|
ABSTRACT Francesca Dominici, Visiting Asst Prof of Biostatistics Measurement error is well recognized to be a potential source of bias on
the estimated association of the adverse effects of the air pollution on
health. Because air pollution is often measured using one or few ambient
monitors, these measurements are surrogate indexes of personal exposure
and can be weak predictors of it.
This paper presents hierarchical modeling strategies for evaluating the
effects of measurement error on estimates of mortality in time series
studies of air pollution, using five epidemiological studies on personal
and ambient concentrations of PM10. The studies present
irregularly
spaced observations, different personal and outdoor sampling of
particles, and different subject characteristics. Our goal is to combine
information across these heterogeneous studies to investigate the
association between average personal exposure and ambient
concentrations, and to estimate the size of bias in estimated
PM10-mortality regression coefficients due to using ambient
monitors rather than personal exposure data about PM10.
At the first stage of our hierarchical model, we introduce
study-specific longitudinal regressions of personal exposures versus
ambient concentrations of particles. Because exploratory analyses have
shown that the association between average personal exposure and ambient
concentration is higher in longest-term variations than shortest-term
variations, we replace the ambient concentration time series with its
decomposition in three time scales variations. At the second and third
stage of the model, we introduce distributions on the longitudinal
regressions parameters to borrow strength between subjects and between
studies respectively.
Finally, to enable the decomposition of ambient concentrations into
three different time scales, provisions must be made to account for
irregularly spaced observations. Under the assumption that ambient
concentrations follow an autoregressive time series model, we fill the
gaps of the missing observations by a Gibbs Sampler. Taking into
account of the heterogeneity across locations, the different sampling
schemes, and the missing data observations, we found that the
measurement error tends to bias the result toward the null hypothesis of
no effect, and that the precision of such estimates is generally
overstated.
Return to Longitudinal/Survival Working Group List | Return to Home Page |
|