Developing a heatwave early warning system for Sweden: evaluating sensitivity of different epidemiological modelling approaches to forecast temperatures

Abstract

Over the last two decades a number of heatwaves have brought the need for heatwave early warning systems (HEWS) to the attention of many European governments. The HEWS in Europe are operating under the assumption that there is a high correlation between observed and forecasted temperatures. We investigated the sensitivity of different temperature mortality relationships when using forecast temperatures. We modelled mortality in Stockholm using observed temperatures and made predictions using forecast temperatures from the European Centre for Medium-range Weather Forecasts to assess the sensitivity. We found that the forecast will alter the expected future risk differently for different temperature mortality relationships. The more complex models seemed more sensitive to inaccurate forecasts. Despite the difference between models, there was a high agreement between models when identifying risk-days. We find that considerations of the accuracy in temperature forecasts should be part of the design of a HEWS. Currently operating HEWS do evaluate their predictive performance; this information should also be part of the evaluation of the epidemiological models that are the foundation in the HEWS. The most accurate description of the relationship between high temperature and mortality might not be the most suitable or practical when incorporated into a HEWS.

Figure 1

Observed vs. 1-day forecast temperatures, Stockholm, 1998–2007. Red line shows the linear estimate for an unbiased fit and the blue the estimated linear relationship between the variables.

Figure 2

Comparison of which models identify heatwave days as risk days. Each grey area represents a period where risk days were identified. The markers describe which models classified each day as a risk day. The date on the x-axis describes the first day of each period with elevated mortality risk.

Figure 3

Risk estimates for the four models using fitted values from observed temperatures and the forecast temperatures on days with temperature above 26 °C. On the x-axis is the estimated risk increases in percent produced by observed temperatures and on the y-axis the estimated risk increase produced by different forecast times for each day in the study period for the: (a) DLNM model; (b) GAM model; (c) PWL model; (d) THR model.

Figure 3

Risk estimates for the four models using fitted values from observed temperatures and the forecast temperatures on days with temperature above 26 °C. On the x-axis is the estimated risk increases in percent produced by observed temperatures and on the y-axis the estimated risk increase produced by different forecast times for each day in the study period for the: (a) DLNM model; (b) GAM model; (c) PWL model; (d) THR model.

Figure 3

Risk estimates for the four models using fitted values from observed temperatures and the forecast temperatures on days with temperature above 26 °C. On the x-axis is the estimated risk increases in percent produced by observed temperatures and on the y-axis the estimated risk increase produced by different forecast times for each day in the study period for the: (a) DLNM model; (b) GAM model; (c) PWL model; (d) THR model.