🚨 New paper in @NatureComms We show that relative temperature extremes (TX90p) as used in many studies can be biased by as much as 50%! The bias arises from the use of too long seasonal windows and can easily be corrected. Details below 1/🧵 doi.org/10.1038/s414... 🧪
TX90p is defined relative to the TX 90th percentile at each location & across the seasonal cycle. (Implicit) assumption: ~10% of days are extreme (in sample) regardless of region and season. We show that this assumption often does not hold and discuss implications. 2/🧵
Running seasonal windows are often used to increase the sample size for the percentile calculation. Many studies use 15- or even 31-day windows, inadvertently mixing seasonal gradients into the threshold. As a result, the extreme frequency is systematically biased low! 3/🧵
Since the seasonal cycle varies across the globe, so does the bias. In some regions, the annual extreme frequency is close to the expected 10%, and in others only 5% (-50% bias). This undermines a fair comparison of extremes and derived heatwave metrics between regions. 4/🧵
The bias weakens with global warming (see paper for details rdcu.be/dBByU) leading to an overestimation of temperature extreme changes. 10% expected baseline frequency -> max. increase x10 -> 100% extremes in the future. In the biased case 👇 the ratio can exceed x13! 5/🧵
Derived metrics such as heatwaves are also affected. We show that the bias also impacts changes in land heatwaves during the warm season as a more impact-focused metric 👇👇 6/🧵
How to eliminate the bias? Only using short windows also comes with problems (see paper for a discussion). We propose to remove the mean seasonal cycle before the percentile calculation. Result: the bias is all but gone 💪 7/🧵
Conclusion: don’t use long seasonal windows in the calculation of extreme threshold without correction! The bias might not obviously show in derived metrics but could still lead to 👉 Pitfalls in diagnosing temperature extremes 👈 8/🧵 doi.org/10.1038/s414...
What about problems in already published results? We focus on the methodological error and its *potential* implications - not on errors in individual studies. For many studies, the impact might be small as other factors dominate but we encourage people to check! 9/🧵
Many thanks to my co-author Aiko Voigt (@aikovoigt.bsky.social), and to Erich Fischer (@erichfischer.bsky.social) and Gabi Hegerl for their input! Thanks also for the helpful comments from three anonymous reviewers and the editor Efi Rousi (@efou.bsky.social) ! 🙏🙏🙏 10/🧵
We have made the code used in the study free to use: github.com/lukasbrunner... Example data are here: doi.org/10.5281/zeno... #openaccess #openscience 🧵/END