Figure 5

Differences between the cervix cancer mortality risk and prediction variance computed by the Bayesian and geostatistical methods. Maps (A, B) highlight the counties where the Bayesian and geostatistical methods differ the most in terms of risk estimates (absolute differences) and prediction variance (relative differences). A diverging color scheme with three breaks (mid-point = 0 or 1) was chosen; counties with values in between the breaks receive a blend of the two break colors. (C) ATA Poisson kriging yields larger risk estimates than BYM model in high-valued areas, while lower risks are predicted in low-valued areas. (D, E) The lognormal hypothesis underlying the BYM model leads to larger prediction variance for larger risk estimates, while the ATA Poisson kriging variance increases in counties with small population at risk and large areas. The size of the dots in scatterplots is proportional either to the size of the population at risk or the area of each county.