Geostatistical analysis of disease data: estimation of cancer mortality risk from empirical frequencies using Poisson kriging

Table 4 Performance comparison of alternative estimators: rank correlation coefficient between estimates and true risk values. Results obtained on average over 100 realizations generated under two different population size scenarios and 3 types of risk map (1 = observed, 2 = smooth, 3 = random). Poisson kriging was conducted with the semivariogram estimated from the underlying risk values (true γ_R(h)) or the simulated mortality rates. Bold numbers refer to best performances outside the ideal case where the true semivariogram of risk is known.

Estimators	WF population			BF population
BREAST CANCER	Scenario 1	Scenario 2	Scenario 3	Scenario 1	Scenario 2	Scenario 3
Observed rates	0.813	0.761	0.805	0.355	0.328	0.199
Population-weighted average	0.792	0.899	0.047	0.623	0.737	-0.001
Global Empirical Bayes	0.751	0.720	0.834	0.129	0.146	0.363
Local Empirical Bayes	0.896	0.922	0.825	0.612	0.708	0.217
Poisson kriging (true γ_R(h))	0.930	0.927	0.822	0.699	0.744	0.257
Poisson kriging	0.929	0.925	0.818	0.654	0.732	0.225
CERVIX CANCER
Observed rates	0.765	0.682	0.770	-0.191	-0.282	0.101
Population-weighted average	0.706	0.886	0.006	0.445	0.650	-0.017
Global Empirical Bayes	0.748	0.664	0.811	0.314	0.225	0.354
Local Empirical Bayes	0.867	0.906	0.789	0.479	0.598	0.189
Poisson kriging (true γ_R(h))	0.905	0.967	0.807	0.593	0.676	0.228
Poisson kriging	0.903	0.961	0.790	0.558	0.611	0.200

ISSN: 1476-072X