Processes and principles for producing credible climate change attribution messages: lessons from Australia and New Zealand

Extreme event attribution (EEA) information is increasingly in demand from climate services. EEA messages can: raise awareness about the effect climate change has already imposed, inform climate change liability conversations, and be combined with climate projections to inform adaptation. However, due to limitations in observations, models and methods, there are barriers towards operationalising EEA in practice. Operational services will need EEA to be done transparently and using preset formats. Here we review recent experience and practice in EEA in Australia and New Zealand with a view to inform the design of an EEA component of climate services. We present a flow chart of the processes involved, noting particular care is needed on the trigger, event definition, and climate model evaluation, with effective stage gates. We also promote the use of tailored causal network diagrams as a standard tool to inform an EEA study and communicate results, with particular care needed for messages on events with lower confidence or complex sets of influences, including tropical cyclones and extratropical cyclones. We suggest that extending EEA to impact attribution is essential for making EEA messages salient but requires an uplift in forming interdisciplinary teams and in granular exposure and vulnerability datasets and is likely to raise new interdisciplinary methodological questions. Finally, we suggest communication of EEA messages can learn more from its origins in medical epidemiology.


Introduction
Extreme event attribution (EEA) seeks to estimate the influence of human activities or other factors, including natural variability, on the probability, severity or other characteristics of an observed extreme weather or climate event.Beyond improving understanding of the response of extremes to different climate forcings, EEA messages have at least four purposes: (1) to raise awareness to a general audience about the effects climate change is already having; (2) to inform analysis of economic risk and the social cost of carbon (3) to inform decisions about building climate resilience; and (4) to address questions of climate change liability, Loss and Damage (L&D) and climate justice (e.g.Schwab et al 2017, Burger et al 2020, Ettinger et al 2021).Demand to operationalise EEA is growing, but EEA is a relatively new and rapidly evolving field, with no agreed standards for many aspects.Creating more guiding principles and protocols for operational EEA can make the practice more efficient, credible and effective, including in the process of selecting and defining

Current practice in ANZ
ANZ have some similarities and some important differences in the extreme events experienced, in the responsibilities and needs of decision-makers and in the current aims of EEA programs.Heat extremes and drought are of interest to all of ANZ.TC impacts are very relevant to northern Australia, and ex-TCs and tropical to extratropical transitions are a strong focus for New Zealand.Impacts from marine heatwaves on ecosystems and fisheries are notable in the East Australia Current region, their downstream effect on New Zealand's climate (Salinger et al 2019), and in tropical reefs in terms of coral bleaching.
As for many places in the world, studies of EEA and changing likelihoods are mostly done opportunistically in ANZ and there is no comprehensive or systematic process or operational service.Scientists in New Zealand have developed a forecast-based attribution system that could be applied operationally at the Met Service (Tradowsky et al 2023), while a suite of methods is also under development in Australia (Hope et al 2022).Though the next step towards routine operations has not been reached in either country, an early trial for Australia is detailed below.New Zealand is aiming for an inventory of events, costs, and hazard attribution assessments but this is not yet a goal in Australian programs.
Observational records in ANZ are similar or better than most places in the world, meaning that context-setting and EEA studies can be done with higher confidence than most regions.The remaining main data gaps in ANZ are a lack of sub-daily data, a sparse observing network in remote areas of Australia and short records in regions with high altitudes (Tait et al 2012, Evans et al 2020, Trewin et al 2020).In the ANZ region, marine records are not robust prior to the advent of satellite retrievals.For TCs, the short record of reliable observations (Chand et al 2019, Courtney et al 2021) makes it difficult to detect a signal from the noise.Simulating Australian climate features and extremes in models includes challenges related to inadequate resolution and other features that may be more prevalent compared to other regions (Lane et al 2023).
There are broadly three categories of methods and models used in climate change context-setting, EEA, and related topics in ANZ: 1. Context setting-various relevant analyses used as part of EEA messaging, including: the analysis of trends in observations; multivariate elucidation using observations and indices to illustrate the relative role of different drivers and of climate change through a simple multiple linear regression model (e.g.

Recent experience in developing narratives
Here we compare and reflect on the lessons in producing credible EEA narratives using two recent examples in ANZ.
The first example is probabilistic risk-based statements on temperature records in Australia in winter 2023: the record warm July in Tasmania (see Hope et al 2024, this issue) and the record warm winter (June-August) in Australia in 2023.These were 'test runs' for an effort by Australia's National Meteorological and Hydrological Service (NMHS: Bureau of Meteorology) to create real-time rapid EEA statements about seasonal and monthly heat records, using multiple methods.These Australian heat events were chosen as test cases for five reasons: (1) month-to season-long average temperature records are routinely reported by the NMHS (2) they are extended heat events with a broad spatial and temporal scale, so we assume high confidence in the direction of influence and the likely order of magnitude of the influence of climate change on the events; (3) multi-method approaches can all be used (simple scaling, statistical and dynamical methods); (4) rapid EEA was possible and the analysis could be completed ready to communicate at the end of the relevant month/season; and (5) there were impacts in agriculture and ecology, including early flowering, but there are no identifiable L&D questions and so discussion of winter warmth is less contentious or potentially inflammatory than for damaging events.
A range of methods were applied to provide contextualisation from observations and seasonal outlooks, unconditioned CMIP-based analysis of FAR and emergence, conditioned ACCESS-MICAS modelling, and the multilinear regression reconstruction of drivers and climate warming.The four methods used all agreed that anthropogenic climate change had likely contributed to the record warm monthly average maximum daily temperature (Tmax) observed in July in Tasmania, but differed in the specifics.Unconditioned methods based on CMIP model data suggested that breaking the maximum temperature record was 17 times more likely when compared to a world without anthropogenic climate change using the FAR framework.Experiments run with the conditioned multi-week to seasonal forecast model (ACCESS-MICAS) found that 0.6 • C of the forecast heat is attributable to climate change.However, Tasmania is a small island relative to the resolution of CMIP models (∼300 km across), the models poorly represented average minimum temperatures (Tmin), and ACCESS-MICAS underestimated the magnitude of the anomalous maximum temperature (0.6 • C compared to 1.4 • C).Also, an observation-based multilinear regression reconstruction of the maximum temperature estimated only a small positive anomaly, with cooling contributions from the developing El Niño, positive Indian Ocean Dipole (IOD), and negative Southern Annular Mode (SAM) apparently largely counteracting the heating contribution from the on-going warming trend, differing from observed relationships (e.g.McKay et al 2023).A subsequent rapid attribution trial of the Australia-wide winter-mean (June-August) record high monthly daily temperature maximum temperature (Tmax) using the same set of methods that was well modelled in CMIP models showed that Australian winters that exceed mean temperatures of 16.5 • C are at least 18 times more likely due to anthropogenic climate change and the Australia-wide average winter temperature.Almost all the forecast anomalous heat was attributable to climate change.While the direction of influence is clear for both warm Tmax records, confidence in the quantification of the effect is much higher for the Australian winter than in Tasmanian winter July case.
Several lessons were learnt through the July and winter 2023 experience.The first lesson was around the timeliness of modelling needed for a multi-model approach.Some of the datasets were available and models could be run nearly instantaneously, other analyses an models required several weeks before the necessary data was obtained; a timeframe that is too long to be included when the NMHS statements about the record are released at the end of the month of the record in question.As such, alternate data sources, models or methods may be required for future rapid attribution analysis, so the results are available in near real-time.The second lesson was that it is useful to have a pre-established assessment of the suitability of each method for likely events across Australia and throughout the year.Finally, nuances between the methods and subtleties between their results, as well as differences between different climate baselines, meant that a lot of care was needed when communicating the results to avoid confusion.While the direction of influence was clear and consistent from all methods, the quantification of the effect was dependent on the method, and the message needs to be worded differently for unconditioned (CMIP-based) and conditioned (ACCESS-MICAS) methods.Pre-defining language ahead of events would help mitigate the communication problem.See Hope et al (this issue) for more detail.
The second example is the WWA study on Cyclone Gabrielle, a slow-moving ex-TC which particularly impacted the east coast of New Zealand's North Island, resulting in one the most damaging weather events in NZ history (Harrington et al 2023).This study was triggered by the severity of recorded impacts and intense public interest: 11 people died from the event and estimates by the New Zealand Treasury place the combined economic costs of both Gabrielle and the Auckland Anniversary floods 3 weeks earlier, at between NZ$9 and $14 billion.
As with most WWA analyses, the study of cyclone Gabrielle was done rapidly and employed multiple data sources, including analysis of a suite of high-resolution regional climate model output following evaluation, as well as non-stationary extreme value analysis of historical rainfall data.These different datasets gave different results: the observational analyses suggested a trend of increasing frequency and intensity of heavy rainfall events in the region, while models predicted no trend.The observation-based analysis was hampered by the lack of access to data for Gabrielle from several high-quality, long-running meteorological stations as a result of storm impacts.Nevertheless, a synthesis of the results was presented with key points (e.g. total rainfall over 48 h was as much as 30% more intense in today's climate than in the preindustrial era), but large uncertainties were attached to each of the presented results.The disagreement between the model-based and observation-based results complicated the communication of the key messages from the rapid attribution study.
Subsequent analysis and developments after the rapid study have also raised further issues.Subsequent sensitivity analysis of the observation-based results suggested a trend in high-rainfall extremes was detectable over the region in question, though the magnitude of this trend was smaller when considering only a smaller subset of weather station records with data dating back to the early 1900s.Also, the results of the rapid analysis were compared to those in a previous highly conditioned forecast attribution analyses for other heavy rainfall events around New Zealand by Tradowsky et al (2023) which found clear evidence that the intensity of the rain which fell from these other flood-inducing events was about 10%-15% larger because of anthropogenic climate change.This previous experience lent weight to the messaging which was emphasised following the release of the rapid Gabrielle attribution study: that the amount of rain which fell during the peak of Cyclone Gabrielle was larger than if an equivalent weather system occurred in a pre-industrial world.Also, subsequent analysis after the rapid WWA study using this highly conditioned attribution framework confirmed such intensification estimates to be correct for cyclone Gabrielle too (Stone et al submitted).

Process for conducting studies
We develop a flow chart for current best practice for producing a rapid EEA statement (figure 1) which draws on the 8-step WWA pathway of van Oldenborgh et al (2021a), and from Philip et al (2020), with modifications to adapt it to a national context and to extend to impacts.This is presented as a working draft of the aim for an operational EEA process.Breaking the process into preset phases, with stage gates (decision points between phases with preset criteria and decided by clear project governance) are useful for the goal of operationalising EEA.
Comprehensive inventories of all events are being pursued, but for operational and rapid studies there must be a choice of which events to study.Decision criteria for the 'trigger' for the event are: (1) climatological rarity (outside of 'normality'), which requires context-setting within high quality observer records; (2) impacts such as property damage and casualties for real-world or public interest and with a desire to improve future responses (not always consistent with 1); and (3) potential for gaining scientific understanding.The likelihood that the EEA study will find a non-null result should not be considered, as this will bias the sampling of events considered.In terms of impacts, WWA thresholds are a declaration of local state of emergency, ⩾100 lives lost or ⩾1000 000 people affected or ⩾50% of total population affected, or else events with a Red Cross international funding appeal.The choice is a little different for studies of national importance compared to the international remit of WWA, so the criteria and quantitative thresholds used in WWA are not the only consideration here.National policy relevance, national level liability claims and adaptation priorities are also relevant, but care must be taken not to politicize the selection process (e.g.deliberately not selecting events with large liability questions, or events that may reveal maladaptation to climate change due to pressure from political interests).Studies from a national operational program also lack the goal of WWA to perform EEA in regions that currently lack studies.Currently, the process in ANZ is ad hoc and for a truly operational service a more objective set of thresholds is required.
Once the trigger is met, event definition must be done carefully and can be the most important part of the process, as differences in the definition give very different messages (e.g.Otto et al 2012, Angélil et al 2014).This includes the variable or index used, but the space and time bounds are especially important (Leach et al 2020).Choices should be informed by the purpose of the study and the needs of the various parties who might fint the EEA messages useful.
Once defined, the next stage gate relies on whether performing the study is achievable (e.g.tools and models are fit for purpose), the study is important to do and contributes to our knowledge.A flow chart covering the assessment of the available tools, knowledge and models is described in Lane et al (2023).Like projections science, decisions for the stage gate on model evaluation in EEA often relies on subjective assessment of model performance in areas such as the seasonal cycle and spatial patterns in the model compared to observations.There is value in moving towards a more objective benchmarking framework, as in the projections space (e.g.Isphording et al 2024).
If the trigger is hit, the event is well defined and the study is achievable, then running multiple methods and models is always of course preferred over a single method (as further supported by recent experience outlined in section 3).
We suggest that effective stage gates are crucial to make an operational practice effective and manageable.Once the trigger has been met, forming the team, and proceeding to event definition should always occur, but at that point the first stage is essential.A second stage gate on extending to impacts and costs should be applied (as noted).These practical steps and flows can be used to guide opportunistic studies, or for rapid or operational work.

Using causal networks
As part of the event definition stage of the process (figure 1), the preconditions, modulators and drivers need to be assessed.A mapping of cascading effects through physical and/or socioeconomic networks is referred to as a causal network (Doblas-Reyes et al 2021, Shepherd andLloyd 2021, Kunimitsu et al 2023) and is a useful tool in EEA.The network can include exposure, vulnerability, and impacts, so is useful in impact attribution (see next section).But it is also useful to make an explicit mapping of the influence of climate variability and climate change through stages of preconditions and proximate causes through to a weather and climate event for EEA (figure 2).Here we suggest that producing a diagram should be standard practice and we present some suggested procedures and principles for this usage.
To produce credible messages and to assess how complete a message can be, the various 'ingredients' need to be examined and then combined into an overall statement that accounts for all the inputs and how they were combined (in linear and non-linear ways).Both the preconditions and proximate causes for the event should be assessed in terms of (1).Direction of influence; (2).Magnitude of influence (relative to the total of event) and (3).Confidence in both.
Two example case studies are shown (figures 3 and 4) that vary in confidence.Both examples show a mapping of all relevant links that are colour-coded for climate variability (black) and climate change (red), either as solid (higher confidence) and dashed (lower confidence).Both magnitude and confidence in an influence are displayed for key links in the diagram using a dial graphic, adapted for this purpose from King et al (2023).
The first example is a simpler and more confident example of the Black Saturday heat wave (figure 3) that occurred in southeast Australia in February 2009.There is higher confidence attached to this example found by published EEA studies on the event (Hope et al 2022, Abhik et al 2023), and given we have a clear understanding of the synoptic setup of the event (i.e.blocking high in the Tasman Sea drawing heat across the continent).The second example is the more complex and less confident example of TCs (figure 4).In this case, we compare two events with contrasting impacts: the coastal flood driven by storm surge plus river flood associated with Severe TC Marcia in February 2015, and the heavy rainfall and riverine flood (no coastal flood) associated with TC Jasper in December 2023.Neither TC is the subject of an EEA study yet.
For the Black Saturday heatwave, a highly confident risk-based attribution statement (assuming 'all else being equal') was possible through shift-stats methods: more than two times more likely in the current climate than a world without human influence.However, there is a potential role of climate change on the dynamics of the weather, and on the drivers of seasonal climate variability such as the E Niño Southern Oscillation (ENSO) that should ideally be acknowledged and accounted for.A FAR based on unconditioned methods (suggesting a value of 68%) and results from conditioned methods both support an attribution of the mean state and weather of 3 • C-4 • C extra warmth, and that ENSO and SAM also contributed but only slightly (Hope et al 2022).It was acknowledged that climate change may have affected ENSO itself, but the effect is less clear but considered likely to be second order in this context.
Overall EEA messages for this extreme event can be quite simply formulated, with the direction of change and confidence in the direction of change well understood.The main decisions are whether to report the risk based and/or the storyline result, then which value to report and with what precision.This contrasts strongly with the TC examples covered next.
Only preliminary EEA analysis has been performed on one storm (Jasper), so here we note how a causal network used in the planning stages of EEA studies.The impacts from these two TCs differ strongly, TC Marcia mainly through coastal floods and wind damage and TC Jasper mainly through inland flooding.Statements regarding this class of event may be possible, with commentary about the attributable trends in relevant factors, including sea level rise, a warmer atmosphere, changes to TC likelihood and intensity and seasonality.However, for EEA statements for these events specifically there are more barriers.Only some of the 'ingredients' of these hazards can be attributed with high confidence, meaning EEA messages would likely be low in confidence as well as incomplete or partial.
For TC Marcia, the effect of climate change on causing >20 cm higher sea level alone could be the subject of a conditional/partial EEA statement, as done for Superstorm Sandy (Strauss et al 2021), and additional commentary on the general effect of a warmer atmosphere on rain rates could be made (separate from dynamical effects of the TC itself).However, the effect of climate change on the TC itself (e.g.TC intensity, structure, and track) is an important, first-order consideration, and changes in these factors due to climate change are less clear (Knutson 2019(Knutson , 2020)).
Even for a conditional statement about higher sea level, there are various pitfalls.The dynamical response of TCs to climate change may only add to the effect of sea level rise, for example where climate change made the TC rapidly intensify before hitting land, or move slower, then this may add to the coastal inundation.In this case, a partial statement based only on sea level could be confidently described as minimum or 'conservative' (in terms of being a likely low estimate).However, the effect of climate change on the dynamics of the TC may conceivably offset the effect of sea level, meaning the conditional statement may not be a likely low estimate.Instead, a comment about the effect of higher sea levels in general may be appropriate, avoiding comments on the specific event.
Turning to TC Jasper, we find similarities and differences with TC Marcia.TC Jasper was a TC off the coast of Queensland, which then weakened as it crossed the coast to sub-cyclone status.However, the resulting tropical low stalled in place and brought near record rainfall and floods to the region over a 5 day period (Emanuel 2024).The main impact was through riverine flooding.Without the importance of sea level rise as an 'ingredient' , statements on the effect of climate change are even more difficult for Jasper than Marcia.Here the only 'ingredient' with high confidence is higher rain rates from a warmer atmosphere, which was found in an initial rapid attribution analysis of the event.However, even with the expectation of higher rain rates based on theory and modelling, there has not yet been a detectable trend in observed TC rainfall rates in the Australian region so current TC attribution statements are almost entirely model based ('attribution without detection' , e.g.Knutson 2019).The possibility of producing a storyline EEA message with any usefulness depends on the ability to gauge the effect of climate change on the TC itself, including a view of the overall effect on likelihood, as well as other aspects of the TC including: • Intensity (peak wind speeds or minimum pressure) • Genesis, rate of intensification, and translation speed • Latitude and other aspects of the track, including the turn before hitting the coast by TC Marcia, and the stalling of TC Jasper • Dynamical aspects of rain rate Unless the model systems used for EEA can produce confident findings on some set of these potential factors, then any EEA messages around TC would be tentative at best.

Attributing impacts
The field of 'impact attribution' enables the impacts (not just the hazard) of extreme events and slow-onset climate change to be attributed (Perkins-Kirkpatrick et al 2022), including costs (e.g.damage to infrastructure) but other impacts as well.For example, extreme heat EEA can be extended into the health impacts of that heat, not just for excess deaths but for a more holistic view including morbidity, hospital admissions, ambulance call outs, aggravation of underlying diseases, fetal and maternal outcomes.It can be useful to extend the causal network concept described above further into the impact and response space, to account for the interactions and feedbacks (e.g.Shepherd and Lloyd 2021).
Attributing impacts to climate change makes EEA statements more salient by drawing out policy implications more clearly, reflected in the demands from Australian government decision-makers surveyed in Machin and Bourbon (2023).Determining EEA of impacts and illustrates the extra level of response now required compared to the past and has implications for future level of response needed, including for: • Infrastructure and planning-including implications for building codes, building design (e.g.Zhang et al 2024).• Human health-adaptation measures and health care to manage impacts of heat and other extremes, e.g.ambulance providers can use an estimate of how many additional ambulances were required during a heatwave or fire event for planning future operations to meet their legislated standard of service.• Emergency response-scope and scale required to address emergencies, e.g.firefighting including national and international resourcing.• Biodiversity, socioeconomic and flow-on or secondary impacts (e.g.vector-borne diseases).
Hence, this extension should be built into the operational flow chart (see above) and performed whenever possible.A prominent recent impact attribution study is of droughts and floods in New Zealand in 2007-2017 (Frame et al 2020a), and there are a few lessons learned from this experience, as well as from the TC Gabrielle study (Harrington et al 2023).Firstly, it is valuable to incorporate multidisciplinary expertise in relevant exposure and vulnerability issues into the study team, ideally through institutional arrangements between relevant organisations.Institutional arrangements should be sensitive to the pressures and demands of each organisation, for example met services, health and other experts are occupied supporting emergency responders during the event, preventing their input to rapid studies.
As a fallback to accessing multidisciplinary expertise, EEA practitioners can use 'off the shelf ' numbers of exposure and damages.Estimates of economic losses are generally available, but increased granularity would mean more nuanced messages can be derived, as was done for Hurricane Harvey (Frame et al 2020, Wehner and Sampson 2021, Smiley et al 2022).Timely release of exposure, vulnerability and impact data enables impact attribution to occur without too much time lag, however there are barriers to this including from processes across state boundaries.Also, some time lag can be unavoidable as it takes time for the full impacts to become known (e.g.full health effects or even socio-economic impacts).For example, after the 2022 flooding event across Queensland and New South Wales, it took 3 months for Insurance Australia to report initial estimates on the insured losses from the event (Insurance Council of Australia 2022) and for a report by the QLD Reconstruction Authority on the damages (Deloitte Access Economics 2022).For these reasons, impact analysis can only be done on non-rapid studies, and EEA services should be aware of the relevant time horizons.Even commentary on impacts in real time must be done with extreme caution, and only qualitatively.
The attribution of impacts is not a simple extension of the meteorological event, with impact attribution having its own features and each type of impact having its own consideration.Indeed, Perkins-Kirkpatrick et al (2022) argue that in every case, impact attribution must be separated from the casual event, regardless of how straightforward it may or may not be to derive the impact.Moreover, it must be kept in mind that additional uncertainties-on top of those that already exist in EEA methods-are introduced with the estimation of impacts.These might include how the impact itself is measured and whether it is one of multiple plausible methods, how the impact method is fitted to climatological data, any assumptions about how non-climatological variables might change over time (e.g.population, human behaviour, land use change), and whether the right attribution method (e.g.FAR, storylines, a 'sliver' approach; Perkins-Kirkpatrick et al 2022) is being employed for the study at hand.All these factors need to be carefully considered on top of the standard EEA decisions when designing an impacts attribution assessment.Moreover, to be done in the most appropriate manner, they must be conducted in close negotiation with specialists of the relative impact, through collaborative arrangements possibly including through forming multisectoral teams of researchers.This can make impacts attribution more challenging (though not impossible), particularly without on-going, coordinated and funded interdisciplinary research efforts.
Moreover, the above considerations in impact attribution compounds uncertainties of the EEA statement, due to inherent assumptions, biases and uncertainties in impact modelling approaches, causing an inflation of uncertainties.This is the case for both top-down (econometrically estimating aggregate impacts on economic output) and bottom-up (quantifying impacts for individual impact channels) approaches that may be used to quantify the impacts of climate change.For example, top-down impact modelling approaches are commonly based on econometric damage functions, which estimate aggregate impacts on economic output.In comparison, bottom-up approaches quantify impacts specifically for individual impact channels (Piontek et al 2021).
Beyond policy implications, impacts attribution provides evidence on the portion of damages that would not have occurred without climate change required for liability discussions (Burger et al 2020).An important aspect of using EEA and impact attribution in a legal context is the minimum threshold of confidence needed, or the 'burden of proof ' in legal terminology (Burger et al 2020, Otto et al 2022b).Extending to attribution of impacts introduces extra considerations that mean the confidence may be lower of the statement may be more conditional compared to EEA.However, EEA with very high confidence in all causal factors is a very high bar that may not be needed as a prerequisite for liability discussions.Commentary and analysis here suggest not only the 'beyond reasonable doubt' but even 'more likely than not' level may meet the bar for civil court cases (Lloyd et al 2021).It has also been proposed that storyline approaches should be acceptable for legal processes (Lloyd and Oreskes 2018).There is also still a role for conditional statements following the 'all else being equal' (or 'ceteris paribus') condition (Lloyd and Shepherd 2020).Also, translation and tailoring of the messages for policymakers may be needed, with an awareness of how the information is used in policy discussions will make the messages more effective.

Communication
Improving various aspects of the EEA procedure can increase the effectiveness of EEA for all its purposes.It is now widely agreed that communication of EEA should aim to be transparent, targeted to the relevant audiences, and follow preset formats wherever possible (van Oldenborgh et al 2021a, Stott and Christidis 2023).It is also widely agreed that including projected future changes in extremes as part of EEA messages adds value to local decision-making (Otto et al 2022a).In practical terms, conclusions and headlines can use the qualitative categories developed by WWA: (1) More likely; (2) Less likely; (3) No effect; and (4) Cannot discern, with communication using pre-set templates at multiple levels: Scientific papers and a report; Key finding/web summary; and press release and briefing (together with a guide for journalists and animated graphics where possible).In section 6 we suggest that extending to impact attribution makes the outputs more salient and policy relevant, so is now valuable.

Confidence
A key decision is whether to communicate messages, and how to handle inconclusive or unclear results.Nearly all attribution studies will be based on some contingent set of assumptions and will contain some ambiguity.The failure to capture and reflect this contingency and ambiguity could have serious ethical implications due to the potential uses of the information in real-world decisions.Additionally, communicating messages poorly or using framing choices that lack understanding of the context-dependent political meanings and particularities that determine policy responses (Lahsen and Ribot 2022) can produce poor outcomes.
To assess confidence, we can draw lessons from previous sections.In section 4 we suggest that effective stage gates of event definition and of model evaluation will prevent studies proceeding if models are not fit for purpose, save wasted effort, and add credibility and trust to studies that are published.If a trigger cannot be demonstrated to be met, if the event definition is unclear, or if the data, models and methods are not fit for purpose, then the confidence in the result will be low, suggesting caution in communicating results.Even if studies do not proceed, it is useful to capture the lessons learned and what further research and understanding would be needed.In section 5 we suggest that a complete and well-defined causal network from climate variability and change to the hazard is an essential tool to inform decisions and effective messaging.If notable causal links cannot be accounted for, this also suggests lowering confidence in any message or not communicating the message, especially where those links may be first order effects on the total, or else carefully constructing the message within the contingency of the study (TC impacts are a notable example in section 5).

Lessons from epidemiology
EEA was strongly inspired by concepts from the field of epidemiology (Allen 2003) and continues to draw ideas as well as explanatory and narrative techniques from it.The archetypal and often-used analogy in risk-based EEA is lung cancer risk in smoking and non-smoking human populations.This compares risks of a climate extreme in a 'population' of Earth climate samples as it is (the 'smoking' group) to a population of Earth climate samples in a counterfactual world where there is no human influence (the 'non-smoking' group).This analogy remains useful and appropriate, and it is useful to draw further lessons between the two fields, here we consider the size of the FAR-risk in EEA, which is equivalent to the strength of association (effect size) in the Bradford-Hill criteria (Hill 1965) in epidemiology.
The communication pathways and potential pitfalls have obvious parallels between EEA with low FAR-risk value and diet studies that have a low effect size (table 1).Messages for both may be of great interest, but overconfident or overly simple messaging about the effect size should be avoided.Poor messaging may in fact end up being damaging to the credibility of the enterprise, as shown through the public discourse around diet studies.Regular red meat consumption was found to be detrimental to health in one study, and this message entered the public consciousness for many years but was not supported by further studies and meta-analysis (Lescinsky et al 2022).Similarly, moderate consumption of red wine was found to be beneficial to health according to one study and was widely reported but subsequently revised (Lombardo et al 2023).Also, consumption of eggs was discouraged for decades due to a perception that it would raise risks associated with cholesterol, but this was similarly discredited.
This suggests that care is needed for EEA and impacts attribution results with a small 'effect size' , and giving no message may be appropriate, since messages entering the public discourse can be oversimplified and then become highly persistent, like the messages about meat, red wine, or eggs in our diet.Conversely, the field of EEA can learn from successful messaging in epidemiology, where EEA of events with large effect sizes such as heatwaves and marine heatwaves can be communicated clearly and strongly and could help drive responses like those to asbestos and tobacco.
While effect size is the most obvious criterion with important parallels between the two fields, there will also be lessons from the other eight in Bradford-Hill in terms of using multi-method approaches, distinguishing between signal and noise, use of multiple lines of evidence and so on.The other eight are: (2).Reproducibility (consistency), (3).Specificity (no other explanation), (4).Temporality (effect follows cause), (5).Dose response (the effect gets stronger with a higher dose), (6).Plausibility (does the result make sense), (7).Coherence (does it agree with other lines of evidence), (8).Analogy (what is the effect of similar factors), and (9).Experiment (experimental evidence backing up the finding).

Conclusions
The principles and processes of EEA to inform multiple purposes are in constant evolution, and recent lessons from ANZ, can contribute to this development.
We suggest that an operational flow chart (figure 1) is useful not only for operational work, but also for opportunistic studies, with particular focus given to triggers, event definition and critical stage gates before running models or proceeding to impact attribution.We also suggest that a causal network diagrams from climate variability and climate change through preconditions and proximate causes to the extreme itself (figures 2-4) is a useful standard tool for all studies, both to assist in designing the study as well as in understanding the completeness and confidence in the communication messages.
Impact attribution is being called for and provides information that is more relevant and more salient for decision makers and liability discussions, so should be pursued whenever possible (section 6).This requires organisational arrangements and networks, as well as multidisciplinary collaboration, to bring together expertise and granular datasets in a timely way.
Communication efforts can take lessons from recent experience, and from the origins of the field in epidemiology.Confident results with a large FAR risk ('effect size') can be communicated clearly and can meet the threshold for liability and legal consideration.Conversely, messages around events with a lower effect size, such as specific TCs or ex-tropical cyclones and their impacts, need very careful attention, and should consider the benefits and risks of communicating anything at all, or the risks of not communicating.

Figure 1 .
Figure 1.Flow chart of proposed recommended processes in selecting, conducting, and communicating extreme event attribution (EEA) in Australia and New Zealand.

Figure 2 .
Figure 2. Generic 'causal network' diagram of influences from climate variability and change on a physical extreme climate event.

Figure 3 .
Figure 3.A causal network diagram for the Black Saturday heatwave 2009, noting the confident links to climate change (solid red), likely or possible links to climate change (dashed red) and links to climate variability (black).The dial graphics show the direction, confidence, and magnitude (relative to the overall extreme) associated with the confident links, where the direction of influence is shown as red arrow, confidence shown on the dial colour scale from low (red) to high (green), and relative magnitude showing the reverse colours (low green to red high).

Figure 4 .
Figure 4.A causal network diagram for (top) the coastal flood caused by TC Marcia and (bottom) the inland flood caused by TC Jasper in Queensland (boxes, arrows and dials as for figure 3).

Table 1 .
Examples of cases with high and low effect size in medical epidemiology and climate change attribution.