Another study at odds with climate alarmism

Nicholas Lewis's recent paper "Objectively combining climate sensitivity evidence" calls into question the Intergovernmental Panel on Climate Change (IPCC AR6) likely range of climate sensitivity to carbon dioxide.

Here is a press release style summary of his paper.
One of the most important conclusions of the recent 6th Assessment Report of the Intergovernmental Panel on Climate Change (IPCC AR6) was to reduce the uncertainty in estimates of climate sensitivity to doubling the amount of carbon dioxide in the atmosphere. Since 1979, the likely range (66% chance) of climate sensitivity has been between 1.5°C and 4.5°C. This range has remained stubbornly wide, until the IPCC AR6 narrowed the likely range to be between 2.5°C and 4.0°C.

A new paper by independent scientist Nic Lewis published in the journal Climate Dynamics challenges the conclusions of the IPCC AR6 about climate sensitivity. Lewis’ analysis reduces the magnitude of climate sensitivity by one third, relative to the range provided by the IPCC AR6. These results suggest that future global warming in response to fossil fuel emissions could be significantly less than has been assumed by policy makers.

In 2015, the World Climate Research Programme convened a Workshop aimed at reducing the uncertainty in estimates of climate sensitivity to increasing carbon dioxide. The Workshop ultimately resulted in publication of a report (a 92 page paper) by many of the participants that thoroughly assessed all lines of evidence (Sherwood et al, 2020). A key result of this paper was to reduce the likely range of climate sensitivity values to 2.6 oC to 3.9 oC. While Lewis was an invited participant to the 2015 Workshop, he was not a coauthor on this paper. The Sherwood et al. paper strongly influenced the IPCC AR6’s assessment of climate sensitivity.

Lewis’ paper critiqued the methods used in the Sherwood et al. paper, finding significant errors, inconsistencies and other shortcomings. Lewis remedied these shortcomings and also revised key input data, almost entirely to reflect more recent evidence. The results of Lewis’ analysis determined a likely range of 1.75 to 2.7oC for climate sensitivity. The central estimate from Lewis’ analysis is 2.16 oC, which is well below the IPCC AR6 likely range. This large reduction relative to Sherwood et al. shows how sensitive climate sensitivity estimates are to input assumptions. Lewis’ analysis implies that climate sensitivity is more likely to be below 2 oC than it is to be above 2.5 oC.

The lower estimates of climate sensitivity determined by Nic Lewis have profound implications for climate models and projections of warming for the 21st century. Climate models used in the IPCC AR6 had values of climate sensitivity ranging from 1.8oC to 5.6oC. The IPCC AR6 judged that some of the climate models had values of climate sensitivity that were too high. Hence the AR6 selected only the climate models with reasonable values of climate sensitivity to be used in projections of 21st century climate change. Lewis’ analysis indicates that a majority of climate models used in the IPCC AR6 have values higher than the likely range.

Here is a link to his paper.

Here are some excerpts.

----------------------------------------

Introduction

The Earth's climate sensitivity is a key measure of the longer-term climate response to external forcing. It is perhaps the most important ill-quantified climate system parameter. In principle, climate sensitivity represents the equilibrium change in mean surface temperature to a doubling of atmospheric CO2 concentration from preindustrial levels, once the deep ocean has reached a stable state. In practice it is normally estimated using some approximate measure, often derived from disequilibrium changes. Climate sensitivity has been estimated from various types of evidence, but none of these has narrowly constrained its value. The first five Assessment Reports by the Intergovernmental Panel on Climate Change (IPCC) relied heavily on estimates of climate sensitivity from global climate model (GCM) simulations. The 1.5–4.5 K likely range for climate sensitivity in the 2013 IPCC Fifth Assessment Report (AR5) was identical to the range presented in the landmark Charney (1979) report, with the great increase in GCM sophistication since 1979 not having led to any narrowing of the climate sensitivity range.

GCMs use semi-empirical approximations (parameterizations) to represent subgrid-scale cloud and convection processes that are known to be critical to determining the model's climate sensitivity, which varies by up to a factor of three among GCMs. In one well regarded GCM, a simple change to how convective precipitation was parameterizedFootnote1 varied its climate sensitivity by a factor of two, with no obvious change in how well the model otherwise performed (Zhao et al. 2016). Changing the order in which the various parameterized atmospheric modules were updated in each time step was found to vary another GCM's climate sensitivity by a factor of up to two, with ambiguity existing regarding the optimum ordering (Donahue and Caldwell 2018). Moreover, the universal use in GCMs of deterministic parameterizations may bias their climate sensitivity upwards (Strommen et al. 2019). Such issues make the reliability of GCM-derived estimates of climate sensitivity questionable.

In the light of such issues, and the further widening of the range of GCM climate sensitivities in the latest (CMIP6) generation of GCMs (Zelinka et al. 2020), the IPCC Sixth Assessment Report (AR6) abandoned the previous reliance on GCM climate sensitivities. Instead, evaluation of climate sensitivity was approached by combining estimates based on different lines of evidence, such as process understanding (feedback analysis), the historical instrumental record, and paleoclimate data.

Combining different lines of evidence should, to the extent that they are independent, enable climate sensitivity to be estimated more precisely than from any single line of evidence (Stevens et al. 2016). A comprehensive attempt to do so was made by Sherwood et al. (2020, henceforth S20), a 92-page study. S20 was conducted under the auspices of the World Climate Research Programme's Grand Science Challenge on Clouds, Circulation and Climate Sensitivity and provides a very detailed investigation of climate sensitivity. As the most influential recent assessment, S20 was cited over twenty times in the relevant AR6 chapter, which approached climate sensitivity estimation on very similar lines to S20, albeit not using its formal probabilistic methods. There are in principle considerable strengths in S20's scientific approach. Its main results were derived by combining understanding from feedback analysis (Process evidence) with evidence from changes since circa 1850 (Historical evidence), and from cold and warm past periods (Paleoclimate evidence)—three lines of evidence that S20 judged to be largely independent.

The contribution the present study makes to estimation of climate sensitivity is three-fold. First, it identifies statistical problems in S20. The main methodological argument is that, when Bayesian methods are used, an Objective rather than a Subjective Bayesian approach should be taken. This means that rather than the investigator choosing the prior distribution, the prior distribution should be mathematically computed, based on the assumed statistical model relating to all the evidence to be analyzed (Bernardo 2009). S20 used a Subjective Bayesian statistical method, with an investigator-selected prior distribution, that has been shown may produce unrealistic climate sensitivity estimation when used to combine differing types of evidence (Lewis 2018), and S20 provided no evidence that it did not do so in this case. Moreover, for all except Process evidence, S20 used a method of estimating likelihoods that turns out to be unsound. This study validates its likelihood estimates by using multiple methods and cross-checking their results. S20's method is shown to often result in serious likelihood underestimation at higher climate sensitivity levels.

The second contribution of this study is that it develops and applies an Objective Bayesian approach to combining differing climate sensitivity evidence, using a mathematically computed prior distribution. The results using the methodology developed and the same input assumptions as S20 are then used to assess what effect the statistical problems identified in S20 have on its results. It is found that they bias S20's estimation of climate sensitivity downwards, although only to a minor extent even at the upper uncertainty bound when all three lines of evidence are combined.

This study's third contribution is to review and where appropriate revise the input assumptions used by S20, paying particular regard to more recent evidence, and to investigate the effect of the revised input assumptions on estimates of climate sensitivity using the developed Objective Bayesian methodology. Some of the revisions to input assumptions relate to the treatment in certain cases of CO2 forcing and/or the warming it causes. This study differs from S20 regarding the appropriate scaling of CO2 forcing, and comparison of warming, where different changes in CO2 atmospheric concentrations are involved, and regarding scaling CO2 forcing where its use requires a different estimation basis from that on which the forcing estimate was derived. The combined effects of the revisions to S20's CO2 related estimates and to other input assumptions result in a major reduction in estimated climate sensitivity.

-------------------------------

Discussion

This study first identifies statistical problems in S20. Using a Subjective Bayesian statistical method involving an investigator-selected prior distribution, as S20 does, may produce unrealistic climate sensitivity estimation when used to combine differing types of evidence (Lewis 2018), even assuming that the data likelihood functions are correct. In this case, I found that the method S20 used for estimating likelihoods for all but Process evidence was in fact unsound, and that it underestimated likelihood at high S levels, substantially so in some cases. I also found that S20 used an uncertainty estimate for PETM CO2 forcing that was a factor of ten too low, due to an apparent coding error, further biasing their likelihood estimate (although not affecting their main results).

This study then develops an Objective Bayesian approach to combining differing climate sensitivity evidence that, unlike the method used in Lewis and Grünwald (2018), is not restricted to dealing with a particular simple statistical model. The approach involves computationally deriving Jeffreys' prior distributions that are designed to maximize the influence of the data on the results and to produce probabilistic estimates that are as close as possible to being confidence intervals, and thus are well calibrated. Three different inferential methods employed for this purpose each provide nearly identical estimated likelihoods and Jeffreys' priors, and final results. This result is very supportive of the validity of the methods used and of the results they produce.

The robustness of S20's results to the use of properly calibrated statistical methods and validly calculated likelihood estimates is then examined, using the Objective Bayesian methods developed in this study. It is shown that while S20's choice of prior and its likelihood misestimation lead to over-constraining of high S levels, based on S20's data-variable assumptions the downwards bias in S20's Baseline combined evidence results is modest: the median S estimate is approximately 0.13 K low, and the 95% uncertainty bound 0.35 K low. However, the bias in S20's No Process results is over twice as large.

The other main contribution of this study is to assess the impact of revising various input data-variable distributions used by S20, by:(i)

adjusting the F2×CO2 value used for inferring S from Process and Historical evidence to reflect the effect of climate feedback changing over GCM abrupt4xCO2 simulations, as should undoubtedly be done;
(ii)

allowing for the CO2 concentration-ERF relationship being slightly non-logarithmic, and estimating the ECS to S ratio in a way that is unaffected by that relationship;
(iii)

changing some of S20's other data-variable estimates to reflect more recent information; and
(iv)

using arguably better justified (albeit not based purely on more recent information), alternative estimates for a few other data-variables.

I find that doing so results in substantially lower and better constrained estimates for S. The median S estimate when combining all lines of evidence, using the Objective Bayesian method and the LGM and mPWP for Paleoclimate evidence, reduces from 3.23 to 2.16 K.

All the revised data-variable estimates are not only defensible but, given the evidence now available, in my view are better justified than S20's original estimates. Moreover, omitting the only revisions dependent, to a greater or lesser extent, on reevaluation of existing evidence only very modestly changes the combined evidence results, with the omission just of the revision of the Historical aerosol forcing having almost no effect on the results.

It therefore currently remains quite plausible that S is below 2 K. The truncation in S20's results of the lower bound for S does not appear justified given the range of data-variable estimates supported by relevant, mainly more recent, studies. There is 36% probability of S being under 2 K, considerably greater than the 26% probability of S exceeding 2.5 K, according to the revised data-variable assumptions 'All combined: Paleo LGM + mPWP' results; they also imply that it is extremely unlikely that S is below 1.5 K, and extremely unlikely that S is above 3.2 K.

The revised data-variable median Historical evidence estimates of Shist and TCR are somewhat higher than the comparable estimates in Lewis and Curry (2018), of 1.66 K and 1.33 K respectively. The excess is mainly due to a stronger aerosol ERF change, even after revising S20's assumptions. Further revising S20's median aerosol ERF to match the change per the AR5 time-series, extended post-2011 using AR6's annual changes, would reduce the Table 8 median Shist and TCR to respectively 1.82 K and 1.40 K. Changing the base period to 1869–1882 to match Lewis and Curry (2018), avoiding the poorly observed 1861–1868 period, would further reduce those estimates, to 1.79 K and 1.37 K. The methane shortwave ERF adjustment, and greater estimated change in radiative imbalance, in AR6 can account for the small remaining differences.