Total Suspended Particulate Matter and Daily
Mortality in Cincinnati, Ohio
Environmental Epidemiology Program, Harvard School of Public Health, Boston, MA
nia, cardiovascular disease, and the elderly
The London smog disaster of December
1952 established that high levels of air- can be replicated. In addition, this study
used diagnostic plots and sensitivity analy-
borne particles and sulfur dioxide pro-
duced large increases in daily death ratesI. ses that have been used in few if any prior
studies to ensure that the association is not
Since the London episode, time-series
confounded by inadequate control for
analyses have associated daily mortality
with daily particle concentrations in
Daily deaths of residents of Hamilton
London (2,3), New York (4,5), Santa
Clara, California (6), Detroit (A), Steuben- County, Ohio, the county containing
Cincinnati (population 873,224 in 1980)
ville, Ohio (8), Philadelphia (9), St. Louis
were extracted from the detail mortality
and eastern Tennessee (10), Utah Valley
tapes of the National Center for Health
(11), Minneapolis (12), and Birmingham,
Statistics for the calendar years 1977-82.
AL (13). Both the air pollution concentra-
Deaths from accidental causes (ICD 9
tions and mortality increases were lower
>800) and deaths occurring outside the
than seen in London in 1952.
county were excluded from the analysis. In
These studies have reported that the
addition to daily deaths, deaths of persons
primary association was with particles and
aged 65 or older and daily deaths from car-
not sulfur dioxide, and none have found
diovascular disease (ICD 390-448) and
any evidence for a threshold. In Santa
from pneumonia (ICD 480-486) were
Clara (6) and the Utah Valley (11), coex-
tabulated. These daily counts were match-
posure with sulfur dioxide was essentially
absent, avoiding any possible confounding ed to the 24-hr average (midnight to mid-
night) of total suspended particulates
with that pollutant. Clearly, airborne parti-
(TSP) from all the population-oriented
cles are associated with daily mortality
independent of sulfur dioxide exposure. monitors within the city. Pollution data
were retrieved from the Environmental
The regression coefficients reported in ....~~~~~~~~~~.
Protection Agency's aerometric data bank.
Santa Clara and Utah were similar to the
For each day, TSP from all available moni- and monthly variations, including the pos-
regression coefficients in the other studies.
This suggests that the other associations sibility that seasonal peaks may vary in
tors were averaged to produce an area wide
were due to the particle and not the sulfur mean. Daily monitoring for TSP began in date and magnitude from year to year.
dioxide exposure. In the other cities, the Multiple categories of temperature and
1977. Only one monitoring station was
regression coefficients of particles in mod- dew point temperature allow for a U-
running every day in the city of Cincinnati
els containing sulfur dioxide were little that year. In 1979-81, two additional shaped relationship with daily deaths, or
changed from their values without simulta- other nonlinearities, if present. A linear
monitors were operated on a daily sched-
neous consideration of sulfur dioxide; they ule. Several additional monitoring stations and quadratic time term were also includ-
also remained statistically significant. In were operational on a one-day-in six-basis, ed to capture any continuous temporal
contrast, the coefficients of sulfur dioxide and were included in the average when trend in mortality over the period.
Several alternative approaches were
were generally much lower in models that available. Data on daily mean temperature
and daily mean humidity were obtained also examined to see how sensitive the air
controlled for particle exposure and were
from the weather station at the Cincinnati
not statistically significant. This indicates pollution results were to the method used
to control for these potential confounders.
that variations in particle concentrations airport.
independent of sulfur dioxide were associ- Counts of rare events, such as dying on The second approach replaced the dummy
a particular day, are not normally distrib- variables with natural splines. This mod-
ated with variations in daily mortality.
eled the dependence of mortality on time
Variations in sulfur dioxide concentrations uted, and were modeled in Poisson regres-
that were independent of particle concen- sions. Substantial seasonal and other long- and temperature using cubic polynomials
fit to intervals of time (or temperature).
term temporal patterns are often found in
trations did not appear to be associated
The cubic polynomials were subject to
daily mortality data. While air pollution
with variations in daily mortality in these
continuity constraints at the boundaries of
may be responsible for part of that pattern,
the intervals. The number of intervals each
other factors clearly play the dominant
The study in Philadelphia provided
variable was divided into was chosen by
some more details on the specific causes of role. Temperature and humidity may be
analysis of deviance. First an initial num-
death that were elevated on high air pollu- related to mortality in a nonlinear way.
The analytic method was chosen to deal
tion days. The relative risks were highest
for pneumonia and cardiovascular disease. with these issues. Address correspondence to J. Schwartz, En-
vironmental Epidemiology Program, Harvard
The basic approach taken in this analy-
They were also highest for the elderly. The
School of Public Health, 665 Huntington Avenue,
sis was to use dummy variables for each of
present analysis sought to examine the rela- Boston MA 02115 USA.
the 72 months in the study period and for
tionship between airborne particles and Joel Schwartz is supported in part by a John D.
8 categories of temperature and dew point
daily mortality in Cincinnati, Ohio. In and Catherine T. MacArthur Fellowship.
temperature. Dummy variables for each
particular, it sought to determine if the Received 21 September 1993; accepted 28 Dec-
month in the study controlled for seasonal
finding of higher relative risks for pneumo- ember 1993.
Environmental Health Perspectives
- * 99.
The final approach used nonparametric old model for air pollution. In addition to
ber of intevals was chosen. For example,
regression. The generalized additive model the statistical test, the smoothed covariate
the initial number of intervals chosen for
(14) allows a Poisson regression to be fit as adjusted plot of daily deaths versus air pol-
time was 24, which means each interval
lution allows a direct graphical examina-
a sum of nonparametric smooth functions
was 6 months long, and was fit with a sep-
tion of the shape of the relationship
of each of the predictor variables. A non-
arate cubic polynomial. Then the number
parametric smoother is a tool for summa- between exposure and outcome.
of intervals was increased (decreasing the
Time series data are often serially corre-
rizing the trend of a response measure-
number of months per interval in our
ment, Y, as a function of one or more pre- lated. Although this may be due to the
example) until the improvement in model
weather, there may be serial correlation
dictor measurements, Xi. . . XP . Most
fit for the increased degrees of freedom
remaining in the residuals of the regression
smoothers are generalizations of weighted
used was not statistically significant. The
models. Ignoring such serial correlation
spline approach has the capability of fitting moving averages. These predict the expect-
ed value of Yat AX as the weighted mean of can lead to biased hypothesis tests. All
dependencies that may be highly nonlinear
regression models were tested for serial cor-
the actual values of the Y's corresponding
and is therefore useful in assuring that any
relation in the residuals, and if any was
association with TSP is not confounded by to all the X's in a symmetric neighborhood
found, autoregressive Poisson models were
around A. The weights decline with dis-
inadequate control for weather or season.
tance from the center of the neighborhood.
Another approach to ensuring that any
Table 1 shows the distribution of air
The properties of such smoothers have
observed association with TSP is not due
pollution, temperature, dew point temper-
been extensively discussed (15,16). Loess
to a coincidence of high TSP days with
(17) was used in this analysis, which fit the ature, and daily mortality in Cincinnati
extreme weather conditions is to repeat the
during the years 1977-82. The distribu-
counts of daily deaths to smooth functions
analysis excluding such days. The basic
tion of counts of daily deaths is illustrated
of temperature, dew point temperature,
analysis was repeated excluding days with
24-hr mean temperature below 100F or day of study, and TSP. Loess uses a run-
ning regression rather than a running mov-
above 80'F. These represent the coldest
ing average, with weights that decline as
and warmest 2.5% of the days, respective-
the cube of the distance from the center of
ly. In addition, the analysis was repeated
the neighborhood. The generalized addi-
using the minimum daily temperature
tive model allows for a test of the improve-
instead of the average daily temperature as
ment in fit compared to a linear no-thresh-
the control variable.
Table 1. Distribution of variables in the analysis: daily mortality and total suspended particulates (TSP) in
Cincinnati, 1977-1982 a
75th Percentile Mean
Variable 25th Percentile Median
Daily deaths 24
TSP (pg/m3) 76 30.8
53 71 93
Temperature, 0F 19.6
Dew point temperature, OF 60
28 44 a
Daily deaths, >age 65
Figure 1. Histogram of daily deaths in Cincinnati,
Pneumonia deaths 1
Cardiovascular deaths Ohio, during the study period, 1977-1982.
Week of Study
Week of Study
Figure 2. (A) Plot of daily deaths in Cincinnati, Ohio, versus time during the course of the study. Each point represents 1 of the 2191 days during the study period.
The line shown is a nonparametric smoothed curve fit to the data and illustrates the seasonal variability in the data. (B) Plot of the residuals of daily mortality
(from the basic regression model without total suspended particulates) versus time. There is no long-term pattern remaining in the data, as indicated by the non-
Volume 102, Number 2, February 1994
Figure 3. (A) Plot of the residuals of daily mortality (from the basic regression model without total suspended particulates) versus temperature. There is no inter-
val of temperature where the model net over- or underestimates daily mortality, as indicated by the nonparametric smoothed curve. (B) Plot of the residuals of
daily mortality (from the basic regression model without total suspended particulates) versus dew point temperature. There is no interval of dew point tempera-
ture where the model net over- or underestimates daily mortality, as indicated by the nonparametric smoothed curve.
in Figure 1, which shows the slight skew- basic model produced no significant im-
showing systematic over- or underestima-
ness expected. An average of 21 persons tion of mortality. provement in fit (p = 0.98). Excluding the
Table 2 shows the relative risk of (RR)
died each day in the metropolitan area, warmest and coldest days also had little
death for a 100 Pg/mi3 increase in TSP
producing an annual death rate of 8.8 per effect on the association between TSP and
from the basic model and the various sensi- daily deaths. The relative risk for TSP in
The basic regression model did a good tivity analyses. TSP was a significant risk that model was also similar to the basic
job of controlling for the seasonal variation factor for mortality in this analysis (RR = model. The test for nonlinearity in the
1.06, 95% CI = 1.03-1.10). The use of TSP association was not significant (p =
in mortality. For example, Figure 2a shows
a plot of daily deaths by week of study. 0.20). Figure 4 shows the nonparametric
natural splines resulted in a more parsimo-
nious model, with 34 intervals of time and
The line through the points is a nonpara- smoothed plot of daily deaths versus TSP
metric smooth. The strong seasonal varia- 2 intervals each of temperature and dew in Cincinnati, after controlling for covari-
point temperature being fit with cubic
tion is obvious. Figure 2b shows the ates.
deviance residuals from the basic regression Table 3 shows the relative risk of death
polynomials within each interval. The dif-
for a 100 pg/m3 increase in TSP concen-
ference in explanatory power of the
model without air pollution. The seasonal
pattern has been removed, as illustrated by dummy variable model was not significant trations for the age and cause specific death
the nonparametric smooth. (p = 0.99), despite its using 39 more rates examined. The same pattern seen in
Philadelphia (9) was evident in Cincinnati.
degrees of freedom. The association of
Figure 3a shows the deviance residuals
mortality with TSP was unchanged in the The relative risk was higher for the elderly
(again from the basic model without air
pollution) plotted against temperature, and spline model (Table 2). Similarly, the gen-
Figure 3b shows the deviance residuals eralized additive model fit the daily mortal-
plotted against dew point temperature. ity data more parsimoniously than the
Again, a good fit seems to have been dummy variable model. Again, the use of
achieved, with no temperature intervals 42 additional degrees of freedom in the
Table 2. Relative risk of death and 95% confidence intervals for a 100 pg/mr3 increase in total suspended a
particulates under different modeling alternatives
95% Confidence interval
Model Relative risk a
Basic modela 1.03-1.10
Natural spline modelb 1.03-1.10
No extreme days 1.02-1.10
Generalized additive models 1.04-1.13
Minimum temperatures 1.03-1.10
ZU 4I HMa ISl
a72 dummy variables for each month of study, 8 categories each of temperature and dew point tempera- Total Suspended Particulates
ture, and linear and quadratic time trend variables.
bSeparate cubic polynomials fit for 34 intervals of time, and 2 intervals each of temperature and dew point Figure 4. Nonparametric smooth curve of daily
temperature. mortality versus total suspended particulate, after
CModel a, excluding days with temperature below 100F or above 80IF. controlling for temperature, dew point tempera-
dNonparametric smooth functions of time, temperature, and dew point temperature. ture, and time in a generalized additive model. No
eModel 1, but with daily minimum temperature instead of daily average temperature. evidence for a threshold is seen in the relationship.
Environmental Health Perspectives
respirable particles. Moreover, the London
Table 3. Relative risk of death and 95% confidence intervals for a 100 pg/m3 increase in total suspended episode of 1952 provides ample demon-
particulates for all-cause mortality, and for selected subgroups by age and cause of death stration of biological plausibility: it is clear
that respirable particles increased mortality
Outcome Relative risk 95% Confidence interval
in that episode, although no mechanism
All deaths 1.06 1.03-1.10
was determined for that case, either.
Deaths, age >65 1.05-1.14
Cardiovascular deaths 1.08 REFERENCES
All models are after adjustment for covariates.
1. HMSO. Mortality and morbidity during the
London fog of December 1952. Report no.
the model in removing seasonal variation
and for deaths from pneumonia and car- 95. on Public Health and Medical Subjects,
from the data. The results are not con-
diovascular disease. Her Majesty's Public Health Service. London:
founded by extreme weather conditions
The association between mortality and Her Majesty's Stationary Office, 1954.
because the association is little changed
TSP seen in Cincinnati was almost identi- 2. Mazumdar S, Schimmel H, Higgins ITT.
when those days are excluded. In addition,
cal to that seen in Philadelphia. Moreover, Relation of daily mortality to air pollution: an
analysis of 14 London winters, 1958/59-
the residual plots indicate that the basic
the relationships between TSP and cause-
1971/72. Arch Environ Health 37:213-220
model did not under- or overestimate the
and age-specific daily deaths in the two (1982).
cities were remarkably similar. For compar- dependence on weather conditions in any 3. Schwartz J, Marcus A. Mortality and air pol-
ison, in Philadelphia, the relative risk for a temperature range. The sensitivity of the lution in London: a time series analysis. Am J
100 pg/m3 increase in TSP was 1.07 (95% results to alternative ways of controlling for Epidemiol 131:185-194(1990).
CI =1.04-1.10) for all-cause mortality, weather was slight. In the absence of a 4. Glaser M, Greenburg L. Air pollution, mor-
tality, and weather: New York City, 1960-
plausible candidate for a major con-
1.11 (95% CI = 0.97-1.27) for pneumonia
1964. Arch Environ Health 22:334-343
founder, the conclusion that the observed
deaths, 1.10 (95% CI = 1.06-1.14) for car- (1971).
association is causal seems reasonable.
diovascular deaths, and 1.10 (95% CI = 5. Schimmel H, Murawski TJ. The relation of
1.06-1.13) for deaths in the elderly. The One way to indirectly assess the possi- air pollution to mortality. J Occup Med
finding of such similar associations in dif- bility of confounding by omitted factors is 18:316-333(1976).
ferent locations makes it less likely that the to examine the relationship in different 6. Fairley D. The relationship of daily mortality
to suspended particulates in Santa Clara
associations could be due to a confounder. locations, as the correlation between those
County, 1980-1986. Environ Health Per-
Smoking habits, occupational expo- factors and air pollution should differ spect 89:159-168(1990).
among cities. If the association is due to an
sure, dietary patterns, and socioeconomic 7. Schwartz J. Particulate air pollution and daily
factors are not potential confounders of omitted or imperfectly controlled con- mortality in Detroit. Environ Res 56:204-
this association because their day-to-day founder, the induced regression coefficient 213(1991).
for particulate air pollution will depend on
variation is small and not correlated with 8. Schwartz J, Dockery DW. Particulate air pol-
lution and daily mortality in Steubenville,
the regression coefficient of the con-
air pollution. More likely, potential con-
Ohio. Am J Epidemiol 135:12-20(1992).
founder with daily mortality and the
founders are epidemics, other factors that 9. Schwartz J, Dockery DW. Increased mortality
produce the seasonal variation in mortality, regression coefficient of the confounder in Philadelphia associated with daily air pollu-
and weather. In Cincinnati, daily mortality with particulate air pollution. If the latter tion concentrations. Am Rev Respir Dis
and respiratory epidemics both peaked in varies widely but the coefficient of partic- 145:600-604(1992).
the winter. In contrast, winter was the sea- ulate air pollution with mortality does not 10. Dockery DW, Schwartz J, Spengler JD. Air
pollution and daily mortality: associations
vary widely, confounding by that factor is
son with the lowest TSP concentrations.
with particulates and acid aerosols. Environ
Hence, these factors are unlikely sources of unlikely. The recent studies of particulate Res 59:362-373(1992).
upward bias in the association. Further- air pollution and daily mortality have all 11. Pope CA, Schwartz J, Ransom M. Daily mor-
more, Figure 2 demonstrates the success of reported relatively similar effect sizes. This tality and PM10 pollution in Utah Valley.
lends credence to their findings, since if Arch Environ Health 42(3):211-217(1992)
the associations were due to confounding, 12. Schwartz J. Particulate air pollution and daily
mortality: a synthesis. Public Health Rev 1
we would expect a greater variability in the
effect size estimates. Figure 5 shows the rel-
13. Schwartz J. Air pollution and daily mortality
Steubenville ative risk of death for a 100 pg/m3 increase in Birmingham, AL. Am J Epidemiol 137:
in TSP concentration from the recent
St. Louis _
studies of air pollution and daily mortality.
Utah | _ _ _
14. Hardle W. Smoothing techniques with imple-
The consistency is remarkable. mentation in S. New York. New York:
Although the mechanism by which air-
15. Gasser T, Muller HG, Mammitsch V. Ker-
borne particles exacerbate illnesses and nals for nonparametric curve estimation. J R
*_ 0.01 '.9--I----w
increase their mortality rates is not under- Stat Soc B 47:238-52(1985).
stood, neither is the mechanism by which
i~~~~ct 16. Cleveland WS, Devlin SJ. Robust locally-
00Pm3inrease in ail toa suspne par- tobacco smoking increase the risk of death weighted regression and smoothing scatter-
from myocardial infarctions. This fact has plots. J Am Stat Assoc 74:829-36(1988).
17. Hastie T, Tibshirani R. Generalized additive
not prevented a conclusion being drawn
models. London:Chapman and Hall, 1990.
from the strong epidemiologic information
Figure 5. Relative risk of all-cause mortality from a 18. Liang KY, Zeger SL. Longitudinal data analy-
in the case of smoking and it should not
100 pg/rn3 increase in daily total suspended par- sis using generalized linear models. Bio-
impede a similar conclusion in the case of
ticulates from studies in different cities. metrika 73:13-22(1986).
Volume 102, Number 2, February 1994