Evaluating the performance of Bayesian smoothing methods for local cluster detection using a simulation design

Lemke D, Hense HW

Zusammenfassung

Introduction: Bayesian smoothing techniques are applied in the field of disease mapping with the aim of removing random noise from the risk estimates and these methods are also increasingly applied in the field of local cluster detection. The objective of this study was to compare the performance, in terms of accuracy and precision, of two empirical (Poisson-Gamma model and Log-normal model) and one hierarchical Bayesian (BYM model) model, using an artificial risk surface.Methods: The artificial risk surface was constructed with the lung cancer incidence rates for both sexes in the age group of 40-79 years as provided for the Regierungsbezirk Münster by the Cancer Registry NRW. An urban and a rural cluster were selected, and the background incidence rates were doubled to create artificial clusters with twofold relative risk (RR) increments. The other areas were postulated to be non-cluster areas consistent with the constant risk hypothesis. Cases were sampled from this risk surface as geocodes using an inhomogeneous Poisson process with 1000 realizations and the simulated observed cancer cases were used to evaluate the sensitivity and specificity of the three smoothing methods at two different spatial scales: census tracts (N=1983, n=43 risk areas) and communities (N=78; n=2 risk areas). As threshold for detecting a cluster area, the lower bound of the 95% credible interval for the RR estimate had to be >1.25 (census tracts) and >1.0 (communities).Results: The two empirical models showed a contradictory behaviour: the PG-model shrunk the risk estimates towards the global mean such that, at census tract level, only an average sensitivity below 5% with an average specificity of 100% was achieved. In contrast, the Log-normal model inflated the risk estimates and the mean sensitivity was above 95% while the mean false positive rate was below 10%. At community level, the PG-model showed an increased sensitivity (+ 15%) while maintaining the very high specificity, whereas for the Log-normal model results were similar to the census tract level. The BYM model failed at both spatial scales to detect any cluster area.Conclusion: None of the Bayesian smoothing techniques was capable of adequately capturing the spatial heterogeneity of a moderate relative risk increase in areas of high or moderate spatial resolution. Therefore, these methods appear inappropriate for use in cluster detection or disease surveillance.