An Updated Review of Event Attribution Approaches

Cheng QIAN, Yangbo YE, Yang CHEN, and Panmao ZHAI

1 Key Laboratory of Regional Climate–Environment for Temperate East Asia, Institute of Atmospheric Physics,Chinese Academy of Sciences, Beijing 100029

2 University of Chinese Academy of Science, Beijing 100049

3 State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, China Meteorological Administration, Beijing 100081

ABSTRACT There have been considerable high-impact extreme events occurring around the world in the context of climate change. Event attribution studies, which seek to quantitatively answer whether and to what extent anthropogenic climate change has altered the characteristics—predominantly the probability and magnitude—of particular events,have been gaining increasing interest within the research community. This paper reviews the latest approaches used in event attribution studies through a new classification into three major categories according to how the event attribution question is framed—namely, the risk-based approach, the storyline approach, and the combined approach.Four approaches in the risk-based framing category and three in the storyline framing category are also reviewed in detail. The advantages and disadvantages of each approach are discussed. Particular attention is paid to the ability,suitability, and applicability of these approaches in attributing extreme events in China, a typical monsoonal region where climate models may not perform well. Most of these approaches are applicable in China, and some are more suitable for analyzing temperature events. There is no right or wrong among these approaches, but different approaches have different framings. The uncertainties in attribution results come from several aspects, including different categories of framing, different conditions in climate model approaches, different models, different definitions of the event, and different observational data used. Clarification of these aspects can help to understand the differences in attribution results from different studies.

Key words: event attribution, extreme events, anthropogenic climate change, risk-based approach, storyline approach, combined approach

1. Introduction

People around the world have heard of or experienced high-impact extreme weather and climate events more frequently in recent years. For example, the extraordinary heatwave on the Pacific coast of the US and Canada during June 2021, the severe growing-period frosts during early April 2021 in central Europe, the record-breaking persistent heavy precipitation (super Meiyu) in the middle and lower reaches of the Yangtze River in China in June–July 2020, and the destructive downpours in July 2021 in Henan Province of China. In the immediate aftermath of a high-impact extreme weather or climate event, the most frequently asked question from governments, media, and public is whether the event was related to climate change. From a climate science perspective, this question falls into the realm of “event attribution”, which seeks to quantitatively answer whether and to what extent anthropogenic climate change has altered the characteristics—predominantly the probability and magnitude—of particular events (Stott et al., 2016;Seneviratne et al., 2021). This information is vital to inform how climate change is worsening severe weathers and further guide the community to better prepare for greater climate-related risks in a future warmer world or to better rebuild our cities and infrastructure for a climate-changed world after a disaster (https://www.technologyreview.com/10-breakthrough-technologies/2020/). In view of these scientific and practical values, a rapid attribution of Tropical Storm Imelda in 2019 was selected as one of the 10 Breakthrough Technologies 2020 by the Massachusetts Institute of Technology (MIT;https://www.technologyreview.com/10-breakthroughtechnologies/2020/).

This paper reviews the currently available methods for event attribution studies from a new classification perspective. We pay particular attention to the ability, suitability, and applicability of these methods in attributing extreme events in China, a typical monsoonal region. We also discuss the sources of uncertainties in attribution results and some new pathways underway in current Key Research and Development projects in China (2018YFC 1507700).

2. Event attribution approaches

According to how the event attribution question is framed, three categories of approaches can be classified and the attribution result depends very much on the framing (Fig. 1). A commonly used framing is the so-called risk-based approach (or probability-based approach)(Seneviratne et al., 2021), which treats the specific event that inspired the study as one case of a class of similar events to the one in question and answers probabilistically the question of to what extent climate change has altered the probability of this class of event (Stott et al.,2004). A similar concept to the shift in the probability due to the shift in the distribution is the change in the magnitude for a given return period (Otto et al., 2012).For example, a 1-in-50-yr event has become more intense in the current climate-changed world. The second and rather different framing is the storyline approach(Shepherd, 2016). This method traces back various factors contributing to the specific event based on physical understanding and takes some of them as preconditions (mostly large-scale circulations), answering deterministically the question of to what extent climate change has altered the magnitude of this specific event (Shepherd, 2016). The third approach is termed here as the combined approach, as it combines the previous two seemingly irreconcilable approaches (Ye and Qian, 2021;Qian et al., 2022) and integrates them into a common framework (Shepherd, 2016). In the following, we review these three categories of approaches in detail, oneby-one.

Fig. 1. Classification in this study of current methods for event attribution research.

2.1 Risk-based approach (or probability-based approach)

Easterling et al. (2016) termed this category of approach as “Oxford school”, because it was proposed by Allen (2003) and implemented by Stott et al. (2004) for the attribution of the European heatwave of 2003. As stated above, this approach answers the question of to what extent anthropogenic climate change has altered the probability (Stott et al., 2004) or magnitude (Otto et al.,2012) of this class of event similar to the one in question.The attribution result states, for example, “anthropogenic climate change made this event type twice as likely,” or“anthropogenic climate change made this event 15%more intense” (Seneviratne et al., 2021). A well-known metric in this approach is the fraction of attributable risk(FAR; Stott et al., 2004). It is calculated by 1-P0/P1,whereP1is the occurrence probability in today’s climate(the factual world), whereasP0is that in a counter-factual world where anthropogenic influences on climate are absent (Stott et al., 2004). A similar concept was originally referred to as the “risk ratio”, but more precisely termed as the “probability ratio” (PR; Fischer and Knutti, 2015).Some studies have simply compared return periods in the factual and counter-factual world. In terms of how to generate the counter-factual world, there are two major categories of the probability-based approach, with and without climate models.

2.1.1Without climate models

This category of the probability-based approach is based on statistical analysis of observations. Specifically,studies compare the statistical characteristics in a recent period when climate change is more intense to those in an earlier period when climate change was weaker, and consider the overall effects of climate change over the analysis period. This method can be further classified into the empirical approach and the flow analogue approach,the latter of which has also been termed as the analoguebased approach (Stott et al., 2016). We review them in detail in the following subsections.

2.1.1.1 Empirical approach

This approach fits a statistical distribution to the observations and estimates the role of climate change through comparing the difference in the return period or probability between a recent period and an earlier period,assuming that climate change only affected the location parameter but not the other parameters for temperature events, or assuming that climate change affected the distribution for precipitation time series or variables related to (the lack of) precipitation or wind (Philip et al., 2020).The fitted distribution can be a generalized extreme value(GEV) distribution (Van Oldenborgh et al., 2015), a generalized Pareto distribution (King et al. 2015), or other distributions (Philip et al., 2020). For example, Van Oldenborgh et al. (2015) analyzed the unusual cold extremes during the winter of 2013/2014 in North America and included the effects of global warming to first order by allowing the location parameter µ in the fitted GEV distribution to vary, assuming a linear dependence on the(lowpass filtered) global mean temperatureTg:µ=µ0+αTg. They found that cold extremes in this region were significantly less likely, with a return period in 2013/2014 of about 12 yr compared to that (about every 4 years) in the colder climate of the 1950s. This kind of method has been applied by Chinese scientists. For example, Zhou et al. (2018) applied it to reflect the influence of global warming on the likelihood of precipitation extremes occurring in July 2016 in Wuhan and China. They found that under the global mean air temperature of 1961, such a type of precipitation extreme was a 1-in-272-yr event; however, it became a 1-in-28-yr event when the observations were shifted up with the global mean air temperature of 2016.

The advantage of this approach is that, as it does not involve climate models at all, it can therefore be used to verify climate model results and is more readily accepted by skeptics of the veracity of climate models; plus, it can also be applied to cases that climate models do not simulate well (Stott et al., 2016). The disadvantage of this approach is that it can only estimate the overall effect of climate change on the extreme event, or assumes that the trend is human-induced; however, the trend in a short interval can also be affected by decadal or multidecadal internal variability (NASEM, 2016).

Thinking specifically about China, regional temperatures, precipitation, and drought can also be affected by multidecadal variability (Qian and Zhou, 2014; Qian,2016), and thus if only data after 1950 are used, it is difficult to clearly distinguish the effects of multidecadal variability and human-induced climate change. Unfortunately, there are few long-term daily observational datasets starting before 1950 available in China, and thus this approach retains some uncertainty if it is used alone in event attribution.

2.1.1.2 Flow analogue approach

Flow analogues, which were previously used in weather forecasting (Lorenz, 1969), can be used to obtain the values of a variable from historical days with large-scale synoptic atmospheric circulation similar to the day of interest during an extreme event (Yiou et al., 2007;Jézéquel et al., 2018). Through a reconstruction from several best analogues of each day, it can be used to estimate either the role of circulation in an extreme event(Jézéquel et al., 2018; Ren et al., 2020b) or the overall climate change over the period considered for an extreme event (Stott et al., 2016), or even both (Ye and Qian,2021). When the atmospheric circulation is fixed to resemble the observation, the difference in the occurrence probability between the current period and an earlier period is due to the thermodynamic effect of climate change (Stott et al., 2016; Ye and Qian, 2021). For example, Ye and Qian (2021) refined the flow analogue method to put the estimation of the contributions of both atmospheric circulation and climate change to an extreme event in the same framework; and they found that the atmospheric circulation explained 71% of the magnitude of the record-breaking extreme precipitation event in the middle and lower reaches of the Yangtze River during the Meiyu season of 2020, and that climate change had increased the occurrence risk of this type of event under similar atmospheric circulation conditions by five times.

This approach belongs to the category of conditional attribution, which limits or constrains the state of one or more parts of the climate system (NASEM, 2016). The strength of this approach is that it is easily adaptable to other regions, and to other events (Jézéquel et al., 2018).As this approach also does not involve climate models, it can be used as a complement to climate model-based approaches to help understand the results from model-based approaches (NASEM, 2016). The disadvantages of this approach are that analogues are selected without any condition on the previous days and do not properly consider the soil–moisture feedback (Jézéquel et al., 2018).Therefore, it might fail to reach the observed anomaly if it is applied to a drought event or heatwave. Some other feedback processes, such as snow/ice albedo feedback,may also be unaccounted for in the flow analogue method.Moreover, this approach cannot fully distinguish the role of human-induced climate change (Shepherd, 2016). In addition, one may argue that climate change may also alter the atmospheric circulation (Stott, 2016), although this is highly uncertain and hard to prove in a particular region. It should be noted that one may also argue that when an event is record-breaking or the dynamic setup leading to the event is unique, there might be no analogue close enough to the observed case. However, in practice, the extreme climate event is usually reconstructed from analogues selected at the daily scale, meaning that a day with a similar synoptic situation historically can normally be found.

In China’s case, although its application is still limited, this approach is adaptable, as was demonstrated in a conditional attribution study of an extreme precipitation event and concurrent extreme temperature event in southern China in 2020 (Ye and Qian, 2021). For any type of extreme event that climate models do not simulate well,this method might be a preferable option.

2.1.2Using climate models

In order to clearly distinguish the role of human-induced climate change, climate models are an indispensable tool to construct the counterfactual world in the absence of anthropogenic forcings. According to the level of conditioning on the state of the climate system used,this type of event attribution approach can be further divided into two sub-approaches—namely, the coupled climate model approach [the least conditional approach considering the combined effect of the thermodynamic change and changes in the large-scale atmospheric circulation (Stott et al., 2004)]; and the atmosphere-only model approach, which is driven by the observed sea surface temperature (SST) and sea–ice concentration (SIC) at the time of the extreme event [SST/SIC-conditioned approach (Christidis et al., 2013; Ciavarella et al., 2018)].Either approach generates data from large ensembles of simulations with and without anthropogenic influences to obtain large samples of the climatic variable under investigation. The prerequisite of these approaches is that these models should be able to perform well. Therefore,model evaluation is important (Trenberth et al., 2015);however, this task is difficult for counterfactual world experiments because it never existed (Shepherd, 2016).

2.1.2.1 Coupled model approach

This method takes the effects of climate change on dynamic and thermodynamic aspects of the event(s) in question as a whole, though sometimes researchers struggle to use the models to reproduce the dynamics in terms of magnitude, position, and configuration amongst multiple circulation agents (especially for monsoonal precipitation extremes over eastern China). As this approach has the least conditions, it estimates overall probabilities from both the thermodynamic change and changes in the large-scale atmospheric circulation. It therefore takes into account the influence of human-induced climate change on changes in the atmospheric circulation.

a. Unconditioned

Two sub-approaches have been widely used. One is carried out through using the optimal fingerprinting method to constrain the signal and then superimposing the natural internal variability onto it, as originally proposed by Stott et al. (2004) using the HadCM3 (Hadley Centre Coupled Model version 3) climate model and later extended by Sun et al. (2014) using multiple models from phase 5 of the Coupled Model Intercomparison Project (CMIP5). This approach has been used in several papers in China (Song et al., 2015; Sun et al., 2016,2018). The other sub-approach does not constrain the signal and uses the climate model output from the CMIP archive directly. Several studies have applied this approach in China (Zhou et al., 2014; Miao et al., 2016;Sun et al., 2019; Wang et al., 2019, 2021).

The advantage of this approach is that there are a number of climate models already available—for example,those in the CMIP5 and CMIP6 archives; and therefore,large samples can be obtained for the extreme event of interest, allowing the generation of robust statistics(NASEM, 2016). The constrained approach using the optimal fingerprinting method also demonstrates a strong linkage between conventional detection and attribution studies on long-term changes and event attribution (Stott et al., 2016). The limitation of this approach is that the credibility of the attribution depends highly on the simulation performance of the climate models used. For small-scale events or those in a region with complex terrain where climate models perform poorly, this approach will result in large uncertainty. The constrained approach using the optimal fingerprinting method is powerful when applied to broad-scale events; however, it is difficult to apply to small-scale events because at that scale the scaling factor in the optimal fingerprinting method cannot be calculated or reliably estimated. The optimal fingerprinting method relies on the assumption of a Gaussian regression residual, which is hard to fulfill at small scales, especially for variables other than temperature such as extreme precipitation.

For China, the scaling factors in optimal fingerprinting detection and attribution of long-term changes in temperature and extreme temperatures at various scales can be estimated and the signals detected (Qian and Zhang, 2019; Sun et al., 2022). Therefore, the constrained approach using the optimal fingerprinting method can be successfully applied in temperature-related events(Sun et al., 2014, 2015, 2016, 2018), but not for other variables. The formal detection and attribution of longterm changes in precipitation and extreme precipitation in China is rather complicated (Sun et al., 2022).

b. Additionally conditioned

It should be noted that there have also been some studies that further conditioned a certain aspect of the climate system to conduct conditional attribution to see their roles as well, after conducting the above-mentioned unconditioned analysis. These studies in China can be classified into three categories: individual forcing-conditioned, climate variability mode-conditioned (typically ENSO-conditioned), and atmospheric circulation-conditioned. The attribution statements of these studies can be,for example, “human-induced climate change and El Niño increased the likelihood of similar events by 10-fold and 50%, respectively.” In terms of individual forcing-conditioned studies, one would use the greenhouse gas (GHG) forcing-only simulation or anthropogenic aerosol forcing-only simulation from the CMIP6 archive(Lu et al., 2021), or the land-use change forcing-only simulation to drive a land surface model (Ji et al., 2020).Of course, other techniques can also be used to isolate the effect of individual forcing alone. For example,Freychet et al. (2019) used a nudging method to constrain the model circulation and only vary aerosol emissions to isolate the effect of the aerosol emission alone,although they did not conduct an event attribution analysis. In terms of climate variability mode-conditioned studies, previous studies basically follow the study of Lewis and Karoly (2013) to pick out El Niño years and estimate the shift in likelihood under El Niño conditions(Sun and Miao, 2018; Yuan et al., 2018; Zhou et al.,2018; Du et al., 2020, 2021). Some studies have also conditioned the Arctic Oscillation (AO; Du et al., 2020).In terms of atmospheric circulation-conditioned studies,they condition the atmospheric circulation to that relevant to the event under investigation, such as the western Pacific subtropical high (WPSH; Zhou et al., 2019,2020). Some studies will overlap using these conditions—for example, Zhou et al. (2019, 2020) used both WPSH-conditioned and local urbanization effect-conditioned analyses, and Du et al. (2020) used both AO-conditioned and ENSO-conditioned analyses. Since the causal factors are sometimes not independent from each other(e.g., AO and ENSO), the attribution results may overlap each other. In addition, as each ENSO event is distinct,this also adds some uncertainty in the attribution results.

2.1.2.2 Atmospheric model approach

A second category of using climate models is the SST/SIC-conditioned approach, which uses atmosphereonly models driven by observed SST, SIC, and external forcings at the time of the extreme event. These models are often called “Atmospheric Model Intercomparison Project” (AMIP) large ensemble runs, such as the operational event attribution system (HadGEM3-A) at the Met Office Hadley Centre (Christidis et al., 2013, 2018).

There have been many successful applications of this approach in China. These studies include by using only one AMIP model, such as the HadGEM3-A model(Burke et al., 2016; Qian et al., 2018; Chen et al., 2019a;Ren et al., 2020a; Zhang W. X. et al., 2020; Hu et al.,2021) or MIROC5 (Ma and Zhu, 2019); using two AMIP models (HadGEM3-A and CAM5; Zhang L. X. et al.,2020); using an AMIP model (HadGEM3-A) and comparing it with CMIP5 models (Li et al., 2018; Sun et al.,2020; Lu et al., 2021); using two AMIP models(CAM5.1 and MIROC5) and comparing them with CMIP5 models (Ma et al., 2017a, b); or even using an AMIP model (HadGEM3-A) and comparing with regional models (weather@home; Sparrow et al., 2018). For example, by using the 525 large ensemble simulations with and without anthropogenic forcings from the HadGEM3-A-N216 model, Qian et al. (2018) found that anthropogenic influences had reduced the likelihood of an extreme cold event in midwinter similar to the recordbreaking one in January 2016 in eastern China by about two-thirds.

The advantage of this approach is that prescribing SSTs in an atmospheric model instead of using coupled models can reduce model biases and enables more ensemble simulations and high-resolution simulations because they are cheaper and faster to run, potentially resulting in a better representation of extreme events, and enhancing the signal-to-noise ratio (Stott et al., 2016). As the SST condition at the time of the event is observed, it is quite suitable for events that are induced by ocean effects on the atmosphere in one direction. The disadvantage of this approach is that when an event is affected by strong ocean–atmosphere coupling, applying an atmosphere-only model might induce bias (Stott et al., 2016).In addition, in counterfactual world experiments to produce the SST in the absence of anthropogenic forcings,the strategy of subtracting the ensemble mean SST of CMIP models (as a measure of the influence of anthropogenic forcings) is often applied. Therefore, the choice of SST and SIC patterns is highly determined by estimates of the climate sensitivity to external human forcing agents (Easterling et al., 2016), and thus influences the attribution results (Easterling et al., 2016; NASEM,2016; Sparrow et al., 2018; Seneviratne et al., 2021).Even using the same SST and SIC forcing pattern, different models may yield different responses of extreme events (Easterling et al., 2016). This adds another source of uncertainty to the attribution conclusion.

It should also be noted that there have been some studies in China (Lu et al., 2020; He et al., 2021) that further conditioned the key atmospheric circulation system to resemble the one at the time of the event to estimate the role of this key circulation in the shift in likelihood, basically following the correlation coefficient method in Christidis and Stott (2015), after conducting an SST/SICconditioned atmosphere-only model analysis. For example, Lu et al. (2020) analyzed the hottest spring event in eastern China in 2018 and reported that anthropogenic forcings may have increased the chance of such an event by 10 times, while the anomalous circulation had a roughly 2-fold impact. It should also be noted that there were some earlier case studies of extreme events in China (published in a special issue of the Bulletin of American Meteorological Society) that did not reach an attribution conclusion owing to the poor performance of the climate model used, and suggested that model biases needed to be reduced to enable successful attribution(Wilcox et al., 2015). Therefore, evaluating the ability of the climate model is a key step when using this approach in event attribution.

2.2 Storyline approach

Easterling et al. (2016) termed this category of approach as “Boulder school.” It is a framing based on analysis of the physical processes and is similar to accident investigation (Shepherd, 2016). The causal chain of factors leading to the event is firstly identified, and the role of each factor is assessed (Trenberth et al., 2015;Shepherd, 2016). This approach is used to quantify the effect of the anthropogenic forcings on the magnitude of the specific extreme event itself, which drives the event attribution analysis (Shepherd, 2016). The specific event and the resulting impacts are perceived by the local-to-regional community more than elsewhere, who will then be more open to believing the impacts of climate change and take this forward into their future planning. For example, Hoerling et al. (2013) used this approach was used. They investigated the 2011 Texas drought/heatwave and found that about 20% of the magnitude of this event resulted from human-induced climate change.

Methods within the storyline approach can be classified into three categories. The first is based entirely on observations, which by definition does not rely on models. For example, Diffenbaugh et al. (2017) proposed a method to separate the contributions of observed longterm trends to individual events. Their results showed that 41% of precipitable water associated with the 2013 Colorado floods was contributed by the trend (which can be roughly considered as the influence of global warming). Ye and Qian (2021) refined the flow analogue method to estimate the role of atmospheric circulation and climate change in the magnitudes of two 2020 extreme events in southern China. They found that the atmospheric circulation explained about 71% and 44% of the magnitude of the extreme precipitation event in the middle and lower reaches of the Yangtze River and the concurrent hot event to its south, respectively; and that under atmospheric circulation patterns similar to the 2020 pattern, climate change increased the magnitude of these two events. The second category is based on global climate models and can be used to study large-scale extreme events (Hoerling et al., 2013; van Garderen et al.,2021; Wang G. M. et al., 2021). A typical study of this type was carried out by Wehrli et al. (2019), who quantitatively estimated the contribution of recent GHG emissions, atmospheric circulation, the land surface, and oceans to five heatwaves from 2010 to 2016 based on the Community Earth System Model. The last category is based on regional climate models, from which high-resolution simulation results can be obtained (Meredith et al., 2015; Michaelis et al., 2019; Reed et al., 2020). One useful method in this category is the “pseudo-global warming” method, in which control simulations are conducted to reproduce the observed extreme events, and then “deltas” are applied to the initial and boundary conditions to re-simulate and assess changes in the characteristics of the event (Michaelis et al., 2019).

The advantage of this approach is that it is very relevant for understanding the origin of an extreme event and helping understand how climate change is playing a role in (thermo) dynamic processes that contribute to that particular event and probably influence its magnitude (Philip et al., 2020). It is therefore helpful for understanding this particular event and for guiding research on its prediction. The method does not depend on the performance of the model in representing the circulation (dynamic processes) reliably (Shepherd, 2016), which is important in a risk-based approach. At regional scales, the effect of climate change includes both thermodynamic and dynamic effects. Therefore, the credibility of the risk-based approach is a challenge where climate models have poor ability in simulating the circulation that dominates the occurrence of a particular event, especially in regions with complex terrain, such as a mountain valleys or canyons (Shepherd, 2016). Under this circumstance, it has been suggested that the storyline approach should be preferred (Trenberth et al., 2015). This advantage makes the method particularly suitable for attributing monsoonal precipitation extremes, even at a local scale, whose occurrence and intensity are rather sensitive to the position,magnitude, and configurations amongst several key circulation agents (Chen and Zhai, 2014, 2015, 2016; Chen et al., 2019b).

The disadvantage of this approach is that it does not take into account the effect that circulation may also be altered by climate change, and is therefore only a partial attribution or conditional attribution (Shepherd, 2016)and cannot report the actual risk of the event occurring,as is done by the unconditional risk-based approach (Otto et al., 2016). The imposed conditions limit an overall assessment of the anthropogenic influences on that event,as the large-scale dynamics can counteract or enhance the thermodynamics (Seneviratne et al., 2021). It has also been argued that the storyline approach does not involve the probability of similar events and is thus useless for future adaptation since the event in question occurred only once (Philip et al., 2020). In addition, the storyline approach may not be relevant to either the assignation of blame, nor planning decisions, in disaster recovery (Otto et al., 2016).

Theoretically, the above advantages and disadvantages are also valid for China. There have barely been any papers published as yet to report work in which this approach was applied in event attribution in China, other than the above-mentioned study by Ye and Qian (2021)based entirely on observations. A recent Key Research and Development project (2018YFC1507700) on event attribution is applying this approach to investigate the extreme precipitation event during the Meiyu season of 2020 and the flood-triggering heavy precipitation in Zhengzhou in 2021, based on regional model simulations.

2.3 Combined approach

There is no right or wrong between the above two reviewed and seemingly irreconcilable approaches (Lloyd and Oreskes, 2018). Actually, they can be used to complement each other and be viewed within a common framework (Shepherd, 2016). Here, we term the approach, which combines the risk-based approach with the storyline approach, as the combined approach. Very few studies have as yet applied (Cheng et al., 2018) or developed (Ye and Qian, 2021) this approach. Among the few, which was one of the outcomes from the abovementioned Key Research and Development project of China, Ye and Qian (2021) developed a flow-analoguebased method to first estimate the contribution of atmospheric circulation to the magnitude of the particular extreme event using a storyline approach, and then estimated the contribution of climate change to the occurrence probability of a similar event using the risk-based approach. In the case of the above-mentioned record-breaking extreme precipitation event in the middle and lower reaches of the Yangtze River during the Meiyu season of 2020, they found that the atmospheric circulation explained 71% of the magnitude of this event and climate change had increased the occurrence risk of such an event under similar atmospheric circulation conditions by five times. The above-mentioned Key Research and Development project is applying and further developing this approach. In this respect, Qian et al. (2022) quantified the contribution of anthropogenic forcings and atmospheric circulation to the heavy rainfall event that occurred in mid-August 2020 in southwestern China. They found that atmospheric circulation resembling that in 2020 can explain about 47% of the observed heavy rainfall intensity, and anthropogenic forcings have roughly doubled the likelihood of such a heavy rainfall event. In terms of China, where the terrain is complex and the causal factors complicated, this approach has the advantages of both the risk-based approach and the storyline approach, and to some extent avoids their disadvantages when used alone.

3. Summary and concluding remarks

In this paper, we have reviewed three major categories of framing and four specific approaches for the riskbased framing for event attribution studies. The advantages and disadvantages of these approaches are summarized briefly in Table 1. They can all be used in China as in other countries, but with different levels of confidence.If the climate model possesses good ability in simulating the atmospheric circulation, the confidence in the attribution from the coupled model approach and atmospheric model approach will be high. If not, the empirical approach, flow analogue approach, storyline approach, or combined approach might be preferable. As China is located in a typical and complex monsoonal region where climate models may not perform well, evaluation of model ability is a key step when using the coupled model approach and atmospheric model approach in this region.

Table 1. Summary of the advantages and disadvantages of each event attribution approach reviewed in this paper

As stated above, different approaches frame the event attribution question differently. Thus, they cannot be compared directly, although some previous studies have applied different approaches for a single event in order to try to validate the result from one approach, viewing their differences as one kind of uncertainty source. For example, for the same case—the Russian heatwave event in 2010—it can be both mostly internally generated in terms of magnitude [mainly natural in origin, as concluded in Dole et al. (2011), which is from a storyline approach perspective] and mostly externally driven in terms of occurrence probability [80% of its probability would not have occurred without the large-scale climate warming since 1980, as concluded in Rahmstorf and Coumou(2011), which is from a risk-based perspective], as explained by Otto et al. (2012). As stated previously, the storyline approach targets the specific event itself;however, the risk-based approach treats the specific event that inspired the study as one case of a class of similar events to the one in question. This class could be similar in magnitude (with intensity equal to or larger than the threshold as commonly used in the FAR calculation), at a slightly different time of the year [for example,Qian et al. (2018) included adjacent pentads and Zhang W. X. et al. (2020) calculated maximum accumulated rainfall over four weeks over the entire summer to avoid selection biases], or in a nearby region [for example,Philip et al. (2020) pooled observations or grid points over a homogeneous region that was larger than that of the events in order to increase the number of independ-ent events]. Even if in the same risk-based framing,quantitative attribution results from different approaches cannot be directly compared either, although qualitative attribution results that anthropogenic forcings have played a role are consistent with each other. For example,for the same case (the July–August record-breaking hot event in northwestern China in 2015), Zhang et al. (2022)estimated that anthropogenic forcings have increased the likelihood of this class of events by approximately 27 and 12 times in the MIROC5 and HadGEM3-A-N216 models, respectively, based on the SST/SIC-conditioned atmosphere-only model approach with prescribed observed SST/SIC boundary conditions and external forcings in 2015. However, with an unconditional coupled model approach from the CMIP5 archive, which was given the external forcing at the 1961–2015 level regardless of the SST/SIC boundary conditions, there is a 21 times increase in the likelihood of similar events due to anthropogenic forcings. Therefore, only when the approach is the same, can quantitative results be compared directly, such as in the 2015 case in China above, i.e., 27 and 12 times in MIROC5 and HadGEM3-A-N216 (Zhang et al.,2022). Indeed, as an extreme event is a low probability event often at the tail of the distribution, a tiny difference in the climate model can result in a huge difference in the estimated FAR or PR. It should also be noted that even if the model used is the same, different definitions of the event can also result in different or even seemingly contradicting attribution conclusions. For example,Zhang W. X. et al. (2020) found that anthropogenic forcing has reduced the probability of summer persistent heavy rainfall (defined as the maximum accumulated rainfall over four weeks from June to August, or Rx28day in short) in central western China similar to that in 2018 by about 47%, but increased that of daily extremes (defined as the maximum 1-day rainfall, or Rx1day in short) by about 1.5 times, based on the same HadGEM3-GA6-based atmosphere-only model. These seemingly contradicting attribution conclusions are due to the different physical processes in the corresponding definition of the event at different timescales (Zhang W.X. et al., 2020). In addition, using different observational data can also introduce uncertainty in the attribution conclusion, which also deserves attention.

To sum up, different approaches frame the attribution question differently. Therefore, every event attribution study should clearly state the framing of the attribution question in their communication of the attribution results in order not to induce seemingly contradictory results. It is also important to clearly state how the event is defined,which is another important source of uncertainty. These ways of communication will help us to integrate and better understand attribution results from different studies.