Testing the efficacy of preventive maintenance

During the service lifetime of a repairable machine, regular preventive maintenance or scheduled overhauls may be carried out. One may wish to test the null hypothesis that the overhauls have no effect on the subsequent sequence of times to failure, or on the costs of the failures (e.g. labour time for repairs). Requirements of a satisfactory statistical test are discussed, and a test based on partial likelihood is introduced. To exemplify its practical application, the test is used to analyse published data on failure of submarine engines, and a dataset collected during a recent study of hospital equipment. Some generalizations that would involve more detailed modelling of the effect of maintenance are discussed. A related permutation test would be a robust alternative test in the case of several identical machines, or as a test for trend.