation analysis (MA). Situations can arise where Z and Y cannot both be measured on an individual unit. Mouse and plant experiments are two examples where measurement of Z requires terminating the animal or plant and Y is to be measured at a later time. We refer to such data as terminal measures data. Another situation may be where one experiment focused on measurement of Z, and the second on measurement of Y, and interest is in combining the data sets to evaluate the ME of Z on Y. This talk discusses the issues involved in analyzing terminal measures data. An imputation approach is proposed to construct an individual-level mediation model that relies on a blocking variable. The technique is illustrated on a mouse data set from two experiments, one where Z was measured and another where Y was measured. Sensitivity of results to a nonestimable partial correlation between Z and Y are discussed.