Genetic coefficients are constants that enable crop simulation models to mimic the phenological and physiological
idiosyncrasies of individual varieties. Methods have been developed for estimating genetic coefficients from research data
or from public variety trials. However, the degree of experimental design regularity present in public trials is largely absent
from their private counterparts. The question therefore arises as to whether estimation techniques developed with public data
will work in the private–sector environment. This was evaluated using a set of 1155 yield observations representing
221 varieties grown in 1997, 1998, and 1999 at 105 site–year combinations as part of the on–going testing program of the
Delta and Pine Land Company.
The data were divided into three groups: varieties with 15 or more observations, those with 5 to 14 observations, and those
with four observations or less. The first group (24 varieties) had sufficient data to support genetic coefficient estimation. Four
of the 24 had 10 or more observations in each of the three years. These four were quarantined to ensure their independence
and formed an Evaluation Set. The remaining variety groups were used to characterize locations. Corresponding weather data
were obtained from the National Climatic Data Center. Soil features were first estimated by an automated procedure based
on reported texture and then adjusted manually within published limits to reproduce maturity group average yields at each
site. The Evaluation Set was not used in this latter process. Optimal variety and site parameters were then estimated by
searching a prediction database pre–calculated by parallel processing. The variety parameters were critical short day length
and an index interrelating a set of phenological attributes. The site parameters were rooting profiles that fine–tuned water
availability. Cross–validation was used to estimate root mean square errors of yield prediction.
Results indicate that it is possible to extract coefficient estimates from commercial data, and that these estimates can be
used to predict outcomes in independent situations. However, those situations must be from the same statistical population
as the original calibration data. The stability of the estimates obtained was strongly dependent on the manner in which
varieties with 5 to 14 observations were utilized. Although more than one mechanism seemed to be at work, decreasing
importance of rooting profile estimation and easing restrictions on free genetic coefficients were associated with improved
parameter stability as new data were added. It was also apparent that actual management use of the resulting estimates would
require better characterization of soils than is currently present in performance trial data.