4vEPA
United States
Environmental
Protection
Agency
Office of
Research
and
Development
EPA 600/R-01/028
June 2001
www.epa.gov/ncea
Spatial Variation in Ozone Concentrations in Phoenix, AZ for 1997
Introduction
Statistical analyses of the human health effects of airborne pollutants based on aggregate
population time-series data have often relied on ambient concentrations of pollutants measured at
one or more central sites in a given metropolitan area. In the particular case of ground-level ozone
(O3) pollution, central-site monitoring has been justified as a regional measure of exposure partly on
grounds that correlations between concentrations at neighboring sites measured over time are usually
high (U.S. EPA, 1996). In analyses where multiple monitoring sites provide ambient O3
concentrations, a summary measure such as an average has thus often been regarded as adequately
characterizing the exposure distribution. Indeed, a number of studies have referred to multiple-site
averaging as the method for measuring O3 exposure (U.S. EPA, 1996).
This report revisits the practice of multiple-site averaging. The results are drawn from an
analysis examining simultaneous mapping of population density and ambient O3 concentrations on
the scale of a single metropolitan area. Because of the ready availability of data associated with a
related research effort, the city selected for this analysis was Phoenix, AZ. The Phoenix metro area
is situated within Maricopa County, the most populous county in Arizona.
Description of Data
Two types of data were retrieved from Internet sites maintained by U.S. government agencies.
The first type of data consists of O3 concentrations reported at seventeen monitoring sites operating
in the greater Phoenix area during calendar year 1997, extracted from the U.S. EPA Aerometric
Information Retrieval System (AIRS) data base. The latitude and longitude of each monitor were
also retrieved from the AIRS data base. Although far from uniformly distributed, the seventeen
monitor sites overlay roughly a 35><30-mile rectangle covering the Phoenix area. The AIRS data
base provides concentrations recorded at each monitoring site by date and hour. For each monitor
reporting data on a given date, the maximum eight-hour moving average was calculated between the
hours of 8 am and 8 pm. Any 8 am - 8 pm data block for a monitor was excluded from consideration
if it did not contain all twelve hourly observations. To identify potential reporting problems, the
daily eight-hour maximum for each monitor was compared to the eight-hour maxima at nearby
monitors and extreme differences were investigated through inspection of the hourly records. The
most common pattern observed was a precipitous drop of O3 levels to near zero reported hourly
-------
concentrations, persisting for a day or more. These cases (a total of twenty-three monitor-days
distributed across six different monitors) were attributed to instrument malfunction and were
excluded from the analysis.
In order to characterize the geographic distribution of the local population on a sufficiently
detailed scale, total (all-ages) population and population-by-age data were obtained for each census
tract in Maricopa County in 1990, using the Decennial Census lookup facility provided on the U.S.
Census Bureau website. Although more current population data would have been preferred, 1990
is the most recent year for which complete census tract data (including distribution by age) were
publicly available at the time of this analysis. Census tracts can vary considerably in geographic size
and some have irregular shapes. To facilitate map placement of the population for each census tract,
the geographic centroid (latitude and longitude) of each tract was also obtained using query features
provided as part of the Decennial Census lookup. The population numbers and centroid for each
census tract were then combined according to the FIPS state-county-tract identifier codes established
by the Census Bureau.
Description of Maps and Spatial Smoothing Approach
The initial objective of the project was to develop a visual characterization of pollution monitor
placement with respect to local population distribution. Applying the graphics capabilities of SPLUS
2000 (MathSoft, Inc.) to the data, detailed population density maps were produced with O3 monitor
locations superimposed. The two-dimensional LOESS smoothing routine available in SPLUS was
then applied to the 8-hour O3 maxima across monitoring sites (span = .75) to generate concentration
isopleths covering much of the densely populated part of Maricopa County. The isopleths are
intended to provide a convenient visual summary of the data across the network of monitoring sites.
By superimposing these isopleths on the population/monitor maps, composite maps were obtained
showing the variation in O3 concentrations across the populated area. (The spatial smoothing routine
was applied to the O3 concentration data on a scale treating one degree latitude as representing the
same distance as one degree longitude. In fact, the distance represented by one degree latitude is
roughly 20 percent greater than the distance represented by one degree longitude at this location on
the globe, which introduced minor distortions in the smoothing process. However, all maps have
been sized so that one inch on the vertical scale represents the same distance in miles as one inch on
the horizontal scale, resulting in proper depiction of greater vertical distance per degree than
horizontal distance per degree.) Notably, Phoenix has several large population clusters of persons
> 65 years of age concentrated on the fringes of the metro area. Therefore, maps were also produced
reflecting the geographic distribution of this special population.
Among the numerous maps generated, significant spatial variation is often apparent in the
concentrations reported across monitoring sites. Figures 1 a - d show a sampling of the types of
spatial patterns observed. In each map the triangle symbols denote the locations of the monitors
reporting on the indicated date. The range of concentrations indicated by the contour lines for each
plot is generally conservative due to the wide-span smoothing, i.e., the contour lines generally span
-------
a narrower numerical range than do the individual concentrations across monitoring sites. Each
square on a map represents a single census tract. The size of each square is proportional to the 1990
all-ages population residing within the corresponding census tract, and the placement of each square
is based on the census tract centroid. The map area represents roughly a 43.1 -mile square. Selected
segments of the Interstate and U.S. highway system are also included for reference.
It might be noted from the maps that several monitors, most notably the four monitors in the
extreme east and northeast of the metro area, are at sites with very low nearby population. These are
nevertheless useful in the spatial analysis since they act as "high-leverage" data points in the
smoothing process. Because the LOESS smoothing procedure available in SPLUS does not
extrapolate outside the range of the coordinates, the coverage area for the concentration isopleths is
restricted to a rectangle roughly 35 miles from east to west and 30 miles from north to south.
The maps for 6/8/1997 and 7/5/1997 (Figures 1 a and d) indicate a "bullseye" pattern with the
highest concentrations observed near the center of the metro area. The map for 6/19/1997 (Figure
1 b) indicates a gradient in O3 concentrations increasing in a general west-to-east direction, with
notable westward intrusions of the concentration zones. The map for 6/28/97 (Figure 1 c) indicates
relative spatial uniformity of concentrations.
Based on Figures 1 a - d, the monitoring network appears to provide reasonably effective
coverage of the all-ages population. A second set of maps for the same four days (Figures 2 a - d),
showing the population distribution of individuals > 65 years of age, provides something of a
contrast to the first set of maps. Given the observed spatial variation in O3 concentrations on
selected days, the monitors do not appear to be ideally situated to allow the assessment of exposure
for this at-risk population. Although there may be some discrepancies created by superimposing
1997 monitor sites on 1990 population maps, there is reason to expect that the maps are
representative of the situation in 1997. Based on more recent U.S. Census Bureau aggregate-level
estimates, population growth in Phoenix proper between 1990 and 1997 was about 20% while in the
remainder of Maricopa County population growth was about 34%. Indications are that the areas to
the northwest (e.g., Peoria and Surprise City) and to the southeast (e.g., Chandler, Gilbert, and
Mesa), which do not fall directly under the monitoring network, experienced this same
disproportionate growth. Based on these growth patterns, it is unclear whether the discrepancy
between the geographic distribution of persons > 65 years of age and that of the general population
persists, although it seems unlikely to be substantially changed from 1990 given that suburban
Phoenix is well-known for its retirement communities.
As a cautionary note, although the cases presented here need not be regarded as particularly
anomalous, neither should they be regarded as representative or typical; they were selected simply
as interesting examples of the types of spatial patterns observed.
-------
33.7-
(A
-------
-112.25 -112.05 -111.85 -111.65
Longitude (decimal degrees)
-112.25 -112.05 -111.85 -111.65
Longitude (decimal degrees)
(a) 6/8/97
(b) 6/19/97
-112.25 -112.05 -111.85 -111.65
Longitude (decimal degrees)
-112.25 -112.05 -111.85
Longitude (decimal degrees)
-111.65
(c) 6/28/97
(d) 7/5/97
Figures 2 (a) - (d). Eight-Hour Maximum O3 Concentrations (ppb) over Greater Phoenix, AZ for selected days in 1997.
Squares denote census tracts and are proportional in size to the 1990 population > 65 years of age. Triangles show
locations of O3 monitors reporting for each day. The maps represent an area approximately 43 miles square.
-------
Loss of Information on Spatial Variability Due to Averaging
Despite the high correlations often observed between O3 concentrations measured over time at
neighboring sites, the correlations are not perfect and the maps presented in the previous section
suggest that the shapes of the concentration isopleths do not always persist from day to day.
Consequently, cross-site averages of O3 concentrations might not fully describe changes in exposure
patterns day to day. This hypothesis can be studied somewhat more systematically by calculating
the cross-site average and standard deviation of O3 concentrations for each day, and then plotting
daily averages against daily standard deviations. If the daily standard deviations exhibit variability
which is unrelated to the daily average concentrations, a loss of information about exposure patterns
will be realized if the average alone is used to characterize O3 exposure. For this phase of the
analysis, the period from January through mid-July 1997 was considered. This period is of particular
interest because it was during this time only that an O3 monitor was in operation in the northwest part
of the metro area. This monitor (visible in all of the maps) often indicated concentrations much
lower than other monitors in the network, and moreover it is the monitor nearest to much of the
population cluster > 65 years of age in the northwest corner of the metro area. (Maricopa County
government staff confirmed that quality control and quality assurance procedures had been followed
regarding data from this monitor.) In order to make the analysis precise, only those dates with
complete hourly readings for all monitors in a fixed subset were admitted into the analysis. Fourteen
monitors operated more or less continuously over the January to mid-July period, and on 47 days all
fourteen monitors reported hourly readings for all twelve hours between 8 am and 8 pm. The
locations of the subset of fourteen monitors used in this analysis are shown in Figure 3.
Since the primary interest is in population exposure, the daily cross-site averages and standard
deviations were population-weighted. Specifically, the all-ages population for each census tract was
assigned to the nearest monitor. Each monitor was then given a weight proportional to its assigned
population. Figure 4 shows the resulting daily cross-site averages plotted against the daily cross-site
standard deviations for the 47 dates with complete hourly data for the subset of fourteen monitors.
The population-weighted cross-site daily averages range between 20.8 and 71.0 ppb, while the
population-weighted standard deviations range between 3.0 and 8.8 ppb. It is apparent from the
scatterplot that for this particular data set, the cross-site averages and standard deviations are weakly
related (sample correlation coefficient r = -.06). At all observed average concentrations the pattern
of ambient concentrations may reflect relatively low or relatively high site-to-site
variability, depending on the particular day. As a specific example, the population-weighted
averages corresponding to Figures 1 c and d are 62.3 ppb and 64.8 ppb, respectively, while the
respective population-weighted standard deviations are 4.2 ppb and 8.6 ppb, indicating a substantial
difference in site-to-site variability for two days with comparable cross-site averages. This
difference is readily apparent in the two figures. Repeating this analysis under a weighting scheme
based on the population > 65 years of age produced qualitatively comparable results, although the
daily standard deviations were somewhat greater than those based on all-ages population weighting.
-------
33.7
-------
Site-to-site Correlation Before and After Seasonal Filtering
Site-to-site correlation of measured pollutant concentrations as a function of distance has been
previously investigated for various airborne pollutants, including O3. Based on data collected from
a network of sixteen monitors across the greater Detroit area in the summer of 1981, the correlations
between O3 concentrations measured at a central city monitor and the other fifteen monitors were
generally high (median r = .78) but also were observed to decrease with distance between monitors
(Kelly et al., 1986). Site-to-site correlations of O3 concentrations measured across a network of four
monitors in this same geographic area over a multi-year period were similarly high (median r = .83)
and decreased gradually as the distance between monitors increased (Lippman et al., 2000).
Similarly, based on O3 concentrations measured across a network of ten monitors in the greater
Chicago metro area from 1985 -1994, site-to-site correlations were generally high (median r = .75)
but clearly decreased with distance (Ito and Thurston, 1999).
Many contemporary environmental epidemiology time-series analyses employ some form of
filtering to remove seasonality and other types of trends from the data prior to the correlational
modeling phase. In some instances, filtering "on both sides" is the preferred approach (Kinney and
Ozkaynak, 1991; Cakmak, etal. 1998). Under this approach the analyst would take steps to detrend
not only the health outcome time-series data but also the pollutant time-series data. In the case of
O3, it stands to reason that site-to-site correlation is driven in part by its strong seasonal profile (Ito,
et al. 1998). Hence, we may anticipate that seasonal detrending is likely to reduce the relatively high
site-to-site correlation commonly observed between monitors across an urban airshed.
An analysis of the site-to-site correlation as a function of distance for the 1997 Phoenix data
suggests the usual high site-to-site correlation even at distances greater than 30 miles. For
consistency, this site-to-site analysis was based on data for the same dates and monitors analyzed in
the previous section. The selected monitors provided enough geographic spread to allow an
evaluation of the distance effect. Figure 5a shows the 91 site-to-site correlations plotted against
distance in miles. The trend appears approximately linear, and an OLS regression line was fit
through the scatter of points (estimated intercept = .98, estimated slope = -.0046).
To examine the effect of seasonal detrending on the site-to-site correlations, the time series of
eight-hour maxima for each monitor was filtered using a wide-span LOESS smoother. The within-
site standard deviations of the filtered concentrations ranged from 6.1 ppb to 9.5 ppb for the fourteen
monitors, indicating that substantial day-to-day variation remained after detrending. The site-to-site
correlations of the filtered concentrations were then plotted against distance and a regression line was
again fit to the scatter of points (estimated intercept = .92, estimated slope = -.0079). The results are
shown in Figure 5b. Not surprisingly, the site-to-site correlations for the filtered concentrations are
weaker and decline more rapidly with distance than do those based on unfiltered concentrations. The
smaller intercept suggests an overall loss of correlation even between neighboring monitors, and the
decline in correlation with distance is -.0079 / -.0046 =1.7 times as fast as a result of seasonal
filtering.
-------
1.0
0.9-
iS 0.8-
o
O
0)
J 0.7
0)
0.6-
0.5
10 20
Distance Between Monitors (miles)
30
Figure 5a. Site-to-site correlation of eight-hour maximum O3 concentrations as a function of distance between O3
monitors. The O3 maximum concentrations have not been seasonally detrended.
1.0
0.9-
c
.0
0)
I
0.8-
0.7-
1
V)
0.6-
0.5
10 20
Distance Between Monitors (miles)
30
Figure 5b. Site-to-site correlation of eight-hour maximum O3 concentrations as a function of distance between O3
monitors after seasonal detrending.
-------
Summary
The results of the data analyses presented in this report indicate that in the Phoenix metro area:
(1) significant spatial variation in daily maximum eight-hour O3 concentrations is occasionally
observed and patterns of spatial variation do not always persist from day to day; (2) monitor
placement may not be ideal for measuring population exposure, particularly for the population of
individuals > 65 years of age; (3) multiple-site averaging does not provide full information on the
distribution of ambient concentrations; and (4) high site-to-site correlations are due in part to
common seasonal variation at all monitoring sites and do not remain nearly as high after seasonal
detrending. Findings (1) and (2) are presented visually and might be considered anecdotal, but
nevertheless may be potentially useful in suggesting general hypotheses. Findings (3) and (4) result
from systematic analyses and may have implications for the statistical evaluation of health effects.
10
-------
References
Cakmak S., Burnett R., and Krewski D. (1998). Adjusting for temporal variation in the analysis of
parallel time series of health and environmental variables. Journal of Exposure Analysis and
Environmental Epidemiology, 8(2): 129-144.
Ito K., Thurston G.D., Nadas, A., and Lippman, M. (1998). Multiple pollutants' spatio-temporal
exposure characterization errors. In Proceedings of a Specialty Conference Cosponsored by the Air
& Waste Management Association and the U.S. Environmental Protection Agency's National
Exposure Research Laboratory, Gary, NC.
ItoK. andThurstonG.D. (1999). A pilot study of the effects of Berkson-type exposure measurement
errors on regression models of mortality and morbidity: final report. Unpublished manuscript.
Kelly N.A., Ferman M.A., and Wolff G.T. (1986). The chemical and meteorological conditions
associated with high and low ozone concentrations in southeastern Michigan and nearby areas of
Ontario. Journal of the Air Pollution Control Association, 36: 150-158.
Kinney P.L. and Ozkaynak H. (1991). Associations of daily mortality and air pollution in Los
Angeles County. Environmental Research, 54: 99-120.
Lippman, M., Ito K., Nadas, A., Burnett, R.T. (2000). Association of particulate matter components
with daily mortality and morbidity in urban populations. HEI Research Report No. 95. Boston:
Health Effects Institute.
U.S. EPA (1996). Air quality criteria for ozone and related photochemical oxidants. Office of
Research and Development, EPA/600/P-93/004cF, Washington, DC.
Disclaimer
This document has been reviewed in accordance with U. S. Environmental Protection Agency policy
and approved for publication. Mention of trade names or commercial products does not constitute
endorsement or recommendation for use.
For additional information contact:
National Center for Environmental Assessment
U.S. Environmental Protection Agency
MD-52
Research Triangle Park, NC 27711
11
------- |