Abstract
Simulated annual temperature series are used to compare seven homogenization procedures. The two that employ likelihood ratio tests routinely outperform other methods in their ability to identify modest (0.33°C; 0.6 standard deviation anomaly) shifts in the mean. The percentage of imposed shifts that are detected by these methods is similar to that based on tests that rely on a priori metadata information concerning the position of potential shifts. These methods, along with a two-phase regression approach, are also best at identifying and placing multiple shifts within a single time series. Although the regression procedure is better able to detect multiple breaks that are separated by relatively short time intervals, in its published form it suffers from a higher-than-expected Type I error rate. This was also found to be a problem with a metadata-based procedure currently in operational use. The likelihood tests are strongly influenced by the presence of trends in the difference series and short (<20 yr) series length.
The ability of a given procedure to detect a discontinuity is predominately influenced by the magnitude of the discontinuity relative to the standard deviation of the data series being evaluated. Data series length, correlation between the test series and its associated reference series, and test series autocorrelation also influence test performance. These features were not considered in previous homogenization method comparisons.
Discontinuities with magnitudes less than 0.6 times the standard deviation of the time series represent the lower limit for homogenization. Based on the most effective homogenization techniques, 10% of the 1.25 standard deviation discontinuities are likely to remain in climatic data series, unless reference station correlations are exceptional or quality station metadata are available.
Corresponding author address: Dr. Art DeGaetano, Cornell University, 1119 Bradfield Hall, Ithaca, NY 14853. Email: atd2@cornell.edu
