A non-correlation-based approach to the search for anthropogenic climate change signals

Predictions of warming by General Circulation Models (GCMs) for a greenhouse-gas enhanced atmosphere have led to the Rio Climate Treaty, the Kyoto Protocol, and other efforts to reduce greenhouse gas emissions. These efforts have been based, to a large degree, on an apparent correlation between the model simulations and surface observations. The ability of the models to reproduce the current climate has been the focus of several research groups, including the Atmospheric Model Intercomparison Project (AMIP). Additional research has focused on whether the trends in the data are pointing in the same direction as the model prognostications for a greenhouse-gas-warmed world. These research endeavors focus on an answer to the question: Can GCMs reproduce the past and current climate with sufficient accuracy that their predictions for future climate conditions are viable?

The most common approach to assess the extent to which observed trends in air temperature are reflected in GCM prognostications of anthropogenic warming has been to evaluate trends in observed and modeled air temperature fields both at and above the Earth's surface. Many of these "fingerprinting" efforts have involved comparing the spatial patterns of air temperature change between observations and GCM simulations in which the effective trace gas concentration of the atmosphere is increased from the pre-industrial baseline. The underlying premise of such analyses is that increased correspondence (over time) between the observed and modeled fields suggest that (1) the models are correctly simulating the current three-dimensional temperature structure of the atmosphere and (2) these changes are likely induced by human activities.

Often, these analyses use the "centred" and "uncentred pattern correlation coefficients" as the primary tools to ascertain the degree of correspondence and its change over time, between observations and the model simulations. In a recent article, Legates and Davis (1999) have demonstrated that such correlation-based approaches are inappropriate and may, in fact, have biased the results of any analyses dependent upon them. Our claims are based on five main problems associated with correlation-based measures: (1) overestimation of the spatial correlation, (2) evaluation of only the general pattern similarity and not the absolute differences between the fields, (3) variation of correlation with choice of a reference period when missing observations are included, (4) over-sensitivity of the statistic to outliers relative to observations near the mean, and (5) standardization of differencing. Legates (Legates and McCabe, 1999) recently examined several more appropriate statistics and provided guidelines for the appropriate use of such model-observation intercomparison statistics. We utilize these guidelines to determine the applicability of such methods in climate change fingerprint detection.

Using these more appropriate, non correlation-based statistics, we re-examine the results of surface air temperature analyses using data from the Jones et al. (1994) air temperature data archive and simulations from the CGCM1 (Canadian Global Coupled Model -- Reader and Boer, 1998) to determine the efficacy of the statistics. The CGCM1 runs were selected because the authors have some familiarity with this model (Legates, 1999) and because it has been chosen for the US National Assessment of Climate Change. However, our paper emphasizes the development and evaluation of a suite of non-correlation-based statistics for potential use in fingerprint detection studies.