Today, I discuss the concept of the “time of emergence” for the detection of a signal of a change in climate in observations and projections. Our early work in this area led to some surprising results (at least to me!) and profoundly shaped how I think about the detection and attribution of changes in the statistics of extreme weather.
In 2010, Bender et al. explored how the incidence of Category 4 and 5 Atlantic hurricanes might change to 2100 under different climate projections. They found that an ensemble mean of 18 projections resulted in an overall decrease in storms but a whopping 81% increase in the frequency of Category 4 and 5 hurricanes from 2020 to 2100.
That projected increase is far larger than assessed by the most recent report of the Intergovernmental Panel on Climate Change (IPCC), which suggested little or no change in category 4 and 5 tropical cyclones globally to 2100: (1)
For a 2°C global warming, the median proportion of Category 4–5 TCs increases by 13%, while the median global TC frequency decreases by 14%, which implies that the median of the global Category 4–5 TC frequency is slightly reduced by 1% or almost unchanged.
Bender et al. 2010 used our research on normalized hurricane losses to estimate that their projected changes to Atlantic hurricane incidence would result in a 30% increase in U.S. damage potential, with decreasing losses from fewer weaker hurricanes offset by the increasing losses from the large increase in the strongest storms.
They also estimated that due to the relative infrequency of Category 4 and 5 storms in the context of large variability it would take 60 or more years before the signal of the projected change would emerge from the background of variability. When the signal of a change in climate becomes detectable is called the “time of emergence.” Over the past decade or so, the notion of the “time of emergence” has motivated a significant literature and an emphasis in the most recent IPCC report.
The IPCC AR6 defines the “time of emergence” as follows:
The emergence of a climate change signal occurs when that signal exceeds some critical threshold (usually taken to be a measure of natural variability; see for example, Hawkins and Sutton, 2012) or when the probability distribution of an indicator becomes significantly different to that over a reference period (e.g., Chadwick et al., 2019; see also Chapter 10 and Section 1.4.2), in which case external anthropogenic forcings can be detected as causal factors. The ‘time of emergence’ (ToE) or ‘temperature of emergence’ is the time or global warming level thresholds associated with this exceedance. Emergence is particularly relevant to impacts, risk assessment and adaptation because human and natural systems are largely adapted to natural variability but may be vulnerable if exposed to changes that go beyond this variability range; this is not to say that changes within natural variability have no impact, as occurrence of damaging extremes proves. Emergence also informs the timing of adaptation measures. The emergence of a change is always relative to a reference period (e.g., the pre-industrial period or a recent past), depending on the framing question. In the former case, the goal is to estimate the amplitude of an anthropogenically driven change while in the latter, it is to estimate the amplitude of change relative to a baseline that is familiar to stakeholders.
In a 2011 paper, we — Ryan Crompton, John McAneney and I — used Bender at al. 2010 as a starting point for an analysis of the time of emergence of the signal of change in hurricane loss data. We asked:
If changes in storm characteristics in fact occur as projected [by Bender et al. 2010], then on what timescale might we expect to detect these effects of those changes in damage data?
We should expect the time of emergence of a signal of change in damage data to take longer than for climate data, because damages introduce more complexities into a time series — notably where a storm makes landfall and the characteristics of exposed loss potential it encounters.
In our 2011 paper we constructed many projections of future U.S. hurricane losses, sampling the distributions of future storms from Bender et al. 2010 and then combining those futures with a loss distribution based on the 106-year normalized loss time series. We then identified the time of emergence from the resulting projected time series of damages — at 90%, 95% and 99% confidence levels.(2)
Our results are shown in the table below.
There are several things to unpack.
First, we were surprised at the results. The shortest emergence timescale was 120 years and the longest was 550 years (at 95% confidence, subtract 40 years for 90%). In other words, if we were to take the projections of Bender et al. 2010 as true — of an 81% increase in Category 4 and 5 storms to 2100 — then it would take more than 200 years to meet the threshold of detection under the IPCC framework for detection and attribution.
Second, not all individual models in our study projected increasing intense hurricanes. Two of the four individual models projected decreases, and detection of decreases also takes a very long time. In either case the large emergence timescales result from the projected changes being relatively small in comparison to documented variability.
Third, for the smallest change (under the HadCM3 model), the emergence timescale is more than 500 years. Note that change — a decrease in damage potential of 9% — is much smaller than the other model projections, and smaller projected changes imply longer emergence timescales. The IPCC AR6 projection of a 1% or no decrease in the global incidence of Category 4 and 5 storms would never be detectable in observational data.
Our results were apparently also a surprise to the broader community. Kerry Emanuel, of MIT, replicated our study by utilizing a set of four different climate model projections for future hurricane incidence from his research (Emanuel et al. 2008). These models projected much larger future changes than did Bender et al. 2010. Even so, Emanuel 2011 confirmed our results, finding from only one model a timescale of emergence of less than 100 years (the values for the four models are 40, 113, 170, and >200 years).
Taken together, these studies arrive at some important results:
- Even assuming very large changes in future hurricane incidence, we should not expect a signal of change to be detectable beyond climate variability under the IPCC framework in either hurricane incidence or in hurricane damage in 2024, 2054, or even 2094.
- Of course, a signal that has not emerged but is assumed to exist, still exists, but at a level that is not detectable in observations. Decision makers who finely judge risks will have to decide the practical importance of a signal that may exist but is not detectable.Source: Knutson et al. 2020
- There exist a very wide range of projections for the future behavior of tropical cyclones (see figure above from Knutson et al. 2020). If you tell me what result you want, I can find you a study in support of that result. The very long emergence timescales mean that it will be very difficult, perhaps impossible, to identify with evolving experience which projections may be more accurate than others.(3)
- The most recent IPCC does not project large (or really any) changes in the incidence of Category 4 and 5 hurricanes to 2100. Therefore, we should expect there to be no signal present today in tropical cyclone behavior — detectable or undetectable.
Wouldn’t it be great if the IPCC systematically summarized the expected emergence timescale for historical and projected changes in all types of extreme events?
Well, we are in luck. The IPCC AR6 did exactly that. We will have a look in Part 4.
1 The IPCC AR6 does not single out the North Atlantic basin.
2 You can find all of the methodological details in our paper.
3 There is a common tendency to aggregate a wide range of modeling studies and report the statistical properties of the set of models (as does Kossin et al. 2020 above). The statistics of such “ensembles of opportunity” are often meaningless and apt to mislead as they are not random samples from a population. Expect a full post on this down the road.