The Impact of Air Pollution on Infant Mortality: Evidence from
Geographic Variation in Pollution Shocks Induced by a Recession

Kenneth Chay and Michael Greenstone

ABSTRACT

	Previous research has documented a statistical association
between differential pollution levels across sites and variation in
adult health outcomes.  However, a heated debate has arisen about
whether these documented associations are causal.  At issue is the
possibility that since air pollution, and pollution changes, are not
randomly assigned across localities, previous studies may not be
adequately controlling for a number of potential confounding
determinants of mortality.  In this study, we use sharp, differential
air quality changes across sites attributable to geographic variation
in the effects of the 1981-82 recession to estimate the relationship
between infant mortality and suspended particulates air pollution.  It
is shown that in the narrow period of 1980-82, there was substantial
variation across counties in changes in particulates pollution, and
that these differential pollution reductions appear to be orthogonal
to changes in other factors which may be correlated with infant
mortality.  Thus, this research design may provide a more credible
basis for evaluating the pollution-infant mortality relationship than
those previously used.  The analysis is based on the most detailed and
comprehensive data available on air pollution, infant births and
deaths, and other potential determinants of infant health outcomes.
We find a significant impact of pollution reductions on decreases in
infant mortality rates at the county level, with a 1 mg/m3 particulate
matter reduction resulting in about 4-8 less infant deaths per 100,000
births (a 0.3-0.5 elasticity).  The estimates are driven mostly by
deaths that occur within one month of birth and are very stable across
a variety of specifications.  The results also reveal a nonlinear
relationship between pollution and infant mortality.  Both of these
relations are masked in "conventional" analyses based on less credible
research designs than the approach used here (e.g., cross-sectional
and "fixed effects" analyses).