• TABLE OF CONTENTS
HIDE
 Front Cover
 Title Page
 Table of Contents
 Acknowledgement
 Introduction
 Adaptability analysis: An...
 Single-factor trials
 Factorial trials: Interactions...
 Technology packages and compon...
 Systems trials
 Design of on-farm research-extension...
 References
 Index














Title: Adaptability analysis
CITATION THUMBNAILS PAGE IMAGE ZOOMABLE
Full Citation
STANDARD VIEW MARC VIEW
Permanent Link: http://ufdc.ufl.edu/UF00081797/00001
 Material Information
Title: Adaptability analysis
Physical Description: Book
Creator: Hildebrand, Peter E.
Publisher: Florida Book Store,
Publication Date: 1996
Copyright Date: 1996
 Record Information
Bibliographic ID: UF00081797
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
Resource Identifier: oclc - 191868015

Table of Contents
    Front Cover
        Page i
        Page ii
    Title Page
        Page iii
        Page iv
    Table of Contents
        Page v
        Page vi
        Page vii
    Acknowledgement
        Page viii
        Page ix
        Page x
    Introduction
        Page 3
        Page 4
        Page 5
        Page 6
        Page 7
        Page 8
        Page 9
        Page 10
        Page 11
        Page 12
        Page 13
        Page 14
        Page 15
        Page 16
        Page 17
        Page 18
        Page 19
        Page 20
        Page 21
        Page 22
        Page 23
        Page 24
        Page 25
        Page 26
        Page 27
        Page 28
    Adaptability analysis: An overview
        Page 29
        Page 30
        Page 31
        Page 32
        Page 33
        Page 34
        Page 35
        Page 36
        Page 37
        Page 38
        Page 39
        Page 40
        Page 41
        Page 42
        Page 43
        Page 44
        Page 45
        Page 46
        Page 47
        Page 48
        Page 49
        Page 50
        Page 51
        Page 52
        Page 53
        Page 54
        Page 55
        Page 56
        Page 57
        Page 58
        Page 59
        Page 60
        Page 61
        Page 62
    Single-factor trials
        Page 63
        Page 64
        Page 65
        Page 66
        Page 67
        Page 68
        Page 69
        Page 70
        Page 71
        Page 72
        Page 73
        Page 74
        Page 75
        Page 76
        Page 77
        Page 78
        Page 79
        Page 80
        Page 81
        Page 82
        Page 83
        Page 84
        Page 85
        Page 86
        Page 87
        Page 88
        Page 89
        Page 90
        Page 91
        Page 92
        Page 93
        Page 94
    Factorial trials: Interactions among factors
        Page 95
        Page 96
        Page 97
        Page 98
        Page 99
        Page 100
        Page 101
        Page 102
        Page 103
        Page 104
        Page 105
        Page 106
        Page 107
        Page 108
        Page 109
        Page 110
        Page 111
        Page 112
        Page 113
        Page 114
        Page 115
        Page 116
        Page 117
        Page 118
    Technology packages and components
        Page 119
        Page 120
        Page 121
        Page 122
        Page 123
        Page 124
        Page 125
        Page 126
        Page 127
        Page 128
        Page 129
        Page 130
        Page 131
        Page 132
        Page 133
        Page 134
        Page 135
        Page 136
        Page 137
        Page 138
        Page 139
        Page 140
        Page 141
        Page 142
        Page 143
        Page 144
        Page 145
        Page 146
        Page 147
        Page 148
        Page 149
        Page 150
        Page 151
        Page 152
        Page 153
        Page 154
    Systems trials
        Page 155
        Page 156
        Page 157
        Page 158
        Page 159
        Page 160
        Page 161
        Page 162
    Design of on-farm research-extension trials
        Page 163
        Page 164
        Page 165
        Page 166
        Page 167
        Page 168
        Page 169
        Page 170
        Page 171
        Page 172
        Page 173
        Page 174
        Page 175
        Page 176
        Page 177
        Page 178
        Page 179
        Page 180
    References
        Page 181
        Page 182
        Page 183
        Page 184
        Page 185
        Page 186
    Index
        Page 187
        Page 188
        Page 189
        Page 190
Full Text




131


ADAPTABILITY ANALYSIS


A METHOD FOR THE DESIGN,
ANALYSIS AND INTERPRETATION OF
ON-FARM RESEARCH-EXTENSION





Peter E. Hildebrand and John T. Russell



To be published by:

Iowa State University Press 1
Ames, Iowa
Spring, 1996


1 Copyright 0 1996 Iowa State University Press


=THE
FLORIDA BOOK STORE
Custom Publishing Service
I $~4I


__i_~_i















ADAPTABILITY ANALYSIS


A METHOD FOR THE DESIGN,
ANALYSIS AND INTERPRETATION OF
ON-FARM RESEARCH-EXTENSION





Peter E. Hildebrand and John T. Russell



To be published by:

Iowa State University Press 1
Ames, Iowa
Spring, 1996


1 Copyright 1996 Iowa State University Press






















Contents



Acknowledgements viii

1. Introduction 3
Science and Agricultural Technology Development 3
Broadly Adaptable vs. Location-Specific Technologies 5
Origin of Adaptability Analysis 7
Purpose. of On-Farm Research-Extension 8
On-Farm Research-Extension and Farming Systems
Research-Extension 8
Recognizing Farm and Environmental Diversity 10
Environments Defined 12
Incorporating Farmers' Perspectives and Evaluations 14
Enhancing Adoption Through Adaptation and Learning 15
Location Specificity and Sustainable Agriculture 16
Problems with Combined Analysis of Variance in
Analyzing On-Farm Trials 17
History of Stability Analysis by Regression 18
Fifty Years of Use by Plant Breeders 18
Usefulness of Stability Analysis in Assessing GxE
Interaction 20
Traditional Breeders' Perspective: Broad Adaptability 21
Several Definitions of Stability 23
Statistical Concerns with Regression Stability Analysis 24
The Modifications to "Stability Analysis" 25
The Book 26













Adaptability Analysis


2. Adaptability Analysis: An Overview 29
Introduction 29
Important Steps in Adaptability Analysis 29
Data Requirements 31
Sources of Data 31
Quality of Data 33
An Example of AA from Brazil 34
Relating Treatment Response to Environment by
Regression on El 35
Comparison of Results from AA Regression and from
ANOVA 42
Relationship between El and Environmental
Characteristics 45
Analysis within Tentative Recommendation Domains
(Verification) 51
Use of Alternative Evaluation Criteria 57
Multiple Extension Recommendations 58
Summary 60

3. Single-Factor Trials 63
On-Farm Maize Variety Trials (Paraguay) 63
Hormone Implants in Beef and Dual-Purpose Cattle (Panama) 73
On-Farm Sorghum Variety Trials (Cameroon) 78

4. Factorial Trials: Interactions Among Factors 95
Maize Variety-By-Fertilizer On-Farm Trial (Malawi) 95
Maize N and P Fertilization Trials (CIMMYT, Mexico) 107
Agronomic Analysis 109
Economic Analysis 109
Recommendations and Multiple Extension Messages 111

5. Technology Packages and Components 119
Bean Systems (Costa Rica) 119
Agronomic Analysis 119
Economic Analysis 124
Risk Analysis 125
Potatoes (International Potato Center, Peru) 130
Technology Package Trial, 1978 130
Technology Package Trial, 1979 135












Contents


Factorial Trial of Components, with Interaction 145
Was More than one Year of Data Necessary? 153

6. Systems Trials 155
Change to Pasture-Intensive Dairy System (New York) 155

7. Design of On-Farm Research-Extension Trials 163
Nature of On-Farm Trials 163
Research Questions 163
Extension Activities and the Learning Environment 165
Designing for Efficiency 167
Farmer Participation 167
Choice of Treatments 169
Choice of Control Treatments 169
Replications vs. Environments 172
Environments vs. Years 173
Data Collection 176
Location Specificity, Biodiversity and Sustainable Agriculture 178

References 181


Index























Acknowledgements


Data used in the examples in the book are taken from previously
published, public domain sources (citations are in the References).
Nevertheless, we wish to thank the persons and organizations whose data
were used. In order of appearance, this includes:

B. K. Singh, Professor at EARTH (Escuela de Agricultura de la
Region Tropical Hdmeda) in Costa Rica for the Manaus, Brazil cowpea
data. Also involved in this research were EMBRAPA and EMATER,
Brazil; the Soils Management CRSP funded by USAID; and the Soil and
Water Science Department; and the Farming Systems Program of the
Institute of Food and Agricultural Sciences, University of Florida.

Federico Poey, President, AGRIDEC, Miami, Florida, for the maize
variety trial data from Paraguay.

James R. Simpson, Professor of Food and Resource Economics,
University of Florida; Michael Sands, Rodale Institute; and Luis
Hertentains, INIAP, David, Panama, for the cattle data from Panama.

For the Cameroon sorghum variety data, Cameroon's Institut de
Recherche Agronomique (IRA), including the Testing and Liaison Unit
in Maroua, where the trials were conducted under the direction of Jerry
J. Johnson and Martin Fobasso. Also, thanks to the National Cereals
Research and Extension Project (International Institute of Tropical
Agriculture and U.S. Agency for International Development).

Art Hansen, Professor of Anthropology, University of Florida; and
Emmanuel Mwango and Benson Phiri, both of the Department of













Acknowledgements


Agricultural Research, Ministry of Agriculture, Malawi, for the Malawi
maize and fertilizer data.

CIMMYT, Mexico, for the maize nitrogen and phosphorus fertilizer
data.

Barbara C. Bellows, Research Associate, SANREM CRSP for the bean
systems data from Costa Rica.

Douglas Horton, CIP, Peru, for the potato technological package and
component data.

L. Toomer and D.L. Emmick, USDA Soil Conservation Service,
Syracuse, New York for the dairy grazing systems data.

The name given to the methodology, and used as the title for the book
- Adaptability Analysis was suggested by K.R. Tefertiller, Professor,
Food and Resource Economics, University of Florida, after reading early
versions of Chapters 1 and 2 of the book. We appreciate his insight and
convincing argument, and acknowledge his contribution.

For the facilities, services and time used during the preparation of this
book, we acknowledge the Food and Resource Economics Department,
Institute of Food and Agricultural Sciences, University of Florida.

Finally, a very special acknowledgement goes to RoseMary Espaillat,
Secretary, Food and Resource Economics Department, University of
Florida, for her patience, skill and good will during the many drafts of
the book and for setting up all the tables and crafting all the graphs
included. The excellent quality of the photo-ready copy is due to her
dedication and efforts. Thanks, RoseMary, from both of us.


















1


Introduction



Agricultural research and extension efforts during the last half of the
twentieth century have met with considerable success. Globally, food
and fiber production has kept pace with a rapidly expanding human
population even as increasingly less favorable agricultural areas are
brought into production. Per hectare yields of the many important crops
in the most favorable areas have increased dramatically during this period
of time. Yet there remain large and significant pockets around the globe
where the effects of international technology generation and diffusion
efforts by means of established research and extension organizations,
whether international, national or non-governmental, have been little felt.

Science and Agricultural Technology Development

Much of the success of this revolution in agricultural technology has
been based on the scientific agricultural research process developed by
such institutions as the United States Department of Agriculture and the
Land Grant University system of the United States and their extensive
networks of agricultural experiment stations, as well as by agricultural
research services in western Europe and by the International Agricultural
Research Centers (IARCs) of the Consultative Group on International
Agricultural Research (CGIAR). The research process at these stations,
which' is largely disciplinary in nature, involves a high degree of
experimental control. There are three reasons for this control: to
minimize variation in non-experimental variables for scientific purposes;
to reduce residual variance for statistical purposes; and to achieve












Adaptability Analysis


maximum expression of treatment variables for disciplinary purposes.
Resulting technologies were high yielding and indicators of the potential
available for agricultural production in the many farming systems of the
world.
During and following World War II the agricultural revolution in the
industrialized world, led by the United States, was characterized by
increasing capital investment with a substitution of tractors for animal
traction, followed by a decrease in draft livestock numbers, then in
amount of pasture on farms. As other livestock became increasingly
limited and availability of manure for fertility scarce, chemical fertilizer
use began to increase rapidly. Crop intensification was manifested in
more limited rotations and in a resulting need for and use of chemical
pesticides of all types. All of these factors, as well as irrigation, allowed
farmers to modify natural environments to make them more similar to
those created on experiment stations for scientific and statistical
purposes. In industrialized countries, these forces were creating an
agriculture with homogeneous conditions over very broad areas as
"innovators" and "early adopters" began buying up the farms of those
"laggards" and "non-adopters" who for one reason or another were not
adopting the technology. Industry, as well, benefitted from the potential
thus created for "broadly adaptable" technologies made possible by the
ability of capital-intensive farmers to create environments suitable for
high-yielding, modern technologies.
In the 1960s there was a great deal of enthusiasm for the potential of
this science-based, industry-supported agriculture to lead a "Green
Revolution" in the developing countries of the Third World. Wheat and
rice led the way, particularly in Asia. By the 1970s enthusiasm was
dampened as it became evident that the greatest yield advances of the
Green Revolution were being made on those farms with the best
environments and greatest resource bases. The technology researchers
and developers who had confidence in the transferability of science-based
and industry-supported agriculture became interested in "yield gap"
(Ghodake and Walker, 1982) or "yield constraint" (De Datta et al.,
1978) studies to ascertain why technologies did not produce the same
results in the field as they did on the stations where they originated.
These researchers correctly surmised that the gaps were at least in part
responsible for the lack of adoption of what to their minds should have
been "broadly adaptable" technologies.
Also during the 1970s, the approach to technology development now
called Farming Systems Research-Extension (FSRE) was beginning to












Introduction


emerge. This methodological process, heavily influenced by social
scientists (including agricultural economists), began to create an
awareness that the problem lay not with "laggard" farmers but with the
fact that small-scale, limited resource, family farmers were unable to
dominate and/or modify their environments in the ways that capital-
intensive, industrialized farmers were. Because of their constraints, these
small-scale farmers had to live and produce within the environmental
conditions caused by the natural endowment and socioeconomic situations
existent on their farms. The diversity of combinations created by these
conditions meant that technologies would have to be more environment
specific than envisioned for the broadly adaptable technologies of
industrialized countries. They would need to fit in both the base
biophysical environmental conditions and the socioeconomic situations of
the producers if they were to be adoptable. That is, the problem of non-
adoption lay not with the potential target farmers but with the nature of
the technology being generated for them.


Broadly Adaptable vs. Location-Specific Technologies

One research tool widely used by plant breeders over the last 30 years
to evaluate genetic material-new lines, varieties and hybrids-is called
"stability analysis". This procedure, to be discussed in more detail later,
provided a means to assess the response of new materials to different
environments and the interaction of different materials to environmental
changes. Because these materials were initially and extensively tested on
experiment stations, yield levels generally reflected highly productive
environments. The highest yielding materials under these conditions
often exhibited large linear coefficients (> 1) when regressed on an index
of environmental quality, indicating that they tended to be very fragile
and break down in poor conditions. Materials that were more robust and
held up in poor conditions tended to yield relatively poorly in the
experiment station conditions, resulting in small linear regression
coefficients (< 1). The higher-yielding materials with linear regression
coefficients approaching one were considered to be the "best", "most
stable" or "most broadly adaptable" (variety B in Figure 1.1).
It has been argued (Hildebrand, 1990a) that rejecting materials that
excelled in the best environments but collapsed in poorer environments
(variety A in Figure l.la), and those that excelled in poor environments
but did not respond to good environments (variety C) was a negative












Adaptability Analysis


0 1 2 3 4 5
ENVIRONMENTAL INDEX, El


0 1 2 3 4 5 6
ENVIRONMENTAL INDEX, El


FIGURE 1.1 A negative interpretation (top) of treatment-by-
environment interaction resulting in the compromise choice of one
broadly adapted treatment, B, and a positive interpretation (bottom) of
the same interaction, resulting in the choice of two treatments, A and C,
specifically adapted to good and poor environments, respectively.












Introduction


interpretation of the results of stability analysis. In essence, it was a
compromise based on the assumption that one "best" variety had to be
selected (variety B), an assumption based on the belief that environments
could not be stratified adequately to allow for specific recommendations
or could be so modified by use of inputs that "poor environments" would
be due only to unpredictable characteristics such as rainfall and pest
incidence. This compromise resulted in the rejection of many materials
that otherwise might have been well adapted to the poor environments
inhabited by most of the world's small-scale farmers.


Origin of Adaptability Analysis

The modifications to traditional stability analysis (described in detail
on page 25) that resulted in Modified Stability Analysis (Hildebrand,
1984), or MSA, included a rethinking of this negative interpretation. To
generate materials for specific environmental situations, those that do
well in "good" environments are accepted for those "good" environments
even if they do not do well in "poor" environments. Conversely, those
that do well in "poor" environments, even if they do not respond to
"good" environments are accepted for those "poor" environments. This
positive interpretation of environmental interaction (Figure 1. Ib) provides
the opportunity to identify not only genetic materials that excel in wide
ranging conditions, but also any other kind of technology being evaluated
as potentially useful.
The incorporation of the evaluation of other kinds of technology
(fertilizers, rotations, intercropping systems, livestock practices, etc.) by
means of MSA was another modification of stability analysis. Thus,
MSA could serve as a comprehensive method for the design, analysis and
interpretation of on-farm research data for specific adaptability. It can
be useful in the developing countries, where tremendous environmental
variability exists, and in industrialized countries where concerns with a
more sustainable agriculture require more location-specific technologies
than have been the norm over the past half century.
Unfortunately, confusion over the word "stability" in the name MSA
has prevented many practitioners from fully comprehending and
exploiting the potential of the method for identifying technology adapted
to specific conditions. Since the method in fact aims at specific
adaptability and not broad adaptability, i.e., not stability across diverse
environments, we propose the name Adaptability Analysis (AA), rather













Adaptability Analysis


than Modified Stability Analysis (MSA), as a more accurate description
of the method.


Purpose of On-Farm Research-Extension

On-Farm Research-Extension and Farming Systems
Research-Extension

Farming systems research-extension (FSRE) is a relatively recent but
increasingly common and familiar approach to the generation, evaluation
and diffusion of agricultural technology, particularly for small-scale,
limited-resource farming systems. The Association for Farming Systems
Research-Extension, founded in 1989, describes the objectives of the
approach as the

development and adoption through participation by farm household
members of improved and appropriate technologies and management
strategies to meet the socioeconomic and nutritional needs of farm families;
to foster the efficient and sustainable use of natural resources; and to
contribute toward meeting global requirements for food, feed, and fiber.

More concretely, there are several characteristics of farming systems
research (FSR) common to almost all variants of the approach. As it
was originally conceived, through the concurrent efforts of many people
working in several different parts of the world in the 1970s, the key
concepts that characterize FSR are that

FSR isfarmer-oriented, seeking to provide technologies relevant
to farmers' goals, needs, and priorities;
FSR is systems-oriented, viewing the farm holistically and focusing
on the interrelationships and interactions among farm sub-systems;
FSR is problem-solving in nature, with the goal of identifying
specific constraints to improved production and developing specific
solutions;
FSR is interdisciplinary, dependent on collaboration between
biological and social scientists;
FSR complements strategic, commodity and disciplinary research,
drawing upon the existent body of knowledge, and adapting the
results of these research efforts to farmers' circumstances;













Introduction


On-farm research is central to FSR, ensuring close collaboration
between researchers and farmers and allowing evaluation of
technologies under the environmental and socioeconomic conditions
in which they will be used;
FSR providesfeedbackfrom farmers, permitting communication of
farmers' points of view to policy makers and station-based
researchers.
(Adapted from Merrill-Sands, 1986)

The addition of an "E" to FSR reflects a recognition that the adaptive,
problem-solving nature of FSR, particularly its on-farm research
component, serves an extension function in itself but more importantly
is enhanced by active collaboration with field-level extension personnel.
Many practitioners of FSRE have come to realize that the
interdisciplinarity of FSRE should include, by definition, linkages that
extend across the separation of research and extension and also the
linkages across the boundaries of biological and socioeconomic research.
The term on-farm research-extension also reflects the basic tenets that 1)
on-farm trials or other on-farm research activities often serve an
extension as well as a research function, and 2) such activities are
uniformly more efficient and effective when carried out as joint research
and extension collaborations.
Although there is in practice considerable diversity among various
efforts-projects or national programs, called FSR or FSRE, most such
efforts use an approach characterized by a number of identifiable stages,
such as description or diagnosis, design, testing and dissemination.
These stages can be summarized as follows:

Diagnosis, or initial characterization of the farming systemss, through
activities such as review of existing information, and informal surveys,
e.g., sondeos, rapid rural appraisals, etc. This characterization
includes
preliminary identification of problems and constraints;
tentative grouping into homogeneous farming systems.'


SMany farming systems practitioners consider homogeneous farming
systems to be synonymous with recommendation domains, reflecting early FSRE
methodology (Hildebrand, 1986, p 52). In this present book, recommendation
domains and homogeneous farming systems are quite different.
Recommendation domains as used here can be described as combinations of












Adaptability Analysis


Planning and design of research activities, including
biological research;
continuing physical and agro-socioeconomic characterization.
Development and evaluation of technologies:
on-station commodity and discipline research;
researcher-managed on-farm trials, including exploratory, site-
specific, and regional trials;
farmer-managed on-farm trials, for 1) evaluation of technologies
by farmers themselves, 2) refined partitioning of recommendation
domains, and 3) initiation of technology diffusion.
Promotion of acceptable technology throughout the appropriate
diffusion domain(s).

In all stages the following activities occur:

Collection, analysis and interpretation of data:
experimental and non-experimental data from trials;
economic farm enterprise records;
other socioeconomic data from surveys.
Iterative, regular reexamination of activities and results, for
continual refinement of recommendation domains;
elucidation, in collaboration with extension, of location-specific
recommendations;
incorporation of farmer and field-extension feedback into the
process;
improved planning of future work.
(Adapted from Hildebrand and Poey, 1985)

Recognizing Farm and Environmental Diversity

Inherent in the FSRE approach to adaptive research-extension is the
recognition that the limited-resource farm households that are important
clientele of FSRE, live and work on farms characterized by a high


environmental conditions and evaluation criteria from information obtained
during the on-farm research-extension effort. This is a concept much more
useful than the older classification of homogeneous farming systems based on
information obtained only during the diagnosis. The definition of homogeneous
farming systems in the diagnosis is valuable to help understand the diversity of
systems in the research domain and as a guide to help locate on-farm trials to
sample this diversity.












Introduction


degree of both biophysical and socioeconomic diversity. FSRE was
developed in large part in response to the failure of previous efforts at
technology development and transfer, efforts which assumed that
technologies "proven" superior on experiment stations would as a matter
of course be recognized as superior and adopted by farmers.
Technologies developed solely under the conditions common on experi-
ment stations are, however, rarely transferable directly to small-scale,
limited-resource farmers. First, the biophysical conditions of stations are
usually extremely favorable to crop and animal production.
Traditionally, selection of research station sites has favored fertile soils
with good structure and water-holding capacity. Over time, continued
application of fertilizers and pesticides has raised even further the basal
productivity of these sites. Experiment stations generally, then, are
unrepresentative even of their geographical area in that they usually
produce consistently high yields as a result of their histories of high-
input, almost "unlimited-resource" management.
The problem, however, is not just that conditions on experiment
stations do not allow identification of technologies that will perform well
under limited-resource, low-input conditions. It is also that the high-
input management regimes characteristic of stations, particularly when
these include irrigation, tend to render the stations more homogeneous,
in effect smoothing over both differences among fields within a station
and also differences among stations. This situation tends to favor the
development of overly delicate technologies, i.e., technologies that
require a certain, specific, narrow (usually high) range of growth factors
in order to perform well. Robust technologies, those capable of good,
or at least adequate results under the broader, more variable, and often
unpredictable range of conditions faced by most farmers, tend to be
systematically disfavored under usual experiment station conditions.
This is not to argue that commodity or discipline (strategic) research
carried out on experiment stations is unnecessary or irrelevant. Such
research is valuable in the assessment of many alternative technologies
under well-controlled conditions, subject to scientifically rigorous
standards. Through factorial or other multi-factor experiments, such
research allows examination and explanation of important interactions
both among experimental factors and between experimental and
environmental factors. It also allows estimation of the upper limits of
the production potential of technologies. On-farm research-extension, as
carried out in the FSRE approach can convey knowledge of important
problems to be addressed by on-station, component research.












Adaptability Analysis


Additionally, FSRE can attempt to adapt the results of this necessary
discipline- and commodity-based experiment station research through
joint research-extension-farmer evaluation of promising technologies
under the range of environmental conditions prevalent in the area in
question.
Sometimes researchers attempt to develop technologies under a range
of environmental conditions by conducting multi-stationtrials. While on-
station multilocational trails are often an essential component of the
technology development process, they have serious limitations. They do
take into account some environmental diversity (the between-location,
bio-physical effect), but they still fall short of evaluating technologies
under the types of conditions farmers face. First, since each station is
usually a collection of good environments, the range of environments
represented will also be higher than the range of environments found on
farms. In addition, the range of environments will probably be much
more narrow. This is because in order for an environment to be "good"
it has to supply some minimum amount of each of many factors required
for crop or animal growth and development. All "good" environments
supply the necessary amounts of all these requirements. On the other
hand, any environment can be "poor" by failing adequately to supply just
any one of these factors. Good environments, then, tend to be relatively
similar, while each bad environment can be bad in its own way. In any
event, few would dispute that most experiment station environments are
more like other station environments than one farm is like other farms.

Environments Defined

An environment, as used in this book, is the product of the entire set
of factors, both biophysical and socioeconomic, that can materially affect
crop and livestock productivity. These factors (collectively, the
environment) in turn affect performance and eventual acceptability by
farmers of a given technology. Many types of factors vary to cause
differences among environments. These include, among others:

Farm type:
irrigated vs. rainfed;
mechanized vs. animal traction vs. hand tillage;
commercial vs. subsistence (or large vs. small);
diversified vs. single-enterprise.













Introduction


Nature of the farm household:
ethnic or cultural background;
hierarchy of authority (female- vs. male-headed, etc.);
importance of off-farm or nonagricultural on-farm employment;
access to and control over production factors and returns to
production.
Climate:
rainfall (amount, distribution, reliability, etc.); temperature;
altitude; humidity, etc.;
Soils:
origin, texture, structure, depth, water-holding capacity, drainage;
fertility, mineral nutrient status, organic matter content, cation
exchange capacity, pH, etc.;
slope, position on toposequence;
color, indigenous classification.
Farmer management:
traditional "farmers' practices" vs. "improved" practices, e.g., use
of organic or chemical fertilizers and pesticides, extension
packages, etc.;
date of seeding, timing and frequency of cultural operations, etc.
rotation, previous crop.
Other.
geographic location, or "agro-ecological zone";
commonly occurring disease and pest pressures;

It is important to realize that environment, as used here, can be but is
by no means always synonymous with "site" or "location" as these terms
are commonly used by researchers. In the case of multilocational on-
station trials, for example, it is recognized that soil or other differences
within a single location require partitioning of this variability by
blocking; in this case each block is in fact a separate environment.
Similarly, on-farm research programs often group trials within research
villages, and comparison between villages results in each village being
considered and called a separate "site" or "location." Yet within each
village there are sometimes several repetitions of a given trial. These are
most often single replicates, i.e., one replicate per farm, of the set of
treatments to be compared. In such a case, each individual set of the
treatments is a separate environment. Note that "environment" is not
even necessarily synonymous with "farm" or even with "field", since in
the case of single-replicate on-farm trials it is possible-if not common-












Adaptability Analysis


for a farm or field to have more than one replicate of the trial, and thus
more than one environment.
In this book, the term environment usually will correspond to the
"block" or replicate in farmer-managed regional or verification trials-
with each farmer testing a single replicate of the treatments (usually in
a completely randomized block). It will be meant to reflect the entire set
of factors that can cause differences in the responses of crops and
livestock to the technology being evaluated.

Incorporating Farmers' Perspectives and Evaluations

Traditionally, two methods have been employed to incorporate
farmers' perspectives and evaluations into the development of agricultural
technology. The first is through farmer feedback about the technology
after viewing "demonstrations", or after some initial adoption during the
extension/diffusion process. The second is through field-day visits by
farmers to experiment station trials. The first of these methods is
necessary in all cases; it tends, however, to happen rather late in the
technology development process. If there are problems with technologies
being developed by research, finding out about them and "going back to
the drawing board" at the extension/diffusion stage results in much lost
time, labor and financial expense.
The second of these traditional methods is also useful and should not
be neglected. But it must be recognized that farmers are not very likely
to give pertinent opinions and evaluations of technologies they see
"demonstrated" under what to them are foreign conditions, conditions
almost always substantially more favorable than their own.
By doing a considerable portion of the development of new
technologies in close collaboration with farmers-through on-farm
research-extension activities-incorporation of farmers' own evaluations
of those technologies will be more timely, thereby reducing wasted time
and effort, and more relevant to their specific circumstances. This can
only help to ensure relevance of the final recommendations eventually
made. To be most effective, farmers' perspectives and judgments should
be elicited continually during their participation in all phases of the
technology development process. Such participation is essential in
identification of priority problems and potential solutions to those
problems (through both formal and informal diagnostic efforts), in design
and implementation of on-farm trials, and in analysis and interpretation
of trial results.












Introduction


It is important that the results of on-farm research-extension, socio-
economic as well as biological, be presented and discussed with farmers,
either individually or in groups. Although it is difficult to translate the
often complex results of trials into simple language that can be
understood by farmers who often have little formal education, the
feedback obtained by farmers, and the value of that feedback in
formulating viable extension recommendations, is well worth the effort.
There is one question researchers should always keep in mind: if a
proposed trial is so complex that its results will not be understood by the
farmers in whose fields or herds it is to be done, is it not too complex
to be an on-farm trial in the first place?

Enhancing Adoption Through Adaptation and Learning

Many innovations extended to farmers are rejected by them because
the learning and mastery of the new technologies involved are difficult,
and positive results are not seen by farmers until after several or many
instances of trial and error. The learning process undergone by farmers
collaborating in on-farm trials is similar to that undergone by other
farmers once new technologies reach the diffusion stage of the
technology development and transfer process. Collaborating farmers
learn, through participation in trials, how best to implement new
technological options and how to modify and adapt them to specific local
conditions in order to get the most out of them. On-farm researchers
learn from these adaptation efforts both how better to design future
technologies aimed at similar farmers, and also how better to interpret
and extend the results of on-farm research.
The fine-tuning (or sometimes the complete overhaul) of technologies
tested under farmers' conditions substitutes to a large degree for the same
process that farmers trying later to adopt the same technologies would
otherwise have to go through. Having a good part of this learning and
adaptation take place in the development and testing phase, rather than
the diffusion phase, spares farmers willing and eager to adopt innovations
the confusion, disappointment and frustration of trying to make poorly
adapted technologies work in their particular circumstances. Avoiding
these negative experiences goes a long way in ensuring the rapid and
lasting adoption of externally introduced agricultural innovations. The
concept of learning curves, how they are affected by improved adaptation
of technologies, and how they affect both the results of on-farm trials












Adaptability Analysis


and the efficiency of the technology development process, will be
discussed later in Chapter 7.

Location Specificity and Sustainable Agriculture

Modern agricultural technology is developed for, and is itself
responsible for developing highly regulated and increasingly
homogeneous ecosystems. Local variation in soil fertility and water
holding capacity, and local differences in rainfall regime and in pest
incidence, are compensated for and to some extent eliminated by
irrigation, mechanization, fertilizers and pesticides (Nguyen and
Anderson, 1991). It is for just this sort of agro-ecosystem that "broadly
adapted" varieties and other agricultural technologies are generally
developed in the richer countries of the world.
In recent years, however, there has arisen an increased awareness that
the ecological and economic costs of creating and maintaining these
highly controlled, homogeneous agricultural environments will not be
sustainable in the long term. Greater attention is currently devoted to the
development of "alternative" or "sustainable" agricultural systems. It is
certain that these systems will have greater diversity from one location
to another than do those whose indigenous variability is dominated by
heavy use of external agricultural inputs.
As agricultural policy makers begin to change the incentives which
have encouraged the use of technologies broadly adapted to these often
artificially produced, superior environments, new technologies will have
to be developed "to conform with the environments where they will be
used, not dominate them (Hildebrand, 1990b, p. 286). This change in
emphasis, and the concurrent imperative to target specific environmental
conditions will make on-farm research-extension activities central to the
development of technologies conducive to more sustainable agricultural
systems. Adaptability Analysis (AA), using environmental
characterization and evaluation criteria appropriate to specific farmers
and production circumstances, will help in ensuring the identification of
technologies specifically adapted to the more variable environmental
conditions of these systems.












Introduction


Problems with Combined Analysis of Variance in Analyzing On-Fann
Dials

The most common analytical method for the analysis of agronomic and
other agricultural trials is the analysis of variance (ANOVA). While a
very powerful and useful tool, analysis of variance is somewhat complex
and sensitive to violations of a number of statistical assumptions; these
assumptions are often not met by the sorts of "messy" data produced
under farmers' conditions, characterized by a high degree of variability,
which is not normally accounted for in the sorts of experiments that lend
themselves to analysis by ANOVA.
The major difficulties with ANOVA for analysis and interpretation of
on-farm trials will be dealt with in detail in Chapter 2. They can be
summarized, however, as follows. First, the important sources of
environmental variability, those likely to affect decisions concerning
specific adaptability of experimental treatments to particular
environmental conditions, must be identified and incorporated into a
specific experimental design ex ante, that is, before the trial is put in the
field. This is a problem in that one of the purposes and great benefits of
on-farm trials is precisely that they help identify those environmental
characteristics liable to determine specific adaptability; these
characteristics are rarely known beforehand. Even if they were known,
however, their incorporation into on-farm trials, as will be seen below,
results in trials that are too large, too complex, and too difficult to
analyze.
The second difficulty with ANOVA is that in order to estimate the
importance of the treatment-by-environment interaction critical to the
analysis of adaptability of different treatments to different environments,
ANOVA requires a suitable error term to test the significance of the
interaction. This is a problem because the types of on-farm trials most
likely to give an accurate and complete range of environmental
conditions are farmer-managed ones, either regional or verification trials,
according to the typology of Hildebrand and Poey (1985). These types
of trials, since they are managed by farmers whose own evaluations of
the treatments reflect the most important analytical criteria, are best and
most profitably implemented with one complete replicate of treatments
per environment (in most cases this means one replicate per field).
Unfortunately, for this sort of trial the treatment-by-environment
interaction cannot be tested by simple ANOVA. There is no error term
with which to test it, since this interaction is itself the residual error













Adaptability Analysis


term, i.e., the term in ANOVA which represents that part of variability
unaccounted for by experimental factors.
Doing on-farm trials with two or several replicates per farm allows for
estimating the treatment by environment interaction, and combined
analysis of variance is a powerful technique for statistically estimating
that interaction. If, however, the implementation of trials is made so
complicated that it jeopardizes the full involvement of farmers, for the
sake of ai analytical method that most extensionists and many
researchers find impossible to do correctly with the computing methods
at their disposal, and the output of which is often not fully or even
adequately interpreted, the potential benefits of on-farm research are lost.
Adaptability Analysis is not proposed as a substitute for ANOVA as
a tool for the design and analysis of on-farm research-extension.
Adaptability Analysis (alone or in combination with ANOVA) does,
however, have significant advantages over ANOVA alone. First, it is
simple and relatively easy to perform and to interpret; second, it allows
evaluation of the critical treatment-by-environment interaction without the
need for the complications inherent in using more than one replication
per on-farm test site. Finally, the output of AA is in a form
understandable and relevant to farmers and extensionists as well as to
researchers.


History of Stability Analysis by Regression

The method of regressing varietal yields at each of many locations on
the mean yield of all varieties at each location was developed to provide
a means of estimating the interaction of plant genotypes and
environments. Before looking at how Adaptability Analysis modifies this
technique, it would be useful to review briefly its history and
development, its strengths and its statistical weaknesses, and the context
in which plant breeders have employed it.

Fity Years of Use by Plant Breeders

Plant breeders have long recognized that the production of a given
variety will vary greatly from one trial location, i.e., environment to
another. Whether different environments result from differences in
climate, soils, disease and pest pressures, or management practices, some
varieties will be better suited to some environments than to others.












Introduction


Useful comparison of varieties within a crop improvement program,
then, requires an accurate estimate of the differential responses of a
number of varieties over a range of locations, an estimate of what
breeders call the genotype-by-environment (GxE) interaction.
Early statistical attempts to analyze GxE interaction among a group of
varieties involved refining the analysis of variance procedure to partition
the variability in a multilocational varietal trial into those variance
components attributable to differences among varieties, those due to
differences among locations, and those due to the interaction of the
genotype and location (environment). These attempts were fruitful not
just in helping breeders to better understand GxE interaction, but also in
advancing the state of the art of ANOVA. They were, however,
characterized by a statistical sophistication that made them difficult to
grasp or apply without a strong background in mathematical statistics.
The use of regression on the mean yields of all varieties at a trial site
dates from the late 1930s when Yates and Cochran (1938) illustrated that
oftentimes a great deal of the variety-by-location interaction in ANOVA
of multilocation trials can be partitioned out and accounted for by
differences in linear regression on mean yield, an artifice frequently of
use in revealing relations between general fertility and varietal
differences (p. 565).
Finlay and Wilkinson (1963) further developed the concept of using
linear regression of varietal yields on site means as a technique for
assessing adaptation in a plant breeding program. They simplified the
technique by using the individual varietal means as the dependent
regression variable (rather than the difference between individual yields
and the overall mean, as Yates and Cochran had done). Oddly, they
used the total of yields at a site, rather than the mean of all yields, as the
independent regression variable; this difference does not, however,
materially change the interpretation of results. More importantly, they
discussed in detail the implications of the technique for determining the
adaptability of varieties to changing environmental conditions, whether
breeding is for specific adaptability, "for closely defined ecological
conditions," or for general adaptability, "for more extensive conditions
that include a considerable range of environments" (p. 743).
The most widely cited publication describing linear regression stability
analysis, and the first one to use the term "environmental index," adapted
the technique for the purpose of identifying a number of "stability
parameters," one of which was the linear regression coefficient, another
the deviations from regression (Eberhart and Russell, 1966). This work,












Adaptability Analysis


perhaps the most cited in the literature, has confused the interpretation
of the results of such regression analyses. The confusion arises first, as
will be seen below, because Eberhart and Russell changed the definition
of "stability," and second because they proposed statistical measures of
varietal stability as indicators of adaptability, resulting in later over-
emphasis on stability itself rather than adaptability-either general or
specific-to environmental conditions. Nevertheless, in one form or
another, the technique of linear regression of varietal yields on site
means has been used by two generations of plant breeders interested in
dealing with the differential response of varieties to a range of
environments.

Usefulness of Stabiity Analysis in Assessing GxE Interaction

Linear regression of varietal yields on mean site yields is only one of
many techniques plant breeders have developed to estimate the
importance of the interaction of plant genotype and environment. As
discussed above, it was initially developed to better explain GxE
interaction in a plant breeding program. Once overall variability has
been partitioned into that due to genotype, that due to environment, and
that due to the interaction of the two, the variability accounted for by the
interaction can be further partitioned into that part of the GxE interaction
due to linear regression on site mean and some unexplained (residual)
variability, essentially the deviations from the linear regression.
Stability analysis by regression was not the first nor last method used
by breeders to quantify GxE interaction. Many newer and much more
complex and sophisticated methods have been developed, and many
statistics identified, but linear regression remains much used. Its
usefulness is that it allows application of familiar and easily interpreted
statistics (the regression coefficient and deviations from regression) to the
assessment of GxE interaction. Freeman (1973), in a comparison of
many statistical methods for the analysis of GxE interaction, considers
that "Of all the techniques discussed, there can be no doubt that, for
geneticists, the most fruitful has been the regression approach" (p. 350).
One result of the success of linear regression for analysis of GxE
interaction was that the relationship of varietal yield with site mean,
represented by the regression line, led to something of a preoccupation
with assessing the stability of varieties across environments and
identifying "stability parameters." A multitude of stability analyses arose
to rival the multitude of techniques for assessing GxE interaction. In












Introduction


fact, since one of the chief reasons for growing varieties in a range of
environments has always been to estimate their stability, analysis of GxE
interaction has by and large evolved into and been supplanted by
"stability analysis".

Traditional Breeders' Perspective: Broad Adaptability

There are two reasons why linear regression stability analysis, and
other types of stability analysis as well, have traditionally been developed
and used by plant breeders to identify broad adaptability in varieties.
The first is that it has been assumed that modern crop production
management methods, i.e., the use of chemical fertilizers, pesticides,
mechanization, irrigation, etc., could in most cases dominate the
indigenous factors that tend to cause yields to be generally low and
highly variable. The second is that breeders have thought it impossible
or at any rate too difficult to identify ex ante the specific environmental
conditions that exist in any given farmer's field and to relate these to the
environmental conditions of the several sites in which varieties are tested
in multilocational trials.
Finlay and Wilkinson (1963) rightly considered that varieties with bi
= 1 had "average stability", but their chief goal was not stability, in
itself, rather varieties "well adapted to all environments." These
varieties would be those with average stability and high mean yields
(Figure 1.2). While recognizing the advantage of specific adaptation to
"closely defined ecological conditions," Finlay and Wilkinson
considered this to be impractical for most rainfed field crops, given the
variability due to season: "Even in a uniform edaphic environment a
considerable degree of general adaptability will be important, because of
the marked fluctuation of climatic conditions from season to season." (p.
743). This is an important concern; in using Adaptability Analysis for
identification of specifically adapted technologies, particular care must
be given to ensure that the range of environments in a trial be representa-
tive of the range of environments that exist over years. This point is
discussed in detail in Chapter 7.
Similarly, Eberhart and Russell (1966) defined an "adapted variety" as
one with high yield, bi = 1, and small deviations from regression.
Much attention has been devoted to their definition of a slope of unity
(along with small deviations from regression) as indicating stability, but
it is less commonly recognized that stability per se was not, in the view
of Eberhart and Russell, the primary goal of a plant breeding program.














Adaptability Analysis


The primary goal was a broadly adapted variety. Varieties with bi < 1
would yield less than the average of all tested varieties (i.e., be poorly
adapted) in high-yielding environments; those with bi > 1 would yield
less than average (and therefore be poorly adapted) to low-yielding
environments. This, as has been discussed, is a negative interpretation
of the linear regression on environmental mean; it gives emphasis to
where varieties will be poorly adapted rather than looking actively for
where they will be well adapted. It is also explicitly intended for
"conditions such as exist for maize in the United States, [where] the
breeder usually wants a variety that does above average in all
environments" (p. 38).






SPECIFICALLY
>1 ADAPIED TO
FAVOURABLE
ENVIRONMENIS

BELOW
AVERAGE STABEITr

1.0 POORLY ADAFIED TO AVERAOB STABUr WEIL ADAPTED TO
S ALLENVIRONMENS AL ENVIRO
0. ABOVE
AVERAGE STABIDIY


SPnamICAuIY ADAPTED
<1 TOUNFAVOURAEBL
ENVIRWOMEIS



VARIETY MEAN YIELD


FIGURE 1.2 Broad and specific adaptation, determined by mean yield
and linear regression coefficient (source: Finlay and Wilkinson, 1963).












Introduction


For most limited-resource farmers in many parts of the world,
however, risk avoidance and above-average yields in poor environments
(and seasons) are far more important than above-average yields in all
environments. In part, this importance derives from the fact that limited-
resource farmers are less able to modify their poor environments with
high levels of inputs. Note that the increased importance of agricultural
sustainability is leading many to recognize that all farmers, not just those
in developing countries, must farm within the capabilities of their
environments and not modify the environments to suit the technology.

Several Definitions of Stability

Considerable confusion has arisen from the fact that the several
methods of stability analysis have engendered a great many different
measures of stability. In a recent review, Lin, et al. (1986) identified
nine stability statistics which they classed into four general groups and
related this classification to three different concepts of stability. Two
groups of statistics are based on sums of squares from ANOVA. The
other two groups of statistics are based on linear regression, the first
group on the regression coefficient, the second on deviation from
regression. The authors give preference, as does Freeman (1973), to use
of the regression coefficient since it gives the shape, or structure, of the
varietal response as well as its variation.
The sheer number of statistics proposed as stability parameters has
caused confusion; even more confounding is that two of the seminal
papers in the development of the regression approach to stability analysis
define stability, with regard to the regression coefficient, differently.
Eberhart and Russell (1966) define "a stable variety" as one with a
regression coefficient (b)1 equal to one, and with deviations from
regression (s2d) equal to zero. Finlay and Wilkinson (1963), on the other
hand, consider a variety with bi = 1 to be a variety of average stability,
one with bi > 1 to be a variety with less than average stability, and one
with bi < 1 to be a variety with greater than average stability. By
extension, a "stable variety" would be one with bi = 0.
It is somewhat unfortunate that most plant breeders followed the lead
of Eberhart and Russell in defining stability as b, = 1, which is not
intuitively obvious. Even more unfortunate, however, is that many
breeders and other researchers have followed the lead both of Eberhart
and Russell and of Finlay and Wilkinson in using regression-based
stability analysis as a means of identifying varieties or other new












Adaptability Analysis


technologies that exhibit "broad adaptability." As will be discussed in
detail in Chapter 2, Adaptability Analysis uses regression of treatment on
an environmental index not to identify those treatments that are "stable"
or that have "broad adaptability," but rather to identify those treatments
that are best adapted to particular environments.

Statistical Concerns with Regression Stability Analysis

Stability analysis by regression of treatments on site means has been
much criticized from a statistical viewpoint. The chief criticism was that
because the environmental index, the mean at each trial site or
environment, is calculated from all of the individual treatment yields, and
is therefore clearly related to each of them, this type of analysis violates
an assumption of least-squares regression that the dependent and
independent regression variables be independent of each other, i.e., that
their error terms not be correlated. The major problem with what
Freeman (1973) calls "the logical difficulties of regressing one set of
variables on another which is not independent of them" (p.343), is that
estimates of bi and other regression statistics, as well as tests of their
significance or of differences between them, are biased.
Despite these concerns, the authors of several reviews of stability
analysis methods, including Freeman (1973) and Lin, et al. (1986), have
ended by maintaining that until multivariate techniques using independent
environmental measures are developed, linear regression on
environmental mean is perhaps the most useful of the currently available
techniques. Its advantages are that it is relatively simple and, more
importantly, that it permits an analysis of the structure, i.e., a graphical
representation, of the treatment-by-environment interaction.
In any event, none of these statistical concerns compromise the
usefulness of Adaptability Analysis. It is of critical importance that
researchers realize that AA is not intended to give statistically exact
estimates, either predictive or descriptive. Altogether too much emphasis
is now placed on the statistical "significance" of the regression statistics
in AA, as will be discussed later. If the goal of on-farm research-
extension were to attach values to new technologies that pretended to
represent their degree of stability or of general adaptability, then AA
would be flawed. For identifying technologies specifically adapted to
given environmental conditions, however, AA, properly understood and
employed, is the simplest and perhaps the most valuable tool currently
available.













Introduction


Those unduly preoccupied with the supposed statistical problems of
AA should keep two thoughts in mind. The first has to do with ultimate
ends:

When any statistical analysis is carried out the basic question always is:
what use is to be made of the results? It is only a secondary question to
enquire: how are the data to be analyzed? Too much concentration on the
second point leads to what has been described as the third kind of statistical
error: the right answer to the wrong question. (Freeman, 1973, p. 350).

The second consideration is concerned with the need to make the best use
of the methods at hand, even if these are not perfect:

Theoretical objections aside, the on-farm researcher often has but two
alternatives: using (an environmental index) El or being unable to make
useful recommendations within a reasonable period of time. And, as
McCullagh and Nelder (1989) point out, "A first, though at first sight, not
a very helpful principle, is that all models are wrong; some, though, are
more useful than others and we should seek those." (Stroup et al., 1993,
p. 169).

Traditional agricultural research and extension methods have long
failed limited-resource farmers in many parts of the world, a failure
reflected by generally poor adoption of technologies supposedly
"improved" but ultimately shown to be poorly adapted to farmers'
conditions. Given this fact, the practical usefulness of AA in identifying
adapted technologies outweighs any statistical imprecisions. Returning
to the analogy of AA as a tool, it can be likened more to a machete than
to a scalpel. While this analogy may dismay some researchers, it should
cause no concern at all to those charged with the rapid development and
dissemination of new technologies to farmers, nor to farmers themselves.

The Modifications to "Stability Analysis"

Several of the modifications to "stability analysis" as used by plant
breeders for many years have been mentioned. The following is a more
complete list of the modifications that were incorporated in MSA and are
now incorporated in Adaptability Analysis, AA.

By making a positive rather than negative interpretation of the
treatment-by-environment interaction, AA encourages the












Adaptability Analysis


development of situation-specific technologies critical to sustainable
agriculture as well as to limited resource farmers.
By emphasizing fuller characterization of on-farm environments for
use in the interpretation of results, AA increases the efficiency of
farmer and extension involvement in the research-extension process.
AA stresses the use of multiple evaluation criteria, including those
most relevant to farmers for whom the technology is being
developed.
The specificity of AA permits the development and use of multiple
extension messages (recommendations) for technology diffusion
purposes.
AA is suited to the evaluation of all kinds of technologies, not just
germplasm.
Measures of risk meaningful to farmers as well as researchers are
incorporated in AA.

Taken together, these characteristics of AA help it serve as a basis for
any on-farm research-extension program for technology development.


The Book

This book is intended to collect, synthesize, and bring to the attention
of a larger audience the most up-to-date thinking in the use of a
comprehensive method for the design, analysis, and interpretation of on-
farm research results. The method has come to be called, for better or
worse, Modified Stability Analysis (Hildebrand 1984), a name which
reflects the nature of its origins, but which unfortunately has tended to
mislead many practitioners of agricultural research and extension as to
its purpose and its full promise.
One of the major purposes of this work, then, is to help correct some
of the misconceptions and misuses of the method, which we propose to
call Adaptability Analysis (AA), because it reflects how the method is
best employed. Another main purpose is to illustrate the broad potential
of Adaptability Analysis through a review of the conceptual
underpinnings of the method and through a series of analyses of real on-
farm data.
To judge from results reported in journals and elsewhere, limits to
adequate exploitation of the method until now have been due in part to
an implicit assumption that it is above all an analysis of stability. This












Introduction


is not the case. The full usefulness of Adaptability Analysis is not-or
is only rarely and incidentally-that it allows identification and
comparison of the stability of agricultural technologies such as improved
varieties and breeds, fertilizer use, and other production practices across
a range of production environments. The advantage of AA is rather that
it is a simple and accurate method of identifying and comparing the
performance of agricultural technologies under specific biophysical
conditions and socio-economic circumstances. In brief, while traditional
stability analysis seeks to identify broad adaptability, i.e., adaptability
to a wide range of environments, AA seeks to help identify specific
adaptability, i.e., adaptability to particular groups of environments.
In Chapter 2 we present an overview of Adaptability Analysis
including a step-by-step analysis of a real-world example from the
Amazon area of Brazil. Chapters 3 through 6 contain a number of
examples of different types of trials for which AA is an appropriate
analytical tool. We hope that these chapters will serve to spur the
imagination of research-extension workers as well as to illustrate the
flexibility and robustness of the method. Finally, Chapter 7 is devoted
to the design of on-farm research amenable to Adaptability Analysis,
making the investment in the research-extension effort more productive
and adapted technologies more rapidly available to farmers.













28 Adaptability Analysis


















2


Adaptability Analysis:
An Overview


Introduction

This chapter presents a method of experimental design, analysis and
interpretation previously called Modified Stability Analysis (MSA),
which we have changed to Adaptability Analysis (AA). MSA was
developed to use the linear regression methods of plant breeders to
analyze farmer-managed, usually single-replication, on-farm trials across
a wide range of conditions, and to incorporate an analysis of risk into
those methods (Hildebrand, 1984). Its advantages are particularly
evident when it is used to identify and develop technologies adapted to
specific on-farm environmental conditions and evaluated by farmers' own
evaluation criteria (Hildebrand, 1990b).
The chapter treats the various steps in AA, the types and quality of
data for AA, the relationship between AA and other methods of analysis,
and the use of risk assessment in AA. An example analysis, taken from
research conducted in the humid tropical zone of Brazil is used to
illustrate key concepts, and to present the various steps in the analysis
and interpretation of on-farm research-extension data.

Important Steps in Adaptability Analysis

The following is an outline of the major steps in the use of
Adaptability Analysis of on-farm trials. Later in this chapter we will
discuss these steps in more detail, using an example of cowpea
fertilization trials for illustration. There is nothing sacred about the













Adaptability Analysis


number of steps; several could have been grouped together as a single
"step," and several could have been divided even further. In the view
of the authors, however, these are the procedures that should be done in
order to ensure getting the most out of the AA method for the efficient
design, analysis, and interpretation of on-farm research-extension.

1. Conduct the trial in function of planned methods of analysis.
Trials to be analyzed by AA should include collection of data to
adequately characterize each environment and to permit calculation
of all relevant evaluation criteria.
2. Calculate the environmental index, EI. El is calculated as the
mean of the yields (for crops, usually in kg or tons per unit land
area). It is an index that is an estimate of each environment's
capacity to produce the crop or livestock product in question under
the treatments tested.
3. Relate treatment response to environment.
3a. Plot observations (data points) for each treatment against El on
separate graphs.
3b. Estimate treatment response to EI, i.e., the relationship of
each treatment to environment. This can by done by regression
or by drawing a line (straight or curved) by hand.
4. Compare the response of treatments to El and estimate the
treatment-by-environment interaction. Again, this can be done
visually or, if computational capacity is available, by statistical
methods such as that described in Stroup et al. (1993). Visually,
interaction will be indicated by crossover or wide divergence of
response lines.
5. If TxEI interaction is indicated, relate El to environmental
characterization and divide environments into potential
recommendation domains. If no clear relationship can be shown
between El and any of the environmental characteristics on which
data were collected, divide environments based on yields of the
"check," i.e., of farmers' current practices.
6. Interpret results and define recommendation domains.
6a. Within each potential recommendation domain, and for each
evaluation criterion, evaluate the relative performance of the
treatments, both in terms of mean differences and of variability
in the means. Means can be reported by themselves (with
standard errors) or compared by ANOVA, if statistical capacity
is sufficient.












Adaptability Analysis: An Overview


It is important to analyze risks associated with each treatment,
for each of the evaluation criteria. Many introduced or new
technologies with higher mean yields (or higher mean returns)
present more risk of low yields or returns as well. A simple
way to estimate risk is by graphing the distribution of lower
confidence intervals associated with each treatment mean.
6b. Produce multiple extension messages for different groups of
farmers for whom different evaluation criteria are relevant.
NOTE: Steps 3 through 6 should be done from the perspective of
multiple evaluation criteria, i.e., criteria important to
different groups of farmers.


Data Requirements

Sources of Data

Data for the analysis of on-farm research-extension can and should
come from a number of sources; each of these sources can supply
important information for improved design, implementation, analysis,
and interpretation of trials for which the AA method is appropriate.

Sondeos are similar to other types of informal "rapid rural appraisal"
(RRA) or "rapid reconnaissance" surveys. They are conducted by
interdisciplinary teams of researchers and development/extension
personnel, and are designed to be low-cost and of short duration. Their
purpose is to orient the work of on-farm research-extension through
interdisciplinary understanding of the farming systems in a target area,
including preliminary delimitation of research, recommendation, and
diffusion domains, and through identification of key problems and
constraints faced by farmers (Hildebrand, 1981; FSSP 1987). Sondeos
can give valuable preliminary evidence of the variability of environ-
mental conditions faced by farmers, of the types of innovations needed
by and acceptable to farmers, and of the types of criteria important to
farmers in the evaluation of new technologies.
Directed surveys ("focused" sondeos, "topical" rapid rural
appraisals) are a useful tool for exploiting the on-going, iterative nature
of the FSRE approach. Similar to conventional sondeos in their use of
interdisciplinary teams, they involve less time and expense. They are
intended to target a narrow, specific aspect of the farming systems of












Adaptability Analysis


interest, such as a particular crop or other farm enterprise. Focused
sondeos or RRAs have been used, for example, to identify farmers'
evaluation criteria for specific crops in Nicaragua (Betanco, etal., 1990),
and to investigate the traditional grain storage practices in North
Cameroon (Wolfson, 1990). They can be of great value in an on-going
technology development effort by providing increasingly detailed and
complete information on farmers' biophysical and socioeconomic
environmental circumstances.
Enterprise records maintained by farmers with technician help provide
data that can be more accurate than data from surveys based on farmer
recall. Simple forms (see, for example, Shaner et al., 1982, pp. 309-
314) can be developed for traditional farm enterprises. This type of data
serves as ex ante information to help design trials and provides the basis
for analyzing modified technology based on current farming practices.
Enterprise records are also useful to help convert plot data to field-size
situations.
On-station trials often furnish the technologies to be tested under
farmers' conditions before they are selected for dissemination to a larger
population of farmers. Socioeconomic analyses of on-station trial results
are often lacking, leading to on-farm trials of technologies which are
clearly inappropriate to farmers' circumstances. Sufficient analysis of
experiment station trials should be done to allow ex ante socioeconomic
analyses of proposed on-farm trial technologies before such trials are put
in farmers' fields.
On-station trials can also furnish data on additional environments and
can be incorporated into on-farm trial data if a "farmers' control"
treatment has been included in the on-station trials and if sufficient data
are available for environmental characterization. To use on-station data,
those treatments that are used in the on-farm trials, and the "farmers
control" are extracted from the on-station data. Each year, and even
each block can be considered as additional environments and added to the
on-farm data.
On-farm trials, particularly single-replicate, farmer-managed on-farm
trials are the usual source of biological and socioeconomic data
susceptible to full exploitation by Adaptability Analysis. For the most
efficient use of trial data, it is imperative that sufficient characterization,
both biophysical and socioeconomic, of each on-farm trial environment
be done. Deciding just how much and what kinds of data need to be
collected is often not immediately evident; making these decisions is
itself an iterative process. Over time, and through trial and error,












Adaptability Analysis: An Overview


researchers and extensionists should establish a minimum data set for on-
farm trials in their research domains.

Quality of Data

Data from the on-farm trials do not require more than one replication
per environment to be amenable to Adaptability Analysis. However,
there are some characteristics of the data which improve its quality and
reliability (Stroup et al., 1993). The number of environments required
for estimation of the significance of treatment-by-environment (TxEI)
interaction should result in at least 20 degrees of freedom in the error
term of ANOVA. Allowing for estimation of both linear and quadratic
responses of the treatments to EI, for four treatments, ten environments
would be required. In a verification trial with only two treatments (for
example, the recommended treatment and the farmer check) about 23
environments is adequate.
These suggestions, of course, are approximate. The appropriate
number of environments is a function of the variance and the required
sensitivity-all case-by-case situations. To allow for adequate verification
of treatment performance within two or three potential recommendation
domains, a minimum of 15 to 20 environments should be aimed for.
Experience in analyzing numerous data sets has indicated that if three
conditions are met in a single year's trial, the estimates of environment-
by-treatment response will be consistent across years. These conditions
(adapted from Stroup et al., 1993, p. 172) are:

1. The range of the environmental index (EI) should be at least as
great as the mean of the index values, i.e.; the ratio of range to
mean should be at least one.
2. The distribution of environments (Els) should be reasonably
uniform from good to poor.
3. The range and distribution of the yields of farmers' current
practices should approximate those normally expected over a
period of years.

It may well be that the distribution of environments is more important
than number so long as some minimum number (again, around 15 to 20)
is achieved. Dividing the range of Els by the overall mean El is a useful
measure of the representativeness of the data. On-farm data usually have
greater range than station trials, even if the station trials are reported












Adaptability Analysis


over a number of years. A quick search resulted in locating the
following ratios:

1. An on-farm potato study (reported in Caldwell and Taylor, 1987,
p. 353) on agronomic practices from 20 farms by the International
Potato Center in the Mantaro Valley of Peru: 1.42/1.
2. An on-farm maize variety study (Poey n.d., p. 58) conducted by
the farming systems program in Paraguay on 24 farms: 1.21/1.
3. An on-farm maize variety and fertilizer use study (Hansen et al.,
1982) conducted on 14 farms by the farming systems team in
Malawi: 1.57/1.
4. The multilocation, on-station, Yates and Cochran study, when
years are combined (as the authors did): 0.55/1; or 0.83/1 when
years and locations are separate.
(Hildebrand 1990a, p. 177)

A very narrow range, resulting in ratios less than one as in the Yates and
Cochran (1938) study, usually indicates the mean yields of the trials were
very high because they were conducted under controlled conditions. It
can mean that only the best farms were selected, or that the multiple
environments represent experiment stations in different areas. Although
these ratios are taken from data resulting from the trials, prior attention
to designing the trial to span a considerable environmental range pays
dividends.


An Example of AA from Brazil

The results of an on-farm trial of phosphate fertilization of cowpea
near Manaus, in the humid tropical zone of Brazil (Singh, 1990; Stroup
et al., 1993), are used below to illustrate AA. The government of Brazil
was supporting colonization in the area and was concerned about the loss
of rain forest. The problem this research addressed was the rapid loss
of fertility, particularly phosphorus, in land cleared for cropping. The
research domain chosen was a colonization area along the Rio Preto da
Eva, some two hours by boat upstream from the market town of Rio
Preto da Eva, which was on a paved road connecting to Manaus.
Primary information was collected by means of a sondeo involving
persons from the Brazilian research and extension agencies and a local
organization. Farmers' knowledge of indigenous technology, agronomic












Adaptability Analysis: An Overview


practices, and land types being used were recorded. An extensive soil
sampling program was carried out to understand soil physical and
chemical characteristics and relate them to farmers' rationale for
assigning a particular cropping pattern to a given land type. Classified
by land use type, some were from primary forest farmed for one year,
others from primary forest farmed for two or for three years, and still
others from secondary forest farmed for one, two or three years; one
field was even classified as waste land.Constraints to maintaining fertility
included distance to market and the availability of cash for purchasing
and transporting inputs. Thirteen on-farm trial sites were selected in the
research domain and four fertilization treatments were applied in a
randomized complete block design with only a single replicate of the
treatments in each environment. The four treatments, based on previous
on-station research were selected:

1) the farmers' local practices (FP), which had essentially no or very
low levels of fertilization, i.e., 0-4 kg/ha P;
2) a "full dose" of triple superphosphate (TSP) plus potassium
containing roughly 0-18-60 kg/ha N-P-K;
3) chicken manure (CM) plus a "half dose" of triple superphosphate,
equivalent to the same total amounts of P and K as the TSP
treatment, but with half the P coming from chicken manure;
4) processed city waste (PCW) plus a half dose of triple superphos-
phate, again with the same levels of P and K, but with half the P
coming from the city waste.

Relating Treatment Response to Environment by Regression on El

Once an on-farm trial is conducted within a given research domain,
relevant evaluation criteria identified, and data necessary for adequate
environmental characterization collected (step 1 on page 30), analysis
begins by plotting the yield of each treatment on the environmental
index, (El). A best fit response line is found for each treatment across
Els, either by simple visual estimation or by determination of the best
regression model (usually either linear or quadratic).
Based on the relative responses across environments, important
treatment-by-environment interactions may be identified. These would
indicate the possible existence within the research domain, and for yield
as a criterion, of more than one recommendation domain for the tested
technologies. Finally El, which can only be known after the data.are













Adaptability Analysis


collected, should be related to the environmental characterization
variables collected in order to allow grouping of farms into one or
another recommendation domain.
Yield data for this trial (t/ha) are in Table 2.1. The yields of all four
treatments plotted on the environmental index are shown in Figure 2.la.
The data show a considerable amount of variability (Cvs were around
25%), so that it is difficult to judge the structure of the response of any
one treatment across Els, due to the overlap of observations from several
treatments. For this reason, it is always advisable to plot the observa-
tions of each treatment individually on El. The yields for the PCW
treatment, for example, are plotted on El separately in Figure 2.1b.
From simple visual examination, the response of PCW across
environments seems to be linear.
In Figure 2.2a, the response of TSP appears also to be linear, while
the response of CM, although having a marked linear component (one
with a slope apparently around zero), seems also to have an upward
"bow" to it, hinting at a possible higher order (e.g., quadratic)
component (Figure 2.2b).
The response of the farmers' local practice (FP) seems to have a
similar, but downward, bow to it, again indicating something more than
a purely linear response to environment (Figure 2.3a). Figure 2.3b
illustrates two alternative regression models to describe the yield
response of the farmers' practices across environments (i.e., across El).
Before evaluating which regression better fits the data, a few words on
the use of the R= statistic are called for here. The R= statistic is directly
dependent, in single-order linear regression models, on the slope of the
regression line. This can be seen from the formula for calculating R= in
such a model:
R2= (b.) (Exy) / Ey

Thus, for two simple linear regression lines with the same deviations
from regression (which is perhaps a better measure of true "goodness of
fit"), the line with the greater slope will have the greater R2. For this
reason, R2 is not of much intrinsic value for comparing regression
lines of different treatments in an analysis. It is, however, a good
statistic to help decide, for a single treatment, which of two or more
alternative regression models (e.g., linear or quadratic) is the best.
Just as some sort of curved response across El tentatively can be
judged to be better than a simple straight-line response by visually


36 "
















TABLE 2.1 Sample spreadsheet, Brazil cowpea on-farm trial data, and regression estimates


TONS/HA*


LOC FP PCW TSP


0.10 0.20
0.00 0.00
0.15 0.50
0.20 0.40
0.50 0.65
0.15 0.50
0.60 1.20
0.70 0.90
1.20 1.50
1.50 1.80
1.45 1.95
2.20 1.90
1.70 1.65


EST. LINEAR REGRESSION VALUES


CM El EP FP PCW TSP


1.30 1.65
1.30 2.00
1.35 1.35
1.20 1.70
1.10 1.50
2.10 2.05
1.60 2.25
2.30 1.80
2.20 1.90
2.10 1.70
2.50 1.90
2.60 1.40
2.65 2.15


0.813 0.660 0.018
0.825 0.681 0.035
0.838 0.701 0.053
0.875 0.766 0.106
0.938 0.879 0.194
1.200 1.440 0.564
1.413 1.995 0.863
1.425 2.031 0.881
1.700 2.890 1.269
1.775 3.151 1.375
1.950 3.803 1.621
2.025 4.101 1.727
2.038 4.151 1.745


0.254
0.271
0.288
0.339
0.424
0.780
1.069
1.086
1.459
1.561
1.799
1.900
1.917


1.259
1.273
1.286
1.327
1.396
1.683
1.916
1.929
2.230
2.312
2.504
2.586
2.600


CM

1.719
1.721
1.723
1.728
1.736
1.773
1.802
1.804
1.842
1.852
1.876
1.887
1.888


EST QUAD REGR.

FP CM

0.139 1.627
0.142 1.639
0.147 1.651 0
0.161 1.686
0.190 1.739
0.395 1.901
0.656 1.960
0.674 1.961
1.145 1.936
1.299 1.910
1.698 1.818
1.886 1.765
1.919 1.756















TABLE 2.1 (Continued)


PP Regression
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


Output:
-1.12808
0.246645
0.894074
13
11


1.409967
0.146328


TSP Regression Output:
Constant 0.369682
Std Err of Y Est 0.227601
R Squared 0.856565
No. of Observations 13
Degrees of Freedom 11


X Coefficient(s)
Std Err of Coef.


1.094407
0.13503


PCW Regression
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


CM Regression
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


Output:
-0.84858
0.204428
0.919291
13
11


1.35756
0.121282


Output:
1.606977
0.284149
0.057473
13
11


0.138066
0.168579


FP QUAD Regression Output:
Constant 0.522392
Std Err of Y Est 0.208385
R Squared 0.931261
No. of Observations 13
Degrees of Freedom 10

X Coefficient(s) -1.24034 0.945128
Std Err of Coef. 1.146142 0.406341


CM QUAD Regression Output:
Constant 0.350798
Std Err of Y Est 0.274238
R Squared 0.201893
No. of Observations 13
Degrees of Freedom 10

X Coefficient(s) 2.155225 -0.71934
Std Err of Coef. 1.508339 0.534751


* FP farmers' practices; PCW
El = environmental index.


= processed city waste; TSP = triple superphosphate; CM = chicken manure;













Adaptability Analysis: An Overview 39

3
2 I-----------------------


AA a 1 0




1 -
*


0. 0

0 I n o I I I
0 0. 1 1. 2 25
ENVIRONMENTAL INDEX, El





3













a I gO I( I I
S0 1 2 2





ENVIRONMENTAL NDEX, E
0 *



1 0

*



0 0.5 1 1.5 2 2.5
ENVIRONMENTAL INDEX, El





FIGURE 2.1 Yields of all treatments (a) and of PCW alone (b), plotted
on EI.













40



3




2





01
0.


Adaptability Analysis


0 0.5 1 1.5 2 2
ENVIRONMENTAL INDEX, El







b


A A
2- A A A A

AA A


1
5 -




0 I
0 0.5 1 1.5 2 2.5
ENVIRONMENTAL INDEX, El


FIGURE 2.2 Yields of TSP (a) and CM (b), plotted on EI.


a

U


U U


" '


i I













Adaptability Analysis: An Overview


3

2.

< 2

91.5

1

0.5

0


a

4

*
*






___ I I I
0.5 1 1. 2
ENVIRONMENTAL INDEX, El

F l


0. 1 1I 2
ENVIRONMENTAL INDEX, El

OBSERVED NEAR QUAIDAIC
Belong


FIGURE 2.3 Yields of FP plotted on El (a) and comparison fit of linear
and quadratic regression of FP on El (b).











Adaptability Analysis


examining the scatter plot of observations, as for CM and FP (Figures
2.2 and 2.3), so also can the best regression fit often be identified
visually. In the case of farmers' practice in the Brazil data, for example,
quadratic regression results in only a 4% improvement in R2 over linear
regression (0.93 vs. 0.89). This is all very well to be able to calculate
and know, but it adds little not told by simple visual comparison of the
linear and quadratic regression lines superimposed on the observations
of FP plotted on El (Figure 2.3b). In the same way, the quadratic
regression of CM (R2 = 0.20) is an improvement over the linear (R2=
0.05).
Note that the R2 for the CM linear treatment is much lower than for
the others; this is due in large part to the fact that the linear slope is so
near zero, as explained above. (See Table 2.1 for regression statistics).
The question of linear versus quadratic equations to describe the
response of each treatment on El is of some intrinsic interest, but is only
crucial if the resulting graphical picture of the TxEI interaction is
materially different (in terms of cross-overs, etc.). Comparing the graph
of linear regression lines for all four treatments (Figure 2.4a) to that of
linear regression lines for TSP and PCW and quadratic lines for FP and
CM (Figure 2.4b), it is apparent that while the "structure" of the
interaction appears quite different, the practical interpretation of the
graphs for the delineation of recommendation domains would be
essentially the same. For both graphs, the cross-over of CM and TSP
occurs at about El = 1.3, and in both graphs TSP is the superior
treatment in environments with El > 1.3 while CM is the best treatment
in environments with El < 1.3. In this case, then, both the linear and
quadratic regressions result in essentially the same interpretation for
deciding upon two potential recommendation domains, assuming that
yield per hectare is the evaluation criterion of interest. Based on the
criteria for data quality, p. 33, although the El range to El mean ratio is
a bit low (0.89), the distribution and overall range of Els is quite good,
so one could expect the relationships shown in Figure 2.4 to be
reasonably consistent over time.

Comparison of Results from AA Regression and from ANOVA

One of the first things to note from the discussion above is how much
more one learns from AA than from a traditional analysis of variance.
The simple analysis of variance for this trial, which has only one
replication per test site (i.e., per environment), is presented in Table 2.2.













Adaptability Analysis: An Overview


1 15
ENVIRONMENTAL INDEX, EI
PP Pcw TSP CE I
100- OE s


1 13.5
ENVIRONMENTAL INDEX, El
PP PCW 551 CM E
M S In sem.ns A


FIGURE 2.4 Regressions of on-farm cowpea fertilization treatments on
EI: linear for all treatments (a), and linear for TSP and PCW with
quadratic for FP and CM (b).











Adaptability Analysis


TABLE 2.2 Simple ANOVA for yield over all locations, Brazil cowpea
data

Source of Variation df Prob. ofF


Environment 12 0.00001
Treatment 3 0.00001
Residual 37



This analysis tells us that there exist some important differences in the
mean yields among the four treatments and among the environments
(which in this ANOVA serve as "blocks"). However, had this ANOVA
indicated no significant treatment effects, it could have been due to
masking by a treatment-by-environment interaction, which is not shown
in this table and cannot be estimated by simple ANOVA (see p. 17).
Because the main effect of treatment is significant, the next step is to
examine the differences among treatments. Some group of contrasts
might be of interest, but the usual (if less statistically defensible) method
is multiple mean comparisons, using a test such as LSD. or Duncan's
Multiple Range Test. Standard mean separations, again assuming only
one recommendation domain, presented in Table 2.3 indicate that in
terms of yield the CM and TSP treatments were essentially equal and
were superior to PCW, which was not significantly different (at a =
0.05) from the farmers' local practices. So far so good, but the fields
on which the trial was conducted were very different (Table 2.4).
Is it likely that cowpea yield in all of these diverse environments,
given the rapid decline in soil fertility known to be a problem in rain
forest land put into cultivation, would respond to fertilization in exactly
the same way from one farm to another? Would one not expect some
sort of treatment-by-environment interaction across these very different
environments? Analysis of variance including the interaction of
treatment and the linear effect of El (Stroup et al., 1993) indicates that
there is in fact a highly significant TxEI interaction that can be explained
by linear regression on El (Table 2.5).











Adaptability Analysis: An Overview


TABLE 2.3 Mean separation of yields, over all locations, Brazil
cowpea data

Teatmentt Mean Field (t/ha)

TSP 1.87 a
CM 1.80 a
PCW 1.01 b
FP 0.80 b

t Means in a column followed by the same letter are not significantly different
at the 5% level of probability, according to Fisher's Protected Least Squares
Difference test.
1 TSP = Full dose of triple superphosphate ; CM = chicken manure + half
dose of TSP; PCW = Processed city waste + half dose TSP; FP = Farmers'
practices.


Note that the significance of the TxEI interaction term in ANOVA tells
us only that the slope of at least one regression line is different from one
or more of the other lines. To see the structure of the interaction, it is
more useful to examine the graph of the regression lines (Figure 2.4).
Notice also that the ANOVA conducted to this point, used only to
illustrate the added value and verify the "significance" of what linear
regression on El shows us, can be dispensed with if capacity does not
exist to do it.

Relationship between El and Environmental Characteristics

Environments can be characterized using both biophysical and
socioeconomic factors that may be at the same time both quantitative and
qualitative in nature. Data obtained for characterizing the environments
in the Amazon example include soils characteristics and a category called
"land type", Table 2.4. The soils characteristics are self explanatory.
Land type refers to the kind of forest that was cleared (P = primary, S
= secondary) and the number of years it has been cropped (1 = first
year, etc.). The term WL refers to land that had been cleared by
bulldozer at the time of colonization and was, essentially, waste land.
Because the data in Table 2.4 have been sorted by EI, it is easy to
assess the relationship between El and these characteristics. Lower Els
















TABLE 2.4 Characterization of environments, Brazil cowpea on-farm trial data


OBSERVATIONS


FARM El LAND TYPE pH ECEC* AISATt P


5.2 3.24 59.8
5.4 2.21 61.2
5.3 1.25 80.1
4.9 1.91 63.2
5.0 1.72 64.0
4.7 1.35 74.1
5.1 2.31 86.6
4.9 0.99 82.3
4.6 2.34 94.9
4.3 1.20 94.5
4.6 1.66 94.2
4.3 1.94 83.4
4.3 1.83 87.5


8.0
12.9
5.0
7.0
10.6
7.6
4.0
2.3
3.4
2.0
4.8
6.1
0.1


REGRESSION ESTIMATES

pH ECEC AL SAT P

1.818 1.729 1.978 1.598
2.050 1.466 1.933 2.075
1.934 1.221 1.335 1.306
1.470 1.390 1.870 1.501
1.586 1.341 1.845 1.851
1.239 1.247 1.525 1.559
1.702 1.491 1.129 1.209
1.470 1.155 1.265 1.044
1.123 1.499 0.866 1.151
0.775 1.209 0.878 1.014
1.123 1.326 0.888 1.287
0.775 1.397 1.230 1.413
0.775 1.369 1.100 0.830


(cmol+ charge/kg soil) effective cation exchange capacity
g/100 g aluminum saturation
Fg/g Mehlich I extractable phosphorus


2.04
2.03
1.95
1.78
1.70
1.43
1.41
1.20
0.94
0.88
0.84
0.83
0.81


PF 1
PF 1
PF2
PF 2
PF 2
SF 1
PF 2
PF 3
SF 2
SF 3
WL
SF 2
SF 3
















TABLE 2.4 (Continued)


)utput:


-4.20945
0.20941
0.83033
13
11


pH Regression C
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom

X Coefficient(s)
Std Err of Coef.


AL Regression Output:
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


1.159151
0.157987


3.872471
0.263772
0.730804
13
11


-0.03168
0.005798


ECEC Regression Output:
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


P Regression Output:
Constant
Std Err of Y Est
R Squared
No. of Observations
Degrees of Freedom


X Coefficient(s)
Std Err of Coef.


0.902806
0.482468
0.099367
13
11


0.254845
0.23133


0.819771
0.356324
0.508751
13
11


0.09733
0.028837











Adaptability Analysis


TABLE 2.5 ANOVA of yield over all 13 locations (including the
interaction between treatment and the linear effect of EI), Brazil cowpea
data

Source of Variation df Prob. ofF

Environment 12 0.00001
Treatment 3 0.00001
Treatment x El (linear) 3 0.00001
Residual 34



(lower-yielding environments) are associated with lower pHs, lower
phosphorus levels, lower ECECs and higher aluminum saturation. If
desired, these relationships can also be graphed and/or estimated by
regression with El being the dependent variable as was done in Figures
2.5 and 2.6 for land types, aluminum saturation, pH and phosphorous.
In this case, the most useful for farmers and extension agents is the land
type characteristic, because farmers in these conditions seldom, if ever,
have detailed soil information on their fields. It can be seen that both the
nature of the forest that was cleared for the field and the number of years
in use are closely associated with El.
For cowpeas in the Manaus trial, it can be seen that the higher-yielding
environments (those with El > 1.3) correspond to fields taken from
primary forest and in first or second year of use, or from secondary
forest and in first year of use. For these situations, if t/ha is the relevant
criterion, the recommendation would be to use TSP, the highest-yielding
treatment for these environments (Figure 2.4). For all other fields, if
cowpea is to be grown and t/ha is the relevant criterion, chicken manure
(CM) would be the amendment recommended. Thus, in this research
domain, two recommendation domains can be tentatively delineated based
on the criterion of t/ha and described in terms usable by both farmers
and extension personnel.
If those environmental characteristics that were collected show no clear
relationship to the Els, the regression of the treatment representing
farmers' current practices (or local varieties, etc.) on El can be used as
a convenient substitute for defining recommendation domains. In Figure
2.4b, for example, the yield of FP corresponding to the cross-over of












Adaptability Analysis: An Overview


LAND TYPE


S2


1.6


1.2

0.
1 0.8


55 0 65 70 75 0 is 9o0 95 100
AI SATURATION
OBSERVED R SQ .7i5



FIGURE 2.5 El plotted on land type (a) and El regressed on aluminum
saturation (b).












Adaptability Analysis


4.5J
PH
OBSERVED R SQ .7


0 2 4 6 1 10 12 14
AVAILABLE P
OBSERVED R SQ .51



FIGURE 2.6 Regression of El on available pH (a) and on phosphorous
(b).


F-


2.4





1.6





O.1


1.6











Adaptability Analysis: An Overview


TSP and CM (where El equals approximately 1.3) is approximately 0.7
t/ha. Therefore the recommendation for farmers who would anticipate
a yield less than 0.7 t/ha would be CM. For those anticipating a cowpea
yield above 0.7 t/ha with their current practices, the recommendation
would be TSP.

Analysis within Tentative Recommendation Domains (Verification)

Having delineated and characterized two tentative recommendation
domains for the criterion t/ha, and by regressing treatment responses to
El, further analysis can be undertaken to help assure that the apparent
differences of treatments within recommendation domains are real. One
means is with analysis of variance and mean separation. Another is with
risk analysis based on a distribution of confidence intervals.

ANOVA or Simple Mean Comparisons. The question addressed by
use of ANOVA to "verify" the existence of different recommendation
domains is the level of statistical significance of treatment differences
within each tentative recommendation domain. In the example, in the
domain of higher-yielding environments (PF,, SF,, and PF2), seven
observations were included in the trial, and in the lower-yielding domain
(PF3, SF2, SF,, WL) there were six. The combined ANOVA across the
two domains is shown in Table 2.6. Combined ANOVA across domains
shows that there are overall differences among treatments (which we
knew from the simple ANOVA in Table 2.2). It also shows that there
is not just a significant TxEI interaction (indicated in Table 2.5), but also
a highly significant treatment-by-domain interaction.
The important term in this analysis of variance is the treatment-by-
domain interaction. This interaction term tells us that the treatment-by-
environment interaction shown but not tested statistically by the
regressions on El in Figure 2.4 is "real." That is, there is very little
probability that the differences in slopes of the regression lines in Figure
2.4 are due just to chance. Note that the advantage of ANOVA is that
it allows one to attach a probability to the estimate of treatment-by-
domain (TxD) interaction, which is related to but different from the
treatment-by-environment (TxEI) interaction shown by regression on El.
The major difference between TxEI and TxD is that El cannot be used
ex ante by farmers in making decisions, while the characteristics which
define domains can.











52 Adaptability Analysis

TABLE 2.6 Combined ANOVA, over tentative land type
recommendation domains, Brazil cowpea data


Source of Variation df Prob. ofF

Domain 1 0.0001
Error a 11 0.0807
Treatment 3 0.0001
Treatment x domain 3 0.0008
Residual error 33

t High-yielding recommendation domain (EI > 1.3) PF1, PF2, SF1,
Low-yielding recommendation domain (El < 1.3) PF3, SF2, SF3, WL


The nature and importance (not just statistical significance ) of the TxD
interaction is shown by analyzing treatment mean differences within each
domain (Table 2.7). Note that if the traditional 0.05 level of probability
is used for determining "significance" of the mean differences, one might
be tempted to doubt the superiority of TSP over CM in the high-yielding
domain and of CM over TSP in the low-yielding domain, relationships
hinted at by the regressions (Figure 2.4). The differences are statistically
different, however, at the 0.10 level of probability. If they had not been
significant at this level, they might have been at 0.20 or 0.30. When
doing ANOVA or means separations as part of AA, one should not
decide whether an effect is "significant" or not at some arbitrary cut-off
point, no matter how supported by convention that cut-off is. Rather,
one should determine at what level the effect is significant, then decide
whether or not that level of significance is important.
Again, although useful in verifying the recommendation domains,
given the difficulties of combined ANOVA, it is not absolutely necessary
as part of AA, if computing facilities are not available. One should
calculate and compare the means themselves. If statistical mean
separation procedures are not available, or if they can only calculate
LSDs or other statistics at the 5% level, reporting the standard errors or
standard deviations should suffice to give an estimate of the confidence
one has in the mean differences within each domain. The means and
standard errors (or standard deviations) are easily calculated, using
whatever tools are available for calculating regressions on EI.












Adaptability Analysis: An Overview


TABLE 2.7 Mean separation of yields, by tentative land type
recommendation domain, Brazil cowpea data

Mean Yield (t/ha)
Treatment* PF1. PF2. SF1 PF3. SF2. SF3. WL
mean std a= a= mean std a= a=
error 0.05 0.10 error 0.05 0.10

TSP 2.28 0.137 a a 1.39 0.146 a b
CM 1.87 0.107 ab b 1.71 0.112 a a
PCW 1.56 0.147 be bc 0.35 0.096 b c
FP 1.34 0.212 c c 0.19 0.069 b c

t Means in a column followed by the same letter are not significantly different
at the 5% or 10% level of probability, by Fisher's Protected LSD test.
t TSP = Full dose of triple superphosphate ; CM = chicken manure + half
dose of TSP; PCW = Processed city waste + half dose TSP; FP = Farmers'
practices.


Risk Assessment. Reporting of means and their variability by potential
recommendation domain should be done in every case. Then, even if
combined analysis of variance or mean separations cannot be done
(which will often be the case if statistical knowledge or computing
facilities are limited), the relative performance of technologies within
specific environmental circumstances, i.e., within domains, can and
should still be further analyzed.
One of the most important analyses to be done once recommendation
domains are tentatively delineated, is analysis of risk to farmers. This
is particularly critical in the lower-yielding recommendation domains.
Minimum levels of performance acceptable to the farmers should be
determined, either through diagnostic activities such as sondeos or formal
surveys of farmers' situations, or through interviews with the farmers
collaborating in the on-farm trialss. In many cases, farmers have, and
will express, some minimum level of performance they must have from
a given crop or animal enterprise. In some circumstances a minimum
yield of a staple food is required merely for the household to survive; in
other circumstances, enough production to ensure family survival plus
income sufficient to school some children is the minimum farmers
expect. It is also not uncommon for farmers to consider how well their











Adaptability Analysis


crops and livestock produce as an indication of their worth, competence,
and standing in the community. Poor performance or outright failure can
be a source of shame to farmers, even if their families survive the
situation. While such motivations are sometimes ignored or minimized
by researchers, their influence in determining the adoption decisions of
farmers is often great. Technologies which on average produce much
above this minimum threshold but which, due to great variability of
performance, have a considerable chance of falling below it, will not be
as acceptable to such farmers as an alternative technology which may on
average produce only somewhat more than the minimum acceptable
level, but which has very little chance of falling below it.
A relatively simple evaluation of the comparative risk of alternative
technologies can be done by calculating and graphing a distribution of
lower confidence limits for the mean of each technology. This should be
done by tentative recommendation domain. Because the consequences of
poor performance are greater in the lower recommendation domain(s),
it is particularly critical to assess risk for the conditions found in this
domain. To calculate distribution of confidence intervals, the
observations only from within the tentative recommendation domain are
used. The sample standard deviation of the data from which the domain
mean is calculated is multiplied by the one-tailed t value corresponding
to the level of probability, and the resulting quantity is subtracted from
the mean:

Lower Confidence Limit = y [(t... ,)) (s) / In]
where y = the treatment mean of observations within
the tentative recommendation domain,
sj = the sample standard deviation associated
with the mean,
n = the number of observations that went into
the calculation of the mean (that is, within
the tentative recommendation domain),
and
p = the probability level (from a one-tailed t
table because interest is only in values
lower than the means).

A table is constructed for the lower confidence limits of the mean at a
number of levels of probability, and the series of lower confidence limits












Adaptability Analysis: An Overview


for each mean is graphed. For the Manaus, Brazil cowpea data for
example, the low-yielding (El < 1.3) recommendation domain, using
t/ha as the evaluation criterion, comprised 6 environments; the mean of
each treatment, then, was derived from 6 observations; therefore n=6.
Table 2.8 presents the mean, sample standard deviation, alpha levels, t
values (and the risk associated with the t probabilities), and the lower
confidence limits for TSP, for this lower-yielding domain. A
comparison of the risks associated with TSP and CM in the lower-
yielding domain is shown in Figure 2.7. Based on this evidence it is
quite clear that CM results in less risk of low yields (t/ha) in the lower-
yielding domain than does TSP. Note from Table 2.7 that the reduced
risk is due to both high mean yield in this domain and less variability
associated with the mean, compared to TSP.


TABLE 2.8 Calculation of lower confidence limits (t/ha) for TSP, low-
yielding domain (PF3, SF,, SF3, WL), Manaus, Brazil cowpea
fertilization trial (n=6, mean = 1.392, Sd = 0.358)

a tw-s Risk(%) Lower Confidence Limit

.250 0.727 25 1.392-(0.358*0.727A/6) =1.285
.200 0.920 20 1.392-(0.358*0.920A/6) 1.257
.150 1.156 15 1.392-(0.358*1.156/V6) =1.223
.100 1.476 10 etc. =1.176
.050 0.015 5 etc. -1.097
.025 3.571 2.5 etc. =1.016
.010 3.365 1 etc. =0.899
.005 4.032 0.5 etc. =0.802
.0005 6.859 0.05 etc. =0.388

Note: The complete table is reproduced in Table 2.9. It should be noted that
the t value should be taken from a one-tailed I table, the values will correspond
to the probability of a larger negative value of t. This probability, expressed as
a percentage, is the column labeled "risk."


In the Manaus study, no estimate of a minimum acceptable level of
cowpea yield was made. For purposes of illustration, we will use an
estimate of 1.2 t/ha as this minimum acceptable level. From Figure 2.7











Adaptability Analysis


0
0 5 10 15 20 25
RISK (% CHANCE OF A LOWER VALUE)
PP TVP CM


FIGURE 2.7 Risk assessment (lower confidence limits), for t/ha; on-
farm cowpea fertilization treatments in low-yielding recommendation
domain (PF3, SF,, SF3, WL).


it can be seen that if TSP is applied, the risk is approximately 10%
(corresponding roughly to one year in ten) that yield would fall below
the minimum acceptable level of 1.2 t/ha. If CM is applied to the
cowpea in this recommendation domain, there is almost no chance
(< 1%, or less than roughly one year in 100) that yield would fall below
the minimum acceptable level. Hence, there is much less risk to the
farmer if CM is recommended than if TSP is the recommendation in the
lower-yielding environments and for the criterion of t/ha. The risk of
low yields from the farmers' practices is very high (certain). This is
reflective of the common practice of cropping at most two years in these
conditions. Notice that this risk analysis adds appreciably to what was
learned by comparing the means (Table 2.7). It allows formulation of
estimates of risk in terms useful to fanners and extensionists. The next
section looks at the effect on recommendation domains when alternative
evaluation criteria are used.











Adaptability Analysis: An Overview


TABLE 2.9 Sample spreadsheet, for confidence limits t/ha), low-
yielding domain (PF3, SF,, SF3, WL), Manaus, Brazil data

Data (t/ha)
Location FP PCW TSP CM El

8 0.10 0.20 1.30 1.65 0.813
13 0.00 0.00 1.30 2.00 0.825
6 0.15 0.50 1.35 1.35 0.838
9 0.20 0.40 1.20 1.70 0.875
2 0.50 0.65 1.10 1.50 0.938
5 0.15 0.50 2.10 2.05 1.200

Domain average 0.183 0.375 1.392 1.708
Domain sample st. dev. 0.169 0.236 0.358 0.275

Prob. of Lower Confidence
Lower Value Limit, t/ha
Alpha (Risk) te., FM CM TSP

0.2500 25.00 0.727 0.133 1.627 1.285
0.2000 20.00 0.920 0.120 1.605 1.257
0.1500 15.00 1.156 0.103 1.579 1.223
0.1000 10.00 1.476 0.081 1.543 1.176
0.0500 5.00 2.015 0.044 1.482 1.097
0.0250 2.50 2.571 0.006 1.420 1.016
0.0100 1.00 3.365 1.331 0.899
0.0050 0.50 4.032 1.256 0.802
0.0005 0.05 6.859 -0.939 0.388


Use of Alternative Evaluation Criteria

The evaluation criterion used to calculate the environmental index EI,
above, is t/ha, the most common criterion used by agronomists in crop
trials and, because it reflects the impact of environment, appropriate in
most cases as the basis for calculating the El. However, few farmers
use this criterion when making production decisions. If either seed,
labor or cash is their scarcest resource, a more appropriate evaluation












Adaptability Analysis


criterion is kg/kg seed, kg/day of labor in a critical period, or kg/dollar
of cash cost, respectively. AA easily lends itself to analysis using
multiple criteria. Figure 2.8 is based on analysis of the cowpea data for
the common farmers' criterion of kg/$ cash cost.
Notice that the same El (usually based on kg/ha yields) is used
regardless of the criterion being evaluated. The EI values used to form
the X-axis for the regression of treatments on El (Figure 2.8a) do not
change. The criteria used on the Y-axis (Figure 2.8a) do change. The
same procedures were used to obtain these relationships as were used to
obtain the relationships based on the researchers' criterion, t/ha. Cash
costs of the treatments were FP = $12, PCW = $208, TSP = $98 and
CM = $127 per hectare. Notice that very different conclusions can
result when the evaluation criteria change. This is important because it
relates to the ultimate recommendations that will be made.
For the farmers' criterion of kg/$ cash cost, the farmers' practices,
without soil amendments, is the most appropriate technology of the four
tested in the higher-yielding environments (again, El > 1.3). Although
both TSP and CM yield more in terms of t/ha, their cost (purchase and
transport) is so high that the return per dollar invested is too low
compared with their current practices. However, in the lower-yielding
domain, iffarmers want or need to produce cowpea in these conditions,
then CM is an appropriate recommendation if achieving 16 kg per dollar
of cash expense three out of four times is acceptable (Figure 2.8b). This
level of production would be equivalent to a cost of approximately 6.25
cents per kg cowpea at the farm. If 14 kg per dollar of cash expense
(7.14 cents per kg) is acceptable, this can be achieved approximately 49
out of 50 times using CM as an amendment in the lower-yielding
environments (Figure 2.8b).
Had characterization of the environments not shown any relation with
recommendation domains, the regression of the treatment representing
farmers' practices (Figure 2.8a) could again be used instead. In this
case, for farmers anticipating yields greater than 20 kg/$ cash cost, the
recommendation would be their own practice. For those expecting lower
yields per $ cash cost, the recommendation would be CM.

Multiple Extension Recommendations

The above analyses can be summarized in a series of extension
recommendations that combine environmental characteristics with specific
evaluation criteria, Table 2.10.











Adaptability Analysis: An Overview


40


loe- mo .. -OW --- - a -
05 1 1. 2
ENVIRONMENTAL INDEX, El
FP Pcw TSP cK M
m ~ usual -


S 10 15 20
RISK (% CHANCE OF A LOWER VALUE)

| 1"


FIGURE 2.8 Estimated responses of the four treatments to environment
(EI) for cowpea in Brazil using a common farmer criterion, kg/$ cash
cost (a), and risk assessment (lower confidence limits) for kg $1 cash
cost in low-yielding recommendation domain (PF,, SF2, SF, and WL)
(b).












Adaptability Analysis


TABLE 2.10 Multiple extension recommendations for four
recommendation domains based on land type and two evaluation criteria,
cowpea near Manaus, Brazil

Land Type
Citerion PF,, PF2, SF, PF,, SF,, SF,, WL

t/ha TSP CM
kg/$ cash cost FP CM



Farmers in the conditions represented by these data from the Amazon
rain forest usually fell the forest from a patch of land, farm it in annual
crops one or two years, then let it return to fallow (perhaps with a
continuing cassava and/or tree crop). These farmers live in isolated
conditions and have little contact with the market. Cash is scarce but
generally land is not. For these farmers, the criterion of kg grain per $
of cash cost is most relevant. Table 2.10 indicates that these farmers
would not invest in the TSP or CM because of its cost even though both
produce more yield measured in t/ha and so long as they have land that
can be taken out of forest and farmed only one or two years. Their best
alternative is to fell another piece of forest. If, however, forest land
becomes scarce and/or the government prohibits the felling of more
forest, then after the second year following primary forest, and after the
first year following secondary forest, CM would be the alternative of
those tested to recommend to them. Because of its cost, TSP would only
be recommended if t/ha were the criterion and then only for the first or
second year of use.


Summary

On-farm trials are increasingly common. Sadly, given the large
variability associated with these trials, and given the standard analytical
tools available to researchers, the results of such trials are all too often
that "no firm conclusions can be reached," or "further trials are needed."
This overview of Adaptability Analysis has shown that AA, used either
as a stand-alone method of experimental design and analysis, or in












Adaptability Analysis: An Overview


combination with older, more established methods, is an effective and
efficient tool in the effort to identify improved technologies and produce
extension messages relevant to a range of farmer circumstances, both
biophysical and socio-economic.
As illustrated by the Brazil cowpea trial, use of AA can quickly and
accurately produce technology recommendations for a variety of
recommendation domains in a way that traditional methods alone cannot.
It does this chiefly through recognizing that the farming systems
research-extension concept of recommendation domain is a function of
the specific socio-economic circumstances that determine what farmers
want from new technologies and the specific biophysical circumstances
of their farm environments, which determine how new technologies
perform. With just a few simple analytical tools, requiring only limited
statistical expertise or computing facilities, AA allows researchers and
extensionists to exploit completely adaptive research done under farmers'
conditions. When these conditions are characterized by a full range of
environmental characteristics (usually biophysical), and interpreted in
light of farmers' multiple evaluation criteria (usually socio-economic),
Adaptability Analysis can result in multiple extension recommendations
appropriate for specific groups of farmers and specific groups of farm
environments (Figure 2.9).











Adaptability Analysis


Extension
Recommendati


-- FP


P .FP


WL


On-Farm
Mule Research-Extension
Multiple in
Criteria Multiple Environments
S t/ha
/ I PFI


SkIg/$

/ha


ke/$
o t/ha


WL


FIGURE 2.9 Multiple extension recommendations, specific to multiple
recommendation domains associated with different biophysical
characteristics and criteria resulting from variable socio economic
conditions.

















3


Single-Factor Trials



As discussed in Chapter II, AA developed from a technique used by
plant breeders for evaluating the stability and adaptability of crop
varieties. Not surprisingly, given that variety trials are probably the
most common type of on-farm trials, one of the most frequent uses of
AA is for determining specific adaptation of new varieties to particular
recommendation domains.

On-Farm Maize Variety Trials (Paraguay)

The first example to be discussed is from the results .of a trial
conducted in a single year in 24 on-farm environments in Paraguay
(Poey, n.d.), Table 3.1. Four maize varieties were tested, with a single
replicate of the four varieties per farm. The quality of this data set is
quite high. The ratio of the range of the El to the overall mean El is
1.2:1, indicating that a wide range of environments was included in the
trial. Furthermore, the range of the farmers' local variety, criollo, spans
what might be expected in normal years, and the distribution of the Els
is quite acceptable. The analysis of variance, over all 24 sites, is
presented in Table 3.2.
The test of significance of the variety-by-environmental index
interaction, which will not be treated here, is explained in detail in
Stroup, et al. (1993); it is also discussed briefly in the introduction. In
this case it means that the slopes of the regression lines reflecting linear
responses of yield to environment in the figures below are significant at
a = 0.0588. If the probability had been much greater (ac greater than














TABLE 3.1 Paraguay on-farm maize variety trial, yield (t/ha). Source: Poey, n.d.
Varieties
Region Location Suwan Guarant Poblaci6n Criollo El
Misiones Ibanez Rojas 1.334 1.383 1.306 1.660 1.421
Cordillera Piribebuy 1.904 1.714 1.602 1.595 1.704
Caazapa Iturbe 2.125 1.549 1.778 1.600 1.763
Misiones San Solano 2.090 1.733 1.895 1.728 1.862
Caazapa E.A. Garay 1.933 1.922 1.775 1.827 1.864
Misiones San Juan 2.357 1.993 1.807 1.641 1.950
Caazapa Maciel 2.421 2.192 1.556 1.690 1.965
Misiones Yacua Sati 2.551 2.103 2.589 1.914 2.289
Cordillera Valenzuela 2.877 1.601 2.558 2.299 2.334
Cordillera Eusebio Ayala 2.701 2.259 2.467 2.243 2.418
Cordillera Isla Pucu 2.516 2.383 2.696 2.261 2.464
Ybycui Caacupe 3.012 2.253 2.363 2.345 2.493
Cnel Bogado Nacional 3.462 2.646 3.262 2.153 2.881
Ybycui Tacuary 3.098 2.670 2.582 3.182 2.883
Concepcion Loreto 4.215 2.738 2.871 2.580 3.101
Cnel Bogado Ypayere 3.184 2.672 3.341 3.443 3.160
Caazapa Yuty 3.735 3.682 3.194 2.790 3.350
Ybycui Sapucai 4.794 3.031 3.744 2.766 3.584
Concepcion Concepcion 4.799 4.128 4.472 4.247 4.412
Ybycui Pereira Que 5.360 4.342 4.583 3.597 4.471
Cnel Bogado Syryryca 4.682 4.503 4.238 4.684 4.527
Caazapa Caazapa 4.960 5.514 3.708 4.136 4.580
Concepcion Horqueta 4.939 4.931. 4.011 4.467 4.587
Concepcion Yby-yau 5.246 5.538 4.781 4.410 4.994












Single-Factor Trials


TABLE 3.2 Analysis of variance, with interaction of varieties and El

Source of Variation df Prob. > F

Environment 23 0.0001
Variety 3 0.0001
VxEI 3 0.0588
Residual 66
C.V. = 12.6%


0.2, for example), it might not be of interest to do the regression
component of AA but rather to assume just one recommendation domain,
and choose the highest yielding variety for all environments (note the
highly significant effect of variety). The inclusion of the treatment-by-El
interaction in ANOVA is not necessary to perform AA; one should not
worry if means of calculating it are not available. Its use here serves
only to illustrate the statistical significance of the interaction shown in the
linear regressions. As always, however, it is the practical rather than
statistical significance which is of most importance when analyzing the
results of on-farm research-extension.
To arrive at an understanding of the response of the yields of the four
varieties to environment, individual treatment yields are plotted on
environmental index. First, plotting the individual yields of all varieties
shows the overall variability and some first ideas on response to
environment (Figure 3.1a). While there appears to be little difference
between the varieties in low-yielding environments (El < 2 t/ha), in
better environments Criollo, the local variety (A), seems to yield
generally less than the other varieties, while Suwan (*) appears to be
relatively high-yielding in all environments, and Guaranf (*) seems to be
high-yielding in the very best environments.
Which is the appropriate regression of yield on environment: linear or
quadratic? Figures 3.1b, 3.2, and 3.3 show the yields of each variety
plotted on EI. They all seem to be generally linear, but it might be
argued from visual estimation alone that the response of Poblaci6n
(Figure 3.1b) and Suwan (Figure 3.2b) curve slightly downward at very
poor and very good environments, while that of Criollo (Figure 3.2a)
and Guaranf (Figure 3.3) curve slightly upward. The regressions are
presented in Figure 3.4a for the linear model and in Figure 3.4b for the













66 Adaptability Analysis

6

a ** "
5- *

-% 4
,3 0






1 -
2





0 1 2 3 4 5 6
ENVIRONMENTAL INDEX, El
SWAN GUARAI POBLAON CIOLLO
a A




b
S -1 3

4O
0
o o
4 0O
00



2-

1-

0 .. I I *I .
0 1 2 3 4
ENVIRONMENTAL INDEX, E

a


FIGURE 3.1 Yields of all varieties (a) and yields of Poblaci6n alone (b)
plotted on Environmental Index (EI).














Single-Factor Trials 67

6

a


4
546



I2

1-


0 I I t 1- 3- -
0 1 2 3 4 5 6
ENVIRONMENTAL INDEX, El










A 4

6 i----------------I- ,---------------


2
0 .




0 1 2 4 6
ENVIRONMENTAL INDEX, El





FIGURE 3.2 Yields of Criollo (a) and Suwan (b) plotted on
Environmental Index (EI).












Adaptability Analysis


6
*
S S


Q'












FIGURE 3.3 Yields of Guaran* plotted on El.
1-












or more, Table 3.3. Is environment, as estimated by EI, related to
0 1 2 3 4 5 6
ENVIRONMENTAL INDEX, El



FIGURE 3.3 Yields of Guaranf plotted on EI.

quadratic model. The linear model predicts Suwan to be superior across
all environments, while the quadratic model predicts that Suwan is
superior (in terms of grain yield at least) especially in the average
environments, while being no worse than the other varieties in both low-
yielding and high-yielding environments. In practical terms, then, if
grain yield per unit land area were the evaluation criterion of most
importance to farmers, Suwan could be recommended to all farmers.
In this particular case, where there seems to be just one
recommendation domain (given the evaluation criterion of t/ha),
environmental characterization is of less importance than if there had
been more than one recommendation domain. It may be desirable,
however, to verify that Suwan is superior in both the "good" and "poor"
environments. One of the environmental characteristics collected in this
trial was years each test field had been cultivated, ranging from 15 to 90
or more, Table 3.3. Is environment, as estimated by EI, related to
number of years in use? Visual evaluation of El plotted on number of
years in use (Figure 3.5) indicates an inverse relationship, which is what
one might logically expect. Linear regression of this relationship shows
a significant negative effect, with number of years in use accounting for
75% of the variability in El among the trial environments (p=0.0001).



0













Single-Factor Trials 69

















O 1 2 $ 4 5 6
ENVIRONMENTAL INDEX, El
SUWAN GUARANI POBLACON CRIOLLO I


6

b



4







1-
&, M,&A = 4, ,
0 1 2 S 4 5 6
ENVIRONMENTAL INDEX, El

SUWAN GUARAMNI fOBLAON CM1DUD M



FIGURE 3.4 Linear (a) and quadratic (b) regression of Suwan, Guaranf,
Poblacidn and Criollo (a), on EI.














TABLE 3.3 Environmental characterization, Paraguay on-farm maize variety trial

El Region Location Years use Soil color Soil texture* Slope
1.421 Misiones lbanez Rojas > 50 It. brown cs rolling
1.704 Cordillera Piribebuy >50 It. brown sic steep
1.763 Caazapa Iturbe 80-100 brown cs rolling
1.862 Misiones San Solano >50 It. brown cs rolling
1.864 Caazapa E.A. Garay 80-100 brown cs rolling
1.950 Misiones San Juan >50 It. brown cs rolling
1.965 Caazapa Maciel 80-100 brown cs rolling
2.289 Misiones Yacua Sati >50 It. brown cs rolling
2.334 Cordillera Valenzuela >50 red sc steep
2.418 Cordillera Eusebio Ayala >50 dk. brown cs steep
2.464 Cordillera Isla Pucu >50 red sc steep
2.493 Ybycui Caacupe 50 dk. brown cs level
2.881 Cnel Bogado Nacional 40 dk. brown sc level
2.883 Ybycui Tacuary 50 dk. brown cs level
3.101 Concepcion Loreto 50 It. brown cs rolling
3.160 Cnel Bogado Ypayere 20-30 red sc rolling
3.350 Caazapa Yuty 30 red sc rolling
3.584 Ybycui Sapucai 10-20 black Ic steep
4.412 Concepcion Concepcion 20 dk. red sc rolling
4.471 Ybycul Pereira Que 20 reddish sc rolling
4.527 Cnel Bogado Syryryca 20-30 red sc rolling
4.580 Caazapa Caazapa 20 red sc level
4.587 Concepcion Horqueta 20 dk. red sc rolling
4.994 Concepcion Yby-Yau 20 dk. red sc rolling

4 cs = clayey sand; sic = sand-loom-clay; sc = sandy clay; Ic = loamy clay










Single-Factor Trials


6

5 b -0.041
SR2= 0.75




12



0 I I I I I I I I
0 10 20 30 40 50 60 70 0o 90 100
YEARS IN USE



FIGURE 3.5 El plotted on years in use.


In the case of these data, analysis of variance, by potential
recommendation domain based on years of use, supports the conclusion
that Suwan is superior in all locations, i.e., that there is just one
recommendation domain if only t/ha is the criterion used (Table 3.4).
Finally, it can be shown from AA that Suwan does not just have
higher average mean yields than the other varieties, it also presents lower
risk to farmers. In the environments with El < 4, for example, there
is negligible risk of yields of Suwan falling below 2.5 t/ha, while the
chance of Guaranf yielding less than this level is around five percent, and
for Poblaci6n, over ten percent (Figure 3.6).
It can be seen from these analyses that AA can indicate the existence
of only a single recommendation domain as well as the existence of more
than one. It can happen, however, that there is just one recommendation
domain using a given evaluation criterion (as for yield/ha in this case)
and more than one domain when using alternative evaluation criteria.












Adaptability Analysis


TABLE 3.4 Mean separationst, over all locations and by tentative
recommendation domain based on years in use, Paraguay maize trials


Variety All locations Years in Use < 51 Years in Use > 50

-Grain yield (kg/ha)
== -= Oa=
0.5 0.30 0.05 0.3 0.05 0.30
Suwan 3.35 a a 4.27 a a 2.26 a a
Guarani 2.90 b b 3.74 b b 1.89 b b
Poblaci6n 2.88 b b 3.63 b bc 2.00 b c
Criollo 2.72 b c 3.45 b c 1.86 b c
LSD 0.215 0.112 0.357 0.185 0.204 0.105

t Means in a column followed by the same letter are not significantly different
by Fisher's Protected LSD test at the 0.05 and 0.30 levels of probability.













02
3.5













0 5 10 15 20 25
RISK (% CHANCE OF A LOWER VALUE)
SUWAN QUARANI POKAOON 1IUOL


FIGURE 3.6 Risk assessment for maize varietal yield in low-yielding
environments (more than 50 years in use).











Single-Factor Trials


Hormone Implants in Beef and Dual-Purpose Cattle (Panama)

The difficulty inherent in on-farm livestock trials has limited the
number of such trials and subsequently limited the number analyzed
using Adaptability Analysis. However, many on-farm livestock trials are
amenable to analysis by AA, as the example below illustrates.
This is a fairly straight forward trial of two treatments conducted on
a large number of farms. A simple, two-treatment trial was conducted
in a large number of farm environments in Panama to evaluate the
efficacy of growth-hormone implants in increasing daily gain of cattle
(Simpson, et al., 1988). There were great differences among herds in
this study, including nature of the system and herd size. Four kinds of
system were included: 1 = dual purpose calves, 2 = calves from a
cow/calf system, 3 = combined cow/calf and fattening (grass), and 4 =
fattening (grass). Seasons during which the animals were in the trial
included RB = beginning of the rains (3/15-6/15), RE = ending of the
rains (6/15-9/15), DB = beginning of the dry season (9/15-12/15), and
DE = ending of the dry season (12/15-3/15).
Both treatments (Zeranol* implant and a control) were included in each
of 44 separate trials under regular farm conditions with the number of
animals being subjected to each treatment varying from two to 20.
Weights were taken 90 days after the implant. In 25 of the trials only
one 90-day cycle was included, 9 covered a 180-day period and thus had
the original implant and one reimplant, 3 covered 270 days, and 7 had
the original and three reimplants. Altogether there are 80 environments
represented in the data, Table 3.5.
The observations and relationship between the treatments are shown in
Figure 3.7. The ratio of the El range (1.42) to the El mean (0.92) is
1.60. The range of the control seems to be a reasonable range to expect
and the distribution is quite uniform. Therefore, these data meet the
three criteria for data quality, so we can have confidence in the
relationships indicated.
There appears to be an advantage to the implant in all environments,
although it becomes minimal where El < 1.0. An attempt to character-
ize the environments is shown in Table 3.6 where season and system
were combined and the average Els for the resulting sets were used to
sort. The two highest average Els were for the cow/calf systems in the
two wet periods, the next highest were for the combined system in the
wet periods, followed by the cow/calf system in the dry seasons. The
availability of dams' milk apparently provided a favorable environment












Adaptability Analysis


TABLE 3.5 Daily gain (pounds) from hormone implant and control
treatments for four systems and four seasons, cattle in Panama

System* Seasont Control Implant El


RB
RB
RE
RB
DE
DB
DE
RB
RB
RB
RE
RE
DE
RE
RB
RE
DE
DB
RE
RE
DB
RB
RE
RE
DB
DE
DB
DE
DE
RE
DB
DB
RE
DE
RB


1.53
1.62
1.49
1.39
1.33
1.30
1.41
1.15
0.96
1.29
1.21
1.21
1.22
1.26
1.25
1.21
1.01
1.17
1.13
1.17
1.11
1.03
1.25
0.60
1.13
0.98
1.16
1.14
1.06
1.13
1.01
1.10
1.03
0.93
0.99


1.76
1.54
1.52
1.50
1.53
1.51
1.32
1.54
1.70
1.25
1.31
1.29
1.27
1.22
1.23
1.27
1.47
1.30
1.31
1.26
1.31
1.39
1.13
1.72
1.19
1.28
1.09
1.05
1.11
1.00
1.09
0.99
1.02
1.09
1.02


1.64
1.58
1.50
1.44
1.43
1.40
1.36
1.34
1.33
1.27
1.26
1.25
1.24
1.24
1.24
1.24
1.24
1.23
1.22
1.21
1.21
1.21
1.19
1.16
1.16
1.13
1.12
1.09
1.08
1.06
1.05
1.04
1.02
1.01
1.00


0











Single-Factor Trials


TABLE 3.5 (Continued)


System* Seasont Control Implant El


RE
DE
RB
DB
RB
RE
DE
DE
DE
RB
DB
RE
RE
RE
RB
DE
DE
DB
RE
RB
DB
RE
RB
RB
RB
RE
DB
DB
DE
DB
RB
DE
DE
DE
RE
RB


1.01
0.87
0.80
0.66
0.79
0.90
1.35
0.96
0.85
0.88
0.83
1.02
0.82
0.79
0.86
0.77
0.72
0.86
0.89
0.74
0.60
0.74
0.73
0.74
0.63
0.56
0.60
0.55
0.68
0.51
0.56
0.53
0.43
0.45
0.50
0.60


1.00
0.98
1.01
1.14
1.00
0.87
0.41
0.80
0.87
0.84
0.88
0.68
0.83
0.85
0.78
0.85
0.87
0.73
0.68
0.82
0.94
0.77
0.75
0.71
0.71.
0.75
0.64
0.65
0.51
0.66
0.60
0.60
0.67
0.64
0.59
0.46


1.00
0.92
0.90
0.90
0.89
0.88
0.88
0.88
0.86
0.86
0.85
0.85
0.82
0.82
0.82
0.81
0.79
0.79
0.78
0.78
0.77
0.75
0.74
0.72
0.67
0.65
0.62
0.60
0.59
0.58
0.58
0.56
0.55
0.54
0.54
0.53









0
76 Adaptability Analysis

TABLE 3.5 (Continued)


System* Seasont Control Implant El

3 DB 0.49 0.52 0.50
1 RE 0.34 0.65 0.49
3 DB 0.52 0.43 0.47
4 RE 0.53 0.19 0.36
3 DE 0.29 0.42 0.35
4 DE 0.26 0.21 0.23
1 DE 0.25 0.19 0.22
1 DE -0.01 0.45 0.22
4 DE 0.08 0.27 0.17
Source: Simpson et al., 1988
* System: 1 = dual, 2 = cow/calf, 3 = combined, 4 = fattening
t Season: RB = rains begin, RE = rains end, DB = dry begins, DE = dry
ends


that helped the calves respond to the implants even in the dry seasons.
Beyond these first six combinations, there was no clear division relating
either to season or system, or the combination of the two. For this
reason, a risk analysis was done for the six top categories in Table 3.6
(calves), then the remaining combinations were grouped into 1) wet
seasons and 2) dry seasons, Figure 3.8.
The superiority of the implant treatment in all cases is evident in the
risk analysis as well. For the calves there is little chance of falling
below a daily gain of 1.25 pounds with the implant. Without the
implant, there is a chance one year in 14 of falling below 1.2 pounds
daily gain in these higher environments. Even in all the poorer
environmental situations the impact of the hormone implant is evident,
Figure 3.8.
Because the cost of the implant required only 3.11 pounds of additional
gain over the 90 days or 0.035 per day, (Simpson et al., 1988) the
implants clearly paid for themselves in all environmental situations so no
further economic analysis is required. If the farmers are willing to take
the time (for which charges were included in the cost estimates) then the
implants can be recommended for all environments included in the trial.












Single-Factor Trials


S 0.2 0.4 0.6 0.1 1 12 1.4 1.6 1.8
ENVIRONMENTAL INDEX, El
CONTROL MLANTS CONTROL IANTS
o U mm .


FIGURE 3.7 Relationship of the hormone implants to the control across
environments, cattle in Panama.

TABLE 3.6 Relationship of season and system to environment

Season* System Average El No. Cases

RE Cow/calf 1.305 4
RB Cow/calf 1.270 2
RB Combined 1.142 5
RE Combined 1.059 5
DE Cow/calf 1.043 5
DB Cow/calf 1.032 2
RB Dual p. 0.976 8
DE Combined 0.952 5
DB Dual p. 0.948 4
DB Fatten 0.919 5
RB Fatten 0.879 4
RE Dual p. 0.847 10
RE Fatten 0.788 3
DB Combined 0.777 5
DE Dual p. 0.639 8
DE Fatten 0.624 5

* RB = rains begin, RE = rains end, DB = dry begins, DE = dry ends












Adaptability Analysis


1A
13





1- 0.8
OOS -
.o. .-..*-*-*.-.-'-'-'- ^ ^ ^ ^ ^ ^^ ^ ^^



0.5 I I I I -
0 5 10 15 20 25
RISK (% TIME BELOW VALUE)
CON, CALVES MP. CALVES CON. WE DIMP. WET CN, DRY IMP. DRY



FIGURE 3.8 Risk analysis, hormone implants, cattle in Panama.


On-Farm Sorghum Variety Trials (Cameroon)

This example, of sorghum varietal testing in northern Cameroon,
illustrates the use of AA for "validation" or "demonstration" on-farm
trials, usually done with large numbers of locations (environments) and
small numbers of treatments. Regression of El on quantitative
environmental characters is used to predict recommendation domains and
compare the results to known patterns of adoption. The usefulness of the
three criteria for data quality (particularly distribution of Els) in deciding
how many years of testing are needed is also illustrated.
The improved short-cycle photoperiod-insensitive variety S35 was
tested during four years of on-farm "pre-extension" trials. These trials
were a collaborative effort between the parastatal cotton development
agency SODECOTON and Cameroon's Institute of Agronomic Research
(IRA). Despite promising results in the on-farm trials and extensive
recommendation of S35 by SODECOTON, widespread adoption was
slow. In some areas S35 was much in demand and well adopted; in
other areas it was poorly accepted by farmers.











Single-Factor Trials


The trials took place from 1984 to 1987 (Testing and Liaison Unit,
1986; Testing and Liaison Unit, 1987; Johnson, 1988). The trials in
1984 were done in two different groups, according to geographical
location, north or south. Varieties were S35, locals and 38-3 in the
northern regions and S35, locals, and E35-1 in the central regions. Tests
in 1985 were separated into two groups with different target ranges of
seeding dates. Varieties in the early-seeded sites were S35, locals, and
S34; in the late seeded sites, varieties were S35, locals, S36 and S20.
Varieties for all sites in 1986 and 1987 were S35, locals, CS54 and
CS61. In 1986 and 1987 there was just one set of tests each year. The
treatment called "locals" was not a single variety; each collaborating
farmer selected a locally adapted variety as a local check.
Each on-farm test was conducted on a 0.25-hectare field (50 m by 50
m), was farmer-managed and monitored throughout the season by a
SODECOTON agent. Between-row spacing was 0.80 m; within-row
spacing was 0.40 m between hills, plants thinned to two per hill. In
1984, recommended seeding dates were 20 June to 10 July, choice of
preceding crop was left to the farmer, and all sites were fertilized with
61-13-25 kg ha-1 N-P-K. In 1985, fertilizer was reduced to just 46 kg
N ha-', as urea, and all sites had cotton as the preceding crop. Test
conditions in 1986 and 1987 were as in 1985 except that recommended
date of seeding was 10 to 20 June, in order to avoid favoring the early
maturing, improved varieties (Johnson, 1988).
Standard analysis of test results, by year and by test group within a
year, are presented in Table 3.7. In 1984 S35 yielded 80 percent more
than locals in the north regions and 90 percent more in the central
regions. In late-seeded tests in 1985, S35 yielded over 20 percent more
than locals. In all the 1985-87 early-seed tests, yields of S35 and locals
were essentially the same. The reason for these differences between
years is probably that 1984 was a very low-rainfall year. On the basis of
this, S35 had been extended as a "drought tolerant" variety. Tests in
1984, however, were also characterized by generally late seeding dates
(Table 3.7), in part because of late rains in many areas, but also because
of logistical difficulties in this first year of tests and because
collaborating farmers were given a somewhat late recommended date of
seeding. The rainfall and date-of-seeding factors are confounded in the
general "year effect" in a combined ANOVA of these trials, making
analysis of variance of limited value for identifying potential
recommendation domains. On the basis of the individual year results,
however, S35 came to be recommended generally as good for seeding late.











Adaptability Analysis


TABLE 3.7 Summary of on-farm variety test results with standard t test
results leading to recommendation of S35 for late seeding

Seasonal Mean
Yieldt (kg/ha) Rainfall Seeding
Year Test Group S35 Local S.E. (mm) Date

1984 North (42 Sites) 1070 598 99 393 3 July
Central (46 Sites) 1573 829 109 359 6 July
1985 Early-Seeded (42 Sites) 1866 1721 Ns 118 504 14 June
Late-Seeded (16 Sites) 1416 1156 160 515 28 June
1986 All (38 Sites) 2164 2128 Ns 145 621 22 June
1987 All (35 Sites) 1889 1825 Ns 123 604 19 June

t Yields of varieties for a year and test group are significantly (*) or not
significantly (us) different at the 5% level, by t test.
Source: Adapted from: Testing and Liaison Unit, 1986; Testing and Liaison
Unit, 1987; and Johnson, 1988)


The Adaptability Analysis presented here is based on the data set of
166 on-farm varietal trials in 1984 and 1985 (Table 3.App) These
include 87 environments in 1984 and 79 in 1985. Some on-farm tests in
1985, not strictly varietal tests in the same program but including S35
and farmers' locals are also included here, illustrating the flexibility and
inclusiveness afforded by AA. Adaptability Analysis was applied to
these data in order 1) to examine the structure of the variety-by-
environment interaction across farm environments; 2) to determine if
more than one recommendation domain existed for these varieties; 3) to
identify those variables identifiable ex ante which could serve to predict
in which recommendation domain a given farm environment would
belong; and 4) to determine whether AA might have predicted in just the
first year or two of tests the patterns of S35 adoption actually seen.
Because each year had a different group of varieties, and because El
depends entirely on the specific varieties included, AA in this example
uses only those varieties common to all years (or all groups of test within
a year), i.e., S35 and the farmers' locals. Linear regression of yields of
S35 and locals on El in 1984 are shown in Figure 3.9. The El : mean












Single-Factor Trials


4,5O0
4,000 *
ssoo -
s,0,0 -

s(K" csS- 0 40


2500
S.000 e


0 500 1,000 100 2,000 2,500 s,000 3,50
ENVIRONMENTAL INDEX, El
35 LOCALS 35 WCALS3



FIGURE 3.9 Regressions of S35 and locals on El for 1984.

ratio (3177 / 1037 = 3.06) is quite good, but the distribution is poor,
with a gap of almost 700 in the upper range of Els. Also of note is the
concentration of Els in the low and mid ranges, including a great number
of EIs near zero. Comparison of the regressions indicate that in this
drought year with a late onset of rains, S35 is superior in yield to locals
over all environments, and that this superiority is greatest in the high-
yielding environments.
In 1985, Figure 3.10, the ratio of El to mean is also quite good
(3065 / 1614 = 1.90). Here, however, the distribution of Els is more
even than in 1984, with a smaller gap between El 2700 and El =
3300. In contrast to 1984, there are many fewer Els, and no
concentration of them, in the range below 500. If the 1985 data have a
flaw, it is that perhaps too few Els are less than 900, since 900 kg/ha is
generally accepted as the average of sorghum yields in northern
Cameroon. Regression indicates again a superiority of S35 (though
smaller) in all environments. There is, however, no convergence of
yields at El approaching zero, but rather a slight convergence at high
Els.
Linear regression of yields on El for the combined data set, Figure
3.11, indicates a greater superiority of S35 over locals compared to the















82


Adaptability Analysis


4.500

4,000

.500


2.500

Z 2,000
1.500


Soo
0


FIGURE 3.10 Regressions of S35 and locals on El for 1985.


4,5OOr-----------------------
4,000

3,500 -

3,000 -
2,= C* * ^^ p- a

2,OOO 0 a

1,500- #

1000 0



0 s00 1,000 1,00 2.00 2,500 S,000
ENVIRONMENTAL INDEX, El
133 LOCAlS 35s LOCALS
.F -I o



FIGURE 3.11 Regressions of S35 and locals on El for 1984-85.


1,000 1500 2000 2,500
ENVIRONMENTAL INDEX, El
35s LOCALS S35 LOCALS
.. D











Single-Factor Tials


1985 data, but again without the convergence of yields at very low Els
seen in the 1984 data alone. Here the ratio of El to overall mean is
again very good (3324 / 1311.8 = 2.53) and the distribution is better
than that of either 1984 alone or 1985 alone.
It should be clear that the number of tests (or years of tests) needed to
get a good picture of the variety-by-environment interaction will depend
on how representative the sample of test environments is of the entire
population of environments, across years and locations, for which
recommendations are to be made. In this case, in a region where year
effects such as rainfall and date of seeding (often highly dependent on the
onset of rains) are extremely important determinants of overall mean
yields, one year of data may not be sufficiently representative.
Is there more than one recommendation domain for S35 as opposed to
locals, given that in each year and in the combined data set S35 is
superior to locals across all environments? It is often posited that in the
case of varieties, limited resource farmers usually want a significant
increase in yields over their own varieties before deciding to adopt a
variety with which they are less familiar. Some claim that a 50%
increase in yields is often required. Based on the 1984 data alone, S35
is seen to produce 50% increase over locals' yields across all
environments (Figure 3.12). Using the larger combined data set,
however, predicted yields of S35 are less than 50% greater than locals'
yields when El is larger than around 1000; S35 produces more than 50%
higher yields when El is less than 1000 (Figure 3.13). Note that at El
= 1000, predicted local yields are around 800 kg/ha, i.e., around the
average for the region. Based on comparing the relative performance of
S35 and locals in proportional rather than absolute terms, then, a case
can be made for two potential recommendation domains.
Regressions of El on date of seeding and on rainfall were done to see
how much of the variability in average yields could be accounted for by
these variables. The results of these regressions are presented in Table
3.8. In the 1984, 1985, and the combined 1984-85 data sets, El was
found to be inversely related to number of days planting following 1
January. This relationship accounted for 17 percent of the variability in
yields for the 1984-85 subset. It accounted for just 13 percent of the
variability of yields in 1985 alone and for only three percent of the
variability of 1984 yields, no doubt because such a high proportion of
tests in 1984 were seeded late.















Adaptability Analysis


1A A-


0 500 1,000 1,500 2,000 2,500 3,000 S,5C
ENVIRONMENTAL INDEX, El



FIGURE 3.12 Yield of S35 as a proportion of locals yield, plotted on
EI, 1984.


3.WO3.SAW


500 1,000 1,500 2000 250
ENVIRONMENTAL INDEX, El


FIGURE 3.13 Yield of S35 as a proportion of locals yield, plotted on
EI, 1984-5.


- I I I I


2.8



2.4



2

1.8


1A



0
0


I I II 1 .


a












Single-Factor Trials


The relationship between El and rainfall is not consistent from one
data set to another (Table 3.8), and the percentage of variability in yields
due to rainfall is much less than that due to date of seeding. Farmers do
have some idea of average rainfall over all years and that some regions
are generally higher yielding than others. They never know, of course,
what the rainfall in any given year will be, but many have indicators to
help them predict rainfall. Also, recommendations could conceivably be
made to farmers on the basis of knowledge of average rainfall at their
location. In this case, however, based on these results of regression of
El on rainfall, such recommendations for S35 versus local varieties
would not be warranted.


TABLE 3.8 Regressions of El (S35 and locals) on date of seeding
(DOS) and on rainfall*, 1984, 1985, and 1984-85

DOS Rainfall
a) 1984 b -11.4 1.04
R2 0.03 0.03
p 0.022 0.088
b) 1985 b -22.6 -1.6
R2 0.13 0.09
p 0.006 0.02
c) 1984-85 b -22.0 0.63
R2 0.17 0.01
p 0.0001 0.126

t Date of seeding = days after 1 January.
$ Rainfall = mm, 1-90 days after seeding.


Farmers can and do, however, make choices among alternative
technologies at the time of seeding based on their knowledge of the
effects of date of seeding. The linear regressions of S35 and locals on
EI, and the inverse relationship of El and date of seeding, led to the
delineation, for this study, of two potential recommendation domains, an
early-seeded one (date of seeding before 26 June) and a late-seeded one
(date of seeding after 25 June). Combined analysis of variance indicated
significant (a = 0.0027) treatment-by-domain interaction. Analysis of












Adaptability Analysis


variance and mean separations within each of these domains indicated a
mere 17% percent increase in yields from S35 compared to locals in the
early-seeded domain, but a 67% percent increase in the late-seeded
domain (Table 3.9). Over all 166 test sites, i.e., for only a single
recommendation domain, S35 resulted in only a 36% percent increase in
yields; this may seem a great increase, but in many cases it may not be
enough to offset other characteristics of S35 which farmers find less
appealing than for the local varieties (e.g., taste, cooking qualities,
resistance to bird and other pest damage).


TABLE 3.9 Mean grain yields,t 1984-85 data, over all locations, and
by recommendation domain based on date of seeding


All Locations Seeded Before 26 June Seeded After 25 June
Variety (166 Environments) (83 Environments) (83 Environments)
kg/ha

S35 1510 a 1655 a 1366 a
Locals 1113 b 1417 b 810 b
LSDo.0 105.3 147.2 145.7

t Means within a column followed by different letters are significantly
different at the 5% level, by Fisher's Protected LSD test.


Risk analysis by distribution of confidence intervals (not illustrated)
showed that in both early- and late-seeded domains, the chance of an
unacceptably low yield depended almost entirely on the difference
between means and very little on differences in the variability in yields
on which the means were calculated.
How well do the proposed early- and late-seeded recommendation
domains match the known patterns in the adoption of S35 in the Center
North Zone of North Cameroon? Ranking the SODECOTON sectors by
average date of seeding (based on seeding dates of the 1984-85 tests,
only two of the sectors with average seeding date after 25 June (dos > 1-
77) are not sectors where adoption of S35 has been either fair or very
good (Table 3.10). The two sectors with very good adoption which












Single-Factor Trials


have, from the data, average seeding dates before 26 June are in fact not
generally late-seeding areas and probably owe their high rates of
adoption to the proximity of a well-run experiment station where farmers
have been able to observe and learn about S35 and its production for
many years. Due to the station, availability of S35 seed is also almost
always good.


TABLE 3.10 Correspondence of average test date of seeding, by
sector, and sectors where S35 has actually been adopted, 1984-85 data
combined

N Sector Mean Date of Seeding
(Days after 1 January)

5 Ardaf 168.6
6 Lara 170.6
7 Bidzar 171.6
7 Karhay I 171.6
6 Dziguilao 171.8
10 Mokong 171.9
9 Gobo 172.3
7 Dana 173.7
6 Guidiguis 174.0
8 Moutouroua 175.1
6 Kourgui 175.2 **
5 Mayo Oulo 176.4
6 Koza 176.5 **
7 Sorawel 177.3
10 Hina 177.3
7 Meme 180.4 **
10 Zongoya 180.7
7 Karhay H 180.7
6 Moulvoudaye 181.2*
8 Mindif 182.1 *
7 Yoldeo 182.3 *
9 Bogo 186.8 *
5 Dogba 191.4*

** Sectors with very good S35 adoption
Sectors with good S35 adoption












Adaptability Analysis


While it is true that this study benefitted much from hindsight, it
indicated that adaptability analysis can be of great value in determining
specific adaptation to groups of farm environments and in delineating
recommendation domains based on characteristics that can be identified
ex ante. Had AA-with appropriate environmental characterization-been
done on the data from these trials a recommendation of S35 only for
those areas where farmers tend to seed late might have led to increased
efficiency in extension efforts and perhaps to improved adoption. Such
a recommendation could have been obtained from just the first two years'
(1984-85) data of this four-year series of trials.
Further, although this example was based on AA with only two
treatments, the very simplicity of the data-combined with an
exceptionally large number of locations and years-is useful in illustrating
the importance of several key procedures in AA. The first of these is the
plotting of data points on El before attempting to find the best regression
fit. In many cases, a hand-drawn approximation of the response will
give as good a picture of the treatment-by-El interaction as any easily
derived regression function.
These data highlight the importance of ensuring that the sample of on-
farm trial environments is broad and representative of the total
population of environments for which recommendations will eventually
be made. Representativeness may be relatively easy to ensure in some
agroecosystems, as with the Manaus cowpea trials in the rainforest zone
in Brazil, where the dominant determinant of mean yields was soil
fertility, determined by land type. In others, however, where important
environmental determinants of yield are very unpredictable across
locations or across seasons, particular and careful attention must be paid.
When environmental characteristics can be quantified, as in this case,
their relation to El can be assessed with linear regression. When they
cannot, e.g., for land type, soil texture class, geographical region, etc.,
correspondence with El can be done by sorting environments and looking
for patterns in the corresponding ranking of the non-quantifiable
environmental variable(s). Once recommendation domains are tentatively
identified on the basis of treatment-by-environment interactions in the
regression portion of AA, analysis of relative treatment performance
within each domain can be done. When performance relative to some
current local practice is of most importance, as will almost always be the
case in variety trials, yields as a proportion of the local variety's yields
will help in delineating potential recommendation domains and in
evaluating performance within each domain.












Single-Factor Trials


TABLE 3.App Grain yields, environmental index, date of seeding (days
after January 1) and rainfall, 1984-85 data from Cameroon

Year S35 Locals El D.O.S. Rainfall


1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984


21
22
140
153
96
260
63
236
127
191
305
363
475
353
512
320
662
970
529
1040
1101
725
550
875
724
1044
1143
1228
1286
943
1309
1543
486
1601
1806


18
130
14
25
87
0
234
85
270
257
148
135
127
272
122
467
227
66
529
43
0
400
605
280
531
331
235
153
243
600
256
97
1214
175
0


19
76
77
89
91
130
148
160
198
224
226
249
301
312
317
393
444
518
529
541
550
562
577
577
627
687
689
690
764
771
782
820
850
888
903


199
177
186
174
209
196
186
185
192
181
185
188
185
190
176
195
199
186
180
181
205
167
195
198
192
186
201
190
193
162
186
175
184
187
205


243
306
396
598
182
155
160
329
243
301
302
278
500
251
390
458
236
309
253
304
360
335
329
579
332
285
462
279
536
317
335
433
339
476
278












Adaptability Analysis


TABLE 3.App (Continued)

Year S35 Locals El D.O.S. Rainfall


1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984


941
1423
1216
961
1261
1584
1722
1070
845
953
1934
2110
1586
1534
1878
1453
1001
1500
1855
1395
1049
1805
2048
1605
1512
2078
1405
2092
1045
1627
928
2085
2265
1085
2042
2697
2264


886
441
659
1000
714
400
269
947
1237
1132
184
29
554
643
303
814
1300
908
565
1038
1384
636
417
889
984
461
1182
508
1593
1084
1784
653
564
1754
824
335
855


913
932
937
980
987
992
995
1008
1041
1042
1059
1069
1070
1088
1090
1133
1150
1204
1210
1216
1216
1220
1232
1247
1248
1269
1293
1300
1319
1355
1356
1369
1414
1419
1433
1516
1559


532
392
437
563
435
411
449
438
507
558
498
429
155
498
321
476
298
260
221
518
389
513


174
196
181
170
184
185
194
193
196
184
189
175
171
175
176
184
189
178
196
199
186
177
187
186
192
186
177
178
177
196
167
168
178
194
188
196
167


328
309
331
298
337
316
347
297
493
325
402
286
461
132












Single-Factor Trials


TABLE 3.App (Continued)


Year S35 Locals El D.O.S. Rainfall


1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1984
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985


1574
1883
1808
2540
1607
2275
2490
2443
1874
2188
2619
2714
2476
3353
4018
130
327
748
770
789
658
1159
1462
1230
571
661
904
571
989
778
715
1414
862
1505
861
885
884


1548
1249
1348
915
1879
1444
1271
1344
2085
1859
1466
1675
2195
2629
2375
427
455
302
300
338
633
234
50
366
1039
950
769
1116
731
1087
1159
606
1238
595
1325
1306
1329


1561
1566
1578
1727
1743
1859
1880
1893
1979
2023
2042
2194
2335
2991
3196
278
391
525
535
563
645
696
756
798
805
805
836
843
860
932
937
1010
1050
1050
1093
1095
1106












Adaptability Analysis


TABLE 3.App (Continued)

Year S35 Locals El D.O.S. Rainfall


1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985
1985


1331
1509
1408
1737
1187
1211
1407
1371
1605
1270
1298
1366
1417
1821
1207
1427
2454
1115
1297
1160
1536
1844
1806
2071
2074
1434
1896
1586
2145
2466
1814
1944
1895
1985
2631
2494
2169


970
801
1005
739
1300
1365
1220
1289
1075
1461
1490
1456
1426
1084
1716
1496
509
1856
1719
1875
1508
1235
1384
1166
1268
1912
1461
2020
1526
1312
2060
2059
2128
2076
1449
1627
2010


1150
1155
1206
1238
1243
1288
1313
1330
1340
1365
1394
1411
1421
1452
1461
1461
1481
1485
1508
1517
1522
1539
1595
1618
1671
1673
1678
1803
1835
1889
1937
2001
2011
2030
2040
2060
2089


550
275
410
571
728
619
636
421
548
477
513
364
462
825
460

523
599
381
638
396
410
485
397
561
425
594
504
374
369
418
644
345
694
526
715
409




University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs