Abstract

The winter climate of the British Isles is characterized by considerable interannual variability, which, because of the general climate sensitivity of a number of health outcomes, places at times considerable pressure on the provision of health services. Seasonal climate forecasts potentially could improve management within the health sector and assist in hedging against the vagaries of climatic variability. For this reason, an exploratory analysis of the potential utility of seasonal climate forecasting for the health sector in the United Kingdom is presented here. Study results revealed that the general level of winter mortality at the monthly to seasonal time scale possesses a strong association with simple descriptors of winter climate such as maximum temperature and the number of days below a given temperature threshold. Because such climate indices can be derived from the output of coupled seasonal climate prediction models, predictions of general levels of mortality may be possible using simple transfer functions that describe winter climate and health associations. Despite the potential one-month-ahead and one-season-ahead predictability of winter mortality levels, the predictability of the key climate indices by coupled climate models is shown to be somewhat limited, which compromises the ability to predict general levels of winter mortality for all months except February.

1. Introduction

The climate of the British Isles is characterized by considerable climatic variability, which has a discernible impact on a variety of economic sectors. Seasonal climate forecasts have the potential to improve economic management and assist in hedging against the vagaries of climatic variability. However, to assess the viability of applying seasonal climate prediction science to the improvement of economic performance and social well-being, the degree to which a particular sector of the economy or society is climate sensitive at the seasonal time scale must be established. Given this, the purpose of this paper is twofold—first, to explore the extent to which winter mortality is sensitive to winter climate variations at the monthly to seasonal time scale and, second, to present an exploratory analysis of the potential utility of seasonal climate forecasting for the U.K. health sector. To date this sector has received little consideration as a candidate for incorporating seasonal climate information into its planning and decision-making process. Winter is focused upon because the levels of excess winter mortality in the United Kingdom, which lie between 20 000 and 50 000 (Department of Health 2001), place the United Kingdom among the most winter-sensitive countries in Europe in terms of mortality (Curwen 1990; Eurowinter Group 1997).

Of chief concern for the United Kingdom are the health outcomes of respiratory, cerebrovascular, and ischemic heart disease (IHD) because these diseases make up a large proportion of the annual disease burden and mortality. They also display a strong seasonality and demonstrate a detectable degree of sensitivity to weather and climate variations (Mackenbach et al. 1992; Eurowinter Group 1997; Pell and Cobbe 1999; Gemmell et al. 2000; McGregor 1999, 2001; van Rossum et al. 2001; McGregor et al. 2004). It is consequently hypothesized that the general level of winter mortality will vary in accordance with interannual variation of the severity of winter because frequent exposure to cold causes a rise in mortality risk factors (Lloyd 1991) through increasing blood pressure and viscosity, vasoconstriction, heart rate, and angina (Morgan and Morgan 1997). As a consequence, the seasonal variation of mortality is often explained in relation to the climatological occurrence of cold weather in winter (Elwood 1993). Moreover, production losses from death in those of working age suffering from diseases such as ischemic heart disease, which is winter-weather sensitive (McGregor 2005), contribute greatly to the overall financial burden placed on the U.K. economy (British Heart Foundation 2000). This fact provides strong motivation for exploring mortality and temperature links at the winter monthly to seasonal time scale, because such links could form the basis of empirical seasonal climate prediction–informed health forecast (HF) models. The output from such models could be used within the framework of an early detection/watch warning system, similar in principle to those being developed for vectorborne diseases in the Tropics (Kovats and Cresswell 1999; Thomson et al. 2003). Such warning systems could help the health sector to make the transition from reactionary to anticipatory planning and could help to prepare the U.K. National Health Service for the health effects of forthcoming harsh winters because winter mortality and hospitalization patterns are similar.

An underlying assumption of any climate-based seasonal HF system will be that antecedent, current, or expected future values of climate variables or diagnostics will provide a fundamental source of predictability for a range of health outcomes. In turn, such predictability will be heavily dependent on whether reliable long-range forecasts of health-sensitive climate variables can be made. Long-range forecasts, specifically for the United Kingdom, are produced routinely by the Met Office using both dynamical and statistical information, with the public forecast for winter 2005/06 of anomalously low temperatures being a recent example. The statistical methods use predictors based on sea surface temperature information. This is also true for empirical seasonal climate predictions systems developed outside the Met Office (Lloyd-Hughes and Saunders 2002; McGregor and Phillips 2004; Qian and Saunders 2003; Wedgbrow et al. 2002). The Met Office dynamical long-range monthly and seasonal forecast systems provide global coverage. Forecasts for the United Kingdom are contained within subsets of that global information. The seasonal temperature and precipitation forecasts out to 6 months ahead have been produced as a research activity for several years, and information has been made available on the Met Office Web site (online at http://www.metoffice.gov.uk) since 2002. The dynamical seasonal forecast information is updated each month, and the amount of information made available is steadily increasing. Forecasts are issued in the form of probabilities for categories (e.g., three above-/near-/below-normal categories); the number of categories was recently increased from three to five. The amount of model data is also increasing; following the development of the European Multimodel Seasonal to Interannual Prediction System (EuroSIP), products based on a combination of information from the European Centre for Medium-Range Weather Forecasts (ECMWF) and Met Office systems have recently been added to the Web site. There is ample evidence that such a multimodel approach increases the skill of forecasts (Doblas-Reyes et al. 2005).

2. Data and methods

The analysis of IHD mortality and climate relationships presented here was undertaken at the monthly to seasonal level because relationships at this time scale could be potentially useful for developing extended-range seasonal climate prediction–informed HF models.

Daily natural all-cause mortality information for 1974–99 was obtained from the Office for National Statistics for the nine government office regions of England. Because there appears to be little variation in the climate and health relationships across England (McGregor et al. 2004), mortality data were extracted for the West Midlands region only for the purposes of this exploratory study (Fig. 1). The mortality data were divided into two categories: total mortality and mortality for over 65 yr of age. For each mortality group, daily mortality data were age standardized to deaths per 100 000 using midyear population estimates obtained from the Office of National Statistics and then were accumulated into monthly [December (D), January (J), and February (F)] and winter-season [December–February (DJF)] totals for the period 1974/75–1998/99. All mortality time series were detrended because statistically significant declines in mortality were found for all mortality groups. Detrending is important because it potentially removes from the analysis nonclimatic confounders resulting from possible medical and technological improvements over the study period. Detrending was achieved by fitting a cubic regression model to the mortality data (Wilks 1995) with time as the independent variable, because this model yielded the best description of the decrease of mortality over the study period when compared with other possible models such as splines. Regression model residuals were added to the mean mortality for the study period to produce a detrended time series for each mortality group.

Fig. 1.

Location of the nine grid points, numbered 1–9, for which predictions from the GloSea prediction system were available.

Fig. 1.

Location of the nine grid points, numbered 1–9, for which predictions from the GloSea prediction system were available.

Temperature was used as the main winter climate index because monthly and seasonal values of variables such as humidity, precipitation, and pressure were found in preliminary analyses to exert little effect on the general levels of mortality at the monthly to seasonal time scale. Daily surface maximum and minimum temperatures were obtained from the British Atmospheric Data Center for seven meteorological stations within the West Midlands. Because temperature displays little variation across the West Midlands, daily temperature data from the seven stations were averaged to construct a regional monthly and seasonal time series for the mean average (Tmean), maximum (Tmax), and minimum (Tmin) temperature over the period of 1974–99. Prior to any further analysis, both the temperature- and age-standardized mortality data were converted to standardized anomalies (z-score transformation) with a mean of 0 and a variance of 1.

Because the general level of monthly and seasonal mortality may be related to the number of “cold” days, as opposed to mean monthly and seasonal temperature, temperature data were used to derive variables that describe the number of days below a range of standardized Tmax, Tmean, and Tmin thresholds: 1.0, 0.5, 0, −0.5 and −1.0 standard deviations (sd), where a negative standard deviation indicates a value below the mean. These thresholds cover, in general, what may be considered as warm (sd > 1.0), mildly warm (1.0 > sd > 0.5), average (0.5 > sd > −0.5), mildly cool (−0.5 > sd > −1.0), and cool to cold (sd < −1.0; i.e., more than 1 sd below the mean) winter conditions. For the “temperature threshold” method, the number of days per month/winter season that had recorded temperatures below the prescribed threshold was counted.

The independent temperature datasets were used as input into jackknife regression analysis (Duchesne and MacGregor 2001) to produce a set of prediction models for total and 65-yr-plus (65-yr+) mortality at the monthly and seasonal time scales. The jackknife technique allows the usefulness of a prediction equation to be tested when the sample size is not large enough to permit a separation of the sample into dependent and independent cases (Chan et al. 1998; Wilks 1995). The technique is reviewed in Wilks (1995) and has been applied widely in the development of prediction algorithms in the fields of applied meteorology and climatology (Chan 1995; Colman and Davey 2003; Demuth et al. 2004; Elsner and Schmertmann 1993; Klotzkbach and Gray 2004; Pagano and Garen 2005). The jackknife technique comprises a number of steps. First, a prediction equation is derived based on all years except one—for example, year 1 of the 25 years in this study. The prediction equation is then used to predict the mortality level of year 1 as well as the error. This two-step procedure is then repeated until all years have been excluded once (24 times in this study). In this way each of the predictions can be considered to be independent. Assuming that the prediction errors are not too large, as indicated by the statistical significance of the correlation between the predicted and observed values of mortality, then the resultant prediction equation can be used for making future predictions (Chan et al. 1998).

Two broad types of prediction models were produced. One used as a set of predictor variables the mean, maximum, and minimum daily temperature (Tmean, Tmax, Tmin) for a given month (D, J, or F) or the overall season (DJF). The second type of model used the number of days below five standardized temperature thresholds as the predictor variable. The statistical significance of the models was assessed by an analysis of variance. The models were subsequently “forced” with hindcasts (historical predictions) of monthly and seasonal Tmean, Tmax, and Tmin and the number of days below a given threshold to produce hindcasts of monthly and seasonal mortality.

Values for the predictor variables were derived from 186-day hindcasts (starting in November and running through the end of February) of temperature for nine grid points covering the wider study area (Fig. 1) from 15 ensembles (15 different runs of the same model; each run possesses slightly different initial conditions) of the Met Office’s Global Seasonal (GloSea) Prediction System for the period of 1987–99. The GloSea Prediction System is based on the Met Office Third Hadley Centre Coupled Ocean–Atmosphere General Circulation Model (HadCM3; Gordon et al. 2000) coupled climate model. It uses a stretched north–south ocean grid, in which a resolution of 1.25° in both the meridional and zonal directions improves to 0.28° in the meridional direction in the Tropics, and it has 40 vertical levels. No flux correction is employed, and an interactive sea ice model is incorporated. The atmospheric component of GloSea is the Third Hadley Centre Atmospheric Model (HadAM3; Pope et al. 2000). HadAM3 has a horizontal resolution of 2.5° latitude by 3.75° longitude and 19 vertical levels. GloSea uses a coastal tiling scheme, which allows the ocean model to have a coastline determined by the ocean grid rather than by the lower-resolution atmosphere grid, giving improved representation of land/sea geography.

Initial ocean conditions for the GloSea hindcasts are obtained by forcing the ocean component with momentum, heat, and freshwater fluxes from the 40-yr ECMWF Reanalysis (ERA-40), after first spinning up from climatological values of temperature and salinity (Levitus and Boyer 1994; Levitus et al. 1995) using ERA-40 flux climatological data. A slow-time-scale relaxation of temperature and salinity to climatological values is applied at all model levels. From 1986 onward the Forecasting Ocean Assimilation Model (FOAM) data assimilation scheme is used to assimilate subsurface temperature data. The FOAM scheme (Bell et al. 2000) uses a form of optimal interpolation. An equatorial pressure gradient bias correction scheme (Bell et al. 2004; Huddleston et al. 2004) is used to address unrealistic deep overturning circulations that can occur near the equator when only thermal data are assimilated into an ocean model.

To generate perturbed initial conditions for the ensemble, the method described in Palmer et al. (2004) is used. The method is the same as that used for real-time forecasting but is adapted for the 15-member ensembles (40 members are used in real-time predictions). The ensemble size for retrospective (15 in this study) and real-time forecasts is constrained by computational resources and represents a compromise between the desirability of a large ensemble to sample likely outcomes and the desirability of a forecast model with good representation of physical processes and high spatial resolution. Because the majority of climate predictability at the intraseasonal to seasonal time scale in the North Atlantic–European sector can be attributed to long-term memory in the ocean surface heat content (Colman and Davey 2003; McGregor and Phillips 2004), the 15 ensembles were generated as a mixture of wind stress perturbations applied during the ocean assimilation phase and SST perturbations applied at forecast initialization time.

To evaluate the degree of predictability of winter climate and gain insight into any interensemble variation in predictability, the GloSea daily predictions for the nine grid points for all 15 ensembles were transformed into standardized monthly and seasonal anomalies (based on ensemble mean and standard deviation) and then were compared with standardized values from the observational record for the West Midlands using correlation analysis. Predictions for nine grid points were compared with observations to establish whether grid points at some distance from the study area offered better or worse predictability when compared with the “local” grid point. Further, probabilistic forecasts of equiprobable tercile categories (below, near, and above normal) of the 15-ensemble mean temperature (Tmean, Tmax, and Tmin) were also generated. These were assessed using relative operating characteristic (ROC) curves, or scores (Stanski et al. 1989; Toth et al. 2003). Only results of the ROC analysis for the lowest tercile are presented here because anomalously high levels of winter mortality are generally associated with below-normal temperatures that occur in the lowest tercile of the temperature distribution for the study area (McGregor 2005; McGregor et al. 2004).

3. Results

Mortality prediction models based on the two methods are presented in Tables 1 and 2. At the monthly (D, J, and F) time scale there are significant inverse associations between Tmean, Tmax, and Tmin and total and 65-yr+ mortality in most cases (Fig. 2a). This result contrasts with the situation at the seasonal level (DJF). Progression through the winter from December to February appears to be associated with a strengthening of the temperature total mortality association such that, of the three individual winter months, February possesses the prediction model with the best fit to the original data (Table 1). Furthermore, the associations between mean temperature conditions and mortality appear to be stronger for total mortality as compared with 65-yr+ mortality for February. For December, however, this is not the case because in this month 65-yr+ mortality bears a closer association with temperature than does total mortality.

Table 1.

Monthly and seasonal jackknife regression models based on Tmean, Tmax, and Tmin. An em-dash indicates no statistical significance at the 0.01 or 0.05 level, a single asterisk indicates statistical significance at the 0.05 level, and a double asterisk indicates statistical significance at the 0.01 level; r2 is an adjusted coefficient of explanation that takes into account serial correlation.

Monthly and seasonal jackknife regression models based on Tmean, Tmax, and Tmin. An em-dash indicates no statistical significance at the 0.01 or 0.05 level, a single asterisk indicates statistical significance at the 0.05 level, and a double asterisk indicates statistical significance at the 0.01 level; r 2 is an adjusted coefficient of explanation that takes into account serial correlation.
Monthly and seasonal jackknife regression models based on Tmean, Tmax, and Tmin. An em-dash indicates no statistical significance at the 0.01 or 0.05 level, a single asterisk indicates statistical significance at the 0.05 level, and a double asterisk indicates statistical significance at the 0.01 level; r 2 is an adjusted coefficient of explanation that takes into account serial correlation.
Table 2.

As in Table 1 but for models based on the number of days below a given standardized temperature threshold.

As in Table 1 but for models based on the number of days below a given standardized temperature threshold.
As in Table 1 but for models based on the number of days below a given standardized temperature threshold.
Fig. 2.

(a) Association between February monthly average standardized temperature and February mortality rate (deaths per 100 000). Regression coefficient (slopeAVG) is −140.368, Y intercept (yintcptAVG) is 1844.403, standard error of regression coefficient (Yintcpt) is 3.836, correlation coefficient (Corr-Coe) is −0.671, and coefficient of determination or r2 (Squ-coeff) is 0.450. (b) Association between the number of days below February standardized average temperature of −1 (units of standard deviation) and February mortality rate (deaths per 100 000). The slopeAVG is 17.488, yintcptAVG is 1756.43, Yintcpt is 3.3836, Corr-Coe is 0.803, and Squ-coeff is 0.645.

Fig. 2.

(a) Association between February monthly average standardized temperature and February mortality rate (deaths per 100 000). Regression coefficient (slopeAVG) is −140.368, Y intercept (yintcptAVG) is 1844.403, standard error of regression coefficient (Yintcpt) is 3.836, correlation coefficient (Corr-Coe) is −0.671, and coefficient of determination or r2 (Squ-coeff) is 0.450. (b) Association between the number of days below February standardized average temperature of −1 (units of standard deviation) and February mortality rate (deaths per 100 000). The slopeAVG is 17.488, yintcptAVG is 1756.43, Yintcpt is 3.3836, Corr-Coe is 0.803, and Squ-coeff is 0.645.

The degree of association between the number of days below a given standardized temperature threshold and mortality is shown in Table 2. For D, J, and F and DJF the standardized temperature threshold displayed in the table indicates the threshold model for which the best fit was obtained from among five possible threshold models. Overall, there is a positive association between the number of days below the various thresholds and mortality (Fig. 2b), indicating that, for all months and the winter season as a whole, higher mortality levels are associated with an increasing number of cool to cold days. Both total and 65-yr+ mortality appear to be most sensitive to the number of days below −1.0 standard deviation for J and F, which represents cool to very cold days. February possesses the strongest association between the number of cool days and mortality, and associations are stronger for total mortality as compared with 65-yr+ mortality. Of interest is December because the threshold value below which mortality demonstrates the greatest sensitivity is higher than it is for January and February. This fact may indicate nonacclimatization (both physiological and behavioral) to cool weather early in the season such that the temperatures in December do not have to be all that low to engender a health response. By January and February people have adjusted to the winter weather and therefore sensitivity has decreased, with the result that much lower temperatures are required to produce high mortality levels.

Of particular note is that the threshold method reveals models with greater fit than the method based on Tmean, Tmax, and Tmin, as is seen clearly for the month of February. Further, the threshold method discloses statistically significant associations at the seasonal (DJF) level for 65-yr+ mortality. Overall, the results of the threshold-method analysis indicate that the number of cool to cold days per month or winter season may be better than the mean Tmean, Tmax, and Tmin as a climate diagnostic for health analysis at the intraseasonal to seasonal time scale. Furthermore, although there is little contrast in the degree of association between the number of days below the standardized threshold values for Tmean, Tmax, and Tmin and 65-yr+ mortality, for total mortality Tmax and Tmean thresholds provide a better explanation of the variability of this health outcome.

Figure 3 presents the range of correlations between the predicted and observed minimum temperature for D, J, F and DJF 1987–99 for the nine grid points and for each of the individual ensemble members. Tmean and Tmax are not shown here because they demonstrate behavior that is similar to that of Tmin. The D, J, and F predictions are made 1 month in advance. For DJF, results are shown for a November and December prediction of DJF minimum temperature. Immediately apparent is the interensemble variability of predictability. Of the three winter months, February appears to be the most predictable, with consistently high correlations between predicted temperature for the nine grid points and observed temperature over the West Midlands (Fig. 3c). The narrow range of correlation coefficients for February also demonstrates that predictions across the nine-gridpoint domain are, in general, equally reliable as indicators of temperature anomalies for the study area. This situation, however, is not the case for December or January (Figs. 3a,b) because some grid points and ensembles provide a much better indication of the likely temperature anomalies over the West Midlands than do others. The grid points that demonstrate a good level of association with the observations in the West Midlands for December and January are generally located to the east of the study area over the North Sea (grid points 3, 6, and 9 in Fig. 1), indicating that, in terms of temperature anomalies, predictions over nearby sea as opposed to land surfaces may be more reliable indicators of anomalous winter conditions. At the seasonal level, there is also considerable interensemble variability of predictability as well as intraensemble variability (Figs. 3e,f). Although not shown here, similar results were found for the temperature thresholds such that stronger associations between the predicted and observed number of days below a given threshold were observed for February for the grid point closest to the study area. However, predictions of the number of days below the range of thresholds for February were inferior to those for Tmean, Tmax, and Tmin.

Fig. 3.

Box plots of interensemble variation in correlation between predicted (1 month ahead) and observed maximum temperature for (a) December, (b) January, and (c) February and (d) November-based and (e) December-based prediction of winter for nine grid points. Horizontal axis is ensemble number (1–15), and vertical axis is correlation between predicted and observed temperature. In the box plots the median, upper, and lower quartile and 90th and 10th percentile values for the correlation coefficients are shown.

Fig. 3.

Box plots of interensemble variation in correlation between predicted (1 month ahead) and observed maximum temperature for (a) December, (b) January, and (c) February and (d) November-based and (e) December-based prediction of winter for nine grid points. Horizontal axis is ensemble number (1–15), and vertical axis is correlation between predicted and observed temperature. In the box plots the median, upper, and lower quartile and 90th and 10th percentile values for the correlation coefficients are shown.

The conclusions drawn regarding intermonthly contrasts in predictability, based on the correlation-based assessment, are corroborated by the results of the ROC analysis of the 15-ensemble mean for Tmin (Fig. 4). At the monthly scale, February is the only month to demonstrate an ROC score (0.63) in excess of 0.5, which is the “no skill” value attained by random forecasts or constant forecasts of the climatological probability of the event (Fig. 4c). Forecasts for the other two winter months possess no skill, but January is marginally better than December. Although not shown here, DJF forecasts initialized in November and December were found to possess no skill.

Fig. 4.

ROC analysis of the 15 ensemble mean for lower quartile of Tmin for (a) December, (b) January, and (c) February. “HIT” refers to hit rate for which predictions are correct, and “FAR” refers to false-alarm rate for which predictions are wrongly made. The dashed line represents the division between skill (above) and no skill (below). The greater is the area under the curve above (under) the dashed line, the better (worse) the skill of the predictions is.

Fig. 4.

ROC analysis of the 15 ensemble mean for lower quartile of Tmin for (a) December, (b) January, and (c) February. “HIT” refers to hit rate for which predictions are correct, and “FAR” refers to false-alarm rate for which predictions are wrongly made. The dashed line represents the division between skill (above) and no skill (below). The greater is the area under the curve above (under) the dashed line, the better (worse) the skill of the predictions is.

To evaluate the degree to which the general level of mortality is predictable, 1-month-ahead hindcasts for February 1987–99 were used as input into the models presented in Tables 1 and 2 to make retrospective forecasts of February mortality. February was focused on because it is the month with the greatest temperature forecast skill. For the hindcasts, predicted temperatures for grid point 5 (Fig. 1) were used because it is the nearest land-based grid point with the best predictability for the West Midlands. Mortality “forecasts” were then compared with the observed mortality. The Tmax and Tmean February models produced forecasts that possess a statistically significant association with the observed total mortality (r = 0.64) and 65-yr+ (r = 0.59) mortality, respectively. The threshold models fared not as well in predicting total mortality, despite the February threshold model (Table 2) having a level of fit that is similar to that of the February Tmean model (Table 1). The inferior performance of the February threshold model may be due to the fact that the 1-month-ahead predictions of the number of days below the various standardized temperature thresholds are not as good as the predictions of the more conventional Tmean, Tmax, and Tmin.

4. Discussion and conclusions

High levels of natural all-cause mortality at the winter monthly to seasonal scale for the West Midlands have been found to be statistically associated with anomalous cold as described by standardized anomalies of Tmin and Tmax and a greater frequency of days below a given temperature threshold. Interannual variations in winter mortality are associated with winter climate variability because periods of anomalously cold weather have a fundamental effect on mortality through increasing, for example, the risk factors of heart disease (Collins et al. 1985; Jehn et al. 2002; Healy 2003; Keatinge et al. 1984; Mercer 2003; Woodhouse et al. 1993). Elderly people with preexisting circulatory problems are particularly vulnerable to the risk factors during cold weather (Neild et al. 1994).

Study results also revealed that the level of association between the monthly climate indices and mortality varies across the months. This finding may indicate that the degree of association between weather and mortality depends on how early or late it is in the winter season. The higher correlation found between mortality and temperature in February is difficult to explain but may be related to a buildup of a pool of susceptible individuals throughout the winter and the fact that February possesses a greater frequency of cool to cold days (62%), as described by the standardized Tmax threshold value of −1.0, as compared with January (50%) and December (38%).

Although the climate indices utilized in this study are able to explain a good proportion of the variability of winter mortality, the utility of the associated mortality prediction models (Tables 1 and 2) for making reliable predictions of the general level of mortality will be very much dependent on the skill of the forecasts of 1-month-ahead values of the requisite climate indices. Based on the results of this study, currently reliable forecasting of mortality appears to be a possibility only for the month of February, mainly because of the fact that this month possesses the greatest degree of predictability of the health-sensitive climate indices. However, bear in mind that only 13 yr have been used to assess the retrospective predictability of mortality. With the availability of hindcasts of winter temperature conditions for the 25 yr used to develop the mortality prediction models, a better idea of mortality predictability will be gained.

A comparison of the two methods for mortality prediction revealed that the models based on Tmean, Tmax, and Tmin produced superior mortality predictions, despite the threshold models having a better fit to the original data. This result is because predictions of the number of threshold days were not as skillful as those for Tmean, Tmax, and Tmin. Therefore, although a good validated prediction model may exist, the efficacy of such a model for prediction is heavily dependent on the accuracy of the predicted values of the input variable. One reason why skillful predictions of temperature may have been obtained for February and not for January and December is that February possesses less variability in temperature when compared with the other two winter months. Further, because winter North Atlantic sea surface temperature (SST) anomalies play a role in reinforcing North Atlantic Oscillation activity in February and into early spring (Rodwell and Folland 2003), improved temperature predictions over Europe for this period might be expected if coupled model–based prediction systems like GloSea can capture such ocean–atmosphere interaction, which normally manifests itself as a pronounced tripole SST pattern in the North Atlantic. In its positive phase, the tripole pattern comprises negative SST anomalies in the subtropical North Atlantic and to the south of Greenland and positive anomalies off the east coast of North America (Davies et al. 1997). That resolution of this dipole pattern is possible has been demonstrated by a recent evaluation of the GloSea prediction system (Graham et al. 2005), which is not only able to capture the tripole SST pattern but also produces good predictions of late-winter and early spring temperature over western Europe when development or intensification of a positive tripole SST pattern occurs subsequent to January.

Despite February demonstrating a fair level of predictability, problems exist for the other winter months as well as for winter as a whole. The problem of winter predictability in the extratropics is widely acknowledged by the seasonal prediction community (Grotzner et al. 1999; Goddard et al. 2001; Marshall et al. 2001; Graham et al. 2000; Liniger et al. 2004; Quan et al. 2004). The winter predictability problem arises because much of the variability in the extratropical atmosphere and oceans is associated with the chaotic behavior of the atmospheric circulation and associated nonlinearities. These factors pose an array of problems for seasonal climate forecasting. Although the winter atmospheric circulation over Europe in general is composed of a number of circulation regimes, such as blocked and zonal modes that persist on occasions well beyond the synoptic time scale, the transition between these regimes is rapid and unpredictable because of nonlinearity of atmospheric dynamics. Therefore, the challenge for seasonal climate prediction in the extratropics is perhaps not to predict the timing of the transition between circulation modes but to estimate the predominant regimes in the forthcoming season (WCRP 1995, 1997). For this reason, because slowly evolving boundary conditions in the North Atlantic, as expressed by SST anomalies, are likely to affect the frequency of the circulation regime in any one winter season, the resolution of the North Atlantic SST tripole pattern, which has a discernable impact on the winter circulation regime over Europe, by coupled ocean–atmosphere models appears to be critical if better winter-season predictions are to be forthcoming. This fact bears significant implications for the application of seasonal climate prediction information to the health sector, because high levels of winter mortality are generally associated with anomalously cold winters (McGregor 2005). Such winters are dominated by blocking regimes, which have a greater frequency of occurrence in years in which the North Atlantic Oscillation is in a negative phase (Shabbar et al. 2001). If improved coupled ocean–atmosphere seasonal prediction systems can deliver information on the likely circulation regime statistics for a forthcoming winter, then worthwhile extended-range predictions of winter mortality and, possibly, hospital admission could be made. Recently developed multisystem forecasts offer some hope in this regard (Doblas-Reyes et al. 2005). However, it should be acknowledged that the societal utility of such predictions will depend very much on seasonal climate forecasters engaging with potential end users to determine their specific forecasting needs (Goddard et al. 2001; Hartmann et al. 2002; Stern and Easterling 1999). Moreover, if extended-range predictions are to be used as the basis for issuing early warnings about impending harsh winters and their related health effects, then, to be effective, a requirement of such warning systems will be availability of a well-designed and well-integrated set of implementable intervention measures (Ebi et al. 2004). Without these measures, well-intentioned winter climate/health warnings will be rendered ineffective and the reduction of the burden of disease attributable to intraseasonal to seasonal climate variability will not be achieved.

REFERENCES

REFERENCES
Bell
,
M. J.
,
R. M.
Forbes
, and
A.
Hines
,
2000
:
Assessment of the FOAM global data assimilation system for real-time operational ocean forecasting.
J. Mar. Syst.
,
25
,
1
22
.
Bell
,
M. J.
,
M. J.
Martin
, and
N. K.
Nichols
,
2004
:
Assimilation of data into an ocean model with systematic errors near the equator.
Quart. J. Roy. Meteor. Soc.
,
130
,
873
893
.
British Heart Foundation
,
2000
:
Coronary heart disease statistics: British Heart Foundation statistics database 2000. 204 pp
.
Chan
,
J. C. L.
,
1995
:
Prediction of annual tropical cyclone activity over the western North Pacific and the South China Sea.
Int. J. Climatol.
,
15
,
1011
1019
.
Chan
,
J. C. L.
,
J-E.
Shi
, and
C-M.
Lam
,
1998
:
Seasonal forecasting of tropical cyclone activity over the western North Pacific and the South China Sea.
Wea. Forecasting
,
13
,
997
1004
.
Collins
,
K. J.
,
J. C.
Easton
,
H.
Belfield-Smith
,
A. N.
Exton-Smith
, and
R. A.
Pluck
,
1985
:
Effects of age on body temperature and blood pressure in cold environments.
Clin. Sci.
,
69
,
465
470
.
Colman
,
A. W.
, and
M. K.
Davey
,
2003
:
Statistical prediction of global sea-surface temperature anomalies.
Int. J. Climatol.
,
23
,
1677
1697
.
Curwen
,
M.
,
1990
:
Excess winter mortality: A British phenomenon?
Health Trends
,
22
,
169
175
.
Davies
,
J. R.
,
D. P.
Rowell
, and
C. K.
Folland
,
1997
:
North Atlantic and European predictability using an ensemble of multi-decadal atmospheric GCM simulations.
Int. J. Climatol.
,
17
,
1263
1284
.
Demuth
,
J.
,
M.
DeMaria
,
J. A.
Knaff
, and
T. H.
Vonder Haar
,
2004
:
Evaluation of Advanced Microwave Sounding Unit tropical cyclone intensity and size estimation algorithms.
J. Appl. Meteor.
,
43
,
282
296
.
Department of Health
,
2001
:
Health Effects of Climate Change in the UK: An Expert Review for Comment.
238 pp
.
Doblas-Reyes
,
P. J.
,
R.
Hagedorn
, and
T. N.
Palmer
,
2005
:
The rationale behind the success of multi-model ensembles in seasonal forecasting. Part II: Calibration and combination.
Tellus
,
57A
,
234
252
.
Duchesne
,
C.
, and
J. F.
MacGregor
,
2001
:
Jackknife and bootstrap methods in the identification of dynamic models.
J. Process Contrib.
,
11
,
553
564
.
Ebi
,
K. L.
,
T. J.
Teisberg
,
J.
Thomas
,
L. S.
Kalkstein
,
L.
Robinson
, and
R. S.
Weiher
,
2004
:
Heat watch/warning systems save lives: Estimated costs and benefits for Philadelphia 1995–98.
Bull. Amer. Meteor. Soc.
,
85
,
1067
1073
.
Elsner
,
J. B.
, and
C. P.
Schmertmann
,
1993
:
Improving extended-range seasonal predictions of intense Atlantic hurricane activity.
Wea. Forecasting
,
8
,
345
351
.
Elwood
,
P. C.
,
1993
:
Temperature and risk factors for ischaemic heart disease in the Caerphilly prospective study.
Br. Heart J.
,
70
,
520
525
.
Eurowinter Group
,
1997
:
Cold exposure and winter mortality from ischaemic heart disease, cerebrovascular disease, respiratory disease, and all causes in warm and cold regions of Europe.
Lancet
,
349
,
1341
1346
.
Gemmell
,
I.
,
P.
McLoone
,
F. A.
Boddy
,
G. J.
Dickinson
, and
G. C. M.
Watt
,
2000
:
Seasonal variation in mortality in Scotland.
Int. J. Epidemiol.
,
29
,
274
279
.
Goddard
,
L.
,
S. J.
Mason
,
S.
Zebiak
,
C. F.
Ropelewski
,
R.
Basher
, and
M.
Cane
,
2001
:
Current approaches to seasonal to inter-annual climate predictions.
Int. J. Climatol.
,
21
,
1111
1152
.
Gordon
,
C.
,
C.
Cooper
,
C. A.
Senior
,
H.
Banks
,
J. M.
Gregory
,
T. C.
Johns
,
J. F. B.
Mitchell
, and
R. A.
Wood
,
2000
:
The simulation of SST, sea ice extents and ocean heat transports in a version of the Hadley Centre coupled model without flux adjustments.
Climate Dyn.
,
16
,
147
168
.
Graham
,
R. J.
,
A. D. L.
Evans
,
K. R.
Mylne
,
M. S. J.
Harrison
, and
K. B.
Robertson
,
2000
:
An assessment of seasonal predictability using atmospheric general circulation models.
Quart. J. Roy. Meteor. Soc.
,
126
,
2211
2240
.
Graham
,
R. J.
,
M.
Gordon
,
P. J.
McLean
,
S.
Ineson
,
M. R.
Huddelston
,
M. K.
Davey
,
A.
Brookshaw
, and
R. T. H.
Barnes
,
2005
:
A performance comparison of coupled and uncoupled versions of the Met Office seasonal prediction general circulation model.
Tellus
,
57A
,
320
339
.
Grotzner
,
A.
,
M.
Latif
, and
A.
Timmermann
,
1999
:
Interannual to decadal predictability in a coupled ocean–atmosphere general circulation model.
J. Climate
,
12
,
2607
2624
.
Hartmann
,
H. C.
,
T. C.
Pagano
,
S.
Sorooshian
, and
R.
Bales
,
2002
:
Confidence builders: Evaluating seasonal forecasts from user perspectives.
Bull. Amer. Meteor. Soc.
,
83
,
683
698
.
Healy
,
J. D.
,
2003
:
Excess winter mortality in Europe: A cross-country analysis identifying key risk factors.
J. Epidemiol. Community Health
,
57
,
784
789
.
Huddleston
,
M. R.
,
M. J.
Bell
,
M. J.
Martin
, and
N. K.
Nichols
,
2004
:
Assessment of wind-stress errors using bias corrected ocean data assimilation.
Quart. J. Roy. Meteor. Soc.
,
130
,
853
871
.
Jehn
,
M.
,
L. J.
Appel
,
F. M.
Sacks
, and
E. R.
Miller
III
,
2002
:
The effect of ambient temperature and barometric pressure on ambulatory blood pressure variability.
Amer. J. Hyperten.
,
15
,
941
945
.
Keatinge
,
W. R.
,
S. R. K.
Coleshaw
,
F.
Cotter
,
M.
Mattock
,
M.
Murphy
, and
R.
Chelliahm
,
1984
:
Increases in platelet and red cell counts, blood viscosity, and arterial pressure during mild surface cooling: Factors in mortality from coronary and cerebral thrombosis in winter.
Br. Med. J.
,
289
,
1404
1408
.
Klotzbach
,
P. J.
, and
W. M.
Gray
,
2004
:
Updated 6–11-month prediction of Atlantic basin seasonal hurricane activity.
Wea. Forecasting
,
19
,
917
934
.
Kovats
,
S.
, and
M.
Cresswell
,
1999
:
El Niño and health.
Tiempo
,
33
.
9–13
.
Levitus
,
S.
, and
T. P.
Boyer
,
1994
:
Temperature. Vol. 4, World Ocean Atlas 1994, NOAA Atlas NESDIS 4, 117 pp
.
Levitus
,
S.
,
R.
Burgett
, and
T. P.
Boyer
,
1995
:
Salinity. Vol. 3, World Ocean Atlas 1994, NOAA Atlas NESDIS 3, 99 pp
.
Liniger
,
M. A.
,
A.
Wolfgang
,
A.
Mueller
, and
C.
Appenzeller
,
2004
:
Seasonal forecasting: Skill assessment from a grid point perspective.
Geophys. Res. Abstracts
,
6
..
Lloyd
,
E. L.
,
1991
:
The role of cold in ischaemic heart disease: A review.
Pub. Health
,
105
,
205
215
.
Lloyd-Hughes
,
B.
, and
M. A.
Saunders
,
2002
:
Seasonal prediction of European spring precipitation from El Nino Southern Oscillation and sea surface temperatures.
Int. J. Climatol.
,
22
,
1
14
.
Mackenbach
,
J. P.
,
A. E.
Kunst
, and
C. W. N.
Looman
,
1992
:
Seasonal variation in mortality in The Netherlands.
J. Epidemiol. Community Health
,
46
,
261
265
.
Marshall
,
R. T.
, and
Coauthors
,
2001
:
North Atlantic climate variability: Phenomena, impacts and mechanisms.
Int. J. Climatol.
,
21
,
1863
1898
.
McGregor
,
G. R.
,
1999
:
Winter ischaemic heart disease deaths in Birmingham, UK: A synoptic climatological analysis.
Climatic Res.
,
13
,
17
31
.
McGregor
,
G. R.
,
2001
:
The meteorological sensitivity of Ischaemic heart disease mortality events in Birmingham, UK.
Int. J. Biometeor.
,
45
,
133
142
.
McGregor
,
G. R.
,
2005
:
Winter North Atlantic Oscillation and ischaemic heart disease mortality in three English counties.
Int. J. Biometeor.
,
49
,
197
204
.
McGregor
,
G. R.
, and
I. D.
Phillips
,
2004
:
Specification and prediction of monthly and seasonal rainfall over the southwest peninsula of England.
Quart. J. Roy. Meteor. Soc.
,
130
,
193
210
.
McGregor
,
G. R.
,
H. A.
Watkin
, and
M.
Cox
,
2004
:
Temperature and ischaemic heart disease mortality seasonality relationships: Implications for climate based health forecasting.
Climatic Res.
,
25
,
253
263
.
Mercer
,
J. B.
,
2003
:
Cold—An underrated risk factor for health.
Environ. Res.
,
92
,
8
13
.
Morgan
,
M. D.
, and
J. M.
Morgan
,
1997
:
Weather and People.
Prentice Hall, 200 pp
.
Neild
,
P. J.
,
D.
Syndercombe-Court
,
W. R.
Keatinge
,
G. C.
Donaldson
,
M.
Mattock
, and
M.
Caunce
,
1994
:
Cold induced increases in erthrocyte count, plasma cholesterol and plasma fibrinogen of elderly people without a comparable rise in Protein C or Factor X.
Clin. Sci.
,
86
,
43
48
.
Pagano
,
T.
, and
D.
Garen
,
2005
:
A recent increase in western U.S. streamflow variability and persistence.
J. Hydrometeor.
,
6
,
173
179
.
Palmer
,
T. N.
, and
Coauthors
,
2004
:
Development of a European Multimodel Ensemble System for Seasonal-to-Interannual Prediction (DEMETER).
Bull. Amer. Meteor. Soc.
,
85
,
853
872
.
Pell
,
J. P.
, and
S. M.
Cobbe
,
1999
:
Seasonal variations in coronary heart disease.
Quart. J. Med.
,
92
,
689
696
.
Pope
,
V. D.
,
M. L.
Gallani
,
P. R.
Rowntree
, and
R. A.
Stratton
,
2000
:
The impact of new physical parametrizations in the Hadley Centre climate model: HadAM3.
Climate Dyn.
,
16
,
123
146
.
Qian
,
B.
, and
M. A.
Saunders
,
2003
:
Seasonal predictability of wintertime storminess over the North Atlantic.
Geophys. Res. Lett.
,
30
.
1698, doi:10.1029/2003GL017401
.
Quan
,
X. W.
,
P. J.
Webster
,
A. M.
Moore
, and
H. R.
Chang
,
2004
:
Seasonality in SST-forced atmospheric short-term climate predictability.
J. Climate
,
17
,
3090
3180
.
Rodwell
,
M. J.
, and
C. K.
Folland
,
2003
:
Atlantic air-sea interaction and model validation.
Ann. Geophys.
,
46
,
47
56
.
Shabbar
,
A.
,
H.
Jianping
, and
K.
Higuchi
,
2001
:
The relationship between the wintertime North Atlantic oscillation and blocking episodes in the North Atlantic.
Int. J. Climatol.
,
21
,
355
369
.
Stanski
,
H. R.
,
L. J.
Wilson
, and
W. R.
Burrows
,
1989
:
Survey of common verification methods in meteorology. World Meteorological Organisation, World Weather Watch Tech. Rep. 8, WMO/TD 358, 114 pp
.
Stern
,
P. C.
, and
W. E.
Easterling
,
1999
:
Making Climate Forecasts Matter.
National Academy Press, 175 pp
.
Thomson
,
M. C.
,
M.
Indeje
,
S. J.
Connor
,
M.
Dilley
, and
N.
Ward
,
2003
:
Malaria early warning in Kenya and seasonal climate forecasts.
Lancet
,
362
,
580
.
Toth
,
Z.
,
O.
Talagrand
,
G.
Candille
, and
Y.
Zhu
,
2003
:
Probability and ensemble forecasts.
Forecast Verification: A Practitioner’s Guide in Atmospheric Science, I. T. Jolliffe and D. B. Stephenson, Eds., John Wiley and Sons, 137–163
.
van Rossum
,
C. T. M.
,
M. J.
Shipley
,
H.
Hemingway
,
D. E.
Grobbee
,
J. P.
Mackenback
, and
M. G.
Marmot
,
2001
:
Seasonal variation in cause-specific mortality: Are there high-risk groups? 25-year follow-up of civil servants from the first Whitehall study.
Int. Epidemiol. Assoc.
,
30
,
1109
1116
.
WCRP
,
1995
:
CLIVAR—A study on Climate Variability and Predictability: Science plan World Climate Research Programme Tech. Note WCRP No. 89, WMO/TD No. 690, 172 pp
.
WCRP
,
1997
:
CLIVAR—A research programme on Climate Variability and Predictability for the 21st century. World Climate Research Programme Tech. Note WCRP No. 101, WMO/TD No. 853, ICPO No. 10, 48 pp
.
Wedgbrow
,
C. S.
,
R. L.
Wilby
,
H. R.
Fox
, and
G.
O’Hare
,
2002
:
Prospect for seasonal forecasting of summer drought and low river flow anomalies in England and Wales.
Int. J. Climatol.
,
22
,
219
236
.
Wilks
,
D. S.
,
1995
:
Statistical Methods in the Atmospheric Sciences. International Geophysics Series, Vol. 59, Academic Press, 464 pp
.
Woodhouse
,
P. R.
,
K.
Khaw
, and
M.
Plummer
,
1993
:
Seasonal variation of blood pressure and its relationship to ambient temperature in an elderly population.
J. Hypertens.
,
11
,
1267
1274
.

Footnotes

* Current affiliation: Benefield Group, Sydney, Australia

+ Current affiliation: Leeds University, Leeds, United Kingdom

Corresponding author address: Professor G. R. McGregor, Department of Geography, King’s College London, Strand, London WC2R 2LS London, United Kingdom. Email: glenn.mcgregor@kcl.ac.uk