of stability analysis, and the reasons to use on-farm testing, should change if we are to move, effectively and efficiently, toward a more .sustainable agriculture.
DISCUSSION
Analysis of variance is an indispensable tool to evaluate replicated research conducted at only one location. It can be used by a skilled statistician to evaluate multilocation research, but it is easy to make errors (in part depending upon whether the locations are considered fixed or random) if the analyst is not skilled (McIntosh, 1983). Furthermore, in the field, but seldom published, it is not uncommon to see an analysis showing significant location-by-treatment interaction that masks mean treatment differences (the concern of Yates and Cochran). The
researcher concludes it is not possible to detect treatment differences (thereby making the results non-publishable) and
recommends either repeating the experiment or controlling the sources of *interaction in a new experiment. 'This latter recommendation leads many researchers to *try to minimize differences among locations by stratification: by environmental control techniques such as irrigation, fertilization, and rigorous
pest control: by excluding low-yielding environments, and/or by rejecting on-farm testing in favor of the experiment station..
The result of the above tendencies is that much testing of germplasn is carried out under favorable conditions and superior environments.' This reduces the range of environments in which the materials are evaluated and limits the validity of the results to similar favorable environments. In discussing pre-1980 corn (Zea
may L.) selection procedures by seed companies, Bradley et al. (1988) state:
"High-yield environments were chosen f or .testing locations because yield was the trait of most interest and- because high yields gave low coefficients of variation. Outliers (values exceeding the researcher's
expect-ed data set) were often excluded f rom the data summaries. Stress environments were avoided when choosing locations or, if they occurred due to seasonal stress, they were often excluded from data summaries...
As a result of past performance testing systems, hybrids
introduced to the product line generally had high yields in high-yielding environments. They also had many
defensive problems such as barrenness, broken stalks, dropped ears, and root-lodging because these traits are most likely to occur in stress environments, which were least likely to have been sampled or reported... Customer acceptance of most hybrids was poor because the marketplace *itself was often used to evaluate the
hybrids" (pp. 34-35).
There are two consequences of this type of approach. First,' the results apply only to farms and farmers with similar, usually
superior environments. The only reason hybrids (or other material) bred and tested under these conditions have had any successful adoption is that many farmers in the developed world, and a select few in developing countries, have had the resources and capability
173