Content uploaded by Joel Schwartz
Author content
All content in this area was uploaded by Joel Schwartz on Oct 24, 2015
Content may be subject to copyright.
ORIGINAL ARTICLE
Ozone trends and their relationship to characteristic weather
patterns
Elena Austin
1,2
, Antonella Zanobetti
1
, Brent Coull
3
, Joel Schwartz
1,4
, Diane R. Gold
1,4
and Petros Koutrakis
1
Local trends in ozone concentration may differ by meteorological conditions. Furthermore, the trends occurring at the extremes
of the Ozone distribution are often not reported even though these may be very different than the trend observed at the mean or
median and they may be more relevant to health outcomes. Classify days of observation over a 16-year period into broad
categories that capture salient daily local weather characteristics. Determine the rate of change in mean and median O
3
concentrations within these different categories to assess how concentration trends are impacted by daily weather. Further
examine if trends vary for observations in the extremes of the O
3
distribution. We used k-means clustering to categorize days of
observation based on the maximum daily temperature, standard deviation of daily temperature, mean daily ground level wind
speed, mean daily water vapor pressure and mean daily sea-level barometric pressure. The five cluster solution was determined
to be the appropriate one based on cluster diagnostics and cluster interpretability. Trends in cluster frequency and pollution
trends within clusters were modeled using Poisson regression with penalized splines as well as quantile regression. There were
five characteristic groupings identified. The frequency of days with large standard deviations in hourly temperature decreased
over the observation period, whereas the frequency of warmer days with smaller deviations in temperature increased. O
3
trends
were significantly different within the different weather groupings. Furthermore, the rate of O
3
change for the 95th percentile
and 5th percentile was significantly different than the rate of change of the median for several of the weather categories.
We found that O
3
trends vary between different characteristic local weather patterns. O
3
trends were significantly different
between the different weather groupings suggesting an important interaction between changes in prevailing weather conditions
and O
3
concentration.
Journal of Exposure Science and Environmental Epidemiology advance online publication, 9 July 2014; doi:10.1038/jes.2014.45
Keywords: criteria pollutants; environmental monitoring; exposure modeling
INTRODUCTION
Regulations and technological improvements implemented over
the past decades have caused marked improvement in air quality
over many areas in the United States. In particular, pollutant trends
over the Northeastern United States have been shown to be
decreasing for many important pollutants including PM
2.5
,NO
2
,SO
2
and CO.
1–3
Trends for median O
3
concentrations have a more mixed
result, with some studies suggesting decreases while others suggest
no significant trends.
4–7
There is also evidence that background
O
3
concentrations have been increasing, perhaps because of an
increase in conditions that are favorable to O
3
formation such as
higher temperature and shifts in the NOx/VOC ratio.
5,8,9
Agood
understanding of the factors driving long-term changes in O
3
concentration is crucial in order to better inform policy decisions.
The strong seasonal patterns in ambient O
3
concentrations
have been well-described worldwide as well as in the North
Eastern United States.
10,11
The seasonality in O
3
concentrations is
related both to the effects of meteorological conditions on rates of
O
3
formation as well as to changes in the availability of precursors.
The meteorological factors relating to O
3
formation are intensity of
solar radiation, cloud cover, temperature, water vapor pressure
and boundary layer height. Changes in prevailing wind directions
exhibit seasonal dependence and also determine the sources and
concentrations of transported O
3
precursors.
12
Local spring-time ozone maxima have been observed at many
remote locations, thought to be less impacted by anthropogenic
emissions.
13,14
Two-main processes are thought to contribute to
these spring time maxima. These are stratosphere-troposphere
exchange of O
3
-rich stratospheric air into the troposphere as well
as increases in solar-radiation acting on O
3
precursors that have
accumulated over the winter period.
15
Profound stratosphere-
troposphere exchange events resulting in significant O
3
intrusion
into the troposphere are generally associated with extreme weather
events, such as thunderstorms, which can be identified through
local weather variables such as local wind speed, water vapor
pressure, barometric pressure and fluctuations in ground-level
temperatures.
16
Stagnation, which generally occurs in the North-
Eastern United States in the height of summer, is characterized by
low wind speed and high ambient temperature, leads to extremely
favorable conditions for O
3
formation.
17
Concentrations of anthropogenic O
3
precursors show a strong
seasonal dependence as well. The decomposition rate of
1
Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts, USA;
2
Department of Environmental and Occupational Health, University of
Washington School of Public Health, Seattle, Washington, USA;
3
Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts, USA and
4
Channing
Laboratory, Department of Medicine, Brigham and Women’s Hospital, Boston, Massachusetts, USA. Correspondence: Dr Elena Austin, Department of Environmental and
Occupational Health, University of Washington School of Public Health, 1959 NE Pacific Street, Seattle, Washington, WA 98195, USA.
Tel.: þ206 685 4471. Fax: þ206 616 2687.
E-mail: elaustin@uw.edu
Received 14 November 2013; revised 14 March 2014; accepted 30 April 2014
Journal of Exposure Science and Environmental Epidemiology (2014), 1 –11
&
2014 Nature America, Inc. All rights reserved 1559-0631/14
www.nature.com/jes
peroxyacetylinitrate, an important sink for both NO
x
and radicals,
is highly dependent on ambient temperature.
18
Isoprene
emissions from plants, a major VOC precursor to O
3
, have also
been shown to be highly temperature dependent.
19
Last, NO
x
concentrations are generally higher in winter because the reduced
oxidative capacity of the atmosphere results in longer residence
time. As these important precursors are strongly influenced by
temperature, this is yet another important facet of the complex
relationship between temperature and O
3
formation.
Larger time-scale changes in weather may also have an impact
on O
3
concentrations. Over the past three decades, clear trends in
weather conditions, including extreme temperature events, total
precipitation and humidity have been observed.
20,21
Furthermore,
changes in synoptic weather patterns, such as changes in the
strength of the North Atlantic Oscillation, El Nin
˜o–Southern
Oscillation, as well as changes in the frequency of the Mid-
Latitude Cyclone may contribute to altering concentrations of
transported air pollutants over New England.
22,23
Considering weather conditions when performing trend analy-
sis for O
3
is therefore crucial. Commonly, temperature and wind
speed are accounted for as a main effect in the regression model
used to calculate trend.
24
Including a wider range of meteoro-
logical variables in a regression model can be problematic
because of issues of correlation between weather variables.
Different approaches have been used to resolve this problem
including smoothing over a larger spatial area,
25
neural network
models,
26
cluster analysis and classification tree approaches
27,28
and probabilistic models.
29
In order to better understand how trends in air pollution are
impacted by short- and long-term scale changes in climate, we
propose to classify days of observation in a 16-year period based
on the multivariate relationship between local weather variables.
Within each of these identified groups, we will perform a trend
analysis of O
3
over the time period. This approach allows us to
identify how the O
3
concentrations are changing independently of
long-term changes in local weather patterns. Differences in trends
within each weather pattern group will also be examined for
quantiles of the O
3
distribution. Observed differences in trend can
then be ascribed to changes that are independent of weather
patterns such as changes in emissions of VOC and NO
x
precursors
to O
3
formation.
METHODS
The different steps undertaken in this Methods section are summarized in
Figure 1.
Data
The weather data used in this study included hourly measurements recorded
at the Boston Logan airport as provided by the National Oceanic and
Atmospheric Administration (NOAA) database. The data used were collected
during 1995–2010 and 24-h averages were estimated for days with at least
75% completeness in the hourly data. From 1995–2010, O
3
data measured at
four sites within the metropolitan Boston, MA, area were obtained from the
US EPA’s aerometric information retrieval system. We computed city-wide
averages to better capture regional distribution. Although the focus of this
paper is the ozone trends, concentrations of NO
2
,NO
x
,SO
2
and BC were
also retrieved from US EPA’s aerometric information retrieval system to aid
interpretation of the results. Trend analysis for these pollutants was
performed, in order to compare with the O
3
results.
The weather variables used in the cluster analysis included maximum
daily temperature, standard deviation of daily temperature, mean daily
ground level wind speed, mean daily water vapor pressure (absolute
water content of the air) and mean daily sea-level pressure. These
weather variables are used to characterize different physical properties
thus their distributions and units are different. For this reason, their values
were converted to z-scores, which were used for clustering. There were
87 days excluded from the analysis because of more than 25% missing
hourly data.
Clustering
Clusters were identified using k-means cluster analysis, Hartigan–Wong
algorithm.
30
We allowed for values of k(number of clusters) to range
between 2 and 10. To ensure that the solution selected is stable and not a
local minimum, the k-means algorithm was run 1000 times with 1000
different random initial seeds. The solution with the lowest sum of squares
within (SSW) was retained as the best solution for the given conditions.
The large number of initial seed values helped to ensure that the solution
did not correspond to a local minimum.
31
For all values of k, the algorithm
converged within 300 iterations.
The number of clusters (k) was selected using three main criteria. The
first was the desire to pick a solution that yielded a reasonable total SSW
value (keeping in mind that SSW decreases as the number of clusters
decreases). The second was a desire to minimize pollutant concentration
variability within the clusters of a given solution and the third was to
identify clusters that were interpretable based on our knowledge of
weather patterns in the Boston area.
The k-means algorithm used was implemented in the k-means function
of the stats package in R v. 2.15.2.
Comparing Clustering Solutions
The Rand Index is a measure of similarity between two different partitions
of the same data set. This index ranges between 0 and 1, where 0 indicates
that two data clusters do not agree on any pair of points and 1 indicating
that the data clusters are exactly the same. The Rand Index represents a
weight of the sites classified together in the two solutions versus the
sites classified separately. In this paper, we used the adjusted Rand Index
in order to compare different clustering solutions. First proposed by
Hubert and Arabie,
32
the adjusted Rand Index corrects the Rand Index for
the random chance that pairs are classified together. Steinley et al.
33
suggested that an adjusted Rand Index greater than 0.9 reflected excellent
agreement, values greater than 0.8 reflected good agreement, values
greater than 0.65 indicated moderate agreement and less than 0.65
indicated poor agreement.
33
Back-Trajectory Analysis
Back-trajectory paths were calculated using the HYSPLIT model (v. 4.9)
developed by NOAA. The meteorological archive used was the Eta Data
Assimilation System with 40 km resolution (EDAS40). For every hour of
every day from 2004 to 2009, an 84-h back-trajectory was computed in
HYSPLIT from the starting coordinates of the sampling site and a vertical
height of 750 m. The vertical movement of air parcels within the system
was modeled using an isentropic assumption.
34–36
Unfortunately, because
the EDAS40 data begins in 2004, it was not possible to obtain earlier back-
trajectory information in the same projection space.
The back-trajectory information begins in 2004 as this was the first year
available at this higher grid resolution. We use these data as a
representative sample of our data set which ranges from 1995 to 2010.
This approach is also supported by the sensitivity analysis, which shows no
great changes in the clustering results over the early and later parts of the
study period as discussed below.
Trend Analysis
Trend analysis of yearly cluster frequency was performed using both a
generalized additive Poisson model with a penalized spline term
representing the year, in order to allow for non-linear trends. The
penalized spline was calculated using the MGCV package in R v. 2.15.2.
O
3
trends were described using a linear regression model, controlling for
seasonality using sine and cosine functions and day of week as an indicator
variable. In addition, a quantile regression model was used to model
differences in trends for different quantiles of O
3
concentrations over time.
This model also controlled for seasonality with sine and cosine functions
as well as day of week using an indicator variable. Quantile regression
was first proposed by Koenker
37
as a robust alternative to traditionally
used mean estimator.
37
In addition, quantile regression allows us to
estimate the trend for different quantiles in the data, allowing us to
observe changes in the extremes of the data set, for example, trends in
baseline pollution and regional smog episodes. Quantile regression was
performed using the quantreg package in R v. 2.15.2.
Ozone trends and characteristic weather patterns
Austin et al
2
Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11 &2014 Nature America, Inc.
RESULTS
Selecting k
We selected the number of clusters (k) to be 5. Figure 2a shows
the fractional change in SSW for increases in kfrom 2 to 10. As k
increases the value of the SSW decreases; however, the decrease
slows significantly after k¼5. In part, this justifies selecting 5 as
the number of clusters.
We also considered the 4 and 6 cluster solution. However, the 5
cluster solution proved to be more interpretable based on general
knowledge of weather patterns throughout the Boston year. The 5
cluster solution showed a clear seasonal pattern, which was a goal
when we were seeking to partition these data.
Cluster Descriptions
The clusters were sorted based on the number of days per cluster,
from largest to smallest. Important meteorological and chemical
characteristics of each cluster are presented in Table 1. The
monthly distribution of each cluster is presented in Figure 3a.
Interpretation of the clusters is based on their mean characteristics.
Clusters 1 and 3 occur primarily in the winter months. Cluster 1
is characterized by high sea-level pressure and low boundary layer
height, whereas cluster 3 is characterized by lower sea-level
pressure, higher precipitation and higher boundary layer. Cluster 1
is higher in NO and BC, whereas cluster 3 is higher in O
3
. The
higher concentration of primary pollutants in cluster 1 may be
Figure 1. Methods summary chart.
Figure 2. Selecting the number of clusters (k). (a) Decrease in SSW. (b) Adjusted Rand Index comparing all solutions to the k¼5 solution.
Ozone trends and characteristic weather patterns
Austin et al
3
&2014 Nature America, Inc. Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11
reflective of decreased mixing on those days due to a lower
boundary height.
Clusters 2 and 4 are primarily in warm weather clusters. Cluster
4 occurred in June, July and August, whereas cluster 2 was more
spread out in time from May to October. The mean daily
temperature of cluster 4 is 23 1C, as compared with 16 1C for
cluster 2, and the water vapor pressure is significantly higher in
cluster 4. The meteorological conditions occurring in Cluster 4
are consistent with summer time regional pollution events
because of photochemical transformation of pollutants. This is
confirmed by the elevated PM
2.5
and O
3
concentrations as well as
very high maximum 8-h O
3
and the low NO concentrations in
cluster 4.
Cluster 5 occurs in the transition months (spring and fall)
and is characterized by moderate temperatures (13 1C) and
moderate O
3
and PM
2.5
. In addition, the standard deviation
of the hourly temperature values is highest in Cluster 5
than other clusters, suggesting strong diurnal temperature
differences.
In order to determine the persistence of different clusters,
we calculated the number of consecutive days belonging to
the same cluster (cluster runs). Figure 3b shows that the
median number of consecutive days ranges between 1 and 2.
There are infrequent episodes that occur where consecutive days
can be as high as 16. Cluster 5 seems to have the fewest such
consecutive days.
Table 1. Summary of the five clusters.
Cluster 1–1523 days Cluster 2–1308 days Cluster 3–1145 days Cluster 4–946 days Cluster 5–834 days
Winter (high
pressure)
Summer
(moderate)
Winter
(low pressure)
Summer
(stagnant)
Transition
Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD)
Temperature (1C) 2 (5) 16 (4) 4 (6) 23 (3) 13 (6)
Max T past 3 days (1C) 7 (6) 21 (6) 9 (7) 26 (4) 16 (7)
Max T today (C) 6 (5) 19 (4) 8 (6) 29 (3) 20 (6)
SD hourly temperature (1C) 2 (1) 2 (1) 2 (1) 3 (1) 4 (1)
Water vapor pressure (mbar) 5 (2) 14 (4) 6 (3) 20 (4) 9 (3)
Sea-level pressure (mbar) 1022 (6) 1016 (6) 1007 (7) 1013 (5) 1017 (7)
Ozone (p.p.b.) 18 (9) 22 (9) 24 (9) 35 (12) 27 (11)
8-H max ozone (p.p.b.) 26 (10) 32 (11) 31 (9) 52 (17) 39 (14)
Wind speed (m/s) 5 (1) 4 (1) 7 (2) 5 (1) 5 (1)
Boundary layer height (m) 686 (329) 620 (226) 1086 (399) 785 (224) 801 (312)
PM2.5 (mg/m
3
) 10 (5) 10 (5) 7 (4) 16 (9) 10 (5)
Sulfur dioxide (p.p.b.) 7 (4) 3 (2) 4 (3) 3 (2) 4 (3)
Nitrogen dioxide (p.p.b.) 23 (7) 19 (6) 17 (6) 19 (7) 21 (8)
Nitrogen monoxide (p.p.b.) 24 (20) 15 (11) 12 (8) 9 (6) 16 (12)
Black carbon (mg/m
3
) 0.9 (0.6) 1.0 (0.6) 0.6 (0.4) 1.1 (0.6) 0.9 (0.6)
Precipitation per day (mm) 1 (5) 3 (10) 6 (12) 2 (7) 1 (4)
Bolded rows represent variables used in clustering.
Monthly
10
10 25
0
0
% of Days
1
2
3
4
5
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5
Consecutive Days in each Cluster
1412108642
Number of Consecutive Days
Figure 3. Distribution of clusters. (a) Monthly distribution. (b) Distribution of consecutive days by cluster.
Ozone trends and characteristic weather patterns
Austin et al
4
Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11 &2014 Nature America, Inc.
Wind Directions
Ground-level hourly wind direction measurements were available
for every day of observation. We compared the percent of hours
within eight direction quadrants for each cluster. Figure 4 shows
the results of this analysis. In cluster 1, 45% of the ground level
wind directions were from the N-W quadrant. For cluster 2, 22% of
the hourly wind directions were from the NE-E direction. Cluster 3
had 55% of the ground wind coming from the SW-NW direction.
Cluster 4 had 60% of the wind coming from the S-W direction and
cluster 5 had ground level wind directions ranging from S to NW.
Because ground level wind direction does not necessarily indicate
the direction of the prevailing winds, and potentially the origin
of transported pollutants, we also wanted to plot hourly
back-trajectories for the days in our observation period.
Figure 5 shows the distance and location from Boston of all the
hourly back-trajectories that were less than 1500 m above ground
level in the 72 h before arriving in Boston. The visualization was
done with the heR.misc package.
38
This figure further
demonstrates the differences between clusters that were
observed in the same season. The winter clusters have very
different profiles with cluster 3 (associated with high boundary
layer height) showing a significant contribution of air masses from
the NW. The summer clusters also have very different profiles with
cluster 4 (regional pollution) showing a strong contribution from
air masses to the West of Boston that is not seen in cluster 2. The
transition season cluster has similar patterns of back-trajectories as
cluster 4 (regional pollution).
Sensitivity Analysis
For the cluster analysis, we wanted to determine the sensitivity of
the solution to the initial conditions. First, we determined how
sensitive the solution was to the number of clusters selected. As
shown in Figure 2b, the 4 and 6 cluster solution moderately agrees
with the 5 cluster solution, whereas for other kvalues, the cluster
solution is significantly different and less interpretable.
We also confirmed that clustering the data from the beginning
of the observation period and the end of the observation period
would not produce significantly different results. For this analysis,
the data set was separated into two parts, one with dates before 1
January 2003 and the second with dates starting at 1 January
2003. The two data sets were clustered into five clusters as
described above. These two solutions were compared with the
initial clustering that was obtained using the entire data set.
Comparing the early and late data sets to the solution obtained
from all the years combined shows excellent comparability
(adjusted Rand index of 0.95 and 0.88, respectively). Therefore,
we conclude that it is appropriate to cluster and interpret the data
over the entire observation period.
We examined how sensitive the solution was to missing days
of observation in the data set. We randomly removed 10% of the
Figure 4. Percent of observed hours within each direction bin by
cluster.
Km
Km
Km
Km
Km
Km
Cluster 1
Cluster 4 Cluster 5
Cluster 2 Cluster 3
Figure 5. Back-trajectory analysis 0–72 h before sampling day and 0–1000 m from ground level.
Ozone trends and characteristic weather patterns
Austin et al
5
&2014 Nature America, Inc. Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11
days (576 days) to create 100 new data sets. We then clustered
these 100 data sets and compared the results to our solution using
the adjusted Rand index. The mean adjusted Rand index across all
the data sets is 0.97 (SD ¼0.06).
Trend in Yearly Weather Patterns
A trend analysis of the yearly frequency of each cluster as a
function of the year of observation was performed. The use
of a penalized spline to model the trend allows for non-linear
relationships. The results are presented in Figure 6. There are no
clear trends for clusters 1, 3 and 4. However, there is strong
evidence that there is a linear decrease in the frequency of cluster
5 (1.3% per year), which corresponds to transition days and a
linear increase in the frequency of cluster 2 (1% per year), which
corresponds to mild summer days.
Pollutant Trends
First, we examined the overall trend in concentrations of O
3
,NO
2
,
SO
2
, BC and PM
2.5
over the entire observation period without
accounting for cluster groupings (Figure 7) using both linear
regression and quantile regression. We further examined the
differences in the trends for O
3
by cluster. The results for O
3
are
presented in Figure 8. The trend values and standard errors are
presented in Table 2. There are important differences in the trends
both across clusters as well as across quantiles for the same
cluster. Across all the clusters, the background concentrations of
O
3
are increasing over time, although the magnitude of this
difference is much smaller for cluster 4 than the other clusters.
The trends in the 95th percentile values of O
3
are quite different
over the different clusters. Cluster 4 has the strongest decline
in the 95th percentile value of O
3
(0.69 p.p.b./year), whereas
cluster 5 has the strongest increase in the 95th percentile value of
O
3
(0.42 p.p.b./year). The differences in trend by cluster were also
examined for the other pollutants, but there were no significant
trend differences by cluster, results are not presented here.
In order to better understand the relative impact of changes in
weather frequency vs changes in concentration within weather
cluster, we calculated the change in O
3
concentration attributable
to each process, over the 16 years of observation, as shown in
Equation 1.
d½O3¼XfdDi
365d½O3i16yþd½O3i
Di
365dgð1Þ
Where d½O3is the change in O
3
over 16 years
dD
i
is the change in the number of days in cluster iover 16
years
½O3iis the median concentration of O
3
within cluster i
d½O3iis the change in concentration of O
3
within cluster iover
the 16-year period
Figure 6. Trend in cluster frequency over time.
Ozone trends and characteristic weather patterns
Austin et al
6
Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11 &2014 Nature America, Inc.
mean (LSE) fit
median (LAE) fit (p>0.05)
median (LAE) fit (p<=0.05)
mean (LSE) fit
median (LAE) fit (p>0.05)
median (LAE) fit (p<=0.05)
mean (LSE) fit
median (LAE) fit (p>0.05)
median (LAE) fit (p<=0.05)
mean (LSE) fit
median (LAE) fit (p>0.05)
median (LAE) fit (p<=0.05)
mean (LSE) fit
median (LAE) fit (p>0.05)
median (LAE) fit (p<=0.05)
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
Ozone Trend
Black Carbon Trend PM2.5 Trend
Sulfur Dioxide Trend Nitrogen Dioxide Trend
1995 2000 2005 2010 1995 2000 2005 2010 1995 2000 2005 2010
1995 2000 2005 20101995 2000 2005 2010
Ozone (ppb)
Sulfur Dioxide (ppb)PM2.5 (ug/m3)
Black Carbon (ug/m3)
Nitrogen Dioxide (ppb)
806040200
010203040
0102030405060
012345
0 10203040
Figure 7. Quantile regression and linear regression of O
3
,NO
2
,SO
2
, BC and PM
2.5
by year.
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
All Days Cluster 1
Cluster 3 Cluster 4 Cluster 5
Cluster 2
Figure 8. Trend analysis of ozone by cluster.
Ozone trends and characteristic weather patterns
Austin et al
7
&2014 Nature America, Inc. Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11
Di
365dis the average number of days (per year) belonging to
cluster i
Results of this analysis are presented in Table 3. In general,
the change in yearly concentration attributable to change
in concentration within cluster is much higher than changes in
concentration attributable to changes in yearly frequency. In
addition, we can see that the change in trend for the baseline is
much higher than for the median. The 95th percentile concentra-
tion of O
3
is decreasing overall, but this is mainly due to important
decreases in the 95th concentration trend of cluster 4. We also
observe that clusters 1 and 2 are both strongly contributing to the
overall increasing trend in O
3
concentration throughout the year.
This analysis is instructive in contrasting how changes in weather
and changes in rates of formation of O
3
are driving the overall
differences being observed.
We determined that including categorical values for month
instead of describing seasonality as a periodic function did not
significantly affect the trend estimates. We also examined whether
the assumption of linearity in the trends of O
3
was reasonable by
also modeling the trends using basis splines. There was no
evidence of strong deviation from linearity.
DISCUSSION
The weather clusters identified represent five main daily weather
types that occur in the Boston area. These categories are based
on the maximum daily temperature, standard deviation of the
daily temperature, ground level wind speed, water vapor pressure
and the sea level pressure. Examining the trends in O
3
concentration within each of these characteristic meteorological
groupings allows for a better understanding of the trend of O
3
over time, independent of trends in climate and changes in the
duration and start of seasonal conditions. Furthermore, our results
demonstrate that modeling mean O
3
values does not provide
adequate information on O
3
trends at the extremes. This has been
previously demonstrated in other instances.
39
The overall trend in
baseline O
3
(0.27±0.07 p.p.b./year) as determined by the quantile
regression of the 5th percentile in O
3
concentration over the
time-period agrees with those of other studies reporting trends in
O
3
baseline over the West Coast of Ireland, (0.31±0.12 p.p.b./
year)
40
and off the West Coast of the USA, 0.5–0.8 p.p.b./year.
9
We know that weather strongly influences O
3
formation and
removal. Examining the trend in O
3
concentrations within the five
identified clusters allows for the investigation of the relationship
changes in O
3
concentrations independent of changes in weather
patterns. Over the 16-year time-period, the sensitivity analysis
confirmed that the cluster types did not change significantly
rather there has been a clear change in the yearly frequency of the
clusters over time. There was a decrease in the frequency of
cluster 5 (1.3% per year) and an increase in the frequency of
cluster 2 (1% per year). Cluster 2 corresponds to warmer days in
the summer with higher nighttime temperatures, whereas cluster
5 corresponds to days in the transition seasons with lower
nighttime temperatures and mild daytime temperatures. There-
fore, we have observed a decrease in transition season days and
an increase of days with summer characteristics in the spring and
fall seasons. Furthermore, cluster 2 corresponds to days with
Easterly winds. Although 16 years of data are a short time to make
any inference about long-term changes in temperature trends,
this observation does coincide with climate change studies that
Table 2. Rate of change of ozone from 1999 to 2010.
Quantile Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Overall
Change
per year
(p.p.b./
year)
Standard
error
Change
per year
(p.p.b./
year)
Standard
error
Change
per year
(p.p.b./
year)
Standard
error
Change
per year
(p.p.b./
year)
Standard
error
Change
per year
(p.p.b./
year)
Standard
error
Change
per year
(p.p.b./
year)
Standard
error
0.05 0.36 0.07 0.38 0.07 0.31 0.11 0.10 0.11 0.24 0.10 0.27 0.03
0.25 0.28 0.05 0.25 0.05 0.26 0.07 0.02 0.11 0.21 0.06 0.24 0.03
0.5 0.30 0.04 0.23 0.05 0.09 0.05 0.06 0.10 0.11 0.08 0.18 0.03
0.75 0.28 0.04 0.11 0.07 0.07 0.05 0.28 0.12 0.18 0.08 0.11 0.03
0.95 0.23 0.07 0.08 0.10 0.15 0.08 0.69 0.25 0.42 0.14 0.02 0.05
Gray entries are not statistically different than 0 (PZ0.05).
Table 3. The change in O
3
(p.p.b.) attributable to changes in cluster frequency and changes in concentration within cluster over the 16-year period
of observation.
Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Overall
change
Change in median O
3
(p.p.b.) over the 16-year period
Due to Dfrequency 0.23 0.66 0.06 0.23 0.86 0.26
Due to Dconcentration 1.30 0.88 0.36 0.01 0.23 2.76
Total 1.53 1.54 0.30 0.24 0.63 2.51
Change in 5th percentile O
3
(p.p.b.) over the 16-year period
Due to Dfrequency 0.07 0.26 0.03 0.12 0.32 0.13
Due to Dconcentration 1.36 0.98 1.05 0.20 0.45 4.04
Total 1.43 1.24 1.03 0.08 0.13 3.91
Change in 95th percentile O
3
(p.p.b.) over the 16-year period
Due to Dfrequency 0.46 1.14 0.10 0.38 1.43 0.31
Due to Dconcentration 1.03 0.39 0.34 1.79 0.44 1.05
Total 1.49 0.75 0.44 2.17 0.99 1.36
Ozone trends and characteristic weather patterns
Austin et al
8
Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11 &2014 Nature America, Inc.
note a decrease in the daily standard deviation of temperature
being recorded during the winter and transition seasons in the
northeast.
41,42
Classifying based on daily weather allows for important trends
to be observed. In particular, baseline concentrations on stagnant
summer days remain unchanged. By contrast, baseline concentra-
tion in mild summer conditions are increasing more rapidly than
data set as a whole (0.38±0.14 p.p.b./year). Winter clusters also
show strong increases in baseline O
3
concentrations. The high
boundary layer winter days (cluster 3) show slight decreases
(0.15±0.14 p.p.b./year) in upper quantiles of O
3
. These days are
characterized by low pressure, air masses from the NW and high
precipitations. The decrease in maximum O
3
concentrations is
possibly due to increases in winter precipitation over the NE that
have been observed in climate change studies.
43
Winter days with
low boundary layer height show an increase of 0.23±0.13 p.p.b./
year in the 95th percentile. These days are associated with low
precipitation and higher concentrations of primary pollutants,
including NO. The increase in the 95th percentile concentrations
of O
3
may indicate that over time on these days, O
3
formation is
increasingly limited by NO
x
concentrations. Ozone concentration
on transition season days have increases in the 95th percentile at
a rate of 0.42±0.26 p.p.b./year. This agrees with observations of
recent increases in spring maxima O
3
concentrations.
13
Significant
decrease in O
3
in the upper quantiles of cluster 4 ( 0.70±
0.55 p.p.b./year) agrees with previous evidence that effective
emission controls have decreased summertime smog episodes
over NE North America.
3
Comparing this clustering method to examining O
3
trends by
season demonstrates the added value in categorizing days based
on observed weather patterns rather than by fixed seasonal time
periods. Figure 9 shows the results of the trend analysis of O
3
by season. The increase in transition season maxima is not at
all visible with this analysis method. Furthermore, differences
between low and high pressure days are also not evident. The
decrease in higher quantile values in the stagnant summer
conditions is captured, however, perhaps because these days
tend to be highly concentrated within the summer months. The
magnitude of the decrease in trend is decreased as compared
with the clustering method. We also performed a second analysis
that additionally controlled for daily temperature as well as season
and observed no significant differences between this model and
the model that only accounted for season.
The results of the trends in frequency and the trends in O
3
concentration within cluster were combined in order to assess
their respective contributions on total O
3
concentration (Table 3).
The overall change in O
3
concentration is strongly driven by
changes occurring within clusters 1 and 2. Both these clusters
have markedly low mixing heights as compared with the other
clusters. In addition, they both have significantly higher concen-
trations of NO. The increase in O
3
observed here appears to be
associated with photochemistry of locally emitted pollutants as
opposed to regional transported pollutants.
In addition, the results of Table 3 suggest that there is an overall
increase in the 95th percentile of O
3
in the transition weather
group. This trend is particularly concerning given recent analyses
suggest increased susceptibility to air pollution in the spring time.
Within 10 Canadian cities, springtime weather was associated with
a stronger response to O
3
among the elderly. In addition, they
observed higher RR estimates in the springtime.
44
Likewise,
stronger associations between air pollution and mortality and
morbidity have been observed in the springtime.
45,46
For comparison we estimated the trends of other pollutants.
The trends of NO
x
,SO
2
, BC and PM
2.5
in this time period are shown
in Figure 7. For these four pollutants, the quantile regression
results indicate a stronger decrease for the higher percentile
values than for the low-level values, indicating that background
concentrations are not changing as quickly as the higher
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
1995 2000 2005 2010
Ozone (ppb)
806040200
Date
Percentiles: 5th, 25th, 50th, 75th, 95th
2000 2005 2010
Ozone (ppb)
806040200
Winter
Summer Fall
Spring
Figure 9. Trends in O
3
by season.
Ozone trends and characteristic weather patterns
Austin et al
9
&2014 Nature America, Inc. Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11
concentrations. This suggests that emission controls are having a
positive impact on high ground level concentrations. Although
not shown here, there is no observable difference in the trends of
these pollutants by cluster, suggesting that the trends in these
pollutants are not weather dependent.
Because the NO
x
/VOC ratio is an important factor in the rate of
formation of O
3
, we wanted to investigate whether the differences
observed for the different weather clusters are related weather-
dependent trends in this ratio. When accounting for daily NO
x
concentration in the trend analysis, the trends for clusters 1, 2, 3
and 5 become insignificant or marginally positive. On the other
hand, the trends of O
3
within cluster 4 remain consistent with
those observed when not accounting for NO
x
. This suggests that
other than in stagnant summer conditions, cluster-specific
changes in O
3
were associated with similar decreases in NO
x
. This
is due to the reaction of ozone with NO. This mechanism is more
important during the winter and the night when the atmosphere
is more stable and the local primary emissions containing NO have
more impact on the transported ozone. However, this mechanism
becomes less important during the summer episodes when ozone
formation is enhanced and there is more vertical mixing.
CONCLUSIONS
The method presented here groups days of observation into
meaningful and interpretable groupings that reflect characteristic
weather patterns. These groupings can then be used to examine
changes in O
3
concentrations over time. Seasonal analysis has
been successfully used to identify differences in pollutant trends
as well as differences in response to air pollution. However, this
method allows for greater precision in identifying the exact
calendar days that represent a cohesive weather grouping. The
use of quantile regression allows for the identification of
important trends that are occurring at the extremes of the
distribution and that may be missed in a mean regression. These
changes that are occurring at the extremes can be extremely
important and yield information that is relevant to policy and
decision making. The analysis demonstrates the effectiveness of
emission control policies to diminish maximum summertime O
3
concentrations. However, it also demonstrates that baseline O
3
concentrations are significantly increasing, particularly in winter
and moderate summer conditions. Furthermore, there is evidence
that there is the need for increased O
3
controls in the transition
months in order to curb the rising trend in maximum values.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
ACKNOWLEDGEMENTS
This publication was made possible by USEPA grant 834798. Its contents are solely
the responsibility of the grantee and do not necessarily represent the official views of
the USEPA. Further, USEPA does not endorse the purchase of any commercial
products or services mentioned in the publication.
Research reported in this publication was supported by the National Institute of
Environmental Health Sciences of the National Institutes of Health under award
number T32ES015459, NIEHS 1R21ES020194 and PPG NIEHS P01 ES009825. The
content is solely the responsibility of the authors and does not necessarily represent
the official views of the National Institutes of Health.
REFERENCES
1 Kelly VR, Lovett GM, Weathers KC, Likens GE. Trends in atmospheric concentration
and deposition compared to regional and local pollutant emissions at a rural site
in southeastern New York, USA. Atmos Environ 2002; 36: 1569–1575.
2 Fann N, Risley D. The public health context for PM2. 5 and ozone air quality
trends. Air Quality. Atmos Health 2013; 6: 1–11.
3 Parrish DD, Singh HB, Molina L, Madronich S. Air quality progress in North
American megacities: a review. Atmos Environ 2011; 45: 7015–7025.
4 Lin CC, Jacob DJ, Fiore AM. Trends in exceedances of the ozone air quality
standard in the continental United States, 1980–1998. Atmos Environ 2001; 35:
3217–3228.
5 Cooper O, Parrish D, Stohl A, Trainer M, Ne
´de
´lec P, Thouret V et al. Increasing
springtime ozone mixing ratios in the free troposphere over western North
America. Nature 2010; 463: 344–348.
6 Bojkov RD, Bishop L, Fioletov VE. Total ozone trends from quality-controlled
ground-based data (1964–1994). J Geophys Res Atmos 1995; 100: 25867–25876.
7 Logan J, Staehelin J, Megretskaia I, Cammas J, Thouret V, Claude H et al. Changes
in ozone over Europe: Analysis of ozone measurements from sondes, regular
aircraft (MOZAIC) and alpine surface sites. J Geophys Res Atmos 2012; 117: 1–23.
8 Chan E, Vet R. Baseline levels and trends of ground level ozone in Canada and the
United States. Atmos Chem Phys 2010; 10: 8629–8647.
9 Jaffe D, Ray J. Increase in surface ozone at rural sites in the western US. Atmos
Environ 2007; 41: 5452–5463.
10 Bloomer BJ, Vinnikov KY, Dickerson RR. Changes in seasonal and diurnal cycles of
ozone and temperature in the eastern US. Atmos Environ 2010; 44: 2543–2551.
11 Logan JA. Tropospheric ozone: seasonal behavior, trends, and anthropogenic
influence. J Geophys Res Atmos 1985; 90: 10463–10482.
12 Schichtel BA, Husar RB. Eastern North American transport climatology during
high-and low-ozone days. Atmos Environ 2001; 35: 1029–1038.
13 Monks PS. A review of the observations and origins of the spring ozone max-
imum. Atmos Environ 2000; 34: 3545–3561.
14 Allen DJ, Dibb JE, Ridley B, Pickering KE, Talbot RW. An estimate of the strato-
spheric contribution to springtime tropospheric ozone maxima using TOPSE
measurements and beryllium-7 simulations. J Geophys Res 2003; 108: 8355.
15 Vingarzan R. A review of surface ozone background levels and trends. Atmos
Environ 2004; 38: 3431–3442.
16 Stohl A, Bonasoni P, Cristofanelli P, Collins W, Feichter J, Frank A et al.
Stratosphere-troposphere exchange: A review, and what we have learned from
STACCATO. J Geophys Res Atmos 2003; 108: 8516.
17 Banta RM, Senff CJ, White AB, Trainer M, McNider RT, Valente RJ et al. Daytime
buildup and nighttime transport of urban ozone in the boundary layer during a
stagnation episode. J Geophys Res 1998; 103: 22519–22544.
18 Sillman S, Samson PJ. Impact of temperature on oxidant photochemistry in
urban, polluted rural and remote environments. J Geophys Res Atmos 1995; 100:
11497–11508.
19 Petron G, Harley P, Greenberg J, Guenther A. Seasonal temperature variations
influence isoprene emission. Geophys Res Lett 2001; 28: 1707–1710.
20 Peterson TC, Zhang X, Brunet-India M, Va
´zquez-Aguirre JL. Changes in North
American extremes derived from daily weather data. J Geophys Res 2008; 113:
D07113.
21 Simmons A, Willett K, Jones P, Thorne P, Dee D. Low-frequency variations in
surface atmospheric humidity, temperature, and precipitation: Inferences from
reanalyses and monthly gridded observational data sets. J Geophys Res Atmos
2010; 115: 1–21.
22 Leibensperger EM, Mickley LJ, Jacob DJ. Sensitivity of US air quality to mid-
latitude cyclone frequency and implications of 1980–2006 climate change. Atmos
Chem Phys 2008; 8: 7075–7086.
23 Doherty R, Stevenson D, Johnson C, Collins W, Sanderson M. Tropospheric ozone
and El Nin
˜o–Southern Oscillation: Influence of atmospheric dynamics, biomass
burning emissions, and future climate change. J Geophys Res 2006; 111: D19304.
24 Lou Thompson M, Reynolds J, Cox LH, Guttorp P, Sampson PD. A review of
statistical methods for the meteorological adjustment of tropospheric ozone.
Atmos Environ 2001; 35: 617–630.
25 Tai A, Mickley L, Jacob D, Leibensperger E, Zhang L, Fisher J et al. Meteorological
modes of variability for fine particulate matter (PM2. 5) air quality in the United
States: implications for PM2. 5 sensitivity to climate change. Atmos Chem Phys
2012; 12: 3131–3145.
26 Hooyberghs J, Mensink C, Dumont G, Fierens F, Brasseur O. A neural network
forecast for daily average PMosub410o/sub4concentrations in Belgium.
Atmos Environ 2005; 39: 3279–3289.
27 Davis J, Eder B, Nychka D, Yang Q. Modeling the effects of meteorology on ozone
in Houston using cluster analysis and generalized additive models. Atmos Environ
1998; 32: 2505–2520.
28 Huang L, Smith RL. Meteorologically-dependent trends in urban ozone. Environ-
metrics 1999; 10: 103–118.
29 Cox WM, Chu S. Meteorologically adjusted ozone trends in urban areas: a
probabilistic approach. Atmos Environ B Urban Atmos 1993; 27: 425–434.
30 Hartigan J, Wong M. A k-means clustering algorithm. J R Statis Soc C 1979; 28:
100–108.
31 Steinley D. K-means cluste ring: A half-century synthesis. Br J Math Stat Psychol
2006; 59: 1–34.
Ozone trends and characteristic weather patterns
Austin et al
10
Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11 &2014 Nature America, Inc.
32 Hubert L, Arabie P. Comparing partitions. J Classif 1985; 2: 193–218.
33 Steinley D. Properties of the Hubert-Arable Adjusted Rand Index. Psychol Methods
2004; 9: 386.
34 Draxler R, Rolph G. HYSPLIT (HYbrid Single-Particle Lagrangian Integrated
Trajectory) model access via NOAA ARL READY website (http://www.arl.
noaa.gov/ready/hysplit4.html). NOAA Air Resources Laboratory, Silver Spring.
2003.
35 Rolph GD, Draxler RR. Sensitivity of three-dimensional trajectories to the
spatial and temporal densities of the wind field. J Appl Meteorol 1990; 29:
1043–1054.
36 Rolph G. Real-time Environmental Applications and Display sYstem (READY)
Website (http://www.arl.noaa.gov/ready/hysplit4.html). NOAA Air Resources Labo-
ratory, Silver Spring. Silver Spring, MD, 2003.
37 Koenker R, Bassett Jr G. Regression quantiles. Econometrica 1978; 33–50.
38 Klepeis NE. Exposure Science Website. http://ExposureScience.Org, 2004.
39 Baur D, Saisana M, Schulze N. Modelling the effects of meteorological variables on
ozone concentration—a quantile regression approach. Atmos Environ 2004; 38:
4689–4699.
40 Derwent R, Simmonds P, Manning A, Spain T. Trends over a 20-year period from
1987 to 2007 in surface ozone at the atmospheric research station, Mace Head,
Ireland. Atmos Environ 2007; 41: 9091–9098.
41 Bradley RS, Griffiths ML. Variations of twentieth-century temperature and pre-
cipitation extreme indicators in the northeast United States. J Clim 2007; 20:
5401–5417.
42 Brown PJ, Bradley RS, Keimig FT. Changes in Extreme Climate Indices for the
Northeastern United States, 1870–2005. J Clim 2010; 23: 6555–6572.
43 Hayhoe K, Wake CP, Huntington TG, Luo L, Schwartz MD, Sheffield J et al. Past and
future changes in climate and hydrological indicators in the US Northeast. Clim
Dyn 2007; 28: 381–407.
44 Vanos JK, Cakmak S, Bristow C, Brion V, Tremblay N, Martin SL et al. Synoptic
weather typing applied to air pollution mortality among the elderly in 10 Cana-
dian cities. Environ Res 2013; 126: 66–75.
45 Zanobetti A, Schwartz J. The effect of fine and coarse particulate air pollution on
mortality: a national analysis. Environ Health Perspect 2009; 117: 898–903.
46 Franklin M, Schwartz J. The impact of secondary particles on the association
between ambient ozone and mortality. Environ Health Perspect 2008; 116: 453.
Ozone trends and characteristic weather patterns
Austin et al
11
&2014 Nature America, Inc. Journal of Exposure Science and Environmental Epidemiology (2014), 1 – 11