A new statistical approach to climate change detection and attribution

Abstract

We propose here a new statistical approach to climate change detection and attribution that is based on additive decomposition and simple hypothesis testing. Most current statistical methods for detection and attribution rely on linear regression models where the observations are regressed onto expected response patterns to different external forcings. These methods do not use physical information provided by climate models regarding the expected response magnitudes to constrain the estimated responses to the forcings. Climate modelling uncertainty is difficult to take into account with regression based methods and is almost never treated explicitly. As an alternative to this approach, our statistical model is only based on the additivity assumption; the proposed method does not regress observations onto expected response patterns. We introduce estimation and testing procedures based on likelihood maximization, and show that climate modelling uncertainty can easily be accounted for. Some discussion is provided on how to practically estimate the climate modelling uncertainty based on an ensemble of opportunity. Our approach is based on the “models are statistically indistinguishable from the truth” paradigm, where the difference between any given model and the truth has the same distribution as the difference between any pair of models, but other choices might also be considered. The properties of this approach are illustrated and discussed based on synthetic data. Lastly, the method is applied to the linear trend in global mean temperature over the period 1951–2010. Consistent with the last IPCC assessment report, we find that most of the observed warming over this period (+0.65 K) is attributable to anthropogenic forcings (+0.67 \(\pm\) 0.12 K, 90 % confidence range), with a very limited contribution from natural forcings (\(-0.01\pm 0.02\) K).

This is a preview of subscription content, log in to check access.

Fig. 1
Fig. 2
Fig. 3
Fig. 4

References

  1. Allen M, Stott P (2003) Estimating signal amplitudes in optimal fingerprinting, Part I: Theory. Clim Dyn 21:477–491. doi:10.1007/s00382-003-0313-9

    Article  Google Scholar 

  2. Allen M, Tett S (1999a) Checking for model consistency in optimal fingerprinting. Clim Dyn 15(6):419–434

    Article  Google Scholar 

  3. Allen M, Tett S (1999b) Checking for model consistency in optimal fingerprinting. Clim Dyn 15(6):419–434

    Article  Google Scholar 

  4. Annan J, Hargreaves J (2010) Reliability of the cmip3 ensemble. Geophys Res Lett 37(L02703). doi:10.1029/2009GL041994

  5. Berliner L, Levine R, Shea D (2000) Bayesian climate change assessment. J Clim 13(21):3805–3820

    Article  Google Scholar 

  6. Bindoff N, Stott P, AchutaRao K, Allen M, Gillett N, D Gutzler D, K Hansingo K, Hegerl G, Hu Y, Jain S, Mokhov I, Overland J, Perlwitz J, Sebbari R, Zhang X (2013) Detection and attribution of climate change: from global to regional. In: Stocker TF, Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, Nauels A, Xia Y, Bex V, Midgley PM (eds) Climate change 2013: the physical science basis. Contribution of working group I to the fifth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge

    Google Scholar 

  7. Boucher O, Randall D, Artaxo P, Bretherton C, Feingold G, Forster P, Kerminen V-M, Kondo Y, Liao H, Lohmann U, Rasch P, Satheesh SK, Sherwood S, Stevens B, Zhang XY (2013) Clouds and aerosols. In: Stocker TF, Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, Nauels A, Xia Y, Bex V, Midgley PM (eds) Climate change 2013: the physical science basis. Contribution of Working Group I to the fifth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge and New York

  8. Brohan P, Kennedy J, Harris I, Tett S, Jones P (2006) Uncertainty estimates in regional and global observed temperature changes: a new data set from 1850. J Geophys Res 111:D12106. doi:10.1029/2005JD006548

    Article  Google Scholar 

  9. Dufresne JL, Bony S (2008) An assessment of the primary sources of spread of global warming estimates from coupled atmosphere-ocean models. J Clim 21(19):5135–5144

    Article  Google Scholar 

  10. Fuller WA (1987) Measurement error models. Wiley, Amsterdam

    Google Scholar 

  11. Fyfe J, Gillett N, Zwiers F (2013) Overestimated global warming over the past 20 years. Nat Clim Change 3(9):767–769

    Article  Google Scholar 

  12. Gillett NP, Arora V, Matthews D, Allen M (2013) Constraining the ratio of global warming to cumulative CO2 emissions using CMIP5 simulations. J Clim 26(18):6844–6858

    Article  Google Scholar 

  13. Hall A, Qu X (2006) Using the current seasonal cycle to constrain snow albedo feedback in future climate change. Geophys Res Lett 33:L03502. doi:10.1029/2005GL025127

    Google Scholar 

  14. Hannart A (2015) Integrated optimal fingerprinting: method description and illustration. J Clim 29(6):1977–1998. doi:10.1175/JCLI-D-14-00124.1

    Article  Google Scholar 

  15. Hannart A, Ribes A, Naveau P (2014) Optimal fingerprinting under multiple sources of uncertainty. Geophys Res Lett 41:1261–1268. doi:10.1002/2013GL058653

    Article  Google Scholar 

  16. Hasselmann K (1979) On the signal-to-noise problem in atmospheric response studies. In: Meteorology of Tropical Oceans. Royal Meteorological Society, pp 251–259

  17. Hasselmann K (1993) Optimal fingerprints for the detection of time-dependent climate change. J Clim 6(10):1957–1971

    Article  Google Scholar 

  18. Hasselmann K (1997) Multi-pattern fingerprint method for detection and attribution of climate change. Clim Dyn 13(9):601–611

    Article  Google Scholar 

  19. Hegerl G, Zwiers F (2011) Use of models in detection and attribution of climate change. Wiley Interdiscip Rev Clim Change 2(4):570–591. doi:10.1002/wcc.121

    Article  Google Scholar 

  20. Hegerl G, Hasselmann K, Cubash U, Mitchell J, Roeckner E, Voss R, Waszkewitz J (1997) Multi-fingerprint detection and attribution analysis of greenhouse gas, greenhouse gas-plus-aerosol and solar forced climate change. Clim Dyn 13(9):613–634

    Article  Google Scholar 

  21. Hegerl G, Zwiers F, Braconnot P, Gillet N, Luo Y, Marengo Orsini J, Nicholls N, Penner J, Stott P (2007) Understanding and attributing climate change. In: Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB, Tignor M, Miller HL (eds) Climate change 2007: the physical science basis. Contribution of working group I to the fourth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge

    Google Scholar 

  22. Hegerl GC, North GR (1997) Comparison of statistically optimal approaches to detecting anthropogenic climate change. J Clim 10(5):1125–1133

    Article  Google Scholar 

  23. Hegerl GC, Hoegh-Guldberg O, Casassa G, Hoerling M, Kovats R, Parmesan C, Pierce D, Stott P (2010) Good practice guidance paper on detection and attribution related to anthropogenic climate change. In: Stocker TF, Field CB, Qin D, Barros V, Plattner G-K, Tignor M, Midgley PM, Ebi KL (eds) Meeting report of the intergovernmental panel on climate change expert meeting on detection and attribution of anthropogenic climate change IPCC working group I technical support unit. University of Bern, Bern

    Google Scholar 

  24. Henderson CR (1953) Estimation of variance and covariance components. Biometrics 9(2):226–252

    Article  Google Scholar 

  25. Huber M, Knutti R (2012) Anthropogenic and natural warming inferred from changes in earth/’s energy balance. Nat Geosci 5(1):31–36

    Article  Google Scholar 

  26. Huntingford C, Stott P, Allen M, Lambert F (2006) Incorporating model uncertainty into attribution of observed temperature change. Geophys Res Lett 33:L05710. doi:10.1029/2005GL024831

    Article  Google Scholar 

  27. IPCC (2007) Climate change 2007: the physical science basis. In: Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB, Tignor M, Miller HL (eds) Contribution of working group I to the fourth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge

  28. IPCC (2013) In: Stocker TF, Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, Nauels A, Xia Y, Bex V, Midgley PM (eds) Climate change 2013: the physical science basis. Contribution of working group I to the fifth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge

  29. Jones G, Stott P (2011) Sensitivity of the attribution of near surface temperature warming to the choice of observational dataset. Geophys Res Lett 38(L21702). doi:10.1029/2011GL049324

  30. Jones GS, Stott P, Christidis N (2013) Attribution of observed historical near surface temperature variations to anthropogenic and natural causes using CMIP5 simulations. J Geophys Res Atmos 118(10):4001–4024

    Article  Google Scholar 

  31. Knutti R, Hegerl G (2008) The equilibrium sensitivity of the earth’s temperature to radiation changes. Nat Geosci 1(11):735–743. doi:10.1038/ngeo337

    Article  Google Scholar 

  32. Knutti R, Abramowitz G, Collins M, Eyring V, Gleckler PJ, Hewitson B, Mearns L (2010) Good practice guidance paper on assessing and combining multi model climate projections. In: Stocker TF, Qin D, Plattner G-K, Tignor M, Midgley PM (eds) Meeting report of the intergovernmental panel on climate change expert meeting on assessing and combining multi model climate projections. IPCC working group I technical support unit. University of Bern, Bern, Switzerland

  33. Le Cam L (1990) Maximum likelihood—an introduction. Int Stat Inst Rev 58(2):153–171. doi:10.2307/1403464

    Article  Google Scholar 

  34. Mitchell J, Karoly D, Hegerl G, Zwiers F, Allen M, Marengo J (2001) Detection of climate change and attribution of causes. In: Houghton et al (ed) Climate change 2001: the scientific basis. Contribution of working group I to the third assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge

  35. Morice C, Kennedy J, Rayner N, Jones P (2012) Quantifying uncertainties in global and regional temperature change using an ensemble of observational estimates: the hadcrut4 data set. J Geophys Res 117(D8). doi:10.1029/2011JD017187

  36. Myhre G, Shindell D, Bréon F-M, Collins W, Fuglestvedt J, Huang J, Koch D, Lamarque J-F, Lee D, Mendoza B et al (2013) Anthropogenic and natural radiative forcing. In: Stocker TF, Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, Nauels A, Xia Y, Bex V, Midgley PM (eds) Climate change 2013: the physical science basis. Contribution of working group I to the fifth assessment report of the intergovernmental panel on climate change. Cambridge University Press, Cambridge, pp 571–657

  37. Rao CR, Kleffe J (1988) Estimation of variance components and applications. North-Holland, New York

    Google Scholar 

  38. Ribes A, Terray L (2013) Application of regularised optimal fingerprinting to attribution. Part II: Application to global near-surface temperature. Clim Dyn 41(11–12):2837–2853. doi:10.1007/s00382-013-1736-6 on line

    Article  Google Scholar 

  39. Ribes A, Terray L, Planton S (2013) Application of regularised optimal fingerprinting to attribution. Part I: Method, properties and idealised analysis. Clim Dyn 41(11–12):2817–2836. doi:10.1007/s00382-013-1735-7

    Article  Google Scholar 

  40. Rotstayn LD, Collier MA, Shindell DT, Boucher O (2015) Why does aerosol forcing control historical global-mean surface temperature change in CMIP5 models? J Clim 28(17):6608–6625

    Article  Google Scholar 

  41. Santer B, Wigley T, Barnett T, Anyamba E (1995) Detection of climate change and attribution of causes. Cambridge University Press, Cambridge

    Google Scholar 

  42. Santer B, Painter J, Mears C, Doutriaux C, Caldwell P, Arblaster J, Cameron-Smith P, Gillett N, Gleckler P, Lanzante J, Perlwitz J, Solomon S, Stott P, Taylor K, Terray L, Thorne P, Wehner M, Wentz F, Wigley T, Wilcox L, Zou CZ (2013) Identifying human influences on atmospheric temperature. Proc Natl Acad Sci 110(1):26–33. doi:10.1073/pnas.1210514109

    Article  Google Scholar 

  43. Shin SIS, Sardeshmukh D (2011) Critical influence of the pattern of tropical ocean warming on remote climate trends. Clim Dyn 36(7–8):1577–1591

    Article  Google Scholar 

  44. Shiogama H, Stone D, Nagashima T, Nozawa T, Emori S (2013) On the linear additivity of climate forcing-response relationships at global and continental scales. Int J Climatol 33(11):2542–2550. doi:10.1002/joc.3607

    Article  Google Scholar 

  45. Stevens B, Bony S (2013) What are climate models missing. Science 340(6136):1053–1054

    Article  Google Scholar 

  46. Stott P, Mitchell J, Allen M, Delworth D, Gregory J, Meehl G, Santer B (2006) Observational constraints on past attributable warming and predictions of future global warming. J Clim 19(13):3055–3069

    Article  Google Scholar 

  47. Taylor K, Stouffer R, Meehl G (2012) An overview of CMIP5 and the experiment design. Bull Am Meteorol Soc 93(4):485–498. doi:10.1175/BAMS-D-11-00094.1

    Article  Google Scholar 

  48. Terray L, Corre L, Cravatte S, Delcroix T, Reverdin G, Ribes A (2011) Near-surface salinity as nature’s rain gauge to detect human influence on the tropical water cycle. J Clim. doi:10.1175/JCLI-D-10-05025.1

  49. van Oldenborgh G, Doblas Reyes F, Hawkins E (2013) Reliability of regional climate model trends. Environ Res Lett 8(1):014,055. doi:10.1088/1748-9326/8/1/014055

    Article  Google Scholar 

Download references

Acknowledgments

The authors are grateful to the two anonymous referees for their constructive comments, which were of great value in improving the paper. Part of this work has been supported by the Fondation STAE, via the project Chavana, and by the Extremoscope and ANR-DADA projects.

Author information

Affiliations

Authors

Corresponding author

Correspondence to Aurélien Ribes.

Appendix

Appendix

Model (1)–(3) as a linear regression Gaussian model

Based on the notation used in (1)–(3), we define

$$\begin{aligned} Z = \left( \begin{array}{c} Y \\ X_1 \\ \vdots \\ X_{n_f}\end{array} \right) , \quad \quad Z^* = \left( \begin{array}{c} X^*_1 \\ \vdots \\ X^*_{n_f}\end{array} \right) , \quad {\text { and }} \quad \varepsilon = \left( \begin{array}{c} \varepsilon _Y \\ \varepsilon _{X_1} \\ \vdots \\ \varepsilon _{X_{n_f}} \end{array} \right) . \end{aligned}$$
(35)

which are vectors of size \(({n_f}+1)n\), \({n_f}n\), and \(({n_f}+1)n\), respectively. Then, (2) and (3) may be written simultaneously as

$$\begin{aligned} Z = A Z^*+\varepsilon , \end{aligned}$$
(36)

where

$$\begin{aligned} A= \left( \begin{array}{cccc} I_n &{} I_n &{} \dots &{} I_n \\ I_n &{} 0 &{} \dots &{} 0 \\ 0 &{} I_n &{} \ddots &{} \vdots \\ \vdots &{} \ddots &{} \ddots &{} 0 \\ 0 &{} \dots &{} 0 &{} I_n \end{array} \right) , \end{aligned}$$
(37)

is the \(({n_f}+1)n \times {n_f}n\) design matrix of the linear Gaussian model (36).

This representation is useful to derive some of the statistical properties of our model. It is used, for instance, in “Appendix 8.2.3”. \(A'A\) has also a simple closed form. However, it is difficult to derive the precision matrix \(\left ( A'A \right )^{-1}\) in closed form, which would be required to deduce, e.g., the MLEs more directly than in Sect. 3.3.

Hypothesis testing details

Goodness of fit tests

The tests proposed in Sect. 3.5 are goodness of fit tests constructed as follows.

If \(Z \in {\mathbb {R}}^n\) is a random vector with distribution \(N(\mu ,{\varSigma })\) under a given model, with \(\mu\) and \({\varSigma }\) known, we will call “goodness of fit” test of this model the deviance test, i.e. the likelihood ratio test (LRT) with respect to a saturated alternative hypothesis \(H_1: Z \sim N(\theta ,{\varSigma })\), where \(\theta \in {\mathbb {R}}^n\) is unknown. The log-likelihood is 0 under this alternative, so the LRT only involves \(-2\) log-likelihood under \(H_0\), the considered model. Therefore the LRT statistic is

$$\begin{aligned} T = (Z-\mu )' {\varSigma }^{-1} (Z-\mu ), \end{aligned}$$
(38)

which follows a \(\chi ^2(n)\) distribution under \(H_0\).

Minimized likelihood under \(H_0: Y^* = 0\) in model (1)–(3)

Under \(H_0: Y^*=0\), each \(X_i^*\) has to be estimated under the additional constraint that \(\sum _{i=1}^{{n_f}} X_i^* = 0\). Under \(H_0\), (6) becomes

$$\begin{aligned} \ell _{H_0}(X^*_i) = Y' {\varSigma }_Y^{-1} Y + \sum _{i=1}^{{n_f}} \left (X_i - X^*_i\right )' {\varSigma }_{X_i}^{-1} \left (X_i - X^*_i\right ). \end{aligned}$$
(39)

The gradient of \(\ell _{H_0}\) with respect to \(X_i^*\) is

$$\begin{aligned} {\varSigma }_{X_i}^{-1} (\widehat{X}_i^* - X_i). \end{aligned}$$
(40)

The gradient of the constraint is \({\mathbbm {1}}_n\), i.e. the “all-one” vector of size n. The theory of Lagrange multipliers imposes that at the constrained minimum, these two gradients are proportional, so that

$$\begin{aligned} {\varSigma }_{X_i}^{-1} (\widehat{X}_i^* - X_i) = \lambda {\mathbbm {1}}_n, \end{aligned}$$
(41)

where \(\lambda\) is some constant. From (41),

$$\begin{aligned} \widehat{X}_i^* = X_i + \lambda {\varSigma }_{X_i} {\mathbbm {1}}_n, \end{aligned}$$
(42)

and thus the constraint \(\sum _{i=1}^{{n_f}} \widehat{X}_i^* =0\) gives

$$\begin{aligned} \lambda {\mathbbm {1}}_n = - {\varSigma }_X^{-1} X, \end{aligned}$$
(43)

where \(X = \sum _{i=1}^{{n_f}} X_i\) and \({\varSigma }_X= \sum _{i=1}^{{n_f}} {\varSigma }_{X_i}\). The maximum likelihood estimate of \(X_i^*\) under \(H_0\) is therefore

$$\begin{aligned} \widehat{X}_i^* = X_i - {\varSigma }_{X_i} {\varSigma }_X^{-1} X. \end{aligned}$$
(44)

Finally, the minimized value of \(-2\) log-likelihood under \(H_0\) is

$$\begin{aligned} \ell _{H_0}(\widehat{X}^*_i) = Y' {\varSigma }_Y^{-1} Y + X' {\varSigma }_X^{-1} X. \end{aligned}$$
(45)

This result is used in “Appendix 8.2.3”, Eq.  (46), to derive the LRT of \(H_0: Y^*=0\) within the statistical model defined by (1)–(3).

Choice of the null and the alternative hypotheses in the detection test

The “detection” test deals with the null hypothesis of no change, i.e. \(H_0: Y^*=0\). The detection test we propose in (18) is a goodness of fit test based on (2) only. In particular, it doesn’t consider (1) and (3). Here, we present two alternatives that might be considered to test the same null hypothesis. Both are LRTs between two nested well-defined hypotheses on the data (YX).

  • In the statistical model (1)–(3), consider the null-hypothesis \(H_0: Y^*=0\) versus the alternative hypothesis \(H_1:\)(1)–(3), ie the same model with an unspecified \(Y^*\). Following (45) and (14), this test would be based on the statistic

    $$\begin{aligned} \ell _{H_0} - \ell _{H_1} = Y' {\varSigma }_Y^{-1} Y + X' {\varSigma }_X^{-1} X - (Y-X)' ({\varSigma }_Y+{\varSigma }_X)^{-1} (Y-X) \sim _{H_0} \chi ^2(n). \end{aligned}$$
    (46)

    Note that the distribution under \(H_0\) is known to be \(\chi ^2(n)\) since \(H_0\) is a linear sub-hypothesis of \(H_1\), and they differ by a dimension of n (see 8.1).

  • In the statistical model (1)–(3), consider the null-hypothesis \(H_0: Y^*=0\) versus the saturated alternative hypothesis. This test is a goodness of fit test as defined above. Such a test would be based on the statistic (see (45))

    $$\begin{aligned} \ell _{H_0} = Y' {\varSigma }_Y^{-1} Y + X' {\varSigma }_X^{-1} X \sim _{H_0} \chi ^2(2n), \end{aligned}$$
    (47)

    where the null-distribution is deduced directly from (2) and (3).

Both of these tests would treat information from Y and X very symmetrically because, given (1), \(Y^*=0\) implies not only that Y is small, but also that X is small (as \(X^*=0\)). Therefore, rejection of these tests may happen because either Y or X are “large”. In the case where X is large while Y is not, detection would not be a consequence of an abnormal observation, but rather the consequence of too large a response simulated in climate models. This, of course, is not consistent with the definition of detection. In our opinion, it is then more appropriate to discuss the first term in the right hand side of (45) separately.

We further argue that there is a fundamental distinction between detection and attribution in this respect. As detection only assesses whether observations are consistent with internal variability, historical simulations by climate models (X) are not required for detection. The only requirement is to quantify internal variability—which is usually done based on other simulations by climate models. Attribution, however, definitively relies on historical simulations to disentangle contributions from different forcings based on some physical knowledge. This fundamental difference is the main reason why we propose to base detection on (2) only, while (1)–(3) are considered as a whole to perform attribution, and in particular to estimate the contributions of individual forcings.

As a last remark, the test defined in (18) may also be considered as a test of the null hypothesis \(H_0: ``Y^*=0\) and (1) does not hold”. Indeed, removing (1) implies that the second term in 45 is zero, and then our test (18) is actually a LRT against a saturated alternative hypothesis.

Estimation of \({\varSigma }_{\text {m}}\) and \({\varSigma }_X\) with unbalanced data

This section deals with the realistic case where models have run ensembles of historical simulations of various sizes. For instance, in the CMIP5 archive, the number of historical simulations with all external forcings varies from 1 to 10 depending on the model considered. Here, we mainly discuss the estimation of \({\varSigma }_{\text {m}}\), which is not explicitly addressed in the main text. We also mention what \({\varSigma }_X\) should be considered in this unbalanced case.

Consistent with Sect. 4, we assume that simulation k of model j can be decomposed as

$$\begin{aligned} w_{jk} = \mu + m_j + \epsilon _{jk}, \quad j=1,\dots ,n_m, \quad k=1,\dots ,n_j, \end{aligned}$$
(48)

where \(m_j \sim N(0,{\varSigma }_{\text {m}})\), and \(\epsilon _{jk} \sim N(0,{\varSigma }_{\text {v}})\), leading to

$$\begin{aligned} w_{jk} \sim N \left (\mu ,{\varSigma }_{\text {m}}+{\varSigma }_{\text {v}}\right ). \end{aligned}$$
(49)

We then introduce

$$\begin{aligned} w_{j.} = \frac{1}{n_j} \sum _{i=1}^{n_j} w_{jk} \sim N \left (0, {\varSigma }_{\text {m}}+\frac{{\varSigma }_{\text {v}}}{n_j} \right ). \end{aligned}$$
(50)

This kind of framework is actually a multivariate linear mixed model, and useful references could be found in Rao and Kleffe (1988). Within such models, the main challenge comes from the estimation of variance components (here \({\varSigma }_{\text {m}}\) and \({\varSigma }_{\text {v}}\)), and no optimal estimator is known in the general unbalanced case (i.e. if the ensemble sizes \(n_j\) are not all equal). We propose to use a method of moments approach very similar to that proposed by Henderson (1953), but with a couple of specific features:

  • Consistent with common practice in climate science, we estimate the fixed effect \(\mu\) as the mean of the ensemble means from each model. In this way each model is given equal weight, disregarding the number of simulations performed. Consequently, we consider \(\widehat{\mu } = \overline{w}= \frac{1}{n_m} \sum _{j=1}^{n_m} w_{j.}\), whereas the common approach in statistics uses \(\widehat{\mu } = w_{..} = \frac{1}{n} \sum _{j,k} w_{jk}\).

  • As a large number of unforced simulations are available to estimate \({\varSigma }_{\text {v}}\) (Table 1), we assume that this term is already known, and only focus on the estimation of \({\varSigma }_{\text {m}}\). Note that within-ensemble differences might also be used to estimate \({\varSigma }_{\text {v}}\).

One has:

$$\begin{aligned} \overline{w}= \frac{1}{n_m} \sum _{j=1}^{n_m} w_{j.} \quad \sim \quad N \left ( \mu ,\frac{1}{n_m} {\varSigma }_{\text {m}}+ \frac{1}{n_m^2} \sum _{j=1}^{n_m} \frac{1}{n_j} {\varSigma }_{\text {v}}\right ). \end{aligned}$$
(51)
$$\begin{aligned} {\text {Var}} \left ( w_{j.} - \overline{w}\right ) = \left (1-\frac{1}{n_m} \right ) {\varSigma }_{\text {m}}+ \left ( \frac{1}{n_j}-\frac{2}{n_m n_j}+\frac{1}{n_m^2} \sum _{j=1}^{n_m} \frac{1}{n_j} \right ) {\varSigma }_{\text {v}},\end{aligned}$$
(52)
$$\begin{aligned} {\text {E}}\ \left ( \sum _{j=1}^{n_m} (w_{j.} - \overline{w})^2 \right ) = (n_m-1) {\varSigma }_{\text {m}}+ \frac{n_m-1}{n_m} \sum _{j=1}^{n_m} \frac{1}{n_j} {\varSigma }_{\text {v}}. \end{aligned}$$
(53)

Using the method of moments, we estimate this quantity with

$$\begin{aligned} SSM = \sum _{j=1}^{n_m} (w_{j.} - \overline{w})^2. \end{aligned}$$
(54)

Finally, given an estimate of \({\varSigma }_{\text {v}}\), we estimate \({\varSigma }_{\text {m}}\) with

$$\begin{aligned} \widehat{{\varSigma }}_{\text {m}} = \frac{1}{n_m-1} \left ( SSM - \frac{n_m-1}{n_m} \sum _{j=1}^{n_m} \frac{1}{n_j} {\varSigma }_{\text {v}}\right )_+, \end{aligned}$$
(55)

where \(A_+\) means that negative eigenvalues of A are set to 0. \(\widehat{{\varSigma }}_{\text {m}}\) is a truncation of a quadratic unbiased estimator, very similar to Henderson (1953) or MIVQUE estimators (Rao and Kleffe 1988). A nice property of this approach is that it can be used even if the \(w_{j.}\) only are known (e.g. the \(w_{jk}\) are not observed). This happens, for instance, if one ensemble of simulations has not been performed, and the response of each model is estimated by subtraction, e.g. \(w^{ANT} = w^{ALL} - w^{NAT}\). In such a case, we could assume, for model j:

$$\begin{aligned} w^{ANT}_{j.} = w^{ALL}_{j.} - w^{NAT}_{j.} \sim N \left ( 0,{\varSigma }_{\text {m}}^{ANT}+\frac{{\varSigma }_{\text {v}}}{n_j^{ANT}} \right ), \end{aligned}$$

where \(n_j^{ANT}\) is defined by

$$\begin{aligned} \frac{1}{n_j^{ANT}} = \frac{1}{n_j^{ALL}} + \frac{1}{n_j^{NAT}}. \end{aligned}$$
(56)

Finally, under this unbalanced assumption, putting (51 52 53) together with \((\mu - w^*) \sim N(0,{\varSigma }_{\text {m}})\) suggests considering

$$\begin{aligned} {\varSigma }_X = \left ( 1 + \frac{1}{n_m} \right ) \widehat{{\varSigma }}_{\text {m}} + \frac{1}{n_m^2} \sum _{j=1}^{n_m} \frac{1}{n_j} {\varSigma }_{\text {v}}. \end{aligned}$$
(57)
Fig. 5
figure5

Same as Fig. 5 based on the “models are centered on the truth” paradigm

Application to global mean temperature using the “models are centered on the truth” paradigm

In this section, we present and briefly discuss the results obtained in the analysis of global mean temperature, when the “models are centered on the truth” paradigm is used instead of the “models are statistically indistinguishable from the truth” paradigm, which was used in Sect. 6.2.

Under this paradigm, outputs from each climate model, \(w_j\), may be regarded as being sampled independently from the same distribution, which is centered on the truth, i.e. \(w_j \sim N(w,{\varSigma }_m)\), where \(w\) is the true value of the simulated parameter, and \({\varSigma }_m\) describes the climate modelling uncertainty on this parameter. The distribution of the multimodel mean is then given by \(\overline{w}\sim N(w, {\varSigma }_m/n_m)\). Finally, \({\varSigma }_X = {\varSigma }_m /n_m\) has to be considered under this paradigm (if internal variability is neglected).

The primary consequence of considering this alternative paradigm is to narrow the climate modelling uncertainty. Consequently, the assessment of consistency between models and observations is usually more demanding, while more weight is given to models in the estimation of individual forcing contributions.

Under this revised assumption, panel a) is unchanged.

In panel b), observations are found to be barely consistent with models (p value 0.09). The expected ALL warming, based on climate models only, is \([+0.74\)K\(, +0.86\)K] (90 % confidence interval). This is a much narrower interval than reported in Sect.  6.2 (\([+0.44\)K\(, +1.16\)K]). The weak consistency with observations mentioned above is then related to internal variability, which has to be added to these numbers. After inference, the estimated past warming lies between \([+0.72\)K\(, +0.83\)K], which is still substantially greater than the observed value of \(+0.65K\). This can be understood as follows: because there is little uncertainty in the estimate provided by climate models, the method considers that internal variability is partly responsible for the low observed value. The overall forced change is then estimated to be higher than that found in raw observations.

In panel c), results are similarly impacted. The expected (climate models only) NAT response is unchanged, and the ANT response is expected to lies within \([+0.67\)K\(, +0.93\)K], which is again quite a narrow interval. After the inference is performed, the ANT response is estimated to be within \([+0.64\)K\(, +0.82\)K]. If compared to the results given in Sect. 6.2, the impact of changing the paradigm is limited. The main impact of changing the paradigm is to discard the lowest values, from \(+0.55\)K to \(+0.64\)K.

Overall, these results suggest that the “models are statistically indistinguishable from the truth” paradigm, which was used in the main text, is more appropriate to ensure consistency between models and observations, and avoids over emphasizing the climate models outputs.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Ribes, A., Zwiers, F.W., Azaïs, J. et al. A new statistical approach to climate change detection and attribution. Clim Dyn 48, 367–386 (2017). https://doi.org/10.1007/s00382-016-3079-6

Download citation

Keywords

  • Detection
  • Attribution
  • Climate change
  • Optimal fingerprint