Performs a careful evaluation by repeated double-CV for multivariate regression methods, like PLS and PCR.
mvr_dcv(formula, ncomp, data, subset, na.action,
method = c("kernelpls", "widekernelpls", "simpls", "oscorespls", "svdpc"),
scale = FALSE, repl = 100, sdfact = 2,
segments0 = 4, segment0.type = c("random", "consecutive", "interleaved"),
length.seg0, segments = 10, segment.type = c("random", "consecutive", "interleaved"),
length.seg, trace = FALSE, plot.opt = FALSE, selstrat = "hastie", ...)
array [nrow(Y) x ncol(Y) x repl] with residuals using optimum number of components
array [nrow(Y) x ncol(Y) x repl] with predicted Y using optimum number of components
matrix [segments0 x repl] optimum number of components for each training set
array [nrow(Y) x ncol(Y) x ncomp x repl] with predicted Y for all numbers of components
SEP over all residuals using optimal number of components
spread of inner half of residuals as alternative robust spread measure to the SEPopt
MAD of residuals as alternative robust spread measure to the SEPopt
MSEP over all residuals using optimal number of components
final optimal number of components
vector of length ncomp with final SEP values; use the element afinal for the optimal SEP
formula, like y~X, i.e., dependent~response variables
number of PLS components
data frame to be analyzed
optional vector to define a subset
a function which indicates what should happen when the data contain missing values
the multivariate regression method to be used, see
mvr
numeric vector, or logical. If numeric vector, X is scaled by dividing each variable with the corresponding element of 'scale'. If 'scale' is 'TRUE', X is scaled by dividing each variable by its sample standard deviation. If cross-validation is selected, scaling by the standard deviation is done for every segment.
Number of replicattion for the double-CV
factor for the multiplication of the standard deviation for the determination of the optimal number of components
the number of segments to use for splitting into training and test
data, or a list with segments (see mvrCv)
the type of segments to use. Ignored if 'segments0' is a list
Positive integer. The length of the segments to use. If specified, it overrides 'segments' unless 'segments0' is a list
the number of segments to use for selecting the optimal number if
components, or a list with segments (see mvrCv)
the type of segments to use. Ignored if 'segments' is a list
Positive integer. The length of the segments to use. If specified, it overrides 'segments' unless 'segments' is a list
logical; if 'TRUE', the segment number is printed for each segment
if TRUE a plot will be generated that shows the selection of the optimal number of components for each step of the CV
method that defines how the optimal number of components is selected, should be one of "diffnext", "hastie", "relchange"; see details
additional parameters
Peter Filzmoser <P.Filzmoser@tuwien.ac.at>
In this cross-validation (CV) scheme, the optimal number of components is determined by an additional CV in the training set, and applied to the test set. The procedure is repeated repl times. There are different strategies for determining the optimal number of components (parameter selstrat): "diffnext" compares MSE+sdfact*sd(MSE) among the neighbors, and if the MSE falls outside this bound, this is the optimal number. "hastie" searches for the number of components with the minimum of the mean MSE's. The optimal number of components is the model with the smallest number of components which is still in the range of the MSE+sdfact*sd(MSE), where MSE and sd are taken from the minimum. "relchange" is a strategy where the relative change is combined with "hastie": First the minimum of the mean MSE's is searched, and MSE's of larger components are omitted. For this selection, the relative change in MSE compared to the min, and relative to the max, is computed. If this change is very small (e.g. smaller than 0.005), these components are omitted. Then the "hastie" strategy is applied for the remaining MSE's.
K. Varmuza and P. Filzmoser: Introduction to Multivariate Statistical Analysis in Chemometrics. CRC Press, Boca Raton, FL, 2009.
mvr
data(NIR)
X <- NIR$xNIR[1:30,] # first 30 observations - for illustration
y <- NIR$yGlcEtOH[1:30,1] # only variable Glucose
NIR.Glc <- data.frame(X=X, y=y)
res <- mvr_dcv(y~.,data=NIR.Glc,ncomp=10,method="simpls",repl=10)
Run the code above in your browser using DataLab