Weather and Climate Extremes 12 (2016) 15–23

Contents lists available at ScienceDirect

Weather and Climate Extremes

journal homepage: www.elsevier.com/locate/wace

Statistical downscaling of regional climate model output to achieve

projections of precipitation extremes
Eric M. Laflamme a,n, Ernst Linder b, Yibin Pan b
a
Department of Mathematics Plymouth State University, MSC 29, 17 High Street Plymouth, NH 03264, United States
b
Department of Mathematics and Statistics University of New Hampshire, W383 Kingsbury Hall, 33 Academic Way Durham, NH 03824, United States

art ic l e i nf o a b s t r a c t

Article history: In this work we perform a statistical downscaling by applying a CDF transformation function to local-
Received 2 July 2015 level daily precipitation extremes (from NCDC station data) and corresponding NARCCAP regional cli-
Received in revised form mate model (RCM) output to derive local-scale projections. These high-resolution projections are es-
30 October 2015
sential in assessing the impacts of projected climate change. The downscaling method is performed on
Accepted 21 December 2015
58 locations throughout New England, and from the projected distribution of extreme precipitation
Available online 22 December 2015
local-level 25-year return levels are calculated. To obtain uncertainty estimates for return levels, three
Keywords: procedures are employed: a parametric bootstrapping with mean corrected confidence intervals, a non-
Statistical downscaling parametric bootstrapping with BCa (bias corrected and acceleration) intervals, and a Bayesian model. In
Generalized Pareto distribution
all cases, results are presented via distributions of differences in return levels between predicted and
Regional climate models
historical periods. Results from the three procedures show very few New England locations with sig-
Bootstrapping
Bayesian analysis nificant increases in 25-year return levels from the historical to projected periods. This may indicate that
Uncertainty quantification projected trends in New England precipitation tend to be statistically less significant than suggested by
many studies. For all three procedures, downscaled results are highly dependent on RCM and GCM model
choice.
& 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction greater frequency of heavy-precipitation and extreme events.

Precipitation extremes are a primary concern as these events
There is great societal interest in assessing the impacts of are typically more impactful than precipitation events alone and
projected climate change, and more specifically, there is an intense are responsible for a disproportionately large part of climate-re-
interest in the impact of change in variability and extreme events lated damages (Kunkel et al., 1999; Easterling et al., 2000; Meehl
that could accompany global climate change predictions (Tebaldi et al., 2000). Natural systems may also be affected by changes in
et al., 2006). Increases in these extremes have already been ob- precipitation extremes, as these events have been shown to cause
served as precipitation events, heat waves, and drought are oc- shifts in ecosystem distributions, to trigger extinctions, and to alter
curring with greater intensity and frequency over the past few species morphology and behavior (Parmesan et al., 2000). Fur-
decades (U.S. Climate Change Science Program (USCCSP), 2008). thermore, extreme rainfall often translates into extreme flooding
and consequently great material and economic losses, erosion and
Other analyses have provided additional evidence that precipita-
damage to crops, collapse of lifeline infrastructure, the breakdown
tion extremes are becoming more extreme and will continue to do
of public health services (Douglas and Fairbank, 2011), fatalities
so in the future (e.g. Zwiers and Kharin, 1998; Groisman et al.,
(Kunkel et al., 1999), and structural damage to dams, bridges, and
1999; Meehl et al., 2000; Tank and Konnen, 2003; Karl and Knight,
coastal roads.
1998; Kharin and Zwiers, 2005). In their survey of recent projec- In this work, we outline a procedure to examine potential
tions of climate extremes provided by global circulation models change in precipitation extremes in New England. In the following
(GCMs), Tebaldi et al. (2006) concluded that models agree with section (Section 2), we discuss methods of statistical/probabilistic
observations over the historical period and that there is a trend downscaling as well as some elements of extreme value theory
towards a world characterized by intensified precipitation, with a germane to our analysis. In Section 3, we describe our data: pre-
cipitation observations for locations throughout New England,
n
Corresponding author.
historical climate model output, and projected series for future
E-mail addresses: emlafl[email protected] (E.M. Laflamme), precipitation. Section 4 outlines our downscaling approach and
[email protected] (E. Linder), [email protected] (Y. Pan). discusses all details of our methods. Section 5 presents our

http://dx.doi.org/10.1016/j.wace.2015.12.001
2212-0947/& 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
16 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23

downscaling results. Lastly, in Section 6, a discussion and conclu- FY (y)=FX (x) with y=FY−1 ( FX (x) )
sion is presented.
where F is a CDF of a climate random variable. Once a mapping has
been defined, it is then applied to large-scale dataset to create a
2. Downscaling local-scale prediction. The method does not take into account the
information of the distribution of the future modeled dataset
Atmosphere-ocean general circulation models, or AOGCMs, are (Michelangeli et al., 2009). Furthermore, the method of quantile
coupled atmosphere and ocean models that simulation weather at mapping cannot provide local-scale quantiles outside the range of
a global scale. AOGCMs are the main component of global climate the historical observations (Michelangeli et al., 2009). Proposed by
models (GCMs) which are the primary tools used to quantify and Wood et al. (2004), the technique was applied to downscale
assess climate change impacts (Wilby and Harris, 2006). However, monthly precipitation and temperature output from a GCM, and
because global weather simulation is so computationally ex- became known as bias-correction and spatial downscaling (BCSD).
pensive, these models provide predictions at an extremely coarse To overcome the shortcomings of the quantile matching
scale (250 KM by 250 KM, in most cases). The issue is that en- methodology, Michelangeli et al. (2009) proposed an extension to
vironmental impact models are sensitive to local climate char- this mapping called the CDF-t. The CDF-t is similar to quantile
acteristics, and the drivers of local climate variation are not cap- mapping as it compares local- and large-scale distributions, but it
tured at the coarse scales of GCMs (Maurer and Hidalgo, 2008). accounts for changes in the large-scale CDF between historical and
That is, GCMs do not provide an accurate description of local cli- future periods. Let X denote a variable from climate model output
mate. To overcome this discrepancy, methods of ‘downscaling’ are and let XC denote the series of the variable over the current, or
applied to produce local-scale climate predictions based on cor- calibration, period. Then, XP denotes the variable projected into
responding GCM scenarios. the future, the time series from runs of the climate model in the
Downscaling appears in two forms: Dynamical and statistical future. Similarly, let YC and YP denote the current and future series
downscaling (or empirical statistical downscaling, ESD). Dynami- for the local-level station. We note that while YC is observed, YP will
cal downscaling is a computationally-intensive technique which need to be predicted or downscaled. Finally, a transformation, T (∙),
makes use of the lateral boundary conditions combined with re- is assumed to exist between the large- and local-scale variable
gional-scale forcings such as land-sea contrast, vegetation cover, such that T (∙):[0,1]→[0,1]. We then have the relationship:
etc., to produce regional climate models (RCMs) from a GCM. RCM
outputs are typically produced over regular geographic grids with (
FYP (x)=T ( F XP (x) )=FYC F X−C1 ( F XP (x) ) ) (1)
scales in the tens of kilometers.
Statistical downscaling (SD), on the other hand, is a computa- where FYP and F XP are the respective empirical CDFs for the local-
tionally less demanding alternative that may be applied to achieve and large-scale prediction, and FYC and F XC are the respective CDFs
a variety of results. Essentially, statistical downscaling is a two- of observed (historical) local-level data and observed large-scale,
step process consisting of 1) the development of statistical re- or regional data. For further details see Michelangeli et al. (2009).
lationships between local climate variables and large-scale pre- The improvement over quantile mapping is that the future, local-
dictors, and 2) the application of such relationships to the output scale distribution is a function of both historical observations and
of large-scale output to simulate local climate characteristics in the large-scale information that may be distributed differently be-
future (Hoar and Nychka, 2008). Statistical downscaling is a rea- tween calibration and projection periods.
listic approach to develop a specific, local-level climate prediction. However, for precipitation data, we are more concerned with
Typically, SD methods are applied to GCM projections, but may the extreme events. In these cases, where the tails, which corre-
also be applied to RCM output as these results may not be re- spond to the extremes or high quantiles, are of primary interest,
presentative for the local climate (Skaugen et al., 2002; Engen- the non-parametric CDF-t is not ideal. Generally speaking, these
Skaugen, 2004). Furthermore, RCM output may simply have in- rare, extreme values result in empirical CDFs for precipitation that
adequate spatial resolution for some impact studies, and hence are heavy-tailed. With few data at the extreme ends of the dis-
additional statistical downscaling must be applied to the dyna- tribution, non-parametric quantile estimates in these tails have
mical model results (Benestad et al., 2007). large variance and they may be strongly influenced by a single
extreme event. Also, observations of historical changes, as well as
2.1. Probabilistic downscaling future projections, confirm that changes in the distributional tails
of precipitation (extremes) may not occur in proportion to changes
This analysis focuses on a method of ‘probabilistic downscaling’ in the mean and may not be symmetric in nature (Kharin and
to project a single variable, extreme precipitation, into the future. Zwiers, 2005; Robeson, 2004; Tank and Konnen, 2003; Easterling
While traditional ESD models the link between large- and local- et al., 2000).
scale variables, probabilistic downscaling is a type of statistical In light of this, Kallache et al. (2011) proposed the XCDF-t
downscaling that models the relationship between large- and lo- technique to downscale the distribution of extremes exclusively.
cal-scale statistical entities. In this case, the statistical entities are The technique is analogous to the CDF-t technique of Michelangeli
the corresponding cumulative distribution functions (CDFs) of the et al. (2009) in that is makes use of the same transformation
large- and local-scale precipitation extremes. In this way, prob- function form (see Eq. (1)) to link large- and local-scale distribu-
abilistic downscaling techniques do not retain the chronology, or tions of climate variables. Unlike the CDF-t method, however, the
exact ordering, of events. However, accurate descriptions of future XCDF-t links estimated parametric distributions of large- and lo-
climate distributions are themselves sufficient predictions as we cal-scale extremes only. To do this, ‘exceedances over a threshold’
do not aim to predict weather, but rather the distribution of a based on extreme value theory (EVT) are used to fit appropriate
weather variable (precipitation extremes). distributions to extremes based on limiting properties of max-
When working exclusively with cumulative distribution func- stable processes (See, for example, Coles, 2001). The framework of
tions, the simplest form of downscaling is what is referred to as EVT allows for more precise estimation of the extreme portions of
‘quantile mapping’ or ‘quantile matching’. This non-parametric distributions.
technique downscales a large-scale value x by selecting a local- For the XCDF-t, F XP , FYC , and F XC are cumulative Generalized
scale value y based on the following: Pareto distribution (GPD) for the extremes of the modeled
 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23 17

predicted, local-level observed (historical), and modeled historical boundary conditions provided by global circulation models, or
series, respectively. Each of the GPD CDFs have the following form: GCMs. Thus, RCMs are highly dependent on their GCM ‘drivers.’ A
−1 total of eight different RCM/GCM combinations were used in
⎛ ξ (z − u) ⎞⎟ ξ conjunction with the NCDC station-level data. Information re-
Fu (z )=1 − ⎜ 1+ ,
⎝ σ ⎠ garding the specific RCMs used is listed below (Table 1). We also
where, for z > u are the exceedances above some sufficiently large note that all GCMs have been forced with the SRES A2 greenhouse
threshold, u, and where ξ and σ are shape and scale model gas emissions scenario for the 21st century (See, Nakicenovoic
parameters, respectively (Pickands, 1975; Balkema and de Haan, et al. (2000) for Special Report on Emissions Scenarios (SRES)
1974). The qualitative behavior of the GPD is dominated by ξ , the commissioned by the IPCC).
shape parameter. For ξ > 0, the distribution of excesses is un- For each of the eight cases, current RCM data, three-hourly
values of 42 climate variables are available for the period 1970–
bounded and has the traditional ‘Pareto’, heavy tail; for ξ < 0, the
1999. Similarly, three-hourly RCM projected model outputs are
distribution has a finite upper bound and resembles an inverted
available for the period 2040–2069. Large-scale precipitation
Weibull-type distribution; and ξ = 0 corresponds to an un-
output was converted from instantaneous flux in units of kg/m2s
bounded, exponential-type distribution.
to three-hourly accumulations in inches. For both model outputs,
Lastly, we note that the XCDF-t method assumes that the re-
current and future (projected), and the station data, respective
lationship between large- and local-scale extreme remains con-
three-hourly and hourly accumulations were aggregated into daily
stant between calibration and prediction period. That is, the
totals of precipitation measured in inches.
models are based on historical data, and there is no guarantee that
the past statistical relationships between different data fields will
hold in the future. This so-called ‘stationarity’ assumption is made
4. Methodological details
with any ESD procedure.
4.1. Data matching

3. Data To downscale the 58 New England locations, each station must

be matched with a corresponding region from the eight RCMs to
Local-level precipitation data, YC , were obtained from the U.S. establish the required XC , XP , and YC series. Although all RCMs are
National Climatic Data Center (NCDC), a climate data archiving and comprised of equally sized 50 km by 50 km grids, these layouts are
retrieval system operated by the U.S. National Oceanographic and not identical. In the majority of cases, stations are matched to
Atmospheric Administration (NOAA) Satellite and Information model output based simply on gridded boundaries. In few coastal
Service (http://www.ncdc.noaa.gov/). Hourly precipitation accu- locations, however, where nearest grids were predominately
mulations reported in hundredths of inches were originally ob- ocean-related, stations were matched to the nearest terrestrial
tained for 69 meteorological stations from all six New England gridpoints. Ultimately, each of the precipitation data from the 58
states covering a period from 1948 to 2010. Because long series stations was matched to eight corresponding historical and future
were required to observe extreme events, few stations with RCM outputs.
shorter precipitation series were omitted from our analysis. In the
end, only stations with continuous measurements between 1970 4.2. Generalized Pareto Distribution and CDF transfer function
and 2000 were retained, a total of 58 locations of the original 69
New England monitoring stations. Following the method of Kallache et al. (2011), a GPD is first
Regional climate model (RCM) output is used for large-scale fitted to the precipitation exceedances associated with the three
precipitation data, both current and projected, or XC and XP from available series, XC , XP , and YC . This requires selecting a threshold
Eq. (1), respectively. Four regional model outputs used in down- for the three series, XC , XP , and YC , to distinguish extreme values.
scaling were acquired from the North American Climate Change The method of Karl et al. (1996) was employed where a fixed, high
Assessment Program (NARCCAP), an international program which percentile was chosen as a threshold. In our case, the 98.5th per-
acts as a custodian for regional climate simulations to be used in centile was chosen as a threshold for all series, at all stations,
impact assessment and research (Mearns et al., 2012). Each of the which equates to 5 observations per year, on average, or about 200
RCM outputs is driven by larger-scale atmospheric and oceanic observations for the entire collection period. Also, we coerce the

Table 1
RCM/GCM combinations used for large-scale model output.

Combination RCM Name Modeling group GCM name

CRCM-CGCM3 Canadian Regional Climate Model (Caya and Laprise, 1999) OURANOS / UQAM 3rd Generation Coupled Global Climate Model
(McFarlane et al., 2005)
CRCM-CCSM Canadian Regional Climate Model (Caya and Laprise, 1999) OURANOS / UQAM Community Climate System Model (Kiehl and Gent,
2004)
HRM3-GFDL Hadley Regional Model 3 / Providing Regional Climates for Hadley Centre Geophysical Fluid Dynamics Laboratory GCM (Del-
Impact Studies (Jones et al., 2003) worth et al., 2006)
HRM3-HADCM3 Hadley Regional Model 3 / Providing Regional Climates for Hadley Centre Hadley Centre Coupled Model, version 3 (Johns
Impact Studies (Jones et al., 2003) et al., 2003)
RCM3-GFDL Regional Climate Model, version 3 (Giorgi et al., 1993a, b; Pal University of California, Santa Geophysical Fluid Dynamics Laboratory GCM (Del-
et al., 2007) Cruz worth et al., 2006)
RCM3-CGCM3 Regional Climate Model, version 3 (Giorgi et al., 1993a, b; Pal University of California, Santa 3rd Generation Coupled Global Climate Model
et al., 2007) Cruz (McFarlane et al., 2005)
WRFP-CGCM3 Weather Research and Forecasting Model (Skamarock et al., Pacific Northwest National 3rd Generation Coupled Global Climate Model
2005) Laboratory (McFarlane et al., 2005)
WRFP-CCSM Weather Research and Forecasting Model (Skamarock et al., Pacific Northwest National Community Climate System Model (Kiehl and Gent,
2005) Laboratory 2004)
18 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23

fitted GPDs, F XC , F XP , and FYC from Eq. (1), to have positive or local-level 25-year return level is identified via application of Eq.
identically zero shape parameters corresponding to heavy and (2) and the corresponding quantile function, respectively. Differ-
exponential tails, respectively. In most cases, the shape parameters ences between historical and predicted values were observed. This
are positive, but few cases yield negative parameter estimates process was repeated 200 times (bootstrap iterations) to yield a
which were subsequently set to zero. This assumption is sup- distribution of differences in 25-year return level estimates. Lastly,
ported by the widely held belief that precipitation data, particu- from this distribution, ‘mean corrected’ bootstrap confidence in-
larly maxima and extremes, consistently appear as heavy-tailed tervals (See, for example, Davison and Hinkley, 1997) were
(e.g. Katz et al., 2002; Smith, 2001). Lastly, to produce a distribu- calculated.
tion of future precipitation exceedances at the local-level, FYP , the
transfer function (1) may be applied to the three fitted GPDs cor- 4.4.2. Nonparametric bias correction and acceleration (BCa) con-
responding to the XC , XP , and YC series. This process is done for fidence intervals
each station and model combination. Just as some skew (asymmetry) is present in the individual
distributions of 25-year return levels, it is likely that the dis-
4.3. Return levels tribution of differences in return levels between predicted and
current periods is also skewed. Furthermore, it is known that
A return level is a high quantile of an extreme value distribu- bootstrap procedures may not be consistent for extreme value
tion that will be exceeded with some known probability. Return problems as there is a tendency for samples to generate shorter
levels are universally understood and relate directly to location- tails than the true sample distribution (Coles and Simiu, 2003). In
specific climate impact assessments such as flooding, potential light of this, a non-parametric bootstrapping procedure with BCa
erosion, etc. As stated in Coles (2001), it is usually more con- (bias-corrected and accelerated) adjustment (Efron and Tibshirani,
venient to interpret extreme value models in terms of quantiles or 1993) was employed. In this procedure, BCa interval endpoints are
return levels, rather than individual parameter values. given by percentiles of the bootstrap distribution, but adjusted to
We extend the work of Kallache et al. (2011) to include climate account for the skew and bias of the data. The actual percentiles
projections in the form of return levels. In their downscaling of used depend on the acceleration and bias correction, which, gen-
extreme precipitation, Kallache et al. (2011) focus on transforming erally speaking, measure the rate of change of the standard error
the entire distribution of large-scale predicted values to unknown, and median bias of our estimator, respectively. In practice, these
smaller-scale predicted values, an approach similar to Miche- values are estimated by repeatedly sampling the data, and re-
langeli et al. (2009). Instead of predicting an entire distribution of sulting BCa intervals are simply calculated.
exceedances, we have chosen to predict a specified return level via For this procedure, the three series, XC , XP , and YC , are re-
the quantile matching of the XCDF-t procedure. sampled with replacement. GPD models are fit to each sample and,
When estimating the q -quantile of YP , we can directly apply the from the fitted parameter estimates and a given q value associated
transformation (1) by noting that, if F X−P1 (q)=xP , and with the return level, both a predicted and historical local-level
25-year return level is identified via application of Eq. (2) and the
( (
ReturnLevel (q)=y^P (q)=FY−C1 F XC F X−P1 (q) )) (2) corresponding quantile function, respectively. The difference be-
tween the historical and predicted return level is then calculated.
since the two quantiles are matched by having the same cu- This resampling, fitting, and calculation process is repeated 500
mulative value. In our case, for the r -year return level, the value q times to produce a distribution of differences in return levels be-
is given by: tween the predicted and historical periods. From this distribution
⎡ ⎤ in differences, the adjusted BCa interval is calculated as given by
1 ⎥
q=1 − ⎢ Efron and Tibshirani (1993, see pg. 185). We note that, compared
⎢⎣ r*ny *λ^u ⎥⎦
to the previous bootstrapping procedure, the procedure including
this bias correction is considerably more computer intensive.
where r is the return level period, ny is the number of observations
per year, and λ^u is the empirical threshold exceedance rate that 4.4.3. Bayesian models and credibility intervals
estimates λμ=Pr (X > u). In our case, r = 25, ny=365.25, and We further extend the analysis of Kallache et al. (2011) by
λ^u=. 015. Point estimates for local-level, predicted 25-year return pursuing a Bayesian model approach. Like the non-parametric
levels are produced from the application of the transfer function bootstrap with BCa approach, this approach is pursued to address
(2) to fitted GPD parameters estimated from the three ser- skewness and produce more accurate and precise intervals than
ies, XC , XP , and YC . the bootstrap procedure. As with our other approaches, pre-
cipitation exceedances are assumed to follow a GPD which is
4.4. Climate change uncertainty quantification asymptotically justified by extreme value theory. The Bayesian
procedure will result in distributions for these extreme value
4.4.1. Parametric bootstrapping and confidence intervals distribution parameters that will serve as a basis for downscaling
We further extend the work of Kallache et al. (2011) to include and ultimately return level estimation. That said, posterior esti-
a parametric bootstrapping procedure to calculate measures of mates are based on assumed prior distributions and results may be
uncertainty associated with return levels. Our parametric boot- sensitive to such specification.
strapping procedure was devised and implemented for all 58 lo- In the Bayesian approach, posterior distributions are obtained
cations (local-level stations) and all 8 RCM/GCM combinations. For using Bayes theorem. For ease of calculation, however, Markov
this procedure, GPD model forms were first fit to XC , XP , and Chain Monte Carlo (MCMC) techniques (see, for example, Gelman
YC series. Then, from each of the three pairs of parameter estimates et al., 2014) are used to simulate the posterior draws of the un-
(a shape and scale for each series), random exceedances were known quantities and obtain simulation-based estimates of pos-
generated. The result was a simulated series of extremes corre- terior parameters of interest. MCMC is often performed under a
sponding to each of the XC , XP , and YC series. For each of these Gibbs sampling framework where a sequence of samples from
simulated series, GPD parameters were refit via maximum like- conditional individual distributions is generated to approximate a
lihood. Next, for these parameter estimates and a given q value joint distribution.
associated with the return level, both a predicted and historical R statistical software (R Core Team, 2013) was used in
 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23 19

conjunction with the open-source software OpenBUGs (see http:// Finally, from this distribution of differences, a 90% credibility
www.openbugs.net) through the ‘R2WinBUGs’ R package to pro- (confidence) interval is extracted using the 5th and 95th
duce posterior distributions for the GPD parameters, ξ and σ . For percentiles.
each location, given the data from the three known series, XC , XP ,
and YC , posterior distributions for the respective parameters
ξ XC , σ XC , ξ XP , σ XP , ξYC , and σ YC were produced. For each series, a beta 5. Results
prior distribution was assumed for shape parameters. That is,
5.1. Return level point estimates
(
ξ XC ~Beta αξ XC , βξ X
C
),
From local-level, predicted return levels and return levels
produced from the local-level historical data (based on the fitted
GPD parameters for the YC data), we produce a bubble plot of the
(
ξ XP ~Beta αξ X P , βξ X
P
), differences (Fig. 1, below) at each of the 58 stations throughout
New England. This figure represents the projected effect of climate
change on extreme precipitation at the local-level via down-
(
ξ YC ~Beta αξ YC ,βξ Y
C
), scaling. In Fig. 1, red and blue circles indicate positive and negative
changes, respectively, and the size of the circles indicates the
where the beta parameters, alpha and beta ( α∙, β∙ ), are set to be magnitude of such change. We note that in this figure and all
1.5 and 2.5, respectively, for each series. These distributions are subsequent figures, dot sizes are on a common scale so that plots
semi-informative as they restrict estimates to values between are comparable.
0 and 1 (thus, fairly non-informative). As stated earlier, pre- We observe that results are sensitive to the choice of GCM
cipitation extremes typically follow a heavy-tailed or exponential- driver (GCM denoted by second part of combination title in lower
tailed distribution, so it is commonplace to assume a constant zero case). Among common RCMs, different GCM drivers yield con-
or slightly positive shape parameter and thus restrict our model to siderably different results (HRM3_gfdl versus HRM3_hadc, for
only such cases, or values deemed ‘sensible.’ example). Also, differences in station-level return levels are highly
For scale parameters, prior distributions were assumed to be dependent on our choice of RCM, and this choice yields con-
uniform (Gelman, 2006), or siderably different and occasionally contradictory results
(RCM3_cgcm versus WRFG_cgcm, for example). Finally, the station
(
σ XC ~Unif. a σ XC , bσ XC , ) results generally exhibit fewer decreases (negative values) than
increases. The largest positive values are consistently observed in
(
σ XP ~Unif. a σ X P , bσ X P , ) northern New Hampshire at the Pinkham Notch location. This is
realistic as this area has the highest elevation among all 58 sta-
tions, is known for extreme weather, and is likely sensitive to
(
σ YC ~Unif. a σ YC , bσ YC . ) climate change projections.

For these parameters, the uniform distribution is defined from

5.2. Parametric bootstrap confidence intervals
0 to 50 ( a and b, respectively). Thus, the prior specification is non-
informative as parameter estimates from traditional model-fitting Fig. 2 below illustrates the variability attributable to down-
were typically found to be between 0.2 and 1.2 for daily rainfall in scaling by observing the 90% bootstrapped lower bound (the 5th
inches. percentile) of the distribution of differences in 25-year return le-
The Gibbs sampler was initialized with starting va- vels between predicted and historical periods. Positive ( 40) lower
lues,ξ (0) ¼0.48 and σ (0) ¼ 3 for each series, values based on previous bounds are marked in red and identify locations with significant
model-fitting. For each station/model combination, individual increases in 25-year return levels from historical to future periods.
samplers, each based on 20,000 iterations, were run for the three Locations marked in blue indicate differences in return levels that
series to produce distinct sets of shape and scale posterior dis- do not show a significant increase.
tributions. MCMC sampled values require a ‘burn-in’ period before When comparing results from the model combinations, we
convergence to an ergodic Markov chain (before the Markov chain notice that the choice of GCM has an influence on the resulting
stabilizes), so 5000 initial iterations were discarded and not used differences in return levels. For example, plots for WRFG_ccsm and
for inference. For ease of calculations, the 15000 remaining itera- WRFG_cgcm (which share a common RCM) show conflicting re-
tions were further thinned to 5000. Thus, for each station/model sults at several locations in southern New England and northern
combination, 5000 sets (ξ XC , σ XC , ξ XP , σ XP , ξYC , and σ YC ) of parameter New Hampshire/Vermont. This suggests that our choice of GCM
estimates are produced. Convergence of the individual chains was does contribute to the variability observed in downscaling results.
then checked by diagnostic commands contained in the R CODA Also, lower bounds for station-level differences in return levels
package. show significant dependence on our choice of RCM. This is illu-
For each station/model combinations, and for each of the 5000 strated in the significant disagreement between results obtained
sets of GPD parameters ( ξ and σ ) fit to the exceedances of the from RCM3_cgcm, CRCM_cgcm, and WRFG_cgcm combinations,
XC , XP , and YC series, a GPD cumulative or quantile function is for example. We conclude that there is little agreement between
calculated as required for application of Eq. (2). For a certain the 8 RCM/GCM combinations in terms of return level projections,
q-quantile of YP , we apply Eq. (2) to produce a single, predicted 25- and both the regional model and GCM driver choice contribute to
year return level estimate. Additionally, for each of these sets of such variability.
parameter estimates, a 25-year return level is calculated for the YC Generally speaking, observing lower bounds for differences in
series, the station-level historical data. The difference in return return levels, we find significant increases at relatively few sta-
levels between predicted and historical series are calculated and, tions, mostly sporadic increases. There may be some weak evi-
after considering all 5000 sets of parameter estimates, a dis- dence of significant increases throughout locations in southern
tribution of differences in return levels between the predicted and New England (Connecticut, Massachusetts, and Rhode Island).
historical periods is produced for each station/model combination. While subtle, this trend is most clearly seen from the WRFG_cgcm
20 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23

CRCM_ccsm HRM3_gfdl RCM3_cgcm WRFG_ccsm

-2.82 0.11 0.8 1.32 4.91

CRCM_cgcm HRM3_hadc RCM3_gfdl WRFG_cgcm

Fig. 1. Difference in 25-year return level estimates between predicted and historical periods (predicted minus historical) for the 8 climate model combinations.

and HRM3_gfdl model combinations. Overall, our results above 5.3. Nonparametric bootstrap and BCa confidence intervals
(Fig. 2) are quite different from station results based on return
level point estimates only, where numerous increases were ob- Fig. 3 below compares the 90% lower bounds (5th percentile) of
served (Fig. 1). These comparisons illustrate the variability attri- the BCa interval for the differences in 25-year return levels be-
butable to the procedure, thus how downscaling adds uncertainty tween predicted and historical periods. As before, positive lower
to our results. bounds are marked in red and identify locations with significant
increases in 25-year return levels from historical to future periods.
From the plot above, the majority of lower bounds are negative,
but some significant increases are observed in northern New

CRCM_ccsm HRM3_gfdl RCM3_cgcm WRFG_ccsm

-4.93 -1 -0.66 -0.29 0.93

CRCM_cgcm HRM3_hadc RCM3_gfdl WRFG_cgcm

Fig. 2. 90% lower bound for distribution of differences in bootstrapped 25-year return levels between predicted and historical periods.
 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23 21

CRCM_ccsm HRM3_gfdl RCM3_cgcm WRFG_ccsm

-6.62 -1.06 -0.62 -0.14 2.43

CRCM_cgcm HRM3_hadc RCM3_gfdl WRFG_cgcm

Fig. 3. 90% lower bound for distribution of differences in BCa 25-year return levels between predicted and historical periods.

Hampshire (WRFG models) and throughout southern New Eng- 5.4. Bayesian credibility intervals
land (CRCM_cgcm and WRFG_cgcm combinations). Generally
speaking, we find that lower bounds of the BCa intervals are The intervals for differences in 25-year return levels produced
comparable to those obtained without the adjustment (Fig. 2). under the Bayesian framework generally tend to resemble those
That is, regardless of method (parametric bootstrap or non-para- produced via non-parametric bootstrapping with BCa adjustment.
metric bootstrap with BCa), increases and decreases follow the Below (Fig. 4) we compare our Bayesian results, the 90% credibility
same broad pattern across locations and model combinations. We intervals for difference in return levels, to the estimates and con-
fidence intervals obtained from both our parametric bootstrap and
note that the BCa adjustment, however, tends to inflate the posi-
non-parametric/BCa procedures. For readability, this plot shows
tive differences. Also, in general, there are more positive differ-
results for the six Connecticut locations and eight model combi-
ences associated with the BCa procedure. This is expected and
nations only.
consistent with our understanding of the BCa, how the procedure
Next, we compare the 90% lower bounds for differences in 25-year
is designed to capture the right-skew of the distribution of dif-
return levels between predicted and historical periods. These plots are
ferences, and how the procedure typically inflates lower con- directly comparable to those produced previously via parametric
fidence limits. Lastly, as before, there is little agreement between bootstrap and non-parametric/BCa methods. In Fig. 5 below, red and
the 8 RCM/GCM combinations, and both the regional model and blue circles indicate positive and negative changes, respectively, and
GCM driver choice contribute to such variability. the size of the circle indicates the magnitude of change.

Parametric bootstrap 1=Mansfield Hollow

BCa 2=Norfolk
10 Bayesian 3=Rockville
4=Storrs
5=Thomaston Dam
6=West Thompson
Return Level (inches)

CRCM_ccsm CRCM_cgcm HRM3_gfdl HRM3_hadc RCM3_cgcm RCM3_gfdl WRFG_ccsm WRFG_cgcm

1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

Fig. 4. Comparison of Bayesian intervals (blue) to parametric bootstrap (black) and BCa (red) intervals.
22 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23

CRCM_ccsm HRM3_gfdl RCM3_cgcm WRFG_ccsm

-4.7 -0.91 -0.63 -0.4 0.42

CRCM_cgcm HRM3_hadc RCM3_gfdl WRFG_cgcm

Fig. 5. 90% lower bound for distribution of differences in Bayesian 25-year return levels between predicted and historical periods.

From the plots above, the vast majority of locations have ne- support this claim. Differences in point estimates between local-
gative lower bounds (blue circles) which indicate differences in level observations and downscaled future estimates show wide-
return levels that do not show a significant increase between spread increases across most of New England, although many of
historical and predicted periods. The few significant increases (red these significant increases are not observed when we consider
circles) observed are primarily found at locations in southeast New measures of uncertainty. From the three techniques used to assess
England (HRM3_gfdl, HRM3_hadc, and WRFG_cgcm) in Con- uncertainty in our estimates (the parametric bootstrap, the non-
necticut, Massachusetts, and Rhode Island. This may provide some parametric bootstrap with BCa adjusted confidence interval, and
weak evidence of a trend in increases in return levels, but, as seen the Bayesian/MCMC framework), few significant increases in 25-
from previous results (parametric and non-parametric bootstrap), year return levels were observed, and only subtle evidence of
these results are dependent on choice of both GCM and RCM. significant increases throughout locations in southern New Eng-
land was identified. These results, however, are highly dependent
on model choice. Certainly, future work will entail developing
6. Conclusion and discussion techniques to decrease sampling variability and achieve more
precise estimates.
Theoretically, this work is heavily influenced by the work of The choice of RCM and GCM driver has been shown to impact
both Michelangeli et al. (2009) and Kallache et al. (2011) who the downscaling procedure substantially. While it was not the
respectively introduced the CDF-t and its parametric extension to primary purpose of this analysis, we have indirectly illustrated the
extremes, the XCDF-t. While this work is made possible by these effect this choice can have on downscaling and climate predic-
procedures, we extend the procedure in a variety of ways. First, we tions. Related to this, in their presentation of guidelines for
use the downscaling procedure to produce predicted, local-level downscaling climate variables, Wilby et al. (2004) note that it is
return level estimates, widely interpretable estimates critical to increasingly recognized that any comprehensive impact study
climate change impact assessment. To produce measures of un- should be founded on multiple GCM (or large-scale) model out-
certainty associated with these return levels, a parametric boot- puts. Schliep et al. (2010) analyzed historical output from six RCMs
strapping procedure was employed. Next, we develop a non- via a spatial Bayesian hierarchical model and found different RCMs
parametric resampling scheme using a BCa (bias correction and yielded substantially different 100-year return level estimates.
acceleration) confidence interval to account for potential skew- Going forward, quantifying the effect of RCM and GCM on pro-
ness/bias. Lastly, to further pursue measures of uncertainty asso- jections should be pursued. Such work may entail a functional
ciated with return levels, a Bayesian estimation framework is de- analysis of variance or functional ANOVA. Related and recent work
veloped. In all cases, for all techniques, results are presented via in this area has been undertaken by Kaufman and Sain (2010) and
comparison between predicted and historical return levels. Sain et al. (2010), for example.
Our results are most comparable to those of Douglas and Fair-
bank (2011) who investigated trends in extreme precipitation for
stations in northern New England. Unlike their work, however, our
References
results, the return levels, consider both predicted model output as
well as current data. In their work, the authors conclude that
Balkema, A., de Haan, L., 1974. Residual life time at great age. Ann. Probab. 2,
coastal areas of northern New England have significant increases 792–804.
in maximum precipitation records. Our results, however, do not Benestad, R.E., Hanssen-Bauer, I., Chen, D., 2007. Empirical Statistical Downscaling.
 E.M. Laflamme et al. / Weather and Climate Extremes 12 (2016) 15–23 23

World Scientific Publishing, Singapore. data: an inter-comparison of two statistical downscaling methods. Hydrol.
Caya, D., Laprise, R., 1999. A semi-implicit semi-lagrangian regional climate model: Earth Syst. Sci. 12, 551–563.
The Canadian RCM. Mon. Wea. Rev. 127, 341–362. McFarlane, N.A., Scinocca, J.F., Lazare, M., Harvey, R., Verseghy, D., Li, J., 2005. The
Coles, S., 2001. An Introduction To Statistical Modeling of Extreme Values. Springer, CCCma third generation atmospheric general circulation model CCCma Internal
London. Report, p. 25.
Coles, S., Simiu, E., 2003. Estimating uncertainty in the extreme value analysis of Mearns, L.O., Arritt, R., Biner, S., Bukovsky, M.S., McGinnis, S., Sain, S., Caya, D.,
data generated by a hurricane simulation model. J. Eng. Mech. 129 (11), Correia Jr., J., Flory, D., Gutowski, W., Takle, E.S., Jones, R., Leung, R., Moufouma-
1288–1294. Okia, W., McDaniel, L., Nunes, A.M.B., Qian, Y., Roads, J., Sloan, L., Snyder, M.,
Davison, A.C., Hinkley, D.V., 1997. Bootstrap Methods and Their Application. Cam- 2012. The north american regional climate change assessment program: over-
bridge University Press, New York. view of phase I results. Bull. Amer. Meteorol. Soc. 93, 1337–1362.
Delworth, T.L., et al., 2006. GFDL's CM2 global coupled climate models - Part 1: Meehl, G.A., Collins, W.D., Boville, B., Kiehl, J.T., Wigley, T.M.L., Arblaster, J.M., 2000.
Formulation and simulation characteristics. J. Clim. 19 (5), 643–674. Response of the NCAR climate system model to increased CO2 and the role of
Douglas, E.M., Fairbank, C.A., 2011. Is precipitation in northern new england be- physical processes. J. Clim. Am. Meteorol. Soc. 13, 1879–1898.
coming more extreme? statistical analysis of extreme rainfall in massachusetts, Michelangeli, P.-A., Vrac, M., Loukos, H., 2009. Probabilistic downscaling ap-
new hampshire, and maine and updated estimates of the 100-year storm. J. proaches: application to wind cumulative distribution functions. Geophys. Res.
Hydrol. Eng. 16 (3), 1. Lett. 36, L11708.
Easterling, D.R., Meehl, G.A., Parmesan, C., Changnon, S.A., Karl, T.R., Mearns, L.O., Nakicenovoic, N. et al., 2000. Special report on emissions scenarios a special report
2000. Is precipitation in northern new england becoming more extreme? sta- of working group III of the intergovernmental panel on climate change, Cam-
tistical analysis of extreme rainfall in massachusetts, new hampshire, and bridge: Cambridge University Press.
maine and updated estimates of the 100-Year storm. Science 289 (5487), Pal, J.S., et al., 2007. Regional climate modeling for the developing world: the ICTP
2068–2074. RegCM3 and RegCNET. Bull. Amer. Meteorol. Soc. 88, 1395–1409.
Efron, B., Tibshirani, R.J., 1993. An Introduction to the Bootstrap. Chapman and Hall/ Parmesan, C., Root, T.L., Willig, M.R., 2000. Impacts of extreme weather and climate
CRC, New York. on terrestrial biota. Bull. Am. Meteorol. Soc. 81, 443–450.
Engen-Skaugen, T., 2004. Refinement of dynamically downscaled precipitation and Pickands, J., 1975. Statistical inference using extreme order statistics. Ann. Stat. 3,
temperature scenarios. DNMI Climate Report 15, 1–20. 119–131.
Gelman, A., 2006. Prior distributions for variance parameters in hierarchical R Core Team, 2013. R: A language and environment for statistical computing. R
models. J. Stat. Anal. 1 (3), 515–533. Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0,
Gelman, A., Carlin, J.B., Stern, H.S., Rubin, D.B., 2014. Bayesian Data Analysis 2. http://www.R-project.org/.
Chapman & Hall/CRC, London. Robeson, S.M., 2004. Trends in time-varying percentiles of daily minimum and
Giorgi, F., Marinucci, M.R., Bates, G.T., 1993a. Development of a second-generation maximum temperature over North America. Geophys. Res. Lett. 31, L04203.
regional climate model (RegCM2). Part I: Boundary-layer and radiative transfer Sain, S.R., Nychka, D., Mearns, L., 2010. Functional ANOVA and regional climate
processes. Mon. Weather Rev. 121, 2794–2813. experiments: A statistical analysis of dynamic downscaling. Unpublished
Giorgi, F., de Canio, G., Bates, G.T., 1993b. Development of a second-generation manuscript. http://www.image.ucar.edu/  ssain/publications/narccap_gfdl.pdf.
regional climate model (RegCM2). Part II: Convective processes and assimila- Schliep, E.M., Cooley, D., Sain, S.R., Hoeting, J.A., 2010. A comparison study of ex-
tion of lateral boundary conditions. Mon. Wea. Rev. 121, 2814–2832. treme precipitation from six different regional climate models via spatial
Groisman, P. Ya, et al., 1999. Changes in the probability of heavy precipitation: hierarchical modeling. Extremes 13, 219–239.
important indicators of climatic change. Clim. Chang. 42, 243–283. Skamarock, W.C., Klemp, J.B., Dudhia, J., Gill, D.O., Barker, D.M., Wang, W., Powers, J.
Hoar, T., Nychka, D., 2008. Statistical downscaling of the community climate system G., 2005. A description of the advanced research version 2. NCAR Technical
model (CCSM) monthly temperature and precipitation projections. White paper Note NCAR/TN-4681STR, p. 88.
preprint, Institute for Mathematics Applied to Geosciences/National Center for Skaugen, T., Astrup, M., Roald, L.A., Skaugen, T.E., 2002. scenarios of extreme pre-
Atmospheric Research, Boulder, CO 80307. cipitation of duration 1 and 5 days for Norway caused by climate change.
Johns, T.C., Gregory, J.M., Ingram, W.J., Johnson, C.E., Jones, A., Lowe, J.A., Mitchell, J. Technical report. Norgesvassdrags-ogenergidirektorat, NVE, Postboks 5091
F.B., Roberts, D.L., Sexton, B.M.H., Stevenson, D.S., Tett, S.F.B., Woodage, M.J., Majorstua, 0301 Oslo, Norway.
2003. Anthropogenic climate change for 1860 to 2100 simulated with the Smith, R.L., 2001. Trends in Rainfall Extremes Technical report, Department of
HadCM3 model under updated emissions scenarios. Clim. Dyn. 20, 583–612. Statistics, University of North Carolina, Chapel Hill NC 27599-3260 USA.
Jones, R.G., Hassell, D.C., Hudson, D., Wilson, S.S., Jenkins, G.J., Mitchell, J.F.B., 2003. Tank, A.M.G. Klein, Konnen, G.P., 2003. Trends in indices of daily temperature and
Workbook on generating high resolution climate change scenarios using PRE- precipitation extremes in Europe, 1946–99. J. Clim., Am. Meteorol. Soc. 16,
CIS, UNDP, p. 32. 3665–3680.
Kallache, M., Vrac, M., Naveau, P., Michelangeli, P.-A., 2011. Non-stationary prob- Tebaldi, C., Hayhoe, K., Arblaster, J.M., Meehl, G.A., 2006. Going to the extremes: an
abilistic downscaling of extreme precipitation. J. Geophys. Res. 116, D05113. inter-comparison model-simulated historical and future changes in extreme
Karl, T.R., Knight, R.W., 1998. Secular trend of precipitation amount, frequency, and events. Clim. Chang. 79 (3–4), 185–211.
intensity in the United States. Bull. Am. Meteorol. Soc. 79, 231–242. U.S. Climate Change Science Program (USCCSP), 2008. Weather and climate ex-
Karl, T.R., Knight, R.W., Easterling, D.R., Quayle, R.G., 1996. Indices of climate change tremes in a changing climate, U.S. Climate Change Science, Program Synthesis
for the United States. Bull. Am. Meteorol. Soc. 77, 279–292. Analysis Production 3.3 〈http://www.climatescience.gov/Library/sap/sap3-3/fi
Katz, R.W., Parlange, M.B., Naveau, P., 2002. Statistics of extremes in hydrology. Adv. nal-report/sap3-3-final-allpdf〉.
Water Resour. 25, 1287–1304. Wilby, R.L., Harris, I., 2006. A framework for assessing uncertainties in climate
Kaufman, C.G., Sain, S.R., 2010. Bayesian functional ANOVA modeling using gaussian change impacts: low-flow scenarios for the River Thames, UK. Water Resour.
process prior distributions. Bayesian Anal. 5, 123–150. Res. 42, W02 419.
Kharin, V.V., Zwiers, F.W., 2005. Estimating extremes in transient climate change Wilby, R.L., Charles, S.P., Zorita, E., Timbal, B., Whetton, P., Mearns, L.O., 2004.
simulations. J. Clim. Am. Meteorol. Soc. 18, 1156–1173. Guidelines for use of climate scenarios developed from statistical downscaling
Kiehl, J.T., Gent, P.R., 2004. The community climate system model, version two. J. methods. Published on-line, supporting material to the IPCC, pp. 27.
Clim. 17, 3666–3682. Wood, A.W., Leung, L.R., Sridhar, V., Lettenmaier, D.P., 2004. Hydrologic implica-
Kunkel, K.E., Andsager, K., Easterling, D.R., 1999. Long-term trends in extreme tions of dynamical and statistical approaches to downscaling climate model
precipitation events over the conterminous United States and Canada. J. Clim. outputs. Clim. Chang. 62, 189–216.
Am. Meteorol. Soc. 12, 2515–2527. Zwiers, F.W., Kharin, V.V., 1998. Changes in the Extremes of the Climate Simulated
Maurer, E.P., Hidalgo, H.G., 2008. Utility of daily vs. monthly large-scale climate by CC GCM2 under CO2 Doubling. J. Clim. Am. Meteorol. Soc. 11, 2200–2222.