TY - JOUR
T1 - Late detection of breast and colorectal cancer in Minnesota counties
T2 - An application of spatial smoothing and clustering
AU - Thomas, Avis J.
AU - Carlin, Brad
PY - 2003/1/15
Y1 - 2003/1/15
N2 - Cancers detected at a later disease stage are associated with significantly higher mortality risk. Lack of uniformity in county-level cancer detection rates is thus of substantial interest to state health departments and other public health professionals. In this paper, we perform several spatial analyses of breast and colorectal cancer detection data for Minnesota counties for 1995-1997. We look for outliers and clusters in the late detection rates using a number of techniques: (i) applying various mapping schemes, (ii) smoothing the data using Bayesian methods implemented via Markov chain Monte Carlo, and (iii) applying maximum likelihood techniques to test for the presence of clusters and to identify the most likely clusters. Our results suggest a fairly uniform spatial pattern in both sets of detection rates. Spatially smoothed rates did not reveal clusters of counties with significantly higher late detection risk, nor were county-level covariates (such as income, education, and race) particularly helpful in explaining the rates. However, our spatial clustering approach (using the scan statistic) did produce statistically significant clusters of counties which may indic te differences of practical importance for public health.
AB - Cancers detected at a later disease stage are associated with significantly higher mortality risk. Lack of uniformity in county-level cancer detection rates is thus of substantial interest to state health departments and other public health professionals. In this paper, we perform several spatial analyses of breast and colorectal cancer detection data for Minnesota counties for 1995-1997. We look for outliers and clusters in the late detection rates using a number of techniques: (i) applying various mapping schemes, (ii) smoothing the data using Bayesian methods implemented via Markov chain Monte Carlo, and (iii) applying maximum likelihood techniques to test for the presence of clusters and to identify the most likely clusters. Our results suggest a fairly uniform spatial pattern in both sets of detection rates. Spatially smoothed rates did not reveal clusters of counties with significantly higher late detection risk, nor were county-level covariates (such as income, education, and race) particularly helpful in explaining the rates. However, our spatial clustering approach (using the scan statistic) did produce statistically significant clusters of counties which may indic te differences of practical importance for public health.
KW - Bayesian methods
KW - Conditionally autoregressive (CAR) prior
KW - Markov chain Monte Carlo (MCMC)
KW - Spatial scan statistic
UR - http://www.scopus.com/inward/record.url?scp=0037438319&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=0037438319&partnerID=8YFLogxK
U2 - 10.1002/sim.1215
DO - 10.1002/sim.1215
M3 - Article
C2 - 12486754
AN - SCOPUS:0037438319
SN - 0277-6715
VL - 22
SP - 113
EP - 127
JO - Statistics in Medicine
JF - Statistics in Medicine
IS - 1
ER -