Staff Paper Series
FOOD AND RESOURCE ECONOMICS DEPARTMENT
Institute of Food and Agricultural Sciences
University of Florida
Gainesville, Florida 32611
Farmer Participation for More Effective Research
in Sustainable Agriculture
Walter W. Stroup', Peter E. Hildebrand2
and Charles A. Francis3
Staff Paper SP91-32
Staff papers are circulated without formal review by the Food and Resource Economics
Department. This paper, a chapter for the proposed American Society of Agronomy Special
Publication: "Technologies for Sustainable Agriculture in the Tropics", is being circulated for
review and comments. Contents are the sole responsibility of the authors.
1 Associate Professor, Department of Biometry, University of Nebraska-Lincoln
2 Professor, Food and Resource Economics Department, University of Florida
3 Professor, Department of Agronomy, University of Nebraska-Lincoln
Senior authorship not assigned.
Farmers have conducted their own research from before plants and animals were domesticated.
However, with the advent of scientifically based agriculture their influence on technology
development waned. Farming systems research-extension (FSRE) methodology was a response to
a concern that Green Revolution technology was bypassing many small, resource-poor farmers in
the Third World. Based on the FSRE-generated concepts of domains (research, recommendation
and diffusion), the unique nature of on-farm research, and its demands on statistical analysis are
examined. On-farm trials differ from on-station trials in two important ways: 1) the objectives are
usually different, and 2) the variablity of on-farm data is more complex and must be addressed with
greater sophistication. Four analysis of variance (ANOVA) models for on-farm research data are
examined and the relationship of ANOVA to modified stability analysis (MSA) is discussed. Means
of incorporating larger farms (both developed and developing countries) into an organized research
and extension effort are examined. Finally, the integration of large and small farms into a combined
research and extension effort is discussed.
FARMER PARTICIPATION FOR MORE EFFECTIVE RESEARCH
IN SUSTAINABLE AGRICULTURE
If developing countries are to meet national food needs and alleviate rural poverty,
millions of small farmers must become active participants in the agricultural
research and development process (Whyte and Boynton, 1983).
Most crops and many predominant agricultural production systems are the result of
empirical research, or trial and error, by generations of farmers working the land. Neolithic farmers
knew much about 1500 different plant species used for food and medicine (Braidwood, 1967).
Vestiges of their traditional subsistence systems still exist in many regions (Francis, 1986b). With
the advent of scientifically based agriculture following World War II, however, farmers' influence on
technology development became less and less.
In the late 1960s and early 1970s, the international development community began to see
a need to reach the many small, resource-poor farmers who were being by-passed by the Green
Revolution. Whyte and Boynton (1983) argued that this meant 1) an increased emphasis upon on-
farm research, 2) greater interdisciplinary collaboration, 3) agricultural bureaucracies that are more
responsive to the interests and needs of small farmers, and 4) small farmers should no longer be
treated simply as passive recipients of what the experts decide is good for them.
To respond to this new clientele, a methodology was needed to efficiently find
environment-specific technologies for large numbers of such farmers. This methodology had to not
only reach farmers in widely varying and often difficult situations who lack the resources required
to dominate the environment, but also:
speed up the technology development, evaluation, delivery and adoption process,
efficiently use scarce institutional resources (those human, physical, and financial
resources of national agricultural research and extension services in developing
In order to accomplish these needs, the methodology required an integrated, multidisciplinary
approach that incorporated farmers, researchers and extension personnel.
Internationally, over the last 20 years, this "real world" or on-farm research for large
numbers of farmers has come to be called Farming Systems Research and Extension (FSRE). In the
broadest sense FSRE involves
rapid diagnosis of farm problems by multidisciplinary teams to provide the basis for
adaptive and descriptive biophysical on-farm research which is supported by
socioeconomic research on-farm and in the farm community,
controlled biophysical research in laboratories and on-station, and
simultaneous dissemination and diffusion of results.
By incorporating farmers from the beginning of technology development -- from problem
diagnosis, through adaptation and evaluation -- FSRE methodology reduces the incidence of
research results that perform poorly on farms (Figure 1), or the rejection on-station of technologies
which might have performed well on farms but were never released (Figure 2).
As methods have developed over time, FSRE is not limited to small farms. Indeed, efficiency in the
use of research resources is enhanced by incorporating farmers from multiple environments. In this
chapter, methods for incorporating large numbers of limited resource farmers into on-farm research
are discussed. Later in the chapter, means of including larger, more commercial farms is covered.
CONCEPTS AND METHODS
Using FSRE methods, farm problems are diagnosed by rapid rural appraisal procedures
(Chambers, 1981) or sondeos (Hildebrand, 1981), that incorporate farmers as active participants
working with multidisciplinary research and extension teams. These methods are flexible and may
or may not use formal questionnaires in the process. Problems encountered are elaborated and
prioritized for research by several methods including those proposed by Tripp and Woolley (1989)
from CIMMYT and CIAT.
An earlier concept that sought homogeneous groups of farms (Hildebrand, 1981; Norman,
1980) has been modified to incorporate the concept of a research domain (Wotoweic, et al., 1988)
which recognizes the fact that farms and farmers are highly variable and targets this variability.
Often research domains are chosen based on biophysical characteristics although they may be
chosen politically. Research domains ideally contain a wide range of environments that are
incorporated as early as possible in the technology screening process. Environments in this context
can be associated with farms, fields or even portions of fields. The use of socioeconomic
considerations in the choice of environments within the research domain enhances efficiency in
technology development and evaluation.
To comprehend a research domain, compare the environment for producing tobacco in the
field on the small, resource-poor farm in north Florida shown in Photo 1 -- with the environment for
raising tobacco in the field on a larger farm, in the same area, but which has enough resources that
it can dominate the environment to a much greater extent, shown in Photo 2 -- and to the
environment for raising tobacco on an experiment station, also in the same area, where it is grown
with few limitations, allowing most environmental factors to be dominated, Photo 3.
All of these environments can be considered part of the same research domain and be
incorporated simultaneously into an integrated technology development, evaluation and diffusion
process for tobacco in north Florida. The nature of on-farm research in research domains is
exploratory, to answer the questions WHAT and WHERE, not why and when. Diverse
environments, such as those shown in north Florida, enhance the exploratory nature of on-farm
research in research domains.
In a research domain, an integrated, multidisciplinary research and extension team conducts
both biophysical and socioeconomic on-farm research and analyzes the results to 1) characterize
the biophysical environments associated with each location, 2) elicit farmers' evaluation criteria
with respect to the technology being evaluated, and 3) define recommendation domains. A
recommendation domain is a unique combination of these environmental characteristics and
Recommendation domains, then, are one or more subsets of a research domain which
target for homogeneity of 1) natural and farmer-created biophysical environments, and 2) farmers'
evaluation criteria for the technology being evaluated. Also modified from previous thinking that
recommendation domains pertained to whole farms (Byerlee, et al., 1982), or cropping or farming
systems (Hildebrand, 1981) are that these logically can refer as well to individual fields on a farm,
or even different locations in the same field. The most important concept is to consider
recommendation domains as environments whose biophysical and socioeconomic characteristics
can be identified.
The nature of on-farm research in recommendation domains is validation, to confirm
answers as to 1) how each alternative (treatment) will respond, and 2) where each alternative is
best, as well as to refine the characterization of the recommendation domains and farmer
evaluation criteria. At this stage, the number of treatments in on-farm trials is limited. Extension
personnel can play an increasingly important role by expanding coverage for evaluation and
enhancing exposure (diffusion) of the technology.
Diffusion domains are informal interpersonal communication networks through which newly
acquired knowledge of agricultural technology normally flows. Knowledge of these networks is
important in helping research and extension personnel locate on-farm trials to target for
The challenge of diagnosis and identification of these several domains is complicated by
how information is collected, analyzed, and evaluated from on-farm trials. We need to clearly
identify where research results can be applied, how broad the recommendations can be, and for
whom these new technologies are appropriate. To be credible for farmers as well as rigorous from
a statistical point of view, results from on-farm research must be analyzed and evaluated according
to valid statistical methods.
ANALYTIC VERSUS ENUMERATIVE STATISTICAL METHODS
Over the past several decades, procedures for the design and analysis of experiments have
been developed and utilized very effectively in agricultural research. Many of these procedures
have become so institutionalized that it is easy to lose sight of the fact that they are only specific
applications of statistical theory to specific experimental conditions -- namely, those of the
agricultural experiment station.
Are the requirements of on-farm trials identical to those of experiment station trials? There
is no good reason to expect they should be. In fact, on-farm trials differ from their on-station
counterparts in two very significant ways: 1) the objectives are typically quite different; and 2) the
variability of the data in an on-farm trial is typically more complex and must be addressed with
greater sophistication than is normally required for an on-station trial.
How do the objectives of on-farm trials differ from on-station research? How does this in
turn affect decisions regarding appropriate statistical methodology? Although probably not obvious
to agricultural researchers, on-farm trials have many statistical similarities to quality improvement
experimentation in manufacturing. Deming (1953, 1975), the statistician whose contributions to
quality in Japanese industry are legendary, distinguishes between two approaches to statistical
analysis: enumerative and analytic. From Deming (1975):
Enumerative. "The action to be taken on the frame depends purely on estimates or
complete counts of one or more specific populations of the frame. The aim of the statistical study
in an enumerative problem is descriptive." Virtually all classical statistical procedures -- t-tests, F-
tests, analysis of variance (ANOVA), standard confidence intervals -- are enumerative in nature.
Analytic. "In which action will be taken on the process or cause-system that produced the
frame studied, the aim being to improve the practice in the future." Only statistical procedures
which involve prediction rather than estimation or hypothesis testing are analytic in nature.
Deming (1975) puts it another way: "A 100 percent sample in an enumerative problem
provides the complete answer to the problem posed for an enumerative problem. .. In contrast, a
100 percent sample of a group of patients, or of a section of land, or of last week's product,
industrial or agricultural, is still inconclusive in an analytic problem. This point, though fundamental
in statistical information for business, has escaped many writers."
Clearly, most on-farm trials have analytic rather than enumerative objectives. Thus, the
literal application of enumerative statistical procedures, many of which form the core of statistical
tradition in agricultural research, is not appropriate for most on-farm trials. For example, the
analysis of variance (ANOVA) can be very useful for interpreting data from on-farm trials.
However, traditional ANOVA places much emphasis on hypothesis testing and significance levels.
These are important in enumerative studies, but essentially irrelevant to analytic studies, where the
emphasis is on prediction and taking action.
Ad hoc statistical procedures are common in analytic studies. While many of these
procedures can be validly criticized using enumerative statistical arguments, these criticisms often
miss the point. Analytic studies are usually conducted with less prior knowledge of and control
over experimental conditions. The choice is frequently between no knowledge and useful, if
imperfect knowledge; conditions of optimality characteristic of enumerative statistical procedures
are simply not an option. Analytic studies typically sacrifice control over variability for a broadened
research domain. This does not make them incorrect or invalid, it just means that the researcher
must understand the trade-offs and choose statistical methods accordingly.
The complex variability in on-farm trials often troubles those trained in traditional statistical
methods for agricultural research. In statistical jargon, these methods are examples of "ordinary
least squares"; their main virtue is that they are easy relativelyy) to do without a computer, which
was a vital consideration in the 1920s and 1930s when they were developed. Their main
drawbacks are the rigid structure and narrow, frequently unrealistic assumptions required of the
data to permit legitimate interpretation. Since on-farm trials rarely satisfy these assumptions, many
have concluded -- falsely -- that they are somehow "statistically improper." In truth, traditional
methods simply cannot accommodate the complexity of on-farm trials.
"Ordinary least squares" theory has long since been supplanted by more versatile methods,
mixed linear model methods (or "mixed model methods" as they will be referred to here) being of
particular importance to on-farm trials. The virtue of mixed model methods is their flexibility; their
drawback is that they generally require a computer. Thus, while mixed model theory has been
around for nearly a half century, it did not become practical to use until the 1970s in developed
countries and the 1980s in most developing countries. By then, more traditional methods were so
deeply entrenched in statistics courses, on experiment stations, and in agricultural research journals
that substantial re-education has either been required or, more correctly, is still required.
Recently, there has been a great deal of interest in applications of mixed model theory in
agriculture. Henderson (1975) developed best linear unbiased predictors (or BLUPs) as an
alternative to more enumerative-type estimators. Perceived at first as an ad hoc procedure, Harville
(1976) put BLUP on sound theoretical footing. A regional publication of the southern Research and
Information Exchange Group in statistics (Southern Regional Bulletin, 1989) contained several
examples of mixed model applications in agriculture. This publication also contained articles by
McLean (1989) and Stroup (1989a) describing mixed model theory and methods.
In the following section, mixed linear models appropriate for on-farm trials are discussed.
These models superficially resemble models used to evaluate on-station data. The goal of this
section is to show how to use mixed model theory to understand the distinction between the
various assumptions that can be made about these models, their effect on the resulting analysis,
and their implications for the on-farm researcher. The larger objective is to empower the on-farm
researcher with a relevant statistical perspective so that design and analysis choices appropriate to
on-farm trials can be made.
THE "TYPICAL" ON-FARM TRIAL
On-farm trials are conducted in a variety of ways, but most have a common basic structure.
The following is a generic description of the essential elements:
Suppose a number of treatments, V, are to be evaluated. Each treatment is observed at F
different farms where the specific biophysical and socioeconomic characteristics of the specific site
on the farm will be characterized. At each farm site, each treatment is "replicated" R times -- the
word "replicated" appears in quotes here because, as will become apparent later in this discussion,
multiple observations on treatments within a farm site may not be true replications. Note that the
term "farm", to be designated in what follows by the letter "F" is generic. The term more
specifically should be interpreted as "environment". In specific trials, "field", "location", "village",
etc. may apply equally.
Schematically, this trial can be represented as in Figure 1. As a starting point for analysis
of this trial, the following mathematical model can be used:
yi = p + fi + r(f), + Vk + vfi + e,, (1)
where y, is the observation on the jh replication of the i* farm for the k* treatment,
p is the overall mean,
f, is the effect of the i* farm,
r(f), is the effect of the j* replication in the i* farm,
vk is the effect of the k* treatment,
vfk is the interaction between the it farm and k* treatment, and
e, is residual variation not accounted for by the above effects.
The analysis of variance (ANOVA) implied by this model has the following general form:
SOURCE OF VARIATION degrees of freedom
FARM F F-1
REP(FARM) R(F) F(R-1)
TREATMENT V V-1
FARM X TREATMENT V x F (F-1)(V-1)
RESIDUAL resid F(R-1)(V-1)
TOTAL FRV 1
This ANOVA has several possible interpretations, depending on the specific objectives of a
given on-farm trial and how the effects in the model are defined as a consequence. In order to
make appropriate use of this ANOVA table, the researcher must be clear about the objectives of
the trial and the nature of the effects being observed. Some useful definitions follow.
Population of inference: The set of elements (e.g. farms) to which the results of the study
are to be applied. This is similar to the concept of a research domain.
Prediction space: Applications of study results from on-farm trials often take the form of
recommendations. Recommendations are based on the predicted behavior of the treatments, either
for the entire population or for various sub-populations. The set of elements (e.g. farms or
environments) to which a prediction is intended to be applicable is called the prediction space. This
is similar to the concept of a recommendation domain.
Random and Fixed Effects: Effects in the study -- treatments, farms, replicationss" -- can be
considered as fixed or random depending on 1) how they are chosen and 2) what prediction space
is appropriate to the objectives of the study. An effect is considered fixed if the levels of a
particular factor are chosen deliberately in advance of the study. In this case identical levels would
be used again were the study to be repeated based on the same prior knowledge, and prediction is
limited to only those levels actually represented in the study. Typically, treatments such as tillage
methods or fertilizer levels in a variety trial would be considered fixed effects. An effect is
considered random if the levels actually observed in the study result from a random sample of a
larger population -- identical levels in a repeat of the study would be exceedingly unlikely.
Prediction in this case is intended to apply to the population of which the levels observed are only
representatives. I he most blatant example of a random effect would be the effect of "replication"
or of residual variation. Many effects are not clearly fixed or random -- the effect of farm site or
environment, for example. Whether an effect is fixed or random has a major impact on the
analysis, as will be demonstrated below.
Most statistical methods texts, e.g. Steel and Torrie (1980) or Snedecor and Cochran
(1980), contain discussions of fixed and random effects. Many texts on the design or planning of
experiments discuss the population of inference, e.g. Cox (1958) or Mead (1988). The reader is
referred to these texts for more detail.
IMPACT OF FIXED OR RANDOM EFFECTS ON ANOVA
In the ANOVA for the on-farm trial given above, it is usually fairly clear that "treatments"
are fixed effects and replicationss" are random. Farms, however, are not so easily categorized.
Different farms may have been selected quite intentionally based on certain criteria: size, income,
technology level, soil type, climatic characteristics, etc. Or they may have been selected at
random from a target population. Actually, these are extremes; usually, farms are selectecdusing a
combination of fixed and random effect tactics. That is, a spectrum of defined conditions must be
represented, but some form of random sampling is done within each condition. Essentially, this
amounts to stratiied random sampling.
It follows that farms are not easily categorized as fixed or random. Usually, in fact, the
"correct" analysis of the on-farm trial will involve some compromise between the analysis with
farm as a fixed effect and the analysis with farms as random. Before examining this
"compromise," it is instructive to look at the appropriate analyses with farms strictly fixed or
If farms are fixed, then the only random components of model (1) are r(f), and ek. Denote
the variance of r(f),j by a2 and the variance of ew by a2. Then the expected values of the mean
squares of the ANOVA are as follows:
SOURCE OF VARIATION EXPECTED MEAN SQUARE
F o2 + FVo,,2 + RV0,
R(F) o2 + FVar,
V a2 + FR,,
Vx F o + R0,
where o,, 0,, and 0, denote variation attributable to the fixed effects f1, vk, and vfk,
If farms are random, then the components f, and vf, from model (1) are also random.
Denote their variances by o,2 and a2, respectively. Then the expected mean squares are:
SOURCE OF VARIATION EXPECTED MEAN SQUARE
F o2 + Ro,2+ Va,2 + RVo,2
R(F) o2 + Vor,2
V a + Ra,2 + FR0,
Vx F o0 + Roa2
These two ANOVA tables imply very different approaches to inference. When farms are
fixed, the data analyst's first concern must be the farm by treatment interaction (V x F), the
magnitude of which is assessed by the F-ratio MS(V x F)/MS(resid). If this F-ratio indicates the
existence of interaction, then effort must be focused on understanding its nature. Even if the
interaction F-ratio appears to be negligible, the data analyst would do well to partition MS(V x F)
into meaningful components, e.g. using contrasts, since important interactions often are masked by
the large number of degrees of freedom associated with the V x F effect (see Snedecor and
Cochran (1980), pp. 304-307).
When farms are considered fixed, the treatment main effect is of interest only if the
interaction effects are negligible, i.e. if it is clear that the same relationships among treatment
means appear to hold for every farm in the population of inference. This is generally not true, but
if it is then the treatment main effect can be evaluated using the F-ratio MS(V)/MS(resid).
When farms are considered random, then test of farm by treatment interaction, which uses
the same F-ratio as above, has a far different interpretation. Specifically, it means that differences
among treatments vary at random by farm. This is quite distinct from the fixed effect case, in
which interaction implies that relative differences among treatments are affected by systematic,
identifiable and repeatable farm characteristics (i.e. the characteristics that motivated the choice of
the farms in the first place). In fact, the test for interaction is not particularly interesting if farms
are random: if oa, is not greater than zero, ten the assumption of random farms is probably
defective. Of interest is the treatment main effect. This is evaluated using the F-ratio MS(V)/MS(V
x F). Its purpose is to verity that iftterences among treatment means, substantial and consistent
enoughh to be seen through the population of inference, over and above random ditterences among
treatment Dy farm, actually exist.
To summarize, if farms are fixed, the F-ratio of primary interest is that for the V x F
interaction, MS(V x F)/MS(resid), or, if the Vx F interaction is negligible then the V main effect is
assessed by MS(V)IMS(resid). If farms are random, the V x F test is of little intrinsic interest
(except to verify the validity of the assumptions); of primary interest is the V main effect, which in
this case has an F-ratio MS(V)IMS(V x F).
In on-farms trials as they are actually conducted, farms are rarely purely fixed or purely
random effects. The above ANOVAs, therefore, are useful as academic exercises to illustrate
issues the farming systems researcher needs to understand, but neither, unmodified, is likely to be
of much use in practice.
PARTITIONING THE FARM BY TREATMENT INTERACTION
In most on-farm trials, the population of inference includes a set of "types of
environments," that the researcher wants to be represented. In the extreme fixed effects case, the
number of types would be F, and thus only one environment per type would be observed. In the
extreme random effects case, there would be exactly one type of environment (or so little would be
known about the environments that typing could not be done prior to conducting the trial) and F
randomly sampled environments per type. Usually on-farm researchers would reject either extreme;
a more realistic design would be to randomly sample a number of environments from each of the
several types of in the population.
If the types of "farms" are very well defined, model (1) could be modified as follows:
yi, = p + t, + f(t),i + r(tf),k + v, + vt, + vf(t),, + e,,, (2)
where t. is the effect of farm type,
f(t), is the effect of farm within type,
vt, is the farm type by treatment interaction,
and other terms follow by extension from model (1).
In model (2) type and treatment would be considered fixed, farm and replication random,
and analysis would proceed accordingly based on the following ANOVA:
SOURCE OF VARIATION df EXPECTED MEAN SQUARE
T T-1 o2 + Raut2+ Voaf2 + RVaf,2 + FRV0,
F(T) T(F-1) o2 + R,, + Vo,2 + RVo,,
R(TF) TF(R-1) o2 + Vor2
V V-1 o~ + Ra7,2 + TFR0,
VxT (T-1)(V-1) o2 + Rarv2 + FRO
Vx F(T) T(V-1)(F-1) o2 + Roa,2
residual TF(R-1)(V-1) o2
The type by treatment (V x T) interaction would be of initial primary interest. Its F-ratio is
MS[V x T]/MS[V x F(T)].
As before, partitioning MS[V x T] into meaningful contrasts would be strongly advisable.
For example, suppose the farm types are:
1. higher rainfall, mechanized
2. higher rainfall, non-mechanized
3. lower rainfall, mechanized
4. lower rainfall, non-mechanized
and the treatments are:
1. standard variety, no fertilizer
2. standard variety, with fertilizer
3. resistant variety, no fertilizer
4. resistant variety, with fertilizer
The type main effect could be partitioned into rainfall and mechanization main effects and a
rainfall by mechanization interaction. The treatment main effect could be partitioned into variety
and fertilizer main effects and a variety by fertilizer interaction. Then the interaction of any of the
three type effects with any of the three treatment effects could be evaluated. For example, a
rainfall by variety effect could be examined to see if the resistant variety is equally advantageous at
lower and higher rainfall. In the unusual case that type by treatment interactions are negligible, the
treatment main effect could be tested using MS[V]/MS[V x F(T)].
Predicted performance of treatments for particular farm types can be obtained using
confidence intervals for the treatment x farm type means. Care should be taken to base the
confidence interval on the correct standard error. Most statistical software packages are poorly
suited to work with mixed linear models such as model (2) without special attention. For a
complete discussion of this issue, see McLean (1989) and Stroup (1989a). Predicted performance
of specific farms within a given farm type for a particular treatment can be obtained by calculating
best linear unbiased e (Henderson, l /b). ihese are not the same as usual sample
means. Again, see McLean (1989) and Stroup (1989a and 1989b) for a full discussion of best
linear unbiased prediction.
A special case of the above analysis occurs when "environmental tyes and their potential
interactions with treatment are not well understood prior to conducting the on-farm trial. In such
cases, the researcher makes an attempt to represent as wide a spectrum of types as possible
within the population of inference but a "clean" partition of the variability among environments into
types and environments within types may not be possible. Indeed, one objective of the research
may be to provide insight concerning which environments favor or disfavor certain treatments and
what features are common to these environments. Various forms of "stability analysis" are
important examples of this approach.
Excellent review articles on stability analysis are available (see Freeman (1973), Hill (1975),
Westcott (1985)). Hildebrand (1984) has adapted the approach for on-farm trials and its use is
demonstrated in the following section. This discussion will be restricted to pointing out its relation
to model (2) above. In Hildebrand's modified stability analysis (MSA), an index for a given
environment (El) is defined as the mean response over all treatments at that farm site. A linear
regression over Els is obtained for each treatment and used as a basis for determining
"recommendation domains," a notion loosely similar (but not identical) to the mixed model concept
of prediction space. In terms of ANOVA, this could be expressed by modifying model (2):
y, = p + fI + r(f), + v, + B(El,) + vfk + ek (3)
where El, is the index of the iP environment, and
Ik is the linear regression coefficient for the k* treatment.
In essence, El in model (3) replaces type in model (2). Also, f, in model (3) is equivalent to
ti + f(t), in model (2) and vfk in model (3) is equivalent to vf(t), in model (2). Since environment
(represented by "F") aside from El, is a random effect, the ANOVA is:
SOURCE OF VARIATION df EXPECTED MEAN SQUARE
F F-1 a2 + R",2 + Vo'2 + RVo,2
R(F) F(R-1) o2 + Va,2
V V-1 o2 + Rou'2 + FR0,
V x El V-1 o2 + Ro2 + FR0Ei
VxF (V-1)(F-2) 02 + Ro,2
residual F(V-1)(R-1) o2
Equality of the Bk can be tested using MS[V x EI]/MS[V x F]. A "significant" F-ratio would
imply that treatments respond unequally to El (and thus to whatever environmental types the El1
imply). This would in turn provide formal justification for predicting that different treatments are
optimal for various "recommendation domains."
There is no reason why the use of environmental indices need be limited to linear
regression. For example, model (3) can easily be extended to
yk = p + fi + r(f),, + v, + Bil(Eli) + B2k(EI,)2 + vf, + eJk, (4)
where Bk is the linear regression coefficient for the k*" treatment, and
B1k is the quadratic regression coefficient for the k'h treatment.
The ANOVA for model (4) would be identical to the ANOVA for model (3) except that an
additional line for V x El2 (or V x El x El) with V-1 degrees of freedom would appear immediately
after V x El and the remaining V x F term would have (V-1)(F-3) degrees of freedom.
The F-ratio MS[V x EI2]IMS[V x F] tests the equality of quadratic regression over El for the
various treatments. Pictorially, this can be visualized as in Figure 4. Note that the quadratic
regressions are quite different for the treatments, although their linear components are similar.
Several authors have noted the limitations of linear-only regression over El, e.g. Westcott (1985).
However, model (4) should make it clear that this restriction is unnecessary. Indeed, model (4) can
be extended to more complex forms of regression over El.
If there is only one "replication" per farm (a discussion of the advantages and
disadvantages of this appears below) then the R(F) and residual terms in the ANOVA have np
degrees of freedom and the result is the following simplified form:
SOURCE OF VARIATION df EXPECTED MEAN SQUARE
F F-1 oa + Vo,2
V V-1 a2 + F,
V x El V-1 a,2 + Foe
V x E2 V-1 a2 + F0~2
V x F (now the residual) (V-1)(F-3) 0o,
Note that this has no impact on the F-ratio used.
The use of El in stability analysis has been widely criticized because the independent
variable El is in fact a function of the dependent variable. Westcott (1985) makes a case for
greater use of independently determined "environmental variables." He also notes that
"environmental measurements are very seldom available." Theoretical objections aside, the on-
farm researcher often has but two alternatives: using 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." Critics often point to the
weaknesses in formal statistical properties of analysis using El. These difficulties clearly exist;
however, a more compelling point is that the researcher often has the El as the ONLY objective
guide to environmental quality. These criticisms would be severe problems if formal, definitive
statistical inference were the objective. It is not. The more important use of this type of analysis
is to obtain preliminary insight regarding the consistency of treatment performance, which fields,
farms or groups of farms appear to be troublesome, what recommendations appear to be
reasonable etc. This sort of analysis is always a starting point, never an end in itself.
For the researcher to make the jump from finding a significant El x treatment interaction
from a model such as (3) or (4) to associating El with predictable future environments or
"recommendation domains" and making reliable treatment recommendations for them obviously
requires a great deal of thought and care (and involves, to a large extent, non-statistical questions,
i.e. why are some El low and others high). Predicted treatment performance for farms included in
the trial can be made using well known best linear unbiased prediction methods. The Els have no
intrinsic meaning, so predictions for fields or farms not included in the trial are only as good as the
researcher's ability to predict which fields or farms will be in which recommendation domain. The
on-farm trial will not by itself generate data suitable for this purpose.
AN IMPORTANT NOTE ON DESIGNING ON-FARM TRIALS
Note that neither MS[R(FT)] nor MS[resid] are ever used in the analysis of the "usual" on-
farm trial i.e. one described by some variation on model (2). The appropriate denominator term for
all tests of interest is MS[V x F(T)]. Why is this important? Both MS[R(FT)] and MS[resid] require
that R, the number of replicationss" per farm, be at least two. However, neither of these terms
has any role in the analysis of the standard on-farm trial. What would happen if only one
replication per farm were observed? Neither MS[R(FT)] nor MS[resid] could be calculated.
However, since neither term plays any role in the analysis, this is no real disadvantage.
It IS important to have as many farms per type as possible. This maximizes the degrees of
freedom for MS[V x F(T)]; since this is the denominator term for all F-ratios of interest, this will
maximize power and, consequently, the usable information available. Thus, it is the FARM that is
the true replication in an on-farm trial, not the "replication" within a farm (hence the motivation for
the quotation marks'). rhis is important because on-farm researchers often have been advised to
replicate within a farm, even for example in Hildebrand and Poey (1985)1 From an ANOVA
viewpoint, we now know this is clearly erroneous advice. Moreover, it is wasteful: the researcher
would be better off observing more farms. Even worse, it abuses the hospitality of the farmer
donating the space for the research to be conducted; the farmer should not have any more land out
of ordinary production than absolutely necessary.
To repeat, in most on-farm trials, the number of farms observed should be maximized.
Replication within a farm should not ordinarily be necessary and is usually wasteful. The only
exception is for the purely "farms as fixed effect" case of model (1), an unlikely, though not
unheard of, on-farm trial design.
MODIFIED STABILITY ANALYSIS
One method for managing research in such different environments as those shown above in
north Florida is with "stability analysis", modified to provide a positive rather than a negative
interpretation to treatment by environment interaction (Hildebrand 1990). Figure 3 shows
hypothetical results of three varieties (as an example of three alternative technologies) that have
been tested over an appropriately wide range of environments. In this hypothetical case, all three
have the same overall mean yield and deviations from regression, s2d, = 0. The linear regression
coefficients are 1.5, 1.0 and 0.5 for varieties A, B and C, respectively. In the absence of other
disqualifying characteristics, variety B (the most generally adaptable according to Finlay and
Wilkinson (1963), or the most stable according to Eberhart and Russell (1966)) would be selected
based on the value of the regression coefficient. The argument against variety A is that because it
has a coefficient much higher than unity, it is too sensitive to environmental change and does
poorly in poor environments. Variety C, because it has a coefficient much lower than unity, is
unable to exploit high-yielding environments. Therefore, variety B, which is not superior in any
environment, is chosen as the best of the three.
Notice that the argument against variety A with a high coefficient, moves from right to left
or toward low environments (it does poorly in poor environments). The opposite is true of the
argument against variety C with a low coefficient, which moves from left to right or toward high
environments (it is unable to exploit good environments). These are negative interpretations which
lead to the selection of variety B, Figure 5.
If the emphasis regarding varieties with a high regression coefficient were toward, rather
than away from the best environments (which variety can exploit the better environments?), variety
A would be selected. Likewise, if for varieties with a low coefficient, emphasis were toward
(rather than away from) the poorer environments (which variety can maintain yield even in poorer
environments?), variety C would be selected, Figure 6. The difference is not one of analytical
procedure, but of a positive rather than a negative philosophy, goal and/or attitude toward
The result of using this approach with modified stability analysis (Hildebrand, 1984) is to
describe recommendation domains within which specific technologies excel (recommend variety A
for the better environments and variety C for the poorer environments, in the above example,
rather than variety B for all environments).
Numbers of locations (environments)
Following models (3) and (4), the number of environments required for estimation of
treatment by environment response in research domains and verification in recommendation
domains is not excessive. In order to have at least 20 degrees of freedom in the error term, and
allowing for estimation of both linear and quadratic responses as in model (4), if 8 treatments are
included in the trial, such as might be used in an exploratory trial in a research domain, 6
environments is an adequate number. For 4 treatments, 10 environments would be required, and
in a verification trial with only two treatments (the recommended treatment and the farmer check,
for example) 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
Numbers of years
Experience has indicated that if three conditions are met, the estimates of environment by
treatment response stabilize in one year. These conditions are:
1. The range of environmental indices (El) should be at least as great as the mean of the
2. The range of environmental indices should approximate what would normally be
expected over a period of years.
3. The distribution of environments should be reasonably uniform from good to poor.
However, it should be remembered that at least two years of data will be available for
estimates if both an exploratory trial (in a research domain) and a validation trial (in a
recommendation domain) are carried out prior to making firm recommendations. Also, preliminary
data often are available from on-station trials, conducted over previous years, as the technology is
being developed. The treatments that are common from among these current and previous trials
can be combined in a single MSA. The data from previous years can also help to verify whether
the range of environments included in a current trial is adequate.
Singh (1990) reports on recent research conducted near Manaus, Brazil, that illustrates
many of these concepts. The on-farm portion of his research was conducted in two small farming
communities in the municipality of Rio Preto da Eva, Amazonas, Brazil, where the government was
initiating a small watershed management program. The Brazilian national agricultural research
institution (EMBRAPA) has a mandate to develop appropriate technology for different farming
conditions in this relatively inaccessible area. Also collaborating in the research were EMATER
(extension) and SEPA, the state development planning entity, TROPSOILS, and the University of
Secondary information regarding indigenous farming practices of the area were collected
from published sources. A rapid appraisal of the area was conducted with a multidisciplinary team
of persons from EMBRAPA, SEPA and EMATER who visited the area on three different occasions.
Farmers' knowledge of indigenous technology, agronomic 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.
Three treatments, based on previous on-station research, were selected for comparison
with farmers' practices (FP) for growing maize (Zea mays L.) and cowpea (Vimna unguiculata).
Only results from the cowpea are reported here. All three treatments with amendments received K
(60 kg ha"1) broadcast. Processed city waste (PCW), chicken manure (CM) and triple super
phosphate (TSP) were applied in 25 cm bands. The cowpea variety IPEAN V-69 was planted in
rows 60 cm apart. Plot size varied from 100-200 square meters. Land preparation and planting
methods consisted of clearing the area by slash and burn, followed by manual land preparation and
planting with sticks.
The project area is inhabited by subsistence farmers who clear land from primary forest (PF)
or secondary forest (SF) and farm it up to three years before abandoning it as waste land (WL).
Cowpea trials were established on 13 locations. Eight were replicated and five were not. Yield
results, averaged across replications where appropriate, and the environmental index for each
location are shown in Table 1. Analysis of variance using the model (4) with R = 1 is shown in
For the criterion Mg ha-', the response of the four treatments to environment, using
modified stability analysis, is shown in Figure 7. It is clear that amendments are needed to
maximize per ha yield from these soils. In the poorer environments (El < 1.32) CM produces the
best results, and in the better environments (El>1.32) TSP is best.
The biophysical characteristics of the better and poorer environments closely follows the
nature of the land being used. That is, the better environments (El> 1.32) are all land taken from
PF and in first or second year of use and SF in first year of use. All other categories (PF,, SF,, SF3,
WL) are in the poorer environments. For farmers whose evaluation criterion is to maximize Mg ha"1,
the research domain can be divided into two recommendation domains. For farmers with PF, and
PF2, the recommendation is to use TSP. Farmers in all other cases should use CM.
Figures 8 and 9 show the MSA results for the alternate evaluation criterion of kg per dollar
of cash cost (kg/$CC), a criterion usually of great importance to farmers in this area who have little
cash to spend for agricultural inputs. Figure 8 shows that for the better environments (here
El> 1.25, but covering the same soil situations) FP is by far the best practice of those tested. In
the first or second year out of primary forest, none of the other tested treatments would be
acceptable to farmers for whom cash is very scarce and therefore need to maximize kg/$CC. For
any other soil situation, however, either TSP or CM could be recommended even if the farmers had
to use scarce cash to purchase the amendments.
The use of Figure 9 narrows the choice somewhat in the poorer environments. CM
produces very stable results compared to TSP which could result in fewer kg/$CC than FP. This
leads to the recommendation of CM as the best choice of those treatments tested in cases where
farmers use, or are forced by circumstances to use fields more than one year out of secondary
forest or two years out of primary forest.
FOCUS ON INDIVIDUAL FARMERS. FARMS. AND FIELDS IN COMMERCIAL AGRICULTURE
Farmers, themselves, continually fine-tune their systems to the specific resource and
infrastructure conditions in which the farm and family are found. In addition to the crop varieties
mentioned earlier, a number of innovations in farm equipment originated with producers. Perhaps
the agricultural machinery industry has been the sector most active in capturing the experience of
farmers and putting this into commercial practices. Many of the current tillage, planting, and
harvesting units have reached their present form through farmer modification of what was on the
market, and then tested and adopted by industry for the next generation of commercial units. The
ridge tillage planters and cultivators are currently going through this phase of farmer modification.
A number of cropping system innovations likewise originated with farmers. Annual windbreaks
have been proven useful to reduce transpiration in crops between the windbreaks, and perennial
windbreaks used to break the wind and trap moisture as snow in the Northern Great Plains.
Alternating strips of different species, such as maize and soybean, have been used by a number of
farmers in the Western Corn Belt. Although there is a growing body of technical research on
experiment stations to validate and quantify the effects of these practices, many of them in fact
originated with farmers in the region. What has been difficult is the rationalization of different
methods used by farmers to test their systems, and those used by scientists trained in a different
In many respects, on-farm research has a great deal in common with industrial statistical
process control. In manufacturing, products are designed in the lab, then prototypes are produced
and evaluated under "real world conditions." During the latter phase, problems are identified when
typical workers, rather than research engineers, attempt to produce the product and when
prospective consumers attempt to use the product. Invariably, they find ways to "break" the
product that would never occur to lab workers. So it is with agricultural research. The experiment
station or greenhouse can be thought of as the agronomic lab. The function of on-farm trials is
identical in agriculture to "real world" testing in manufacturing. Real farmers will surface problems
not encountered by experiment station workers; the research process is not complete until this is
DESIGNS FOR RESEARCH ON COMMERCIAL FARMS
In a recent symposium of the American Society of Agronomy in San Antonio, there were
many presentations about how research is being conducted on farms. The examples appeared to
fall into one of two categories. First was the replicated trial with relatively small plots in which the
university researcher developed an agenda, designed treatments and plots in the field, collected
most of the data and interpreted the results. The farmer was a participant in providing land and
some cultural operations during the season, but was not an active part of the planning or the
evaluation process. This role for the farmer meets most of the reasons for locating plots on farms
as listed by Lockeretz (1987).
In contrast, a second approach was essentially an extension of the farming systems
research/extension philosophy and method (Hildebrand and Poey, 1985), where farmers were
primary participants in the setting of a research agenda, search for relevant treatments, layout and
implementation of the trial, and interpretation and use of results. The latter approach provides an
environment in which the methodologies given by Taylor (1990) for on-farm research can be
implemented: use of multidisciplinary research teams (including the farmer), whole-farm analysis of
results where appropriate, design of long-term plots and treatments, and synthetic as well as
analytical approaches to use of data. In the United States, the former approach has been favored
by researchers from land grant universities, while the latter has been part of the agenda of farmer
groups and other non-profit organizations. The proponents of each approach find it difficult to
communicate at times with others who do not share their definitions of what constitutes research,
since each group has a relatively clear mind set of what is meant by "on-farm research", while in
fact these definitions are quite divergent.
When a university trained scientist uses the term "research", there is an assumption of an
explicit and testable hypothesis, replicated treatments in a randomized pattern in a standard design,
homogeneity of variances among treatments, control of experimental conditions, and relative
uniformity of the experimental area -- or some blocking pattern to handle variation in the field.
These are the normal assumptions connected with the analysis of variance, and although they are
not always strictly adhered to we often make the assumption that they are being met. Saying that
"standard statistical procedures were followed" implies all of the above even if the researcher (or
farmer) did not really understand the statistical thinking very well.
Many of these criteria are not recognized nor understood by most farmers. They prefer
trials that are fairly close to the home farm or under similar conditions or both, that have plots large
enough to use commercial equipment, that show visible differences among treatments, that can
reduce costs or increase profits, or that solve a constraint that was already perceived on their farm
or in the area (Francis, 1986a). In the real world we encounter comparisons from one year to the
next, from one field to another, from one farm to a neighbor's, or among strips in a field that have
different treatments (eg. varieties or hybrids) with no replication. Although these comparisons do
not meet the criteria recognized by the scientist to qualify as credible or valid research, the results
are no less meaningful to many farmers. We do find that careful explanation in an extension
meeting of some of the criteria used by researchers, for example replication and randomization of
treatments, leads to a fairly quick understanding of the need for these methods and the importance
for repeatability of the experience.
Are these two definitions of "research" mutually exclusive, or is there some middle ground
where farmer creativity, land, and resources can be utilized for credible on-farm research? Over the
past several years, there has been substantial work on large plots with few treatments, replication
and randomization, and standard statistical analysis. Long strip designs used to compare two or
three treatments were described by Thompson (1990), and are currently being used by a number of
the members of the Practical Farmers of Iowa, among other groups. Rzewnicki et al. (1988)
summarized these trials from Iowa as well as some from farms and from experiment stations in
Nebraska. With plots that ranged from 200 to 1200 feet long by four to eight rows wide and three
to six replications per treatment, they found coefficients of variation from less than 1.0 to about 10
percent; the CVs were frequently less than five percent. Practical researchers who are familiar
with the variation in most field experiments find these levels very acceptable.
How is it possible that such large plots have low CVs? Although we are only now testing
these hypotheses by comparing large and small plots from the same field (Shapiro et al., 1989,
1990), it appears that a long and narrow plot goes across a range of variability in the field. A plot
located adjacent with the same dimensions crosses the same gradient, and at any one point there
is relatively less difference between the strips than there is across the gradient in each long plot.
Thus the potential exists for planting contrasting treatments side by side, allowing use of full sized
commercial equipment and having a highly visible comparison, while still meeting the requirements
of replication and randomization. This would appear to be one option for an individual farmer to
collect credible data for one site in one year, and use standard statistical techniques such as
analysis of variance, t-test, or paired comparisons to evaluate the trials. In one set of comparisons,
the Clay County Corn Growers in Nebraska planted maize hybrids in unreplicated strip plots in four
areas in the county, with similar conditions and the same hybrids in each test. Analyzed with
farms as replications, there were CVs from three to four percent over the five years of the tests
(Rzewnicki et al., 1988). This opens the possibilities for individuals or groups of farmers to work in
a cooperative research network and to develop a credible set of comparisons for use by them and
by others. Each farmer becomes a part of the research and extension network, since these plots
are used for field tours and the data for extension meetings before the next planting season.
Results from these large replicated or unreplicated trials in Nebraska represent one approach
that can be taken by farmers in a highly mechanized, large farm situation. It is a challenge to the
practical researcher or applied extension person to explain the basic characteristics of the trials,
and to work directly with farmers in developing the research agenda.
INTERPRETATION AND EXTRAPOLATION OF RESULTS
The zone across which such results can be applied depends on how many sites were used
for the trials, the soil and climatic characteristics of the sites, whether similar results could be
expected from other sites in the region, and how credible or repeatable the results are from the
experiments. Several dimensions of this question have been discussed above. In statistical terms,
the potential for application of results across a range of environments depends on the significance
of the specific technology by environment interaction. An example is the testing of hybrids across
locations, and measuring the genotype by location interaction. When this is low, it is relatively
easy to recommend one or a few hybrids across a wide area; when the interaction is large, there is
a high degree of site specificity and need for unique choices for different locations.
It is important to consider the effects of replications, years, and locations in contributing to
the value of results. Increasing number of locations and environments had little effect beyond
about eight on the magnitude of the standard error of a genotype mean (Saeed et al., 1984).
Increasing number of years from one to two substantially reduced the standard error of the mean,
while adding an additional year had minimal effect. Likewise, increasing number of replications has
little effect on the standard error. The influence of additional locations or environments is much
greater than either adding years or replications to an experiment in order to reduce the variance of
a mean, thus increasing the potential for detecting statistical differences among treatments in the
experiment. Although it is less expensive to add replications in a single location, this is relatively
ineffective in increasing the potential to detect differences. This is consistent with the above
discussion on need for a large number of locations or environments for testing, and the relatively
smaller need for replication in one site. The concept of single replications and a large number of
locations is a cornerstone of current commercial hybrid testing strategies (Bradley et al., 1988).
The efficiency of this procedure in a testing program has recently been described (Dofing and
Francis, 1990). Replication at one site does improve the precision of measurement at that site in
that year. But with multilocational (multiple environment) on-farm testing, the relevant variance is
that among locations or environments. Therefore, multiple replications at one location contribute
little to the potential extrapolation from that site to others, or to other years.
The challenge for the individual farmer is to decide what information really applies to his or
her site, given the abundance of results from trials that are available from industry, university, or
private sources. The better a farmer is able to characterize the farm and the individual fields, and
the better the description of the conditions under which data were collected in other sites, the
easier it will be to decide which data or recommendations are relevant. This is a practical way of
defining recommendation domains, a topic already explored. The best place to look for relevant
data is within the same recommendation domain as that where the field is located. It should be
apparent that these domains are not defined only by geographical location, by soil type, or by any
single factor. Likewise, it is possible that a single farm may encompass several domains. It is
important to understand the concept, and to use this information to best access the most
appropriate data before making production technology decisions.
PARTICIPATORY MODEL FOR RESEARCH AND EXTENSION
On-farm research trials and demonstrations for extension purposes have long been a staple
component of comprehensive investigation and development programs in agriculture. Some of the
reasons have been described above. There are even more compelling reasons today why research
with individual farmers and groups of producers makes sense (Francis et al., 1990). There are
limited research and extension budgets, with an increasing focus of federal funds in the U.S. on
basic work at the expense of applied research. This is a trend that is being followed by national
research programs around the world, as scientists become better prepared for basic investigations
and the glamour of genetic engineering and high technology solutions pervades the scientific
community. In contrast to the range of ecological situations where farmers produce crops, the
research establishments have relatively few experiment stations. Much of the research performed
on these stations is reductionist in nature, with limited regard for the incorporation of new
innovations into the total farming system. For these reasons, there is comparative advantage to
conducting at least some of the research in a wider array of sites with collaborating farmers.
Another compelling reason for working directly with individual farmers and groups relates to
distance from the controlled research site to the farm where results will be applied. This "distance"
may take several forms (Francis et al., 1990). Geographic space in miles or kilometers from one
site to another is the most commonly used measure of distance. Farmers are willing to travel
certain distances to visit other sites, depending on culture and infrastructure (Rzewnicki, 1991).
More important, perhaps, is the "ecological distance" from one site to another. For example, a low
lying area with poor drainage and heavy soils may be a very short physical distance from a well
drained, lighter soil on a hillside, yet the soil conditions, appropriate cultural practices, and crops or
varieties that are appropriate may be quite distinct. Finally, there may be "conceptual" or
"psychological distances" between researcher and farmer, based on differences in education or
experience, and these need to be bridged in order to effect a working partnership and a fully
participatory system of research and extension. On- farm activities among people who have
mutual respect for each others' talents and potentials to contribute can help to overcome these
The potentials of a participatory network of farmers and researchers can perhaps best be
illustrated through use of an example. The unique contributions of the farmer in the total research
process is highlighted. A number of additional examples, especially in farmer contributions to ideas
for weed management, were recently summarized by Francis and Doll (1991).
Maize yield response to nitrogen in crop rotations. In order to study the effect of nitrogen
applications on maize yields in continuous maize and sorghum compared to rotation of these
cereals in Nebraska, a network of about thirty farmers was established to work with a project of
the University of Nebraska. There has been great concern about the energy costs of this input in
maize production, as well as potential for nitrate contamination of ground water supplies that are
frequently used for human and animal consumption. Supported in part through a grant from the
Nebraska Energy Office, a university technician established contact with a number of farmers,
many of whom were members of the Nebraska Sustainable Agriculture Society. All were interested
in more efficient use of nitrogen, and in finding ways to quantify the effects of a cereal-legume
rotation on response to this important nutrient.
In cooperation with farmers, fields and experimental sites were chosen, soil samples were
taken, and lab test results discussed. Together the team determined realistic yield goals and
developed nitrogen budgets considering all sources of this major nutrient. Each farmer thus derived
a conservative but optimum level of nitrogen for the coming season. In most fields this N rate and
a one-half rate were included, and in some fields a zero rate as well. On many fields these
treatments were applied in replicated strips across the entire field. For the years 1988, 1989, and
1990 there was a predicted yield response to rates of 80 to 150 pounds N/acre, although the
actual economic optima were lower in most fields (Franzluebbers, 1991). Following soybean,
sweet clover, or alfalfa there was no economic response to applied nitrogen by either maize or
sorghum under rainfed conditions. In irrigated fields, there was no economic response of maize
yields to nitrogen if the maize followed alfalfa. The conclusions from this three-year project, as
interpreted by farmers and project personnel, was that nitrogen generally is over applied under
many conditions in Nebraska.
Statistical analysis followed the standard procedures described above in model (1). Where
replicated treatments were present, the analysis and comparison were conducted on each farm. In
a number of cases there was a single replication per farm, and these results were pooled with the
replicated sites using a single mean per treatment per location, and locations used as replications.
Regression analysis was used to compare the response of cereal yields to applied nitrogen in
rotation compared to continuous culture. Grain yield response in maize and sorghum to applied
nitrogen was measured at 29 sites with continuous cropping and at 57 sites in rotation with
legumes and small grains. Continuous maize responded to nitrogen up to about 80 kg N/ha, the
maximum level in the trials. Maize and sorghum following small grains or legumes showed only a
modest response in some cases, not statistically significant, and not economically sound because
the cost of nitrogen plus application was not offset by the increase in yield. This type of analysis
is useful for grouping results of like treatments across sites.
To reach more farmers with this information, eight meetings were scheduled jointly by the
Sustainable Agriculture Society and Nebraska Extension in early 1990. The objectives of the trials
and methods were described, tables or figures presented, and the meeting turned over to farmers
to interpret the data and derive results. A lively discussion ensued about results from the trials and
how to apply them to specific fields. The university staff present were valuable as resource
people, helping to explain why or why not crops were responding in specific situations. But the
farmers were deriving their own recommendations, and the extension specialists were able to
empower the producers to make these decisions. During the next summer of 1990, many of the
trial fields were used for field tours and discussions on site. Farmers were in charge of describing
what happened. These are both examples of participatory extension practices.
Should the farmer put more confidence in results from his or her own field trial, or from the
aggregate analysis across sites? It is usually appealing to have one's own data from a field on the
farm, where the cultural practices are known and the results appear to uniquely fit that farm.
Whether these are the best data to use to predict next year's results depends on the similarity of
cultural practices, hybrids, and soil conditions across the range of sites, and how likely those sites
represent the potential range of possible rainfall events that may occur over a number of years.
Since rainfall is the most limiting factor in most Nebraska sites each year in rainfed crop culture, it
is possible that the mean performance over several similar sites will be a better predictor of next
year's situation than the results from the single farm. We are seeking data and a method to
analyze this situation. The decision by an individual farmer at the moment is a judgement call, and
the best that we can do is to provide tools to help improve that judgement.
RESEARCH-EXTENSION AGENDA: LARGE AND SMALL FARMERS
The approach and examples presented in this chapter illustrate the potential of an emerging
paradigm, or shift in patterns of research and extension activities. This applies to both large and
small farmers. Efficient research of the type being discussed here depends on recognition and full
representation of the research domain or inference space. Scientists need to recognize that on-
station trials are useful for research and development, but limited for surfacing knowledge about
production realities. Both large and small farmers need to recognize that they represent these
realities, and in a major way can contribute to their identification and solution.
Once the inference space is well understood, it is possible to follow with carefully designed
activities to solve the problems associated with the primary constraints to sustainable production
using resources and information from both large and small farms. Data gathered in-situ is critical,
and on-farm trials that include the widest possible variety of farms are essential. Failure to include
this range of environments can result in compromised research (bad for the scientist) that may
produce flawed information and incorrect production decisions (bad for farmers). The need for
many environments and a wide range of participants should be clear.
Large-scale farmers often have the resources to contribute to the research process, and can
easily grasp the relevance of self-generated information to their own operations. Using the
principles described above, their efforts can be more productive if combined across locations, and
at times combined with data from small-scale farm trials. A wide coalition of farmers, researchers,
and extension specialists can bring together the resources needed to broaden the range of
environments over which results can be collected. If large farmers are incorporated into a
technology evaluation network with small farmers, the length of time required for testing can be
reduced because environments (locations) can substitute substantially for years. This can lead to
greater efficiency in use of scarce resources, both from farmers and the government. It may well
be possible for many of the larger farmers, in conjunction with the national research and extension
organizations, to pick up costs associated with on-farm research that will benefit them, and at the
same time enhance the development of technology for all farmers.
The role of the research and extension specialists in this system needs to be clearly
recognized. The upgrading of farmer participation and input into the research process does not
devalue the scientists' role, but rather expands their capacity to recognize and work with real world
problems that limit production and the range of solutions that may be available to solve them. The
role of research specialists can, in fact, become more focused on describing the "why" behind
questions in agriculture, ecology, biology and sustainability. The role of extension specialists can
be to catalyze the exchange of information among a number of credible and relevant sources. The
research process can be a rigorous statistical exercise, and it is possible to determine where results
can be applied in the appropriate recommendation domains.
Assumptions about roles of different players, cooperation among farmers and scientists,
relevance of information from various sources, and ultimate objectives of the systems need to be
recast. This is the paradigm shift described above. Many agricultural infrastructures are set up
with good intentions, but fail to produce anticipated results due to inadequate communication,
limited scientific literacy among specialists and farmers, and a strained relationship between those
who develop theoretical knowledge and those who focus on practical application. This is a
problem in both developing and developed countries. The farming systems paradigm, and
especially the on-farm research approaches described here can offer enormous potential that will
benefit national agricultural infrastructure as well as sustainable agricultural production systems.
Bradley, J. P., K. H. Knittle, and A. F. Troyer. 1988. Statistical methods in seed corn product
selection. J. Prod. Agric. 1:34-38.
Braidwood, R. J. 1967. Prehistoric men. Scott, Foresman and Co., Glenview, Illinois.
Byerlee, D., L. Harrington and D.L. Winkelmann. 1982. Farming systems research: Issues in
research strategy and technology design. J. Am. Econ. Assoc. 5:899-904.
Chambers, R. 1981. Rapid rural appraisal: Rationale and repertoire. Public Administration and
Chambers, R., A. Pacey, and L. A. Thrupp. 1989. Farmer first: Farmer innovation and agricultural
research. Intermediate Technol. Publ., London.
Cox, D.R. 1958. The Planning of experiments. Wiley, New York.
Deming, W.E. 1953. "On the distinction between enumerative and analytic surveys." Journal of the
American Statistical Association, 48: 244-255.
Deming, W.E. 1975. "On probability as a basis for action." American Statistician, 29: 131-137.
Doing, S. M., and C. A. Francis. 1990. Efficiency of one-replicate yield testing. J. Prod. Agric.
Eberhart, S.A., and W.A. Russell. 1966. Stability parameters for comparing varieties. Crop Sci.
Finlay, K.W., and G.N. Wilkinson. 1963. The analysis of adaptation in a plant breeding
programme. Aust. J. Agric. Res. 14:742-754.
Francis, C. A. 1986a. Designs for research: the quest for credibility. Seminar, MTDI Workshop
on Small Farm Enterprises, Ames, Iowa, May 12.
Francis, C. A., editor. 1986b. Multiple cropping systems. Macmillan Publ. Co., New York.
Francis, C. A., J. King, J. DeWitt, J. Bushnell, and L. Lucas. 1990. Participatory strategies for
information exchange. Amer. J. Alternative Agr. 5:000-000 (in press)
Francis, C. A., and J. D. Doll. 1991. Participatory research and extension strategies for
sustainable agricultural systems. Weed Science (submitted).
Franzluebbers, A. 1991. On-farm results of fertilizer response in crop rotations. M.S. Thesis,
Univ. of Nebraska, Lincoln.
Freeman, G.H. 1973. "Statistical methods for the analysis of genotype-environment interaction."
Heredity, 31: 339-354.
Gilbert, E. H., D. W. Norman, and F. E. Winch. 1980. Farming systems research: a critical
appraisal. MSU Rural Devel. Paper No. 6, Dept. Agr. Econ., Michigan State Univ., East Lansing,
Harrington, L.W. and R. Tripp. 1984. Recommendation domains: A framework for on-farm
research. CIMMYT Economics Program Working Paper 02/84. CIMMYT, Mexico.
Harville, D.A. 1976. "Extension of the Gauss-Markov theorem to include the estimation of random
effects." Annals of Statistics, 4: 384-395.
Henderson, C.R. 1975. "Best linear unbiased prediction under a selection model." Biometrics, 31:
Hildebrand, P.E. 1981. Combining disciplines in rapid appraisal: The sondeo approach.
Agricultural Administration. 8:423-432.
Hildebrand, P.E. 1984. "Modified stability analysis of farmer managed, on-farm trials." Agronomy
Journal, 76: 271-274.
Hildebrand, P.E. 1988. Technology diffusion in farming systems research and extension.
Hildebrand, P.E. 1990. Modified stability analysis and on-farm research to breed specific
adaptability for ecological diversity. In: Kang, M.S. Genotype-by-environment interaction and
plant breeding. Louisiana State University. pp.169-180.
Hildebrand, P. E., and F. Poey. 1985. On-farm agronomic trials in farming systems research and
extension. Lynn Rienner Publ, Inc., Bolder, Colorado. 162 pp.
Hill, J. 1975. "Genotype-environment interactions a challenge for plant breeding." Journal of
Agricultural Science, 85: 477-493.
Lockeretz, W. 1987. Establishing the proper role for on-farm research. Amer. J. Alternative Agr.
McCullagh, P. and Nelder, J.A. 1989. Generalized linear models. Chapman and Hall, New York.
McLean, R.A. 1989. "An introduction to general linear models." in applications of mixed models in
agriculture and related disciplines. Southern Cooperative Series Bulletin No. 343, pp. 9-38.
Louisiana Agricultural Experiment Station, Baton Rouge.
Mead, R. 1988. The Design of experiments: statistical principles for practical application.
Cambridge University Press, Cambridge, UK.
Norman, D. W. 1980. Farming systems approach: relevance for the small farmer. Rural Devel.
Paper No. 5, Dept. Agr. Economics, Michigan State Univ., East Lansing, Michigan.
Plucknett, D. L., and N. J. H. Smith. 1986. Historical perspectives on multiple cropping. Ch. 2 in:
Multiple Cropping Systems, C. A. Francis (editor). Macmillan Publ. Co., Inc., New York. pp. 20-39
Rhoades, R. 1989. The role of farmers in the creation of agricultural technology. Ch. 1.1 in:
Farmer first: Farmer innovation and agricultural research, R. Chambers, A. Pacey, and L. A. Thrupp,
editors. Intermed. Technol. Publ., London. pp. 3-9.
Rzewnicki, P. E., R. Thompson, G. W. Lesoing, R. W. Elmore, C. A. Francis,A. M. Parkhurst, and R.
S. Moomaw. 1988. On-farm experiment designs and implications for locating research sites.
Amer. J. Alternative Ag.r 3:168-173.
Saeed, M., C. A. Francis, and J. F. Rajewski. 1984. Maturity effects on genotype x environment
interactions in grain sorghum. Agron. J. 76:55-58.
Shapiro, C. A., W. L. Kranz, and A. M. Parkhurst. 1989. Comparison of harvest techniques for
corn field demonstration. Amer. J. Alternative AGr. 4:59-64.
Shapiro, C. A., A. M. Parkhurst, and W. L. Kranz. 1990. Update on the statistical implications of
experimental units for on-farm trials. Agron. Abstr. p. 30.
Singh, B.K. 1990. Sustaining crop phosphorus nutrition of highly leached oxisols of the Amazon
Basin of Brazil through use of organic amendments. PhD Dissertation. University of Florida.
Snedecor, G.W. and Cochran, W.G. 1980. Statistical methods, 7th edition. Iowa State University
Steel, R.D.G. and Torrie, J.A. 1980. Principles and procedures in statistics: a biometrical
approach, 2nd edition, McGraw-Hill, New York.
Stroup, W.W. 1989a. "Predictable functions and prediction space in mixed linear models." In:
Applications of mixed models in agriculture and related disciplines. Southern Cooperative Series
Bulletin No. 343, pp. 39-48. Louisiana Agricultural Experiment Station, Baton Rouge.
Stroup, W.W. 1989b. "Why mixed models?" in Applications of mixed models In: Agriculture and
related disciplines. Southern Cooperative Series Bulletin No. 343, pp. 1-8. Louisiana Agricultural
Experiment Station, Baton Rouge.
Taylor, D. C. 1990. On-farm sustainable agricultural research: lessons from the past, directions
for the future. J. Sustainable Agr. 1(2):43-87.
Thompson, R. 1990. The Thompson-farm on-farm research. Rodale Institute, Emmaus,
Pennsylvania. 68 pp.
Tripp, R., and J. Woolley. 1989. The planning stage of on-farm research: Identifying factors for
experimentation. Mexico, D.F. and Cali, Colombia: CIMMYT and CIAT.
Westcott, B. 1985. ""Some methods of analysing genotype-environment interaction." Heredity, 56:
Wotowiec, P., S. Poats and P.E. Hildebrand. 1988. Research, recommendation and diffusion
domains: A farming systems approach to targeting. Chapter 6 In: Poats, S., M. Schmink and A.
Spring. Gender issues in farming systems research and extension. Westview Press, Inc. Boulder,
Table 1. Cowpea yield (Mg/ha) across environments,
Rio Preto da Eva, Amazonas, Brazil, 1989
LOCATION FP PCW TSP CM El
8 0.10 0.20 1.30 1.65 0.81
13 0.00 0.00 1.30 2.00 0.82
6 0.15 0.50 1.35 1.35 0.84
9 0.20 0.40 1.20 1.70 0.88
2 0.50 0.65 1.10 1.50 0.94
5 0.15 0.50 2.10 2.05 1.20
4 0.60 1.20 1.60 2.25 1.41
1 0.70 0.90 2.30 1.80 1.42
11 1.20 1.50 2.20 1.90 1.70
12 1.50 1.80 2.10 1.70 1.78
3 1.45 1.95 2.50 1.90 1.95
10 2.20 1.90 2.60 1.40 2.02
7 1.70 1.65 2.65 2.15 2.04
Source: Singh (1990).
Table 2. ANOVA, Cowpea response to environment,
SOURCE OF VARIATION df MEAN SQUARE Pr > F
Location 12 0.9470 0.0001
Treatment 3 3.8127 0.0001
Env*Trt 3 0.9923 0.0001
Env*Env*TrT 3 0.1266 0.1840
Residual 30 0.0736
Figure 1. Possible results of on-station testing.
MAY HAVE PERFORMED
WELL ON FARMS
Figure 2. Possible results of on-station testing.
Figure 3. Representation of a typical on-farm trial.
2 3 4 5 6 7
Figure 4. Illustration of treatment by environment interaction.
o I I I I I
0 1 2 3 4 5 6
ENVIRONMENTAL INDEX, El
Figure 5. Hypothetical results of variety testing over range of environments.
ENVIRONMENTAL INDEX, El
Figure 6. Negative interpretation of the response of varieties to environment resulting in choice
of variety B for "broad adaptation".
0 1---I I I I I
0 1 2 3 4 5
ENVIRONMENTAL INDEX, El
Figure 7. Positive interpretation of the response of varieties to environment resulting in a choice
of variety A for the better environments and of variety C for the poorer environments.
1 1.2 1.4 1.6 1.8 2 2.2
ENVIRONMENTAL INDEX, El
WL SF, SF, PF, SF PF2 PF,
Figure 8. Cowpea response (MG/HA) of four treatments to environment,
Manaus, Brazil 1990 (Singh)
1 1.2 1.4 1.6 1.8 2
ENVIRONMENTAL INDEX, El
SWL SF, SF, PF, SF, PF2 PF,
Figure 9. Cowpea response (KG/$CC) of four treatments to environment
Manaus, Brazil, 1990 (Singh)
. I I.1
Z f I
z : :\
SLU I I
r 80 8
100 I ---
-10 -5 0 5 10 15 2(
Figure 10. Stability of three treatments in cowpea for poor environments
Manaus, Brazil, 1990 (Singh)
p~. -j- l"
Yr -'----k-' -
-4Y ~ -,*---- .
The publications in this collection do
not reflect current scientific knowledge
or recommendations. These texts
represent the historic publishing
record of the Institute for Food and
Agricultural Sciences and should be
used only to trace the historic work of
the Institute and its staff. Current IFAS
research may be found on the
Electronic Data Information Source
site maintained by the Florida
Cooperative Extension Service.
Copyright 2005, Board of Trustees, University