Impact of Monitoring Network Design on Exposure Prediction and Measurement Error

• Activity Number: 351
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Adel Lee*+ and Lianne Sheppard
• Companies: and University of Washington, Department of Biostatistics
• Keywords: Preferential sampling ; Geostatistical modeling ; Air pollution monitoring ; Network design ; Measurement error

Abstract

Preferential sampling (PS) has been defined in the context of geostatistical modeling as dependence between the sampling locations and the process that describes the spatial structure of the data. PS can occur when networks are designed to find high values. For example, networks based on the US Clean Air Act monitor how often air quality standards are exceeded. The effect of PS has been illustrated in the literature most often by studying its impact on the fitted spatial model and the resulting prediction biases. In this work we show that PS also affects the measurement error that results from using predictions in a second stage analysis on the association between the predicted values and some outcome. We design a simulation study based on national monitoring data from the US Environmental Protection Agency. A universal kriging model is used to predict exposures and linear regression to model the outcome. We find that PS can greatly affect the validity of the estimate of the regression parameter of interest by inflating the measurement error. We identify conditions under which the universal kriging and linear regression model are most susceptible to the adverse affects of PS.