A Potential Outcomes, and Typically More Powerful, Approach Than 'Cochran-Mantel-Haenszel' and Related Methods

• Activity Number: 294
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Cassandra Wolos Pattanayak*+ and Donald B Rubin and Elizabeth Zell
• Companies: Harvard University and Harvard University and CDC
• Keywords: Cochran-Mantel-Haenszel estimator ; confidence interval ; potential outcomes ; Rubin Causal Model

Abstract

In public health studies, outcome measures such as the odds ratio, rate ratio, or efficacy are often estimated across strata to assess the overall effect of active treatment versus control treatment. Patients may be partitioned into blocks by experimental design or into subclasses based on estimated propensity scores to improve observed covariate balance across treatment groups. We propose tests and intervals for these estimands that, in a finite sample, are typically more powerful than Cochran-Mantel-Haenszel (Cochran, 1954; Mantel-Haenszel, 1959) and analogous (Tarone, 1981) procedures. The new methods multiply impute missing potential outcomes within the Rubin Causal Model (Holland, 1986; Rubin, 1974, 1978) so that estimands can be directly estimated. The underlying assumptions are often reasonable, especially when strata are based on highly predictive covariates. These methods are usually similar to classical ones for drawing inferences about a population from which patients in the study are considered a random sample. The typically more powerful tests and typically shorter intervals are especially relevant for safety studies because smaller differences can be detected.