A quantile regression analysis of daily North American temperatures, 1979--2013

Monday, 15 December 2014
Martin Tingley1, Karen A McKinnon2, Andrew N Rhines2 and Peter J Huybers2, (1)Pennsylvania State University Main Campus, University Park, PA, United States, (2)Harvard University, Cambridge, MA, United States
The contribution of average warming to increases in high temperature extremes and decreases in low temperature extremes is well established, but the role of changes in other aspects of the temperature distribution is unclear. By inferring separate slopes for each quantile of the distribution, quantile regression allows trends in the center of the temperature distribution to be disentangled from trends in other aaspects of the distribution. In particular, as a change in the properties of one tail of the distribution -- such as the hottest 10% of temperatures becoming hotter-- does not affect the 50th percentile, quantile regression can be useful in diagnosing changes in temperature extremes that are not attributable to changes in the center.

Applying quantile regression to explore linear trends in daily summertime and wintertime maximum and minimum temperature observations from North American weather stations, we find spatially coherent patterns of change that vary across quantiles of the distribution, indicative of changes in second or higher moments of the distribution. We repeat the analysis on both the NCEP-DOE Reanalysis II and the ERA Interim Reanalysis, and spatially smooth and interpolate all results to a common grid to facilitate comparisons. The reanalysis products do not reproduce key patterns seen in the observations, and feature higher magnitude trends of both signs, particularly for the lower quantiles of maximum temperatures.