Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2016 May;27(3):347-55.
doi: 10.1097/EDE.0000000000000450.

A Bayesian Method for Cluster Detection With Application to Brain and Breast Cancer in Puget Sound

Affiliations
Free PMC article

A Bayesian Method for Cluster Detection With Application to Brain and Breast Cancer in Puget Sound

Albert Y Kim et al. Epidemiology. .
Free PMC article

Abstract

Cluster detection is an important public health endeavor, and in this article, we describe and apply a recently developed Bayesian method. Commonly used approaches are based on so-called scan statistics and suffer from a number of difficulties, which include how to choose a level of significance and how to deal with the possibility of multiple clusters. The basis of our model is to partition the study region into a set of areas that are either "null" or "non-null," the latter corresponding to clusters (excess risk) or anticlusters (reduced risk). We demonstrate the Bayesian method and compare with a popular existing approach, using data on breast, brain, lung, prostate, and colorectal cancer, in the Puget Sound region of Washington State.

Figures

Figure 1
Figure 1
Study region with Seattle (solid line) and Tacoma (dashed line) metropolitan regions highlighted and the city of Mount Vernon marked with a dot.
Figure 2
Figure 2
Maps of income unadjusted (left column) and income adjusted (right column) expected counts of brain cancer.
Figure 3
Figure 3
Maps of income unadjusted (left column) and income adjusted (right column) standardized morbidity ratios for brain cancer.
Figure 4
Figure 4
Maps of income unadjusted (left column) and income adjusted (right column) multiple cluster scan statistic results for brain cancer.
Figure 5
Figure 5
Maps of income unadjusted (left column) and income adjusted (right column) expected counts of breast cancer.
Figure 6
Figure 6
Maps of income unadjusted (left column) and income adjusted (right column) standardized morbidity ratios for breast cancer.
Figure 7
Figure 7
Maps of income unadjusted (left column) and income adjusted (right column) multiple cluster scan statistic results for breast cancer.
Figure 8
Figure 8
Wide and narrow distributions on the relative risk. Under the null (no clusters/anti-clusters), relative risks θ are assumed to arise from the narrow prior, so that there is still a small amount of “wobble” about 1. Under the alternative (at least one cluster/anti-cluster), the relative risks are assumed to arise from the wide prior, so that there is greater variation. Note that the θ scale is logarithmic.
Figure 9
Figure 9
Maps of income unadjusted (left column) and income adjusted (right column) posterior probabilities of brain cancer cluster membership.
Figure 10
Figure 10
Prior/posterior probabilities of the number of brain cancer clusters/anti-clusters.
Figure 11
Figure 11
Sensitivity of posterior probabilities of the number of brain cancer clusters/anti-clusters to π0. Tracing horizontal lines across the plot give the set of posterior probabilities for that π0 value.
Figure 12
Figure 12
Maps of income unadjusted (left column) and income adjusted (right column) posterior probabilities of breast cancer cluster membership.
Figure 13
Figure 13
Prior/posterior probabilities of the number of breast cancer clusters/anti-clusters.
Figure 14
Figure 14
Sensitivity of posterior probabilities of the number of breast cancer clusters/anti-clusters to π0. Tracing horizontal lines across the plot give the set of posterior probabilities for that π0 value.

Similar articles

See all similar articles

Publication types

Feedback