Regression coefficients in PLS DA

enzo · January 9, 2021, 4:36am

Hi again,

I would like to know how I could to get the regression coefficients in PLS DA.

Thanks very much

Enzo

enzo · January 9, 2021, 1:40pm

Hi,

I think a got it. In plsda code, line 604. Matrix B.hat isn’t it?

Is this expression to calculate coefficient regression equivalent to Bpls= (W(P’W)^-1)Q’ ??
where Q is Y-loadings, W is X-weights, P’ is X-loadings

B.hat[[i]] = sapply(1 : ncomp[i], function(x){Wmat[, 1:x] %% solve(t(Pmat[, 1:x]) %% Wmat[, 1:x]) %*% t(Cmat)[1:x, ]}, simplify = “array”)

Thanks

Enzo

github.com

mixOmicsTeam/mixOmics/blob/master/R/predict.R

# ========================================================================================================
# This function makes a prediction of a 'newdata' by using the results of 'object'.
# Depending on the class of the object (mint).(block).(s)pls(da) (16 classes so far), the input data is different
# and the preparation of the data is different - different scaling for instance if object="mint...."
# However, the prediction formula is the same for all classes, thus only one code
# ========================================================================================================

#' Predict Method for (mint).(block).(s)pls(da) methods
#' 
#' Predicted values based on PLS models. New responses and variates are
#' predicted using a fitted model and a new matrix of observations.
#' 
#' \code{predict} produces predicted values, obtained by evaluating the
#' PLS-derived methods, returned by \code{(mint).(block).(s)pls(da)} in the
#' frame \code{newdata}. Variates for \code{newdata} are also returned. Please
#' note that this method performs multilevel decomposition and/or log ratio
#' transformations if needed (\code{multilevel} is an input parameter while
#' \code{logratio} is extracted from \code{object}).
#' 
#' Different prediction distances are proposed for discriminant analysis. The

This file has been truncated. show original

kimanh.lecao · January 10, 2021, 10:48pm

Yes, that is the one.
From a normal PLS/PLS-DA model you could extract the loadings and loadings.star coefficients too, see recent post Loading values in plsda same as loading.weight in pls package - #2 by kimanh.lecao

Kim-Anh