Articles Assessing Confounding, Effect Modification, and Thresholds in the Association between Ambient Particles and Daily Deaths Joel Schwartz Environmental Epidemiology Program, Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts, USA were not available in enough of the cities to I examined the relationship between daily deadt and airborne partides in 10 U.S. cities with vary- allow examination ofthat variable. ing dimatic conditions and seasons in which partide concentraions were high. Airbome partiles Social and economic factors were extract- were associated with significant increases in daily deaths [0.67% increase for a 10 pgWm3 increase ed from the 1990 decennial Census (7) for in particles; 95% confidence intervl (CI), 0.52-0.81%]. This association was the same in summer use as potential effect modifiers. The variables and winter. To examine potential confounding by other pollutants, I regrsed city- and season- used were the unemployment rate, the per- specific effect sizes against the relationship between airborne particles and other pollutants. centage of the population living below the Controiing for other pollutants did not substantialy (or siicany) chnge the estimated effect poverty level, the percentage of the popula- ofairborne particles. Socioeconomic differences between cities likewise did not modify the effict. tion with a college degree, and the percentage The increase in daily death that occurred out ofhospitals (0.89% per 10 pg/m3; CI, 0.67-1.10%) ofthe population that was nonwhite. was substantially greater than the increase in deaths in hospitals (0.49%; CI, 0.314.68%). This is The assignment of PM1O exposure raised consistent with results previously reported in Philadelphia, Pennsylvania, and suggests that the a number of issues. Many of the locations particle-associated deaths are not just being brought forward by a few days. It is also nt have more than one monitoring location, with recent animal and human studies ofthe mehnisms ofpartide toxcity. Key work airborne but typically only one monitor operates on a particles, air pollution, climate mortality. Environ Helth Pepect 108:563-568 (2000). [Online daily basis, with the others operating every 3 May 20001 third or sixth day. If data from all of the btp://ehpnetl.niebs.nib.gpv/doas/2000108p563-568schwarn/abstracth.nm monitors were simply averaged, the daily mean would change on days when new monitors were included merely because their Studies on four continents have reported in potential effect modifiers for particulate air annual average differs from the monitoring associations between daily concentrations of pollution. Among these are social and eco- station that operates on a daily basis. ambient particles and daily deaths (1,4. The nomic factors that may represent differences The variance of PM10 measurements also magnitude of the regression coefficients var- in underlying risk. For example, income has can differ from monitoring location to mon- ied, but were remarkably similar compared to been shown to be a potent predictor of life itoring location. Day-to-day changes in epidemiologic studies of other exposures. expectancy. These factors differ among cities which monitors are included in the daily Several arguments have been made to ques- in the United States, and these diffferences average would also result in changes in the tion the relevance ofthese findings for public can be used to explore their role as effect day-to-day variation in the exposure measure health and preventive measures. It has been modifiers for the impact ofairborne particles. that do not represent true changes in expo- argued that the deaths are occurring in per- sure, but only changes in the sampling of sons who were already seriously ill and who Methods monitors. To remove these influences, I used would have died in a few days anyway. It has Data. I selected 10 U.S. cities with approxi- the following algorithm. The annual mean been argued that air pollution is responsible mately daily PM1O (particulate matter < 10 was computed for each monitor for each for the deaths, but that airborne particles are pm) monitoring to provide a reasonable year and subtracted from the daily values of not the responsible agent; rather, other pollu- number of locations for a combined analysis. that monitor. I then standardized these daily tants confound the particle findings. It has The cities were New Haven, Connecticut; deviances from each monitor's annual aver- also been argued that the partide associations Pittsburgh, Pennsylvania; Birmingham, age by dividing by the standard deviation for only exist at higher concentrations, and Alabama; Detroit, Michigan; Canton, Ohio; that monitor. The daily standardized devia- therefore, most days are below a presumed Chicago, Illinois; Minneapolis-St. Paul, tions for each monitor on each day were threshold for effect; hence public health Minnesota; Colorado Springs, Colorado; and averaged, producing a daily averaged stan- interventions to lower exposure would have Spokane and Seattle, Washington. Daily dardized deviation. I multiplied this by the no impact on most days. deaths in the metropolitan county containing standard deviation of all of the centered Two recent papers addressed the first each city were extracted from National monitor readings for the entire year and argument by showing that the association Center for Health Statistics mortality tapes added back in the annual average of all of between daily deaths and airborne particles (6) for the years 1986-1993. I also computed the monitors. This gave a daily average persisted after accounting for any short-term separate daily counts of deaths in the hospi- PMIO concentration for each day in each displacement of (reduced time until) deaths tals and deaths out of hospitals. Minneapolis (3,4). In this paper, I address the latter two and St. Paul were combined and treated as Address correspondence to J. Schwartz, Environ- issues in a multiple-city analysis ofparticulate one city. Daily weather data were obtained mental Epidemiology Program, Department of air pollution and daily deaths. I also indirect- for the same years, from the nearest airport Environmental Health, Harvard School of Public ly address the first issue by an analysis strati- weather station, and daily concentrations of Health, 665 Huntington Avenue, Boston, MA fied by location ofdeath. PM1O, sulfur dioxide, ozone, and carbon 02115 USA Telephone: (617) 432-1245. Fax: (617) et al. that monoxide were obtained for those years from 277-2382. E-mail: [email protected] Recently, Sunyer (5) reported This work was supported in part by NIEHS persons with a previous emergency room visit the U.S. Environmental Protections Agency's grant ES 07410 and by an EPA PM Research for chronic obstructive pulnmonary disease Aerometric Information Retrieval System Center Award. (COPD) had a greater risk of air pollution- (AIRS) monitoring network (Research Received 30 June 1999; accepted 14 January induced mortality. In general, there is interest Triangle Park, NC). Nitrogen dioxide data 2000. Environmental Health Perspectives * VOLUME 1081 NUMBER 61 June 2000 563 Articles . Schwartz city. I then computed the mean of the PM1o criterion used was to choose the parameter for simplicity, a Gaussian outcome and imag- concentration on the day of death and the for each variable that minimized Akaike's ine that Xt is the concentration on day t of day preceding death to use as my exposure Information Criterion (14). the pollutant that is causally associated with index. Most studies have found that a 2-day PM1O was treated as having a linear asso- the outcome YI. Hence average is a better predictor ofmortality than ciation with daily mortality in this analysis to Yt = + 131Xt + error.[1 a single day's exposure. Rather than optimiz- facilitate the combination of coefficients PhO ing in each location, I used the same 2-day across cities. Robust regression was used to Xt is correlated with another pollutant, Z4 average to ensure comparability. reduce sensitivity to outliers in the dependent which is not causally related to Y,. Therefore Analytical methods. For each city, a gen- variable. To reduce sensitivity to outliers in I may write eralized additive Poisson regression was fit the pollution variable, the baseline analysis Xt = 0y + yZt + error. [2] (8,, modeling the logarithm of the expect- was restricted to days when PM1O levels were ed value of daily deaths as a sum of smooth < 150 pg/m3, the currently enforced ambient What happens if Zt is used as the exposure functions of the predictor variables. The standard. This also ensures that the results are variable instead ofXA? Substituting Equation generalized additive model allows regressions unambiguously relevant to questions of revi- 2 into Equation 1 1 have to include nonparametric smooth functions sion ofthose standards. Yt= P0 +P1y0+ + error. [3] to model the potential nonlinear dependence Assessment ofconfounding. Confounding PlyZt ofdaily admissions on weather and season. It is usually assessed by induding the potential I have confounding by the omitted covariate assumes that confounder in the regression. This is a prob- Xp and the coefficient of Zt will be propor- log[E(I1] = PO = S,(XI) + ... +Sp(X), lem for air pollution epidemiology because tional to yi, the slope of the association atmospheric patterns, such as the height of between Xt and Zr This can be illustrated by where Yis the daily count of deaths, E(Y) is the inversion layer, tend to produce parallel some simple simulations. Figure 1 shows the the expected value of that count, the X. are increases and decreases in all air pollutants. results of a simulated example where one the covariates and the Si are the smooth (i.e., This creates considerable collinearity, and variable has a true association with the out- continuously differentiable) functions. For hence instability, in the estimated regression come, and the second variable does not but is the Si I used loess (10, a moving regression coefficients.