EPA/600/4-85/046
July 1985
Application of the
Microenvironment Monitoring
Approach to Assess Human
Exposure to Carbon Monoxide
Naihua Duan
Rand Corporation
Santa Monica, CA 90405
EPA Contract No.
63-02-4058
EPA Project Officer
Harold Sauls
ENVIRONMENTAL MONITORING SYSTEMS LABORATORY
OFFICE OF RESEARCH AND DEVELOPMENT
U.S. ENVIRONMENTAL PROTECTION AGENCY
RESEARCH TRIANGLE PARK, NC 27711

-------
NOTICE
This document has been reviewed in accordance with
U.S. Environmental Protection Agency policy and
approved for publication. Mention of trade names
or commercial products does not constitute endorse-
ment or recommendation for use.

-------
- iii
PREFACE
This study was conducted to apply the microenvironment monitoring
approach to assess human exposure to carbon monoxide, using data from
the Washington Urban Scale Study carried out in the winter of 1982-1983
by Research Triangle Institute, and the CO Microenvironment Study
conducted in the winter of'1983 by Battelle Columbus Laboratories, under
the auspices of the U.S. Environmental Protection Agency. Additional
technical background on the modeling of human exposure to air pollution
is available in N. Duan, Models for Human Exposure to Air Pollution, The
Rand Corporation, N-1884-HHS/RC, 1982.
This report should be of interest to researchers who want to model
human exposure to air pollution in terms of activity patterns.
The analysis reported herein was performed pursuant to Contract No.
68-02-4058 with the U.S. Environmental Protection Agency.

-------
-iv -
SUMMARY
This study applies the microenvironment monitoring (MEM) approach
to estimate human exposure to carbon monoxide (CO), with activity time
data from the Washington Urban Scale Study and CO concentration data
from the CO Microenvironment Study. The estimated MEM exposures are
then compared with estimated exposures based on the personal monitoring
(PM) approach (the PM exposures).
For the specific data being used in this study, the MEM exposures
are about 40 percent higher than the PM exposures. However, despite
that discrepancy, the MEM exposure is found to be a powerful predictor
for the PM exposure. On the log scale, the MEM exposure has the correct
span relative to the PM exposure; the discrepancy between the two sets
of estimates is found to be a constant drift.
Two major factors could explain the discrepancy between the MEM and
the PM exposures. First, the CO Microenvironment Study might have
oversampled microenvironments with high CO concentrations. For example,
the commuting routes in the sample were selected to be the "ones
considered to be heavily travelled and predicted to have high CO
exposures during rush hour periods" (Mack et al., 1984). Second,
Wallace, Thomas, and Mage (1984) found a discrepancy between the
carboxyhemoglobin (COHb) levels estimated from breath measurements and
those estimated from the PM exposures, indicating that the PM exposures
might underestimate the actual exposure. One of their possible
explanations for the underestimation is a "consistent decline in the
readings" on the CO monitor "as the battery approached the end of its
charge."
Given the exploratory nature of the data used in this study, the
results reported here should be considered as an illustration and should
be generalized only with caution to future exposure studies. Based on
the experience from this study, in future exposure studies, the other
version of the microenvironment type (MET) approach, enhanced personal
monitoring (EPM), should be given a higher priority when feasible. When
the MEM approach is the only feasible one, the sampling of

-------
microenvironments should be carried out with either the weighted
sampling scheme or the simulated human activity scheme.
In the estimation of exposures, it is assumed that there is already
a scheme to classify microenvironments into METs. Part of this study
also examined how to choose an appropriate classification scheme. In
particular, Duan's (1981) criterion was applied to the microenvironments
measured in the CO Microenvironment Study to evaluate various schemes
for grouping similar microenvironments into METs.

-------
- vi -
ACKNOWLEDGMENTS
The author would like to express his appreciation to the following
colleagues for their support during the course of this study: T. D.
Hartwell from Research Triangle Institute and Gregory Mack from Battelle
Columbus Laboratories for their help in clarifying the data from the
Washington Urban Scale Study and the CO Microenvironment Study; Allan
Abrahamse and Adele Palmer from The Rand Corporation, for their
thoughtful review of an earlier draft of this report, which resulted in
a substantial improvement in its organization; Wayne Ott and Lance
Wallace from U.S. Environmental Protection Agency, for their thoughtful
comments on an earlier draft; David Holland from the U.S. Environmental
Protection Agency, for valuable help in the delivery of the data used in
this study, and thoughtful discussions and suggestions throughout the
course of this study; and Harold Sauls from the U.S. Environmental
Protection Agency, the Project Officer for this study, for valuable
guidance on the conduct of this study.

-------
-yvii -
CONTENTS
PREFACE 		iii
SUMMARY 	;		!	i V
ACKNOWLEDGMENTS 	;	Vi
TABLES 	i	ix
GLOSSARY 		X
Section
.1. INTRODUCTION 	'		1
Exposure Assessment 		2
Review of the MET Approach 		5
Methods for Estimating Exposure 		11
II. ACTIVITY TIME DATA 		14
Washington Urban Scale Study 		14
METs and Activity Segments 		15
Quality of Activity Time Data 			18
Startup Time, Total Time, and Summary Statistics 		19
III. CO CONCENTRATION DATA 		23
Microenvironment Study 		23
Summary Statistics for the MET Concentrations 		24
IV. ESTIMATED EXPOSURES 		28
Application of the Convolution Method 				28
Estimated Exposures Based on MEM 		31
Estimated Exposures Based on PM 		32
Comparison of Exposure Distributions 		34
Comparison of Individual-specific Exposures 		36
Conclusion 		41
V. EVALUATION OF MET CLASSIFICATION SCHEMES 		42
Elementary METs 		43
Duan's Criterion 		44
Empirical Results 		48
VI. FUTURE APPLICATIONS OF THE MET APPROACH 		50
Enhanced Personal Monitoring and Microenvironment
Monitoring 		50
Sampling of Microenvironments 		51
Study Designs and Data Collection 		53
Further Use of the Activity Database 		54

-------
--vi ij -
Appendix
A.	DEFINITIONS OF NETs 		57
B.	QUALITY CRITERION FOR THE ACTIVITY TIME DATA 			61
C.	FREQUENCY DISTRIBUTIONS OF ACTIVITY TIME 		64
D.	SENSITIVITY ANALYSIS 		73
E.	DUAN'S CRITERION 			'		82
F.	ESTIMATION OF PARAMETERS IN DUAN'S CRITERION 		84
REFERENCES 		87

-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
16
17
19
20
22
22
25
26
29
31
33
33
35
35
37
39
48
TABLES
Activity Time for Modes of Travel and Types of Shop ....
Average Concentration for Modes of Travel and
Types of Shop 	
Cross Tabulation for the Two Criteria 	
Frequency Distribution of Startup Time 	
Frequency Distribution of Total Time 	
Summary Statistics for Standardized Activity Times 	
Summary Statistics for CO MET Concentrations Based on MEM
Summary Statistics for CO MET Concentrations Based on PM
Correlation Between MET Concentration and MET Time 	
Summary of Imputed MET Concentrations 	
Summaries for MEM and PM Exposures 	
Percentiles of MEM and PM Exposures 	
Summaries of log MEM and Exposures 	
Percentiles of log MEM and PM Exposures 	
MSE and GAIN for MEM Exposures 	
Regressions of PM Exposures on MEM Exposures 	
Ranking of MET Decompositions 	

-------
- X
c.d. f
CO
COHb -
MEM
MEM-C
MEM-H
MET
MSE
NAAQS
NEM
PEM
PM
ppm
SHAPE
GLOSSARY
cumulative distribution function
carbon monoxide
carboxyhemoglobin
microenvironment monitoring
the convolution method applied to microenvironment monitoring data
the hybrid method applied to microenvironment monitoring data
microenvironment type
mean squared error
national ambient air quality standards
NAAQS exposure model
personal exposure monitor
personal monitoring
parts per million
simulation of human air pollution exposures

-------
- 1 -
I. INTRODUCTION
This study applies the microenvironment monitoring (MEM) approach,
one of the two versions of the microenvironment types (MET) approach
(also called the indirect approach), to estimate human exposure to
carbon monoxide (CO), using activity time data from the Washington Urban
Scale Study and CO concentration data from the CO Microenvironment
Study. The estimated exposures based on the MEM approach (the MEM
exposures) are then compared with estimated exposures based on the
personal monitoring approach, also called the direct approach (the PM
exposures).
For the specific data used in this study, the MEM exposures are
about 40 percent higher than the PM exposures. However, despite the
discrepancy, the MEM exposure is found to be a powerful predictor for
the PM exposure. On the log scale, the MEM exposure has the correct
span relative to the PM exposure; the discrepancy between the two sets
of exposure estimates is found to be a constant drift. Further details
on these results are discussed in Sees. II-IV.
Two major factors could explain the discrepancy between the MEM and
the PM exposures. First, the CO Microenvironment Study might have
oversampled microenvironments with high CO concentrations. For example,
the commuting routes in the sample were selected to be the "ones
considered to be heavily travelled and predicted to have high CO
exposures during rush hour periods." (Mack et al., 1984). Second,
Wallace, Thomas, and Mage (1984) found a discrepancy between the COHb
levels estimated from breath measurements and those estimated from PM
exposures, indicating that the latter might underestimate the actual
exposure. One of their possible explanations for the underestimation is
a "consistent decline in the readings" on the CO monitor "as the battery
approached the end of its charge."
Given the exploratory nature of the data used in this study, the
results reported here should be considered as an illustration and be
generalized only with caution to future exposure studies. Based on the
experience from this study, in future exposure studies the enhanced

-------
- 2 -
personal monitoring (EPM) approach should be given a higher priority
when feasible. When the MEM approach is the only feasible approach, the
sampling of microenvironments should be carried out with either the
weighted sampling scheme or the simulated human activity scheme.
In the estimation of exposures, it is assumed that a scheme to
classify microenvironments into METs has already been chosen. Part of
this study also examined how to choose an appropriate classification
scheme. In particular, Duan's (1981) criterion has been applied to the
microenvironments measured in the CO Microenvironment Study to evaluate
various schemes to group similar microenvironments into microenvironment
types (METs).
EXPOSURE ASSESSMENT
Until recently, human exposure to air pollution could be assessed
only with fixed-site ambient monitoring data. Typically people residing
in the same neighborhood near a monitoring station were treated as
homogeneous receptors fixed at the location of the monitoring station.
Recent field studies with personal exposure monitors (PEMs) have found
this approach inadequate for such pollutants as carbon monoxide, which
are spatially variable or have nonambient sources or sinks. For such
pollutants it is important to take into account people's mobility and
activities in assessing their exposures.
With the recent development of PEMs, it became feasible to
incorporate the mobility and activities into exposure assessment,
especially for CO, for which several reliable, continuous PEMs are
available. There are two general approaches to exposure assessment
using PEMs. The first is the personal monitoring (PM) approach, also
called the direct approach, in which human subjects are sampled from the
target population and are equipped with PEMs for a certain time to
measure their exposures directly. This approach is taken in the
Washington Urban Scale Study, the details of which are discussed in Sec.
II. The advantages of this approach are its simplicity and its freedom
from modeling assumptions. The main disadvantage is its cost, too high
for large scale investigations.

-------
- 3 -
An alternative approach to assess exposure is the MET approach in
which pollutant concentration data are combined with or enhanced by
activity time data.1 There are two ways in which the MET approach can
be implemented: the enhanced personal monitoring (EPM) approach, and
the microenvironment monitoring approach. The latter approach was taken
in the CO Microenvironment Study, the details of which are discussed in
Sec. III.
Assessment of Individual Exposures
In many situations it is important that the estimated exposure be
attributable to the exposed individuals. For example, in
epidemiological studies to quantify the health effects of air pollution,
one needs to know who is exposed to what levels of pollution. In that
situation the estimated exposures must be close to the actual exposures.
In this study we don't know the actual exposures; both the MEM
exposures and the PM exposures are estimates and are subject to error.
Nevertheless it is still preferable that the two estimated exposures be
close to each other; if the two estimated exposures are substantially
different from each other, at least one of them must be substantially
different from the unobserved actual exposure.
The mean squared error (MSE) will be used as the criterion to
measure how close the two estimated exposures are to each other. The
criterion is defined as follows:
MSE = I. (MEM. - PM.)2/n ,	(1)
111
where MEM^ is the MEM exposure for the ith unit (say, person-day), PM^
is the PM exposure for the ith unit, n is the total number of units.
The square root of MSE can be interpreted as the average magnitude
of the difference between the two estimated exposures.
In certain other situations it is important only to distinguish
people exposed to higher levels of air pollution from people exposed to
lower levels. For example, one might be interested in qualifying the
health effects of air pollution--i.e., in determining qualitatively the
*Duan (19S1); Duan (1982); Ott (1981)

-------
- 4 -
existence of health effects. In such situations it is important only to
rank people's exposures, it is not necessary to know the exact levels of
exposures. In such situations it is necessary only that the estimated
exposure be a significant predictor for actual exposure. Therefore we
can use the strength of the regression of the actual exposure on the
estimated exposure as the criterion for evaluating the estimated
exposure. The estimated exposure might be substantially biased from the
actual exposure, but so long as the two measures are well correlated,
the estimated exposure can be used as a proxy for actual exposure in
qualifying health effects.
In this study the actual exposures are unknown and therefore cannot
be used to estimate the regression relationship between the estimated
exposures and the unobserved actual exposures. Instead the PM exposure
is used as the benchmark; the PM exposure is regressed on the MEM
exposure2 to test for the relationship between the two estimated
exposures. If the MEM exposure is not a good predictor for the PM
exposure, it is unlikely that it will be a good predictor for the
unobserved actual exposure.
Assessment of Exposure Distributions
For certain other purposes, such as risk assessment, it is
necessary only to estimate the distribution of exposures, but it is not
necessary to identify each individual's exposures, although it is useful
to have such exposures available for reference.
To take an extreme case, consider a population with two
individuals, one exposed to 10 ppm of CO, another exposed to 1 ppm. If
the dose-response relationship has already been established from other
sources, the risk assessment can be done without knowing each
individual's exposure. For example, assume that it is already known
that exposure to levels higher than 9 ppm is harmful. The goal for risk
assessment is to estimate what fraction of the target population is
exposed to harmful levels. In this case the only information relevant
to the risk assessment is that half of the hypothetical population is
2The PM exposure is taken in this analysis as the standard method;
the MEM exposure is taken as the new method to be calibrated relative to
the standard method.

-------
- 5 -
exposed to 10 ppm and the other half is exposed to 1 ppm. It is not
necessary to distinguish which individual is exposed to which level.
(To take the case even further, suppose an exposure assessment method
wrongly estimated the first individual's exposure to be 1 ppm while his
actual exposure is 10 ppm, and wrongly estimated the second individual's
exposure to be 10 ppm while his actual exposure is 1 ppm, the result
should be judged satisfactory for risk assessment, because the exposure
distribution is correct.)
REVIEW OF THE MET APPROACH
The MET approach combines MET-specific pollutant concentration data
and activity time data to estimate exposures. This approach
incorporates information about people's mobility and activity. Two
different ways can be used to implement the MET approach, the EPM
approach and the MEM approach. Detailed discussion of these approaches
was given in Duan (1981, 1982).
Microenvironments and METs
A microenvironment is a chunk of air space with homogeneous
pollutant concentrations. The integrated exposure of a specific
individual may be represented as a weighted sum of concentrations in the
icroenvironments in which he spent time during a given study period,
say a 24-hour period, weighted by the amount of time he spent in each
microenvironment:
E. = Z.Z. x x.. ,	(2)
i J J ij
where E^ is the ith person's exposure during the study period, the
microenvironment concentration 2f. is the pollutant concentration in the
J
jth microenvironment, the microenvironment time T.. is the time the ith
ij
person spent in the jth microenvironment. The summation is taken over
all microenvironments that some individual in the target population
might spend some time in.
If the microenvironment times t.. and microenvironment
ij
concentrations 2f have been measured, Eq. (2) may be used to calculate
exposures E^. However, in any realistic situation, there are far too
m

-------
- 6 -
many microenvironments to be measured extensively. To make good use of
data collection resources, microenvironments of a similar nature must be
grouped in terms of pollutant concentrations. For example, all indoor
microenvironments might be grouped together, and all outdoor
microenvironments might be another group. Such groups of similar
microenvironments will be referred to as microenvironment types (METs).
In terms of the METs, integrated exposures may be reformulated as
follows:
E. = I.C.. x T ,	(3)
1 k ik lk
where the MET time T., is the total amount of time the individual spent
ik
in microenvironments belonging to the kth MET, and the MET concentration
is the average concentration the individual is exposed to during the
time he spent in microenvironments belonging to the kth MET. The MET
time is related to the microenvironment times as follows:
T. = 1.6 x t . . ,	(4)
ik j jk ij
6^ = 1 if jth microenvironment belongs to kth MET,
0 otherwise.
The MET concentration C., is related to the microenvironment
ik
concentrations as a weighted average of microenvironment concentrations
2f in microenvironments belonging to the kth MET, weighted by the
fraction of time spent in each microenvironment belonging to this MET:
C., = 1.6 x w. x y. ,	(5)
ik j jk ijk j
where w.., = t../T., .
ijk ij ik
Equation (3) will be referred to as the time-weighted summation formula.
The discussion has focused on integrated exposure. Maximum one-
hour and eight-hour exposure may be dealt with similarly. For each one-
hour or eight-hour period, the time-weighted summation formula may be
used to estimate the integrated exposure for the time period. The

-------
- 7 -
maximum may then be taken over all one-hour or eight-hour periods of
interest.
Implementation of the MET Approach
The the main advantage of the MET approach is that concentration
data and activity data are combined from two different sources. In
certain situations one of the two databases might already be available
or is inexpensive to obtain. In certain situations both databases might
already be available. For example, the sources of human exposures to
nitrogen dioxide and CO are fairly similar. Both are mainly determined
by combustion sources. Therefore the same set of METs can be used for
both pollutants, with little or no modifications. The activity data
collected in the CO studies can therefore be used as the activity data
for a future study of human exposure to nitrogen dioxide. If another
database becomes available for MET concentrations for nitrogen dioxide
or can be collected inexpensively, it will be feasible to combine the
existing activity database with the nitrogen dioxide concentration
database to assess human exposure to nitrogen dioxide without the need
to collect additional activity data.
There are two approaches to implement the MET approach, depending
on how the concentration data are collected. In the first approach, the
MET concentrations are obtained from personal monitoring. In the second
approach, the MET concentrations are obtained from microenvironment
monitoring. With either approach, the activity data have to be
collected from human subjects, using either a diary or a recall survey.
Enhanced Personal Monitoring
The personal monitoring approach collects pollution concentration .
data on a.sample of human subjects and derives each individual's
exposure directly from the measured concentrations. If activity time
data are also collected, either on the same sample of human subjects or
on a different sample of human subjects, the concentration data from
personal monitoring can be combined with activity time data to estimate
exposure using the time-weighted summation formula Eq. (3). This
approach is called the enhanced personal monitoring (EPM) approach.3
3Duan (1982) referred to this as the continuous personal monitoring
approach.

-------
- 8 -
To use the personal monitoring data in the EPM approach, it is
necessary to have continuous measurements of pollutant concentrations'*
to derive the MET concentrations to be used in the time-weighted formula
Eq. (3). Therefore the EPM approach requires that reliable continuous
PEMs be available. Furthermore, for the PEMs to be used by untrained
human subjects, they have to be small, lightweight, and easy to operate.
For carbon monoxide, several reliable PEMs satisfy the above
requirements and EPM is a feasible way to implement the MET approach.
With the EPM approach, a sample of human subjects is equipped with PEMs.
Participants keep a diary of their activities. From the continuous
measurements and diaries, it is possible to determine at any instant the
MET the participant was in and the instantaneous concentration the
participant was exposed to. The average concentration the participant
was exposed to during the time he spent in that MET may then be derived.
That is the MET concentration needed for the time-weighted summation
formula.
The diaries the participants record during the monitoring phase
provide some activity data, which can be supplemented by additional
activity data. One inexpensive way to expand the activity database is
to collect diary data on the same participants on additional days. For
example, participants may be enrolled in the study for a week, record
their daily activities, and be equipped with PEMs for one or two days
during the study period. Additional participants may only fill out the
diaries but not participate in the monitoring. The marginal cost in
collecting the additional diaries should be low compared with collecting
the monitoring data. Furthermore, existing activity data can be
combined at little or no cost with the activity data collected on the
participants who are equipped with PEMs.
"With the "pure" personal monitoring approach one needs only
integrated measurements for the assessment of integrated exposures.

-------
- 9 -
Microenvironment Monitoring
An alternative way to implement the MET approach is
microenvironment monitoring.5 Instead of MET concentration data with
personal monitoring, a number of microenvironments may be sampled in
each MET, with research staff or trained technicians sent to the sampled
microenvironments to monitor those microenvironments directly.
Duan (1982) noted that the distribution of microenvironment
concentrations Y. can be different from the distribution of MET
J
concentrations	except for a special case in which for each MET,
each individual visits only one microenvironment of that type, and each
microenvironment belonging to that MET has the same probability of being
visited. This special case is unlikely to be satisfied, because people
might visit several microenvironments belonging to the same MET, or some
microenvironments will have higher probabilities of being visited. When
the special case described above is not satisfied, Duan (1982) showed
that the microenvironment concentrations are usually more variable than
the MET concentrations.
To use microenvironment concentrations to derive valid estimates of
MET concentrations, it is necessary to obtain a valid sample of
microenvironments. For the sample to be valid, the target population is
not the population of microenvironments belonging to a certain MET.
Instead, it is the coincidence (intersection) of microenvironments and
people. For example, consider a MET consisting of all office spaces.
Assume also that each identifiable room is sufficiently homogeneous and
can be regarded as a microenvironment. If the collection of all
microenvironments were the target population, a roster of all office
spaces could be the frame for the sampling, or an area probability
sample. However, some of these office spaces might be empty most of the
time, some might be crowded most of the time. To obtain a valid sample
of microenvironments, it is necessary to adjust for such differences.
One possibility discussed in Duan (1982) is to count or estimate the
number of people present at each sampled microenvironment and use these
counts as weights. The crowded microenvironments get larger weights in
5This approach was referred to as the replicated microenvironment
monitoring approach in Duan (1982).

-------
- 10 -
the analysis. Another possibility is to simulate real human activities
using available diary data.
Comparison of EPM and MEM Approaches
Compared with the MEM approach, the EPM approach has the advantage
of avoiding the difficult task of sampling microenvironments. In
essence, with the EPM approach, the participants sample the
microenvironments for the investigator.
The difficulty in the sampling of microenvironments is the main
disadvantage of the MEM approach. If all microenvironments belonging to
the same MET have exactly the same concentration, or have very similar
concentrations, this is not a problem. Any reasonable sample of
microenvironments will give a good result. However, concentrations are
more likely to vary substantially from microenvironment to
microenvironment, even though they belong to the same MET. In this
situation the appropriate sampling of microenvironments is important.
The EPM approach is very demanding on hardware. As of now, CO
appears to be the only pollutant for which the necessary continuous PEM
is available for this approach.
The main advantage of the MEM approach is that it does not require
miniature continuous PEMs. Because the monitoring is conducted by
research staff or trained technicians, it is possible to use portable
continuous monitors, which are inconvenient for the untrained human
subjects to use. (Of course it does not hurt to have PEMs available.)
Another disadvantage of the EPM approach is the need to sample and
monitor human subjects. This disadvantage is probably secondary,
because it is already necessary to sample human subjects to obtain the
diary, data. However, the inconvenience to the human subject in the
collection of the diary information is likely to be minor relative to
the collection of the monitoring data; with the latter the participant
needs to wear the monitor for an extended time.
The MEM approach has the advantage of not needing human subjects in
the monitoring. However, human subjects must still be used to collect
the activity data, unless there is a suitable activity database already
available.

-------
- 11 -
Another disadvantage of the MEM approach is that it might be
difficult to access certain microenvironments sampled for monitoring.
With the EPM approach it might be easier for the human subjects to
access such microenvironments.
METHODS FOR ESTIMATING EXPOSURE
The MET concentration data and the MET time data can be combined in
several ways to estimate exposure. If one is interested only in average
exposure, one can use the average-time weighted summation formula and
estimate average exposure by
E = IkCk x Tk,	(6)
where E is the average exposure, is the average MET concentration for
the kth MET, and T^ is the average MET time for the kth MET. This
method implicitly assumes that the MET concentrations and MET times are
uncorrelated, because otherwise the correct formula for average exposure
based on averaging both sides of the time-weighted summation formula (3)
over the individuals should contain an additional covariance term:
* = V5*" \ + Cov«v V-
The assumption that MET concentrations and MET times are uncorrelated is
not unusual. It is implicitly assumed in all models for human exposure
using the MET approach, including SHAPE (Ott, 1981), the convolution
method, and the hybrid method. The assumption basically rules out
people's response to air pollution such as staying away from high
concentration METs during polluted days.
For most purposes the mere estimation of average exposure is
inadequate, and it is necessary to estimate exposure distribution or
individual exposures. There are several ways of doing this. One
approach is to use simulation models such as SHAPE (Ott, 1981), in which
the concentration and activity data are summarized by parametric (or
nonparametric) probabilistic distributions, human activity and
concentration data are simulated from those probabilistic distributions,

-------
- 12 -
and the simulated data are used to estimate exposures. This type of
approach generally assumes that the concentration and time are
independent to facilitate the simulation.
Another approach is the convolution method proposed in Duan (1981).
With this approach, units (e.g., person-days) are paired from the
activity database with units (e.g., days) from the concentration
database to form convoluted units (e.g., person-days), and the exposure
for each convoluted unit is estimated using a time-weighted summation
formula similar to Eq. (3).
E. = I, C , x T ,	(7)
lm k mk lk
where E^ is the exposure combining the ith unit in the activity
database and the mth unit in the concentration database, C , is the MET
mk
concentration for the mth unit in the concentration database in the kth
MET, and T^ is the MET time for the ith unit in the activity database
in the kth MET.
To illustrate the application of Eq. (7), consider a study that has
43 days of MEM data, combined with a sample of 705 persons, each
providing one day of activity diary. If the ith person in the activity
sample spent the day according to {T^} and was exposed to concentrations
{Cm> in the METs encountered during that day, he will receive exposure
E. . As independence is assumed between the MET concentrations and
lm
times, each of the 43 concentration vectors	is equally likely for
each of the 705 participants. With the convolution method, the
exposures E are derived for each of the 43 x 705 = 30,315 pairings of
persons and days in the two databases. Each such pairing is one
convoluted person-day. If the number of convoluted person-days is too
large, we can sample from them.
This method also requires that the concentration and time be
independent. Duan (1982) showed that the distribution of exposures
estimated from the convolution method is an unbiased estimate of the
distribution of actual exposures and is a function of the empirical
c.d.f.s for the MET concentrations and the MET times. Because the
empirical c.d.f. is the efficient estimator of the true c.d.f. in the
sense of being the minimum variance unbiased estimator, the exposure

-------
- 13 -
distribution estimated using the convolution method is also efficient in
the same sense.
An alternative to the convolution approach was suggested by Harold
Sauls in a private communication. This method can be viewed as a
hybridization between the average-time weighted summation formula Eq.
(6) and the convolution method Eq. (7). With this hybrid method, the
average MET concentration in each MET is used to estimate the exposure
for each unit (day or person-day) from the activity database by
E. = I.C. x T .	(8)
i k k lk
This method ignores the variability in exposures due to the variability
in MET concentrations. If all microenvironments belonging to the same
MET have the same concentration, this method is preferable to the
convolution method because of its simplicity. If the microenvironments
belonging to the same MET vary substantially, this approach is likely to
underestimate the variability of the exposure distribution. The
estimated exposures based on this approach will be referred to as the
hybrid exposures.

-------
- 14 -
II. ACTIVITY TIME DATA
WASHINGTON URBAN SCALE STUDY
A population-based study on CO exposure was conducted during the
winter of 1982-1983 in the Washington, D.C. metropolitan area.1 Details
on this study are available in Hartwell et al. (1984a and b) and
Whitmore et al. (1984).
An area probability sample of human subjects was enrolled for one
day each in this study. The participants filled out activity diaries
giving the activities they were engaged in during each time period. The
activities are entered in the diaries as activity segments, where each
activity segment is defined to be the time period between two reported
changes in activities in the activity diary. The participants'
exposures to CO were measured using PEMs, which gave the average
concentration for each activity segment.
The participants in the Urban Scale Study were selected from a
probability sample. To extrapolate from the sample to the target
population, it is necessary to weight the individual observations by the
sampling weights based on sampling probabilities. In preliminary
analysis, the summary statistics based on the weighted and the
unweighted procedures were compared. The weighting did not have a major
effect on the results. For example, the average time spent in car
commuting differs by about 2 percent between the weighted and the
unweighted estimates. Because the primary goal of this study is to
compare the estimated exposures based on the the MEM and the PM
approaches, the extrapolation to the target population is not crucial.
To simplify the analysis it was decided not to weight the individual
observations.
*A similar study was conducted in the Denver metropolitan area at
about the same time. Details on this study are available in Johnson
(1984).

-------
- 15 -
METS AND ACTIVITY SEGMENTS
In the Washington Urban Scale Study each participant fills out
activity diaries for one day. During this sampling day, whenever there
is a new activity--e.g., the participant stops reading a newspaper in
the living room (end of an old activity) and goes outside for a walk
(beginning of a new activity)--the participant is required to record the
start time of the new activity and describe it. The period between two
entries in the activity diary is referred to as an activity segment.
Each activity segment is regarded as one microenvironment.
Based on information available, activity segments are grouped into
seven METs: packing, public transportation, private car, pedestrian,
shops, offices, and others. The rest of this section gives the heuristic
definitions of these METs. Detailed definitions are given in Appendix
A.
The MET parking is restricted to indoor parking because only indoor
parking concentration data are available from the CO Microenvironment
Study. The MET public transportation includes both bus and metrorail.
Because both buses and metrorails are monitored in the Microenvironment
Study, it is possible to consider them as distinct METs. However, in
the evaluation of MET classification schemes to be discussed in Sec. V,
it was found not to be beneficial to distinguish these two METs;
therefore public transportation is considered as one MET without further
refinement.
The MET private car includes private cars, trucks, motorcycles, and
vans. It is debatable whether this MET should be restricted to the
narrow definition including private cars only. (Only private cars are
monitored in the Microenvironment Study.) The four modes of travel were
grouped into one MET for two reasons. (1) The amount of time spent in
trucks, motorcycles, and vans is very small compared with the amount of
time spent in private cars. The top part of Table 1 gives the average
amount of time spent in each of these modes of travel. The total amount
of time spent in the four modes of travel is 1.623 hours per person per
day, out of which only 0.106 hours belong to the three modes other than
private car, less than 7 percent of the total. (2) The MET
concentrations based on PEM for those four modes of travel are roughly

-------
- 16 -
Table 1
ACTIVITY TIME FOR MODES OF TRAVEL
AND TYPES OF SHOP
Fraction
MET
Mode/Type
Average Time
of MET (%)
Car
Car
Truck
Motorcycle
Van
1.517 hr
0.069
0.002
0.035
93.47
4.25
0. 12
3. 16

Total
1.623
100.00
Pedestrian
Walking
Jogging
Biking
0.254
0.007
0.008
94.42
2.60
2.97

Total
0.269
100.00
Shops
Stores
Mall
0.369
0.015
96.09
3.91

Total .
0.384
100.00
similar.. The top part of Table 2 gives the average concentrations along
with the standard errors for the averages. The difference between car
and truck is small (about 1 ppm) and statistically insignificant (t =
0.73). The difference between car and van is not small (about 3 ppm)
and is statistically significant (t = 3.67). However, only seven people
reported using a van in their travel. The difference between car and
motorcycle is about 2 ppm; the standard error and t-statistics for this
difference are not available because only one person reported using a
motorcycle in his travel.
The MET pedestrian includes walking, biking, and jogging. It is
again debatable whether jogging and biking should be grouped with
walking into one MET. Table 1 shows that the amount of time spent
jogging and biking is very small (less than 6 percent) compared with
time spent walking. The difference in concentrations between walking
and jogging is very small (less than 0.1 ppm) and statistically
insignificant (t = 0.09). The difference between walking and biking is
about 2 ppm and is statistically significant (t = 2.09). However, only

-------
- 17 -
Table 2
AVERAGE CONCENTRATION FOR MODES OF TRAVEL
AND TYPES OF SHOP
Average
MET
Mode/Type
Na
Concentration
SEb
Car
Car
592
5 .1
0.22

Truck
22
6.3
1.67

Motorcycle
1
3.0


Van
7
2.1
0.79
Pedestrian
Walking
220
2.3
0. 16

Jogging
6
2.3
0. 78

Biking
5
4.0
0.82
Shops
Stores
225
2.2
0. 17

Malls
11
1.8
0.54
aThe number of participants who used this
mode/type during the sampling period.
^Standard error of the average concentration.
five people reported biking during the sampling period. Therefore they
are combined into one MET.
The MET shops consist of the activity segments reported as stores,
shopping malls, and theaters in malls. The amount of time spent in the
malls is small relative to the time spent in stores (less than 5
percent). The difference in concentration is very small (less than 0.5
ppm) and statistically insignificant (t = 0.65). Therefore they are
combined into one MET.
The MET offices consists of activity segments reported as offices.
The MET other is a residual category for activity segments not
considered above. The main component of activity segments in this MET
is home. Because there are no microenvironment monitoring data
corresponding to these activity segments in the Microenvironment Study,
this MET cannot be refined any further.

-------
- 18 -
QUALITY OF ACTIVITY TIME DATA
During the preliminary analysis of the activity time data from the
Urban Scale Study, the quality of the data was found not to be uniform.
Some participants took great care in providing a detailed and presumably
accurate diary of their activities. Some participants reported
questionable data. For example, some participants did not report
spending any time sleeping during the sampling period. The same
observation was made in Hartwell et al. (1984a and b).
To identify the participants who provided more reliable diary data,
two criteria were developed. The first is based on the participants'
success or failure in following the skip logic in the activity diary and
will be referred to as the skip-logic criterion. The second is based on
the consistency checks that were developed during the preparation of the
activity database (Hartwell et al., 1984a, Section 5.4.2.2) and will be
referred to as the consistency criterion. The details on these criteria
are given in Appendix B.
The two criteria are strongly associated, indicating they measure
some common trait--presumably the quality of the diary data. Table 3
gives the cross tabulation for the two criteria. Participants who pass
the consistency criterion are more likely to pass the skip-logic
criterion, and participants who fail the consistency criterion are also
more likely to fail the skip-logic criterion. The X2 statistics for
association in this 2x2 table is 14.027, with one degree of freedom.
The P-value is 0.0002, indicating the association is highly significant.
The two criteria are used to stratify the participants into
different levels of reliability. There are 705 participants in the
sample--the "whole sample." A subset of 361 participants passed the
skip-logic criterion--i.e., followed the skip logic correctly. This
pool will be called the "good sample." A subset of 127 participants
passed both the skip-logic criterion and the consistency criterion.
This pool is the "best sample."
Applying all analyses to be discussed in the rest of this section
and in later sections simultaneously on all three samples for comparison
serves as a sensitivity analysis to examine whether the lack of
uniformity in the quality of the activity data results in any difference

-------
- 19 -
Table 3
CROSS TABULATION FOR THE
TOO CRITERIA
Skip-logic
Fail Pass	Total
Fail 267 234 501
Cons istency
Pass 77	127 204
Total 344 361 705
in the conclusion. As it turns out, practically all of the analysis
results are insensitive to the sample chosen. Therefore the results are
given for the whole sample in the text, and for the "good" and the
"best" samples in Appendix B.
STARTUP TIME, TOTAL TIME, AND SUMMARY STATISTICS
Startup Activity
Most participants reported an activity segment labeled "begin
diary," indicated by the activity code 87 in Hartwell et al. (1984a and
b). Table 4 gives the frequency distribution of this activity. Some
participants took hours to get started in their diaries. We interpret
this as an oversight in the instruction. These participants failed to
realize that "begin diary" is an instantaneous event that is over as
soon as it is recorded, and a new activity should be reported
immediately. For example, the participant might be watching TV in the
evening when he started the diary, and continued watching TV following
the start of the diary. The correct diary should be an activity segment
with zero (or nearly zero) duration for "begin diary," followed by
another activity segment for watching TV, followed by the next activity
such as going to bed. Many of the participants reported "begin diary"
as a time-consuming activity--for example, the participant in the
earlier example might report the time he spent watching TV after the
start of the diary as part of the activity segment "begin diary," and
won't report another activity transition until he went to bed. The

-------
- 20 -
Table 4
FREQUENCY DISTRIBUTION OF STARTUP TINE
Dens itya
Range
Count
(Percent per minute)
0 (minute)
228
32.3
1
3
0.4
2
30
4.3
3
25
3.5
4
23
3.3
5
1'3
1.8
6
22
3.1
7
14
2.0 (median)
8-10
47
2.2
11 - 15
46
1.3
16 - 25
47
0.7
26 - 45
39
0.28
46 - 75
40
0. 19
76 - 120
29
0.09
121 - 180
38
0.09
181 - 240
30
0.07
241 - 300
22
0.05
301 - 360
4
0.009
361 - 420
4
0.009
420 - (b)
1

The density is defined as the percentage
((count/705)x100 percent) divided by the length
of the range (the number of minutes covered by this
range), and is given in the unit "percent per minute."
For example, 6.7 percent of the participants fall
inside the range 8-10 minutes, which covers three
minutes, therefore the density for this range is
6.7%/3 minutes = 2.2%/minute.
^The largest nine observations are (in decreasing
order): 496, 387, 382, 380, 363, 358, 323, 314, and
312, all in minutes.

-------
- 21 -
actual activity for this activity segment is missing. Therefore the
activity segments "begin diary" were deleted from all further analyses
and the subsequent activity segment treated as the real beginning of the
diary.
Total Time
The total amount of time each participant reported on during the
survey is not always exactly 24 hours. Table 5 gives the frequency
distribution of the total time reported in the diary. The total time is
skewed toward the lower end, because of the deletion of the startup
time. Some participants' total time is appreciably shorter than 24
hours.
To make the activity times comparable from participant to
participant, all activity times are standardized relative to 24 hours.
SAT = RAT * 24/Total time,
where SAT = standardized activity time, RAT = reported activity time,
total time = reported total time, all measured in hours. Unless noted
otherwise, all activity times in the rest of this report are given as
standardized activity times.
Summary Statistics
Table 6 gives the summary statistics for the standardized activity
times. Further details on the distribution of activity times are given
in Appendix C.

-------
- 22 -
Table 5
FREQUENCY DISTRIBUTION
OF TOTAL TIME
Range3 Count Density^
Minimum	17.0
17
-
18
5
0.7
18
-
19
4
0.6
19
-
20
24
3.4
20
-
21
34
4.8
21
-
22
44
6.2
22
-
23
94
13.3
23
-
23.5
77
21.8
23.
,5
- 24
171
48.5
24
-
24.5
136
38.6
24.
.5
- 25
62
17.6
25
-
26
44
6.2
26
-
27
7
1.0
27
-
28
3
0.4
Maximum	27.483
£
Total time in hours
including lower endpoint and
excluding upper endpoint.
^Densities in percentages
per hour range. See Note a
in Table 4 for the definition
of density.
Table 6
SUMMARY STATISTICS FOR STANDARDIZED
ACTIVITY TIMES
MET
Mean3
SDb
Public
0. 100
0.423
Private car
1.517
1.591
Pedestrian
0.269
0. 707
Parking
0.084
0.766
Shop
0.384
1.064
Office
3.051
3.914
£
Average activity time in hours.
^Standard deviation of activity
times in hours.

-------
- 23 -
III. CO CONCENTRATION DATA
MICROENVIRONMENT STUDY
The CO Microenvironment Study was conducted in the Washington, D.C.
metropolitan area during the winter of 1983. The main part of the study
focused on the measurement of commuting microenvironments, including
parking garages, driving an automobile, riding a bus, riding a train,
and walking. The detailed design and preliminary results from the study
are given in Flachsbart (1982a and b, 1984) and Mack et al. (1984).
For automobile commutes, the study identified eight routes that
"collectively extend 160 miles, about 8.6% of the total length (1,853
miles) of Washington's arterials and freeways." (In 1980, the
Washington metropolitan area had 9,432 miles of streets and roads,
including arterials, freeways, and locals.) The routes selected were
"ones considered to be heavily traveled and predicted to have high CO
exposures during rush hour periods." (Mack et al. 1984.)
Although the routes were chosen to be representative of the
arterials and freeways, they might not be representative of all routes
traveled by the general population. The empirical analysis found that
for the commuting METs, the MET concentrations from the CO
Microenvironment Study are substantially higher than corresponding MET
concentrations based on personal monitoring from the Urban Scale Study.
This could be due to the oversampling of arterial routes in the study.
A Commuter Study Links Data Base was constructed from the commuting
part of the Microenvironment Study. Each commuting route was divided
into links ranging from one-half to three miles, each link being a
physically distinct segment of the route, and will be regarded as a
microenvironment. Additional links were coded for parking garages.
For quality assurance, several commuting trips used collocated
monitors or inside/outside pairs. Preliminary results on monitor
accuracy and monitor precision were given in Mack et al. (1984). This
study restricts attention to the primary monitor in the paired
monitoring.

-------
- 24 -
The Microenvironment study also included monitoring on some indoor
microenvironments--shopping centers and offices. Additional monitoring
was also conducted on walking microenvironments. The pedestrian data
are combined with those from the commuting part of the study and
analyzed together.
The Microenvironment Study did not give a comprehensive coverage of
all microenvironments commonly encountered. One major exclusion was the
home microenvironments. A residual MET, referred to as the MET other,
consists of all microenvironments not covered in the Microenvironment
Study. For the exposure estimation to be discussed in Sec. IV, the
microenvironment monitoring data will be supplemented with personal
monitoring data from the Urban Scale Study for the microenvironments not
covered in the Microenvironment Study.
SUMMARY STATISTICS FOR THE MET CONCENTRATIONS
MET Concentrations Based on MEM
For each MET other than the MET other, the measurements from the
Microenvironment Study are aggregated into daily averages, which are
used as the MET concentrations in further analysis. A total of 43 days
were measured in the Microenvironment Study, from January 1 through
March 18, 1983.
Table 7 gives the summary statistics for the MET concentrations for
the six METs. As was expected, the concentrations in parking garages
are very high. The average concentration exceeds the one-hour federal
standard level of 35 .ppm. The concentration in private cars is also
fairly high. The average concentration exceeds the eight-hour federal
standard level of 9 ppm. Public transportation, walking, and shops have
moderate levels averaging about 5 ppm. Offices have low levels,
averaging about 2 ppm.
MET Concentrations Based on PM
An alternative set of estimates of MET concentrations can be
derived from the personal monitoring data in the Urban Scale Study. For
each activity segment reported, the exposure for that activity segment
is computed as the product of the duration of the activity segment and

-------
- 25 -
Table 7
SUMMARY STATISTICS FOR CO MET
CONCENTRATIONS BASED ON MEM
MET
Meana
SDb
N'c
Parking
44.
.55
32.
.36
29
Pedestrian
4,
.95
2.
.07
13
Public
5 .
.34
3.
. 12
24
Private car
11.
.39
3.
.11
32
Shop
4.
.20
1,
.54
9
Office
2.
.29
0.
.86
8
3
Average of the MET con-
centrations given in ppm.
^Standard deviation of the
MET concentrations given in
ppm.
Number of days on which
MEM data was available for
this MET.
its average CO concentration. For each participant and for each MET,
the exposures from the activity segments belonging to that MET are
summed as the total exposure for that MET. The total exposure in the
MET is divided by the total amount of time in the MET to get the MET
concentration.
For certain activity segments, the CO concentrations are not
available, possibly because of monitor failure. Those activity segments
are not included in the calculation of the MET concentrations. To
assess the effect of those missing data, the amount of time belonging to
such activity segments is calculated for each participant and for each
MET. For three METs--namely, shops, parking, and public transportat ion--
none of the participants had any activity segments with missing CO
concentration data. For the other three METs, for some participants
some of the activity segments did not have CO concentrations. However,
the amount of time for those activity segments is very small. For the
MET private car, the average amount of time per participant for which CO
concentration is missing is 0.004 hours. This is less than one-half of
1 percent of the average time of 1.623 hours spent in this MET. For the

-------
- 26 -
MET office, the average amount of time without CO concentration is 0.002
hours, very small compared with the average time of 3.051 hours in this
MET. For the MET pedestrian, the average amount of time without CO
concentration is 0.001 hours, again very small compared with the average
time of 0.269 hours in this MET. The activity segments with missing
concentration data therefore have very little effect.
Table 8 gives the summary statistics for the average MET
concentrations based on personal monitoring.
Comparison of MET Concentrations
The MET concentrations based on PM are substantially lower than the
correponding MET concentrations based on MEM, especially in the
commuting METs. (See Tables 7 and 8.) The most dramatic difference of
all is the MET parking, in which there is a fourfold difference between
PEM and MEM. The average MET concentration for private cars based on
MEM is more than twice the corresponding average concentration based on
personal monitoring.
Two major factors could explain the discrepancy between the MET
concentrations based on MEM and personal monitoring. First, for some
METs, such as the commuting METs, the Microenvironment Study might have
oversampled microenvironments with higher CO concentrations. For
Table 8
SUMMARY STATISTICS FOR CO MET
CONCENTRATIONS BASED ON PMa
MET
Mean
SD
Parking
9.60
12.6
Pedestrian
2.29
2. 35
Public
3. 10
2.65
Private car
5.08
5. 18
Shop
2. 19
2.47
Of f ice
1.82
2. 73
aThe summary statis-
tics are based on 705
participants in the
Urban Scale Study.

-------
- 27
example, the commuting routes selected were the "ones considered to be
heavily traveled and predicted to have high CO exposures during rush
hour periods." (Mack et al., 1984.) For future studies using the MEM
approach, alternative sampling strategies should be considered.
Second, Wallace et al. (1984) found a discrepancy between the COHb
levels estimated from breath measurements and those estimated from the
exposures based on personal monitoring, indicating that the PEMs might
underestimate the actual concentrations. One of their explanations for
the underestimation is that there is a "consistent decline in the
readings" on the CO monitor "as the battery approached the end of its
charge." Wallace et al. (1984) reported that "a study is underway at
EMSL-RTP to explain the behavior of PEMs after 24 hours of monitoring
and determine the battery effect on the calibration." The result of
that study should throw light on the nature of the discrepancy.

-------
- 28 -
IV. ESTIMATED EXPOSURES
This section discusses the estimation of personal exposures using
the- microenvironment monitoring approach and the personal monitoring
approach. For the MEM approach, both the convolution method and the
hybrid method are used. With either method, there is a substantial
discrepancy between the MEM and PM exposures. The MEM exposures are
about 40 percent higher. However, despite the discrepancy, the MEM
exposure is found to be a powerful predictor for the PM exposure. On
the log scale, the MEM exposure has the correct span relative to the PM
exposures; the discrepancy between the two estimated exposures is found
to be a constant drift.
APPLICATION OF THE CONVOLUTION METHOD
The convolution method discussed in Sec. I is an efficient method
of combining concentration data and activity time data to estimate
exposures. The hybrid method is a simpler method. This section will
discuss several empirical issues encountered in the application of the
two methods: the underlying assumption about independence between MET
concentrations and MET times, how to deal with the METs that were not
monitored in the Microenvironment Study, and the imputation of MET
concentrations on days when the MET is not measured in the
Microenvironment Study.
Independence and Correlation
Both the convolution and the hybrid methods assume that MET
concentration and MET time are stochastically independent. The same
assumption is implicitly made in other modeling approaches. The amount
of data available in this study does not permit examination of the
independence assumption in detail. Instead attention is restricted to
the weaker assumption that MET concentration and MET time are
uncorrelated. (Independence implies being uncorrelated, but being
uncorrelated does not necessarily imply independence.) The data from
the Urban Scale Study were used to estimate the correlations between MET

-------
- 29 -
time and MET concentration based on PEM. The results, given as Table 9,
indicate that for the Urban Scale Study, MET time and MET concentrations
are uncorrelated.1 This lack of correlation supports the application of
both the convolution method and the hybrid method.
Other Microenvironments
As was discussed in Sec. Ill, the microenvironment monitoring in
the Microenvironment Study was not comprehensive and did not cover all
METs usually encountered in general human activities. Those
microenvironments not covered in the Microenvironment Study will be
grouped into one MET and referred to as the MET other. For example, no
MEM data were collected in the home microenvironments, where most people
spend most of their time. Therefore the MEM data must be supplemented
with personal monitoring data for microenvironments without
microenvironment monitoring data.
Table 9
CORRELATION BETWEEN MET CONCENTRATION
AND MET TIME
MET
Pa_.
tb"
Parking
0.05
0.6
Public
0.09
0.5
Private car
-0.05
0.3
Pedestrian
0.02
0.8
Shops
-0.02
0.7
Office
CM
O
O
1
0.8
The correlation between MET
concentration and MET time for the
specific MET.
^The t statistic for the
null hypothesis p=0.
1The correlations estimated from the Washington Urban Scale Study
data are cross - individual. (Each person in the study is observed for
only one day.) The result here still allows for the possibility that
MET concentrations and MET times for the same individual might be
correlated across days. The latter correlation can be tested only if
there are multiple days on the same individual, such as in the Denver
Urban Scale Study (Johnson et al., 1984).

-------
- 30 -
With PM data substituting for MEM data for the MET other, the time-
weighted summation formula (3) is revised as follows. For the ith
person and the jth day, the exposure is estimated by
E. . = I.T., * C.. + T. x C. ,	(9)
ij k lk jk lr lr
where the summation is taken over all METs covered in the
Microenvironment Study,	is the MET time for the ith person in the
kth MET covered in the Microenvironment Study, C.. is the MET
Jk
concentration for the jth day in the kth MET. The last term in Eq. (9)
gives the exposure from the MET other, referred to as the rth MET.
is the MET concentration for the ith person in the MET other, based on
personal monitoring. T^^ is the MET time in the MET other for the ith
person.
The formula (8) for the hybrid approach is similarly modified as
follows:
E. = I. C, x T., + C x T. .	(10)
i k k lk r lr
Imputation of MET Concentrations on Missing Days
On any one of the sampling days, the microenvironment monitoring
was not conducted on all the METs. It is therefore necessary to impute
the missing MET concentrations.
The imputation procedure is based on the relationship among MET
concentrations. Even though the MET concentration for the kth MET is
not observed on the jth day, the MET concentrations for the other METs
may be used to impute the missing MET concentration if the MET
concentrations are related to each other.
An imputation procedure similar to the EM-algorithm (Dempster,
Laird, and Rubin, 1977) is used to impute the missing MET
concentrations. First the missing values are replaced by the sample
mean based on observed values for the same MET. For each MET the MET
concentration is regressed on the other MET concentrations. In other

-------
- 31 -
words, is regressed on C^, C^, etc.	is regressed on C^, C^, etc.
where is the concentration in the first MET, is the concentration
in the second MET, etc. From each estimated regression model, the
predicted concentrations are used to replace the ones originally
missing. (The observed MET concentrations are retained and not replaced
by the predicted values.). The algorithm is then cycled back to update
the regression models. The algorithm is cycled through four iterations.
The summary statistics for the final iteration are given in Table 10.
Compared with the corresponding summary statistics in Table 7 based on
MET concentrations, the average concentrations are similar, but the
standard deviations for the imputed concentrations are lower. This is a
common phenomenon in imputation procedures. The predicted values do not
take into account the variation not explained by the fitted regression
models.
Table 10
SUMMARY OF IMPUTED MET CONCENTRATION'S
MET
Mean
SD
Parking
46.82
28.50
Pedestrian
4.82
1.19
Public
5.10
2.68
Private car
11.26
2. 70
Office
2.30
0.37
Shops
4.30
0.75
ESTIMATED EXPOSURES BASED ON MEM
The exposures for 30,315 convoluted person-days were estimated
based on the whole sample of all 705 participants in the Urban Scale
Study and the 43 days in the Microenvironment Study. The summary
statistics for the MEM exposures based on the convolution method are
given as the first row in'Table 11 and the first column in Table 12.
The average exposure is just over 2 ppm-days. The distributions are
highly skewed.

-------
- 32 -
The summary statistics for the logarithmic transformation of the
MEM exposures based on the convolution method are given as the first row
in Table 13 and the first row in Table 14. The average log exposure is
about 0.5 log(ppm-day). The log concentrations are somewhat skewed to
the left.
The summary statistics for the hybrid exposures are given as the
second row in Tables 11 and 13, and the second column in Tables 12 and
14. The average and median exposures for the two methods are very
similar. The hybrid exposures have a much narrower spread, indicating
that neglecting the variability in the MET concentrations has an
important effect on the exposure distribution for our data. In Tables 7
and 8 the standard deviations for the MET concentrations were found to
be fairly high for several of the METs.
ESTIMATED EXPOSURES BASED ON PM
An alternative set of exposure estimates can be derived using the
personal monitoring data available from the Urban Scale Study. In
particular,
E. = I.T.. x c.. + T. x c. ,	(11)
i k lk lk lr lr
where the summation is taken over all METs covered in the
Microenvironment Study, T^ is the MET time for the ith person in the
kth MET covered in the Microenvironment Study,	is the MET
concentration for the ith person in the kth MET. The last term in Eq.
(11) gives the exposure from the MET other, referred to as the rth MET.
C. is the MET concentration for the ith person in the MET other, based
lr	r	>
on personal monitoring. T is the MET time in the MET other for the
ith person.
In a comparison of Eqs. (9) and (11), it can be seen that the
difference between the exposure estimates based on the two approaches is
the replacement of the MET concentration based on PM,	for the MET
concentration based on MEM, C., .
jk

-------
- 33 -
Table 11
SUMMARIES FOR MEM AND PM EXPOSURES
Method	Mean3	SDb	IQRC Skew0*	Kurte
NEM-Cf~	2.29	2.22	1.47	9.47	175.0
MEM-HS	2.29	1.63	0.77	9.39	114.4
PM	1.59	1.63	1.52	3.11	16.7
Average of the estimated exposures
in ppm-days.
^Standard deviation of the estimated
exposures.
c
Interquartile range of the esti-
mated exposures.
^Skewness of the estimated exposures
e
Kurtosis of the estimated exposures
^MEM exposure using the convolution
method.
cr
MEM exposure using the hybrid
method.
Table 12
PERCENTILES OF MEM AND PM EXPOSURES
(ppm-day)
Percentile MEM-C MEM-H PM
1
5
10
25
50
75
90
95
99
0.05	1.20	0.05
0.45	1.20	0.10
0.75	1.34	0.22
1.28	1.70	0.58
1.89	2.06	1.17
2.75	2.47	2.10
3.90	2.99	3.30
5.08	3.63	4.49
10.00	6.90	7.54

-------
- 34 -
The personal monitoring approach provides 705 estimated exposures,
one for each person in the Urban Scale Study monitored for one day. The
summary statistics for the PM exposures are given as the third row of
Table 11 and the third column of Table 12. The average PM exposure is
about 1.5 ppm-days. The distribution is somewhat skewed, with skewness
being about 3.
The log PM exposure is given as the third row of Table 13 and the
third column of Table 14. The average log PM exposure is about 0
log(ppm-day). The distribution is slightly skewed toward the left, with
the skewness coefficient being about -0.7. The PM exposures are roughly
approximated by the lognormal distribution.
COMPARISON OF EXPOSURE DISTRIBUTIONS
For certain purposes such as risk assessment, it is only necessary
to estimate the distribution of exposures without identifying each
individual's exposure.
The comparison between the two sets of summary statistics for the
estimated exposures given in Tables 11 and 12 indicates that the two
distributions are substantially different. The average MEM exposure is
about 40 percent higher than the average PM exposure. The difference is
highly significant (t = 6.69 for the convolution method, t = .8.01 for
the hybrid method). The percentiles for the MEM exposures based on the
convolution method are consistently higher than the corresponding
percentiles for the PM exposures through the entire range of their
distributions. The percentiles in the lower range for the hybrid
exposures are higher than the corresponding percentiles for the PM
exposures, and the opposite is true for the upper range. This indicates
that the distribution of hybrid exposures is closer to the center than
the distribution for the PM exposures.
The two-sample Kolmogorov-Smirnov test (Smirnov, 1939; Massey,
1951) for the difference between the MEM and PM exposure distributions
is highly significant (P < 0.0000001 for both methods).
The comparison between the summary statistics for the log estimated
exposures also indicates major differences between the MEM and PM
exposures. The average log MEM exposure is significantly higher than
the average log PM exposure.

-------
Table 13
SUMMARIES OF LOG MEM AND EXPOSURES
Method
Meana
SDb
IQRc
Skewd
Kurte
MEM-C
0.56
0.82
0. 76
-1. 38
4.85
MEM-H
0. 74
0.36
0.37
1.81
8.71
PM
-0.02
1. 10
1.29
-0.68
0.32
£
Average of the estimated log
exposures in log(ppm-day).
^Standard deviation of the estimated
log exposures.
c
Interquartile range for the esti-
ated log exposures.
^Skewness of the estimated log
exposures.
0
Kurtosis of the estimated log
exposures.
Table 14
PERCENTILES OF LOG MEM AND PM EXPOSURES
(log (ppm-day))
Percentile MEM-C MEM-H PM
1
-2.98
0.18
1
o
o
5
1
o
v£>
0. 18
-2.33
10
-0.29
0.29
-1.53
25
0.25
0.53
-0.55
50
0.64
CM
o
0. 15
75
1.01
0.90
0.74
90
1.36
1.09
1. 19
95
1.63
1.29
1.50
99
2.30
1.93
2.02

-------
- 36 -
The discrepancy between the MEM exposure distribution and the PM
exposure distribution is consistent with the discrepancy shown in Tables
7 and 8, indicating that the MET • concentrations based on MEM are
substantially higher than the corresponding MET concentrations based on
personal monitoring. The discrepancy can be attributed to the
oversampling of high concentration microenvironments in the
Microenvironment Study and the underestimation of concentrations in the
personal monitoring in the Urban Scale Study.
COMPARISON OF INDIVIDUAL-SPECIFIC EXPOSURES
Mean Squared Error
For certain situations, such as the quantification of health
effects of air pollution, the estimated exposures must be close to the
actual exposures. The actual exposures are unknown in this study. Both
the MEM exposures and the PM exposures are estimates and are subject to
errors. As was discussed in Sec. I, the two estimated exposures are
compared in terms of how close they are to each other. The mean squared
error (MSE) given in Eq. (1) is used as the criterion.
To interpret the criterion MSE, the percentage gain was also
derived from the ignorant estimate of zero:
Gain = (1 - MSE^MSEq) x 100% ,
where MSE^ = Average of (PM exposure - MEM exposure)2,
MSEq = Average of (PM exposure - 0)2.
The percentage gain measures how much better the MEM exposure is than
the ignorant estimate that all exposures are zero. The criterion can
also be interpreted as the percentage of the sum of squares in the PM
exposures that is explained by the MEM exposures,2 analogous to the R2
statistic in the usual regression analysis.
2The criterion GAIN could be defined alternatively as the
percentage of sum of squares in the MEM exposures that is explained by
the PM exposures, but since the burden of proof lies on the MEM
exposures, GAIN is defined as given in the text.

-------
- 37 -
The MSE and GAIN results are given in Table 15. The mean squared
error indicates that there is a substantial difference between the MEM
and the PM exposures. On the original scale, the square root of MSE is
about 2 ppm-days, indicating that on the average the MEM exposure is
about 2 ppm-days away from the PM exposure. This is large compared with
the average PM exposure of 1.6 ppm-day and the standard deviation of
about 1.6 ppm-days, as given in Table 11. The percentage gains for both
methods are around 16 percent, indicating that there is a lot of
variation left unexplained by the MEM exposure.
On the log scale, the square root of MSE is about 0.9 log(ppm-
day) for the convolution method, indicating that on the average the MEM
exposure is about 0.9 log(ppm-day) away from the PM exposure. This is
also large compared with the average log PM exposure of about 0.0
log(ppm-day), and the standard deviation of about 1.1 log(ppm-day), as
given in Table 13. The square root of MSE for the hybrid method, about
1.3, is even larger. The percentage gain for the convolution method is
about 35 percent, indicating that the log MEM exposures based on the
convolution method explain a substantial fraction of variation in the
log PM exposures. The percentage gain for the hybrid approach is only
about 4 percent.
Table 15
MSE AND GAIN FOR MEM EXPOSURES
Convolution	Hybrid
Gain	Gain
Scale MSE (Percent)	MSE (Percent)
Original 4.32^ 17.1	4.45a 15.8
Log	0.79b 34.1	1.6?b 4.2
3
MSE given in (ppm-day)2.
^MSE given in (log (ppm-day))2

-------
- 38 -
The large mean squared errors are consistent with the discrepancy
shown in Tables 7 and 8, indicating that the MET concentrations measured
from MEM are substantially higher than the corresponding concentrations
measured from personal monitoring.
MEM Exposure as a Predictor for PM Exposure
For certain situations such as qualifying the health effects of air
pollution, it is only necessary that the estimated exposure be a good
predictor of actual exposure. For such situations the appropriate way
to assess the validity of the estimated exposure is to exaimine the
regression relationship between the actual and estimated exposures. The
slope coefficient in the regression relationship must be significant,
indicating that the estimated exposure predicts the ranking of actual
exposures, even though the magnitude might be off. Furthermore, the
slope coefficient should be close to one, and the intercept coefficient
close to zero, implying that the estimated exposures are properly
calibrated relative to the actual exposures.
In this study we don't know the actual exposures, therefore cannot
estimate the relationship between the estimated exposures and the
unobserved actual exposures. As discussed in Sec. I, the PM exposure is
used as the benchmark and the regression relationship tested for between
the two estimated exposures, regressing the PM exposure on the MEM
exposure.
The results for the regression of PM exposures on the MEM exposures
are given as Table 16. On the original scale, the regression results
show a very significant relationship between the PM and the MEM
exposures. The convolution method gives a more significant slope
coefficient than the hybrid method. This indicates that even though the
MET concentrations from MEM and PM are substantially different, the MEM
exposures are still useful for predicting the ranking of the PM
exposures. In other words, given that a certain individual's MEM
exposure is high, it is reasonable to expect that his PM exposure is
also high. For health effect studies in which the main focus is on
qualifying the existence of health effects, this result indicates that
the MEM can be used as a proxy for the PM exposure.

-------
- 39 -
Table 16
REGRESSIONS OF PM EXPOSURES ON MEM EXPOSURES
(t-statistics given in parentheses)
R2
Method
Scale
Intercept
Slope
(Percent)
Convol
Original
0.528
(7.70)
0.466
(21.64)
39.9

Log
-0.601
(-19.35)
1.053
(33.44)
61.3
Hybrid
Original
1.011
(9.84)
0.254
(6.94)
6.4

Log
-0.667
(-7.39)
0.879
(8.02)
8.4
The R2 statistic for the convolution method is about 40 percent,
indicating that the MEM exposure is not only a significant predictor for
the PM exposure but is also an informative predictor, explaining an
important fraction of the variability in the PM exposure. The hybrid
method has a much smaller R2.
With the convolution method, the slope coefficient in this
regression is about 0.5, substantially smaller than one, and the
intercept coefficient is about 0.5 ppm-day, significantly larger than
zero. This indicates that the MEM exposures are not well calibrated
relative to the PM exposures.
For simplicity the estimated regression model will be approximated
as follows:
PM exposure = 0.5 + 0.5 * MEM exposure.
At low levels (less than 1 ppm-day), the MEM exposure underestimates the
PM exposures. For example, for an individual with MEM exposure equal to
zero, the regression model predicts that his actual exposure is probably
about 0.5 ppm-day. At higher levels (more than 1 ppm-day), the MEM
exposure overestimates the PM exposure. For example, for an individual

-------
- 40 -
with MEM exposure equal to 10 ppm-day, the regression model predicts
that his PM exposure is probably about 5.5 ppm-day, substantially lower
than the MEM exposure. Because the average MEM exposure is about 2 ppm-
day, for most people the MEM exposure overestimates the PM exposure
according to the regression model.
It is possible to recalibrate the MEM exposures using the
regression model as follows:
Recalibrated estimate = 0.5 + 0.5 x MEM exposure.
The feasibility of such a recalibration in practice remains to be
studied. In future applications of the MEM approach, the PM exposures
for such a recalibration might not be available. Moreover, because the
PM exposures might be estimated on a small sample, it is not clear that
the recalibration will improve the precision of the result.
On the log scale, the regression results also show a significant
relationship between the MEM exposure and the PM exposure, indicating
that the MEM exposures successfully predict the ranking of the PM
exposures. The convolution method gives a more significant slope
coefficient than the hybrid method. The R2 statistic for the
convolution method is about 60 percent, indicating that the log MEM
exposure is fairly powerful in explaining an important, fraction of the
variability of the log PM exposure. The hybrid method gives a much
smaller R2.
With the convolution method, the slope coefficient in the log-
scale regression is very close to one, the difference is not
statistically significant at the conventional 5 percent level (t =
1.68). This indicates that the span of the MEM exposures is well-
calibrated relative to the PM exposures. The intercept coefficient is
about -0.6 log(ppm-day), significantly less than zero, indicating that
the MEM exposure consistently overestimates the PM exposure. For
simplicity of discussion, the regression models are approximated as
follows:
log PM exposure = -0.6 + log MEM exposure.

-------
- 41 -
The approximate regression model indicates that at all levels, the
log MEM exposures overestimate the log PM exposure by about
0.6 log(ppm-day). This can be interpreted as follows: The log MEM
exposures have the correct span relative to the log PM exposures but
have a constant drift -0.6 log(ppm-day). In terms of the original
scale, this means that the PM exposure is proportional to thie MEM
exposure with the proportionality factor exp(-0.6) = 0.55:
PM exposure - 0.55 x MEM exposure.
It is also possible to recalibrate the log MEM exposures using the
regression model as follows:
Recalibrated estimate = 0.6 + log MEM exposure.
As was noted earlier in the discussion on the recalibration on the
original scale, the validity of such a recalibration remains to be
established.
CONCLUSION
The discrepancy found in this study between the MEM exposures and
the PM exposures, in terms of both individual-specific exposure and
exposure distribution, is probably specific to the data used in the
current study and should be generalized to future exposure studies only
with caution. Given the imperfect sampling of microenvironments and the
problems in the personal monitoring discussed earlier, it is impressive
that the MEM exposure still turns out to be a successful predictor for
the PM exposure, especially on the log scale, on which the MEM exposure
based on the convolution method has the correct span relative to the PM
exposure, and the discrepancy is restricted to a constant drift.
The convolution method is preferable to the hybrid method for the
data used here. There is much variability in the MET concentrations in
these data, which is unfavorable to the hybrid approach.

-------
- 42 -
V. EVALUATION OF MET CLASSIFICATION SCHEMES
The analysis in the earlier part of this report assumes the MET
classification scheme, which classifies all microenvironments usually
encountered into seven METs: parking, pedestrian, public trans-
portation, private cars, offices, shops, and other. Conceivably
other classification schemes could also be used. This section discusses
the basis for choosing this particular MET classification scheme.
Duan (1981) recommended that the grouping of the microenvironments
into METs should be carried out in two stages: a preliminary
classification stage and an evaluation stage. In the first stage, the
researcher should identify a profile of potentially useful METs. Those
METs are the minimal chunks that can be identified from the information
available, and will be called elementary METs. The second stage will
evaluate the elementary METs and consider whether each of them should be
analyzed as a distinct MET on its own or should be combined into coarser
groupings referred to as composite METs.
From the data from the Microenvironment Study and the Urban Scale
Study, eight elementary METs can be identified: parking, pedestrian,
bus, rail, private cars, offices, shops, and other. A criterion
developed in Duan (1981) is applied to evaluate the above METs. The
results are as follows:
1.	The best decomposition of microenvironments covered in the
Microenvironment Study is to separate commuting from business.
The composite MET commuting consists of the elementary METs
parking, buses, rails, private cars, and pedestrian. The
composite MET business consists of the elementary METs shops
and offices.
2.	After this primary decomposition, the next best decomposition
is to separate commuting into parking and in-transit.
3.	After the first two decompositions, the next best decomposition
is to separate business into shops and offices.

-------
- 43 -
4.	The next best decomposition is to separate the MET in-transit
into vehicles and pedestrian. The composite MET vehicle
consists of the elementary METs public tran sport at ion and
private cars.
5.	The next best decomposition is to separate the MET vehicle into
public transportation and private vehicles.
6.	The least effective decomposition is to separate the MET public
transportation into buses and rails.
7.	The elementary MET other has to be kept separate from all other
elementary METs because there is no concentration data from the
Microenvironment Study for microenvironments in this MET.
Based on the above results, buses and rails were combined into one
composite MET, public transportation, giving the seven METs used in the
exposure analysis in earlier sections.
ELEMENTARY METs
Fairly detailed information was collected in both the
Microenvironment Study and the Urban Scale Study to characterize the
microenvironments and activity segments being measured. Based on the
information common to the two studies, eight elementary METs may be
identified: parking, pedestrian, buses, rails, private cars, offices,
shops, and other. The definitions of parking, pedestrian, private cars,
offices, shops, and other are given in Sec. II. Buses and rails are
self-defined. Further details on the definitions of the elementary METs
are given in Appendix A.
The identification of the eight elementary METs given above did not
use all the information available from the two studies. For example,,
the Micr.oenvironment Study provides additional information on the
in-transit microenvironments such as average speed of vehicle. With
only the MEM data, the information on speed may be used to refine the
elementary MET private car into, for example, private car at a low speed
(say, below 35 mph), private car at a medium speed (say, between 35 and
50 mph), and private car at a high speed (say, above 50 mph). However,
the speed information is not available from the Urban Scale Study, so

-------
- 44 -
those refined METs cannot be identified for activity time data from the
Urban Scale Study. In the Urban Scale Study, it is known only that the
participant is in a private car during a certain activity segment, but
the vehicle's speed is unknown. Because any elementary MET to be
considered has to be identifiable both for the Microenvironment Study
and for the Urban Scale Study, the speed information cannot be used to
refine the elementary MET private car.
DUAN'S CRITERION
Duan (1981) developed a criterion to evaluate MET classification
schemes that has become known as Duan's criterion. This section will
give a new interpretation of this criterion in terms of effective sample
s ize.
The criterion is based on the estimation of average exposure when
implementing the MET approach using the enhanced personal monitoring
(EPM) approach. In the context of EPM, assuming a sufficient amount of
additional diary information to practically eliminate the part of the
variability in the estimated average exposure due to the variability in
the MET times, Duan (1981) showed that decomposition of a coarser MET
into finer METs always decreases the variance of the estimated average
exposure. Duan's criterion (DC) is defined in terms of the amount of
decrease in this variance as follows:
DC = n x (Var - Var r),
c	I
where n is the number of person-days in the monitoring sample, Var^ is
the variance of the estimated average exposure based on the coarser MET
classification, and Var^ is the variance of the estimated average
exposure based on the finer MET classification after the decomposition
being considered. The criterion can be interpreted as the decrease in
variance per observation.
The mathematical form of the criterion is rather complicated and is
given in Appendix E. Following are several of its important features:

-------
- 45 -
•	The more different the concentrations in the two refined METs
are, the more gain there is from the refinement.
•	The criterion may be factored into two terms, the first term
determined entirely by the distribution of the MET times, the
second term determined entirely by the distribution of the MET
concentrations. Therefore the two factors in the criterion may
be evaluated separately. For example, one may assess the first
term from a population-based survey or activity diary and the
second term from MEM data.
•	The more time spent in the two finer METs combined, the more
gain there is from the refinement.
•	The more variable the time allocation is between the two finer
METs, the more gain there is from the refinement. The
variability in exposure comes from two sources: the
variability in MET concentrations and the variability in MET
times. The EPM approach incorporates additional activity data
to eliminate or reduce the variability in the average exposure
due to the variability in MET times. Therefore the value in a
refined MET classification depends on the variability in MET
times. If there is no variability in MET times, then there can
be no gain from the decomposition.
PIESS
Duan's criterion may be interpreted in terms of the percentage
increase in effective sample £ize (PIESS). There are two ways to
increase effective sample size. The first is to increase the sample
size. The second is to use a more efficient analytic method (such as
the more refined MET classification) that gives a more precise result by
decreasing the variance of the estimated quantity. The PIESS with a
more efficient analytic method (such as the finer MET classification) is
the percentage increase in sample size that would be required to achieve
the same precision with the less efficient analytic method (such as the
coarser MET classification).

-------
- 46 -
It can be shown that the PIESS in the kth decomposition (imposed
after the first k - 1 decompositions have been imposed) is given by
PIESSk = DCk/(Var(E) - DC - DC - ... - DCk) x 100 percent,
where Var(E) is the variance of exposure, DC^ is the criterion for the
first decomposition,	DC^ is the criterion for the kth
decomposition.
The following is a hypothetical example:
n = 100,
Var(E) = 10,
DC x = 5 ,
DC2 = 2.
If one does not use the MET approach and simply monitors the exposure
for n = 100 subjects with PEM, the variance for the estimated average
exposure is
Var(E)/100 = 0.1 .
If one uses the MET approach and imposes the first decomposition, say
decompose indoor and outdoor as two distinct METs, the variance of the
estimated average exposure is reduced to
(Var(E) - DC )/100 = 0.05 .
The PIESS for this decomposition is
DC^/(Var(E) - DC^) = 100 percent.
For the coarser method (not using the MET approach) to achieve the same
precision, it is necessary to increase the sample size by 100 percent.
When that is done and 200 subjects are monitored with the coarser
method, the variance of the estimated average exposure is

-------
- 47 -
Var(E)/200 = 0.05 ,
the same as the variance based on the MET approach with the first
decompos it ion.
Further imposition of the second decomposition, say decompose
indoor into indoor with gas stove in use and indoor without gas stove in
use, the variance of the estimated average exposure is further reduced
to
(Var(E) - DC - DC2)/100 = 0.03.
The PIESS for this decomposition is
DC^/CVarCE) - DC^ - DC^) = 67 percent.
For the coarser method (using only the first decomposition) to
achieve the same precision, it is necessary to increase the sample size
by 67 percent. In other words, 100 observations with both
decompositions is equivalent to 167 observations with only the first
decomposition. The latter is equivalent to 334 observations without the
MET approach.
Relative Criterion Versus Absolute Criterion
Duan's criterion (as well the PIESS interpretation of this
criterion) compares the relative benefits from various classifications
to determine which one is preferable. It can be applied to rank the
relative benefits from various decompositions. The process can thought
of as cutting a pie. Duan's criterion can be used to decide which is
the best first cut, which is the best second cut, etc. However, where
to stop cannot be determined from this criterion.
If other sources permit determination of the maximum number of METs
that can be used in a study--for example, the data logger might only
allow six METs to be distinguished — the criterion may be applied to
determine the best cuts or decompositions that will lead to the six most
efficient METs. Other situations might not permit determination of how

-------
- 48 -
many METs to use from other sources. In such situations it will be
desirable to develop an absolute criterion to determine where to stop.
EMPIRICAL RESULTS
Duan's criterion was applied to the concentration data from the
Microenvironment Study and the activity time data from the Urban Scale
Study to rank the relative benefits from various MET decomposition
schemes. The details on the estimation of the criterion are given in
Appendix F.
Table 17 summarizes the final ranking of the MET decompositions and
gives Duan's criterion and PIESS for each MET decomposition. The
largest improvement in Duan's criterion and PIESS is achieved by the
decomposition that distinguishes commute and business. (Because of the
limitations in the Microenvironment Study, it is possible to address
only the business part of noncommute, namely shops and offices.) This
decomposition results in nearly a one-third improvement in effective
sample size.
Table 17
RANKING OF MET DECOMPOSITIONS
MET 0 a
MET lb
MET 2b
DCC
PIESSd
(Percent)
All
Commute
Bus iness
364.6
31.1
Commute
Parking
In-transit
89. 1
8.2
Bus iness
Shops
Offices
15.8
1.5
In-trans it
Vehicles
Pedestrian
12.6
1.2
Vehicles
Public
Private car
6.2
0.6
Public
Bus
Rail
0.05
0.005
£
MET 0: The prototype MET being decomposed.
bMET 1, MET 2 : The two new METs.
CDuan's criterion.
^Percentage increase in effective sample size.

-------
- 49 -
Duan's criterion for this decomposition is about six times the
value of the criterion for the next best decomposition of commute into
parking and in-transit. This means that it takes about six
decompositions similar to the latter to match the improvement in the
former decomposition. In other words, the former decomposition is
equivalent to six decompositions like the latter.
The decomposition of business into shops and offices compares
favorably with several finer decompositions of commute, even though the
concentrations in the METs offices and shops are both low compared with
commuting METs. This is mainly because much more time is spent in the
MET business than in the commuting METs. Therefore in considering MET
decompositions it is important to recognize METs in which people spend a
good deal of their time, even though the concentrations in these METs
might be low. Hartwell et al. (1984a) noted that even though the indoor
CO level is generally lower than that encountered during commuting,
those METs contribute more than half of the total exposure.
Most of the decompositions result in a nontrivial improvement in
PIESS, except for the decomposition of public transportation into bus
and rail. It takes more than a hundred decompositions similar to this
decomposition to match the improvement in other decompositions, mainly
because the average amount of time spent in these METs is very small.
Therefore this decomposition will be ignored in the rest of this
analysis.1
lDuan's criterion does not permit determination of where to stop.
The decision may be viewed as a decision that seven METs is the maximum
number of METs that can be used or that a PIESS substantially less than
1 percent does not warrant the creation of an extra MET.

-------
- 50 -
VI. FUTURE APPLICATIONS OF THE MET APPROACH
ENHANCED PERSONAL MONITORING AND
MICROENVIRONMENT MONITORING
There are two ways to implement the MET approach, enhanced personal
monitoring (EPM) and microenvironment monitoring (MEM). If the former
approach is taken, the sampled participants will sample the
microenvironments for the study using his own activity pattern;
therefore there is no need to be concerned about the sampling of
microenvironments. In the latter approach, the research staff have to
determine the microenvironments to be measured; therefore, an
appropriate sampling scheme must be devised for the microenvironments.
The main disadvantage of EPM relative to microenvironment
monitoring is the need to use human subjects in the monitoring and the
need to use PEMs capable of continuous measurements. When those factors
do not apply--e.g., if a future exposure assessment study using the MET
approach will be conducted in conjunction with a personal monitoring
study with PEMs capable of continuous measurements--enhanced personal
monitoring will be preferable. The apparent advantage is the avoidance
of the difficult task of sampling microenvironments. So long as the
human subjects are chosen to be representative of the target population,
they will automatically generate a representative sample of
microenvironments during their monitoring.
There are certain situations in which microenvironment monitoring
will be the better choice. An obvious case is when there are no PEMs
capable of continuous measurements. In that situation it might still be
possible to conduct microenvironment monitoring using portable (but not
truly personal) monitors. Another case is when a comprehensive activity
database already exists. In that case it is possible to avoid human
subjects entirely by using microenvironment monitoring.
A third situation for which microenvironment monitoring might be
chosen is when the primary goal is not to assess exposure but to
generate data to feed into exposure models such as SHAPE or NEM. In
those situations it is still important to design an appropriate sampling

-------
- 51 -
scheme for microenvironments. Otherwise the distribution of
microenvironment concentration generated as input to the models might
not be appropriate.
SAMPLING OF MICROENVIRONMENTS
For the rest of this section it is assumed that microenvironment
monitoring has been decided as the more appropriate way to implement the
MET approach for a future exposure study. For each MET, one needs to
determine the microenvironments belonging to that MET to be sampled.
It was noted in Sec. I that the appropriate target population is
not the population of microenvironments such as all office spaces, but
the intersection of people and the microenvironments. There are two
possible schemes to generate a representative sample of this target
population. The first is the weighted sampling scheme. The second is
to simulate real human activities, using an existing activity database.
The two approaches are not exclusive. They may be applied simul-
taneously in the same study. For home and office METs, the weighted
sampling scheme might be preferable. For commuting METs, simulated
activity might be preferable. If feasible it would also be desirable
to collect some personal monitoring data at the same time for comparison.
Weighted Sampling Scheme
With this approach, the microenvironments are sampled as if they
were the target population, then the number of people present in the
sampled microenvironments are counted or estimated as weights to adjust
the difference between the two target populations. This approach was
first discussed in Duan (1982). It is valid only if each person in the
period of interest will visit only one microenvironment belonging to the
MET of interest.
The assumption of visiting only one microenvironment of the same
type will probably apply to certain METs but not to others. For
example, it probably applies to the METs home and office for most
people. However, it might not apply to certain METs such as commuting
METs--a substantial fraction of people probably visit more than one such
microenvironment during the period of interest. How multiple visits to

-------
- 52 -
microenvironments of the same type will affect the validity of this
approach remains to be studied empirically.
The best way to sample microenvironments in this approach is to use
area probability sampling, with the modification that the count of
people present at the microenvironment be incorporated as part of the
sampling weights.
Simulate Real Human Activities
An alternative to the weighted sampling scheme is to abandon the
direct sampling of microenvironments and sample people instead. This
approach requires that there already is an activity database providing
sufficient detail on the location and characteristics of the micro-
environments reported in the diaries. This is not necessarily a
restrictive requirement, because one type of situation in which
microenvironment monitoring is preferable to enhanced personal
monitoring is when there is an existing activity database.
With this approach, diaries from the activity database are sampled
and an attempt made to simulate the reported activity paths. For
example, assume that we are interested in the MET commuting, to which
the weighted sampling scheme might not apply because of the problem of
visiting multiple microenvironments in the same MET. Diaries will be
sampled from the activity database, and the commuting routes described
in the diaries laid out along with descriptions of the conditions of the
routes--e.g., type of vehicle, ventilation status, speed, and presence
of smokers. (It might be easier to collect part of this information on
a general level from a recall survey rather than from diaries.) The
research staff can then set out with the monitoring instruments to
measure the MET exposure this person would receive given the described
commuting path and conditions. It may be impossible to match all
aspects of a microenvironment based on information collected from the
diaries and recall surveys. Nevertheless it should be possible to match
the main relevant features.

-------
- 53 -
STUDY DESIGNS AND DATA COLLECTION
Collection of Additional Diary Information
If primary data collection for activity patterns will be used in a
future study, it is preferable to collect the diary information over a
longer period of time to enhance the information on activities. The
marginal cost for collecting additional diaries from a person already in
the sample is likely to be low. If enhanced personal monitoring will be
used, it will be preferable to collect the diary information for a few
days before passing out the PEM to ensure that the participant is
already familiar with the diary instrument so that the quality of the
diary information during the monitoring period will be good.
Closed Format for the Diary
Hartwell et al. (1984a) recommended that the open-ended format in
the diaries in the Urban Scale Study be replaced by a closed format in
future studies to improve the quality of the activity data. In the
closed format the participant is given a list of METs with instructions
how to determine which MET a certain activity belongs to, and is
instructed to report the specific MET he is in during each activity
segment. This change will improve the quality of the activity data to
be collected.
Hartwell et al. (1984a) also recommended that a simplified version
of the diary be built into the data logger so that when the participant
keys in the MET at each change in activity, the activity will be
recorded automatically along with the concentration. This will probably
improve the quality of the activity data. The hardcopy diary and
electronic diary can be compared during analysis for validation,
avoiding a substantial fraction of the mismatches between the diary and
the monitoring data segments.
Match Between Activity and Monitoring METs
If microenvironment monitoring will be used in future studies along
with primary diary data collection, it will be desirable to match the
METs in the design phase so that the METs monitored can be matched with
METs defined from the diaries. In particular, if the closed format for

-------
- 54 -
the diary is adopted, the METs in the diaries should match with the METs
in the microenvironment monitoring.
Overlap Between METs in Microenvironment Monitoring
If microenvironment monitoring will be used in future studies, it
will be desirable to design the monitoring schedule so that there will
be sufficient overlap in time between each pair of MET to allow for the
examination of the relationship among the MET concentrations.
FURTHER USE OF THE ACTIVITY DATABASE
The activity database from the Urban Scale Study is not perfect.
However, it is the first of its kind, providing a population-based
sample of real human activities containing information relevant to
combustion-related air pollution. As was noted in Sec. I, one advantage
of the MET approach is that an existing activity database can be used
for other purposes. For example, the current activity database can be
used in a future exposure study on nitrogen dioxide. The nonambient
sources of nitrogen dioxide are basically the same as the nonambient
sources for CO, namely, indoor combustion (smoking, gas stove) and
commuting. The activities identified in the current activity database
are probably sufficient for characterizing METs relevant for nitrogen
dioxide. Therefore a future study on human exposure to nitrogen dioxide
can be carried out with little need for new activity data. One
possibility would be to use the monitors available to conduct a
microenvironment study for nitrogen dioxide and apply the convolution
method to estimate nitrogen dioxide exposure.
Such a nitrogen dioxide study using the current activity database
can serve as a useful pilot study for the future application of the
continuous PEM for nitrogen dioxide under development. The information
generated can help evaluate the MET classification scheme developed in
this study for nitrogen dioxide, and will also be useful in identifying
the subpopulation at high risk.
The current activity database can also be used to provide updated
input to existing exposure models such as SHAPE and NEM. Before the
Urban Scale Study, the exposure models had to rely on synthetic activity
patterns based on the best input information then available. With the

-------
- 55 -
activity database from the Urban Scale Study, it is possible to provide
updated activity information for the exposure models, using either the
realized activity paths or statistical summaries from the current
activity database to recalibrate the statistical distribution used as
~
input to the exposure models.

-------
- 57 -
Appendix A
DEFINITIONS OF METS
The heuristic definitions of METs used in this study are given in
Sec. II. This appendix gives the detailed definition rules for the
elementary METs, including parking, pedestrian, buses, rails, private
cars, offices, shops, and other. The composite MET public trans-
portation considered in the exposure estimation is the union of the
elementary METs buses and rails.
DEFINITIONS OF ELEMENTARY METS FROM
THE MICROENVIRONMENT STUDY
Parking Garage
From the Microenvironment Study, a microenvironment is defined to
be a parking garage when the microenvironment is measured as part of the
commuter part of the Microenvironment Study, and the link number is
given in the Commuter Study Links Data Base either as 31, indicating the
microenvironment is inside a parking garage at the beginning of the
trip, or as 38, indicating the microenvironment is inside a parking
garage at the end of a trip.
Bus
From the Microenvironment Study, a microenvironment is defined to
be a bus when the microenvironment is measured as part of the commuter
part of the Microenvironment Study, and the link number is given as
1-20, indicating this is a roadway link; the vehicle code is given as 6,
indicating the vehicle is a bus.
Rail
A microenvironment is defined to be rail with the same criterion as
was used to define bus, except that the vehicle is given as 7 instead of
6, indicating the vehicle is rail.

-------
- 58 -
Pedestrian
There are two groups of pedestrian microenvironments measured in
the Microenvironment Study. In the commuter part of the study, some
pedestrian data are collected. Those microenvironments are identified
by the same criterion that was used for bus and rail, except that the
vehicle code is 8, indicating the "vehicle" is pedestrian.
Some additional pedestrian data were collected along with the shop
microenvironments during the noncommute phase of the Microenvironment
Study. Those data are given as the Walking Survey Database and have
been combined with the other pedestrian data in the analysis.
Private Car
The majority of the microenvironments measured in the commuter part
of the Microenvironment Study are private cars. Those microenvironments
are identified with the same criterion that was used for the other
roadway METs, except that the vehicle codes are given as VI through V5,
indicating the microenvironment is one of the five survey vehicles.
Shops
The shop microenvironments are measured in the noncommute part of
the Microenvironment Study. The data are given as the Shopping Center
Survey Database.
Offices
The office microenvironments are measured in the noncommute phase
of the Microenvironment Study. The data are given as the Office
Building Database. The analysis is restricted to the measurements taken
between 7 a.m. and 6 p.m., when the majority of the population spend
time in this MET.
Other
There are no MEM data from the Microenvironment Study for the
microenvironments belonging to the MET other.

-------
- 59 -
DEFINITIONS OF METS FROM THE URBAN SCALE STUDY
Parking
An activity segment is defined to be parking if the location of
activity is given in Hartwell et al. (1984a and b) as 0661, indicating
the location is indoors-garage.
Bus
An activity segment is defined to be bus if in Hartwell et al.
(1984 a and b) the activity code is given as 1, indicating the activity
is "transit, travel;" the location of activity is given as 0100,
indicating the location is "in transit;" and the mode of travel is given
as 0300, indicating the trip was made in a bus.
Rail
An activity segment is defined to be rail using the same criteria
as for bus except that the mode of travel should be 0500, indicating the
trip was made in "train/subway."
Car
An activity segment is defined to be car using the same criterion
for bus except that the mode of travel should be 0200, 0400, 0663, or
0664. Mode of travel 0200 indicates the trip is made in a car. Mode of
travel 0400 indicates the trip is made in a truck. Mode of travel 0663
indicates the trip was made on a motorcycle. Mode of travel 0664
indicates the trip was made in a van.
Pedestrian (Noncar)
An activity segment is defined to be noncar, corresponding to the
elementary MET pedestrian in Sec. Ill, using the same criterion for bus,
except that the mode of travel should be 0100, 0661, or 0662. The mode
0100 indicates walking, the mode 0661 indicates jogging, and the mode
0662 indicates biking.

-------
- 60 -
Shops
An activity segment is defined to belong to the MET shops if the
location code is 0400 or 0664. Location 0400 indicates the activity
segment was stores. Location 0664 indicates the activity segment was
shopping malls/theater in malls.
Offices
An activity segment is defined to belong to the MET offices in
Hartwell et al. (1984a and b) if the location code is given as 0300,
which indicates the location is an office.
Other
All other activity segments are assigned to the MET other.

-------
- 61 -
Appendix B
QUALITY CRITERION FOR THE ACTIVITY TIME DATA
This appendix presents the detailed definitions of the two quality
criteria for activity time data--namely, the skip-logic criterion and
the consistency criterion.
SKIP-LOGIC CRITERION
The activity diary requires the participant to respond differently
to two questions depending on whether the current activity segment is
indoor or outdoor (the latter includes in-transit activity). If the
current activity is indoor, the participant should give a valid response
to the following questions:
•	Is there a garage attached to the building?
•	Is there a gas stove in use?
For both questions, "don't know" is a valid response. If the
participant did not give a valid response to those questions (indicated
by a missing value code in the database), he is judged to have made an
error in the skip logic.
If the current activity is outdoor (including in-transit), the
participant should skip those two questions. If the participant fails
to do so and responds "yes," "no," or "don't know" to those questions,
he is determined to have made an error in the skip logic. (The correct
response to these questions for those activity segments is no response,
indicated by a missing value code in the database.)
Table B.l gives the frequency distribution of the error rate in
following the skip logic for the whole sample of 705 participants. As
was noted earlier, 361 participants made no errors in the skip logic.
Those participants constitute the "good" sample.

-------
- 62 -
Table B.l
FREQUENCY DISTRIBUTION OF ERROR RATES'*
IN FOLLOWING THE SKIP LOGIC IN THE"
ACTIVITY DIARIES
Density0
(Percent per
Ranged
Count
unit range)
0
361
(51.2)d
0 - 0.05
123
348.9
0.05 - 0.1
109
309.2
0.1 - 0.2 '
62
87 .9
0.2 - 0.3
25
35.5
0.3 - 0.4
8
11.3
0.4 - 0.5
6
8.5
0.5 - 0.6
4
5.7
0.6 - 0.7
4
5.7
0
1
o
00
3
4.3
Maximum
0.
,7857
^The fraction of activity segments
on which a skip logic error is found.
^The ranges include the upper and
exclude the lower endpoint.
CThe density is given as percent per
unit range — that is, the percentage in the
range divided by the width of the range.
See the discussion of density in Table 4,
note a.
^The density over this range is
infinite; the value given in parentheses
is the actual percentage of respondents who
pass this criterion.

-------
- 63 -
The skip-logic criterion is inversely related to the number of
reported activity segments, namely the more activity segments a
participant reports, the more chance he has of making an error. As can
be seen from Table B.l, some participants rejected from the good sample
made skip-logic errors on a very small fraction of activity segments.
It is plausible that the activity data based on their diaries are still
fairly reliable. However, for the sake of simplicity, the "good sample"
is restricted to the participants who made no errors at all to avoid
ambiguity.
CONSISTENCY CRITERION
During the preparation of the activity database before this study,
several inconsistencies in the reported activity data were found,
including those between reported activities and locations and those
between activity time and monitoring time. The details of those
inconsistencies were given in Hartwell et al. (1984a and b) , Section
5.4.2.2.
Such inconsistencies are interpreted as the participant's failure
to fill out the diary correctly, indicating that the activity data from
this participant might be less reliable. Therefore the consistency
checks are another criterion to test for the participant's reliability.

-------
- 64 -
Appendix C
FREQUENCY DISTRIBUTIONS OF ACTIVITY TIME
Tables C.l through C.7 give the frequency distributions for the
activity times for each of the elementary METs. Generally speaking, the
activity distributions are similar among the three samples.
The prevalence of bus and rail users is low. (See Tables C.l and
C.2.) For both bus and rail, a few participants report spending several
hours in those METs. Those participants are more frequent in the whole
sample and the good sample than in the best sample. For example, one
participant in the whole sample reported spending more than four hours
in the rail. The longest rail times for the good and the best sample
are both less than two hours. The validity of those reported long bus
and rail activities are questionable. In this case the skip-logic and
consistency criteria were successful in eliminating those questionable
participants from the best sample and the good sample.
Most people spend some time in the MET private car (see Table C.3).
The distribution of time spent in this MET is again fairly skewed, with
a few participants spending a great deal of time in this MET. The
maximum amount of time spent in this MET is 9-1/2 hrs for the best
sample, 12 hrs for the .good sample, and 15 hrs in the whole sample.
Again the quality criteria appear to have succeeded in eliminating
participants spending questionably long times in this MET.
About one-fourth to one-third of the participants spend some time
in the MET pedestrian (see Table C.4). One participant in the whole
sample reported spending close to 11 hrs in this MET. The quality
criteria again were successful in eliminating this questionable
participant from the good and the best sample. The maximum time in this
MET is about four hours for the good sample and about two hours for the
best sample.
Only a few participants reported spending some time in the MET
parking (see Table C.5). A few participants reported spending over ten
hours in this MET.

-------
- 65 -
About half of the participants report spending time in the MET
office (see Table C.6). One participant in the whole sample reported
spending over 18 hours in this MET. For all three samples the
distribution of time spent in this MET appears bimodal, with one cluster
of participants spending a small amount of time (a few hours or less
than one hour) and another larger cluster of participants who spend
about eight hours. Comparison of the distributions among the three
samples indicates that the distribution in the best sample is clustered
around eight hours more than in the other two samples. There are also a
few reported workaholics spending over 12 hours in this MET; again more
of these unusual participants are in the whole sample than in the good
sample and the best sample.
About one-third to one-fourth of the participants spend some time
in the MET shops (see Table C.7). Most of those are fairly short, but a
few participants report up to eight or nine hours in this MET.
The activity times have been standardized as was discussed in
Sec. II.

-------
- 66 -
Table C.l
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET BUS

Whole Sample
Good
Sample
Best
Sample


Dens ityb

Dens ity

Density


(Percent

(Percent

(Percent
Range3
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
661
(93.8)c
335
(92.8)c
121
(95.3)c
0 - 0.25
2
1.1
1
1.1
0
0.0
0.25 - 0.5
11
6.2
5
5.5
0
0.0
0.5 - 0.75
14
7.9
10
11.1
2
6.3
0.75 - 1
2
1.1
1
1.1
1
3.1
1 - 1.5
8
2.3
5
2.8
2
3.1
1.5 - 2
4
1.1
2
1.1
0
0.0
2-3
1
0. 1
1
0.3
1
0.8
3 - 4
' 1
0. 1
1
0.3
0
0.0
4 - 5
0
" "o.o
0 ~
0.0
o"
0.0
5 - 6
1
0.1
0
0.0
0
0.0
Maximum
5 .
139
3.
262
2.
163
aThe ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table	4, note a.
c
The density for zero hour is infinity.	The value given
is the percentage for participants with no time in this MET.

-------
Table C.2
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET RAIL

Whole Sample
Good
Sample
Best
Sample


Dens ityb

Density

Density


(Percent

(Percent

(Percent
Range3
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
670
(95.0)c
341
(94.5)C
124
(9 7 . 6)c
"0 - 0.25
3
1.7
2
2.2
1
3.1
0.25 - 0.5
12
6.8
6
6.6
1
3.1
0.5 - 0.75
6
3.4
3
3.3
0
0.0
0.75 - 1
4
2.3
2
2.2
0
0.0
1 - 1.5
8
2.3
7
3.9
1
1.6
1.5 - 2
1
0.3
0
0.0
0
0.0
2 - 3
0
0.0
0
0.0
0
0.0
3-4
0
0.0
0
0.0
0
O'TO
4-5
1
0.1
0
0.0
0
0.0
Maximum
4.
.287
1.
478
1.
387
aThe ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table 4, note a.
c
The density for zero hour is infinity. The value given
is the percentage for participants with no time in this MET.

-------
Table C.3
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET PRIVATE CAR

Whole Sample
Good Sample
Best
Sample


Dens ityb

Dens ity

Density


(Percent

(Percent

(Percent
Range3
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
100
(14.2)c
56
(15.5)c
22
(17 .3)C
0 - 0.25
16
9.1
8
8.9
2
6.3
0.25 - 0.5
39
11.1
22
24.4
7
22.0
0.5 - 0.75
55
31.2
31
34.3
8
25 .2
0.75 - 1
68
38.6
32
35.5
11
34.6
1 - 1.5
139
39.4
65
36.0
18
28.3
1.5 - 2
93
26.4
49
27. 1
21
33. 1
2-3
103
14.6
53
14.7
24
18.9
3 - 4
47
6.7
23
6.4
7
5.5
4 - 5
16
2.3
10
2.8
4
3.1
5 - 6
10
1.4
4
1.1
0
0.0
6 - 7
6
0.9
1
0.3
0
0.0
7 - 8
3
0.4
1
0.3
0
0.0
8 - 9
4
0.6
3
0.8
2
1.6
9-10
2
0.3
1
0.3
1
0.8
10 - 11
1
0.1
0
0.0
0
0.0
11 - 12
2
0.3
2
0.6
0
0.0
12 - 13
0
0.0
0
0.0
0
0.0
13 - 14
0
0.0
0
0.0
0
0.0
14 - 15
0
... „
0.0
0
0.0
0
0.0
15 - 16

0. 1 "
0
0.0
0
0.0
Maximum
15.
.252
11
.991
9.
523
ihe ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table 4, note a.
The density for zero hour is infinity. The value given
is the percentage for participants with no time in this MET.

-------
_Z ^1.- .
Table C.4
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET PEDESTRIAN
Whole Sample Good Sample	Best Sample
Densityb	Density	Density
(Percent	(Percent	(Percent
Range3	Count per hr) Count per hr) Count per hr)
0 hrs
480
(68.1)C
255
(70.6)c
95
( 7 4 . 8 ) c
•0 - 0.25
47
26. 7
23
25.5
8
25.2
0.25 - 0.5
57
32.3
26
28.8
6
18.9
0.5 - 0.75
34
19.3
15
16.6
3
9.4
0.75 - 1
29
16.5
15
16.6
4
12.6
1 - 1.5
27
7.7
15
8.3
8
12.6
1.5 - 2
13
3.7
6
3.3
2
3.1
2 - 3
10
1.4
3
0.8
1
0.8
3 - 4
5
0.7
3
0.8
0
0.0
4-5
2
0.3
0
0.0
0
0.0
5 - 6
0
0.0
0
0.0
0
0.0
6-7
0
0.0
0
0.0
0
0.0
7 - 8
0
0.0
0
0.0
0
0.0
8 - 9
0
0.0
0
0.0
0
0.0
9 - 10
0
0.0
0
0.0
0
0.0
10 - 11
1
0. 1
0
0.0
0
"0.0
Maximum	10.935	3.867	2.362
^he ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table 4, note a.
CThe density for zero hour is infinity. The value given
is the percentage for participants with no time in this MET.

-------
- 70 -
Table C.5
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET PARKING

Whole Sample
Good Sample
Best
Sample


Dens ityb

Dens ity

Density


(Percent

(Percent

(Percent
Range3
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
646
(91.6)c"
338
(93.6)c
122
(96. l)c
0 - 0.25
34
19.3
15
16.6
3
9.4
0.25 - 0.5
10
5.7
4
4.4
0
0.0
0.5 - 0.75
3
1.7
2
2.2
1
3.1
0.75 - 1
3
1.7
0
0.0
0
0.0
1 - 1.5
2
0.6
0
0.0
0
0.0
1.5 - 2
2
0.6
0
0.0
0
0.0
2 - 3
0
0.0
0
0.0
0
0.0
3 - 4
1
0.1
0
0.0
0
0.0
4 - 5
1
0.1
0
0.0
0
0.0
5 - 6
0
0.0
0
0.0
0
0.0
6 - 7
0
0.0
0
0.0
0
0.0
7 - 8
0
0.0
0
0.0
0
0.0
8 - 9
0
0.0
0
0.0
0
0.0
9 - 10
1
0.1
0
0.0
0
0.0
10 - 11
1
0.1
1
0.3
1
0.8
11 - 12
0
0.0
0
0.0
0
0.0
12 - 13
1
0.1
1
0.3
0
0.0
Maximum
12.
.689
12.
.689
10.
.417
The ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table 4, note a.
c
The density for zero hour is infinity. The value given
is the percentage for participants with no time in this MET.

-------
Table C.6
FREQUENCY DISTRIBUTION FOR ACTIVITY TIME FOR THE MET OFFICE

Whole Sample
Good Sample
Best
Sample


Density^

Density

Dens ity


(Percent

(Percent

(Percent
Range3
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
356
(50.5)c
188
(52. l)c
64
(50.4)c
0 - 1
50
7 . 1
21
5.8
6
4.7
1 - 2
19
2.7
5
1.4
2
1.6
2 - 3
20
2.8
7
1.9
2
1.6
3-4
9
1.3
2
0.6
0
0.0
4 - 5
15
2.1
6
1. 7
3
2.4
5 - 6
13
1.8
9
2.5
0
0.0
6 - 7
34
4.8
20
5.5
6
4.7
7-8
5 7
8.1
27
7.5
10
7.9
8 - 9
63
8.9
36
10.0
16
12.6
9-10
45
6.4
28
7.8
11
8.7
10 - 11
15
2.1
8
2.2
5
3.9
11 - 12
5
0.7
2
0.6
2
1.6
12 - 13
2
0.3
1
0.3
0
0.0
13 - 14
1
0.1
1
0.3
0
0.0
14 -
1

0
0.0
0 "
0.0
Maximum
18.
.339
13.
145
11,
.556
£
The ranges include the upper and exclude the lower endpoint.
^See the discussion on density in Table 4, note a.
Q
The density for zero hour is infinity. The value given
is the percentage for participants with no time in this MET.

-------
- 72 -
Table C.7
FREQUENCY DISTRIBUTION" FOR ACTIVITY TIME FOR THE MET SHOPS

Whole Sample
Good
Sample
Best
Sample


Density'3

Density

Density


(Percent

(Percent

(Percent
Ranged
Count
per hr)
Count
per hr)
Count
per hr)
0 hrs
471
(66.8)c
260
(72.0)c
94
(74.0 )c
0 - 0.25
48
27.2
25
27.7
8
25.2
0.25 - 0.5
45
29.0
19
21.1
4
12.6
0.5 - 0.75
¦ 29
16.5
15
16.6
5
15 . 7
0.75 - 1
40
22.7
15
16.6
8
25.2
1 - 1.5
30
8.5
11
6.1
4
6.3
1.5 - 2
11
3. 1
2
1. 1
0
0.0
2 - 3
13
1.8
4
1.1
1
0.8
3 - 4
4
0.6
2
0.6
0
0.0
4-5
4
0.6
3
0.8
1
0.8
5 - 6
3 ,
0.4
1
0.3
1
0.8
6 - 7
1 '
0.1
0
0.0
0
0.0
7-8
1
0. 1
0
0.0
0
0.0
8 - 9
" 4 "
0.6
4
1.1""
1
0.8 "
9-10
1
0.1
0
0.0
0
0.0
Maximum	9.789	8.547	8.121
^he ranges include the upper and exclude the lower endpoint.
^S^e the discussion on density in Table	4, note a.
c
The density for zero hour is infinity.	The value given
is the percentage for participants with no	time in this MET.

-------
- 73 -
Appendix D
SENSITIVITY ANALYSIS
This appendix compares the results given in the text across the
three samples defined in terms of the quality criteria--namely, the
whole sample of 705 available participants, the good sample of 361
participants who passed the skip-logic criterion, and the best sample of
127 participants who passed both criteria. (The results in the text
were based on the whole sample.)
ACTIVITY TIMES FOR THE ELEMENTARY METS
Table 4 gave the summary statistics for the activity time in the
various METs. Table D.l compares the activity time summaries across the
three samples. Generally speaking, there is very little difference
among the activity times across the three samples.
ESTIMATED EXPOSURES
Section IV discussed the MEM and PM exposures. Tables D.2 through
D.13 compare the summary statistics for the estimated exposures across
the three samples. Tables D.2-D.5 compare the summaries for the MEM
exposures based on the convolution method. Tables D.6-D.9 compare the
summaries for the MEM exposures based on the hybrid method. Tables
D.10-D.13 compare the summaries for the PM exposures. The estimated
exposures are similar across the three samples.
Duan's criterion and PIESS
Section V discussed Duan's criterion and PIESS. Tables D.14 and
D.15 compare those measures across the three samples. The qualitative
nature of the results is similar across the three samples.

-------
- 74 -
Table D.1
SUMMARY STATISTICS FOR STANDARDIZED ACTIVITY TIMES FOR
ELEMENTARY METS
(Time in hours)
Whole Sample Good Sample Best Sample
MET	Mean SD	Mean	SD	Mean	SD
Bus	0.060	0.319	0.065	0.293	0.051	0.258
Rail	0.040	0.241	0.043	0.207	0.017	0.132
Private car	1.517	1.591	1.485	1.609	1.503	1.475
Pedestrian	0.269	0.707	0.226	0.528	0.198	0.443
Parking	0.084	0.766	0.075	0.864	0.089	0.925
Shop	0.384	1.064	0.331	1.077	0.298	0.983
Office	3.051	3.914	3.201	3.996	3.547	4.190

-------
Table D.2
SUMMARIES FOR THE MEM EXPOSURES BASED
ON THE CONVOLUTION METHOD
Sample
Mean3
SDb
Skew0
Kurtd
Whole
2.287
2.215
9 .471
175.0
Good
2.225
2.370
10.98
205. 1
Best
2.067
2.308
11.89
208.6
Average of the MEM exposures
in ppm-days.
^Standard deviation of the MEM
exposures.
Q
Skewness of the MEM exposures.
^Kurtosis of the MEM exposures.
Table D.3
PERCENTILES OF MEM.EXPOSURES BASED ON
THE CONVOLUTION METHOD
Percentile Whole Sample Good Sample Best Sample
1
0.051
0.050
0.050
5
0.452
0.395
0.275
10
0.746
0.706
0.487
25
1.281
1.252
1.255
50
1.894
1.875
1.840
75
2. 753
2.665
2.533
90
3.902
3.671
3.459
95
5 .082
4.457
3.887
99
9.997
10.78
6.131

-------
- 76 -
Table D.4
SUMMARIES FOR LOG MEM EXPOSURES BASED
ON THE CONVOLUTION METHOD
Sample
Mean3
SDb
Skewc
Kurtd
Whole
0.555
0.815
-1.381
4.858
Good
0.516
0.835
-1.452
4.904
Best
0.432
0.868
-1.528
4.241
Average of the log MEM exposures
in log(ppm-day).
^Standard deviation of the log
MEM exposures.
Skewness of the log MEM
exposures.
^Kurtosis of the log MEM
exposures.
Table D.5
PERCENTILES OF LOG MEM EXPOSURES BASED ON THE
CONVOLUTION METHOD
Percentile Whole Sample Good Sample Best Sample
1
-2.976
-2.996
-2.996
5
-0.794
-0.929
-1.290
10
-0.293
-0.348
-0.719
25
0.247
0.225
0.227
50
0.639
0.629
0.610
75
1.013
0.980
0.929
90
1.362
1.300
1.241
95
1.626
1.495
1.358
99
2.302
2.378
1.813

-------
- 77 -
Table D.6
SUMMARIES FOR THE HYBRID EXPOSURES
Sample
Mean3
SDb
Skewc
Kurt^
Whole
2.289
1.628
9.387
114.4
Good
2.247
1.771
10.25
124.9
Best
2.286
1.856
8.808
90. 12
£
Average of the hybrid exposures.
^Standard deviation of the hybrid
exposures.
Q,
Skewness of the hybrid
exposures.
^Kurtosis of the hybrid
exposures.
Table D.7
PERCENTILES OF THE HYBRID EXPOSURES
Percentile Whole Sample Good Sample Best Sample
1
1.202 ppm
1.202
1.202
5
1.202
1.202
1.202
10
1.340
1.278
1.258
25
1.706
1.628
1.557
50
2.063
2.051
2.093
75
2.479
2.465
2.541
90
2.996
2.958
3.007
95
3.635
3.365
3.356
99
6.905
6. 151
16.86

-------
- 78 -
Table D.8
SUMMARIES FOR THE LOG
HYBRID EXPOSURES
Sample
Meana
SDb
Skewc
Kurtd
Whole
0. 740
0.361
1.806
8.709
Good
0.717
0.362
1.930
10.80
Best
0.724
0.384
1.839
9 . 796
Average of the log hybrid
exposures.
^Standard deviation of the log
hybrid exposures.
c
Skewness of the log hybrid
exposures.
^Kurtosis of the log hybrid
exposures.
Table D.9
PERCENTILES OF THE LOG HYBRID EXPOSURES
(log (ppm))
Percentile Whole Sample Good Sample Best Sample
1
0.184
0. 184
0. 184
5
0. 184
0.184
0. 184
10
0.293
0.246
0.229
25
0.534
0.487
0.443
50
0. 724
0. 718
0. 738
75
0.908
0.902
0.932
90
1.097
1.084
1. 101
95
1.291
1.213
1.211
99
1.932
1.817
2.825

-------
- 79
Table D.10
SUMMARIES FOR THE PM EXPOSURES
Sample
Mean3
SDb
Skewc
Kurtd
Whole
1.593
1.634
3.112
16.74
Good
1.514
1.572
3.047
14. 15
Best
1.411
1.395
2.906
13.40
Average of the PM exposures in
ppm-days.
^Standard deviation of the PM
exposures.
c
Skewness of the PM exposures.
^Kurtosis of the PM exposures.
Table D.11
PERCENTILES OF THE PM EXPOSURES
Percentile Whole sample Good sample Best sample
1
0.050
0.050
0.050
5
0.097
0.092
0.092
10
0.218
0.163
0. 151
25
0.577
0.505
0.423
50
1. 165
1.154
1. 155
75
2.103
2.010
1.833
90
3.295
3.106
2.968
95
4.494
3.888
3.738
99
7.541
10.14
9.318

-------
- 80 -
Table D.12
SUMMARIES OF LOG PM EXPOSURES
Sample
Mean3
SDb
Skewc
Kurtd
Whole
-0.017
1.096
-0.678
0.315
Good
-0.082
1. 118
-0.675
0. 168
Best
-0.123
1.090
-0.688
0. 101
£
Average of the log PM exposures
in log(ppm-day).
^Standard deviation of the log PM
exposures.
c
Skewness of the log PM exposures.
^Kurtosis of the log PM exposures.
Table D.13
PERCENTILES OF THE LOG PM EXPOSURES
Percentile Whole Sample Good Sample Best Sample
1
-2.996
-2.996
-2.996
5
-2.330
-2.389
-2.389
10
-1.525
-1.814
-1.889
25
-0.550
-0.683
-0.861
50
0. 153
0. 143
0. 144
75
0.743
0.698
0.606
90
1. 192
1.133
1.088
95
1.503
1.358 .
1.319
99
2.020
2. 316
2.232

-------
Table D.14
DUAN'S CRITERION FOR VARIOUS MET DECOMPOSITIONS
MET 0a
MET lb-
MET 2.b
Whole
Good
Best
A110
Commute
Business
364.6
378.4
436.2
Commute
Parking
In-transit
89 . 1
73.3
65.9
Business!0
Shops
Offices
15.8
16.5
17.6
In-transit
Vehicles
Pedestrian
12.6
10.8
13.8
Vehicles
Public
Private car
6.2
6.2
3.1
Public^
Bus
Rail
0.05
0.05
0.02
aMET 0: The coarser MET being decomposed.
^MET 1, MET 2: The two new METs.
c
The correlation between the MET concentrations in
the two finer METs cannot be estimated and is assumed
to be zero. The resulting Duan's criterion and PIESS
should be viewed as a lower bound.
^The correlation between the MET concentrations in
the two finer METs cannot be estimated and is assumed
to be one. The resulting Duan's criterion and PIESS
should be viewed as a upper bound.
Table D.15
PERCENTAGE INCREASE IN EFFECTIVE SAMPLE SIZE
MET 0*
MET lb
MET 2b
Whole
Good
Best
All0
Commute
Business
31.1
36.2
63.7
Commute
Parking
In-trans it
8.2
7.5
10.7
Business0
Shops
Offices
1.5
1.7
2.9
In-transit
Vehicles
Pedestrian
1.2
1.1
2.4
Vehicles
Public
Private car
0.6
0.7
0.5
Public1^
Bus
Rail
0.005
0.005
0.003
aMET 0: The coarser MET being decomposed.
bMET 1, MET 2: The two new METs.
c
The correlation between the MET concentrations in
the two finer METs cannot be estimated and is assumed
to be zero. The resulting Duan's criterion and PIESS
should be viewed as a lower bound.
^he correlation between the MET concentrations in
the two finer METs cannot be estimated and is assumed
to be one. The resulting Duan's criterion and PIESS
should be viewed as a upper bound.

-------
- 82 -
Appendix E
DUAN'S CRITERION
The criterion defined and used in Sec. V for decomposing a coarser
MET into two finer METs, denoted as MET^ and MET^, is shown in Duan
(1981) to have the following expression:
DC = DCt x DCC,
where DC? = (yj + y^)2 * Var (R^,
DCC = (y^ - y2C)2 + (Zn + I22 - 2 x Zi2),	(E. 1)
T
V = average MET time in MET^,
T
y 2 = average MET time in MET^,
R1 = V(T1 + V'
T^ = MET time in MET^,
T2 = MET time in MET2>
Q
y = average MET concentration in MET^,
C
y 2 = average MET concentration in MET^,
Z^ = variance of MET concentrations in MET^,
1^2 = variance of MET concentrations in MET2,
Z^2 = covariance between MET concentrations in MET^ and MET2.
This expression was used in the calculation of Duan's criterion in the
analysis reported in Sec. V.

-------
- 83 -
The factor DC^, measures how different the concentrations can be
between the two METs. The factor can also be expressed equivalently as
DCC = E(C1 - C2)2,	(E.2)
where C^, are the MET concentrations in the two METs. Empirically,
the expanded formulation given in (E.l) is preferable than the more
concise formula (E.2) because the two MET concentrations might not be
observed simultaneously in microenvironment monitoring, as was discussed
in Sec. V. With formula (E.2), the factor cannot be estimated in such
situations. With formula (E.l), it is still possible to estimate part
of the formula--namely, the terms other than	t*ie term ^12'
requires the two MET concentrations to be simultaneously observed, can
be bounded, say, between zero and 0^ x o^, where 0^ and 0^ are the
standard deviations of the MET concentrations in the two METs. (This is
equivalent to bounding the correlation between zero and one. It is
assumed that the correlation cannot be negative.)

-------
- 84 -
Appendix F
ESTIMATION OF PARAMETERS IN DUAN'S CRITERION
The estimation of Duan's criterion requires the estimation of
several summary parameters for activity times and MET concentrations.
ACTIVITY TIME
For activity times, two parameters are needed: the average of the
total amount of time spent in the two METs by the individuals in the
sample, and the variance of the ratios
Ri til/(til + ti2)'
where t is the amount of time the ith person spent in the first MET,
and t is the amount of time the person spent in the second MET.
For each person, the activity times t. „ and t.^ are calculated by
ll	i2
summing the durations from activity segments associated .with the MET
being considered. The average of the total amount of time is calculated
directly. For the other parameter, the variance of the ratios R^, the
ratio may be calculated only for individuals who spent some time in at
least one of the two METs--i.e., t., + t.„ > 0. The ratios and the
11 i2
variance are calculated using those people only. This procedure
implicitly assumes that if the people who did not spend any time in
these two METs on the sampling day spend some time in these two METs on
another day, the variance of their ratios R is the same as the one based
on observed ratios. This assumption cannot be tested using the data
available from this study. However, if there are multiple days of
activity times from the same individual, such as in the Denver CO Study
(Johnson, 1984), it will be possible to test this assumption.
MET CONCENTRATIONS
For MET concentrations, the following parameters are needed:
Vi^ = average MET concentration in the first MET,
^2 = average MET concentration in the second MET,

-------
- 85 -
I = variance of the MET concentrations in the first NET,
^22 = variance of the MET concentrations in the second MET,
p^2 = correlation between the MET concentrations in the two METs.
If both METs are elementary as was defined in Sec. V, these
parameters can estimated directly. The average and variance parameters
are estimated from days on which microenvironment monitoring was
conducted in these elementary METs. The estimates are the same as the
summary statistics given in Table 7. This procedure implicitly assumes
that the MET concentrations are second-order stationary--i.e., they have
the same mean and variance on different days. The correlation is
estimated from the days on which microenvironment monitoring was
conducted in both METs. This procedure also assumes second-order
stationarity. Certain pairs of METs--e.g., bus and rail--were never
measured on the same day for more than one day; therefore the
correlation cannot be estimated. In those situations both the
uncorrelated case (p = 0) and the perfectly correlated case (p = 1) are
considered and Duan's criterion derived for both cases, the former as an
upper bound and the latter as a lower bound.
When one or both of the METs is composite and not elementary--
i.e., is an aggregation of more than one elementary METs--the MET
concentration must be defined on each day before estimating the
parameters. For example, the MET vehicle consists of three elementary
METs: private cars, buses, and rails. There are several ways the MET
concentrations for "vehicle" can be estimated, which will be discussed
below.
Equation (5) expressed MET concentration as the average of
microenvironment concentrations belonging to this MET, weighted by the
amount of time the person spent in each microenvironment. With some
algebraic manipulation, this definition leads to an analogous expression
for MET concentration in a composite MET in terms of the MET
concentrations in its component elementary METs. The MET concentration
for a composite MET is the average of MET concentrations in the

-------
- 86 -
component elementary METs, weighted by the amount of time the individual
spent in each elementary MET:
C., = I.C. . x T. ,/Z.T. . ,	(F. 1)
ik j ij ij j ij
where is the MET concentration in the kth MET, a composite MET,
consisting of the elementary METs {J}, the summations are summed over
these elementary METs; for each elementary MET, say the jth, C is the
MET concentration and T.. is the time spent in that MET.
ij
One way to estimate the MET concentration for a composite MET such
as vehicle is to take the average of all microenvironment concentra-
tions measured as part of this MET. Under this definition, the MET concen-
trations in the elementary METs are weighted by the amount of time the
investigators spent in each elementary MET, instead of the amount of time
the general population spent. For example, the investigators might
spend half of their time in private cars and half of their time in public
transportation. If they average the microenvironment concentrations
they measured as their estimate for the composite MET vehicle, the
estimate is equally weighted by private cars and public transportation.
However, the general population might not allocate their time between
private cars and public transportation in the same way.
When there are no activity data from a human population available,
the direct averaging of microenvironments described above might be the
only method feasible. During the earlier phases of this analysis, this
method was used to estimate MET concentrations. The results are
basically the same as those based on the more accurate method.
For the present study, the activity data are available from the
Urban Scale Study, which can be used to give better estimates of MET
concentrations for composite METs. The final analysis uses the average
amount of time the Urban Scale Study sample (the best sample) spent in
each elementary MET as weights when estimating the MET concentration for
a composite MET, using the formula (F.l) given above. All the results
given in the text of this report are based on this method.
Once the MET concentrations are estimated for the composite METs,
the concentration parameters are estimated in the same manner as for
elementary METs.

-------
- 87 -
REFERENCES
Dempster, A. P., N. M. Laird, and D. B. Rubin, "Maximum Likelihood from
Incomplete Data via the EM Algorithm (with Discussion)," Journal of
the Royal Statistical Society, Vol. 39, No. 1 (Series B) , 1977, pp.
1-38.
Duan, Naihua, "Micro-Environment Types: A Model for Human Exposure to
Air Pollution," Technical Report No. 47, SIMS, Stanford University,
Stanford, California, May 1981.
	, Models for Human Exposure to Air Pollution, The Rand Corporation,
N-1884-HHS/RC, July 1982. Also published in Environment
International, Vol. 8, 1982, pp. 305-309.
Flachsbart, Peter G., "Development of a Field Methodology and Model to
Estimate Commuter Exposures to Air Pollution Using Personal Exposure
Monitors," Progress Report No. 1, Environmental Monitoring Systems
Laboratory, Research Triangle Park, North Carolina, September 30,
1982a.
	, "Development of a Field Methodology and Model to Estimate
Commuter Exposures to Air Pollution Using Personal Exposure Monitors,"
Progress Report No. 2, Environmental Monitoring Systems Laboratory,
Research Triangle Park, North Carolina, February 17, 1984.
	, "Field Survey Procedures for Monitoring Carbon Monoxide Exposures
of Commuters in the Washington Metropolitan Area," Environmental
Monitoring Systems Laboratory, Research Triangle Park, North Carolina,
September 1982b.
Hartwell, T. D., C. A. Clayton, D. M. Michie, R. W. Whitmore, H. S.
Zelon, S. M. Jones, and D. A. Whitehurst, "Study of Carbon Monoxide
Exposure of Residents of Washington, D.C. and Denver, Colorado, Part
I," RTI/2390/00-OIF, Research Triangle Institute, Research Triangle
Park, North Carolina, January 1984a.
	, "Study of Carbon Monoxide Exposure of Residents of Washington,
D.C. and Denver, Colorado, Part II: Appendices," RTI/2390/00-OIF,
Research Triangle Institute, Research Triangle Park, North Carolina,
January 1984b.
Johnson, Ted, "A Study of Personal Exposure to Carbon Monoxide in
Denver, Colorado," Environmental Monitoring Systems Laboratory,
Research Triangle Park, North Carolina, January 1984.
Mack, Gregory A., Peter A. Flachsbart, Charles Rodes, James E. Howes,
Gerry G. Akland, Harold Sauls, "Microenvironmental Approach to
Estimation of Carbon Monoxide Exposures to Commuters," 1984. (To
appear in the Journal of the Air Pollution Control Association.)

-------
- 88 -
Massey, Frank J., Jr., "The Distribution of the Maximum Deviation
between Two Sample Cumulative Step Functions," Annals of Mathematical
Statistics, Vol. 22, 1951, 125-128.
Ott, Wayne R., "Computer Simulation of Human Air Pollution Exposures to
Carbon Monoxide," paper presented at the 74th Annual Meeting of the
Air Pollution Control Association, Philadelphia, Pennsylvania, June
21-26, 1981.
Smirnov, N., "On the Estimation of the Discrepancy between Empirical
Curves of Distributions of Two Independent Samples," Bulletin
Mathematique de l'Universite de Moscou, Vol. 2 (1939), fasc.2.
Wallace, Lance A., Jacob Thomas, and David T. Mage, "Comparison of COHb
Predicted from CO Exposures with End-tidal Breath Estimates of COHb
from the Denver-Washington, D.C., Human Exposure Studies," paper
presented at the 77th Annual Meeting of the Air Pollution Control
Association, San Francisco, California, June 24-29, 1984.
Whitmore, Roy, Sheldon M. Jones, and Martin S. Rosenzweig, "Final
Sampling Report for the Study of Personal CO Exposure,"
RTI/2390/02-OIF, Research Triangle Institute, Research Triangle Park,
North Carolina, January 1984.

-------
j TECHNICAL REPORT DATA
(Please read Instructions on the reverse before completing)
1. REPORT NO. 2.
EPA/600/4-85/046
3. RECIPIENT'S ACCESSION NO.
PBS5 2 289 55'rR
4. TITLE AND SUBTITLE
APPLICATION OF THE MICROENVIRONMENTAL MONITORING
APPROACH TO ASSESS HUMAN EXPOSURE TO CARBON
MONOXIDE
5. REPORT DATE
Julv 1985
6. PERFORMING ORGANIZATION CODE
7. AUTHOR(S)
Naihua Duan
8. PERFORMING ORGANIZATION REPORT NO.
9. PERFORMING ORGANIZATION NAME AND ADDRESS
Rand Corporation
1700 Main Street
Santa Monica, CA 90406
10. PROGRAM ELEMENT NO.
11. CONTRACT/Oe^mOXXOC-
68-02-4058
12. SPONSORING AGENCY NAME AND ADDRESS
Environmental Monitoring Systems Laboratory
Office of Research and Development
U.S. Environmental Protection Agency
Research Triangle Park, NC 27711
13. TYPE OF REPORT AND PERIOD COVERED
14. SPONSORING AGENCY CODE
EPA/600/08
15. SUPPLEMENTARY NOTES
16. ABSTRACT
Exposure estimates based on monitoring carbon monoxide in microenvironment6
are compared to exposure estimates based on personal monitoring with individual,
portable monitors. Methods of calculation are reviewed and discussed, and re-
sults of calculations are presented. These data indicate that population expo-
sure estimates based on data from the Washington Microenvironment Study, com-
bined with people's activity data from the Washington Urban Scale Study, are
I about forty percent higher than estimates based on personal monitoring data
from the Urban Scale Study. The former set of exposure estimates is found to
be a good predictor of the latter. Nevertheless, generalizations of these
findings to other data bases are not valid at this time.
17. KEY WORDS AND DOCUMENT ANALYSIS
a. DESCRIPTORS
b.IDENTIFIERS/OPEN ENDED TERMS
c. COSATl Field/Group



18. DISTRIBUTION STATEMENT
RELEASE TO PUBLIC
19. SECURITY CLASS (This Report)
UNCLASSIFIED
21 . NO. OF PAGES
99
20. SECURITY CLASS (This page)
UNCLASSIFIED
22. PRICE
$ U.s~b
EPA Form 2220-1 (Rev. 4-77) previous edition is obsolete
i

-------