A Bayesian normal homogeneity test for the detection of artificial discontinuities in climatic series

A Bayesian Normal Homogeneity Test (BNHT) for the detection of artificial discontinuities in climatic series is presented. The test is simple to use and allows the integration of prior knowledge on the date of change from various sources of information (e. g. metadata or expert belief) in the analysis. The performance of the new test was evaluated on synthetic series with similar statistical properties as observed total annual precipitation in the southern and central parts of the province of Quebec, Canada. Different priors were used to investigate the sensitivity of the test to the choice of priors. It was found that (1) high-prior probability of no change yields low false detection rates on the homogeneous series; (2) the test has a very high power of detection on series with a single shift (the best power of detection if compared with previous methods applied to the same synthetic series); (3) shifts having a small magnitude are detectable with a low prior probability of no change and (4) when applied to series with multiple shifts with a segmentation procedure and a high probability of no change, the test proved to be performing well in detecting multiple shifts (as performing as the best techniques previously applied to the same synthetic series). An example of application to total annual precipitation in Quebec City, Canada is also presented to illustrate (1) a case for which the results are not affected by the choice of the prior parameters and (2) a case for which information about potential changes found in the metadata was integrated in the analysis and allowed the detection of a change that would not have been detected with a non-informative prior. Copyright (C) 2009 Royal Meteorological Society