SEMINAR
Partitioning Evidence of Statistical Association: A Case Study of Air Pollution and Mortality
Holly Janes, Ph.D.
Johns Hopkins Bloomberg
School of Public Health
Wednesday, February 21, 2007
3:30pm, Moos 2-690
Minneapolis Campus
Abstract:
Spatio-temporal studies of the association between air pollution and mortality
are challenged by the strong effects of unmeasured confounders. Some examples
are seasonality and changes in government policy that cause trends in air pollution
and mortality over time. Motivated by two recent studies, the National Morbidity
and Mortality Air Pollution Study (NMMAPS) and the Medicare Air Pollution Study
(MCAPS), we propose decomposing the measure of association into orthogonal components,
roughly corresponding to distinct scales of spatio-temporal variation. We show
that typical regression models combine these components with a specific set
of weights. Viewing the constituents and their weights helps the scientist to
understand a given model and to compare competing models. Excess variation among
the components is evidence of the influence of unmeasured confounders acting
at particular spatio-temporal scales. We develop a test to detect unmeasured
confounding. Using NMMAPS data for the New York City elderly population, we
find very different evidence of association between air pollution and mortality
across several distinct temporal scales, suggesting the influence of unmeasured
confounders. Our analysis of the MCAPS data also reveals unmeasured confounding,
as evidenced by the different associations between trends in air pollution and
mortality at different spatial scales. Typical regression models combine these
very different components into a single summary measure. Our methods represent
a more transparent way of summarizing association, and motivate the development
of more confounding-robust measures of association. This approach is useful
more broadly for partitioning evidence of statistical association and testing
for confounding by omitted covariates in studies of exposures and outcomes that
vary in space and time.
A social tea will be held at 3:00 P.M. in A434 Mayo. All are Welcome.
For more details contact 612-624-4655 or see http://www.biostat.umn.edu/seminar_academic.html