The observed frequency, intensity, and duration of some extreme weather events have been changing as the climate system has warmed. Such changes in extreme weather events also have been simulated in climate models, and some of the reasons for them are well understood. For example, warming is expected to increase the likelihood of extremely hot days and nights (Figure S.1). Warming also is expected to lead to more evaporation that may exacerbate droughts and increased atmospheric moisture that can increase the frequency of heavy rainfall and snowfall events.
The extent to which climate change influences an individual weather or climate event is more difficult to determine. It involves consideration of a host of possible natural and anthropogenic factors (e.g., large-scale circulation, internal modes of climate variability, anthropogenic climate change, aerosol effects) that combine to produce the specific conditions of an event. By definition, extreme events are rare, meaning that typically there are only a few examples of past events at any given location.
Nonetheless, this relatively new area of science—often called event attribution—is rapidly advancing. The advances have come about for two main reasons: one, the understanding of the climate and weather mechanisms that produce extreme events is improving, and two, rapid progress is being made in the methods that are used for event attribution. This emerging area of science also has drawn the interest of the public because of the frequently devastating impacts of the events that are studied. This is reflected in the strong media interest in the connection between climate change and extreme events, and it occurs in part because of the potential value of attribution for informing choices about assessing and managing risk and in guiding climate adaptation strategies. For example, in the wake of a devastating event, communities may need to make a decision about whether to rebuild or to relocate. Such a decision could hinge on whether the occurrence of an event is expected to become more likely or severe in the future—and, if so, by how much.
The ultimate challenge for the science of event attribution is to estimate how much climate change has affected an individual event’s magnitude1 or probability2 of occurrence. While some studies now attempt to do this, most consider classes of events that are similar to the event that has been observed. Irrespective of whether a specific
1 In this report “magnitude” and “intensity” are used synonymously.
2 In this report “probability” and “frequency” are used synonymously.
event or a class of events is studied, results remain subject to substantial uncertainty, with greater levels of uncertainty for events that are not directly temperature related. The conclusions drawn also depend, in general, on choices made when selecting the events, framing the questions asked about the role of climate change, designing the modeling setup, and selecting statistical tools to quantify uncertainty.
More and more event attribution studies are being published every year, and study results are increasingly requested very quickly after events occur. Some of the study methods are still relatively novel, however, and there are a range of views about how to conduct and interpret the analyses. This report examines the science of attribution of specific extreme weather events to human-caused climate change and natural variability3 by reviewing current understanding and capabilities. It assesses the robustness of the methods for different classes of events and attribution approaches, provides guidance for interpreting analyses, and identifies priority research needs (the full statement of task can be found in Appendix A). This study is sponsored by the David and Lucile Packard Foundation, the Heising-Simons Foundation, the Litterman Family Foundation, the National Aeronautics and Space Administration (NASA), the National Oceanic and
3 In this report, the term “natural variability” encompasses both externally forced variations other than anthropogenic as well as the chaotic component of the atmosphere that is not externally forced. See Glossary.
Atmospheric Administration (NOAA), and the U.S. Department of Energy (DOE), with additional support from the Arthur L. Day Fund of the National Academy of Sciences.
Event attribution approaches can be generally divided into two classes: (1) those that rely on the observational record to determine the change in probability or magnitude of events, and (2) those that use model simulations to compare the manifestation of an event in a world with human-caused climate change to that in a world without. Most studies use both observations and models to some extent—for example, modeling studies will use observations to evaluate whether models reproduce the event of interest and whether the mechanisms involved correspond to observed mechanisms, and observational studies may rely on models for attribution of the observed changes.
Some types of observation-based approaches to event attribution use the historical context in order to determine changes in the rarity of an observed event based on long-term data. For example, this might involve comparing the statistical probability of an event in today’s climate to its probability in some previous time several decades earlier when the concentration of anthropogenic greenhouse gases (GHGs) was much lower. In practice, historical observations are often not available for a long enough period to enable a reliable statistical evaluation of whether there has been a significant change in event frequency or intensity.
Another observational approach is based on analyzing the characteristics of a given weather event (e.g., the large-scale circulation pattern) and looking for historical analogues in order to determine how meteorologically similar events have changed. These studies might compare the amount of rainfall in the current event to similar past events to estimate how the long-term increases in atmospheric temperature and moisture affected the event. As such, this approach does not address how climate change may have influenced the conditions that gave rise to a particular weather pattern. Some studies have also diagnosed the frequency of circulation states in order to determine if these may explain or counteract any change in extreme events. In general, it will be challenging to attribute any such changes to anthropogenic climate change.
Weather and climate model-based approaches to extreme event attribution compare model-simulated weather and climate phenomena under different input conditions: for instance, with and without human-caused changes in GHGs. Many studies rely on coupled atmosphere-ocean climate models, while others may use global atmospheric models, regional models, or models that are constructed specifically to represent a particular class of weather events, such as hurricanes. Multiple simulations can be
conducted to test how changes in sea-surface temperature (SST), the levels of atmospheric CO2 or aerosols, or other variables affect the extreme event of interest. Simulations are often repeated many times with small changes in the initial atmospheric or other conditions to estimate some uncertainties and sensitivities. Figures S.2 and S.3 provide examples of model-based attribution for the extreme heat events in Russia during the summer of 2010 and the extreme flooding events in England and Wales during the autumn of 2000, respectively.
Many studies have used climate models to understand just how unusual observed conditions are with respect to the distribution of possible conditions in a world that is unperturbed by humans. Models are often used to estimate the probability of occurrence of an event with human-caused climate changes (p1) and without these changes (p0). These estimated probabilities are often used to estimate the fraction of attributable risk (FAR)—FAR = (p1 – p0)/p1—or the risk ratio (RR)—RR = p1/p0. These model-based estimates of attributable risk or RR hinge on the model used being able to reliably simulate both the event in question and any changes in this event that may occur due to human-caused climate change or another considered factor.
Some recent studies also have used models to attempt to follow the evolution of a particular extreme weather event—for example, through the use of a set of short-term forecasts using a weather model. This allows detailed study of particular extreme events with a model capable of representing those specific events with fidelity and quantification of the effect of certain aspects of climate change (e.g., increased moisture-holding capacity of a warmer atmosphere) in which there is high confidence. Such studies cannot fully address frequency of occurrence because the results are highly conditional both on the initial state of the atmosphere and land surface that is specified to the model and on the specific sea-surface conditions that prevailed at the time of the event. With these constraints, it may be possible to estimate changes in event magnitude or changes in the frequency of exceedance above or below a given event magnitude, conditional on all else that is required to be specified to make the short-term forecasts. It is not possible, however, to study whether the likelihood of the occurrence of similar initial states and sea-surface conditions has changed.
Event attribution is more reliable when based on sound physical principles, consistent evidence from observations, and numerical models that can replicate the event. The ability to attribute the causes of some extreme event types has advanced
rapidly since the emergence of event attribution science a little more than a decade ago, while attribution of other event types remains challenging. In general, confidence in attribution results is strongest for extreme event types that
- have a long-term historical record of observations to place the event in an appropriate historical context;
- are simulated adequately in climate models4; and
- are either purely meteorological in nature (i.e., the event is not strongly influenced by built infrastructure, resource management actions, etc.) or occur in circumstances where these confounding factors can be carefully and reliably considered.
4 By “adequately” the committee means that, at a minimum, climate models used for event attribution need to accurately capture the spatial patterns and variability of relevant climate-related phenomena. See Table S.1 and Box 4.1 for the committee’s assessment of the capabilities of climate models to simulate each event type.
Non-meteorological factors can limit the accuracy of model simulations of extreme events and confound observational records. Drought and wildfire are examples of events for which non-meteorological factors can be especially challenging in attribution studies.
Furthermore, confidence in attribution results that indicate an influence from anthropogenic climate change is strongest when there is an understood and robustly simulated physical mechanism that relates a given class of extreme events to long-term anthropogenic climate changes such as global-scale temperature increase or increases in water content of a warmer atmosphere.
More frequent occurrences of extreme heat and less frequent occurrences of extreme cold are examples of changes that are consistent with increasing global mean temperatures.
Using this set of criteria (i.e., sound physical principles, consistent evidence from observations, and numerical models that can replicate the event) the committee assessed their confidence in event attribution capabilities for different extreme event types, as illustrated in Figure S.4 and Table S.1.
Confidence in attribution findings of anthropogenic influence is greatest for those extreme events that are related to an aspect of temperature, such as the observed long-term warming of the regional or global climate, where there is little doubt that human activities have caused an observed change. For extreme heat and cold events in particular, changes in long-term mean conditions provide a basis for expecting that there also should be related changes in extreme conditions. Heavy rainfall is influenced by a moister atmosphere, which is a relatively direct consequence of human-induced warming, though not as direct as the increase in temperature itself. The frequencies and intensities of tropical cyclones and severe convective storms are related to large-scale climate parameters whose relationships to climate are understood to varying degrees but, in general, are more complex and less direct than are changes in either temperature or water vapor alone. Nevertheless, atmospheric circulation and dynamics play some role in the development of an extreme event, which is different for different event types. Changes in atmospheric circulation and dynamics are generally less directly controlled by temperature, less robustly simulated by climate models, and less well understood.
Event attribution can be further complicated by the existence of other factors that contribute to the severity of impacts. For example, while many studies have linked an increase in wildfires to climate change, the risk of any individual fire depends on past forest management, natural climate variability, human activities in the forest,
TABLE S.1 This table, along with Figure S.4, provides an overall assessment of the state of event attribution science for different event types. In each category of extreme event, the committee has provided an estimate of confidence (high, medium, and low) in the capabilities of climate models to simulate an event class, the quality and length of the observational record from a climate perspective, and an understanding of the physical mechanisms that lead to changes in extremes as a result of climate change. The entries in the table, which are presented in approximate order of overall confidence as displayed in Figure S.4, are based on the available literature and are the product of committee deliberation and judgment. Additional supporting information for each category can be found in the text of Chapter 4, summarized in Box 4.1. The assessments of the capabilities of climate models apply to those models with spatial resolutions (100km or coarser) that are representative of the large majority of models participating in the Coupled Model Intercomparison Project Phase 5 (CMIP5). Individual global and regional models operating at higher resolutions may have better capabilities for some event types, but in these cases, confidence may still be limited due to an inability to assess model-related uncertainty. The assessments of the observational record apply only to those parts of the world for which data are available and are freely exchanged for research. Most long records rely on in situ observations, and these are not globally complete for any of the event types listed in this table, although coverage is generally reasonable for the more densely populated parts of North America and its adjacent ocean regions.
| = high
|Capabilities of Climate Models to Simulate Event Type||Quality/Length of the Observational Record||Understanding of Physical Mechanisms That Lead to Changes in Extremes as a Result of Climate Change|
|Extreme cold events|
|Extreme heat events|
|Extreme snow and ice storms|
|Severe convective storms|
and possibly other factors, in addition to any exacerbation by human-caused climate change.
Confidence in attribution analyses of specific extreme events is highest for extreme heat and cold events, followed by hydrological drought and heavy precipitation. There is little or no confidence in the attribution of severe convective storms and extratropical cyclones. Confidence in the attribution of specific events generally increases with an increased understanding of the effect of climate change on the event type. Gaps in understanding and limitations in the historical data lead to differences in confidence in attribution of specific events among different event types.
Attribution of events to anthropogenic climate change may be complicated by low-frequency natural variability, which influences the frequencies of extreme events on decadal to multidecadal timescales. The Pacific Decadal Oscillation and the Atlantic Multidecadal Oscillation are examples of such variability. Characterization of these influences is uncertain because the observed record is too short to do so reliably or to assess if climate models simulate these modes of variability correctly.
Given the relative newness of the event attribution field, standards have not yet been established for how to present results, which can make their interpretation difficult, particularly if conflicting evidence is available. Most event attribution studies are subject to substantial uncertainty. Results also hinge on how the event that is analyzed is defined, the specific questions that are posed, the assumptions made when analyzing the event, and the data, modeling, and statistical tools used to conduct the analysis. It is therefore essential to communicate the event definition, event attribution questions, assumptions, and choices clearly when reporting on the outcome of an event attribution study. The technical nature of this information makes it challenging to accurately communicate results, uncertainties, and limitations to the broader public.
There is no single best method or set of assumptions for event attribution, as these depend heavily on the framing of the question and the amount of time available to answer it. Time constraints may themselves affect framing and methodological choices by limiting analyses to approaches that can be undertaken quickly.
A definitive answer to the commonly asked question of whether climate change “caused” a particular event to occur cannot usually be provided in a deterministic sense because natural variability almost always plays a role. Many conditions must align to set up a particular event. Extreme events are generally influenced by
a specific weather situation, and all events occur in a climate system that has been changed by human influences. Event attribution studies generally estimate how the intensity or frequency of an event or class of events has been altered by climate change (or by another factor, such as low-frequency natural variability). Thus, examples of questions that the scientific community can attempt to address include:
- “Are events of this severity becoming more or less likely because of climate change?”
- “To what extent was the storm intensified or weakened, or its precipitation increased or decreased, because of climate change?”
Statements about attribution are sensitive to the way the questions are posed and the context within which they are posed. For example, when defining an event, choices must be made about defining the duration of the event (when did it begin and when did it end) and the geographic area it impacted, but this may not be straightforward for some events (e.g., heat waves). Furthermore, different physical variables may be studied (e.g., drought might be characterized by a period with insufficient precipitation, excessively dry soil, or reduced stream flow), and different metrics can be used to determine how extreme an event was (e.g., frequency, magnitude). Whether an observation- or model-based approach is used, and which observations and/or models were available for studying the event, will also constrain the sorts of questions that can be posed.
Attribution studies of individual events should not be used to draw general conclusions about the impact of climate change on extreme events as a whole. Events that have been selected for attribution studies to date (e.g., events affecting areas with high population and extensive infrastructure attract the greatest demand for information from stakeholders) are not a representative sample. Also, events that are becoming less likely because of climate change (e.g., cold extremes) will be studied less often because they occur less often than events whose frequency is increasing because of climate change. Furthermore, attribution of individual events is generally more difficult than characterizing the statistical distribution of events of a given type and its dependence on climate. For example, it may be possible to make confident statements about how some class of extreme events is expected to change because of human-induced climate change, while at the same time an attribution study of an individual event of that type may be unable to make a confident statement about the human influence on that one specific event. Thus, for all of these reasons, counts of available attribution studies with any positive, negative, or neutral results are not expected to give a reliable indication of the overall importance of human influence on extreme events.
Unambiguous interpretation of an event attribution study is possible only when the assumptions and choices that were made in conducting the study are clearly stated and uncertainties are carefully estimated. The framing of event attribution questions, which may depend strongly on the intended application of the study results, determine how the event will be studied and can lead to large differences in the interpretation of the results. Event attribution studies presented in the following manner are less likely to be misinterpreted:
- Assumptions about the state of one or more aspects of the climate system at the time of the event (e.g., SST anomalies, atmospheric circulation regimes, specific weather situations) are clearly communicated.
- Estimates of changes in both magnitude and frequency are provided, with accompanying estimates of uncertainty, so users can understand the estimated degree of change from the different perspectives.
- Estimates of changes in frequency are presented as a risk ratio—that is, in terms of the ratio of the probability of the event in a world with human-caused climate change to its probability in a world without human-caused climate change. Equivalently, one can compare the return periods of the event (i.e., how rarely an event occurs) in the world without climate change to that in the world with climate change.
- The impact of assumptions (e.g., of how estimates of changes in magnitude and frequency depend on SST anomalies or atmospheric circulation regimes) is discussed.
- Statements of confidence accompany results so users understand the strength of the evidence.
Bringing multiple scientifically appropriate approaches together, including multiple models and multiple studies helps distinguish results that are robust from those that are much more sensitive to how the question is posed and the approach taken. For example, robust attribution analyses typically show that the results are qualitatively similar across a range of event definitions, acknowledging that quantitative results are expected to differ somewhat because of differences in definition. Utilizing multiple methods to estimate human influences on a given event also partially addresses the challenge of characterizing the many sources of uncertainty in event attribution.
Examples of multiple components that can lead to more robust conclusions include:
- Estimates of event probabilities or magnitudes based on an appropriate modeling approach that has been shown to adequately reproduce the event and its circumstances, such as the dynamic situation leading to the event.
- Reliable observations against which the model has been evaluated and that give an indication of whether the event in question has changed over time in a manner that is consistent with the model-based attribution.
- Assessment of the extent to which the result is consistent with the physical understanding of climate change’s influence on the class of events in question.
- Clear communication of remaining uncertainties and assumptions made or conditions imposed on the analysis.
Improving Extreme Event Attribution Capabilities
Continued research efforts are necessary to increase the reliability of event attribution results, particularly for event types for which attribution is presently poorly understood. Some of this research is covered in the ongoing work to understand the connection between climate change and long-term statistics of extremes. Improvements in attribution capability for all event types require improvements in observations, models, theoretical understanding of the links between climate change and extremes, and analysis techniques.
A focused effort to improve understanding of specific aspects of weather and climate extremes could improve the ability to perform extreme event attribution. Because extreme event attribution relies strongly on all aspects of the understanding of extremes and their challenges, the committee endorses the recommendations identified in the white paper sponsored by the World Climate Research Programme “WCRP Grand Challenge: Understanding and Predicting Weather and Climate Extremes” (Box S.1; Zhang et al., 2014) as necessary to make advances in event attribution.
In particular, this committee recommends research that aims to improve event attribution capabilities, which includes increasing the understanding of
- the role of dynamics and thermodynamics in the development of extreme events;
- the model characteristics that are required to reliably reproduce extreme events of different types and scales;
- changes in natural variability, including the interplay between a changing climate and natural variability, and characterization of the skill of models to represent low-frequency natural variability in regional climate phenomena and circulation;
- the various sources of uncertainty that arise from the use of models in event attribution;
- how different levels of conditioning (i.e., the process of limiting an attribution analysis to particular types of weather or climate situations) lead to apparently different results when studying the same event;
- the statistical methods used for event attribution, objective criteria for event selection, and development of event attribution evaluation methods;
- the effects of non-climate causes—such as changes in the built environment (e.g., increasing area of urban impervious surfaces and heat island effects, land cover changes), natural resource management practices (e.g., fire suppression), coastal and river management (e.g., dredging, seawalls), agricultural practices (e.g., tile drainage), and other human activities—in determining the impacts of an extreme event;
- expected trends in future extreme events to help inform adaptation or mitigation strategies (e.g., calculating changes in return periods to show how the risk from extreme events may change in the future); and
- the representation of a counterfactual world that reliably characterizes the probability, magnitude, and circumstances of events in the absence of human influence on climate.
Research efforts targeted specifically at extreme events, including event attribution, could rapidly improve capabilities and lead to more reliable results. In particular, there are opportunities to better coordinate existing research efforts to further accelerate the development of the science and improve and quantify event attribution reliability. Also, it would be beneficial to encourage interdisciplinary research at the interface between the climate, weather, and statistical sciences to improve analysis methods. Event attribution capabilities would be improved with better observational records, both near–real time and for historical context. Long, homogeneous observed records
are essential for placing events into a historical context and evaluating to what extent climate models reliably simulate the effect of decadal climate variability on extremes.
Event attribution could be improved by the development of transparent community standards for attributing classes of extreme events. Such standards could include an assessment of model quality in relation to the event/event class. They also could include use of multiple lines of evidence, developing a transparent link to a detected change that influences events in question and the clear communication of sensitivities of the result to how the question of event attribution is asked.
Systematic criteria for selecting events to be analyzed would minimize selection bias and permit systematic evaluation of event attribution performance, which is important for enhancing confidence in attribution results. Studies of a representative sample of extreme events would allow stakeholders to use such studies as a tool for understanding how individual events fit into the broader picture of climate change. Irrespective of the method or related choices, it would be useful to develop a set of objective event selection and definition criteria. This would help to reduce selection bias and, in some cases, lead to methodological improvements. This also is a prerequisite for the development of a formalized approach to evaluating event attribution results and uncertainty estimates, similar to the existing approaches used to evaluate weather forecasts.
Event Attribution in an Operational Context
As more researchers begin to attempt event attribution, their efforts would benefit from coordination to make sure that there is a systematic approach and that uncertainties are explored across methods and framing. Event attribution can benefit from links to operational numerical weather prediction where available. Some groups are moving toward the development of operational extreme event attribution systems to systematically evaluate the causes of extreme events based on predefined and tested methods. Objective approaches to compare and contrast the analyses among multiple different research groups based on agreed event selection criteria are yet to be developed.
In the committee’s view, attributes of a successful operational event attribution system would include the following:
- objective event-selection criteria to reduce selection bias so stakeholders understand how individual events fit into the broader picture of climate change;
- provision of stakeholder information about causal factors within days of an event, followed by periodic updates as more data and analysis results become available;
- clear communication of key messages to stakeholders about the methods and framing choices as well as the associated uncertainties and probabilities; and
- reliable assessments of performance of the event attribution system through evaluation and verification processes utilizing observations and seasonal forecasts and skill scores similar to those used routinely in weather forecasting.
Some future event attribution activities could benefit from being linked to an integrated weather-to-climate forecasting effort on a range of timescales. The development of such an activity could be based on concepts and practices within the Numerical Weather Prediction community. Ultimately the goal would be to provide predictive (probabilistic) forecasts of future extreme events at lead times of days to seasons or longer, accounting for natural variability and anthropogenic influences. These forecasts would be verified and evaluated using observations, and their routine production would enable the development and application of appropriate skill scores. The activity would involve rigorous approaches to managing and implementing system enhancements to continually improve models, physical understanding, and observations focused on extreme events. Although situating some future event attribution activities in an integrated weather-to-climate forecasting effort would lead to more coordination, the committee encourages continued research in event attribution outside of an operational context to ensure further innovation in the field.
The ability to understand and explain extreme events in the context of climate change has developed very rapidly over the past decade. In the past, a typical climate scientist’s response to questions about climate change’s role in any given extreme weather event was, “We cannot attribute any single event to climate change.” The science has advanced to the point that this is no longer true as an unqualified blanket statement. In many cases, it is now often possible to make and defend quantitative statements about the extent to which human-induced climate change (or another causal factor, such as a specific mode of natural variability) has influenced either the magnitude or the probability of occurrence of specific types of events or event classes. The science behind such statements has advanced a great deal in recent years and is still evolving rapidly. Still further advances are necessary, particularly with respect to evaluating and communicating event attribution results and ensuring that event attribution studies
meet the information needs of stakeholders. Further improvement will depend not only on addressing scientific problems specific to attribution but also on advances in the basic underlying science, including observations, modeling, and theoretical understanding of extreme events and their relation to climate change.