Revisiting The Exceptional Pacific North West Heatwave Part 3

Science Paper Demolishes The WWA Rapid Attribution Analysis

A study was published in November 2022. It says this about the World Weather Attribution rapid analysis which the media obligingly picked up on and amplified as ‘the scientific truth’ thereafter (readers need only read the bold):

The World Weather Attribution initiative (WWA, https://www.worldweatherattribution.org/) performed analyses of the PNW heatwave within weeks of the event and made three rapid attribution statements, which were later published in the scientific literature (Philip et al., 2021). First, they stated that the observed temperatures recorded were “virtually impossible without human-caused climate change.” Second, after estimating that the observed temperatures had a return period of approximately 1,000 years, they stated that such annual maximum daily maximum temperatures (TXx) “would have been at least 150 times rarer without human-induced climate change.” Third, they went on to state that “this heatwave was about 2°C hotter than a 1 in 1,000-year heatwave that at the beginning of the industrial revolution would have been.”

Philip et al. (2021) found that the 2021 spatially averaged temperatures from the ERA5 reanalysis (Hersbach et al., 2020) exceeded the upper bound of an out of sample (not including 2021) non-stationary generalized extreme value (GEV) distribution fit to data from 1950 to 2020. They then took a practical approach while acknowledging the limitations and included the 2021 values, estimating that the current return period of the PNW heatwave was about 1,000 years. Comparing this return period to that obtained under preindustrial temperatures, they concluded that the probability of the PNW heatwave was increased by a factor of 390. Further analysis of climate model simulations and their expert judgment caused them to conclude that the probability of the observed temperature was increased by at least 150 as their final synthesis attribution statement.

We repeat this non-stationary GEV analysis on individual station data from 1950 to 2020 instead of averaging over the WWA study region. In each single-station analysis, we use a GEV distribution with a location parameter linearly dependent on a sum-total forcing variable for five well-mixed greenhouse gases to accommodate non-stationarity (e.g., Risser et al. (2022)), which imposes a non-linear time trend in the GEV model (Section S1 in Supporting Information S1).

Figure 2a shows the Bayesian expectation of the upper bound for daily maximum temperatures for the 1950–2020 GHCN station data. Stations where the observed 2021 values exceed the expectation of the upper bound (“+”) reveal that most of the heatwave's maximum temperatures are outside of the range of the GEV model. Figure 2b shows the 2021 out of sample return times for the GHCN stations, where many stations realized return times in excess of 2,000 years during the 2021 PNW heatwave. The probability of 2021 temperatures exceeding this GEV upper bound (Figure 2c) further illustrates that the out of sample GEV fails to describe the 2021 PNW heatwave. Including the 2021 temperatures in the GEV fitting procedure extends the upper bounds to include these values in the distribution, but the distributions are a poor fit to the rest of the data. Using a χ² goodness-of-fit test, the p-values calculated without 2021 values are generally greater than 0.2, demonstrating strong evidence of an underlying GEV distribution. However, the p-values calculated when 2021 temperatures are included are less than 0.05, indicating that the distribution is significantly different from GEV. Figure 1b, constructed by binning all GHCN station data from 1920 to 2020 (blue) and 2021 (red), further suggests that the temperatures of the 2021 heatwave are drawn from different distributions than previous years that is not accounted for by the time-dependent greenhouse gas covariate. The above evidence suggest that the critical GEV assumption of independent and identically distributed (i.i.d.) data is violated when 2021 temperatures are included.

Given that an in-sample GEV distribution is a poor fit to the GHCN data and that the combined effects of the atmospheric blocking pattern and anomalous AR were likely very rare if not unique, we conclude that there should be little confidence in attribution statements based on in-sample GEV formulations. Philip et al. (2021) argued that the temperatures reached during the PNW heatwave were “virtually impossible” without climate change. However, this conclusion is not supported from a purely Granger causal inference perspective (Ebert-Uphoff & Deng, 2012; Hannart et al., 2016). Granger causality in this sense means that knowledge of greenhouse gas concentrations would inform about the probability of the 2021 heatwave temperatures. But due to the failure of the non-stationary GEV methodology to construct a well-fit in-sample distribution that includes the 2021 temperature values, and the fact that the out-of-sample distribution does not reach the magnitude of the 2021 event, no statement about the role of greenhouse gases should be made from this technique. The statistical analysis presented here only supports an attribution statement that these temperatures were virtually impossible under any previously experienced meteorological conditions, with or without global warming.

Given this evidence that the outlier 2021 temperatures are not drawn from the same distribution as all previous TXx values, we further conclude that TXx is not an appropriate attribution variable for this event. Hence, analysis of TXx from the CMIP global climate models is also inappropriate for understanding this event. Examination of the CMIP database for extreme temperatures of much greater rarity than TXx, either through longer block sizes or high thresholds provides a more sound basis to construct a distribution more suitable to describe events as rare as the PNW heatwave than do distributions of annual maxima. It is clear that ensemble sizes in the standard CMIP database are not large enough to construct such distributions and provide robust attribution statements. However, recent developments of larger climate model ensemble simulations (up to 40 individual realizations) do provide some insight as to the rarity of such events (McKinnon & Simpson, 2022).

I made these points in my own analysis of the WWA rapid attribution study here, back in July 2021:

I said at the time:

It’s almost certain that WWA will choose to define this extreme event only with reference to extreme daytime temperatures in the regional Pacific Northwest.

But this does not stop the intrepid team at WWA from torturing the data to fit an extremely dubious statistical distribution. Yes, they actually squeeze this glaringly unusual and extreme departure from normal into a new statistical time series and by so doing they arrive at a highly improbable estimation of the return time for such an extreme event purportedly based upon this statistical time series. I'm actually gobsmacked. They own up to their sins in the paper, which is at least honest, but of course the media coverage (with the assent and cooperation of the authors themselves) is exceptionally dishonest, conveying the impression that this 'scientific' study revealed a strong link between this event and global warming.

They deliberately chose an area where the heat was particularly extreme and they shoe-horned this extreme event to fit a highly improbably statistical series to arrive at a highly improbable estimate of a return time. Now, even a thousand years doesn't sound that scary, but with projected global warming of two degrees in 20 years time (0.8 degrees hotter than now), which 'could happen', we could then be seeing a heatwave like this every 5 or 10 years. Friederike Otto, one of the paper's authors, explains:

The ‘speed of science’ is a bit slow to be honest, but at least it does catch up with these fraudsters. Too late to reverse the propaganda and misinformation put out by the press though.