This study suggests that both estimation methods provide a useful way to characterize the spatial aspects of this birth outcome. The literature has strongly advocated the use of GIS in surveillance of the maternal environment and its impact on birth outcomes [12, 25–33]. Yet, there are relatively few efforts to integrate and/or compare analytical techniques. SaTScan and Spatial filtering cluster estimation methods produced noticeably different results from the same individual level birth data. SaTScan clusters are likely to differ from Spatial filtering clusters in terms of population characteristics and geographic area within clusters. Using the two methods in conjunction could provide more detail about the population and spatial features contained with each type of cluster.
First, the two methods yielded many significant spatial clusters of low birthweight in Shelby County. We fully expected to find these clusters since the birth outcomes in this Midsouth area are historically among the nation's worst. The annual infant mortality rate in parts of Memphis frequently exceed 16 per 1,000 live births as opposed to about 10 per 1,000 live births in the state and about 7 per 1,000 live births nationally . In 2002, 14% of children in Memphis were born prematurely as opposed to 12% nationally, while approximately 12% of children were born with low birth weight as opposed to 8% nationally . Since 1996, about 40% of child deaths under the age of 18 resulted from prematurity .
Second, Spatial filtering clusters appear to cover much more geographic area han did SaTScan clusters. As we discussed earlier, this is likely due to the difference in basic assumptions between the two models. The two methods differ in their estimation of significance; SaTScan accounts for multiple testing of the highest likelihood of occurrence for all possible cluster locations and sizes while Spatial filtering does not. In addition, there may be somewhat of a "ceiling effect" with SaTScan. This maximum value ensures that the detected clusters, regardless of their location and size, are clusters detected without any pre-selection bias. The maximum allowed value of a spatial cluster does not mean that one has to pre-specify the size of a cluster before running an analysis. It simply means the largest allowed cluster would contain 50% of the at-risk population in the study area. This maximum value is reasonable because a cluster is expected to concentrate in certain areas of the study region. If a cluster covers most of the study area, then the location and size of the study area is no longer meaningful in that study area. Consequently, SaTScan clusters have an inherent but adjustable "cap" on cluster size whereas Spatial filtering is somewhat less limited. The tendency or capacity of Spatial filtering to yield clusters with considerably more geographic variability than SaTScan raises the issue of sample generalizability. Although not demonstrated in this study, the potential for greater variability of the characteristics of births within a Spatial filtering cluster may provide some analytic benefits.
Third, the maternal and familial characteristics of births contained within the two methods were remarkably alike. Additionally, changing the level of geographic scale resulted in very similar patterns between the two methods with respect to these characteristics. As the level of scale increased the sample became increasingly heterogeneous. We know that geographic scale is an important consideration in such investigations irrespective of the analytic method used . This study found that clusters of low birthweight in Shelby County might extend into less impoverished, better educated, and more ethnically diverse communities.
We do not contend that either cluster estimation method is inherently superior. Instead this study underscores the need for an exploratory, integrative, and multi-scalar approach to assessing geographic patterns of disease, since different methods identify different patterns. First, both methods should be compared again using a Poisson approach. A temporal analysis of low birth weight could be conducted. Secondly, the two methods should be compared using different forms of chronic morbidity (i.e., congenital anomalies). Since other diseases differ in prevalence, population, and/or spatial characteristics, the results of the two methods might differ accordingly. Finally, the two methods should also be compared with respect to additional spatial characteristics. For example, the density or distribution of point sources of pollutants within either type of cluster could be examined on varying geographic scales.