Climate Change and Vector-Borne Diseases in Europe
Climate Change and Vector-Borne Diseases in Europe
The dengue dataset was primarily developed for a study of the effects of weather on dengue incidence across Mexico. Dengue data comprised state-specific monthly reports of laboratory confirmed dengue cases, retrieved from the Mexican Health Secretariat (http://www.epidemiologia.salud.gob.mx/anuario/html/anuarios.html, last accessed June 2014) for the period January 1985 to December 2007. Monthly average minimum and maximum temperatures and monthly precipitation for each state were provided by the Mexican National Meteorological Service. Monthly mean humidity was retrieved from the National Centers for Environmental Prediction and National Center for Atmospheric Research (NCEP/NCAR) "Reanalysis 1" (http://www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis.pressure.html, last accessed June 2014). Yearly Gross domestic product (GDP) per capita (PPP in constant 2005 international dollars) was obtained from the World Bank at the national level (http://data.worldbank.org/country/mexico, last accessed June 2014). State-specific GDP estimates were computed as previously described. The proportion of people living in urban areas was retrieved from the Mexican Chamber of Deputies (http://www.cefp.gob.mx/intr/bancosdeinformacion/estatales/indicadores_socioeconomicos/is003.xls, last accessed June 2014). Population density was calculated by normalising population to state area size. Table 1 presents the summary statistics for these variables.
Generalized Additive Models (GAMs) are semi-parametric extensions of the generalized linear model (GLM), where the linear predictor ΣβjXj is replaced by a sum (hence the name additive) of smooth functions of covariates Σsj(Xj). Like in GLMs, GAMs allow the exploration of nonlinear data structures in the context of exponential family distributions (e.g. Poisson and Binomial), and use link functions to establish relationships between the mean of the outcome variable and the predictors. Unlike GLMs, GAMs automatically identify and estimate the optimal degree of nonlinearity of the model directly from the data. In our study, the expected number of dengue cases E(yti) ≡ μit for State i at time t was assumed to follow an overdispersed Poisson distribution described by:
where g(.) is a log link function of the expectation μit ≡ E(yti) with yti denoting the time series of dengue counts. The logarithm of the population (ξ) at time t and state i is included as an exposure variable to standardise the dengue data by population. Weather has a delayed effect on dengue incidence. Therefore, we specified our j-th meteorological variables Xjti within biologically and physically plausible time lags based on literature reports in Mexico. Weather variables comprised average monthly minimum (Tmin) and maximum (Tmax) temperatures, monthly precipitation (Precip) and average monthly relative humidity (Humid). All weather parameters were lagged 1 and 2 months (Tmin1:2, Tmax1:2, Precip1:2, Humid1:2). The term sj(.) corresponds to univariate smooth functions defined by penalized cubic regression splines. We adjusted our model for the effects of socioeconomic variables Zkit represented by GDP per capita, the proportion of the population living in urban settlements and population density. Socioeconomic variables entered the model linearly. Analyses were conducted in R version 2.15.0.
Many epidemiological datasets are likely to be dominated by long-term and seasonal trends. Therefore, adjusting the regression models for these patterns is necessary to separate them from the effects of weather parameters on the health variable. Our model does not account for seasonal trends as seasonality for Europe is unlikely to be similar to Mexico given the wider range of temperatures between summer and winter. Although mosquito presence is a key factor in the epidemiology and occurrence of the disease, to our knowledge there are no state-specific long-term time series of mosquito presence across Mexico. Consequently, data on mosquito presence could not be incorporated into our model. The GAM-estimated relationships between dengue, weather and socioeconomic development in Mexico were then used to project dengue fever risk across Europe.
European climate data were retrieved from the regional climate model COSMO-CLM (CCLM), forced with output from the coupled atmosphere–ocean global climate model (GCM) ECHAM5/MPIOM. These regional simulations represent aerosol and GHG forcing according to the A1B scenario of the Special Report on Emissions Scenarios (SRES) of the IPCC. A1B corresponds to a projected increase in global surface temperature of 2.8°C in 2090–2099 (relative to 1980–1999) and a likely range of up to 4.4°C. It assumes rapid economic growth, rapid introduction of efficient technologies, convergence among regions and a balance across energy sources. The regional climate data correspond to the period 1961–2100, with a domain covering the entire European continent at a resolution of about 18 × 18 km. Data were re-scaled to a grid cell size of 10 × 10 km for the purpose of this study. The same four monthly climatic variables (Tmin, Tmax, Precip, Humid) lagged 1 and 2 months, as used for model calibration with the Mexican data, were calculated over four time periods, (a) baseline 1961–1990, (b) short-term scenario 2011–2040, (c) medium-term scenario 2041–2070, and (d) long-term scenario 2071–2100.
GDP per capita data were retrieved from EUROSTAT (in Euros) and converted into constant 2005 international dollars to be concordant with the Mexican data used for model calibration. Country level data from the World Development Indicators dataset (http://databank.worldbank.org/data/Databases.aspx, last accessed June 2014) were disaggregated to NUTS-3 level (Nomenclature of territorial units for statistics) by using the NUTS-3 level shares for each country as calculated from EUROSTAT. Then, an areal weighting approach was employed to convert the NUTS-3 data into the 10 × 10 km grid (see Additional file 1 http://www.biomedcentral.com/1471-2458/14/781/additional). Areal weighting is commonly used to transform administrative boundary data to raster format, whereby each grid cell is assigned a value according to the percentage of its area covered by the overlying administrative region.
The proportion of population living in urban areas and total population data were retrieved from the GEOSTAT 2006 population grid dataset of the European Forum for Geostatistics (EFGS) (http://www.efgs.info/, last accessed June 2014) at a spatial resolution of 1 × 1 km. Urban clusters were defined by two criteria first each grid cell of 1 × 1 km must have a minimum population density of 300 people per km and second clusters of adjoining grid cells must accommodate at least 5000 people, in line with the definitions used by the European Commission. The total number of urban population for each 10 × 10 km grid cell was extracted and divided by total population to obtain proportion of population in urban area (see Additional file 1 http://www.biomedcentral.com/1471-2458/14/781/additional). Due to the lack of projections both in terms of SRES scenario and spatial detail, the socioeconomic variables were held constant at their mean value for baseline conditions in order to isolate the effects of climate.
The model was used to project monthly dengue cases, which were aggregated to calculate the average number of cases per year for each time period. These were used to generate dengue risk maps using ArcGIS 10.1 (http://www.esri.com/software/arcgis, last accessed June 2014). In total, four maps were produced corresponding to the time periods of study. Identical class sizes were applied across all four time periods in order to ensure that value changes could be observed over time. The map series employ a bipolar hue progression ranging from green (no/low risk areas) to bright red (areas with the highest dengue risk). Moreover, a second map series was generated, that normalises dengue cases by total population to derive dengue incidence. We used colours ranging from blue for no/low risk to cherry brown for high risk areas. In addition, standard error for each grid cell was calculated. Standard error values were subjected to the same aggregation and averaging procedure as for dengue number of cases and dengue incidence and were used to produce maps of uncertainty.
Methods
Mexican Data
The dengue dataset was primarily developed for a study of the effects of weather on dengue incidence across Mexico. Dengue data comprised state-specific monthly reports of laboratory confirmed dengue cases, retrieved from the Mexican Health Secretariat (http://www.epidemiologia.salud.gob.mx/anuario/html/anuarios.html, last accessed June 2014) for the period January 1985 to December 2007. Monthly average minimum and maximum temperatures and monthly precipitation for each state were provided by the Mexican National Meteorological Service. Monthly mean humidity was retrieved from the National Centers for Environmental Prediction and National Center for Atmospheric Research (NCEP/NCAR) "Reanalysis 1" (http://www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis.pressure.html, last accessed June 2014). Yearly Gross domestic product (GDP) per capita (PPP in constant 2005 international dollars) was obtained from the World Bank at the national level (http://data.worldbank.org/country/mexico, last accessed June 2014). State-specific GDP estimates were computed as previously described. The proportion of people living in urban areas was retrieved from the Mexican Chamber of Deputies (http://www.cefp.gob.mx/intr/bancosdeinformacion/estatales/indicadores_socioeconomicos/is003.xls, last accessed June 2014). Population density was calculated by normalising population to state area size. Table 1 presents the summary statistics for these variables.
Model Calibration
Generalized Additive Models (GAMs) are semi-parametric extensions of the generalized linear model (GLM), where the linear predictor ΣβjXj is replaced by a sum (hence the name additive) of smooth functions of covariates Σsj(Xj). Like in GLMs, GAMs allow the exploration of nonlinear data structures in the context of exponential family distributions (e.g. Poisson and Binomial), and use link functions to establish relationships between the mean of the outcome variable and the predictors. Unlike GLMs, GAMs automatically identify and estimate the optimal degree of nonlinearity of the model directly from the data. In our study, the expected number of dengue cases E(yti) ≡ μit for State i at time t was assumed to follow an overdispersed Poisson distribution described by:
where g(.) is a log link function of the expectation μit ≡ E(yti) with yti denoting the time series of dengue counts. The logarithm of the population (ξ) at time t and state i is included as an exposure variable to standardise the dengue data by population. Weather has a delayed effect on dengue incidence. Therefore, we specified our j-th meteorological variables Xjti within biologically and physically plausible time lags based on literature reports in Mexico. Weather variables comprised average monthly minimum (Tmin) and maximum (Tmax) temperatures, monthly precipitation (Precip) and average monthly relative humidity (Humid). All weather parameters were lagged 1 and 2 months (Tmin1:2, Tmax1:2, Precip1:2, Humid1:2). The term sj(.) corresponds to univariate smooth functions defined by penalized cubic regression splines. We adjusted our model for the effects of socioeconomic variables Zkit represented by GDP per capita, the proportion of the population living in urban settlements and population density. Socioeconomic variables entered the model linearly. Analyses were conducted in R version 2.15.0.
Many epidemiological datasets are likely to be dominated by long-term and seasonal trends. Therefore, adjusting the regression models for these patterns is necessary to separate them from the effects of weather parameters on the health variable. Our model does not account for seasonal trends as seasonality for Europe is unlikely to be similar to Mexico given the wider range of temperatures between summer and winter. Although mosquito presence is a key factor in the epidemiology and occurrence of the disease, to our knowledge there are no state-specific long-term time series of mosquito presence across Mexico. Consequently, data on mosquito presence could not be incorporated into our model. The GAM-estimated relationships between dengue, weather and socioeconomic development in Mexico were then used to project dengue fever risk across Europe.
European Data and Dengue Fever Risk Modelling
European climate data were retrieved from the regional climate model COSMO-CLM (CCLM), forced with output from the coupled atmosphere–ocean global climate model (GCM) ECHAM5/MPIOM. These regional simulations represent aerosol and GHG forcing according to the A1B scenario of the Special Report on Emissions Scenarios (SRES) of the IPCC. A1B corresponds to a projected increase in global surface temperature of 2.8°C in 2090–2099 (relative to 1980–1999) and a likely range of up to 4.4°C. It assumes rapid economic growth, rapid introduction of efficient technologies, convergence among regions and a balance across energy sources. The regional climate data correspond to the period 1961–2100, with a domain covering the entire European continent at a resolution of about 18 × 18 km. Data were re-scaled to a grid cell size of 10 × 10 km for the purpose of this study. The same four monthly climatic variables (Tmin, Tmax, Precip, Humid) lagged 1 and 2 months, as used for model calibration with the Mexican data, were calculated over four time periods, (a) baseline 1961–1990, (b) short-term scenario 2011–2040, (c) medium-term scenario 2041–2070, and (d) long-term scenario 2071–2100.
GDP per capita data were retrieved from EUROSTAT (in Euros) and converted into constant 2005 international dollars to be concordant with the Mexican data used for model calibration. Country level data from the World Development Indicators dataset (http://databank.worldbank.org/data/Databases.aspx, last accessed June 2014) were disaggregated to NUTS-3 level (Nomenclature of territorial units for statistics) by using the NUTS-3 level shares for each country as calculated from EUROSTAT. Then, an areal weighting approach was employed to convert the NUTS-3 data into the 10 × 10 km grid (see Additional file 1 http://www.biomedcentral.com/1471-2458/14/781/additional). Areal weighting is commonly used to transform administrative boundary data to raster format, whereby each grid cell is assigned a value according to the percentage of its area covered by the overlying administrative region.
The proportion of population living in urban areas and total population data were retrieved from the GEOSTAT 2006 population grid dataset of the European Forum for Geostatistics (EFGS) (http://www.efgs.info/, last accessed June 2014) at a spatial resolution of 1 × 1 km. Urban clusters were defined by two criteria first each grid cell of 1 × 1 km must have a minimum population density of 300 people per km and second clusters of adjoining grid cells must accommodate at least 5000 people, in line with the definitions used by the European Commission. The total number of urban population for each 10 × 10 km grid cell was extracted and divided by total population to obtain proportion of population in urban area (see Additional file 1 http://www.biomedcentral.com/1471-2458/14/781/additional). Due to the lack of projections both in terms of SRES scenario and spatial detail, the socioeconomic variables were held constant at their mean value for baseline conditions in order to isolate the effects of climate.
Mapping of Dengue Fever Risk
The model was used to project monthly dengue cases, which were aggregated to calculate the average number of cases per year for each time period. These were used to generate dengue risk maps using ArcGIS 10.1 (http://www.esri.com/software/arcgis, last accessed June 2014). In total, four maps were produced corresponding to the time periods of study. Identical class sizes were applied across all four time periods in order to ensure that value changes could be observed over time. The map series employ a bipolar hue progression ranging from green (no/low risk areas) to bright red (areas with the highest dengue risk). Moreover, a second map series was generated, that normalises dengue cases by total population to derive dengue incidence. We used colours ranging from blue for no/low risk to cherry brown for high risk areas. In addition, standard error for each grid cell was calculated. Standard error values were subjected to the same aggregation and averaging procedure as for dengue number of cases and dengue incidence and were used to produce maps of uncertainty.