Impact of Exposure Misclassification in Air Pollution Epidemiology: Effect of Error Type

GRETCHEN T GOLDMAN (1), James A Mulholland (1), Armistead G Russell (1), Matthew J Strickland (2) and Paige E Tolbert (2)

(1) Georgia Institute of Technology, (2) Emory University

     Abstract Number: 143
     Last modified: November 6, 2009

Abstract
Introduction: Exposure measurement error is inherent to large population epidemiologic studies of ambient air pollution health effects and this error can have significant implications for interpretation of results. In ongoing time-series studies of ambient air pollution and emergency department (ED) visits for respiratory and cardiovascular diseases, we have been exploring the impact of measurement error on the log-linear regression of population-weighted average levels of twelve ambient air pollutants (NO$_2, NO$_x, SO$_2, CO, O$_3, PM$_(10), PM$_(2.5), and PM$_(2.5) components sulfate, nitrate, ammonium, organic carbon and elemental carbon) and daily ED counts through Monte Carlo simulation of error. The modeled measurement error is inclusive of instrument precision error and error resulting from the lack of correlation between monitors over space. The latter is the larger component of measurement error, particularly for primary air pollutants.

Methods: Two extremes in the conceptual framework of error type are classical error, in which measurement error is independent of the true value, and Berkson error, in which actual exposures vary independently about an average measured value. Our primary error model has been developed from collocated instrument data and geostatistical semivariogram analysis and includes both Berkson and classical error. Here, we employ two additional error models, a purely Berkson model and a purely classical error model, to investigate the effect of measurement error type on health risk assessment. To compare across pollutants, risk ratios are typically expressed per interquartile range (IQR) of the pollutant of interest; thus, we consider impacts of measurement error not only on the slope from Poisson regression but also on the IQR.

Results and Discussion: As expected, we found that classical error results in the largest attenuation of the slope, whereas Berkson error results in no attenuation of the slope estimate; our primary model provided results that were within the range of these extreme case values. Differences between models in the bias to the null per IQR were not so large, however, due to contrasting impacts of IQR on the bias. The ambient measured with error (from simulations) has a larger IQR than the true ambient (base case) in the classical error model whereas, in the Berkson error model, the population-weighted average ambient (base case) has a smaller IQR than the true individual ambient levels (from simulations). Average bias to the null estimates due to measurement error using the Berkson and classical models were 3.5% (1.7% standard deviation) and 4.8% (1.7%), respectively, for the secondary pollutant O$_3, 12.1% (3.8%) and 19.3% (3.3%) for PM$_(2.5) organic carbon of mixed primary and secondary origin, and 21.8% (5.3%) and 38.5% (4.5%) for the primary pollutant CO. Our primary error model results, incorporating both Berksonian and classical features, tend to yield bias to null results between these results. In light of these differences in measurement error associated with each pollutant, both magnitude and type of error should be considered in assessing the impact of measurement error in time-series epidemiologic studies of air pollution.