Bounds on the Proportion of Negative Response in Cross-Over Designs with Binary Outcomes

• Activity Number: 582
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Edwin Andong Ndum*+ and Gary L. Gadbury
• Companies: ACT, Inc. and Kansas State University
• Keywords: Bounds ; Crossover design ; Potential outcomes ; Probability of negative effect ; Causal effect ; binary outcomes

Abstract

Cross-over designs allow for a subject's outcome on more than one treatment to be observed over multiple periods. However, it is impossible to measure a true individual causal effect from observed data because only one outcome response to a treatment is observable at any particular time. Often, the difference between the proportion of subjects benefiting on a new and standard treatment is interpreted as the probability that an individual will benefit on the new treatment. This interpretation can be misleading when the effects of a treatment vary widely across subjects, i.e., subject-treatment interaction is present in the population. This talk considers bounds for a nonestimable probability of negative individual effect when the average effect is positive. The potential outcomes framework is used to quantify consequences of subject-treatment interaction. Initially, we construct bounds from the crossover design, TC-CT, with binary outcomes. Subsequently, we find tighter bounds using the TC-CT-TT-CC design in which some subjects are allowed to stay on the same treatment across the two periods. Concepts are illustrated with an example.