Title: Exposure Assessment for Health Effect Studies: Insights from Air Pollution Epidemiology
1Exposure Assessment for Health Effect Studies
Insights from Air Pollution Epidemiology
Lianne Sheppard University of Washington Special
thanks to Sun-Young Kim, Adam Szpiro
2Background
- Most epidemiological studies assess the effects
of an exposure on a disease outcome by estimating
a regression parameter (e.g. relative risk) - Models condition on exposure
- A complete set of pertinent exposure measurements
typically are not available - gt Need to use an approach to estimate (predict)
exposure - Health results are affected by the quality of the
exposure estimate - Exposure assessment for epidemiology should be
evaluated in the context of the health effect
estimation goal
3Typical approach
- Estimate or predict exposure as accurately as
possible - Plug in exposure estimates into a health model
estimate health effects - Challenges
- Health effect estimate is affected by the nature
and quality of the exposure assessment approach - Health effect estimate may be
- Biased
- More variable
- Typical analysis does not account for uncertainty
in exposure prediction gt inference not correct
4Goals
- Advance understanding of environmental and
occupational exposure assessment for use in
epidemiological research - Focus on the air pollution epidemiology
application because certain features of exposure
may be better understood than other applications
5Air pollution exposure framework
- Hypothesized personal exposure model
- Personal exposure EP ambient source (EA)
non-ambient source (EN) - EA ambient concentration (CA) a
- Ambient concentration occurs both outdoors and
indoors due to the infiltration of ambient
pollution into indoor environments - a f o(1-f o)Finf is the ambient exposure
attenuation factor - Ambient attenuation is a weighted average of
infiltration (Finf), weighted by time spent
outdoors (f o) - Exposure of interest long-term ambient source
(EA)
6Air pollution exposure assessment for
ambient-source long-term exposure
- Individual time-activity and building specific
infiltration information typically not available - Ambient concentration data are readily available
from EPA - Limited number of fixed locations
- Rich in time (often daily or hourly measurements)
- Monitor siting criteria are pollutant dependent
for some pollutants monitors are sited away from
sources - Even with rich data, models are needed to predict
concentrations at locations without monitors - Collection of additional concentration data to
better predict spatially varying concentrations
should focus on representing - Design space
- Geographic space
7Air pollution concentration prediction
- Spatially varying concentrations are typically
predicted using - Land use regression
- Kriging or other spatial smoothing approach
- Nearest monitor
- Air pollution concentration modeled using
universal kriging includes - Mean model (design space)
- Geographically defined (spatially varying)
covariates Land use regression - New covariates derived from physically-based
deterministic models - Variance model (geographic space)
- Spatial smoothing
8Spatial Correlation Structure
- Variance model recognizes that nearby residuals
are correlated - Example Exponential geostatistical variogram
model
Incorporating spatial correlation into the model
will improve spatial predictions
9Concentration prediction comments
- With limited concentration data, a
spatio-temporal model is needed for air pollution
concentration data - The success of a prediction model depends upon
- The structure in the underlying exposure surface
- The availability of data to capture this
structure
10Need For Spatio-Temporal Model
AQS Monitor in Azusa (060370002)
AQS Monitor in Long Beach (060371301)
Log NOx (ppb)
Home Outdoor Monitor in Long Beach (notional)
11Examples of spatial surfaces
- Spatial surface of five exposure models (lighter
higher concentration)
12Plug-in exposure health effect estimates
- Predicted exposure is used as the covariate in
the health effect regression model - The quality of the exposure model affects the
quality of health effect estimates - Exposures that can be predicted well (e.g. those
with large-scale spatial structure) yield health
effect estimates with good properties regardless
of prediction approach - Less predictable exposure surfaces yield health
effect estimates with poorer properties - Attenuation bias
- Large standard errors
13Plug-in exposure health effect estimates
Exposure with structure captured by the
predictions
True exposure vs. nearest monitor
True exposure vs. kriged
14Plug-in exposure health effect estimates
Exposure with little structure captured by the
predictions
True exposure vs. nearest monitor
True exposure vs. kriged
15Health effect estimates
- Applications Note that comparing results from
different exposure predictions gives only one
realization of the relationship between health
effect estimates - This is very limited information
16Application example
- Relative risk of detectable aortic calcium for a
10 ug/m3 increase in PM2.5 (Allen et al 2009) - Kriged exposure 1.06 (.96, 1.16)Nearest
monitor 1.05 (.96, 1.15)
17Comments about health effect estimates
- Even with true (known) exposures, health effect
estimates have uncertainty - Uncertainty of health effect estimates increases
as predicted exposure becomes more smooth (less
variable) - Predictions (modeled exposures) only represent a
fraction of the variation in true exposure - Health effect estimates can be evaluated by
assessing their - Bias
- Variance
- Coverage Percent of 95 confidence intervals
that cover the true value
- or Mean square error (variance bias2)
18Health effect estimates example
19Conclusions
- Capture as much of the pertinent underlying
exposure variation as possible in the exposure
model - Health effect estimate is affected by the nature
and quality of the exposure assessment approach
20Health effect estimates for exposures other than
air pollution
-
- Is the underlying exposure framework clear?
- Challenges predicting exposure
- Less data (no existing regulatory monitoring
network) - Cant capture complex structures (such as
spatio-temporal variation) - How well do exposure data represent individuals
with no data? - Many sources of variation, often without
measurable determinants
21Comments
- Study design is a critical feature
- Linkage between the design, the key aspects of
exposure, and the pertinent health outcome? - Does the design focus on spatial variation
(cohort studies) or temporal variation (time
series studies)? - Multiple testing and potential for reporting bias
- Evaluation of multiple exposure prediction
approaches is yet another opportunity for
epidemiologists to cherry-pick results - Predictions are more smooth than data
- gt decreased exposure variation in health analyses
22Research needs
- What are the important exposure features to
capture for health effect estimates? Consider - Sources of variation in underlying true exposure
and their relevance for the health outcome - Study design
- Exposure data that are feasible to collect
- Alignment of these features
- How many exposure measurements are needed?
- Exposure data are often much more limited than
health data - What are the best inputs to the exposure models?
- Approaches to health effect estimation to give
good inference - Good coverage 95 CI covers the true value 95
of the time