摘要：Performance measurement is increasingly viewed as an essential component of environmental and public health protection programs. In characterizing program performance over time, investigators often observe multiple changes resulting from a single intervention across a range of categories. Although a variety of statistical tools allow evaluation of data one variable at a time, the global test statistic is uniquely suited for analyses of categories or groups of interrelated variables. Here we demonstrate how the global test statistic can be applied to environmental and occupational health data for the purpose of making overall statements on the success of targeted intervention strategies. In a previous study, Enander et al. 1 evaluated the compliance of the Rhode Island automotive refinishing industry sector with state and federal environmental, health, and safety regulations. Baseline and postintervention data were collected on 24 performance indicator variables at 82 randomly sampled facilities. The Fisher exact test was used in making statistical comparisons of proportions between baseline and postintervention data on each performance indicator. In addition to assessing each P value against the nominal .05 level of significance to determine the indicators that showed improvement, the authors used a modified Bonferroni adjustment to control the overall type 1 error rate. We describe global test statistics that we used to reach an overall conclusion about auto body compliance in performance categories identified by Enander et al. This statistical approach represents an additional method that public and environmental health professionals can use to analyze categorical data and make overall statements on the success of targeted intervention strategies. The global test statistic was originally proposed by O'Brien, 2 with further work by Pocock et al. 3 In clinical trials or in the context of an inspection program, the objective is to show improvements in postsetting indicator variables. The traditional procedure is to test all variables simultaneously via the Hotelling T 2 approach when the variables have a multivariate normal distribution. However, the T 2 test is 2 sided, and therefore conclusions on overall improvement cannot be made. In Bonferroni procedures, conclusions are made one variable at a time. In reviewing global test procedures, Geller 4 and others have noted that the intercorrelations among variables can range from moderate to high. In the Bonferroni methods, information about the relationship between variables is not used, and thus there is a loss of power. In the case of highly correlated variables, global procedures are much less conservative than the Bonferroni procedures. A variable-specific method is more appropriate in answering questions relating to effectiveness as measured by individual variables, whereas a global method might be more appropriate in answering questions relating to effectiveness as measured by uniform improvement across all variables. Thus, the 2 methods complement each other. To assist in computing global statistics, the University of Rhode Island's Computer Science and Statistics Department has posted the program online at http://www.cs.uri.edu/∼dymo-van . Because the program uses MATLAB, users need to have access to this package.
