The Health Effects Institute (HEI) has
recently released an announcement that Johns Hopkins University investigators
of the National Morbidity, Mortality and Air Pollution Study (NMMAPS) have
updated their previous estimates of the mortality effects of acute exposure to
particulate air pollution. (See http://www.healtheffects.org/Pubs/NMMAPSletter.pdf) Below find frequently asked questions and answers
provided by the Hopkins NMMAPS team that help to better understand the issues
involved. More details are available from the web-site: http://www.biostat.jhsph.edu/biostat/research/update.main.htm What
is the concern? The
team carrying out the National Morbidity, Mortality and Air Pollution Study or
NMMAPS has published estimates of the health effects of air pollution that were
obtained by application of a method called generalized additive models (GAMs)
implemented in the S-plus statistical software. A GAM is a valuable methodology
for estimating an effect of one or more variables (here, air pollution, weather
and time) on an outcome (here, daily deaths) when we cannot assume that the
relationship takes a particular functional (e.g., linear) form. Because of its
flexibility, it is widely used in air pollution research and other
applications. The GAM estimation procedure, as implemented in standard
statistical software (as, for example, in S-PLUS , Stata, R and SAS) relies
upon default convergence criteria. We accepted the default convergence criteria
in the S-PLUS version 3.4 and recently discovered that they were inadequate to
produce optimal estimates and could introduce upward bias. We have re-done
our analyses with more stringent convergence criteria for the GAM estimation
procedure and found that estimates for individual cities changed by small
amounts and that the estimate of the average particulate pollution effect
across the 90 largest U.S. cities changed from a 0.41% increase to a 0.27%
increase in daily mortality per 10 micrograms per cubic meter of PM10. As an independent verification, we have
also used a similar parametric model (GLM) and distinct software
(glm in S-PLUS 3.4) and obtained a pooled estimate of 0.22%, similar to the
value from GAM. See (http://www.healtheffects.org/Pubs/NMMAPSletter.pdf)
for further discussion. Do the NMMAPS conclusions
change? The quantitative estimates
change, but the major
conclusions do not.
Each of
these findings is unchanged in our re-analysis using the stricter convergence
criteria. But
doesn’t the decrease from 0.41% to 0.27% represent a 35% decrease in your
assessment of the effect of PM10 on mortality? Yes, this
is one way to look at our result. However, the implications may be less than is
implied by a 35% reduction in the relative rate. Although we use relative changes to measure the association of
air pollution with mortality, for public health purposes, it is the change in
absolute risk associated with pollution that is more important. In our case,
the change in risks due to implementing the improved algorithm is 0.41-0.27=0.14%
per 10 micrograms per m3 of PM10. To
understand these changes, consider the city of Baltimore where there are
roughly 20 deaths per day or 7,300 deaths per year. If we could reduce PM10
in Baltimore from the current average value of 35 down to 25 micrograms per m3,
our prior estimate of 0.41% corresponds to saving 30 lives per year from the
acute effects alone. Our updated estimate would correspond to 20 deaths, a
change of 10 deaths per year. Having heard so
much about the health effects of particulate air pollution from the news, I am
surprised that there are only 20 deaths per year attributable to particle
exposure in a city the size of Baltimore.
The
20 deaths are attributable to acute exposure only. Time
series studies like NMMAPS compare mortality within the same population on
higher versus lower pollution days and can therefore only estimate the effect
of shorter-term elevations in pollution. But persons in Baltimore and other
cities are also chronically exposed, that is exposed day-in and day-out. Other
long-term “cohort” studies estimate the combined effects of chronic and acute
exposure by comparing rates of death across U.S. cities, statistically
controlling for personal characteristics such as age and smoking. The major U.S. cohort
studies are the American Cancer Society Study (Pope
et al. 2002) and the Six Cities Studies (Dockery
et al, 1993). They estimate an increase in total mortality of roughly 4%
and 5% per 10 microgram per m3 increase in the long-term level of particles,
respectively. This is an order of magnitude higher than found in time
series studies. Cohort studies also have been used to predict the number
of lives saved from a pollution reduction program since both acute and chronic
exposures will change. Then why are
the time series studies useful if they only estimate the acute effects and the
cohort studies can estimate the combined acute and chronic effects? The
time series studies like NMMAPS contribute important information in identifying
whether particles acutely cause illness and death, presumably because persons
with underlying heart and lung disease are at risk. By
comparing mortality from day to day within the same population, time series
studies are less subject to “ecologic bias”. The time series studies also
provide evidence relevant to scientific questions that support a causal
relationship of particles with mortality including: the effects of
co-pollutants, cause-of-death-specific pollution effects, exposure measurement
error, existence of thresholds, and geographic variations in the pollution
effects that might point toward the toxic component of the particles. How did you
discover the convergence problem with the GAM function in S-plus? As a
follow up to our main NMMAPS analyses, we continued to explore the
sensitivity of the results to several alternative modeling approaches. In doing
so, we came across a situation where the software package unexpectedly
calculated the same relative rate estimates for particulate pollution even
though we were substantially changing the adjustment for confounding factors. This hint led us to examine the
implementation of the algorithm line-by-line and to the finding that the
default convergence criteria were not adequate for our problem. Using
default settings has been the standard of practice in environmental
epidemiology until we started investigating this issue. Does this
mean the all time series studies of air pollution will have biased results? No. First, the convergence problem only
occurs if two or more smooth curves are in the GAM. In our case, we used smooth
curves for time, temperature, and dew point temperature. Furthermore, the size
of the bias depends on two key factors: (1) the size of the log-relative rate;
and (2) the degree of correlation between the pollution variable and the smooth
functions of the other confounders. We cannot predict the direction or size of
the bias, but it tends to be relatively more important for smaller effects when
the correlation above is larger. |