Exploratory comparison of methods for combining failure-rate data from different data sources

Thirteen methods are considered for use in pooling failure-rate data from different data sources. A Bayesian approach is taken in which two distinct sources of variation are assumed to be present; namely, prior variation between data sources and statistical error variation within each data source. An exploratory Monte Carlo simulation is used to compare the performance of the methods when used to construct both pooled point and 90% interval estimates of the failure-rate. The results indicated that those methods based on simple averaging techniques are satisfactory when only a small number of data sources are to be pooled. When there are fifteen or more data sets to be pooled, more sophisticated methods, which incorporate additional model structure, are superior. An example is given to illustrate the use of each of the proposed methods