J7.2
Tracking Climate Change: Ed Epstein's Influence and Mentorship (Invited Presentation)
![](images/flag.png)
![](images/blue_ribbon.png)
My first direct involvement with this problem was in 1998 near the end of my long tenure at CAC/CPC and was entirely practically motivated; I wanted to account separately for the effects of climate change and ENSO in formulating seasonal forecasts. CPC was implicitly and partially doing this with a 10-year retrospective average referred to as the Optimal Climate Normal (OCN), but the method had several important shortcomings. I took an entirely different approach, the hinge model, whose seed, I realized in hindsight, was sown by Ed Epstein's 1982 paper “Detecting Climate Change.” Furthermore, my belief in 1998 that climate change was not only detectable but large at regional scales over the United States with clear seasonality, is unquestionably partially a consequence of the clarity and credibility of the message in Ed's 1982 paper. In early 1998 the community consensus supported global-scale warming, but only a minority acknowledged substantive impacts already at local and regional scales.
In the 1982 paper, Ed recognized the possibility that global warming may have begun after 1975. He then proposed simple statistical models for climate change, including one that anticipated the hinge in several respects, and used these to develop likelihood ratios for the purpose of determining how long it would take to be confident from the record that it was indeed taking place. He concluded that 10 years would be sufficient in all cases. Coincidently, this would roughly correspond to Ed's tenure at CAC where he was my colleague and mentor.
The 1982 paper made a substantial impression on me, but during Ed's time at CAC he helped set the stage for my current work in two other ways. First, almost on his arrival at CAC, Ed and I were asked by Tom Karl to help him quantify how unusual the set of U.S. winters from 1975-76 through 1982-83 were. This was the period that many (including myself) now believe was the onset of modern warming, but also began with the coldest set of U.S. winters in the 20th century. We concluded that the return period for such an unusual set of winters was over 1000 years (Karl et al., 1984). This work further enhanced my awareness of the global transitions that took place in the mid-1970s. Second, while at CAC Ed taught me how important it was for CAC operations to accurately and smoothly estimate climate normals, through discussions and the work he produced to do so (five papers that Huug van den Dool addresses in his symposium paper).
The most viable alternative approaches for tracking temperature normals for the purpose of nine-month lead (next year) seasonal forecasts based on independent data tests (Wilks, 2013; Wilks and Livezey, 2013) are a 15-year retrospective mean and the so-called hinge fit. The latter is a piecewise-linear and –continuous model of climate change where the normal does not change (is stationary) from 1940 to 1975, but then changes linearly thereafter. The model's form, including a 1975 change point, obviously harks back to both Epstein (1982; note the similarity to Case 2 in Fig. 2) and Karl, Livezey, and Epstein (1984). Its smoothness conforms to normals developed in Ed's other climate papers. The hinge model has been criticized because of the arbitrary (but well-motivated) choices made in specifying the pre-change point slope and the change point year. Wilks (2013) and Wilks and Livezey (2013) both demonstrate that these choices are close to optimal.
Relative forecast performance depended on the temperature data set. The 15-year OCN overwhelmingly performed best in terms of squared error on inhomogeneous mega-climate division series (areally-averaged records contaminated by some non-climatic discontinuities and trends). The performance of the hinge relative to the OCN improved somewhat on station data but considerably on homogeneous station data. For the latter case, each method performed best in 5 out of 12 regions/seasons examined, suggesting a hybrid approach. Overall for all data sets, the hinge had smaller bias (generally cold for the OCN).
For detrending and signal separation, for example separating ENSO and climate change signals, a running 15-year normal is inadequate for several reasons and the hinge fit or a technique with similar attributes has to be used. The OCN not only does not remove a lot of “natural variability” but also requires false data at the beginning and end of records to apply to the full record. Thus the hinge fit should be used to develop appropriate estimates of probabilities and conditional (say to ENSO) probabilities.