Effects of exposure measurement error in the analysis of health effects from traffic-related air pollution

Abstract

In large epidemiological studies, many researchers use surrogates of air pollution exposure such as geographic information system (GIS)-based characterizations of traffic or simple housing characteristics. It is important to evaluate quantitatively these surrogates against measured pollutant concentrations to determine how their use affects the interpretation of epidemiological study results. In this study, we quantified the implications of using exposure models derived from validation studies, and other alternative surrogate models with varying amounts of measurement error on epidemiological study findings. We compared previously developed multiple regression models characterizing residential indoor nitrogen dioxide (NO(2)), fine particulate matter (PM(2.5)), and elemental carbon (EC) concentrations to models with less explanatory power that may be applied in the absence of validation studies. We constructed a hypothetical epidemiological study, under a range of odds ratios, and determined the bias and uncertainty caused by the use of various exposure models predicting residential indoor exposure levels. Our simulations illustrated that exposure models with fairly modest R(2) (0.3 to 0.4 for the previously developed multiple regression models for PM(2.5) and NO(2)) yielded substantial improvements in epidemiological study performance, relative to the application of regression models created in the absence of validation studies or poorer-performing validation study models (e.g., EC). In many studies, models based on validation data may not be possible, so it may be necessary to use a surrogate model with more measurement error. This analysis provides a technique to quantify the implications of applying various exposure models with different degrees of measurement error in epidemiological research.

Figure 1. Distribution of estimated odds ratios per interquartile increase in concentration using various models of simulated indoor NO₂ concentrations given different true odds ratios

^a gold standard is defined as the simulated indoor concentrations white boxes: True OR = 1.05 cross-hatch boxes: True OR = 1.50 grey boxes: True OR = 2.00 solid line = median, boxes = interquartile range, whiskers = 10^th and 90^th percentiles

Figure 2. Distribution of estimated odds ratios per interquartile increase in concentration using various models of simulated indoor PM_2.5 concentrations given different true odds ratios

^a gold standard is defined as the simulated indoor concentrations white boxes: True OR = 1.05 cross-hatch boxes: True OR = 1.50 grey boxes: True OR = 2.00 solid line = median, boxes = interquartile range, whiskers = 10^th and 90^th percentiles

Figure 3. Distribution of estimate odds ratios per interquartile increase in concentration using various models of simulated indoor EC concentrations given different true odds ratios

^a gold standard is defined as the simulated indoor concentrations white boxes: True OR = 1.05 cross-hatch boxes: True OR = 1.50 grey boxes: True OR = 2.00 solid line = median, boxes = interquartile range, whiskers = 10^th and 90^th percentile

Figure 4. Power to detect positive associations under a surrogate model vs. a gold standard given different R² thick solid line: R²

thick dotted-dashed line: R² = 0.90, thick dashed line: R² = 0.75, thin dotted-dashed line: R² = 0.42^a, dotted line: R² = 0.28^b, thin dashed line: R² = 0.10, thin solid line: R² = 0.05^c