Human-driven climate change has caused a wide range of extreme weather events to become more frequent in recent decades. Although increased and intense periods of extreme weather are expected consequences of anthropogenic climate warming, it remains challenging to rapidly and continuously assess the degree to which human activity alters the probability of specific events. This study introduces a new framework to enable the production and communication of global real-time estimates of how human-driven climate change has changed the likelihood of daily weather events. The framework's multi-method approach implements one model-based and two observation-based methods to provide ensemble attribution estimates with accompanying confidence levels. The framework is designed to be computationally lightweight to allow attributable probability changes to be rapidly calculated using forecasts or the latest observations. The framework is particularly suited for highlighting ordinary weather events that have been altered by human-caused climate change. An example application using daily maximum temperature in Phoenix, AZ, USA, highlights the framework's effectiveness in estimating the attributable human influence on observed daily temperatures (and deriving associated confidence levels). Global analyses show that the framework is capable of producing worldwide complementary observational- and model-based assessments of how human-caused climate change changes the likelihood of daily maximum temperatures. For instance, over 56 % of the Earth's total land area, all three framework methods agree that maximum temperatures greater than the preindustrial 99th percentile have become at least twice as likely in today's human-influenced climate. Additionally, over 52 % of land in the tropics, human-caused climate change is responsible for at least five-fold increases in the likelihood of preindustrial 99th percentile maximum temperatures. By systematically applying this framework to near-term forecasts or daily observations, local attribution analyses can be provided in real time worldwide. These new analyses create opportunities to enhance communication and provide input and/or context for policy, adaptation, human health, and other ecosystem/human system impact studies.
Many weather and climate events that were rare just a century ago have become more frequent in recent decades
Extreme event attribution is a relatively new and rapidly growing field within climate science
The concept of “rapid attribution”
We have developed a global framework to quantify whether and how much human-caused climate change has changed the likelihood of daily local weather events from the preindustrial climate to today. Our goal is to enable daily attribution assessments that support and frame climate change communication for a broad range of users and audiences from the very start of an event. The approach is designed to be (1) rigorously based on existing principles and methods in attribution science (as described primarily in
The framework is intended to complement existing attribution approaches (i.e., WWA rapid attribution). Multiple methods are employed to perform attribution calculations. The framework also quantifies uncertainty for each method, either directly through resampling or indirectly by taking advantage of the method's underlying data structure (e.g., intermodel uncertainties characterized by an ensemble). Synthesis across results from multiple methods then informs a final attribution assessment.
This study is meant as an illustrative introduction to this new real-time-capable attribution system. Section
We construct a methodological framework to make comprehensive worldwide assessments of the role of human-driven climate change in local daily weather events, at predefined spatial scales and for predefined variables, and draw inspiration and guidance from rapid and traditional attribution studies. This section provides a high-level overview of the attribution framework; a glossary of key study terms is provided in the Supplement (Table S3).
The approach streamlines existing, mature techniques described in the National Academies of Sciences (NAS) report on the “Attribution of Extreme Weather Events in the Context of Climate Change” Note that we only compute attribution estimates for events that are exhibiting a statistical response to global mean temperature changes; physical (i.e., meteorological and climatological) conditions during an event are not explicitly considered in our methods (Sect.
We quantify attribution estimates by contrasting event likelihoods from an observed or modeled “forced” distribution (which has been influenced by human activity since the late 19th century) of a single Earth system state variable of interest (temperature, soil moisture, humidity, etc.) to a defined “counterfactual” distribution of that variable
Diagram of this study's multi-method approach to quantifying local climate attribution estimates. The two observation-based methods begin by (1) relating the local temperature (orange) to GMST (gray) to obtain
Each method offers different lines of evidence for the extent to which changes in a state variable are attributable to anthropogenic climate change
The observation-based methods begin by characterizing how the variable of interest (in this study we consider the daily maximum temperature) changes as a function of GMST (Fig.
The state variable we use to illustrate the application of the new framework is the daily maximum temperature ( We note that state variables that are strongly affected regionally by dynamics, such as precipitation, are inappropriate for study with this method without further modification to account for non-linearity and dynamical variability
Observed
We take the monthly GMST time series from the Met Office Hadley Centre/Climatic Research Unit Temperature data set, version 5
Daily surface maximum temperatures and mean temperatures are drawn from the global climate model output of the Coupled Model Intercomparison Project phase 5
Before model simulations can be used in attribution analyses – especially when evaluating the probabilities of specific absolute temperatures – they must be bias adjusted. We apply a trend-preserving bias adjustment method developed by
The observation-based methods first determine the relationship between the state variable of interest and GMST (arrow 1 in Fig.
In the median-scaling method, a set of scale factors,
The quantile-scaling method mirrors the median-scaling method, except that monthly scale factors are calculated and distributions are scaled over a set of quantiles derived from the daily data. We find temperatures associated with each of 30 quantiles – chosen to be analogous to the number of days in an average month – roughly equally spaced between 0.01 and 0.99 (the full set of quantiles is in Sect. S1 in the Supplement), resulting in 30 annual quantile time series of maximum temperature,
In each method, as described in Appendix B, the resulting scale factors for a given month and location are used to translate a monthly climatological distribution according to the difference between the climatological mean GMST and a target GMST. The target GMST for the forced distribution is
Median- and quantile-scaling methods have different assumptions and tradeoffs. Median-scaling follows a traditional perspective of climate change's influence on temperature in that GMST warming will cause a linear shift in the state variable distribution while its shape remains fixed (
For our demonstration in this study, we use the median- and quantile-scaling methods to translate 31-year climatological distributions of observed daily
In each method, we implement uncertainty analyses to produce distributions of attribution estimates. This allows the attribution framework to provide not only median attribution estimates but also confidence intervals quantifying the robustness of attribution estimates and enabling inter-method statistical comparisons (Sect.
Trends and correlations in local At each month and location, find the median time series (as described above), Where For each A 3-year resampling window is chosen as a balance between retaining the year-over-year autocorrelation represented in the climate trend while still capturing a representative range of interannual variability. Compute the associated scale factor by regressing the resulting time series against annual GMST (Appendix B). Pool each of the
Resampled quantile-scaling distributions are found with the same sequence, replacing the median annual time series with each quantile annual time series,
The model-based method defines the forced and counterfactual distributions (arrow 4 in Fig.
Next, we find the calendar year when each model's historical plus projected 31-year centered-running-mean of GMST exceeds
Following a hazard-based framing, the attribution framework uses coupled changes in GMST and maximum temperature exceedance probabilities to determine the extent and confidence level of attributable human-influence on daily temperature events. At a given absolute temperature or quantile, the ratio of exceedance probabilities between the forced and counterfactual distributions provides an estimate of how that quantile or temperature has shifted because of human-caused climate change. We quantify this shift using the (exceedance) probability ratio
At each global location and for each of the forced and counterfactual distributions, these probabilities can be calculated by integrating daily exceedances of an absolute temperature threshold. We either prescribe this temperature directly or infer it from a prescribed quantile. Integration is performed using the counterfactual climatology of each distribution set (i.e., via median-scaling, quantile-scaling, and modeling methods) and can be done in the context of monthly, seasonal, and annual units of analysis. For each given threshold and context, we calculate PR over the 31 years of the paired forced/counterfactual distributions. Because we seek the climatological PR of any given year (rather than a specific year), our final attribution estimate for each method is given by the mean over the 31 individual-year PR values. We fully describe our calculations that quantify PR based on discrete exceedance counts in Appendix C.
PR uncertainties are determined by calculating PR values using the full distributions of either resampled scale factors and their paired forced/counterfactual distributions (observation-based methods) or each individual forced climate model distribution against the pooled counterfactual distribution (model-based method). To determine the statistical significance of attribution from the resulting PR distributions, we compare 95 % confidence intervals of each method against a null hypothesis that exceedance probabilities should be the same between the forced and counterfactual climates, i.e.,
Discrete exceedance count calculations (Appendix C) cannot appropriately quantify probability changes in the extreme tails of the 31-year forced and counterfactual temperature distributions. This is because there are too few temperature observations exceeding or subceeding (i.e., the antonym of “exceed”; see
Attribution of temperatures above the critical quantile could be performed using generalized extreme value distributions, as is often done in traditional attribution studies of extreme events
We use three examples to illustrate the temporal and spatial performance of our attribution framework. We first present an example from Phoenix, AZ, USA, that shows how our three methods work in practice over a month of daily data and how their results can be combined to make a final attribution assessment. We then extend our case study to include attribution estimates at various locations around the world on a single day. Finally, we consider the global spatial fingerprint of human-caused climate change by looking at worldwide probability ratios for each location's 99th percentile maximum temperature.
As an example application of our real-time attribution system, we use each method to assess the attribution of Phoenix
For each daily observation, we use each of the three framework methods to calculate probability ratios from July forced and counterfactual distributions (i.e., over the July monthly unit of analysis). Daily PR values illustrate how the probability of meeting or exceeding Phoenix's daily observed
We first analyze Phoenix PR values calculated from each method over each monthly unit of analysis (January through December) at the 95th percentile of each counterfactual distribution. Seasonal cycles of PR medians and 95 % confidence intervals (from the observation-based method PR distributions) and individual-model PRs are plotted in Fig.
Seasonal cycles of probability ratios at the 95th percentile in Phoenix, calculated with the observation-based median-scaling (purple) and quantile-scaling (orange) methods and the model-based method (black). Purple and orange dots and bars show the median and 95 % confidence intervals for the observation methods, respectively; each black dot shows the PR of an individual model from the CMIP5 ensemble. The black dashed line at
The model-based PR seasonal cycle exhibits a pattern consistent with other model attribution studies, showing an increasing seasonal temperature amplitude as the climate warms
We now explore our framework's attribution assessment of daily
Application of the attribution framework to July 2016 in Phoenix, Arizona, USA.
Empirical cumulative distribution functions (CDFs) from the observation-based (orange and purple curves) and model-based (black/gray curves) methods illustrate the attribution framework as it relates to Phoenix temperatures in July (Fig.
July 2016 probability ratios (Fig.
There is good agreement on the magnitude – and perfect agreement on the sign – of July Phoenix PR values from the framework's multiple methods. This leads to a consistent and statistically significant result of increasing frequency associated with every daily
A final cohesive assessment of these framework results depends on the particular desired application. For example, the percentage of days with significantly attributable
We further explore the framework results by examining the attribution estimates at multiple locations on a single day. Using Berkeley Earth gridded observations on 27 July 2016, we compare the Phoenix attribution calculations (Figs.
Median and 95 % confidence intervals (CI) of probability ratios (PRs) associated with
Table
In Asunción and Bengaluru, the three methods significantly disagree on the magnitude of attributable changes, with smaller model-based estimates than those from the observation-based methods. This lack of multi-method consensus (values marked with the superscript
PR values on 27 July 2016 show the wide variation of results that can arise from daily attribution calculations. Weather noise drives much of this local observed temperature variability, while the signal of anthropogenic warming acts to increase the baseline maximum temperatures and change the likelihoods of each daily temperature being observed. For instance, relatively high local temperatures in Bengaluru and Nairobi are associated with PR values ranging from 2.2 to 173, whereas relatively common local temperatures in Asunción, Cape Town, and Mildura are associated with attribution estimates that are either barely significant or insignificant, indicating little to no human influence on the likelihood of their maximum temperature on 27 July 2016. Warsaw and Phoenix attribution estimates are modest, indicating a clear and attributable human-driven increase in the probabilities of their warmer-than-average local temperatures on 27 July 2016 (Sect.
Though not exhaustive, these examples illustrate the framework's capacity to provide a broad range of location-specific estimates of daily climate attribution, given a grid of observed or forecast maximum temperatures. When combined with environmental conditions from global forecasting models, future framework applications will use this capability to support concurrent and immediate worldwide operational estimates of attributable daily weather events.
Globally resolved probability ratios (unitless) calculated with the observation-based
Moving on from these to location-specific examples to a complete global scale, we now demonstrate results from the attribution framework by mapping the probability ratios calculated at the annual 99th percentile of each counterfactual distribution. Probability ratios
We find a coherent pattern of much higher probability ratios in the tropics that is consistent with time-of-emergence studies. In tropical regions, the anthropogenic signal of climate change dominates over small-amplitude weather noise, allowing human-influenced temperature trends to be detected earlier than in mid- and high latitudes
This study has detailed the development of a joint observational- and model-based (i.e., multi-method) framework to generate real-time estimates of the role that human-caused climate change plays in producing local daily temperatures around the globe. The framework is designed to be flexible across data sources and climatological state variables (especially those tied to climate warming through thermodynamics, e.g., temperature, sea level, and soil moisture;
Our methods are informed by state-of-the-art attribution guidelines from National Academies of Sciences report
Some procedural concessions have been made in this implementation. The framework is strictly statistical and is not able to consider the physical environment (e.g., synoptic conditions) during an event. It is currently unable to compute attribution estimates for dynamically driven extremes
Several limitations warrant future investigation and improvement. The framework's observation-based attribution methods assume a historical linear-scaling relationship between the state variable and global mean surface temperature. Some research studies have shown that this relationship can be nonlinear, especially for high temperature extremes in the tropics
This study used a single state variable, the daily maximum temperature, to illustrate the attribution framework in action and a single set of observations and models – Berkeley Earth and CMIP5 – on a shared coarse grid. Because of this coarse grid, small-scale climate events in some regions may not be well represented. To limit biases arising from this scale discrepancy, extending this approach to a finer-scale grid or point-based observations may be required – particularly in operational contexts requiring location specificity. By design, the framework is agnostic of data sources and may be flexibly extended or adapted with different model ensembles
The multi-method approach herein could be applied to estimate attributable probability changes for other well-observed and well-modeled climatological state variables (e.g., precipitation), providing appropriate adaptations (e.g., variable-specific bias adjustment). Extending our framework to integrate other environmental state variables, particularly those that may exhibit weak or nonlinear relationships with global mean temperature changes, could require future work that differs from variable to variable.
Global climate attribution studies remain underdeveloped and underexplored. While projects like the World Weather Attribution initiative have provided prominent rapid assessments of the influence of human-induced climate change on daily weather characteristics, these studies are often ad hoc and geographically biased towards Europe or North America. A recent study
Our sample global attribution analyses show a consistent pattern of strongly attributable human influences on maximum temperatures across the tropics. These results highlight important links between global inequities, climate data, and attribution of extreme temperatures. Although the tropical time series from Berkeley Earth are sufficiently long –
Public perceptions of climate risk are strongly tied to the effects of extreme weather
Following
Detrend Transfer the simulated climate change signal for every distribution quantile from Given a single daily observation, For each Use parametric quantile mapping to adjust the distribution of values in Comparing over the calibration period, the estimated root mean square error between Restore the trend subtracted from
In this study, bias adjustment is applied to each model's GMST time series and monthly
Following
An observed climatological distribution of the state variable over an arbitrary time period (taken herein to be 1985–2015) is given by
Note that the scaling approach retains the full range of interannual variability present in the underlying climatology.
At a given quantile in the counterfactual distribution,
Our final estimate of the attributable probability ratio is given by the mean over every year's individual probability ratio values, i.e.,
Scripts and code to recreate our analyses are available for direct download from GitHub (
The supplement related to this article is available online at:
DMG, AP, BHS, and FELO developed the methodology. DMG and KH developed the software, validated the project, conducted the formal analysis, and led the investigation. DMG wrote the draft and was assisted by AP, BHS, KH, and FELO during the review and editing stages. DMG and AP visualized the project, while AP, BHS, and FELO conceptualized it. KH curated the data, BHS administered the project, and AP supervised.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors thank Claudia Tebaldi and two anonymous reviewers, for helpful comments which improved this work. We also thank Lukasz Tracewski and Dan Dodson, for maintaining the computing resources on which the climate attribution framework was developed and tested.
This research has been supported, in part, by The Schmidt Family Foundation/The Eric and Wendy Schmidt Fund for Strategic Innovation.
This paper was edited by Mark Risser and reviewed by two anonymous referees.