全球变化知识资源中心

Data sources and quality.

An initial search of the Web of Knowledge was carried out to identify factors influencing EWD rates. The search encompassed all years and excluded non-English-language articles. Many combinations of search terms were explored (Supplementary Information). Secondary searches were carried out on references cited by articles discussing EWDs and their causes.

Articles relating to the causes of EWDs were identified as targets for data retrieval. These were supplemented by source data: temperature data were obtained from the UK’s Meteorological Office–Hadley Centre Central England Temperature data set³⁰; and economic, social, population and mortality data from the Office for National Statistics. When the required data were unavailable, such as those relating to housing quality or to government initiatives to combat EWDs, a web search using Google was initiated, which was focused on information held in government departments, agencies and other organizations holding specialist housing data sets (Supplementary Information).

Data were collected for the period from 1951 to 2011 and a full list of all data sources used is provided in Supplementary Table 1. Population (and a subset aged over 65 years old), excess winter mortality and incidence of influenza-like illness in England and Wales were documented. Daily mean temperatures for central England were also obtained. These data, being representative of a national geographical mean³⁰, were sufficient for our study of national trends. We therefore collected demographic, EWD, influenza incidence and winter temperature data representative of England and Wales. Certainly different regions will experience different temperatures, but they are all highly correlated to this central England value. Also, EWDs are surprisingly stable across regions, for example with the number of EWDs in Cornwall, the mildest part of England, being nearly identical to that for England and Wales.

A range of extrinsic factors influencing seasonal, temperature-related mortality were also documented. Specific factors were excluded only when their impacts or characteristics were already represented within another factor (for example, income level versus percentage household expenditure on fuel). We recorded expenditure on heating as a percentage of income, policy initiatives aimed at combating EWDs (cold weather payment, winter fuel allowance, warm front) and four key housing quality factors, each focused on a particular housing characteristic affecting health in winter (availability of inside toilet, access to tapped hot water, central heating and double glazing). Data were drawn from the domestic energy fact file, the UK housing energy fact file, the English housing survey and the Halifax housing data set, which contains a reliable description of the changing condition of the UK housing stock. Housing quality improvements were assumed to be linear between available years. This assumption was validated for double glazing, where data for the full period were available. For statistical analyses, all four housing quality measures were combined by simple averaging into a single measure.

Statistical analyses.

EWDs were expressed as a function of population over 65 years old (as about 90% of total EWDs occur in this age group⁹) to remove changing demographics as a factor. The raw daily temperature data were transformed into two measures: number of days per winter period below 5 °C (a measure of winter cold intensity); and number of days per winter below 5 °C and showing a 4 °C drop from the previous day (a measure of volatility within cold spells, defined as one or more days below 5 °C). Cold days were calculated for the same period as EWDs, namely 1 December to 31 March. Correlation analysis was used to determine the interdependence of variables. Linear multifactor regression analysis identified those factors associated independently with EWDs. This was achieved by carrying out a series of regressions removing the least significant factor at each repeat until only highly significant factors remained. We also ensured that the amount of variation explained by the fitted model, R-square, remained relatively stable. To explain the trend over time in more detail, a moving average method was employed with a period of ten years. This allowed the smoothing of the data to eliminate most of the year-to-year variation. Linear multifactor regression analysis was applied to the smoothed data. To assess year-to-year variation, the relevant data were detrended by removal of the time component and analysed by linear multifactor regression. Using separate data sets, that is, a smoothed data set and a detrended data set, allowed the elimination of confounding time-dependent factors when addressing the two specific questions of: what is causing the long-term trend in decreasing EWDs; and what is causing the short-term year-to-year variation in EWDs. We also tested for correlation breakdown between EWDs and the factors with strong year-to-year variation. To further explain factors associated with EWDs, data were split into subsets characterized by key changes in factors, or the introduction of new factors (such as a policy initiative).