On Two-Sample McNemar Test

• Activity Number: 368
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Jim Xiang*
• Companies: Ortho McNeil Pharmaceutical
• Keywords: Hypothesis testing ; Paired binary data ; Pseudo likelihood ; Asymptotic normality ; Crossover design

Abstract

It is common in clinical trials we collect data in terms of existence of disease symptoms before and after a treatment. For example, fever defined as oral temperature >38 C is a common symptom associated with community acquired pneumonia (CAP). To measure the effectiveness of a treatment, we often compare the proportion of patients with symptom at the baseline to that after the treatment. McNemar (1947) proposed a test statistic for comparing two matched proportions when the outcome is dichotomous. In a randomized clinical trial, one group of patients with CAP received levofloxacin and the other group of patients received the active control. When treatment effects were compared in terms of fever resolution, a two-sample McNemar test proposed by Feuer and Kessler (1989) was performed. In this talk, we show that their test inflates type I error in hypothesis testing, and propose a new two-sample McNemar test that is superior in terms of preserving type I error. Simulations are conducted to compare both type I error and the statistical power. Real data sets from clinical trials are considered. Limitation of the two-sample McNemar test is also discussed.