INTERNATIONAL JOURNAL OF CLIMATOLOGY
Int. J. Climatol. 20: 489–501 (2000)
PREDICTIONS OF CLIMATE CHANGE OVER EUROPE USING
STATISTICAL AND DYNAMICAL DOWNSCALING TECHNIQUES
JAMES MURPHY*
Hadley Centre for Climate Prediction and Research, Meteorological Office, Bracknell, Berks, UK
Recei6ed 10 No6ember 1998
Re6ised 10 July 1999
Accepted 25 July 1999
ABSTRACT
Statistical and dynamical downscaling predictions of changes in surface temperature and precipitation for 2080–2100,
relative to pre-industrial conditions, are compared at 976 European observing sites, for January and July. Two
dynamical downscaling methods are considered, involving the use of surface temperature or precipitation simulated
at the nearest grid point in a coupled ocean–atmosphere general circulation model (GCM) of resolution 300 km
and a 50 km regional climate model (RCM) nested inside the GCM. The statistical method (STAT) is based on
observed linear regression relationships between surface temperature or precipitation and a range of atmospheric
predictor variables. The three methods are equally plausible a priori, in the sense that they estimate present-day
natural variations with equal skill.
For temperature, differences between the RCM and GCM predictions are quite small. Larger differences occur
between STAT and the dynamical predictions. For precipitation, there is a wide spread between all three methods.
Differences between the RCM and GCM are increased by the meso-scale detail present in the RCM.
Uncertainties in the downscaling predictions are investigated by using the STAT method to estimate the grid point
changes simulated by the GCM, based on regression relationships trained using simulated rather than observed values
of the predictor and the predictand variables (i.e. STAT–SIM). In most areas the temperature changes predicted by
STAT–SIM and the GCM itself are similar, indicating that the statistical relationships trained from present climate
anomalies remain valid in the perturbed climate. However, STAT–SIM underestimates the surface warming in areas
where advective predictors are important predictors of natural variability but not of climate change. For precipitation,
STAT–SIM estimates the simulated changes with lower skill, especially in January when increases in simulated
precipitation related to a moister atmosphere are not captured. This occurs because moisture is rarely a strong enough
predictor of natural variability to be included in the specification equation.
The predictor:predictand relationships found in the GCM do not always match those found in observations. In
January, the link between surface and lower tropospheric temperature is too strong. This is also true in July, when
the links between precipitation and various atmospheric predictors are also too strong. These biases represent a likely
source of error in both dynamical and statistical downscaling predictions. For example, simulated reductions in
precipitation over southern Europe in summer may be too large. © British Crown Copyright 2000.
KEY WORDS: Europe; climate change; downscaling; climate models; surface temperature; precipitation
1. INTRODUCTION
The problem of ‘downscaling’ output from global general circulation models (GCMs, current grid size
typically 300 km) to obtain information for localized areas has received increasing attention in recent
years (see Wilby and Wigley, 1997), motivated by the requirement of policy-makers for detailed regional
scenarios of climate change (e.g. Department of the Environment, 1996). Many downscaling techniques
are based on statistical relationships linking observations of local variables to the observed atmospheric
circulation. These relationships are then applied to the circulation simulated by a GCM in order to
generate predictions of local climate (e.g. Karl et al., 1990; von Storch et al., 1993). The use of such
* Correspondence to: Hadley Centre for Climate Prediction and Research, Meteorological Office, London Road, Bracknell, Berks
RG12 2SY, UK.
© British Crown Copyright 2000

J. MURPHY490
methods is motivated by an assumption that GCMs simulate the large-scale atmospheric circulation better
than they simulate surface climate elements such as surface temperature and precipitation (e.g. Palutikof
et al., 1997), because the latter are particularly sensitive to subgrid-scale processes (convection, cloud
formation, turbulent transports, etc) which can only be represented approximately in the models.
The alternative is dynamical downscaling, in which predictions of (say) site-specific temperature or
precipitation are derived from values of surface temperature or precipitation simulated at nearby GCM
grid points. Here the assumption is that predictions based explicitly on relevant physical processes will be
better than those in which the underlying physics is only represented implicitly, via empirical relationships.
In practice, even dynamical downscaling will require an empirical element because the grid box predictor
variables are effectively spatial means for an area of 300300 km2 (e.g. Osborn and Hulme, 1997),
whereas the predictands are local values influenced by spatial heterogeneities in the regional physiography
(orography, coastlines, soil and vegetation types, etc). This mismatch in scales can be addressed by nesting
a high resolution regional climate model (RCM) inside the GCM (e.g. Giorgi et al., 1993; McGregor and
Walsh, 1993; Jones et al., 1995). Unfortunately, the resolution of RCMs (typically 50 km) is not yet
high enough to capture all aspects of the regional forcing. In addition, most research groups only have
a few years experience in developing and using RCMs, which is not long enough to optimize the
performance of such complex tools.
Despite their potential limitations, the use of downscaling schemes is essential if plausible regional
scenarios are to be constructed for impact assessments (Watson et al., 1998). It is, therefore, important to
compare statistical and dynamical downscaling methods in order to understand the reasons for the
differences in their predictions and, if possible, to identify the optimum technique.
Kidson and Thompson (1998) and Murphy (1999) have recently performed such comparisons by
assessing downscaling estimates of variability observed in the recent historical record. In both cases,
relevant model integrations were forced with time series of analyses of the observed atmospheric
circulation in order to remove the influence of model circulation biases. Kidson and Thompson (1998)
compared statistical and RCM estimates of daily and monthly temperature and precipitation observed at
stations in New Zealand, and found that, on average, the methods gave similar levels of skill. Murphy
(1999) compared three methods of downscaling values of surface temperature and precipitation observed
at 976 European observing stations. The predictor variables were:
(i) temperature or precipitation simulated at the nearest GCM grid point to the target location;
(ii) temperature or precipitation simulated at the nearest grid point of an RCM nested inside the GCM;
and
(iii) values of various atmospheric variables, including regional temperature, moisture, properties of the
near-surface wind and vertical stability, and large-scale patterns of variability in mean sea level
pressure. Subsets of these were converted into predictions of local temperature or precipitation, using
linear regression.
The dynamical methods (i) and (ii) were compared with the statistical method (iii) in terms of the ability
to predict interannual variations in monthly means. All three methods gave roughly equal levels of skill,
supporting Kidson and Thompson’s (1998) result. Average explained variances were higher in winter than
in summer and were higher for temperature than for precipitation. The dynamical methods were also
compared by evaluating errors in the predictions of daily distributions and the climatological mean and
variance of monthly averages. The RCM yielded improved skill relative to the GCM in estimates of both
climatological mean temperature and the daily and interannual variability of precipitation.
In this paper, Murphy’s (1999) study is extended to consider downscaling predictions of climate change.
For simplicity, only changes in multi-annual monthly means are considered. The intention is to determine
the spread of changes that can be generated from a single GCM simulation, using plausible downscaling
techniques. In order to explain differences between the statistical and the dynamical predictions a further
experiment is carried out in which the statistical method is used to predict the changes simulated by the
GCM. This provides an opportunity to assess the extent to which simple statistical regression equations,
calibrated from natural variability, can reproduce climate changes driven by changes in radiative forcing
© British Crown Copyright 2000 Int. J. Climatol. 20: 489–501 (2000)