Cluster analysis in geographical epidemiology: the use of several statistical methods and comparison of their results

Background: The increasing interest in environmental epidemiology has been followed by the development of many statistical tests for detecting disease clustering near a point source. The objectives of this study were to compare several tests to detect disease clustering, among which modelisation using Markov Chain Monte Carlo methods. Methods: We compared six statistical methods for detecting disease clustering of bladder cancer around an industrial centre of Isere (France) for the period 1983-1997: Stone's test, score test, and two log-linear modelisations (with and without corrections for extra-Poisson variations) using two ways of parameters estimation (maximum likelihood and Markov Chain Monte Carlo methods). Results: The results of the Stone test and the score test are not in favour of a higher risk of bladder cancer around the considered point source. The conclusions brought by the log linear modelisations are the same, but the results obtained using the Markov Chain Monte Carlo Method are very dependant of prior distributions determined for the different parameters. Conclusion: Markov Chain Monte Carlo methods, which allow taking into account complex geographical effects, seem well adapted to cluster analysis in geographical epidemiology. However, they remain difficult to implement.