sof_pc {funcharts}

Scalar-on-function linear regression based on principal components

Description

Scalar-on-function linear regression based on principal components. This function performs multivariate functional principal component analysis (MFPCA) to extract multivariate functional principal components from the multivariate functional covariates, then it builds a linear regression model of a scalar response variable on the covariate scores. Functional covariates are standardized before the regression. See Capezza et al. (2020) for additional details.

Usage

sof_pc(
  y,
  mfdobj_x,
  tot_variance_explained = 0.9,
  selection = "variance",
  single_min_variance_explained = 0,
  components = NULL
)

Arguments

y

A numeric vector containing the observations of the scalar response variable.

mfdobj_x

A multivariate functional data object of class mfd denoting the functional covariates.

tot_variance_explained

The minimum fraction of variance that has to be explained by the set of multivariate functional principal components retained into the MFPCA model fitted on the functional covariates. Default is 0.9.

selection

A character value with one of three possible values:

if "variance", the first M multivariate functional principal components are retained into the MFPCA model such that together they explain a fraction of variance greater than tot_variance_explained,

if "PRESS", each j-th functional principal component is retained into the MFPCA model if, by adding it to the set of the first j-1 functional principal components, then the predicted residual error sum of squares (PRESS) statistic decreases, and at the same time the fraction of variance explained by that single component is greater than single_min_variance_explained. This criterion is used in Capezza et al. (2020).

if "gcv", the criterion is equal as in the previous "PRESS" case, but the "PRESS" statistic is substituted by the generalized cross-validation (GCV) score.

Default value is "variance".

single_min_variance_explained

The minimum fraction of variance that has to be explained by each multivariate functional principal component into the MFPCA model fitted on the functional covariates such that it is retained into the MFPCA model. Default is 0.

components

A vector of integers with the components over which to project the functional covariates. If this is not NULL, the criteria to select components are ignored. If NULL, components are selected according to the criterion defined by selection. Default is NULL.

Value

a list containing the following arguments:

• mod: an object of class lm that is a linear regression model where the scalar response variable is y and the covariates are the MFPCA scores of the functional covariates,

• mod$coefficients contains the matrix of coefficients of the functional regression basis functions,

• pca: an object of class pca_mfd obtained by doing MFPCA on the functional covariates,

• beta_fd: an object of class mfd object containing the functional regression coefficient \beta(t) estimated with the scalar-on-function linear regression model,

• components: a vector of integers with the components selected in the pca model,

• selection: the same as the provided argument

• single_min_variance_explained: the same as the provided argument

• tot_variance_explained: the same as the provided argument

• gcv: a vector whose j-th element is the GCV score obtained when retaining the first j components in the MFPCA model.

• PRESS: a vector whose j-th element is the PRESS statistic obtained when retaining the first j components in the MFPCA model.

References

Capezza C, Lepore A, Menafoglio A, Palumbo B, Vantini S. (2020) Control charts for monitoring ship operating conditions and CO2 emissions based on scalar-on-function regression. Applied Stochastic Models in Business and Industry, 36(3):477–500. doi:10.1002/asmb.2507

Examples

library(funcharts)
data("air")
air <- lapply(air, function(x) x[1:10, , drop = FALSE])
fun_covariates <- c("CO", "temperature")
mfdobj_x <- get_mfd_list(air[fun_covariates], lambda = 1e-2)
y <- rowMeans(air$NO2)
mod <- sof_pc(y, mfdobj_x)