Alcohol mortality: a comparison of spatial clustering methods

Abstract

The identification of spatial clusters of alcohol mortality can be a key tool in identifying locations that are suffering from alcohol-related problems or are at risk of experiencing those types of problems. This study compares two methods for identifying statistically significant spatial clusters of county-level alcohol mortality rates in New York. One method utilizes a local indicator of spatial association to determine which groups of neighboring counties have rates that are significantly related to each other. The other method is a spatial scan technique that calculates a maximum likelihood ratio of cases relative to the underlying population to identify the group of counties that rejects the null hypothesis of “no clustering”. The results show that because each technique bases its cluster detection on its own criteria, different counties are selected by each method. However, the overlap of the selections indicates that the two analytic methods illustrate different elements of the same clusters. Consequently, these spatial analytic techniques are seen as complimentary and are best used in tandem rather than individually. These findings suggest that multiple methods are a preferred approach to identifying clusters of alcohol-related mortality at the county level.

Introduction

Alcohol has long been recognized as causing significant deleterious health effects. In its most recent report to the United States Congress, the National Institute on Alcohol Abuse and Alcoholism (NIAAA) detailed the current understanding of the interactions between alcohol use and abuse and various social, cultural, psychological, and biological factors (U.S. Department of Health and Human Services, 1997). Alcohol is involved in as many as 43.6 percent of automobile crash fatalities, which itself is the leading cause of death in the United States for persons aged 34 and under. Cirrhosis of the liver, primarily caused by drinking, is the 11th leading cause of death in the US. Alcohol has been shown to increase the risk for some coronary and vascular disease, birth defects, and many types of cancers. It is also associated with a significant number of homicides, suicides, assaults, cases of domestic violence, and accidents at work and home. Alcohol misuse and alcohol dependence, which is a compulsive need to consume alcohol, have major roles in social and economic problems such as absenteeism from work, divorce, family violence, crime, unemployment, and homelessness (NIAAA, 1994; U.S. Department of Health and Human Services, 1997). Alcohol use is clearly associated with a substantial amount of morbidity and mortality.

Although the geographic study of alcohol-related phenomena is in its relative infancy, there is strong indication that spatial patterns are likely to exist (Wieczorek & Hanson, 1997b; Gruenewald et al., 1996). Patterns of alcohol use and problems are known to be related to such factors as socioeconomic status, gender, age, ethnicity, ancestry, and sociocultural setting (U.S. Department of Health and Human Services, 1997). Cultural and human geographers have recognized that many of these factors result in non-random geographic distributions. Outside of the United States, geographic analysis of alcohol use and problems is also relatively uncommon. Previous research has examined cultural impacts on consumption (MacAndrew & Edgerton, 1969), national variation in consumption (Hupkens, Knibbe, & Drop, 1993), and the role of specific types of alcohol outlets as related to drunk driving in Australia (Gruenewald, Stockwell, Beel, & Dyskin, 1999). Some reports have made use of mapping to show variation in alcohol-related mortality (McGlashan, 1979), but the use of spatial analysis to study alcohol problems is still relatively uncommon in epidemiology. The review of geographic information systems and spatial analytic methods in epidemiology by Moore and Carpenter (1999) did not include any citations of studies on alcohol use or health consequences. Therefore, methods of geographic analysis are likely to provide additional insight and understanding of alcohol use and related social and health problems.

In general, geographic analysis has been found to be valuable in many areas of epidemiology (Ricketts, Savitz, Gesler, & Osborne, 1994; Aangeenbrug, Leaverton, Mason, & Tobin, 1997). The utilization of cluster techniques to determine whether spatial or space–time clusters of cancers, especially rare diseases such as brain cancer, is widespread (Aldrich, Krautheim, Kinee, Wanzer, & Tibara, 1997; Caldwell, 1989). Geographic patterns, such as spatial clusters, are especially relevant for the planning and delivery of health-related services (Ricketts et al., 1994). Spatial distributions of health-related factors are useful for identifying where services are most needed, the level of services required, and the type of services that are most appropriate. Spatial patterns provide important insights into the role of social, cultural, and environmental influences on morbidity and mortality. These exploratory spatial analyses can assist in identifying the etiology of specific diseases by suggesting hypotheses inductively. Furthermore, the geographic study of disease over time is an effective method of public health surveillance. New disease hotspots can be identified quickly, and public health responses can be utilized to minimize morbidity and mortality.

This paper examines two methods for detecting statistically significant spatial clusters: Local indicators of spatial association (LISA), as defined by Luc Anselin (1995a), and Martin Kulldorff's spatial scan statistic, which uses a moving window with a continuously varying radius and a likelihood ratio calculation (Kulldorff, 1997). Anselin's method has begun to be used in an epidemiological setting. Ord and Getis (1995) demonstrated the properties of the G_i statistic, similar to LISA but without Anselin's rigorous connection to a global spatial statistic (the Moran's I), on the concentration of AIDS around San Francisco. Wieczorek and Hanson (1997a) have conducted a preliminary analysis on the usefulness of LISA when exploring alcohol mortality patterns. Kulldorff's recent enhancement of the spatial scan approach has been demonstrated on breast cancer rates (Kulldorff, Feuer, Miller, & Freedman, 1997), leukemia cases (Kulldorff & Nagarwalla, 1995; Hjalmars, Kulldorff, Gustafsson, & Nagarwalla, 1996), and sudden infant death syndrome (Kulldorff, 1997). However, the differences between the two approaches and the clusters that they reveal have not been explored and are not fully understood.

This paper compares these methods by exploring a case study of alcohol-explicit mortality rates at the county level for New York State. We examine and compare the clusters identified as significant by the two methods, and we explore the reasons for similarities and differences. Specifically we answer three questions:

• What counties are identified as significantly clustered by the LISA and the spatial scan statistic?

• Are the hotspots identified by each method the same or different?

• What are the reasons, in the methods and in the data, for the similarities or differences?

The expectations are that both methods should identify the same hotspots of high mortality. Additionally, the LISA should identify areas of clustered low mortality, and where counties with high or low rates are surrounded by the opposite. Because of the differences in the statistical formulae, however, it is likely that slightly different specific counties will be selected for the hotspots. Comparing the hotspots and specific counties identified by the LISA and the spatial scan will illustrate what is being revealed by each method.