Getting Started with plsRglm

In plsRglm: Partial Least Squares Regression for Generalized Linear Models

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  message = FALSE,
  warning = FALSE,
  fig.width = 7,
  fig.height = 5,
  dpi = 150
)

set.seed(123)
library(plsRglm)

plsRglm provides partial least squares regression for linear and generalized linear models, repeated k-fold cross-validation, bootstrap utilities, and support for incomplete predictor matrices. This vignette is the practical starting point for the current package API. The companion vignette vignette("plsRglm", package = "plsRglm") keeps the longer historical case studies and algorithmic notes.

Core Fitting Workflows

plsR() is the dedicated interface for ordinary PLS regression. plsRglm() extends the same ideas to generalized linear and ordinal models, and can also fit modele = "pls" through the shared interface.

Linear PLS with matrix and formula interfaces

data(Cornell)
XCornell <- Cornell[, 1:7]
yCornell <- Cornell$Y

pls_fit_matrix <- plsR(yCornell, XCornell, nt = 3, verbose = FALSE)
pls_fit_formula <- plsR(Y ~ ., data = Cornell, nt = 3, pvals.expli = TRUE, verbose = FALSE)

pls_fit_formula$InfCrit
coef(pls_fit_formula)

The fitted model stores the extracted components (tt), the loadings (pp), the coefficients on the original predictors (Coeffs), and information-criterion summaries (InfCrit).

Generalized PLS models

data(aze_compl)
logit_fit <- plsRglm(y ~ ., data = aze_compl, nt = 3, modele = "pls-glm-logistic", verbose = FALSE)

logit_fit$InfCrit
head(predict(logit_fit, type = "response"))

family_fit <- plsRglm(
  Y ~ .,
  data = Cornell,
  nt = 2,
  modele = "pls-glm-family",
  family = gaussian(link = "log"),
  verbose = FALSE
)

family_fit$family$family
family_fit$family$link

plsRglm() supports predefined model shortcuts together with a custom-family entry point:

plsRglm(Y ~ ., data = Cornell, nt = 3, modele = "pls")
plsRglm(Y ~ ., data = Cornell, nt = 3, modele = "pls-glm-gaussian")
plsRglm(Y ~ ., data = Cornell, nt = 3, modele = "pls-glm-inverse.gaussian")
plsRglm(y ~ ., data = aze_compl, nt = 3, modele = "pls-glm-logistic")
data(pine)
plsRglm(round(x11) ~ ., data = pine, nt = 3, modele = "pls-glm-poisson")
plsRglm(x11 ~ ., data = pine, nt = 3, modele = "pls-glm-Gamma")
plsRglm(Quality ~ ., data = bordeaux, nt = 2, modele = "pls-glm-polr")
plsRglm(
  Y ~ .,
  data = Cornell,
  nt = 3,
  modele = "pls-glm-family",
  family = gaussian(link = "log")
)

Ordinal responses are handled through modele = "pls-glm-polr". As with MASS::polr(), the response should be an ordered factor:

data(bordeaux)
bordeaux$Quality <- factor(bordeaux$Quality, ordered = TRUE)
polr_fit <- plsRglm(Quality ~ ., data = bordeaux, nt = 2, modele = "pls-glm-polr", verbose = FALSE)

head(predict(polr_fit, type = "class"))

Cross-Validation and Model Choice

Use cv.plsR() for ordinary PLS regression and cv.plsRglm() for generalized models. Both provide repeated k-fold cross-validation and integrate with summary() and cvtable().

cv_pls <- cv.plsR(Y ~ ., data = Cornell, nt = 3, K = 4, NK = 2, verbose = FALSE)
cv_pls_summary <- cvtable(summary(cv_pls))

cv_pls_summary
plot(cv_pls_summary)

cv_logit <- cv.plsRglm(
  y ~ .,
  data = aze_compl,
  nt = 3,
  K = 4,
  NK = 2,
  modele = "pls-glm-logistic",
  verbose = FALSE
)
cv_logit_summary <- cvtable(summary(cv_logit, MClassed = TRUE))

cv_logit_summary
plot(cv_logit_summary)

For generalized models, summary(..., MClassed = TRUE) exposes miss-classification information when it is relevant.

Prediction and Missing Data

Incomplete predictor matrices are a core package feature, both during fitting and during prediction.

data(pine)
data(pine_sup)
data(pineNAX21)

pred_fit <- plsRglm(
  x11 ~ .,
  data = pine,
  nt = 3,
  modele = "pls-glm-family",
  family = gaussian(),
  verbose = FALSE
)

pine_sup_small <- pine_sup[1:3, 1:10]
pine_sup_small[1, 1] <- NA

predict(pred_fit, newdata = pine_sup_small, type = "response", methodNA = "missingdata")
predict(pred_fit, newdata = pine_sup_small, type = "scores", methodNA = "missingdata")

missing_train_fit <- plsR(x11 ~ ., data = pineNAX21, nt = 3, verbose = FALSE)
missing_train_fit$na.miss.X

When newdata contains incomplete rows, methodNA = "missingdata" treats all prediction rows with the missing-data scoring rule, while methodNA = "adaptative" switches between complete-row and incomplete-row formulas automatically.

Bootstrap Utilities

bootpls() and bootplsglm() wrap the boot package for PLS and PLS-GLM models. The default resampling schemes differ:

• bootpls() defaults to (y, X) resampling with typeboot = "plsmodel".
• bootplsglm() defaults to (y, T) resampling with typeboot = "fmodel_np".

For a lightweight vignette render, the examples below use a small number of resamples and request non-BCa confidence intervals.

boot_pls <- bootpls(pls_fit_formula, R = 20, verbose = FALSE)
dim(boot_pls$t)
confints.bootpls(boot_pls, indices = 2:4, typeBCa = FALSE)

boot_logit <- bootplsglm(logit_fit, R = 20, verbose = FALSE)
dim(boot_logit$t)
confints.bootpls(boot_logit, indices = 1:4, typeBCa = FALSE)

The plotting helpers boxplots.bootpls() and plots.confints.bootpls() can be applied directly to these bootstrap objects when a graphical summary is helpful.

Further Reading

• Use vignette("plsRglm", package = "plsRglm") for the historical applications and algorithmic note.
• Use the function help pages for lower-level weighted constructors such as PLS_lm_wvc() and PLS_glm_wvc().
• Use ?cv.plsRglm, ?bootplsglm, and ?predict.plsRglmmodel for the full argument reference.

sessionInfo()

plsRglm documentation built on June 17, 2026, 5:06 p.m.