The results indicate that between 1990 and 1998, four geographic regions were identified with excess mortality rates in Texas that were statistically significant. With respect to suspected excess mortality, the regions detected with excess breast cancer mortality were consistent with those presented in the analyses of the 1970–1994 data by the US National Cancer Institute  and the 1990–1997 data reported by Zhan.  The results rendered supporting evidence that most counties that were previously suspected of having elevated breast cancer mortality do indeed have excessive cancer mortality. The present study additionally identified West Texas counties as having excess mortality from breast cancer in the Hispanic female population that persisted for 9 years, which was not previously reported. The relative risk of this cluster was at the modest level of 1.18. Nevertheless, this region had the highest in relative risk, with the longest temporal persistence among detected potential clusters of all racial groups in this study. Based on this finding and on the comparisons of LLRs for the primary suspected clusters from both scan trials (non-Hispanic Whites LLR = 35.00 vs. Blacks LLR = 10.01 vs. Hispanic LLR = 29.01 vs. Others LLR = 8.30), it was determined that the Hispanic and non-Hispanic White female populations in the regions detected with clusters had the highest burden of breast cancer mortality, as evidenced by both temporal persistence and spatial concentration.
The verification of breast cancer excess mortality over time may prove beneficial for health policy and planning. For instance, the state of Texas has yet to reach the Healthy People 2010 Objective of 16.3 deaths per 100,000 females in the population, as provided by the U.S. Centers for Disease Control and Prevention (i.e. Objective 3-3, "reduce the breast cancer death rate", http://www.healthypeople.gov/Document/html/tracking/od03.htm). Spatiotemporal analysis such as that described in this study will be instrumental in planning and reaching the projected objective. For instance, the present analysis underscored the two regions with multiple racial groups that bear the persistent burden of breast cancer mortality, and detected a potential 9-year persistence of excess breast cancer mortality with the highest relative risk in the Hispanic population. Both regions carry a disproportional burden of excess mortality and warrant further investigation and policy intervention. The results of spatiotemporal analysis quantified disease burden over time by both spatial concentration (as determined by p values, LLRs and relative risks) and temporal persistence (as determined by the duration of detected clusters), which presented another perspective of measuring health disparity. It contributed to an understanding of the persistent burden of the disease across space and time, as well as aiding in determining whether the mortality burden that may have persisted into the current decade.
Several research notes arose from this study and warrant elaboration. First, as identified in the present study, the very modest relative risks that occurred over a large region of contiguous counties in Texas did not necessarily meet the strict definition of "clusters" of epidemic intensity. Compared with previous studies using SatScan for cluster detection, [16, 19] the relative risks revealed in the present study were apparently lower, and no localized, hot-spot clusters (with constant, high risks in the clusters) that persisted over time were detected. On the other hand, breast cancer may have a substantial developmental period and may have potential risk/vulnerability factors, such as the stage at diagnosis, access to treatment, and the exposure to environmental toxic wastes that are not fully understood. These potential contributors were not accounted for in the present study. Given that many of these causes and risk factors may have operated over various time scales, the mortality examined here is only an endpoint in that process. While early detection of cancer is generally beneficial to survival, there is controversy over the effectiveness of breast cancer screening in reducing mortality.[20, 21] Ideal interventions may also target modifiable risk factors that exist above and beyond the windows of space or time considered here. Nevertheless, this study offered baseline descriptions of persistently elevated breast cancer deaths in Texas, which may serve as a point of departure for policy deliberation and health resource allocation. Second, although this study focused primarily on statistically significant excess mortality, it by no means suggested that those non-statistically-significant regions of excess mortality were less important. To be statistically significant at the 0.05 or 0.01 level, outcome measures had to satisfy the Poisson distribution model and all independent variables of this study, including space, time and age, had to fit into the model simultaneously, and produce a large LLR as a result of spatial-temporal analysis. For example, the potential cluster detected among Blacks between 1991–1996 in Gulf Coast Texas (RR = 1.15, p = 0.12) was for all age groups. However, the results may become statistically significant if analysis was conducted with the stratification of certain age groups, such as among Black females aged 25 to 40. Therefore, the p-value derived is construed as an indicator, suggesting the level of excess mortality that calls for further investigation. Third, the choice of county level analysis entailed the strengths and weaknesses intrinsic to this level of aggregation. Although sub-county level (such as census tracts or block groups) of analysis may be preferred in cancer analysis,  we chose the county-level data because this level of aggregation was used in other studies on detecting breast cancer clusters,[4, 15] and also because the disproportional demographic distribution of Texas population made sub-county level analysis less feasible. For example, there were seven border counties that averaged fewer than ten breast cancer deaths, and had a population of less than 900 residents during the study period. The rates based on these small numbers of events and small population sizes tend to be unpredictable and often inflated. In particular, several of the above sparsely populated counties were in the summary choropleth maps of the age-adjusted mortality rates by race presented in Figures 1 to 3. These inflated mortality rates tend to produce visual bias, as these are the counties that attract more visual attention due to the intensive colour shading. Readers are advised to use caution when trying to interpret health outcomes, including excess mortality in these sparsely-populated counties. Fourth, with reference to data management, we found that most spreadsheet programs are limited in accommodating data in a worksheet (for example, 65,536 rows or records). Conventional spreadsheet programs are insufficient for storing the aggregated data of multiple years required in the present study (N > 195 k). Instead of using spreadsheet programs, we recommend the use of a relational database for data management for spatial-temporal analysis using SatScan.
As observed in Jacquez and Greiling [23, 24] of the Journal, several methodological issues involving spatial analysis warrant consideration when using SatScan for cluster analysis. Among other potential limitations are the assumption that the clusters are cylindrically-shaped and the constraints that are attributable to centroids and the edge effects of the scan method. As the authors advised, the scan statistic is but one tool that one may bring to bear on the study of geographic variation in cancer. Nevertheless, th e shape of cluster detection in SatScan may be enhanced over time, and improved methods of utilizing SatScan for cluster analysis are emerging [25, 26]. As users of this program, we found that it affords a great opportunity of analyzing the unique geo-demographic composition of Texas data. Particularly appealing is the fact that the program is in the public domain. It provides an opportunity for the integration (as a calculation engine) with other mapping programs, such as EpiMap, freely available from the CDC. Currently, the authors are working to develop an integrated solution using these two programs as a health surveillance system for Texas counties.