Skip to main content
Figure 8 | International Journal of Health Geographics

Figure 8

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

Figure 8

Impact of modeling approach on the prediction accuracy and the precision of the probability intervals (realization #50). Fifty realizations of the spatial distribution of cervix cancer mortality rates in Region 2 were simulated and then analyzed using a Bayesian (BYM model) and a geostatistical (point and area-to-area Poisson kriging) approach. Results for the 50th realization are presented. Top scatterplots (A, B) illustrate that the geostatistical risk estimates are better correlated with true risk values (smaller Mean Absolute Error of prediction, MAE) than the Bayesian estimates. Plot of the fraction of true mortality risk values falling within probability intervals (C), and the width of these intervals versus the probability p (D). The goodness statistic measures the similarity between the expected and observed fractions in the accuracy plots (best if closer to 1). Narrower probability intervals (i.e. smaller mean widths) indicate more precise models of uncertainty. (E) Ratio of accuracy and PI-width curves; whenever both ratios exceed one (black dashed line), the geostatistical PI is narrower than the Bayesian PI, while including a larger fraction of true values.

Back to article page