A consistent statistical model selection for abrupt glacial climate changes

• Published in: Climate dynamics 2019-06, Vol.52 (11), p.6411-6426
• Main Authors: Lohmann, Johannes, Ditlevsen, Peter D.
• Format: Article
• Language: English
• Subjects: Analysis; Autocorrelation; Autocorrelation function; Autocorrelation functions; Bayesian analysis; Climate; Climate change; Climate cycles; Climate models; Climate system; Climate variability; Climatic changes; Climatology; Computation; Computer simulation; Data processing; Earth and Environmental Science; Earth Sciences; Forecasts and trends; Geophysics/Geodesy; Glacial climates; Glacial periods; Ice cores; Mathematical models; Oceanography; Oscillations; Probability theory; Qualitative analysis; Residential density; State of the art; Statistical analysis; Statistical inference; Statistical methods; Statistical models; Statistics; Stochasticity
• ISSN: 0930-7575; 1432-0894
• DOI: 10.1007/s00382-018-4519-2

The most pronounced mode of climate variability during the last glacial period are the so-called Dansgaard–Oeschger events. There is no consensus of the underlying dynamical mechanism of these abrupt climate changes and they are elusive in most simulations of state-of-the-art coupled climate models. There has been significant debate over whether the climate system is exhibiting self-sustained oscillations with vastly varying periods across these events, or rather noise-induced jumps in between two quasi-stable regimes. In previous studies, statistical model comparison has been employed to the NGRIP ice core record from Greenland in order to compare different classes of stochastic dynamical systems, representing different dynamical paradigms. Such model comparison studies typically rely on accurately reproducing the observed records. We aim to avoid this due to the large amount of stochasticity and uncertainty both on long and short time scales in the record. Instead, we focus on the most important qualitative features of the data, as captured by summary statistics. These are computed from the distributions of waiting times in between events and residence times in warm and cold regimes, as well as the stationary density and the autocorrelation function. We perform Bayesian inference and model comparison experiments based solely on these summary statistics via Approximate Bayesian Computation. This yields an alternative approach to existing studies that helps to reconcile and synthesize insights from Bayesian model comparison and qualitative statistical analysis.