Impact of preferential sampling on exposure prediction and health effect inference in the context of air pollution epidemiology.

Preferential sampling has been defined in the context of geostatistical modeling as the dependence between the sampling locations and the process that describes the spatial structure of the data. It can occur when networks are designed to find high values. For example, in networks based on the U.S. Clean Air Act monitors are sited to determine whether air quality standards are exceeded. We study the impact of the design of monitor networks in the context of air pollution epidemiology studies. The effect of preferential sampling has been illustrated in the literature by highlighting its impact on spatial predictions. In this paper, we use these predictions as input in a second stage analysis, and we assess how they affect health effect inference. Our work is motivated by data from two United States regulatory networks and health data from the Multi-Ethnic Study of Atherosclerosis and Air Pollution. The two networks were designed to monitor air pollution in urban and rural areas respectively, and we found that the health analysis results based on the two networks can lead to different scientific conclusions. We use preferential sampling to gain insight into these differences. We designed a simulation study, and found that the validity and reliability of the health effect estimate can be greatly affected by how we sample the monitor locations. To better understand its effect on second stage inference, we identify two components of preferential sampling that shed light on how preferential sampling alters the properties of the health effect estimate.