United States
Environmental Protection
Agency
Office of Air Quality
Planning and Standards
Research Triangle Park NC 27711
EPA-450/4-81-015
February 1981
Air
U.S. Environmental Protection
Agency Intra-Agency Task
Force Report on Air
Quality Indicators
-------
EPA-450/4-81-015
February 1981
U.S. Environmental Protection Agency
Intra-Agency Task Force Report
on Air Quality Indicators
by
W.F. Hunt, Jr. (chairman), G. Akland, W. Cox, T. Curran,
N. Frank, S. Goranson, P. Ross, H. Sauls, and J. Suggs
U.S. Environmental Protection Agency
Office of Air, Noise, and Radiation
Office of Research and Development
Office of Planning and Management
Region 5
Prepared for
U.S. ENVIRONMENTAL PROTECTION AGENCY
Office of Air, Noise, and Radiation
Office of Air Quality Planning and Standards
Research Triangle Park, NC 27711
February 1981
U.S. Environmental Protection Agency
Region V, Library
230 South Dearborn Street
Chicago, Illinois 60604
-------
This report is issued by the Environmental Protection Agency to report technical data of
interest to a limited number of readers. Copies are available - in limited quantities - from
the Library Services Office (MD-35), U.S. Environmental Protection Agency, Research
Triangle Park, North Carolina 27711; or, for a fee, from the National Technical Infor-
mation Service, 5285 Port Royal Road, Springfield, Virginia 22161.
Publication No. EPA-450/4-81-015
11
U,S. Environmental Protection Agency
-------
CONTENTS
Figures
Tables
1. Overview 1-1
1.1 Recommendations 1-1
1.2 Future work 1-4
2. Known Measures of Data Uncertainty (Precision and
Accuracy) 2-1
2.1 Need for qualifying air pollution data 2-1
2.2 Schedule for implementing the precision and
accuracy program 2-3
2.3 Uses of precision and accuracy data 2-3
2.4 References 2-4
3. Detecting and Removing Outliers 3-1
3.1 Causes of outliers 3-1
3.2 Statistical procedures for identifying outliers 3-2
3.3 Recommended treatment of outliers 3-6
3.4 Conclusions 3-7
3.5 References 3-7
4. Area of Coverage and Representativeness 4-1
4.1 Background 4-1
4.2 Network description 4-3
4.3 Representativeness 4-4
4.4 References 4-6
5. Data Completeness and Historical Continuity 5-1
5.1 Data completeness 5-1
5.2 Historical completeness 5-9
5.3 References 5-15
6. Statistical Indicators and Trend Techniques 6-1
6.1 Statistical indicators 6-1
6.2 Trend techniques 6-7
6.3 References 6-12
ill
-------
CONTENTS (continued)
7. Inferences and Conclusions
7.1 Background
7.2 Case studies
7.3 Long-term solutions
7.4 References
8. Data Presentation 8-1
8.1 Concepts to be displayed 8-1
8.2 Chart types and uses 8-2
8.3 Classification of data 8-29
8.4 Input parameters, data transformations, and
statistical comparisons 8-30
8.5 Audience applicability 8-30
8.6 Caveats and suggestions 8-32
8.7 Available plotting resources 8-33
8.8 Guidance for selection of charts 8-33
8.9 Summary and recommendations 8-34
8.10 Future issues 8-34
8.11 References 8-35
9. Continuity of Year-To-Year Reports 9-1
9.1 Method changes 9-1
9.2 NAMS network changes 9-1
IV
-------
FIGURES
Number Page
1 Empirical Histograms of All Possible Annual
Averages Based on One or Two Months of Data
Per Quarter Compared With Theoretical Fre-
quency Distributions Defined by Statistical
Model for 1973 SO2 at the New York City
Laboratory Site 5-8
2 Theoretical Probability Distribution of Annual
Mean S02 With 10 Sites in Each Year 1974-1976
and 15 Sites in 1977 and 1978 5-13
3 Comparison of SO2 Trends at Bayonne, N.J.,
With Regulations Governing % Sulfur Content
in Fuel 7-3
4 Weekly Average CO and Wind Speed in Richmond,
VA From January 4 to February 28, 1974 7-5
5 CO Air Quality From 18 Monitoring Sites and
Motor-Vehicle Gasoline Consumption for N.J.
From 1972 Through 1976 7-7
6 Quarterly TSP Maximum Values in Region VI
From 1972 to 1977 Illustrating the Effect
of the 1977 Dustorm 7-9
7 Satellite Views of February 23-25, 1977, Dust-
storm at Succeeding Time Periods 7-10
8 Status and Trends in Air Quality in Colorado 8-4
9 Intersite Correlation Test Data 8-5
10 Magna, Utah, Day 3, 0.50 Probability Ellipses
of the West-East and South-North Wind Com-
ponents for Three Cluster Types. Winds From
the West and South are Positive 8-6
11 Twenty-Four Hour TSP Values, 1972 8-7
12 Air Quality Data, 24-h TSP Concentration
Values, October 15, 1976 8-9
-------
FIGURES (continued)
Number Page
13 Maximum 1-h 03 Values/Day, 1977 (SAROAD
Site 141220002P10) 8-10
14 Oxidant Trends Adjusted for Meteorology 8-11
15 Annual Average and Second-High Day TSP
Values, 1970-75 8-12
16 Population Exposure Distributions of Annual
Mean TSP for 1970 and 1976 in City of
Chicago 8-13
17 Ambient CO Concentration and Gasoline
Consumption, 1972-77 8-14
18 Comparison of Monthly GM, 12-mo Running GM,
and Predicted Monthly Means (By Double
Moving Average Method), 1964-74 8-15
19 Trends in CO Levels in New York's 45th Street
Station, April-January, 1975-77 8-16
20 Trends in PSI Levels, 16 Cities, 1973-76 8-17
21 Actual vs. Potential Emissions for Illinois,
Tons/Year 8-19
22 Number of Days Per Year That the TSP Primary
Standard or Alert Level was Exceeded,
Colorado 8-20
23 Regional Changes in Metropolitan Population
Exposures to Excess TSP Levels, 1972-1977
(Width of Each Regional Column is Propor-
tional to its Metropolitan Population) 8-21
24 Trends of Annual Mean TSP Concentrations
From 1970 to 1976 at 2350 Sampling Sites 8-22
25 Wind Rose Pattern 8-23
26 Source Category Contributions to Particulate
Air Pollutants 8-24
27 Air Quality Status (TSP) and Trends in 25
Largest Urban Areas in EPA Region 5 8-25
vi
-------
FIGURES (continued)
Number Page
28 Isopleths of TSP Concentrations (yg/m3) in
EPA Region V and Iowa for October 15, 1976 8-26
29 Air Quality Status, Colorado, 1972 8-27
30 Three Dimensional Plot of NC>2 Concentrations,
ug/n\3, November 1973 8-28
Vii
-------
TABLES
Number Page
1 Special Checks and Audits for Estimation
of Precision and Accuracy 2-2
2 Concentration Ranges for Automated Analyzer
Audits 2-2
3 Outlier Tests 3-9
4 NADB Validity Criteria 5-2
5 Continuous Measurement Summary Criteria 5-3
6 Data Completeness for Continuous S02
Monitoring, 1973 5-4
7 Accuracy of Particulate Sampling Frequencies
and Averaging Intervals 5-6
8 Deviation of Observed Annual Mean and Maximum
From True Values Among 42 TSP Sites,
1976-78 5-6
9 Annual Means 5-14
10 Weekly Average CO Concentrations (ppm) and
Windspeeds (mph) in Richmond, VA, January
4-February 28, 1974 7-4
11 Multiple Parameter Listing, 1979, ug/m3 8-3
12 Weekly TSP Maximums at City Sites, 1979,
Vig/m3 8-3
13 Region 5 Monthly Averages by Site Type, 1979,
Ug/m3 8-3
14 Outline of Input Parameters, Data Transforma-
tions, and Statistical Comparisons 8-31
Vi 11
-------
1. OVERVIEW
The Intra-Agency Task Force on Air Quality Indicators was established to
recommend standardized air quality indicators and statistical methodologies for
presenting air quality status and trends in national publications. As a first
step, the members of the task force identified topics of concern and prepared a
series of issue papers on these topics; these papers discuss the background and
current status of each issue, develop recommendations, and identify areas that
need additional work. These individual papers make up the remaining sections
of this document.
To put the activities of the task force in perspective, it should be no-
ted that on May 10, 1979, EPA promulgated regulations for ambient air quality
monitoring and data reporting. These regulations were a result of the ground-
work of the Standing Air Monitoring Work Group (SAMWG), and reflect EPA's con-
cerns about data quality, timeliness, and consistency from one area to another.
Specific provisions in these regulations instituted the routine reporting of
precision and accuracy information to aid in characterizing data quality, the
designation of specific sites in state and local agency monitoring networks to
be used in national trends analyses, and the increased standardization of sit-
ing criteria to ensure greater uniformity. All ambient air quality data re-
ceived by EPA should reflect these changes by 1981; the data bases currently
used in EPA's analyses are in transition. In a sense, the monitoring communi-
ty has identified and begun to implement improvements in the air quality data
bases, and now those responsible for analyzing the data must ensure that the
best use is made of these improvements.
1.1 RECOMMENDATIONS
In developing recommendations, two common concerns were apparent. The
first involved data bases that do not yet exist (e.g., precision and accuracy
information). Since it is premature to recommend how these data should be
used in reporting air quality, the Task Force has simply identified the group
1-1
-------
that should take the initiative in developing methods for using the informa-
tion. The second concern involved the relative merits of standardization.
The data from a given air quality study can be analyzed by a wide variety of
statistical techniques. In many cases, different approaches are equally ac-
ceptable. In fact, in certain cases, there are statistical techniques that
are recommended but seldom applied. The important point here is that an em-
phasis on standardization should not discourage the development and applica-
tion of new techniques. Consequently, these recommendations should be view-
ed as a set of general principles rather than a set of inflexible rules. If
a particular approach satisfies the intent of these recommendations, it is
satisfactory; if it does not, an explanation should be included in the analy-
sis to say why alternative techniques were used.
The recommendations are grouped in four categories: data base, data
analysis, data interpretation, and data presentation. In each category, both
general and specific points are presented.
1.1.1 Data Base
In general, each analysis should indicate what data were used. For small
studies, specific sites can be named; for large studies, it will suffice to in-
dicate the data source (e.g., the National Air Data Branch, NADB) and the se-
lection criteria used to choose sites. Specific recommendations are listed be-
low.
1. Precision and Accuracy (Section 2) - EPA does not require the sub-
mission of precision and accuracy information until 1981; therefore
no guidance on its use will be given at this time. It is recommend-
ed that ORD/EMSL take the lead in developing procedures for using
this information.
2. Data Screening (Section 3) - Statistical procedures for detecting
outliers are available, and some have been incorporated into SAROAD.
Under the new monitoring regulations and management plan for the
National Air Monitoring Stations (NAMS), users will eventually be
able to assume that NAMS data quality has been verified. In the in-
terim, however, the user should apply appropriate screening proce-
dures to the data for any small-scale analysis; for large-scale
analyses, the user may rely on robust statistical techniques that
will minimize the potential impact of anomalous data.
3. Site Selection (Sections 4 & 5) - The NAMS will provide a usable,
quality assured, standardized data base for trendsparticularly
national trends. Composite values of NAMS data will provide a
1-2
-------
useful index for national trends assessment. In the interim, the
user must select sites on the basis of specific criteria which en-
sure adequate completeness and seasonal balance. The criteria
should be stated clearly in the analysis.
1.1.2 Data Analysis
The analysis should be structured so that results are stated in terms of
statistical significance. Analyses that have no statistical basis should in-
dicate that they do not and why they do not. Those analyses with a statisti-
cal basis should indicate the statistical approach used. More specific recom-
mendations concerning data analysis follow.
1. Choice of Summary Statistics (Section 6) - Summary statistics should
reflect the appropriate air quality standard and not be biased due
to sample size. If an analysis requires the use of a statistic that
is biased with sample size, care must be taken to ensure that compar-
isons over time or across sites are not affected by differences in
sample sizes.
2. Comparability of the Data Base (Sections 5 & 6) - Any trends analy-
sis should be structured so that results are not attributable to the
data base varying with time. Interpolated data may be used for the
visual presentation of trends, but trend statistics should be based
on actual data unless the effect of interpolation can be quantified.
3. Trend Techniques (Section 6) - Standard statistical techniques such
as Chi-square, nonparametric regression, aligned-rank tests, analy-
sis of variance, and time series are all acceptable means of assign-
ing probability statements to trends analyses. The primary concern
is that the tests used are statistical in nature, not which tests
are used. However, EPA groups need to take a more active role in
applying various techniques to air data to assess the relative mer-
its of different procedures.
1.1.3 Data Interpretation
An analyst can facilitate the interpretation of air quality monitoring
data from the existing NAMS network by using other sources of information that
help explain why an air quality trend has or has not taken place or why there
are significant differences between sets of air quality data. To better assess
the effectiveness of EPA's emission control program, the agency should collect
data on all variables that impact air quality in at least two major urban areas,
1.1.4 Data Presentations
Data presentations should be consistent with the analysis, and should be
adequately labeled so that they can stand alone.
1-3
-------
1. Choice of Scales - Distortion of scales should be avoided. In gen-
eral , the pollutant concentration axis should start at zero concen-
tration.
2. Distinguish No-Data Cases - Presentations involving shading (e.g.,
maps) should clearly indicate cases with no data as a distinct cate-
gory.
1.2 FUTURE WORK
In view of the transition occurring in the air quality data bases, it is
recommended that this task force be continued during 1981, when NAMS data and
precision and accuracy data will be received initially by EPA. During this
time, the following tasks should be performed:
1. Assessment of Statistical Manpower - To ensure that the task force
recommendations can actually be implemented and are not merely idealized
goals, it will be necessary to appraise the available and planned techni-
cal resources. A breakdown of resources needed should be provided, with
particular attention to the in-house resources needed to provide continu-
ity and technical guidance.
Lead Group: 0PM
Target Date: March 81
2. Use of Precision and Accuracy Information - The eventual use of -
precision and accuracy information needs to be better defined. Although
these data are not currently being received by EPA, similar preliminary
information is available to EMSL. These data should be examined, and a
plan should be developed on how this type of information can be incorpo-
rated into EPA's use of air quality data. Consideration should be given
to the feasibility of eventually establishing national performance stand-
ards for precision and accuracy data.
Lead Group: ORD/EMSL
Target Date: July 81
3. Use of Site Information in Trends Analyses - An important feature of
the NAMS data base is the detailed information available describing indi-
vidual sites. To routinely make efficient use of this information, it
will be necessary to identify the relevant site parameters and to develop
computer software to link the site information with the air quality in-
formation. Attention should be given to stratifying the data into broad
classifications needed for more refined data analysis.
Lead Group: OAQPS/MDAD
Target Date: September 81
1-4
-------
4. Assessment of Statistical Software - Several statistical software
packages are available that could be used for air quality data analysis.
To ensure that EPA statistical manpower is efficiently utilized an as-
sessment should be made of what statistical software is applicable. In
particular, this assessment should determine: (1) what statistical pack-
ages are being used and to what extent and (2) if the best statistical
packages are being employed and if not, why not.
Lead Group: ORD/EMSL/RTP
Target Date: September 81
5. Presentation of Data - Because the Task Force is recommending specif-
ic types of data presentations a pilot study should be initiated to ensure
that these recommendations are feasible to implement on a routine basis.
Attention should be given to identifying computer programs that would
facilitate these presentations and if gaps exist to develop the necessary
programs to the extent possible with existing resources.
Lead Group: Region V
Target Date: September 81
1-5
-------
2. KNOWN MEASURES OF DATA UNCERTAINTY (Precision and Accuracy)
Concepts of precision and accuracy have been the subject matter for many
presentations and publications. ASTM Committee E-ll on quality control of ma-
terials discussed definitions and implications of these two ideas over a 10-
year period. Complete agreement on the meanings of precision and accuracy is
unlikely to be found in the literature on scientific measurement systems. In
assessments of the quality of ambient air data, EPA uses estimates of preci-
sion to characterize the relative capability of the monitoring system to re-
peat its results when measuring the same thing and estimates of accuracy to
2
characterize the closeness of an observation to the "truth."
Beginning January 1, 1981, EPA will require all state and local report-
ing agencies to calculate precision and accuracy estimates in a prescribed man-
ner and to qualify all data entered into the EPA data bank with quarterly pre-
cision and accuracy estimates. Some, but not all, of the components necessary
2
for estimating precision and accuracy (Appendix A, 40-CFR-58) are available.
Currently, collocated samplers and single concentration precision checks provide
data which can be used to estimate the precision of manual and automated sam-
plers, while audits of flew rates and laboratory analytical measurements provide
data which can be used for estimating accuracy. Perhaps improved methods for
estimating these parameters will be developed as a result of the increased at-
tention to assessing air quality data.
Table 1 presents the types and frequencies of special checks for precision
and accuracy by pollutant measurement method; Table 2 displays the concentration
range for each audit level.
2.1 NEED FOR QUALIFYING AIR POLLUTION DATA
Over many years, the air pollution data bank has grown into a gigantic
body of computerized records of concentrations of pollutants measured at sites
across the Nation at points in time. Necessarily, great amounts of attention
and expense have been devoted to devising systems to process the data into a
2-1
-------
TABLE 1. SPECIAL CHECKS AND AUDITS FOR ESTIMATION OF
PRECISION AND ACCURACY
Precision
Accuracy (local audit)*
Automated analyzers
(S02> CO, N02, 03)
Type check
Frequency
Scope
One concentration
Biweekly
All monitoring instruments
Manual methods
Type check
S02
N02
TSP
Frequency
Scope
Collocated samplers at two
sites
Each monitoring day
Two sites (of high concen-
tration)
Three or four concentrations
25% of the analyzers each quar-
ter; at least one per quarter
All analyzers each year
Type of audit
Flow
NA
NA
One level
25% of the sites
each quarter; at
least once per
quarter
All sites each
year
Analytical
Three levels
Three levels
NA
Each analysis-
day; at least
twice per
quarter
NA
*See Table 2 for audit levels.
TABLE 2. CONCENTRATION RANGES FOR AUTOMATED ANALYZER AUDITS
Concentration range, ppm
Audit level
1
2
3
4
S02, N02, 03
0.03-0.08
0.15-0.20
0.40-0.45
0.80-0.90
CO
3-8
15-20
40-45
80-90
2-2
-------
computerized retrievable form. Also, high priority has been given to the de-
velopment and refinement of a system for collecting data from state and local
agencies and sending the data through EPA Regional Offices to the NADB com-
puter system. Inevitably, air pollution control administrators and affected
industries are concerned about the quality of the data that result from this
great expenditure of resources, especially pollutant measurements that may
represent exceedances of National Ambient Air Quality Standards (NAAQS) or
that have other serious implications.
As a result of continued programs of quality assurance and technological
improvements in monitoring and analysis, today's ambient air data are doubt-
lessly more representative of true concentrations than those of some years ago.
However, these improvements in data quality cannot be quantified because no
routine, standardized data assessment program has been in effect. Implementa-
tion of the program described above should provide the means to evaluate prog-
ress in measuring and recording ambient air data and should give managers a
higher level of confidence in making decisions based on air pollutant measure-
ments.
2.2 SCHEDULE FOR IMPLEMENTING THE PRECISION AND ACCURACY PROGRAM
All designated NAMS sites are to begin operation January 1, 1981 {Appen-
dixes A and E, 40-CFR-58). On July 1, 1981, the first quarterly report is due
into EMSL; the first quarterly summary report from EMSL to the EPA Regional
Offices is due September 1, 1981; and the first annual report is scheduled for
July 1, 1982. The remaining State and Local Air Monitoring Stations (SLAMS)
are to be phased into the Precision and Accuracy Reporting System as soon as
possible after January 1, 1981. The distinction of SLAMS and NAMS is of lit-
tle relevance here, since precision and accuracy data relate only to agencies.
2.3 USES OF PRECISION AND ACCURACY DATA
Confidence in conclusions concerning air quality trends will be more
scientifically defensible with the precision and accuracy data because of the
increased capability to test for statistical significance in trends analyses.
Also, additional interpretive insights may be gained; for example, if accuracy
does not change significantly, the quality assurance program may be ruled out
as a factor affecting trends.
2-3
-------
For many trends study purposes, the analyst must set an arbitrary limit
(e.g., +15%) on sites to be included in an analysis. The analyst must make
judgments in light of the purpose and the user requirements. In all cases,
the analyst should test to determine any effect of precision and accuracy in-
formation on conclusions.
It is premature at this time to attempt to develop definitive criteria
for using precision and accuracy data in determining trends, nonattainment
status, and so forth. Inclusion of precision and accuracy data in analyses
may appear to add to the overall levels of uncertainty, but experience over
time is needed to establish criteria for setting limits on the data to be
used in a particular situation.
Precision and accuracy data will permit comparisons of data quality
within and between monitoring networks and within and between States and re-
gions. Also, precision and accuracy data together with data from EPA's Am-
bient Air Performance Audit Program should provide the means for effective
evaluations of analytical laboratories hired by State and local agencies.
A great deal of interest centers on the possible uses of precision and
accuracy data to report probability intervals about peak and mean estimates
of air quality data. If the peak value is not an outlier, there is no prob-
lem with the probability interval. Long-term (3 years) studies may be re-
quired to verify the assumptions that the sample of precision and accuracy
information is representative and that extreme values are within the popula-
tion. There should be no problem in reporting probability intervals for
mean values. In any case, the assessed quality of air pollution measurements
should be included in all relevant reports and publications.
2.4 REFERENCES
1. Precision Measurement and Calibration, Statistical Concepts and Proce-
dures. National Bureau of Standards Special Publication 300. Volume 1.
February 1969.
2. Rhodes, R. C. Precision and Accuracy Data for State and Local Air Moni-
toring Networks: Meaning and Usefulness. Paper presented at the 73rd
APCA Annual Meeting, Montreal, Canada, June 1980.
3. Appendix A. Quality Assurance Requirements for State and Local Air Moni-
toring Stations (SLAMS). FR Vol. 44, No. 92, pp. 24574-27581, May 10,
1979.
2-4
-------
3. DETECTING AND REMOVING OUTLIERS
An outlier is an observation that does not conform to the pattern estab-
lished by other observations. This pattern may be a scatter plot, frequency
histogram, time series, or simple listing. The parent population from which
the observations are drawn is usually assumed to behave according to a certain
probabilistic model, and the observations are usually assumed to be drawn at
random or at least independently. The agreement between the observations and
the parent population depends both on the correctness of the underlying assump-
tions and on certain aspects governing the selection process. Outliers should
not be isolated from other problems of statistical analysis; they should be in-
cluded among anomalies such as nonadditivity, nonconstant variance, bias, tem-
poral drift, or wrong model specification. A correction in another problem
can often solve the outlier problem.
3.1 CAUSES OF OUTLIERS
Three basic ways an outlier can occur are (1) mistakes in readings, (2)
wrong model specification, and (3) rare deviation.
Mistakes in readings can occur during any stage of data processing, but
the more common mistakes (e.g., transcription errors) usually occur early in
the processing; during data coding or key punching, transcription errors often
go unnoticed. Unusual readings from instruments may be caused by power fail-
ure or surges, improper calibration, breakdowns, torn filters, contamination,
chemical interaction, leaks, and so forth. Not adhering to an experimental
plan or design can affect recorded data. Mistakes occur both in totally auto-
mated data gathering processes and in those that rely on human consideration
or intervention.
With the exception of glaring mistakes that have obvious explanations,
model specification usually is the basis for deciding if a discordant observa-
tion is an outlier. In fact, most tests currently being used routinely by EPA
2
to detect outliers are actually testing some type of nonnormality. For ex-
ample, the null hypothesis of all the tests listed in Table 3, with the
3-1
-------
exception of the gap test, is that the population from which the observations
are being drawn is specified by a normal distribution model. References 3 and
4 describe additional tests which specify normal, lognormal, or Weibull models.
Detection of an outlier by any one of these tests is paramount to declaring
the parent population has a different distribution than was originally assumed.
Of course, the final conclusion would be that the observation is an outlier.
In any parent population, some observable events have a low probability
of occurring. These events are usually associated with the extremes (tails)
of the distribution. Although the probability of obtaining one of these ob-
servations is small, it is possible for such an event to occur. Being a rare
occurrence, the observation is almost always treated as an outlier.
3.2 STATISTICAL PROCEDURES FOR IDENTIFYING OUTLIERS
Statistical procedures for identifying outliers are used for different
reasons. One reason is to justify what would have been done anyway (i.e., to
reject observations that an experienced investigator would intuitively reject).
Another reason for using statistical tests is to provide algorithms to the
2 5
computer ' for scanning large sets of data that would be impractical to scan
by visual inspection. At present, 5122 SLAMS, 1269 NAMS, and 68 Inhalable
Particulate Network (IPN) sites are producing data which require validation.
In addition, special studies such as the Philadelphia IP Study, the National
Forest Ozone study, and the Across Freeway Study (LACS) are collecting large
quantities of continuous and 24-hour data. Whether the data are collected
routinely or for special studies, a procedure for detecting outliers is neces-
sary to strengthen the validity of conclusions reached during data analysis.
Table 3 lists several statistical tests for objectively screening a set
of data and identifying possible outliers. These tests are currently being
applied to routine data storage systems within EPA. These tests, except for
the studentized range, the studentized t-test, and the Shewhart control chart
require no external or independent parameter estimates; conclusions are based
solely on the data at hand. Except for the Gap Test, these tests for out-
liers are tests for normality. All tests are aimed at high values relative
to some measure of spread based on the sample. Where routine screening of
data is necessary, a battery of several tests is advisable.
The Dixon ratio test was the first of many standard statistical procedures
which have been found to work well in air data screening. Some work better
3-2
-------
than others, depending on sampling frequencies and durations. For example,
the Shewhart control chart was found superior to the Dixon ratio in screening
2
24-hour measurements. These and other screening procedures constitute the
Air Data Screening System (ADSS), which is now being implemented in 27 States
2
through the Air Quality Data Handling System.
In another screening program, several statistical procedures are current-
2
ly being applied to SAROAD data. With the exception of the gap test, these
tests have one thing in common--they all assume that the observations are sam-
ples from a single normal population with specific location and shape parame-
ters. Tables and charts are used to identify values that have a low probabili-
ty of occurring if all observations were taken from the same population; these
values are flagged for further investigation.
To demonstrate how outliers can be identified, several tests, in Table 3
are applied to the following set of mass data in the 0-15 micron size range
gathered over a 5-month period in Birmingham, Alabama, using a dichotomous sam-
pler.
24-hour samples
Date
08/01/79
08/07/79
08/25/79
08/31/79
09/06/79
09/12/79
09/30/79
10/06/79
10/12/79
10/18/79
10/24/79
10/30/79
11/05/79
11/23/79
12/05/79
12/11/79
yg/m3
43.8
66.7
17.8
45.3
92.6
34.8
38.4
47.5
64.7
36.2
30.6
16.9
36.6
15.3
101.8
35.0
Monthly values
Average (x)
43.4
55.3
39.2
26.0
68.4
Range (R)
48.9
57.8
47.8
21.3
66.8
3-3
-------
DJxon Ratio2'3
X - X
Calculated: R99 = ^ - ^ = im'o " ?H = 0-418.
t-C. A - A.} iUl.O - i/.O
Tabulated: rQ gQ (i.e., a = 0.10) = 0.454 for n = 16.
Thus: R22 < r0>9Q.
Shewhart Control Chart2'3
Range: LCLR = D3R = (0)(43.95) = 0.
UCLR = D4R" = (2.57)(53.95) = 112.95.
In this example, the largest integer less than the average was used (n = 3).
Average: LCL9 = I - A9R = 40.95 - (1.02)(43.95) < 0 (use LCL7 = 0 in this
A £ A
case).
= I + A2R = 40.95 + (1.02)(43.95) = 85.78.
Thus the monthly average (68.4) and range (66.8) for December are not consid-
ered outliers.
Chauvenet's Criterion^
A = 45.25
S = 25.00
= 101.8 - 45.25 ,, 9(.
S 25.00
The critical value of C for sample size n = 16 is 2.154 with an a-level of
0.016 for a one-tailed test, thus 101.8 is an outlier. Chauvenet's criterion
is not recommended for samples of n < 7 for a two-tailed test and for samples
of n < 4 for a one-tailed test because it tends to flag too many valid obser-
vations.
Grubbs Test7'8
Calculated: SS = 9372.26, SS16 = 5961.16, and SS15 lg = 3161.25.
Thus: Lj_ = ~^- = 0.64 and l_2 = ^>16 = 0.34
Tabulated: Lj = 0.576 and L2 = 0.405 for n = 16.
3-4
-------
Therefore, at the a = 0.05 significance level, only the highest value (101.8)
can be rejected as an outlier.
Coefficient of Skewness
n o
n £ (Xn. - X)
Calculated: B = ~~ = 1.0
I (X.-X)2 3/2
1=1 1
Normally, this test for skewness is not used for n small (say < 25); however,
the value of B exceeds the tabulated value 0.71 for n = 25 (the tabulated val-
ue for n = 16 will be < 0.71) and therefore, normality is rejected at the a =
0.05 significance level.
After excising 101.8 and 92.6, normality can no longer be rejected; both
values are outliers based on this test.
g
Studentized T-Test
The standard deviation (S) required for testing Studentized deviates was
estimated from the August, September, and October data to be 21.52. November
and December data are grouped together because there are only two observations
3
in each month. The suspected outlier is 101.8 yg/m .
T X" " X 101.8 - 47.18 _ o ,-,
1 S 21.52 *-w'
The critical value (a = 0.05) for examining one outlier among four samples,
relative to the prior S based on 11 degrees of freedom, is tQ 05(4,11) = 2.24.
is
9
3
Thus 101.8 yg/m is an outlier.
Studentized Range"
The S required for the Studentized range test was estimated from the same
set of data as the Studentized t-test. November and December data are grouped
together as before, and the suspected outlier is again 101.8 yg/m .
Calculated: W = Xn - Xj = 101.8 - 15.3 = 86.5.
Tabulated: W = qQ Q5(4,11)S = 4.26(21.52) = 91.68
Thus: 101.8 is not an outlier.
Some tests are more suited to hourly than to 24-hour data. Table 3 summarizes
the recommended applications of the commonly used tests. For examples of how
other tests are applied, consult the references.
3-5
-------
References 3 and 4 describe nonparametric tests for comparing two or more
data sets to aid in identifying a data set that does not conform to the pattern
established by the other data sets. Reference 4 also discusses the use of non-
parametric tests for a single sample. However, identifying an outlier by a non-
parametric test is inconsistent with our definition of an outlier; an outlier
is an atypical observation because it does not conform to a model which we hypo-
thesize to describe the data. If we have no model in mind, it is difficult to
describe what we mean by an outlier. Hence, nonparametric techniques are recom-
mended only for the comparison of data sets.
3.3 RECOMMENDED TREATMENT OF OUTLIERS
No observation should be rejected solely on the basis of statistical tests,
since there is always a predictable risk (i.e., the a-level) of rejecting "per-
fectly good" data. Compromises and tradeoffs are sometimes necessary, especial-
ly in routinely scanning large amounts of data; in these situations, no statis-
tical rule can substitute for the knowledge and judgement of an experienced
analyst who is thoroughly familiar with the measurement process.
In any data-collecting activity, all data must be recorded along with any
notes that may aid in statistical analysis. Gross mistakes should be corrected
if possible before performing calculations in the final analysis; if a mistake
cannot be explained or corrected, it is not always wise to discard the reading
as though it had never occurred. Further experimental work may be needed, since
a gross error in the observations can bias the analysis in all but the most ro-
bust statistical procedures. In addition, data collected under differing condi-
tions should not be combined for identifying outliers unless the experiment was
designed (e.g., factorial design) to handle such data with an appropriate model
during final analysis.
In performing the calculations in the final statistical analysis, the
question of what weight, if any, to assign a discordant value is difficult to
answer in general terms. An acceptable explanation for an outlier should pre-
clude any further use of the value. Sometimes it is obvious that an observa-
tion does not belong even though there is no explanation for its existence.
Although the value should not be used in further calculations, it should be
mentioned in the final report.
q
Experienced investigators differ greatly on these matters. Excluding
"good" data may not be as serious as including "bad" data and then excluding
3-6
-------
any questionable observations in further calculations at the risk of losing
information on estimates and introducing some bias. Reducing sample variation
(increasing precision) may be preferred over introducing a slight bias, es-
pecially if the bias is theoretically estimable. Using robust air quality
indicators that are not strongly influenced by outliers may avoid many argumen-
tive, subjective decisions.
3.4 CONCLUSIONS
The data analyst should not assume that all outliers represent erroneous
values. Some outliers occur because the analyst has used the wrong probabilis-
tic model to characterize the data. For example, outlier tests based on as-
sumptions of normality may be inappropriate for nonnormal data sets. In addi-
tion, there is always a finite chance that an extreme value will occur natural-
ly. The analyst should carefully investigate these possibilities before dis-
carding outliers. The use of robust air quality indicators is recommended
since they decrease the need for detecting outliers. Reference 4 discusses
several outlier tests, their performance, and provides a brief summary of com-
parative information for each test, including pertinent references.
3.5 REFERENCES
1. International Encyclopedia of Statistics. Vol. 2. The Free Press, New
York, 1978. pp. 1039-1043.
2. Guideline Series. Screening Procedures for Ambient Air Quality Data.
EPA-450/2-78-037, July 1978^
3. Nelson, A. C., D. W. Armentrout, and T. R. Johnson. Validation of Air Mon-
itoring Data. EPA-450/2-78-037, August 1980.
4. Barnett, Vic and Toby Lewis. Outliers in Statistical Data. John Wiley
and Sons, Inc. New York, New York, 1978.
5. Northrop Services, Inc. A Data Validation Program for SAROAD. Cont. No.
68-02-2566. ESG-TN-78-09, December 1978.
6. Dixon, W. J. Processing Data for Outliers. Biometrics. 9(l):74-89,
March 1953.
7. Grubbs, F. E., and G. Beck. Extension of Samples, Sizes and Percentage
Points for Significance Tests of Outlying Observations. Technometrics
14(4):847-854, November 1972.
3-7
-------
8. Tietjen, G. L., and R. H. Moore. Some Grubbs-Type Statistics for the De-
tection of Several Outliers. Technometrics 14(3):583-597. 1972.
9. Natrella, M. G. Experimental Statistics. Chap. 17. National Bureau of
Standards Handbook 91, 1963.
10. Duncan, A. J. Quality Control and Industrial Statistics. Richard D.
Irwin, Inc., 1965.
11. Quality Assurance Handbook for Air Pollution Measurement Systems. Volume I,
EPA-600/9-76-005. U.S. Environmental Protection Agency, Environmental Moni-
toring and Systems Laboratory, Research Triangle Park, N.C., January 1976.
3-8
-------
CO
u
u
Q£
U
H
C
cr
u
5
I
o c
*£
C *-"
ft! "
E u
U Q.
Ol D.
CC IB
I/I
Ol
3
O
_
IB
JS OJ
C 01
« s
VI IB
C Ol
O 3 1. VI
VI IS IB
Ol > V> 3
*&. ""«
E in 3 ^
«0-^'D^-
ift «/>»- ^
. *D > C
_^[j"> .*. ^_
1 [^ ^?
51W5
i»
£
VI
01
J3
5
i *
IB
«p»
Ow
O) "PO
«/> I.
Oi b E
01 ai «
u j: in 01 vi
0 4J C
C >>» C
§o ** ^ o
*J L» «*
+* c a> **
sKllp-
^g-S.-hl
> EP- Ol vi C
5 « P--P- o
< (J > Q.-0 c
E i
« i.
in « 0^
C P- 01 VI
« *» 01
It- > «JJ=
O U Ol
at § a -olc
5 ,2
IB "x 15 "~ 2
> g > in
E E cttn
"-in O -p- O o
X 01 U C VIP-
Sp- 01 -P- J3
a. v> E 3 E
VI O
" " . >
I. U-
T 5 u
C C -! *> CO
x x x o xi
» » O"^ ro» in
^^ ^^ CM * PM M PO r\j
x. v|x vlx v^< v|
1 C, 1 C. 1 C. 1 C.
CV CV CV CV
x ro'x cox 'x «
^i ^5 ^5 ^i
< PH CM CM
1 1 1 1
C C C C
X X X X
1 1 1 1
c e c c
X X X X
1 1 I I
O P* -H CM
J^
*
*o
2
g
X
o
«n
0.2
Sf
«o
I/) V»
0>JC|
a> CNJ]
5 *^-
A
x
^- I/I
5S
c c
£ £
1-3 Si
O u in-p-
>*- P- *J VI
- *!§ js
* +*~ 5
p- IB
vi N o in
ai «I C
p 4J C
J3 TJ C 0
CO
W IB O Ol
C
w 'fX x)
p- O C
U V 01
Ol ** Ci
toc?5-o
ai oi£ c in i-S
"f~ *^ I- VI *" ^3
in-p- oi<«- Ol *J^O CP-
*> E > Ol C 3 S
3 « a. « E
OP- ai E c E c
coincaicoo
OIP- O C VI-P-
p-L^L.incoiCOi3
OI 3OI3-p-CjO
> C in-O O-* O-E-P--P-
OlBC-p-tBCUS^I.
^ u A a > 4^ vi in p *J
& "O 4-* £
01 OI-P- IB
> vi in E v>
at CO *J *j =
""e- 1 1 «*~ §
>«- £ - .p. 4-)l
o « ai i cooi
M 01 IB L. C
01 fO P** * 4-1 4-* *P-
cnit- u o o vi cxi
co a t- u coc
« > 4J 4-* O O 01
s- vi IB c c u a.
O. O O DO)
aisp- u u curt -a
01 0 P- i.
ios.toi-i.raa>-
U. Ol L. CD OI 4J1OC
Ol Ol S CX « «
>p- > O CX OP--O
IB « O t 3 3COI
- U > 01 «l
p- XI p
10 5
gff "
u
in 4J o
3 C ^
* ^^
in c
f in L. in
i IB o ai
i
e-o -o
U * ft)
5 CM
P- IB ««. IB
S 4JPH *J
in p v»4
0) 5
H- C
Om^
« V» «sk
-p- p in «-H vi
U Ol 4J 4J
ai > vi v in
iS'p? * g£
g, tp- l~
i- at *> s ' _e
^~ ^j Oy«r_ E ^ Q
C ^3 O E OJ
O *» 01 01 « 1- C
4-> > c VI 3 O
O P--P- 4J u
01 ai >> IB u 01 in
3t-oip-cfooiai
"10 o.'o. oi ai 5*5.
> E B *o 4^ ^ " 01 E
OaC3Cl8>iB
VI 1 O
01-0 -i
4-> p- V. 1 Ol
in Q. « x co>
01 E XI IB X
en to C J-
L. VI IB > 01 -<
5 4J C W B >CU)I
P- C VI -p- « -p-
II II 1
c
X I/I IO IX t»
D< 1
xC
1*-*
0
"c
o
11
"E
u
I/I
a>
c
ai
ID
JC
o
Ol
3-9
-------
o
c
0
u
CO
LU
_1
"o c
Ol O
O
01
10 01
"V O. V»
,o!o-o
g-Sx
i in
in X k.
a, g_ra_a;
« C 44
> in a) 3
1- 3 O
f. in er
i a a>
i .a in ID
in i
O VI
. 01
O "O C
k, 0)1
O J= 3
>t- 44 in
in
o U. 10
- 0
in CD
ai in
f 44 It.
£1 C O
5*0 C X
IB «"* +* i»
O 3 10
u -.a E
OlO C
a. L. o
(/> «B 44 c
i
k.
i«- ai k.t
o -O in 44 «
3 0 E
44 C -5 O
U 44 C
ai c
c m
01 C Ol X
ji in c
C.Q L. O
3 O Ol-
io T3 44
1- C«t- C 3
x IO O 3 J9
Oi in
H
in a
VI
f
44 U-
» O
Ol k.
O C
23
5 S
X C
CM
^ CM
D< ***
1 D<
f- 1
X
«~ X
e M i -i
CM!
e "~
i
CD
o
gs
u v>
si
o in
o
in
si
i
k. 44
S."u
u
°°c-
01 44 a
I C 4i
io o m
44 Q. 01
3
IO O1
U 44
oi e
o. 01 a
Vrt U O
I
0) O Ol
01 44 > C 1
k. Ol Ol *. 3
10 .C Ol X3 X^
T3 k.'Z
O O 44 k. Ol 44
o 44 a «ITB in
+4 -a ci-
ts oi c 3-0
01 01 k. io
3k. 44 .«^-
IO VI C «
E io Ol *^ C o
o aif *i o
< U E ** « C 44
i i
a. >* i c i
3 4» c a
a e o c
. 44 O tflf
« OJ44
03 L. i.-
3 > CL Q.V
t- oi EO
ID -o E a
> o m 44
T> C C
44 in ai k. io cio
Oli- eo -S-O 01 C
Ol O. k. c 01 k. 01
k. E 01 «J4-> S- Q.S-
4*foaoio
in a in E u-o *4
I II
X IX l/l
^^,
1
i L>
c
x 1
i
c
1
1
"S
N
44
1
I/I
ID
0
§
I/I
1
O
> in
f in
l 1 Ol l
io c .c ai
1 C 44 k.
44 10 O
k. 3
o -Q « oi e
It. V. M 44 O
k. 44 m
< 44 C E 44
in in eo
01 O -M
< -o a. m >
J3 01 01
SgiJU"0
ID k. C T3 k. OL eoi-
Q.-- 01 01 44 3
I/) 44 Q.^ in 0*
§1
in - k>
3 t! In p
"eo 0.-5 O
> 10 C
Ol
44 O i» O
X O Ol
1^ 44 *O 44
Ifc.
° E
*4 a
Ol in
in
c
IQ
H.
e o
r
44
oi oi ai
3 3 in
*<0 "<0 0
> in >
ffl in
x in c c ai
1 c 1 2"cL
1 II
x x z
i^
X
1
c
X
1
=»
"l
IO
k.
o
Ol
N
44
1
I/)
in
in ai
01 3
3 i
IO
IO >
i-l CM
(
^ It
> O
CL O
C 44
°t
a §
oi m
tn in
,10
E «
in E
D 3 O
en o c
i i 1
Si k. in i it- ttj oi Q. x
en ai 3 c O Ek.>ai g 44
o> o 3 *- n V o 5-r-
44mk. c 44 k. a. s vi-
io > oi o Q. E « in g
c *~ k.'a.44 c oi in k.
e >> a. E Q. c ^ oik.o
oi IOEX oi*4i«j:oic
^44OIM344 ^ .44-Q
ai cj= m-r- oi cx c t-
^SS*J-J'II3'5 «-S5-Sc
a CL. E 10 44 k.c
oi foaio oik,33o
44Q.
Ill
1-1 CM
1 1
XXX
i-l CM
1 1
x*~ x*~
1 1
X X
I 1
o~ a~
t i
~ S
M
c
i
0
44
1
01
IO
3-10
-------
T3
0)
C
CO
CO
1
o c
0) o
O f-
C 4->
§10
u
o J^
u Q.
Of O.
a: re
in
a>
10
,_
ID
U
"C
U
4j
in
a>
s-
£
in
in
(O
m
in
s
1*-
0
c
o
c
1
|
1
u
4^
in
0)
r
i
ID
C
4-»
in
t- i «e o
O 4-> 4J
>i ai i c c
i 4-> 0) c 01 i oi in
ID 3 f- «r- 0) OJ 4J 2 O)
T> O i- C 4-> 4- OJ .O 4-> E
m nj « a. o>-»- 10 »-
re .c u c i 4-> c 4->
r »- 4-> -r- C i- IO
C 0. C - -X "0 .C 10
O »- 3 i- O. in C +J
4-9 OJ O. in ID UJ ro 4->
I/I* ID 4-> ID 4-» E H- O. QJ 1C
OI3 4J I/I : O X L.
3 E e a> 01 z a> o in
"io »''" t- .c a>
-c Ti oi Si £ .c cntJ 0 > £
J.5X S o-o^^i S §
V)
£
O O> >, ID »-
t- en io 3 o
o E cr
« 4-» 1/1 a>
* c in u
r. OJ OJ C ID in
I/I Of ID OJ
0) U XI Q) D. 1-
i O) ID E ID
ID i n. f- cr
r ""* -0 "S«t-
m "O in aj 1/1 o
t- C 4J in 3
o ie c 3 in
oj * OJ E
a. ri o ai i-3
i/> _i a.ja 10 m
in -o t
§"
ID a. o
> E c
o c
ai u o
E en ID 3
^ *"*
i- a OIT-
Ot f I.
OJ o in
832^
OJ 1
1 i- 1 1 1 I/I
o>f>< o>a> o > 3
O ID S X UIDf UIDr
4J 0) 4-> IB
mOJinv) inaj »inOJ>
^ 4-> r j; O> ^4-!C3 L»4-*O
ID Q. en (!) a -^t ID c
= S- E .0 3 t- -o re 3 i-
croiaf cro3> cr o en
inn-in ai int-r- inv-c
o > o 4-> T-
O OJ *C 4-* r* OJ OOJOJOJ OOJ3
§u ojm3 E !/)
1 1
i in
CIO
to
II
CM
_J
«
4J
in
QJ
Jo
1
1
II i OJ *n *J
4^ £^ *. 4^ 1 Q) t/) (U f- Q..C *- **-«!->
>» 3 t c o »*-» a; *c +J
t-
«* C i/>D>(>nC(L)l/)L. dJ QJ U. (O
o g"1* .-^^ S^^S^S C 5! Q.
4J t. W CL^ -D OJ -,- L. Q.£ *-> O CT1
*« £ 3 -S OJ C X^«* ° ^ « ° "" U **
^i'^rf^'i'go's'i ?§>£! 1
^^Q.StfU^&f^^iTICi'S'S
i i
C ID
O) xt c in
o o io v a.
OJ f » > OJ
I- 4-> in ^
ID T- OJ O>4J
it- in >t- > ID
0 a. o re-c
in ID f 4^
en in
r OJlt- fc.
10 in *- o OJ
*- 01 4-> 4-*
O f- i- ID
OJ <»- O OJ
&«- -^ i-
ul 4-> J3 o en
C 1
in 4-> ai in i n -r-
ai f en c -o s.
3 en L. o c E 4->
> f-iDOJ-^OJOVI
ID U^-.c^ O.&--T-
> 4-> ID OJ I*- "CJ
OJ O > "O
OJfoOl-cin>
§in f »t-
O CLTJ J2 O1 Q.4J
4-> ID Ol O C E C
i- 01 s. -r- m oj c
Ocn lat-Emco
c ID a. o> 3 o->-
0> »- E -C V) 4J O.4J
c >><- o 4-> in c x 3
Of OL>oiaa>ajJ3
14-1 XI
Ct--C OJ3 1 COO.
f-t-COIA JQQJS- 4->ID
O. in L. 3 £ t- 4-> CLf IB 0)
ID -^ f 4J4-> O JZ ID
CT)*O OJO incJZID 4->r
^> *fr-<*-O4->0! 1*-
10 >>4-> O T3 in >>r OJ>£
O C T) Cf»f- 4->,
»*-CO fc.f OJ o--^-^-4J.c
OOJ4JJC OJIOWO 4-> r SJI4J
3 u 4->f-racaiaj-«- wen
4-*a>o^ Sc ** roojf-a>
C*4-4J OJ S-CO4->C4-> Ol O)
OJ 3>>--<- lOOt-OOJID t-ID-CH-
r lO^Or CJ.O.4^LDO'^ O.>4-^O
II II V
^i r< O.
^
1
-------
4. AREA OF COVERAGE AND REPRESENTATIVENESS
4.1 BACKGROUND
Over the decade of the 1970's the number of ambient air pollutant moni-
tors increased dramatically from approximately 1800 in 1970 to approximately
8000 in 1979. The lack of uniform criteria for station locations, probe
siting, sampling methodology, quality assurance practices, and data handling
procedures resulted in data of unknown quality. In preparing "national" air
quality trends data bases for the major pollutants, the number of monitoring
sites changed constantly, reflecting the growth in State monitoring networks
during this period. A conglomeration of different types of sitesrural, in-
dustrial, residential, commercial, etcevolved with no national plan. Some
urban areas of the country had extensive monitoring while others did not.
In October 1975, at the request of the Deputy Administrator of EPA, a
Standing Air Monitoring Work Group (SAMWG) was established. The Work Group
was to critically review and evaluate current air monitoring activities and
to develop air monitoring strategies which would be more cost effective,
would help to correct identified problems, would improve overall current op-
erations, and would adequately meet projected air monitoring goals. Members
of the Work Group represent State and local air pollution control agencies
and EPA program and regional offices.
SAMWG's review indicated that the current ambient monitoring program is
basically effective in providing information for support of State Implementa-
tion Plan (SIP) activities. Several areas were identified where deficiencies
existed, however. The principal areas where corrections are needed are sum-
marized below.
1. Lack of uniformity in station location and probe siting, sampling
methodology, quality assurance practices, and data handling proce-
dures have resulted in data of unknown quality.
2. Existing regulations coupled with resource constraints do not allow
State and local agencies sufficient flexibility to conduct special
purpose monitoring studies.
4-1
-------
3. Resource constraints and the diversity of data needs frequently re-
sult in untimely or incomplete reporting of air quality data. Ade-
quate air quality data for national problem assessments and routine
trend analyses are in some cases not available to agency headquar-
ters until 12-18 months after each calendar quarter.
4. In some cases, data are being reported to the EPA central data bank
from more stations than are absolutely necessary for adequate assess-
ment of national control programs and analysis of pollutant trends.
The recommendations of SAMWG later became the basis for the Federal moni-
ii
to:
2
toring regulations promulgated on May 10, 1979. These regulations require EPA
o set stringent requirements for a refined national monitoring network
in areas with high population and pollutant concentrations to provide
a sound data base for assessing national trends;
o give the States flexibility to use resources freed from SIP monitor-
ing work to meet their own needs;
o establish uniform criteria for siting, quality assurance, equivalent
analytical methodology, sampling intervals, and instrument selection
to assure consistent data reporting among the States;
o establish a standard national pollutant reporting index and require
its use for major metropolitan areas; and
o require the submission of precision and accuracy estimates with air
quality data to enable better interpretation of data quality.
These regulations should produce a streamlined, high-quality, more cost-effec-
tive national air monitoring program.
The States are required to establish a network of stations to monitor
pollutants for which National Ambient Air Quality Standards (NAAQS) have been
established. Each network is to be designed so that stations are located in
all areas where the State and the EPA Regional Office decide that monitoring
is necessary. The stations in the network are termed State and Local Air
Monitoring Stations (SLAMS).
Data summaries from the network are to be reported annually to EPA. Data
from a subset of SLAMS to be designated as National Air Monitoring Stations
(NAMS) are to be reported quarterly to EPA.
4-2
-------
4.2 NETWORK DESCRIPTION
4.2.1 The SLAMS Network
The SLAMS network should be designed to meet a minimum of four objectives:
1. To determine the highest concentrations expected in each area covered
by the network;
2. To determine representative concentrations in areas of high popula-
tion density;
3. To determine the impact of ambient pollution levels from significant
sources or source categories; and
4. To determine background concentrations.
Each monitoring site is required to be identified by location and type of sur-
roundings as well as by monitoring objective and spatial scale of representa-
tiveness. The spatial scale of representativeness is described in terms of the
physical dimensions of the air parcel sampled by the monitoring station through-
out which actual pollutant concentrations are reasonably similar; the scale
adjectives are micro, middle, neighborhood, urban, regional, national, and glob-
al.
4.2.2 The NAMS Network
The NAMS stations are selected from the SLAMS network to emphasize urban
and multisource areas. The primary objective for NAMS is to monitor areas where
pollutant levels and population exposure are expected to be highest, consistent
with the averaging time of the NAAQS. Accordingly, NAMS fall into two catego-
ries:
1. Stations in area(s) of expected maximum concentrations; and
2. Stations with poor air quality and high population density but not
necessarily in area(s) of expected maximum concentrations.
For each urban area where NAMS are required, both categories of stations must
be established. If only one NAMS is needed to monitor suspended particulates
(TSP) and sulfur dioxide (SO^), the first category must be used. The NAMS are
expected to provide superior data for national policy analyses, for trends, and
for reporting to the public on major metropolitan areas. Only continuous in-
struments will be used at NAMS to monitor gaseous pollutants.
4-3
-------
Siting requirements and definitions vary with pollutant, but NAMS are re-
quired only in urban areas with populations of at least 50,000 where pollutant
concentrations are known to exceed the secondary NAAQS. The number of urban
areas monitored will vary with pollutant, as indicated below.
Pollutant
Total suspended particulates (TSP)
Sulfur dioxide (S02)
Carbon monoxide (CO)
Ozone (03)
Nitrogen dioxide (NOz)
Urban
areas
212
160
77
85
33
NAMS
monitors
636
244
121
208
65
The NAMS network is being selected on the basis of experience with past
monitoring data and primarily attempts to measure the highest pollutant levels
associated with each urban area. The word "associated" indicates that, with
transport-effected pollutants such as 03 or TSP, the highest levels may occur
in areas of low population densities. The location selection should be periodi-
cally reviewed, based on historical monitoring data, meteorological conditions,
and changes in emission patterns. For example, there is definite evidence that
the highest 03 levels in the Los Angeles basin have been moving upwind in the
last few years, probably because of the changing composition of tailpipe pollu-
tants. Changing fuel usage and new construction may have also contributed to
this shift.
4.3 REPRESENTATIVENESS
The question has often been asked as to what constitutes a "national"
trend? In previous National Air Quality, Monitoring and Emission Trends Re-
ports,3'4'5 all sites with data available in the National Aeromatic Data Bank
(NADB), that could meet an historical data completeness criteria, were select-
ed for the national trend. In the case of total suspended particulate (TSP)
as many as 3000 sites could meet an historical completeness criteria. These
sites come from a variety of networks representing urban areas, rural areas,
and large point sources. Geographical coverage was largely weighted by popu-
lation; that is, the more populated areas generally had more monitoring sites.
In analyzing the national trend each site was weighted equally.
In contrast to TSP, the automotive related pollutantscarbon monoxide
(CO) and ozone (O-)--had less than 250 trend sites meeting historical complete-
5
ness criteria as late as 1977. Although California had a disproportionate
4-4
-------
number of these trend sites, each site was weighted equally in developing the
national trend.
In each of the three cases, the sites used for analyzing the national
trend represented a mixture of siting criteria. Consequently, it was difficult
to determine what the resulting trend really represented.
The major difficulty in selecting sites that define a "national" trend is
determining what mix of sites represents national air quality. How should geo-
graphic distribution be determinedby population or by land mass? What types
of sitescenter city, suburban, residential, commercial, industrial, rural,
remote, etc.--should be in the national trend and what proportions are appro-
priate?
4.3.1 The NAMS Solution
A partial solution to this dilemma lies in the NAMS network. Part of
the problem with a "national" trend is a semantic onehow is it defined? The
NAMS can resolve this problem because they caji be defined. For each of the cri-
teria pollutants, NAMS are located in either the areas of expected maximum con-
centration or the areas of high population density. Consequently, the NAMS lend
themselves to stratification into these two site populations. Additional siting
information is available through the NAMS management information system, which
will allow for the use of covariate data (such as traffic counts in the case of
CO) for the first time.
The NAMS are particularly appropriate for characterizing national trends
for urban sites located in areas of expected maximum concentration or high popu-
lation density. For example, the CO NAMS located in areas of expected maximum
concentration make up a clearly defined population and could serve as a good in-
dicator of the success (or failure) of the automotive emission control program.
4.3.2 The SLAMS and Detailed Urban Area Analyses
While the NAMS can be used to define national urban trends in areas of
expected maximums or high population density, the SLAMS can be used for detail-
ed urban area analyses. If one is trying to determine changes in air quality
in an urban area, then an examination of both spatial and temporal change is in
order. The SLAMS networks lend themselves to these types of analyses. EPA has
published several guidelines useful in determining spatial and temporal trends;
they should be consulted before initiating these types of analyses.
4-5
-------
4.4 REFERENCES
1. Monitoring and Air Quality Trends Report, 1974. U.S. Environmental Pro-
tection Agency, Office of Air Quality Planning and Standards. Research
Triangle Park, N.C. Publication No. EPA-450/1-76-001. February 1976.
2. Federal Register, Vol. 44, February 8, 1979, pp. 8202-8237.
3. National Air Quality and Emission Trends Report, 1975. U.S. Environmental
Protection Agency, Office of Air Quality Planning and Standards. Research
Triangle Park, N.C. Publication No. EPA-450/1-76-002. November 1976.
4. National Air Quality and Emissions Trends Report, 1976. U.S. Environmen-
tal Protection Agency, Office of Air Quality Planning and Standards, Re-
search Triangle Park, N.C. Publication No. EPA-450/1-77-002.
5. National Air Quality, Monitoring and Emissions Trends Report, 1977. U.S.
Environmental Protection Agency, Office of Air Quality Planning and Stand-
ards, Research Triangle Park, N.C. Publication No. EPA-450/2-78-052.
December 1978.
6. Guideline on Procedures for Constructing Air Pollution Isopleth Profiles
and Population Exposure Analysis. U.S. Environmental Protection Agency,
Office of Air Quality Planning and Standards, Research Triangle Park,
N.C. Publication No. EPA-450/2-77-024a. October 1977.
7. Users Manual for Preparation of Air Pollution Isopleth Profiles and Popu-
lation Exposure Analysis. U.S. Environmental Protection Agency, Office
of Air Quality Planning and Standards, Research Triangle Park, N.C.
Publication No. EPA-450/2-77-024b. October 1977.
4-6
-------
5. DATA COMPLETENESS AND HISTORICAL CONTINUITY
Data completeness for summary statistics and historical continuity for
trend data are two concerns of the data analyst. In both cases, the quantity
of data available affects the uncertainty of the analytical results. In cur-
rent practice, criteria are invoked in the data screening steps to select min-
imally acceptable data sets. This section explores the effect of missing data
on the uncertainty of analytical results, and discusses criteria for ensuring
adequate data completeness and historical continuity.
5.1 DATA COMPLETENESS
5.1.1 Background and Purpose
The number of air quality values produced by ambient monitors is usually
fewer than the maximum number possible. Missing values result from the inter-
mittent sampling schedules for manual methods, instrument failure, downtime for
calibration, or human error. The temporal balance and the sample size of the
resultant data set can seriously affect the validity of the sample and the un-
certainty of its summary statistics.
Published criteria have been used by EPA to establish the validity of data
sets for summarizing and analyzing air quality data. Such criteria should mini-
mize the uncertainty associated with air quality summary statistics. Unfortu-
nately, for each pollutant the same completeness criteria are used for all sum-
mary statistics, despite the fact that the uncertainty associated with a summa-
ry statistic varies with the type of statistic and its variance properties.
In a sense, validity criteria provide a capability for identifying data
produced by poorly operating instruments and for screening data samples that
may otherwise yield misleading or incorrect estimates of air quality levels.
Data requirements defined by the criteria should involve the type of summary
statistic (e.g., annual mean or maximum daily average) as well as its intended
application (e.g., trends analysis or status assessment with respect to the
standard). In general, fewer data are needed to determine an annual average
5-1
-------
than a short-term statistic, and fewer data per year are needed to estimate a
long-term trend than a yearly status.
Currently, the NADB uses different validity criteria for different time
periods and sampling approaches, as shown in Table 4. The criteria for each
sampling approach are intended to consider both the characteristics in the data
collection and the primary objective of the monitoring; however, for specific
applications there are inconsistencies in the criteria for different approaches.
This section discusses only the criteria for summary statistics for periods of
3 to 12 months.
TABLE 4. NADB VALIDITY CRITERIA1
Continuous Sampling (1-h, 2-h, and 4-h data)
Quarterly statistics - 75% or 1642 hours
Annual statistics - 75% or 6570 hours
Intermittent Sampling (24-h data)
Quarterly statistics
Five samples per quarter
If 1 month has no values, at least two
values in other months
Annual statistics - four valid quarters
5.1.2 Origin of NADB Criteria
Criteria for intermittent sampling were formulated on the basis of a bi-
weekly sampling schedule which was used by the National Air Surveillance Net-
work (NASN) up to 1972. Nehls and Akland recommended that more stringent cri-
2
ten a be applied when the sampling schedule is every 3rd or 6th day.
For continuous sampling, the origin of the 75 percent criteria is not
3
clear. Early Federal Air Quality Publications used a 50 percent criterion to
report summary statistics. Nehls and Akland suggested 75 percent completeness
and a more structured basis for summary statistics, as shown in Table 5; these
criteria require as few as 3402 hourly observations, or 39 percent of the pos-
sible hours.
5-2
-------
TABLE 5. CONTINUOUS MEASUREMENT SUMMARY CRITERIA2
Time interval
3-h running average
8-h running average
24-h
Monthly
Quarterly
Yearly
Minimum requirement
3 consecutive hourly values
6 hourly values
18 hourly values
21 daily averages
3 consecutive monthly averages
9 monthly averages with at least
2 monthly averages per quarter
5.1.3 Characteristics of Air Quality Sampling
Intermittent 24-hour sampling generally follows a fixed systematic or
pseudo-random schedule which is intended to provide representative coverage of
a time period with only a fraction of the total possible observations. By de-
sign, data from intermittent sampling can provide good estimates of an annual
mean, but may severely underestimate the peak values. Trying to improve the
estimate of the peak observation by intensifying the sampling during periods
of high pollution levels is known as episode monitoring. When these unsched-
uled episode data are combined with the scheduled data, bias can be intro-
duced into summary statistics.
Early intermittent sampling for TSP and other pollutants by the NASN was
biweekly on a modified random schedule which yielded 20 samples a year. Later
the schedule was modified to ensure equal representation for each day of the
week. After 1972, samples were collected every 12th day yielding a maximum of
30 to 31 samples a year. EPA's current recommended sampling schedule for TSP
is once every 6 days. Most agencies seem to follow this schedule. Among 2882
sites with 1978 data that meet the NADB validity criteria, over 60 percent pro-
duced 40 to 60 samples, and less than 10 percent sample more often than 1 in 3
days.
Continuous hourly monitoring, by providing a more complete representation
of air quality, should provide much better estimates of the true annual mean
and short-term peak values. In reality, continuous monitoring is often incom-
plete, and can yield biased estimates of long-term behavior. Missing data of-
ten occur in blocks of consecutive days or weeks. Sometimes a pollutant such
as 03 is monitored during only part of the year; only seasonal statistics are
appropriate in these cases.
5-3
-------
TABLE 6. DATA COMPLETENESS
FOR CONTINUOUS S02 MONITORING, 1973
Minimum quarterly
completeness (%)
25
26-32
33-49
50-66
67-74
75-100
Total
Number of sites
Annual completeness
<75%
41
11
40
14
2
108
>75%
4
18
10
84
116
Total
41
11
44
32
12
84
224
As an example of actual monitoring performance, data completeness for
1973 S02 data is presented in Table 6. Of 224 sites, only 116 reported 75
percent of the total hourly observations. According to the percentage of com-
pleteness for 1973, 56 of the 108 sites not meeting the 75 percent annual com-
pleteness criterion had at least one-third of the total observations in each
calendar quarter.
5.1.4 Characteristics of Air Quality Data
Air quality data are known to exhibit temporal components generally in
the form of diurnal, weekly, or seasonal cycles. The variances for each com-
ponent may be different, thus an arbitrarily selected sample of the entire
series may not yield unbiased, minimum variance statistical estimates if the
sample does not represent each portion of each time component equally.
For example if a diurnal pattern always shows the highest concentrations
between 0800-1200 hours, but the sampling does not include any observations
during these hours, simple sample statistics will have a negative bias. A sam-
ple requires temporal balance to be representative.
With a seasonal pattern of nonuniformly distributed data, the more extreme
the seasonal variation, the larger the potential bias. In addition, the larger
the sampling imbalance across the seasons, the larger the potential bias. If a
seasonal pattern shows one quarter with concentrations twice those of the other
three quarters, the range of possible bias in a mean is 88 percent to 136 per-
cent if one-tenth of the total observations are in a single quarter.
5-4
-------
Temporal balance is essential only if simple unweighted statistics are
used. Unbiased estimates can be obtained from unbalanced samples by using ap-
propriate sample stratification and by applying weighting factors. This method
is particularly effective in determining unbiased estimates for a given sample
size when the variance components are different within sampling strata.
The problems of unbalanced samples were considered in the establishment
of the original validity criteria for TSP and other 24-hour data. The recom-
mendations of Nehls and Akland for continuous data also considered these prob-
lems. Unfortunately, the current criteria for intermittent sampling do not
ensure balance when sampling more frequently than once in 2 weeks; more strin-
gent criteria or weighting factors need to be considered. In addition, the
current NADB criteria for continuous sampling may not ensure balance while re-
quiring more observations than necessary. Fewer observations could provide
9
better estimates if the sampling recommendations were adopted.
5.1.5 Estimates of Air Quality
The variation in a statistical estimate of air quality depends on the
temporal variation of the data, the size of the data sample, the temporal dis-
tribution of the sample within the year, and how the data are combined in de-
veloping the estimate.
Data completeness has varying impacts on different sample statistics. In
general, the more data available, the better the estimate. For a given level
of uncertainty, however, fewer observations will be needed to estimate a mean
than, say, a maximum value. For data with cyclical components, temporally bal-
anced samples with less data will yield better estimates than unbalanced samples
with more data. With the use of weighted averages, more data will usually yield
better estimates, even if they are not balanced.
Intermittent Data - Nehls and Akland investigated the accuracy of annual
means estimated from systematic samples. Table 7 shows how accuracy degrades
with decreasing sampling frequency based on particulate data collected almost
continuously in Philadelphia over 9 years. As expected, the percent error in-
creases with decreasing sampling frequency. For a given sampling frequency,
the error for a quarterly mean is approximately twice that for the annual mean.
5-5
-------
TABLE 7. ACCURACY OF PARTICULATE SAMPLING FREQUENCIES AND
AVERAGING INTERVALS
Sampling
frequency
(k) days
2
3
4
5
6
7
8
9
Percent error compared to daily sampling
Yearly
1.4
1.7
2.3
2.7
3.2
8.7
3.8
5.0
Quarterly
2.3
3.2
4.6
6.0
6.5
10.7
8.1
8.9
Monthly
3.9
6.5
8.7
10.3
12.0
15.1
14.9
16.0
Empirical evidence of the effect of sample frequency on summary statistics
is presented in Table 8, based on TSP data from 14 sites monitoring continuous-
ly for 3 years, 1976-78. Estimates of annual averages and annual maximums when
sampling every other day, every 3rd day, every 6th day, and every 12th day are
compared with the averages and maximums obtained from daily sampling. Errors
of the estimates for the annual maximum are typically 4 to 6 times larger than
the errors in the mean. Even sampling every other day, the sample maximum un-
derestimates the annual maximum by 9 percent.
TABLE 8.
DEVIATION OF OBSERVED ANNUAL MEAN AND MAXIMUM FROM TRUE VALUES
AMONG 42 TSP SITES, 1976-78
Average percentage error, (%}
Sampling
frequency
1 in 12
1 in 6
1 in 3
1 in 2
Mean
7.1
4.1
2.1
1.9
Maximum
30.1
22.0
13.6
8.9
Continuous Data - The variability of an annual mean derived from continu-
ous data can be examined by assuming that the data consist of 1 or 2 months of
complete data in each calendar quarter. This assumption corresponds to the ex-
treme situations in which the quarterly completeness is 33 percent or 67 per-
cent.
5-6
-------
Variance of the annual mean Var(x) based on complete months of data is
expressed as
Var(x)=f^ (i-f). £(1-1), (1)
where f is the fraction of months sampled in each quarter (f = n/3 and 1-f is
the finite population correction), n is the number of months in a quarter, and
2
a is the variance among true monthly means within a quarter. For simplicity,
the number of days in a month and the number of months sampled per quarter are
assumed to be equal. To generalize this approach, this model could be con-
structed to consider subsamples of any number of discrete blocks of consecutive
days within each calendar quarter.
The variance of an annual mean based on monthly data defined by Equation 1
depends on the variability among monthly means within a quarter and on the num-
ber of months sampled in a quarter. Thus, for 1 month of data per quarter, the
standard deviation (SD) of the mean is 0.41 times the SD of monthly means. For
2 months of data per quarter, the SD of the mean is 0.2 times the SD of monthly
means.
The model based on monthly data was evaluated by comparing the empirical
distribution of annual means (based on 1 or 2 months of actual data) with the
theoretical distribution defined by Equation 1. Complete S02 measurements ob-
served at the NYC laboratory site during 1973 were used to generate all 81
possible subsamples and corresponding annual means. Using the average quarter-
2
ly estimate of variance among months equal to (21.4) , the estimates of vari-
ance for the sample means using 1 month or 2 months of data per quarter are
2 2
(8.8) and (4.4) respectively. Normal distributions corresponding to these
variance estimates were compared to the histogram based on all 81 possible
annual averages; there was good agreement, as indicated in Figure 1. This ex-
ample demonstrates that data with one-third of the observations in each calen-
dar quarter can produce reasonable sample estimates.
The analysis above assumes that the variance is the same in each sampling
strata (quarter). If variances are different, the sampling period with the
highest variability should have the best sampling representation.
5.1.6 Summary and Recommendations
Data completeness criteria for producing summary statistics are desirable
to ensure representative estimates of air quality. Ideally, different criteria
5-7
-------
o
0.1
g. °-05
60
TWO MONTHS PER QUARTER
70
80
90
100
110 120
0.1
0.05
Ul
o:
ONE MOUTH PER QUARTER
70 80 90 100 110
S02 CONCENTRATION (yg/m3)
120
Figure 1. Empirical histograms of all possible annual averages based on
one or two months of data per quarter compared with theoretical
frequency distributions defined by statistical model for 1973
S02 at the Mew York City laboratory site.
5-8
-------
should be used to screen data for different uses; however from a practical
point of view, a single criterion specifying minimum data completeness for one
statistic (e.g., a mean) will probably be used for all applications. The cur-
rent NADB criteria for summarizing data need to be revised. The recommenda-
tions of Nehls and Akland appear reasonable for intermittent data; the existing
criteria are reasonable as long as weighted statistics are used to correct for
the potential sampling imbalance. Methods for specifying the required strata
and numbers of observations should be resolved in future work.
5.2 HISTORICAL COMPLETENESS
5.2.1 Background and Purpose
Important to trend analysis is data base preparation. If one is analyz-
ing the trend at only one monitoring site, all data from that monitor can be
considered. When analyzing a group of monitors, a specific time frame repre-
sented by the group should be used, and the historical data completeness of
each candidate monitoring site should be examined to determine suitability for
trends analysis within the time frame.
Since air quality trend analysis focuses on year-to-year variation (as
opposed to within-year variation), emphasis is usually on comparison of annual
statistical indicators derived from complete or "valid" data sets. Thus, the
number and temporal distributions of the annual statistical indicators are im-
portant; this characteristic is termed "historical completeness."
The data series should be as complete as possible, but missing values are
permissible. Air quality monitoring does not always produce complete data
records. Instrument failure and data processing problems are two of the many
reasons for missing data. In many cases, emphasis is on compliance assessment
rather than long-term monitoring. Historical completeness criteria are used
during the data screening process to identify the largest possible data base
from which a representative sample can be drawn for trend analysis.
5.2.2 Characteristics of Missing Data in Trend Analysis
The number of missing values permitted depends on the objective of the
trend analysis, the trend technique, and the variance components of the data.
In this context, variance components include error due to incomplete sampling,
instrument error, meteorological fluctuations, and departures of the observed
5-9
-------
trend from an assumed underlying pattern. The sensitivity of a trend tech-
nique improves with more data; thus the analyst desires the maximum amount of
data at the largest number of sites. The sensitivity also depends on the vari-
ability of the trend. If, for example, a monotonic trend were to exist at a
group of sites, a minimum of two observations selected at random from each site
would be sufficient to categorize the trend. Similarly, if the analysis objec-
tive were to detect a long-term shift without the need to categorize the year-
to-year pattern, a few widely spaced data points would be adequate. If, how-
ever, the temporal patterns were more complex or were masked by high data vari-
ability, more data would be needed to separate the trend from other variance
components, and more complete temporal records would be needed to obtain an ac-
curate year-by-year trend.
5.2.3 Historical Completeness Criteria
At a minimum, each site in the trend analysis must have two data points
representing the time series. To ensure that these data points are widely
spaced, the time period may be divided into two segments. For example, if the
6-year period 1972-77 were divided into 1972-74 and 1975-77, a site could be
selected if it produced one valid year of data in the first and in the second
time segments; this procedure ensures that data from the start and end of the
time period are represented.
A data selection approach often used by EPA is based on the historical
completeness of quarterly data. A criterion commonly used in the analysis of
national trends is four consecutive calendar quarters with valid data in each
of 2 time segments. Because the last one or two quarters of data can be miss-
ing from the most recent years due to late reporting, this approach can yield
more current estimates than the approach based on complete annual data. In
addition, using quarterly data to derive annual summary statistics minimizes
the bias caused by within-year sampling imbalance (Section 5.1).
Using historical completeness criteria can increase the number of candi-
date monitors many fold, and thus help obtain a more geographically representa-
tive sample. Of the 4000 TSP monitors reporting data to the NADB between 1972
and 1977, 2661 met the aforementioned criteria based on valid annual data, and
2737 met the criteria based on valid quarterly data. The trend period was di-
vided into two time segments--1972-74 and 1975-77. Using three 2-year time
5-10
-------
segments and valid annual estimates, 1737 would qualify, but only 417 had valid
annual data in each year between 1972 and 1977.
If a representative sample of stations with valid data for all years is
available, additional sites with missing data may not be needed or may only be
needed to improve the detectability of trends or to confirm the existence of
trends by using a larger population.
5.2.4 Historical Completeness in Data Presentation
Historical completeness of data is particularly important for graphical
presentations of trend results. Trend lines based on a group of monitoring
stations are commonly used in air quality analysis. Such an aggregate trend
line must be based on the same number of sites for each data point (e.g., year)
to minimize bias caused by a changing data base; this requirement is satisfied
by using every-year stations or by estimating all missing values. Estimating
missing values is not necessary for trend assessment, but it is for data pres-
entation.
In general, EPA characterizes the air quality trend in a defined geograph-
ic region using data for n years from m monitors. These data can be arranged
in the matrix:
Monitor 1
Monitor 2
Monitor m
Year 1
X21
xml
Year 2
X12
X22
Xm2
. . . Year n
xln
... x2n
x. . ...
1J
' xmn
where value x. occurs at monitor i in year j. The presentation of trends re-
quires an average regional value of x for each year. If the matrix is complete,
a reasonable estimate of the average value for year j is the arithmetic mean
of each column,
/s , m
yi = m E xii'
j in .j_i ij
However, this estimate may be biased if the matrix has missing values.
Two methods have proved useful in estimating yearly means when the data
matrix is incomplete. In the one method, missing values are estimated by
linear interpolation before the column means are calculated. This method
5-11
-------
produces reasonable estimates when the underlying trend is linear. The other
method assumes each cell in the matrix can be characterized by a general linear
model :
Mi
where
y is a constant--the grand mean of all the responses which would be
obtained if there were no errors;
M. is a term peculiar to the ith monitor, and is independent of year;
Y. is a term peculiar to the jth year, and is independent of monitor;
3 and
e-. denotes the experimental error associated with x-jj, and it is as-
J sumed to be a normal random variable with standard deviation ae.
Computer programs can be used to generate least squares estimates of these
parameters regardless of the number of missing values in the data matrix.
Each yearly mean is estimated as
/N /S. A.
u . = y + Y.. (4)
J J
Statistical methods are available for determining confidence intervals for the
yearly means and for the differences between yearly means.
One advantage of the generalized linear model is that it explicitly re-
lates the number of missing yearly values to the uncertainty of the estimates
of the yearly means. Working backwards through the model, the analyst can
specify a desired confidence interval, calculate the number of permissible
missing values, and adjust the data base accordingly.
5.2.5 Effect of Historical Completeness on Trends Analysis
The impact of missing data on trends analysis is empirically examined by
sampling from a group of 30 S02 sites with 75 percent complete data each year
of 1974-78. Theoretical 90 percent and 50 percent probability intervals based
on sampling one-third of the annual means in each of the first 3 years and one-
half in each of the last 2 years are shown in Figure 2. This method of sampling
is analogous to selecting sites independently for each year, and the intervals
show the distribution of means in each year.
The smallest variability is in the last 2 years because of the larger num-
ber of observations used to calculate those means. The distribution of means
5-12
-------
42
40
38
36
34
32
30
28
E
^ 26
3-
~ 24
P 22
§ 20
| 18
s
S 16
8 W
12
10
8
6
4
2
o
»
-
-
-
i
-j- -r
_. i-L
* * T T
\ \^ \ 1
r
1 1 M
95th % ile
1 75th % ile
* MEAN (50th % ile)
J 25th % ile
1 5th % ile
i i
1974 1975 1976 1977 1978
YEAR
Figure 2. Theoretical probability distribution of annual mean
S02 with 10 sites in each year 1974-1976 and 15 sites
in 1977 and 1978.
5-13
-------
is sufficiently variable for a subsample of means to show an incorrect upward
trend. The extremes occur when a site effect exists and when different sites
are used at the start and end of the time period. The highest and lowest an-
nual means from the aforementioned subsamples are shown in Table 9.
TABLE 9. ANNUAL MEANS
1974 1975 1976 1977 1978
Lowest
Highest
14.5 21.4 11.3 12.9 14.5
56.9 56.8 58.7 35.5 37.3
The situation is less ominous if an additional constraintthat the same sites
be used in the first 3 years and in the last 2 yearsis imposed as a require-
ment for historical continuity.
Sampling from a finite population dictates that the distribution of means
for each year of the first 3 years will be independent of the distribution of
means for 1 of the last 2 years. If there is a site effect, other combinations
will be dependent. The most extreme results will occur if the distribution of
means in the first and last years are independent; for example, there is a 25
percent chance that the subsample-derived mean of the first year will be less
than the true value and that the subsample-derived mean for the last year will
be greater than the true value. The extreme value for the first year could be
14.5 and that for the last year could be 37.3, but values for the intermediate
years would counterbalance the extremes at the end years because of the site
effect. Example values for the three intermediate years are 33.7, 42.2, and
14.9, respectively. Thus, historical continuity minimizes the chances of an
incorrect result, but this rare combination of missing values may still pre-
vent the detection of the true downward trend.
5.2.6 Conclusions and Recommendations
Historical completeness criteria are useful screening tools for selecting
sites for trend analysis. Historical data should be as complete as the environ-
mental data base will permit in order to establish a representative sample which
is large enough to detect trends. Trend analysis should keep track of the ex-
tent of the historical completeness, and should report separate findings accord-
ingly. Techniques such as linear interpolation and the generalized linear model
5-14
-------
should be applied to trend data with missing values to minimize bias due to site
effects for the purpose of displaying graphical results.
5.3 REFERENCES
1. Aeros Manual Series, Volume III. Summary and Retrieval. EPA-450/2-76-009a,
July 1977.
2. G. J. Nehls and G. G. Akland. Procedure for Handling Aerometric Data. J.
of Air Pollution Control Association. 23:180 (1973).
3. Air Quality Data from the National Air Sampling Networks and Contributing
State and Local Networks, 1964-1965. United States Department of HEW, 1966.
5-15
-------
6. STATISTICAL INDICATORS AND TREND TECHNIQUES
This section attempts to answer two commonly asked questions: What is
the air quality? How is it changing? Throughout the section, it is assumed
that the principles and procedures of the previous sections have been follow-
ed; that is, the data sets have been screened for outliers and the sites have
been examined to ensure representativeness. Separate subsections are devoted
to each topicstatistical indicators and trend techniques, but a certain
amount of interplay is involved between the two. For example, when a partic-
ular statistical indicator is used to summarize the data, a natural followup
concern is how this indicator has changed over time.
6.1 STATISTICAL INDICATORS
The term "statistical indicator" is used in this section in a fairly gen-
eral sense to include any statistic which summarizes air quality data for a
particular time period. Technically, a statistic is a "summary value calculat-
ed from a sample of observations." For some air quality applications, we may
have not merely a sample but also the entire population; therefore, the comput-
ed value is not technically a statistic. For the purposes of this section,
statistical indicator is used in either case.
This subsection discusses the background and purpose of statistical indi-
cators for air quality, the types of indicators frequently used, what proper-
ties are desirable, and the relative merits of various indicators as well as
future needs.
6.1.1 Background and Purpose
A continuous air quality monitor can produce a pollutant measurement for
each hour of the day every day of the year. This means as many as 8760 con-
centration values for a single pollutant at a single site. The volume gf data
is further increased by the number of pollutants measured at a site and by the
number of sites within an area. Such a quantity of data requires some type of
data reduction to conveniently summarize the data. A wide variety of summary
6-1
-------
statistics could be used. The proper choice depends on the intended purpose
of the information and on an awareness of the characteristics of the data.
While many specific purposes could be listed for air quality statistical
indicators, for most practical applications there are basically two: (1) in-
dicating status with respect to standards and (2) evaluating trends. An air
quality standard is an absolute frame of reference for an indicator. If an
air quality statistic is not compared to a standard, it is usually compared to
a value from some other time period under the general heading of trend analy-
sis. A third purpose is using the statistic as a basis of comparison with data
from other sites or cities; however, most remarks that apply to choosing an in-
dicator for trend analysis also apply to these comparisons.
An important point in any discussion of statistical indicators is that
many air quality standards are structured in terms of concentration limits not
to be exceeded more than once a year. This places a premium on information in
the upper tail of the distribution. Accordingly, the average and median val-
ues commonly used as summary statistics in many fields are often of little in-
terest in air quality data analyses. In fact, only the extreme values may be
of interest for comparisons with standards.
6.1.2 Types of Statistical Indicators
A wide variety of statistical indicators are being used in air quality
data analyses. Some reflect standard statistical choices such as means or
percentiles. The highest or second highest value or the number of times the
level of the standard is exceeded is often used due to the importance of the
higher concentration values in air quality management. For example, the 1-hour
NAAQS for CO specifies a level of 9 ppm not to be exceeded more than once a
year. Thus, the second highest hourly value or the number of hourly values
greater than 9 ppm would equivalently indicate whether or not the site meets
the standard.
Other statistical indicators represent a compromise between focusing on
the peak concentrations and introducing more stability into the indicator.
These indicators may use either upper percentiles or knowledge of typical pol-
lutant patterns to construct a useful index for trends. Examples include the
average of daily maximum 03 measurements for the 03 season or the number of
times a level (other than the standard) is exceeded.
6-2
-------
Although the previously discussed indicators are typically used to sum-
marize data for one pollutant at one site, they can be adapted to accommodate
data for one pollutant at several sites within an area. A further generaliza-
tion is to incorporate data for several pollutants into a single index. To do
this requires some method of normalizing the individual pollutant measurements
so that there is some basis for scaling their relative contributions to the in-
2 3
dex value. The Pollutant Standards Index (PSI) is a method for doing this. '
An ideal indicator would incorporate both air quality data and population
data to provide a measure of population exposure to various air pollution lev-
els. Efforts have been made in developing such measures, but the degree of
spatial and temporal resolution required for both data sets makes these beyond
the current state of the art. However, simplifying assumptions may be intro-
4
duced to obtain rough approximations for these types of measures.
6.1.3 Desirable Properties
Certain properties are desirable for a statistical air quality indicator.
Some properties are desirable on a purely intuitive basis; others involve tech-
nical or practical considerations.
Clarity - Because the purpose of an indicator is to convey information,
there is an advantage in using an indicator that is easily understood. Although
understanding is partially a function of the intended audience, simple data pre-
sentations should use more easily understood indicators than those appropriate
for a detailed examination of alternative control strategies if the more detail-
ed analysis requires a more complicated indicator. Clarity does not necessari-
ly mean that the indicator is simple to compute. For example, the computations
for evaluating the PSI require linear segmented functions, but the final result
is relatively easy to comprehend.
Independence of Sample Size - Another desirable property, on an intuitive
basis, is that the indicator be independent of sample size. For example, sam-
pling TSP every 6th day commonly results in approximately 60 measurements a
year. Sampling every day of the year may result in 365 (or 366) measurements.
Use of the maximum value, the second maximum, or the number of times the stand-
ard level is exceeded creates a problem with different sample sizes. A site
that samples once every 6th day has only a l-in-6 chance of measuring the an-
nual maximum; a site that samples every day obviously measures the maximum.
Any indicator that does not account for varying sample sizes can be misleading.
6-3
-------
Robustness - A property that should be mentioned is robustnessthat is,
the indicator is not unduly influenced by a few outliers. How desirable this
property is for air quality data analysis depends on the type of study. In
air pollution work, the higher concentrations are often the most important;
therefore, the median concentration may be robust but irrelevant. If the pur-
pose of the study is to examine trends resulting from an overall emission re-
duction, the robustness of the indicator may be desirable.
Precision and Accuracy - A statistical indicator can have all of the
above properties and yet fail to adequately summarize the data. Therefore,
precision and accuracy must be included as desirable properties. Accuracy im-
plies that the indicator is unbiased; precision measures the variability of
5
the estimate. A recent paper by Johnson discusses these properties for cer-
tain air quality indicators.
Feasibility - A practical evaluation of potential indicators must con-
sider the feasibility of implementing a particular choice. An indicator may
be highly desirable but not feasible to implement due to the present state of
the art or to lack of information in the air quality data bases. This does
not mean that a promising indicator should be ignored because it is difficult
to implement; such a situation can highlight where additional work is needed.
Relevance - After all of the above properties have been considered, the
final test of an air quality indicator is its relevance to the purpose of the
study. Does the indicator sufficiently characterize the data so that the re-
sults may be easily translated into a clear statement of the information in
the air quality data?
6.1.4 Discussion of Candidate. Indicators
Consideration of all possibilities is not practical when discussing can-
didate indicators. It is more convenient to delineate certain classes of in-
dicators: peak values (highest and second highest measurements), mean values
(arithmetic and geometric), percentiles, exceedance statistics, design value
statistics, multisite indicators, multipollutant indicators, and seasonal in-
dicators.
Peak Value Statistics - This class includes indicators such as the maxi-
mum and the second highest concentrations which have been used because they
are consistent with the "once per year" type of air quality standard. These
6-4
-------
are not independent of sample size. For less than everyday sampling, they
have a negative bias because they underestimate the true value. Therefore
it is difficult to recommend this class of statistics unless an adjustment
is made for sample size. A possible alternative approach is the use of ex-
pected peak values estimated by fitting distributions to the data; this ap-
proach yields statistics independent of sample size.
Mean Values - TSP, SCL, and N02 have NAAQS's involving annual means (ar-
ithmetic for S02 and NCL; geometric for TSP), so the use of an annual mean for
these pollutants is fairly natural. For CO and 03, the annual mean is not par-
ticularly useful for assessing status. Because the primary CO control strategy
involves motor vehicle emission reductions, the annual mean may be adequate for
trend analyses; nevertheless, changes in higher concentrations, should be exam-
ined. The typical diurnal and seasonal patterns for 03 suggest that the an-
nual mean is of little value.
Percentiles - Use of percentiles is one means of adjusting for differ-
ences in sample sizes and affording a degree of protection against a few ex-
treme values. If higher concentrations are of interest, little will be gained
by using lower or midrange percentiles. Upper percentiles are more useful;
the 90th, 95th, and 99th percentiles provide an adequate range for trends analy-
ses.
Exceedance Statistics - This class of statistics includes both the fre-
quency and the relative frequency that an air quality level (NAAQS or other)
was exceeded. This type of statistic is intuitively appealing, and is rela-
tively easy to understand. If the level of the standard is used, the results
relate directly to status assessment. Statistics involving the recorded num-
ber of exceedances (rather than the percentage of exceedances) depend on the
sample size; with less than complete sampling, they underestimate the true val-
ue. As with peak value statistics, it would be advisable to use an adjustment
to account for incomplete sampling. This type of correction was incorporated
7 8
into EPA's recently revised 0, standard. '
A change from one to two exceedances does not mean that the air quality
has become 100 percent worse. Caution is needed in interpreting this type of
statistic because it exaggerates percentage changes. Johnson considered the
relative precision of exceedances and the 90th percentile for 03 data, and
concluded that the 90th percentile was more desirable for trend analyses.
6-5
-------
Design Value Statistics - For a site that fails to meet a standard, the
design value is the concentration that must be reduced to the level of the
standard for site compliance. Design values computed for sites that meet the
standard would be less than or equal to the standard. The design value is a
convenient summary statistic, but it is not easy to compute because it may in-
Q
volve factors other than the actual air quality concentrations. Although in-
formative, these statistics are difficult to recommend for general use because
of the complexities of their estimations.
Multisite Indicators - An indicator that combines data from several sites
q
(e.g., PSI) can be useful. A recent paper by Cox and Clark suggests areawide
indicators for examining trends for regional-scale pollutants such as CL. A
potential problem with multisite indicators is that data from cleaner sites can
mask higher concentrations at other sites. The seriousness of this problem de-
pends on the purpose of the study.
Multipollutant Indicators - An indicator that combines data for several
pollutants should scale the measurements so that the contribution of each pol-
lutant is appropriate. The PSI was developed for this purpose, and is recom-
2 3
mended as the standard indicator for these types of applications. ' It pro-
vides a convenient air quality indicator for daily reporting. A multipollutant
indicator is not recommended for characterizing trends because of the difficul-
ty of interpreting the results; for example, improvement in one pollutant may
mask degradation in another. The logical step in interpreting the results is
to examine trends for each pollutant.
6.1.5 Conclusions
When the purpose of the analysis is to compare standards, an indicator
that relates directly to the standard is needed. For pollutants with annual
mean standards, the annual mean also suffices for trend analyses. For pollu-
tants with only peak value standards, peak value and exceedance statistics
are acceptable for trends if adjusted for sample size; upper percentiles (90th,
95th, and 99th) should also be considered. For pollutants (e.g., 03) with
clear seasonal peaks, statistics based on data for only the peak season are
acceptable. In all cases, the data should be summarized for an averaging time
that corresponds to the averaging time of the standard.
Multisite indicators may warrant further study, particularly for areawide
pollutants. In national trend assessments, data are available for hundreds of
6-6
-------
sites; but because many subsets of these sites are correlated, interpretation
of the results is complicated. Multisite indicators could be used to aggre-
gate the data for appropriate subregions as an intermediate step in evaluating
trends.
Another area that warrants further attention is the development of popu-
lation exposure indicators. Such indicators could provide not only a useful
technical tool but also an effective means of conveying information to the
general public.
6.2 TREND TECHNIQUES
A question of interest is whether air quality has changed over time. The
search for the answer results in a variety of analyses categorized as trend as-
sessment. Air quality trends analyses can vary in complexity from a simple nu-
merical comparison of statistics from two time periods to a detailed time-series
analysis incorporating the effects of meteorology and emission control programs.
This subsection is a general discussion of trend techniques for air quali-
ty data. More detailed information on certain techniques is contained in EPA's
Guidelines for the Evaluation of Air Quality Trends. Another useful report,
Methods for Classifying Changes in Environmental Conditions, has recently been
prepared in conjunction with the development of EPA's environmental profiles
effort.11
6.2.1 Background and Purpose
Because of the general public's interest in air pollution, questions con-
cerning air quality are fairly common. Air pollution summaries in units such
as micrograms per cubic meter need a useful frame of reference to make the re-
sult more meaningful. The NAAQS's are one basis of comparison; another is how
air quality values have changed over time, which is a convenient practical way
to provide a perspective on what the numbers mean.
Both the general public and the technical community have a vested interest
in air pollution control programs, and trends analyses are frequently used to
evaluate the effectiveness of these programs. These analyses can be quite so-
phisticated. For example, ambient CO trends during the mid-1970's in New Jersey
were the result of both the Federal Motor Vehicle Control Program (FMVCP) and
an Inspection/Maintenance (I/M) Program instituted by the State; this result was
further complicated by the residual impact of the gasoline shortage in 1974-75.
6-7
-------
These factors, in addition to the possible influence of varying meteorology,
rapidly escalated the level of detail needed in an analysisnot only with re-
spect to the trend techniques used but also with respect to the types of data
needed.
Another consideration in air quality trends analyses is the scope of the
study. The study may involve data from a single site, or it may involve a na-
tional data sample from hundreds or thousands of sites. A technique that may
be reasonable for an individual site may not be feasible for a more extensive
data set.
6.2.2 Types of Trend Techniques
A wide variety of trend techniques have been used for air quality trends
analyses. This section briefly discusses nonstatistical and statistical ap-
proaches.
Nonstatistical - The usefulness of graphs (discussed in more detail in
Section 8) in any data analysis should not be underrated. At its simplest lev-
el, the purpose of a trend analysis is to determine what patterns are present
12
in the data. Box-plots, time-series plots, histograms, and so forthall
serve as convenient guides to the data analyst. In many cases, a careful choice
of the proper graphical technique suffices to indicate trends (e.g. Whittaker
Henderson,13 Box-plots,14'15 and two-way tables16'17'18).
A second nonstatistical approach is a simple numerical comparison between
summary statistics for two time periods. This is sometimes modified so that
trends are categorized as "no change" if the change fails to exceed some speci-
fied limit. This seemingly arbitrary cutoff may involve an underlying stati-
stical rationale and an assumed variance component.
Statistical - The statistical technique may merely be a comparison of two
annual means by using a standard t-test. When a sequence of data values is
19
available, regression analysis may be used. This may be either parametric or
nonparametric. ' Trends analysis using regression basically assumes a linear
trend over time. In many practical situations, this assumption may not be rea-
sonable, so alternative procedures (e.g., analysis of variance (ANOVA)) may pro-
vide a more flexible framework for indicating change. Because of the time de-
pendencies in sequences of air pollution data, time-series models have been used
for air quality trends analyses. Intervention analysis is one of the techniques
20
used to examine control strategies.
6-8
-------
The above techniques apply to a single series of data points. For analy-
ses of data from several sites, the results at each site may be summarized as
either up or down, and then analyzed as a contingency table; this technique as-
sumes that the sites are independent.
6.2.3 Desirable Properties
From a practical viewpoint, a desirable trend technique would be one which
is intuitively easy to understand, feasible to implement, based on reasonable
assumptions, and capable of providing results that are easily interpreted.
Intuitively Easy to Understand - People are more comfortable if the
rationale for the analytical technique is intuitively understandable. A tech-
nique may always be implemented on a "black box" basis: the user submits the
required input data, and the answer appears as the output. For the result to
be useful, however, the user must be confident that the technique will work.
This does not necessarily mean that the technique itself must be simple; it
means that the underlying concepts must be easily explained. For example, an
autoregressive, integrated, moving average model can be presented with a bar-
rage of notations involving forward and backward difference operators. Yet,
the need for a model that incorporates seasonal and diurnal patterns and depend-
encies from one hour to the next is more easily understood.
Feasible to Implement - It is a truism to state that the best trend tech-
nique available will be ignored unless it is feasible to implement. For air
quality analyses, several considerations are involved in evaluating feasibili-
ty. Because there are limited data histories in many practical situations, the
technique must be applicable to relatively short time periods (e.g., 2 to 5
years). For an analysis of data from a few sites, one may compute many types
of summary statistics; for national analyses, choices must be limited to those
readily available from NADB, which essentially limits the selection to certain
standard quarterly or annual statistics. (Monthly statistics are not stored,
and would have to be computed from raw data.) The number of sites involved
also places practical constraints on the amount of computer time and analyst
time available per site.
Based on Reasonable Assumptions - Every statistical procedure has an un-
derlying model requiring that certain assumptions be satisfied. These assump-
tions should be met by the air quality data for a test to be useful; therefore,
6-9
-------
the stipulations should be structured so that the air quality data can reasona-
bly be expected to meet them.
Results Easily Interpreted - After the analysis is complete, the success
of a technique depends on whether the results can be easily interpreted. This
is one advantage of a statistical technique, as opposed to a nonstatistical
technique, for assessing change. The probability assigned by the statistical
procedure is an aid in interpreting the results; it also affects the type of
statistical technique that should be used. For instance, if the annual rate
of change is the variable of practical interest, a procedure such as regres-
sion should be used since this is the type of information it produces.
6.2.4 Discussion of Candidate Techniques
Important in any discussion of candidate trend techniques for air quality
data is the interpretation of the final result. Because of the strong season-
al ity of certain pollutants (e.g., CU), the question of interest often involves
trends for a particular season. From a practical viewpoint, if the effective-
ness of a control strategy is of concern, only the peak months of the year may
be of interest. Because the NAAQS's are stated in terms of annual status as-
sessments (with the exception of Pb, which is quarterly), analyses are often in
terms of annual summary statistics. Use of an annual summary statistic reduces
the number of data points available for the analysis; but if the trend in this
summary statistic is the item of interest, the analysis should be structured
around the statistic.
The remainder of this subsection briefly discusses broad classes of trend
techniques, and indicates comparative strengths or weaknesses.
Graphical Procedures - Graphical presentations (discussed in more detail
in Section 8) are included here to emphasize that they are a useful first step
in any data analysis, not merely a final step for presentation purposes.
Numerical Comparisons - A simple comparison of summary statistics from two
time periods is of minimal use in most cases because it provides no frame of
reference for what is a normal change from one year to the next. Any underly-
ing statistical rationale for categorizing changes as up, down, or as no change
should be stated in the analysis. If the results of these comparisons are ag-
gregated over many sites for a contingency table analysis, they are acceptable.
Statistical Comparisons - Statistical comparisons may be made using
either parametric or nonparametric techniques. A typical parametric technique
6-10
-------
would be a t-test to compare 2 years of data, although the presence of sea-
sonal ity may artificially inflate the error term and reduce the sensitivity of
the test. Nonparametric tests may be used to avoid the assumptions that the
errors are normally distributed. The Wilcoxon signed-rank test allows for sea-
sonal ity, and is therefore a useful nonparametric technique. Aligned-rank
tests have been suggested for assessing changes in environmental data; how-
ever, experience with these tests for ambient air quality trends is limited at
this time.
Regression - Both parametric and nonparametric regressions provide an es-
timate of the rate of change as well as a statistical significance test. In
assessing air quality trends, the rate of change is often important because it
relates directly to the effectiveness of a control strategy rather than simply
stating whether an increase or decrease has occurred. Because of seasonal pat-
terns in the data, regression is normally applied to an annual statistic or a
summary statistic for the peak season; this is a disadvantage since only one
value a year is used. Regression also tests for a linear trend, which is use-
ful either when the user is interested in a linear trend or in a relative mea-
sure of net change over the time period; but if a finer resolution of the pat-
tern is desired, regression may not be adequate.
Analysis of Variance - The general class of analysis of variance (ANOVA)
models offers considerable flexibility for trends analysis. The user can in-
troduce month, season, and year terms into the model if a single site is being
examined. For an areawide analysis, a site-effect term can be added as well as
interactions and even covariates. ANOVA is not restricted to linear trend as-
sumptions, and multiple comparison tests are available to test for significant
differences among effects of interest.
Time-Series Analysis - A basic assumption in many statistical techniques
is that successive measurements are independentthat is, the value of a par-
ticular measurement does not depend on past measurements. In view of the
diurnal and seasonal patterns often present in air quality data, one may use
time-series techniques for these problems. Applications of time-series tech-
20
niques for air quality analyses have used Box-Jenkins techniques, Fourier
21 22
analysis, and polynomials to remove the seasonal components. Intervention
analysis techniques seem appropriate for examining control strategies and other
20
types of intervention. While these techniques have the advantage of produc-
ing more information, they are also more difficult to apply; consequently, it
6-11
-------
is difficult to recommend them for routine large-scale analyses.
6.2.5 Conclusions
At the present time, it does not seem feasible to recommend specific trend
techniques. Since trends analyses are done for different purposes, the level
of resolution required in the answer may vary from a simple up-down classifica-
tion to an actual determination of the percentage improvement associated with
a specific control strategy. However, it does seem advisable to encourage the
use of statistical tests as an objective means of classifying trends.
Areas that warrant further examination include the intervention/time-series
approach to see if the implementation of this type of analysis can be simplified
so that its use can become more routine. Also, more experience is needed in ap-
plying the aligned-rank test to air quality data.
6.3 REFERENCES
1. Kendall, M. G., and W. R. Buckland. A Dictionary of Statistical Terms.
Hafner Publishing Company, Inc., New York, N.Y., 1971.
2. Guideline for Public Reporting of Daily Air Quality Data - Pollutant
Standards Index (PSI). No. 1-2-044. Office of Air Quality Planning and
Standards, U. S. Environmental Protection Agency, Research Triangle Park,
N.C., August 1976.
3. Federal Register 44(92)-.27598, May 10, 1979.
4. Frank, N. H., et al. Population Exposure: An Indicator of Air Quality
Improvement. 70th Annual Meeting of the Air Pollution Control Association,
Toronto, Canada, June 1977.
5. Johnson, Ted. Precision of Quantile and Exceedance Statistics. American
Society of Quality Control Technical Conference Transactions, Atlanta, Ga.,
1980.
6. Johnson, T., and M. Symons. Extreme Values of Weibull and Lognormal Dis-
tributions Fitted to Ambient Air Quality Data, Paper No. 80-71.4, Annual
Meeting of the Air Pollution Control Association, Montreal, Canada, June
1980.
7. Federal Register 44(28):8282, February 8, 1979.
8. Curran, T. C., and W. M. Cox. Data Analysis Procedures for the Ozone
NAAQS Statistical Format. Journal of the Air Pollution Control Associa-
tion 29(5):532, May 1979.
9. Cox, W. M., and J. B. Clark. An Analysis of Ozone Concentration Patterns
Among Eastern U.S. Urban Areas (draft).
6-12
-------
10. Guidelines for the Evaluation of Air Quality Trends, OAQPS No. 1.2-015.
U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, N.C., February 1974.
11. Methods for Classifying Changes in Environmental Conditions. VRI-EPA
7.4-FR3 0-1 (draft). Vector Research, Inc., Ann Arbor, Michigan,
February 1980.
12. Tukey, J. W. Exploratory Data Analysis. Addison-Wesley Publishing Com-
pany, Reading, Mass., 1977.
13. Spirtas, R., and H. 0. Levin. Characteristics of Particulate Patterns
1957-1966. U.S. Department of Health, Education, and Welfare, National
Air Pollution Control Administration, Raleigh, N.C., March 1970.
14. National Air Quality and Emissions Trends Report, 1976. Publication No.
EPA-450/1-77-002. U.S. Environmental Protection Agency, Office of Air
Quality Planning and Standards, Research Triangle Park, N.C., December
1977.
15. National Air Quality and Emissions Trends Report, 1977. EPA-450/2-78-052.
U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, N.C., December 1978.
16. Phadke, M. S., et al. Los Angeles Aerometric Data on Oxides of Nitrogen
1957-72. Technical Report No. 395. Department of Statistics, University
of Wisconsin, Madison, Wise., 1974.
17. Tiao, G. C., et al. Los Angeles Aerometric Ozone Data. Tech. Report No.
346. Department of Statistics, University of Wisconsin, Madison, Wise.,
1973.
18. Tiao, G. C., et al. Los Angeles Aerometric Carbon Monoxide Data. Tech.
Report No. 377. Department of Statistics, University of Wisconsin, Madi-
son, Wise., 1974.
19. The National Air Monitoring Program: Air Quality and Emissions Trends.
Annual Report Volume 1. Pub. No. EPA-450/l-73-001a. U.S. Environmental
Protection Agency, Office of Air Quality Planning and Standards, Research
Triangle Park, N.C., July 1973.
20. Box, G. E. P., and G. C. Tiao. Intervention Analysis with Applications to
Economic and Environmental Problems. Journal of the American Statistical
Association 70:708-79, 1975.
21. Phadke, M. S., M. R. Grupe, and G. C. Tiao. Statistical Evaluation of
Trends in Ambient Concentrations of Nitric Oxide in Los Angeles. Environ-
mental Science and Technology, Vol. 12, No. 4, April 1978.
22. Horowitz, J., and S. Barakat. Statistical Analysis of the Maximum Concen-
tration of an Air Pollutant: Effects of Autocorrelation and Non-stationarity.
Atmospheric Environment 13:811-818, 1979.
6-13
-------
7. INFERENCES AND CONCLUSIONS
This section focuses on the problem of identifying cause-and-effect rela-
tionships after a statistically significant trend or difference between data
sets has been determined. Previous sections have discussed the problems of
data completeness, historical trends criteria, precision and accuracy of data,
screening data for outliers, siting and representativeness, statistics for
analyzing air quality data, and trend techniques. Assuming that all of these
problems can be resolved in a particular analysis, the problem of interpreta-
tion remains.
7.1 BACKGROUND
At the present time, most air quality data available for analysis are
submitted to NADB. The major enhancement to this system is the current de-
signation of NAMS--a refined national monitoring network in areas with large
populations and high pollutant concentrations. Each NAMS will meet uniform
criteria for siting, quality assurance, equivalent analytical methodology,
sampling intervals, and instrument selection to assure consistent data report-
ing among the States. Precision and accuracy data will be available with the
air quality data for the first time.
Though the enhancements achieved through the monitoring regulations will
provide a much sounder national air quality data base for decisionmakers, the
data are still seriously deficient in determining cause-and-effect relation-
ships between air quality, emission changes, meteorology, and the impact of un-
anticipated events such as fuel shortages. Explaining the causes of air quali-
ty changes is difficult without information to supplement basic air quality
data. For example, what is the impact of a locals regional, or national gaso-
line shortage on ambient air pollution levels? What is the probable impact of
fuel switching or tampering with automotive emission control devices on air
quality levels?
A case in point is the recent controversy over 03 trends in Los Angeles.
A significant increase was observed in 0, levels between 1977 and 1978. An
7-1
-------
2
attempt was made by the South Coast Air Quality Management District to deter-
mine how much of the increase was due to meteorology; they used a meteorologi-
3
cal index developed by Zeldin; stated that a significant increase in 0, levels
remained after they had adjusted the data for meteorology; and speculatively
attributed it to the breakdown of catalytic control devices. What was needed
was concomitant information on emission changes; independent estimates of fuel
switching, tampering, and so forth; and more complete meteorological data.
The problem in making inferences and conclusions is largely the lack of
supplemental information needed to interpret air quality data analysis and
trends. This lack could be improved in two ways: (1) the data analyst could
try to find concomitant information (e.g., weather data collected at airports
by the National Oceanic and Atmospheric Administration), or (2) a long-term
solution would be to apply the principles of experimental design and enhance
the NAMS in two or more major urban areas with supplemental information on
meteorology and emission changes. The first approach is discussed in the fol-
lowing section on case studies, and the second is discussed in the section on
long-term solutions.
7.2 CASE STUDIES
Over the years, a number of attempts have been made to explain ambient
air quality trends in terms of changes in emissions, meteorological conditions,
or instrument measurement practices (e.g., changes in laboratory procedures or
calibration). The approach can best be illustrated by examples, as in the fol-
lowing four case studies.
7.2.1 Comparison of SOg Trends and the %S Content Regulations in Distillate
and Residual Fuel Oil in Bayonne, N.J.
The impact of regulations for controlling the sulfur content of fuels on
ambient S00 levels was illustrated in Bayonne, N.J., by comparing ambient S09
4
levels with the effective dates of regulations limiting the sulfur content
(Figure 3). The EPA report which presented the analysis stated that the im-
provement in S02 air quality at this site can be attributed primarily to regu-
lations which became effective in New Jersey and New York during 1968-72. The
trend at the Bayonne site is consistent with the national trend in SO^ during
this time period, and is probably due to changes in the allowable sulfur con-
tent of fuel.
7-2
-------
£50
EDO
550
500
450
| 400
£ 350
ff 300
250
200
150
100
50
' I ' ' ' I ' ' ' I
24-HOUR PRIMARY STANDARD (365)
| I
MAXIMUM24-HOUR CONCENTRATION
QUARTERLY LINE
TREND UNE
ANNUAL PRIMARY STANDARDS (80)
_1 J i I I Ti I I I I I
\--xx ^
I I I I i I I
19S7
139
1969
1S70
YEAR
isn
1372
1973
S/1/68COAL*1%S
RESIDUAL OIL-KS|
UGHTOIL-CJSS
|PRE-1958FUEL-L55SJ
ID/1/70 RESIDUAL OIL « US S
LIGHT OIL-0.«J
JO/1/71 ALL OIL-0.35 S
Figure 3. Comparison of S02 trends at Bayonne, N.J., with regulations
governing % sulfur content in fuel.
7-3
-------
The analysis could have been improved, and the conclusions strengthened
if appropriate meteorological data had been available to be treated as a co-
variate. As it stands, however, the graphical presentation suggests a reason-
able cause-and-effect relationship between the decreasing levels of SCL and
the increasing restrict!veness of regulations limiting the sulfur content of
fuel.
7.2.2 Impact of Gasoline Shortage on CO in Richmond, Va.
An analysis of CO data in Richmond, Va., illustrates the problem of in-
5
terpreting ambient data. In 1974, an EPA analysis was undertaken to evaluate
the effect, if any, of the energy crisis on CO levels. The CO monitoring site
in downtown Richmond was representative of the influence of commuter traffic
patterns.
The period of time chosen for the analysis was the last 4 weeks (28 days)
of January and the 4 weeks of February. This time period was expected to re-
flect the most severe period of the gasoline shortage and to minimize the po-
tential anomalies in traffic patterns associated with the Thanksgiving, Christ-
mas, and New Year's holidays. The windspeed data recorded at the R. E. Byrd
International Airport served as an approximate indicator of the windspeed at
the site. The data were presented as weekly averages to compensate for the
differences in daily traffic patterns (Table 10).
TABLE 10. WEEKLY AVERAGE CO CONCENTRATIONS (ppm) AND WINDSPEEDS (mph)
IN RICHMOND, VA, JANUARY 4-FEBRUARY 28, 1974
Week
1/04-1/10
1/11-1/17
1/18-1/24
1/26-1/31
2/01-2/07
2/08-2/14
2/15-2/21
2/22-2/28
Hourly
average
2.92
1.94
3.07
2.88
2.28
2.31
2.08
1.73
Daily
8-h max.
4.45
3.13
4.79
4.56
3.54
3.76
3.09
2.81
6 to 9 a.m.
and
4 to 7 p.m.
average
4.50
3.18
4.52
4.20
3.56
3.40
3.00
2.82
Average
daily
windspeed
5.90
7.87
6.63
6.21
7.24
8.06
9.13
9.67
The CO data showed a downward trend during this 8-week period; however,
the average windspeed increased during this period (Figure 4). The decline
7-4
-------
3.0
I.
o.
a:
t
LU
o
o
2.0
1.0
I T
I i 1 T
AVERAGE CO
' TJlND SPEED
I I I I I
1
I I
30
20
a.
CO
10
1 2 3
JANUARY 1974
678
FEBRUARY 1974
WEEK
Figure 4. Weekly average CO and wind speed in Richmond, VA
from January 4 to February 28, 1974.
7-5
-------
in CO levels may, therefore, have been due to the increase in average wind-
speeds, which is indicative of greater dilution.
The interdependence was seen in the statistical analysis of the data. If
the change in windspeed was ignored, four parametersdaily average, maximum 8-
hour average, rush-hour average, and nonrush-hour averageshowed statistically
significant decreases, although the rush-hour decrease was less apparent in the
averages. These findings were based on ANOVA. When windspeed was introduced
into the analysis as a covariate, using analysis of covariance, none of the
changes in the above parameters were significant. Therefore, although the data
at this site indicated a decline in CO levels during this period, the associat-
ed increase in windspeeds made the cause of the decline difficult to assess.
Failure to detect significant trends after adjusting for windspeed is not
entirely unexpected. The variability associated with CO measurements and the
relatively brief duration of the gasoline shortage make it possible for the ef-
fect to go unnoticed, unless the monitoring site itself was in precisely the
right location to detect changes due to alterations in traffic patterns.
This example illustrates one problem in trying to use data collected for
one purpose for yet another purpose. The assumption that the windspeed at the
airport reflects the windspeed at the downtown Richmond site may not be cor-
rect. Ideally, the windspeed should be measured at the CO monitor so the ana-
lyst could be more confident in interpreting the results.
7.2.3 Trends in CO in New Jersey
An analysis of CO data in New Jersey illustrates an attempt to relate
statewide CO trends to gasoline consumption. Figure 5 was prepared by the N.J.
Department of Environmental Protection to illustrate the progress made in re-
ducing State CO levels from 1972 through 1976. The dates for the initiation
of the two phases of their inspection/maintenance (I/M) program are shown.
New Jersey indicated that CO levels continued to improve despite an overall in-
crease in gasoline consumption.
While progress in CO levels is clearly evident, the effectiveness of the
New Jersey I/M program is confounded with the overall effectiveness of the CO
emission reductions attributed to the Federal Motor Vehicle Control Program
(FMVCP). Ideally, it would be desireable to separate the effects of the two
control programs; this could be accomplished with a designed experiment. Con-
comitant meteorological data would be desirable to determine if changes in
meteorology would aid in explaining the statewide CO trend.
7-6
-------
Q.
a.
LU
es
I
CM
CO
IMPLEMENT
PHASE I
I/M PROGRAM
GASOLINE
CONSUMPTION
AMBIENT
CO
IMPLEMENT
PHASE II
I/M PROGRAM
300
275
CD
u>
O
C3
LU
et
O
S
250 o
CM
iI
z:
o
ii
Q.
225
o
o
o
<
CJ3
200
1972 1973 1974 1975
TIME, years
1976
1977
Figure 5. CO air quality from 18 monitoring sites and motor-vehicle
gasoline consumption for N.J. from 1972 through 1976.
7-7
-------
7.2.4 The Central Plains Drought
On February 24, 1977, the extremely dry soil conditions in the Central
Plains and a strong frontal system resulted in dust being stirred up and trans-
ported east. Figure 6, a modified box-plot, shows peak TSP levels in Region VI
(Central Plains) by quarter from 1972 through 1977; the dramatic increase in
the first quarter of 1977 is obvious on this graph. Monitoring sites through-
out Texas, Oklahoma, and Arkansas reported high TSP levels during this February
duststorm. Several sites recorded daily values of more than 1000 yg/m ; a sin-
gle value of this magnitude would increase the annual geometric mean at a site
by 10 percent.
In this example, the high TSP levels reported throughout Texas, Oklahoma,
and Arkansas are substantiated by satellite pictures taken February 23-25, 1977
(Figure 7).
7.2.5 Conclusions and Recommendations
With the existing NAMS network, the best an analyst can do to facilitate
the interpretation of air quality monitoring data is to seek other sources of
information to help explain why an air quality trend has or has not taken place
or why there is a significant difference between sets of data. The four case
studies illustrate the importance of meteorology, emission changes resulting
from control programs, and the impact of extraordinary events (e.g., the dust-
bowl of February 1977). A long-term solution is application of the principles
of experimental design to the collection of air quality and appropriate con-
comitant meteorological data. Such a solution is discussed in the following
section.
7.3 LONG-TERM SOLUTIONS
One solution to the problem of trying to assess the effectiveness of EPA's
control programs would be to select two or more urban areas and to collect con-
comitant information on a continuing basis to aid in answering both anticipated
and unanticipated questions about effectiveness. The objective would be to de-
termine cause-and-effect relationships between air quality and emissions after
isolating the effects of meteorology, these relationships could be determined
if data could be collected to estimate the impact of gasoline shortages, fuel
switching, automobile control device tampering, contributions of various sources,
-------
_H> -
01
o
-M
OJ
r~~
CTi
S- O
t/1
4->
o
-I- I
cot-^
01 en
o: i-i
c O
(D
E OJ
X O)
Q- CD
00 C
>,*
S- -P
Ol (/)
+J 3
S- i
ro i
3 -i-
cr
O)
§o
o
C\J O
tiu/Brt
dSl
7-9
-------
-o
O)
co
o
s_
o
-4->
to
-t->
to
3
-a
CT1
CM to
I -O
CO O
C\J !-
S-
>, <1)
S_ CL
to
3 0)
S- E
J3 T-
QJ !->
U_
to
OJ
O)
03
CO
O)
S-
CD
7-10
-------
changes in instrument calibration, degree of human exposure, extent of fugitive
dust, and so forth.
Air quality, monitoring, and emissions data would be collected according
to an experimental design to test stated hypotheses. Each of the criteria pol-
lutantsISP, S02, CO, CL, and N02--would be measured in the two or more urban
areas along with IP and sulfates. Sources of emission changes to be consider-
ed in developing the experimental design include but are not limited to: (1)
changing economic conditions; (2) increased or decreased refinery capacity; (3)
fuel switching; and (4) growth patterns (e.g., the number of dry cleaning estab-
lishments).
At a minimum, experimental designs would use analysis of variance (ANOVA)
and the analysis of covariance (COVA) to test hypotheses, and would use confi-
dence intervals about means in question to display significant findings. Other
appropriate statistical techniques would also be used to test hypotheses.
A contract has been let to explore this long-term solution in greater de-
tail. The objective is to pose policy questions and then to determine what
supplemental information would be needed to answer the questions in two or more
urban areas. The associated costs are also to be determined. When the report
on this contract is completed, its principal findings will be summarized and
appended to this report.
7.4 REFERENCES
1. Federal Register. Vol. 44, May 10, 1979, pp 27588-27604.
2. Davidson, A. et. al., Air Quality Trends in the South Coast Air Basin.
South Coast Air Quality Management District, El Monte, Calif. June 1979.
3. Zeldin, M. D., and D. M. Thomas, Ozone Trends in the Eastern Los Angeles
Basin Corrected for Meteorological Variations. Presented at the Inter-
national Conference on Environmental Sensing and Assessment, Las Vegas,
Nevada, 1975.
4. Monitoring and Air Quality Trends Report, 1972. Pub. No. EPA-450/1-73-004.
U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, N.C. December 1973.
5. Monitoring and Air Quality Trends Report, 1973. Pub. No. EPA-450/1-73-007.
U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, N.C. October 1974.
6. New Jersey Department of Environmental Protection Annual Report: July 1,
1976-June 30, 1977. New Jersey Department of Environmental Protection,
Trenton, N. J. (in preparation).
7-11
-------
7. Use of Meteorological Data in Air Quality Trend Analysis. EPA-450/3-78-024.
U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, N.C. May 1978.
7-12
-------
8. DATA PRESENTATION
The purpose of this section is to review well-established data presenta-
tion techniques applicable to aerometric measurements and to provide guidance
on selection of the forms most suitable to the scope of the problems to be
analyzed. Selection is based on criteria such as audience applicability,
spatial and temporal classification, and availability of computerized statis-
tical and graphical resources. Therefore the discussion is focused on (1)
the fundamental concepts to be displayed, (2) chart types, (3) input data
forms, (4) statistical classifications, (5) audiences, (6) caveats/enhance-
ments, and (7) available plotting resources. Finally, displays that meet
these requirements are included.
8.1 CONCEPTS TO BE DISPLAYED
A simple but accurate presentation of statistical measures of large data
bases or parametric relationships contributes greatly to the reader's under-
standing of the data. The most common displays of data on air quality and
source emissions include (1) current status of the pollution problem, (2) sta-
tistical or descriptive trend, (3) impact of one or more parameters on another,
(4) comparison of two or more groups or classes for a specific parameter, (5)
relation of component parts to the whole, and (6) spatial and temporal patterns.
Statistical measures may be chosen not only from individual monitoring-site
or emission point-source populations of averages, medians, percent!les, maximums
or standards exceedances but also from aggregates (weighted or unweighted) of
these sites on a city, county, region, state, national, or global scale.
Data presentation should be carefully planned. The charts or graphs, as
well as the statistical displays, may be easily distorted (intentionally or
unintentionally) by compression or expansion of scales, by nonuniform or broken
gridding, by line and shading optical illusions, and by cluttered or complex in-
formation. The following subsection introduces the most frequently used dis-
plays.
8-1
-------
8.2 CHART TYPES AND USES
8.2.1 Tabular Summaries
Summarizing data in tabular format is an important first step in data
analysis. The validity and accuracy of the data should be certified before
applying summary techniques. The tabular summary provides a permanent record
of the individual data items which can be further analyzed in future research
projects.
Aerometric examples of this form are NADB's monthly listings of hourly
air quality values and yearly listings of intermittent (daily or composite)
values. These data are also summarized by NADB in the yearly and quarterly
frequency distribution listings. Similarly, point-source emissions data are
submitted to NADB by the National Emissions Data System (NEDS).
Tabular summaries provide comparisons of several parameters categorized
at one or more levels. Examples are listings of TSP, SCL, and NCL by day of
the year (Table 11), weekly TSP/maximum values at several monitoring sites
(Table 12), and TSP monthly averages by site type (Table 13).
Finally, either certain qualitative measures may not be readily applica-
ble to graphing or more information is provided in a smaller document. One
example is a display of current attainment status and trends using a variety
of symbols and colors, as shown in Figure 8.
Tabular listings or summaries should be limited in most cases to technical
support documents. For regional or national publications, these summaries
should be limited to comparisons of one or two parameters and about five class-
es or categories. Tedious or complex tables often contribute to the reader's
confusion and misunderstanding; moreover, descriptions of statistical tech-
niques used in producing these tables are often either left out or limited in
the text discussion.
8.2.2 Point Charts
Point charts are used primarily to describe situations where scatter or
clustering are important. Since these displays for large data bases would be
almost impossible to draw manually, computerized techniques have been applied.
Rapid display of data to detect outliers, parametric relationships, or data
distribution is clearly a benefit to the statistician. Common uses are corre-
lation or cluster analysis (Figures 9 and 10), plotting air quality concen-
trations by year (Figure 11), and comparison of air quality for several
8-2
-------
TABLE 11. MULTIPLE PARAMETER LISTING, 1979, yg/nr
Yr/mo/day
79/01/01
79/01/02
79/01/04
79/12/31
TSP
125
86
145
SO 2
35
43
50
83
NO 2
45
52
96
TABLE 12. WEEKLY TSP MAXIMUMS AT CITY SITES, 1979, yg/m3
Week
beginning
Jan 7
Jan 14
*
Dec 29
1979 Max
1979 Avg
Sites
4th & Market
120
163
*
103
236
125
6th & Vine
50
No sampling
.
*
52
99
55
18th Jackson
70
99
*
63
136
86
City
max
120
163
103
236
125
City
avg
80
131
73
187
77
TABLE 13. REGION 5 MONTHLY AVERAGES BY SITE TYPE, 1979, yg/m3
Month
Jan
Feb
Dec
Total
Commercial
Sites
Obs
Avg
Industrial
Sites
Obs
Avg
Residential
Sites
Obs
Avg
Remote
Sites
Obs
Avg
8-3
-------
County
Oefferson
La Plata
Larimer
Las An i mas
Logan
Mesa
Moffat
Montezuma
Montrose
Morgan
Otero
Pitkin
P rowers
Pueblo
Routt
Weld
TSP
LlJ
O
^ ft
b
o
o
iC>
o
o
o
E>°
*
*
OL)>
s$>
B^
so,
o
NO 2
CO
s>
5>
°x County
['^ Adams
Alamosa
Arapahoe
Archuleta
Boulder
Clear Creek
Delta
Denver
Douglas
Eagle
El Paso
Fremont
Garfield
Gunnison
Huerfano
O
TSP
[>'
0,
[*)>
&
O
&>
t.
^
s>8
^
w+
[£>
0
so-
O
0
wo,
0
0
o
CO
^
t
It
Ox
L5J
s>
O
a/ Status based on annual
mean only
l~n Ho evidence standard exceeded
[if] Exceeds primary standard
| Exceeds alert level
] j Increasing trend (deterioration)
| J^ Ho apparent trend
Decreasing trend (improvement)
Figure 8. Status and trends in air quality in Colorado.
8-4
-------
cc QJ
200,
150
100
90
80
70
60
50
40
30
25
20
15
10
10 15 20 25 30 40 50 60 70 80 90100
CONCENTRATION, yg TSP/m3
(Site 397140020H01)
150 200
Figure 9. Intersite correlation test data.
-------
. PHY 3'
Mr 85
Nfi
MM
I'M
I - KOI
t.BJ
9 - CIO«
« - IS
i - «£S1 N)N£ COR*
C . SDUTM HIKC CMf
nc»N
i I
1C.70 1C.1C
IB.OC ic.oe
s.-re
c.ee
C.01
.
c.ci
4-1!
».*!
t - en
s.n
s.st
t.Tt
t.e;
t.TI
c.oc
D.7I
D.OJ
S.1J
»*.«
&.»»
-I.SI
J.«
e.u
C.OB
O.DI
-J.M
K-«
(.OC
DO CWIf
« sou Mi«a -
LMEL « »-*«
Figure 10. Magna, Utah, day 3, 0.50 probability ellipses of the west-east and
south-north wind components for three cluster types. Winds from the west and
south are positive.
-------
600
500
400
300
200
100
PRI
SEC
TSP
- - - tr
o o o o
o° °
° oo
o ° o o o
-, o o o
o o 0 o
o I
FEE APR JUN AUG OCT
MONTH
Figure 11. Twenty-four hour TSP values, 1972.
-------
categories (Figure 12).
Point charts usually are accompanied by some "best-fitting" solid line to
describe annual trend or diurnal or seasonal pattern. The integrity of the in-
dividual points is the foremost reason for displaying data in this way.
8.2.3 Line or Curve Charts
The line or curve chart is perhaps the most widely used method of present-
ing summarized data graphically. The chart is also the most easily constructed
manually. The most common uses in displaying air quality or emissions data are:
1. Data coverage over a long time period (Figure 13).
2. Emphasis on movement rather than on actual amount (Figure 14).
3. Comparison of several series (same measurement unit) on same chart
(Figure 15).
4. Trends in frequency distribution; e.g., population exposure
(Figure 16).
5. Use of the multiple amount scale (Figure 17).
6. Estimates, forecasts, interpolation, or extrapolation (Figure 18).
8.2.4 Surface Charts
The simple surface or band chart depicts a single trendline with shading
or crosshatching filling in the area between trend and base lines to enhance
the picture of the trend. Thus, in this chart type:
1. The magnitude of the trend is emphasized,
2. A cumulative series of components of a total trend is depicted, and
3. Certain portions of the chart are accented for a specific purpose.
Figure 19 exemplifies classification of CO data by concentration level and by
number of monitored days.
8.2.5 Column/Bar Charts
Column/bar charts are intuitively simple for most readers to follow since
they accent discrete dates or categories with comparative heights of the columns/
bars for one main statistic. The primary purpose is to depict numerical values
of the same type over a given time period (i.e., multiple years), as shown in
Figure 20. Other uses are:
8-8
-------
800
700
600
500
400
300
200
100
o TSP
PRIMARY
SECONDARY
fee,
1 ft0 °c
§ JD
iS_ C3
o
D
QD
O
c
o
o
o o
o
o
o
o
£L
°- *J*U °
fa
B
0
D
o
o
11 IA
MI Mfl
STATE
OH
WI
Figure 12. Air quality data, 24-h TSP concentration values
October 15, 1976. "
8-9
-------
PRIMARY
» « 0,
FEE
APR
JUN
AUG
OCT
Figure 13. Maximum 1-h 03 values/day, 1977
(SAROAD site 14jl220002.P10).
8-10
-------
300
280
260
240
220
8 200
o
ii
<
180
x
o
OBSERVED AIR QUALITY -
A ADJUSTED AIR QUALITY
- - AIR QUALITY AS A FUNCTION
OF EMISSIONS
NUMBER OF DAYS
ABOVE 200 MJ/
(BASINWIDE TOTAL)
ANNUAL AVERAGE OF
DAILY 1-HOUR MAXIMUM
(AZUSA, PASADENA,
POMONA COMPOSITE)
(O
o
280 s.
>>
260 "
240
220
200
180
1965 1967 1969 1971
YEAR
1973
1975
Figure 14. Oxidant trends adjusted for meteorology.
8-11
-------
ANNUAL STANDARD
PRIMARY 24-H
ALERT 24-H
SECONDARY 24-H
ANNUAL AVERAGE
2ND MAX DAY
x-
D-
C/0
200
100
1970
1971
1972
YEAR
1973
1974
1975
Figure 15. Annual average and second-high day TSP values,
1970-75.
8-12
-------
-< o
D_
o
Q-
100
80
60
40
20
40
60
75 80
100
120
ANNUAL TSP CONCENTRATIONS, yg/m2
Figure 16. Population exposure distributions of annual mean TSP for 1970
and 1976 in city of Chicago.
8-13
-------
Q.
Q.
CD
OL
UJ
275
250
225
OL
CJ3
I
CM
O
k1
t
o.
21
rD
oo
o
o
o
t/>
<
CJ3
200
1972
1973
1974
1975
1976 1977
YEAR
Figure 17. Ambient CO concentration and gasoline
consumption, 1972-77.
8-14
-------
MONTHLY GEOMETRIC MEAN
12-MONTH RUNNING GEOMETRIC MEAN
1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
Figure 18. Comparison of monthly GM, 12-mo running GM, and
predicted monthly means (by double moving average method),
1964-74.
8-15
-------
JMBER OF DAYS BY CONCENTRATION INTERVAL
i « i ro
tn o <_n o
CD O CD O O
-
--.
1 1 TOTAL
IIIHIIIIII 9-13 ppm
1' ... 1 "n-17 nnm
1
--j.hrt:::; :;r.;i::
S::ii h:!:l:i!:»iSil.. - . -
::::j:;j::r:i;u::::!:;i::.'!:::-.:::::~;::::
mm^m^^^^
\%%ffi% >17 ppm
|y^y^>yg777??v3^Kg
1975
1976
1977
Figure 19. Trends in CO levels in New York's 45th Street Station,
April-January, 1975-77.
8-16
-------
Dm in
BO
700
ISO
100
50
l
»
: 0
1
; 200
; iso
)
5 100
>
1 M
! o
200
150
100
50
0
PSI Interval
100-200
PSI Interval
.. 200-300
Boston
-
Cincinnati
-
.
Denver
-
,-rr||tf PSI intervil ^
:ffiS >snn t.
250
200
150
100
50
0
200
150
100
50
0
200
150
100
50
0
Z2% both PSI Intervals
%22 200-300 and >300
Chicago
-
i
1
!lllll|
111
Dayton. Ohio
Fresno
_.
IIIIHI
1973 1974 1975 1976
1B3 1974 1975 1976
Figure 20. Trends in PSI levels, 16 cities, 1973-76.
8-17
-------
1. Comparison of numerical values of the same type for several cate-
gories (Figure 21),
2. Comparison of two or three independent series over time such as
grouped columns, subdivided columns, and three dimensions (Fig-
ure 22),
3. Display of increases or decreases, losses or gains, or deviation
from requirement or norm (Figure 23),
4. Display of ranges of maximal and minimal values for a series (Fig-
gure 24).
The wind rose is a familiar application of the bar graph in circular form
(Figure 25).
Caveats to be considered in preparation of these charts are irregular
time sequences, spacing between columns/bars, scale breaks, shading, and the
ordering or sequencing of items represented by the columns/bars.
8.2.6 Pie/Sector Charts
The familar pie/sector chart in the form of a circle compares component
parts, and shows their relation to the whole. Source emission categorization
lends itself well to this chart type (Figure 26), as does population exposure
(Figure 27). The pie chart is often used with line, column, bar, or map dis-
plays to exhibit geographic or categorical components of trends.
8.2.7 Map Charts
Map charts are attention-getters, and they are most applicable to en-
vironmental statistics, especially air quality. National, regional, State,
county, and city maps may depict formation and transport of air pollutants in
a real-time, dynamic sense. Moreover, within boundaries, current year-of-
record and trends may be shown through isopleths, symbols, or shading. Fig-
ures 28 and 29 demonstrate two applications of the treatment of data in the
dynamic (isopleth) and static modes.
8.2.8 Three-Dimensional Charts
Advanced computerized graphic techniques have made the three-dimensional
chart type a viable alternative to three-way tables and to restrictive two-
dimensional plane representations. The most significant applications have been
to dispersion modeling and contour mapping (Figure 30).
8-18
-------
- 5
5 4
ce:
uu
Q_
~ 3
o
_ ACTUAL (NEDS) AND ILLEI (IEPA DATA) ,
| ACTUAL
1 POTENTIAL
ILLEI
TSP
S02
NO,
HC
CO
Figure 21. Actual vs. potential emissions for Illinois, tons/year.
-------
County (city)
Archuleta (Pagosa Springs)
Boulder (Longmont)
Denver (Denver)
Eagle (Vail)
Fremont (Canon City)
Garfield (Rifle)
Moffat (Craig)
Otero (Rocky Ford)
Pitkin (Aspen)
Prowers (Lamar)
Pueblo (Pueblo)
Routt (Steamboat Springs)
Weld (La Salle)
Primary level exceeded
Alert level exceeded
10
1976
1977
1978
Figure 22. Number of days per year that the TSP primary
standard or alert level was exceeded, Colorado.
8-20
-------
O)
o
cu
Q.
00
O
D-
X
CL.
O
Q-
100
90
80
70
60
50
40
30
20
10
0
1172
I I
CU DECREASE M EXMMIK
i HKMEMf M tXfOSUKt '
1IT7TMEll*E
10
8 7 6
WEST
5 4
REGION
EAST
Figure 23. Regional changes in metropolitan population exposures to
excess TSP levels, 1972-1977 (width of each regional column is
proportional to its metropolitan population).
8-21
-------
90TH PERCENTILE
75TH PERCENTILE
COMPOSITE AVERAGE
MEDIAN
25TH PERCENTILE
10TH PERCENTILE
110
100
E
^ 90
f
^
o
^ 80
H-
§ 70
2^
o
o
uj 60
i
§ 50
Qi
2 40
Q
UJ
a
g 30
:r>
t/o
-j 20
«a:
r^
o
- 10
0
_
1
1
A
w-
A
1
i
i
ii
A
A
*
T 1
1
1
-
i
1
1
A
-
J
i
1 1 ~~
A
-»
T_
I-
^
|
1970 1971 1972 1973 1974 1975 1976
YEAR
Figure 24. Trends of annual mean TSP concentrations from 1970 to 1976 at
2350 sampling sites.
8-22
-------
0-3 4-7 8-12 13-18 19-24
SPEED CLASSES (mph)a
0123456789 10
SCALE, %
Bias removed and calms distributed.
Figure 25. Wind rose pattern.
8-23
-------
1.5 1.5
REGION 127
Totol. 15,387 Tons
REGION 128
Total. 51,993 Tons
REGION 129
Total: 99,189 Tons
2.5<1
80
K
E
Y
REGION 130
Total. 6,574 Tons
% FUEL COMBUSTION (residential,
utilities, industrial, commercial,
institutional)
\ INDUSTRIAL PROCESS
TRANSPORTATION
% SOLID WASTE
% MISCELLANEOUS (slash
burning, solvent evaporation loss I
Adopted from u S Environmental Protection
Agency, 1972 Notional Emissions Report
REGION 131 REGION 132 REGION 133
Total: 42,807 Tons Total: 34,855 Tons Totol: 42,722 Tons
Figure 26. Source category contributions to particulate air pollutants,
8-24
-------
TOTAL SUSPENDED PARTICIPATES (TSP)
IONS OF PEOPLE AFFECTED
0123 4567
ILLINOIS
INDIANA
MICHIGAN
MINNESOTA
OHIO
WISCONSIN
8 9 10
Figure 27. Air quality status (TSP) and trends in 25 largest urban areas in
EPA Region 5. Pie chart depicts the percent of population exposed to levels
greater than NAAQS for TSP in Region 5. Bar chart is estimated number of
people exposed to these exceedances on a state-by-state basis.
8-25
-------
100
100
Figure 28. Isopleths of TSP concentrations (yg/m3)
in EPA Region V and Iowa for October 15, 1976.
8-26
-------
K///VJ Insufficient data (<75% of maximum possible observa-
tions)
''I No evidence primary standard exceeded
P-PCO^ Primary standard exceeded
Alert standard exceeded
Figure 29. Air quality status, Colorado, 1972.
8-27
-------
E
cn
c
o
01
u
o
o
o s-
01
03 >
C O
O Z
O)
(1)
o
CO
0)
3
8-28
-------
8.3 CLASSIFICATION OF DATA
One of the early goals for standardizing the codes to be applied national-
ly to air quality monitoring sites was selectivity for analysis. The SAROAD
codes provide time classification, general geographic classification, and site/
neighborhood classification. However, only the first two are currently avail-
able at NADB for user applications. A profile of environmental quality on any
geographic basis will require more unique ways of data classification and
analysis in the future.
8.3.1 Time Categories
Overall, the NAAQS's dictate the following categories as most important:
year (TSP, S02, N02), quarter (Pb), day (TSP, S02), 8 hours (CO), 3 hours (S02),
and 1 hour (03, CO). However, trends analysis techniques could require classi-
fication by hour of the day; day of the week; week; and month for multiple year
comparisons. Autocorrelation techniques may require other nonstandard time in-
tervals and hence a more flexible data base retrieval system. Research needs
will require further analysis of long- and short-range transport, and time se-
lection will be critical for trajectory analysis.
8.3.2 Geographic Categories
Summary statistics are usually retrieved from NADB on a monitoring or
site point-source basis. All individual sites may be retrieved for a city,
county, Air Quality Control Region (AQCR), and State for a given time span.
Aggregation statistics are not yet standard selection options, except in
special computer routines not available to all users. Nationwide reports
either summarize individual site statistics for all States or select certain
geographic regions, urban areas, or special interstate areas. Likewise, re-
gional and state reports aggregate individual site statistics from all sites
or from selected counties or cities. Multicounty, county, and city agencies
have more opportunity to deal with specific subcounty areas in annual quali-
ty reports.
8.3.3 Site/Neighborhood Categories
This largely unexplored categorization has the potential to be a most in-
teresting summary in future regional and national reports. It is now a part
of the National and Regional Environmental Profile (NREP) effort. The new
8-29
-------
monitoring regulations require that sites be classified according to neighbor-
hood and monitoring objective. Site types published previously were industrial,
commercial, residential, and agricultural within center city, suburban, near
urban, rural, and remote areas. The new classifications are microscale, middle
scale, neighborhood scale, and urban scale. Once established, NAMS will be
categorized and summarized in any of the classification schemes.
8.4 INPUT PARAMETERS, DATA TRANSFORMATIONS, AND STATISTICAL COMPARISONS
Table 14 outlines the details of data types which could be summarized.
Note that effort was made to delineate individual air quality and emissions
statistics from aggregate statistics and counting summaries. Counting sum-
maries are generally used internally as management tools. Only a few such
summaries should be related in RNEP's or in reports to the Administrator. Much
work remains on standardizing data transformation techniques. In addition,
the techniques of maximum-likelihood estimation, application of the general
linear model (GLM) regression and analyses of variance, and spectral (time-
series) analysis could be applied in more cases than they are presently.
8.5 AUDIENCE APPLICABILITY
The most important consideration in data presentation is applicability
to a specific audience. For example, the report to the Administrator may
cover not only air quality assessment but also the end results of EPA pro-
grams and regulatory policies. The data needs are national in scope for the
environmental assessment, perhaps with highlighted urban areas on criteria
and noncriteria pollutant issues. Both air quality and emissions data, cur-
rent status and trends, would be given on a national and possibly a region
basis; this level would require more graphic presentation than a technical/
management reportfor example, a graphical summarization of data resulting
from singular national events (volcanic eruptions or widespread duststorms).
A report to the Regional Administrator may be merely a subset of the re-'
port to the national Administrator, but with emphasis on regional policies.
The summary statistics are usually given state by state, by selected major ur-
ban areas and by regions.
Reports to other government agencies and in response to.congressional and
(to a large extent) public inquiries are usually limited to specific areas
throughout a State. Environmental profiles provide an easily understandable
8-30
-------
TABLE 14. OUTLINE OF INPUT PARAMETERS, DATA TRANSFORMATIONS, AND
STATISTICAL COMPARISONS
Individual ("raw") air quality data
All hourly concentration values
All daily concentration or "index" values
Maximum/minimum site values
Average/median site values
Quantiles by site
Individual emissions data
Process
Stack
Plant
Aggregate air quality data for given category
Maximum/minimum values
Average or median of maximum/minimum data values
Weighted average/median of averages/medians
Maximum/minimum quantiles (e.g., max all 90th percentiles)
Average/median quantile of like quantiles
Aggregate emissions data for given category
Maximum/minimum process
Maximum/minimum stack
Maximum/minimum plant
Average/median process
Average/median stack
Average/median plant
Counting summaries for given statistic
Number of sites, cities, counties, States, regions
Number of processes, stacks, plants, cities, etc.
Data transformations
Distributional (e.g., Weibull, logarithmic)
Regression
Analysis of variance (ANOVA)
Spectral analysis and smoothing techniques
Example statistical comparisons by site, geographic, and time categories
Total number of pollutant sites
Number of sites exceeding annual standard
Number of sites exceeding short-term standards
Counties with data
Counties exceeding standards
Days exceeding primary standard
PSI class distribution
Site averages by year
City averages by year
Counties (urban areas, etc.) with significant pollutant trends up,
down, no change
Emission estimates by major source category
Emission density vs. population and land use
8-31
-------
graphic summary of environmental status and trends, and solve the problem of
repetitive requests.
Air quality standards "violations" and trends on a pollutant/county or
urban area basis (using colored maps) are the main tools of the RNEP effort.
The most common questionWhere are the least polluted areas of the United
States?--has not been addressed directly in EPA reports because of the com-
plexity of issues regarding sufficiency, representativeness, and comparability
of data. However, profiles with an added impetus to incorporate the PSI will
at least indirectly address this issue.
Scientific and technical audiences may have stringent requirements for
level of data summary and documentation of the statistical analysis techniques.
8.6 CAVEATS AND SUGGESTIONS
The following should be considered in planning graphics displays:
1. Descriptive or statistical parameters required?
2. Time to develop a summary from individual data?
3. Cost for preparation of graphics?
4. Data quality assured, valid, and carefully analyzed?
5. Proper emphasis on purpose of chart?
6. Chart not too light, heavy, confusing, complex?
7. Color/shading required?
8. Lettering styling and size appropriate?
9. Scale heights distorted?
10. Data gaps in chart treated correctly?
11. Too much gridding and/or lettering?
12. Grid units appropriate?
13. Grid points clearly identified?
14. Use of pictorial symbols and descriptive titles?
15. Chart self-explanatory?
8-32
-------
8.7 AVAILABLE PLOTTING RESOURCES
Graphic display packages are available at both the USEPA National Com-
2
puter Center (NCC), Research Triangle Park, N.C., and the Washington Computer
Center (WCC) Washington, D.C. "Stand-alone" systems include Integrated Plotting
Package (IFF), Harvard Graphic^ Tektronix, Calaomp, and Statistical Analysis Sys-
tem (SAS). Subroutines are available for user program interface under most of
these systems. Currently, statistical analysis and graphics programs developed
in some EPA Regional Offices are being documented for use by other regions;
these include PSI, box-plot, and pollution/wind roses.
8.8 GUIDANCE FOR SELECTION OF CHARTS
No single standardized list of charts should be mandatory for displaying
aerometric data; however, the following steps should be used to guide the selec-
tion of the most relevant charts.
1. Will the audience be interested in technical details and require
follow-up documentation concerning both data base and data analysis
methodology? If so, individual data or tabular summary may be suf-
ficient. If graphics are needed, see steps 2-12 below.
2. Select the geographic scale of the summary: national, regional,
AQCR, SMSA, county, urban area, city, township, or monitoring
site/source.
3. Select concept to be charted: trend, current status, parameter(s)
vs. parameters),, one parameter vs. several categories, composi-
tion of components, or maps.
4. Select time class: hour of day; day of week; month; calendar
quarter; season; year; or multiple year.
5. Select period of analysis: start year/month/day/hour and end year/
month/day/hour.
6. Select statistical summaries: individual, summation, or aggregation.
7. Select statistical analysis technique: descriptive statistics; dis-
tributional; regression; ANOVA; spectral analysis; or trend analysis.
8. Select site type, if desired: industrial, commercial, residential,
and so forth.
9. How can the analysis be accomplished? Manually (small data base),
hand calculator/computer, or large computer.
10. If a computer is necessary, are there statistical analysis and/or
graphics systems available? Check NCC, WCC, and Regional Office.
8-33
-------
11. Retrieve data according to selection criteria above.
12. In preparation of the chart, follow Section 8.7 caveats.
8.9 SUMMARY AND RECOMMENDATIONS
This section discussed established data display techniques applicable to
aerometric measurements, and provided guidance for determining the formats most
suitable for the concepts to be displayed, the audience applicability, the in-
put data classification and analysis, and the plotting resources. The discus-
sion was limited to data resulting from the air monitoring and emissions inven-
tory processes. The purpose of such data is to provide quantitative insights
to help abate air pollution, manage natural resources, plan environmental
programs, and inform the public.
The discussion did not consider topics such as sufficiency, representa-
tiveness, and data validity, and it assumed data to be efficiently accessible.
Since new statistical analysis techniques are covered in preceding sections,
this section discusses only analysis by currently available techniques.
Finally, the section did not attempt to standardize graphic formats; it
merely provided guidance and criteria, and gave example displays of suitable
data presentations. Graphics must accurately portray the data, help readers
understand the data, and help captivate interest. The guidance and criteria
in this section are not meant to stifle the innovation and creativity of
those preparing these artforms.
8.10 FUTURE ISSUES
Section 8 will be modified as new statistical techniques are used to
yield different parametric relationships and thus different graphics. Manda-
tory standardized analysis techniques may result in standardized computer
graphics programs. Questions to be addressed in the future are:
1. How stringent will USEPA be on data completeness and "representative-
ness"? Should we display quality assurance data? If so, should we
exclude source-impacted monitoring sites, or restrict analysis to
NAMS?
2. How can special "success stories," nonattainment status, or other
policy/regulatory issues be displayed in environmental assessment
documents? Should they be included?
3. will population exposure analysis be used in future reports to the
Administrator? Will population exposure software be easily used by
Regional Offices?
8-34
-------
4. How frequently will the RNEP's be published? How do the profiles
relate to the report to the Administrator, to Congress, and to the
National Air Quality and Emissions Trends Report! Should the pro-
file format change from year to year and perhaps cover selected
urban areas or regions in "off" years?
8.11 REFERENCES
1. Spear, Mary E. Practical Charting Techniques. McGraw-Hill, Inc., New
York, N.Y., 1969.
2. NCC User Reference Manual.
8-35
-------
9.0 CONTINUITY OF YEAR-TO-YEAR REPORTS
This section focuses on two major topics which can seriously affect the
continuity of data reported throughout time. First are changes in the opera-
tional definitions for measurements of pollutants and in the statistical in-
dicators or techniques. Second are shifts in the data base.
The discussion assumes that detailed data, as opposed to summary values,
will continue to be available from NAMS. It also assumes that as data needs
change with time, published reports will reflect the new priorities rather
than maintain format merely for historical continuity.
9.1 METHOD CHANGES
Data analysts preparing air quality reports should continually check
for changes in measurement techniques. These changes may require the adjust-
ment of the raw data for past years to maintain consistency in the data base.
The data base should not prevent the adoption of a new statistical indi-
cator (for example, a change from the median to the mean as an indicator of
the central tendency of a distribution). Availability of detailed data for
past years should make it possible to compute values of the new indicator for
data of the past. There should be no reason for presenting trend tables with
old indicator values for past years and new indicator values for the present
year. Similarly, introduction of a different statistical method should be
accompanied by a presentation of its applicability to data of the past.
9.2 NAMS NETWORK CHANGES
The discussion above assumes that the number and location of monitors
providing data is held constant over the years for which the trend is studied.
However, the NAMS network will probably undergo many changes in years to come
as more is learned about the nature and effects of the various pollutants, as
funding levels change, and as source configurations are altered. If summari-
zation of data is carried out on the urban level (as recommended in Section 4
instead of by monitoring site, effects of changes in the network will be
9-1
-------
minimized. If an index of air quality is adopted, it is recommended that the
subset of NAMS used for index determination remain unchanged over a relatively
long period.
There are a number of ways in which changes in the NAMS network could oc-
cur; each would have a distinct effect on the continuity across years. In par-
ticular, there are three changes that could occur even if the subset of NAMS
used for index determination remains fixed.
1. The stations could be resited within an urban area,
2. The network in an urban area could be expanded, or
3. The network in an urban area could be contracted.
Effects of these changes can be minimized by the procedures given below.
A station should be resited only if there is hard evidence to justify it.
If resiting occurs, data from both the old and new sites should be gathered con-
currently for one year to firmly establish the differences between the two sites
and to assist in correcting any misleading conclusions formed using data from
the old site.
A change in location for a station used to report the maximum concentration
for an area should not affect the continuity of the data base; in fact, the maxi-
mum monitoring site should be reviewed periodically to assure a reasonable ap-
proximation of the maximum concentration associated with the urban area.
It is difficult to foresee plausible reasons for resiting those stations
located according to the criteria of high population density and poor air quali-
ty, unless there is a considerable change in air quality patterns over the urban
area. In this case, the resulting "discontinuity" of the data base would re-
flect a discontinuity in the measured phenomenon and should be noted.
If the NAMS network in an urban area is expanded, past values of indices
and trend analyses can be adjusted by relating concurrent values from the old
network stations to those from the new combined network.
If the NAMS network in an urban area is contracted, all indices and trend
computations for past years should be redone using only the data from the small-
er network.
9-2
-------
TECHNICAL REPORT DATA
(Please read Instructions on the reverse before completing)
1. REPORT NO.
EPA-450-4/81-015
3. RECIPIENT'S ACCESSION-NO.
February 1981
4. TITLE AND SUBTITLE
5. REPORT DATE
U.S. Environmental Protection Agency
Intra-Agency Task Force Report on Air Quality Indicators
6. PERFORMING ORGANIZATION CODE
7 AUTHOR(S)
4.F. Hunt, Jr., (Chairman), G. Akland, W. Cox, T. Curran
N. Frank, S. Goranson, P. Ross. H. Sauls and J. Suggs
8. PERFORMING ORGANIZATION REPORT NO
9. PERFORMING ORGANIZATION NAME AND ADDRESS
U.S. Environmental Protection Agency
)ffices of: Air, Noise and Radiation, Research and
Development and Planning and Management
EPA Region 5
10. PROGRAM ELEMENT NO.
11. CONTRACT/GRANT NO.
12. SPONSORING AGENCY NAME AND ADDRESS
J.S. Environmental Protection Agency
Office of Air, Noise and Radiation
Office of Air Quality Planning and Standards
Research Triangle Park, N.C. 27711
13. TYPE OF REPORT AND PERIOD COVERED
Fnrrp P
14. SPONSORING AGENCY CODE
15. SUPPLEMENTARY NOTES
16. ABSTRACT
, The Intra-Agency Task Force on Air Quality Indicators was established to recommend
standardized air quality indicators and statistical methodologies for presenting air
Duality status and trends in national publications. This report summarizes the
^commendations of the Task Force grouped into four categories: data base, data analysi
Jata interpretation and data presentation. The report includes the position papers
repared by the Task Force members dealing with precision and accuracy data, detecting
nd removing outliers, area of coverage and representativeness, data completeness and
nstorical continuity, statistical indicators and trend techniques, inference and
inclusion, data presentation, and continuity of year-to-year reports.
Chic
7.
KEY WORDS AND DOCUMENT ANALYSIS
DESCRIPTORS
b.IDENTIFIERS/OPEN ENDED TERMS
c. COSATl Field/Group
ntra-Agency Task Force on Air Quality Indie
ir Quality Indicators
tatistical Methodologies
iata Base
iata Analysis
iata Interpretation
ata Presentation
recision and Arr.urarv Oat.a
a tors
Outliers
Representativeness
Data Completeness
Statistical Indicators
Trend Techniques
Continuity of y
;ar-to-year reports
8. DISTRIBUTION STATEMENT
Release Unlimited
19. SECURITY CLASS (This Report)
Unclassified
21. NO. OF PAGES
20. SECURITY CLASS (This page)
Unclassified
22. PRICE
EPA Form 2220-1 (9-73)
------- |