JUNE 2013 SEILER ET AL.

1303

Likely Ranges of Climate Change in Bolivia

CHRISTIAN SEILER
Earth System Science Group, Wageningen University and Research Centre, Wageningen, Netherlands, and Department
of Climate Change and Environmental Services, Fundaci

on Amigos de la Naturaleza, Santa Cruz de la Sierra, Bolivia

RONALD W. A. HUTJES
Earth System Science Group, Wageningen University and Research Centre, Wageningen, Netherlands

PAVEL KABAT
International Institute for Applied Systems Analysis, Laxenburg, Austria, and Earth System Science Group,
Wageningen University and Research Centre, Wageningen, Netherlands

(Manuscript received 30 July 2012, in final form 30 November 2012)

ABSTRACT

Bolivia is facing numerous climate-related threats, ranging from water scarcity due to rapidly retreating
glaciers in the Andes to a partial loss of the Amazon forest in the lowlands. To assess what changes in climate
may be expected in the future, 35 global circulation models (GCMs) from the third and fifth phases of the
Coupled Model Intercomparison Project (CMIP3/5) were analyzed for the Bolivian case. GCMs were vali-
dated against observed surface air temperature, precipitation, and incoming shortwave (SW) radiation for the
period 1961–90. Weighted ensembles were developed, and climate change projections for five emission
scenarios were assessed for 2070–99. GCMs revealed an overall cold, wet, and positive-SW-radiation bias and
showed no substantial improvement from the CMIP3 to the CMIP5 ensemble for the Bolivian case. Models
projected an increase in temperature (2.58–5.98C) and SW radiation (1%–5%), with seasonal and regional
differences. In the lowlands, changes in annual rainfall remained uncertain for CMIP3 whereas CMIP5 GCMs
were more inclined to project decreases (29%). This pattern also applied to most of the Amazon basin,
suggesting a higher risk of partial biomass loss for the CMIP5 ensemble. Both ensembles agreed on less rainfall
(219%) during drier months (June–August and September–November), with significant changes in interannual
rainfall variability, but disagreed on changes during wetter months (January–March). In the Andes, CMIP3
GCMs tended toward less rainfall (29%) whereas CMIP5 tended toward more (120%) rainfall during parts of
the wet season. The findings presented here may provide inputs for studies of climate change impact that assess
how resilient human and natural systems are under different climate change scenarios.

1. Introduction forest (Rammig et al. 2010). Droughts and floods associ-

ated with El Ni~ no–Southern Oscillation (ENSO) events
The Plurinational State of Bolivia is facing numerous
affect thousands of people and lead to economic losses
climate-related threats. In the Andes, rapidly retreating
equivalent to millions of U.S. dollars (UNDP 2011). As
glaciers affect the supply of drinking water, agricultural
a developing country with one-third of the labor force
production, and the provision of energy from hydro-
working in the agricultural sector (http://www.ine.gob.
power (Bradley et al. 2006). In the lowlands, a reduction
bo), Bolivia is considered to be extremely vulnerable to
of rainfall as is projected by some general circulation
climate change (World Bank 2010).
models (GCMs) may lead to the partial loss of the Amazon
Given this exposure to climate-related threats, a re-
gional analysis of climate change projections is of great
interest. The Fourth Assessment Report (AR4) of the In-
Corresponding author address: Christian Seiler, Earth System
Science Group, Wageningen University and Research Centre,
tergovernmental Panel on Climate Change (IPCC) docu-
Droevendaalsesteeg 3, 6708 PB Wageningen, Netherlands. mented climate change projections from multiple GCMs
E-mail: [email protected] belonging to the third phase of the Coupled Model

DOI: 10.1175/JAMC-D-12-0224.1

Ó 2013 American Meteorological Society

1304 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

Intercomparison Project (CMIP3) for the South Ameri- conjunction with three lateral boundary conditions
can continent (Christensen et al. 2007). AR4 found that (Boulanger et al. 2010). To our knowledge though, cli-
GCMs had a cold bias in the Andes and a slightly warm mate change scenarios related to CLARIS have so far
bias in the tropical lowlands. Averaged over tropical only been published for southern South America (Nunez
South America (128N–208S and 828–348W, hereinafter et al. 2009).
AMZ), this bias was estimated with an annual median Previous analyses of climate change scenarios men-
of 0.68C. Climate change projections in the same region tioned above are of mainly subcontinental scale and
showed a median increase in annual temperatures of include very few Bolivian stations for model evaluation,
3.38C and an interquartile range of 2.68–3.78C under the making it difficult to apply this information to Bolivia.
intermediate emission scenario ‘‘A1B.’’ For precipitation, Also, an evaluation of the most recent generation of
the same report found a wet bias in the Andes and a dry GCMs (CMIP5; see section 2) is still missing. The coarse
bias in the tropical lowlands, estimated with an annual spatial resolution of GCMs limits the validity of model
median of 28% for the AMZ region (Christensen et al. results in the Andes. Given the potential disagreement
2007, their Fig. S11.26 and Table S11.1). This dry bias among models on directional changes of precipitation,
was likely due to an underestimation as well as an ex- however, a comprehensive evaluation of GCMs is a nec-
tension too far southward of the intertropical conver- essary precursor to any downscaling effort. We therefore
gence zone. In most of tropical South America, GCMs analyzed projections from 35 different GCMs, including
did not agree on the directional change of annual rain- models from CMIP3 and CMIP5, as well as five emission
fall, resulting in a median change of 0% and an inter- scenarios. In doing so, we assessed for the Bolivian case
quartile range from 23% to 16% in the AMZ region. 1) how well GCMs reproduce historical climate pat-
This lack of agreement was confirmed by numerous terns, 2) what changes in climate means and variability
other studies that focused on the Amazon region (e.g., may be expected, and 3) to what extent there are dif-
Li et al. 2006; Jupp et al. 2010). Seasonal changes were ferences between the CMIP3 and CMIP5 ensembles.
more certain, with more rainfall from January until Our results may provide input for climate change impact
March and less rainfall from July until September in assessments, exploring the probability of climate-related
major parts of the Amazon, including Bolivia’s northern threats such as water scarcity or Amazon dieback in
lowlands (Vera et al. 2006). In the Bolivian Altiplano, Bolivia.
multiple GCMs revealed a warm and wet bias (Seth The following sections describe our study area, methods,
et al. 2010). Temperature projections under the high- results, and discussion. The method section outlines the
emission scenario ‘‘A2’’ ranged from 158 to 168C data and emission scenarios, as well as approaches re-
(monthly medians), and rainfall projections tended to- lated to model validation, ensemble weighting, and mul-
ward less annual rainfall, with decreases from May to timodel agreement. Model skills in reproducing historical
September and increases during March and April. climate patterns and future projections of temperature,
Numerous attempts have been made to dynamically precipitation, and shortwave (SW) radiation are pre-
downscale global climate change scenarios for tropical sented in the results section. Our discussion interprets
South America (e.g., Soares and Marengo 2009; Urrutia the results in the context of existing literature and elab-
and Vuille 2009; Marengo et al. 2010). Projections in- orates on the principal findings.
clude temperature increases from 68 to 88C, with higher
values during June–August (JJA), enhanced moisture
2. Study area
transport from the Amazon to the La Plata basin that
leads to rainfall deficiency in the former and rainfall Bolivia is a tropical country measuring more than
excess in the latter, and a more intense hydrological 106 km2. With its main altitudinal divisions being low-
cycle with more rainfall in December–February (DJF) lands (,800 m MSL), Andean slopes (800–3200 m MSL),
and less rainfall in JJA in the Bolivian lowlands. Pro- and highlands (Altiplano; .3200–6500 m MSL), Bolivia’s
jections are based on very few lateral boundary condi- climate varies with increasing altitude from tropical to
tions [mainly the Hadley Centre Atmosphere Model, cold desert climate, with annual mean surface air tem-
version 3 (HadAM3)], however, limiting their robust- peratures ranging from 08 to 308C (Fig. 1). Rainfall
ness in a region with little agreement among GCMs on ranges from ,300 to .3000 mm yr21 and varies from
the directional change of rainfall. Other efforts include high to low from the northern Andean slopes, northern
the La Plata Basin Regional Hydroclimate Project lowlands, southern lowlands, and southern Andean slopes
(CLARIS LPB; http://www.claris-eu.org), which aims to the Altiplano, with a north–south division roughly at
at predicting regional climate change impacts on the 188S. Much higher values (.5000 mm yr21) along the
La Plata basin using several regional climate models in northern Andean slopes are documented by SENAMHI
JUNE 2013 SEILER ET AL. 1305

FIG. 1. (a) Location of Bolivia, (b) surface elevation (m MSL), (c) the three regions studied in this work, (d) annual mean air tem-
perature (8C) from 1961 to 1990, (e) annually accumulated precipitation (mm yr21) from 1961 to 1990, and (f) mean SW radiation (W m22)
from 1979 to 1990. Black dots denote locations of meteorological stations.

(2009) from stations not included in this research. The enhanced precipitation with a southeastward extension
austral summer (DJF) and winter (JJA) coincide with toward the Atlantic Ocean, referred to as the South
the wet and dry seasons, respectively. Incoming SW Atlantic convergence zone (SACZ). At the same time,
radiation ranges from 160 to 260 W m22, with higher the release of condensational heat over the Amazon and
values in the Altiplano and southern lowlands. Seasonal Andean slopes leads to the formation of the upper-level
differences in temperature, precipitation, and SW radi- Bolivian high pressure system at 200 hPa (;12 km MSL)
ation vary among regions, with the largest seasonal dif- (Lenters and Cook 1997), which further enhances mois-
ferences being 88C in the Andes, 180 mm month21 in the ture advection from the Amazon to the Bolivian low-
northern lowlands, and 112 W m22 in the southern low- lands and highlands (Vuille 1999). An upper-tropospheric
lands, respectively. To account for the spatial gradients trough near the coast of Northeast Brazil forms as a re-
in climate, we stratified Bolivia into three regions [north- sponse to the rising air motion over the continent (Silva
ern lowlands (NLL), southern lowlands (SLL), and Andes and Kousky 2012). The chaco low, Bolivian high, LLJ,
(AND); see Fig. 1c], that roughly characterize a warm SACZ, and upper-tropospheric trough form the main
and humid climate, a warm and dry climate, and a cold components of the South American monsoon system
and dry climate, respectively. Given the coarse spatial (Zhou and Lau 1998), affecting rainfall in DJF. In aus-
resolution of GCMs, Andean slopes were merged partly tral winter (JJA), the chaco low and the SACZ dissipate,
with the lowlands regions and partly with the Andes leading to less moisture transport from the north. Cold
region. fronts from the southern polar regions penetrate into
Climate patterns in Bolivia are shaped by the fol- the Bolivian lowlands, leading to low temperatures and
lowing synoptic scale systems: In austral summer (DJF), to limited precipitation when colliding with warm trop-
a low pressure system called the thermal chaco low in- ical air masses (Garreaud 2000; Ronchail and Gallaire
tensifies at 258S, enhancing the easterly trade winds to 2006; SENAMHI 2009). The Bolivian high dissipates
transport moisture from the northern tropical Atlantic and westerly winds prevail in Bolivia, preventing mois-
Ocean into the continent. Deflected by the Andes, mois- ture transport from the lowlands to the Andes in JJA
ture is transported southward from the Amazon River (Vuille 1999).
region to the subtropical plains of southeastern South Sources of climate variability in Bolivia include 1) the
America by a low-level jet (LLJ) located at ;925–850 hPa Pacific decadal oscillation (PDO), 2) ENSO, and 3) the
(;1 km MSL) and less than 100 km east of the Andean Antarctic Oscillation (Garreaud et al. 2009; Seiler et al.
slopes (Marengo et al. 2004). This pattern leads to 2013). Climate variability may lead to extreme events,
TABLE 1. Overview of the GCMs used in this study for the validation period of 1961–90 and for emissions scenarios SRES B1, A1B, A2, RCP4.5, and 85. Not all GCMs were available for
all scenarios. GCMs available at the time of writing and used in this study are denoted with an ‘‘x.’’
1306

1961–90 20C3M SRES (2070–99) RCP (2070–99)

Country GCMs Identifier CMIP (historical) B1 A1B A2 4.5 8.5
Norway Bjerknes Centre for Climate Research Bergen Climate Model, version 2.0 A 3 x x x x — —
(BCCR-BCM2.0)
Canada Canadian Centre for Climate Modelling and Analysis (CCCma) Coupled B 3 x x x x — —
Global Climate Model, version 3.1 [CGCM3.1(T47)]
Canada CGCM3.1(T63) C 3 x x x — — —
France Centre National de Recherches M
et
eorologiques Coupled Global Climate D 3 x x x x — —
Model, version 3 (CNRM-CM3)
Australia Commonwealth Scientific and Industrial Research Organisation model, E 3 x x x x — —
mark 3.0 (CSIRO-Mk3.0)
Australia CSIRO-Mk3.5 F 3 x x x x — —
United States Geophysical Fluid Dynamics Laboratory Climate Model, version 2.0 G 3 x x x x — —
(GFDL-CM2.0)
United States GFDL-CM2.1 H 3 x x x x — —
United States Goddard Institute for Space Studies Atmosphere–Ocean Model (GISS-AOM) I 3 x x x — — —
United States GISS ModelE coupled with the Hybrid Coordinate (HYCOM) ocean model J 3 x — x — — —
(GISS E-H)
United States GISS ModelE coupled with the Russell ocean model (GISS E-R) K 3 x x — x — —
China Flexible Global Ocean–Atmosphere–Land System Model, gridpoint version 1.0 L 3 x x x — — —
(FGOALS-g1.0)
Russia Institute of Numerical Mathematics Coupled Model, version 3.0 (INM-CM3.0) M 3 x x x x — —
France L’Institut Pierre-Simon Laplace Coupled Model, version 4 (IPSL-CM4) N 3 x x x x — —
Japan Model for Interdisciplinary Research on Climate, version 3.2 (high resolution) O 3 x x x — — —
[MIROC3.2(hires)]
Japan MIROC (medium resolution) [MIROC3.2(medres)] P 3 x x x x — —
Germany–Korea ECHAM and the global Hamburg Ocean Primitive Equation model (ECHO-G) Q 3 x x x x — —
Germany ECHAM5/Max Planck Institute Ocean Model (MPI-OM) R 3 x x x x — —
Japan Meteorological Research Institute Coupled Atmosphere–Ocean General S 3 x x x x — —
Circulation Model, version 2.3.2 (MRI-CGCM2.3.2)
United States Community Climate System Model, version 3.0 (CCSM3) T 3 x x x x — —
United States Parallel Climate Model (PCM) U 3 x x x x — —
United Kingdom Met Office–Hadley Centre Coupled Model, version 3 (UKMO-HadCM3) V 3 x x x x — —
United Kingdom UKMO-Hadley Centre Global Environmental Model, version 1 W 3 x — x x — —
(UKMO-HadGEM1)
Canada Second Generation Canadian Earth System Model (CanESM2) a 5 x — — — x x
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY

France CNRM Coupled Global Climate Model, version 5 (CNRM-CM5) b 5 x — — — x x

Russia INM Coupled Model, version 4.0 (INM-CM4) c 5 x — — — x x
France IPSL Coupled Model, version 5, coupled with the Nucleus for European d 5 x — — — x x
Modelling of the Ocean (NEMO) model, low resolution (IPSL-CM5A-LR)
Japan MIROC, version 5 (MIROC5) e 5 x — — — x x
VOLUME 52
JUNE 2013 SEILER ET AL. 1307

including droughts and floods. Droughts mainly occur in

RCP (2070–99)
8.5

11
x
x
x

x
x
x
the southern lowlands and in many regions in the Alti-
plano from June to August (CONARADE 2010), whereas
floods happen from January to March mainly in the sa-

4.5

12
x

x
x
x

x
x
x
vannas of the northern lowlands (Bourrel et al. 2009) but
also in the catchment areas of Lake Titicaca and Poop
o
(UNDP 2011) in the Andes.
A2
—
—
—
—

—
—
—
18
SRES (2070–99)

Meteorological observations reveal that Bolivia’s cli-

mate is currently warming at a rate of 0.18C (10 yr)21,
A1B
—

—
—
—

—
—
—
22
with larger increases in the Andes and during the dry
season (Seiler et al. 2013). Rainfall increased from 1965
B1
—
—
—
—

—
—
—
21
to 1984 (12% in DJF and 18% in JJA) and decreased
from 1985 to 2004 (24% in DJF and 210% in JJA),
1961–90 20C3M

roughly following the pattern of the PDO.

(historical)

35
x
x
x
x

x
x
x

3. Methods
a. Data
CMIP

1) OBSERVATIONS
5

5
5
5

5
5
5

Meteorological observations consisted of surface air

Identifier

temperature, precipitation, and incoming shortwave ra-

h

k
g
f

diation at the surface. Temperature and precipitation

measurements were provided by the Bolivian National
TABLE 1. (Continued)

Service of Meteorology and Hydrology (SENAMHI)

Max Planck Institute Earth System Model at base resolution (MPI-ESM-LR)

GFDL Earth System Model with Generalized Ocean Layer Model (GOLD)

and were homogenized by Seiler et al. (2013). Meteo-

rological stations with sufficient data for the period
1961–90 were selected from this homogenized dataset,
resulting in 25 stations with temperature measure-
GISS ModelE2 coupled with Russell ocean model (GISS-E2-R)

ments and 59 stations with precipitation measurements

(Figs. 1d,e). Stations contained at least 50% and, on
average, 77% and 83% of daily temperature and pre-
cipitation measurements, respectively. There were
four stations with slightly less than 50% of data in re-
MIROC, Earth System Model (MIROC-ESM)

HadGEM2 (Carbon Cycle) (HadGEM2-CC)

gions with low station density. Daily data were converted

HadGEM2 (Earth System) (HadGEM2-ES)
GCMs

to monthly data if no more than 3 days were missing in

ocean component (GFDL-ESM2G)

a month.
Monthly mean SW radiation was obtained for the
MRI, version 3 (MRI-CGCM3)

period 1979–90 from the Modern-Era Retrospective

Analysis for Research and Applications (MERRA), a
National Aeronautics and Space Administration (NASA)
reanalysis for the satellite era using the Goddard Earth
Observing System Data Assimilation System, version 5
(GEOS-5; Rienecker et al. 2011). The data belong to
a collection of datasets (‘‘Obs4MIPs’’) that has been
organized according to the CMIP5 model output re-
quirements and made available on the Earth System
Grid gateway (http://pcmdi3.llnl.gov).
United Kingdom
United Kingdom

2) GENERAL CIRCULATION MODELS

United States

United States
Country

Germany

We used a total of 35 different GCMs, with 23 GCMs

Japan

Japan

corresponding to CMIP3 and 12 GCMs corresponding

Sum

to CMIP5. CMIP3 GCMs contributed to AR4 (Pachauri

1308 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

and Reisinger 2007), and CMIP5 data will contribute modeled with observed monthly values averaged over
to the IPCC’s Fifth Assessment Report (AR5), expected 1961–90 and plotted the correlation coefficient, centered
for publication by 2014. Changes from CMIP3 to CMIP5 root-mean-square error (RMSE), and standard deviation
GCMs include changes in emission scenarios (see sec- of each GCM in a Taylor diagram (Taylor 2001). Alti-
tion 3b), improved model physics, finer spatial resolu- tudinal differences between grid cells and meteorological
tion, and additional processes related to the oceanic stations may lead to large differences in temperature. To
and terrestrial carbon cycle, aerosols, atmospheric correct for this orographic effect, we calculated local
chemistry, and ice sheets. All GCMs were validated lapse rates for each individual GCM (3.18C km21 on av-
against observational data for the period 1961–90, and erage) and used these lapse rates in conjunction with the
future climates were evaluated for the period 2070–99. models’ surface elevation to bring modeled temperatures
The number of GCMs for future climates depended on to the altitude of the meteorological stations. In addition
their availability at the time of writing and therefore to Taylor diagrams, we used box-plot diagrams to assess
varied among scenarios, with 21, 22, 18, 12, and 11 whether errors were random or systematic.
GCMs for IPCC Special Report on Emissions Sce-
d. Ensemble weighting
narios (SRES) B1, A1B, A2, RCP4.5, and RCP8.5,
respectively (Table 1). To make a more valid com- To compare modeled and observed temperature,
parison between CMIP3 and CMIP5 projections, we precipitation, and SW radiation, we calculated skill
also used a subsample consisting of 12 CMIP3 GCMs, scores Sk,l for each GCM (index k) and each variable
including only predecessors of the 12 CMIP5 GCMs (index l) according to Taylor (2001):
(viz., B, D, G, J, M, N, O, P, R, S, V, and W; see Table 1).
4(1 1 Rk,l )4
b. Emission scenarios Sk,l 5 , (1)
sfk,l )2 (1 1 R0k,l )4
sfk,l 1 1/^
(^
Our analysis covered three SRES emission scenar-
ios for the CMIP3 GCMs (Nakicenovic et al. 2000) where R is the correlation coefficient, R0 is the maxi-
and two representative concentration pathway (RCP) mum attainable correlation coefficient, and s ^ f is the
emission scenarios for the CMIP5 GCMs (Moss et al. standard deviation of the GCM divided by the standard
2010). The SRES scenarios are based on socioeconomic deviation of the observation. After quantifying Sk,l for
storylines and differ in the resulting atmospheric carbon each GCM and variable, Sk,l was normalized for each
dioxide (CO2)-equivalent concentrations, which include variable by dividing by the sum of all Sk,l from all GCMs,
the net effect of all anthropogenic forcing agents. We leading to scores between 0 and 1:
used SRES scenarios B1, A1B, and A2 with the respec-
tive atmospheric CO2-equivalent concentrations of 600, Sk,l
S^k,l 5 . (2)
850, and 1250 ppm by the year 2100. For comparison, 35
CO2-equivalent concentration for the year 2005 is esti- å Sk,l
k51
mated to be 375 ppm (Pachauri and Reisinger 2007).
RCPs project future radiative forcings without de- The final weight of each GCM w ^ k was then taken as
fining new socioeconomic scenarios. RCPs are compat- the normalized product of the three normalized scores
ible with the full range of stabilization, mitigation, and from temperature, precipitation, and SW radiation:
reference emissions scenarios available in the current
wk
scientific literature with an adequate separation of the ^k 5
w , (3)
35
radiative forcing pathways. There are four RCPs in
total (RCP2.6, -4.5, -6.0, and -8.5), each named after
å wk
k51
their radiative forcing reached by 2100. We used
RCP4.5 and RCP8.5, corresponding to radiative forc- where
ings of 4.5 and 8.5 W m22 by 2100, respectively. The
3
P
resulting CO2-equivalent concentrations in the year
wk 5 S^k l .
,
(4)
2100 for RCP4.5 and RCP8.5 are 650 and 1370 ppm, l51
respectively (Moss et al. 2010).
Weights were used to develop box plots of likely
c. Validation
ranges of climate change by calculating the weighted
We quantified the ability of each GCM to reproduce 5th, 17th, 50th, 83rd, and 95th percentiles of projected
historical temperature, precipitation, and SW radiation yearly and monthly changes for each region. The IPCC
patterns for Bolivia. For this purpose we compared considers changes to be likely if their probability of
JUNE 2013 SEILER ET AL. 1309

FIG. 2. (left) Taylor and (right) box-plot diagrams of (a),(b) monthly mean air temperature, (c),(d) monthly
precipitation, and (e),(f) monthly SW radiation. Black uppercase and blue lowercase letters correspond to GCMs
from CMIP3 and CMIP5, respectively. Horizontal bars in the box plots present the median, the boxes give the
interquartile ranges, and the whiskers show the minimum and maximum values. White boxes correspond to CMIP3,
and gray-shaded boxes show CMIP5 GCMs.
1310 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

FIG. 3. GCM scores for reproducing temperature (dark gray), precipitation (light gray), and SW radiation (white), along with the
corresponding final weights (black).

occurrence is estimated at 66% (Mastrandrea et al. 2010), Fcalc 5 max(s21 /s22 , s22 /s21 ) , (5)
corresponding to the range enclosed by the 17th and 83rd
percentiles. Changes in the mean were tested for signifi- where Fcalc is the calculated F value and s2 is the vari-
cance using a weighted t test with a probability level of ance, with 1 and 2 denoting the two periods 1961–90 and
95%. 2070–99, respectively. Changes in the mean were tested
e. Multimodel agreement for significance with the t test:

We assessed whether GCMs projected significant , !1/2

changes in interannual variability and whether GCMs s22 s21
tcalc 5 (x2 2 x1 ) 1 . (6)
agreed on the directional changes of rainfall and SW n2 n1
radiation. After regridding all GCMs to a common
spatial resolution of 28 3 28, changes in interannual Last, we plotted for each grid cell the number of
variability were tested for significance with the F test: GCMs with significant changes in interannual variability

FIG. 4. Changes in annual mean temperature (8C) of the weighted ensemble comparing 1961–90
with 2070–99 for NLL, SLL, and AND under emission scenarios SRES B1, A1B, A2, RCP4.5,
and RCP8.5. The central line within each box represents the weighted median value of the
model ensemble. The top and bottom of each box show the weighted 83th and 17th percentiles,
enclosing 66% of the data, and the top and bottom of each whisker display the weighted 95th
and 5th percentiles, respectively. Gray boxes denote statistically significant changes with a 95%
probability (t test).
JUNE 2013 SEILER ET AL. 1311

and significant increases and decreases in precipitation

and SW radiation.

4. Results
a. Validation and ensemble weighting
GCMs revealed an overall cold bias with RMSE
ranging from 38 to 78C and high correlations ranging
from 0.85 to 0.95 (Fig. 2). Most GCMs overestimated
spatial and/or temporal variability, with higher standard
deviations (58–138C) relative to the observed overall
standard deviation of 78C.
For rainfall, GCMs revealed an overall wet bias with
RMSE ranging from 40 to 140 mm month21. Correla-
tions strongly varied among GCMs, with coefficients
ranging from 0.3 to 0.9. Most GCMs modeled higher
standard deviations (60–160 mm month21) relative to the
observed overall standard deviation of 75 mm month21.
For SW radiation, GCMs revealed an overall positive
SW radiation bias with RMSE ranging from 15 to
40 W m22. As for rainfall, correlations strongly varied
among GCMs, with coefficients ranging from 0.5 to
0.96. Most GCMs had higher standard deviations (30–
50 W m22) relative to the observed overall standard
deviation of 35 W m22.
GCMs differed most with respect to precipitation,
differed by less for radiation, and differed least for tem-
perature, but none of the GCMs performed best in all
three variables (Fig. 3). The highest and lowest final
weights differed by a factor of 9. The 10 best GCMs
consisted of 7 CMIP3 and 3 CMIP5 GCMs, with the
Japanese CMIP3 model MRI-CGCM2.3.2 having ach-
ieved the highest weight. Some GCMs improved from
CMIP3 to CMIP5 (e.g., MIROC-ESM), some hardly
changed (e.g., MPI-ESM-LR), and others worsened
(e.g., CanESM2) in reproducing Bolivia’s historical cli-
mate. The mean weights of CMIP3 and CMIP5 were
about equal (0.029 and 0.028, respectively), while the
mean weight of the 12 GCMs from the CMIP3 sub-
sample had a slightly higher weight of 0.035. Hence, the
skill to reproduce the historical climate of Bolivia has
on average not improved from CMIP3 to CMIP5.
b. Temperature projections
Changes in temperature were statistically significant
throughout Bolivia and coincided with changes in CO2
concentrations, with increasing values from B1, RCP4.5,
A1B, A2, to RCP8.5. Weighted median changes of annual
FIG. 5. Changes in monthly mean temperatures (8C) from 1961–90
mean temperature were 2.68, 2.98, 3.88, 4.48, and 5.68C for to 2070–99 for emission scenarios SRES B1, A1B, and A2, as well
emission scenarios B1, RCP4.5, A1B, A2, and RCP8.5, as RCP4.5 and RCP8.5. Also given are observed mean tempera-
respectively, with slightly higher medians in the Andes tures (OBS) from 1961 to 1990 for each region.
(Fig. 4). Changes in monthly temperatures were generally
higher during the late dry season (August–November) and
1312 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

FIG. 6. Number of GCMs per grid cell agreeing on significant (a)–(c) increases and (d)–(f) decreases in annually accumulated precipitation.
(g)–(i) Mean changes (%) of annually accumulated precipitation (not weighted), and (j)–(l) number of GCMs per grid cell agreeing on
significant changes in interannual variance in August. The results apply to SRES emission scenarios (left) A1B, (center) RCP4.5, and (right)
RCP8.5. The CMIP3 ensemble was reduced to 12 GCMs, using models with corresponding versions in CMIP5 (B, D, G, J, M, N, O, P, R, S, V,
and W). The periods of comparison are 1961–90 and 2070–99. Significance was tested with a t test with a 95% probability.

were less high during the wet season (December–February) SRES B1, and the number of GCMs with significant
(Fig. 5). Monthly weighted median changes ranged from increases and decreases was about equal (not shown).
28 (SRES B1) to 88C (RCP8.5) in the northern lowlands Under SRES A1B and A2, about one-half of the GCMs
and from 28 (SRES B1) to 68C (RCP8.5) in both southern projected significant changes, but also here there was no
lowlands and Andes. Weighting GCMs had no major im- clear preference for either increases or decreases. This
pact on the resulting temperature scenarios. result also applied to the reduced SRES sample of 12
c. Precipitation projections GCMs, which only included predecessors of the CMIP5
GCMs used in this study (Figs. 6a,d). Under RCP4.5,
1) ANNUAL PRECIPITATION
most GCMs projected no significant changes; among the
A count of the number of GCMs that agree on statis- remaining GCMs, however, there were more models
tically significant changes in annual precipitation revealed with statistically significant decreases than increases in
that most GCMs projected no significant changes under both the northern and southern lowlands (Figs. 6b,e).
JUNE 2013 SEILER ET AL. 1313

FIG. 7. As in Fig. 4, but for relative changes in annually accumulated precipitation (%).

The periods of comparison are 1961–90 and 2070–99. in the lowlands, as well as in most of the Amazon.
Under RCP8.5, more than one-half of the GCMs (up This was most evident in October under RCP8.5 with
to 7 of 11 GCMs) projected significant decreases in the 6–10 of 11 GCMs projecting statistically significant
lowlands, with mean changes of up to 215% (Fig. 6i). decreases throughout most of the Amazon basin and
This tendency for less rainfall was not just restricted no model projecting significant increases (not shown).
to the Bolivian lowlands, but was also present in large In Bolivia the decrease in monthly precipitation was
parts of the Amazon basin, with mean decreases of up strongest in the northern lowlands with changes of
to 220% at the equator (Figs. 6g–i). Most GCMs from the median by 229 mm month21 (219%) in November
both ensembles did not project significant changes in under RCP8.5 (Fig. 8). This decrease was accompanied
the interannual variability of annual precipitation (not by significant changes in interannual variability, pro-
shown). jected by more than one-half of the GCMs from both
The decrease of annual rainfall projected by the ensembles, mainly during JJA (up to 10 of 12 GCMs
CMIP5 ensemble in the Bolivian lowlands was largest under RCP8.5; see Figs. 6j–l). During the wet months
for RCP8.5 (29% median and 217% 5th percentile; see of January–March, CMIP3 GCMs projected mainly
Fig. 7). In the Andes, the CMIP3 ensemble tended to- an increase in rainfall in the lowlands (up to 8 of
ward slightly less annual rainfall (23% median and 27% 18 GCMs under SRES A2), as well as in most of the
5th percentile) while the CMIP5 ensemble tended toward Amazon basin (not shown). In Bolivia this increase
slightly more annual rainfall (13% median and 15% was biggest under SRES A2 (115 mm month21 and
83rd percentile). The projections of the individual GCMs 18%) and was absent for the CMIP5 ensemble (Fig. 8).
varied so much, however, that changes of the weighted In the Andes, the CMIP3 ensemble tended toward less
ensembles were not statistically significant. Weighting rainfall (211 mm month21 and 29%) while the CMIP5
GCMs had no major impact on the projected changes. ensemble tended toward more rainfall (124 mm month21
In summary, changes in annual rainfall remained un- and 120%) in parts of the wet season. For annual
certain in the lowlands for CMIP3 while CMIP5 GCMs totals, the projections of the individual GCMs varied
were more inclined to project decreases (29%) there, as so much that changes in the weighted ensembles were
well as in most of the Amazon basin. In the Andes, not statistically significant. Weighting GCMs had no
CMIP3 GCMs tended toward slightly less annual rain- major impact on the changes projected. In summary,
fall (23%) while CMIP5 tended toward slightly more both ensembles agreed on less rainfall (219%) in
annual rainfall (13%). the lowlands during drier months (JJA and SON),
with significant changes in interannual rainfall vari-
2) MONTHLY PRECIPITATION
ability, but disagreed on changes during wetter months
Counting the number of GCMs agreeing on statis- (JFM). In the Andes, CMIP3 GCMs tended toward less
tically significant changes in monthly precipitation monthly rainfall (29%), while CMIP5 tended toward
revealed that both ensembles were clearly more in- more (120%) monthly rainfall during parts of the wet
clined to project less rainfall from July to November season.
1314 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

d. Shortwave radiation projections

Under all scenarios, at least one-half of the GCMs
predicted a significant increase and only a very few
predicted a decrease in annual SW radiation in the
northern lowlands (not shown). This signal became
stronger from the SRES to the RCP scenarios, with the
latter predicting more annual SW radiation in the southern
lowlands as well. In the Andes, no clear signal was vis-
ible under the SRES scenarios while under the RCP
scenarios most GCMs predicted a significant increase.
The medians of the regional changes in annual SW
radiation of the weighted ensemble were positive across
all regions and scenarios, with 66% of the ensemble
predicting an increase under emission scenarios SRES
B1, RCP45, and RCP85 (Fig. 9). Weighted median
changes in the annual mean SW radiation ranged from
1% to 3%, with likely changes as big as 111%.
Projected changes in monthly SW radiation showed
the biggest increases for scenarios and seasons with de-
creases in rainfall. Increases in the median were as large
as 20 W m22 (9%) in September in the northern low-
lands, 9 W m22 (3%) in February in the southern low-
lands, and 13 W m22 (5%) in December in the Andes
(Fig. 10). Despite the overall trend for more radiation,
changes were not statistically significant.

5. Discussion
We validated 35 GCMs against observed surface air
temperature, precipitation, and incoming SW radiation
and analyzed climate change projections from five emis-
sion scenarios, distinguishing among three climatologi-
cally contrasting regions in Bolivia. GCMs revealed an
overall cold, wet, and positive SW radiation bias and
showed no substantial improvement from the CMIP3
to the CMIP5 ensemble for the Bolivian case. Models pro-
jected an increase in temperature (2.58–5.98C) and SW ra-
diation (1%–5%), with seasonal and regional differences.
In the lowlands, changes in annual rainfall remained un-
certain for CMIP3 while CMIP5 GCMs were more inclined
to project decreases (29%). This result also applied to most
of the Amazon basin, suggesting a higher risk of partial
biomass loss for the CMIP5 ensemble results. Both en-
sembles agreed on less rainfall (219%) during drier months
(JJA and SON), with significant changes in interannual
rainfall variability, but disagreed on changes during wetter
months (JFM). In the Andes, CMIP3 GCMs tended
FIG. 8. As in Fig. 5, but for changes in monthly precipitation toward less rainfall (29%) while CMIP5 tended toward
(mm month21). more (120%) rainfall during parts of the wet season.
Our approach included the following five limitations:
1) We are aware that climate change projections have
only limited validity in mountainous regions because
JUNE 2013 SEILER ET AL. 1315

FIG. 9. As in Fig. 4, but for relative changes in annually accumulated SW radiation (%).

of the coarse spatial resolution of GCMs. High- distributed. We therefore tested significance for
resolution regional climate modeling would there- critical months with the nonparametric Mann–
fore be more appropriate in the Andes. Forcing a Whitney–Wilcoxon test, which did not lead to very
regional climate model with multiple lateral bound- different results. We were unfortunately not able
ary conditions, however, is very resource intensive, to implement this test for the weighted ensemble but
leading to very few model runs being available currently only for the unweighted ensemble, and therefore we
(e.g., Soares and Marengo 2009; Urrutia and Vuille adhered to the t test for weighted samples.
2009; Marengo et al. 2010). Most of these regional 5) The observation period of SW radiation lasted only
climate change scenarios are based on very few lateral from 1979 to 1990. We assumed that this period was
boundary conditions (mainly HadAM3), limiting their sufficiently long for obtaining mean values represen-
robustness in a region with little agreement among tative for the complete period 1961–90.
GCMs on the directional change of rainfall. We there-
fore consider our approach to be a necessary pre- Our results confirm numerous findings from Christensen
cursor to future downscaling efforts in the region. et al. (2007), Soares and Marengo (2009), Urrutia and
2) The stratification of Bolivia into three regions (north- Vuille (2009), and Marengo et al. (2010), including
ern lowlands, southern lowlands, and Andes) corre- 1) a cold and wet bias for the Andean region, 2) stronger
sponded to the coarse spatial resolution of most temperature increases during the austral winter, 3) great
GCMs but strongly simplified the true heterogeneity uncertainty associated with the directional change of
of the country. The merger of parts of the Andean annual rainfall, and 4) a tendency for an intensification
slopes with the Altiplano into one region neglected of the hydrological cycle for CMIP3 GCMs, with more
the contrasting precipitation regimes caused by the rainfall during the wet season. For SW radiation, our
steep orography of the Andes. Given the coarse re- positive bias agreed with the positive bias of 6 W m22 for
solution, however, such processes were hardly re- the global average found for multiple CMIP3 GCMs,
produced, making our stratification reasonable. which was most likely determined by processes in the
3) The final weights assigned to each GCM were to some cloud-free atmosphere rather than by an anomalous
extent subjective. We combined the weights related absorption through clouds (Wild 2008).
to temperature, precipitation, and SW radiation by Differences between projections from the CMIP3 and
forming the product, whereas a sum of the weights CMIP5 ensembles may emerge from differences in the
would have been an equally valid option. We preferred emission scenarios, resolution, or processes. Given the
the product over the sum because it led to stronger lack of a common scenario, both ensembles cannot be
differences among weights. Because weighting the compared directly. To make a comparison nevertheless,
ensembles had no major impact, however, uncer- Knutti and Sedl
acek (2013) calibrated the simple Model
tainties related to choice of method may be neglected. for the Assessment of Greenhouse Gas Induced Climate
4) The significance of monthly changes was tested with Change (MAGICC) to 19 CMIP3 models and ran it for
the Welch t test for weighted samples, despite the the RCP scenarios. On a global scale, CMIP5 projections
fact that for some cases the data were not normally seem to be largely consistent with CMIP3 projections.
1316 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 52

The same study, however, reveals different rainfall pro-

jections for both ensembles in the Amazon region,
suggesting that differences are not due to differences
in the scenarios alone but also to the addition of new
processes. Additional analysis will be required to de-
termine which processes are responsible for differences
in rainfall projections.
The projected changes may heavily affect human and
natural systems in Bolivia. Climate change impact studies
are required to further assess the potential implications
for different sectors under different scenarios, however.
Such studies could address potential risks for drinking
water supply from glaciers, hydropower, agricultural
production, and ecosystem stability. Given the large
uncertainty of rainfall projections, it will be essential to
incorporate a wide range of climate models in these
studies. Furthermore, it is recommended to try to identify
the factors that lead to a decrease in rainfall as well as
changes in interannual rainfall variability during JJA
in the Amazon. This process would require a detailed
analysis of how well GCMs reproduce the synoptic-scale
systems of South America and of how these systems
change under scenarios of climate change.
To conclude, we hope that this research has con-
tributed to a better understanding of climate change
projections in Bolivia and has provided a basis for a
discussion on climate change impacts and adaptation.
Our findings may provide inputs to further assess how
resilient human and natural systems are under dif-
ferent climate change scenarios.

Acknowledgments. This research was supported by

the Departmental Pilot Program of Adaptation to Cli-
mate Change (PDACC) as well as the Raising the Alert
about Critical Feedbacks between Climate and Land
Use Change in Amazonia project (AMAZALERT).
PDACC is carried out by the Fundaci
on Amigos de la
Naturaleza (FAN) as well as the departmental govern-
ment of Santa Cruz and is funded by the embassy of the
Netherlands. AMAZALERT is jointly funded by the
European Seventh Framework Programme and national
organizations. We thank the Bolivian SENAMHI for
the provision of the meteorological data. MERRA data
used in this study were provided by the Global Modeling
and Assimilation Office at the NASA Goddard Space
Flight Center. We acknowledge the World Climate Re-
search Programme’s Working Group on Coupled Mod-
eling, which is responsible for CMIP, and we thank the
FIG. 10. As in Fig. 5, but for changes in monthly SW radiation
(W m22). climate modeling groups (listed in Table 1 of this paper)
for producing and making available their model output.
For CMIP, the U.S. Department of Energy’s Program for
Climate Model Diagnosis and Intercomparison provides
coordinating support and leads development of software
JUNE 2013 SEILER ET AL. 1317

infrastructure in partnership with the Global Organiza- University Press, 570 pp. [Available online at http://www.ipcc.
tion for Earth System Science Portals. We are grateful for ch/ipccreports/sres/emission/index.php?idp50.]
Nunez, M., S. Solman, and M. Cabr
e, 2009: Regional climate
the comments from three anonymous reviewers who
change experiments over southern South America. II: Climate
have helped us to improve the quality of this paper. change scenarios in the late twenty-first century. Climate Dyn.,
32, 1081–1095.
REFERENCES Pachauri, R. K., and A. Reisinger, Eds., 2007: Climate Change 2007:
Synthesis Report. IPCC, 104 pp. [Available online at http://
Boulanger, J., and Coauthors, 2010: A Europe–South America
www.ipcc.ch/publications_and_data/publications_ipcc_fourth_
network for climate change assessment and impact studies.
assessment_report_synthesis_report.htm.]
Climatic Change, 98, 307–329.
Rammig, A., and Coauthors, 2010: Estimating the risk of Amazo-
Bourrel, L., L. Phillips, and S. Moreau, 2009: The dynamics of
nian forest dieback. New Phytol., 187, 694–706.
floods in the Bolivian Amazon basin. Hydrol. Processes, 23,
Rienecker, M., and Coauthors, 2011: MERRA: NASA’s Modern-
3161–3167.
Bradley, R., M. Vuille, H. Diaz, and W. Vergara, 2006: Threats to Era Retrospective Analysis for Research and Applications.
water supplies in the tropical Andes. Science, 312, 1755–1756. J. Climate, 24, 3624–3648.
Christensen, J., and Coauthors, 2007: Regional climate projections. Ronchail, J., and R. Gallaire, 2006: ENSO and rainfall along the
Climate Change 2007: The Physical Science Basis, S. Solomon Zongo valley (Bolivia) from the Altiplano to the Amazon
et al., Eds., Cambridge University Press, 847–940. basin. Int. J. Climatol., 26, 1223–1236.
CONARADE, 2010: Plan de atenci
on de la emergencia human- Seiler, C., R. W. A. Hutjes, and P. Kabat, 2013: Climate variability
itaria y agropecuaria por sequia para el chaco Boliviano and trends in Bolivia. J. Appl. Meteor. Climatol., 52, 130–146.
(Humanitarian and agricultural emergency care plan for SENAMHI, 2009: Climatolog
ıa de Bolivia. Servicio Nacional de
drought for the Bolivian chaco). Estado Plurinacional de Meteorologia e Hidrologia Internal Rep., 27 pp.
Bolivia Consejo Nacional para la Reducci
on de Riesgos y Seth, A., J. Thibeault, M. Garcia, and C. Valdivia, 2010: Making
Atenci
on de Desastres y Emergencias Tech. Rep., 50 pp. sense of twenty-first-century climate change in the Altiplano:
[Available online at http://defensacivilpotosi.com/convenios/ Observed trends and CMIP3 projections. Ann. Assoc. Amer.
PLAN%20SEQUIA%202010.pdf.] Geogr., 100, 835–847.
Garreaud, R., 2000: Cold air incursions over subtropical South Silva, V. B. S., and V. E. Kousky, 2012: The South American mon-
America: Mean structure and dynamics. Mon. Wea. Rev., 128, soon system: Climatology and variability. Modern Climatology,
2544–2559. S.-Y. Wang and R. R. Gillies, Eds., InTech, 123–152.
——, M. Vuille, R. Compagnucci, and J. Marengo, 2009: Present- Soares, W., and J. Marengo, 2009: Assessments of moisture fluxes
day South American climate. Palaeogeogr. Palaeoclimatol. east of the Andes in South America in a global warming sce-
Palaeoecol., 281, 180–195. nario. Int. J. Climatol., 29, 1395–1414.
Jupp, T., P. Cox, A. Rammig, K. Thonicke, W. Lucht, and Taylor, K., 2001: Summarizing multiple aspects of model perfor-
W. Cramer, 2010: Development of probability density functions mance in a single diagram. J. Geophys. Res., 106, 7183–7192.
for future South American rainfall. New Phytol., 187, 682–693. UNDP, 2011: Tras las huellas del cambio clim
atico en Bolivia:
Knutti, R., and J. Sedl
acek, 2013: Robustness and uncertainties Estado del arte del conocimiento sobre adaptaci
on al cambio
in the new CMIP5 climate model projections. Nat. Climate climatico—Agua y seguridad alimentaria (On the trail of
Change, 3, 369–373. climate change in Bolivia: State of the art of knowledge on
Lenters, J., and K. Cook, 1997: On the origin of the Bolivian high adaptation to climate change—Water and food security). Pro-
and related circulation features of the South American cli- grama de las Naciones Unidas para el Desarrollo Tech. Rep.,
mate. J. Atmos. Sci., 54, 656–678. 144 pp. [Available online at http://cambioclimatico-pnud.org.
Li, W., R. Fu, and R. Dickinson, 2006: Rainfall and its seasonality bo/paginas/admin/uploaded/traslashuellas.pdf.]
over the Amazon in the 21st century as assessed by the cou- Urrutia, R., and M. Vuille, 2009: Climate change projections for
pled models for the IPCC AR4. J. Geophys. Res., 111, D02111, the tropical Andes using a regional climate model: Tempera-
doi:10.1029/2005JD006355. ture and precipitation simulations for the end of the 21st cen-
Marengo, J., W. Soares, C. Saulo, and M. Nicolini, 2004: Clima- tury. J. Geophys. Res., 114, D02108, doi:10.1029/2008JD011021.
tology of the low-level jet east of the Andes as derived from Vera, C., G. Silvestri, B. Liebmann, and P. Gonz
alez, 2006: Climate
the NCEP–NCAR reanalyses: Characteristics and temporal change scenarios for seasonal precipitation in South America
variability. J. Climate, 17, 2261–2280. from IPCC-AR4 models. Geophys. Res. Lett., 33, L13707,
——, and Coauthors, 2010: Future change of climate in South doi:10.1029/2006GL025759.
America in the late twenty-first century: Intercomparison of Vuille, M., 1999: Atmospheric circulation over the Bolivian Alti-
scenarios from three regional climate models. Climate Dyn., plano during dry and wet periods and extreme phases of the
35, 1073–1097. Southern Oscillation. Int. J. Climatol., 19, 1579–1600.
Mastrandrea, M., and Coauthors, 2010: Guidance note for lead Wild, M., 2008: Short-wave and long-wave surface radiation bud-
authors of the IPCC Fifth Assessment Report on consistent gets in GCMs: A review based on the IPCC-AR4/CMIP3
treatment of uncertainties. Intergovernmental Panel on Cli- models. Tellus, 60A, 932–945.
mate Change Guidance Note, 5 pp. [Available online at World Bank, 2010: Adaptation to climate change—Vulnerability
https://www.ipcc-wg1.unibe.ch/guidancepaper/ar5_uncertainty- assessment and economic aspects: Plurinational State of Bolivia.
guidance-note.pdf.] World Bank Tech. Rep., 92 pp. [Available online at http://
Moss, R., and Coauthors, 2010: The next generation of scenarios for climatechange.worldbank.org/sites/default/files/documents/
climate change research and assessment. Nature, 463, 747–756. EACC_Bolivia.pdf.]
Nakicenovic, N., and Coauthors, 2000: IPCC Special Report: Zhou, J., and K. Lau, 1998: Does a monsoon climate exist over
Emissions Scenarios—Summary for Policymakers. Cambridge South America? J. Climate, 11, 1020–1040.