Epidemiologic studies of air pollution are attempts to exploit various dimensions of exposure variability that are observable in real-world settings. Because all risk factors cannot be explicitly controlled for by design and because they may be correlated with air pollution, potential confounding is a prominent concern. Statistical models are commonly used in attempts to isolate or estimate associations between health endpoints and air pollution while “controlling” for other risk factors that are potential confounders. Nevertheless, the elimination of residual confounding is never guaranteed because of the potential for inadequate modeling and because of unmeasured, poorly measured, or even unknown risk factors. In an article published in this issue of Epidemiology, Janes et al1 propose an approach to assess unmeasured confounding in a study of time trends in air pollution and mortality. We discuss here the issues of potential confounding in epidemiologic studies of air pollution that use different dimensions of exposure variability, and we comment on the contribution of the Janes et al1 approach in a broader context.
STUDIES OF SHORT-TERM TEMPORAL VARIABILITY
In air pollution epidemiology, the most exploited dimension of exposure variability is short-term temporal variability. Early studies of severe air pollution episodes in Meuse Valley (Belgium),2 Donora (PA),3 and London (England), observed dramatic, large increases in deaths and illness due to respiratory and cardiovascular disease associated with several days of extremely high pollution. However, even in these studies, confounding has been a concern. For example, elevated mortality rates persisted for weeks after the 1952 London episode, but the potential confounding role of an influenza epidemic is still being debated over 50 years later.4,5
Short-term temporal variability has also been analyzed in hundreds of daily time-series studies of mortality counts, hospitalizations, and other health endpoints.6 These studies are unlikely to be confounded by cigarette smoking, socioeconomic factors, or other factors that do not change day-to-day in correlation with air pollution. Associations of day-to-day changes in air pollution with respiratory and cardiovascular mortality counts have been estimated using statistical models that control for time-dependent variables such as long-term time trends, seasonality, and weather.6 Largely motivated by attempts to minimize the potential for confounding, increasingly rigorous statistical modeling techniques have been developed7–9 and large multicity daily time-series studies have been conducted.10,11 Another methodologic innovation, the case-crossover design, also has been used to evaluate effects of short-term pollution exposure by structuring the analysis so that potential confounders such as day of week, seasonality, and other time-dependent variables are controlled for by design rather than by statistical modeling.12,13 There continue to be concerns that residual confounding is responsible for the small estimated effects of short-term air pollution exposure. There is also concern that, for at least some of the studies, the effect estimates are biased downward due to overly aggressive fitting of time trends and control for multiple time-dependent variables.
STUDIES OF LONGER-TERM TEMPORAL VARIABILITY
Epidemiologic studies of air pollution can also attempt to evaluate longer-term temporal variability in exposure. Examples include several studies of natural experiments or interventions. Abrupt reductions in air pollution that accompanied a 13-month closure of a steel mill in Utah Valley were associated with reductions in hospitalizations and mortality.14 Reductions in respiratory and cardiovascular deaths coincided with substantial reductions in air pollution resulting from a ban of the use of coal burning in Dublin, Ireland.15 Reductions in seasonal mortality were observed in the first 12 months after restrictions on sulfur content of fuel in Hong Kong.16 Small but statistically significant drops in mortality were associated with an 8-month copper-smelter strike that resulted in sharp reductions in sulfate particulate matter and related air pollutants across 4 Southwest states—even while controlling for time trends, mortality in bordering states, and nation-wide mortality counts for influenza/pneumonia, cardiovascular, and other respiratory deaths.17 These studies evaluated the effects of temporal changes in pollution over a period of 8 or more months, but they tried to minimize confounding by restricting the analysis to well-defined natural experiments or interventions while controlling for long-term time trends and other potential confounders.
STUDIES OF LONG-TERM SPATIAL VARIABILITY
Another important source of exposure variability is spatial (or cross-sectional), such as differences in long-term (1 year or more) average ambient concentrations across metropolitan areas or differences in average concentrations across smaller communities or neighborhoods within metropolitan areas. Population-based cross-sectional studies in the 1970s and 1980s found that mortality rates across U.S. metropolitan areas were associated with fine-particulate-matter air pollution.18,19 Although investigators attempted to control for population-average differences in various demographic, socioeconomic, and other ecologic variables, these studies were largely discounted because of the inability to control for individual risk factors such as cigarette smoking and related concerns about confounding.
Several prospective cohort studies have reported that risks of cardiopulmonary mortality are strongly associated with spatial differences in long-term average concentrations of fine particulate air pollution (PM2.5) even after controlling for cigarette smoking and numerous other individual risk factors.20–26 Extensive reanalyses22 and extended analyses23–25 of early prospective cohort studies demonstrated that the PM2.5-associated mortality risk estimates were remarkably robust to various modeling specifications and to control for individual risk factors (including age, sex, race, smoking, alcohol use, marital status, education, body mass, occupational exposures, and diet). Even analyses that included spatial smoothing23 or ecologic socioeconomic variables22,24 found little evidence of spatial or “contextual neighborhood” confounding.
JANES ET AL1 STUDY OF LONG-TERM TIME TRENDS
In this issue of Epidemiology, Janes et al1 report a unique and interesting study design that attempts to evaluate longer-term time trends in mortality and air pollution. They conduct a 3-year (2000–2002) monthly time-series regression analysis using data from U.S. Medicare enrollment files to estimate the association between monthly mortality rates for U.S. counties and average PM2.5 concentrations for the previous 12 months. They estimate a regression model that includes a smooth function of time, county-specific indicators, and 2 measures of air pollution that represent different scale components of variation—the national trend in annual average PM2.5, and local trends defined as county-specific deviations in PM2.5 from national trends. They find associations of PM2.5 with mortality at the national scale, but no association at the local scale. They contend that the mortality association with long-term time trends are more likely to be confounded at the national scale than on the local scale and suggest that if the “association with national trends is set aside, there is little evidence of an association between long-term exposure to PM2.5 and mortality.”
We agree with the authors that it is difficult to make strong statistical inferences regarding trends in death rates that reflect long-term trends in pollution because of potential confounding of other slowly time-varying factors. Janes and et al1 present a clever approach, but from the perspective of the broader literature, their conclusions are overstated. They suggest that the statistical framework of their study draws from both cohort studies of long-term exposure and multisite time-series studies of short-term exposure. Unfortunately their analysis tells us little or nothing about unmeasured confounding in those and related studies because the methodology of Janes et al1 largely excludes the sources of variability that are exploited in those other studies. By using monthly mortality counts and lagged 12-month average pollution concentrations, the authors eliminate the opportunity to exploit short-term or day-to-day variability. Also, by including county-specific indicators, they eliminate the opportunity to exploit long-term average spatial variability—which makes up the large majority of the exposure variability in this data set and which serves as the basis for most of the epidemiologic studies of long-term air pollution exposure. If data were available, it might be possible to analyze much longer time series to increase exposure variation in time trends, but the potential for confounding over time would likely be even greater.
In at least one important way, the analysis of Janes et al1 is akin to the natural experiment or intervention studies. They are evaluating temporal changes in exposure over time periods of several months to a year or so. However, the natural experiment or intervention studies are strengthened by the fact that they can exploit abrupt changes in pollution exposure that is well-defined in both time and space.
It is interesting to note that Janes and colleagues, in another analysis,27 have used the same 2000–2002 U.S. Medicare mortality data and county-level PM2.5 data for the same time period to conduct a cross-sectional analysis that exploited the long-term average spatial variability in the data. They conducted the analysis for the same counties as previous cohort studies and found similar or even larger associations of PM2.5 and mortality. Although they are using basically the same data, it is not surprising, that their paper, reported in this issue of Epidemiology, observes less robust results. Their methodology eliminated the long-term average spatial variability as well as the short-term daily variability, and, given that they have only a 3-year study period with only small changes in air pollution, they have very little exposure variability to exploit. Such conditions are clearly not ideal for robust epidemiological studies.
We must be cautious about interpreting long-term trends of improving air quality and trends in reduced mortality rates as causal. However, the overall epidemiologic evidence for nonspurious associations between cardiopulmonary health and air pollution is much broader than addressed in Janes et al.1 In fact, because of concerns about confounding by other slowly time-varying factors, the use of long-term time trends as the primary source of exposure variability has been largely avoided in most air pollution epidemiology studies.
No single observational epidemiologic study can guarantee the elimination of all residual or unobserved confounding. With regard to the air pollution epidemiology literature, multiple studies using various study designs and dimensions of exposure variability have been conducted. There is no obvious single or common confounder—not cigarette smoking, demographic and socioeconomic factors, or weather and seasonal factors. It's possible that these air pollution studies have experienced “epidemiological bad luck” with multiple confounders that are inadequately controlled for and coincidentally cause spurious correlations that are somewhat coherent across the different study designs. To be consistent with the overall epidemiologic results, these confounders must correlate with pollution across various dimensions of time and space and be stronger risk factors for cardiopulmonary disease than for other diseases. The most likely explanation of recent air pollution epidemiology is that air pollution, especially fine particulate and related pollution, has measurable effects on cardiopulmonary health.
1. Janes H, Dominici F, Zeger S. Trends in air pollution and mortality: an approach to the assessment of unmeasured confounding. Epidemiology
2. Firket J. The cause of the symptoms found in the Meuse Valley during the fog of December, 1930. Bull Acad R Med Belg
3. Schrenk HH, Heimann H, Clayton GD, et al. Air Pollution in Donora, Pa.: Epidemiology of the Unusual Smog Episode of October 1948
. Washington, DC: Public Health Service; 1949.
4. Bell ML, Davis DL. Reassessment of the lethal London fog of 1952: novel indicators of acute and chronic consequences of acute exposure to air pollution. Enivron Health Perspect
5. Bell ML, Davis DL, Fletcher T. A retrospective assessment of mortality from the London smog episode of 1952: the role of influenza and pollution. Environ Health Perspect
6. Pope CA III, Dockery DW. Health effects of fine particulate air pollution: lines that connect. Air Waste
7. Pope CA III, Kalkstein LS. Synoptic weather modeling and estimates of the exposure-response relationship between daily mortality and particulate air pollution. Environ Health Perspect
8. Samet JM, Zeger SL, Kelsall JE, et al. Does weather confound or modify the association of particulate air pollution with mortality? An analysis of the Philadelphia data, 1973–1980. Environ Res
9. Dominici F, Sheppard L, Clyde M. Health effects of air pollution: a statistical review. Int Stat Rev
10. Samet JM, Dominici F, Curriero FC, et al. Fine particulate air pollution and mortality in 20 U.S. cities, 1987–1994. N Engl J Med
11. Analitis A, Katsouyanni K, Dimakopoulou K, et al. Short-term effects of ambient particles on cardiovascular and respiratory mortality. Epidemiology
12. Janes H, Sheppard L, Lumley T. Case-crossover analyses of air pollution exposure data: referent selection strategies and their implications for bias. Epidemiology
13. Schwartz J. The effects of particulate air pollution on daily deaths: a multi-city case-crossover analysis. Occup Environ Med
14. Pope CA. Particulate pollution and health: a review of the Utah Valley experience. J Expo Anal Environ Epidemiol
15. Clancy L, Goodman P, Sinclair H, et al. Effect of air-pollution control on death rates in Dublin, Ireland: an intervention study. Lancet
16. Hedley AJ, Wong CM, Thach TQ, et al. Cardiorespiratory and all-cause mortality after restrictions on sulphur content of fuel in Hong Kong: an intervention study. Lancet
17. Pope CA, Rodermund DL, Gee MM. Mortality effects of a copper smelter strike and reduced ambient sulfate particulate matter air pollution. Environ Health Perspect.
, in press.
18. Lave LB, Seskin EP. Air pollution and human health. Science
19. Özkaynak H, Thurston GD. Associations between 1980 U.S. mortality rates and alternative measures of airborne particle concentration. Risk Anal
20. Dockery DW, Pope CA III, Xu X, et al. An association between air pollution and mortality in six U.S. cities. N Engl J Med
21. Pope CA III, Thun MJ, Namboodiri MM, et al. Particulate air pollution as a predictor of mortality in a prospective study of U.S. adults. Am J Respir Crit Care Med
22. Krewski D, Burnett RT, Goldberg MS, et al. Reanalysis of the Harvard Six Cities Study and the American Cancer Society Study of Particulate Air Pollution and Mortality
. Special Report
. Cambridge: Health Effects Institute; 2000.
23. Pope CA III, Burnett RT, Thun MJ, et al. Lung cancer, cardiopulmonary mortality, and long-term exposure to fine particulate air pollution. JAMA.
24. Jerrett M, Burnett RT, Ma R, et al. Spatial analysis of air pollution and mortality in Los Angeles. Epidemiology
25. Laden F, Schwartz J, Speizer FE, et al. Reduction in fine particulate air pollution and mortality: extended follow-up of the Harvard Six Cities Study. Am J Respir Crit Care Med
26. Miller KA, Siscovick DS, Sheppard L, et al. Long-term exposure to air pollution and incidence of cardiovascular events in women. N Engl J Med
27. Eftim SE, Janes H, McDermott A, et al. A comparative analysis of the chronic effects of fine particulate matter. Johns Hopkins University, Dept. of Biostatistics Working Paper 121. 2006. Available at http://www.bepress.com/jhubiostat/paper121