multivarious

This package is intended to provide some basic abstractions and default implementations of basic computational infrastructure for multivariate component-based modeling such as principal components analysis.

The main idea is to model multivariate decompositions as involving projections from an input data space to a lower dimensional component space. This idea is encapsulated by the projector class and the project function. Support for two-way mapping (row projection and column projection) is provided by the derived class bi-projector. Generic functions for common operations are included:

• project for mapping from input space into (usually) reduced-dimensional output space
• partial_project for mapping a subset of input space into output space
• project_vars for mapping new variables (“supplementary variables”) to output space
• reconstruct for reconstructing input data from its low-dimensional representation
• residuals for extracting residuals of a fit with n components.

The package now also includes a mixed-model path for operator-valued ANOVA. With mixed_regress(), each named fixed-effect term in a repeated-measures design can be extracted as an effect_operator, then analyzed with the same core verbs:

• effect for named term extraction
• components and scores for interpretable effect axes
• reconstruct for effect contributions in original variable space
• perm_test for omnibus and rank inference
• bootstrap for subject-level stability

The broader calibration harness for this path lives at experimental/mixed_effect_operator_calibration.R, with batch outputs saved under experimental/results/ when you run the simulation grid locally.

Installation

You can install the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("bbuchsbaum/multivarious")

Example

This is a basic example which shows you how to solve a common problem:

library(multivarious)
#> 
#> Attaching package: 'multivarious'
#> The following objects are masked from 'package:stats':
#> 
#>     residuals, screeplot
#> The following objects are masked from 'package:base':
#> 
#>     transform, truncate
## basic example code

Mixed effect operators

set.seed(1)

design <- expand.grid(
  subject = factor(seq_len(6)),
  level = factor(c("low", "mid", "high"), levels = c("low", "mid", "high")),
  KEEP.OUT.ATTRS = FALSE
)
design$group <- factor(rep(c("A", "B"), each = 9))

level_num <- c(low = -1, mid = 0, high = 1)[as.character(design$level)]
group_num <- ifelse(design$group == "B", 1, 0)
subj_idx <- as.integer(design$subject)
b0 <- rnorm(6, sd = 0.5)

Y <- cbind(
  b0[subj_idx] + level_num + rnorm(nrow(design), sd = 0.15),
  group_num + rnorm(nrow(design), sd = 0.15),
  level_num * group_num + rnorm(nrow(design), sd = 0.15),
  rnorm(nrow(design), sd = 0.15)
)

fit <- mixed_regress(
  Y,
  design = design,
  fixed = ~ group * level,
  random = ~ 1 | subject,
  basis = shared_pca(3),
  preproc = pass()
)

E <- effect(fit, "group:level")
pt <- perm_test(E, nperm = 19, alpha = 0.10)

ncomp(E)
#> [1] 0
ncomp(pt)
#> [1] 0

Albers theme

This package uses the albersdown theme. Vignettes are styled with vignettes/albers.css and a local vignettes/albers.js; the palette family is provided via params$family (default ‘red’). The pkgdown site uses template: { package: albersdown }.