Generalized Cochran-Mantel-Haenszel Test Statistics for Correlated Categorical Data

Abbott Laboratories, North Chicago, Illinois, United States
Communication in Statistics- Theory and Methods (Impact Factor: 0.27). 05/1996; 26(8). DOI: 10.1080/03610929708832016
Source: CiteSeer


S Three new test statistics are introduced for correlated categorical data in stratified R Theta C tables. They are similar in form to the standard generalized Cochran-Mantel-Haenszel statistics but modified to handle correlated outcomes. Two of these statistics are asymptotically valid in both many-strata (sparse data) and large-strata limiting models. The third one is designed specifically for the many-strata case but is valid even with a small number of strata. This latter statistic is also appropriate when strata are assumed to be random. Key words: correlated categorical data, generalized Cochran-Mantel-Haenszel statistics, sparse data. 1 INTRODUCTION In multicenter clinical trials, the responses are often recorded on a discrete scale, such as stages of disease severity or levels of improvement following an intervention. In addition, the categorical responses may be correlated because of repeated or multiple measurements on each individual or subsampling from clusters such as...

Full-text preview

Available from:
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: This paper proposes two new Mantel-Haenszel test statistics for correlated binary data in 2 x 2 tables that are asymptotically valid in both sparse data (many strata) and large-strata limiting models. Monte Carlo experiments show that the statistics compare favorably to previously proposed test statistics, especially for 5-25 small to moderate-sized strata. Confidence intervals are also obtained and compared to those from the test of Liang (1985, Biometrika 72, 678-682).
    Preview · Article · Jan 1998 · Biometrics
  • [Show abstract] [Hide abstract]
    ABSTRACT: Two features commonly exhibited by randomized trials of health promotion interventions are cluster randomization and stratification. Ignoring correlations between individuals within clusters can lead to an inflated type I error rate and hence a P-value which overstates the significance of the result. This paper compares several methods for analysing categorical data from a stratified cluster randomized trial. We propose an extension of a method from survey sampling that uses the design effect to reduce the effective sample size. We compare this with three methods from Zhang and Boos that extend the standard Cochran-Mantel-Haenszel (CMH) statistic by using appropriate covariance matrices, and with a bootstrap method. The comparison is based on empirical type I error rates from a simulation study, in which the number of clusters randomized is small, as in most public health intervention studies. The method that performs consistently well is one of the Zhang and Boos extensions of the standard CMH statistic.
    No preview · Article · Dec 2002 · Statistics in Medicine
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Multilevel data are a commonly encountered phenomenon in many data structures. Modelling such data requires careful consideration of the association between underlying variables at each level of the data structure. This requires the use of effective univariate techniques prior to modelling. However, currently no univariate tests are used to handle this situation. This paper presents the modification and novel application of a test developed by Zhang and Boos for testing the association between categorical variables measured on clusters of observations, for examining initial association in a multilevel framework. Zhang and Boos have used a SAS/IML programme (unpublished) for performing their test. This paper presents an R function for the application of the test, which will be freely available to users, since R is an open source software. The function is tested on a dataset from the medical field pertaining to respiratory disease severity of patients, attending several different clinics. The explanatory variables pertaining to this study are Age, Gender, Duration and Symptom, while the response variable indicating the severity of the diagnosis made is termed Diagnosis. The results indicate that when the experimental units show low levels of correlation within clusters with respect to a particular explanatory variable, the test performs similarly to the Standard Cochran Mantel Haenszel (CMH) test. When the corresponding correlation is high, the Generalized CMH (GCMH) test results in a smaller p-value than the Standard CMH test. Of the four variables, only Symptom and Duration are significant with respect to association with Diagnosis.
    Full-text · Article · Jun 2012 · Journal of the National Science Foundation of Sri Lanka
Show more