Assessing Uncertainty in Climate Index Computations: Sensitivity Analyses and Methodological Considerations

Applied research efforts employing climate indices as part of climate change impacts studies can involve a number of “Large Dataset” sources. Some data sources represent past conditions, such as station observations, gridded observation-based datasets, and reanalysis products. Other model-based data sources, including global climate models and earth system models, dynamically downscaled regional models, and statistically downscaled / bias corrected projections, represent both simulated past conditions and multi-decadal future projections. Experimental design considerations call for a review of candidate datasets to determine if they contain all the required variables at the temporal resolution necessary to compute the desired climate indices. All else being equal, the probability of encountering data suitability challenges increases as the number of variables used to compute an index increases and the temporal resolution becomes finer. If direct computation of a climate index is not possible, one may opt to explore ways to estimate the index from the available data, either via an established formula or developing and validating a new approximation method.
Ideally, one would like to perform apples-to-apples comparisons of climate indices across different data sources. Such comparisons would include quantification of uncertainties associated with the individual data sources’ input variables and assessing how those uncertainties propagate through the computation of climate indices. However, for various reasons, comprehensive and quantitative uncertainty analyses are often not feasible. Not wanting to allow “perfect be the enemy of good,” applied research studies can proceed and include in their discussions qualitative caveats noting the presence of unquantified uncertainties. Yet, there is a risk that some stakeholders and clients may infer that an unquantified uncertainty can be assumed to be negligible, when in fact it may not be inconsequential for their decision-making purposes. When comprehensive uncertainty assessments can not be conducted, sensitivity analyses sometimes can provide information about the importance of otherwise unquantified uncertainty sources relative to some known and generally appreciated uncertainties. Here we present some examples of sensitivity analyses that aim to partially fill the gap between comprehensive uncertainty analyses and verbal caveats. The sensitivity to approximating daily maximum heat index values from daily measures rather than higher frequency data is examined using observational and model data sets. Also, for other climate indices, we explore the sensitivity of future statistically downscaled projections to the choice of gridded observations used for bias correction purposes and to other methodological choices one might make in a climate impacts study that makes use of multiple data sources. There can be value in comparing the magnitude of those sensitivities to the more generally appreciated sources of climate projection uncertainties popularized in an oft cited Bulletin of the AMS paper by Hawkins and Sutton (2009) (i.e., internal [unforced] variability, climate model [response] uncertainty, and scenario uncertainty).
Reference: Hawkins, E., and R. Sutton, 2009: The Potential to Narrow Uncertainty in Regional Climate Predictions. Bulletin of the American Meteorological Society.