Detection of clusters of a rare disease over a large territory: performance of cluster detection methods

For many years, the detection of clusters has been of great public health interest. Several detection methods have been developed, the most famous of which is the circular scan method. The present study, which was conducted in the context of a rare disease distributed over a large territory (7675 cases registered over 17 years and located in 1895 units), aimed to evaluate the performance of several of the methods in realistic hot-spot cluster situations.

All the methods considered aim to identify the most likely cluster area, i.e. the zone that maximizes the likelihood ratio function, among a set of cluster candidates. The circular and elliptic scan methods were developed to detect regularly shaped clusters. Four other methods that focus on irregularly shaped clusters were also considered (the flexible scan method, the genetic algorithm method, and the double connected and maximum linkage spatial scan methods). The power of the methods was evaluated via Monte Carlo simulations under 27 alternative scenarios that corresponded to three cluster population sizes (20, 45 and 115 expected cases), three cluster shapes (linear, U-shaped and compact) and three relative risk values (1.5, 2.0 and 3.0).

Three situations emerged from this power study. All the methods failed to detect the smallest clusters with a relative risk lower than 3.0. The power to detect the largest cluster with relative risk of 1.5 was markedly better for all methods, but, at most, half of the true cluster was captured. For other clusters, either large or with the highest relative risk, the standard elliptic scan method appeared to be the best method to detect linear clusters, while the flexible scan method localized the U-shaped clusters more precisely than other methods. Large compact clusters were detected well by all methods, with better results for the circular and elliptic scan methods.

The elliptic scan method and flexible scan method seemed the most able to detect clusters of a rare disease in a large territory. However, the probability of detecting small clusters with relative risk lower than 3.0 remained low with all the methods tested.