172
K. M. Sørensen et al.
location or year as segments. The segments should be representative for what
you would like the model to be able to predict.
2. Do not use full cross-validation (leave out one sample at a time) unless you have
very few samples [44]!
3. Replicates must always end up in the same cross-validation segments!
4. If you have several experimental design factors, try to use the different factors
as segments in the cross-validation, i.e., batch, year, variety, location, etc.
5. If different models are approximately equally good, be conservative and select
the model that is most parsimonious (i.e., uses fewest latent factors).
6. If different cross-validation methods suggest different numbers of latent factors,
try to repeat a method, where samples are randomly split into (relatively few)
segments.
The application of other more advanced validation methods like double crossvalidation, permutation and Monte Carlo testing often adds complementary insight
of the model performance (Westerhuis et al. 2008). However, in most cases following
the rules above will be adequate to validate the multivariate models. While validation
allows for the assignment explained of variance in the PCA models, it is primary
when it comes to regression/prediction models that validation becomes crucial for
assigning measures of accuracy to the models in terms of bias, variance, confidence
intervals, prediction errors, etc., and to the determination of number of components.
7.7 Variable Selection in Regression
Too much data—too little information!
—Harald Martens, Norwegian chemometrician
The spectral range of NIR instruments is typically determined by hardware components such as optical materials, light sources and detectors. The spectral region for
a given instrument may thus not be optimal for your application. The multivariate
advantage has been amply demonstrated for PCA and PLS applications to NIRS
data, but how much multivariate is enough and how much is too much?
In principle, two or a few, covarying neighboring variables should suffice to
provide the multivariate advantage. Despite the high redundancy in NIR spectra,
the multivariate methods can often be improved by variable selection. The primary
reason for the improvements is the reduced number of interferences in the reduced
set of variables, but also because of the fact that the data structure in the NIR region
has different behaviors across the NIR region, in particular in the shortwave (SW)
NIR region, the first overtone region and the combination tone region. Feeding all
this variation to PCA or PLS may deteriorate the performance. It can be an advantage
to get rid of the irrelevant spectral regions and spectral regions which contain mostly
K. M. Sørensen et al.
location or year as segments. The segments should be representative for what
you would like the model to be able to predict.
2. Do not use full cross-validation (leave out one sample at a time) unless you have
very few samples [44]!
3. Replicates must always end up in the same cross-validation segments!
4. If you have several experimental design factors, try to use the different factors
as segments in the cross-validation, i.e., batch, year, variety, location, etc.
5. If different models are approximately equally good, be conservative and select
the model that is most parsimonious (i.e., uses fewest latent factors).
6. If different cross-validation methods suggest different numbers of latent factors,
try to repeat a method, where samples are randomly split into (relatively few)
segments.
The application of other more advanced validation methods like double crossvalidation, permutation and Monte Carlo testing often adds complementary insight
of the model performance (Westerhuis et al. 2008). However, in most cases following
the rules above will be adequate to validate the multivariate models. While validation
allows for the assignment explained of variance in the PCA models, it is primary
when it comes to regression/prediction models that validation becomes crucial for
assigning measures of accuracy to the models in terms of bias, variance, confidence
intervals, prediction errors, etc., and to the determination of number of components.
7.7 Variable Selection in Regression
Too much data—too little information!
—Harald Martens, Norwegian chemometrician
The spectral range of NIR instruments is typically determined by hardware components such as optical materials, light sources and detectors. The spectral region for
a given instrument may thus not be optimal for your application. The multivariate
advantage has been amply demonstrated for PCA and PLS applications to NIRS
data, but how much multivariate is enough and how much is too much?
In principle, two or a few, covarying neighboring variables should suffice to
provide the multivariate advantage. Despite the high redundancy in NIR spectra,
the multivariate methods can often be improved by variable selection. The primary
reason for the improvements is the reduced number of interferences in the reduced
set of variables, but also because of the fact that the data structure in the NIR region
has different behaviors across the NIR region, in particular in the shortwave (SW)
NIR region, the first overtone region and the combination tone region. Feeding all
this variation to PCA or PLS may deteriorate the performance. It can be an advantage
to get rid of the irrelevant spectral regions and spectral regions which contain mostly