1. multivariate genotype - environment data
Analysis of
multivariate genotype - environment data
using Non-linear Canonical Correlation Analysis
Hans Pinnschmidt
Danish Institute for Agricultural Sciences
Division of Crop Protection
Cereal Plant Pathology Group
Denmark
Archived at

2. Background Objectives
BAROF WP1 data: multivariate measurements on 86 spring barley genotypes in 10 environments (2 years: 2002 & 2003, 3 sites: Flakkebjerg, Foulum, Jyndevad, 2 production systems: ecological & conventional).
Objectives
Multivariate characterisation of genotypes with emphasis on yield-related properties.

3. } derive information on
factors:
genotype environment
G1 E1
. .
. Ej . Gi
variables:
X1(i,j) ... Xm(i,j)
parameters
Xm(i)1 ... Xm(i)p
Xm(j)1 ... Xm(j)p
variables:
yield
1000 grain weight
grain protein contents
culm length
date of emergence
growth duration
mildew severity
rust severity
scald severity
net blotch severity
disease diversity
weed cover
broken panicles & culms
lodging
parameters:
raw data
mean/median/max./min.
rank/relative values
main effects
interaction slopes
raw data adjusted for E/G main effects/slopes (residuals)
IPCA scores
SD/variance }
derive information on
general properties, specificity, stability/variability

4. Non-linear Canonical Correlation Analysis (NCCA): an optimal scaling procedure suited for handling multivariate data of any kind of scaling (numerical/quantitative, ordinal, nominal).

5. Non-linear Canonical Correlation Analysis (NCCA)
data treatment: quantitative variables (vm) were converted into ordinal variables with n categories (v v1n, ..., vm1 ... vmn).

6. Non-linear Canonical Correlation Analysis (NCCA)
is based on multivariate contingency tables containing frequency counts. G1 . Gi
E Ej
Vm Vmn

7. Non-linear Canonical Correlation Analysis (NCCA):
main “dimensions” ( principal components) are determined
“loadings” of variables ( overall correlation) are computed
“category centroids” are quantified
“object scores” ( principal component scores) are computed

8. Characterisation of environments
based on data adjusted for
G main effects (= residuals)

10. Flakkebjerg 2002:
high rust &
1000 grain weight
late sowing
Foulum 2002 conventional
& Jyndevad 2003 ecological:
high mildew & lodging
Flakkebjerg 2003:
high yield, net blotch &
panicle breakage
low mildew & lodging
Jyndevad 2002 ecological:
low yield, 1000 grain weight,
weed infestation, protein content

11. Characterisation of genotypes
based on data adjusted for
E main effects (= residuals)

13. dimension 5 (sq. root) dimension 1 (sq. root)
high yield & 1000 grain weight
low protein content & lodging
high mildew
low net blotch & disease diversity
low yield &
1000 grain weight
low mildew

14. genotypes & environments based on:
Characterisation of
genotypes & environments
based on:
raw data
data adjusted for E main effects
data adjusted for G & E main effects
( G x E interaction)

15. low yield, 1000
grain weight, weed
infestation &
net blotch
high mildew
high rust
late emergence
high yield, 1000 grain weight & net blotch
low mildew
low rust
short culms
early emergence

16. high yield & 1000
grain weight
low protein content
& lodging
low net blotch &
disease diversity
high mildew
low yield &
1000 grain weight
high protein content

17. little lodging
high panicle breakage
high yield &
1000 grain weight
low protein content
low yield &
1000 grain weight
high protein content
much lodging

18. Conclusions & outlook
NCCA is an “intuitive” method good for “visualising” the main features in multivariate data of various scales.
NCCA is useful for obtaining an overall orientation of G properties and E characteristics.
Future work:
Refinements to obtain a better synopsis of E-specific performance of G’s as related to their property profiles.
Include AMMI- and clustering (biclassification) results in NCCA, organise data as environment-specific sets of variables.

19. Characterisation of genotype performance in
individual environments based on:
raw yield- and disease data
disease main effects of G’s
environmental disease variability of G’s
(= standard deviation of E adjusted data)