Abstract:

Utilisation of plant resistance for insect pest management
requires identification of resistant genotypes and development of
cultivars combining high levels of resistance with other desirable
usually screened under suitable conditions to identify those with
resistance to the insect pest in question using appropriate measures.
A basic limitation to conventional selection methods is the
choice of cut-off point(s) for efficient partitioning of genotypes into
resistance/susceptible groups and the assumption that resistance traits
contribute equally and independently to yield loss. This makes it
difficult to effectively distinguish and/or delineate genotypes into
resistance/susceptible groups and may thus hinder selection and
consequently delay or derail the development of resistant genotypes.
This study considered objective statistical approaches to the
classification and selection problems. The aim was to optimise the
selection, by maximising the differences between the selected group
and the other group(s) while minimising the differences within the
selected group. Procedures for partitioning crop genotypes into
resistance classes, based on observed resistance levels, were
proposed and the potential for their use in selecting genotypes was
evaluated.
characteristics. Consequently, a large number of genotypes are
For the single-trait selection, the three categories of procedures
graphical method where plots of ordered means
against the corresponding normal scores were used to determine the
optimum number of groups and their limits. The second is based on
cluster analysis techniques. In the third procedure, classification
models were formulated and hypothesis testing using the likelihood
ratio and the quasi bayesian test statistics facilitated the identification
of the most plausible points to subdivide the entries under study into
distinct groups.
Cluster analysis (CA) and likelihood ratio test (LRT) procedures
were investigated for multi-trait selection, focussing on
interrelationships among the resistance variables. Use of principal
component analysis (PCA) generated indices (principal components
(PCs)) was compared with some 3 commonly used indices, viz, rank
summation index (RSI), Smith-Hazel index (SHI) and Dobie's index of
susceptibility (I).
The practical application of the procedures was investigated
using two groups of real field data and the comparison of their relative
performance was made through simulation studies.
The results show that the hypothesis testing procedure, overall,
performed best. The LRT statistic plots for the 2-group model, in
addition to locating the optimum cutpoint for the model, gave a clear
picture of the entire structure of the data, especially when distinct
groups existed. Its power to detect the 1-group model, as shown
1
investigated include
of the quasi bayesian test statistic gave poor results.
Cluster Analysis, in general, gave fairly similar results to the
LRT procedure. The method was fairly robust to distributional
assumptions and easy to use.
Normal score plots were fairly precise at determining the
optimum number of groups in the data and, used along with other
procedures, the plots gave satisfactory and useful results.
The results further showed that, for multivariate classifications,
the appropriate procedure and index to adopt depended heavily on
intercorrelations among the variables. Simulation studies indicated
that classification was greatly facilitated by the existence of high
correlations among variables. In this case, promising results were
obtained using PCs as the basic variables for classification, in
comparison to RSI, SHI and I.
Overall, the optimal conditions for all the classification
procedures included equal groups sizes and high intercorrelations
among variables.
Application of the proposed procedures to real data
demonstrated that the conventional methods of pre-setting selection
conditions, in form of group limits, selection intensity, etc., could lead
to enormous loss of useful materials.
It was concluded that, other than being alternatives for solving
the classification and subsequent selection problems, the procedures
from