Using GLMs to simulate daily rainfall under scenarios of climate change

Nadja Leith December 8, 2005

Contents
1 Introduction 2 Generalised Linear Models for daily rainfall 3 Data 3.1 3.2 Rainfall data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Atmospheric data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 3.2.2 3.3 Climate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reanalysis data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 3 3 3 4 6 6 11 16 16 17 27 28 31

Atmospheric predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4 Model building 5 Simulation 5.1 5.2 1961-1990 period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2071-2100 period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6 Conclusion 7 Appendix A 8 Appendix B

Introduction

There is a pressing need to understand the possible eects of climate change. One area which requires attention is the eect of a changing global climate on local scale rainfall. For example, in order to run hydrological rainfall-runo models for future ood risk assessment we need to be able to simulate realistic sequences of daily and sub-daily rainfall that are consistent with projected changes in climate. This document outlines work undertaken for DEFRA project FD2113, Work Package 1.2. The aim of this work is to establish a methodology for simulating realistic sequences of future daily rainfall conditional on atmospheric information from deterministic climate models. We employ Generalised Linear Models (GLMs), introduced in Section 2, to relate daily rainfall to atmospheric predictors. The GLMs are tted to observed daily rainfall and the corresponding atmospheric reanalysis data, with this relationship then being applied to Climate Model output. The data and climate model output are discussed in Section 3, model tting is addressed in Section 4 and simulation results are presented in Section 5. Section 6 presents the conclusions of this work.

Generalised Linear Models for daily rainfall

The work presented here makes use of the GLIMCLIM software (Chandler, 2002) in which daily rainfall, denoted Yt , is modelled in two stages, both of which use a GLM. Firstly, as in Chandler and Wheater (2002), rainfall occurrence is modelled using logistic regression. Let pi denote the probability of rain at a single site on day i, and let xi denote a corresponding predictor vector for that site which includes atmospheric measures. Then the occurrence model is ln(pi /(1 pi )) = xT . i Wet day rainfall amounts are modelled using a gamma distribution with a common dispersion parameter. Conditional on predictor vector zi , the mean rainfall for the site on the ith wet day is mi with ln(mi ) = zT . i Such models can be tted simultaneously to data from a number of sites. In order to simulate realistic multi-site rainfall sequences GLIMCLIM allows for spatial dependence, resulting from large scale weather systems, by constructing a joint distribution of rainfall at all sites which respects the marginal distributions from the GLMs (Yang et al., 2006b). However, in the work reported here dierent sites are modelled separately, as this work package is primarily focused on single site rainfall. Modelling the sites separately allows dierent sites to have dierent dispersion parameters, ensuring that even if dierent sites have the same mean rainfall on a particular day, they may have dierent values for rainfall variance. Predictors considered for both the occurrence and amounts models include periodic functions along with specic month eects to represent seasonality, and transformations of rainfall values on previous days to account for persistence. Atmospheric predictors must also be included if the models are to simulate realistic daily rainfall sequences under future climate 2

scenarios (the types of atmospheric predictors used will be discussed in Section 3). Finally, interactions between the predictors are considered. This allows for some covariates to modulate the eect of others. For example, it will be seen in Section 4 that in the model for rainfall amounts the eect of the rainfall on the previous two days varies with season. This is appropriate because convective and frontal rainfall, which dominate in summer and winter respectively, result in dierent degrees of autocorrelation. In the work reported below, the choice of covariates was based on the Likelihood Ratio Statistic (see Dobson (1990) for further details). Standard residual analyses were also used to check the models and suggest improvements. These resulted, for example, in the inclusion of a February eect beyond the sine and cosine waves in the occurrence model of Section 4. Once a nal model has been selected its properties can be obtained via simulation. This is required because the nonlinear structure of the models does not allow the properties to be obtained analytically.

3
3.1

Data
Rainfall data

Models have been tted separately to daily rainfall data from three gauges (at Heathrow, Birmingham and Manchester airports). Data from the 1961-1990 period were used to t the models (this corresponds to the control period of the climate models). Data were missing at Heathrow for January to August of 1988 and February 1989. Birmingham airport had missing observations for December 1983 and August 1989. The records for Manchester airport were complete. It should be noted that during the mid-1970s there was a conversion from imperial to metric measurement units (all rainfall values in this work are in millimetres), resulting in a change of measurement resolution. To minimise the eect of this, it is appropriate to threshold the rainfall data in such a way that small rainfall amounts are set to zero throughout the period of record (Yang et al., 2006a). Here, a threshold of 0.5mm has been used. This threshold is subtracted from all rainfall values with Yt > 0.5 before tting, so that the GLMs are t to rainfall data Yt = max(0, Yt 0.5). Data simulated using the GLMs then have the threshold added back to any non-zero values. The simulated rainfall will therefore contain no values between zero and 0.5mm.

3.2

Atmospheric data

The aim of this work is to simulate future rainfall conditional on projections of future climate. The GLMs are t using observed atmospheric data and are then simulated conditional on climate model output for the 2071-2100 period. We therefore need to identify atmospheric variables that are closely related to rainfall, and are also reproduced reasonably by climate models. 3

Model type Institution GCM Canadian Centre for Climate Modelling and Analysis GCM Commonwealth Scientic and Industrial Research Organisation GCM Max-Planck-Institute GCM Hadley Centre RCM Danish Meteorological Institute RCM Swedish Meteorological and Hydrological Institute RCM Hadley Centre Table 1: Details of employed climate models.

Model name cgcm2 csiromk2 echam4 hadcm3 hirham rcao hadrm3p

3.2.1

Climate Models

Deterministic climate models attempt to represent explicitly the physical processes underlying climate. Both General Circulation Models (GCMs) and Regional Climate Models (RCMs) are used in this work. The specic models used are summarised in Table 1. RCM data were obtained from the PRUDENCE (Prediction of Regional scenarios and Uncertainties for Dening EuropeaN Climate change risks and Eects) project ([Link] This was an investigation which produced high-resolution (grid distances of between 0.2 and 0.5 ) climate simulations for the 1961-1990 (control) and 2071-2100 periods, using RCMs from a number of climatological institutions across Europe. At boundaries all models were forced with an atmosphere only global model which was in turn forced by sea surface temperatures (SSTs). For the future 2071-2100 period, SSTs were simulated using a General Circulation Model (GCM) without ux adjustment (HadCM3). All results presented here are based on the A2 transient green-house gas scenario as specied by the Intergovernmental Panel on Climate Change (IPCC), in which total annual Carbon Dioxide emissions in 2100 are approximately 3.5 times greater than in 2000 (IPCC, 2001, Appendix II SRES Tables). For reasons of convenience, RCM years have 360 days only (30 days per month). RCMs developed by three institutions were considered (under the advice of David Stephenson): the Danish Meteorological Institute (DMI), the UK Hadley centre (HC) and the Swedish Meteorological and Hydrological Institute (SMHI). The DMI and SMHI RCMs were forced using a dierent atmosphere only model to the HC RCM. As previously mentioned, all three RCMs considered here are forced using SSTs from the same GCM. Yet due to the extremely complex nature of the climate system and the resulting uncertainties in model formulation and parameterisation, dierent GCMs can give very dierent projections for future climate. Rowell (2004) observes that the uncertainty in projected surface air temperature over the UK due to RCM formulation is relatively small in comparison to the uncertainty resulting from the formulation of the driving GCM. Therefore, in order that this source of uncertainty be addressed, we consider output from GCMs developed by four dierent institutions: the Canadian Centre for Climate Modelling and Analysis (CCCma; cgcm2), Commonwealth Scientic and Industrial Research Organisation (CSIRO; csiromk2), Max-Planck-Institute (MPI; echam4) and the Hadley Centre (HC; hadcm3). The data were obtained from the Statistical DownScaling Model (Wilby and 4

Dawson, 2001) (SDSM) website ([Link] cocwd/SDSM/[Link]) on a grid of 2.5 latitude by 3.75 longitude. The CCCma and CSIRO models both have 365 days per year and run from 1961 to 2100, the MPI and Hadley Centre models have 360 days per year and run from 1961 to 2100 and 1961 to 2099 respectively. Again, all output used is conditional on the A2 transient green-house gas scenario, so emission scenario uncertainty is not addressed. It is noted that the relative humidity values from CSIRO, MPI and particularly the CCCma model should be treated with caution having been derived algorithmically from pressure, temperature and specic humidity (Harris, 2004). Harris (2004) observes that calculation of relative humidity from other CCCma outputs did not result in realistic values, suggesting it may be inappropriate to attempt to derive relative humidity from existing variables. One possible explanation oered for this is the inappropriateness of the calculated saturation vapour pressure. Given that the aim of this work is to simulate daily rainfall it would be legitimate to suggest using the climate model rainfall output directly. We do not take this approach for a number of reasons. Firstly, climate model rainfall output is a spatial average but single site values are often required. Secondly, impact assessment studies ideally require a number of possible rainfall sequences and this cannot be easily achieved using climate models due to computational costs. Finally, climate models cannot explicitly resolve the processes controlling rainfall. Instead, they build empirical relationships between rainfall and other atmospheric variables, a technique called parameterisation. It has been found that dierent climate models give very dierent projections for rainfall, even at large spatial and temporal scales, giving a lower level of condence in the rainfall realisations. Stefan Hagemann and Daniela Jacob (personal communication, 2004) found that the dierent RCMs in the PRUDENCE project gave both negative and positive proportional changes in rainfall between the control and future periods over the Elbe and Rhine catchments. Frei et al. (2003) report that the PRUDENCE RCMs underestimated the average intensity over the Alps by 16-42%. Nevertheless, it is sensible to compare future rainfall properties obtained via the GLM rainfall simulations with those based on climate model rainfall outputs directly. If the two dierent approaches give similar results, this may improve condence in our projections of future rainfall. In Section 5.2 we will compare the values of mean daily rainfall obtained via simulation of the GLMs with the values based on the climate model rainfall outputs. We have discussed the limitations of climate model rainfall output. There are also limitations of climate model output for other variables, although in general these are not so serious as for rainfall. Due to the chaotic nature of the climate system climate models cannot generate actual climate sequences, but aim to simulate realisations with climate like properties (von Storch and Zwiers, 1999; Smith, 2002). In other words, the best any climate model can hope to do is to produce a sequence which is, in terms of its properties, indistinguishable from the true trajectory (at least within the period of interest). In reality there are two further issues to consider when using climate model output, and while previous discussion in the literature has focused on GCMs the same arguments can be applied to RCMs. Firstly, it is likely that properties ([Link] and variance) of model outputs may match those of observations only when aggregated over large temporal scales. The IPCC (2001,

p. 473) suggest that GCMs provide credible simulations of climate, at least down to subcontinental scales and over temporal scales from seasonal to decadal. Prudhomme et al. (2002) note that despite the belief that GCMs model atmospheric variables (such as the sealevel pressure) better at a daily time step than they model rainfall, there is still no consensus about the reliability at that time scale. Secondly, the IPCC (2001, p. 750) note that studies rarely use climate model outputs directly because biases are too great. Therefore, the output from any climate model run to investigate the eect of anthropogenic forcings is always compared to the output from a control climate model run (after both have had spin-up time). Changes in mean climate between the baseline period and the future are often calculated either as the dierence or as the ratio between the simulated forced climate and the simulated control climate (IPCC, 2001, p. 751). The IPCC note that ratios are often used with those surface variables that are either positive or zero (e.g. precipitation and pressure), while for other variables (e.g. temperature) the dierence is usually used. These relative or absolute changes (often stratied by season) are then used in conjuction with the observed data for the baseline period to create a future climate scenario. Both these forms of manipulation are consistent with the assumption that a simulated climatological mean is linearly biased. Furthermore, it seems reasonable to extend this idea and to assume that the observed or projected value of some atmospheric variable, Xto (where t indexes at an appropriate temporal scale), has the same distribution as + Xts , where Xts is the corresponding climate model output for time t. 3.2.2 Reanalysis data

The NCEP (Kalnay et al., 1996) and ERA40 (at [Link] mnth/) reanalysis data sets are derived by feeding quality controlled observations into a physical model, which then outputs gridded values of many dierent variables. While a number of atmospheric variables (e.g. temperature) are closely related to input observations, other variables (e.g. precipitation) are not constrained by actual observations for that variable. This means that the reanalysis data sets are more reliable for some variables than for others.1 The reanalysis data play the role of observed atmospheric data in the work presented here, and are used to t the GLMs relating observed rainfall to observed atmospheric variables.

3.3

Atmospheric predictors

In this section we discuss how the reanalysis data and climate model output were manipulated in order to provide atmospheric predictors for the GLMs. The GLMs are tted using both the NCEP and ERA40 data in order to assess whether the model parameter estimates are robust to dierences in the reanalysis data. As the tted models will then be simulated conditional on the GCM/RCM output, atmospheric predictors were limited to those that are reasonably represented by the climate models. Therefore temperature, sea level
NCEP rainfall data were not included in the set of potential atmospheric predictors due to quality issues (Chandler, 2000).
1

CCCma

CSIRO

0.6

0.4

Sealevel pressure

Sealevel pressure

0.2

0.0

0.2

2070

2075

2080

2085 Year

2090

2095

2100

0.2

0.0

0.2

0.4

SB W

EE SE

0.6

SB W

EE SE

2070

2075

2080

2085 Year

2090

2095

2100

MPI

Hadley

0.6

0.4

Sealevel pressure

Sealevel pressure

0.2

0.0

0.2

2070

2075

2080

2085 Year

2090

2095

2100

0.2

0.0

0.2

0.4

SB W

EE SE

0.6

SB W

EE SE

2070

2075

2080

2085 Year

2090

2095

2100

Figure 1: Annual mean sea-level pressure standardised with respect to the 1961-1990 control period, for grid squares with centre points 50 N, 0 (Southern England, SE), 52.5 N, 0 (Eastern England, EE), 52.5 N, 3.75W (Wales, W), 55 N, 3.75 W (Scottish Borders, SB). pressure, relative humidity were considered (specic humidity and temperature being too highly correlated in the reanalysis data to both be used). All reanalysis data and climate model output for these variables were regridded onto the SDSM NCEP grid, measuring 2.5 latitude by 3.75 longitude. As measures of humidity were only available for the HC RCM, this realisation was used in rainfall simulation from all RCMs. However, humidity output for DMI and SMHI may be available in the future. In Section 2 it was discussed that dierent models will be t for dierent sites. This allows the use of atmospheric predictors corresponding to an area local to the rainfall site. This is particularly important as the projected changes in atmospheric values are dierent for dierent grid squares, as shown in Figures 1 to 3. For example, Figure 2 shows that the CSIRO model projects greater increases in temperature for the more southerly grid squares. Atmospheric predictors corresponding to a particular site were calculated using a weighted average of nearby grid squares to allow for the non-continuous nature of the data. The weighting was based on the assumption that the atmospheric values for a particular grid square corresponded to its centre point. Let (xi , yi) be the coordinates for the centre of grid square i, let (xs , ys ) be the coordinates of the site, and let d(xs , xi ) and d(ys , yi ) be the east-west and north-south distances (in kilometres) respectively. Then the weight for grid square i is proportional to w(xs , xi ) + w(ys , yi) where w(xs , xi ) = 275 d(xi , xs ) if d(xi , xs ) 275 0 if d(xi , xs ) > 275

CCCma

CSIRO

0.2 0.4 0.6 0.8 1.0 1.2 1.4

SB W

EE SE

0.2 0.4 0.6 0.8 1.0 1.2 1.4

SB W

EE SE

Temperature

2070

2075

2080

2085 Year

2090

2095

2100

Temperature

2070

2075

2080

2085 Year

2090

2095

2100

MPI

Hadley

0.2 0.4 0.6 0.8 1.0 1.2 1.4

SB W

EE SE

0.2 0.4 0.6 0.8 1.0 1.2 1.4

SB W

EE SE

Temperature

2070

2075

2080

2085 Year

2090

2095

2100

Temperature

2070

2075

2080

2085 Year

2090

2095

2100

Figure 2: Annual mean temperature standardised with respect to the 1961-1990 control period, for grid squares with centre points 50 N, 0 (Southern England, SE), 52.5 N, 0 (Eastern England, EE), 52.5 N, 3.75W (Wales, W), 55 N, 3.75 W (Scottish Borders, SB).

CCCma

CSIRO

SB W

EE SE

SB W

EE SE

0.5

Relative humidity

Relative humidity
2070 2075 2080 2085 Year 2090 2095 2100

0.0

0.5

1.0

1.0
2070

0.5

0.0

0.5

2075

2080

2085 Year

2090

2095

2100

MPI

Hadley

SB W

EE SE

SB W

EE SE

0.5

Relative humidity

Relative humidity
2070 2075 2080 2085 Year 2090 2095 2100

0.0

0.5

1.0

1.0
2070

0.5

0.0

0.5

2075

2080

2085 Year

2090

2095

2100

Figure 3: Annual mean relative humidity standardised with respect to the 1961-1990 control period, for grid squares with centre points 50 N, 0 (Southern England, SE), 52.5 N, 0 (Eastern England, EE), 52.5 N, 3.75W (Wales, W), 55 N, 3.75 W (Scottish Borders, SB).

CCCma CSIRO MPI HC NCEP

Yearly average sealevel pressure

0.4
1960

0.2

0.0

0.2

0.4

1980

2000

2020 Year

2040

2060

2080

2100

Figure 4: Yearly average sea-level pressure at Heathrow, standardised with respect to the 1961-1990 mean and standard deviation. and equivalently for w(ys , yi). The cut-o of 275 was chosen as this is approximately the maximum distance between the centre points of two grid squares. Hence, if a site is located exactly at the centre of a grid square, the values of atmospheric variables are obtained solely from that square. To illustrate the resulting atmospheric sequences, Figures 4, 5 and 6 show yearly averages from the dierent GCMs for sea-level pressure, temperature and relative humidity respectively at Heathrow. Given that climate model years do not have the same number of days as calendar years it would be inappropriate to use daily atmospheric variables in our GLMs. Therefore, monthly mean values of the spatially averaged data were obtained for the 1961-1990 period. Such a degree of temporal smoothing will still allow for simulation of rainfall typical of weather systems that occur on smaller temporal scales, as the structure of such rainfall events will be incorporated into the variance and autocorrelation structure of the GLMs. The monthly atmospheric values were then standardised by month (i.e. with respect to the monthly means and standard deviations for the 1961-1990 period). This standardisation was done for two reasons. Firstly, it was desired to deseasonalize the predictors as seasonality will be included explicitly elsewhere in the GLMs. Secondly, it accounts for the climate o model bias. To explain this, let Xym be the NCEP or ERA40 monthly mean value of o a particular atmospheric variable in year y and month m; the values of Xym have been s observed for the control period but, obviously, not for the future. Similarly let Xym be the GCM/RCM simulated monthly mean value of the atmospheric variable in year y and o o month m; here, values are available for both control and future periods. Let Xm and Sm o be the month dependent sample mean and standard deviation for the Xym in the control s s period, and similarly for Xm and Sm . Thus the GLMs are tted using atmospheric predictors 9

1.0

CCCma CSIRO MPI HC NCEP

Yearly average temperature

0.4
1960

0.2

0.0

0.2

0.4

0.6

0.8

1980

2000

2020 Year

2040

2060

2080

2100

Figure 5: Yearly average temperature at Heathrow, standardised with respect to the 19611990 mean and standard deviation.

Yearly average relative humidity

0.5

0.0

1.0

CCCma CSIRO MPI HC NCEP

1960

1980

2000

2020 Year

2040

2060

2080

2100

Figure 6: Yearly average relative humidity at Heathrow, standardised with respect to the 1961-1990 mean and standard deviation.

10

o o o (Xym Xm )/Sm . Now let us assume, as discussed in Section 3.2.1, that the GCM/RCM s o outputs are biased so that + Xym follows the same distribution as Xym . It then follows that o o s s o s (Xym Xym )/Sym ( + Xym Xym )/Sym .

where denotes equality of distribution. Clearly,

and so
o o s s o s (Xym Xym )/Sym (Xym Xym )/Sym .

Therefore, to simulate rainfall using based on the GCM/RCM output the tted GLMs are s s s s conditioned on Zym = {(Xym Xym )/Sym }.

Model building

In this section we describe the construction of GLMs for rainfall at the three selected sites (see Section 3.1). This work aims to provide a methodology for using GLMs to simulate daily rainfall, and the predictors included in the models presented here should provide a guide to suitable predictors for other sites. Therefore in this work, if a predictor is found to be signicant at any site, it is included in the GLMs for all sites. The predictors for Heathrow, and the corresponding parameter estimates and standard errors obtained when tting using the NCEP and ERA40 reanalysis data, are listed in Tables 2 to 5. The models for other sites can be found in the appendix. We let Yt denote the rainfall value on day t, I(Y [t k] > 0) be an indicator of rainfall occurrence k days previously at the site and I(Y [t k] > 0 : k = 1 to n) be an indicator of rain on each of the previous k days. Both the NCEP and ERA40 tted models are fairly similar. The interactions between I(Y [t k] > 0 : k = 1 to 2) and the sine and cosine components in both the tted amounts models imply a seasonally varying autocorrelation structure: the dependence on the previous two days rainfall occurrence is greater in winter than in summer. This agrees with the results of Chandler and Wheater (2002) and is expected because there are fewer long-lasting frontal weather systems in summer in mid-latitudes. Interestingly, the opposite eect is observed in the occurrence model: here, the dependence on the previous days rainfall is greater in summer. Such a surprising result is not observed in either the Manchester or Birmingham occurrence models. The Heathrow models show that autocorrelation for both occurrence and amounts varies with sea level pressure, and the parameter estimates suggest that dependence between rainfall on successive days is greater in the more stable weather systems associated with higher pressure. Although the amounts model explains only 3.2 percent of the variance in daily rainfall, it will be illustrated in Section 5.1 that this apparently small degree of predictive power captures the seasonal and inter-annual variation in rainfall properties well. The tted GLMs for the three sites are examples of models that could also be tted to data from other sites. We note that when tting and then extrapolating a rainfall amounts 11

Main Eects Parameter Estimate Constant -1.28 Sea-level pressure -0.41 Temperature 0.02 Relative humidity 0.10 I(Y [t 1] > 0) 0.93 I(Y [t 2] > 0) 0.49 I(Y [t 3] > 0) 0.17 I(Y [t k] > 0 : k = 1 to 2) -0.55 ln(1 + Yt1 ) 0.17 Daily seasonal eect, cosine component 0.22 Daily seasonal eect, sine component 0.05 February eect -0.55 Interactions Daily seasonal eect, cosine component, -0.12 with I(Y [t 1] > 0) Daily seasonal eect, sine component, with 0.10 I(Y [t 1] > 0) February eect with I(Y [t 1] > 0) 0.58 Daily seasonal eect, cosine component with 0.07 temperature Daily seasonal eect, sine component with 0.004 temperature February eect with temperature 0.21 Daily seasonal eect, cosine component with -0.05 relative humidity Daily seasonal eect, sine component with 0.05 relative humidity February eect with relative humidity -0.16 Sea-level pressure with I(Y [t 1] > 0) 0.20

Standard Error 0.04 0.03 0.03 0.03 0.08 0.06 0.05 0.09 0.04 0.04 0.04 0.19 0.07 0.07 0.29 0.04 0.03 0.17 0.03 0.03 0.17 0.05

Table 2: Logistic occurrence model for Heathrow, tted using the NCEP reanalysis data.

12

Eect Parameter Estimate Constant 1.33 Sea-level pressure -0.08 Temperature -0.002 ln(1 + Y [t 1]) 0.08 ln(1 + Y [t 2]) 0.06 I(Y [t k] > 0 : k = 1 to 2) -0.15 I(Y [t 1] > 0) 0.15 Daily seasonal eect, cosine component -0.22 Daily seasonal eect, sine component -0.14 Interactions I(Y [t k] > 0 : k = 1 to 2) with Daily 0.18 seasonal eect, cosine component I(Y [t k] > 0 : k = 1 to 2) with Daily -0.11 seasonal eect, sine component Daily seasonal eect, cosine component with -0.12 sea-level pressure Daily seasonal eect, sine component with 0.02 sea-level pressure I(Y [t k] > 0 : k = 1 to 2) with sea-level 0.12 pressure

Standard Error 0.03 0.03 0.02 0.03 0.03 0.07 0.07 0.03 0.03 0.07 0.06 0.03 0.03 0.05

Table 3: Gamma amounts model for Heathrow, tted using the NCEP reanalysis data. This model accounts for 3.2% of the variance. The dispersion parameter for this model was estimated to be 0.73.

13

Eect Parameter Estimate Constant -1.28 Sea-level pressure -0.44 Temperature 0.001 Relative humidity 0.06 I(Y [t 1] > 0) 0.94 I(Y [t 2] > 0) 0.49 I(Y [t 3] > 0) 0.18 I(Y [t k] > 0 : k = 1 to 2) -0.55 ln(1 + Yt1 ) 0.16 Daily seasonal eect, cosine component 0.23 Daily seasonal eect, sine component 0.05 February eect -0.56 Interactions Daily seasonal eect, cosine component, -0.13 with I(Y [t 1] > 0) Daily seasonal eect, sine component, with 0.11 I(Y [t 1] > 0) February eect with I(Y [t 1] > 0) 0.56 Daily seasonal eect, cosine component with 0.09 temperature Daily seasonal eect, sine component with 0.02 temperature February eect with temperature 0.19 Daily seasonal eect, cosine component with -0.02 relative humidity Daily seasonal eect, sine component with 0.02 relative humidity February eect with relative humidity -0.01 Sea-level pressure with I(Y [t 1] > 0) 0.20

Standard Error 0.04 0.03 0.03 0.03 0.08 0.06 0.05 0.09 0.04 0.04 0.04 0.19 0.07 0.07 0.29 0.04 0.03 0.16 0.04 0.03 0.16 0.05

Table 4: Logistic occurrence model for Heathrow, tted using the ERA40 reanalysis data.

14

Eect Parameter Estimate Constant 1.33 Sea-level pressure -0.08 Temperature -0.02 ln(1 + Y [t 1]) 0.08 ln(1 + Y [t 2]) 0.06 I(Y [t k] > 0 : k = 1 to 2) -0.15 I(Y [t 1] > 0) 0.15 Daily seasonal eect, cosine component -0.22 Daily seasonal eect, sine component -0.14 Interactions I(Y [t k] > 0 : k = 1 to 2) with Daily 0.18 seasonal eect, cosine component I(Y [t k] > 0 : k = 1 to 2) with Daily -0.11 seasonal eect, sine component Daily seasonal eect, cosine component with -0.13 sea-level pressure Daily seasonal eect, sine component with 0.02 sea-level pressure I(Y [t k] > 0 : k = 1 to 2) with sea-level 0.12 pressure

Standard Error 0.03 0.03 0.02 0.03 0.03 0.07 0.07 0.03 0.03 0.07 0.06 0.03 0.03 0.05

Table 5: Gamma amounts model for Heathrow, tted using the ERA40 reanalysis data. This model accounts for 3.2% of the variance. The dispersion parameter for this model was estimated to be 0.73.

15

model, constraints are required on the coecients corresponding to previous days rainfall if the simulated rainfall is to remain within reasonable bounds. For example, a three way interaction between temperature, ln(1 + Y [t 1]) and seasonality was also found to be signicant when included in the amounts model presented here. However, future simulations from a model including this interaction escalated, at times, to unrealistic values of more than a meter of rainfall in a day. This was a result of the projected increase in temperature, as this in turn increased the eect of Yt1 on t . Given this problem the interaction was not included in the nal model. This highlights the need to consider carefully the structure of models that are to be used for extrapolation purposes. Given that extrapolation of tted models is fundamental to this project, we suggest models should be robust and capture the main features of rainfall at the possible expense of detail.

5
5.1

Simulation
1961-1990 period

To investigate the performance of the tted models, rainfall is simulated for the 1961-1990 period and the properties of these sequences compared with those of the observations. Such model checking is required since small errors in model specication may cumulate over a long period of time (Yang et al., 2006b). Furthermore, simulation allows us to assess the performance of the occurrence and amounts models when used together. Figures 7 through to 15 are checks on the t of the NCEP tted models presented in Section 4. Given the similarity between the NCEP and ERA40 tted models, the latter will not be considered further here. 200 simulations of the GLMs were run for the control period, conditional on the NCEP data. Figures 7 to 9 investigate how well the GLMs simulate total annual rainfall at each site. The rst plot in each Figure is a Probability Integral Transform (PIT) (Dawid, 1984) which can be used to assess the calibration of modelled distributions. The PITs show the number of years the observed total annual rainfall fell in between certain percentiles of the corresponding simulated distributions. If the observed values had actually come from the simulated distributions the results should look like samples from a uniform distribution. The second plots in Figures 7 to 9 show the simulated distributions of total rainfall in each year, along with the corresponding observed values. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the black line shows the observed value (thicker lines indicate missing data that have been simulated ten times conditional on the observed data as described in Yang et al. (2006b)). If the simulations are realistic the observed values should look as if they have been sampled from the distributions. The notable inter-annual variability in the simulated distributions (e.g. the shift upwards in 1974) is due to the NCEP explanatory variables. Based on these plots the models appear to perform well; the PITs are convincing as 30 independent realizations from the uniform distribution. Figures 10 to 12 show the simulated distributions of total summer and winter rainfall at the dierent sites and the corresponding observed values (black line). Again, if the 16

Probability integral

0 1 2 3 4 5 6 7

Frequency

0.0

0.2

0.4 Percentiles

0.6

0.8

1.0

Total annual rainfall

mm

0
1960

400

800

1200

1965

1970

1975 Year

1980

1985

1990

Figure 7: i) Probability Integral Transform for annual rainfall at Manchester: observed values as a sample from simulated distributions. ii) Simulated distributions of total annual rainfall in Manchester. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values. simulations are realistic the observed values should look as if they have been sampled from the seasonal distributions. All three Figures show that the simulated distributions of seasonal totals are consistent with the observed values. Finally, Figures 13 to 15 show the simulated distributions of various monthly summary statistics and the corresponding observed values (black line). There is generally a good match to the summary statistics, particularly to the mean and the proportion of wet days. The fact that the simulated standard deviation, standard deviation when wet and maximum are generally too small at all sites in summer shows that the use of a constant dispersion parameter estimate throughout the year is not ideal.

5.2

2071-2100 period

The ultimate aim of this work is to simulate a range of future rainfall sequences that are consistent with the future climates projected by the climate models. Therefore, 200 simulations were run for the 2071-2100 period for each GCM and RCM, using the NCEP tted GLMs presented in Section 4. To condense the results, monthly rainfall summary statistics were computed for each simulation. Figures 16, 17 and 18 show, for each GCM, the simulated distributions of December daily rainfall mean, 5-day maximum and variance at Heathrow for each year between 2071-2100. The bands in the gures represent the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles of the simulated distributions. The irregular nature of the distributions is 17

Probability integral

Frequency

0.0

0.2

0.4 Percentiles

0.6

0.8

Total annual rainfall

mm 0
1960

400

800

1200

1965

1970

1975 Year

1980

1985

1990

Figure 8: i) Probability Integral Transform for annual rainfall at Birmingham: observed values as a sample from simulated distributions. ii) Simulated distributions of total annual rainfall in Birmingham. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

Probability integral

0 1 2 3 4 5 6

Frequency

0.0

0.2

0.4 Percentiles

0.6

0.8

1.0

Total annual rainfall

mm

0 200
1960

600

1000

1965

1970

1975 Year

1980

1985

1990

Figure 9: i) Probability Integral Transform for annual rainfall at Heathrow: observed values as a sample from simulated distributions. ii) Simulated distributions of total annual rainfall in Heathrow. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values. 18

Total summer (JJA) mean areal rainfall

mm

0 100
1960

300

500

1965

1970

1975 Year

1980

1985

1990

Total winter (DJF) mean areal rainfall

mm

0 100
1960

300

500

1965

1970

1975 Year

1980

1985

1990

Figure 10: Simulated distributions of total seasonal rainfall at Manchester (top:summer, bottom:winter). The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

Total summer (JJA) mean areal rainfall

0
1960

100 200 300 400

mm

1965

1970

1975 Year

1980

1985

1990

Total winter (DJF) mean areal rainfall

0
1960

100 200 300 400

mm

1965

1970

1975 Year

1980

1985

1990

Figure 11: Simulated distributions of total seasonal rainfall at Birmingham (top:summer, bottom:winter). The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

19

Total summer (JJA) mean areal rainfall

0
1960

100 200 300 400

mm

1965

1970

1975 Year

1980

1985

1990

Total winter (DJF) mean areal rainfall

mm

0 100
1960

300

500

1965

1970

1975 Year

1980

1985

1990

Figure 12: Simulated distributions of total seasonal rainfall at Heathrow (top:summer, bottom:winter). The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

Mean

Std Dev

P(wet)

Mean when wet

0.55

3.5

0.50

3.0

0.45

2.5

Proportion

mm

mm

0.40

2.0

mm 0.35 0.30
2 4 6 Month 8 10 12

1.5

1.0

6 Month

10

12

6 Month

10

12

6 Month

10

12

SD when wet

Maximum

Lag 1 ACF

Lag 2 ACF

0.3

Correlation

Correlation
2 4 6 Month 8 10 12

0.1

6 Month

10

12

6 Month

10

12

0.05

0.0

0.00

20

0.05

40

0.10

mm

mm

0.2

60

0.15

80

0.20

100

0.25

120

0.4

6 Month

10

12

Figure 13: Simulated distributions of monthly summary statistics at Manchester. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

20

Mean

Std Dev

P(wet)

Mean when wet

0.50

2.6

6.0

2.4

5.5

2.2

5.0

0.45

Proportion

2.0

4.5

0.40

mm

mm

1.8

4.0

mm 0.35 0.30 0.25

1.6

3.5

1.4

3.0

1.2

2.5

6 Month

10

12

6 Month

10

12

3.5

4.0

4.5

5.0

5.5

6.0

6.5

7.0

6 Month

10

12

SD when wet

Maximum

Lag 1 ACF

Lag 2 ACF

80

0.3

60

Correlation

Correlation
2 4 6 Month 8 10 12

40

20

0.1

6 Month

10

12

6 Month

10

12

0.05

0.0

0.00

0.05

0.10

mm

mm

0.2

0.15

0.20

0.25

100

0.4

6 Month

10

12

Figure 14: Simulated distributions of monthly summary statistics at Birmingham. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

Mean

Std Dev

P(wet)

Mean when wet

0.45

0.40

2.0

0.35

Proportion

mm

mm

1.5

0.30

mm 0.25 0.20
2 4 6 Month 8 10 12

1.0

6 Month

10

12

6 Month

10

12

6 Month

10

12

SD when wet

Maximum

Lag 1 ACF

Lag 2 ACF

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35

80

60

Correlation

Correlation
2 4 6 Month 8 10 12

40

20

6 Month

10

12

6 Month

10

12

0.05

0.00

0.05

0.10

mm

mm

0.15

0.20

0.25

100

10

6 Month

10

12

Figure 15: Simulated distributions of monthly summary statistics at Heathrow. The bands correspond to the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles and the thick line shows the observed values.

21

Simulated mean daily rainfall, conditional on CCCma output.

Simulated mean daily rainfall, conditional on HC output.

mm

mm 2 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

2075

2080

2085 Year

2090

2095

2100

Simulated mean daily rainfall, conditional on CSIRO output.

Simulated mean daily rainfall, conditional on MPI output.

mm

mm 2 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

2075

2080

2085 Year

2090

2095

2100

Figure 16: Simulated distributions of December mean daily rainfall at Heathrow, 2071-2100 a result of the GLMs being simulated conditional on only one realization of each climate model. These plots show that conditioning a rainfall simulation on a single climate model realization will underestimate the variability in rainfall properties at any particular time point. This is because, even if a climate model simulates sequences with the correct properties, no individual realisation provides the atmospheric values that will actually be observed in the future. However, there is also a second issue to consider. Figures 19 and 20 show the mean annual rainfall and proportion of wet days at Heathrow, calculated from the 200 simulations for each GCM. Visual inspection of these plots suggests that the annual rainfall statistics simulated using the dierent GCMs have dierent distributions, rather than just being dierent realizations from the same distribution. For example, the mean rainfall and proportion of wet days are generally higher for rainfall simulated using the CSIRO GCM atmospheric predictors, as compared to the rainfall simulated using the other three GCMs. This is not surprising given the dierences in the GCM projections shown in Figures 4 to 6. Tables 6 to 9 summarise the GLM simulated rainfall in December and June at Heathrow and Manchester respectively for the 2071-2100 period. They also show the rainfall means corresponding to the rainfall output generated by the climate models. As for the atmospheric variables, GCM/RCM precipitation corresponding to a particular site has been calculated using a weighted average of grid squares (see Section 3.3). As is standard in the literature, s we have assumed climate model rainfall output to have a multiplicative bias (i.e. Xym o follows the same distribution as Xym ) and so proportional changes were calculated from climate model outputs, with these changes then being applied to the observed 1961-1990 22

Simulated five day maximum daily rainfall, conditional on CCCma output.

Simulated five day maximum daily rainfall, conditional on HC output.

100

mm

50

mm 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

50

100

2075

2080

2085 Year

2090

2095

2100

Simulated five day maximum daily rainfall, conditional on CSIRO output.

Simulated five day maximum daily rainfall, conditional on MPI output.

100

mm

50

mm 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

50

100

2075

2080

2085 Year

2090

2095

2100

Figure 17: Simulated distributions of December 5-day maximum rainfall at Heathrow, 20712100
Simulated variance of daily rainfall, conditional on CCCma output.
150 150

Simulated variance of daily rainfall, conditional on HC output.

100

mm

50

mm 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

50

100

2075

2080

2085 Year

2090

2095

2100

Simulated variance of daily rainfall, conditional on CSIRO output.

Simulated variance of daily rainfall, conditional on MPI output.

100

mm

50

mm 0
2070 2075 2080 2085 Year 2090 2095 2100

0
2070

50

100

2075

2080

2085 Year

2090

2095

2100

Figure 18: Simulated distributions of December daily rainfall variance at Heathrow, 20712100 23

2.5

CCCma CSIRO

MPI HC

Mean Rainfall 1.0

1.5

2.0

2075

2080

2085 Year

2090

2095

2100

Figure 19: Simulated annual mean daily rainfall at Heathrow, 2071-2100 daily rainfall mean for each site. We cannot compare the GLM simulated variances and ve day maxima with those generated directly by the climate models, as climate models are generally considered to be area averages while we are looking a single site rainfall. Tables 6 and 8 show that both the GLM simulations and the climate models project wetter winters at Heathrow and Manchester airports in the 2071-2100 period with respect to the 1961-1990 period. Tables 7 and 9 show that both the GLM simulations and the climate models generally suggest drier summers in the future period, although the GLM approach almost always projects a greater decrease in mean rainfall. The slight dierences between the rainfall means from the two dierent approaches reect the dierences in the methodology used to relate rainfall to atmospheric variables. In favour of the GLM approach, we note that the climate model parameterisations are not constructed separately for each grid square; also, the climate model rainfall values are area averages whereas point rainfall is of more relevance for hydrological applications. Tables 6 to 9 also show that there are signicant dierences in the rainfall statistics simulated using the GLMs when conditioned on dierent GCMs/RCMs. The standard errors in Tables 6 to 9 were calculated using the result presented in Appendix 8, assuming stationarity within months for the 2071-2100 period. Although this is technically incorrect, since the atmospheric predictors from the climate models are not stationary in this period (see, for example, Figure 5), resulting simulated trends in the rainfall statistics appear minimal so the standard errors will be reasonably accurate. On the basis of these tables we conclude that no single GCM or RCM can be assumed to reproduce the properties of future climate.

24

0.50

CCCma CSIRO

MPI HC

Proportion of wet days

0.20
2070

0.25

0.30

0.35

0.40

0.45

2075

2080

2085 Year

2090

2095

2100

Figure 20: Simulated annual proportion wet days at Heathrow, 2071-2100

Climate Model

Daily mean GLM mean (s.e.) CM mean CCCma (GCM) 2.18 (0.14) 2.15 CSIRO (GCM) 2.76 (0.14) 2.41 MPI (GCM) 2.18 (0.16) 2.36 HC (GCM) 2.21 (0.18) 2.05 DMI (RCM) 2.08 (0.12) 2.21 SMHI (RCM) 2.12 (0.12) 2.30 HC (RCM) 2.25 (0.15) 2.19 Observed 1961-1990 1.81 (0.16)

Daily variance GLM mean (s.e.) 16.27 (1.28) 20.82 (1.29) 16.85 (1.61) 18.36 (1.92) 16.11 (1.23) 16.60 (1.22) 17.78 (1.56) 13.76 (2.24)

5-day max GLM mean (s.e.) 28.74 (1.37) 34.22 (1.25) 28.69 (1.74) 29.39 (1.94) 28.25 (1.23) 28.75 (1.26) 29.77 (1.61) 27.61 (2.53)

Table 6: Summary of the rainfall distribution in December at Heathrow airport. The climate model statistics are for the 2071-2100 period, with the observed 1961-1990 values shown for comparative purposes. The GLM mean values are based on the GLM simulations and the CM mean values correspond to the rainfall output produced directly by the climate models.

25

Climate Model

Daily mean GLM mean (s.e.) CM mean CCCma 1.06 (0.08) 1.56 CSIRO 1.13 (0.08) 1.97 MPI 1.10 (0.10) 1.03 HC (GCM) 0.98 (0.07) 1.45 DMI (RCM) 0.90 (0.06) 1.25 SMHI (RCM) 0.88 (0.06) 1.37 HC (RCM) 1.12 (0.09) 1.23 Observed 1961-1990 1.73 (0.19)

Daily variance GLM mean (s.e.) 10.95 (0.65) 11.28 (0.64) 10.79 (0.08) 9.95 (0.66) 9.27 (0.60) 9.05 (0.60) 11.11 (0.78) 22.17 (4.33)

5-day max GLM mean (s.e.) 18.74 (0.94) 19.58 (0.90) 18.65 (1.16) 17.54 (0.94) 16.63 (0.88) 16.21 (0.92) 19.11 (1.11) 29.31 (3.13)

Table 7: Summary of the rainfall distribution in June at Heathrow airport. The climate model statistics are for the 2071-2100 period, with the observed 1961-1990 values shown for comparative purposes. The GLM mean values are based on the GLM simulations and the CM mean values correspond to the rainfall output produced directly by the climate models.

Climate Model

Daily mean GLM mean (s.e.) CM mean CCCma 2.92 (0.13) 2.94 CSIRO 3.34 (0.10) 3.58 MPI 3.19 (0.14) 3.24 HC (GCM) 2.82 (0.15) 2.82 DMI (RCM) 2.83 (0.09) 2.87 SMHI (RCM) 2.85 (0.09) 3.04 HC (RCM) 2.98 (0.12) 2.88 Observed 1961-1990 2.53 (0.23)

Daily variance GLM mean (s.e.) 22.24 (1.04) 25.67 (0.90) 23.99 (1.08) 21.74 (1.25) 21.73 (0.80) 22.29 (0.87) 23.06 (1.04) 21.00 (3.80)

5-day max GLM mean (s.e.) 36.44 (1.07) 41.54 (0.83) 38.35 (1.17) 36.46 (1.25) 35.97 (0.81) 36.19 (0.84) 37.12 (1.08) 34.91 (2.85)

Table 8: Summary of the rainfall distribution in December at Manchester airport. The climate model statistics are for the 2071-2100 period, with the observed 1961-1990 values shown for comparative purposes. The GLM mean values are based on the GLM simulations and the CM mean values correspond to the rainfall output produced directly by the climate models.

26

Climate Model

Daily mean GLM mean (s.e.) CM mean CCCma 1.05 (0.10) 2.09 CSIRO 1.60 (0.10) 2.23 MPI 1.33 (0.13) 1.40 HC (GCM) 1.38 (0.08) 1.88 DMI (RCM) 1.01 (0.07) 1.60 SMHI (RCM) 0.99 (0.07) 1.74 HC (RCM) 1.28 (0.12) 1.59 Observed 1961-1990 2.20 (0.22)

Daily variance GLM mean (s.e.) 10.11 (0.88) 14.39 (0.87) 12.34 (1.09) 12.99 (0.72) 9.62 (0.67) 9.49 (0.70) 12.16 (1.05) 17.97 (3.13)

5-day max GLM mean (s.e.) 18.36 (1.28) 27.44 (0.92) 21.33 (1.44) 24.70 (0.85) 18.26 (0.97) 17.66 (1.02) 20.96 (1.41) 31.69 (2.83)

Table 9: Summary of the rainfall distribution in June at Manchester airport. The climate model statistics are for the 2071-2100 period, with the observed 1961-1990 values shown for comparative purposes. The GLM mean values are based on the GLM simulations and the CM mean values correspond to the rainfall output produced directly by the climate models.

Conclusion

In the work presented here we have tted GLMs relating observed single site rainfall to observed sea-level pressure, temperature and relative humidity. The observed atmospheric variables were obtained from either NCEP or ERA40 reanalysis data, and results were insensitive to which data set was used. It is noted that interactions involving autocorrelation and trending terms should be avoided if the models are to simulate realistic rainfall under future atmospheric conditions. The GLMs performed well when simulating rainfall for the 19611990 model tting period; interannual rainfall variability and monthly summary statistics were generally successfully reproduced. However, the assumption of a constant dispersion parameter throughout the year was not ideal as it resulted in the models underestimating rainfall variability in summer. Sequences of future rainfall were simulated by conditioning the models on GCM or RCM atmospheric outputs. The GLM simulations of future rainfall suggest drier summers and wetter winters in the 2071-2100 period with respect to the 19611990 period. These results agree with the rainfall output directly by the climate models, although the GLMs lead to drier summer projections. It was shown that more than one realisation of atmospheric predictors is needed from each climate model in order to obtain a comprehensive range of plausible rainfall simulations. Furthermore, conditioning the GLMs on dierent climate models gave signicantly dierent projections of future rainfall properties. These issues will be addressed in further work to be undertaken for FD2113, which will provide a methodology for simulating a comprehensive range of future daily rainfall sequences that accommodate the uncertainty in projections from deterministic climate models.

27

Appendix A

Logistic occurrence model for Birmingham airport t using the NCEP reanalysis data. Main effect: -----------Constant slp temperature relative humidity I(Y[t-1]>0) I(Y[t-2]>0) I(Y[t-3]>0) I(Y[t-k]>0: k=1 to 2) Daily seasonal effect, cosine component Daily seasonal effect, sine component Smooth February effect Ln(1+Y[t-1]) 2-way interactions: ------------------Daily seasonal effect, cosine component with I(Y[t-1]>0) Daily seasonal effect, sine component with I(Y[t-1]>0) Smooth February effect with I(Y[t-1]>0) relative humidity with Daily seasonal effect, cosine component relative humidity with Daily seasonal effect, sine component relative humidity with Smooth February effect temperature with Daily seasonal effect, cosine component temperature with Daily seasonal effect, sine component temperature with Smooth February effect slp with I(Y[t-1]>0) Coefficient -----------1.122993 -0.466950 0.004546 0.079479 0.912695 0.444489 0.108103 -0.497323 0.163514 0.052204 -0.569312 0.222263 Coefficient ----------0.055864 0.110895 0.290003 -0.051803 0.021715 -0.134119 0.032619 -0.009333 0.379504 0.282116 Std Err ------0.0375 0.0335 0.0249 0.0269 0.0794 0.0629 0.0456 0.0899 0.0415 0.0414 0.1796 0.0404 Std Err ------0.0631 0.0630 0.2726 0.0355 0.0332 0.1581 0.0352 0.0328 0.1632 0.0468

Gamma model for Birmingham airport t using the NCEP reanalysis data. Main effect: 28 Coefficient Std Err

-----------Constant slp temperature Ln(1+Y[t-1]) Ln(1+Y[t-2]) I(Y[t-k]>0: k=1 to 2) Daily seasonal effect, cosine component Daily seasonal effect, sine component I(Y[t-1]>0) 2-way interactions: ------------------slp with I(Y[t-k]>0: k=1 to 2) slp with Daily seasonal effect, slp with Daily seasonal effect, I(Y[t-k]>0: k=1 to 2) with Daily seasonal effect, I(Y[t-k]>0: k=1 to 2) with Daily seasonal effect,

----------1.385830 -0.070117 0.035137 0.098816 0.054616 -0.136968 -0.136556 -0.129932 -0.017867 Coefficient ----------0.007374 -0.073739

------0.0306 0.0241 0.0198 0.0307 0.0271 0.0664 0.0320 0.0322 0.0645 Std Err ------0.0443 0.0285 0.0289 0.0584 0.0580

cosine component -0.031641 sine component 0.041937 cosine component -0.019790 sine component

Logistic occurrence model for Manchester airport t using the NCEP reanalysis data. Main effect: -----------Constant slp temperature relative humidity I(Y[t-1]>0) I(Y[t-2]>0) I(Y[t-3]>0) I(Y[t-k]>0: k=1 to 2) Daily seasonal effect, cosine component Daily seasonal effect, sine component Smooth February effect Ln(1+Y[t-1]) 2-way interactions: ------------------Daily seasonal effect, cosine component with I(Y[t-1]>0) 29 Coefficient -----------1.087429 -0.314110 -0.009573 0.064493 1.123724 0.618078 0.147290 -0.730473 0.114228 -0.005382 -0.680694 0.271478 Coefficient ----------0.094655 Std Err ------0.0390 0.0343 0.0265 0.0295 0.0792 0.0632 0.0450 0.0886 0.0423 0.0421 0.1885 0.0390 Std Err ------0.0617

Daily seasonal effect, sine with I(Y[t-1]>0) Smooth February effect with I(Y[t-1]>0) temperature with Daily seasonal effect, temperature with Daily seasonal effect, temperature with Smooth February effect relative humidity with Daily seasonal effect, relative humidity with Daily seasonal effect, relative humidity with Smooth February effect slp with I(Y[t-1]>0)

component

0.005066 0.362890 0.102876

0.0617 0.2770 0.0372 0.0320 0.2220 0.0371 0.0323 0.2082 0.0451

cosine component 0.020711 sine component 0.563228 -0.065280 cosine component 0.057973 sine component -0.142363 0.115326

Gamma model for Manchester airport t using the NCEP reanalysis data. Main effect: -----------Constant slp temperature Ln(1+Y[t-1]) Ln(1+Y[t-2]) I(Y[t-k]>0: k=1 to 2) Daily seasonal effect, cosine component Daily seasonal effect, sine component I(Y[t-1]>0) 2-way interactions: ------------------slp with I(Y[t-1]>0) slp with Daily seasonal slp with Daily seasonal I(Y[t-1]>0) with Daily seasonal I(Y[t-1]>0) with Daily seasonal Coefficient ----------1.380710 -0.059167 0.013540 0.078862 0.022942 0.004166 -0.158346 -0.174798 0.048001 Coefficient ----------0.030049 -0.033750 effect, cosine component -0.042745 effect, sine component 0.106037 effect, cosine component 0.006836 effect, sine component 30 0.0494 0.0494 0.0256 Std Err ------0.0301 0.0289 0.0187 0.0265 0.0242 0.0564 0.0384 0.0387 0.0592 Std Err ------0.0366 0.0254

Appendix B

This method calculates standard errors for Y for a non-independent sample of random variables Yi. Let there exist a distribution for X, in this case the atmospheric predictors, and a distribution for Y |X, in this case a rainfall statistic. Assume we sample nx independent Xs and for each X we sample ny conditionally independent Y s. Let Y i be the mean of the Y s corresponding to the ith X and let Y be the overall mean of the Y s. Then V ar(Y ) = V ar( = Y 1 + Y 2 + ... + Y nx ) nx

1 V ar(Y 1 + Y 2 + ... + Y nx ) n2 x 1 (V ar(Y 1 ) + V ar(Y 2 ) + ... + V ar(Y nx )) = n2 x and V ar(Y i ) = E[V ar(Y i |Xi )] + V ar[E(Y i |Xi )] 1 = E[V ar(Y |Xi )] + V ar[E(Y |Xi )] ny where Y is generated conditional on Xi . Then V ar(Y ) = 1 1 [nx ( E[V ar(Y |Xi )] + V ar[E(Y |Xi )])] 2 nx ny E[V ar(Y |Xi )] V ar[E(Y |Xi )] + = ny nx nx

Thus the standard errors were estimated by calculating the sample mean and variance of the rainfall summary statistics in a particular month of each year, and then taking the sample variance and mean of these values over the years.

References
Chandler, R. (2002). GLIMCLIM: Generalized linear modelling for daily climate time series. Technical Report 227, University College London. Available at [Link] Chandler, R. E. (2000). A comparison of NCEP and observational rainfall data, 1958-1993. Technical report, University College London. Chandler, R. E. and Wheater, H. S. (2002). Analysis of rainfall variability using generalized linear models: A case study from the west of Ireland. Water Resources Research, 13(10):art. no. 1192. Dawid, A. P. (1984). Statistical theory: The prequential approach. Journal of the Royal Statistical Society, A, 147(2):278292. 31

Dobson, A. (1990). An Introduction to Generalized Linear Models. Chapman and Hall, London. Frei, C., Christensen, J., Deque, M., Jacob, D., Jones, R., and Vidale, P. (2003). Daily precipitation statistics in regional climate models: Evaluation and intercomparison for the European Alps. Journal of Geophysical Research - Atmospheres, 108(D3). Harris, I. (2004). Climate scenarios for water resource planning; evaluating gcm products for downscaling regional scenarios. Technical Report X1-045/2B, Environment Agency. IPCC (2001). Climate Change 2001 the scientic basis. Cambridge University Press, Cambridge. Third Report of the Intergovernmental Panel on Climate Change. Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., Deaven, D., Gandin, L., Iredell, M., Saha, S., White, G., Woollen, J., Zhu, Y., Chelliah, M., Ebisuzaki, W., Higgins, W., Janowiak, J., Mo, K. C., Ropelewski, C., Wang, J., Leetmaa, A., Reynolds, R., Jenne, R., and Joseph, D. (1996). The NCEP/NCAR 40-year reanalysis project. Bull. Amer. Meteorol. Soc., 77:437471. Prudhomme, C., Reynard, N., and Crooks, S. (2002). Downscaling of global climate models for ood frequency analysis: where are we now? Hydrological Processes, 16:11371150. Rowell, D. P. (2004). An initial estimate of the uncertainty in UK predicted climate change resulting from RCM formulation. Technical Report 49, Hadley Centre. Smith, L. A. (2002). What might we learn from climate forecasts. PNAS, 99(1):24872492. von Storch, H. and Zwiers, F. (1999). Statistical Analysis in Climate Research. Cambridge University Press. Wilby, R. L. and Dawson, C. W. (2001). Statistical Downscaling Model: SDSM. Yang, C., Chandler, R., Isham, V., and Wheater, H. (2006a). Quality control for daily observational rainfall series in the UK. Water and Environment Journal. to appear. Yang, C., Chandler, R., Isham, V., and Wheater, H. (2006b). Spatial-temporal rainfall simulation using generalized linear models. Water Resources Research. to appear.

32