A statistical approach to the phasing of atmospheric reorganizationand sea ice retreat at the onset of Dansgaard-Oeschger events underrigorous treatment of uncertainties

Keno Riechers, Niklas Boers2020 · DOI

TL;DR

This paper introduces a rigorous statistical framework to analyze the timing of abrupt climate shifts (Dansgaard-Oeschger events) using paleoclimate proxy data. By carefully accounting for uncertainties in data, the study finds that the previously observed tendency for atmospheric changes to precede sea ice retreat is likely a stochastic feature, not a causal relationship, suggesting a more complex trigger mechanism for these events.

What Problem Does This Paper Address?

Dansgaard-Oeschger (DO) events are abrupt climate shifts observed in paleoclimate records, characterized by rapid warming, atmospheric reorganization, and sea ice retreat. While previous studies using proxy data like ice core concentrations of Calcium (Ca²⁺, linked to atmospheric circulation) and Sodium (Na⁺, linked to sea ice extent) suggested a temporal lag – with atmospheric changes preceding sea ice changes – the physical mechanism and causality remain debated. A key challenge is the inherent uncertainty in precisely dating these abrupt transitions in noisy proxy records. This paper aims to rigorously assess the relative timing of Ca²⁺ and Na⁺ transitions at the onset of DO events, specifically addressing whether the observed tendency for Ca²⁺ to lead Na⁺ is a robust indicator of causality or a statistical artifact due to data uncertainties.

Methods

The authors employ a sophisticated statistical approach to handle the uncertainties in paleoclimate proxy data. Key methods include:

  • Bayesian MCMC Ramp-Fitting: This algorithm is used to detect transition onsets in noisy time series, providing not just a single date but a probability distribution for the transition time. This is crucial for quantifying uncertainty.
  • Probabilistic Transition Onset Detection: The algorithm returns an "uncertain sample" of transition onset lags, where each lag is itself a random variable reflecting measurement uncertainty.
  • Rigorous Uncertainty Propagation: A core contribution is the development of a framework to propagate these uncertainties through statistical tests. This involves deriving probability distributions for test statistics and p-values, rather than relying on point estimates.
  • Hypothesis Testing on Uncertain Samples: The paper adapts standard hypothesis tests (t-test, Wilcoxon-signed-rank test, bootstrap test) to work with these uncertain samples, focusing on testing the null hypothesis of a simultaneous transition (zero lag) or a Na⁺ lead.
  • Population Mean Inference: The study infers the properties of an underlying "population" of DO event lags from the limited sample, accounting for both measurement uncertainty and the inherent randomness of the events.

Datasets

The study utilizes high-resolution Ca²⁺ and Na⁺ concentration time series from the NGRIP ice core, covering the period from 60 to 10 kyr BP. These data were previously published by Erhardt et al. (2018) and cover 24 DO event onsets. The time resolution varies from 2 to 4 years.

Key Insights and Ideas

  • Uncertainty is Paramount: The paper strongly emphasizes that ignoring or inadequately handling uncertainties in transition onset detection can lead to misleading conclusions. Rigorous propagation of these uncertainties fundamentally changes the interpretation of the data.
  • Stochasticity vs. Causality: The observed tendency for Ca²⁺ to lead Na⁺ transitions, previously interpreted as a potential causal link, is shown to be statistically indistinguishable from a purely stochastic process when uncertainties are properly accounted for.
  • No Exclusive Trigger: The findings contradict the idea that either atmospheric reorganization (Ca²⁺) or sea ice retreat (Na⁺) exclusively triggers the other. The presence of events where Na⁺ leads Ca²⁺, and the overall lack of a guaranteed systematic lead/lag, suggests a more complex, possibly external, forcing mechanism for DO events.
  • A Two-Level Randomness: The study highlights a "two-level randomness": the inherent variability of the climate system (population distribution) and the uncertainty in measuring these events from proxy data (individual sample uncertainty).

Key Results

  • Non-Significant Results After Uncertainty Propagation: While hypothesis tests on expected sample values suggested a significant Ca²⁺ lead, rigorous propagation of uncertainties to p-values resulted in non-significant findings for all three tests (t-test, WSR, bootstrap) at a 0.05 significance level.
  • Probability of Ca²⁺ Lead: The probability that the population mean lag is greater than zero (indicating a Na⁺ lead on average) is estimated at 22%. The probability of the sample mean exceeding zero is 13%.
  • Tendency, Not Guarantee: The data still shows a tendency for Ca²⁺ to lead Na⁺, with approximately 13% of events exhibiting a Na⁺ lead on average. However, this tendency is not guaranteed, and the probability of at least 10 out of 16 events showing a Ca²⁺ lead is around 50%.
  • Control Group Validation: Analysis using a control group with alternative realizations of the data confirmed the robustness of the statistical framework and the results.

Limitations and Future Work

  • Proxy Interpretation: The authors acknowledge that the interpretation of Na⁺ as a sole sea ice proxy is debated, and new results are anticipated. Further refinement of proxy interpretation is needed.
  • Model Studies: The paper calls for sustained efforts in conducting model studies, particularly those that can be directly compared with proxy records (proxy-enabled modeling), to improve understanding of the temporal order of events.
  • Other Proxy Variables: The authors suggest that similar statistical analyses should be applied to other proxy variables to gain a more comprehensive understanding of the phasing of DO events.
  • Systematic Pattern of Na⁺ Leads: While the study identifies events with a Na⁺ lead, it does not reveal a systematic pattern for these events, suggesting that further investigation is warranted.

Summary updated

Abstract

For previous glacial intervals, concomitant shifts in different proxy records from Greenland ice cores indicate repeated abrupt climate transitions comprising – among others – abrupt warming, a sudden reorganization of the atmospheric circulation, and a retreat of perannial sea ice. The physical mechanism underlying these so-called Dansgaard-Oeschger (DO) events remains debated. Former studies have made an effort to deduce the progression of temperature, circulation, and sea-ice changes at the onset of DO events from paleoclimate proxy records to constrain potential triggering mechanisms. In this context, recent research reports on systematically delayed transitions in Na+ concentrations and δ18O values compared to Ca2+ concentrations and the annual layer thickness by about one decade. This is interpreted as a temporal lag of sea ice retreat and Greenland warming with respect to atmospheric reorganization at the onset of DO-events. Here, we present a comprehensive statistical analysis of the relative phasing of DO transitions in Ca2+ and Na+ concentration records from the NGRIP ice core for the period 60–10 kyr BP. Regarding the time lags identified in this period as a sample generated from an unknown population, we derive probability density functions for the sample and population mean and test the null-hypothesis of a simultaneous transition. Special attention was paid to the uncertainties inherent to the transition onset detection in noisy data. Their rigorous propagation changes the test results from significant to non-significant and therefore a purely stochastic origin of the observed tendency for Ca2+ to lead the transition cannot be ruled out. In fact, we show that the data is very likely to comprise both: DO events that were led by a Ca2+ transition, as well as events led by a Na+ transition. Together, these findings clearly contradict a systematic lead or lag between the DO transitions in the two proxies, and the apparent Ca2+ lead should therefore not be interpreted as indication of a causal relationship. Under the assumption that all DO events followed the same physical mechanism and that the proxy interpretation holds true, the we conclude that at DO transition onsets, neither was the atmospheric reorganization caused by sea ice retreat, nor was the sea ice retreat triggered by atmospheric reorganization.
BibTeX
@misc{riechers2020statistical,
  author = {Keno Riechers and Niklas Boers},
  title = {{A statistical approach to the phasing of atmospheric reorganizationand sea ice retreat at the onset of Dansgaard-Oeschger events underrigorous treatment of uncertainties}},
  year = {2020},
  doi = {10.5194/cp-2020-136},
}
Figure