Multiple change-point analysis of disease incidence rates

Monitoring disease frequency in the livestock industry has become feasible by use of the recently developed health monitoring databases, Basic problems in analysing incidence rate (incidence density) over calendar time are to detect changes and to estimate the times and amounts of change. We discuss a number of plotting techniques and associated statistical hypothesis tests for retrospective detection of single and multiple change-points, The tests are modifications of well-known tests either for testing uniformity of a sample, or for comparing two Poisson variates by a likelihood-ratio test or a Pearson chi-square test. Firstly, tests for a change-point at a fixed time are modified to conservative tests for one change-point in a given finite set of possible change-points in a single interval. For the problem of finding multiple change-points we then study sequential strategies similar to procedures for selecting regressor variables in multiple linear regression. In particular, a modified forward selection technique is shown to perform well in two examples with disease incidence data from a Danish pig health and production monitoring system. In one of the data sets there is a large number of cases and fairly small changes in the rates, while the other data set has a smaller total number of cases but large changes in the rates.