  SEQ CHAPTER \h \r 1 

UNITED STATES ENVIRONMENTAL PROTECTION AGENCY

WASHINGTON, D.C.  20460

	

April 02, 2007

MEMORANDUM

SUBJECT:  	Transmittal of Meeting Minutes of the FIFRA Scientific
Advisory Panel Meeting Held January 9 - 12, 2007 on the Review of Worker
Exposure Assessment Methods.

TO:		Anne Lindsay, Acting Director

Office of Pesticide Programs

FROM:	Myrta R. Christian, Designated Federal Official

FIFRA Scientific Advisory Panel

Office of Science Coordination and Policy

THRU:	Steven Knott, Executive Secretary

FIFRA Scientific Advisory Panel

Office of Science Coordination and Policy

Clifford J. Gabriel, Ph.D., Director	

Office of Science Coordination and Policy

Attached, please find the meeting minutes of the FIFRA Scientific
Advisory Panel open meeting held in Arlington, Virginia on January 9 -
12, 2007.  This report addresses a set of scientific issues being
considered by the Environmental Protection Agency pertaining to the
Review of Worker Exposure Assessment Methods.

Attachment

cc:

James B. Gulliford					Frank Sanders

James J. Jones				           Betty Shackleford

William Jordan					Enesta Jones

Margie Fehrenbach					Douglas Parsons

Janet Andersen					Vanessa Vu (SAB)

Steven Bradbury					Jeff Evans

William Diamond					Jeff Dawson

Debbie Edwards					Cassi Walls

Richard Keigwin					David J. Miller

Oscar Morales					Mathew Crowley

Tina Levine						OPP Docket

Jack Housenger					

Lois Rossi						

					

						

		

FIFRA Scientific Advisory Panel Members

Steven Heeringa, Ph.D. (FIFRA SAP Chair)

John R. Bucher, Ph.D., D.A.B.T.

Janice Elaine Chambers, Ph.D., D.A.B.T.

Stuart Handwerger, M.D.

Kenneth J. Portier, Ph.D.

FQPA Science Review Board Members

  SEQ CHAPTER \h \r 1 Henry T. Appleton, Ph.D.

Dana B. Barr, Ph.D.

Brian Curwin, Ph.D.

Paul Y. Hamey, M.Sc.

Cynthia J. Hines, M.S., CIH

Brian J. Hughes, Ph.D., M.P.H., D.A.B.T.

Dallas E. Johnson, Ph.D.

David Kim, Ph.D.

Andrew J. Landers, Ph.D.

Chensheng Lu, Ph.D.

Peter D.M. Macdonald, D.Phil., P.Stat.

William J. Popendorf, Ph.D.

Mark G. Robson, Ph.D., M.P.H., A.T.S.

SAP Minutes No. 2007-03

A Set of Scientific Issues Being Considered by the

Environmental Protection Agency Regarding:

REVIEW OF WORKER EXPOSURE ASSESSMENT METHODS

JANUARY 9 - 12, 2007

FIFRA Scientific Advisory Panel Meeting,

held at the 

Environmental Protection Agency Conference Center

Arlington, Virginia

NOTICE

These meeting minutes have been written as part of the activities of the
Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA), Scientific
Advisory Panel (SAP).  The meeting minutes represent the views and
recommendations of the FIFRA SAP, not the United States Environmental
Protection Agency (Agency).  The content of the meeting minutes does not
represent information approved or disseminated by the Agency.  The
meeting minutes have not been reviewed for approval by the Agency and,
hence, the contents of these meeting minutes do not necessarily
represent the views and policies of the Agency, nor of other agencies in
the Executive Branch of the Federal government, nor does mention of
trade names or commercial products constitute a recommendation for use.

The FIFRA SAP is a Federal advisory committee operating in accordance
with the Federal Advisory Committee Act and established under the
provisions of FIFRA as amended by the Food Quality Protection Act (FQPA)
of 1996.  The FIFRA SAP provides advice, information, and
recommendations to the Agency Administrator on pesticides and
pesticide-related issues regarding the impact of regulatory actions on
health and the environment.  The Panel serves as the primary scientific
peer review mechanism of the EPA, Office of Pesticide Programs (OPP),
and is structured to provide balanced expert assessment of pesticide and
pesticide-related matters facing the Agency.  Food Quality Protection
Act Science Review Board members serve the FIFRA SAP on an ad hoc basis
to assist in reviews conducted by the FIFRA SAP.  Further information
about FIFRA SAP reports and activities can be obtained from its website
at   HYPERLINK "http://www.epa.gov/scipoly/sap/" 
http://www.epa.gov/scipoly/sap/   or the OPP Docket at (703) 305-5805. 
Interested persons are invited to contact Myrta R. Christian, SAP
Designated Federal Official, via e-mail at christian.myrta@epa.gov.

In preparing the meeting minutes, the Panel carefully considered all
information provided and presented by EPA, Health Canada, California
EPA, the Agricultural Handler Exposure Task Force (AHETF) and the
Antimicrobial Exposure Assessment Task Force II (AEATF II), as well as
information presented by public commenters.  This document addresses the
information provided and presented by these groups within the structure
of the charge.

 TABLE OF CONTENT

  TOC \o "1-2" \h \z \u    HYPERLINK \l "_Toc162076219"  PARTICIPANTS	 
PAGEREF _Toc162076219 \h  2  

  HYPERLINK \l "_Toc162076220"  INTRODUCTION	  PAGEREF _Toc162076220 \h 
4  

  HYPERLINK \l "_Toc162076221"  PUBLIC COMMENTERS	  PAGEREF
_Toc162076221 \h  6  

  HYPERLINK \l "_Toc162076225"  SUMMARY OF PANEL DISCUSSION AND
RECOMMENDATIONS	  PAGEREF _Toc162076225 \h  7  

  HYPERLINK \l "_Toc162076226"  PANEL DISCUSSION AND RECOMMENDATIONS	 
PAGEREF _Toc162076226 \h  12  

CHARGE   HYPERLINK \l "_Toc162076227"  1:  Data Needs	  PAGEREF
_Toc162076227 \h  12  

CHARGE   HYPERLINK \l "_Toc162076228"  2:  Passive Dosimetry	  PAGEREF
_Toc162076228 \h  18  

CHARGE   HYPERLINK \l "_Toc162076229"  3:  Passive Dosimetry vs.
Biomonitoring	  PAGEREF _Toc162076229 \h  25  

CHARGE   HYPERLINK \l "_Toc162076230"  4:  Normalization of Exposure by
Amount of Active Ingredient Handled (AaiH)	  PAGEREF _Toc162076230 \h 
29  

CHARGE   HYPERLINK \l "_Toc162076231"  5:  Within-worker and
Between-worker Variability	  PAGEREF _Toc162076231 \h  34  

CHARGE   HYPERLINK \l "_Toc162076232"  6:  Sample Size: Number of Sites
and Subjects per Scenario/Activity	  PAGEREF _Toc162076232 \h  39  

  HYPERLINK \l "_Toc162076233"  REFERENCES	  PAGEREF _Toc162076233 \h 
44  

  HYPERLINK \l "_Toc162076234"  APPENDICES	  PAGEREF _Toc162076234 \h 
45  

  HYPERLINK \l "_Toc162076235"  APPENDIX A:  Definitions and
Abbreviations	  PAGEREF _Toc162076235 \h  45  

  HYPERLINK \l "_Toc162076236"  APPENDIX B:  Expanding the Concept of
Grading Data	  PAGEREF _Toc162076236 \h  48  

  HYPERLINK \l "_Toc162076237"  APPENDIX C:  A Critique of the AHETF
Study Design	  PAGEREF _Toc162076237 \h  51  

 

SAP Minutes No. 2007-03

A Set of Scientific Issues Being Considered by the

Environmental Protection Agency Regarding:

REVIEW OF WORKER EXPOSURE ASSESSMENT METHODS

JANUARY 9 - 12, 2007

FIFRA Scientific Advisory Panel Meeting,

held at the 

Environmental Protection Agency Conference Center

Arlington, Virginia

Steven G. Heeringa, Ph.D.				Myrta R. Christian, M.S

FIFRA SAP Chair                              		Designated Federal
Official

FIFRA Scientific Advisory Panel        	        	FIFRA Scientific
Advisory Panel

Date:  	April 02, 2007				Date:  April 02, 2007

Federal Insecticide, Fungicide, and Rodenticide Act

Scientific Advisory Panel Meeting

January 9 - 12, 2007

REVIEW OF WORKER EXPOSURE ASSESSMENT METHODS

PARTICIPANTS

FIFRA SAP Chair

Steven G. Heeringa, Ph.D., Research Scientist & Director for Statistical
Design, University of Michigan, Institute for Social Research, Ann
Arbor, MI

Designated Federal Official

Myrta R. Christian, M.S., FIFRA Scientific Advisory Panel, Office of
Science Coordination and Policy, EPA

FIFRA Scientific Advisory Panel Members

John R. Bucher, Ph.D., D.A.B.T., Deputy Director, Environmental
Toxicology Program, NIEHS, Research Triangle Park, NC

Janice E. Chambers, Ph.D., D.A.B.T., William L. Giles Distinguished
Professor & Director, Center for Environmental Health Sciences, College
of Veterinary Medicine, Mississippi State University, Mississippi State,
MS

Stuart Handwerger, M.D., Professor of Pediatrics, University of
Cincinnati Children's Hospital Medical Center, Cincinnati, OH

Kenneth M. Portier, Ph.D., Program Director, Statistics, American Cancer
Society, Statistics and Evaluation Center, Atlanta, GA

FQPA Science Review Board Members

  SEQ CHAPTER \h \r 1 Henry T. Appleton, Ph.D., National Toxicologist,
U.S. Forest Service, Arlington, VA

Dana B. Barr, Ph.D., Chief, Pesticide Laboratory, Centers for Disease
Control and Prevention,

National Center for Environmental Health, Atlanta, GA

Brian Curwin, Ph.D., Associate Research Fellow, National Institute of
Occupational Safety and Health, Cincinnati, OH

Paul Y. Hamey, M.Sc., Principal Scientist, Human Exposure Assessment,
Pesticides Safety Directorate, York, United Kingdom

Cynthia J. Hines, M.S., C.I.H., Senior Research Industrial Hygienist,
National Institute of Occupational Safety and Health, Cincinnati, OH

Brian J. Hughes, Ph.D., M.P.H., D.A.B.T., Toxicologist, Michigan
Department of Agriculture

Pesticide and Plant Pest Management Division, Lansing, MI

Dallas E. Johnson, Ph.D., Professor Emeritus, Department of Statistics,
Kansas State University, Manhattan, KS

David Kim, Ph.D., Research Fellow, Harvard University, Department of
Environmental Health, Harvard School of Public Health, Boston, MA

Andrew J. Landers, Ph.D., Director, Application Technology Group and
Associate Professor

Cornell University, Department of Entomology, Geneva, NY

Chensheng Lu, Ph.D., Assistant Professor, Emory University, Department
of Environmental and Occupational Health, Rollins School of Public
Health, Atlanta, GA

Peter D.M. Macdonald, D.Phil., P.Stat., Professor of Mathematics and
Statistics, McMaster University, Department of Mathematics and
Statistics, Hamilton, Ontario, Canada

William J. Popendorf, Ph.D., Professor, Utah State University,
Department of Biology, Logan, UT

Mark G. Robson, Ph.D., M.P.H., A.T.S., Director of the New Jersey
Agricultural Experiment Station and Professor of Entomology, Rutgers
University, New Brunswick, NJ

INTRODUCTION

The Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA),
Scientific Advisory Panel (SAP) has completed its review of the Worker
Exposure Assessment Methods.  Advance notice of the meeting was
published in the Federal Register on October 27, 2006.  The review was
conducted in an open Panel meeting held in Arlington, Virginia, from
January 9 – 12, 2007.  Dr. Steven G. Heeringa chaired the meeting. 
Myrta R. Christian served as the Designated Federal Official.

  SEQ CHAPTER \h \r 1 The FIFRA SAP met to consider and review the
Worker Exposure Assessment Methods.  The Agency issued its first
occupational exposure testing guidelines in the early 1980s.  These
guidelines were intended to standardize the methodology used to conduct
the studies necessary to allow the Agency to determine the potential
exposures and consequent risks associated with the activities
surrounding the use of pesticides.  These activities include handling
pesticides (i.e. mixing, loading and applying) as well as working in
treated sites following pesticide applications (e.g., harvesting,
thinning, weeding; servicing cooling towers).  In the early 1990s, two
databases--the Pesticide Handlers Exposure Database (PHED) and the
Chemical Manufacturers Association (CMA) database were constructed in
order to estimate exposures resulting from mixing/loading/applying
pesticides.  The data assembled for use in these databases were taken
from published literature as well as from industry studies submitted to
the Agency.  These databases have been used as the main sources for
estimating occupational exposures to workers handling pesticides for
both registration and reregistration actions.  Since the early 1980s,
the Agency has been using a scenario-based approach in its assessments
for estimating exposures for occupational pesticide handlers (e.g.,
mixers, loaders, and applicators).  This approach is consistent with the
Agency's guidelines for exposure assessment which can be found on the
EPA website at   HYPERLINK
"http://cfpub.epa.gov/ncea/cfm/recordisplay.cfm?deid=15263" 
http://cfpub.epa.gov/ncea/cfm/recordisplay.cfm?deid=15263 .

Over the years since the issuance of the exposure guidelines, scientific
issues have been raised about the accuracy of exposure estimates based
on data developed using these methods.  In addition, recent protocols
for the generation of new agricultural pesticide handler exposure data
are being generated by a pesticide industry task force and were reviewed
by the Agency's Human Studies Review Board (HSRB) (see   HYPERLINK
"http://www.epa.gov/osa/hsrb/files/june2006finaldraftreport82806.pdf" 
http://www.epa.gov/osa/hsrb/files/june2006finaldraftreport82806.pdf  
for further information).  The board raised questions concerning the
scientific merits of the proposed protocols.

Given the scientific issues that have been raised regarding occupational
pesticide exposure estimates and study protocols, including the recent
comments from the HSRB, at this time EPA asked the FIFRA Scientific
Advisory Panel (SAP) to evaluate, in detail, issues associated with
certain methodologies used to generate exposure studies and the
procedures used to develop exposure estimates.  As part of the
background for the SAP meeting, the Agency developed a case study that
details the procedures and data the Agency uses to evaluate 6 exposure
scenarios that are common in agriculture.  These data can be found in
the existing Pesticide Handlers Exposure Database.

The Charge to the FIFRA SAP focused on the following issues: data needs
(e.g., availability of data in the Pesticide Handlers Exposure
Database); sample collection methods (e.g., whole-body dosimetry,
handwashing, facial/neck wipes, and biological monitoring [BM]); unit
exposure (e.g., relating the amount of exposure to the amount of
chemical active ingredient handled); and sample size issues (e.g.,
inter-/intra-worker variability and representativeness).

The agenda for this SAP meeting included presentations from the Health
Effects Division (HED) and Antimicrobial Division (AD) in the OPP.  In
addition, presentations were provided by Health Canada, Pest Management
Regulatory Agency; California EPA, Department of Pesticide Regulation;
Agricultural Handler Exposure Task Force (AHETF) and the Antimicrobial
Exposure Assessment Task Force II (AEATF II).

Dr. Tina Levine (Director, HED, OPP) offered opening remarks at the
meeting.

	PUBLIC COMMENTERS

Oral statements were presented as follows:

Rebeckah Adcock, on behalf of the Pesticide Policy Coalition, American
Farm Bureau Federation

Larry Olsen, Ph.D., on his own behalf

Andrew Moore, on behalf of the National Agricultural Aviation
Association

Pamela Roa, Ph.D., on behalf of Farmworker Justice

Written statements were provided by:

Richard Fenske, Ph.D., MPH.  University of Washington's School of Public
Health and Community Medicine

  SEQ CHAPTER \h \r 1 SUMMARY OF PANEL DISCUSSION AND RECOMMENDATIONS

Measuring dermal exposures is not simple, but is rather one of the most
complex tasks in the field of exposure assessment.  The Panel was
frequently complimentary toward the efforts and presentations made by
the Agency and both of the task forces in dealing with these important
and complex issues.

Because terminology is so important within the topic of dermal hazards
from chemicals, a glossary of terms was prepared and is included as
Appendix A to this Panel Report to help readers understand the
particular definition of "exposure" and its relation to "dose" used in
this context and the distinction among the terms “bias,”
“uncertainty,” and “variability.”  In exposure and risk
assessment, bias and uncertainty need to be minimized; conversely,
variability needs not to be minimized but examined.  Therefore, when
evaluating dermal exposure data, EPA should consider (1) the validity of
the sampling and analytical methods in terms of bias, (2) the magnitude
of the uncertainty inherent within the methods, and (3) the ability of
the resulting information to illuminate the variability.  These issues
are integral to virtually all of the Panel's responses.

1)  Data Needs

The Panel agreed with the Agency’s concern about the limitations of
the existing PHED exposure database.  Furthermore, they concluded that
additional data could significantly improve the Agency’s ability to
assess worker exposure.  They listed eight limitations within PHED
including its inconsistent data quality; a patch-work of methods, some
with high uncertainty and data censoring; a high level of "clustering,"
and an inadequate number of samples and diversity within some scenarios.
 Panelists also expanded on the following three broad weaknesses within
PHED.

The inclusion within PHED of studies where either not all parts of the
body were monitored or a substantial number of exposures were
undetectable do not allow the results to yield accurate exposure
statistics of interest for regulatory assessments.  New Agricultural
Handlers Exposure Database (AHED) data and software may be able to
correct these weaknesses.  The AHED study design will also include more
reliable exposure assessment methods (especially of the hands; see also
Charge #2) and newer ("modern") pesticide application equipment and
techniques (see also Charge #4).  The probable uncertainty in the
calculated exposure values creates an unrecognized weakness.  A
rationale is presented in Appendix B to expand the current concept of
“data grading” based only on the analytical method to include the
“probable uncertainty” of the calculated exposure level.  Such a
broader grading scheme could help users interpret exposure values better
as well as create a direct means for the Agency to demonstrate both the
weakness of the existing PHED and the improvements that should result
from a new database.

The Panel viewed the selection criteria proposed by AHETF and AEATF to
be reasonable for generating exposure data for using in exposure
assessments, with the following caveats.  The monitoring duration
requirement may be too stringent.  Some provision to allow the inclusion
of data from settings where only short-term uses are the norm may need
to be added.  The criteria to use biomonitoring data only if primate
dermal absorption and pharmacokinetic data exist for the chemical may
also be too restrictive.  The Panel suggested that “extrapolation”
factors appropriately estimated from rat and porcine models to humans
should be allowed.  Better justification is needed to either include or
exclude air sampling from the protocol, and the criteria for sampling an
“inspirable” aerosol needs to encompass large droplets or particles.
 Further, the array and location of patches (if they are used) should be
standardized. 

Finally, the Panel noted (here and in our response to Charge #6) the
need to generate a database that documents the frequency and the
duration with which handlers in general are exposed to pesticides in the
scenarios being considered, as a complement to the database of the
intensity of exposures of participants in the studies being contemplated
herein. 

2)  Passive Dosimetry

The response to this charge is divided into four parts that address
bias, a correction factor for handwashing, a correction for breakthrough
of dosimeters, and complementary uses of biomonitoring and dosimetry.

The Panel concluded that although a bias may exist, no bias between
dermal exposure monitoring and biological monitoring could be detected
in large part because of the statistical uncertainty inherent in the
exposure data (see also Appendix B).  The uncertainty for patch
dosimeters can be a factor of 4× to7× due to the calculation needed
to scale up from deposition onto a patch of circa 40 cm² to a body-part
area of circa 1000 cm² and the potential variability in the spatial
pattern of dermal deposition onto any given body-part.  In comparison,
the probable uncertainty when using whole-body dosimeters that cover
circa 90% of the body is likely to be no more than about 1.5×.  A
similar analysis in our response to Charge #3 will show that the
probable uncertainty of a dermal exposure value derived from
biomonitoring could range from ±20× to ±100×.  

The Panel was slightly more equivocal about a need to correct
handwashing for its efficiency at recovering pesticides from the skin. 
Existing data clearly indicate that adsorption of certain pesticides can
occur within a matter of minutes after the exposure, that hand wiping
underestimates dermal exposure more than does hand washing, and that
recovery efficiency is really not a constant.  A first-order kinetic
model of adsorption that depends upon the pesticide’s KO/W or
octanol-water partition coefficient was suggested.  Limited results
using this model indicate that the interval from initiating exposure
until washing the hands can be important when measuring pesticide
handler exposure over the planned 4 to 8-hour day.  However, others on
the Panel pointed out that the accuracies of either modeling or
experimental data could be confounded in field conditions by the effects
of repetitive (multiple) rinsing or washing that can change the skin’s
absorption rate, either enhancing or decreasing recovery from the skin. 
Also wearing chemical protective gloves will decrease the importance of
hand exposures categorically.  Biomonitoring presents an interesting but
often impractical approach to correcting handwash data due to its
inclusion of other routes of exposure and the impact of the additional
burden on the participant.  Overall, the Panel recommended that use of a
hand washing technique should be accepted in AHETF or AEATF studies if
it is supported by either laboratory data and/or a model that predicts
and can correct for its efficiency over the sampling time for the
pesticide being studied. 

The Panel concluded that generally there is no need to correct
whole-body dosimeters worn under outer garments for the breakthrough of
residues.  Patch dosimeters should have an impervious backing to prevent
breakthrough, but they may have an unacceptably large probable
uncertainty to meet the proposed uses of AHED or Biocide Handlers
Exposure Database (BHED).  The use of a whole-body dosimeter placed
directly against the skin is recommended, but no sure means of detecting
its breakthrough from liquid saturation was identified.  

The Panel also concluded that biological monitoring can be complementary
to dermal exposure monitoring to detect if not estimate the amount of
breakthrough from a whole body dosimeter (WBD).  Therefore, the Panel
generally supports an Agency proposal that biomonitoring may be included
in any sampling plan, but recommended that it not be required because of
its potential to bias participant selection.  Furthermore, the detection
of breakthrough should not be grounds to discard the sample, again to
avoid selection bias. 

3) Passive Dosimetry versus Biomonitoring

The relationship between passive dosimetry and biomonitoring depends in
part upon what form of dosimetry is being used and upon the question
being asked.  Within this caveat, the Panel concluded that passive
dosimetry can generate data that can be used to predict worker exposure
for a wide variety of scenarios and activities.

The above conclusion was supported by two analyses.  A basic analysis of
the pathways taken by the pesticide from exposure to excretion was used
to show that concurrent whole-body dosimetry and biomonitoring will
interfere with each other, precluding any expected correlation between
their data.  The second analysis was used to show that the “probable
uncertainty” in dermal exposure values calculated from either
biomonitoring or patch dosimetry   SEQ CHAPTER \h \r 1 are sufficiently
large to weaken correlations using data from independent studies.  Thus,
the agreement in the data presented to the Panel is about as good as can
be expected and is sufficient to support the Agency’s conclusion that
a passive dosimetry-based approach can generate data that can be used to
develop relatively predictive estimates of worker exposure for a wide
variety of scenarios and activities.  

4)  Normalization of Exposure by Amount of Active Ingredient Handled
(AaiH)

Most Panel members agreed that the data shown to the Panel did not
consistently support a linear relationship between exposure and AaiH.  A
linear relationship seems intuitively logical, but a physical rationale
should be developed to support that hypothesis (or other hypotheses) in
all scenarios.  Several good reasons were given as to why a linear
relationship might exist but not be detectable within the PHED data. 
Some arguments were presented to accept and/or explain an apparent
non-linear relationship between AaiH and exposure.  And some factors
other than AaiH were suggested that might be better correlated with
exposure than AaiH.

For dermal exposure to be proportional to AaiH implies that a
consistently small fraction of the amount of pesticide handled is
deposited onto the handler’s skin.  The Panel proffered three reasons
why so few of the exposure examples from PHED were found to be
proportional to AaiH.  The “ecological fallacy” is the mistaken
assumption that all members of a group have the same characteristics as
the group at large.  The “engineering fallacy” is the mistaken
assumption that all work practices and equipment within a scenario are
in fact the same.  Either fallacy   SEQ CHAPTER \h \r 1 might cause
proportionality to apply within a closely defined portion of a scenario
(such as within a particular cluster), but fail to apply across the
whole scenario.  Alternatively, the uncertainty within the data might
just be too large to permit a linear proportionality to be detected,
even if it exists.  

Several Panel members pointed out that strict adherence to
proportionality is not dictated.  If a nonlinear relationship exists
between AaiH and dermal exposures to pesticides, then the log-log
regression coefficient that gives the best fit of the relationship
should be used, whether it is equal to 1 or not.  And some other
Panelists took the position that at least in certain scenarios, no
correlation at all should be expected between AaiH and exposure, and
went on to suggest other factors that might be better predictors of
exposure such as the number of times a handler contacts a contaminated
surface or pesticide residues on surfaces that are at steady-state or
saturated, regardless of the AaiH.

Some suggested that the variability in field conditions will make it
difficult for a well-designed observational study such as that proposed
to illuminate a clear predictive relationship.  An array of short
documents that put forth a plausible argument based on physical
mechanisms applicable to each scenario might comprise an alternative way
to justify any of the above assumptions and could also help in designing
"purposive non-random sampling" to support any of these predictive
correlates of exposure.  The cluster-based sampling strategy is being
optimized to test linearity (see Charge #6).  In scenarios where
correlation with a variable other than AaiH is hypothesized, the degree
of cross-correlation between these variables and AaiH needs to be
considered.

5)  Within-worker and Between-worker Variability

The Panel agreed that exposure data collected from observational studies
has the potential to address all three potential sources of variation
identified in the background documents: within-handler, between-handler,
and between-study.

Much of the Panel felt that within-handler variability should be
de-emphasized because repeated measurements on any one handler are
likely to demonstrate low within-handler correlations and that
between-handler variability can predict the distribution of long-term
cumulative exposures better than within-handler variability.  The Panel
also recognized that repeated measurements may introduce a selection
bias within participants and the narrow time window for pesticide
application can make obtaining repeated measurement data infeasible for
some scenarios.  Also a decision to include repeated measures could
raise ethical or other issues at a Human Studies Review Board. 

At the same time, some on the Panel were concerned that EPA might be
missing an opportunity to obtain at least some limited repeated
measurements.  One suggestion was to analyse within-handler data
currently within PHED to provide some further quantitative evidence to
support a later decision.  The Panel was somewhat split among members
who suggested that repeated measurements require a large increase in
both effort and analytical costs and those that felt that adding limited
repeated passive dosimetry measurements on the same or next day would
provide the most information for the least additional investment.  The
Panel recommended that EPA and the AHETF evaluate whether within- and
between-worker variability might be evaluated for selected scenarios
where application frequency and logistics are favourable.  Statistical
techniques were also suggested using statistical methods that
incorporate cost information to see how to get the most precision for
the least total cost.  

6) Sample Size: Number of Sites and Subjects per Scenario/Activity

While the appropriate number of Monitored Units is integral to the goals
set for any database, the goals set by the Agency and by the AHETF may
be in conflict.  The goal of the Agency is to “adequately represent
the range of exposure of people who engage in a particular scenario and
activity.”  The goals of AHETF are to be able to estimate exposure
after normalizing by the amount of active ingredient handled [AaiH]
within a proposed factor [K] of 3 and to be able to “distinguish”
between complete proportionality and complete independence of exposure
and AaiH.  

To meet the accuracy goal, the number of sampling units will depend in
part on the K factor.  The Panel felt that a default value of K = 3
seems reasonable, although it need not be a fixed value and might best
be varied among scenarios.  The number of sampling units under the
cluster-based sampling plan is also quite sensitive to the specified
geometric standard deviation [GSD] and intra-class correlation
coefficient [ICC].  Therefore, EPA and the AHETF should consider
building in one or two “check points” after a certain amount of new
data is collected to evaluate assumptions about the GSD and ICC and
refine the scenario sample sizes.  In order to understand the robustness
of the cluster design to the assumption of log-normality and its impact
on calculated sample size, the Panel also recommended that sample size
simulations be performed using an alternative skewed distribution for
concentration values, such as a Gamma distribution.

Overall, the Panel believes that the recommendation to have 5 handlers
per cluster with approximately 5 clusters for each scenario/activity
seems reasonable at this point in time.  However, because the variation
in ICCs observed to date comes from sparse data, the Panel again
recommends building in one or two “check points” to consider
adjustments in the numbers of clusters and monitoring units (MUs) within
clusters to be sampled.  At the same time, clear guidelines are needed
on how to add new clusters that guard against the potential to
“parse” the target population into more clusters just to limit
sample size requirements.  More clarity was desired about how a cluster
is defined, whether that definition needs to be scenario dependent, and
the notion that “geographic differences” are important for
establishing clusters. 

The Panel questioned the adverse effect that was being imposed on sample
selection by the secondary goal proposed by AHETF to be able to
elucidate a potential predictive relationship between exposure and AaiH.
 Is it feasible to span the desired range of AaiH without biasing the
selection of application equipment and/or work practices?  Will the
“bias toward conditions that might yield higher exposures” conflict
with the assumption that the purposive sample of the MUs approximates a
probability sample from the target population?  The conclusion of a
detailed critique of the AHETF sampling plan contained within Appendix C
is that the lack of a database that documents the distribution of
real-world tasks, activities, or pesticide usage may make it very
difficult to judge or compensate for the biases introduced by the
sampling selection design being proposed (see also the last comment in
our response to Charge #1).  Further clarification is needed about this
sampling bias and how it might affect the distribution of exposure
values and subsequent uses of the database.  

The AEATF Study Plan deals with a very different situation and appears
much more amenable to experimental control.  The proposal to take 15
monitoring units initially seems adequate to give an overview, and yet
in this case it should be feasible to increase the sample size for any
scenario at a future date if more observations are needed.  The Panel
also suggested that the AEATF undertake a pilot study to compare the
results of studies conducted at one simulated site versus three field
locations.

PANEL DISCUSSION AND RECOMMENDATIONS

Agency Charge

1)  Data Needs

EPA believes that many studies within our current database have
limitations.  In some cases, the Agency is lacking data to address
modern pesticide application equipment and techniques.  EPA believes
that additional data could significantly improve our ability to estimate
and better characterize the range of worker exposure with greater
certainty. 

	Please comment on these limitations and EPA’s conclusion that
additional data could improve significantly the Agency’s ability to
assess worker exposure.  Also, please comment on the selection criteria
proposed by the AHETF and AEATF in their respective submissions for
evaluating the extent to which existing data would meet EPA’s exposure
assessment needs.

Panel Response 

The Panel agreed with the Agency’s concern about the limitations of
the existing PHED exposure database and added a number of concerns of
their own.  Furthermore, the Panel also concluded that additional data
could improve significantly the Agency’s ability to assess worker
exposure.  The limitations of the current PHED database are summarized
below.

Inadequate QA/QC in many of the available data and the limitations of
the current data grading criteria to adequately depict the uncertainty
within the results (see Appx. B).

Inclusion of data based on sampling methodologies (e.g., patch
dosimeters and outer whole-body dosimeters) that yield exposure
estimates with significantly higher levels of uncertainty than data
based on inner whole-body dosimeters.

Inclusion of data with high amounts of data censoring coupled with
treating all censored data as LOD/2 and not informing users of the
database about the extent of censoring.

Inclusion of some data based on incomplete dermal sampling (i.e., not
from the entire body) that requires a method (algorithm) that combines
data from different individuals to assemble a complete dermal exposure
value.

The existence and unknown effect of high levels of "clustering" (a lot
of data from one study) within the data that comprise many scenarios
(see also Charge #6).

The short sampling period of some of the data in selected scenarios
resulting in more uncertainty when scaling up to full day exposures.

An inadequate number of measurements in many scenarios resulting in less
reliable interpretations and extrapolations.

A potential lack of “representativeness” due to the absence of
modern work practices and equipment in the exposure database (and the
possible inclusion of some older work practices that may now be less
common or no longer used).

Lack of diversity of test conditions within some scenarios due to a
combination of limited numbers, clustering, the age of the data, and
evolving technologies.

The Panel agrees that these limitations decrease our confidence that
PHED can reliably estimate exposures for pesticide handlers in all
handling scenarios.  The ability of such a database to be able to
estimate exposure has become an essential part of a regulatory risk
assessment that ensures there is a sufficient margin between the likely
exposure and the toxicological endpoint of concern.  Thus, there is a
need to be able do this consistently with a degree of confidence that
protects the health of handlers while permitting products that present
acceptable risks into the market for the benefit of growers, industry,
and consumers.    SEQ CHAPTER \h \r 1 While PHED has served its initial
purpose, its goals have evolved.    SEQ CHAPTER \h \r 1 In addition to
the above limitations, Panelists elaborated on the following three broad
weaknesses within PHED.

The first weakness is the structure of the database and the algorithms
necessitated by that structure.  These reflect the fact that many of the
original data came from studies where either not all parts of the body
were monitored or exposures were undetectable.  As a consequence of the
structure, extrapolation algorithms hidden within the software are often
necessary to utilize the data to estimate exposure.  While the structure
and algorithms may have been expedient at the time, analyses of these
data do not represent individual exposures or yield accurate exposure
statistics or confidence limits for those statistics.  This is true for
both the central tendency (arithmetic mean) and higher exposure values
(e.g., 95th percentiles), both of which may be of interest for
regulatory assessments.  The AHED software represents an opportunity to
correct this deficiency and to calculate estimates of the mean and
higher percentiles and their associated confidence intervals.

The data from the existing PHED presented on p. 33 of the Agency's
Review is a good (although perhaps inadvertent and not fully explained)
example of the weakness of relying on the current data and algorithm
used to accommodate censored data.  As a result of PHED's use of
incomplete and/or censored data from mixed sources using various
methods, it predicts that hand exposure without gloves (0.0095 mg/lb
a.i.) is equal to or slightly less than exposure with gloves (0.0097
mg/lb a.i).  Such a result is completely counter to a logical
expectation of reality.  Continued reliance on such outcomes may not
only lead to incorrect decisions for handler protection but also weakens
the policy of the Agency to require the use of appropriate protective
gloves.

Several Panelists encouraged the Agency to adopt more sophisticated
statistical methods of dealing with censored data (e.g., MLE) that would
yield less statistically biased estimates of the distribution of
exposures.  In addition (and at the very least) an indication of the
degree of censoring should be included in the output of the exposure
assessment database.  Indeed, there is a strong body of scientific
opinion, with much agreement at the international level, that the
characterization of both the variability and uncertainty in exposure
assessments should be transparent.

The second weakness relates to the actual data.  It is clear from the
materials submitted to the Panel that the limitations of PHED could
severely hamper the Agency’s ability to adequately assess pesticide
handler exposures.  In addition to problems with conducting a reliable
exposure assessment using PHED for those scenarios where the data
quality (QA/QC) or the number of monitoring units is low, much of the
study methodology within PHED is dated.  The studies are a minimum of 12
years old, while some are as much as 30 years old.  Study designs and
the sampling and analytical methodologies for measuring pesticide
exposure have improved over the years, resulting in exposure to a
greater proportion of the body being measured, vastly improved limits of
detection for pesticides in sample media, less censored data (i.e.,
non-detectable data), and improved overall quality assurance of exposure
assessment data.  As evidence of the continuing evolution of methods,
the Panel discussed relatively recent studies that show the rapid dermal
adsorption of some active ingredients and the analyses of probable
uncertainty to be described in response to Charge #2, that suggest even
some current data within AHED that are based on hand washing should be
interpreted with caution.

The applicability of older data to newer ("modern") pesticide
application equipment and techniques presents some interesting dilemmas.
 Clearly, exposure cannot be assessed using PHED in scenarios with new
technologies, work practices, or product formulations that are not
reflected within PHED.  This fact justifies some new studies.  But if
much old equipment is still in use (a likely possibility given farming
traditions), then the old PHED data (despite their limitations) are
applicable, have value, and should not be discarded or entirely
abandoned.  It should be possible to test for a statistical difference
between new and old data.  One would then need to decide whether such a
difference was due either to better study methods or to safer
technologies.  If the latter were true, the Agency would have to decide
whether, for instance, the lower exposure data that might result from
the utilization of newer equipment and techniques would not be
applicable to the older but still functioning equipment and techniques. 
Such an outcome might necessitate creating, for example, a new use
category and possibly restricting the use of certain high-toxicity
chemicals to those newer equipment and techniques that limit exposures. 
Underlying these dilemmas is the lack of a database describing the
actual distribution of equipment and techniques among current users (see
further comments on this point in the last portion of this response). 
The evolution of equipment and work practices within a scenario will
also affect our response to Charge #4.

The third weakness pertains to the unrecognized impact that the
assumptions, statistical variations, and calculations implicit within
each dermal exposure assessment method have on the uncertainty in the
resulting exposure values.  Assigning a "Data Grade" to each study is a
good concept.  However, expanding the concept of grading to the end
result (the measured exposure level) will help users to better interpret
the resulting exposure values as well as create a direct means for the
Agency to demonstrate both the weakness of the existing PHED and the
improvements that should result from a new database.  The only grading
criterion that is currently integrated into PHED is the quality of the
analytical method (defined within PHED by a combination of “mean % Lab
recovery" and the "coefficient of variation (CV) for Lab recovery" or
the “mean % Field recovery" or “mean % Storage [recovery]
Stability.")  While the percent recovery contributes to the uncertainty
of the resulting exposure data, Appendix B explains why the CV is the
more important of these two parameters.  Parts of the Panel’s response
to Charges #2 and #3 will expand on this same theme to explain why
calculations implicit within the sample collection method (e.g., PD
versus BM) have a much larger effect on the uncertainty of the resulting
exposure value than the CV of the analytic method and should therefore
be integrated into the Data Grade.  The same principle (with less
quantitative measures than CV) could also be applied to other data
attributes such as the lack of complete body monitoring, the lack of
complete urine collected (if biomonitoring data were included), and the
impact of the simplistic method currently used to account for
undetectable samples (due to censoring).  (If censoring is not
integrated into the grading criteria, at the very least the fraction of
undetectable samples should be an output to users of PHED and AHED.) 
These other measures of data quality are likely to be better indicators
of the poor quality of the existing PHED data than the current Data
Grade based only on the analytical method and could be used by the
Agency to help justify the need for new exposure data whose uncertainty
can be improved by orders of magnitude.

In 2003 the   SEQ CHAPTER \h \r 1 International Life Sciences Institute,
Risk Science Institute, convened an international workshop to consider
how to conduct   SEQ CHAPTER \h \r 1 probabilistic assessments of worker
exposure to agricultural pesticides.  This workshop brought together
exposure assessors, modelers, toxicologists, and statisticians.  After
considering case studies developed for that workshop, it became apparent
to participants that the PHED contained so much unexplained variation
(likely due to the limitations in the data and mixed study protocols)
that the objective could not be achieved reliably.  Consequently, it was
concluded that more robust representative data are required to attempt
to fulfill the objective.

The Panel agreed with the Agency’s conclusion with regard to the
limitations of PHED and that additional worker exposure data could
vastly improve their ability to assess worker exposure, provided that
the studies to collect additional data are designed in such a way as to
address the shortcomings of PHED.  It is important that the new studies
be representative of the diverse nature of handlers and use settings,
minimize bias, monitor a significant proportion of the handler’s
working day, use methods that decrease the uncertainty associated with
unavoidable extrapolations, and use a sufficient number of monitoring
units to allow some meaningful separation of the effects of uncertainty,
natural variations, and bias within the results.  Furthermore, it
appears from the supporting documentation presented to the Panel that
the proposed generation of exposure data from AHETF and AEATF is
designed to address the limitations of PHED and the needs of the Agency.

A recent Human Studies Review Board review of the proposal to collect
new data questioned the need for new human exposure studies, citing that
the Agency did not clearly demonstrate the need for new data.  This
Panel is clearly of the opinion that additional worker exposure data
collected on human volunteers under field conditions and label
requirements on chemicals that have been approved by the Agency are
necessary.

AHETF and AEATF study selection criteria

The Panel has also been asked to comment on the selection criteria
proposed by AHETF and AEATF for using existing exposure data to meet the
exposure assessment needs of the Agency.  While the criteria outlined in
documents submitted to the Panel by the AHETF and AEATF appear to be
reasonable for including existing data into regulatory exposure
assessments, the Panel expressed concerns regarding the selection
criteria for future data including those that follow and those regarding
“clustering” and “sample selection” discussed in our response to
Charge #6.

The monitoring duration requirement may be too stringent.  For instance,
the requirement to monitor for at least one-half day is unlikely to
capture all high intensity and short-term dermal exposures.  Some
provision to allow the inclusion of data from settings where only
short-term uses are the norm may need to be added to the criteria.  

The criteria for the use of biomonitoring data may also be too
restrictive.  The AEATF states in their criteria document that
biomonitoring data would only be acceptable if primate dermal absorption
and pharmacokinetic data exist for the chemical being monitored.  The
point is made that extrapolation parameters must be available for the
study to be selected by the AEATF.  However, there is a paucity of
studies that have “extrapolation” factors to humans, and most are
estimated from rat and porcine models (e.g., McDougal, 2002 and
Williams, 1996).  The requirement that only primate dermal absorption
data be included in this database may be too stringent.  There are
mathematical ways of treating rodent and porcine data to make them more
applicable to human dermal absorption, namely Fick’s law of diffusion
and adjusting for the thickness of the skin.  At the same time, the use
of pharmacokinetic and dermal absorption data to back-calculate from
urinary excretion to dermal exposure that can be used in the generic
database is likely to introduce a very large amount of uncertainty into
the generic exposure assessment, as will be discussed further in Charge
#3.

μm droplets, not all of which are “inspirable.”  It should also be
remembered that although large droplets or particles may not be respired
into the lung, they may be deposited in the nasal region or mouth where
they are available for absorption.

To the degree that patch dosimeters are allowed, the Agency should
standardize the array of patches that must be used in the assessment of
dermal exposures to pesticides, in particular the location of each
patch.  The current PHED includes some data that relied on the skills
and observations of the researcher to determine the areas of the body to
monitor, which introduced the biases of the researcher and a lack of
consistency across studies.  And finally, some means should be
incorporated within the AHED to allow estimating dermal exposures for
people of different shapes and sizes from either patch or whole-body
dosimeter data.

Other databases

Neither the current PHED nor the proposed AHED or BHED include data that
document the distribution of tasks, activities, or pesticide use
information within any given exposure scenario.  The variance observed
within an exposure database is the combined result of statistical
uncertainty and imprecision within the assessment methods and natural
variability within an exposure scenario.  Better interpretation of the
observed variability in exposure and the adequacy of an exposure
database require (1) measuring the intensity of exposure with consistent
methods (which is the only one of the three that should happen within
the new proposals), (2) examining the frequency of exposure, and (3)
identifying the duration of exposure.  The Panel is unaware of any
database that contains the latter two important descriptors of exposure
scenarios.  Although we are not suggesting that a requirement to
generate such a database be placed on the task forces, such a
descriptive database would greatly increase confidence in any exposure
database that will serve as the basis for conducting adequate and
reliable risk assessments (see also Appendix C and our response to
Charge #6).

Agency Charge

2)  Passive Dosimetry

The common approach for conducting dermal exposure monitoring studies
relies on the use of whole-body dosimetry, handwashing, and facial/neck
wipes.  In some cases, biological monitoring is also used as an
alternative method.  Exposure estimates in Agency risk assessments,
however, typically rely on “to the skin” measurements (i.e.,
potential dose) coupled with dermal absorption data or dermal toxicity
studies in order to calculate risks.  The Agency believes that these
methods are complementary and that they can provide appropriate
estimates for exposure assessment but that the results directly relate
to the reliability of the inputs used.  Please comment on the Agency’s
conclusion regarding passive dosimetry and biological monitoring,
including whether a systematic bias exists in either approach.

Based on the information presented, the Agency has particular concerns
over three specific aspects of how these studies are conducted including
(1) the possible need to correct for the efficiency of the handwashing
technique; (2) compensating for absorption of residues through the skin
during sample collection periods; and (3) the breakthrough of residues
under whole-body dosimeter garments.  Please comment on the need to
systematically account for residue losses due to these potential method
biases.  If there is a need, please describe how these corrections
should be accomplished in a way that could reduce uncertainties in the
resulting exposure estimates.

Panel Response

This response is divided into four parts.  The first three parts each
address a “numbered” portion of the charge.  A discussion of the
complementary nature of biological monitoring [BM] and passive dosimetry
[PD] comprises the last part.

1) 	Is there a systematic bias in either approach?

Bias between dermal exposure monitoring and biological monitoring is not
detectable within the data presented to the Panel.  Statistical
uncertainty is at the crux of this response.  Other less comprehensive
evidence in the published literature suggests that a bias may exist, but
if a bias does exist within PHED, it is much smaller than the
uncertainty of the two methods.  The level of uncertainty in the
estimates of dermal exposure using passive dosimetry will be discussed
below; the level of uncertainty in the estimates of dermal exposure
using biological monitoring [BM] will be discussed in our response to
Charge #3.

Any form of dosimeter will be affected by imprecision within the method
used to calculate exposure.  The uncertainty in exposure measurements
derived from patch dosimeters is due, for example, to the effect of
scaling up from deposition onto a patch of circa 40 cm² to a body-part
area of circa 1000 cm² and the variability of dermal deposition onto
any given body part.  The interaction of these two factors as defined by
Equation 2.1 has the same effect on the probable uncertainty in the
resulting exposure value as spike recovery had in Equation B.1 in
Appendix B.  (The variability in absorption and metabolism will have a
similar effect on the exposure value calculated from urinary excretion,
as will be described in the response to Charge #3).

                          Eqn. 2.1

The variability in the spatial deposition onto the skin of a given
handler is generally unknown.  Placing two patch dosimeters on each
portion of the body only covers between 2% and 4% of each body part,
resulting in a scaling factor of 25-50.  The use of patch dosimeters and
gloves (a "whole body-part dosimeter") means that dosimeters cover about
6% percent of the whole body.  Figure 2-1 allows one to see how even a
small amount of variability in the spatial pattern of dermal deposition
will cause a large probable error when using patch dosimeters.  

 

Figure 2-1.  The interaction of the CV in dermal deposition and the
fraction of the body covered by dosimeters upon the magnitude of the
Probable Uncertainty

In comparison, whole body dosimeters cover circa 90% of the body, and
even a large variability in the uncovered area (face and neck) can only
have a relatively small effect on the probable uncertainty in the
resulting dermal exposure value.  Any time the overall efficiency of
such dosimeters is at least 70% or better, the analyses presented in
Figure 2-1 indicate that the uncertainty in the estimates of dermal
exposure using whole-body dosimeters is quite a bit smaller than the
uncertainty using either patch dosimeters or biomonitoring.

The implications of these levels of uncertainty will be discussed more
fully in Charge #3, but the simple answer to the question herein is that
statistical uncertainty precludes the detection of any bias between PD
and BM data within the data presented to the Panel.  The rationale to
believe that a bias may exist in either approach to assess dermal
exposure is discussed in the next two subsections of this Response.

2)	Is there a need to correct handwashing for its efficiency or to
compensate for the absorption of residues through the skin during sample
collection periods?

Overall, the Panel believes that if a correction method for adsorption
and/or absorption can be derived that can significantly decrease the
uncertainty of the resulting measurement at a reasonable cost and within
approved human studies guidelines, then it should probably be applied. 
However, to answer this second question, one has to realize that the
skin is more than a filter that just slows absorption.  Skin comprises a
complex system of layers onto which pesticides may reside, adhere,
diffuse, or even be metabolized.  Pesticide reaching the skin may first
reside on the skin from which it can be washed with relative ease (very
analogous to a “dislodgable pesticide residue” on foliage).  A
pesticide may also be temporarily adsorbed onto the top layer of the
skin (the stratum corneum) before being absorbed through the skin and
into the body.  The degree of adsorption and the rate of absorption
either far enough into or through the stratum corneum where the
pesticide is not removable with a weak solvent determine the efficiency
of handwashing as an assessment method.  Added to these physiologic
mechanisms is the varying degree of importance of hand exposure within a
scenario and its mitigation by the use of chemical protective gloves.

Adsorption would be expected to be pesticide- (and possibly
formulation-) dependent and to be related to the ability of the
particular pesticide to adhere to the stratum corneum.  The fraction
adsorbed would be expected to vary based upon the time the pesticide
remains on the skin prior to washing, the amount of pesticide that was
deposited onto the skin (and perhaps the rate of deposition), and both
intra- and inter-personal variability.  The efficiency of handwashing
would depend upon the ability of the solvent (e.g., water, alcohol,
detergent) to remove the chemical from the skin or to promote its
absorption into the skin which may vary based upon the physical and
chemical properties of the pesticide and the handwashing protocols used.
 

Existing data clearly indicate that adsorption and absorption of certain
pesticides can occur within a matter of minutes after the exposure has
occurred.  For example, data presented in Fenske and Lu (1994) show that
several handwashing solvents recover less than 50% of chlorpyrifos from
the skin immediately after exposure and recover only about 20% from the
hands one hour after exposure.  The handwashing efficiency data
summarized in Table 1 of a review by Brouwer et al., 2000 cited in the
EPA Review range from 23 to 96%.  The result ranges from a negligible
bias to a four-fold underestimation of the amount of pesticide measured
on the skin.  

In contrast to hand washing, much of the hand wipe data presented (such
as the 10% mean recovery with a CV of 33% for azinphos-methyl in Table
3-4 of the Review (from Fenske et al., 1999) and the many chemicals with
circa 50% mean recovery with a similar CV in Table 2 of the review by
Brouwer et al. (2000) indicates that hand wipes may be both more biased
and more variable than hand washes.  The combined impact that lower
levels of recovery efficiency and higher variability has upon the
probable uncertainty of the resulting measured hand exposure as depicted
in Figure 2-1, led some Panelists to recommend that hand wiping not be
used in deference to either hand washing or gloves.

 

The duration of the monitoring interval (p. 38 of the Review) appears to
be a factor affecting both glove dosimeters and handwashing.    SEQ
CHAPTER \h \r 1 An analysis was conducted by one Panelist of the dermal
recovery via hand washing data cited in the Review’s Table 3-7 for
captan and Table 3-8 for chlorpyrifos.  In both cases, the efficiency of
recovering a single ("spike") deposit onto the hands appears to follow
the same exponential model of the form shown in Equation 2.2 that
results if one assumes that the rate of adsorption is a constant
fraction of the dose to the hands.

(−time / HalfLife)        Eqn. 2.2

Using the above data in this model yields half-lives for the recovery by
handwashing of slightly over 2 hours for captan and about 1 hour for
chlorpyrifos.  Morever, the HalfLife for these two chemicals [in hours]
is approximately "5.1/log KO/W" with a coefficient of variance of only
about (5%.  This type of model was then used to generate a predictive
equation for wash recovery following an assumed uniform rate of dermal
exposure over the work period, as shown in Equation 2.3.

         Eqn. 2.3

Equation 2.3 indicates that if exposure is constant with time, recovery
via washing will reach a "steady state" in two to three half-lives. 
These facts suggest that the interval from initiating exposure until
washing the hands can be important if monitoring from the planned
minimum of 4 hours to a nominal maximum of an 8-hour day, as proposed by
AHETF.  Equation 2.3 predicts that waiting this long to wash the hands
of handlers of these particular chemicals would result in recoveries
that would underestimate exposure by factors of 3( to 5(.  Thus, the use
of a hand washing technique in AHETF studies should be accepted if it is
supported either by a predictive model or by animal or human laboratory
recovery data for retention times extending up to the maximum sampling
time for the pesticide being studied (see also the Panel’s comment
regarding primate dermal absorption data as part of the AHETF and AEATF
study selection criteria near the end of our response to Charge #1). 

Others on the Panel pointed out a weakness to both the modeling and
experimental approaches.  The accuracy of a correction using either
approach could be confounded by the effects of repetitive (multiple)
rinsing or washing of the skin of the hands during a work shift for data
collection, to attend to personal hygiene needs, or for other purposes. 
This repetitive rinsing or washing raises the possibility of changing an
individual’s absorption rate due to alterations in the physiological
nature of the outer layers of their skin (i.e., desaponification or
protein binding), co-solvent effects, or other mechanisms.  Some of
these alterations can be expected to enhance dermal penetration rates,
causing handwash data to under-estimate exposure, while other
alterations could increase the recovery from the skin, potentially
causing handwash data to over-estimate exposure.  The uncertainty
associated with hand washing represents a major topic for research.  The
above reasons might lead the AHETF to favor the use of cotton glove
dosimeters to assess hand exposures in future AHED monitoring studies;
however, they are not without their own limitations. 

An alternative approach to using one of the methods described above to
correct handwash data would be to quantify the amount of absorbed dose
based upon excreted metabolites and pharmacokinetic information and add
this to the passive dosimetry estimates.  Unfortunately, because
biomonitoring provides data that are independent of the route of
exposure, other routes of exposure unrelated to handling would be
included which (if not accounted for) might overestimate the total
dermal dose due to the handling task.  Also, the biomonitoring approach
is not applicable to pesticides that do not have a reliable biomarker or
where sufficient pharmacokinetic information is lacking.  The burden to
the participant becomes larger if they are requested to provide 24-hour
urine samples over a period of days, which might bias participant
selection in some way.

Further data are needed for better quantification of chemical absorption
through human skin in vivo.  Currently, the Agency uses the “%
absorbed” (or absorption factor) to quantify the amount of dermal
exposure that is taken up into the body.  This “% absorbed” is based
on empirical observations for a narrowly defined exposure scenario (see
McDougal, 2002).  However, the amount of absorption is dependent upon
the intensity of the exposure; for larger exposures, the “%
absorbed” is less.  The reason for this is that only the layer of
chemical in direct contact with the skin is available for uptake while
the entire exposure is included in the calculation of fraction absorbed.
 Further studies should be conducted to estimate the percent absorbed
for a range of exposure levels.

Is there a need to correct dosimetry for the breakthrough of residues
under whole-body dosimeter garments?

Breakthrough on a dosimeter could occur either slowly, due to the
migration of a collected pesticide residue due to a mechanism like
particle filtration or permeation or rapidly, due to the saturation of
the dosimeter with a liquid.

Early patch dosimeters had an impervious backing to prevent
breakthrough;   SEQ CHAPTER \h \r 1 however, as shown via Figure 2-1,
the variability in exposures over individual parts of a handler's body
is likely to be sufficiently high in application settings to make the
probable uncertainty of dermal exposures calculated from patches
unacceptably large.  The same analyses depicted in Figure 2-1 showed
that the probable uncertainty associated with whole-body dosimeters was
much more acceptable.

Some researchers used the handler’s outer clothing as a dosimeter, in
which case breakthrough could occur both from liquid saturation and as a
result of the penetration of dry pesticide filtering through the
clothing due to air movements.  Various penetration factors were
suggested ranging from perhaps 5% to 20% with a nominal default value of
10%, implying a default efficiency of 90% with a CV of (100%.  While the
correction for this level of breakthrough is slight, the probable
uncertainty in the resulting calculated exposure is about (2(, as would
be shown in Figure 2-1 if a CV of (100% were included.

A whole-body dosimeter placed directly against the skin would minimize
the slow penetration due to air movement.  However, it would still be
susceptible to breakthrough from liquid saturation.  Unfortunately, one
cannot put an impervious backing on a whole-body dosimeter to protect
against liquid breakthrough without creating a heat-stress hazard to the
wearer.  Moreover, no obvious quantitative criterion for liquid
saturation has been identified, and a qualitative criterion such as
"visible saturation of the outer clothing or exposed whole-body
dosimeter" is probably not reliable.  Patch dosimeters placed
strategically underneath the whole body dosimeter are unlikely to detect
saturation, should it occur, because saturation is unlikely to be
uniform, as discussed in part 1 above.  Detecting and responding to
saturation via biomonitoring is discussed further in part 4 below.

Is biological monitoring complementary to dermal exposure monitoring
(i.e., dosimetry and handwashing)?

Biomonitoring is possibly one of the few viable approaches available to
at least detect if not estimate the amount of breakthrough from a whole
body dosimeter, especially breakthrough due to saturation.  In theory,
the Panel supports the proposal (Review p. 61) that "biomonitoring be
included in any sampling plan" as a validation that the passive
dosimetry collected virtually all of the handler's exposure.  However,
the Panel also foresees that requiring concurrent biomonitoring could
severely restrict study participants to those with no prior (or
near-term subsequent) exposure to the chemical, introducing a
potentially serious bias in the results.  Thus, the Panel recommends
that concurrent biomonitoring not be required. 

Should biomonitoring be performed and if it were to detect breakthrough,
the preferred response is to add the additional dose estimated using a
PBPK or another exposure-excretion model to the exposure calculated from
passive dosimetry.  Discarding all samples associated with a
breakthrough greater than some to-be-defined threshold (e.g., 30%) would
prevent the inclusion of data suspected to be low, but excluding
saturated samples that may not recur in a replacement assessment would
bias the data downward.  If biomonitoring were to detect breakthough but
a correction cannot be estimated, the occurrence of saturation should be
reported both within the data entered and to a user of the database’s
output (much like an earlier recommendation in response to Charge #1 to
report the number of undetectable samples). 

If biomonitoring were used as a complement to dermal exposure
monitoring,   SEQ CHAPTER \h \r 1 then caution should be exercised in
the use of creatinine.  Traditionally, creatinine has been used either
to correct for dilution in a spot urine sample or to check for
completeness of a 24-hr urine sample.  An individual’s creatinine
excretion rate can vary by age, race/ethnicity, physical condition, and
creatinine should only be used to correct for urine dilution for
metabolites excreted in a manner similar to creatinine.  Thus,
creatinine may not always yield an appropriate correction for urine
dilution.  The use of a PBPK model and two timed-interval urine samples
as an alternative to creatinine will be discussed in our response to
Charge #3.  

Agency Charge

  SEQ CHAPTER \h \r 1 3) Passive Dosimetry vs. Biomonitoring

EPA believes that a comparison of exposure estimates derived from data
collected through biomonitoring with data collected through passive
dosimetry is the most appropriate way to assess the predictive nature of
a passive dosimetry-based approach for estimating worker exposure. 
Please comment on the strengths and limitations of this kind of
comparison for judging the potential utility of passive dosimetry data
in conducting exposure assessments.

EPA has conducted such a comparison using available data and believes
that the comparison shows sufficient concordance of estimates based on
biomonitoring data and passive dosimetry data to support the conclusion
that a passive dosimetry-based approach can generate data that can be
used to develop relatively predictive estimates of worker exposure for a
wide variety of scenarios and activities.  Please comment on the
adequacy of the analysis to support EPA’s conclusion.

Panel Response

  SEQ CHAPTER \h \r 1 A basic schematic flow chart is presented in
Figure 3-1 as a visual aid to this response, to depict the relationships
between dermal exposure and the chemical at its various steps en route
to excretion.  This figure tries to depict these relationships for three
passive dosimetry [PD] options: skin washing (hands or face/neck), patch
dosimeters, and whole-body dosimeters.  Each option is represented by a
column of boxes and arrows.  The top box in each column represents the
work activity that results in a dermal exposure.  The width of each
arrow varies (although not to scale) to represent the amount of the
chemical passing from step to step.  Thus, the width of the first arrow
is the same in each column because it represents the same exposure.  In
general the width of the arrow decreases as the chemical passes through
or around clothing and/or dosimeters, through the skin, and through the
metabolic pathways of the body.  The amount of urinary excretion
potentially measurable via biological monitoring [BM] is at the bottom
of each column.

 

		Figure 3-1.  Depiction of the decreasing mass of chemical passing
through various levels of dosimetry, the skin, and the body to urinary
excretion that could be used to extrapolate back to dermal exposure

  SEQ CHAPTER \h \r 1 The first column depicts washing only (typically
of the hands and possibly wiping the face and neck, although wiping has
been shown to recover less pesticide than washing as discussed in Charge
#2).  The Panel’s response to Charge #2 also discussed the
calculations necessary to scale-up the analytically measured amount of
chemical removed via hand washing to estimate the dermal exposure. 
Although hand washing is typically used in conjunction with either patch
or whole body dosimeters, the small change in the width of the arrow in
the first column of Figure 3-1 that reaches the skin reflects the
expectation that washing protected hands will not greatly reduce the
total amount of pesticide absorbed and eventually excreted.  The second
column depicts patch dosimetry in which only a small portion of the
pesticide that reaches the skin is captured, chemically analyzed, then
mathematically scaled-up to estimate the dermal exposure.  Because patch
dosimeters cover only a fraction of the skin, they do not greatly reduce
the amount of chemical that reaches the skin, the fraction of that
dermal deposition that is absorbed to become an internal dose, or the
fraction of that amount that is excreted.  The third column depicts
whole body dosimetry in which (ideally) all of the chemical that would
have reached the skin is captured (except typically that reaching the
face, neck, and hands), chemically analyzed, and reported as dermal
exposure.  A good whole body dosimeter should greatly reduce the dermal
exposure that reaches the skin, the absorbed dose, and the amount
excreted (probably reduced by more than the arrow width shown).

The depictions in Figure 3-1 illuminate two relationships between
passive dosimetry and biomonitoring.  First, the sequence of events
presented in Figure 3-1 causes some forms of dosimetry to reduce the
amount of chemical that can be deposited onto the skin, absorbed, and
excreted.  This interaction between methods would interfere with any
expected correlation between   SEQ CHAPTER \h \r 1 the exposure (as
represented by PD data) and biomonitoring (as represented by BM data). 
Second, the large uncertainty in the calculated dermal exposure
characteristic of biomonitoring and patch dosimetry will increase the
scatter of both variables and decrease the ability of a statistical
regression to detect an existing correlation, especially over a
relatively small range of exposure levels.  A good portion of the
Panel's response to Charge #2 discussed the uncertainty associated with
passive dosimetry, especially washing and patch dosimetry.  A good
portion of the rest of the Panel's response to this question will
elaborate on the   SEQ CHAPTER \h \r 1  uncertainty of dermal exposure
predicted by biomonitoring.

Several Panelists agreed with the Agency (Review p. 41-42) that
biomonitoring may well allow a more reliable prediction of toxicological
risk from the mass of a particular chemical that reaches the body's
internal tissues (including the target organ for adverse health
effects).  However, the goal of the database is to predict exposure of
any pesticide used within a scenario.  The longer chain of events and
the effect of multiple sources of uncertainty depicted in Figure 3-1
between exposure and excretion caused the majority of the Panel to
believe that probable uncertainty of dermal exposure values estimated
from biomonitoring is greater than similar dermal exposure values
estimated from passive dosimetry.  In other words, the dermal exposure
estimated from the use of good dermal dosimetry (represented by whole
body dosimeters and either head patches or a head-neck wipe and either
hand wash adjusted for adsorption or gloves) is subject to fewer
assumptions and less uncertainty than is the dermal exposure estimated
by a back-calculation from urinary excretion.  The latter is subject to
the uncertainty of inter-personal variations in the rate of dermal
absorption, the rate of metabolism and excretion, and the existing body
burden from recent prior exposures.  Any variance in these rates or
percentages either among participants within a study or scenario or
within a participant due to heat stress and/or their work rate (or
inter-species differences if applicable) will have a magnified effect on
the uncertainty in the calculated dermal exposure.  The interaction of
the variability from both of these calculations can be interpreted by
Equation 3.1 in a manner analogous to the interaction between CV for Lab
Recovery and % Lab.Recovery in Equation B.1 and between the variability
in dermal deposition and percent of body covered by dosimeters in
Equation 2.1. 

 	             Eqn. 3.1  

The range of mean % of the dermal deposition that would be excreted in
the examples cited in the comparison by the AHETF (slide #16 by John
Ross and Graham Chester "Comparison of Human Dosimetry and Biomonitoring
Data") is 0.18% to 8.9% (the product of "Human Dermal Abs. (%)" and
"Excretion Fraction (%)").  The Agency agrees (Review p. 42) that the
"rate of urinary excretion can vary considerably among individuals for
many reasons."  A CV of ±10% is probably minimal; the chlorpyrifos
urinary data on p. 46 of the Review has a mean of 1.3% of the dermal
dose excreted but a CV of 65%.  As shown in Figure 3-2, the resulting
probable uncertainties in a calculated dermal deposition from such
urinary excretion data is easily more than one order of magnitude and
could approach two orders of magnitude.

 

Figure 3-2.  The interaction of the CV of the fraction of the dermal
exposure or actual dermal deposition that could be recovered in the
urine upon the magnitude of the Probable Uncertainty

The combination of statistical uncertainties and the potential effect of
interferences probably contribute to the ±10× dispersion from the 1:1
regression line for individuals with concurrent monitoring in slide #24
by John Ross and Graham Chester and an even wider range of dispersion in
their slide #26.    SEQ CHAPTER \h \r 1 The fact that their slide #22
shows a dispersion of only about ±3× from the 1:1 regression line is
probably the result of exposures being based on PD and BM values from
independent studies and group averaging.  Although some Panelists would
have preferred to see the effect of adding all the data (including that
rejected for well defined reasons) into the regressions, the above
analysis of uncertainties indicates that the agreement in the data
presented to the Panel is about as good as can be expected and is
sufficient to support the conclusion that a passive dosimetry-based
approach can generate data that can be used to develop relatively
predictive estimates of worker exposure for a wide variety of scenarios
and activities.    SEQ CHAPTER \h \r 1 Another recently published
analysis of the well-characterized herbicide 2,4-D by Durkin et al.
(2004) sponsored by the U.S. Forest Service and EPA is offered as
evidence of further agreement between these analytical methods.

The large uncertainties associated with biomonitoring are much, much
larger than any of the uncertainties in Figure B-3 associated with
acceptable "% Lab. Recovery" values defined by the PHED Grading Criteria
discussed in Charge #1 and Appendix B.  They are also much larger than
the uncertainties in Figure 2-1 associated with whole-body dosimetry,
however, generally smaller than the other uncertainties in Figure 2-1
associated with patch and glove dosimeters.  For biomonitoring to yield
an estimate of exposure with a probable uncertainty close to the ca.
(3× dispersion noted above for PD versus BM data, the urinary recovery
must be in the range of 20 to 40% of the dermal exposure (values more
easily visualized in Figure 2-1 than in Figure 3-2).  Thus, while the
grouped results of the two methods are in good agreement, the analyses
presented in Figures 2-1 and 3-2 indicate that, in general,
biomonitoring data do not create as certain a measure of dermal exposure
as do passive dosimetry data and that whole body dosimeters are strongly
recommended over patch dosimeters.

However from a broader perspective, other divisions of EPA have recently
adopted physiologically-based pharmacokinetic (PBPK) models for risk
assessment (e.g., for methylene chloride).  PBPK models can be used to
quantify the relationship between exposure and the absorption,
distribution, metabolism, and elimination of a chemical based on two
urinary voids within a known interval (rather than requiring full
24-hour excretions to be collected).  Such models can also be used to
estimate the dose to the target-tissue, for which there is no
alternative approach other than conducting invasive animal studies.  In
fact, statements on page 35 and 42 of the Review indicate that the
Agency views biological monitoring as a good method "to quantify
absorbed dose."  "Such an estimate of absorbed dose, which avoids
potential confounding from assumptions of dermal penetration or
inhalation retention, may be more useful in assessing risk than
route-specific doses estimated from passive dosimetry."  The technical
development of the simplified pharmacokinetic model is ongoing within
the EPA Office of Research and Development, and the current model
development for chlorpyrifos is relatively mature at this point in time.
 Therefore, even though we do not currently have validated PBPK models
for performing reverse-dosimetry using biomonitoring data, the Panel
believes that such models will become available in the future.

Agency Charge

  SEQ CHAPTER \h \r 1 4)  Normalization of Exposure by Amount of Active
Ingredient Handled (AaiH)

The normalization of exposure by AaiH - the unit exposure - has, since
the mid-1980s, been the principle relationship underlying the use of
exposure data in the Agency's pesticide handler exposure assessments. 
It is based on the assumption that the two variables are proportional. 
That is, if one doubles the amount of pesticide they handled or applied,
the resultant exposure will be doubled as well.

The Agency is unsure whether the results of our exploratory work showing
that proportionality between exposure and AaiH is reasonable in some but
not all cases, is a function of limitations of the data within PHED or
whether this relationship is in fact not a reasonable assumption for all
scenarios.  It may be the case that an additional ancillary variable
(e.g., boom length, # of tank mixes, or # de-couplings in a closed
loading system), in addition to or in place of AaiH, may improve the
predictive capabilities of our exposure model.

Though it is recognized that neither the studies in our current database
nor the proposed studies by the AHETF were designed for the primary
purpose of examining proportionality between exposure and AaiH or to
determine the extent to which other parameters influence exposure,
compared with our current database, the Agency believes that the
proposed AHETF studies will generate data that will reinforce the
assumption of proportionality between exposure and AaiH or,
alternatively, inform the applicability of another variable as a more
appropriate predictor of exposure.

 

Based on the themes presented on this topic including its historical
precedent, its application in risk assessment and subsequent risk
management decisions, the Agency’s exploratory work using the six PHED
scenarios in the case study, and the study design and objectives of the
AHETF, please comment on the assumption of proportionality between
exposure and AaiH, as a default.  Also, please provide comments on
whether the proposed AHETF study design is adequate to evaluate
proportionality between exposure and AaiH?  What other parameters should
AHETF study designs measure in order to improve the prediction
capabilities of our exposure model?

Panel Response

  SEQ CHAPTER \h \r 1 Most Panel members agreed that the data shown did
not consistently support a linear relationship between exposure and
AaiH.  A linear relationship between AaiH and exposure seems intuitively
logical, but a physical rationale should be developed to support that
hypothesis (or others) in all scenarios.  Several good reasons were
given why a linear relationship might exist but not be detectable within
the PHED data.  Some arguments were presented to accept and/or explain
an apparent non-linear relationship between AaiH and exposure.  And some
arguments were presented that suggest factors other than AaiH that might
be better predictors of exposure.  The following paragraphs elaborate on
these various perspectives.

  SEQ CHAPTER \h \r 1 For exposure within a scenario to be proportional
to AaiH implies that a consistently small fraction of the amount of
pesticide that workers handle is deposited onto their skin.  It seems
logical to expect exposure to be proportional to AaiH under certain
circumstances (such as open cab airblast applications using similar
application equipment moving at similar speeds and in similar wind
conditions).  However, a description of the physical mechanism that
dilutes the a.i. handled into a small but constant fraction of the AaiH
that actually reaches the handler in each scenario has not been made. 
Developing a written array of hypotheses based on physical mechanisms
applicable to each scenario would seem like a good place to start.

In theory, such an argument is merely an extension of the finding that
absorbed pesticide dose increases proportionally with exposure, as
depicted in Figure 4-1.  Data obtained for three pesticides [(atrazine
(Lu, et al., 1997), diazinon (Lu, et al., 2006), and chlorpyrifos (Lu,
et al., 2007)] clearly demonstrated the proportionality between exposure
and dose.  By extending the exposure-dose continuum to the left in
Figure 4-1, the Agency is extrapolating the proportionality as seen in
the exposure-dose relationship to the one involving AaiH and exposure. 

Figure 4-1.  The proposed continuum of the amount of active ingredient
(pesticide) handled [AaiH] to the exposure and to the absorbed dose

In practice, according to the examples from the PHED database presented
to the Panel, the assumption that exposure is proportional to AaiH is
only valid in a few scenarios but is more often invalid.  Three reasons
were proffered by the Panel to explain why such a proportionality was
not found in the PHED data: an “ecological fallacy,” an
“engineering fallacy,” and the statistical uncertainty intrinsic to
the experimental data.

The “ecological fallacy” is the mistaken assumption that all members
of a group have the same characteristics as the group at large.  The
analyses provided by EPA contained several examples of a proportional
relationship being observed within a scenario but not when examining
each study within that scenario separately.  Such a pattern could be
caused by some of methodological or study design differences described
in Charges #1 and #3 or by unique conditions such as clean-up and repair
activities encountered or assessed in some but not all studies.  If unit
exposures are going to be applied in a risk assessment/management
context, it then should hold across most if not all studies and
scenarios.

The reverse may also occur, where proportionality applies within a
closely defined portion of a scenario (such as within a particular
cluster); however, the proportionality fails to apply across a wider
range of work practices or equipment within that same scenario.  Such a
finding may constitute what another Panel member termed an
“engineering fallacy.”  For example, ground boom sprayers may range
from small or almost “antique” machines to state-of-the-art modern
large self-propelled machines with induction bowls, clean water supplies
for washing gloves before removal, glove lockers, automatic folding
booms, and with the operator and the equipment controls positioned in
closed air-conditioned cabs.  If the unit exposures are derived from
“antique” machines, we might expect the value to be over-protective
in the case of modern equipment.  However, if the unit exposures are
derived from modern equipment, they would not be adequately protective
of the smaller, less advanced equipment.  And if the unit exposures are
derived from a mix of antique and modern equipment, they are unlikely to
support a proportionality assumption across the full range of AaiH. 
Therefore, it is important to consider the specific circumstances used
to generate the data in some detail.

Yet another Panel member pointed out that the failure of the regressions
of PHED data presented to the Panel to support a strictly linear
proportionality between exposure and AaiH should not necessarily lead to
rejecting the existence of such a linear hypothesis.  This view is
justified by the results of the analyses of the probable uncertainty
presented as part of our response to Charges #2 and #3.  These analyses
indicate that the probable uncertainty in exposure values derived from
patch dosimetry or biomonitoring data were circa ± one order of
magnitude.  Thus, rather than viewing each individual pair of data as a
point in a scatter plot to be tested for correlation, the data would be
better portrayed as an array of vertical lines ("lines" because the
probable uncertainty of the AaiH values are virtually zero or perhaps
±1% at most).  Given the imprecision of the Y values for regression,
the necessary range of X values (AaiH) must be at least 100( to yield a
truly significant correlation, as proposed by the AHETF study design and
addressed further by the Panel in Charge #6.

In contrast to the above justifications not to reject proportionality,
several arguments were presented to accept a nonlinear relationship
between exposure and AaiH.

Several Panel members questioned the need for strict adherence to
proportionality.  For instance, in PBPK modeling (and in models for many
other fields), nonlinear scaling laws are determined empirically by
estimating the regression coefficients from a log-log regression
analysis, e.g., cardiac output is proportional to body weight to the
0.75 power.  Similar nonlinear relationships might apply to using AaiH
to predict dermal exposures to pesticides.  To apply this viewpoint, the
log-log regression coefficients that give the best fit of the
relationship between AaiH and exposure should be used, whether or not
they are equal to 1.  In a scenario where an increase in the amount of
active ingredient handled is expected to result in an increased
exposure, whether the ratio is 1:1 could depend upon the amount of
technology used when handling pesticides as well as other factors that
one may or may not be able to control.

Physical mechanisms may also predict or be used to explain nonlinear
relationships.  For instance, it may be notable (  SEQ CHAPTER \h \r 1
or given the probable uncertainty in the exposure values previously
described herein, the following observation may just be serendipitous)
that in virtually all of the "other-than-hand" examples presented in the
background document (p. 90-97), the slope of both outer dosimeters was
less than 1 while the slope for the inner dosimeters was greater than 1.
 One can rationalize this bracketing of 1 (unity) if one assumes that
the   SEQ CHAPTER \h \r 1 AaiH per unit of time are approximately equal
within each of the scenarios but that the transfer of pesticide from the
outer clothing to the inner dosimeters is delayed in time (as has been
observed in some experimental field studies).  In this case, the longer
duration scenarios (which also handled more a.i.) would appear to have
more than a linearly proportional dose, i.e., a power greater than
unity.

And finally, some Panelists took the position that at least in certain
scenarios, no correlation should be expected between AaiH and exposure,
and went on to suggest other factors that might be better predictors of
exposure in these situations.

In some scenarios, discrete events might contribute the predominant
fraction of a handler’s total exposure.  One such situation is where
exposure results mainly through contact with a contaminated surface or
with pesticide residues on surfaces that are at steady-state or
saturated, regardless of the AaiH.  For example, a mechanical transfer
device used for loading liquids might limit exposure to concentrated
residue left on the dry-break coupling which is independent of the
amount transferred (but of course is affected by the concentration of
active ingredient in the formulation or tank mix).  Another example
might be a situation where the user is protected in a closed cab when
making ground boom applications so that exposure mainly occurs when
handling contaminated boom / nozzles and the outside of the cab (door
handle), e.g., (Kline et al., 2003).  The majority of a mixer/loader's
exposure might occur each time they handle a bag of a solid formulation
(independent of the volume of that bag or the amount (fraction) of that
bag they actually dispensed), each time they open a can or jug of liquid
formulation (independent of the volume dispensed), or at the moment that
the concentrate is added to the water diluents (independent of the
volume or mass involved).  Under all these scenarios, exposure may be
proportional to the number of discrete events but may not be
proportional to the AaiH.

  SEQ CHAPTER \h \r 1 The previously mentioned short documents that put
forth a plausible argument based on physical mechanisms might also
identify other potential predictors of exposure.  For instance, a
time-based rate of the activity (e.g., speed or rate of application such
as AaiH/hour) might be a predictive factor when the applicator is moving
away from the immediate, relatively localized point of application such
as a horizontal boom (cf., an airblast sprayer), an airplane, or
possibly even if on foot (as in a hand wand application).  The swath of
a ground boom might be another correlate because it determines the
distance from the applicator from which part of that AaiH is released. 
A combination of application rate and ventilation rate (and not
necessarily the amount applied) might underlie the ambient dilution of
an aerosol spray used indoors.

Looking forward, the Agency’s desire to investigate the
proportionality between exposure and the AaiH in future studies is
warranted.  The AHETF also has recognized the need to verify the
“accepted” assumption that AaiH has a 1:1 relationship with
exposure.  To test this hypothesis, studies should be conducted to
evaluate the relationship of exposures under variable AaiHs while
controlling for confounding variables.  If one is able to spread out the
amounts of ingredient handled within each cluster, then one should be
able to estimate the relationship (whether linear or not) between
exposure and the amount of active ingredient handled for each scenario,
and use this information to more accurately estimate exposure for a
given scenario/activity.  If the cluster-based sampling strategy is
used, then the AHETF’s analysis clearly shows the desirability of
obtaining the maximum within-cluster range in the amount of a.i.
handled.  Achieving a range in the AaiH as high as 100 fold will have
the effect of reducing the needed sample size as compared to a 10-fold
difference.  However, given the AHETF’s proposed threshold of 5 lbs
a.i., the 100-fold difference may be achievable for some pesticides
(e.g., certain herbicides applied by ground boom to row crops) but not
for others.  To the degree that the AHETF plans to assess the
correlation of exposure to other variables such as the number of acres
or the duration of mixing or application or the number of events such as
contact with contaminated surfaces or tank mixes, the degree of
potential cross-correlation between these variables and AaiH should also
be considered.  Such cross-correlations may preclude a strictly
observational study from illuminating all other significant predictors. 
In the meanwhile, the Agency is encouraged to develop an array of short
documents that put forth a plausible argument based on physical
mechanisms that would justify either using a default assumption of a
linear relationship between exposure and AaiH and could help in
designing "purposive non-random sampling” to support other predictive
correlates of exposure.

In conclusion, Panel members agree that a great many factors are
associated with field conditions, e.g., application techniques,
equipment types, meteorological conditions, formulations of pesticide.  
 SEQ CHAPTER \h \r 1 A well-designed observational study such as that
proposed by the AHETF may illuminate the relationship between exposure
and AaiH, but a controlled experimental study beyond the scope of the
studies envisioned herein may be a better way to ascertain this
relationship.  Given this dilemma, the Agency may wish to consider
establishing its own criteria for the strength of evidence needed either
to accept or to depart from the existing default assumption of direct
proportionality.  In addition, the AHETF may wish to consider a more
strategic allocation of their resources by focusing their efforts on
fewer exposure scenarios but employing sufficient monitoring units to
establish a lack of proportionality with greater certainty.

Agency Charge

5)  Within-worker and Between-worker Variability

The proposed AHETF study design does not include true worker replicates
and is not intended to examine the issue of variability within workers. 
The AHETF notes that to appropriately investigate this issue would
require significantly more sampling and resources.  They propose,
however, that their single-day exposure distribution results can be used
to evaluate longer term multiple day exposures by placing reasonable
limits on expected intra-class correlation coefficients (ICC):  they
indicate that, from their own research and review of the literature, the
ICC is likely to be between 0.3 and 0.5 over relatively short periods of
time (e.g., seasonal), and likely to be even lower over longer periods
of time.  

Please comment on the AHETF’s approach to estimating the number of
samples (MU) needed to determine within worker variability and their
conclusion on the importance of measuring such variability in their
proposed studies.

Panel Response

The Panel agreed that exposure data collected from observational studies
has the potential to address all three potential sources of variation
identified in the background documents: within-handler, between-handler,
and between-study.  Within-handler variability is defined as the
variation among different measurements on the same individual doing the
same or a similar task under similar environmental or other conditions
(or what is referred to as “repeated measurements” in the background
document).  Between-handler variability is defined as the variation
among different individuals doing the same or a similar task, possibly
under the same but typically under differing environmental or other
conditions.  Between-study variability is defined as the variation among
different individuals doing the same or a similar task under different
environmental or other conditions but at either different locations
(separated by miles rather than meters) or times (separated by days
rather than hours).  Often between-study variability is confounded with
between-handler variability since researchers may have to go to
different locations and/or different times to find handlers doing the
same task.  It may also be confounded with within-handler variability if
the same handler is measured at multiple time points or locations.  A
necessary and basic component of any quantitative risk assessment is a
good measure of the variability expected from independent handlers doing
the same or a similar task under similar conditions.  In some assessment
scenarios, the variability term required may be the sum of
between-individual and between-study variability.

The term “repeated measurements” may have a different meaning for
different researchers.  In statistics, a repeated measurement would
occur if one unique handler were to do the same task multiple times and
his/her exposure were measured separately for each repetition of the
task.  Other researchers use the phrase repeated measurements to refer
to more than one measurement of a task, regardless of whether the
measurements are on one or many individual handlers.  Still other
researchers may think of repeated measurements as different tasks
measured individually on one handler.  The concern in all of these
definitions is that measurements cannot be considered truly independent.
 The issue is further confounded if the less defined term
“replicates” is used, although for many, replicate measurements
imply independence of responses.

The important issue for design and analysis of exposure studies is the
potential for exposure measurements to be correlated.  The responses
from different handlers doing the same or similar task are often assumed
to be independent (uncorrelated), that is, the exposure of one handler
is not expected to affect or be affected by the exposure of another
handler.  Within-handler measurements on the other hand are typically
assumed to be correlated.  The “simple sample variance” computed
from all applicable exposure measurements will be an underestimate of
the true risk-related variability unless the expected correlation of
measurements in the data is taken into account in the estimation
methodology.  If the database consists of only uncorrelated
measurements, as would be the case where within-handler data were
specifically excluded, then the simple variance would be an acceptable
estimator of variability for the risk assessment.

The AHETF proposal argues, fairly strongly, that the within handler
source of variation is unimportant and/or too expensive to measure given
the objective of the resulting data to support benchmark or minimal
adequacy requirements for Tier I and Tier II risk assessments.  The
proposal also suggests that measurements be taken in studies (clusters
of measurements) that are linked to specific locations and times.  This
design can also result in significant but moderately-sized correlations
among within-study measurements.  The concern with within-study
correlation is that handlers doing the same or similar tasks at one site
and time may produce similar exposure values because the measurements
are taken under common environmental or other conditions.  The measure
of similarity used to quantify within-study variability is the
“intra-class correlation due to clustering” (referred to herein as
the ICC, cf., the intra-class correlation coefficient in the charge) and
the range of interclass correlations for measurements taken over short
periods of time was reported in the background documents to be between
0.3 and 0.5.

The true model for MU exposures (Equation 5.1 below) is a modification
of Equation (1) in the Procedures for Determining the Required Number of
Clusters and Monitoring Units per Cluster to Achieve Benchmark Adequacy
(AHETFb, 2006).

  				Eqn. 5.1

  where 

        Eijk   = 	the exposure obtained for MU j in cluster i in
repeated measure k

        Hijk   = 	the amount of a.i. handled for MU j in cluster i in
repeated measure k

        Qijk   = 	the exposure for MU j in cluster i in repeated measure
k normalized by amount of a.i. handled

        GMQ = 	the population geometric mean for normalized exposure

        Ci     = 	a random effect for cluster (study or condition) i

        Wij   = 	a random effect of MU j in cluster i

        Rijk   = 	a random effect of MU j in cluster i for repeated
measure k

In this model, all three random effects Ci, Wij and Rijk are assumed to
be normally distributed with means of zero and variances of VC, VW and
VR, respectively.  Note that the distribution of random effects could
alternatively be parameterized using a total variance term [V], an
intra-class correlation due to clusters term [ICC], and a within-handler
correlation term [Rww].  The formulation reported in Equation 5.1 is to
clarify potential confusion that might exist about the definitions of
the ICC and the Rww terms.

It was pointed out that regardless of the source of the data,
within-worker variation will always be confounded with errors in the
monitoring technique and the chemical analysis (referred to as
“probable uncertainty” within responses to Charges #1, 2, and 3). 
That is, one can never really measure the true residual error.

The Panel felt that the AHETF arguments to de-emphasize within-handler
variability in section 5.3 (AHETFa, 2006, Technical Summary Document)
are clear and compelling.  In particular, the AHETF argues that:

The combination of the probable uncertainty inherent in exposure
measurements and the typical influence of uncontrolled environmental
factors on the measured exposure would result in repeated measurements
that would be expected to demonstrate low correlation or Rww values
between 0.2 and 0.4 (page 19 of AHETFa, 2006).  The argument for low Rww
values is derived from limited published literature and not on an
analysis of relevant PHED data.

The between-handler data which will populate the AHED database is
expected to support Tier I and Tier II risk assessments that focus on
cumulative exposures over long time periods.  The distribution of
individual long-term cumulative exposures will be best described by the
between-handler distribution regardless of whether the Rww is 0 or 1.

The between-handler data distributions could be used to simulate both
within-handler and between-handler variability in any probabilistic
(Monte Carlo) risk assessment by specifying and drawing from a
distribution of Rww values such Rww is between 0 and 1.

In addition, the Panel noted that expecting to conduct repeated
measurement on each handler would constrain the eligibility of handlers
to participate, thus introducing a selection bias.

One Panel member estimated the within-worker correlation coefficient
using the repeated measures data presented in Figure 5-1 of the EPA
Review document.  Variance components were estimated using a one-way
random effects model with 10 individuals having from 2 to 6 repeated
measurements for a total of 39 observations.  The within-worker variance
component was estimated as 0.38, and the between-worker variance
component as 2.5, resulting in an estimated within-subject correlation
coefficient of 0.9.  This estimate was significantly greater than 0, the
value assumed by AHETF.  This result supports the conclusion that little
additional information would be gained from repeated sampling of one
individual.

While there was little interest among Panel members in increasing
dramatically the total number of measurements taken per scenario by
requiring repeated measurements for every handler in every scenario,
there seemed to be some concern that EPA might be missing an opportunity
by not pressing or investing in some limited repeated measurements.

Current literature on within-handler correlation is small and
problematic.  The Panel suggested that an analysis of the within-handler
data currently within PHED could serve as a starting point for
understanding within-handler correlation.  For instance, it would be
prudent to attempt to estimate Rww using PHED data (as mentioned above
using Figure 5-1 of the EPA Review document) for different exposure
scenarios.  The analysis would be purely exploratory, providing some
evidence to support the assumption that the within-worker correlation is
relatively small and/or bounded.  Additional data might be necessary to
provide evidence-based justification for limiting the range of Rww,
something that might be needed if indeed the AHETF approach is used to
incorporate within-handler variability in a future assessment.  Any
probabilistic risk assessments will want to incorporate both
within-handler and between-handler variability.  The AHETF approach is
always going to be weaker than an approach that is based on estimates
that are backed by actual data.

Those Panel members with in-the-field experience performing the types of
studies being considered suggested that there are significant challenges
to getting good data and that requiring repeated measurements can result
in large increases in both effort and analytical costs.  Others felt
that adding limited repeated measurements should be relatively cheap,
especially when compared to the cost of starting a new study (new
location and time) and or recruiting new handlers.  Repeating
measurements on the same or next day would provide the most information
for the least additional investment.  The need for information on the
costs associated with the various aspects of sampling was pointed out a
couple of times in the discussion.  Cost information would help to
better inform the decision on repeated measurements.  There are
statistical techniques that can be used to adjust sampling efforts among
the three variance components to essentially get the most precision for
the least total cost.

There was also recognition among Panel members that the time window for
pesticide application is often narrow in agricultural situations and
that the number of tasks per worker per pesticide per year may be few. 
These constraints substantially limit both the number of sampling
opportunities and the number of eligible workers, especially within the
same geographic region (since all applicators are usually working within
the same time window), making repeated measurement studies infeasible
for some scenarios.  Given these pressures, the Panel recommends that
unless repeated measurement data are specifically allowed and properly
handled within AHED, in cases where a participant withdraws from a study
that a new worker be recruited rather than using a previously sampled
worker.

The discussion on cost and timing did not reduce some Panelists’
interest in seeing repeated measurements performed in at least a couple
of important scenarios.  Although a repeated measurements sampling
strategy may not be possible for all exposure scenarios, the Panel
recommended that EPA and the AHETF determine whether within- and
between-worker variability might be evaluated for selected scenarios
where application frequency and logistics are favorable.  As an example
of a scenario where repeated measurements might be important, one Panel
member suggested orchard programs where repeated chemical applications
are often performed every 7 to 10 days by the same handler.  One Panel
member noted that when EPA and other researchers use these data to
examine potential predictive determinants of exposure, the best data for
identifying predictive exposure factors is to have measurements taken
under different levels of the suspected factor on the same individual
(i.e., to use the individual as the block or its own control).  Thus,
some limited repeated measurements on handlers in the database could
result in more powerful identification of predictive exposure factors.

Repeated measurements data need not be balanced for subsequent data
analysis when using modern mixed model software.  Although analyses by
such mixed model statistics can estimate within handler variability,
between-handler variability, correlations, overall average exposure, and
associated confidence intervals, such software (or this ability) may not
be compatible with the type of software used for PHED or being
contemplated for AHED/BHED.  Thus, the presence of just a few workers
with repeats in the database might raise practical data management and
statistical analysis issues. A number of questions will need to be
addressed in the AHED interface to ensure that exposure estimates and
uncertainties are valid.  These questions include: Should the repeat
measurements be considered independent?  Should a mixed model be used to
estimate overall average exposure and associated confidence intervals? 
Or should the repeats be removed (or automatically averaged) and simple
data analysis techniques be used?

One Panel member argued strongly that repeat exposure studies on the
same and different handlers are needed to identify the biological
differences between and within handlers that are important in
interpreting biomonitoring data (were such BM data to be collected). 
Differences in metabolism, body weight, age, gender, BMI, and ethnicity
can account for much of the between-individual variability in
biomonitoring data.  But this variability will need to be gauged against
the within-individual variability estimated using measurements from task
replicates on the same handler.  Repeated exposure studies will also
support an understanding of the relationship between passive dosimetry
and biomonitoring results.  That Panel member suggested that based on
the estimated Rww, only a few repeated samples; say 2 or 3 would be
sufficient to provide this understanding.

The Panel discussed extensively those factors relating to repeated
measures that could give rise to ethical or other issues at a Human
Studies Review Board review.  Risk would be increased if a handler were
asked to do something they would not normally do, use pesticides that
they would not normally be handling, or use amounts of pesticide that
they might not usually use.  Some opportunities for repeated
measurements may not add risk, but further justification to HSRB would
be needed if the handler were asked to do the same task multiple times
if it were something that might be typical for the scenario but that
with timing and or amounts would not be typical in that handler’s
normal work assignment.  Another issue was the use of scripting to
achieve repeated measures.  Scripting takes the handler outside of
his/her usual mode of work and hence has the potential to change risks. 
The Panel suggests that these issues be carefully addressed by the
AHETF.  Although unrelated to repeated measures, some of the measurement
techniques such as the use of whole-body dosimeters, create a burden on
the handler not typically encountered in a set of tasks.  This
additional burden might be of concern to the Human Studies Review Board.

Agency Charge

6)	Sample Size: Number of Sites and Subjects per Scenario/Activity

The Agency’s goal is to ensure that monitoring studies rely on sample
sizes that adequately represent the range of exposure of people who
engage in a particular handler scenario and activity.  It is also
recognized that occupational monitoring studies are costly and have many
logistical obstacles.  The Agency is also concerned about limiting the
numbers of participants in these types of studies in accordance with the
ethical requirements described in Subpart K (40CFR26) and the recent
criteria outlined by the Agency’s Human Studies Review Board.  The
Agency’s current guidelines recommend 15 monitoring units for each
scenario.  In addition, the AHETF has provided a rationale for the
number of samples in their study design. 

Please comment on the uncertainties associated with the Agency’s and
AHETF’s recommended number of monitoring units.  In your comments,
please include any recommendations you may have regarding specific
statistical analyses that may assist the Agency in developing better
understanding of these uncertainties and characterizing them in a
complete and transparent manner in Agency assessments based on these
data. 

Panel Response

To design a monitoring program that may be used for a variety of
regulatory purposes by various organizations challenges the developers
to anticipate all possible future applications while keeping costs in
line.  The Panel agrees with the background criteria given in the first
paragraph of the charge above.  The Panel also agrees that the current
PHED database needs to be updated and modernized.  Our response to this
charge is divided into four main parts.  The first part comments on the
uncertainties associated with the Agency’s and AHETF’s recommended
number of monitoring units.  The second and third parts comment on the
planned clustering within the study design.  And the fourth part
comments on sample selection and bias within the study design.  Several
comments made during discussion regarding statistical analyses are in
Appendix C or were covered herein in our response to Charge #4.

Recommended Number of Sampling Units

The appropriate number of Monitored Units is integral to the goals of
database users.  The   SEQ CHAPTER \h \r 1 Panel noted that the
benchmark objectives for data adequacy as established by the AHETF,
listed on slide 17 in the presentation by Larry Holden entitled
“Summary of Statistical Issues for the AHETF Monitoring Program:
Sampling Methods and Sample Sizes,” may not support the goal of the
Agency, stated in the above Charge, to “adequately represent the range
of exposure of people who engage in a particular scenario and
activity.”    SEQ CHAPTER \h \r 1 The AHETF’s primary and secondary
“benchmark objectives” for data adequacy were to meet for all
exposure scenarios a degree of data accuracy to within K-fold when
exposures were normalized by amount of active ingredient (a.i.) handled,
with K proposed to be 3; and for some scenarios, users of the data
should be able to “distinguish” between complete proportionality and
complete independence of exposure and amount of a.i. handled,
respectively.

The first benchmark objective is to estimate the parameters of the
distribution of dermal exposure to an adequate level of precision.  The
criterion chosen (that the upper 95% confidence bound for the parameter
be no more than K times the parameter and the 95% lower confidence bound
be no less than the parameter divided by K) makes sense under the
lognormal assumption.  A closely related criterion, giving similar
results, is to require that the upper 95% confidence limit be no more
than K2 times the lower 95% confidence limit (K above times K below
equals K2); this might be easier to communicate and has the advantage of
not requiring the true parameter value to be explicitly in the formula.

The Panel also discussed the need to think strategically about the
allocation of resources and to establish sampling priorities for
scenarios.  We were told that regulatory personnel have not had
difficulty in specifying what, for them, would be an acceptable value of
K.  A default value of K = 3 seems reasonable, although it need not be a
fixed value.  Scenarios with higher exposure might warrant allocation of
more MUs.  Alternatively, the K value could be larger for a scenario
with a larger than typical MOE, permitting an acceptably smaller sample
size.  In addition, if EPA has a strong need to estimate exposure levels
in the upper tail of the exposure distribution (the 90th percentile for
example), more samples will be required than suggested in the background
documents.

A similar cautionary note would apply if the EPA uses the 95th
percentile for risk assessment in the future, in which case the
AHETF’s analyses indicate that under cluster-based sampling the sample
size required to estimate the 95th percentile within a 3-fold accuracy
is quite sensitive to the specified GSD and ICC.  If the actual GSD or
ICC is greater than anticipated, then much larger sample sizes will be
needed to achieve the desired accuracy.  EPA and the AHETF should
consider building in one or two “check points” after a certain
amount of new data is collected to evaluate assumptions about the GSD
and ICC so that any needed refinements to scenario sample sizes can be
made.  The examples shown to the Panel assumed that the most extreme
upper percentile of exposure that anyone would want to estimate was the
95th, in which case K = a 3-fold relative accuracy can likely be
achieved with 5 clusters and five monitoring units per cluster.  This
means 25 monitoring units per scenario.  However, this sample size will
be inadequate if at a future time it is necessary to estimate the 99.9th
percentile.  One Panelist noted that it appears that 11 or 12 clusters
would be needed to achieve K = a 3-fold accuracy for the 99.9th
percentile.  Thus, the total number of monitoring units would be more
than doubled to 55 or 60.

One Panelist gave three reasons why it may be advisable to have at least
50 monitoring units per scenario: to estimate upper percentiles of
exposure, to make effective comparisons among scenarios, and the
possibility of measuring within-worker variation.  This Panelist thinks
that the value of the database will be greater in the future if costs
are controlled now by making a thoughtful choice of scenarios in which
to sample heavily rather than by using small samples in all scenarios. 
If at a future date a sample size is found to be inadequate for
regulatory purposes, it will be impossible to return and get more
observations that are consistent with the original sample.  It will be
much easier to do a complete study of new scenarios as they are needed.

If the data will be used to compare scenarios (e.g., to compare
different application methods with the same pesticide), then the design
needs to be considered more as a stratified sample and there have to be
enough observations within each stratum to make the test powerful enough
to be worthwhile.  If the sample size meets the first benchmark
objective, it may also be good enough for this objective, but it would
be worth checking this out.

The Panel also noted that all of the sample size calculations presented
are dependent upon the assumption of lognormal distributions, and any
attempt to estimate extreme upper percentiles from a small sample will
be an extrapolation into the tail of the assumed distribution.  The
larger the sample, the more robust will be the ability to validate that
assumption and support the conclusions.  To understand the robustness of
the cluster design (see below) to the assumption of lognormal
concentration values, the Panel recommended that sample size simulations
be performed using an alternative skewed distribution for concentration
values, such as a Gamma distribution.

Clustering within a Scenario

The discussion in the Procedures for Determining the Required Number of
Clusters and Monitoring Units per Cluster to Achieve Benchmark Adequacy
(AHETFb, 2006) was relatively straightforward, clear, proper, and
representative of good statistical thinking.  The Panel compliments Dr.
Holden on creating a clear conceptual model for the sampling process and
following it through to the particulars of the sampling design.  The
cluster sampling design proposed by the AHETF makes good sense, as there
are cost savings in sampling a number of pesticide handlers in a single
field operation.

The Panel notes that if one is going to fix (set) the total number of
monitoring units, then it is generally better to have more clusters
within each scenario/activity and fewer numbers of handlers within each
cluster than it is to have fewer clusters and more handlers per cluster.
 The usual practice in survey design when there is an intra-class
correlation within clusters is to consider the costs of getting to a
cluster relative to the costs of sampling monitoring units within a
cluster.  The optimal cluster size and the number of clusters can then
be chosen to minimize the variance of a parameter estimate subject to
constraints on the total cost.  For this study, AHETF has determined
from experience that the intra-class correlation is modest in size and
that it is often practical to monitor five pesticide handlers within a
cluster.  

The Panel has no problems with the values of geometric standard
deviation and intra-class correlation used in their examples.  However,
the variation in intra-class correlations observed to date comes from
sparse data and variability in monitoring methods and can’t be
attributed to specific scenarios.  The Panel would expect that as more
monitoring data are collected, it will become evident that some
scenarios may have very different intra-class correlations from others,
and adjustments in the numbers of clusters and monitoring units within
clusters that are sampled may need to be considered (see prior comment
on “check points”).

The Panel also noted that for a variety of reasons not all clusters are
likely to come in neat units of five MUs.  Similarly, it may be very
difficult to assess each scenario with only five clusters for each
scenario.  The science will be greatly served by not requiring five
clusters for each handler task.  The AHETF might explore simulations
where the average cluster size is five but with some variability in
cluster size to assess the robustness of results with respect to cluster
size.

Given that analyses presented by the AHETF indicated that greater sample
size efficiency was generally achieved by increasing the number of
clusters rather than increasing the number of monitoring units per
cluster, clearer guidelines are needed for cluster selection so that the
addition of new clusters will achieve some desired degree of dispersion
on the variable of interest and to guard against “parsing” the
population of interest into more clusters just to limit sample size
requirements without improving stratification or representation.

To summarize the above comments, the Panel believes that the Agency’s
and AHETF’s recommendation to have 5 handlers per cluster with
approximately 5 clusters for each scenario/activity is reasonable at
this point in time.  The most important issue for the Agency to consider
is what value of K for a K-fold accuracy is appropriate and reasonable.

Cluster Selection

The Panel noted that the reports being reviewed had information about
what data will be collected, but there was not much information about
how the data will be collected.  It is important to consider such
questions as:

How will the clusters within each scenario/activity be selected?

How will the monitoring units within a cluster be selected?

Will AHETF have control over the amount of ingredient handled?

When will the data be collected?

With respect to the first two bullets above, the Panel suggests that the
definition of a cluster and how clusters will be selected be clarified
and tightened (see also the above comment on “parsing”).  For
example, clarify whether clusters are defined by crop, state, county
within state, crop-state combination, geographic region, etc.  Would the
cluster definition need to be scenario dependent?  Also, the Panel
suggests that the AHETF provide additional evidence to support the
notion that “geographic differences” are important for establishing
clusters.

Sample Selection and Bias

With respect to the third bullet above, there was some concern expressed
that the sample selection was being adversely affected by the secondary
benchmark objective proposed by AHETF to be able to elucidate predictive
relationships with exposure such a proportionality with AaiH.  Their
"Technical Summary Document" (p. 17, 21, and especially p. 40-41)
describes their plan to use "purposive non-random sampling to achieve a
diversity of major factors likely to influence exposure" which they list
as "the amount of active ingredient handled, number of unique workers,
and number of geographic locations."  These concerns start with
feasibility but extend to bias and the ultimate value of the database.

With regard to feasibility, their page 41 is not clear with respect to
what aspect of handling will (or can) be varied to achieve their example
range of 5 to 2000 pounds of a.i. in "a period of time that is
representative of a full workday," viz., from at least 4 to a maximum of
8 hours.  Can the same type of application equipment span this range? 
Will work practices be the same across this range?  What predictive
artifacts are introduced by attempting to maximize the range of AaiH? 
And is manipulating the conditions they select to maximize AaiH worth
the loss of representativeness within the resulting data?  Throughout
these efforts to maximize variability, the Panel recommends that
currently approved practices be used and that maximum amount under
currently approved limits not be exceeded.  

The AHETF Technical Summary Document refers to sampling such that there
is a “bias toward conditions that might yield higher exposures.” 
This bias is in conflict with the statement that users “must also
assume that the purposive sample of the MUs approximates some type of
probability sample from the target population.”  Further clarification
is needed on this sampling bias and how it might affect exposure
distributions and subsequent uses of the database.  A detailed critique
of the AHETF sampling plan is contained in Appendix C.  The conclusion
of that critique is that the underlying distribution of pesticide use
conditions must be considered either during sample selection or during
data analysis.  The impact of this bias is further complicated by the
lack of a database that documents the distribution of tasks, activities,
or pesticide use information within any given exposure scenario from
which to judge or compensate for the biases introduced by the sampling
selection design being proposed (see also the last comment in our
response to Charge #1).

Recommendations Regarding AEATF

The AEATF Study Plan is dealing with a very different situation and is
much more amenable to experimental control.  In particular, it should be
feasible to increase the sample size for any scenario at a future date
if more observations are needed.  The proposal to take 15 monitoring
units initially is adequate to give an overview.  For probabilistic
assessments and determination of exposure at extreme upper percentiles,
15 units will not be enough.

The Panel was also asked whether the AEATF study should be conducted at
one simulated site or three field locations.  The simplest way to answer
this is to try both options a few times in a pilot study and compare the
results.  Perhaps three field sites should be treated as blocks or
strata rather than clusters.  In summary, the most important difference
between the two studies is the possibility of increasing the sample size
at a future date.



REFERENCES

  SEQ CHAPTER \h \r 1 Brouwer DH, Boeniger MF, Van Hemmen J.  2000. 
Hand Wash and Manual Skin Wipes.  Ann. Occup. Hyg. 44(7):501-510.

Durkin P, Hertzberg R, Diamond G.  2004.  Application of PBPK Model for
2,4-D to Estimates of Risk in Backpack Applicators.  Environmental
Toxicology and Pharmacology 16:73-91.

  SEQ CHAPTER \h \r 1 Kline AA, Landers AJ, Hedge A, Lemley AT, Obendorf
SK, Dokuchayeva T.  2003.  Pesticide exposure levels on surfaces within
sprayer cabs.  Applied Engineering in Agriculture 19(4):397-403.

Lu C, Anderson LC, Morgan MS, Fenske RA.  1997.  Correspondence of
salivary and plasma concentrations of atrazine in rats under variable
salivary flow rate and plasma concentration.  J. Toxicol. Environ.
Health 52:317-329.

Lu C, Rodríguez T, Funez A, Irish R, Fenske RA.  2006.  The assessment
of occupational exposure to diazinon in Nicaraguan plantation workers
using saliva biomonitoring.  Ann. New York Acad. Sci. 1076:355-365.

Lu C, Rodriguez T, Thetkathuek T, Funez A, Pearson M.  2007.  Using
Salivary Biomarker in Exposure and Risk Assessments for Organophosphorus
Pesticides: Possibilities and Pitfalls. (submitted).

McDougal JN and Boeniger MF.  2002.  Methods for assessing risks of
dermal exposures in the workplace.  Critical Reviews in Toxicology
32:291-327.

Williams PL, Thompson D, Qiao G, Monteiro-Riviere N, Riviere JE.  1996. 
The use of mechanistically defined chemical mixtures (MDCM) to assess
mixture component effects on the percutaneous absorption and cutaneous
disposition of topically exposed chemicals.  Toxicology and Applied
Pharmacology 141:487-496.

APPENDICES

  SEQ CHAPTER \h \r 1 Appendix A:  Definitions and Abbreviations

This glossary of terms was prepared not only because terminology is so
important within the topic of dermal hazards from chemicals but more
specifically to help assure that the responses of the Panel are both
internally consistent and properly interpreted.  Most of the terms in
this glossary come from the Review of Worker Exposure Assessment Methods
document prepared for the Panel by the U.S. Environmental Protection
Agency, Health Canada, and the California Environmental Protection
Agency (referred to herein as the "Review").  

Absorption

  Factor  = 	A measure of the flux or amount of chemical that crosses a
biological boundary such as the skin (% of the total exposure that is
absorbed).   p.38

    SEQ CHAPTER \h \r 1 AHED  = 	Agricultural Handlers Exposure Database
to be developed by the AHETF.

  BHED  = 	Biocide Handlers Exposure Database to be developed by the
AEATF.

  bias  =  	One form of a bias is a consistent or overall difference
between the result and the true value being estimated, sometimes called
a systematic bias.  Results from a known bias (such as sample recovery
efficiency or representative measurements) can be adjusted for by a
simple calculation.  A bias exists for virtually all of the dosimetry
and biomonitoring methods discussed in this report.

Another form of bias is in sample selection in which some members of the
population are more likely to be included than others.  While the
existence of sample selection bias can be identified, its effect and
methods to adjust the results are not always known.

biological

monitoring  =	[BM or biomonitoring] "is usually employed to quantify the
absorbed dose (also referred to as body burden)."  The Agency accepts
monitoring data based on the collection of biological media such as
urine or blood.  Biomarker data can also be used for predicting
exposures.  In addition to biological sampling media, the Agency also
requires that additional supporting pharmacokinetics and/or
pharmacodynamic information be submitted that can be used to develop
exposure and/or risk estimates. p. 35.  For the purposes of this
discussion, biomonitoring was restricted to urinary excretion.

By definition, dose estimates resulting from biological monitoring
"integrate exposure across all routes.  Such an estimate of absorbed
dose, which avoids potential confounding from assumptions of dermal
penetration or inhalation retention, may be more useful in assessing
risk than route-specific doses estimated from passive dosimetry." 
p.41-42

BM  =	See biological monitoring.

  SEQ CHAPTER \h \r 1 CV  = 		coefficient of variation calculated as the
standard deviation divided by the group's mean (sometimes also called a
coefficient of variance).

Dose  =	the amount of chemical absorbed from exposure to a pesticide in
a given scenario.  p. 21

Dermal

  exposure  =	defined by U.S. EPA (1996b) as the process of pesticide
residue deposition onto the skin, as well as the measurement of the
deposited residue.  p.38

Exposure  = 	Amount deposited on the surface of the skin that is
available for dermal absorption or amount that is inhaled, also referred
to as potential dose.  p. 21  See also Dermal exposure.

Factor-set  = 	a large number or combination of factors that
characterize a specific setting, e.g., climatic conditions, combinations
of equipment, task times, etc.  As used in Appendix C, each unique
combination of factors is denoted by the symbol Si. 

  SEQ CHAPTER \h \r 1 inner

  dosimeter  =a dosimeter worn under the handler's work clothing and
held in some way against the skin.

LOAEL  =	Dose level in a toxicity study, the lowest dose level where an
adverse effect occurred in the study (mg pesticide active ingredient/kg
body weight/day).  p.38

MOE  =	Margin of exposure, value used by the Agency to represent risk or
how close a chemical exposure is to being a concern (unitless).   p.38

  SEQ CHAPTER \h \r 1 MU  =		Monitoring Unit or one person whose
exposure is assessed via measurements such as biomonitoring or passive
dosimetry.

NOAEL  =	Dose level in a toxicity study, where no observed adverse
effects occurred in the study (mg pesticide active ingredient/kg body
weight/day).  p.38

  SEQ CHAPTER \h \r 1 outer 

  dosimeter  =	a dosimeter worn outside of regular work clothing (in
some cases, the regular work clothing has comprised the outer
dosimeter).

passive 

  dosimetry  =	[PD] "employs some sort of physical monitor that traps
residues from the surface of the skin (i.e., they absorb or remove, such
as a dosimeter or a wash) to determine dermal exposure (i.e., also
referred to as a potential dose).  Passive dosimetry methods (e.g.,
patches, gloves, dosimeters or washes).  p. 35  For the purposes of this
discussion, passive dosimetry [PD] included hands or face/neck washing,
patch dosimeters, and whole-body dosimeters.

Patches  =	Various forms of absorbent pads (usually made of gauze but
sometimes of alpha-cellulose) placed on the body at fixed locations.

PD  =	see Passive Dosimetry. 

PHED  =	Pesticide Handlers Exposure Database consists of dermal and
inhalation exposure measurements compiled by EPA beginning in the late
1980s from a wide variety of scenario specific pesticide handler
exposure studies.

Probable

  uncertainty =The expected diversity of the result based on an analysis
using propagation of error theory.  Its magnitude is based on the
precision of the individual measurements used to calculate a result and
the form of the calculations being made.  For linear calculations,
probable uncertainty may be characterized by a standard deviation,
coefficient of variation, or geometric standard deviation after
adjusting for the multiplication or division factor used to achieve the
result, such as after adjusting for sample recovery from a dosimeter or
in urine.

  variability  =	The diversity of the population being studied.  The
true variability of a population is only known if repeated measures of
whatever is being studied can be made with an accurate and completely
precise method.  Variability is usually characterized in terms of the
observed variation or geometric standard deviation of a random sample of
such measurements.  

WBD  =	A whole-body dosimeter preferably composed of full-length cotton
underwear, cf., a WBD composed of coveralls and worn as outer clothing. 
The latter suffers the combined problem of not retaining all of the
pesticide deposited onto them and passing through them some fraction of
the pesticide that was deposited.

Appendix B: Expanding the Concept of Grading Data

The current data grade (p. 31 of the Review) is based (in part) on two
relevant criteria: the "% Lab recovery" and the "CV for Lab recovery"
which is also a percent.  The current “Grading” evaluation of PHED
data is based on these two criteria independently, i.e., a lower
letter-grade is assigned if the data exceed either criterion.  The
effect of this interpretation is shown by the small shaded rectangles in
Figure B-1 below.  In fact, these two criteria are two characteristics
of the same data that interact and can be combined within the concept of
“data grade” to indicate the magnitude of the probable uncertainty
within the resulting exposure value, as calculated by the equation
below.  The “magnitude of uncertainty” is shown as a multiplying
factor herein but could be viewed as a percent by subtracting one and
multiplying by 100 (as a percent, it would be analogous to a CV, but
interpreting values that exceed several hundred percent was deemed
easier as a multiplier such as (4( than as the equivalent (500%).

 	Eqn. B.1

 

Figure B-  SEQ Figure \* ARABIC  1 .  A narrow view of graded data
quality

Each of these two criteria can contribute to the magnitude of the
uncertainty within the resulting calculated exposure; however, only
together do they define the magnitude of that statistically probable
uncertainty.  Figure B-2 shows an array of possible values of these two
criteria that have the same magnitude of probable uncertainty
represented by the relatively large shaded triangles comprising each
grade.  One can discern several points from this figure.  One is the
broader range of possible criteria that result in the same quantitative
effect on the probable uncertainty as an indicator of data quality.  The
Agency could consider broadening their criteria for "Data Grade" by
using the above equation without decreasing the precision of their
resulting data.  

 

Figure B-  SEQ Figure \* ARABIC  2 .  A broader view of graded data
quality

Second, one can see in Figure B-2 that the CV of the Recovery has a
greater effect on the Grade than the mean Lab. Recovery.  In fact, one
can see that a sufficiently precise estimate of a poor sample recovery
can yield a smaller probable uncertainty in exposure than an imprecise
estimate of a good sample recovery.  An extension of this observation is
that the reliability of the result (as characterized by the magnitude of
the probable uncertainty) cannot be defined by only specifying a mean
recovery (any value on the X-axis in either figure) without also
specifying a CV.  

 

Figure B-  SEQ Figure \* ARABIC  3 .  The impact of the CV of the mean
Lab. Recovery on 

the magnitude of Probable Uncertainty

This observation is perhaps more visible in Figure B-3 in which the
Y-axis is the probable uncertainty and the data's CV is a parameter. 
The Agency’s current definition of Grade is still depicted as the
shaded zones in the lower right, and the suggested extended definition
is shown as horizontal dashed lines at constant values of probable
uncertainty.  Notice that one cannot predict the variability of the
result by only specifying a value on the X-axis (a % Lab. Recovery)
without also specifying a CV.  Thus, the Agency should consider adding
CV limits to their Field Recovery and Storage Stability criteria.

A third observation is that the magnitude of the probable uncertainties
shown in these figures (even using the Agency's worst current Data Grade
of "D") are all much smaller than the range of variabilities in the data
discussed within their Review (such as a 3× range in dermal absorption
or differences between passive dosimetry and biomonitoring).  The
magnitude of these probable uncertainties will also be seen to be much
smaller than other uncertainties when the same principle behind Equation
B-1 is applied to passive dosimetry data in Charge #2 and to urinary
biomonitoring data in Charge #3.  The same principle (with less specific
measures of CV) could also be applied to other data attributes such as
the lack of complete body monitoring in much of the PHED dosimetry data,
the variability within the calculated results implicit within each form
of passive dosimetry, and the impact of the simplistic method currently
used to account for undetectable samples (due to censoring).  Any one of
these quantitative attributes of the data within PHED probably has a
greater impact on the variability within the results than does sample
recovery but are currently not part of the grading criteria. 
Furthermore, incorporating these other measures of data quality into the
“data grade” is likely to illuminate the poor quality of much of the
existing PHED data and the potential for the new AHED data to improve
that quality by orders of magnitude.

Appendix C:  A Critique of the AHETF Study Design

The Panel’s understanding of risk assessment is that the exposure
value used in the risk equation is expected to be “representative”
of the average exposure that would be experienced by the population
potentially exposed to the chemical.  For probabilistic risk
assessments, individual exposure values are drawn from a distribution of
exposures that are expected to describe the distribution of long term
average exposures for individuals in the population potentially exposed
to an active ingredient.  The Agency should look at the proposed
sampling design through the lens of its “representativeness.”

The first assumption made (by AHETF) is that a surrogate cluster
sampling model that assumes underlying random selection can be used to
estimate sample sizes even though the proposed sampling methodology does
not advocate random sampling for clusters.  The second assumption (by
AHETF) relates to the normality of variance components in the
nested-effects linear model on log-normalized exposure.  The real
concern herein is with the acceptability of using the surrogate random
sampling model.

The discussion in Sections 5.1 and 5.2 of the AHETF Technical Summary
background document (AHETFa, 2006) is excellent in that it provides a
good framework for thinking about sampling for exposure assessment.  The
following will use a slight modification of their conceptual model to
illustrate a technical concern with the sampling protocol that is being
proposed. 

The goal of the AHED dataset is the estimation of the true exposure, E
for a specific handler task. To collect these data, AHETF proposes a
cluster sampling or hierarchical sampling design in which clusters (or
studies) are essentially examinations of handlers performing the handler
task of interest at specific locations and times.  As mentioned in the
background documentation, there exist a very large number of potential
studies.

Figure C-1.  AHETF proposed cluster sampling design

Conceptually each Ci is characterized by specific settings for a large
number of factors, e.g., climatic conditions, equipment combinations,
task times, etc.  Denote each unique combination of factors by Si for
factor-set.  (Note: in the discussion before the Panel, the word
“scenario” was used for “factor-set,” but this caused some
confusion and has been changed here.  The word scenario as used by AHETF
and EPA applies to a handler task.)  In theory, if one knew all of the
conditions that affected exposure one could compute a true average
exposure concentration Ei for each factor-set.  Each cluster/study is
essentially a replicate of some set of factors. Since many of the
factors that impact exposure are continuous, theoretically there are an
infinite number of factor-sets and hence there are an infinite number of
potential studies.

Figure C-2.  AHETF study design conceptualized as a set of factor
conditions

One way to think of the true exposure for a handler task within a
cluster would be as the average of exposures for the handler task across
all possible factor-sets. 

  								Eqn. C.1

But this equation assumes that each possible factor-set has an equal
probability (or frequency) of occurrence, and we know this is not true. 
Pest application tasks have certain climatic conditions that define when
they must or can be performed.  Some types of equipment are more common
than others.  For example, the factors-set contained within the blue
circle in Figure C-2 might represent the more common conditions.  If one
knew the relative frequency or probability of each factor-set, denoted
as Pi, then the true exposure could be estimated as a weighted average.

  					Eqn. C.2

Since the true condition space is continuous, the most mathematically
appropriate way to describe the true exposure is given in the integral
part of Equation C.2.  The term dF(s) essentially describes the relative
probability for each possible factor-set, s, in the total factor space
S.

What does the surrogate random sampling model mean in terms of clusters
C and factor-set S?  With a random selection of clusters, one
essentially selects factor-sets at random for inclusion in the study,
and their proportion within the sample will be their relative frequency
within the population of interest.  This also means that averaging the
study-specific average exposures should produce an unbiased estimate of
the true handler task exposure.  That is, Equation C.3 is an unbiased
estimate of what we need for the risk assessment.

  						Eqn. C.3

From Figure C-2, simple random sampling would result in relatively more
of the sampling effort placed in the blue circle than would be placed
outside of it.

Now consider the “Diversity” sampling approach as proposed by AHETF.
 The approach does not include randomness.  Through thoughtful
considerations of location and time, it is possible that a large number
of factor-sets will be examined.  In fact, the background document to
the Panel seemed to indicate that the locations and times would be
selected to ensure that different conditions would be represented.  The
problem with the Diversity approach is that the relative frequency of
scenarios will not necessarily be considered in the selection of the
scenarios.  Hence, it is possible (if not likely) that only one of a
really common (high relative frequency) factor-set will be included in
the sample set at the same time as one really rare (low relative
frequency) factor-set is included.  The dots and clusters in Figure C-3
depict an array of sampled conditions chosen for their diversity that
are not representative of the relative frequency of factor-sets.  The
relatively common factor conditions found within the blue circle has
small representation compared to the uncommon conditions outside the
blue circle.  When the sample average of estimated exposures is
computed, estimated exposure for the rare scenario is weighted equally
with the estimated exposure for the common scenario, and as a result the
sample average will be a biased estimate of the true exposure.

Figure C-3.  Non random sampling can result in samples that do not
properly represent the distribution of conditions

The issue is even worse if one wants to estimate a distributional
parameter of the true exposure like the standard deviation or an upper
percentile.  The non-probability sample will not produce a faithful
estimate of the population distribution.  Worse, one cannot even predict
the direction of the bias.  The study design would produce overestimates
if the over-represented rare factor-sets produce high exposures, or the
design would produce underestimates if over-represented rare factor-sets
produce low exposures.

This then is the basis for the statement made by the AHETF statistician
that “non-random sampling means that statistical methods alone are
insufficient for generalizing to the target population.”  Most
statisticians and many risk assessors are aware of this problem.  This
problem is not new.  Almost every environmental dataset has this
problem.  The question is whether we want to support the creation of
another environmental dataset with this problem.

AHETF acknowledges the above problems in the background document and
points out that rarely are the true relative frequencies of the
factor-sets known.  At the same time, it is not possible to create a
simple random sample that is guaranteed to appropriately represent a
specific scenario.  The goal of the “Diversity” sampling approach
proposed for populating AHED is to “achieve a diversity of major
factors that are likely to influence exposure” and to attempt to
“capture the major aspects of” the actual distribution of exposures.
 In essence, AHETF will attempt to identify specific Ci to sample that
are “representative” of the whole set of possible conditions such
that the distribution of exposures from the Diversity sample is
approximately equal to the distribution of exposures appropriately
weighted for all factor-sets.

Statisticians have heard this kind of sampling proposal many times but
have never seen a true success.  It is actually impossible to
purposively define a sample that produces a distribution of exposures
that duplicates the true population distribution when one has no
knowledge of the true population distribution to start with.  Rare
events are seldom given proper consideration and common events are often
under represented.  Selecting to get true representation does not work. 
Randomness in selection must be used somewhere in the design to even
approach an unbiased estimate.

So, is this really a hopeless situation?  Not necessarily.  EPA and
AHETF have at this point the opportunity to rethink these issues and
possibly come up with some new approaches that might get them closer to
their stated goals.  While more creative thinking may come up with a
number of feasible alternatives, consider the following approach.

Create a list of all of the factors that are known to impact exposure
levels within a specific scenario.  The list may be long but is not
infinite.

Rank-order the factors by their expected magnitude of impact on exposure
variation.  A Delphi approach might be used with a Panel of expert risk
assessors to accomplish this ranking. 

Select the top two to four factors, and identify for each factor two to
three categories or levels.

Create the set of all possible combinations of factor levels.  Consider
these combinations as strata of the population of interest.  In a sense,
these become the factor-sets of interest, Si*. 

Next assign a weight, wi, to each factor-set, Si* that approximates its
relative frequency in the population.  Sampling theory tells us that
these weights don’t have to be exact for us to gain large improvements
in estimator precision.  Here again, the use of a Delphi approach with a
Panel of agricultural experts could help.

Selection of studies (two options).

Option 1 (see Figure C-4): Select at-random studies and/or MUs for each
factor-set such that the relative number of exposure estimates obtained
for the factor-set equals its weight.  The population exposure estimate
is the average of the estimated exposures for the MUs.

Option 2: (see Figure C-5): Select at random a fixed number of studies
or MUs for each factor set and assign each the factor set weight, wi. 
The population exposure estimate is the weighted average of the
estimated exposures for the MUs.

Figure C-4.  Allocation of studies and MUs to factor-set strata
according to their relative weights.

Figure C-5.  Uniform allocation of studies and MUs to each factor-set
strata.

This approach incorporates both “representation” and
“randomness” into the creation of the database and should result in
an average exposure estimate that is less biased than the “average”
that would be obtained from the Diversity sampling protocol. 

The above approach is quite similar to what AHETF is actually proposing.
 The major difference is that in the approach outlined here an attempt
will have been made first to map out the possible condition space in a
rough categorized way, to assign relative importance to each category,
and finally to sample according to that relative importance.

A number of issues with this approach remain to be addressed by the
Agency and AHETF statisticians.  For example, if a uniform sampling plan
(Option 2) were used, the sample weights would be used to estimate
simple statistics such as the mean or standard deviation.  Much more
complex is the use of these weights to estimate the upper percentiles
for exposure, to estimate the exposure distribution, or to test the
exposure distribution for a specific distributional form. 

It was pointed out that many seemingly unrelated variables are
correlated in pesticide application studies (e.g., the number of acres
sprayed, type of PPE, use of a tractor with a cab, etc.).  It is also
acknowledged that sufficient information on condition factors would be
needed to understand how factors co-vary.  A two-dimensional
stratification based on Factor A and B would be much less effective if
Factors A and B were highly correlated.  It would be better to use two
factors that are relatively uncorrelated in the definition of the
strata.  In this case, each factor could be seen as a surrogate or
representative of a large set of correlated factors (kind of like a
principal component).

Finally, the interface between the User and the AHED dataset should
reflect not only the data contained within the database but also the
sampling design used to collect that data.  Hence, if a uniform sampling
design were used, weighted estimates and correspondingly appropriate
tests for distributions should be presented as a result of a User query.

 A similar but more vexing problem would result if the data showed that
newer equipment and modern techniques actually resulted in higher
exposures.

 While it is possible to measure that variability by cutting a WBD into
small pieces (about the size of a patch) and analyzing each piece
separately, such a study is more amenable to research than to a new data
requirement.

 Despite such a good fit and the fact that several Panelists noted that
KO/W is a part of other models for dermal absorption rate (although here
it is only modeling adsorption), an analysis of dermal recovery for only
two insecticides each at only two points in time (nominally zero and one
hour) is not sufficient to validate this model for all other chemicals.

ഀ׆Ā⍰⸁⼀㄀$摧瀿X

2

F

²

Ä

ò

ó

þ

ÿ

"

F

g

ƒ

—

¨

®

µ

¸

¹

º

â

ã

摧瀿X

!

"

1

6

E

F

S

X

f

g

r

w

x

‚

ƒ

‘

—

¡

º

â

j§

@

hÁ

옍)

@

@

&

@

&

@

&

@

옍

옍

hú

hú

hú

hú

(

)

«

hú

hú

!

`

a

⑈币梄愁̤摧寔Ü਀«

¬

À

Á

$

Ø	@

Ø	@

Ø	@

&

@

&

@

&

@

&

@

&

@

&

@

/

0

1

2

3

4

.

/

0

1

2

4

␃愃̤摧奅Xༀ4

>

?

@

A

U

V

k

o

y

{

|

~

ƒ

„

¦

§

¸

Á

Â

Å

ê

ë

ï

ð

ñ

B

C

G

i

j

v

w

{

|

}

L

M

$

@

@

Æ

@

$

@

@

@

@

@

@

@

@

@

@

@

@

@

@

@

@

Ø	@

hú

hú

hú

hú

hú

hú

 skin and in particular have very different adsorbent and absorbent
properties than human skin, b) some pesticides cannot be analytically
recovered well from cotton, c) breakthrough can happen with gloves, d)
it may be difficult for applicators to do some tasks while wearing the
gloves, and e) wearing cotton gloves may modify the handler’s
behavior.

  Technically, Equation 3.1 only approximates the result of a
propagation of error analysis of the two steps of absorption and
excretion as only one step, but the result is not far off because the
fraction of the dermal exposure that is absorbed is generally much lower
than the fraction of the absorbed dose that is excreted.

 PAGE   

 PAGE   2 

 

OFFICE OF 

PREVENTION, PESTICIDES, AND TOXIC SUBSTANCES

