A guest post by Nic Lewis
When the Lewis & Crok report “A Sensitive Matter” about climate sensitivity in the IPCC Fifth Assessment Working Group 1 report (AR5) was published by the GWPF in March, various people criticised it for not being peer-reviewed. But peer review is for research papers, not for lengthy, wide-ranging review reports. The Lewis & Crok report placed considerable weight on energy budget sensitivity estimates based on the carefully considered AR5 forcing and heat uptake data, but those had been published too recently for any peer reviewed sensitivity estimates based on them to exist.
I am very pleased to say that the position has now changed. Lewis N and Curry J A: The implications for climate sensitivity of AR5 forcing and heat uptake estimates, Climate Dynamics (2014), has just been published, here. A non-paywalled version of the paper is available here, along with data and code. The paper’s results show the best (median) estimates and ‘likely’ (17–83% probability) ranges for equilibrium/effective climate sensitivity (ECS) and transient climate response (TCR) given in the Lewis & Crok report to have been slightly on the high side.
Our paper derives ECS and TCR estimates using the AR5 forcing and heat uptake estimates and uncertainty ranges. The analysis uses a global energy budget model that links ECS and TCR to changes in global mean surface temperature (GMST), radiative forcing and the rate of ocean etc. heat uptake between a base and a final period. The resulting estimates are less dependent on global climate models and allow more realistically for forcing uncertainties than similar estimates, such as those from the Otto et al (2013) paper.
Base and final periods were selected that have well matched volcanic activity and influence from internal variability, and reasonable agreement between ocean heat content datasets. The preferred pairing is 1859–1882 with 1995–2011, the longest early and late periods free of significant volcanic activity, which provide the largest change in forcing and hence the narrowest uncertainty ranges.
Table 1 gives the ECS and TCR estimates for the four base period – final period combinations used.
Table 1: Best estimates are medians (50% probability points). Ranges are to the nearest 0.05°C
AR5 does not give a 95% bound for ECS, but its 90% bound of 6°C is double that of 3.0°C for our study, based on the preferred 1859–1882 and 1995–2011 periods.
Considerable care was taken to allow for all relevant uncertainties. One reviewer applauded “the very thorough analysis that has been done and the attempt at clearly and carefully accounting for uncertainties”, whilst another commented that the paper provides “a state of the art update of the energy balance estimates including a comprehensive treatment of the AR5 data and assessments”.
Earlier sensitivity studies based on observed warming during the instrumental period (post 1850) have generally used forcing estimates derived from one or more global climate models (GCMs), particular published studies and/or the forcing estimates given in AR4 – which were only for a single year. Our paper appears to be the first to use the comprehensive set of forcing time series and uncertainty ranges provided in AR5. They are based on careful assessments of forcing estimates from various published studies, and are likely to become widely used in observationally-based sensitivity studies.
There is thus now solid peer-reviewed evidence showing that the underlying forcing and heat uptake estimates in AR5 support narrower ‘likely’ ranges for ECS and TCR with far lower upper limits than per the AR5 observationally-based ‘likely’ ranges of: 2.45°C vs 4.5°C for ECS and 1.8°C vs 2.5°C for TCR. The new energy budget estimates incorporate the extremely wide AR5 aerosol forcing uncertainty range – the dominant contribution to uncertainty in the ECS and TCR estimates – as well as thorough allowance for uncertainty in other forcing components, in heat uptake and surface temperature, and for internal variability. The ‘likely’ ranges they give for ECS and TCR can properly be compared with the AR5 Chapter 10 ‘likely’ ranges that reflect only observationally-based studies, shown in Table 1. The AR5 overall assessment ranges are the same.
The CMIP5 GCMs used for AR5 all have ECS values exceeding 2°C, whereas 70% of our preferred main results ECS probability lies below that level, and over 90% lies below the 3.2°C mean ECS of CMIP5 models. The 33 CMIP5 models with suitable archived data show TCR values exceeding our preferred best estimate of 1.33°C in all but one case, with an average TCR exceeding the top of our 1.8°C ‘likely’ range.
The Otto et al 2013 energy budget study (of which I was an author) used ensemble mean forcing data derived from simulations by CMIP5 climate models, with an adjustment reflecting the difference between CMIP5 models’ aerosol forcing and the AR5 best estimate. Our preferred TCR best estimate exactly matches the lowest uncertainty Otto et al estimate, that using a 2000–09 final period; their estimate using a long 1971–2009 period was almost the same. The Otto et al ECS best estimates based on data for those two periods are slightly higher than ours based on 1995–2011 and 1987–2011 data. For 2000–2009 that is mainly because heat uptake estimates over that period vary considerably and Otto et al used a high estimate. For 1970–2009 it is likely mainly due to a large mismatch in volcanic forcing with the common base period of 1860–79 used for all its estimates.
Our preferred ECS and TCR estimates are closely in line with those from other recent instrumental-observation studies based on warming over the bulk of the instrumental period that use spatiotemporal temperature data (not just GMST) to make their own estimate of aerosol forcing, and do not use a prior distribution that strongly pushes their ECS or TCR estimate towards values above where the data fits best. Those studies are: Aldrin et al. (2012), Ring et al. (2012), Lewis (2013) and Skeie et al. (2014).
Some recent studies (e.g., Huber and Knutti, 2014) seek to imply that ECS and/or TCR estimates in line with ours and those in the above-mentioned studies, which point to CMIP5 GCMs being oversensitive, are strongly influenced by the low increase in surface warming this century. Rogelj et al (2014) named four studies in this regard (Schmittner et al 2011, Aldrin et al 2012, Lewis 2013 and Otto et al 2013). The claim is factually incorrect in respect of all four studies. And for our study, ECS and TCR estimates using final periods from 1971 or 1987 to 2000, 2001, 2002 or 2003 differ very little from those using data to 2011.
It has been claimed that incomplete coverage of high-latitude zones in global temperature datasets biases down their estimate of the rate of increase in GMST. However, over the long periods involved in this study there is no evidence of any such bias. The increase in GMST per the published HadCRUT4v2 global dataset, used in the study, exceeds rather than underestimates the area-weighted average of the calculated increases for ten separate latitude zones, which method gives a full weighting to each zone.
Scientists who work on or with global climate models (GCMs) tend to be suspicious of observationally-based ECS and TCR estimates that lie well below the values indicated by most GCMs. One of their arguments is that most ECS estimates based on warming during the instrumental period are actually of effective climate sensitivity, which reflects climate feedbacks during the period studied, and that many GCMs show feedbacks changing and effective climate sensitivity increasing over time, so that energy budget and other instrumental observation-based estimates of ECS will be too low.
However, it is standard to estimate the equilibrium sensitivity of coupled GCMs from regressing radiative imbalance on GMST over a 150 year simulation involving an artificial scenario with an abrupt quadrupling of CO2 concentration. The vast majority of such regression plots show no evidence of the strength of climate feedbacks changing measurably over time, apart from (for about half the GCMs) during the first year or two when the climate system is undergoing very rapid adjustment to the quadrupling of CO2. If such behaviour in that period reflects reality, there would be an effect on the relationship of energy budget ECS estimates to the regression-based estimate of equilibrium sensitivity, but it would be extremely small and possibly negative.
It is claimed in a model-based study (Shindell, 2014) that spatially inhomogeneous forcings, principally from aerosols and ozone, lead to substantial underestimation of TCR by global energy budget and similar methods, with the levelling off of aerosol forcing in recent decades exacerbating the underestimation. ECS estimation could also be affected. Whilst the underlying arguments may well be valid, based on real world data their effects on TCR estimation appear minor – on my rough calculations, only a few percent.
Going the other way, AR5 assesses the overall effects on forcing of land use change as equally likely to be positive or negative, but its forcing estimates only include the negative albedo effect. If land use forcing is centred at zero, but with doubled uncertainty, reflecting AR5’s assessment, the preferred period combination best estimate for ECS drops to 1.54°C, with a 95% bound of 3.5°C. The TCR best estimate falls to 1.27°C, with a 95% bound of 2.3°C.
The study does not assume any possible contribution to the increase in GMST from indirect solar influences not allowed for in the AR5 forcing estimates, or from natural internal climate variability affecting ocean heat uptake and/or forcing.
Figure 1: Estimated PDFs for ECS and TCR, and their reciprocals, for the four period combinations
The paper’s ECS and TCR estimates, and the uncertainty associated with them, could also be presented in the form of probability density functions, as in panels a and b of Figure 1. The PDFs are skewed due principally to the dominant uncertainty, that in forcing, affecting the denominator of the fractions used to estimate ECS and TCR.
Skewed PDFs can be misleading. With an unchanged best estimate (the median or 50% probability point, typically identical or very close to the estimate obtained disregarding uncertainties), the PDF mode (location of its peak) moves away from the best estimate as data uncertainties increase. Reparameterizing ECS and TCR as their reciprocals (being the climate feedback parameter in the case of ECS, ignoring uncertainty in ), as in the lower panels c and d of Figure 1, results in much less skewed PDFs and avoids the mode shifting misleadingly as the width of the PDF varies. If these reciprocal parameterisations had been used in climate science, much of the misunderstanding about Bayesian inference and choice of priors that has held back progress in observationally-based estimation of ECS would likely not have occurred. When estimation errors for a parameter have fixed, symmetrical Gaussian or similar distributions – as those in panels c and d approximately do – then few people would normally dispute use of a uniform prior for it. The PDFs for ECS and TCR shown in panels a and b – which embody highly non-uniform priors – can be obtained from those for 1/ECS and 1/TCR in panels c and d by applying the standard change-of-variables formula.
Postscript (dated 24 September 2014)
A commenter asked for an extended table giving comparisons with other recent studies as well. Such a table follows. The numbers are from my own calculations. Two estimates are given for Aldrin et al (2012), one based on a uniform-in-ECS prior and the other on a uniform-in-1/ECS prior. I believe the 1/ECS prior to provide more objective ECS estimates. The TCR estimates for Lewis (2013) are derived from its ECS and ocean effective diffusivity estimates using an empirical formula. The Skeie et al (2014) TCR estimates are based on a distribution fitted to the mean and range given in the paper. Ring et al (2012) did not give any uncertainty ranges and so is not included. Its ECS best estimates varied from 1.4 to 2.0 C depending on surface temperature dataset used. These are the only other recent studies that form their own inverse estimate of aerosol forcing using data resolved at least hemispherically (or use an estimate thereof that is consistent with the AR5 best estimate) and which do not use a prior distribution that strongly pushes their ECS or TCR estimate towards substantially higher values than found in the new paper. (Full references for these various studies are given in the paper.)
I would also like to clarify, in relation to the Lewis and Crok report “A Sensitive Matter”, that (as with all GWPF reports) it was sent for comment and review to all members of the GWPF Academic Advisory Council. Accordingly, whilst it was not published in a peer reviewed journal, the criticism it received for not being peer-reviewed was inaccurate. It had undergone peer review, albeit under a different (but arguably typically more rigorous) procedure from that of a journal.