TPLSr (version 1.0.3)

TPLS: Fit a TPLS model to data

Description

Fit a TPLS model to data

Usage

TPLS(X, Y, NComp = 50, W = 0, nmc = 0)

Arguments

X

n-by-v data matrix of real numbers. Rows correspond to observations (trials) and columns to variables (e.g., fMRI voxels).

Y

n-by-1 Vector of real numbers. Can be binary (0/1) for classification model, or can be continuous.

NComp

Maximum number of partial least squares component you want to use. Default is 50, and this is on the safe side for fMRI.

W

n-by-1 vector of positive observation weights.

nmc

A switch to skip mean-centering. Default is off (0). Only turn it on (1) when the data is already mean-centered and you want to save memory space by not creating another copy of the data for mean-centering.

Value

A TPLS object that contains the following attributes. Most of the time, you won't need to access the attributes.

• NComp: The number of components you specified in the input

• W: Normalized version of the observation weights (i.e., they sum to 1)

• MtrainX: Column mean of X. Weighted mean if W is given.

• MtrainY: Mean of Y. Weighted mean if W is given.

• scoreCorr: Correlation between Y and each PLS component. Weighted correlation if W is given.

• pctVar: Proportion of variance of Y that each component explains.

• betamap: v-by-NComp matrix of TPLS coefficients for each of the v variables, provided at each model with NComp components.

• threshmap : v-by-NComp matrix of TPLS threshold values (0~1) for each of the v variables, provided at each model with NComp components.

Examples

# NOT RUN {
# Fit example TPLS data with a TPLS model
# Load example data (included with package).
X = TPLSdat$X
Y = TPLSdat$Y

# Fit the model, with default options (50 components, no observation weights)
TPLSmdl <- TPLS(X,Y)

# Make in-sample prediction at threshold of 0.5 and at all possible components
pred <- TPLSpredict(TPLSmdl,1:50,0.5,X)

# Look at the correlation between prediction and Y.
# This is in-sample prediction. Ergo, the model with most components will have the highest
# predictive correlation. In practice, you should choose the number of components and
# threshold using cross-validation. See example for TPLS_cv
cor(Y,pred)

# Extract the predictor for a model with 25 PLS components and threshold at 0.7 (just cuz)
betamap <- makePredictor(TPLSmdl,25,0.5)

# This is the intercept
betamap$bias

# These are the coefficients for the original variables
betamap$betamap
# }