How does Poisson kriging compare to the popular BYM model for mapping disease risks?

Table 7 Performance comparison of geostatistical and Bayesian estimators: proportion of false positives.

Estimators	Lung cancer		Cervix cancer
LOW RISK THRESHOLD (RR = 1)	Average	% best result	Average	% best result
BYM model	0.071	30	0.077	12
Point Poisson kriging (adjacent counties)	0.069	20	0.066	14
ATA Poisson kriging (adjacent counties)	0.070	30	0.063	24
ATA Poisson kriging (32 neighbors)	0.072	20	0.061	50
HIGH RISK THRESHOLD	RR = 1.1		RR = 1.25
BYM model	0.069	20	0.080	4
Point Poisson kriging (adjacent counties)	0.066	14	0.065	12
ATA Poisson kriging (adjacent counties)	0.064	36	0.064	32
ATA Poisson kriging (32 neighbors)	0.064	30	0.060	52

Results obtained on average over 50 realizations generated for Regions 1 and 2. Poisson kriging was conducted using either adjacent counties (same neighbors as BYM Model) or the 32 closest counties in terms of distance between population-weighted centroids. ATA kriging accounts for the shape and size of the counties in the analysis. The proportion of false positives is derived from Receiver Operating Characteristic (ROC) curves computed using a low (Relative Risk, RR = 1) and high risk threshold (RR = 1.1, RR = 1.25). Bold numbers refer to best performances: few false positives. The second column gives the percentage of realizations where the particular method yields the best results.

ISSN: 1476-072X