This paper, authored by Dukpa Kim, Tatsushi Oka, Francisco Estrada, and Pierre Perron, provides a rigorous statistical framework to settle long-standing debates in climate science regarding the drivers of global warming and the reality of the mid-90s "warming hiatus."
TL;DR
The authors develop new multivariate statistical tests to identify "common breaks" in time-series data, proving that major shifts in global temperature trends—specifically the 1960s warming surge and the 1990s slowdown—align perfectly with shifts in anthropogenic radiative forcing. This synchronization provides powerful evidence that human activity, rather than natural variability, is the primary driver of these trend changes.
What Problem Does This Paper Address?
Climate scientists have long observed that global temperatures don't follow a straight line; they move in "segments" with different growth rates. Two periods are particularly controversial:
- The 1960s "Onset": A sharp increase in the rate of warming.
- The 1990s "Hiatus": A period where warming appeared to slow down, leading some to argue that global warming had "paused" or was merely the result of natural internal noise.
The challenge is proving causality. In time-series analysis, two variables can look related simply by chance (spurious correlation). However, if you can prove that the exact dates when the trend changes (the "breaks") are identical across both temperature and human-induced radiative forcing, the case for a causal link becomes much stronger. This paper addresses the lack of robust multivariate tests capable of handling "joined segmented trends"—where the trend lines are continuous but change slope at specific, unknown points.
Methods
The authors treat the climate as a multivariate system of equations where temperatures and radiative forcing are dependent variables.
- Joined Segmented Trends: Unlike standard structural break models that might allow for "jumps," this model assumes the trend is continuous (joined) at the break date, which is more physically realistic for climate data.
- New Statistical Tests: They introduce three primary tests to detect common breaks across different series:
- Quasi-Likelihood Ratio (LR) Test: A standard approach, but one that is "non-pivotal" (its distribution is messy and depends on nuisance parameters).
- GLS-Wald and OLS-Wald Tests: Corrected versions of the Wald test. The OLS-Wald is computationally efficient, while the GLS-Wald accounts for correlation between different equations (e.g., how Northern and Southern hemisphere temperatures might fluctuate together).
- Bootstrap Procedure: Because the asymptotic distributions of these tests are complex, the authors use a bootstrap method to generate reliable p-values, ensuring the tests don't "hallucinate" breaks due to serial correlation in the noise.
- Filtering: They use the Atlantic Multidecadal Oscillation (AMO) and North Atlantic Oscillation (NAO) to "purge" the temperature data of low-frequency natural variability, allowing the underlying secular trend to emerge.
Datasets
The study utilizes several high-profile temperature and forcing datasets:
- Temperature Series: HadCRUT4 (1850-2014), GISS-NASA (1880-2014), Berkeley Earth, and the Karl et al. (2015) dataset. These include global, Northern, and Southern hemispheric data.
- Radiative Forcing: Data from GISS-NASA (1880-2010), specifically focusing on Well-Mixed Greenhouse Gases (WMGHG) and Total Radiative Forcing (TRF).
- Modes of Variability: AMO and NAO indices used for filtering natural noise from the temperature series.
Key Insights and Ideas
- Commonality as Evidence of Causality: The "aha moment" is that common break dates across independent physical processes (forcing vs. temperature) mitigate the risk of spurious correlation. If the "thermostat" (forcing) and the "room" (temperature) change at the same time, the thermostat is likely the cause.
- The Importance of Filtering: The authors show that once you account for natural oscillations like the AMO, the "breaks" in temperature become much clearer and align much better with radiative forcing.
- Bivariate Power: Testing temperature and forcing together (bivariate) provides significantly more statistical power than testing temperature alone. This allows the authors to detect the 1990s slowdown even in relatively short or noisy datasets.
Key Results
- The 1960s Break: The tests confirm a common break in the early 1960s (roughly 1963–1966), marking the "onset of sustained global warming" driven by a surge in greenhouse gas forcing.
- The 1990s "Hiatus": The paper formally rejects the idea that the post-1990s slowdown was just natural variability. They find a common break around 1990–1992 where the rate of growth in both forcing and temperature significantly decreased.
- Validation: For the filtered series, the null hypothesis of "no common break" was rejected across almost all pairs of temperature and forcing data, providing a robust, multi-source confirmation of these trend shifts.
Limitations and Future Work
- Computational Intensity: The GLS-Wald and LR tests are computationally demanding, especially as the number of breaks ($m$) increases, following an $O(T^m)$ complexity.
- Serial Correlation Sensitivity: While the bootstrap helps, the tests can still suffer from size distortions if the serial correlation in the noise is extremely high.
- General Applicability: The authors note that while they applied this to climate science, these multivariate joined-trend tests are of "independent interest" and could be applied to economics or other fields where structural shifts are suspected to be shared across multiple variables.
Summary updated