1 AUGUST 2021 SANTER ET AL.

6281

Using Climate Model Simulations to Constrain Observations

BENJAMIN D. SANTER,a STEPHEN PO-CHEDLEY,a CARL MEARS,b JOHN C. FYFE,c NATHAN GILLETT,c QIANG FU,d
JEFFREY F. PAINTER,a SUSAN SOLOMON,e ANDREA K. STEINER,f FRANK J. WENTZ,b MARK D. ZELINKA,a
g
AND CHENG-ZHI ZOU
a
Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California
b
Remote Sensing Systems, Santa Rosa, California
c
Canadian Centre for Climate Modelling and Analysis, Environment and Climate Change Canada,
Victoria, British Columbia, Canada
d
Department of Atmospheric Sciences, University of Washington, Seattle, Washington
e
Department of Earth, Atmospheric, and Planetary Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts
f
Wegener Center for Climate and Global Change, University of Graz, Graz, Austria
g
Center for Satellite Applications and Research, NOAA/NESDIS, Camp Springs, Maryland

(Manuscript received 5 October 2020, in final form 29 March 2021)

ABSTRACT: We compare atmospheric temperature changes in satellite data and in model ensembles performed under
phases 5 and 6 of the Coupled Model Intercomparison Project (CMIP5 and CMIP6). In the lower stratosphere, multidecadal
stratospheric cooling during the period of strong ozone depletion is smaller in newer CMIP6 simulations than in CMIP5 or
satellite data. In the troposphere, however, despite forcing and climate sensitivity differences between the two CMIP
ensembles, their ensemble-average global warming over 1979–2019 is very similar. We also examine four properties of
tropical behavior governed by basic physical processes. The first three are ratios between trends in water vapor (WV) and
trends in sea surface temperature (SST), lower-tropospheric temperature (TLT), and mid- to upper-tropospheric tem-
perature (TMT). The fourth property is the ratio between TMT and SST trends. All four ratios are tightly constrained in
CMIP simulations but diverge markedly in observations. Model trend ratios between WV and temperature are closest to
observed ratios when the latter are calculated with datasets exhibiting larger tropical warming of the ocean surface and
troposphere. For the TMT/SST ratio, model–data consistency depends on the combination of observations used to estimate
TMT and SST trends. If model expectations of these four covariance relationships are realistic, our findings reflect either a
systematic low bias in satellite tropospheric temperature trends or an overestimate of the observed atmospheric moistening
signal. It is currently difficult to determine which interpretation is more credible. Nevertheless, our analysis reveals
anomalous covariance behavior in several observational datasets and illustrates the diagnostic power of simultaneously
considering multiple complementary variables.
KEYWORDS: Climate change; Satellite observations; Climate models; Ensembles; Model evaluation/performance

1. Introduction benchmark simulations have exposed and in some cases reduced

systematic errors in model representation of many different as-
Since publication of the first assessment report of the
pects of Earth’s climate (Gates et al. 1999; Randall et al. 2007;
Intergovernmental Panel on Climate Change (IPCC) in 1990,
Flato et al. 2013; Sperber et al. 2013; Bellenger et al. 2014).
there have been major improvements in our ability to model
In tandem with advances in modeling, there have been im-
the climate system (Randall et al. 2007; Trenberth et al. 2007;
provements in the forcings used in model simulations of his-
Flato et al. 2013; Hartmann et al. 2013). Thirty years ago, the
torical climate change (Solomon et al. 2011; Fyfe et al. 2013;
climate science community performed single simulations with
Schmidt et al. 2014; Checa-Garcia et al. 2018). Observations
a small number of pioneering atmosphere–ocean models.
have also improved with advances in the ability of scientists to
Today, more complex Earth system models (ESMs) are used to
identify and adjust for residual inhomogeneities in the data
generate large multimodel and single-model ensembles of
(Wentz and Schabel 1998; Mears et al. 2003; Mears and Wentz
simulations (Kay et al. 2015; Fyfe et al. 2017; Eyring et al. 2019;
2005; Fu and Johanson 2005; Karl et al. 2006, 2015; Po-Chedley
Deser et al. 2020). Over the last several decades, standard
et al. 2015; Hausfather et al. 2017; see section 6). This evolution
of models, forcings, and observations is ongoing.
Denotes content that is immediately available upon publica- The Fifth IPCC Assessment Report, published in 2013, re-
tion as open access. lied on CMIP5 simulations performed with roughly four dozen
models (Taylor et al. 2012). The upcoming sixth IPCC assess-
ment will evaluate output from a larger collection of CMIP6
Supplemental information related to this paper is available models and an expanded set of experiments (Eyring et al. 2016,
at the Journals Online website: https://doi.org/10.1175/JCLI-D-20- 2019). Our interest here is in comparing atmospheric temper-
0768.s1.
ature changes in CMIP5, CMIP6, the latest satellite data
(Mears and Wentz 2017; Zou and Wang 2011; Spencer et al.
Corresponding author: Benjamin D. Santer, [email protected] 2017), and a state-of-the-art reanalysis of weather observations

DOI: 10.1175/JCLI-D-20-0768.1
Ó 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright
Policy (www.ametsoc.org/PUBSReuseLicenses).
Unauthenticated | Downloaded 08/08/22 12:49 AM UTC
6282 JOURNAL OF CLIMATE VOLUME 34

with a weather forecast model (Simmons et al. 2020). We seek to the Microwave Sounding Unit (MSU) and the Advanced
determine 1) whether there are important differences between Microwave Sounding Unit (AMSU; Mears and Wentz 2017;
atmospheric temperature changes in CMIP5 and CMIP6 and 2) Spencer et al. 2017; Zou et al. 2018). Microwave emissions are
whether models and observations show consistency in well- proportional to the temperature of broad atmospheric layers.
understood physical constraints on tropical behavior: the ampli- By measuring at different microwave frequencies, MSU and
fication of tropical warming with increasing height, and the ratios AMSU provide estimates of temperatures at different heights.
between trends in tropical water vapor and trends in temperature Here, we analyze TLT, TMT, and the temperature of the lower
at different levels. We show that the combination of these con- stratosphere (TLS).
straints provides new information on model–data consistency. We rely on TLS and TMT datasets produced by Remote
There are several reasons for our focus on atmospheric tem- Sensing Systems (RSS; Mears and Wentz 2016), NOAA’s
perature. First, discrepancies between modeled and observed at- Center for Satellite Applications and Research (STAR; Zou
mospheric temperature changes have received scientific and and Qian 2016), and the University of Alabama in Huntsville
political attention for over 20 years (National Research Council (UAH; Spencer et al. 2017). Only RSS and UAH supply TLT
2000; Karl et al. 2006; Thorne et al. 2011; Fu et al. 2011; Po- measurements. We use the most recent dataset versions: RSS
Chedley and Fu 2012; U.S. Senate 2015; Santer et al. 2017a,b; 4.0, STAR 4.1, and UAH 6.0. The University of Washington
McKitrick and Christy 2020; Po-Chedley et al. 2021). Determining (UW) also produces a TMT dataset, but this is available for the
the causes of these differences remains a priority. Second, esti- tropics only (Po-Chedley et al. 2015). We did not use UW TMT
mates of atmospheric temperature from satellites have recently data for the present study.
undergone important revision, primarily due to improved un- We consider three different versions of the RSS atmospheric
derstanding of the effects of drifts in satellite orbits and instrument temperature data. As noted in Mears and Wentz (2017), ‘‘a
calibration (Po-Chedley et al. 2015; Mears and Wentz 2016, 2017; total of nine MSU instruments cover the period from 1978 to
Zou and Qian 2016; Zou et al. 2018; Spencer et al. 2017). 2005, followed by a series of AMSU instruments that began in
Reanalysis models and data assimilation systems have also mid-1998 and continue to the present’’ (p. 7695). MSU and
evolved (Hersbach et al. 2020; Simmons et al. 2020). Our goal is to AMSU do not measure at the same microwave frequencies;
reassess model–data consistency in the light of these improve- different plausible choices can be made in merging their esti-
ments to observations, models, and external forcings. mated brightness temperatures.
The structure of our paper is as follows. Sections 2 and 3 Mears and Wentz (2016) employed three approaches to
introduce the observational and model data analyzed in our merge MSU and AMSU data:
study. Section 4 discusses basic features of atmospheric tem-
1) MSU and AMSU measurements were used during the
perature time series and trends. Trend comparisons are over
merge period from mid-1998 to 2003.
the full satellite era (1979–2019), a period of stratospheric
2) Only AMSU data were used after 1999. MSU data were
ozone depletion (1979–2000), and a period of ozone recovery
excluded after 1999.
(2001–2019). Section 5 examines the relative magnitudes of
3) MSU data were used after 1999. AMSU data were excluded
forced and unforced temperature changes on different time
before 2003.
scales, and considers whether observed changes are consistent
with results from the forced simulations. The statistical meth- These approaches are referred to subsequently as ‘‘base-
odology in section 5 follows Santer et al. (2011) and is provided line,’’ ‘‘AMSU merge,’’ and ‘‘MSU merge,’’ respectively, and
in the online supplemental material (SM) with only minor are described in more detail in the SM. In sections 5 and 6, we
modifications. Section 6 focuses on the covariability of differ- address the question of whether these three RSS datasets yield
ent aspects of tropical climate change. We examine ratios be- different statistical inferences regarding the correspondence
tween tropical trends in column-integrated water vapor (WV) between simulated and observed measures of climate change.
and sea surface temperature (SST), WV and the temperature All satellite temperature datasets analyzed here are in the
of the lower troposphere (TLT), WV and the temperature of form of monthly means on the same 2.583 2.58 latitude–
the mid- to upper troposphere (TMT), and between TMT and longitude grid. Near-global averages of TLS, TMT, and TLT
SST. These four ratios are compared in observations and were calculated over areas of common coverage in the RSS,
multimodel and single-model ensembles. Prospects for using UAH, and STAR datasets (82.58N–82.58S for TLS and TMT,
such covariability information to constrain divergent observa- and 82.58N–708S for TLT). At the time this analysis was per-
tions are considered in section 7. Appendixes A and B provide formed, satellite temperature data for full 12-month years were
information regarding the calculation of synthetic satellite tem- available for the 492-month period from January 1979 to
peratures and the adjustment of tropospheric layer-average December 2019.
temperature for stratospheric influence.
b. SST data
Section 6 considers two ratio statistics involving SST. The
2. Observational data first is R{WV/SST}, the ratio between tropical trends in WV and
SST (Wentz and Schabel 2000; Held and Soden 2006; Mears
a. Satellite temperature data
et al. 2007; Mears and Wentz 2016). The second is R{TMT/SST},
Since late 1978, NOAA polar-orbiting satellites have mon- the ratio of tropical TMT and SST trends (Wentz and Schabel
itored the microwave emissions from oxygen molecules using 2000; Santer et al. 2005; Po-Chedley et al. 2015). We seek to

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6283

determine whether simulated and observed values of these Maritime Continent. To avoid the introduction of trend biases
ratio statistics are consistent, and how model–data agreement arising from coverage changes, we imposed a ‘‘fixed coverage’’
is affected by structural uncertainty in observed SST data. This mask; that is, our analysis of the satellite WV data was re-
uncertainty arises from differences in raw data, the methods stricted to the subset of grid points with continuous coverage
used to adjust raw data for known inhomogeneities, treatment over the 384-month analysis period. After regridding model
of sea ice, and the decisions made in merging information from WV data to the observational grid, the same fixed coverage
ship-based measurements, buoys, floats, and satellites (Karl mask was applied to all model simulations of historical climate
et al. 2006, 2015; Morice et al. 2012; Hausfather et al. 2017). change.
We quantify structural uncertainty in SST data by calculating
d. Reanalysis data
R{WV/SST} and R{TMT/SST} with four commonly used observa-
tional records: Reanalyses employ an atmospheric numerical weather
forecast model with no changes over time in the model itself
1) Version 2 of the Centennial In Situ Observation-Based
(Bengtsson and Shukla 1988; Kalnay et al. 1996). They provide
Estimates of the Variability of SST and Marine Meteorological
a well-tested framework for blending and constraining assim-
Variables (COBE; Hirahara et al. 2014).
ilated weather information from different sources; each source
2) Version 5 of the NOAA Extended Reconstructed SST
is typically characterized by different accuracy and different
dataset (ERSST; Huang et al. 2017).
temporal and spatial coverage.
3) Version 1 of the Hadley Center Sea Ice and SST dataset
The ERA5 product of the European Centre for Medium-
(HadISST; Rayner et al. 2003).
Range Weather Forecasts (ECMWF) recently superseded the
4) Version 4 of the Hadley Center SST dataset (HadSST;
ERA-Interim. ERA5 was generated with a high-resolution
Kennedy et al. 2019).
version (;31-km horizontal resolution, 137 vertical levels) of
All datasets except HadSST are spatially complete over the the ECMWF operational forecast model and a 4D variational
ocean domain of interest (208N–208S). data assimilation system (Hersbach et al. 2020). According to
Simmons et al. (2020), ERA5 exhibited ‘‘a pronounced cold
c. Satellite water vapor data
bias for the years 2000–06’’ (p. 1).
The satellite WV data used here were produced by RSS and ERA5.1, which spans the affected 2000–06 period, corrects
are from 11 different satellite-based microwave radiometers this error and yields ‘‘analyses with better global-mean tem-
(Wentz 2013). The procedures for intercalibrating and merging peratures in the stratosphere and uppermost troposphere than
information from these instruments and for estimating uncer- provided by ERA5’’ (Simmons et al. 2020, p. 1). Inclusion of
tainties in satellite WV trends are described in detail elsewhere ERA5.1 results allows us to test whether blending model and
(Mears et al. 2018). The WV retrievals are based on mea- observational information in a state-of-the-art reanalysis frame-
surements of microwave emissions from the 22-GHz water work provides layer-average atmospheric temperature trends
vapor absorption line. The distinctive shape of this line pro- similar to those available from actual RSS, STAR, and UAH
vides robust retrievals. The signal-to-noise ratio (S/N) for de- satellite data. We also examine WV and SST1 trends in ERA5.1,
tecting moistening in the lower troposphere by a measurement and we consider if the ‘‘within reanalysis’’ covariance relationships
of water vapor is several times larger than for MSU-based between tropical WV, SST, TLT, and TMT trends are similar to
measurements of air temperature (Wentz and Schabel 2000). those in other observational datasets and in CMIP models.
Relative to WV information from radiosondes and early re-
analysis products, the RSS WV dataset was judged by Trenberth
3. Model output
et al. (2005) to provide the most credible estimate of means,
variability, and trends over oceans. a. CMIP5 simulations
While alternative satellite WV datasets exist, they span
We used model TLS, TMT, TLT, SST, and WV output from
substantially shorter time intervals than the RSS WV data
phase 5 of the Coupled Model Intercomparison Project
(Jiang et al. 2019). At the time our analysis was performed,
(CMIP5) (Taylor et al. 2012). The description of the CMIP5
RSS WV data were available for the 384 months from January
datasets provided in the next two paragraphs follows Santer
1988 to December 2019. Since our primary interest is in mul-
et al. (2017a).
tidecadal changes in WV, we focus here on the RSS product.
Our focus here is on three different types of CMIP5 nu-
Due to the high emissivity of the land surface, the RSS WV
merical experiment: 1) simulations with estimated historical
retrievals are provided over oceans only. We analyze WV
changes in human and natural external forcings, 2) simulations
trends spatially averaged over tropical oceans (208N–208S),
with twenty-first century changes in greenhouse gases and
where there is well-understood covariability between temper-
anthropogenic aerosols prescribed according to representative
ature and atmospheric moisture (Wentz and Schabel 2000;
Held and Soden 2006; Mears et al. 2007; O’Gorman and
Muller 2010).
Because of changes in satellite capabilities, footprint size, 1
SSTs in ERA5 were prescribed using version 2 of the HadISST
and rain and land masking, the spatial coverage of the RSS WV dataset until August 2007, and thereafter with data from the
data changes over time. This results in the systematic addition Operational Sea Surface Temperature and Ice Analysis (OSTIA).
of grid cells with WV data in the western Pacific and near the See Table 7 in Hersbach et al. (2020).

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6284 JOURNAL OF CLIMATE VOLUME 34

concentration pathway 8.52 (RCP8.5; Meinshausen et al. 2011), satellite era (Riahi et al. 2017), so the choice of scenario is un-
and 3) preindustrial control runs with no changes in external likely to affect our model-versus-data comparisons.
influences on climate. In the case of TMT, TLT, SST, and WV, we analyzed 166
Most CMIP5 historical simulations end in December 2005. realizations. For reasons discussed in section 3c, the sample
RCP8.5 simulations were initiated from conditions of the climate size was smaller for TLS (116 extended HIST realizations
system at the end of the historical run. To avoid truncating performed with 21 models). Further details of the CMIP6 ex-
comparisons between modeled and observed climate change tended HIST and control simulations are provided in Tables S3
trends in December 2005, we spliced together output from the and S4, respectively.
historical simulations and the RCP8.5 runs. We refer to these
c. Large initial condition ensembles
spliced simulations subsequently as ‘‘extended HIST’’ runs.
In total, we analyzed 123 individual extended HIST realiza- Large initial condition ensembles (LEs) are valuable tools
tions performed with 28 different CMIP5 models. We excluded for separating forced and unforced climate change (Deser et al.
models that did not consider the scattering and absorption of 2012; Fyfe et al. 2017; Deser et al. 2020). Individual LE
radiation by stratospheric volcanic aerosols (Santer et al. 2013), members are generated with the same model and external
and therefore lack short-term lower-stratospheric warming sig- forcings, but are initialized from different conditions of the
nals after the eruptions of El Chichón in 1982 and Pinatubo in climate system. Each LE member provides a unique realiza-
1991. Including these models in the calculation of multimodel tion of the ‘‘noise’’ of natural internal variability superimposed
average (MMA) temperature changes would bias the MMA on the underlying climate ‘‘signal’’ (the response to the
estimate of volcanic TLS signals. changes in forcing). Typical LE sizes range from 30 to 100.
Details of the start dates, end dates, and lengths of the his- We used four different LEs to quantify uncertainties in tem-
torical integrations and RCP8.5 runs are given in Table S1 in perature and WV trends arising from multidecadal internal
the online supplemental material. Table S2 provides informa- variability. Two LEs applied CMIP5 historical forcing until 2005
tion on the 36 CMIP5 preindustrial control runs used to cal- and CMIP RCP8.5 forcing thereafter. The other two LEs relied
culate climate noise estimates. The control integrations allow on CMIP6 historical forcing until 2014 and SSP5 forcing from
us to determine S/N characteristics of atmospheric tempera- 2015 to 2100. The CMIP5 LEs were performed with version 1 of
ture changes (see section 5). the Community Earth System Model (CESM1; Deser et al.
2012) and with version 2 of the Canadian Earth System Model
b. CMIP6 simulations
(CanESM2; Fyfe et al. 2017; Swart et al. 2018). The CESM1 and
We also analyze sea surface temperature and atmospheric CanESM2 LEs consist of 40 and 50 members, respectively. The
temperature and moisture from model simulations performed two 50-member CMIP6 LEs relied on version 5 of CanESM
under phase 6 of CMIP. These simulations rely on newer (CanESM5; Swart et al. 2019; Fyfe et al. 2021) and on version 6 of
versions of CMIP5 models, often with more comprehensive the Model for Interdisciplinary Research on Climate (MIROC6;
representation of Earth system processes (Eyring et al. 2016), Tatebe et al. 2019). All four LEs used different strategies for
and with contributions from modeling groups that did not initialization of the individual ensemble members.5
participate in CMIP5. Efforts were made in CMIP6 to improve The CanESM5 LE exhibits anomalous aperiodic 1–2-month
the representation of external forcings with known systematic lower-stratospheric warming events in certain ensemble mem-
errors in CMIP5, such as volcanic and solar forcing in the early bers, an issue that is actively under investigation. These warming
twenty-first century (Solomon et al. 2011; Kopp and Lean 2011; events are sufficiently large to influence decadal-time scale TLS
Ridley et al. 2014; Schmidt et al. 2014; Gillett et al. 2016). trends but have minimal impact on decadal variability in tro-
At the time this research was performed, the CMIP6 archive pospheric temperature (or on the regression-based removal of
was still being populated with model simulation output. For stratospheric influence on TMT; see appendix B). We therefore
preindustrial control runs, output was available from 30 dif- excluded the CanESM5 LE from the multimodel analysis of
ferent models. For the analysis of forced simulations, the CMIP6 TLS trends, but used CanESM5 TLS data to remove
CMIP6 historical runs3 from 22 different models were spliced stratospheric influence from CanESM5 TMT data, and included
with results from scenario integrations. CanESM5 LE results in the multimodel analysis of TMT, TLT,
Multiple Shared Socioeconomic Pathway (SSP) scenarios WV, and SST.
were available for splicing (Riahi et al. 2017). We chose the SSP5
scenario here.4 SSP5 most closely approximates the radiative 4. Temperature time series and trends
forcing in the CMIP5 RCP8.5 simulation. The differences in
radiative forcing between the five SSPs are very small over the a. Lower stratosphere
Figure 1a shows time series of near-global averages of TLS.
The lower stratosphere cools over the full satellite era in all
2
RCP8.5 has radiative forcing of approximately 8.5 W m22 in
2100, eventually stabilizing at roughly 12 W m22.
3 5
The CMIP6 historical runs typically end in December 2014. Differences include the selected starting year for the simula-
4
In some publications this scenario is referred to as SSP5–8.5 tion, the strategy for perturbing initial conditions, and whether
because it reaches radiative forcing of 8.5 W m22 by 2100. We perturbations were applied to the atmosphere only or to the at-
adopt the SSP5 nomenclature of Riahi et al. (2017) here. mosphere and the ocean.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6285

solar radiation and outgoing longwave radiation by strato-

spheric volcanic aerosols (Robock 2000; Shine et al. 2003).
The CMIP6 multimodel average has an unrealistically small
TLS signal after El Chichón (Figs. 1a and 2). Based on the
MMA root-mean-square (RMS) errors between observed and
simulated volcanic TLS signals, the TLS response to El Chichón
is better captured in CMIP5 (Figs. 3a,c). For Pinatubo, the MMA
RMS error is smaller in CMIP6 (Figs. 3b,d). These CMIP5-
versus-CMIP6 differences are significant at the 5% level for the
El Chichón signal, but not for the Pinatubo signal (see the
online SM).
Volcanic signal differences in CMIP5 and CMIP6 arise from
multiple factors. These include differences in the type and time
history of information used for prescribing historical changes
in volcanic aerosol loadings, the aerosol optical properties, and
the implementation of these properties in calculating volcanic
radiative forcing (Thomason et al. 2018). Rather than pre-
scribing volcanic aerosol, at least one CMIP6 modeling group
calculated volcanic aerosol loadings based on observed estimates
of volcanically produced SO2 (Mills et al. 2016; Danabasoglu
et al. 2020). Separating and quantifying the impact of these
different factors on volcanic temperature signals requires sys-
tematic numerical experimentation (Rieger et al. 2020; Fyfe
et al. 2021).
Recent studies suggest that the Montreal Protocol led to a
partial recovery of lower-stratospheric ozone and TLS in the
early twenty-first century (Solomon et al. 2016, 2017; Philipona
et al. 2018; Petropavlovskikh et al. 2019; Banerjee et al. 2020).
All model and observational TLS datasets analyzed here ex-
hibit behavior consistent with ozone recovery: pronounced
global-mean cooling of the lower stratosphere over the ozone
FIG. 1. Time series of monthly mean near-global averages of the
depletion portion of the satellite record, defined here as the
temperature of (a) the lower stratosphere (TLS), (b) the mid- to period from 1979 to 2000, followed by weaker cooling or near-
upper troposphere (TMT), and (c) the lower troposphere (TLT). zero trends over the period of ozone recovery from 2001 to
For TLS and TMT, observations are the average of the RSS 2019 (Solomon et al. 2017; Philipona et al. 2018; Steiner et al.
‘‘baseline’’, STAR, and UAH satellite datasets and the ERA5.1 2020; Mitchell et al. 2020). No individual CMIP5 or CMIP6
reanalysis. Since STAR does not produce a TLT dataset, the ob- realization has larger lower-stratospheric cooling in the ozone
servational average for TLT was calculated with RSS ‘‘baseline’’, recovery period than in the ozone depletion period (Fig. 4).
UAH, and ERA5.1 only. CMIP5 synthetic satellite temperatures This underscores the fact that the nonlinear behavior of TLS
were computed from 123 realizations of historical climate change over the satellite era is dominated by the response to ozone
(‘‘extended HIST’’) performed with 28 models. For CMIP6, 116
forcing, not by multidecadal internal variability (Solomon
extended HIST realizations were used for TLS and 166 realizations
for TMT and TLT (performed with 21 and 22 models, respec-
et al. 2017).
tively). All temperature changes are defined as anomalies relative The multimodel average TLS trends for the ozone depletion
to climatological monthly means over 1979–2019. TMT is adjusted and recovery periods are (respectively) 20.368 and 20.078C
for the contribution it receives from stratospheric cooling (see decade21 in CMIP5 and 20.268 and 20.068C decade21 in
appendix B). Calculation of the multimodel average (MMA) in- CMIP6. During the ozone depletion period, the larger multi-
volves first averaging over realizations of an individual model, then model average lower-stratospheric cooling in the older CMIP5
averaging over models. simulations is in better accord with satellite TLS trends, which
range from 20.428 to 20.498C decade21. This is partly due to
observational datasets and model extended HIST simulations. the larger (negative) ozone-induced stratospheric radiative
The main cause of this cooling is human-induced depletion forcing in CMIP5 (Checa-Garcia et al. 2018).
of stratospheric ozone, with a smaller contribution from an- Other factors may also contribute to reduced lower-
thropogenic increases in atmospheric CO2 (Solomon 1999; stratospheric cooling in CMIP6 over 1979–2000. These fac-
Ramaswamy et al. 2006; Thompson et al. 2012; Aquila et al. tors include CMIP5-versus-CMIP6 differences in forcing from
2016; Maycock et al. 2018; Bandoro et al. 2018). Satellite-era tropospheric ozone (Checa-Garcia et al. 2018), volcanoes (see
decreases in TLS are punctuated by large episodic warming above) and stratospheric water vapor (Keeble et al. 2021),
signals after the major eruptions of El Chichón in 1982 and possible differences in the forced response of tropical upwell-
Pinatubo in 1991. Warming arises from absorption of incoming ing (Ball et al. 2020), and whether ozone in models was

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6286 JOURNAL OF CLIMATE VOLUME 34

FIG. 2. Time series of monthly mean anomalies of the temperature of the lower stratosphere (TLS) in CMIP6
extended HIST simulations. Results are for 21 individual CMIP6 models (in gray) and for the RSS ‘‘baseline’’
satellite data (in red). (bottom right) The CMIP6 multimodel average. All anomalies are spatially averaged over
82.58N–82.58S and are defined relative to climatological monthly means over 1979–2019. The number of extended
HIST realizations is indicated in parentheses. Vertical lines denote the times of maximum lower-stratospheric
warming in the RSS ‘‘baseline’’ data after the eruptions of El Chichón and Pinatubo.

prescribed or calculated with interactive ozone chemistry (Lin average TLS trends from the two CMIP phases are more similar
and Ming 2021). in their zonal-mean structure, except at high latitudes in the SH
The zonal-mean structure of TLS trends may provide some (Fig. 5b). More detailed analyses and more systematic numerical
diagnostic clues. Over the ozone depletion period, the smaller experimentation will be required to quantify the relative con-
global-mean lower-stratospheric cooling in the CMIP6 MMA tributions of forcing, response, chemistry, and dynamics to dif-
(relative to the CMIP5 MMA) arises primarily from the tropics ferences between CMIP5 and CMIP6 TLS trends (Solomon
(Fig. 5a). During the ozone recovery period, the multimodel et al. 2017; Checa-Garcia et al. 2018; Fyfe et al. 2021).

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6287

FIG. 3. Root-mean-square (RMS) differences between simulated and observed volcanic signals in lower-
stratospheric temperature in (a),(b) CMIP5 and (c),(d) CMIP6 models. RMS differences were calculated for
24-month periods after (a),(c) the 1982 eruption of El Chichón and (b),(d) the 1991 Pinatubo eruption. The obser-
vational target is the RSS ‘‘baseline’’ TLS time series, spatially averaged over 82.58N–82.58S. Blue dots denote RMS
values from individual realizations of the CMIP5 and CMIP6 extended HIST runs. Horizontal bars are average RMS
differences for individual models. The dashed vertical lines are the multimodel average RMS differences, calculated
by first averaging RMS values over a model’s individual realizations, and then averaging over models.

b. Troposphere larger response to anthropogenic aerosol forcing in CMIP6

(Flynn and Mauritsen 2020; Fyfe et al. 2021).
Multidecadal warming of the global troposphere is a ubiq-
In the four single-model large ensembles, the spread of TMT
uitous feature of the observations and all CMIP5 and CMIP6
and TLT trends arising from internal variability is substantial,
forced simulations (Figs. 1b,c). Over the full satellite era, the
spanning 31%–47% of the trend spread in the CMIP5 and
MMA tropospheric warming rate is very similar in CMIP5 and
CMIP6 multimodel ensembles (Fig. 6a).6 These results are
CMIP6 (0.288 and 0.298C decade21, respectively). This holds
consistent with other recent comparisons of LE spread to
both for TMT and TLT (Fig. 6a). The similarity of the CMIP5
and CMIP6 results is noteworthy given that CMIP6 has a larger
number of models with higher transient climate response
(TCR) and higher effective climate sensitivity (ECS) (Zelinka 6
This percentage represents (sLE /sCMIP ) 3 100, where sLE is the
et al. 2020; Flynn and Mauritsen 2020; Meehl et al. 2020). An standard deviation of the sampling distribution of trends in an in-
independent analysis of surface temperature supports our dividual CMIP5 LE or CMIP6 LE and sCMIP is the standard de-
finding: despite higher average TCR and ECS in CMIP6, the viation of the sampling distribution of ensemble-mean trends in the
MMA historical surface warming rate is comparable in older corresponding CMIP5 or CMIP6 multimodel ensemble contain-
and newer generations of CMIP models, possibly due to a ing the LE.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6288 JOURNAL OF CLIMATE VOLUME 34

UAH-derived ratio, but all other observational datasets have

R{TMT/TLT} values significantly less than one. The decisions
made by RSS in merging MSU and AMSU have limited impact
on TLT trends and substantial impact on TMT trends (Fig. 6a),
thereby introducing large spread in the three RSS-derived
R{TMT/TLT} values. The RSS ‘‘MSU merge’’ dataset, which has
the largest global-mean TMT trend, is closest of the three to
the model expectations of R{TMT/TLT} (Fig. 6b).
It is now recognized that there were systematic deficiencies
in the early twenty-first century solar and volcanic forcing used
in CMIP5 (Kopp and Lean 2011; Solomon et al. 2011; Flato
et al. 2013; Schmidt et al. 2014). Efforts were made to improve
representation of both forcings in CMIP6 (Eyring et al. 2016;
Gillett et al. 2016; Thomason et al. 2018; Rieger et al. 2020). We
find, however, that CMIP5 and CMIP6 multimodel average
trends in TMT are virtually identical over 2001–19 (Fig. 7).
Since other external forcings also changed between these two
generations of models (Checa-Garcia et al. 2018; Fasullo et al.
2021, manuscript submitted to Nat. Climate Change), isolating
the climate impact of improvements in volcanic or solar forcing
FIG. 4. Least squares linear trends in near-global average lower- is challenging. Such diagnosis will benefit from simulations in
stratospheric temperature over ozone depletion and ozone recov- which the same physical climate model is run with different
ery periods (1979–2000 and 2001–19, respectively). Model results versions of individual forcings (Fyfe et al. 2021; Fasullo et al.
are from 123 and 116 extended HIST simulations performed with 2021, manuscript submitted to Nat. Climate Change).
28 different CMIP5 and 21 different CMIP6 models, respectively. Tropospheric trends in ERA5.1 exhibit several notable dif-
CMIP5 trends include results from the 40-member CESM1 and ferences relative to the satellite datasets (Hersbach et al. 2020).
50-member CanESM2 large ensembles (LEs). CMIP6 trends in- Reanalysis TMT trends are smaller than in all satellite datasets
corporate the 50-member MIROC6 LE. Observed estimates of over 1979–2000 and larger than in all satellite datasets over
TLS trends rely on satellite data (RSS, STAR, and UAH) and the
2001–19 (Fig. 7). Over the 2002–18 period covered by Global
ERA5.1 reanalysis. Three different versions of the RSS data are
shown. The 1:1 line (with trends of equal size over the ozone de-
Positioning Satellite (GPS) radio occultation measurements,
pletion and ozone recovery periods) is marked in purple. The both GPS data and radiosondes yield trends in the middle
shaded ellipses are the 2s confidence intervals for each of the three troposphere that are in reasonable accord with the ERA5.1
LEs. For information on spatial averaging and calculation of results (Steiner et al. 2020).
multimodel averages, refer to Fig. 1. While the satellite data analyzed here are derived from
measurements of microwave emissions alone, ERA5.1 uses a
state-of-the-art 4D-variational data assimilation system to con-
multimodel ensemble spread (Mitchell et al. 2020; Po-Chedley strain a weather forecast model with a wide range of multivari-
et al. 2021). able measurements from satellites, radiosondes, and surface
Observed trends in global-mean tropospheric temperature stations (Hersbach et al. 2020; Simmons et al. 2020). Detailed
range from 0.138 to 0.198C decade21 for TMT and from 0.138 to observing system experiments can help to understand the impact
0.218C decade21 for TLT (Fig. 6a). For TLT, over 84% of the of different features of the assimilation system and assimilated
total number of CMIP5 and CMIP6 extended HIST realiza- data (Bormann et al. 2019). Such studies will be useful in rec-
tions analyzed here have trends exceeding the largest obser- onciling the trend differences found here and elsewhere (Steiner
vational result; the corresponding figure is 91% for corrected et al. 2020) between microwave sounders and ERA5.1.
TMT trends. Related work suggests that the smaller observed
warming is partly due to an unusual manifestation of natural
5. Signal-to-noise properties and model–data signal
internal variability. Model realizations with phasing and am-
differences
plitude of internal variability similar to the observations yield
global-mean and tropical tropospheric temperature trends that In previous statistical comparisons of modeled and observed
are closer to satellite results (Po-Chedley et al. 2021). temperature changes, discussion often focused on the appro-
In all individual extended HIST realizations, the ratio priateness of different comparison periods (Santer et al. 2011).
R{TMT/TLT} between global-mean trends in TMT and TLT is This can be uninformative if attention is restricted to a short
close to unity (Fig. 6b). This narrow range occurs despite dif- segment of the overall temperature record. Here we analyze
ferences in external forcings, ECS, and internal variability in atmospheric temperature changes over all NL maximally
the multimodel and single-model ensembles, and despite dif- overlapping L-year periods (see the SM). We consider four
ferences in the patterns of warming in TMT and TLT (Santer different values of L: 10, 20, 30, and 40 years. For each value of
et al. 2019). The 5th–95th percentile ranges of the CMIP5 L, sampling variability is reduced by averaging over all NL
and CMIP6 R{TMT/TLT} sampling distributions encompass the individual measures of temperature change. As we show

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6289

FIG. 5. Zonal-mean trends in monthly mean lower-stratospheric temperature over the (a) ozone depletion and
(b) ozone recovery periods. For information regarding the numbers of CMIP5 and CMIP6 models and extended
HIST realizations, calculation of multimodel averages, spatial averaging, and observational data, refer to Fig. 1.

below, examining behavior averaged over a particular time within the respective CMIP5 and CMIP6 distributions of forced
scale can have diagnostic value. trends. The same is true for observed TLS trends on longer 30-
Figure 8 shows two different types of statistic: trends and and 40-year time scales (Figs. 8c,d). For TMT and TLT, how-
regression coefficients. Results are from individual observa- ever, only observed datasets with larger tropospheric warming
tional datasets and from distributions of statistics in forced and rates are within the model 30- and 40-year distributions of forced
unforced simulations. trends. The UAH-inferred warming on these time scales is in-
Consider the trend results first. Rows 1–3 of Fig. 8 display variably smaller than model expectations (Figs. 8g,h,k,l).
trends in TLS, TMT, and TLT, respectively, for our four se- Amplification of warming with increasing height is a well-
lected values of the time scale L. With increasing L, the am- known and well-understood property of the tropical atmo-
plitude of internally generated trends decreases. As a result, sphere (Stone and Carlson 1979; Santer et al. 2005; Held and
the standard deviations of the forced and unforced trend dis- Soden 2006). Figures 8m–p display one measure of tropical
tributions decrease. For all three atmospheric layers, forced amplification behavior—the regression coefficient b{TMT:TLT}
and unforced trend distributions are completely separated at between time series of tropical ocean averages of TMT and
L 5 40 years (Figs. 8d,h,l). This is a simple visual illustration of TLT. All model and observational values of b{TMT:TLT} are
the dependence of signal and noise on time scale, and of the greater than 1, indicating that temperature changes in the mid-
difficulty in their separation on shorter, noisier time scales of 1– to upper troposphere exceed those in the lower troposphere.
2 decades (Santer et al. 2011). The means and widths of the CMIP5 and CMIP6 sampling
Despite the evolution in model complexity and resolution distributions of b{TMT:TLT} are relatively insensitive to increases
between CMIP5 and CMIP6, the sampling distributions of in L, and show substantial overlap for the forced and unforced
unforced atmospheric temperature trends are remarkably similar runs. The model results imply that b{TMT:TLT} is both invariant to
in the two generations of coupled models. The same is true for time scale and insensitive to forcing, and that it may impose a
the sampling distributions of forced trends on 10- and 20-year robust, physically based constraint on observations (Santer et al.
time scales. On longer 30- and 40-year time scales, however, 2005; Held and Soden 2006).
small differences are apparent in the distributions of forced Observational values of b{TMT:TLT} show a number of in-
tropospheric temperature trends in CMIP5 and CMIP6. These teresting features. First, the ERA5.1 and RSS ‘‘MSU merge’’
may arise because CMIP5 and CMIP6 do not have identical results are well within the range of model expectations on all
multidecadal evolution of certain external forcings (Checa- four time scales considered here. In terms of this tropical am-
Garcia et al. 2018; Fyfe et al. 2021). plification metric, therefore, there is no fundamental discrep-
Figure 8 also provides information on the consistency be- ancy between simulations and all observations.
tween global-mean temperature trends in observations and the Second, as in the model simulations, b{TMT:TLT} is invariant
extended HIST simulations. On shorter 10- and 20-year time to time scale for UAH, ERA5.1, and the RSS ‘‘MSU merge’’
scales, all observed TLS, TMT, and TLT trends are contained case. While the three RSS sensitivity tests have almost identical

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6290 JOURNAL OF CLIMATE VOLUME 34

FIG. 6. (a) Scatterplot of linear trends in near global-mean lower-tropospheric temperature

(TLT) and mid- to upper-tropospheric temperature (TMT) and (b) histograms of the TMT/
TLT trend ratio. All trends are over 1979–2019. TMT is corrected for lower-stratospheric
cooling. The multimodel averages include information from the 50- and 40-member CanESM2
and CESM1 LEs (for CMIP5) and from the 50-member CanESM5 and MIROC6 LEs (for
CMIP6). The shaded ellipses in (a) are the 2s confidence intervals for each LE. Because TLT
is not produced by STAR, the STAR TMT trend is plotted as a horizontal line in (a). Selected
isopleths of equal values of the TMT/TLT trend ratio are denoted by dashed gray lines in (a).
For further details of CMIP5 and CMIP6 realizations and models, calculation of multimodel
averages, spatial averaging, observational data sources, and fits to histograms, refer to caption
of Fig. 1 and the online supplemental material (SM).

b{TMT:TLT} values for L 5 10 years (Fig. 8m), the RSS baseline value is due to a change in the method used by the UAH group
and ‘‘AMSU correct’’ datasets yield regression coefficients to estimate TLT (Spencer et al. 2017). The impact of this
that decrease in size as L increases, and are generally outside change7 was to increase the height of the effective weighting
the range of model results for 30- and 40-year time scales (Figs. function for TLT, thus decreasing the vertical separation be-
8o–p). On these longer time scales, the maximally overlapping tween the TLT and corrected TMT weighting functions. This
L-year windows always sample the 1998–2003 transition be-
tween earlier and more advanced microwave sounders, and
thus are more likely to reflect the impact of different merging 7
The change involved transitioning from a multiangle to a
choices on amplification behavior (see section 2a). multichannel method for calculating TLT. Spencer et al. (2017)
Third, the UAH b{TMT:TLT} value is ;1.1 on all four time scales regard the latter as a ‘‘more robust method of (T)LT calcula-
and is smaller than almost all model results. The anomalous UAH tion’’ (p. 121).

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6291

ratios as R{WV/SST}, R{WV/TLT}, and R{WV/TMT}, respectively.

The relationship between temperature and saturation vapor
pressure changes is governed by the Clausius–Clapeyron (C-C)
equation (Iribarne and Godson 1981). If relative humidity re-
mains approximately constant as temperature increases, C-C
predicts the increase in columnar content of WV (Wentz
and Schabel 2000; Held and Soden 2006; Mears et al. 2007;
O’Gorman and Muller 2010).
The fourth property we examine, the trend ratio R{TMT/SST},
is a measure of the amplification of tropical SST changes in the
tropical troposphere. Its behavior is governed by moist ther-
modynamics (Stone and Carlson 1979; Held and Soden 2006).
The ratio R{TMT/SST} provides information that differs from
that of b{TMT:TLT}, the regression-based amplification metric
considered in section 5.9
In a climate model, these four ratios are internally and
physically consistent. The observed covariability of tropical
WV, tropospheric temperature, and SST should also exhibit
internal and physical consistency. As we show below, however,
observed R{WV/SST}, R{WV/TLT}, R{WV/TMT}, and R{TMT/SST} values
FIG. 7. As in Fig. 4, but for linear trends in global average mid- to can be inconsistent for certain combinations of observed data-
upper-tropospheric temperature (TMT) over 1979–2000 (x axis) sets, and may depart noticeably from model expectations.
and over 2001–19 (y axis). TMT is corrected for the influence of Such departures can have at least three explanations. First,
lower-stratospheric cooling. While Fig. 4 excluded TLS results WV, tropospheric temperature, and SST are measured inde-
from the 50-member CanESM5 LE because of anomalous TLS pendently by different instruments on different satellites and/
variability, TMT trends from the CanESM5 LE are minimally af- or measurement platforms. Each variable has different mea-
fected by this anomalous variability and are included here. The 1:1 surement accuracy and errors. These measurement differences
line (with TMT trends of equal size over the two periods) is marked can affect the estimated covariability between multidecadal
in purple. Simulated TMT trends are larger in the second analysis
trends in WV, tropospheric temperature, and SST.
period in approximately 90% of the realizations. In satellite data,
trends in the two periods are of roughly equivalent size.
Second, the tropospheric temperature and SST datasets
analyzed here were generated by multiple research groups. In
the case of TMT and TLT, each research group uses different
leads to a smaller amplification ratio. To maintain continuity procedures to adjust for drifts in satellite orbits and instrument
with previous tropical amplification studies (Santer et al. calibration, to merge measurements from multiple satellites,
2017b) and to increase the amplification signal, the model, and to merge brightness temperatures estimated from earlier
RSS, and ERA5.1 results shown here do not use the new UAH and more recent types of microwave sounders. For SST, groups
approach for calculating TLT. use different methods to blend information from ships, buoys,
drifting floats, and satellites, to adjust for changes over time in
how SST measurements were made, and to infill SSTs in data-
6. Covariability of different aspects of tropical climate
sparse regions. The decisions made in adjusting tropospheric
change
temperature and SST for these known nonclimatic influences
Properties of the climate system that are controlled by well- can affect trends (Karl et al. 2006, 2015; Hausfather et al. 2017;
understood physical mechanisms and are tightly constrained in Mears et al. 2011; Mears and Wentz 2016, 2017; Zou and Qian
model simulations may be useful for reducing large uncer- 2016; Zou et al. 2018; Spencer et al. 2017; Po-Chedley et al.
tainties in observed temperature trends (Santer et al. 2005). 2015), and can therefore influence the estimated covariability
We consider four such properties here. The first three prop- between real-world temperature and WV changes (or between
erties are ratios between tropical WV trends and trends in observed trends in SST and TMT). Trends in satellite WV data
tropical SST, TLT, and corrected TMT.8 We refer to these

9
The term b{TMT:TLT} was useful for examining whether the TMT
8
Because satellite WV data are available over ocean only, we and TLT time series produced by an individual research group
computed R{WV/TLT} and R{WV/TMT} using ‘‘ocean only’’ TLT and yielded internally consistent estimates of amplification behavior.
TMT trends. Horizontal temperature gradients are weak in the Notably, b{TMT:TLT} used TMT and TLT information from the same
tropical free troposphere, so whether we use TLT and TMT trends microwave sensors flown on the same satellites; in contrast, observed
calculated over ocean only or over land and ocean has minimal values of R{TMT/SST} provide information on the physical consistency
impact on our results. To be consistent in terms of the domain between multidecadal trends in SST and TMT measurements that
analyzed, the TMT trends in R{TMT/SST} also rely on data averaged are processed by different research groups, and that are obtained
over tropical oceans only. using different types of measurement platforms.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6292 JOURNAL OF CLIMATE VOLUME 34

FIG. 8. Trends and regression coefficients in CMIP5, CMIP6, and observations. Maximally overlapping L-year trends were calculated
from time series of monthly-mean, near-global spatial averages of (a)–(d) TLS, (e)–(h) TMT, and (i)–(l) TLT. (m)–(p) The regression
coefficient b{TMT:TLT}, a measure of amplification of warming in the tropical troposphere, was computed with maximally overlapping L-
year time series of monthly mean TMT and TLT, spatially averaged over ocean areas between 208N and 208S. The four selected time scales
shown here are (left to right) 10, 20, 30, and 40 years, respectively. Histograms of these L-year trends and regression coefficients are shown
for CMIP5 and CMIP6 extended HIST simulations and for preindustrial control runs. Histograms are weighted to account for model
differences in the number of extended HIST simulations or in control run length. For each histogram, results are normalized by the total
number of trend or regression coefficient samples. Fits to the model trend and b{TMT:TLT} distributions were performed with kernel density
estimation (see the SM). The vertical lines for the observed trends and regression coefficients are the averages across the maximally
overlapping L-year analysis periods. For trends in TMT, the RSS ‘‘MSU merge’’ and STAR results are almost identical.

are also sensitive to dataset construction choices (Mears et al. (Santer et al. 2005) and between temperature and WV (Mears
2018), but we currently have uncertainty estimates from RSS et al. 2007) are well captured by models (see section 7).
only.10 Figure 9 shows scatterplots of the individual trend compo-
Third, models may have incomplete or inaccurate repre- nents of the four ratio statistics. For each statistic, model results
sentation of the basic physics driving observed tropical cova- are tightly constrained in the CMIP5 and CMIP6 multimodel
riability relationships on multidecadal time scales. This seems ensembles. At least 96% of the variance in simulated WV
unlikely (Held and Soden 2006), particularly given the fact trends (plotted on the y axis in Figs. 9a–c) and in simulated
that on interannual time scales, observed tropical covariability TMT trends (plotted on the y axis of Fig. 9d) is explained by
relationships between surface and tropospheric temperature simulated trends in the independent (x axis) variable. This
indicates that the four covariance relationships of interest here
are relatively insensitive to model differences in the applied
historical forcings, the temperature and WV responses to these
10
We do not use the reanalysis-derived WV trend in estimating forcings, and the properties of simulated multidecadal internal
structural uncertainties in observed WV trends. Other research has variability. A related inference is that even though most of the
found possible problems with WV trends inferred from reanalysis mass of atmospheric water vapor resides in the lower tropo-
products (Bengtsson et al. 2004; Wang et al. 2020). sphere, simulated tropical SST, TLT, and TMT trends impose

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6293

FIG. 9. Scatterplot of tropical trends in (a) WV and SST, (b) WV and TLT, (c) WV and corrected TMT, and
(d) corrected TMT and SST. Trends are over 1988–2019, the period of availability of observed WV data from seven
different microwave radiometers (Mears et al. 2018), and were calculated with WV, TLT, TMT, and SST data
averaged over tropical oceans (208N–208S). For ERA5.1, the location of the purple pentagonal symbols is based on
reanalysis data only (i.e., the ERA5.1 WV trend in (a) is plotted against the ERA5.1 SST trend). All other ob-
servational symbols provide information on the joint variability between trends in different climate variables es-
timated by different research groups. In (a), for example, the satellite WV trend (which is available from RSS only)
is plotted against observed SST trends from ERSST, HadISST, COBE, and HadSST. In (b) and (c), the RSS WV
trend is plotted against TLT and TMT trends from five and six different satellite datasets, respectively. In (d), there
are four different observed SST trends and five different satellite TMT trends, yielding 4 3 5 combinations of SST
and TMT trends (plus the symbol denoting the relationship between the ERA5.1 TMT and SST trends). The x axis
position of observational symbols in (d) reflects the observed SST trend; the y axis position depends on the observed
TMT trend. The regression fits and slopes were estimated with orthogonal distance regression and are given
separately for CMIP5 and CMIP6 results (see the SM). The ERSST and HadSST trends are almost identical; this is
why the solid blue diamond and red star symbols overlap in (a) and (d).

similar constraints on simulated tropical WV trends—that is, amplifies with increasing height, the slope of the regression
there is no evidence that (on multidecadal time scales) SST or between temperature trends and moisture trends decreases.
TLT explains noticeably more of the WV variance than TMT. The regression slope for simulated tropical SST and TMT
The regression fits to the CMIP5 and CMIP6 trends are trends (1.6 for both CMIP5 and CMIP6; see Fig. 9d) is also
8.5% and 8.7% 8C21 for WV and SST, 6.3% and 6.4% 8C21 for consistent with moist adiabatic lapse rate (MALR) expecta-
WV and TLT, and 5.3% and 5.5% 8C21 for WV and TMT tions (Fu et al. 2004).
(Figs. 9a–c, respectively). The decrease in regression slope in Unlike the model covariance relationships in Fig. 9, all four
the progression from Figs. 9a to 9c reflects the fact that tropical sets of observed covariance relationships show substantial
temperature changes closely follow a moist adiabatic lapse spread. The tight clustering of model expectations and the large
rate (Stone and Carlson 1979). As the magnitude of warming observational uncertainty are clearer if we directly compare

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6294 JOURNAL OF CLIMATE VOLUME 34

FIG. 10. Histograms of the ratios between the model trends plotted in each of the four panels of Fig. 9. Results are
shown for (a) R{WV/SST}, (b) R{WV/TLT}, (c) R{WV/TMT}, and (d) R{TMT/SST}. Observational trend ratios in (a)–(c) are
plotted as vertical lines. In (d), trends from each of the five satellite TMT datasets analyzed here (the three RSS
versions, STAR, and UAH) can be paired with four different observed SST trends (from ERSST, HadISST, COBE,
and HadSST), yielding 5 3 4 different observed values of R{TMT/SST}, plus one value for the ratio between the
ERA5.1 TMT and SST trends (see Fig. 9 caption). Observed R{TMT/SST} values in (d) are plotted in six rows. There is
one row for each of the five satellite TMT datasets and one row for the reanalysis. The vertical spacing and y axis
location of rows is nominal; the vertical ordering of rows reflects the size of the observed tropical TMT trend over
1988–2019. The largest TMT trend (in the STAR dataset) has the largest y axis offset in (d). For details regarding fits
to the model histograms and histogram weighting, refer to the SM. Because the ERSST and HadSST trends are
almost identical, the ERSST-based trend ratios in (a) and (d) have been offset vertically. Other observational
results with similar ratios have also been offset for the sake of clarity.

trend ratios (Fig. 10).11 This comparison reveals that observed For all three ratios involving WV trends, there is minimal
SST and tropospheric temperature datasets with the largest overlap between simulations and observations; observed ra-
tropical warming over 1988–2019 have R{WV/SST}, R{WV/TLT}, and tios generally exceed model expectations. For R{WV/SST}, only
R{WV/TMT} ratios closest to the model results (Figs. 10a–c). the COBE SST trend leads to a result consistent with model
expectations (Fig. 10a). For both R{WV/TLT} and R{WV/TMT},
observed trend ratios are larger than almost all of the 289
model results (Figs. 10b,c).12 The agreement between model
11
The lowest and highest observational values for R{WV/SST},
R{WV/TLT}, R{WV/TMT}, and R{TMT/SST} vary by factors of 1.6, 1.7, 1.8,
and 2.9, respectively. The larger range for R{TMT/SST} arises because
12
there is appreciable observational uncertainty in both the numer- For each ratio, there are 123 values for CMIP5 and 166 for
ator and denominator of the ratio. In the three ratios involving CMIP6. For R{WV/TLT} and R{WV/TMT}, only 4 and 3 of the 289 ex-
WV, the structural uncertainty of observed trends can be estimated tended HIST realizations (respectively) have scaling ratios ex-
in the denominator only. ceeding the smallest observed value.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6295

simulated and observed values of R{WV/TLT}, R{WV/TMT}, and

R{TMT/SST}. For the latter ratio, results are averaged over the
individual Z-scores arising from structural uncertainty in ob-
served SST trends. The Z-scores in Fig. 11b involve R{WV/SST}
and R{TMT/SST}, with R{TMT/SST} results averaged over the in-
dividual Z-scores arising from structural uncertainty in satellite
tropospheric temperature trends. The reanalysis results for all
four ratios are shown in Fig. 11c.13
Under the assumption that the model-generated distribu-
tions of the four ratios are realistic representations of the
true (but uncertain) real-world covariance relationships, the
Z-scores allow us to make certain inferences about the likeli-
hood that individual observed SST and tropospheric temper-
ature datasets are consistent with model expectations and with
other observations. In Fig. 11a, for example, STAR and RSS
‘‘MSU merge’’—the datasets with the largest observed tropo-
spheric warming trends—are closest to the model expectations
of WV/tropospheric temperature trend ratios, and therefore
have the smallest Z-scores for R{WV/TLT} and R{WV/TMT}. In
contrast, the muted tropospheric warming in UAH leads to
R{WV/TLT} and R{WV/TMT} values that are significantly larger
than model expectations, thus leading to large UAH Z-scores
for these two ratios. Based on R{WV/TLT} and R{WV/TMT} alone,
therefore, we might infer that the smaller tropical tropospheric
warming trend is UAH is less credible.
FIG. 11. Normalized differences (Z-scores) between observed This inference assumes that the observed trend in tropical
scaling ratios and the mean of model scaling ratio distributions. WV is accurate. A substantially smaller observed WV trend
(a) Results for tests of R{WV/TLT} ratios based on five different
would decrease the UAH-derived R{WV/TLT} and R{WV/TMT}
satellite TLT datasets and for tests of R{WV/TMT} and R{TMT/SST}
ratios based on six different satellite TMT datasets. (b) Results ratios, bringing them in closer agreement with model expec-
involving tests of R{WV/SST} and R{TMT/SST} with four different ob- tations. Since we do not have estimates of the observed WV
served SST datasets. (c) Z-scores relying on tests performed with trend from multiple research groups, it is difficult to assess the
observed trend ratios computed using reanalysis data. All Z-scores likelihood that the true (but uncertain) real-world WV trend is
were calculated using the scaling ratio data in Fig. 10. For each ratio markedly smaller than the RSS WV trend estimate.
tested, the observed ratio is subtracted from the mean of the By considering the R{TMT/SST} ratio, however, we can bring
CMIP5 or CMIP6 sampling distribution of the ratio. These dif- in independently monitored observed SST data. This allows
ferences are normalized by the CMIP5 or CMIP6 standard devia- us to explore the constraint that observed SST trends impose
tion of the ratio’s sampling distribution; CMIP5 and CMIP6 on the size of observed TMT trends. All four observed
Z-scores are then averaged. For the R{TMT/SST} ratios in (a), there is
SST datasets, when considered in combination with the
an additional averaging step: each satellite TMT trend can be
paired with four different observed SST trends, yielding four dif- UAH TMT trend, lead to UAH-based R{TMT/SST} ratios that
ferent Z-scores (see rows in Fig. 10d). We average these four values are significantly smaller than climate model expectations
per TMT dataset. Likewise, each observed SST trend in (b) can be (Fig. 10d).14
paired with 5 different satellite TMT trends, yielding five different In summary, the reduced tropical tropospheric warming in
values of R{TMT/SST} (see columns in Fig. 10d). We average these UAH is not supported by 1) an independent estimate of at-
five values per SST dataset. The brown bars are average Z-scores mospheric moistening from satellite data, 2) all independent
for different types of scaling ratio. Observed ratios to the left of the estimates of observed sea surface warming, or 3) all model
dashed purple are within two standard deviations of the model expectations of R{WV/TLT}, R{WV/TMT}, and R{TMT/SST}.
trend ratio sampling distributions. The above analysis focused on comparing simulated and
observed measures of tropical covariability. It is also of interest
and observed R{WV/SST} values is closer, but depends on the to compare modeled and observed values of the individual
selected combination of observed TMT and SST datasets components of these covariability metrics. In the case of WV,
(Fig. 10d).
We calculated Z-scores to summarize and synthesize the
information in Fig. 10. For each observed ratio in Fig. 10, the 13
In our analysis of ERA5.1 data, we do not ‘‘pair’’ ERA5.1
Z-score is simply the difference between the observed result temperature or moisture trends with trends in other datasets; we
and the mean of the CMIP5 or CMIP6 multimodel average consider only the internal and physical consistency of temperature
ratio, normalized by the CMIP5 or CMIP6 standard deviation and moisture trends within the reanalysis.
of the sampling distribution of the ratio in question. The Z- 14
All four UAH-based R{TMT/SST} ratios are outside of the 5th–
scores in Fig. 11a are measures of the consistency between the 95th percentile range of model results.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6296 JOURNAL OF CLIMATE VOLUME 34

21% of the WV trends over 1988–2019 in the 289 CMIP5 and advantage (relative to single-variable comparisons) that results
CMIP6 extended HIST simulations are smaller than the are less sensitive to model-versus-observed differences in the
satellite-estimated WV trend in Fig. 9a. For SST, TLT, and phasing of internal variability (Santer et al. 2005; Po-Chedley
TMT trends over the same period, only 17%, 12%, and 12% of et al. 2021).
the model results are smaller than the largest observed trend We find significant differences between simulated and
(Figs. 9a–c, respectively). observed values of all trend ratios involving water vapor and
There are multiple interpretations of this finding. One in- tropospheric temperature. Observed ratios exceed model
terpretation is that the higher level of consistency between expectations in most cases (Figs. 10a–c). Observed datasets
simulated and observed tropical WV trends reflects a system- with larger warming of the tropical ocean surface and
atic low bias in observed tropical TLT and TMT trends over tropical troposphere yield ratios of R{WV/SST}, R{WV/TLT},
1988–2019. An alternative explanation is that the satellite WV and R{WV/TMT} that are closer to model results. Ratios be-
trend is overestimated. It is difficult to discriminate between tween moisture and temperature changes calculated with
these two possibilities without additional information, such as the UAH and HadISST datasets, which both have muted
well-quantified estimates of uncertainties in observed WV tropical warming over 1988–2019, are at least 10 standard
trends from different research groups. deviations removed from model expectations (Fig. 11).15
One interesting feature of Fig. 9 relates to the behavior of For R{TMT/SST}, model–data consistency depends on the se-
the ERA5.1. As noted above, the CMIP models show tight lected combination of observed datasets used to estimate
coupling between the multidecadal trends in tropical WV, SST, TMT and SST trends (Fig. 10d).
and tropospheric temperature. In contrast, the agreement be- One interpretation of our findings is that they are due to a
tween the CMIP expectations and reanalysis-based trend ratios systematic low bias in satellite tropospheric temperature
is noticeably better for R{WV/SST} than for either R{WV/TLT} or trends; that is, the size of the observed tropical moistening
R{WV/TMT} (cf. the relative distances from the regression lines signal is greater than can be explained by the independently
of the purple pentagonal symbols in Figs. 9a–c). This closer observed warming of the tropical troposphere. Alternately,
agreement is reflected in the lower Z-score for the ERA5.1- the observed atmospheric moistening signal may be over-
based R{WV/SST} ratio in Fig. 11c. Our results imply that some estimated. Given the large structural uncertainties in observed
aspect or aspects of the assimilation system or assimilated data tropical TMT and SST trends, and because satellite WV data
(Hersbach et al. 2020) may be affecting the internal and are available from one group only, it is difficult to determine
physical consistency of tropical temperature and moisture which interpretation is more credible.
trends in the reanalysis. What we can say with confidence, however, is that decisions
regarding how to merge MSU and AMSU TMT data have
substantial impact on observed tropical TMT trends. This is
7. Conclusions
evident from the three RSS sensitivity tests examined here
Relative to CMIP5, the more recent CMIP6 models have (Mears and Wentz 2016). These sensitivity tests point toward
higher resolution (on average), more complete numerical merging decisions as a significant contributory factor to un-
portrayal of Earth’s climate system, and nominally improved certainties in observed R{WV/TMT} and R{TMT/SST} trend ratios
representation of external forcings (Eyring et al. 2016). These (Figs. 10c,d).
advances do not guarantee improved agreement between Three further points are relevant to the question of whether
simulations and observations. This is apparent in at least two the model–observed differences in Figs. 10a–c are mainly due
aspects of model performance analyzed here: lower-stratospheric to underestimated observed tropospheric temperature trends
cooling over the ozone depletion period and the stratospheric or to an overestimated satellite WV trend. First, independent
temperature response to the El Chichón eruption. Understanding estimates of tropospheric temperature change from GPS ra-
why these features are more accurately represented in CMIP5 will dio occultation (RO) and radiosondes suggest that over the
require more systematic diagnostic efforts to disentangle evolu- 2002–18 period of overlap between MSU/AMSU and GPS-
tionary changes in models from evolutionary changes in model RO, tropospheric warming is smaller in microwave sounders
forcings (Fyfe et al. 2021). than in GPS-RO or radiosondes (Steiner et al. 2020). Second,
The development of satellite temperature datasets remains a there is some evidence that observational uncertainties may
work in progress. Adjustments for known nonclimatic factors be smaller in satellite WV data than in satellite tropospheric
can have significant impact on observed trends in tropospheric temperature data (Wentz 2013; see section 2c). Third, when
temperature, as well as on basic physical properties related to
tropospheric warming (Karl et al. 2006; Mears et al. 2011;
Mears and Wentz 2016, 2017; Zou and Qian 2016; Zou et al. 15
2018; Spencer et al. 2017; Po-Chedley et al. 2015). Multimodel To bring the UAH-derived value of R{WV/TMT} into agreement
with the regression slope of ;5.4% 8C21 estimated from Fig. 9c
and single-model large ensembles tightly constrain four such
would require that the RSS WV trend of 1.46% decade21 was
physical properties: the ratio between tropical trends in WV roughly a factor of 2 smaller. Such an error is well outside the WV
and SST, WV and TLT, WV and TMT, and TMT and SST. trend uncertainty assessed by RSS (Mears et al. 2018). An error
These are denoted here by R{WV/SST}, R{WV/TLT}, R{WV/TMT}, by a factor of ;2 in the observed WV trend would also be required
and R{TMT/SST}, respectively. Comparing modeled and ob- to obtain agreement between the HadISST-derived value of R{WV/
21
served values of such basic covariance relationships has the SST} and the regression slope of ;8.6% 8C in Fig. 9a.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6297

the individual trend components of our four trend ratios are APPENDIX B
examined, the agreement between models and observations is
better for WV and SST trends than for TMT or TLT trends. Method Used for Correcting TMT Data
These three lines of evidence, taken together with the results of
Trends in TMT estimated from microwave sounders receive a
the RSS sensitivity tests, suggest that underestimated observed
substantial contribution from the cooling of the lower strato-
tropospheric warming is plausible. This inference is predicated on
sphere (Fu et al. 2004; Fu and Johanson 2004, 2005; Johanson
the assumption that the model-based covariance constraints are
and Fu 2006). In Fu et al. (2004), a regression-based method was
realistic.
developed for removing the bulk of this stratospheric cooling
While our analysis does not definitively resolve the cause or
component of TMT. This method has been validated with both
causes of significant differences between modeled and observed
observed and model atmospheric temperature data (Fu and
tropospheric warming trends, it does illustrate the diagnostic
Johanson 2004; Gillett et al. 2004; Kiehl et al. 2005). We cal-
power of simultaneously considering multiple complementary
culated two different versions of corrected TMT, the first with
variables (Wentz and Schabel 2000). Our study also highlights
latitudinally fixed and the second with latitudinally varying
the strong internal and physical consistency between the model
regression coefficients. We refer to these subsequently as
constraints derived from multidecadal tropical trends in WV,
TMT1 and TMT2, respectively. The main text discusses cor-
TMT, and SST. Examining additional independently monitored
rected TMT1 only, and does not use the subscript 1 to identify
constraints may be helpful in reducing the currently large un-
corrected TMT.
certainties in observations of tropical climate change.
The regression equation applied in Fu and Johanson (2005)
for calculating corrected TMT is
Acknowledgments. We acknowledge the World Climate
Research Programme’s Working Group on Coupled Modelling,
TMT 5 a24 TMT 1 (1 2 a24 )TLS . (B1)
which is responsible for CMIP, and we thank the climate mod-
eling groups for producing and making available their model For TMT1, we use a24 5 1.1 at each latitude. For TMT2, a24 5
output. For CMIP, the U.S. Department of Energy’s Program 1.1 between 308N and 308S, and a24 5 1.2 poleward of 308. This
for Climate Model Diagnosis and Intercomparison (PCMDI) is consistent with how we have calculated TMT1 and TMT2 in
provides coordinating support and led development of software previous work (Santer et al. 2017b).
infrastructure in partnership with the Global Organization for The advantage of TMT2 is that lower-stratospheric cooling
Earth System Science Portals. This work was performed under makes a larger contribution to TMT trends at mid- to high
the auspices of the U.S. Department of Energy (DOE) by latitudes. The latitudinally varying regression coefficients in
Lawrence Livermore National Laboratory under Contract DE- TMT2 remove more of this extratropical cooling. We prefer to
AC52-07NA27344. At LLNL, B.D.S., S.P.-C., M.D.Z., and J.P. use the more conservative TMT1 here. In practice, the choice
were supported by the Regional and Global Model Analysis of TMT1 or TMT2 has minimal influence on the statistical
Program of the Office of Science at the DOE. S.P.-C. was also significance of differences between the modeled and observed
supported under LDRD 18-ERD-054. At M.I.T., S.S. was partly statistics of interest here (temperature trends and a regression-
supported by NSF-AGS Grant 1848863. All primary satellite, based measure of the amplification of warming with increasing
reanalysis, and model temperature data sets used here are pub- height in the tropical atmosphere).
licly available. Synthetic satellite temperatures calculated from
model simulations and the ERA 5.1 are provided at https:// REFERENCES
pcmdi.llnl.gov/research/DandA/. We thank Adrian Simmons at
Aquila, V., W. H. Swartz, D. W. Waugh, P. R. Colarco, S. Pawson,
ECMWF for assistance with ERA 5.1 data and internal re-
L. M. Polvani, and R. S. Stolarski, 2016: Isolating the roles of
viewers at NOAA and CCCma for helpful comments.
different forcing agents in global stratospheric temperature
changes using model integrations with incrementally added
APPENDIX A single forcings. J. Geophys. Res., 121, 8067–8082, https://doi.org/
10.1002/2015JD023841.
Ball, W. T., G. Chiodo, M. Abalas, and J. Alsing, 2020: Inconsistencies
Calculation of Synthetic Satellite Temperatures from
between chemistry climate model and observed lower strato-
Model Data
spheric trends since 1998. Atmos. Chem. Phys., 20, 9737–9752,
We use a local weighting function method developed at RSS https://doi.org/10.5194/acp-20-9737-2020.
to calculate synthetic satellite temperatures from CMIP5 and Bandoro, J., S. Solomon, B. D. Santer, D. Kinnison, and M. Mills,
CMIP6 output and from the ERA5.1 reanalysis (Santer et al. 2018: Detectability of the impacts of ozone-depleting sub-
2017b). At each grid point, simulated temperature profiles stances and greenhouse gases upon global stratospheric ozone
accounting for nonlinearities in historical forcings. Atmos. Chem.
were convolved with local weighting functions. The weights
Phys., 18, 143–166, https://doi.org/10.5194/acp-18-143-2018.
depend on the grid-point surface pressure, the surface type
Banerjee, A., J. C. Fyfe, L. M. Polvani, D. Waugh, and K.-L.
(land, ocean, or sea ice), and the selected layer-average tem- Chang, 2020: A pause in Southern Hemisphere circulation
perature (TLS, TMT, or TLT). The local weighting function trends due to the Montreal Protocol. Nature, 579, 544–548,
method provides more accurate estimates of synthetic satellite https://doi.org/10.1038/s41586-020-2120-4.
temperatures than use of a global-mean weighting function, Bellenger, H., E. Guilyardi, J. Leloup, M. Lengaigne, and
particularly over high-elevation regions. J. Vialard, 2014: ENSO representation in climate models:

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6298 JOURNAL OF CLIMATE VOLUME 34

From CMIP3 to CMIP5. Climate Dyn., 42, 1999–2018, https:// ——, V. Kharin, B. D. Santer, R. N. S. Cole, and N. P. Gillett, 2021:
doi.org/10.1007/s00382-013-1783-z. Significant impact of forcing uncertainty in a large ensemble
Bengtsson, L., and J. Shukla, 1988: Integration of space and of climate model simulations. Proc. Natl. Acad. Sci. USA, 118,
in situ observations to study global climate change. Bull. e2016549118, https://doi.org/10.1073/pnas.2016549118.
Amer. Meteor. Soc., 69, 1130–1143, https://doi.org/10.1175/ Gates, W. L., and Coauthors, 1999: An overview of the results of
1520-0477(1988)069,1130:IOSAIS.2.0.CO;2. the Atmospheric Model Intercomparison Project (AMIP I).
——, S. Hagemann, and K. I. Hodges, 2004: Can climate trends be Bull. Amer. Meteor. Soc., 80, 29–55, https://doi.org/10.1175/
calculated from reanalysis? J. Geophys. Res., 109, D11111, 1520-0477(1999)080,0029:AOOTRO.2.0.CO;2.
https://doi.org/10.1029/2004JD004536. Gillett, N. P., B. D. Santer, and A. J. Weaver, 2004: Stratospheric
Bormann, N., H. Lawrence, and J. Farnan, 2019: Global observing cooling and the troposphere. Nature, 432, 1, https://doi.org/
system experiments in the ECMWF assimilation system. Tech. 10.1038/nature03209.
Memo 839, European Centre for Medium-Range Weather ——, and Coauthors, 2016: The Detection and Attribution Model
Forecasts, 24 pp., https://doi.org/10.21957/sr184iyz. Intercomparison Project (DAMIP v1.0) contribution to CMIP6.
Checa-Garcia, R., M. I. Hegglin, D. Kinnison, D. A. Plummer, and Geosci. Model Dev., 9, 3685–3697, https://doi.org/10.5194/gmd-
K. P. Shine, 2018: Historical tropospheric and stratospheric 9-3685-2016.
ozone radiative forcing using the CMIP6 database. Geophys. Hartmann, D. L., and Coauthors, 2013: Observations: Atmosphere
Res. Lett., 45, 3264–3273, https://doi.org/10.1002/2017GL076770. and surface. Climate Change 2013: The Physical Science Basis.
Danabasoglu, G., and Coauthors, 2020: The Community Earth T. F. Stocker et al., Eds., Cambridge University Press, 159–254.
System Model version 2 (CESM2). J. Adv. Model. Earth Syst., Hausfather, Z., K. Cowtan, D. C. Clarke, P. Jacobs, M. Richardson,
12, e2019MS001916, https://doi.org/10.1029/2019MS001916. and R. Rohde, 2017: Assessing recent warming using instru-
Deser, C., A. Phillips, V. Bourdette, and H. Teng, 2012: Uncertainty mentally homogeneous sea surface temperature records. Sci.
in climate change projections: The role of internal variability. Adv., 3, e1601207, https://doi.org/10.1126/sciadv.1601207.
Climate Dyn., 38, 527–546, https://doi.org/10.1007/s00382-010- Held, I. M., and B. J. Soden, 2006: Robust responses of the hy-
0977-x. drological cycle to global warming. J. Climate, 19, 5686–5699,
——, and Coauthors, 2020: Insights from Earth system model initial- https://doi.org/10.1175/JCLI3990.1.
condition large ensembles and future prospects. Nat. Climate Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis.
Change, 10, 277–286, https://doi.org/10.1038/s41558-020-0731-2. Quart. J. Roy. Meteor. Soc., 146, 1999–2049, https://doi.org/
Eyring, V., S. Bony, G. A. Meehl, C. A. Senior, B. Stevens, R. J. 10.1002/qj.3803.
Stouffer, and K. E. Taylor, 2016: Overview of the Coupled Hirahara, S., M. Ishii, and Y. Fukuda, 2014: Centennial-scale sea
Model Intercomparison Project Phase 6 (CMIP6) experi- surface temperature analysis and its uncertainty. J. Climate,
mental design and organization. Geosci. Model Dev., 9, 1937– 27, 57–75, https://doi.org/10.1175/JCLI-D-12-00837.1.
1958, https://doi.org/10.5194/gmd-9-1937-2016. Huang, B., and Coauthors, 2017: Extended Reconstructed Sea
——, and Coauthors, 2019: Taking climate model evaluation to the Surface Temperature, version 5 (ERSSTv5): Upgrades, valida-
next level. Nat. Climate Change, 9, 102–110, https://doi.org/ tions, and intercomparisons. J. Climate, 30, 8179–8205, https://
10.1038/s41558-018-0355-y. doi.org/10.1175/JCLI-D-16-0836.1.
Flato, G., and Coauthors, 2013: Evaluation of climate models. Iribarne, J. V., and W. L. Godson, 1981: Atmospheric Thermodynamics.
Climate Change 2013: The Physical Science Basis. T. F. Stocker D. Reidel, 276 pp.
et al., Eds., Cambridge University Press, 741–866. Jiang, J., T. Zhou, and W. Zhang, 2019: Evaluation of satellite and
Flynn, C. M., and T. Mauritsen, 2020: On the climate sensitivity and reanalysis precipitable water vapor data sets against radio-
historical warming evolution in recent coupled model en- sonde observations in central Asia. Earth Space Sci., 6, 1129–
sembles. Atmos. Chem. Phys., 20, 7829–7842, https://doi.org/ 1148, https://doi.org/10.1029/2019EA000654.
10.5194/acp-20-7829-2020. Johanson, C. M., and Q. Fu, 2006: Robustness of tropospheric
Fu, Q., and C. M. Johanson, 2004: Stratospheric influences on temperature trends from MSU Channels 2 and 4. J. Climate,
MSU-derived tropospheric temperature trends: A direct error 19, 4234–4242, https://doi.org/10.1175/JCLI3866.1.
analysis. J. Climate, 17, 4636–4640, https://doi.org/10.1175/ Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis
JCLI-3267.1. Project. Bull. Amer. Meteor. Soc., 77, 437–471, https://doi.org/
——, and ——, 2005: Satellite-derived vertical dependence of 10.1175/1520-0477(1996)077,0437:TNYRP.2.0.CO;2.
tropical tropospheric temperature trends. Geophys. Res. Lett., Karl, T. R., S. J. Hassol, C. D. Miller, and W. L. Murray, Eds.,
32, L10703, https://doi.org/10.1029/2004GL022266. 2006: Temperature trends in the lower atmosphere: Steps
——, ——, S. G. Warren, and D. J. Seidel, 2004: Contribution of for understanding and reconciling differences. U.S. Climate
stratospheric cooling to satellite-inferred tropospheric tem- Change Science Program and the Subcommittee on Global
perature trends. Nature, 429, 55–58, https://doi.org/10.1038/ Change Research. National Oceanic and Atmospheric
nature02524. Administration, 164 pp.
——, S. Manabe, and C. M. Johanson, 2011: On the warming in the ——, and Coauthors, 2015: Possible artifacts of data biases in the
tropical upper troposphere: Models versus observations. Geophys. recent global surface warming hiatus. Science, 348, 1469–1472,
Res. Lett., 38, L15704, https://doi.org/10.1029/2011GL048101. https://doi.org/10.1126/science.aaa5632.
Fyfe, J. C., K. von Salzen, J. N. S. Cole, N. P. Gillett, and J.-P. Kay, J. E., and Coauthors, 2015: The Community Earth System
Vernier, 2013: Surface response to stratospheric aerosol Model (CESM) large ensemble project: A community re-
changes in a coupled atmosphere–ocean model. Geophys. Res. source for studying climate change in the presence of internal
Lett., 40, 584–588, https://doi.org/10.1002/grl.50156. climate variability. Bull. Amer. Meteor. Soc., 96, 1333–1349,
——, and Coauthors, 2017: Large near-term projected snowpack https://doi.org/10.1175/BAMS-D-13-00255.1.
loss over the western United States. Nat. Commun., 8, 14 996, Keeble, J., and Coauthors, 2021: Evaluating stratospheric ozone
https://doi.org/10.1038/ncomms14996. and water vapor changes in CMIP6 models from 1850–2100.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6299

Atmos. Chem. Phys. Discuss., 21, 5015–5061, https://doi.org/ Mitchell, D. M., Y. T. E. Lo, W. J. M. Seviour, L. Haimberger, and
10.5194/acp-21-5015-2021. L. M. Polvani, 2020: The vertical profile of recent tropical
Kennedy, J. J., N. A. Rayner, C. P. Atkinson, and R. E. Killick, temperature trends: Persistent model biases in the context of
2019: An ensemble data set of sea surface temperature change internal variability. Environ. Res. Lett., 15, 1040b4, https://
from 1850: The Met Office Hadley Centre HadSST.4.0.0.0 doi.org/10.1088/1748-9326/ab9af7.
data set. J. Geophys. Res., 124, 7719–7763, https://doi.org/ Morice, C. P., J. J. Kennedy, N. A. Rayner, and P. D. Jones, 2012:
10.1029/2018JD029867. Quantifying uncertainties in global and regional temperature
Kiehl, J. T., J. Caron, and J. J. Hack, 2005: On using global climate change using an ensemble of observational estimates: The
model simulations to assess the accuracy of MSU retrieval HadCRUT4 data set. J. Geophys. Res., 117, D08101, https://
methods for tropospheric warming trends. J. Climate, 18, doi.org/10.1029/2011JD017187.
2533–2539, https://doi.org/10.1175/JCLI3492.1. National Research Council, 2000: Reconciling Observations of
Kopp, G., and J. L. Lean, 2011: A new, lower value of total solar Global Temperature Change. National Academy Press, 169 pp.
irradiance: Evidence and climate significance. Geophys. Res. O’Gorman, P. A., and C. J. Muller, 2010: How closely do
Lett., 38, L01706, https://doi.org/10.1029/2010GL045777. changes in surface and column water vapor follow Clausius-
Lin, P., and Y. Ming, 2021: Enhanced climate response to ozone Clapeyron scaling in climate change simulations? Environ.
depletion from ozone-circulation coupling. J. Geophys. Res., Res. Lett., 5, 025207, https://doi.org/10.1088/1748-9326/5/2/
126, e2020JD034286, https://doi.org/10.1029/2020JD034286. 025207.
Maycock, A. C., and Coauthors, 2018: Revisiting the mystery of Petropavlovskikh, I., S. Godin-Beekmann, D. Hubert, R. Damadeo,
recent stratospheric temperature trends. Geophys. Res. Lett., B. Hassler, and V. Sofieva, Eds., 2019: SPARC/IO3C/GAW
45, 9919–9933, https://doi.org/10.1029/2018GL078035. Report on Long-term Ozone Trends and Uncertainties in the
McKitrick, R., and J. Christy, 2020: Pervasive warming bias in CMIP6 Stratosphere. SPARC Rep. 9, GAW Rep. 241, WCRP-17/2018,
tropospheric layers. Earth Space Sci., 7, e2020EA001281, https:// https://doi.org/10.17874/f899e57a20b, 99 pp.
doi.org/10.1029/2020EA001281. Philipona, R., and Coauthors, 2018: Radiosondes show that after
Mears, C. A., and F. J. Wentz, 2005: The effect of diurnal correction decades of cooling, the lower stratosphere is now warming.
on satellite-derived lower tropospheric temperature. Science, J. Geophys. Res., 123, 12 509–12 522, https://doi.org/10.1029/
309, 1548–1551, https://doi.org/10.1126/science.1114772. 2018JD028901.
——, and ——, 2016: Sensitivity of satellite-derived tropospheric Po-Chedley, S., and Q. Fu, 2012: Discrepancies in tropical upper
temperature trends to the diurnal cycle adjustment. J. Climate, tropospheric warming between atmospheric circulation models
29, 3629–3646, https://doi.org/10.1175/JCLI-D-15-0744.1. and satellites. Environ. Res. Lett., 7, 044018, https://doi.org/
——, and ——, 2017: A satellite-derived lower-tropospheric at- 10.1088/1748-9326/7/4/044018.
mospheric temperature dataset using an optimized adjustment ——, T. J. Thorsen, and Q. Fu, 2015: Removing diurnal cycle
for diurnal effects. J. Climate, 30, 7695–7718, https://doi.org/ contamination in satellite-derived tropospheric temperatures:
10.1175/JCLI-D-16-0768.1. Understanding tropical tropospheric trend discrepancies.
——, M. C. Schabel, and F. J. Wentz, 2003: A reanalysis of the MSU J. Climate, 28, 2274–2290, https://doi.org/10.1175/JCLI-D-13-
channel 2 tropospheric temperature record. J. Climate, 16, 3650– 00767.1.
3664, https://doi.org/10.1175/1520-0442(2003)016,3650:AROTMC. ——, B. D. Santer, S. Fueglistaler, M. D. Zelinka, P. Cameron-
2.0.CO;2. Smith, J. F. Painter, and Q. Fu, 2021: Natural variability con-
——, B. D. Santer, F. J. Wentz, K. E. Taylor, and M. Wehner, 2007: tributes to model–satellite differences in tropical tropospheric
The relationship between temperature and precipitable water warming. Proc. Natl. Acad. Sci. USA, 118, e2020962118,
changes over tropical oceans. Geophys. Res. Lett., 34, L24709, https://doi.org/10.1073/pnas.2020962118.
https://doi.org/10.1029/2007GL031936. Ramaswamy, V., M. D. Schwarzkopf, W. J. Randel, B. D. Santer,
——, F. J. Wentz, P. Thorne, and D. Bernie, 2011: Assessing un- B. J. Soden, and G. L. Stenchikov, 2006: Anthropogenic and
certainty in estimates of atmospheric temperature changes natural influences in the evolution of lower stratospheric
from MSU and AMSU using a Monte-Carlo estimation cooling. Science, 311, 1138–1141, https://doi.org/10.1126/
technique. J. Geophys. Res., 116, D08112, https://doi.org/ science.1122587.
10.1029/2010JD014954. Randall, D. A., and Coauthors, 2007: Climate models and their
——, D. K. Smith, L. Ricciardulli, J. Wang, H. Huelsing, and F. J. evaluation. Climate Change 2007: The Physical Science Basis.
Wentz, 2018: Construction and uncertainty estimation of a S. Solomon et al., Eds., Cambridge University Press, 589–662.
satellite-derived total precipitable water data record over the Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V.
world’s oceans. Earth Space Sci., 5, 197–210, https://doi.org/ Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global
10.1002/2018EA000363. analyses of sea surface temperature, sea ice, and night marine air
Meehl, G. A., C. A. Senior, V. Eyring, G. Flato, J.-F. Lamarque, temperature since the late nineteenth century. J. Geophys. Res.,
R. J. Stouffer, K. E. Taylor, and M. Schlund, 2020: Context for 108, 4407, https://doi.org/10.1029/2002JD002670.
interpreting equilibrium climate sensitivity and transient cli- Riahi, K., and Coauthors, 2017: The Shared Socioeconomic Pathways
mate response from the CMIP6 Earth system models. Sci. and their energy, land use, and greenhouse gas emissions im-
Adv., 6, eaba1981, https://doi.org/10.1126/sciadv.aba1981. plications: An overview. Global Environ. Change, 42, 153–168,
Meinshausen, M., and Coauthors, 2011: The RCP greenhouse gas https://doi.org/10.1016/j.gloenvcha.2016.05.009.
concentrations and their extensions from 1765 to 2300. Climatic Ridley, D. A., and Coauthors, 2014: Total volcanic stratospheric
Change, 109, 213–241, https://doi.org/10.1007/s10584-011-0156-z. aerosol optical depths and implications for global climate
Mills, M. J., and Coauthors, 2016: Global volcanic aerosol prop- change. Geophys. Res. Lett., 41, 7763–7769, https://doi.org/
erties derived from emissions, 1990–2014, using CESM1 10.1002/2014GL061541.
(WACCM). J. Geophys. Res., 121, 2332–2348, https://doi.org/ Rieger, L. A., J. N. S. Cole, J. C. Fyfe, S. Po-Chedley, P. Cameron-
10.1002/2015JD024290. Smith, P. J. Durack, N. P. Gillett, and Q. Tang, 2020: Quantifying

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

6300 JOURNAL OF CLIMATE VOLUME 34

CanESM5 and EAMv1 sensitivities to Mt. Pinatubo vol- monsoon: An intercomparison of CMIP5 vs. CMIP3 simula-
canic forcing for the CMIP6 historical experiment. Geosci. tions of the late 20th century. Climate Dyn., 41, 2711–2744,
Model Dev., 13, 4831–4843, https://doi.org/10.5194/gmd-13- https://doi.org/10.1007/s00382-012-1607-6.
4831-2020. Steiner, A., and Coauthors, 2020: Observed temperature changes
Robock, A., 2000: Volcanic eruptions and climate. Rev. Geophys., in the troposphere and stratosphere from 1979 to 2018.
38, 191–219, https://doi.org/10.1029/1998RG000054. J. Climate, 33, 8165–8194, https://doi.org/10.1175/JCLI-D-19-
Santer, B. D., and Coauthors, 2005: Amplification of surface tem- 0998.1.
perature trends and variability in the tropical atmosphere. Stone, P. H., and J. H. Carlson, 1979: Atmospheric lapse rate
Science, 309, 1551–1556, https://doi.org/10.1126/science.1114867. regimes and their parameterization. J. Atmos. Sci., 36,
——, and Coauthors, 2011: Separating signal and noise in atmo- 415–423, https://doi.org/10.1175/1520-0469(1979)036,0415:
spheric temperature changes: The importance of timescale. ALRRAT.2.0.CO;2.
J. Geophys. Res., 116, D22105, https://doi.org/10.1029/ Swart, N. C., S. T. Gille, J. C. Fyfe, and N. P. Gillett, 2018: Recent
2011JD016263. Southern Ocean warming and freshening driven by green-
——, and Coauthors, 2013: Human and natural influences on house gas emissions and ozone depletion. Nat. Geosci., 11,
the changing thermal structure of the atmosphere. Proc. 836–841, https://doi.org/10.1038/s41561-018-0226-1.
Natl. Acad. Sci. USA, 110, 17 235–17 240, https://doi.org/ ——, and Coauthors, 2019: The Canadian Earth System Model
10.1073/pnas.1305332110. version 5 (CanESM5.0.3). Geosci. Model Dev., 12, 4823–4873,
——, and Coauthors, 2017a: Causes of differences in model and https://doi.org/10.5194/gmd-12-4823-2019.
satellite tropospheric warming rates. Nat. Geosci., 10, 478–485, Tatebe, H., and Coauthors, 2019: Description and basic evaluation
https://doi.org/10.1038/ngeo2973. of simulated mean state, internal variability, and climate
——, and Coauthors, 2017b: Comparing tropospheric warming in sensitivity in MIROC6. Geosci. Model Dev., 12, 2727–2765,
climate models and satellite data. J. Climate, 30, 373–392, https://doi.org/10.5194/gmd-12-2727-2019.
https://doi.org/10.1175/JCLI-D-16-0333.1. Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of
——, J. Fyfe, S. Solomon, J. Painter, C. Bonfils, G. Pallotta, and CMIP5 and the experiment design. Bull. Amer. Meteor. Soc.,
M. Zelinka, 2019: Quantifying stochastic uncertainty in 93, 485–498, https://doi.org/10.1175/BAMS-D-11-00094.1.
detection time of human-caused climate signals. Proc. Natl. Thomason, L. W., and Coauthors, 2018: A global space-based
Acad. Sci. USA, 116, 19 821–19 827, https://doi.org/10.1073/ stratospheric aerosol climatology: 1979–2016. Earth Syst. Sci.
pnas.1904586116. Data, 10, 469–492, https://doi.org/10.5194/essd-10-469-2018.
Schmidt, G. A., D. T. Shindell, and K. Tsigaridis, 2014: Reconciling Thompson, D. W. J., and Coauthors, 2012: The mystery of recent
warming trends. Nat. Geosci., 7, 158–160, https://doi.org/ stratospheric temperature trends. Nature, 491, 692–697,
10.1038/ngeo2105. https://doi.org/10.1038/nature11579.
Shine, K. P., and Coauthors, 2003: A comparison of model-simulated Thorne, P. W., J. R. Lanzante, T. C. Peterson, D. J. Seidel, and K. P.
trends in stratospheric temperatures. Quart. J. Roy. Meteor. Shine, 2011: Tropospheric temperature trends: History of an
Soc., 129, 1565–1588, https://doi.org/10.1256/qj.02.186. ongoing controversy. Wiley Interdiscip. Rev.: Climate Change,
Simmons, A., and Coauthors, 2020: Global stratospheric temper- 2, 66–88, https://doi.org/10.1002/wcc.80.
ature bias and other stratospheric aspects of ERA5 and Trenberth, K. E., J. Fasullo, and L. Smith, 2005: Trends and
ERA5.1. Tech. Memo 859, European Centre for Medium- variability in column-integrated atmospheric water vapor.
Range Weather Forecasts, 40 pp. Climate Dyn., 24, 741–758, https://doi.org/10.1007/s00382-
Solomon, S., 1999: Stratospheric ozone depletion: A review of 005-0017-4.
concepts and history. Rev. Geophys., 37, 275–316, https:// ——, and Coauthors, 2007: Observations: Surface and atmo-
doi.org/10.1029/1999RG900008. spheric climate change. Climate Change 2007: The Physical
——, J. S. Daniel, R. R. Neely, J.-P. Vernier, E. G. Dutton, and L. W. Science Basis. S. Solomon et al., Eds., Cambridge University
Thomason, 2011: The persistently variable ‘‘background’’ Press, 235–336.
stratospheric aerosol layer and global climate change. Science, U.S. Senate, 2015: Data or dogma? Promoting open inquiry in
333, 866–870, https://doi.org/10.1126/science.1206027. the debate over the magnitude of human impact on Earth’s
——, P. J. Young, and B. Hassler, 2012: Uncertainties in the evolu- climate. Hearing before the U.S. Senate Committee on
tion of stratospheric ozone and implications for recent temper- Commerce, Science, and Transportation, Subcommittee
ature changes in the tropical lower stratosphere. Geophys. Res. on Space, Science, and Competitiveness, 114th Congress,
Lett., 39, L17706, https://doi.org/10.1029/2012GL052723. first session, 8 December 2015, https://clio.columbia.edu/
——, D. J. Ivy, D. Kinnison, M. J. Mills, R. R. Neely III, and catalog/12267036.
A. Schmidt, 2016: Emergence of healing in the Antarctic Wang, S., T. Xu, W. Nie, C. Jiang, Y. Yang, Z. Fang, M. Li, and
ozone layer. Science, 353, 269–274, https://doi.org/10.1126/ Z. Zhang, 2020: Evaluation of precipitable water vapor from five
science.aae0061. reanalysis products with ground-based GNSS observations.
——, and Coauthors, 2017: Mirrored changes in Antarctic ozone Remote Sens., 12, 1817, https://doi.org/10.3390/rs12111817.
and stratospheric temperature in the late 20th versus early 21st Wentz, F. J., 2013: SSM/I Version-7 Calibration Report. Tech. Rep.
centuries. J. Geophys. Res., 122, 8940–8950, https://doi.org/ 011012, 46 pp., available at http://www.remss.com/missions/
10.1002/2017JD026719. ssmi/ (see link under ‘‘References’’ therein).
Spencer, R. W., J. R. Christy, and W. D. Braswell, 2017: UAH ——, and M. Schabel, 1998: Effects of orbital decay on satellite-
version 6 global satellite temperature products: Methodology derived lower-tropospheric temperature trends. Nature, 394,
and results. Asia-Pac. J. Atmos. Sci., 53, 121–130, https:// 661–664, https://doi.org/10.1038/29267.
doi.org/10.1007/s13143-017-0010-y. ——, and ——, 2000: Precise climate monitoring using comple-
Sperber, K. R., H. Annamalai, I.-S. Kang, A. Kitoh, A. Moise, mentary satellite data sets. Nature, 403, 414–416, https://
A. Turner, B. Wang, and T. Zhou, 2013: The Asian summer doi.org/10.1038/35000184.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC

1 AUGUST 2021 SANTER ET AL. 6301

Zelinka, M. D., T. A. Myers, D. T. McCoy, S. Po-Chedley, P. M. ——, and H. Qian, 2016: Stratospheric temperature climate record
Caldwell, P. Ceppi, S. A. Klein, and K. E. Taylor, 2020: Causes of from merged SSU and AMSU-A observations. J. Atmos. Oceanic
higher climate sensitivity in CMIP6 models. Geophys. Res. Lett., Technol., 33, 1967–1984, https://doi.org/10.1175/JTECH-D-16-
47, e2019GL085782, https://doi.org/10.1029/2019GL085782. 0018.1.
Zou, C.-Z., and W. Wang, 2011: Inter-satellite calibration of ——, M. D. Goldberg, and X. Hao, 2018: New generation of U.S.
AMSU-A observations for weather and climate applications. satellite microwave sounder achieves high radiometric stabil-
J. Geophys. Res., 116, D23113, https://doi.org/10.1029/ ity performance for reliable climate change detection. Sci.
2011JD016205. Adv., 4, eaau0049, https://doi.org/10.1126/sciadv.aau0049.

Unauthenticated | Downloaded 08/08/22 12:49 AM UTC