Abstract:
As one of a series of studies on the effect of atmospheric particles on human mortality, Schwartz (1993) analyzed mortality data from Birmingham, Alabama, as a function of small particles in the atmosphere (PMlO), meteorology, and systematic time trends. His overall conclusion was that when the meteorological and systematic time-trend effects were adjusted for, there is a statistically significant effect of PMlO on mortality. This is a specific and short-term effect: high particles, typically measured through three-day averages, have an immediate impact on daily mortality rates. The findings of Schwartz (1993) mirrored those of a number of other studies by Schwartz and co-workers, which together have contributed to a widely discussed belief that there is a causal relationship between particles and mortality. An independent study by Samet et al. (1995) verified the numerical correctness of the results but did not examine Schwartz's selection of statistical models for this data set.
We re-examine the whole question, using the same initial data as Schwartz but incorporating a wider range of meteorological variables. When we use the same variables as included by Schwartz, we obtain similar results to his. But when we use alternative models we obtain different conclusions. In particular, when humidity is included among the meteorological variables (it is excluded in the analysis by Schwartz), we find that the PMlO effect is not statistically significant.
We. perform a series of analyses. The first set of analyses introduce models that fit daily death counts to covariates (including PMlO) through a variety of standard linear regression methods. The second set duplicates the methodology in Styer et al. (1995) used in the study of Chicago and Salt Lake City. A third set of analyses employs Poisson regression. In all instances, we find no significant effect due to PMlO. The results we find here are consistent with those of Styer et al.: model selection is critical in making conclusions about the effect of particulates on mortality, requiring consistent, defensible approaches to assure reliable interpretations.
Keywords:
Causal inference, model selection, observational data, PMlO, Poisson regression, semiparametric modelling, cubic splines.
