20-year return period for the rst GEV corresponds to a different return period in the
second GEV. Plotting these new estimated return periods gives the relevant distribution
and these are displayed for the three grid boxes in Fig. 3 (in all cases, corresponding to
the AM series calculated for 30-day durations). Denoting the curves in Fig. 3 as f .¿ /,
the fractional area under f .¿ / between two return periods ¿1 and ¿2 is the probability of
a new return period (derived from two independentsamples of annual maxima from the
same underlying GEV distribution) being between these values.
In Fig. 3, the thin vertical lines plotted represent ¿
control
2:5
, which satis es p.¿ >
¿
control
2:5
/ D 97:5%. This number is at the lower end of the 95% region within which
¿ is expected to be, corresponding to the probability-density function f .¿ /. That is,
when using the bootstrap simulation to generate different possible pairs of GEV curves
from the RCM control simulation of AM, setting ¿ D 20 years for the rst simulation
and calculating the equivalent return period from the second simulation, on 97.5% of
occasionsthis value will be greater than ¿
control
2:5
. This then allows signi cance tests to be
made of differences between the 20-year return period of the RCM control simulation
and that of observations (called ¿
obs) and between the RCM control simulation and
RCM future simulation (called ¿
future). The latter comparison requires an adjustment to
1614 C. HUNTINGFORD et al.
TABLE 2. COMPARISON OF OBSERVED RETURN
PERIODS (¿
obs) AND FUTURE RETURN PERIODS
(¿
future ) OF 30-DAY PRECIPITATION-EVENT ANNUAL MAXIMA WITH THOSE FROM THE REGIONAL CLIMATE
MODEL CONTROL SIMULATION (¿
control 2:5
) (SEE TEXT
FOR DETAILS)
Location ¿ ¿
control 2:5 ¿
obs ¿
future
Lewes 5 2.40 11.4 1.52
Shrewsbury 5 2.42 6.71 1.52
York 5 2.36 6.22 1.69
Lewes 10 3.46 21.2 1.86
Shrewsbury 10 3.48 15.1 1.99
York 10 3.34 9.05 2.33
Lewes 20 4.55 35.4 2.26
Shrewsbury 20 4.63 36.4 2.77
York 20 4.24 12.1 2.33
Lewes 40 5.54 54.2 2.73
Shrewsbury 40 5.73 95.7 4.07
York 40 4.94 15.1 2.61
the methodology above. As the future RCM simulation has just 20 values, when creating
the ‘pairings’ the rst GEVs are based upon a sample of 30 values but the second set of
curves utilize a sample of 20 values. In reality, it is found that the resultant distribution
is very slightly broader and, as such, the differences are ignored.
To assess signi cance, the change in the 20-year return period between the RCM
control simulation and observations (dashed lines, giving ¿
obs) and the RCM control
simulation and the RCM future simulation (dash-dot lines, giving ¿
future) are presented
in Fig. 3. These values are also presented in Table 2, along with identical calculations
for other return periods apart from 20 years. In all cases, it is found that the future
RCM-predicted return period is signi cantly less than would be expected from the RCM
control distribution, but that there is no signi cant difference between the return period
estimated from the RCM control and the observed GEVs (Table 2). It is also found
(although not presented here) that in all cases, ¿
obs < ¿
control
97:5 and ¿
future < ¿
control
97:5
.