Events
The fourteen extreme weather events considered in this paper are listed in Tables 1 and 2. These events all occurred in New Zealand during the decade from mid-2007 to mid-2017. The list is restricted to weather events for which a clear link is established between the meteorological event and damages, and for which anthropogenic climate change attribution analyses are available. Note that the attribution analysis was performed on rainfall, not the actual flood magnitude itself. It was pluvial flooding in these events that led to the damage, as is the standard in the small, high-slope catchments in New Zealand, so precipitation intensity should serve as an accurate proxy. For the floods, the event duration that was used to estimate the FAR was chosen to be the longest period over which the rainfall was extreme, since the damages accrue from the full pulse of excess water. Drought durations vary, and last several months.
Table 1 Insured damages associated with flooding events (resulting from extreme rainfall) in New Zealand over the period 2007–2017 for the 12 costliest events where it has been possible to form estimates regarding the FAR of the rainfall event. Estimates of insurance costs for each event are in the column second from right in CPI-adjusted 2017 NZ$. These data were collected from the private insurers operating in New Zealand, so only include the types of assets that are typically insured, and only where insurance was actually purchased. These mostly include residential housing and privately owned commercial property. It does not include publicly owned property (such as council buildings) and infrastructure such as roads. The right-most column contains the insured losses attributable to climate change, i.e. the insurance loss estimate multiplied by the FAR of the event. The insured cost estimates are from the Insurance Council of New Zealand; no uncertainty estimates are available (obtained on 17/05/19). Full economic costs are expected to be far higher. See text for discussion Table 2 Financial losses associated with the two costliest droughts (in CPI-adjusted 2017 NZ$) that occurred in New Zealand during the period 2007–2017, along with estimated FAR of these events. Estimates of economic costs for each event are in the column second from right, and are based on economic computable general equilibrium modelling done by the New Zealand government, and the right-most column shows the economic losses attributable to climate change, i.e. the economic loss estimate multiplied by the FAR of the event. Note that the quantities are different from those used in estimating the costs associated with floods (insured losses). In this case, for droughts, the losses represent a more comprehensive assessment of impact on the economy. Along with those for floods (Table 1), these cost estimates are also uncertain. See text for discussion Climate modelling experiments
For the flood events, we assume a step-change damage function, where no damage occurs except when precipitation exceeds a specified threshold, with a fixed amount of damage then occurring irrespective of how much the precipitation exceeds that threshold. That threshold is defined as the amount that occurred during each of the identified events. Strictly speaking, we should use delta functions, i.e. the observed amount of damage only occurs for the observed event magnitude (Harrington 2017). However, we adopt the step-change functions because it corresponds to the most common method of calculating FARs and because it affects the FAR estimates calculated in our study by only about 10%. Climate station rainfall observations from NIWA’s National Climate Database (NIWA 2017) are used, provided there is at least 40 years of observations available. In most cases, these were continuous 40-year periods and exclusion of locations with one or more missing years made negligible difference to the results. At each observing location impacted by the event, extreme value theory is used to estimate an annual exceedance probability as described below in Section 3.5. The event duration was chosen to be the longest period over which the rainfall was extreme, since the damages accrue across the period over which the rain falls. In a New Zealand context, two points are worth noting: first, catchments are small, so floods are usually pluvial events; second, it is not usually just the most extreme peak that does the damage, but the full event.
The basis for our inferences about extreme precipitation comes from extremely large ensembles of simulations of a regional climate model from the “weather@home” experiment (Massey et al. 2015) for the Australia/New Zealand (ANZ) region (Black et al. 2016). In this setup, the Met Office HadAM3P model (Pope et al. 2000) is run globally, providing lateral boundary conditions to the regional model HadRM3P (Jones et al. 2004), which covers a domain spanning Australia, New Zealand and much of Indonesia at approximately 50-km resolution (Rosier et al. 2015; Black et al. 2016). The model has been shown to reproduce many large- and medium-scale features of weather and climate with a good degree of accuracy (Black et al. 2016). In particular, model rainfalls are captured well, albeit with some dry bias in the extremes, and associated synoptic conditions, such as atmospheric rivers, are also realistically simulated (Rosier et al. 2015).
For this study, we use very large initial condition ensembles to represent many different realizations of possible weather under varying climate conditions. Two climate forcing ensembles are compared: (1) all forcings (ALL), which includes sea surface temperatures (SSTs) from the “Operational Sea Surface Temperature and Sea Ice Analysis” (Donlon et al. 2012), as well as greenhouse gases, aerosol, ozone, solar, and volcanic forcings; (2) “natural forcings only” (NAT) where the anthropogenic contribution to the forcings has been removed; this produces many realisations of the weather possible in a hypothetical world without anthropogenic influence on climate. Estimates of what SST fields might have looked like in the NAT world were constructed using SST changes (“delta SSTs”) taken from different global coupled models from the Coupled Model Intercomparison Project Phase 5 (CMIP5) (Taylor et al. 2012), using results from “HistoricalNat” runs subtracted from “Historical” runs. These delta SSTs were then subtracted from the SST fields for the year in question of simulation. To try to span the uncertainty in this estimation, delta SSTs from several different GCMs were used, together with a multi-model mean. The setup is very similar to that described by Schaller et al. (2016). Other forcings such as greenhouse gas concentrations, ozone and aerosols are set to pre-industrial levels, as described in Schaller et al. (2016) and Black et al. (2016). Paired ensembles of “ALL” and “NAT” forced simulations were available for the years 2013, 2014 and 2015.
This basic experiment setup has been used in many pieces of recent climate research, and is a recognised part of the climate change detection and attribution literature; interested readers should see Bindoff et al. (2013), or the recent review article by Stott et al. (2016).
The study’s assessments of droughts are based primarily on Harrington et al. (2016), in which the CMIP5 ensemble of models (Taylor et al. 2012) was investigated using self-organising maps (Hewitson and Crane 2002; Gibson et al. 2017) to estimate FAR for such events.
Costs
Cost estimates used here are for the episodic floods and droughts, and ignore chronic costs resulting from gradual climate trends or trends in nuisance flooding. As we discussed earlier, the cost figures for floods account only for direct insured damages, while the cost estimates for droughts measure the annual indirect loss. The figures for insured damages associated with floods are from the Insurance Council of New Zealand (2017). These flooding costs represent a significant underestimate of the full financial and economic impacts of these rainfall events as they do not include losses in economic activity in the aftermath of the events, nor emergency response costs that prevented damage to insured properties. The estimates of economic losses for the two droughts are from government estimates (Ministry of Agriculture and Fisheries 2009; Treasury 2013).
The cost figures for insured damages from floods are reliably measured, as the ICNZ data include the actual amount of payments made by all the ICNZ member companies (ICNZ membership accounts for maybe 95% of overall commercial insurance available in New Zealand). However, while precise, these figures under-state the total economic impact (including both loss and damage) associated with these events. Even the damages under-represent the full extent of damages from extreme rainfall, as they do not include uninsured damages. Uninsured damage includes damage to transportation, communication and utility networks, which are typically uninsured, public and community-owned buildings which are almost never insured commercially, and those commercial and residential buildings that were not under current insurance cover at the time of the event. This latter category is not very large, as insurance penetration rates in New Zealand are quite high (more than 95% for residential buildings, and 70–75% for commercial ones). In addition, the measure of damage (insured and uninsured) does not include economic losses, such as lost production because of infrastructure failures, the costs associated with medical and pastoral care of affected communities, and the costs of disaster relief and clean-up.
The costs associated with droughts, in contrast, are estimates obtained from modelling of the indirect economic losses. These will be very sensitive to the modelling assumptions that are used, and to the exact specification of the shock with which the model is perturbed. Two modelled estimates of the cost of the 2013 drought are available, and are fairly similar (0.6 and 0.7% of annual GDP).
As noted, direct damages and indirect losses are not the same (the former is a flow, the latter is a stock measure) and the two are therefore not comparable. In this specific case, the indirect loss, as estimated for droughts, is probably the more comprehensive figure, as the extent of direct damage to assets from drought is fairly minimal. Nonetheless, the drought cost figures are also much more speculative as they are modelled rather than counted. So, in this case, we observe a tradeoff between a more comprehensive estimate (droughts) and a more accurate one (floods).
The absence of accurate assessments of economic costs associated with the extreme events, including both damages and losses, is probably the largest source of uncertainty in this report. In the absence of full and accurate accounting of economic damages and losses associated with extreme rainfall events, we have drawn on insured losses as the most readily available data. Fuller analysis of the economic impact of weather-related disasters that will tally the range of both damages and losses is required for a better understanding of the total impact of extreme weather, in New Zealand and elsewhere. These estimates of complete costs (damages and losses) are not available consistently anywhere else, either, so their collection or estimation should be a priority in furthering the research agenda outlined in this paper.
Fraction of attributable risk
In estimating the risk of events attributable to anthropogenic climate change we use the “Fraction Attributable Risk” metric, defined as
$$ \mathrm{FAR}=1-{P}_0/{P}_1 $$
where P0 is the probability of an event occurring in the absence of human influence on climate, and P1 the corresponding probability in a world in which human influence is included. While strictly speaking for this study P0 and P1 should be the likelihoods of a certain event magnitude (Harrington 2017), here we use exceedance probabilities because of their established usage in event attribution studies, their better sampling, and their only small bias (about a 10% overestimate for the events in this study). FAR is thus the fraction of the risk that is attributable to human influence (Bindoff et al. 2013). An “event” in this context occurs when a specified threshold is exceeded for some measure of interest, such as cumulative rainfall over a 3-day period. In the rainfall examples in this study, the synoptic conditions for the rainfall events identified within the weather@home/ANZ simulations were found to be consistent with the events that actually occurred, meaning that the simulated event produced the event-above-threshold in a manner physically consistent with observations (Rosier et al. 2015). To calculate extreme rainfall FARs, we follow the practices established in Pall et al. (2011), and, in Australasian contexts, Rosier et al. (2015) and Black et al. (2016). The estimates of FAR in this study are, for the most part, indicative estimates based on the current state of knowledge aimed at providing an approximate order of magnitude estimate of the costs of current climate change. It is acknowledged that, as a metric, FAR has some limitations. Christiansen (2015) has shown, for example, that as the extreme high end of the distribution is approached, FAR tends towards one for Gaussian-distributed variables but towards zero for heavy-tailed distributions. Our modelled rainfall maxima distributions tend to be heavy-tailed, and the interpretation of what could well be underestimated values of FAR is not straightforward. A different measure of likelihood change, such as the risk ratio, might actually be more interpretable; however, we found FAR to be a useful metric in the context of this study, especially given the high level of uncertainty in many of the estimates used here (in particular, the costs).
While in general extreme precipitation at midlatitudes increases with the thermodynamic contribution from the Clausius-Clapeyron relationship (Pall et al. 2011), recent research (Risser et al. 2019, in prep) has demonstrated that changes in weather patterns associated with extremes can compete with this effect. In some instances, at some locations, the net effect can be a reduction in extreme precipitation (Pfahl et al. 2017). This would imply a FAR < 0, i.e. where the risk of extreme events in some places has been reduced because of human influence. There is nothing in our framework to prevent the inclusion of events which have become less likely as a result of human influence on the climate, i.e. events associated with reduced rather than increased costs (but with the Fraction of Attributable Decrease in Risk measure, (Wolski et al. 2014)). However, because our analysis is based on events that have occurred, we have an expected selection bias against events that have not occurred because they are less likely under anthropogenic forcing (i.e. FAR < 0). We only examine two types of events in this study though, and calculate similar moderate FAR values for all wet events and similar moderate FAR values for both dry events. This suggests that the issue of a selection bias is not important for the types of events considered here. Nevertheless, further work is needed to investigate the importance of this selection bias in New Zealand when considering a more expansive list of event types (cold events in particular). In this paper, we focused on the most damaging rainfall events from an insurance perspective, without prior consideration of whether FARs might be positive or negative, or near zero.
Floods
We multiply the FAR by the cost estimate to obtain attributable costs. For some of the events considered, we have simulations for the SST patterns that characterize the year in question. Resourcing constraints mean that we only had three full years of simulations on which to draw—the years 2013 to 2015. As a result, we have restricted the analysis to considering only the impact of climate change on events to the 2007–2017 period, on the basis that these three ensembles provide a reasonable starting point for the quantification of FARs in the context of New Zealand’s climate (Risser et al. 2017).
Observed rainfall was analysed at stations within the catchment of a flooding event for three event durations: 1 day, 3 days and 5 days (none of the events lasted more than 5 days). All events involved heavy precipitation immediately before the flooding, which is generally the case for flooding events in New Zealand. The rainfall associated with each flooding event is the largest over the recent 14-year period for the relevant area and storm duration. The event probability was estimated for each station after fitting a generalized extreme value (GEV) distribution to each station annual maxima series for the three durations. This was done separately for each season where the annual maxima series was based on rainfall observed within that season. The annual exceedance probability for each event was determined at nearby stations based on the fitted GEV distribution for the relevant season. The event probability was calculated as the average over the chosen stations and the duration selected as the longest period over which the event was extreme. Finally, we focused the analysis on the most financially significant flooding events over the period July 2007–June 2017, considering only those events with insurance losses in excess of an arbitrary threshold of 15M in NZ$ (CPI-inflation-adjusted to 2017 NZ$).
In assessing the FARs of each event, we first ensured that the modelled FARs validated towards the tail of the distribution—this is to reflect the uncertainty in estimating the rarity of the actual event. We examined FAR maps for extreme precipitation events at each integer threshold across 1–4% annual exceedance probabilities inclusive, for each season. We then determined a FAR value for each event by examining maps of FAR at the location and percentile determined from the observations, for the appropriate season in which the event occurred, and using the map of FAR computed for the appropriate rainfall duration (1, 3 or 5 days). Where we had the explicit year of modelling for an event (i.e. for events in 2013, 2014 or 2015), we used the model ensembles from the year in question to compute FAR; where we did not, we used results from the 3 years of modelling (2013–2015) pooled.
Based on the FARs presented in Table 1, we estimate that major-flood insured costs attributable to anthropogenic influence on climate are currently somewhere in the vicinity of $140M for this decade. This is what anthropogenic climate change cost the insurance sector in these events. This figure of $140M is understated, as it does not include the attributable costs of more minor floods, and these costs will very likely increase over time, given our finding that virtually all rainfall FARs were substantially positive. As stated earlier, nor does it include all the damages and losses associated with these floods that were not insured. Further investigation into additional smaller flooding events, and extension of the analysis framework to include storm damage and flooding associated with cyclones (especially the four tropical cyclone remnants that hit New Zealand in 2017–2018), an accounting of uninsured damages, and estimates of indirect economic losses would almost certainly represent a large net increase over these numbers.
Droughts
The drought FARs are based primarily on Harrington et al. (2016), in which the CMIP5 ensemble of models (Taylor et al. 2012) was investigated using self-organising maps (Hewitson and Crane 2002; Gibson et al. 2017). Two droughts are considered, one in 2007/2008 and one in 2012/2013. In the case of the latter, estimates of the FAR depend considerably on how the drought is characterized, and this yields a range of plausible FARs, from 40% (using monthly pressure anomalies or the frequency of blocking high pressure systems) to 15% (using precipitation deficits or daily circulation properties). There is a robust anthropogenic increase in the likelihood of observing those SOM nodes which occurred frequently during the 2013 drought. In addition, the SOM approach in Harrington et al. (2016) also represents a method of circumventing difficulties in evaluating anthropogenic changes in drought likelihood, particularly for those locations where precipitation-temperature coupling mechanisms are less significant. Since the overall FAR of the drought will be some (uncertain) combination of the individual FAR components, the choice of a low but plausible FAR of 20% represents a conservative choice.
The 2007/2008 drought occurred on the back of a significant El Niño event. Contributions from El Niño events may not add linearly with contributions from anthropogenic climate change. Without a full study of 2007/2008 conditions, we cannot be sure that the anthropogenic contribution is similar in the 2007/2008 and 2012/2013 cases, since the relative frequency of different synoptic patterns contributing to the drought may be different in El Niño years compared with neutral years and compared with La Niña years (Risser et al. 2017). Joint effects could well be additive, increasing the FAR, since there are reasons to believe that the frequency of ENSO conditions is increasing as a result of climate change (Wang et al. 2017). In this study, we have chosen a moderate value of 15%, but higher values are highly plausible.
Costs associated with the 2012/2013 drought have been estimated at NZ$1.5 billion by the New Zealand Treasury based on a reduced growth in gross domestic product, compared with a hypothetical year without drought (Treasury 2013). With a FAR of 20%, this yields excess costs of the drought due to anthropogenic climate change of NZ$300M. We note that the two drought events analysed were not the only droughts occurring in any part of New Zealand during this decade (2007–2017). We focus on these two events as these events are the only ones for which there is an available quantification of their economic losses. Other extreme weather or extreme weather–related hazards that might have FAR ≠ 0 associated with them—such as storm (damage from winds and storm surges), hailstorms, wildfires, frosts or tornadoes—were all omitted from the analysis because no peer-reviewed attribution studies have yet been conducted for these events occurring in New Zealand.
While developing a simple estimate of the insured losses from floods and the economic costs of droughts which are associated with climate change, the study is not comprehensive, and represents a significant underestimate of the full range of losses of climate change in New Zealand. At the end of the paper, we describe the prospects for including other hydrometeorological events within the framework, as well as secondary systems and cascading impacts.
Treatment of uncertainties
The economic impact cost estimates and FAR estimates in this report are uncertain, with uncertainties around cost probably being the larger of the two. Uncertainty estimates for FARs are presented in the table. Uncertainties associated with the heavy precipitation FARs, estimated by examining the FAR maps for the appropriate location, season and percentile, were estimated as usually likely to be around ± 0.2. In some cases—notably the Otago flood event, which is the subject of a paper in preparation—this means that the FARs even have some considerable likelihood of being negative, i.e. there is a likelihood that the event in question has become less likely because of climate change, even though the balance of evidence suggests certain aspects within the event (namely the intensity of rainfall) were likely made more severe because of climate change. Uncertainty in the drought FARs was estimated by attempting to combine information from Harrington et al. (2016), who examined various metrics (e.g. high surface pressure, low precipitation) individually. A combination of these influences would likely lead to higher FARs than is implied by each component individually; this in part contributes to the uncertainty estimate (Table 2) being asymmetric. The uncertainties about economic impact costs arise first and foremost because in both cases (droughts and floods), we only have a quantification of a part of the overall impacts (economic losses and insured damages, respectively). In addition, the quantities we do use are estimates of the true quantities, and the estimates of economic losses for droughts in particular are subject to substantial uncertainties around them. In fact, the Reserve Bank of New Zealand estimated that the economic loss of the 2013 drought event could be twice the magnitude we used in our analysis here (Kamber et al. 2013).
An additional uncertainty arises through our use of a rainfall, as against flooding, attribution. The scale of flooding clearly depends on other factors, such as antecedent soil moisture conditions, as well as rainfall. In New Zealand, it does seem a reasonable assumption that rainfall could provide an acceptable proxy for flooding in assessing the attribution results; however, it is acknowledged that we do not currently know the scale of the uncertainty introduced by this approximation. Future work, requiring substantially increased resources, is necessary to investigate this.