UNITED STATES ENVIRONMENTAL PROTECTION AGENCY

WASHINGTON, D.C.  20460

	OFFICE OF

	PREVENTION, PESTICIDES, AND

	TOXIC SUBSTANCES

September 8, 2006

MEMORANDUM

SUBJECT:  	Transmittal of Meeting Minutes of the FIFRA Scientific
Advisory Panel Meeting Held June 13 - 15, 2006 on the Analysis of a
Natural Refuge of Non-Cotton Hosts for Monsanto's Bollgard II Cotton.

TO:		James J. Jones, 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 June 13 - 15,
2006.  This report addresses a set of scientific issues being considered
by the Environmental Protection Agency pertaining to the Analysis of a
Natural Refuge of Non-Cotton Hosts for Monsanto's Bollgard II Cotton.

Attachment

cc:

James B. Gulliford					Frank Sanders

Susan Hazen					Richard Keigwin

Margaret Schneider 				William Jordan

Amy Farrell						Douglas Parsons

Anne Lindsay					Enesta Jones

Margie Fehrenbach					Vanessa Vu (SAB)

Janet Andersen					Alan Reynolds 

Debbie Edwards					Sharlene Matten

Steven Bradbury					Leonard Cole

William Diamond					Betty Shackleford

Arnold Layne						OPP Docket

Tina Levine						

Lois Rossi	

FIFRA Scientific Advisory Panel Members

Steven G. Heeringa, Ph.D. (Chair of the FIFRA SAP)

Kenneth M. Portier, Ph.D. 

Janice E. Chambers, Ph.D.

Gary Isom, Ph.D.

FQPA Science Review Board Members

Gary P. Fitt, Ph.D.

Fred L. Gould, Ph.D.

Charles A. Guse, M.S.

David G. Heckel, Ph.D.

Terrance M. Hurley, Ph.D.

Anthony R. Ives, Ph.D.

Michael J. Livingston, Ph.D.

Thomas W. Sappington, Ph.D.

John C. Schneider, Ph.D.SAP Minutes No. 2006-03

A Set of Scientific Issues Being Considered by the

Environmental Protection Agency Regarding:

ANALYSIS OF A NATURAL REFUGE OF NON-COTTON HOSTS FOR MONSANTO'S BOLLGARD
II COTTON

JUNE 13 - 15, 2006

FIFRA Scientific Advisory Panel Meeting,

held at the   HYPERLINK "http://www.ichotelsgroup.com/h/d/hi/1/en/rates"
 Holiday Inn - National Airport ,

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 the Agency presenters, as well as
information presented by public commenters.  This document addresses the
information provided and presented by the Agency within the structure of
the charge.

TABLE OF CONTENTS 

  TOC \o "1-3" \h \z \u    HYPERLINK \l "_Toc145481507"  PARTICIPANTS	 
PAGEREF _Toc145481507 \h  5  

  HYPERLINK \l "_Toc145481508"  INTRODUCTION	  PAGEREF _Toc145481508 \h 
6  

  HYPERLINK \l "_Toc145481509"  PUBLIC COMMENTERS	  PAGEREF
_Toc145481509 \h  7  

  HYPERLINK \l "_Toc145481521"  SUMMARY OF PANEL DISCUSSION AND
RECOMMENDATIONS	  PAGEREF _Toc145481521 \h  8  

  HYPERLINK \l "_Toc145481522"  PANEL DISCUSSION AND RECOMMENDATIONS	 
PAGEREF _Toc145481522 \h  12  

       HYPERLINK \l "_Toc145481523"  SAMPLING AND METHODOLOGY	  PAGEREF
_Toc145481523 \h  12  

        CHARGE   HYPERLINK \l "_Toc145481524"  1:  Adequacy of pheromone
trapping	  PAGEREF _Toc145481524 \h  12  

        CHARGE   HYPERLINK \l "_Toc145481525"  2.  Adequacy of temporal
replication of tbw data	  PAGEREF _Toc145481525 \h  18  

        CHARGE   HYPERLINK \l "_Toc145481526"  3.  Adequacy of tbw
numbers caught and analyzed	  PAGEREF _Toc145481526 \h  24  

       HYPERLINK \l "_Toc145481527"  STATISTICAL ANALYSES	  PAGEREF
_Toc145481527 \h  26  

        CHARGE   HYPERLINK \l "_Toc145481528"  4.  Appropriateness of
pooling tbw data	  PAGEREF _Toc145481528 \h  26  

        CHARGE   HYPERLINK \l "_Toc145481529"  5.  Statistical analyses
of tbw data	  PAGEREF _Toc145481529 \h  31  

       HYPERLINK \l "_Toc145481530"  EFFECTIVE REFUGE CALCULATION AND
MODELING	  PAGEREF _Toc145481530 \h  34  

        CHARGE   HYPERLINK \l "_Toc145481531"  6a:  Estimates of
effective and natural refuges for CBW	  PAGEREF _Toc145481531 \h  34  

        CHARGE 6  HYPERLINK \l "_Toc145481537"  b:  Estimates of
effective and natural refuges for TBW	  PAGEREF _Toc145481537 \h  38   
HYPERLINK \l "_Toc145481539"  

        CHARGE 7:  Adequacy of modeling cbw and tbw to assess
effectiveness of natural      refuge, and importance of preselection for
Cry1Ac resistance	  PAGEREF _Toc145481539 \h  43  

        CHARGE   HYPERLINK \l "_Toc145481545"  8.  Durability of
two-gene resistance affected by presence of one-gene resistance	 
PAGEREF _Toc145481545 \h  56  

       HYPERLINK \l "_Toc145481546"  OVERALL DATA/RESULTS INTERPRETATION
  PAGEREF _Toc145481546 \h  61  

        CHARGE   HYPERLINK \l "_Toc145481547"  9.  Summary assessment by
region of adequacy of natural refuges	  PAGEREF _Toc145481547 \h  61  

  HYPERLINK \l "_Toc145481548"  REFERENCES	  PAGEREF _Toc145481548 \h 
69  

  HYPERLINK \l "_Toc145481549"  APPENDICES	  PAGEREF _Toc145481549 \h 
76  

     Appendix 1  Correct Derivation of Equation 7  HYPERLINK \l
"_Toc145481550"  	  PAGEREF _Toc145481550 \h  76  

       HYPERLINK \l "_Toc145481551"  Appendix 2  Model Uncertainty	 
PAGEREF _Toc145481551 \h  79  

       HYPERLINK \l "_Toc145481552"  Appendix 3  Evolution of Resistance
Hotspots	  PAGEREF _Toc145481552 \h  86  

       HYPERLINK \l "_Toc145481553"  Appendix 4  A Spatially Explicit
2-Loci Model of Resistance Evolution	  PAGEREF _Toc145481553 \h  90  

       HYPERLINK \l "_Toc145481554"  Appendix 5  Post meeting comments
from one Panel member regarding:   County-level Variation in Non-cotton
Cultivated Hosts of the Tobacco Budworm	  PAGEREF _Toc145481554 \h  99  

 

SAP Minutes No. 2006-03

A Set of Scientific Issues Being Considered by the

Environmental Protection Agency Regarding:

ANALYSIS OF A NATURAL REFUGE OF NON-COTTON HOSTS FOR MONSANTO'S BOLLGARD
II COTTON

JUNE 13 - 15, 2006

FIFRA Scientific Advisory Panel Meeting,

held at the   HYPERLINK "http://www.ichotelsgroup.com/h/d/hi/1/en/rates"
 Holiday Inn - National Airport ,

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:  September 8, 2006  				Date:  September 8. 2006

Federal Insecticide, Fungicide, and Rodenticide Act

Scientific Advisory Panel Meeting

June 13 - 15, 2006

ANALYSIS OF A NATURAL REFUGE OF NON-COTTON HOSTS FOR MONSANTO'S BOLLGARD
II COTTON

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

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

Gary E. Isom, Ph.D., Professor of Toxicology, School of Pharmacy &
Pharmacal Sciences, Purdue University, West Lafayette, IN

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

FQPA Science Review Board Members

Gary P. Fitt, Ph.D., Assistant Chief, CSIRO Entomology, Australian
Cotton CRC

Indooroopilly, Queensland, Australia

Fred L. Gould, Ph.D., William Neal Reynolds Professor of Agriculture,
North Carolina State University, Raleigh, NC

Charles A. Guse, M.S., Research Associate, Department of Crop Sciences

University of Illinois, Urbana, IL  

David G. Heckel, Ph.D., Director, Max-Planck-Institute of Chemical
Ecology, Beutenberg Campus, Jena, Germany

Terrance M. Hurley, Ph.D., Associate Professor, Department of Applied
Economics

University of Minnesota, St. Paul, MN  

Anthony R. Ives, Ph.D., Professor, Department of Zoology, University of
Wisconsin-Madison, Madison, WI

Michael J. Livingston, Ph.D., Agricultural Economist, U.S.D.A., Economic
Research Service, Washington, D.C.

Thomas W. Sappington, Ph.D., Research Entomologist, U.S.D.A.,
Agricultural Research Service, Corn Insects & Crop Genetics Research
Unit, Iowa State University, Ames, IA  

John C. Schneider, Ph.D., Professor/Research Entomologist, Department of
Entomology & Plant Pathology, Mississippi State University, Mississippi
State, MS  

INTRODUCTION

The Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA),
Scientific Advisory Panel (SAP) has completed their review of the
Analysis of a Natural Refuge of Non-Cotton Hosts for Monsanto's Bollgard
II Cotton.  Advance notice of the meeting was published in the Federal
Register on March 31, 2006.  The review was conducted in an open Panel
meeting held in Arlington, Virginia, from June 13 to June 15, 2006.  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
Analysis of a Natural Refuge of Non-Cotton Hosts for Monsanto's Bollgard
II Cotton.  The Agency was seeking input from the Scientific Advisory
Panel on whether a natural refuge of non cotton hosts is an effective
refuge to delay the potential for tobacco budworm resistance to the
proteins (Cry1Ac and Cry2Ab2) expressed in Bollgard II® cotton. 
Monsanto Company had submitted an application for the extension of the
FIFRA section 3 registration of the plant-incorporated protectants (PIP)
Bacillus thuringiensis Cry2Ab2 protein and the genetic material
necessary for their production [PV-GHBK11] in event MON 15985 cotton and
Bacillus thuringiensis Cry1Ac protein and the genetic material necessary
for their production [PV-GHBK04] in event MON 15985 cotton.  This
product is intended to provide protection against tobacco budworm,
cotton bollworm, pink bollworm, loopers, armyworms, and other
lepidopteran insects.  The data submitted included the productivity of
tobacco budworm on each alternative host, timing and synchrony of
production on each alternative host, the spatial and temporal scale of
alternative hosts, and modeling efforts to simulate the likelihood of
resistance under different regional scenarios.

	The agenda for this SAP meeting included an introduction of the issues
under consideration provided by Mr. Leonard Cole (Biopesticides and
Pollution Prevention Division (BPPD), OPP).  Issues related to the
tobacco budworm sampling and gossypol analysis were provided by Mr. Alan
Reynolds (BPPD, OPP).  Issues related to effective refuge calculations
and modeling for tobacco budworm and cotton bollworm were presented by
Dr. Sharlene Matten (BPPD, OPP).

	Dr. Janet Andersen (Director, BPPD, OPP) offered opening remarks at the
meeting.

	In preparing these meeting minutes, the Panel carefully considered all
information provided and presented by the Agency presenters, as well as
information presented by public commenters.  This document addresses the
information provided and presented at the meeting, especially the
response to the Agency’s charge.

PUBLIC COMMENTERS

Oral statements were presented as follows:

Graham Head, Ph.D., on behalf of Monsanto Corporation

Mr. Kenneth B. Hood on behalf of Perthshire Farms in Gunnison, MS

B. Roger Leonard, Ph.D., on behalf of Louisiana State University, LSU Ag
Center, and selected cotton organizations in LA

Phillip Roberts, Ph.D., on his own behalf

Richard T. Roush, Ph.D., on his own behalf

Nicholas P. Storer, Ph.D., on behalf of Dow AgroSciences LLC

Mr. Mike Tate on behalf of National Cotton Council

Michael F. Treacy, Ph.D., and Sidney W. Hopkins, Ph.D., on behalf of
Hopkins Agricultural Services, Inc.

Mr. Ray Young on behalf of Louisiana Ag Consultant's Organization

Written statements were provided by:

Craig A. Abel, Ph.D., Agricultural Research Service, U.S. Department of
Agriculture

Michael Adang, Ph.D., University of Georgia

Michael A. Caprio, Ph.D., Mississippi State University

Mr. David Dunlow, President, North Carolina Cotton Producers
Association, Inc.

Mr. Allen B. Helms, Jr., National Cotton Council of America

James C. Jennings, Ph.D., U.S. Biotechnology Regulatory Affairs

Michele C. Mara, Ph.D., and Nicholas E. Piggott, Ph.D., North Carolina
State University

Mr. Bruce Niderhauser, President, North Carolina Crop Consultants
Association

Richard T. Roush, Ph.D., University of California Division of
Agriculture and Natural Resources, Davis, CA

Nicholas P. Storer, Ph.D., Dow AgroSciences

SUMMARY OF PANEL DISCUSSION AND RECOMMENDATIONS

Sampling and Methodology

	

Monsanto relied entirely on pheromone traps to sample populations of
tobacco budworm (TBW), and therefore based all conclusions on analyses
of males only.  The Panel recognizes that both male and female movement
affects resistance evolution, and that our knowledge of TBW movement
patterns and behavior is incomplete.  While the Panel agreed that
sampling TBW populations with pheromone traps was a logical and valid
approach in principle, caution must be exercised in interpreting data
based on pheromone trap sampling, because there are numerous
uncertainties arising from our limited knowledge of TBW movement and
trapability.  For example, the geographic scale to which the trap
captures can be extrapolated remains unknown.  Furthermore, the captures
provide no information about the proportion of moths that leave their
natal habitat, which has a large effect on the rate of resistance
development.

	The Panel identified a number of uncertainties and potential biases
associated with the sampling design itself.  The distribution of traps
among production regions and counties was highly variable and
non-random, with sampling intensity in some regions being very low.  The
sequential subsampling strategy biased estimates of the proportion of
non-cotton-fed males upward.  Of substantial concern to the Panel is
that the novel gossypol analysis technique itself has not been validated
by other laboratories, the threshold of detectability was not reported,
and several key assumptions associated with its use and data
interpretation have not been tested.  Given that the gossypol data are
the foundation of Monsanto's petition, it is critical that EPA 
scrutinize the technique and its assumptions thoroughly.

	Annual variation in effectiveness of unstructured refuge can arise from
annual variation in per-plant insect production from alternative hosts,
density of wild hosts, and percent of total acreage planted to various
non-Bt host crops.  Variable weather can have a significant effect on
temporal and spatial availability and quality of wild host plants.  The
Panel agrees that one year of data in Texas and Tennessee are
insufficient to assess the stability and adequacy of unstructured refuge
in those areas.  The Panel is concerned that the proportion of natural
refuges and alternative hosts may be generally low in parts of the
”MidSouth” region (Arkansas, Louisiana, and Mississippi, in addition
to Tennessee and East Texas).  Some Panel members agreed that two years
of data from North Carolina and Georgia indicate that natural refuge and
continued cultivation of a high proportion of non-Bt corn in these
states could maintain resistance management in the absence of structured
refuge. 

	The Panel noted several sampling biases apparent in the estimation of
the proportion of TBW natural refuge, which were generated by low trap
captures and Monsanto's handling of those situations.  Exclusion of
dates and locations with zero captures introduces a downward bias in
estimates of proportion natural refuge (Rnat) for the MidSouth region. 
Although the "worst-case" counties targeted by Monsanto for sampling as
a group have relatively complete data, they in fact may not be worst
case, since the true worst cases cannot be identified due to lack of
data.

Statistical Analyses

	Monsanto’s approach to pooling the gossypol data was to use a simple
multiple-test method.  The decision to pool was a global one in that
every sample date within a month for a sample location was pooled, or
every sample location within a county was pooled.  While there are some
advantages to this method, it has some significant disadvantages as
well, and the Panel concluded that there are better, more powerful
statistical methods available for determining the appropriateness of
pooling.  The Panel emphasized the need to incorporate biological
justifications into decisions to pool, where to pool, and when to pool,
rather than relying on geopolitical spatial boundaries or calendar month
temporal boundaries.  The Panel suggested an approach to analysis where
two or more alternative generalized linear mixed effects models are
created about plausible alternative hypotheses and to use formal
statistical tests to determine whether any of these alternative models
is significantly better at fitting the data in hand than the null
hypothesis model.  Benefits to this approach include the ability to 1)
estimate the variance components for the deviations from mean percent;
2) determine if these variance components can be related to other
covariates, and 3) incorporate and test for the presence of correlation
in responses one might expect from repeated measurements in time.  None
of these three are possible with the multiple-testing approach.  The
analysis of gossypol fraction differences among counties and months
using the linear logistic model also could be formulated more
appropriately as a generalized linear mixed effects model.  The results
may suggest less pooling, different spatial pooling, and/or different
temporal pooling than Monsanto's multiple-test analysis.  The Panel
agreed that the investment in this more-complex and formal analysis is
warranted, because the estimates of gossypol fraction produced from the
analysis form the basis for subsequent refuge size estimates.

	Although Monsanto supplied only data pooled across dates within month
for each trap and then across traps within county, the Panel performed a
preliminary generalized linear mixed effects model analysis on these
data to address EPA’s charge to “describe” statistical analyses
quantifying variation in the natural refuge across locations and time. 
This analysis was intended to be illustrative of the linear mixed
effects model approach recommended by the Panel, but not intended to be
definitive.  It is not definitive because of problems with the data
(selection biases and spatially inadequate sampling) and because
alternative, equally well-justified, ecologically-relevant schemes for
pooling states into regions other than the East and MidSouth regions
used in this analysis can be envisioned.  In spite of these limitations,
several results are relevant to EPA’s charge to the Panel:  Rnat
declined with Month (June, July, August), and intensity of agricultural
activity (“Hills” vs. “Flats”, see below) and Year (2004, 2005)
were not statistically significant effects in the model.  

Effective Refuge Calculation and Modeling

	In its calculation of Rnat for CBW, Monsanto removed non-Bt cotton moth
production from both the numerator and denominator of the effective
refuge calculation.  The Panel determined that this equation instead
should omit this parameter only from the numerator, and consequently
Monsanto’s equation for Rnat leads to overestimates that can be
substantial.  Based on data reported in Monsanto’s petition,
overestimates of the natural refuge for CBW were largest for Georgia
(37%) and East Texas (44%).

	Furthermore, although Monsanto included the effects of insecticide
sprays in its model for CBW at the request of the 2004 SAP, it did not
do so for the TBW model.  Without correction for spraying in non-Bt
cotton, Monsanto's equations can substantially overestimate the amount
of Rnat for TBW, with the magnitude of the correction depending on the
proportions of the three refuge options used.  Monsanto provided
insufficient data to make these corrections.  Estimates of Rnat were
below 0.05 for some “worst-case” counties in Mississippi and
Louisiana, even without adjusting for insecticide-induced mortality in
Bt-cotton refuges.  With this adjustment, estimates of Rnat could be
even lower.

	For both CBW and TBW, the calculations assume that males are well-mixed
at the spatial scale of variation in habitat types.  Strong
circumstantial evidence suggests this probably is not true for TBW. Rnat
will be overestimated in counties with extensive cotton production,
because  underestimates of the proportion of trapped male TBW
originating from non-Bt cotton will lead to overestimates of Rnat.  This
could be an important source of overestimation if the gossypol assay
gives false negatives, a concern of the Panel given Monsanto's
description of this technique.

	The estimates of Reff and Rnat are imprecise, due to uncertainty in the
estimates of the parameters in the equations.  Imprecision is possibly
large for TBW in those counties used as scenarios for modeling, because
the estimates of Reff and Rnat for these counties are low.  Thus the
Panel notes the necessity of calculating confidence intervals for
estimates of Reff and Rnat.

The Panel agrees that Monsanto’s simple deterministic model has
identified the geographic regions where there is very little risk of
resistance developing (e.g., Georgia).  It also, therefore, identifies
the regions where the risk of resistance developing is greater (e.g.,
the MidSouth region).  However, the model as executed cannot adequately
assess these risks because years to resistance and product efficacy are
insufficient for risk evaluation.  Proper assessment of risk in these
regions requires acquisition of more data, a more robust statistical
analysis of the data, and a more detailed approach to modeling that
includes both spatial and temporal variability in natural refuge.  In
predicting resistance evolution to Bollgard II, there is not only
uncertainty in the estimation of parameters used in models, but also
“model uncertainty.”  Structural or model uncertainty is difficult
to assess, because it may depend on subtle assumptions made in modeling
that have large impacts on model predictions.  The only way to address
model uncertainty is to analyze multiple models that as a group
encompass a range of assumptions about resistance evolution.  Three
principle assumptions of Monsanto's model were challenged, specifically
widespread dispersal of the pests, socio-economic factors affecting the
market share of the products, and the single–locus resistance per
toxin receptor with no cross resistance.  These assumptions will likely
mean that Monsanto is overestimating the time to resistance..  On the
other hand, two broad assumptions made by Monsanto – that there are no
fitness costs associated with resistance, and that resistance
corresponds to a single locus per toxin – might lead Monsanto to
underestimate the time to resistance.  Determining the net effect of the
numerous simplifying assumptions made in the Monsanto model requires a
more rigorous modeling effort.

	Two important dimensions to the question of mosaics of single and dual
gene products are product market share and temporal variability of the
market share.  Monsanto addressed the product market share dimension but
not the temporal changes that should be expected as individual products
lose efficacy.  The Panel conducted a controlled experiment in which the
temporal variability of the mosaic was held constant while varying
product market shares.  The results revealed an opportunity cost of
increasing the market share of the pyramided product in order to reduce
selection for the shared toxin.  This opportunity cost is an increase in
selection pressure for the pyramided toxin.  In a second experiment
where total market share was held constant with increasing temporal
variability in the adoption of the pyramided product, the single toxin
product is on the market longer, which speeds the evolution of
resistance to the shared toxin.  While these experiments suggest that a
quick transition to pyramided toxin products may not always be the best
strategy due to the opportunity cost of increasing selection for the
pyramided toxin, slow transition rates can only be supported by
relatively heavy selection pressure on the pyramided toxin.  The Panel
agreed that the majority of papers published over the past decade using
a variety of modeling strategies have found that quicker transitions to
pyramided toxin products are much preferable in terms of resistance
management.

Overall Data/Results Interpretation

	The Panel cautions EPA that Monsanto’s model is a non-spatial model
implemented deterministically and that this technique may be applicable
to a very limited set of specific geographic situations.  The Panel’s
acceptance of the results for a particular cropping system and pest
species should not be interpreted as a precedent for future
registrations.

	There are many uncertainties, caveats, and assumptions evident
throughout the modeling and analyses presented by Monsanto and revealed
by the Panel discussion around the questions posed by EPA.  Most of
these by themselves might not in fact prove dangerous to IRM for CBW and
TBW, but collectively they represent unacceptably high levels of
uncertainty, especially for the MidSouth region.

	The key data that Monsanto uses to assess natural refuges comes from
the gossypol assay for identifying the non-cotton fraction of the TBW
population.  While this technique is innovative and potentially very
valuable, the Panel had concerns about its validity, accuracy and
repeatability.  Although the Panel received a description of the
analytical technique, it was not complete and important questions
concerning the validity of the technique were identified.  The Panel
thus recommends additional review of the technique by EPA staff,
publication of the assay method in a peer-reviewed journal, performance
and publication of experiments to mimic the conditions experienced by
trapped males prior to analysis, and validation of the methodology by
independent laboratories.  If the gossypol technique withstands critical
scrutiny, the following comments will apply.

	The Panel has noted some potentially serious errors and biases in
Monsanto's calculation of natural refuge for TBW and CBW which must be
addressed.  In addition, the Panel would prefer a more integrated and
comprehensive statistical analysis of the spatial and temporal
variability of refuge estimates and moth trap data to tease out crucial
details regarding the appropriate spatial regions and the critical
temporal periods that pose the most risk.  Despite the potential biases,
some Panel members concluded that for North Carolina and Georgia there
are significant and reliable non-cotton refuges present that should be
adequate to manage Bt resistance in TBW associated with cotton systems
involving Bollgard II cotton.  Current evidence for adequate natural
refuge in the MidSouth region is not convincing.  Because resistance
likely will evolve in areas with little effective refuge, such areas are
of particular concern.  When the estimated proportion of natural refuge
for CBW and TBW is low (5-10%), higher levels of uncertainty attach to a
number of assumptions and calculations for the proportion of effective
refuge.  Estimates of natural refuge below 5% or even (depending on
refuge option) below 1% are not uncommon in the MidSouth.  Alabama was
not sampled at all and, as a transitional state between the East and
MidSouth regions, must be sampled before a recommendation can be made. 
Tennessee and East Texas require additional sampling because both were
sampled only one year.  Other ecologically distinct production areas in
Texas likewise must be sampled.  Only with additional information can an
informed judgment be made regarding the stability and adequacy of the
natural refuge in these areas.

PANEL DISCUSSION AND RECOMMENDATIONS

	The specific issues to be addressed by the Panel are keyed to the
Agency's background documents, references, and Agency's charge
questions.

Sampling and Methodology

Agency Charge

1.  The Panel is asked to comment on the pheromone sampling strategy
employed by Monsanto in which only male tobacco budworm (TBW) were
trapped.

Is this an appropriate sampling strategy?  Can inferences about female
TBW be derived from data gathered exclusively with males?

Panel Response

Summary

	After a review of current knowledge of dispersal behavior of TBW and
other heliothines, the Panel considers the appropriateness of pheromone
sampling for making inferences about the availability of susceptible
insects from natural and managed refuges.  The Panel points out that
despite intensive sampling in some areas, other cotton-producing areas
of equal interest were under-represented and thus the overall picture is
incomplete.  For the better-sampled areas, the Panel feels that despite
the existence of certain biases that remain because females were not
sampled, it is the abundance and distribution of the refuge-generated
male population that is of greater importance, and so trapping males is
an appropriate strategy.  The evidence for an adequate supply of
refuge-generated males does not depend on the total trap counts as much
as it does on the fraction of the trapped sample that tested negative in
the gossypol assay.  There is a bias inherent in the subsampling method
used by Monsanto to estimate this fraction from traps that produced too
many moths to test individually.  Moreover, from the standpoint of
determining the fraction of a trapped sample that originated from
cotton, there are additional potentially serious biases that have not
been adequately addressed.  Extrapolating the results of the gossypol
analysis from the trapped sample to the population as a whole depends on
the validity of assumptions that have not been sufficiently tested,
according to the information provided by Monsanto.

Background

	Pheromone trapping uses chemically-synthesized components of the
female-produced male-attraction pheromone to lure and trap males of a
given species.  It enables the collection of male moths with a
relatively low effort as traps can be left at the same locations for
prolonged periods of time.  Males are usually trapped at night when they
are actively seeking females for mating. Females are not attracted to
the traps.  When the traps are visited and contents removed, the total
number of males since the last visit can be recorded, and the moths are
available for analysis, although these can be in poor condition or even
dead for several days.  Many factors, including weather conditions and
trap location in the landscape, influence how many moths are caught in a
given night by a particular trap.  The actual number trapped also
depends on the effective area from which moths are attracted; on how
attractive the trap is (i.e., the per-moth probability of being trapped
given that they are in that area); and on the number of moths in that
area.  Although large fluctuations in trap catch probably reflect large
fluctuations in the number of moths in the average effective trapped
area, a more precise statement is usually not possible because of the
large variability of other conditions.  It is not generally accepted by
entomologists that trap catches provide a robust and accurate measure of
absolute abundance in a given area.

	The question is whether this is an appropriate sampling strategy for
the purposes of evaluating the durability of transgenic cotton to
resistance development in the pest.  Trap catches were used by Monsanto
to infer two different things:

1)  Abundance of the pest population over time.  These include
comparisons over the course of the season as well as between different
counties and states.  A high local abundance and/or high dispersal is
inferred from a high trap count.  

2)  The composition of the sample.  Among the trapped moths, the ratio
of two different types, classified according to some physical or
chemical analysis, is used to infer the type of plant the moth consumed
when it was a larva.  These include carbon isotopes reflecting the C3 or
C4 photosynthetic pathway, presence of gossypol reflecting consumption
of cotton, or presence of cotinine reflecting consumption of tobacco.

	In the following sections, the Panel addresses these two issues as they
relate to current knowledge of moth abundance and dispersal patterns,
and how reliably and over what scale these are indicated by pheromone
trap counts.  The Panel then considers whether the sampling intensity
and frequency employed by Monsanto was adequate.  Finally, the
appropriateness of this sampling strategy to estimating the likely
contribution of males originating from non-cotton host plants to the
refuge population is considered.

Overall dispersal patterns of adult moths.

	The appropriateness of male pheromone trapping as a sampling technique
should be evaluated in light of what is known about dispersal.  Male and
female dispersion across landscapes are related, at least to the extent
that males having arrived in a cotton field from elsewhere may indicate
that females have moved there as well.  However, Monsanto's categorical
statement that dispersal behavior of males and females is similar for
TBW (  HYPERLINK
"http://www.regulations.gov/fdmspublic/custom/jsp/search/searchresult/do
cumentSearchResult.jsp" \l "#" \o "EPA-HQ-OPP-2006-0217-0013"  Gustafson
 et al. 2005) is inaccurate because this remains an open question.  In
most heliothine species both sexes can undertake extensive pre-mating
movements (Fitt 1991).  Females usually do not reach sexual maturity and
“call” by releasing pheromone until the second or third night after
emergence.  An assessment of the local area and available crops
encountered by newly emerged moths as they search for nectar sources and
potential oviposition sites probably plays a significant part in how far
moths disperse (Fitt 1991) and whether they undertake a truly migratory
movement out of the region altogether (Fitt 1989, Fitt et al 1995). 
Nonetheless, the scale of local movements will usually allow moths to
“sample” the local environment much more broadly than the natal
field where they emerge.  Likewise these species typically display a
period of nectar feeding, oviposition and short-range flights
immediately after dusk each night of their lives which will further
re-distribute moths outside of fields to adjacent habitats or other
fields.

	Differential mobility of males and females might be envisaged since
males actively search for stationary females.  During mating, which
occurs from 1-2 hours after dusk until 3-4:00 am, females are inactive,
releasing pheromone from near the tops of plants, while males engage in
characteristic high-speed, directed flights in search of pheromone
plumes (Fitt 1989).  However, the characteristic mate searching flights
displayed by males when casting for pheromone plumes is almost always
constrained within a field or habitat/crop patch, at least for some
heliothines.  For example, in a field with a large strip of corn
embedded in cotton, males of Helicoverpa armigera cast back and forth
above the corn.  When they crossed the transition between corn and
cotton they flew 5-15 m into the cotton before quickly rebounding and
flying back above the corn.  This phenomenon was observed simultaneously
on both edges of a 2-m tall 5-ha block of corn in a 20-ha block of
cotton (Fitt unpublished).

	Mark-recapture studies have been conducted to characterize the pattern
of adult TBW dispersal and to compare dispersal of males and females. 
Schneider et al. (1989) estimated that about 70% of emerged marked males
moved greater than 18 km (i.e., out of the sampling arena) without being
trapped.  Extrapolation of dispersal curves suggested that all released
males would be within 50 km of the edge of the study area.  The same
study compared the movement of males and females in a large-scale,
mark-release-capture experiment in the Mississippi Delta region. 
Progeny of released females were distributed at least as far from the
release area as were released/pheromone trap-captured males.  Thus, it
appears that female TBW can move as much or more than males under at
least some conditions.  Schneider (1999) used mark-recapture data from
four different years to estimate median movement of males ranging from
9.3 – 23.2 km per generation, and calculated an effective sampling
area of about 20 ha/trap.  They observed patchiness in trap captures at
the scale of several km, but a mostly uniform distribution over a scale
of tens of km.  

	F-statistics based on allozyme variation led Korman et al. (1993) to
conclude that the average diameter of a local TBW population was only 8
km or less.  Data from genetic markers in general indicate low genetic
structuring (and therefore high gene flow) across wide geographic areas.
 At the same time, "typical" gene flow seems to be temporally dynamic,
highest in the spring, and restricted later in the season as evidenced
by increasing FST's (genetic differentiation) as the season progresses
(Han and Caprio 2004).  This pattern of apparent decreased gene flow
during the summer is supported by observations of increasing pyrethroid
resistance as the season progresses, with a drop in resistance at the
beginning of the following year (Luttrell et al. 1991, Sparks et al.
1993, Leonard et al. 1995, Bagwell et al. 2000).  The difference in
per-generation movement observed by Schneider et al. (1989) and
Schneider (1999), where releases were made early in the season, and that
deduced by Korman et al. (1993), where moths were collected in late June
and early July, could be explained by this phenomenon as well (Schneider
1999, Han and Caprio 2004).

Is pheromone trapping an appropriate sampling strategy?  

	The Panel generally agreed that the use of pheromone traps was a
logical and valid way to sample from extensive populations of CBW and
TBW.  Pheromone traps are widely used for broad scale population
monitoring of heliothines, but as noted by many Panel members pheromone
traps are not utilized for making management decisions in the
immediately adjacent crop, because of the inconsistent relationship
between males captured in a pheromone trap and the number and
reproductive activity of females in adjacent fields.  Monsanto has
relied on Leonard et al (1989) to justify the use of pheromone traps on
the basis that there is a positive relationship between trap catches and
egg densities in nearby cotton.  However, the positive relationship
found in that study applied not on a trap-by-trap, field-by-field basis;
but rather over a large local region (5-6.5 km radius), when TBW and CBW
species, 8-10 pairs of traps, and 30-40 fields were all pooled.  The
same conclusion can be drawn in Australia where pheromone traps are
poorly related to egg densities in adjacent fields.  The Panel agreed
that the main utility of pheromone trapping for insect pest and
resistance management is to follow general trends and to obtain an idea
of relative population levels for an area.  In essence the validity of
Monsanto’s data collection is not reliant on a tight coupling of
pheromone trap catches to local population dynamics.  The key point is
that analysis of moths from traps placed adjacent to cotton is to
demonstrate that a significant proportion of those moths have developed
on non-cotton hosts (based on C3/C4 analysis for CBW and gossypol
analysis for TBW).  But the general lack of relationship between trap
captures and local dynamics does indicate that conclusions drawn about
host history are valid only at some geographic scale above that of the
trapping radius, a scale that remains undefined.  Thus the Panel agreed
that the use of traps to collect males as an indicator of the host
source, but not the geographic origin, of local populations is
appropriate, unless host source somehow affects trapability (considered
below).

	However, the distribution of traps among production regions and
counties was highly variable and non-random. Monsanto deliberately
sampled in regions of interest with high cotton production.  In spite of
this intended emphasis, the intensity of sampling in some regions is
very low.  For example, there is no sampling in Alabama. Louisiana,
which grows as much cotton as the whole of Australia and has a “cotton
intensity” score similar to or greater than North Carolina (Monsanto
Report 1, Table 3), was sampled with only four traps in one year and
five in the next. Trapping intensities used in North Carolina, Arkansas,
and Mississippi (about 40 trap locations in each) seem more reasonable
and reliable.  Since the western parts of the Cotton Belt (Louisiana,
Texas) appear to have the lowest percentage of natural refuge and the
greatest month to month variation (e.g., July versus August 2004 in
Bossier County) we need more sampling points for more certainty about
the estimates of refuge.

	Three additional points apply to the conclusion about the general
appropriateness of the male pheromone-trapping strategy:

1) Movement of both males and females affects resistance evolution, and
it must be acknowledged that until additional experimental work is
devoted to tracking females, our knowledge remains incomplete.  The
Panel felt that with a modest amount of additional effort, Monsanto
could have obtained data to determine whether the host-plant history of
males captured in pheromone traps is representative of females in the
same area sampled by the trap.  Females could be caught with a light
trap or by hand, and gossypol content compared to that of males captured
at the same time.  Such a test would not have to be replicated as
extensively as the pheromone trapping, although the comparisons should
be replicated over time at each location during the season.  A few
paired comparisons in a single year across the different regions should
be sufficient to address this question.  The presence of differences at
this point would indicate more extensive experiments are needed to
characterize the female-specific patterns further, find out why they
differ from the males, and assess the consequences to interpretation of
the data and to resistance development.

Mitigating this concern to some extent, the Panel considered that male
dispersal is more important than that of females from the standpoint of
evaluating the distribution of moths relative to refuges.  Given the
efficacy of Bollgard and Bollgard II for TBW it is reasonable to assume
that any moths managing to emerge from a Bt cotton field are likely to
be resistant.  Emergent adults are likely to remain concentrated in the
vicinity of the natal field throughout most of the cotton season while
the crop remains acceptable for oviposition (Farrow and Daly 1987, Fitt
et al. 1989, Schneider 1999, 2003, Han and Caprio 2004, Zalucki and
Furlong 2005).  What is thus required is an abundance of unselected
males moving from the refuges into the Bt cotton area, to be chosen by
resistant females for mating in preference to a resistant male. Thus,
given what is known about TBW adult behavior, it is the mobility of
males originating in refuges the Panel is most concerned about.

2)  The data collected on males give no information about the proportion
that leaves their natal habitat, which has a large effect on the rate of
resistance in all of the models.  Based on his literature review for
Monsanto, Benedict (2005) comes to the conclusion that "pheromone traps
are unlikely to accurately represent the absolute densities of males
emerging from the crop they are placed in...”.  Monsanto purposely
placed most of its traps next to a cotton field.  Therefore it is
impossible to know whether and what proportion of captured males with a
cotton host-history emerged from the adjacent cotton field or were
immigrants from unknown distances.  If the Monsanto estimate of insects
coming from non-cotton sources is biased, it is likely to be biased
towards showing a proportion coming from cotton that is higher than the
true proportion.

3)  Even if the proportion of males coming from non-Bt cotton vs.
non-cotton is the same at the scale of a county, this does not imply
that males are moving at the scale of the county.  Instead, this pattern
could arise even if males show limited dispersal distances, provided the
distribution of habitat types is relatively uniform across the county.

Use of pheromone traps results to infer composition of the sample.

	The question here is whether the analytical method for detecting
gossypol used on a subsample of trapped males provides an accurate
estimate of the fraction of the total TBW in the sampled area that
completed their larval development on cotton.  There are at least three
plausible sources of bias in the estimation:

1) The sequential subsampling strategy itself, as described by Monsanto,
exhibits a bias.  From the traps producing many males, the analysis was
done in batches of ten.  If the initial results indicated a fraction of
cotton-feeding males was greater than 90%, another batch was analyzed. 
Thus the number of males analyzed from a given trap is not independent
of the frequency of cotton-feeding males.  This produces a biased
estimate in favor of the fraction of non-cotton-fed males.

2) There could be a difference between males and females with respect to
the analytical method for detecting gossypol.  The reproductive tissues
of males and females are quite different and may accumulate or retain
gossypol in different amounts.  The two sexes have different activity
patterns and this might also affect retention of gossypol and a
differential rate of change with the aging of the moth.  The amount of
gossypol available to the analytical method may differ between the
sexes.  Monsanto provided no data on the sexes of the individuals used
in the laboratory validation experiments.  If these are primarily
females, they may not be appropriate for calibrating measurements on
field-collected males.

3) There could be other sources of variation in the amount of detectable
gossypol in field-collected males of different ages, collected in
different weather conditions, and dead for different lengths of time. 
These could bias the estimation of cotton-feeders among non-trapped
females as well as non-trapped males in the sampled areas.  This is the
most significant potential bias, as it would affect all of the estimates
of the non-cotton-source males and hence the estimates of the natural
refuge.  For example, if the amount of gossypol in a freshly-collected,
cotton-fed male was high enough to be detected by the method, but the
amount in a cotton-fed male that had been dead in the trap for five days
was not high enough, the degree of bias would depend on the unknown
distribution of arrival times of moths into the trap.

	This last potential source of bias is particularly troubling because
the supplementary information provided by Monsanto (draft manuscript by
Orth, Head and Mierkowski, EPA-HQ-OPP-2006-0217-0013) raises several
questions about the suitability of the analytical method to detect
gossypol in moths.  This is not a quantitative method.  It does not
permit the estimation of the absolute amount of gossypol in a sample by
means of a standard curve.  The detection method depends on the rate of
fragmentation of the Schiff base in the electrospray apparatus, which in
general is not a controllable process and furthermore depends on the
body mass of the moth.  The authors acknowledge problems with the
internal standard as well.  Since quantification is not possible with
this method, the authors use classification criteria that they state are
dependent on the actual apparatus used as well as the absolute amount of
internal standard.  These criteria are evaluated using laboratory-reared
insects that are freeze-dried immediately upon emerging as adults, and
kept frozen until analysis.  These preservation conditions are very
different from those experienced by the field-collected males from the
pheromone traps.  No data are provided to test whether comparable levels
of gossypol detected in the laboratory reared insects would be present
and detectable in such field-collected samples.

	The Panel recommends that EPA ensure that the appropriate technical
expertise is enlisted to evaluate whether the analytical method for
gossypol could be biased, and to determine what sort of supporting
evidence would need to be provided for its validation.

Agency Charge

2.  Monsanto’s TBW sampling and gossypol analyses were conducted over
a two year period (2004 and 2005).  For several states (Tennessee and E.
Texas) data were collected in only one year.  The trends between seasons
were generally consistent, although no statistical/correlation analysis
was performed.  

The Panel is asked to comment on what uncertainties exist from using
data collected from this time period (i.e., 2 years for North Carolina
and Georgia and 1 year for Tennessee and E. Texas) to adequately assess
the potential of natural refuge (i.e., non-cotton hosts) as a substitute
for structured refuge (i.e., non-Bt cotton)?  

Panel Response

Summary

The structured non-Bt cotton refuge gives some assurance that a minimum
% of refuge is present and most importantly that it is interspersed
among the Bt cotton fields.

In determining the adequacy of an unstructured refuge we must consider
that the variation in the refuge provided by wild hosts and non-cotton
crops can be influenced by the following factors:

Year to year variation in quality and insect production from the wild
hosts on a per plant basis.

Year to year variation in the density of these wild hosts.

Year to year variation in percent of total acreage planted to Bt and
non-Bt cultivars of other TBW and CBW host crops.

The evidence from the eastern parts of the Cotton Belt (North Carolina
and Georgia) combined with other published or submitted studies on host
use and relative productivity of different crops (Jackson et al.
submitted, EPA-HQ-OPP-2006-0217-0013) indicates that both species
utilize a broad range of non-cotton hosts in this area, and more
importantly that the relative proportion of moths generated in natural
refuges or non-cotton crops is significant.

Data on non-cotton hosts of TBW in the MidSouth indicate that the size
of the non-cotton refuge may be small in some areas and, because of
potential spatial and year-to-year variation in TBW production from
these hosts, there is a need for more years of data from many MidSouth
locations.

Discussion

	Monsanto has presented data from gossypol analysis of adult budworms
over a two year period for most locations and one year for other
locations.  There is some variation from year to year but the
differences are relatively small.  The question arises of whether the
relative numbers of budworms produced in the unstructured refuges can be
expected to be similar in future years.  There are a number of factors
that must be considered. Some of them fall into the following
categories:

Year to year variation in quality and insect production from the wild
hosts on a per plant basis.

Year to year variation in the density of these wild hosts.

Year to year variation in percent acreage planted to Bt and non-Bt
cultivars of cotton and other host crops.

Each of these will be discussed:

Year to year variation in quality and insect production from the wild
hosts on a per plant basis.

	There is a considerable amount of historical information on the wild
plant species that serve as hosts for the TBW.  However, most of these
studies have been qualitative in nature or restricted to a single year
of sampling, as pointed out by Benedict (2005) (e.g., Neunzig 1969). 
One recent study in Mississippi examined TBW larval and pupal production
from Velvetleaf over a series of years (Carlos Blanco, unpublished data,
Figs. 2-1, 2-2).  The year-to-year variation revealed in this study is
dramatic.  If the results of this study are indicative of expectations
for variation found in TBW production from other wild hosts then we
expect enough variation among single wild hosts to influence the overall
number of TBW from the total set of wild hosts.  Mueller and Phillips
(1983) provide data on larval numbers for TBW on 4 wild hosts over a
two-year period, but the two sites (one each year) were about 70 miles
from each other.  Their data show substantial variation as in the Blanco
study.  It is important to recognize that a high proportion of larvae on
wild hosts can be parasitized and that this varies spatially and
temporally (Mueller and Phillips 1983, Norris and Kogan 2005), so
density of larvae may be a poor indicator of adult production. 
Stadelbacher (1981) examined wild host use by TBW over a 12 year period.
 Unfortunately, he only reported average numbers of larvae over all
years.

 

Fig. 2-1.  Tobacco budworm larval production in cotton and velvetleaf,
Washington Co., MS.  Data from Carlos Blanco (unpublished).

 

Fig. 2-2.  Tobacco budworm moth production in cotton and velvetleaf,
Washington Co., MS.  Data from Carlos Blanco (unpublished).

2)	Year to year variation in the density of these wild hosts.

	Data on year-to-year variation in abundance of wild hosts typically
have not been collected for pest management decisions.  However, many
wild hosts of TBW are known to grow in roadside areas and in
Conservation Reserve Program (CRP) lands.  Research in the 1980s in the
MidSouth identified the primary early season hosts of TBW, and a program
was proposed to decimate populations of these species by changes in
management of roadside areas and other weed control measures (see
Mueller et al. 1984).  This brings up the fact that general changes in
the management of roadside vegetation not directed at wild host
management could impact the densities of wild hosts.

3)	Year to year variation in percent acreage planted to Bt and non-Bt
cultivars of  

cotton and other host crops.

	Monsanto presents data from Federal Government analyses showing that
the proportion of crop types grown in the Southeast and MidSouth have
been relatively stable since 1995.  It is worth examining cropping
patterns over a longer period.

	Kennedy and Storer (2000) examined variation in the number of hectares
grown to wheat, corn, soybean, and cotton over a 15-year period from
1980 through 1995.  The acreage planted to some crops more than doubled
or were halved in a single year-to-year period (Tables 2-1, 2-2).  Over
the 15-year period, the number of hectares of cotton in North Carolina
and Georgia increased more than 10 fold.  In Mississippi, there was a
more than six-fold change in corn acreage.

Table 2-1.  Maximum year-to-year changes in production area (hectares X
1000) of four agronomic crops in each of three states during the period
1980 through 1995.  (Modified from Kennedy and Storer 2000.)

Table 2-2.  Minimum/Maximum number of hectares (X 1000) planted with
four agronomic crops in each of three states during the period 1980
through 1995.  (Modified from Kennedy and Storer 2000.)

	Although relative suitability of crop plants to pests is not expected
to vary as much as that of wild hosts because of genetic uniformity and
cultural farming practices, there can be substantial variation in
year-to-year TBW production from cotton.  Schneider (2003) estimated the
percentage contribution of cotton to the overwintering population of TBW
over 7 years (1996-2002) in northwest Mississippi.  Annual variation in
the contribution of cotton was high: (mean + SD 9.1 + 10.6%; range
0-29%;  n = 7).  If variability in host use during the period of
selection for adaptation to Bt cotton (June-August) were similar to that
observed for September, then confidence in the one to two years of data
that is currently available for the MidSouth is insufficient. 

	If there is an increase in profitability of one of the crop hosts of
TBW or CBW, it is expected that the acreage of this crop will increase
and this could affect the size of the refuge.  Data from China shows
that as Bt cotton increased the profitability of growing cotton, the
acreage planted to cotton in certain provinces increased (Table 2-3)
(Kongming Wu, unpublished data).

Table 2-3.  The planting history of Bt cotton and other host crops of
the bollworm during 1998-2005 in Anci County, Hebei Province and Xiajin
County, Shandong Province, China.  (From K. Wu, unpublished.)

Location	Year	Conventional

Cotton (%)	Bt

Cotton (%)	Maize

(%)	Peanut

(%)	Soybean

(%)	Total area









	

	Stable isotope data from Monsanto’s application and from Gould et al.
(2002) indicate that a large proportion of CBW are developing as larvae
on corn during specific periods of the growing season.  Therefore,
non-Bt corn is serving as the major refuge for CBW.  It is important to
note that some corn in the South and in the Midwest is Bt corn that has
a single Bt toxin, Cry1Ab.  Currently, the proportion of Bt corn in the
Midwest and especially in the South is low, but this may change as new
cultivars are produced that have stacked herbicide tolerance, Cry1Ab,
and another Bt toxin for corn rootworm.  Instead of serving as a refuge,
corn could be a selection agent for Bt resistance.  Data in Gould et al.
(2002) as well as from other sources (Sparks et al. 1975, Hartstack et
al. 1982, Hendrix et al. 1987) indicate that CBW is migrating into the
MidSouth from corn and other hosts in Mexico early in the growing
season.  Late in the season CBW seems to be migrating from the Midwest
to the South.  There is a need for more data on this phenomenon because
such migration means that calculations of refuge must include the
proportion of Bt corn in the Midwest and in Mexico.

	Even with all of the uncertainties above, some Panel members believe
that the data from North Carolina and Georgia offer substantial evidence
that an adequate unstructured refuge currently exists because of the
overwhelming percentage of TBW and CBW moths with a non-cotton history. 
Furthermore, this outcome is easily explained, indeed expected, because
of the current presence of large acreages of alternative preferred hosts
in these regions, namely tobacco and peanuts.  These two crops make up a
substantial proportion of the acreage, but this may change with recent
federal buy-out programs and changes in crop stabilization programs. 
Additionally, the proportion of Bt corn in the Carolinas and Georgia is
currently small.  If that proportion increases, the non-cotton refuge in
these areas could decrease substantially.

	A single year of data from areas of Texas and Tennessee are
problematic, and given the low readings in several areas of the Delta,
more extensive spatial and temporal data would be in order there as
well.  Unpredictable weather events or patterns can influence many
variables including host crop planting date, timing of alternative host
availability progressively through the season because of effects on host
maturation, length of generations, dispersal (including proportion
engaging in facultative long-distance migration), and local flight
behavior.  Monsanto is rightly concerned to use "worst case" scenarios
for its model, but one year of data cannot represent a typical range of
weather scenarios, much less worst case weather scenarios.  BPPD (2006)
expressed concern about the adequacy of two years of sampling, mainly
because land use patterns could change, but temporal availability of
certain alternative hosts due to weather is also important.

	The Texas data seem particularly inadequate.  Four of the five counties
sampled are near one another from the same area of East Texas, and the
other is from the Coastal Bend area.  Other ecological areas of Texas
should be sampled, such as the Lower Rio Grande Valley and the Texas
High Plains.  It is especially important that the latter be tested
because it is such a dry environment, and it is hard to imagine that
alternative local hosts could play much role there for TBW.  Benedict's
(2005) review of the literature indicates that from mid-June on, cotton
is the main host, and the only abundant host, in the Delta region of
Mississippi.  As pointed out by Benedict (citing Sparks et al. 1993),
pyrethroid resistance has been slow to develop in North Carolina,
presumably because there are abundant alternative hosts that provide
refuge from these insecticides.  In contrast, the biggest problems with
pyrethroid resistance in TBW are in the Mississippi Delta, which implies
there is little natural refuge available.  In volume 3 of Monsanto's
petition (Head et al. 2005), it is pointed out that pyrethroid tolerance
in CBW may be building in Texas and Louisiana as well, again bringing
into question the effectiveness of natural refuges in these areas.

Agency Charge

3.  In some counties/states, extremely low numbers of TBW were trapped,
with some traps collecting only one insect.  In Tennessee, TBW numbers
were so low that data were not reported at all for 2004.  In addition,
cotton monitoring efforts have been recently hampered by low
availability of TBW samples (possibly due to a suppressive effect of Bt
cotton).  

Do low overall numbers of TBW trap captures in some areas affect the
ability to assess the effectiveness of natural refuge for IRM?  What
conclusions, if any, should be drawn from the failure to capture
Bt-susceptible TBW at particular sites?

Panel Response

EPA (BPPD 2006) expressed a concern that low numbers of TBW could mean
that insufficient numbers of susceptible moths would be available to
mate with resistant moths emerging from a Bt cotton field.  The absolute
numbers of susceptible moths emerging in an area by itself is not what
matters, only the proportion of the local population that is emerging
from Bt vs. refuge.  If the global population is low, then few eggs are
going to be laid in Bt cotton, and correspondingly fewer resistant
insects will emerge.  What is important is that those that do emerge
from Bt cotton are still overwhelmed proportionally by susceptible
individuals, even though their absolute numbers may be low.  The primary
problem with a low TBW population density is that it prevents assessment
of the relative contribution of the natural refuge compared to non-Bt
cotton.

The causes of low trap catch are uncertain.  At least 5% to 20% of
cotton in all counties growing Bt cotton is non-Bt cotton.  Hence, a
population of TBW sufficiently large to be detected with pheromone traps
would be expected in counties with historically detectable TBW
populations.  The rarity of TBW in some of these counties brings into
question the efficacy of currently required structured refuge – at
least under some circumstances.  Alternatively, although TBW pheromone
traps are fairly sensitive measures of the presence of males seeking
mates in a region, absence of catches does not necessarily indicate a
lack of moth population (see Panel response to Charge 1).  It is also
possible that there has been a regionwide decline in TBW numbers due to
the overall replacement rate decreasing below 1.0 owing to a high
percent of the population dying on Bt cotton.

	Using the generalized linear mixed model described in the Panel
response to Charge 5, multiple analyses of Rnat were performed with
minimum sample size (MinSS) set to 1, 2, 5, 10, and 20 adults.  [The
Panel notes that similar results are obtained using sample size (i.e.,
number of moths tested) as a continuous factor in SAS PROC GLIMMIX
analysis of Rnat.].  The resulting least square mean estimates of Rnat
for each County by Month combination were used in turn as the dependent
variable in a SAS PROC GLIMMIX analysis with Class variables Region and
Month, Continuous variable Ln[MinSS], and interaction terms Region X
Month and Ln[MinSS] X Region(Month).  For the data divided into three
regions (East, “HillsMidSouth”, and “FlatsMidSouth”; see Panel
response to Charge 5), Region and Ln[MinSS] X Region contributed
significantly to variation in Rnat because of differences between the
East region and the two MidSouth regions but not because of any
difference between the two MidSouth regions.  The results of a two
region (East and MidSouth) analysis are given in Fig. 3-1.  This
analysis shows that Rnat declines at a significantly faster rate with
Ln[MinSS] for the MidSouth region than the East region, and Rnat for the
East region does not vary significantly with Ln[MinSS].  In the
MidSouth, the higher estimates of proportion natural refuge at lower
minimum sample sizes suggests that small sample sizes cause a net
upwards bias in the estimates of proportion natural refuge.

 

Fig. 3-1.  Effect of minimum sample size (Ln scale) on least square mean
estimates of the proportion of pheromone-trapped TBW originating from
natural refuge by oviposition month and region.

The Panel notes several sampling biases apparent in estimation of the
proportion of TBW natural refuge, which were generated by low trap
captures and Monsanto's handling of those situations: an upward bias for
Rnat estimates in the East region due to the conditional sampling
protocol used and a net bias of indeterminate sign in the MidSouth
region due to the existence of both upward and downward biases of
unknown relative magnitudes.

1) Sites or counties with extremely low abundance of TBW (as measured by
pheromone trap catches) were excluded from subsequent analyses.  Given
that Rnat June-August, the period of greatest TBW reproduction in
cotton, is higher for the MidSouth region the lower the minimum sample
size used in the estimation process (Fig. 3-1), it appears that Rnat is
higher in areas where the population density of TBW is lower.  10.6%
(22/207) of the county X month sampling combinations were excluded from
Monsanto's data set because 0 moths were captured.  Thus, these
exclusions introduce a downward bias in estimates of Rnat for the
MidSouth region.

2) Sampling bias for counties and traps within counties with the
“highest and most consistent” catches.  The Applicant states the
following (p. 12, Head and Gustafson 2005): “Analyses [for gossypol in
TBW] were focused upon those counties, and trap locations within
counties, where numbers were highest and most consistent throughout the
season.”  Given the negative correlation between numbers of
pheromone-trapped moths caught and estimated Rnat, sampling bias for
counties and for traps within counties with the “highest and most
consistent” catches introduces a downward bias in estimates of Rnat in
the MidSouth region.

3) Additional sets of moths analyzed where the non-cotton contribution
was 10% or less. The Applicant states the following (p. 12, Head and
Gustafson 2005): “For the locations analyzed, moths were analyzed in
sets of 10 for each trap-date combination.  Where the non-cotton
contribution was 10% or less for any trap-date combination, additional
sets of 10 moths were analyzed, if available.”  This protocol makes
sample size conditional on the observations and thus introduces an
upward bias in estimates of Rnat for both the East and MidSouth regions.

Statistical Analyses

Agency Charge

4.  Monsanto used the Fisher’s Exact Test to determine whether the
gossypol data could be pooled.  Data were pooled for individual traps
(i.e., for multiple collection dates for each month) and for counties
(i.e., including all traps within a county for each month).   

The Panel is asked to comment on Monsanto’s approach to pooling the
gossypol data.   

Panel Response

	In assessing Monsanto’s approach to pooling the gossypol data, the
Panel felt that more could have been done in the data analysis to assess
the appropriateness of pooling.  The analysis approach used by Monsanto
consisted of a large number of independently performed Chi Square tests
in which the P-value associated with the test was computed using an
exact-enumeration technique (Fisher’s exact method).  A number on the
Panel felt that while this approach was appropriate for the specific
pooling task it was inadequate for addressing some of the other issues
that came up in the pooling discussion.

	Of importance to understanding the conclusions of the Panel’s
discussion is a clear understanding of the overall goals of the
statistical analysis related to pooling.  One goal was the determination
of whether count data from multiple dates within a month for a sample
location could be pooled, and, conditional on a decision that pooling
was appropriate for all locations, the subsequent determination of
whether to pool count data for sampling locations within a county (for a
given month).  The decision to pool was a global one in that every
sample date within a month for a sample location was pooled or every
sample location within a county was pooled.  The decision by Monsanto
based on their analysis was to pool both sample days and sample
locations, resulting in a dataset for subsequent analysis that consisted
of counts of gossypol-positive and gossypol-negative sample results for
each county for each sampling month.

	The goals of the statistical analysis can be formalized in a set of
hypotheses.  The null hypothesis of the overall analysis is that for
each county/month combination there is one gossypol distribution,
indexed by its average gossypol fraction.  A number of alternative
hypotheses are conceptually considered, one being that locations within
counties (after pooling within sample days within locations for a month)
have varying gossypol distributions (or varying average gossypol
fractions) and the other being that sample days within sampling
locations have significantly different gossypol distributions (or
varying average gossypol fractions).

	The multiple-test method used by Monsanto has the following favorable
properties with regard to the issue of pooling:

The individual tests are simple to perform and easy to understand.

County/month tests are examined individually and hence one can easily
identify conditions where there are strongly significant differences.

The two alternative hypotheses identified above can be examined in two
separate sets of multiple tests allowing different decisions on each
issue and allowing consideration of the issue of pooling locations
within county/month combinations to be conditional on the decision to
pool sample days within sample locations.

The level at which a difference is considered to be statistically
significant can be adjusted to take into account the potential for Type
II errors.  One Panel member suggested that setting the threshold for a
significant difference at P = 0.20, for example, while resulting in more
significant differences would also reduce the chances that sites within
counties or sample days within sites that are truly different are missed
and hence are pooled.

	The multiple-test method used by Monsanto has the following unfavorable
properties with regard to the issue of pooling:

The method does not take into account experiment-wise error (most
statistical methods texts have a chapter on multiple comparisons in
which experiment-wise error is discussed, see for example Ott and
Longnecker 2001, Chapter 9 or Zar 1996, Chapter 11).  Consider that each
of the individual tests (in this case each Chi Square test) has a
certain probability of resulting in a wrong conclusion (say each is
performed at the Type I error probability of 0.05).  When a large number
of these tests are performed on a population where in fact there are no
real differences (the null hypothesis of the pooling analysis) then we
would expect to see a fraction of the tests being statistically
significant.  So if the Type I error for each individual test was 0.05,
we would expect to see about 5% of the multiple test results to be
significant.  Statistical analysis methods that account for
experiment-wise error typically do one of the following: i) reduce the
Type I error of the individual tests so that only a few very significant
test results are considered important, or ii) change the structure of
the individual tests so only the larger differences are considered
significant.

The method does not provide a formal way of directly testing the two
alternative hypotheses.  The conclusion to pool is subjective,
determined primarily by looking at the number of significant tests. 
Different individuals might make different assessments from the same
data.

Second level hypotheses cannot easily be tested.  For example it is not
easy to determine if pooling over years or some spatial pooling other
than at the county level is appropriate.

The method does not take into account all of the data collected in the
decision making process since sample days with very low counts
(including zero counts) and sample locations with low counts were not
included in the analysis.

= 0.05, for 2004, 5 of 102 (4.9%) tests of dates within months for
individual traps were statistically significant, and at = 0.20, 21
of 105 (20%) tests were statistically significant.  For 2005, at =
0.05, 9 of 170 (5.3%) tests of dates within months for individual traps
were statistically significant and at = 0.20, 34 of 170 (20%)
tests were statistically significant.  These combined results suggest
that pooling sample days within months at individual sites may not be a
problem.  But, for pooling across traps within each county, at =
0.05 for 2004, 3 of 38 (7.9%) tests of dates within months for
individual traps were statistically significant and at = 0.20, 12
of 38 (31.5%) tests were statistically significant.  For 2005, at
= 0.05, 7 of 60 (11.7%) tests of dates within months for
individual traps were statistically significant and at = 0.20, 15
of 60 (25%) tests were statistically significant.  These combined
results provide much less support for pooling across traps within a
county.

	A number of Panel members pointed out the need for biological relevance
to support pooling, and that pooling to county level or state level
(i.e., to geopolitical boundaries) may not be appropriate.  There was
concern that some localized (temporal or spatial) differences that could
be important might be lost in a decision to pool to the county by month
level.  Some argued that the county level was too large, based on gene
flow estimates of < 8 km for TBW during the time when cotton is the
favored host (Korman et al. 1993), and therefore the decision to pool,
when to pool, and where to pool requires much more analysis than that
provided in the Monsanto report.  It was also pointed out that the
question of whether unstructured refuges can safely replace structured
refuges has little to do with what the data collectively say about
pooling, or for that matter what they say on average for the Cotton
Belt.  The issue is what can be extracted from these data to address the
question of unstructured refuges at appropriate temporal and spatial
scales.  Finally, it was recommended that Monsanto provide a biological
justification for pooling that goes beyond simply a data analysis.

	One Panel member suggested the use of a moving fixed-length time window
(say four or six weeks) to allow assessment of different temporal moving
period lengths on the tests results (e.g., either simple smoothing of
within sample site temporal data, possibly with non-parametric methods,
or incorporation of autocorrelation into the error structure of the
generalized linear model).  This could allow determination of
appropriate pooling time periods.  Another Panel member suggested that
more appropriate spatial pooling might result from incorporating the
distances among sample locations as a factor in the statistical analysis
models (e.g., a model that incorporates geospatial components, spatial
autocorrelations, kriging, and possibly discrete zonation).  In this
way, appropriate spatial pooling boundaries could be determined and/or
changes in spatial pooling boundaries could be determined. 

	The Panel concluded that there are better, more powerful statistical
methods that are available for examining the alternative hypotheses
related to pooling.  The sampling plan described is hierarchical with
states and months specified, then counties within states that were
specifically selected for their large cotton acreage.  Within sampled
counties a number of sample trap locations were identified and these
traps were monitored and sample moths collected on a weekly basis.  The
null hypothesis described above can be the basis of a model that assumes
counts of gossypol moths in each month by state by county combination
has its own (Binomial) distribution, indexed by the average gossypol
percent (the Binomial success fraction), and that this distribution is
the same for all locations within the county and all sample days within
sample locations in that county.  This can be further formulated in a
fixed effects generalized linear model (McCullagh and Nelder 1989,
Dobson 2001).  The alternative hypotheses also can be formulated as a
Binomial model, but in this case it is assumed that counts at sample
locations within counties have gossypol fractions that vary about the
expected county mean gossypol fraction, or that sampling dates also have
gossypol fractions that vary from the average for that sampling
location.  These hypotheses can be formulated as extensions of the
generalized linear model in which the two additional variance components
are viewed as potential covariates explanatory to the observed
(Binomial) fractions.  The model that addresses the alternative
hypotheses is a mixed effects generalized linear model (Lee et al.
2006).  Thus the analysis approach suggested by multiple Panel members
was to create two or more alternative generalized linear mixed effects
models about plausible alternative hypotheses and use formal statistical
tests to determine whether any of these alternative models is
significantly better at fitting the data in hand than the null
hypothesis model.  It is important to the success of this analysis that
the choices made regarding the various model features (e.g., using a
logit link with a Binomial distribution assuming extra dispersion to
account for the data being of counts of successes (gossypol positive)
assuming Poisson arrivals at the trap) match features of the sampling as
well as the biology of the moths.  Alternate model formulations (e.g.,
via GEE methods, Dobson 2001) were also discussed by the Panel but
specific details were not given.

	While the models suggested above can be complex, the approach is one
that is at the foundation of most statistical tests.  In addition, there
are some benefits to this model fitting/testing approach that cannot be
obtained from the separate test approach, the primary ones being 1) the
ability to estimate the variance components for the deviations from mean
percent; 2) the ability to determine if these variance components can be
related to other covariates (e.g., comparison across years, the topic of
Charge 5); and finally 3) these models have the ability to incorporate
and test for the presence of correlation in responses one might expect
from repeated measurements in time.  None of these three are possible
with the multiple-testing approach.

	The analysis of gossypol fraction differences among counties and months
using the linear logistic model (page 122, Head and Gustafson 2005)
could also be formulated more appropriately as a generalized linear
mixed effects model.  To a certain extent, the county and month effects
should be viewed as random effects in the model which would lead to
slightly different statistical tests. 

	There were some additional comments regarding the use of mixed effects
generalized linear models:

The protocol for choosing sample moths for the gossypol analysis
involved the sequential addition of sample moths in batches of 10 in
situations where the fraction of gossypol positive moths was small. 
There was concern that this process had the potential of biasing the
results of the individual Chi Square test.  As the overall sample size
increases for a given difference in gossypol fraction between the two or
more groups (sample days within a location or sample locations within a
county), the Chi Square test is more likely to reject the result.  In
addition, once the decision is made to pool, those sample days having
more overall samples will have larger weight in the final pooled sample,
essentially resulting in lower average gossypol fraction estimates. 
These same data used in the generalized linear mixed effects model will
not have that same effect because of the way the model performs
estimation and statistical tests.

Without the sample day count data it was not possible to actually fit
the proposed models and hence it was not possible to determine if this
model comparison approach would work for these data.  It is possible
that the model comparison approach might not be successful for some
alternative models, primarily due to lack of balance in the achieved
sampling plan.  In particular, the methods used to estimate the variance
components of the mixed effects general linear model might not.  In this
case, the multiple-testing approach used by Monsanto would be the basis
for the analysis. 

A number of Panel members mentioned concerns with potential lack of
power for any statistical analysis, primarily due to the low numbers of
moths measured for many sample days and sample locations.  Power
addresses the ability of the sampling design to identify significant
differences, in this case significant time-to-time or
location-to-location variability in gossypol fraction when it actually
exists.  Low sampling counts work against power in this case.  Much more
sampling would have been needed to have high power for this study.  The
generalized linear model analysis will have higher power for the
pooling-related tests than can be achieved with the multiple-test
approach.

The generalized linear model approach, being based on a Binomial
distribution for gossypol counts, allows appropriate computing of the
uncertainty related to the gossypol fractions.  If the model related to
the null hypothesis is accepted as the best description of the data
(i.e., the decision is to pool across sampling dates and sampling
locations within county and month) the resulting pooled estimates will
have appropriately computed standard errors.  This approach will result
in slightly different confidence bounds on the gossypol fraction.

All observed data can be included in the generalized linear model
approach, including zero counts of gossypol individuals on a sample day
and counts of one.  These data will not have a large influence on the
model results but any information available collectively from these
sites can be extracted through the model. 

 

Annual variability cannot be measured with only two years.  While the
difference between the two years can be tested directly using the
model-based approach, the relative importance of variability from
year-to-year cannot be compared to county-to-county or site-to-site
variability without additional years of sample data.

There was little discussion of how to handle issues of missing data in
the statistical analysis.  Since the original trap data were not
available it was not clear how zero counts (trap actually checked and no
moths observed) were handled differently from missing data (trap not
checked).  There are more complex statistical models available that
incorporate missing data information into the generalized linear model
(see Helsel 2005 for an introduction to this topic), but these were not
discussed in detail.

	The more formal statistical analysis suggested by the Panel represents
a significant amount of additional work for Monsanto, but it was the
consensus of the Panel that this investment is warranted.  It was noted
by the Panel that the gossypol fraction estimates produced from this
analysis form the basis for subsequent refuge size estimates; namely,
the PNBTC parameter used to estimate ENC in equation 8 (Gustafson and
Head 2005), that is subsequently used to estimate Rnat in equations 9
and 10 (Gustafson and Head 2005).  The Rnat parameter forms the basis
for the request to eliminate structured refuges for Bollgard II cotton. 
The formal statistical analysis may suggest less pooling and/or a
different spatial and/or temporal pooling plan resulting in estimates of
PNBTC that are not necessarily county and month estimates.  If the
decision is not to pool, the PNBTC estimates across sampling dates and
sample locations could be formulated as realizations of a
spatial/temporal random process with distribution estimated via the
generalized mixed effects linear model.  These estimates would form the
basis of a probabilistic (stochastic) approach to the modeling of
refuges and substantially change how the overall refuge analysis results
are considered.

	Finally, the Panel suggested that if these types of sampling studies
are to become more common and used as the basis of future decisions for
resource management of genetically modified crops, EPA might wish to
develop and publish recommendations for appropriate sampling protocols.

Agency Charge

5.  Monsanto did not conduct any statistical analyses comparing the two
sampling years (2004 and 2005).  The Panel is asked to comment on
whether valid comparisons (on a qualitative basis) can be made between
the two years without statistical analyses?  Please describe any
meta-statistical analysis that could improve the overall understanding
of the effectiveness of natural refuge across locations and across time.

Panel Response

Summary

	The Panel concluded that valid comparisons between the two years is not
possible via a meta-analysis.  However, it is possible using a more
formal statistical testing model based on a generalized linear mixed
effects model that takes into account ecologically relevant spatial
scales and incorporates temporal autocorrelation.  A preliminary
analysis using the data available to the Panel concluded the following:

Analysis of variation in Rnat including spatial variation at an
ecologically relevant, regional scale shows that Region (“East” and
“MidSouth”), Month (June, July, and August), and County(Year X
Region) made statistically significant contributions.  Rnat for the East
region was higher than that for the MidSouth region and declined for
both regions with oviposition month.  Intensity of agricultural activity
within the MidSouth region did not affect Rnat.

Analysis of variation in Rnat including spatial variation at an
ecologically relevant, regional scale shows that Year (2004 and 2005)
and Year X Region did not make statistically significant contributions.

Discussion

	As described in the Panel response to Charge 4, Monsanto’s count data
for TBW host plant use are best analyzed using a generalized linear
mixed effects statistical model (e.g., as implemented in SAS PROC
GLIMMIX, SAS Institute, Inc. 2006).  Which function of the count data
should serve as the dependent variable depends on the question to be
addressed.  Whether to pool the data across potential sources of
variation should be based on an analysis of PC, the
binomially-distributed proportion of cotton host use.  In contrast, the
question of which sources of variation significantly affect Rnat should
be based on an analysis of Rnat rather than PC.  Although Rnat is a
function of PC, calculation of estimates of Rnat and its standard errors
from least square mean estimates and standard errors of PC is not
possible because the latter are not available under the assumption that
County is a Random effect (see below).  Ideally, the pooling issue
should be resolved in a statistically rigorous fashion before an
analysis of Rnat is performed, but Monsanto supplied only data pooled
across dates within month for each trap and then across traps within
county.  Fortunately, these pooled data appear to be adequate to address
EPA’s charges to the Panel (see Panel response to Charge 6).

	Instead of retaining the spatially arbitrary division of the county
data among the seven states represented in the data set, the county data
were divided into three, ecologically more relevant regions: all
counties in North Carolina and Georgia [“East”, highly suitable
non-cotton crop hosts available (tobacco and peanuts) and variable
degrees of agricultural intensity]; selected counties in Texas,
Tennessee, Arkansas, Mississippi, and Louisiana [“MidSouth Hills”,
less suitable non-cotton crop hosts available (soybean) and a lower
degree of agricultural intensity]; and selected counties in Texas,
Tennessee, Arkansas, Mississippi, and Louisiana [“MidSouth Flats”,
less suitable non-cotton crop hosts available (soybean) and a higher
degree of agricultural intensity].  “Higher degree of agricultural
intensity” is defined as at least 15% of landscape cropped.  The
counties included in MidSouth Hills are as follows: 2004 Arkansas
(Little River), Louisiana (Bossier), Mississippi (Carroll, Chickasaw,
Clay, Grenada, Lee, Lowndes, Madison, Monroe, Noxubee)]; and 2005
[Arkansas (none), Louisiana (Bossier, Rapides), Mississippi (Carroll,
Chickasaw, Clay, Itawamba, Lee, Lowndes, Monroe, Noxubee, Prentiss),
Tennessee (Haywood, Carroll, Fayette, Gibson), Texas (Austin, Burleson,
Fort Bend)].  All other counties in these states were included in
MidSouth Flats.  Note that while there are good ecological reasons for
the geographic division described above, other geographic divisions
could be conjectured based on slightly different ecological assumptions,
and this conjectured division could also be analyzed in the manner
described below.

	PC was calculated for all available Month X County combinations,
including those for which the host use of only a single adult was
determined.  Rnat was calculated under the following restrictions: (1)
cotton was present in the county (three exclusions—all for Little
River County in Arkansas), (2) the fractions of the landscape area
planted to Bt (ABtC) and non-Bt (ANBtC) cotton were reported (five
exclusions—all in Tennessee), and (3) PC ( ANBtC/ABtC (no exclusions).

	SAS PROC GLIMMIX (SAS version 9.1 for Windows, © 2002-2003, SAS
Institute Inc, Cary, NC, USA) was used to analyze Rnat.  The variable
Rnat is binomially distributed (range 0-1), so the conditional
probability distribution of the data was set to binomial.  County within
Region X Year was treated as a Random effect in the analysis model.  A
first order, autoregressive covariance structure was specified to deal
with the lack of independence of observations within County X Month. 
Linear and quadratic Month trends were tested using contrasts.  The
results are given in Fig. 5-1.  Because interaction terms involving
Region were not statistically significant and the MidSouth Hills and
MidSouth Flats regions were not significantly different, the least
square mean estimates for a MidSouth region combining the two are also
shown in Fig. 5-1.

 

Fig. 5-1.  Proportion of natural refuge for different regions estimated
from pheromone-trapped TBW by month.  Estimates for MidSouth (combined)
uses estimates for MidSouth Hills and MidSouth Flats.

	For the period of time during which TBW is subject to selection for
counteradaptation to Bt cotton, the average level of Rnat is much higher
for the East region than for the MidSouth region.  This difference may
result from the presence of peanut and/or tobacco production in the East
and the absence of correspondingly suitable non-cotton crop hosts in the
MidSouth.  During July and August in the MidSouth region, Rnat averages
ca. 0.3, which is considerably greater than 0.05, the non-Bt cotton
structured refuge requirement.  However, as discussed in the Panel
response to Charge 7, counteradaptation to Bt cotton may develop locally
and spread from so-called “hotspots”.  Rnat estimates for the
counties in the MidSouth region that were singled out by Monsanto as
“worst-case” examples (i.e., lowest Rnat levels) are frequently
below 0.05 (see Panel response to Charge 6).  Consequently, the observed
regional mean estimates of Rnat are inadequate to demonstrate
Monsanto’s contention that natural refuges alone can prevent
counteradaptation of the TBW to Bt cotton in the MidSouth region.

	None of the models using Year (2004 and 2005) or Year X Region
variation in Rnat were statistically significant when the models
included spatial variation at an ecologically relevant regional scale. 
This suggests that, at least for this limited data set, annual variation
was small compared to other sources of variation.  This analysis can
only be considered preliminary due to the small number of years
considered, and these results should not detract from the broader Panel
recommendation that more years of data should be required for evaluating
questions of this gravity.

Effective Refuge Calculation and Modeling

Agency Charge

6.  Monsanto has corrected their calculation of effective refuge size
presented in Gustafson and Head, 2004 based on the Agency’s (BPPD,
2004) and June 2004’s Federal Insecticide, Fungicide, and Rodenticide
Act (FIFRA) Scientific Advisory Panel’s (SAP) recommendations (SAP,
2004).   Modifications to the calculation of the effective refuge size
involved removing the assumption of constant effective refuge size and
explicitly accounting for the lower production of CBW and TBW in cotton
where survival of these insects is reduced.  Estimation of the effective
refuge now assumes a regionally specific annual cycle of effective
refuge size, according to data collected in alternative host studies of
CBW   ADDIN EN.CITE
<EndNote><Cite><Author>Head</Author><Year>2004</Year><RecNum>70</RecNum>
<record><rec-number>70</rec-number><ref-type
name="Report">27</ref-type><contributors><authors><author>Graham
Head</author><author>Richard
Voth</author></authors></contributors><titles><title><style
face="normal" font="default" size="100%">A final report on studies to
assess production of </style><style face="italic" font="default"
size="100%">Helicoverpa zea</style><style face="normal" font="default"
size="100%"> from alternate host plants and from the external unsprayed
non-Bt cotton refuge for Bollgard</style><style face="superscript"
font="default" size="100%">®</style><style face="normal" font="default"
size="100%">
cotton</style></title></titles><dates><year>2004</year></dates><pub-loca
tion>St. Louis, Missouri</pub-location><publisher>Monsanto
Company</publisher><isbn>MSL-19238, MRID
46222401</isbn><urls></urls></record></Cite></EndNote> (Head and Voth,
2004)  and TBW   ADDIN EN.CITE
<EndNote><Cite><Author>Head</Author><Year>2005</Year><RecNum>130</RecNum
><record><rec-number>130</rec-number><ref-type
name="Report">27</ref-type><contributors><authors><author>Graham
Head</author><author>David
Gustafson</author></authors></contributors><titles><title><style
face="normal" font="default" size="100%">Production of </style><style
face="italic" font="default" size="100%">Heliothis virescens
</style><style face="normal" font="default" size="100%">from alternative
host plants and the role of these host plants as natural refuge for
Bollgard II</style><style face="superscript" font="default"
size="100%">®</style><style face="normal" font="default" size="100%">
cotton</style></title></titles><dates><year>2005</year></dates><pub-loca
tion>St. Louis, Missouri</pub-location><publisher>Monsanto
Company</publisher><isbn>MSL-20123</isbn><urls></urls></record></Cite></
EndNote> (Head and Gustafson, 2005) .  These data were combined with 
corn planting estimates on either the regional scale for CBW, or
county-scale for TBW, to estimate effective (i.e., current (structured
non-Bt cotton + non-cotton) and natural (non-cotton only) refuge sizes
for each of what were conservatively assumed to be six annual
generations for each pest.  

Estimation of the relative number of CBW adult moths produced by each of
the five sub-compartments is given by the following equation:  Mij = Aij
Eij LBij LSij    (Equation 1).  

[M is the number of adult moths produced per unit area of the region; A
is the proportion of the region occupied by the crop type of interest; E
is the relative (to cotton, i.e., Ecotton=1) number of effective eggs
(eggs that would produce adults in the absence of B.t. or pyrethroid
sprays) laid in the crop type; LB is the fraction of larvae surviving in
the presence of the B.t. crop; LS is the fraction of larvae surviving a
pyrethroid insecticide spray on the crop; the subscript i refers to the
compartment (B for B.t. or R for refuge); and the subscript j refers to
the particular crop type within the compartment (1 = cotton, 2 = corn, 3
= other C3 host crop).]

The effective refuge, Reff, is defined as the proportion of adult moths
that would have been produced in the refuge compartment (non-Bt cotton,
non-Bt corn, non-cotton C3 crops) in the absence of any induced larval
mortality:  

    (Equation 5; used when CBW populations were actively feedings in
cotton, Generations 3-5)

Effective refuge estimations for all of the “non-cotton” generations
are given by:

   (Equation 6)

The natural refuge component (i.e., non-cotton C3 crops + non-Bt corn
components) of the total effective refuge is as follows:  

  (Equation 7)

      

The Agency asks the SAP to comment on the estimated CBW effective and
natural refuge calculations.

Pooled, county-level estimates of the percent cotton-reared TBW moths
were combined with county-level landcover information to estimate the
current effective refuge and natural refuge for each county per month. 
The relative TBW productivity of non-cotton areas within a county for a
specific month is given as:

   (Equation 8)

The current effective refuge (non-Bt cotton + non-cotton hosts) for TBW
is defined as the proportion of TBW moths actually produced in the
effective refuge compartment prior to selection by Bt cotton:

   (Equation 9)

The estimated natural refuge (non-cotton hosts) for TBW is given by the
following equation:

   (Equation 10)

The Agency asks the SAP to comment on the estimated TBW effective and
natural refuge calculations.

Panel Response

Summary

1) Equation 7 overestimates the amount of natural refuge for CBW, Rnat,
with the overestimates largest for Georgia (37%) and East Texas (44%).

2) Equations 8 and 9 overestimate the amount of effective refuge, Reff,
and natural refuge, Rnat, for TBW by not accounting for possible
insecticide application in structured refuges.  Generally, the size of
the overestimation is roughly equal to LSNBtC, the ratio of survival of
TBW in non-Bt cotton (including from insecticide spraying) to their
survival on non-cotton hosts.  Thus, if LSNBtC = 20%, the true estimates
of Reff and Rnat are roughly 20% of the values given by Monsanto.  The
overestimation can be considerably less severe when ANBtC/ABtC << PNBtC
(e.g., Lenoir Co., North Carolina, and Mississippi Co., Arkansas).

3) Underestimates of PNBtC will lead to overestimates of Rnat.  This
could be an important source of overestimation if the gossypol assay
gives false negatives (see Panel response to Charge 1) and hence
underestimates of PNBtC.

4) The estimates of Reff and Rnat are imprecise, due to uncertainty in
the estimates of the parameters in the equations.  Imprecision is
possibly large for TBW in those counties used as scenarios for modeling
(see Panel response to Charge 7), because the estimates of Reff and Rnat
for these counties are low (see Panel response to Charge 3).  Monsanto
does not provide enough information to assess the level of uncertainty,
and it is unclear whether the sampling intensity from the present study
is sufficient to estimate this uncertainty.

Discussion

	Specific issues regarding the calculations for (a) CBW and (b) TBW are
presented below followed by (c) a discussion of Panel concerns regarding
parameter uncertainty.

a) CBW

	The Panel had several specific concerns about the equations used to
calculate Reff and Rnat for CBW.

Equation 1 

This equation assumes that movement rate of males from all habitat types
is the same.  If males are more likely to disperse from a given type of
habitat, this will have the same consequences for the estimates of Reff
and Rnat as increasing the number of eggs laid in these habitats, E. 
The consequences of bias in estimates of E are described below.

Equation 2

Monsanto provides evidence that two pyrethroid applications are common
for cotton regardless of whether or not it is Bt cotton.  However,
conventional and Bt cotton differ in the proportion of acres that
receive pyrethroid applications.  Based on this information, Monsanto
assumes the survival rate of CBW on cotton is LSi1 = [Ti1 ((1/3) +
(2(1-Ki1)/3))] + (1 - Ti1).  Equation 2 can be interpreted as a weighted
average of the survival rate over three generations of CBW when two out
of the three generations receive pyrethroid applications. 
Alternatively, Monsanto could have employed the assumptions LSi11 = 1
for the first generation of CBW on cotton, and LSi123 = Ti1(1-Ki1) + (1
- Ti1) for the second and third generations.  These alternative
assumptions are more consistent with what is being described in the
field.  Note that LSi11 > LSi1 > LSi123, which means Monsanto’s
methodology will underestimate production of CBW moths for the first
cotton generation and overestimate production for the second and third
generations.  Therefore, Monsanto’s methodology will underestimate the
temporal variability of CBW moth production and effective refuge
calculations.  Simulations conducted by one Panel member in preparation
for this SAP meeting, using the model reported in Hurley et al. (2006),
suggest this type of increased temporal variation can speed the
evolution of resistance (see also Ives and Andow 2002).  These results
are not detailed in this report due to the modest differences that were
observed.  These modest differences suggest Monsanto’s simplifying
assumption may be a reasonable, although not necessarily conservative,
approximation.  Still, Monsanto could strengthen the credibility of
their results by incorporating the type of temporal variability in
pyrethroid applications observed in the field into their model.

Equation 5

Biases in the estimates of Aij , Eij , LBij , and LSij have different
consequences depending on the habitat considered.  Specifically,
underestimates of all of the parameters for refuges (in cotton, corn,
and other C3 host crops) will underestimate Reff and Rnat.  Conversely,
underestimates of all of the parameters for Bt cotton and Bt corn will
overestimate Reff and Rnat.

Equation 7

	Monsanto (Gustafson and Head 2005) defines the natural refuge as

 , 				(Equation 6-1)

which removes non-Bt cotton moth production from both the numerator and
denominator of the effective refuge calculation.  This equation instead
should omit MR1 only from the numerator to give

 				(Equation 6-2)

Monsanto’s equation for Rnat leads to overestimation that can be
substantial.  Based on the data reported in Tables 2 and 5 (Gustafson
and Head 2005), Monsanto’s estimate of the natural refuge (RnatCBW)
will be 37, 12, 6, and 44% higher than the natural refuge estimate given
by Equation 6-2 for Georgia, Mississippi, North Carolina, and Texas. 
Appendix 1 gives a correct derivation of Equation 7.

Additional concerns

	These equations are applied at the scale of entire states.  This
assumes either that the population of CBW is completely mixed at the
regional scale (i.e., a given trapped insect has the same probability of
coming from any location in the region), or that the distribution of
habitats is uniform across the region at the scale of insect movement
(i.e., for a trapped insect, the distribution of habitat types in the
specific area from which it was produced is the same as the distribution
of habitat types throughout the entire region).  Data are presented to
show that CBW often travel long distances.  Nonetheless, it is very
difficult to determine from these data whether the CBW population is
completely mixed at the scale assumed by equations 1-7; this is
discussed further under Charge 1.  We note, however, that Bollgard II is
not strongly high-dose for CBW, and this may affect the impact of
incomplete mixing on the rate of resistance evolution.

	A related concern is brought up by EPA's review (p. 14, BPPD 2006).  No
information is given about whether traps were located next to Bt or
non-Bt fields.  The basic assumption behind equations 1-7 is that the
CBW population is completely mixed with respect to habitat types. 
Therefore, under this assumption there should be no effect of trap
placement on the habitat of origination of the moths.  This concern of
the EPA reviewers suggests that they do not believe that CBW males are
completely mixed at the regional scale.

b) TBW

Equations 9-10

	Calculations for TBW are similar to CBW, but assume that the survival
of TBW from Bt crops is zero.  Also, the effect of differential survival
in non-cotton hosts and non-Bt cotton including that due to insecticide
spraying (which was included for CBW after a request by the SAP 2004)
was not included in the case of TBW.  Including differential survival
and expressing ENC in terms of measured parameters (e.g., PNBTC) gives 

 					(Equation 6-3)

and

 					(Equation 6-4)

where LSNBTC is the survival of TBW from pesticide application and other
mortality factors in non-Bt cotton relative to survival in non-cotton
hosts.

  for Monsanto’s seven “worst-case” counties for the oviposition
months June, July, and August, shows that Rnat in the MidSouth region is
not infrequently below 0.05 for LSNBtC = 1, and is generally below 0.05
for LSNBtC = 0.2 and 0.1 (Figs. 6-1, 6-2, and 6-3, respectively; Gibson 

 

Fig. 6-1.  Proportion of natural refuge, Rnat, calculated by month from
pheromone trap-captured TBW for worst-case counties in six states, under
the condition LSNBtC = 1.  

 

Fig. 6-2.  Proportion of natural refuge, Rnat, calculated by month from
pheromone trap-captured TBW for worst-case counties in six states, under
the condition LSNBtC = 0.2.  

 

Fig. 6-3.  Proportion of natural refuge, Rnat, calculated by month from
pheromone trap-captured TBW for worst-case counties in six states, under
the condition LSNBtC = 0.1.  

County, Tennessee, excluded due to lack of data for June and August). 
Even for counties in the East, Rnat can be below 0.05 for LSNBtC = 0.1
(Fig. 6-3).  Also, it is apparent that Rnat was unusually low in Austin
County, Texas, relative to other MidSouth counties in June 2005. 
Whether this is a difference peculiar to East Texas or represents year
to year variability that may also occur in other areas is an open
question.

	These calculations, of course, depend on the differential level of
mortality caused by spraying, LSNBtC.  If, for example, most of the
structured refuge is in the form of 5% external unsprayed refuge, then
LSNBtC will be close to 1, and the incorrect formulas used by Monsanto
will not grossly overestimate Reff and Rnat.  As a result, detailed
information is needed about the deployment of refuges and the intensity
of insecticide use to properly calculate Reff and Rnat.

	Some discussion arose amongst Panelists about how to incorporate
insecticide spraying in refuges into equations 9-10, and the Panelists
agreed on the following argument.  The correction for spraying in non-Bt
cotton affects the calculations for Reff and Rnat mainly through
changing the productivity of TBW from non-Bt cotton relative to the
potential productivity of TBW from Bt cotton that would occur in the
absence of mortality from Bt.  The importance of the contrast between
productivity in non-Bt cotton vs. the potential productivity in Bt
cotton is due to the manner in which the refuge reduces the rate of
resistance evolution.  While susceptible TBW can only reproduce in the
refuge, resistant TBW can also reproduce in Bt cotton fields. 
Therefore, the definition of refuge (as correctly used by Monsanto) is
the ratio of productivity of susceptible TBW from refuges to the
potential productivity of resistant TBW from all habitats.  To make this
explicit, consider the equation for Reff in terms of M, defined (as in
Monsanto’s equation 1 for CBW) as the production of TBW from different
habitats:

 			(Equation 6-5)

Here, values of EBTC and ENBTC are the densities of eggs oviposited in
the different habitats, and mortality from Bt, LB, has been excluded
because resistant TBW will not suffer from this mortality.  Note that
the term involving ABTC in the denominator, EBTC ABTC, does not include
LSNBTC.  Therefore, including spraying mortality in non-Bt cotton
refugia reduces the estimate of Reff by accounting for the reduction in
production of susceptible TBW from non-Bt cotton relative to the
potential production of resistant TBW from Bt cotton.  Correcting the
equations for spraying in non-Bt cotton also changes the estimates of
ENC, because spraying will decrease the production of TBW from non-Bt
cotton relative to non-cotton refuge.  Nonetheless, it is the contrast
between reproduction of resistant TBW from Bt crops and susceptible TBW
from non-Bt crops that is responsible for the overestimates of Reff and
Rnat in Monsanto’s equations.

Sensitivity of TBW calculations to parameter biases

	For TBW the effects of biases in parameter estimates on Reff and Rnat
can be summarized using Equations 6-3 and 6-4 above:

ANBTC – If the estimates of land-area cover of non-Bt cotton are high,
then the estimates of Reff and Rnat will be high as well.

ABTC – If the estimates of land-area cover of Bt cotton are high, then
the estimates of Reff and Rnat will be low.

LSNBTC – If the estimates of the survival from spraying insecticide in
non-Bt cotton are high (or, as assumed by Monsanto, 1), then the
estimates of Reff and Rnat will be high.

PNBTC – If the estimates of the proportion of (non-Bt) cotton-reared
insects as opposed to those originating from non-cotton wild hosts are
high, then the estimates of Reff and Rnat will be low.  This effect on
Rnat deserves special attention.  As discussed in Charge 1, the Panel is
concerned that the gossypol assays are giving false negatives, and
therefore underestimating PNBTC.  Underestimating PNBTC will
overestimate Rnat.  Therefore, until the gossypol assay is validated,
the estimate of Rnat should be considered an overestimate, although it
is unclear how much of an overestimate it is.  Note that this source of
overestimation of Rnat is separate from the source of overestimation
caused by ignoring insecticide application in non-Bt cotton described
above.

	Like the calculations for CBW, the calculations for TBW assume that
males are well-mixed at the spatial scale of variation in habitat types.
 In fact, it is unlikely that TBW males from non-Bt cotton and
non-cotton hosts are well-mixed during the cotton growing season, at
least in Louisiana.  The seasonal pattern of variation in the frequency
of resistance to pyrethroid insecticides as this resistance became
established in Louisiana (low at the beginning of the year and
increasing as the growing season progressed only to return the next
spring to a level nearly as low as the level of the previous spring)
(Leonard et al.1995, Bagwell et al. 2000) is strong circumstantial
evidence that the population of TBW selected for resistance to the
insecticides in cotton was isolated during the cotton production season
from a very large TBW population not under such selection, and that the
populations were panmictic (unstructured or random-mating populations)
in the fall and/or spring.  This interpretation is supported by
increased genetic structuring (decreased gene flow) of local TBW
populations during the cotton-growing season in western Mississippi
measured with random amplified polymorphic DNA markers (Han and Caprio
2004).  In counties with extensive cotton production, this will tend to
bias the estimates of PNBtC upwards, resulting in underestimation of
Rnat.

(c) Parameter uncertainty

	The calculations make numerous assumptions about the parameters A, E,
LS, LB, and P that are used.  They also assume that these parameters are
known without error.  To discuss the issue of parameter uncertainty, we
will focus here on the calculations for TBW (corrected as described
above), noting that similar issues arise for the calculations for CBW.

Uncertainty in the estimates is difficult to assess but could be large. 
Monsanto does not provide information about the uncertainty of some
parameters.  Although Monsanto provides confidence intervals for the
estimates of P (see note below), these confidence intervals do not
directly translate into confidence intervals for Reff or Rnat.  It is
possible, however, to estimate or approximate confidence intervals in
Reff and Rnat using the statistical approaches described in Charge 4 if
uncertainty of all parameter estimates is known.

 .  If this occurred for the counties used in Monsanto’s modeling
scenarios, Monsanto would overestimate the size of the refuge by 100%. 
This example demonstrates that Monsanto needs to calculate confidence
intervals for their estimates of Reff and Rnat to account for the range
of expected refuge sizes in risk calculations.

	Note on confidence intervals for PNBtC:  The formula for the upper
limit of the 95% CI for Rnat (p. 121, Gustafson and Head 2005) is
incorrect as presented, but the correct formula apparently was used in
the calculations.  The lower confidence limit given (p. 121, Gustafson
and Head 2005) is an approximation that fails when all the TBW test
negative for gossypol.  For a total of n individuals tested, the correct
formula is: LL = 0.051/n (Louis 1981).  This correction results in a
small increase (≤ 7.1%) in LL.

  assuming AVE[PNBtC] = 0.5, ANBTC = 0.05, and ABTC = 0.95; PNBTC is
calculated assuming a binomial sample of 10 males.

Agency Charge

7.  Monsanto examined the durability of each of the three Bt cotton
products (i.e., Bollgard, WideStrike, and Bollgard II) individually and
together in the marketplace using a three-gene model. The Bt protein,
Cry1Ac, is common to all three products.  The presence of each of these
products in the marketplace selects for potential resistance to Bollgard
cotton, expressing only the Cry1Ac protein, and also selects for
resistance to the other two products through the common selection for
Cry1Ac resistance.  The products vary greatly in the rate at which they
select for resistance to Cry1Ac because of the presence of additional
insecticidal proteins in Bollgard II (Cry2Ab2) and in WideStrike
(Cry1F).  

The three-gene model for insect resistance evolution used in this study
is based on a conceptual model similar to that proposed by Dow
AgroSciences (DAS) for its product, WideStrike cotton, and was reviewed
by a recent U.S. EPA Scientific Advisory Panel (SAP)   ADDIN EN.CITE
<EndNote><Cite><Author>U.S.
EPA</Author><Year>2004</Year><RecNum>156</RecNum><record><rec-number>156
</rec-number><ref-type name="Electronic
Source">12</ref-type><contributors><authors><author>U.S.
EPA,</author></authors><secondary-authors><author>Scientific Advisory
Panel</author></secondary-authors></contributors><titles><title><style
face="normal" font="default" size="100%">Transmittal of minutes of the
FIFRA Scientific Advisory Panel meeting held June 8-10, 2004: product
characterization, human health risk, ecological risk, and insect
resistance management for </style><style face="italic" font="default"
size="100%">Bacillus thuringiensis</style><style face="normal"
font="default" size="100%"> (Bt) cotton
products</style></title><secondary-title>SAP report no.
2004-05</secondary-title></titles><volume>2005</volume><number>December
1,
2005</number><dates><year>2004</year></dates><pub-location>Washington,
D. C.</pub-location><publisher>United States Environmental Protection
Agency</publisher><urls><related-urls><url>http://www.epa.gov/scipoly/sa
p/2004/june/final1a.pdf</url></related-urls></urls></record></Cite></End
Note> (SAP, 2004) .  However, the SAP questioned some of the
mathematical details of the DAS model and Monsanto has made some changes
to address the SAP’s concerns.  As shown in schematic form in Figure
3, Appendix 2, the three-gene model is based on the following
assumptions concerning the mechanism of activity of the three commercial
Bt cotton products (Bollgard, Bollgard II, and WideStrike cotton):

The Cry1Ac toxin, present in all three products, binds to two receptors,
60% to receptor A and 40% to receptor B. 

The Cry1F toxin, present only in WideStrike cotton, binds exclusively to
receptor A.

The Cry2Ab2 toxin, present only in Bollgard II cotton, binds exclusively
to receptor C.

a)    CBW.  Based both on the intrinsic durability of each of the three
B.t. cotton products (Figure 4, Appendix 2) and the three-gene modeling
analyses for all three Bt cotton products together in the marketplace
(Table 14, Appendix 2), Bollgard II retained the highest level of
efficacy against CBW in all scenarios (all regions).  Given the
assumptions of the three-gene model and its limitations, there is likely
enough effective natural refuge to be sufficient to delay the evolution
of resistance to Bollgard II cotton for more than 25 years (not a
precise number of years) under all plausible scenarios in all four
regions (Table 14, Appendix 2).  This is because of the relatively high
mortality of individuals heterozygous to Cry1Ac resistance in the
presence of Cry2Ab2, as compared to WideStrike.  WideStrike is
intermediate in many scenarios because of the shared binding receptor
between Cry1F and Cry1Ac and the likelihood of cross-resistance is
greater.  Bollgard is weakest in all scenarios because there is no high
dose for CBW and it is a single-gene product.  Monsanto’s models
predict that CBW resistance to Bollgard cotton will evolve in less than
the 30 year horizon in the Georgia, Mississippi, and E. Texas regions in
most scenarios except for 2-C (Bollgard = 0.1; Bollgard II = 0.8;
WideStrike = 0.1).  Resistance always took at least 30 years to evolve
to all three Bt cotton products in the North Carolina region in all
scenarios, even the natural refuge scenarios.  When Bollgard cotton
acreage is minimized, Bollgard II and WideStrike longevity is maximized
(Table 14).  Large amounts of Bollgard II cotton in the marketplace
increased the durability of both Bollgard and WideStrike (Table 14,
Appendix 2).  Uncertainties in the pheromone captures, estimation of
adult productivity, carbon isotope analyses, spatial analysis,
estimation of effective refuge calculation, degree of shared binding
affinity of Cry1Ac to receptor A and B, genetics of resistance,
resistance mechanism, initial resistance allele frequency, and other
modeling assumptions affect the precision and accuracy of the modeling
predictions.  Monsanto’s modeling also does not consider pre-selection
for Cry1Ac resistance.  Ten years of selection pressure (since 1996) for
resistance to Cry1Ac has already occurred.  Field resistance to Cry1Ac
places additional selection pressure on the Cry2Ab2 component of
Bollgard II cotton.

Given the assumptions and uncertainties in Monsanto’s CBW modeling
efforts, the Agency asks the SAP to comment on the utility of the
modeling to predict the effectiveness of natural (non-cotton C3 crops +
non-Bt corn) vs. current effective refuge (non-Bt cotton + non-Bt corn 
+ non-cotton C3 crops) to manage CBW resistance to the  toxins expressed
in Bollgard II.  Discuss the impact of pre-selection for Cry1Ac
resistance on the modeling output.

b)  TBW.  The intrinsic durability of all three Bt cotton products is
much greater for TBW than for CBW because of the “high dose” of
Cry1Ac for TBW expressed in all three products.  In virtually all cases,
all three products retained their efficacy (i.e., no resistance) for
more than 30 years (maximum time for the simulation) even if all cotton
in a region is planted to that product and no structured refuge is
required (i.e., all natural refuge).  The only exceptions occur for
Bollgard cotton in Tennessee and Mississippi.  Given the assumptions of
the three-gene model and its limitations, there is likely enough
effective natural refuge to be sufficient to delay the evolution of
resistance to Bollgard II cotton for more than 30 years (i.e., the time
horizon of the model, not to be interpreted as a precise number of
years) under all plausible scenarios in all four regions.  This is due
to the extremely high efficacy of Cry1Ac against TBW, and the fact that
Cry1Ac is present in all three Bt cotton products.  In the state with
the lowest natural refuge for TBW, Mississippi (see Table 13, Appendix
2), resistance to Cry1Ac and Cry1F evolved after 21 years in scenario
1-N if the structured refuge requirements for Bollgard and WideStrike
cotton were removed.  Uncertainties in the pheromone captures, gossypol
analyses, spatial analysis, estimation of effective refuge calculation,
degree of shared binding affinity of Cry1Ac to receptor A and B,
genetics of resistance, resistance mechanism, initial resistance allele
frequency, and other modeling assumptions affect the precision and
accuracy of the modeling predictions.  Monsanto’s modeling also does
not consider pre-selection for Cry1Ac resistance.  Ten years of
selection pressure (since 1996) for resistance to Cry1Ac has already
occurred.  Field resistance to Cry1Ac places additional selection
pressure on the Cry2Ab2 component of Bollgard II cotton.

Given the assumptions and uncertainties in Monsanto’s TBW modeling
efforts, the Agency asks the SAP to comment on the utility of the
modeling to predict the effectiveness of natural (non-cotton hosts) vs.
current effective refuge (non-Bt cotton + non-cotton hosts) to manage
TBW resistance to the Bt toxins expressed in Bollgard II cotton. 
Discuss the impact of pre-selection for Cry1Ac resistance on the
modeling output.

Panel Response

Introduction

The Panel cautions EPA that acceptance of a simple non-spatial model
used deterministically in this registration package could be interpreted
as a precedent for future registrations.  The panel advises EPA to
clearly state under what conditions use of such models is appropriate. 
The specific model developed by Monsanto and the method used for its
analysis were intended to address a very specific modification to an
existing registration.  The results clearly identify geographic regions
where resistance evolution poses minimal risk, but the model and
analyses are inadequate for evaluating the risk in any of the other
regions or scenarios.

This is a very simple analytical model intended to address a very
specific question:  Would the removal of structured refuge significantly
increase the risk of developing resistance to the three Cry toxins
currently used in the environment?  As with any model, the quality of
the output and the ability to analyze output for the purpose of
answering a question depends on 1) the choice of parameters that
dictates the structure of the model, 2) the quality of the parameter
estimates used as inputs to the model, and 3) the validity of the
assumptions represented by the structure, inputs and method of
execution.

The Panel challenged all three aspects of the Monsanto model.  Panel
responses to Charges 1 – 6 examined many of these issues in greater
detail.  In this section we focus on evaluating Monsanto’s claim that
natural refuges are sufficient to protect against resistance evolution
in the “worst case” scenarios for each of the regions they consider.

Difficulties Interpreting Model Results

Monsanto’s model, as presented to the Panel, appears to have
identified some geographic regions where there is very little risk of
resistance developing.  By identifying regions of little risk, it also
identifies the regions where the risk of resistance developing is
greater.  However, the model (as executed) cannot adequately assess risk
in these regions, such as in Mississippi and East Texas regions.  Proper
assessment of risk in these regions requires acquisition of more data, a
more robust statistical analysis of the data, and a more detailed
approach to modeling that includes both spatial and temporal
variability.

Monsanto has taken a cafeteria approach to modeling this system.  They
have used a simple analytical model and truncated execution at 30 years.
 Truncating execution is a constraint typically applied to, and more
appropriate for, stochastic simulation models that are prohibitively
computer intensive.  While 30 years may represent a reasonable (and
perhaps long) product life, Monsanto’s decision to truncate execution
at 30 years renders all of the TBW and much of the CBW output inadequate
for assessing the risk of resistance evolution if structured refuges
were removed.

Tables 14 and 15 (Appendix 2 in Gustafson and Head 2005) present the
years to resistance and the efficacy at 30 years for each product by
region.  Years-to-resistance is a relative index or qualitative method
of assessing risk, a point made clear in the EPA presentation.  A
secondary perspective on risk is provided by the product efficacy. 
However, by truncating execution of the model at 30 years, product
efficacy becomes the only perspective.  The TBW results (Table 15) show
no qualitative differences at all and cannot be evaluated.  The CBW
results (Table 14) show qualitative differences for some scenarios but
the 30-year efficacies suggest that resistance would develop soon after
the 30-year time horizon.

The efficacy of a product does not translate well to a resistance allele
frequency, especially when the product contains more than one toxin,
affects more than one locus or when cross resistance is not involved. 
The 30-year allele frequencies for each locus would provide a third
perspective for assessing risk and aid in our overall understanding of
the process of evolution to transgenic crops.

Given the simplicity of the analysis, this model should have been run
until resistance developed regardless of the number of generations or
years.  If Monsanto wants to demonstrate that resistance never develops
under a particular scenario, then allele frequencies for resistance for
each locus need to be provided showing that there has been no change
over some period of time (either from initial allele frequencies or some
stable equilibrium allele frequency over an extended period of time).

In predicting resistance evolution to Bollgard II, there is not only
uncertainty in the estimation of parameters used in models (as described
in previous sections), but also “model uncertainty,” as is always
the case.  Model uncertainty refers to very different predictions that
can arise from different models.  Unfortunately, while there are
statistical methods to account for parameter uncertainty, it is much
more difficult to formally assess model uncertainty before resistance
evolution occurs.  Of course, once resistance has arisen, it is possible
to validate models against the data.  For example, Livingston et al.
(2002) show that a simple genetic model can characterize the evolution
of pyrethroid resistance in CBW and TBW field populations.  Nonetheless,
in the absence of such empirical data for Bt cotton, and in particular
for a two-toxin product like Bollgard II, model uncertainty can only be
addressed by careful mathematical examination of different models to
understand why they give different results.

	The model which Monsanto used for its analyses (Gustafson and Head
2005) is based on Caprio (1998).  Like all models, this model makes
numerous simplifying assumptions about the biology of TBW and the
processes that affect resistance evolution.  Exploration of this model
(Appendix 2) demonstrates that seemingly subtle differences in the
assumptions not only can have a large impact on predictions about the
rate of resistance evolution, but can also change which of the
parameters have the greatest influence on the rate of resistance
evolution.  This highlights the problem that model uncertainty affects
not only the quantitative predictions about the rate of resistance
evolution, but also the key factors that must be measured to make these
predictions.

	A suite of well-understood models of resistance to a pyramid Bt product
does not yet exist.  Furthermore, key issues like spatial structure,
linkage disequilibrium, and differential movement of males and females
have not been explored in detail.  This makes model uncertainty high.

Parameter Estimates

Parameter uncertainty is a common problem in risk analysis that can
often lead to incorrect analysis and inappropriate conclusions. 
However, there are several methodologies for incorporating parameter
uncertainty and exploring the sensitivity of the results to differences
in parameter estimates.  The uncertainties associated with many of the
parameter estimates have been discussed under Charges 1-6.  In general,
these uncertainties can be resolved with the collection of more years of
data and better spatial sampling of the regions.  Even without more
data, a more appropriate statistical analysis should provide a better
understanding of the hosts from which moths developed and the geographic
and temporal distribution of moth abundance in each region.  The most
significant challenge and the greatest uncertainty involves the gossypol
analysis.  All estimates of moth production from non-cotton refuges are
dependent on the validity of this technique (see Panel response to
Charge 3).  Confirmation of the gossypol analysis as a means of
assessing the proportion of moth trap catches originating from cotton
hosts would have the greatest impact on the confidence of risk
assessment regardless of the modeling techniques used.

Both analytical and simulation models have been used in the past to
assess the risk of resistance to plant-incorporated protectant
compounds.  Deterministic and stochastic approaches have been used with
both kinds of models.  Monsanto’s model is a simple analytical model
using perhaps the simplest deterministic approach.  This technique makes
broad assumptions concerning the year-to-year stability of the landscape
characteristics.  Monsanto explicitly defines the proportions of each
product as constant over the 30-year time horizon (see socio-economic
factors below).  Implicitly, this model also assumes year-to-year
stability in the proportions of crop types, the proportions of
continuous and rotated crops, and the spatial distribution of various
hosts throughout the landscape.  Similarly, the biology and behavior of
the insect in response to the landscape is assumed to be stable from
year-to-year.

An alternative to an analytical model is to simulate the system in
detail.  Risk can be assessed using either modeling technique by varying
the parameter estimates across a range of values.  Using a deterministic
approach, inputs can be chosen as fixed values representing the mean (or
median as appropriate) as well as the high and low values representing
some confidence interval (assumed 95% CI).  Monsanto used this basic
approach assuming that the lowest moth trap catches in each region
represents the greatest risk of refuge failure.

Conceptually this is an appropriate methodology but (i) the statistical
analysis of the trap data, (ii) uncertainty about the similarity of male
and female dispersal behavior, (iii) uncertainty about the accuracy of
the gossypol analysis, and (iv) unknown uncertainty in other parameters,
does not permit us to assess the lower bounds of the confidence interval
with any accuracy.  If these issues are resolved, we recommend a
multi-pronged approach involving several modeling strategies, including
simple deterministic models and more-complex simulation models. 
Mississippi and East Texas are regions that require particular attention
in these modeling efforts.

Assumptions

	Three principle assumptions of the model were challenged by the Panel,
specifically, widespread dispersal of the pests, socio-economic factors
affecting the market share of the products, and the single locus per
toxin receptor without cross-resistance.  The expected changes to these
assumptions have the potential for resistance to evolve more rapidly
than Monsanto’s results suggest.  Detailed evaluation of these
assumptions follow.

	The Panel also notes that two of the assumptions may be considered
conservative because the literature suggests that resistance to
transgenics involves multiple loci and may be associated with fitness
costs.  The anticipated effects of incorporating less conservative
estimates are described after the challenges.

Spatial and Temporal Variability of Dispersal

	Figures 1 and 2 in Benedict (2005) illustrate the magnitude of
differences in moth trap catches and host origin throughout the season
and between regions.  Confidence in the assumption that TBW is
dispersing widely across the landscape varies throughout the season.  If
there are periods when dispersal is limited and mating moths are derived
from the local landscape, resistance can evolve in a local population as
a "hotspot."  Single-locus spatially explicit models of resistance
evolution (e.g., Peck et al. 1999, Storer et al. 2003, Sisterson et al.
2004) show that when refuges become rare, resistance is much more likely
to occur in hotspots of Bt fields that are relatively isolated from
refuges.  Generations of low dispersal can play a disproportionate role
in the rate of resistance evolution.  The very low FST values reported
in the literature for population differentiation (based on neutral
genetic markers), indicating high gene flow in general for TBW (Han and
Caprio 2004, Benedict 2005), could exist even in the face of the
oscillating temporal dispersal patterns over the season in which a
resistance allele is increasing in frequency.  While low FST values can
reflect the effects of high levels of gene flow at neutral loci caused
by migration over a long period of time, something very different can be
happening to allele frequencies at loci under strong selection pressure.

	The structure of the Monsanto model is non-spatial and therefore
assumes that the distribution of refuge and Bt fields and the dispersal
behavior of the adult insect are similar throughout the landscape. 
These are not realistic assumptions.  A model incorporating a spatially
explicit landscape and requiring 2-loci for resistance was developed for
this report by one Panel member (see Appendices 3 and 4).  As found with
single-locus models (Peck et al. 1999, Storer et al. 2003, Sisterson et
al. 2004), in this 2-locus model hotspots can occur when refuges (either
natural or structured) are rare.  When hotspots occur, resistance
evolution can be much more rapid than in the absence of hotspots.  This
poses a challenge for IRM, because strategies must be designed to
eliminate all possible areas where hotspots might arise; a hotspot in a
single county might cause resistance that spreads throughout a region. 
The model shows that hotspots can occur even when males are broadly
dispersive provided females have limited dispersal; therefore, direct
information about female movement is needed to assess the risk of
hotspots.  Finally, in the model hotspots seem to overwhelm the effects
of a fitness cost of resistance.  Although fitness costs to resistant
genotypes (i.e., reduced fitness in the absence of selection from Bt)
generally slow resistance evolution (see "Fitness costs" below), this
effect was minimal when the model produced hotspots. 

The Panel includes this model to emphasize several points: (1) Model
structure is critically important; Monsanto’s non-spatial model cannot
produce hotspots and therefore might ignore an important component of
resistance evolution.  (2) Simple changes in model assumptions can lead
to dramatically different outcomes. (3) Our ability to determine
appropriate model structure is limited by the lack of basic information
needed to understand resistance evolution, especially in a 2-loci
system.  Because the appropriate model structure cannot be determined,
the Panel cautions against relying on any one model or modeling
technique for evaluating resistance to Bollgard II.

Socio-economic factors

The assumption that the market share of each product would remain stable
over a 30-year time horizon is unrealistic.  If a particular product
fails to control a pest, we should expect growers to change their choice
of control measures.  Monsanto’s product market share scenarios do not
account for this important socio-economic factor.  For example, in
scenario 3-N for Texas (see Table 14 in Gustafson and Head 2005)
resistance to product A (WideStrike) evolves in 19.5 years, resistance
to product B (Bollgard) evolves in 12.8 years, while resistance to
product C (Bollgard II) does not evolve within the 30-year time horizon.
 The efficacy of Bollgard II at 30 years suggests that resistance will
evolve soon after execution of the model was truncated.  It is
unrealistic to assume that growers would continue to use the two failed
products for 17 (Bollgard) and 10 (WideStrike) years.  Continued use of
these products is equivalent to using these areas as refuge for product
C (Bollgard II).  

A more realistic assumption is that product market shares will decline
for products A and B and that growers will replace these products with
either an increase in the market share of product C or resort to some
alternative control measure such as pesticides (another alternative
refuge).  However, Monsanto assumes that agricultural producers plant
the maximum allowable proportion of Bt cotton at the state level which,
as time-series data on adoption show, is not likely to be the case over
any 30-year period.  This assumption leads to the simulation result that
Cry1 and Cry2 resistance will evolve more rapidly than is likely to be
the case in reality.  TBW results are unlikely to be affected by
including a grower behavioral response.  However, from a quantitative
perspective, including the grower’s behavioral response (Livingston et
al. 2006) could dramatically change the time to resistance for CBW
results.

Potential for Metabolic Cross-Resistance

One Panel member challenged the assumption that resistance to each toxin
is associated with an independent single locus.

The three-gene model for insect resistance evolution considered by
Monsanto makes assumptions about the action and interaction of
resistance genes that do not fully reflect the resistance mechanisms
reported in the scientific literature.  Thus modifications of the model
to incorporate such resistance mechanisms might yield different
predictions of the durability of Bt cotton.  Three assumptions made by
Monsanto are discussed below:  1) independent toxin action, 2) absence
of minor resistance mechanisms, and 3) absence of a sequestering
mechanism.

1)  Toxins may not act independently.  In the model presented by
Monsanto, survivorship of an insect on a single toxin is represented as
a logistic function of the toxin-receptor complex, and toxins are
assumed to act independently so that the total survivorship is the
product of the single-toxin survivorships.  However, some studies show
that toxins do not act independently, but instead synergistically or
antagonistically (Tabashnik 1992).  An example of antagonistic action,
in which the mortality caused by a combination of toxins is lower than
predicted by the mortality of each singly, is given by Liao et al.
(2002).  In this study, the median lethal concentration (LC50) of Cry1Ac
and Cry2Aa presented together in various ratios was considerably less
than the LC50 predicted by models of independent action of the separate
toxins, for a susceptible strain of Helicoverpa armigera.  This would
imply that the multiplicative aspect of Monsanto's model could be
underestimating the probability of survivorship on Bollgard II cotton of
insects that have resistance to one of the two toxins.

2)  The evolution of "minor" resistance mechanisms may influence
selection on "major" resistance genes.  In the model presented by
Monsanto, the only mutations that affect resistance are in the receptors
for the specific toxins.  This "target-site" resistance has been firmly
established as a major resistance mechanism, and mutations in one of the
receptors (a 12-domain cadherin protein expressed in the midgut
epithelium) have been identified that cause high Cry1Ac resistance in
three lepidopteran species.  (These are TBW (Gahan et al. 2001), pink
bollworm Pectinophora gossypiella (Morin et al. 2003), and the old-world
cotton bollworm Helicoverpa armigera (Xu et al. 2005)).  However,
additional resistance mechanisms are known, such as changes in proteases
that activate the protoxin (Oppert et al. 1997, Li et al. 2004) or
changes in abundances of other binding targets for Cry1Ac such as
alkaline phosphatase (Jurat-Fuentes et al 2005).  These are "minor"
resistance mechanisms that acting alone would not permit survival on
transgenic cotton.  But by reducing the amount of toxin that has access
to the receptor, they may contribute to resistance, especially in
individuals that are heterozygous for the mutation in the receptor
itself.

3)  A novel, recently-proposed resistance mechanism of "toxin
sequestration" could have profound effects on resistance development on
single-gene and pyramided cotton varieties.  Gunning et al. (2005)
claimed that overproduction of esterases in the "silver selected" strain
of Helicoverpa armigera was responsible for Cry1Ac resistance in that
strain.  They proposed that esterases could bind to the toxin and keep
it from interacting with its receptor in the midgut.  Overproduction of
esterases is well-established as a mechanism of resistance to
organophosphorus insecticides in aphids and mosquitoes.  Each esterase
subunit binds a single molecule of the insecticide in its active site,
sequestering the insecticide and keeping it from reaching
acetylcholinesterase, its target in the nervous system.  Gunning et al.
(2005) proposed the same sort of sequestration was operating in the
"silver selected" strain.  Such a mechanism would be expected to be
energetically costly, especially if a 1:1 ratio of sequestering esterase
to Bt toxin were required.  However, if the esterase-toxin complex
aggregates with other toxin molecules, such a ratio might not be
required.  Instead, there might be a general precipitation of toxin,
such as induced by elastase TPP-75 (a digestive protease) produced in
the midgut of the spruce budworm Choristoneura fumiferana, which
precipitates Cry1Aa at a 1% molar ratio (Milne et al. 1998).  

The report by Gunning et al. (2005) is highly controversial, and
although it was published under peer review in a respected journal, the
results have not been independently confirmed by other workers to date. 
If they are, and if direct evidence is produced for an esterase present
in the midgut lumen in sufficient quantities to prevent Cry1Ac from
binding to its receptor(s) in the midgut, this information would have to
be considered in formulating models of resistance development.  In
addition, if such an esterase could sequester other toxins, such as
Cry1F or Cry2Ab, this single-resistance mechanism could protect the
insect from pyramided varieties of Bt cotton as well as Bollgard.

The consequences of this putative new sequestration mechanism for the
modeling would require addition of a term describing the esterase-toxin
complex.  If this complex could be formed for any Bt toxin, the
three-gene model would probably reduce to a single-gene model where an
increase in the resistance allele frequency would have equal impacts on
the efficacy of Bollgard, WideStrike, and Bollgard II cotton.  Such a
resistance gene, if it existed, would be essentially immune to the
benefits of pyramiding in reducing the threat of resistance.  Moreover,
in this case pre-selection with Bollgard could facilitate
cross-resistance to the pyramided varieties.

Single Locus per Toxin

Here, “single locus per toxin” is defined to mean that it only takes
an allele at a single locus to confer resistance to a single toxin, and
there is no possible cross resistance from such alleles.  Monsanto
modeled resistance to each toxin as a single-locus problem affecting the
independent binding sites.  Therefore, the pyramided products containing
two toxins require resistance to develop at two independent loci. 
Excluding the possibility of metabolic cross-resistance presented in the
above scenario, single locus resistance for each toxin (as opposed to
resistance to each toxin requiring multiple resistance alleles at
multiple loci) represents the fastest possible path to resistance
evolution in the TBW where both Cry1Ac and Cry2Ab are considered to each
be at a “high dose” (i.e., if phenotypic resistance is completely
recessive at all loci involved in resistance).  This is not the case
with CBW where even Bollgard II is not a high dose and any resistance
allele with a small effect on tolerance of one or both toxins is
expected to be selected for.  In some cases the forms of resistance that
have been developed in laboratory colonies appear to involve multiple
loci and the aggregation of several minor forms of resistance.  Alves et
al. (2006) assessed the number of loci involved in resistance to the
Cry1Ab for two laboratory-derived resistant strains of Ostrinia
nubilalis, the European corn borer.  One method of assessment suggests
that resistance involves at least 10 different loci while a second
assessment technique suggests there are more than 20 loci involved.  In
contrast, the same techniques imply that resistance in the diamondback
moth, Plutella xylostella, is related to a single locus (which may
explain why this insect exhibits resistance in the field).

We cannot assume that the number of loci involved in Bt resistance in O.
nubilalis is directly translatable to either CBW or TBW.  However, field
level resistance has not evolved in any of these three pests and allele
frequencies for resistance remain at or below the levels observed when
transgenics were first introduced.  Obviously, we cannot assume that
each toxin requires a specific combination of alleles at 10 completely
independent loci.  But, we can assume that resistance to each toxin
involves more than one independent locus.  

A two-toxin product with resistance conferred by a single independent
locus for each toxin is probably comparable to a single toxin product
with resistance conferred by two independent loci.  Given this
assumption, the direction of this effect is to delay resistance and the
magnitude of the effect is probably exponentially related to the number
of loci involved.  By modeling resistance as a single locus for each
toxin, the approach is both reasonable and conservative for TBW.  If a
multi-locus resistance were modeled for each toxin the effect would
delay resistance in TBW and the magnitude of the delay would overwhelm
most if not all of the uncertainties mentioned.  This would not be the
case for CBW.

Fitness Costs

The model was run without fitness costs.  This is also a conservative
assumption, perhaps very conservative (except possibly if hotspots
occur; see Appendix 3).  The first paragraph of Gustafson and Head
(2005), Section 4.4 summarizes the literature evidence that Cry1Ac
resistance is associated with fitness costs.

Fitness costs can take many forms, reduced fecundity, delayed
development, and reduced overwintering survival to mention but a few of
the possibilities.  Fitness costs are effectively selection against the
trait of interest once accounting for the direct fitness benefits of the
trait under selection (i.e., factoring out the benefits of resistance to
Bt toxins in Bt fields).  The consequences of including fitness costs in
models of the evolution of resistance dramatically change the possible
outcomes.  If fitness costs are included there are four potential
outcomes:

Resistance could be delayed.

The allele frequency could reach a stable equilibrium.  This type of
system operates as a cline, a gradient of opposing forces that reaches a
balance point with fitness costs selecting against resistance while
toxins select for resistance.  . 

The population can become extinct.

The population can persist but the allele is extirpated from the
population.

With some notable exceptions like the diamondback moth (Plutella
xylostella), resistance to the Cry complex of toxins has been developed
in laboratory colonies.  Resistant individuals from these colonies
exhibit developmental delays and a reduction in biomass when compared to
susceptible individuals of comparable age not exposed to the toxin. 
These are qualitative descriptions that imply metabolic fitness costs.  

Peck et al. (1999) simulated Bt resistance evolution for TBW in North
Carolina.  Developmental delays during the growing season caused a
temporal separation of the adults emerging from refuge and Bt fields. 
Although this kind of separation may not fit the typical definition of a
fitness cost, it resulted in positive assortative mating among resistant
individuals and increased the rate of resistance evolution.  However, as
the magnitude of the developmental delays was increased, the time to
resistance evolution became less predictable.  Differences in the
synchrony of the diapausing life-stage with the end of the growing
season often resulted in selection against the individuals that had been
selected for resistance earlier in the growing season.  When these
relationships were combined with allele frequencies below 0.03 the
allele was often extirpated from the population.

Carrière et al. (2001) attempted to estimate the effect of diapause
asynchrony quantitatively in Pectinophora gossypiella.  Three different
laboratory colonies exhibiting different levels of Bt resistance were
reared without exposure to the toxin under photoperiods that should
induce diapause.  Individuals that pupated were not in diapause and
would not survive the winter.  Individuals that did not pupate were
assumed to be in diapause.  Alternative host availability for this pest
in Arizona is so limited that cotton is the only viable host.  So,
individuals that emerge prior to spring planting are unlikely to find
hosts for their offspring.  Homozygous resistant individuals suffered a
71% reduction in spring emergence when compared to homozygous
susceptible individuals.  In addition, the strains with the greatest
level of resistance suffered the greatest overwintering mortality.

Caprio (2006) submitted a public comment to the Panel presenting a
cursory evaluation of the effects of fitness costs for the MidSouth
region (Mississippi).  Figure 7 in that submission shows that 25% of the
model runs led to resistance evolution in less than 15 years in the
absence of structured refuge, while Figure 8 shows that approximately 5%
of the model runs led to resistance evolution with a structured refuge. 
Fitness costs in this model were implemented as a reduction in fecundity
ranging from 0.0 to 0.5 with a most likely value of 0.1 but the
mechanism of selection against resistance is irrelevant to the question.
 These fitness costs are still well below the losses reported by
Carrière et al. (2001).  Also submitted to EPA for the Panel was a
second model (Gould et al. 2006) showing that when there is a high dose
of each of two toxins, the resistance alleles to the 2 toxins never
reach a problematic level as long as there is a recessive fitness cost
of 0.05 and a 10% refuge.  Even if selection with a single toxin
cultivar initially raises the frequency of alleles for one toxin above
0.90, after introduction of the two toxin cultivar the frequency of both
alleles reach a very low equilibrium level.  However, the Gould et al.
model is deterministic and must be viewed with caution.

Model runs that result in extinction have been reported since IRM
modeling began (Onstad and Guse 1999, Peck et al. 1999, Sisterson and
Tabashnik 2005).  Peck’s model includes trivial fitness costs as an
offset for the increase in allele frequency due to mutation.  The Onstad
and Guse model includes no explicit fitness costs.  The scenarios in
which the population is driven to extinction result from fitness costs
that are an emergent property of the complex interactions that occur
between the definition of the landscape, the biological and behavioral
characteristics of the target pest and the rarity of the resistance
allele.  Resistance management strategies include an assumption that 1)
extinction is not a probable outcome for chronic pest populations and 2)
resistance will evolve in response to the widespread use of transgenic
crops.  Modelers have therefore focused on the situations and scenarios
that lead to resistance and discounted the conditions that result in
extinction and extirpation of the allele for resistance. Caprio’s
(2006) submission is important because it illustrates the proportion of
model runs that do not conform to these two assumptions about the
outcome.  Nonetheless, Monsanto does not argue that Bollgard II is
likely to cause the extinction of either CBW or TBW, and given this, the
relevance of simulations leading to extinction is uncertain.

Monsanto references several other papers suggesting that fitness costs
are associated with resistance but these three examples illustrate that
fitness costs can delay and perhaps even prohibit the evolution of
resistance to Cry toxins.  There are many potential mechanisms for
fitness costs but regardless of the mechanism, fitness costs represent
one of the best hypotheses for explaining why field populations have not
developed resistance in the past decade.  By not including fitness
costs, Monsanto has taken the most conservative approach to modeling
this system.  If fitness costs were included in the model, the magnitude
of the delay effect could overwhelm the opposing effect of most of the
uncertainties presented.  On the other hand, one Panel member indicated
that the effect of hotspots can easily overwhelm a fitness cost of
resistance and lead to rapid resistance evolution despite high fitness
costs as described in Appendix 3 (see also Table A3-1).

Pre-Selection

Given initial allele frequencies for resistance that are very low, in a
range approximately from 0.001 to 0.01, the time until resistance can be
observed in the field population is an inverse exponential function of
the initial allele frequency.  That is, when the initial allele
frequency is very low, minor increases in the initial allele frequency
result in significantly shorter times to resistance.  Selection for
resistance that may have occurred prior to removal of structured refuge
would result in an increase in the allele frequency for resistance in
the field population.  So, the most conservative answer to the question
is that in this range of allele frequencies any increase shortens the
time to resistance dramatically.

The model uses an initial allele frequency of 0.002 based on Gould et
al. (1997).  In this study, male TBW moths were collected in the
MidSouth region (Mississippi, Louisiana and East Texas) prior to the
introduction of transgenic crops.  The estimated resistance allele
frequency was actually lower than the value used in this model (0.0015
with a 95% CI ranging from 0.0004 to 0.0041).  The study employed a
resistant laboratory colony derived from TBW collected earlier in North
Carolina.  A best approximation of the allele frequency for resistance
in the North Carolina population is 0.001 (Gould et al. 1997). 
Uncertainty about this value was strictly qualitative.

Burd et al. (2003) estimated the allele frequency for resistance in CBW
collected in North Carolina in 2000, several years after transgenics
were in use.  Allele frequencies for resistance to the Cry1Ac toxin
averaged 0.00043 with a 95% CI ranging from 0.00001 to 0.00239.  The
frequency of the resistance allele for the Cry2Aa toxin averaged 0.00039
with a 95% CI ranging from 0.00001 to 0.00216.

It must be noted that the rarity of the allele makes it extraordinarily
difficult to detect.  The 95% confidence intervals reflect this
uncertainty.  While the introduction of transgenics has significantly
changed the spatial and temporal properties of the landscape there is
little evidence that any increase in resistance allele frequency has
occurred in field populations in the US over the past decade.  There is
evidence of increasing resistance to Cry1Ac for H. armigera in China
where there are no structured refuges (Kongming Wu, F. Gould, et al.,
unpublished; Table 7-1).  Monsanto’s model was run conservatively at
the mid to high range for allele frequencies for both pests.

Table 7-1.  Relative average development ratings for six-day old larvae
of Helicoverpa armigera F1 and F2 generation female lines in populations
from Anci, Hebei Province and Xiajin, Shandong Province, China.  (From
K. Wu, unpublished.)



	Relative average development rating per line









Means (+ SE) with different letters are significantly different (P <
0.05; LSD test).  Capital letters in same column in same location
indicate difference in different years.  Small letters in same row
indicate differences between generations in same year.

 a LF-R is a Bt-resistant laboratory strain.  SS1 and SS2 are
Bt-susceptible laboratory strains.

Agency Charge

8.  Modeling suggests that the overall durability of Bollgard II cotton
can be enhanced if Bollgard cotton is removed from the marketplace. 
This conclusion is supported by other researchers who examined the
benefit of managing resistance evolution to two toxins with dissimilar
modes of action using a pyramided approach (Zhao et al., 2005; Roush,
1998; Livingston et al., 2004; Hurley, 2000; Caprio, 2005).  On the
other hand, the concurrent use of single- and two-gene Bt plants can
offer exposed populations a “stepping stone” to develop resistance
to both proteins.  Bollgard, Widestrike, and Bollgard II cotton exist in
a mosaic in southeastern cotton growing regions, with Bollgard
dominating the total acreage.   In 2004, Bollgard cotton acreage
accounted for >95% of all Bt cotton acreage in the U.S. (see Head et
al., 2005, MRID# 467172-03).  Encouraging the adoption of Bollgard II
will increase the overall durability of all three Bt cotton products. 
From an insect management point of view, removal of Bollgard cotton from
the marketplace would benefit the two-gene products, Bollgard II and
WideStrike.   

The Panel is asked to address the implications for selection for CBW and
TBW resistance if the mosaic of single gene and dual gene products
remains in the marketplace for a number of years.  How would selection
pressure be reduced if the single gene product is removed from the
marketplace gradually (e.g., >3 years) or rapidly (e.g., ≤3 years)
over a period of years?

Panel Response

	Two important dimensions to the question of mosaics of single and dual
gene products are product market share and temporal variability. 
Monsanto addressed the product market share dimension by looking at how
resistance evolved for seven scenarios characterized by different
product market shares.  Temporal variability was not explored by
Monsanto.

 	To explore these two dimensions, it is useful to conduct a controlled
experiment.  In the first experiment, one can hold the temporal
variability of the mosaic constant, while varying product market shares.
 For example, one can assume that the same amount of each product is
planted every year, while allowing the proportion of each product’s
market share to vary from one treatment to the next (Monsanto’s
analysis).  In the second experiment, one can hold the total market
share constant for each product, while allowing the temporal variability
of market shares to increase.  For example, one can assume that the
proportion of cotton acres represented by each product does not change
over a specified period of time (e.g., 200 generations), while allowing
the proportion of market shares for each product in each generation to
vary from one treatment to the next.  In preparation for this SAP
meeting, these two experiments were conducted with the deterministic two
gene/two product simulation model reported in Hurley et al. (2006) using
the parameters reported in Table 8-1.  In the second experiment, four
examples of temporal variation in product market shares were explored,
holding the total product market shares constant at 0.5 over 200
generations.  These examples are reported in Figure 8-1.  These
simulations were designed to demonstrate the general effects of product
market share and temporal variability in product market share.  As such,
they were not calibrated to any specific pests, products, or
agro-ecosystem.

Table 8-1.  Controlled parameter values for simulation model examples.





Proportion of Refuge

0.05

Refuge Survival Rate for all Genotypes

1.0

Recombination Factor

0.50

Proportion of Non-Random Mating

0.0

Shared Toxin (e.g., Cry 1Ac)

                      Susceptible Homozygote Survival Rate



0.0

Heterozygote Survival Rate

0.02

Resistant Homozygote Survival Rate

1.0

Initial Frequency of Resistant Alleles

1.0x10-3

Pyramided Toxin (e.g., Cry 2Ab2)

Susceptible Homozygote Survival Rate



Figure 8-1.  Example of temporally variable transition rates with equal
market shares for single and pyramid toxin products (these transition
rates were used for the analysis reported in Table 8-3).

	In the first experiment, reducing the market share of the single toxin
product slows the evolution of resistance to the pyramided toxin, which
replicates Monsanto’s results (see Table 8-2).  For resistance to the
pyramided toxin, there was not a monotonic relationship between market
share and the evolution of resistance.  Initially, increasing the market
share of the pyramided product speeds the evolution of resistance.  But
once the market share of the pyramided product is high enough, further
increases slow the evolution of resistance.  The explanation of these
results has to do with several countervailing effects.  Increasing the
market share of the pyramided product reduces selection for the shared
toxin, which increases the time to resistance for the shared toxin. 
Increasing the time to resistance for shared toxin allows dual modes of
action to operate for a longer period of time in the pyramided product,
which is positive for resistance management.  Increasing the market
share of the pyramided product increases selection for the pyramided
toxin because a greater proportion of the insect population is exposed
to this toxin.  This increased selection pressure reduces the time to
resistance for the pyramided toxin, which is negative for resistance
management.  Increasing the market share of the pyramided product also
means that there will be less refuge for the pyramided toxin once
resistance evolves to the single toxin product because the single toxin
product serves as refuge for the pyramided toxin after resistance to the
shared toxin has evolved.  This effect also reduces the time to
resistance for the pyramided toxin, which is negative for resistance
management.  This simulation illustrates the opportunity cost of
increasing the market share of the pyramided product in order to reduce
selection for the shared toxin.  This opportunity cost is an increase in
selection pressure for the pyramided toxin.

Table 8-2. Example of the relationship between product market share and
time to resistance.



	Single Toxin 

Variety

(e.g., Bollgard)	Pyramided

Variety

(e.g., Bollgard II)	Shared

Toxin 

(e.g., Cry 1Ac)	Pyramided

Toxin





	Note:  When resistance failed to emerge after 200 generations, the
proportion of resistant alleles is reported in parentheses.

	For the second experiment, high temporal variability speeds the
evolution of resistance for both toxins (Table 8-3).  This experiment
provides one interpretation of what has been called the “stepping
stone” effect.  Holding total market share constant, increased
temporal variability in the adoption of the pyramided product means that
the single toxin product is on the market longer, which speeds the
evolution of resistance to the shared toxin.  This means the pyramided
product will effectively become a single toxin product sooner, which
speeds the evolution of resistance to the pyramided toxin.  This
“stepping stone” effect is negative for resistance management and
will be more pronounced when the market share of the pyramided product
is relatively small.

Table 8-3.  Example of the relationship between the variability of
product market shares (see Figure 8-1) and time to resistance.



	

Product Market Share Variability	Shared Toxin

(e.g., Cry 1Ac)	Pyramided Toxin





Note:  When resistance failed to emerge after 200 generations, the
proportion of resistant alleles is reported in parentheses.

	While this experiment suggests that a quick transition to pyramided
toxin products may not always be the best strategy due to the
opportunity cost of increasing selection for the pyramided toxin, slow
transition rates can only be supported by relatively heavy selection
pressure on the pyramided toxin.  The Panel agreed that the majority of
papers published over the past decade using a variety of modeling
strategies (from simple deterministic models to complex stochastic and
spatially explicit models) have found that quicker transitions to
pyramided toxin products is preferable in terms of resistance
management.

	The Panel noted that mosaics and slow transitions to pyramided products
could negatively impact minor as well as major resistance mechanisms. 
Such mechanisms include alterations in midgut protease activity (Li et
al. 2005) and changes in the abundances of proteins that bind Bt toxins
such as aminopeptidases and alkaline phosphatase (Jurat-Fuentes and
Adang, 2004).  These minor mechanisms cannot by themselves enable
survival of insects on transgenic cotton, but may supplement resistance
levels in heterozygotes for target-site mutations.  Products expressing
a single Bt toxin are more likely to select for these minor mechanisms
than pyramided toxin products because of the higher toxicity of the
latter.  These mechanisms may then boost resistance levels of insects
carrying mutations for the receptor to the pyramided toxin.  This would
accelerate resistance development, relative to the models utilized by
Monsanto that only consider target-site resistance.

	While the majority of the literature that explores the effects of the
transition from single to pyramided toxins finds slow transitions have a
negative impact on resistance management, there are counter examples.  A
stochastic, spatially explicit simulation model (Storer et al. 2003;
Livingston, Storer, Gould, Kennedy, and Van Duyn, unpublished) shows
that quicker transition rates to pyramided toxins slow resistance to the
shared toxin, while increasing resistance to the pyramided toxin.  This
result is an example where there is strong selection for the pyramided
toxin relative to the shared toxin, which is partly attributable to the
resistance fitness costs assumed in the model.  Livingston et al. (2006)
used an extended version of this model that incorporates the behavioral
response of a representative cotton producer and Cry2Ab2 and pyrethroid
resistance evolution, and found that grower behavior can affect
management of Cry1Ac and Cry2Ab2 resistance in CBW.  This result is
driven by two factors: economic and biological.  In the model, the
cotton producer chooses between cotton varieties based on profitability.
 When resistance to the shared toxin (Cry1Ac) evolves in the CBW
metapopulations, the representative farmer switches from planting only
Bollgard (the single toxin product) to planting only Bollgard II (the
pyramided toxin product), which reduces Cry1Ac selection pressure.  A
0.025 fitness cost of carrying the Cry1Ac resistance allele,
conventional insecticide (pyrethroids) use, and the presence of non-Bt
corn and soybean fields, which provide unstructured refuge, subsequently
allows susceptibility in CBW to Cry1Ac to re-evolve.  Furthermore,
profit received by the representative eastern North Carolina producer is
highest when the structured refuge requirement is removed for both
Bollgard and Bollgard II, resistance to Cry1Ac evolves very slowly and
the Cry1Ac resistance allele frequency declines when the producer
switches to Bollgard II, and Cry2Ab2 resistance never evolves.

	The Panel generally agreed that the weight of evidence supports a quick
transition to pyramided products when there is no mechanism for
resistance fitness cost.  One Panel member pointed out that it is for
this reason that the Australian cotton industry developed an agreed
policy of rapid transition from single to pyramided Bt cottons.  To
further protect the future benefits of pyramided products, this policy
imposed a cap on the area of single toxin cotton until pyramided
varieties were ready for release.  However, the Panel also noted that
resistance fitness costs and economic factors that affect grower
behavior could promote the transition from single to pyramided toxin
products.  One Panel member emphasized that it is important to
incorporate grower behavior and what seed producers are doing with
Monsanto’s Cry toxins.  For example, in eastern North Carolina,
cottonseed varieties with the Cry1Ac gene that growers actually plant
are available only when stacked with Monsanto’s Roundup Ready gene;
and cottonseed varieties expressing Cry1Ac and Cry2Ab2 that growers
actually plant are available only when stacked with Monsanto’s Roundup
Ready Flex gene, the latter allowing growers to use Roundup longer
during the growing season and more effectively.  Livingston et al.
(2006) incorporated these important aspects of the eastern North
Carolina situation in their analysis, and the results strongly suggest
that the behavior of cotton growers and seed companies likely will
expedite transition from Bollgard to Bollgard II.

	While Monsanto’s analyses were consistent with the views of the
panelists, Monsanto’s analyses only contributed to the recommendations
of the Panel and did not sway the Panel.  This is important to note,
because it illustrates the importance of having multiple, independent
lines of evidence in making scientific recommendations.

Overall Data/Results Interpretation

Agency Charge

9.  There are three major variables to evaluating structured refuge for
Bt crops:  a) production of a sufficient number of susceptible insects
relative to any resistant survivors of the Bt crop, b) proximity of the
refuge to the transgenic crop to facilitate random mating between
susceptible (from the refuge) and resistant (from the Bt crop) insects,
and c) developmental synchrony of the refuge with the transgenic crop to
promote random mating.  

Given Monsanto’s sampling, gossypol analysis, spatial and temporal
analyses, and modeling evaluation, the Agency asks the panel to comment
on whether Monsanto’s analysis scientifically supports the conclusion
that natural refuge will be comparable to the effectiveness of
structured refuge for management of TBW resistance to the Bt proteins
expressed in Bollgard II cotton for each of the four regions:  the
Carolinas, Georgia, Mississippi Delta, and Texas.

Panel Response

General Comments

	Firstly, the Panel would like to compliment Monsanto on the extent and
quality of information provided to support their case that natural
refuge will be sufficient for resistance management for Bollgard II
cotton.  A number of aspects of the research significantly improve
understanding of the ecology of Heliothine pests associated with Bt
cotton systems.

	Nonetheless, there are many uncertainties, caveats, and assumptions
evident throughout the modeling and analyses presented by Monsanto and
revealed by the Panel discussion around the questions posed by EPA. 
Given the magnitude of the potential consequences of the decision EPA
must make, the Panel must apply the most rigorous scientific questioning
to all aspects of the proposal, the data, and the conclusions.

The Gossypol Technique

	One particular area of concern relates to the innovative gossypol
analysis developed by Monsanto.  The validity, accuracy and
repeatability of this technique for identifying the non-cotton fraction
of the TBW population is central to Monsanto’s data presentation and
argument for a natural refuge.  Although the Panel received a
description of the analytical technique, the description was incomplete.
 We thus highlight the following concerns:

The proportion of non-cotton moths is determined by the inability of the
analytical assay to detect gossypol in the bodies of field-collected
male moths in pheromone traps.  Since the assay does not provide a
quantitative estimate of the amount of gossypol actually present, it is
potentially liable to significant biases that have not yet been
experimentally evaluated.

For example, if the gossypol technique has a high detection threshold,
it is quite possible that a number of samples would be classified as
being from non-cotton sources when in fact they simply have gossypol
values that are below the detection level of the method.  We thus need
more information about the sensitivity of the method.

If gossypol concentrations decline in male moths from cotton that have
remained dead for a week in a pheromone trap in the field, there could
be a significant underestimation of the fraction of moths originating
from cotton.  This would project into a significant overestimation of
the size of the natural refuge in the models, with a concomitant
underestimate of the time taken to develop resistance in the pests. 
This bias factor could equal or exceed the other sources of variation of
concern to the Panel, such as the variability between regions or on
different dates.

	Although Monsanto gave some verbal responses to these concerns during
the public meeting, the Panel recommends a rigorous examination of the
gossypol analysis technique.  This should include at a minimum:

A review by EPA staff with appropriate chemistry expertise;

Publication of the assay method in a peer-reviewed journal;

Performance and publication of experiments to mimic the conditions
experienced by trapped males prior to analysis;

Validation of the methodology by independent laboratories.

	If the innovative gossypol analysis does stand rigorous examination,
the Panel notes that it will be of considerable value for ongoing
research by others working in this field.

	In the remainder of our assessment we have assumed that the gossypol
analysis is accurate and sensitive, and that the estimates of non-cotton
moths are reliable.

Estimates of Natural Refuge and Modeling of Resistance Risk

	Monsanto has clearly incorporated a number of elements of conservatism
throughout their analysis, explicitly utilizing a "worst case" approach
in many parameters in their modeling.  The Panel questions whether
collectively the data and modeling allow a sweeping conclusion that
natural refuge is sufficient to delay resistance to Bollgard II across
the four Cotton Belt regions outlined in the submission (Fig. 9-1) –
North Carolina (NC), Georgia (GA), Mississippi Delta (MS), East Texas
(TX) – at a spatial scale within these regions at which resistance is
most likely to develop.

	Monsanto’s simulations indicate that with current estimates of
natural refuge, development of Bt resistance in TBW to Bollgard II was
unlikely within 30 years in any of the four regions.  Monsanto does not
argue that natural refuge will be strictly comparable to the
effectiveness of structured refuge, except to the extent that it will
delay resistance for at least 30 years.  Because all simulations were
truncated at 30 years, it is not possible for the Panel to judge the
relative effectiveness of resistance management with or without
structured refuge.  That said, a predicted durability of at least 30
years would seem more than adequate for biotechnology-based products. 
Monsanto’s decision to use worst-case estimates for input parameters
is conservative and may partially mitigate problems associated with
parameter and model uncertainty.  However, the sensitivity of model
output to parameter values and model assumptions makes this difficult to
assess.

 

Fig. 9-1.  Modified from Fig. 2 in Gustafson and Head (2005),
incorporating modifications by Reynolds (2006).  Caption in Gustafson
and Head (2005) reads, "Definition of CBW modeling regions, which
include all counties with some Bt. cotton sales (2003-5) and adjacent
counties with at least one of the following crops according to USDA
estimates for 2004: corn, cotton, peanuts, or soybeans."

 	The Panel noted some potentially serious biases in the calculation of
natural refuge for TBW and CBW (see Panel response to Charge 6), which
should be addressed by Monsanto.  Despite the potential biases in
estimates of natural refuge, some Panel members concluded that for
Georgia and North Carolina, Monsanto has demonstrated that there are
significant and reliable non-cotton refuges present as part of the
cropping system and the wider environment, and this should be adequate
to manage Bt resistance in TBW associated with cotton systems involving
Bollgard II cotton.

	Following this FIFRA SAP meeting, a Panel member provided additional
analysis and comments regarding the validity of extrapolating data from
areas sampled by Monsanto in North Carolina and Georgia to areas that
were not sampled.  Such comments were not considered or reviewed by the
Panel during the meeting, and are being provided as an appendix to these
meeting minutes (Appendix 5).

	Other regions are more problematic.  While the simulations suggest
long-term durability against TBW in all regions, the Panel does not
believe that one year of data for East Texas and quite limited and
variable sampling through parts of the MidSouth are sufficient to
establish the reliability of adequate natural refuge for TBW.  Texas
especially must be sampled more thoroughly, both spatially and
temporally.  Areas representing different ecozones in that state must be
sampled.  One year of data does not provide coverage for the somewhat
unpredictable rainfall in that region which may cause high variability
in alternative host availability and phenology.  For some parts of
Texas, we have zero years of data, such as the Lower Rio Grande Valley
and the High Plains.  It was unclear to the Panel whether the Monsanto
amendment covered all of Texas, given that west Texas and the High
plains are currently part of a pink bollworm eradication program
utilizing Bollgard cotton.  Nonetheless, we highlight that the High
Plains are very dry with most cotton being irrigated.  The literature
indicates (Benedict 2005) that there are almost no significant
alternative hosts available to TBW in that area.  Given the wide variety
of agroecozones in Texas, it is not possible to extrapolate data on
natural refuge from East Texas to the Texas High Plains, the Lower Rio
Grande Valley, the Winter Garden Region, the Rolling Plains, etc.

	The variability and uncertainties highlighted above in calculations of
natural refuge and its utility in minimizing the risk of Bt resistance
evolution indicate that Monsanto’s data analysis could be much more
comprehensive.  As noted in its response to Charge 5, the Panel would
prefer a more integrated and comprehensive analysis of the spatial and
temporal variability of refuge estimates and moth trap data to tease out
crucial details regarding the appropriate spatial regions and the
critical temporal periods that pose the most risk.

Worst-case scenarios

	Throughout its submission, Monsanto indicates that it has simulated the
worst-case scenario with regard to the adequacy of natural refuge.  One
Panel member noted that resistance is likely to evolve in one area and
then spread and that by picking counties with the lowest effective
refuge, Monsanto is in fact not selecting the “worst-case” scenario,
but instead the “most-likely” scenario for resistance development. 
Little re-assurance can be gained from a risk assessment perspective by
picking the areas with lowest effective refuge.  The true worst-case
scenario involves first selecting the areas with lowest effective refuge
and then using worst-case scenarios for other parameters applied for
these areas.

Assurance of Natural Refuge

	Current requirements for the registration of Bollgard and Bollgard II
include monitoring for Bt resistance evolution (currently done by USDA
Stoneville for Bt cotton).  Resistance will be easier to prevent than
remediate after it has arisen.  By removing structured refuge, Monsanto
will remove farmer control of effective refuges and transfer that
“responsibility” to other farmers and the natural environment.  How
can Monsanto guarantee that natural refuge will always be there and be
adequate?  Who would be responsible for that assurance?  The Panel
suggests that if the structured refuge requirement were removed, a
comprehensive monitoring effort would have to extend beyond resistance
monitoring per se to include monitoring levels of effective refuge
through estimates of area, moth productivity, or ongoing gossypol
analyses of moths trapped in cotton-growing areas.

	This leaves open the residual issue of how areas of non-cotton host
crops for TBW (tobacco/peanuts/soybean) or CBW (largely maize/soybean)
might change in the future if refuge requirements are dropped for
Bollgard II cotton.  Will more of the maize/peanut/soybean acreage be
switched to cotton with the result that natural refuge will decrease? 
The modeling analysis of Gustafson et al. (2005) includes simulations of
significantly reduced soybean area, and the authors conclude that it
would have little impact on effective refuge for CBW.  However, they do
not include the assumption that the area taken out of soybeans is
replaced with Bollgard II cotton.  Furthermore, no consideration is
given to the possibility that the percent acreage of corn planted to Bt
cultivars will not increase as stacked transgenic corn cultivars become
more prevalent.

Specific Comments on the three key components required for an effective
refuge

a) Production of a sufficient number of susceptible insects relative to
any resistant survivors of the Bt crop. 

	A number of different modeling approaches suggest that pyramided plants
like Bollgard II, including two highly efficacious proteins, require
smaller effective refuges than a single gene product and those quite
small refuges can delay resistance for considerable periods.  Based on
the use of the gossypol technique and carbon isotope ratios, and the
various calculations of areas of potential refuge, Monsanto has clearly
demonstrated that a significant proportion of moths across all regions
are generated from non-cotton sources.  This point is more fully
discussed in the Panel's response to Charge 2.  Setting aside for the
moment the inherent uncertainties in all simple models and the lingering
uncertainties about temporal and spatial variability of natural refuge,
Monsanto’s modeling of the durability of Bollgard II against TBW
utilized worst case estimates of natural refuge from the data set and
indicated resistance could be delayed for at least 30 years by reliance
on natural refuge.  Nonetheless the refuge data also demonstrate that
the currently structured refuge crops of non-Bt cotton are contributing
a significant proportion of both TBW and CBW moths to populations across
the Cotton Belt and must thus be contributing to the delay of resistance
over and above that provided by natural refuge.  What is a sufficient
amount of refuge?  EPA's having already accepted a 5% structured refuge
as adequate for resistance management for Bollgard cotton could suggest
in simplistic terms that any credible estimates for natural refuge above
this level would represent a comparable or at least acceptable
protection against resistance.

	Because resistance will likely evolve in areas with little effective
refuge, these areas are of particular concern.  When the estimated
proportion of effective refuge for CBW and TBW is low (< 10%), higher
levels of uncertainty attach to a number of assumptions and calculations
for the proportion of effective refuge:

1) The estimates of effective refuge become less certain as the
estimates become smaller, because in areas with high Bt cotton usage
(small refuges), few insects are collected (see Panel response to Charge
3).  Thus, when estimates of effective refuge are low, they are also
less precise.  This might be addressed to some extent by additional,
more-intensive sampling as suggested under Charges 2 and 3.

2) When the proportion of effective refuge is low, models of resistance
evolution become more sensitive to these values.  For example, the
difference in time to resistance between the cases of 25% and 20% refuge
is often small compared to the difference between the cases of 10% and
5% refuge.

3) Model uncertainty (different predictions produced by structurally
different models) increases when the amount of refuge is small.

4) Resistance can evolve in “hotspots” where the proportion of
refuge area is small.  This is shown by numerous models of resistance
evolution that include spatially explicit landscapes (unlike the
Monsanto model presented).

5) Small amounts of effective refuge likely will coincide with greater
variability in the amount of refuge through the growing season. 
Therefore, depending on weather conditions, there could be generations
that effectively experience no refuge.

6) If there is little effective refuge, the amount of refuge could be
highly variable due to future changes in cropping patterns,
non-agricultural land use, application of insecticides or other insect
control measures in effective refuge, and environmental change as
discussed above.

	While any one of these areas of uncertainty might not in fact prove
dangerous to IRM, as a group they represent unacceptably high levels of
uncertainty.  All of these areas of uncertainty increase and compound
each other when the area of effective refuge is below 5-10%.

	Spatially explicit modeling (Peck et al. 1999, Storer et al. 2003, and
Sisterson et al. 2004, 2005) demonstrates that, with certain assumptions
about pest mobility, resistance development can commence as hotspots in
the landscape.  In general the evidence for wide-scale mobility of TBW
and CBW might mitigate that effect; however, we do not know the
geographic scale of such hotspots for TBW.  It might be the county
scale, but possibly is smaller than that based on what we know from the
literature about dispersal and gene flow during the cotton-host season
(e.g., Korman et al. 1993, Sparks et al. 1993, Leonard et al. 1995,
Schneider 1999, Han and Caprio 2002, 2004).  The more comprehensive data
analysis suggested under Charges 4 and 5 will assist in identifying the
appropriate spatial and temporal scale for pooling of data and for
identifying gaping holes in the provision of natural refuge that might
serve to generate resistance hotspots.  These “gaping holes” might
reflect inadequate spatial sampling or temporal patterns driven by
climate whereby natural refuge appears adequate in a region one year,
but inadequate the next due to differences in the timing or intensity of
rainfall.  This might well be an issue for the MidSouth and Texas region
where natural refuge is lower and climatic extremes may lead to
significant change in the adequacy of natural refuge.  For these reasons
we argue that further data are necessary to characterize the level of
variability to be expected, so that an informed judgment can be made
regarding the suitability of natural refuge.  Although weather is also
variable in North Carolina and Georgia, the contribution of alternative
hosts is so great that this region may be well-buffered in the provision
of natural refuge.

b) Proximity of the refuge to the transgenic crop to facilitate random
mating between susceptible (from the refuge) and resistant (from the Bt
crop) insects.

	With structured refuge it is possible to specify and regulate the
proximity of refuges to Bt crops.  It is not possible to manage or to
assure proximity when relying on natural refuge, which might derive from
patches of alternative crops and wild hosts which are unlikely to be
distributed uniformly.  Spatial structure of the crop and refuge
environment will interact with mobility of the pest.  Both CBW and TBW
are potentially highly mobile insects (Benedict 2005, Schneider et al.
1989) relative to the spatial patterning of hosts and non-hosts in the
cropping regions where they occur.  This does not necessarily mean that
all moths move extensively every generation and in fact there is good
evidence that they do not (Hayes 1991, Korman et al. 1993, Sparks et al.
1993, Leonard et al. 1995, Schneider 1999, 2003, Bagwell et al. 2000,
Han and Caprio 2002, 2004).  In some parts of the Cotton Belt there may
be the potential for hotspots of resistance to develop where TBW
populations are regularly associated with Bt cotton and lack localized
refuge.  Monsanto’s trapping of male moths in pheromone traps adjacent
to cotton crops indicates at least that those moths which had been
generated on non-cotton hosts have moved into the vicinity of a cotton
crop. 

	Monsanto has done a sound job in seeking to quantify land-use patterns
in the counties where data were collected, and in most areas sampled the
combined analyses clearly indicate that non-cotton hosts do contribute a
proportion of moths to local populations.  However, the counties sampled
are not always representative of all counties within a given state. 
Also, with only 1 or 2 years of data it is difficult to assess the
reliability of natural refuge in proximity to Bt cotton in some parts of
the western end of the Cotton Belt.

c) Developmental synchrony of the refuge with the transgenic crop to
promote random mating

	There have always been questions about the synchrony of heliothine
populations on Bt crops and refuges.  For CBW and H. armigera in
Australia and elsewhere, it appears that survivors on Bollgard cotton
develop more slowly than those on conventional refuge crops. 
Development rates, and therefore adult emergence times, of both CBW and
TBW can differ substantially depending on the host plant species
(Lukefahr and Martin 1964, Nadguada and Pitre 1983, Hayes 1988).  This
immediately raises the specter of lack of synchrony of emergent
refuge-generated moths with those from Bt crops.  TBW displays quite
distinct generational patterns across seasons in the Cotton Belt.  Even
so, generational peaks span 2-3 weeks and get progressively broader
through the season due to partially overlapping generations.  The
temporal scale used in Monsanto’s analysis (monthly) is probably
appropriate for these species.  The more comprehensive generalized
linear modeling of the refuge estimates suggested by this Panel would
provide a more acceptable assessment of the spatial and temporal
variability in refuge adequacy.

Conclusions

	The Panel understands that both Monsanto and cotton industry
representatives would prefer a single management strategy to accompany
Bollgard II cotton across the Cotton Belt, as is the case with Bollgard
and Widestrike cotton.  However, based on the data provided, the Panel
does not believe a single conclusion regarding adequacy of natural
refuge can be applied across the whole Cotton Belt “from Texas to
North Carolina.”

	With the caveat that the accuracy of the gossypol technique must be
validated, some Panel members support removal of the structured refuge
requirement for Bollgard II cotton in North Carolina and Georgia.  The
refuge requirement should not be removed in Alabama and parts of Texas
because these areas were not sampled.  The refuge requirement should not
be removed in Tennessee and East Texas because of insufficient data. 
The refuge requirement should not be removed in Mississippi, Arkansas,
and Louisiana because the data presented suggest that the natural refuge
may be inadequate.  A reassessment of these recommendations, based on
additional data from the excluded areas/states, could be warranted.

REFERENCES

Alves, A. P., T. A. Spencer, B. E. Tabashnik, and B. D. Siegfried. 2006.
Inheritance of resistance to the Cry1Ab Bacillus thuringiensis toxin in
Ostrinia nubilalis (Lepidoptera: Crambidae).  J. Econ. Entomol. 99:
494-501.

Bagwell, R., D. Cook, J. Adamczyk, B. Leonard, and S. Micinski. 2000.
Status of insecticide resistance in budworm and bollworm in Louisiana
during 1999, pp. 914-918. In: Proceedings of the Beltwide Cotton
Conferences, National Cotton Council, Memphis, TN.

Benedict, J. H. 2005. Biology and dispersal of the cotton bollworm and
tobacco budworm in North America, pp. 51-119.  In: G. Head and D.
Gustafson, Production of tobacco budworm from alternative host plants
and the role of these host plants as natural refuge for Bollgard II®
cotton.  EPA-HQ-OPP-2006-0217-0010.

BPPD. 2004.  Technical review of Monsanto's submission:  "Impact of
Effective refuge size and typical insecticide use practices on model
predictions of years to resistance of tobacco budworm and cotton
bollworm to Bollgard cotton."  S. Matten memorandum to L. Cole dated
April 22, 2004.

BPPD. 2006.  EPA review of data and support materials submitted by
Monsanto to amend the insect resistance management requirements for
Bollgard II Bt cotton.  A. Reynolds, and S. Matten memorandum to L.
Cole, May 16, 2006.  EPA-HQ-OPP-2006-0217-0007.

Burd, A. D., F. Gould, J. R. Bradley, J. W. Van Duyn, and W. J. Moar.
2003. Estimated frequency of nonrecessive Bt resistance genes in
bollworm, Helicoverpa zea (Boddie) (Lepidoptera: Noctuidae) in eastern
North Carolina.  J. Econ. Entomol. 96: 137-142.

Caprio, M. A. 1998. Two-compartment model for insect resistance.
http://www.msstate.edu/Entomology/PGjava/ILSImodel.html 

Caprio, M. A. 2006. An independent assessment of the risk of the
evolution of resistance in Helicoverpa zea to Bollgard II cotton with
natural and structures refuges in the MidSouth. 
EPA-HQ-OPP-2006-0217-0018.2.

Carrière, Y., C. Ellers-Kirk, A. L. Patin, M. A. Sims, S. Meyer, Y.-B.
Liu, T. J. Dennehy, and B. E. Tabashnik. 2001. Overwintering cost
associated with resistance to transgenic cotton in the pink bollworm
(Lepidoptera: Gelechiidae).  J. Econ. Entomol. 94: 935-941. 

Dobson, A. J. 2002. An Introduction to Generalized Linear Models, Second
Edition. Chapman & Hall/CRC Press LLC, Boca Raton, FL.

Farrow, R. A., and J. C. Daly. 1987. Long-range movements as an adaptive
strategy in the Genus Heliothis (Lepidoptera: Noctuidae): a review of
its occurrence and detection in four pest species.  Aust. J. Zool. 35:
1-24.

Fitt, G. P. 1989. The ecology of Heliothis  species in relation to
agro-ecosystems.  Annu. Rev. Entomol. 34: 17-52.

Fitt, G. P. 1991. Host selection in Heliothinae. Chapter 7, in
"Reproductive Behaviour in Insects - Individuals and Populations", J.
Ridsdill-Smith and W. Bailey (eds.), Chapman and Hall.

Fitt, G. P., M. P. Zalucki, and P. H. Twine. 1989. Temporal and spatial
patterns in pheromone-trap catches of Helicoverpa spp. (Lepidoptera:
Noctuida) in northern New South Wales, Australia: implications for pest
management.  J. Econ. Entomol. 83: 1827-1836.

Fitt, G. P., M. L. Dillon, and J. G. Hamilton. 1995. Spatial dynamics of
Helicoverpa populations in Australia: simulation modelling and empirical
studies of adult movement.  Computers Electronics Agric. 13: 177-192.

Gahan, L. J., F. Gould, and D. G. Heckel 2001. Identification of a gene
associated with Bt resistance in Heliothis virescens.  Science 293:
857-860.

Gould, F., A. Anderson, A. Jones, D. Sumerford, D. G. Heckel, J. Lopez,
S. Micinski, R. Leonard, and M. Laster. 1997. Initial frequency of
alleles for resistance to Bacillus thuringiensis toxins in field
populations of Heliothis virescens.  Proc. Natl. Acad. Sci. USA 94:
3519-3523.

Gould, F., N. Blair, M. Reid, T. L. Rennie, J. Lopez, and S. Micinski.
2002. Bacillus thuringiensis-toxin resistance management: stable isotope
assessment of alternate host use by Helicoverpa zea. Proc. Natl. Acad.
Sci. 99: 16581-16586.

Gould, F., M. B. Cohen, J. S. Bentur, G. G. Kennedy, J. Van Duyn. 2006.
Impact of small fitness costs on pest adaptation to crop varieties with
multiple toxins: a heuristic model.  J. Econ. Entomol. (in press).

Gunning, R.V., H. T. Dang, F. C. Kemp, I. C. Nicholson, and G. D.
Moores. 2005. New resistance mechanism in Helicoverpa armigera threatens
transgenic crops expressing Bacillus thuringiensis Cry1Ac toxin.  Appl.
Environ. Microbiol. 71: 2558-63.

Gustafson, D. and G. Head. 2004. Impact of updated alternative host and
pyrethroid effectivness data on model predictions of cotton bollworm
resistance development to Bollgard® cotton. Monsanto Company, Report
number MSL-19326. St. Louis, Missiouri.

Gustafson, D. I., and G. P. Head. 2005. Modeling the impact of natural
refuge on the evolution of tobacco budworm and cotton bollworm
resistance to Bollgard II® cotton.  EPA-HQ-OPP-2006-0217-0011.

Gustafson, D. I., G. P. Head, and M. A. Caprio. 2005. Modeling the
impact of alternative hosts and pyrethroid sprays on Helicoverpa zea
(Boddie) adaptation to Bollgard® cotton.  J. Econ. Entomol.
(submitted).  Included in supporting materials from Monsanto to EPA,  
HYPERLINK
"http://www.regulations.gov/fdmspublic/custom/jsp/search/searchresult/do
cumentSearchResult.jsp" \l "#" \o "EPA-HQ-OPP-2006-0217-0013" 
EPA-HQ-OPP-2006-0217-0013 .

Han, Q., and M. A. Caprio. 2002. Temporal and spatial patterns of
allelic frequencies in cotton bollworm (Lepidoptera: Noctuidae). 
Environ. Entomol. 31: 462-468.

Han, Q., and M. A. Caprio. 2004. Evidence from genetic markers suggests
seasonal variation in dispersal in Heliothis virescens (Lepidoptera:
Noctuidae).  Environ. Entomol. 33: 1223-1231.

Hartstack, A. W., J. D. Lopez, R. A. Mueller, W. L. Sterling, E. G.
King, J. A. Witz, and A. C. Eversull. 1982. Evidence of long range
migration of Heliothis zea (Boddie) into Texas and Arkansas. Southwest.
Entomol. 7: 188-201.

Hayes, J. L. 1988. A comparative study of adult emergence phenologies of
Heliothis virescens (F.) and H. zea (Boddie) (Lepidoptera: Noctuidae) on
various hosts in field cages.  Environ. Entomol. 17: 344-349.

Hayes, J. L. 1991. Dynamics of nocturnal activity of moths in the
Heliothis complex (Lepidoptera: Noctuidae) in cotton.  J. Econ. Entomol.
84: 855-865.

Head, G., and D. Gustafson. 2005. Production of tobacco budworm from
alternative host plants and the role of these host plants as natural
refuge for Bollgard II® cotton.  EPA-HQ-OPP-2006-0217-0010.

Head, G., and R. Voth. 2004. A final report on studies to assess
production of Helicoverpa zea from alternate host plants and from the
external unsprayed non-Bt cotton refuge for Bollgard® cotton. Monsanto
Company, Report number MSL-19238, MRID 46222401. St. Louis, Missouri.

Head, G., M. C. McCann, and J. W. Mullins. 2005. Scientific and economic
justification for not requiring structured cotton refuges for Bollgard
II® cotton in the U.S. Cotton Belt from Texas to the East Coast. 
EPA-HQ-OPP-2006-0217-0012.

Helsel, D. R. 2005. Nondetects and Data Analysis, Statistics for
Censored Environmental Data. Wiley and Sons, Inc. New Jersey.

Hendrix, W. H., T. H. Mueller, J. R. Phillips, and O. K. Davis. 1987.
Pollen as an indicator of long-distance movement of Heliothis zea
(Lepidoptera: Noctuidae).  Environ. Entomol. 61: 1148-1151.

Hurley, T. M.  2000.  Research report:  A Bioeconomic evaluation of the
gradual introduction of multiple toxin Bt corn.  Prepared for Monsanto
Company.  February 2000.

Hurley, T.M., S. Secchi, and R. Hellmich. 2006. Managing European Corn
Borer Resistance to Bt Corn with Dynamic Refuges.  In: Agricultural
Biotechnology: Economics and Policy (R.E. Just, J.M. Alston, and D.
Zilberman, eds.), New York, Springer Publishers. (In press)

Ives, A. R., and D. A. Andow. 2002. Evolution of resistance to Bt crops:
directional 

selection in structured environments.  Ecol. Lett. 5: 792-801.

Jackson, R. E., J. R. Bradley, J. Van Duyn, B. R. Leonard, K. C. Allen,
R. Luttrell, J. Ruberson, J. Adamcyzk, J. Gore, D. D. Hardee, R. Voth,
S. Sivasupramaniam, J. W. Mullins, and G. Head. 2006. Regional
assessment of Helicoverpa zea (Lepidoptera: Noctuidae) populations on
cotton and non-cotton crop hosts.  Environ. Entomol. (submitted).
(EPA-HQ-OPP-2006-0217-0013)

Jurat-Fuentes, J. L., and M. J. Adang. 2004. Characterization of a
Cry1Ac-receptor alkaline phosphatase in susceptible and resistant
Heliothis virescens larvae.  Eur. J. Biochem. 271: 3127-35.

Kennedy, G. G., and N. P. Storer. 2000. Life systems of polyphagous
arthropod pests in temporally unstable cropping systems.  Annu. Rev.
Entomol. 45: 467-493.

Korman, A. K., J. Mallet, J. L. Goodenough, J. B. Graves, J. L. Hayes,
D. E. Hendricks, R. Luttrell, S. D. Pair, and M. Wall. 1993. Population
structure in Heliothis virescens (Lepidoptera: Noctuidae): an estimate
of gene flow.  Ann. Entomol. Soc. Am. 86: 182-188.

Lee, Y., J. A. Nelder, and Y. Pawitan. 2006. Generalized Linear Models
with Random Effects, Unified Analysis via H-likelihood.  Monographs on
Statistics and Applied Probability 106, Chapman & Hall/CRC LLC, Boca
Raton, FL.

Leonard, B. R., J. B. Graves, G. E. Burris, A. M. Pavloff, and G.
Church. 1989. Heliothis sp. (Lepidoptera: Noctuidae) captures in
pheromone traps: species composition and relationship to oviposition in
cotton.  J. Econ. Entomol. 82: 574-579.

Leonard, B. R., C. A. White, and J. B. Graves.  1995.  Insecticide
resistance frequencies in overwintering and field-collected tobacco
budworms, pp. 967-971. In 1995 Proceedings Beltwide Cotton Conf.,
National Cotton Council, Memphis, TN.

Li, H., B. Oppert, R. A. Higgins, F. Huang, K. Y. Zhu, and L. L.
Buschman. 2004. Comparative analysis of proteinase activities of
Bacillus thuringiensis-resistant and -susceptible Ostrinia nubilalis
(Lepidoptera: Crambidae).  Insect Biochem. Mol. Biol. 34: 753–762.

Li, H., B. Oppert, R. A. Higgins, F. Huang, L. L. Buschman, J. R. Gao,
and K. Y. Zhu. 2005. Characterization of cDNAs encoding three
trypsin-like proteinases and mRNA quantitative analysis in Bt-resistant
and -susceptible strains of Ostrinia nubilalis.  Insect Biochem. Mol.
Biol. 35: 847-60.

Liao, C., D. G. Heckel, and R. Akhurst 2002. Toxicity of Bt insecticidal
proteins for Helicoverpa armigera and Helicoverpa punctigera
(Lepidoptera:  Noctuidae), major pests of cotton.  J. Invert. Pathol.
80:55-63.

Livingston, M. J., G. A. Carlson, and P. L. Fackler. 2002. Use of
mathematical models to estimate characteristics of pyrethroid resistance
in tobacco budworm and bollworm (Lepidoptera: Noctuidae) field
populations.  J. Econ. Entomol. 95: 1008-1017.

Livingston, M. J., G. A. Carlson, and P. L. Fackler. 2004.  Managing
resistance evolution in two pests to two toxins with refugia.  Amer. J.
Agric. Econ. 86: 1-13.

Livingston, M. J., N. P. Storer, J. W. Van Duyn, and G. G. Kennedy.
2006. A structured refuge requirement may not be needed to promote
economic efficiency in eastern North Carolina.  J. Agric. Appl. Econ.
(in press).

Louis, T. A.  1981.  Confidence intervals for a binomial parameter after
observing no successes.  Amer. Statistician 35: 154.

Lukefahr, M. J., and D. F. Martin. 1964. The effects of various larval
and adult diets on fecundity and longevity of the bollworm, tobacco
budworm, and cotton leafworm.  J. Econ. Entomol.  57: 233-235.

Luttrell, R. G., K. Knighten, W. F. Kitten, G. L. Andrews, F. A. Harris,
and J. Reed. 1991. Monitoring pyrethroid resistance in the tobacco
budworm in Mississippi: implications for resistance management. 
Southwest. Entomol. Supplement 15: 5-26.

McCullagh, P., and J. A. Nelder. 1989. Generalized Linear Models, Second
Edition. Chapman & Hall/CRC Press LLC, Boca Raton, FL.

Milne, R., T. Wright, H. Kaplan, and D. Dean. 1998. Spruce budworm
elastase precipitates Bacillus thuringiensis delta-endotoxin by
specifically recognizing the C-terminal region.  Insect Biochem. Mol.
Biol. 28:1013-1023.

Morin, S., R. W. Biggs, M. S. Sisterson, L. Shriver, C. Ellers-Kirk, D.
Higginson, D. Holley, L. J. Gahan, D. G. Heckel, Y. Carrière, T. J.
Dennehy, J. K. Brown, and B. E. Tabashnik. 2003.  Three cadherin alleles
associated with resistance to Bacillus thuringiensis in pink bollworm. 
Proc. Natl. Acad. Sci. USA 100: 5004-5009.

Mueller, T. F., and J. R. Phillips. 1983. Population dynamics of
Heliothis-spp in spring weed hosts in southeastern Arkansas:
survivorship and stage-specific parasitism.  Environ. Entomol. 12:
1846-1850.

Mueller, T. F., V. E. Harris, and J. R. Phillips, Jr. 1984. Theory of
Heliothis (Lepidoptera: Noctuidae) management through reduction of the
1st spring generation – a critique.  Environ. Entomol. 13: 625-634.

Nadguada, D., and H. Pitre. 1983. Development, fecundity, and longevity
of the tobacco budworm (Lep.: Noctuidae) fed soybean, cotton, and
artificial diet at three temperatures.  Environ. Entomol. 12: 582-586.

Neunzig, H. H. 1969. The biology of the tobacco budworm and the corn
earworm in North Carolina.  North Carolina Agr. Expt. Sta. Tech. Bull.
196.

Norris, R. F., and M. Kogan. 2005. Ecology of interactions between weeds
and arthropods.  Annu. Rev. Entomol. 50: 479-503.

Onstad, D. W., and C. A. Guse. 1999. Economic analysis of transgenic
maize and nontransgenic refuges for managing European corn borer
(Lepidoptera: Pyralidae).  J. Econ. Entomol. 92: 1256-1265.

Oppert, B., K. J. Kramer, R. W. Beeman, D. Johnson, and W. H. McGaughey.
1997. Proteinase-mediated insect resistance to Bacillus thuringiensis
toxins.  J. Biol. Chem. 272: 23473-23476.

Ott, R. L., and M. Longnecker. 2001. An Introduction to Statistical
Methods and Data Analysis, Fifth Edition. Duxbury, Pacific Grove, CA.

Peck, S. L., F. Gould, and S. P. Ellner. 1999. Spread of resistance in
spatially extended regions of transgenic cotton: implications for
management of Heliothis virescens (Lepidoptera: Noctuidae).  J. Econ.
Entomol. 92: 1-16.

Reynolds, A. 2006. National refuge proposal for Bollgard II Bt cotton:
EPA overview and IRM considerations.  EPA-HQ-OPP-2006-0217-0039.

Roush, R. T. 1998. Two-toxin strategies for management of insecticidal
transgenic crops:  can pyramiding succeed where pesticide mixtures have
not?  Phil. Trans. Roy. Soc. 353: 1777-1786.

SAP. 2004. Meeting Minutes of the FIFRA Scientific Advisory Panel
Meeting, June 8-10, 2004.  A Set of Scientific Issues Being Considered
by the U.S. Environmental Protection Agency Regarding:  Product
Characterization, Human Health Risk, Ecological Risk, and Insect
Resistance Management for Bacillus thuringiensis (Bt) Cotton Products. 
United States Environmental Protection Agency, Washington, D. C.  
HYPERLINK "http://www.epa.gov/scipoly/sap/2004/june/final1a.pdf" 
http://www.epa.gov/scipoly/sap/2004/june/final1a.pdf .

SAS Institute, Inc. 2006. Littell, R. C., G. A. Milliken, W. W. Stroup,
R. D. Wolfinger, and O. Schabenberger.  SAS for Mixed Models, Second
Edition. SAS Press, Cary, NC, US.

Schneider, J. C. 1999. Dispersal of a highly vagile insect in a
heterogeneous environment. Ecology 80: 2740-2749.

Schneider, J. C. 2003. Overwintering of Heliothis virescens (F.) and
Helicoverpa zea (Boddie) (Lepidoptera: Noctuidae) in cotton fields of
northeast Mississippi.  J. Econ. Entomol. 96: 1433-1447.

Schneider, J. C., R. T. Roush, W. F. Kitten, and M. L. Laster. 1989.
Movement of Heliothis virescens (Lepidoptera: Noctuidae) in Mississippi
in the spring: Implications for area-wide management.  Environ. Entomol.
18: 438-446.

Sisterson, M. S., and B. E. Tabashnik. 2005. Simulated effects of
transgenic Bt crops on specialist parasitoids of target pests.  Environ.
Entomol. 34: 733-742.

Sisterson, M. S., L. Antilla, Y. Carrière, C. Ellers-Kirk, and B. E.
Tabashnik. 2004. Effects of insect population size on evolution of
resistance to transgenic crops.  J. Econ. Entomol. 97: 1413-1424.

Sisterson, M. S., Y. Carrière, T. J. Dennehy, and B. E. Tabashnik.
2005. Evolution of resistance to transgenic crops: interactions between
insect movement and field distribution.  J. Econ. Entomol. 98:
1751-1762.

Sparks, A. N., R. D. Jackson, and C. L. Allen. 1975. Corn earworms:
capture of adults in light traps on unmanned oil platforms in the Gulf
of Mexico.  J. Econ. Entomol. 68: 431-432.

Sparks, T. C., J. B. Graves, and B. R. Leonard. 1993. Insecticide
resistance and the tobacco

budworm: past, present and future.  Rev. Pestic. Toxicol. 2: 149-183.

Stadelbacher, E. A. 1981. Role of early-season wild and naturalized host
plants in the buildup of the F1 generation of Heliothis zea and H.
virescens in the Delta of Mississippi.  Environ. Entomol. 10: 766-770. 

Storer, N. P., S. L. Peck, F. Gould, J. W. Van Duyn, and G. G. Kennedy.
2003. Spatial processes in the evolution of resistance in Helicoverpa
zea (Lepidoptera: Noctuidae) to Bt transgenic corn and cotton in a mixed
agroecosystem: A biology-rich stochastic simulation model.  J. Econ.
Entomol. 96: 156-172.

Tabashnik, B. E. 1992. Evaluation of synergism among Bacillus
thuringiensis toxins.  Appl. Environ. Microbiol. 58: 3343-3346.

USDA, NASS. 2006. Agricultural Statistics Data Base: State and county
data. (http://www.nass.usda.gov/).

Xu, X. J., L. Y. Yu, and Y. D. Wu. 2005. Disruption of a cadherin gene
associated with resistance to Cry1Ac delta-endotoxin of Bacillus
thuringiensis in Helicoverpa armigera.  Appl. Environ. Microbiol. 71:
948-954.

Zalucki, M. P., and M. J. Furlong. 2005. Forecasting Helicoverpa
populations in Australia: a comparison of regression based models and a
bioclimatic based modelling approach.  Insect Sci.  12: 45-56.

Zar, J. H. 1996. Biostatistical Analysis, Third Edition. Prentice-Hall
Inc., Upper Saddle River, NJ.

Zhao, J.-Z., J. Cao, H. L. Collins, S. L. Bates, R. T. Roush, E. D.
Earle, and A. M. Shelton. 2005. Concurrent use of transgenic plants
expressing a single and two Bacillus thuringiensis genes speeds insect
adaptation to pyramided plants.  Proc. Natl. Acad. Sci. USA 102: 
8426-8430.

APPENDIX 1

Correct Derivation of Equation 7

To understand the mistake that Monsanto made in their derivation of
Equation 7, it is useful to reconstruct Monsanto’s equations from a
slightly different perspective.  Consider a landscape of area (.  This
area is divided into five distinct compartments: Bt cotton ((B1), Bt
corn ((B2), non-Bt cotton ((R1), non-Bt corn ((R2), and non-cotton C3
((R3) crops such that ( = (B1 + (B2 + (R1 + (R2 + (R3.  Let moth
production per unit area for each of these crops prior to Bt selection
be (ij for i = B, R and j = 1, 2, 3.  Total moth production for this
landscape can then be written as m = (B1(B1 + (B2(B2 +(R1(R1 + (R2(R2 +
(R3(R3.  Average moth production per unit area is M = m/(.

Effective refuge (Reff) is defined as the proportion of moths produced
by non-Bt and non-cotton C3 crops (i.e., R1, R2, and R3 crops), 

  

 

 , which is exactly the same as Monsanto’s Equation (5).  

Natural refuge (Rnat) consists of the proportion of moths produced on
non-cotton, non-Bt hosts (i.e., R2 and R3 crops) such that 

  

  

 .  

  Rnat, such that RnatCBW > Rnat.  

Figure A1-1 illustrates the compartments used by Monsanto to calculate
the effective refuge.  When Monsanto calculates natural refuge using
Equation (7), it effectively eliminates the non-Bt cotton compartment
(Figure A1-2).  By doing this Monsanto effectively assumes a reduction
in the area of interest: (’ = (B1 + (B2 + (R2 + (R3, instead of (. 
This assumption can create a bias because Mij is calculated per unit
area based on Aij= (ij/( in Figure A1-1, not (ij/(’ in Figure A1-2. 
How much bias can this assumption produce?  Based on the data reported
in

Figure A1-1.  Crop acreage compartments for Monsanto’s effective
refuge calculation.

Figure A1-2.  Crop acreage compartments for Monsanto’s natural refuge
calculation.

Tables 2 and 5 (Gustafson and Head 2005) assuming LBB1 = LBB2 = 1 prior
to Bt selection, Monsanto’s estimate of the effective refuge (RnatCBW)
will be 37, 12, 6, and 44% higher than the natural refuge estimate given
by Rnat for Georgia, Mississippi, North Carolina, and Texas,
respectively.

For the current effective refuge, Reff, simulation, Monsanto only needs
their correct estimate of the effective refuge, so their result should
not be biased.  This is not true for natural refuge simulations.  For
natural refuge, Rnat, simulations, Monsanto keeps the area of cotton
fixed, while deleting structured refuge for Bollgard II for the
simulations reported in Table 14 and all structured refuge for
simulations reported in Figure 4.  This effectively reduces non-Bt
cotton acreage, while replacing it with Bt cotton acreage.  If ( is the
proportional reduction in non-Bt cotton acreage, effective and natural
refuge become 

  

 

and 

  

 , 

 Rnat such that RnatCBW > Rnat for ( ( 0.  This result suggests the
upward bias in Monsanto’s natural refuge estimate will be either
systematically increasing or decreasing in the market share of Bollgard
II depending on the relative moth productivity of Bt and non-Bt cotton
prior to Bt selection.  The data in Tables 2 and 5 suggest EB1 LBB1 LSB1
> ER1 LBR1 LSR1 for all regions such that increasing Bollgard II market
share, increases the upward bias in Monsanto’s natural refuge
calculation.



APPENDIX 2

Model Uncertainty

The Caprio model (Caprio 1998) used by Monsanto does not explicitly
incorporate population densities.  Modifying the model to include
population densities and density-dependent population growth rates
creates a model that violates some reasonable assumptions about the
biology and ecology of insects like TBW.  This illustrates a difficulty
in assessing models that, even though simple in structure, nonetheless
have properties that are not transparent.  This illustrates the problem
of model uncertainty.

A general “2-patch” model of resistance evolution

To explore this issue, we produced a model (code attached below) for
single-locus resistance based on the following assumptions:

1. A fraction Q of the suitable habitat (in which females lay eggs) is
refuge, and a fraction 1–Q is Bt.

2. Following emergence, males and females disperse randomly, so that the
proportion of males and females in refuge and Bt fields are Q and 1 –
Q.

3. Mating is random, and the sex ratio is 1:1.

4. Bt is high dose.  Therefore, survival from Bt of homozygous resistant
larvae is sRR = 1, while 1 >> sRS ≥ sSS.  Genotype does not affect
survival of larvae in refuge.

5. The net reproductive rate of larvae in refuge (i.e., the survival of
larvae times the number of female offspring produced per female larva
that survives to adulthood) is given by a function fR(xR), where xR is
the density of larvae. 

6. The net reproductive rate of larvae in Bt fields that survive the Bt
toxin is given by a function fBt(xBt), where xBt is the density of
larvae that survive Bt toxin.  fBt(xBt) is a continuous function (but
not necessarily monotonically decreasing).  For notational convenience,
let fBt(xBt) = FBt in the limit as xBt approaches zero.  Thus, FBt is
the net reproductive rate of larvae that survive Bt toxins when the
density of surviving larvae in Bt fields is very low.  FBt is assumed to
be a constant.

7. In the absence of a resistance allele (i.e., for a purely susceptible
population), the mean density of insects approaches some fixed value X. 
Because the model does not include temporal environmental variability,
this condition will be satisfied whenever a purely susceptible
population is persistent (i.e., does not go extinct and does not
increase in density to infinity).

	Under these assumptions, the following mathematical result holds (see
Ives and Andow 2002 for methods).  Let p(t) denote the frequency of the
resistance allele in generation t.  Then the asymptotic rate of increase
of the frequency of the resistance allele in the limit as p(t)
approaches zero is given by

 

Here, “asymptotic rate of increase” means the rate of increase after
transient fluctuations in the relative frequencies of the resistance
allele in Bt and refuge fields have dampened out.  Although this
approximation strictly holds only in the limit as p(t) approaches zero,
numerical studies show it generally to be accurate for p(t) < 0.1.  For
p(t) > 0.1, resistance normally occurs in only a few generations for
most reasonable parameter values, so the breakdown in the approximation
when p(t) > 0.1 does not greatly affect conclusions that can be drawn.

	The interesting consequence of this result is that the asymptotic rate
of resistance evolution does not depend on the function fR(xR), the
survival of larvae in the refuge and the fecundity of the surviving
females.  It also does not depend on the functional form of fBt(xBt),
only on fBt(xBt -> 0) = FBt.  Therefore, the theorem holds for a fairly
broad class of models.

	A key conclusion for models that conform to assumptions 1-7 is the
importance of FBt, the net reproductive rate of larvae that survive Bt
toxins when the density of surviving larvae in Bt fields is very low. 
The higher FBt, the more rapid resistance evolution.  Unfortunately, it
is very difficult to estimate FBt in the field.

Comparison with the Caprio model (1998). 

	The Caprio model (1998) forms the basis of the model Monsanto uses to
assess the durability of Bollgard II.  The Caprio model considers a case
such as Bollgard, in which there is a single resistance locus and only
two types of habitat, Bt crop and refuge.  For simplicity, here we will
address this model, rather than the multi-patch model presented by
Monsanto, although the general comments should apply to both.  Also, we
focus on the high-dose case corresponding to TBW.

	To compare to the Caprio model, we derived a model that explicitly
accounts for densities and conforms to assumptions 1-7 above.  For
population growth in Bt fields, we make the simple assumption that if
larvae survive Bt, they have density-independent population growth; in
other words, fBt(xBt) = FBt for all densities xBt.  For population
growth in the refuge fields, we make the simple assumption that the
population densities of larvae surviving in the refuge have a fixed
value xrefuge.  Thus, density-dependent population growth in the refuge
is strong enough to bring populations up to carrying capacity in the
refuge every generation.  This carrying capacity can be affected by
insecticide spraying in the refuge, however, such that if the
insecticide leaves only a fraction k of insects alive, the carrying
capacity is reduced to k xrefuge.  We selected these assumptions about
fBt and fR only because they are simple; as discussed above, other
assumptions give essentially the same results.  We will refer to this as
the FBt-fixed model, since it assumes a constant value of FBt.

	Figure A2-1A compares times to resistance computed from the Caprio
model (red dashed line) to those predicted by the FBt-fixed model just
described (black line).  In the Caprio model, the time to resistance is
much more sensitive to the amount of refuge Q than in the FBt-fixed
model.

	To understand why these models differ, we constructed a modified
version of the Caprio model that explicitly incorporates population
densities and satisfies assumptions 1-5 and 7 above.  It turns out to be
impossible to incorporate density into the Caprio model in a way that
satisfies all assumptions 1-7.  In the modified Caprio model, the
reproductive rate of larvae in Bt fields that survive Bt is 1/(kQ).  In
other words, to formulate the Caprio model to include densities, the
value of FBt must be inversely proportional to the survival of larvae
from spraying in the refuge, k, and the proportion of refuge fields, Q. 
The fact that this modified Caprio model gives the same predicted rates
of resistance evolution is shown in Figure A2-1A by the correspondence
between the lines for the modified Caprio model (green line) and the
original Caprio model (red line).  The dependence of FBt on kQ in the
modified Caprio model is shown in Figure A2-1B.

 

Fig. A2-1. (A) Generations to resistance (from resistance allele
frequency of 0.002 to 0.5) and (B) corresponding values of FBt for
resistant insects in Bt crops versus the proportion refuge Q.  The red
dashed line is the Caprio model (1998).  The dotted green line is the
model with fixed size of the insect population in the refuge, xrefuge =
Q, and FBt of resistant insects in Bt crops the same as in the refuge
(FBt = 1/Q).  The solid black line has a constant FBt of 50.  Other
parameter values are: sRR = 1, sRS = 0.001, sSS = 0.001, k (survival
from insecticide) = 1, and R (proportion of insects leaving natal
habitat) = 1.

The models make different qualitative predictions about factors
affecting resistance evolution

	The FBt-fixed model makes different predictions from the Caprio model
about the consequences of spraying insecticides in refuges.  For the
FBt-fixed model, the rate of resistance evolution is not sensitive to
insecticide spraying in the refuge (Fig. A2-2).  This seemingly
unintuitive pattern is a direct mathematical consequences of the
theoretical result about the asymptotic rate of resistance evolution
described above, in which the rate of resistance evolution does not
depend on the reproduction rate of insects in refuges.  This contrast
between models illustrates how implicit assumptions in models can change
qualitative conclusions drawn from the models.

 

Fig. A2-2.  (A) Generations to resistance (from resistance allele
frequency of 0.002 to 0.5) and (B) corresponding values of FBt for
resistant insects in Bt crops versus the survival of insects from
insecticide in the refuge, k.  The red dashed line is the Caprio model
(1998).  The dotted green line is the model with fixed size of the
insect population in the refuge, xrefuge = kQ, and FBt of resistant
insects in Bt crops the same as in the refuge (FBt = 1/kQ).  The solid
black line has a constant FBt of 50.  Other parameter values are: Q =
0.1, sRR = 1, sRS = 0.001, sSS = 0.001, and R (proportion of insects
leaving natal habitat) = 1.

	Although the lack of effect of insecticide spraying may seem
unintuitive, it has a simple explanation (Ives and Andow 2002) using the
following reasoning:

(i) The rate of resistance evolution depends on the proportion of the
susceptible population that is killed by Bt.  

(ii) In the high-dose case when the frequency of a resistance allele is
low, essentially the entire population is in the refuge.

(iii) The proportion of the susceptible population killed by Bt equals
the proportion of females that leaves the refuge and oviposits in Bt
fields.

(iv) Reducing the population size in the refuge does not change the
proportion of females leaving the refuge.

(v) Therefore, reducing the size of the population in the refuge through
spraying will not change the rate of resistance evolution.  

While this argument begins to break down mathematically when the
frequency of the resistance allele reaches 0.1, by this time there will
be only a very few more generations before the resistance allele reaches
0.5 and control fails.

Matlab code for the modified Caprio (1998) model

% Caprio.m

 

% written by Anthony R. Ives, 9 June 2006

 

% produces a model similar to Caprio 1998, but with explicit insect

% densities and a two patch environment (removing the third "patch" of

% dispersing insects in Caprio 1998

 

clear

clf

 

% baseline values 

R=1;        % proportion of insects leaving natal habitat before mating

m=1;        % proportion of insects leaving mating habitat after mating

Q=.05;      % proportion of refuge

k=1;        % survival from insecticde in refuge

sRR=1;      % survival of RR genotypes in Bt crop

sSS=.001;

sRS=.001;

L=1;        % survival of all genotypes in refuge (excluding
insecticide)

F1=50;      % fecundity of females in Bt crops

x0=k;       % population size of insects in refuge

 

% iterate over three cases:

% 1 - Caprio assumption of population growth

% 2 - fixed fecundity in Bt crops

% 3 - fecundity in Bt crops = Q*k

for flag=1:3

    output=[];

    for z=.02:.01:.2

        Q=z;

        

        % this allows males and females to have different proportions

        % leaving Bt and refuge fields

        R1m=R;

        R1f=R;

 

        R2m=R;

        R2f=R;

 

        m1m=m;

        m1f=m;

        m2m=m;

        m2f=m;

 

        % survivals

        ORR=[sRR 0;0 k*L];

        ORS=[sRS 0;0 k*L];

        OSS=[sSS 0;0 k*L];

 

        % initial densities X and frequencies P

        X=k*[0;Q];

        P=.002*[1;1];

 

        t=1;

        while (P'*X)/sum(X) < 0.5

 

            t=t+1;

            

            % set population size in refuge equal to k*Q

            x0=k*Q;

            

            % pre-mating movement of alleles in males

            z1s=((1-R1m)+(1-Q)*R1m)*X(1);

            z1d=(1-Q)*R2m*X(2);

            z2s=((1-R2m)+Q*R2m)*X(2);

            z2d=Q*R1m*X(1); 

            Mm=[z1s/(z1s+z1d) z1d/(z1s+z1d);z2d/(z2s+z2d)
z2s/(z2s+z2d)];

 

            % pre-mating movement of alleles in females

            z1s=((1-R1f)+(1-Q)*R1f)*X(1);

            z1d=(1-Q)*R2f*X(2);

            z2s=((1-R2f)+Q*R2f)*X(2);

            z2d=Q*R1f*X(1); 

            Mf=[z1s/(z1s+z1d) z1d/(z1s+z1d);z2d/(z2s+z2d)
z2s/(z2s+z2d)];

 

            % movement of alleles and mating to give genotypic
frequences

            WRR=(Mf*P).*(Mm*P);

            WRS=(1-Mf*P).*(Mm*P)+(Mf*P).*(1-Mm*P);

            WSS=(1-Mf*P).*(1-Mm*P);

 

            % redistribution of densities during premating dispersal

            d11=((1-R1f)+(1-Q)*R1f);

            d12=(1-Q)*R2f;

            d21=Q*R1f;

            d22=((1-R2f)+Q*R2f);

 

            D=[d11 d12;d21 d22];

            X=D*X;

            

            % genotypic densities

            XRR=WRR.*X;

            XRS=WRS.*X;

            XSS=WSS.*X;

 

            % post-mating movement of females

            d11=((1-m1f)+(1-Q)*m1f);

            d12=(1-Q)*m2f;

            d21=Q*m1f;

            d22=((1-m2f)+Q*m2f);

 

            % postmating movement and selction

            D=[d11 d12;d21 d22];

            XRR=ORR*D*XRR;

            XRS=ORS*D*XRS;

            XSS=OSS*D*XSS;

 

            % calculation of new frequencies and densities

            P=(XRR+XRS./2)./(XRR+XRS+XSS);

            X=XRR+XRS+XSS;

            

            % 1 - Caprio assumption of population growth

            if flag==1

                X=X./sum(X);

            end

 

            % 2 - fixed fecundity in Bt crops

            if flag==2

                X(1)=F1*X(1);

                X(2)=x0;

            end

 

            % 3 - fecundity in Bt crops = Q*k

            if flag==3

                F1=1/(Q*k);

                X(1)=F1*X(1);

                X(2)=x0;

            end

        end

        output=[output;z t F1];

    end;

 

    c='rkg';

    figure(1)

    subplot(2,1,1)

    plot(output(:,1),output(:,2),[c(flag),'-'])

    xlabel('Proportion refuge')

    ylabel('Gens To Failure')

    hold on

    

    if flag==2 | flag==3

        subplot(2,1,2)

        plot(output(:,1),output(:,3),[c(flag),'-'])

        xlabel('Proportion refuge')

        ylabel('R0')

        axis([0 .2 0 60])

        hold on

    end

end

hold off

APPENDIX 3

Evolution of Resistance Hotspots

	The hotspot phenomenon can be evaluated only in a model that includes
explicit spatial descriptions of the landscape.  A spatially explicit
2-loci model (described in more detail in Appendix 4) demonstrates that
hotspots and rapid evolution are possible for Bollgard II.  The model
used by Monsanto is spatially implicit, since insects are assumed to
disperse uniformly over the region modeled.  Even though CBW and TBW are
highly mobile, it is unlikely that TBW are well-mixed over the regions
considered during the cotton growing season (see Panel response to
Charge 6).  

The spatially explicit model is designed to illustrate some key factors
that affect the rate of resistance evolution.  It is not designed to
make detailed predictions or incorporate all of the complexity likely to
affect resistance evolution.  Therefore, we did not incorporate many
elements of realism included in the spatially implicit Monsanto model. 
In particular, we consider only Bt and refuge fields, and assume that
the proportion of Bt and refuge fields in the environment does not
change through time.  Monsanto assumes that both CBW and TBW only use
cotton in half of their 6 annual generations.  Rather than having
generations with and without selection, here we simply assume that there
are 3 selected generations per year.  We assume refuge fields are
distributed randomly over space; in reality, refuges will likely be
clustered, leaving relatively larger areas of contiguous Bt fields where
hotspots of resistance are likely to occur.  Finally, we assume that the
resistance alleles at both resistance loci have initial frequencies of
0.002, and that one of the toxins (Cry 2Ab2) is much more effective than
the other (Cry 1Ac), as is the case for CBW.  See Appendix 4 for more
details.

Two different versions of the model were produced corresponding to the
Caprio 1-locus model and the FBt-fixed model described in Appendix 2. 
The only difference between these models is the assumption about
density-dependence.  In the Caprio version of the spatial model, the
reproductive rate of larvae in Bt fields that survive Bt is 1/(kQ).  In
the FBt-fixed model, FBt = 50.

	Figure A3-1 shows the spatial distribution of allele frequencies
produced in a simulation of the Caprio version of the model when there
is 5% refuge and both males and females have a dispersal radius of 6
fields.  The simulation shows a “hotspot” of resistance evolution in
which the allele frequency of both resistance alleles reach a peak. 
This type of hotspot occurs, loosely speaking, because there is a
limited number of insects dispersing from the refuge into a region of
contiguous Bt fields.  Although in this model hotspots are generated by
limited dispersal, hotspots can be generated in other ways.  For
example, hotspots will occur even when refuges are within dispersal
range of Bt fields if the numbers of susceptible insects from the refuge
entering the Bt fields is low.  This possibility is apparent from the
information provided by Monsanto showing high variation in the number of
insects captured in different traps.  Therefore, hotspots can occur even
when the dispersal distance is much greater than 6 fields, as used in
this model.

Figure A3-1: Years to crop failure due to resistance evolution in the
spatially explicit 2-loci Caprio model (Appendix 4) in which the
effective refuge area is 5%, and males and females have a dispersal
radius of 6 fields. There is no cost of resistance.  In this simulation,
resistance failure occurred in 4 years.  Other parameter values are: saa
= 0, sAa = 0.001, saa = 1, sBB = 0.1, sBb = 0.2, and sbb = 1.

	 Table A3-1 gives the number of years required for resistance evolution
for several models and assumptions about dispersal.  The first model is
the spatially explicit version of the model based on Caprio (1998). 
When there is limited dispersal of both males and females (labeled
“limited dispersal”) or global dispersal of males and limited
dispersal of females (labeled “limited female dispersal”), hotspots
occur when the effective refuge size, Reff, is 5% but not 10%.  This
illustrates that the likelihood of hotspots can be very sensitive to the
effective refuge size.  Note that the hotspots occur even when there is
global dispersal of males.  While there is considerable information
showing male dispersal is high, female dispersal is less well understood
(see Charge 1).  In the model, limited female dispersal is sufficient to
cause hotspots even when males show global dispersal. 

The FBt-fixed version of the model in which the maximum reproduction
rate of resistant females in Bt fields is FBt = 50 similarly shows
hotspots when there is 5% refuge.  We considered a further case in which
all male and female insects leave their natal fields if they emerge in
refuge (Rrefuge = 1), but 20% of males and female remain in the Bt
fields if they successfully emerge there (RBt = 0.8).  This corresponds
to the case in which males and females remain in suitable habitat.  This
case similarly shows hotspots.

Table A3-1. Years to crop failure due to resistance evolution in the
spatially explicit 2-loci model (Appendix 4).  Values corresponding to
evolution in “hotspots” are given in bold.  “Limited dispersal”
corresponds to a maximum dispersal distance of 6 fields, and “limited
female dispersal” corresponds to global male dispersal and a maximum
dispersal distance of 6 fields for females.  Parameter values common to
all simulations are: saa = 0, sAa = 0.001, saa = 1, sBB = 0.1, sBb =
0.2, and sbb = 1.







Caprio 2-loci







FBt = 50







FBt = 50

RBt = 0.8







FBt = 50

RBt = 0.8

Rrefuge = 1

† Values are averages of 10 simulations, with the minimum and maximum
years to failure given in brackets.

*1 In this set of simulations, the median number of years to resistance
was 4, and 1 of the 10 was > 334 years.

*2 In this set of simulations, the median number of years to resistance
was 14, and 2 of the 10 were > 334 years.

	As a final case, we included a cost of resistance in which the survival
of larvae homozygous and heterozygous for the resistant allele to the
most effective toxin (Cry 2Ab2) have reduced survivals of 0.8 in Bt and
refuge fields compared to the survivals of 1 for susceptible larvae in
refuge fields.  This represents a strong cost of resistance that
increases Bt crop durability to over 1300 years when insects are
completely mixed (as assumed by Monsanto).  However, when there is
limited dispersal of females only, the median durability of Bt crop is
14 years.  This illustrates that hotspots can overwhelm a cost of
resistance, because within the hotspot (unlike the refuge) there is
little competition between resistant and susceptible insects. 
Therefore, the cost of resistance has little effect to slow resistance
evolution.

	Note that the times to resistance produced by this model are not
inconsistent with the results from a detailed, spatially explicit model
presented by Caprio (2006) for the Panel.  This model assumes that the
effective, unstructured refuge is approximately 10% and there is a cost
of resistance.  Caprio reports that 25% of simulations without a
structured refuge led to resistance in less than 25 years.  In Table
A3-1 the case with a cost of resistance and Reff = 0.10 gives an average
time to resistance of 33 years.  In general, the Caprio model provided
to the Panel (Caprio 2006) seemed to predict shorter times to resistance
than the model presented by Monsanto (Gustafson and Head 2005), although
the reason for this difference is unclear.  The models differ in so many
ways, lack of strong concordance might be expected.

	We use this modeling exercise only to make three points: (1) hotspots
can occur for a two-toxin product (Bollgard II) , as they do for a
one-toxin product (Peck et al. 1999, Sisterson et al. 2004), and lead to
rapid resistance evolution, (2) hotspots can occur even when males are
broadly dispersive when females have limited dispersal, and (3) hotspots
may overwhelm the effects of a cost of resistance.  However, this model
suffers from model uncertainty just like the model used by Monsanto. 
Rather than a completed model that can give quantitative predictions,
instead we view it as only a starting point to develop a greater
understanding of resistance evolution to a two-toxin product. 
Unfortunately, our current level of understanding resistance evolution
for two-toxin products is too limited to make quantitative predictions.

	The Panel realizes that the EPA would like a model that is rigorously
validated and capable of making predictions about the rate of resistance
evolution.  Unfortunately, for the case of Bollgard II, such a model
does not exist.  Therefore, the Panel feels that it is best to
acknowledge our ignorance rather than hide it.  While it might be
tempting to use the complexity of predicting the rate of resistance
evolution as an excuse to rely on simple models and simple analyses,
this runs the risk of ignoring processes that might lead to rapid
resistance.

	A particular limitation of the 2-loci model presented here is that it
is an “infinite population model” in which population densities are
modeled rather than individual insects.  In effect, infinite population
models include fractional individuals, and model the mean frequency of
resistance and the mean density of insects.  For example, if there are
two resistance loci with resistance alleles at frequency 0.001, then the
probability of a doubly homozygous resistant individual (assuming
completely random mating) is 10-12.  Thus, the population size would
have to be roughly 1012 before there was a reasonable chance of finding
a doubly resistant individual.  Preliminary simulations show that
infinite population models likely overestimate the time to resistance
rather than underestimate it relative to models that explicitly
following integer individuals and individual alleles.  Therefore,
hotspots would be more likely in more-realistic simulation models. 
Nonetheless, this is an issue for further research.

APPENDIX 4

A spatially explicit 2-loci model of resistance evolution

This Appendix derives a spatially explicit version of Monsanto’s base
model.  Monsanto does not provide equations necessary to reconstruct
their model, so several specific assumptions about their model had to be
made.

Two versions of the spatially explicit model were derived.  The first is
based on Caprio (1998).  We also used a modified version was also
derived in which both the population size in the refuge, xrefuge, and
the per capita population growth rate of resistant females in Bt fields,
FBt, were fixed.  To simplify the presentation, we assume that there are
only Bt and refuge fields, rather than distinguishing the multiple types
of refuges.  Furthermore, we assume that the proportion of refuge in the
environment does not change through time.  Because only half the
generations of either CBW or TBW are likely to be subjected to selection
for Bt resistance, we assume that 50% of the generations are unselected.
 This is approximated by assuming there are 3 generations per year, all
of which are selected.  We do not present a formal analysis of the
model, and it has not been independently proofed.  Nonetheless, code is
provided at the end of the Appendix.

Below we first describe a spatially implicit model that should be
comparable to Monsanto’s model (although without incorporating
multiple habitat types and variation in habitat types through time). 
This spatially implicit model could be modified and used to verify
Monsanto’s calculations.  Here, it is used to verify the computer code
for the spatially explicit model, since both models give the same output
when the spatially explicit model includes “global” dispersal of
insects over the entire region.

	

1.  Spatially implicit model

	The spatially implicit model assumes that if males and females disperse
from their natal fields, they disperse evenly throughout a region and
settle in Bt and refuge fields in proportion to the area of Bt and
refuge fields.  In Monsanto’s model, all males and females are assumed
to disperse from their natal field, although here we generalize by
including the assumption of Caprio (1998) that a proportion R of insects
disperse while 1–R remain in their natal fields.  After dispersal,
mating is random and females oviposit in the fields in which they mate. 
Because the Monsanto model assumes all insects disperse, post-mating
dispersal makes no difference to the model, so for simplicity we have
not included it.

	Selection occurs on the larvae within fields.  Rather than use the
formula provided by Monsanto to calculate the survivals of the 9
possible genotypes, here for simplicity we assume that the survival of
2-loci genotypes is equal to the product of survivals for each locus
separately.  Thus, if the survival of genotype Aa to one toxin (e.g.,
Cry 1Ac) is sAa and the survival of genotype Bb to a second toxin (e.g.,
Cry 2Ab2) is sBb, then the survival of the genotype AaBb to both toxins
is sAaBb = sAasBb.  In the model, genotypes are state variables rather
than gene frequencies, because strong selection will generate linkage
disequilibrium even when alleles of different loci segregate
independently.

2.  Spatially explicit model

	The spatially explicit model includes the same processes modeled within
Bt and refuge fields as in the spatially implicit model.  These
processes, however, are mapped onto space.  Space consists of a 50 x 50
grid of cells with “wrap-around” boundaries (i.e., on a torus). 
Cells are assigned randomly as either Bt or refuge, with the probability
that a cell is refuge being Q.  When insects disperse from cells, they
are distributed among neighboring cells.  In particular, up to a maximum
distance of n (selected as a parameter in the model), the dispersing
population is distributed by a geometric distribution in all directions
from their natal field.

	Note that the way in which refuges are dispersed in the spatially
explicit model is conservative, in that refuges are spread uniformly
throughout space.  If refuges were clustered, there would be larger
areas of Bt fields without nearby refuges, and hotspots would be more
likely.

	To include a cost of resistance, we assume that survivals of resistant
genotypes are reduced in both Bt and refuge fields.  To compute the cost
for different genotypes, we assume survivals are multiplicative in the
same manner as survivals from Bt.

Matlab code for the spatially implicit model

% Caprio2loci.m

% Tony Ives, 16 June 2006

 

% This is for EPA use only and is copyrighted by Tony Ives

 

clear

 

%size of grid

N=50;

 

% flag2=1 for fixed R0, flag2=2 for Caprio model

flag2=2;

 

% reproductive rate in Bt fields (under flag2=1)

F1=50;

 

%dispersal of males (rm) and females (rf) from Bt (1) and refuge (2)
fields

R1m=1;

R1f=1;

 

R2m=1;

R2f=1;

 

% proportion of refuge

Q=.05;

x0=Q;

 

% spraying survival in refuge

k=1;

 

% initial gene frequencies

p1=.002;

p2=.002;

 

s1=[0 .001 1];

s2=[.1 .2 1];

Z=s1'*s2;

Z=Z(:);

    

sAABB=Z(1);

sAaBB=Z(2);

saaBB=Z(3);

sAABb=Z(4);

sAaBb=Z(5);

saaBb=Z(6);

sAAbb=Z(7);

sAabb=Z(8);

saabb=Z(9);

 

s=[sAABB sAaBB saaBB sAABb sAaBb saaBb sAAbb sAabb saabb]; 

 

% calculate V for mating of genotypes

VV(:,:,1)=[1 .5 0;.5 .25 0;0 0 0];

VV(:,:,2)=[0 .5 1;.5 .5 .5;1 .5 0];

VV(:,:,3)=[0 0 0; 0 .25 .5;0 .5 1];

for i2=1:3

    for i1=1:3

        V(:,:,3*(i2-1)+i1)=kron(VV(:,:,i2),VV(:,:,i1));

    end

end

 

% add cost of resistance

s1=[1 .8 .8];

s2=[1 1 1];

 

Z=s1'*s2;

k=Z(:);

 

O1=diag(k.*s);

O2=diag(k);

 

% initially assume genotypes at H-W and unlinked

Wi=[(1-p1)^2 2*p1*(1-p1) p1^2];

Wj=[(1-p2)^2 2*p2*(1-p2) p2^2];

    

W=Wi'*Wj;

W=W(:);

 

W1=W;

W2=W;

 

X=[10^-4;x0];

 

P1=p1;

P2=p2;

 

t=0;

output=[0 P1 P2 0];

while (P1 < 0.5 | P2 < 0.5) && X(1) < x0/2 && t < 2000

 

    t=t+1;

    % pre-mating movement of males

    z1s=((1-R1m)+(1-Q)*R1m)*X(1);

    z1d=(1-Q)*R2m*X(2);

    z2s=((1-R2m)+Q*R2m)*X(2);

    z2d=Q*R1m*X(1); 

    Mm=[z1s/(z1s+z1d) z1d/(z1s+z1d);z2d/(z2s+z2d) z2s/(z2s+z2d)];

    Mm=kron(Mm,eye(9));

 

    % pre-mating movement of females

    z1s=((1-R1f)+(1-Q)*R1f)*X(1);

    z1d=(1-Q)*R2f*X(2);

    z2s=((1-R2f)+Q*R2f)*X(2);

    z2d=Q*R1f*X(1); 

    Mf=[z1s/(z1s+z1d) z1d/(z1s+z1d);z2d/(z2s+z2d) z2s/(z2s+z2d)];

    Mf=kron(Mf,eye(9));

 

    Wm=Mm*[W1;W2];

    Wf=Mf*[W1;W2];

 

    % mating

    WW=[];

    for i=1:2

        Wmi=Wm(9*(i-1)+1:9*i);

        Wfi=Wf(9*(i-1)+1:9*i);

        for j=1:9

            WW=[WW;sum(sum((Wfi*Wmi').*V(:,:,j)))];

        end

    end

    W1=WW(1:9);

    W2=WW(10:18);

 

    % redistribution of densities during premating dispersal

    d11=((1-R1f)+(1-Q)*R1f);

    d12=(1-Q)*R2f;

    d21=Q*R1f;

    d22=((1-R2f)+Q*R2f);

 

    D=[d11 d12;d21 d22];

    X=D*X;

 

    % selection

    Wn1=O1*W1;

    Wn2=O2*W2;

 

    W1=Wn1./sum(Wn1);

    W2=Wn2./sum(Wn2);

 

    % density-dependent population growth

    if flag2==1

        BtXincrease=F1*sum(Wn1);

    else

        BtXincrease=(sum(Wn1)/sum(W1))*x0/(X(2)*sum(Wn2)/sum(W2));

    end

        

    X(1)=BtXincrease*X(1);

    X(2)=x0;

 

    % compute frequencies

    P1=(X(1)*(sum(W1([3 6 9]))+.5*sum(W1([2 5 8]))) + ...

        X(2)*(sum(W2([3 6 9]))+.5*sum(W2([2 5 8]))))/(X(1)+X(2));

    P2=(X(1)*(sum(W1([7 8 9]))+.5*sum(W1([4 5 6]))) + ...

        X(2)*(sum(W2([7 8 9]))+.5*sum(W2([4 5 6]))))/(X(1)+X(2));

 

    output=[output;t P1 P2];

end

GensToFailure=t

YearsToFailure=ceil(t/3)

 

figure(1)

semilogy(output(:,1),output(:,2),'k',output(:,1),output(:,3),'b')

xlabel('Time (generations)')

ylabel('Allele frequencies')

hold on

Matlab code for the spatially explicit model

% Caprio2lociSpatial.m

% Tony Ives, 16 June 2006

 

% Modified from a single-locus model written by Nic Lehmann-Ziebarth and

% Tony Ives

 

% This is for EPA use only and is copyrighted by Tony Ives

 

clear

 

%size of grid

N=50;

 

% flag2=1 for fixed R0, flag2=2 for Caprio model

flag2=2;

 

% reproductive rate in Bt fields (under flag2=1)

F1=50;

 

%dispersal of males (rm) and females (rf) from Bt and refuge fields

rm1=1;

rm2=1;

rf1=1;

rf2=1;

rm=[rm1 rm2];

rf=[rf1 rf2];

 

% proportion of refuge

Q=.05;

x0=Q;

 

% spraying survival in refuge

k=1;

 

% male dispersal: (1) horizontal/vertical; (2) diamond; (3) global
dispersal

disperseflagm=2;

 

% female dispersal: (1) horizontal/vertical; (2) diamond; (3) global
dispersal

disperseflagf=2;

 

%nearest n neighbor movement

n=6;

 

% initial gene frequencies

p1=.002;

p2=.002;

 

s1=[0 .001 1];

s2=[.1 .2 1];

%s2=[1 1 1];

 

Z=s1'*s2;

Z=Z(:);

 

sAABB=Z(1);

sAaBB=Z(2);

saaBB=Z(3);

sAABb=Z(4);

sAaBb=Z(5);

saaBb=Z(6);

sAAbb=Z(7);

sAabb=Z(8);

saabb=Z(9);

 

s=[sAABB sAaBB saaBB sAABb sAaBb saaBb sAAbb sAabb saabb]'; 

 

% calculate V for mating of genotypes

VV(:,:,1)=[1 .5 0;.5 .25 0;0 0 0];

VV(:,:,2)=[0 .5 1;.5 .5 .5;1 .5 0];

VV(:,:,3)=[0 0 0; 0 .25 .5;0 .5 1];

for i2=1:3

    for i1=1:3

        V(:,:,3*(i2-1)+i1)=kron(VV(:,:,i2),VV(:,:,i1));

    end

end

 

% add cost of resistance

%s1=[1 .8 .8];

s1=[1 1 1];

s2=[1 1 1];

 

Z=s1'*s2;

C=Z(:);

 

% set up survivals

O1=k.*C.*s;

O2=k.*C;

 

O=[O1 O2];

AABB=O(1,:);

AaBB=O(2,:);

aaBB=O(3,:);

AABb=O(4,:);

AaBb=O(5,:);

aaBb=O(6,:);

AAbb=O(7,:);

Aabb=O(8,:);

aabb=O(9,:);

 

% set up dispersal matrix Mm

if disperseflagm==1,

    %xx is movement rate to 4 nearest squares.

    %Distribution over squares is pseudo geometric (scale by 1/2)

    M=zeros(N,N);

    xx=(8*(1-.5^n))^(-1);

    for w=1:n

        xw=xx*.5^(w-1);

        M=M+xw*(diag(ones(N-w,1),w)+diag(ones(w,1), N-w));

    end

    Mm=M+M';

end

% set up dispersal matrix Mf

if disperseflagf==1,

    %xx is movement rate to 4 nearest squares.

    %Distribution over squares is pseudo geometric (scale by 1/2)

    M=zeros(N,N);

    xx=(8*(1-.5^n))^(-1);

    for w=1:n

        xw=xx*.5^(w-1);

        M=M+xw*(diag(ones(N-w,1),w)+diag(ones(w,1), N-w));

    end

    Mf=M+M';

end

 

if disperseflagm==2,

    switch n,

        case 2

        load 'M50_n2'

    case 3,

        load 'M50_n3'

    case 4,

        load 'M50_n4'

    case 5

        load 'M50_n5'

    case 6

        load 'M50_n6'

    end

    Mm=M;

end

 

if disperseflagf==2,

    switch n,

        case 2

        load 'M50_n2'

    case 3,

        load 'M50_n3'

    case 4,

        load 'M50_n4'

    case 5

        load 'M50_n5'

    case 6

        load 'M50_n6'

    end

    Mf=M;

end

 

%This creates random distribution of Bt sites over the field

Btgrid=1+(rand(N,N)<Q);

 

O=zeros(N,N,9);

O(:,:,1)=AABB(Btgrid);

O(:,:,2)=AaBB(Btgrid);

O(:,:,3)=aaBB(Btgrid);

O(:,:,4)=AABb(Btgrid);

O(:,:,5)=AaBb(Btgrid);

O(:,:,6)=aaBb(Btgrid);

O(:,:,7)=AAbb(Btgrid);

O(:,:,8)=Aabb(Btgrid);

O(:,:,9)=aabb(Btgrid);

 

%grid of male and female movement proportions

RM=rm(Btgrid);

RF=rf(Btgrid);

 

% grid of initial densities

X=10^-4*ones(N,N);

X(Btgrid==2)=x0;

 

% initially assume genotypes at H-W and unlinked

Wi=[(1-p1)^2 2*p1*(1-p1) p1^2];

Wj=[(1-p2)^2 2*p2*(1-p2) p2^2];

    

W=Wi'*Wj;

W=W(:);

 

for i1=1:N

    for i2=1:N

        WW(i1,i2,:)=W;

    end

end

W=WW;

 

t=0;

Tmax=10^3;

Plist=[];

while t < Tmax && (p1 < .5 | p2 < .5)

    t=t+1;

    

    % dispersal of male alleles

    for i=1:9

        Xs=(1-RM).*W(:,:,i).*X;

        Xd=RM.*W(:,:,i).*X;

    

        if disperseflagm==1,

            %disperse males and alleles to 4 nearest cells

            Xd=(Mm*Xd+Xd*Mm);

        elseif disperseflagm==2,

            %This is diamond dispersal

            Xd=reshape(Mm*Xd(:),N,N);

        else

            %global dispersal

            Xd=mean(mean(Xd))*ones(N,N);

        end

        Xm(:,:,i)=Xs+Xd;

    end

    Wm=Xm./repmat(sum(Xm,3),[1 1 9]);

 

    % dispersal of female alleles and females

    for i=1:9

        Xs=(1-RF).*W(:,:,i).*X;

        Xd=RF.*W(:,:,i).*X;

    

        if disperseflagf==1,

            %disperse males and alleles to 4 nearest cells

            Xd=(Mf*Xd+Xd*Mf);

        elseif disperseflagf==2,

            %This is diamond dispersal

            Xd=reshape(Mf*Xd(:),N,N);

        else

            %global dispersal

            Xd=mean(mean(Xd))*ones(N,N);

        end

        Xf(:,:,i)=Xs+Xd;

    end

    Wf=Xf./repmat(sum(Xf,3),[1 1 9]);

    Xf=sum(Xf,3);

    

    %mating

    for i1=1:N

        for i2=1:N

            for j=1:9

                WWm=reshape(Wm(i1,i2,1:9),9,1);

                WWf=reshape(Wf(i1,i2,1:9),9,1);

                W(i1,i2,j)=sum(sum((WWf*WWm').*V(:,:,j)));

            end

        end

    end

    

    % pre-selection allele frequencies

    Wo=W;

    

    %selection

    for i=1:9

        Wn(:,:,i)=W(:,:,i).*O(:,:,i);

    end

    surv=sum(Wn,3)./sum(W,3);

    W=Wn./repmat(sum(Wn,3),[1 1 9]);

    

    % compute new densities

    X=surv.*Xf;

    if flag2==1

        X(Btgrid==1)=F1*X(Btgrid==1);

    else

        X(Btgrid==1)=(1/(k*Q))*X(Btgrid==1);

    end

 

    X(Btgrid==2)=x0;

 

    % compute gene frequencies

    p1=0;

    p2=0;

    for i1=1:N

        for i2=1:N

            p1=p1+X(i1,i2)*(sum(W(i1,i2,[3 6 9]))+.5*sum(W(i1,i2,[2 5
8])));

            p2=p2+X(i1,i2)*(sum(W(i1,i2,[7 8 9]))+.5*sum(W(i1,i2,[4 5
6])));

 

            P1(i1,i2)=sum(Wo(i1,i2,[3 6 9]))+.5*sum(Wo(i1,i2,[2 5 8]));

            P2(i1,i2)=sum(Wo(i1,i2,[7 8 9]))+.5*sum(Wo(i1,i2,[4 5 6]));

        end

    end

    p1=p1/sum(X(:));

    p2=p2/sum(X(:));

    

    Plist=[Plist; t p1 p2 mean(mean(X))];

    

end

 

% printed output

GensToFailure=t

YearsToFailure=ceil(t/3)

meanX=mean(X(:))

 

if 1

    % graph of gene frequencies through time

    figure(1)

    semilogy(Plist(:,[2 3]))

    xlabel('time')

    ylabel('allele frequency')

 

    % graph 3D picture of gene frequencies

    figure(100)

    subplot(2,1,1)

    surf(P1)

    view(2)

    axis([1 N 1 N 0 1])

 

    subplot(2,1,2)

    surf(P2)

    view(2)

    axis([1 N 1 N 0 1])

end

Appendix 5

Post meeting comments from one Panel member regarding:

 County-level Variation in Non-cotton Cultivated Hosts of the Tobacco
Budworm

	In response to the charge to assess the adequacy of natural refuges by
region (Charge 9), one Panel member provided the following additional
analyses and comments concerning the adequacy of Monsanto's spatial
sampling scheme for extrapolating to areas that were not sampled.  Such
comments were provided by the Panel member after the meeting and were
not considered for review by the Panel, thus they do not reflect a
consensus Panel position.

Monsanto's assertion that non-cotton hosts constitute an adequate
natural refuge to delay counteradaptation by TBW to Bollgard II cotton
in the East region (i.e., Monsanto's "North Carolina" and "Georgia"
regions; see Fig. 9-1 under Charge 9, or Fig. 2 in Gustafson and Head
2005) depends in large part, at least for the states north of Georgia,
on the data Monsanto provided on host use by TBW from five counties in
North Carolina (three counties for two years and two counties for one
year).  Cropping data from the United States Department of Agriculture,
National Agricultural Statistics Service (USDA, NASS 2006) show that
these sampled counties are not as representative of the entire South
Carolina, North Carolina, and Virginia area as one could desire (Fig.
A5-1).  In South Carolina and North Carolina, there exist multi-county
areas containing tens of thousands to a hundred thousand acres of cotton
in which (tobacco + peanut) acreage constitutes < 5% of the (cotton +
tobacco + peanut) acreage: P(T+P) < 0.05.  In the counties sampled,
P(T+P) ( 0.2 except for P(T+P) = 0.10 for Halifax County in which data
were collected for a single year.  In Virginia, cotton acreage totaled
93,000 acres in 2005, and there were no counties with large cotton
acreages for which P(T+P) < 0.05.  However, there is a narrow-necked
peninsula of two counties on the east side of the Chesapeake Bay with
3,000 acres of cotton and no tobacco or peanut acreage.  A peninsula in
North Carolina, more broadly attached to the mainland than the one in
Virginia, contained 23,000 acres of cotton and no tobacco or peanut
acreages in 2005.  The restriction of gene flow between a peninsular
population and its mainland population could significantly accelerate
the rate of development of counteradaptation.  Thus, there are
county-scale or larger areas in all three of these states that may not
have sufficient non-cotton, cultivated hosts to warrant extrapolation
from TBW host use data acquired in the counties sampled in North
Carolina.  Consequently, North Carolina should receive additional
sampling, and both South Carolina and Virginia should be sampled
directly.

In Georgia in 2005, 1,210,000 acres of cotton were grown primarily in
the southeastern two thirds of the state along with 750,000 acres of
peanut and 16,000 acres of tobacco.  There were no counties with
significant cotton acreage for which P(T+P) < 0.05.  In Florida in 2005,
cotton (102,000 acre), peanut (76,000 acre), and tobacco (4,000 acre)
were grown in the panhandle area contiguous with agriculturally similar
areas in Georgia and Alabama (see below); and P(T+P) >> 0.05 for all
counties.  Thus, significant refuges of non-cotton, cultivated hosts
appear to be present in cotton production areas of both Georgia and
Florida.  However, Florida should be sampled directly to test the
expected outcome.

	In Alabama in 2005, no tobacco was grown; and peanut (225,000 acre) was
grown in 13 counties in the extreme southern and southwestern parts of
the state only.  A total of 193,000 acres of cotton was produced in this
area, but P(T+P) >> 0.05 for all counties.  In the rest of Alabama, but
primarily in the northern half of the state, 357,000 acres of cotton
were grown.  In addition, a total of 25,300 acres of cotton was grown in
three counties on the southern border of Tennessee contiguous with the
cotton production area in Alabama.  Only 425 acres of tobacco was
produced in one of these Tennessee counties.  Thus, southern Alabama
appears to have a significant refuge of non-cotton, cultivated hosts,
but central and northern Alabama is apparently much more similar to
Mississippi.  However, both of these areas in Alabama should be sampled
directly.

 

Fig. A5-1.  County-level variation in proportion cultivated hosts of
tobacco budworm (TBW)  that are non-cotton with respect to locations of
Monsanto's sampling to quantify non-cotton vs. cotton  host use by TBW
in North Carolina and South Carolina in 2005.  Proportion of acreage
planted to cotton, tobacco, and peanut that was planted to tobacco or
peanut (P(T+P)) is written above the county name and the cotton acreage
harvested is written below the county name (USDA, NASS 2006).  P(T+P) <
0.05 in the counties shaded red.  Monsanto’s sampling for TBW males
for host plant analysis was performed in the counties shaded blue.

 Following this FIFRA SAP meeting, a Panel member provided additional
analysis and comments regarding the validity of extrapolating data from
areas sampled by Monsanto in North Carolina and Georgia to areas that
were not sampled.  Such comments were not considered or reviewed by the
Panel during the meeting, and are being provided as an appendix to these
meeting minutes (Appendix 5).

 Following this FIFRA SAP meeting, a Panel member provided additional
analysis and comments regarding the validity of extrapolating data from
areas sampled by Monsanto in North Carolina and Georgia to areas that
were not sampled.  Such comments were not considered or reviewed by the
Panel during the meeting, and are being provided as an appendix to these
meeting minutes (Appendix 5).

 Following this FIFRA SAP meeting, a Panel member provided additional
analysis and comments regarding the validity of extrapolating data from
areas sampled by Monsanto in North Carolina and Georgia to areas that
were not sampled.  Such comments were not considered or reviewed by the
Panel during the meeting, and are being provided as an appendix to these
meeting minutes (Appendix 5).

.

 Following this FIFRA SAP meeting, a Panel member provided additional
analysis and comments regarding the validity of extrapolating data from
areas sampled by Monsanto in North Carolina and Georgia to areas that
were not sampled.  Such comments were not considered or reviewed by the
Panel during the meeting, and are being provided as an appendix to these
meeting minutes (Appendix 5).

 PAGE   

 PAGE   2 

 PAGE   

 PAGE   

 PAGE   3 

 PAGE   

 PAGE   

 PAGE   98 

l

VELVETLEAF  

COTTON 

Jul  

2000  

1500  

1000  

500  

0  

COTTON  

COTTON

2003  

2002  

Aug  

2000  

1500  

1000  

500  

0  

Oct  

2000  

1500  

1000  

50‰ഠ」ഠ഍佃呔乏ഠ愍഍佃呔乏ഠ嘍䱅䕖䱔䅅⁆ഠ䌍
呏佔⁎഍〲㔰ഠ㈍〰‴഍〲㌰ഠ㈍〰′഍〸〰‰ഠ㐍〰
〰†഍ര㠍〰〰†഍〴〰‰ഠ」഍〸〰‰ഠ㐍〰〰†഍
ര䰍牡慶⽥慨഍䕖噌呅䕌䙁†഍䕖噌呅䕌䙁†഍〸〰‰
ഠ㐍〰〰†഍ര㈍〰‴ഠ嘍䱅䕖䱔䅅⁆ഠ䴍瑯獨栯ൡ嘍
䱅䕖䱔䅅⁆ഠ

‹



 

,

C

S

g

v

‰

˜



¹

Î

ß

ò

ü

-

T

v

™

º

à

ý

h]6

ᄉ脈栖㙝

ᄉ脈栖㙝

h]6

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄉ脈栖㙝

ᄏ脈栕礊u栖㙝

ᄉ脈栖㙝

hcb

 hcb

hcb

@

hcb

hcb

hcb

hcb

hcb

h

h

 h

h

h

hqU

hqU

h	n

h	n

h	n

G

G

G

G

G

/	’

/	’

/	’

	㜀$␸䠀$摧灲%

ᄀ킄㜂$␸䠀$葠ː摧怡®

摧ᕏÓ

4

:

F

L

P

w

z

|

}



¬

/¬



°

萏ɶ萑ﻲ葞ɶ葠ﻲ摧䮣å

฀슄༁ᲄ㜂$␸䠀$葝ǂ葞Ȝ摧拲ú̀Ĥ␷㠀$⑈愀Ĥ摧㹝Q

T

&

&

hqU

hqU

j

h¸

h	n

  h¸

 h¸

 h¸

 h¸

 

*

+

=

A

|

~



摧

h

 h

 h

h

 h

 h

萑ː葠ː摧嫸øᤀ

kd

਀&䘋

਀&䘋

਀&䘋

hH

hH

hqU

B

萏Ũ葞Ũ摧㧼	㜀$␸䠀$摧埂Ð	㜀$␸䠀$摧ᲛË

hH

‘

’

hH

h	n

hH

#

5

?

u

•

½

¾

Í

Ó

Ô

è

é

ü

h	n

h	n

4

C

`

u

{

‚

’

“

ž

Ÿ

Ì

Í

á

é

ê

õ

ù

ú

D’

 hH

hH

hH

jƒ

␱㜀$␸䠀$摧ᆋ²ᰀ

h©

hqU

)

0

ž

Ÿ

 

¤

¥

hqU

ᤀVELVETLEAF  

COTTON  

v

i

v

r

u

s

e

d

i

c

i

t

c

e

s

n

I

B

0

6

0

5

0

4

0

3

0

2

0

1

0

1

8

.

0

6

.

0

4

.

0

2

.

0

0

A

l

a

v

i

v

r

u

s

e

d

i

c

i

t

c

e

s

n

I

0

8

0

6

0

4

0

2

0

1

8

.

0

6

.

0

4

.

0

1

2

.

0

0

B

0

6

0

5

0

4

0

3

0

2

0.8

0.6

0.4

0.2

0

August

July

June

MidSouth (combined)

MidSouth Flats

MidSouth Hills

East

       Proportion Natural Refuge (SEM)

Oviposition Month

0

1

0

2

.

0

5

1

.

0

1

.

0

5

0

.

0

0

e

g

u

f

e

r

n

o

i

t

r

o

p

o

r

P

0

5

1

0

0

1

0

5

0

2

.

0

5

1

.

0

1

.

0

5

0

.

0

0

e

g

u

f

e

r

n

o

i

t

r

o

p

o

r

P

A

   Proportion Natural Refuge 

August 

July 

June 

5

1

0.8

0.6

0.4

0.2

0

10

1

2

20

MidSouth  

East  

August 

July 

June 

Minimum Sample Size  

Map prepared by William Sansing   HYPERLINK
"mailto:william.sansing@ssrc.msstate.edu" 
william.sansing@ssrc.msstate.edu  Mississippi State University

Aug. 17, 2006

 

TX region = E.TX, OK,

W.LA, KS, W.AR

MS region = MS, LA,

AR, W.TN, KY, MO

GA region = GA,

mid-TN, FL, AL

NC region = NC,

SC, VA

Moths/ha

Oviposition Month  

GA: Decatur  

NC: Lenoir  

TX: Austin  

AR: Mississippi  

LA: Bossier  

MS: Yazoo  

June  

July  

August  

0.001  

0.01  

0.1  

1  

0.05  

 

 

Proportion Natural Refuge

 

0.001  

0.01  

0.1  

1  

0.05  

GA: Decatur  

NC: Lenoir  

TX: Austin  

AR: Mississippi  

LA: Bossier  

MS: Yazoo  

Oviposition Month  

June  

July  

August  

 

 

Proportion Natural Refuge

 

0.001  

0.01  

0.1  

1  

0.05  

GA: Decatur  

NC: Lenoir  

TX: Austin  

AR: Mississippi  

LA: Bossier  

MS: Yazoo  

Oviposition Month  

June  

July  

August  

