Monitoring change in spatial patterns of disease: comparing univariate and multivariate cumulative sum approaches

Abstract

Prospective disease surveillance has gained increasing attention, particularly in light of recent concern for quick detection of bioterrorist events. Monitoring of health events has the potential for the detection of such events, but the benefits of surveillance extend much more broadly to the quick detection of change in public health. In this paper, univariate and multivariate cumulative sum methods for disease surveillance are compared. Although the univariate method has been previously used in the context of health surveillance, the multivariate method has not. The univariate approach consists of simultaneously and independently monitoring the disease rate in each region; the multivariate approach accounts explicitly for any covariation between regions. The univariate approaches are limited by their lack of ability to account for the spatial autocorrelation of regional data; the multivariate methods are limited by the difficulty in accurately specifying the multiregional covariance structure. The methods are illustrated using both simulated data and county-level data on breast cancer in the northeastern United States. When the degree of spatial autocorrelation is low, the univariate method is generally better at detecting changes in rates that occur in a small number of regions; the multivariate is better when change occurs in a large number of regions.