A rigorous statistical framework for estimating the long-term health effects of air pollution

暴露于空气污染造成的不利健康影响在全世界都是众所周知的，并对财政和公共卫生产生重大影响。例如，在英国，空气污染估计会使预期寿命减少6个月，相应的健康成本每年高达190亿英镑。历届英国政府都采取行动，通过立法（例如，英国空气质量战略，2007年）和设立空气污染医学影响委员会来调解空气污染的有害影响。在过去30年中，为评估空气污染对健康的影响进行了大量流行病学研究，其中大多数研究集中在几天高浓度的影响上。很少有研究集中在长期接触的影响上，这种影响可以通过比较多年来居住在小地理区域（如选区）的人口的污染水平和健康状况来评估。然而，进行这样的研究是一项复杂的任务，重要的是流行病学家能够使用适当的统计方法来准确量化污染对健康的影响。特别是，许多因素将影响健康状况不佳的格局，包括污染程度和社会经济因素。然而，后者中的一些将是未知的或无法测量的，它们的存在将导致健康数据的时空相关性。这种相关性很可能在空间上是局部的，因为地理上相邻地区的健康状况不佳程度的相似性将取决于生活在这些地区的人口之间的相似性。不考虑这些未知因素将有可能使估计的污染与健康关系产生偏差，因此本项目的主要挑战之一是制定一种统计方法来解决这一问题。另一个关键挑战将是准确估计每个地方人口和年份的个别空气污染物水平。这是困难的，因为污染的时空格局往往是由大气过程驱动的，而大气过程本身又受气象过程的影响。更复杂的是，这些过程通常彼此具有非线性影响，这就排除了在实践中经常采用的线性插值方法的使用。多种污染物和整体空气质量对健康的影响也知之甚少，因为量化这些影响需要多变量污染模型，难以拟合和分析。该项目将使用英国气象局开发的最先进的气象、气候和空气质量模型，以产生可靠的空气污染估计。这个项目的主要目的是为解决这些问题的健康和污染数据建立和测试一个单一的综合模型，从而能够估计总体空气污染对健康的影响。第二个目的是量化忽略这些问题对估计的污染-健康关系的影响（偏差）。健康模型将需要在小区域健康数据中提供局部时空相关性的准确表示，而污染模型将需要根据需要，在任何空间和时间分辨率上提供对个别污染物和整体空气质量的不确定性的估计和测量，以与健康数据保持一致。重要的是，使用正式的统计框架使我们能够进行进一步的创新：即衡量气候变化对健康和空气污染的影响。这将通过利用确定性气候模式的输出来预测未来气候条件下的空气污染水平，并在综合健康和污染模式中使用这些预测水平来实现。总的来说，该提案通过开发和测试一套新的统计模型，概述了迄今为止尝试过的最详细的健康和空气污染数据之间的联系。

The adverse health effects resulting from exposure to air pollution are well known across the world, and have a substantial financial and public health impact. For example, in the UK air pollution is estimated to reduce life expectancy by 6 months, with corresponding health costs of up to £19 billion each year. Successive UK governments have acted to mediate against the harmful effects of air pollution, by introducing legislation (e.g. UK Air Quality Strategy, 2007), and setting up the Committee On the Medical Effects of Air Pollution. Numerous epidemiological studies have been conducted to assess the health impact of air pollution over the last 30 years, most of which have focused on the effects of a few days of high concentrations. Much less research has focused on the effects of long-term exposure, which can be assessed by comparing the levels of pollution and ill health in populations living in small geographical regions, such as electoral wards, over a number of years.However, conducting such a study is a complex task, and it is important that epidemiologists have access to appropriate statistical methods to accurately quantify the health impact of pollution. In particular, numerous factors will affect the pattern in ill health over space and time, including pollution levels and socio-economic factors. However, some of the latter will be unknown or unmeasured, and their existence will induce spatio-temporal correlation into the health data. This correlation is likely to be localised in space, as the similarity of the levels of ill health in geographically adjacent areas will depend on the similarity between the populations living in those areas. To not account for these unknown factors will risk biasing the estimated pollution-health relationship, and thus one of the key challenges of this project is to develop a statistical approach to address this issue. The other key challenge will be to accurately estimate the levels of individual air pollutants for each local population and year. This is difficult, because the spatio-temporal pattern in pollution is often driven by atmospheric processes, which themselves are influenced by meteorological processes. A further complication is that, often, these processes have non-linear effects on each other, which excludes the use of linear interpolation methods often adopted in practice. The effects of multiple pollutants and overall air quality on health are also poorly understood, as quantifying these requires multivariate pollution models which are hard to fit and analyse. This project will use state-of-the-art meteorological, climate and air quality models developed by the Met Office to produce reliable air pollution estimates.The main aim of this project is to create and test a single integrated model for health and pollution data that addresses these issues, thus allowing the effects of overall air pollution on health to be estimated. A secondary aim is to quantity the impact (bias) that ignoring these issues has on the estimated pollution-health relationship. The health model will need to provide an accurate representation of the localised spatio-temporal correlation in small-area health data, while the pollution model will need to provide estimates and measures of uncertainty for individual pollutants and overall air quality at any spatial and temporal resolution, as required to align with the health data. Importantly, the use of a formal statistical framework allows us to make a further innovation: namely, to measure the effect of climate change on health and air pollution. This will be achieved by using output from deterministic climate models to project air pollution levels under future climate conditions, and using those projected levels in the integrated health and pollution model. Overall, this proposal outlines the most detailed linkage of health and air pollution data yet attempted, by developing and testing a set of novel statistical models.