Representational mapping model (ReNA-Map)

For each ROI/searchlight, fits a reduced-rank regression from seed feature vectors (X) to ROI item-level patterns (Y), summarizing mapping rank and fit.

Usage

repmap_model(
  dataset,
  design,
  repmap_des,
  key_var,
  rank = "auto",
  max_rank = 20,
  ridge_lambda = NULL,
  ...
)

Arguments

dataset: An mvpa_dataset.
design: An mvpa_design.
repmap_des: Output of repmap_design.
key_var: Column or formula giving item identity (e.g. ~ ImageID).
rank: "auto" for rrpack::cv.rrr rank selection, integer for fixed rank, or 0 for no mapping (zero map; not an identity transform).
max_rank: Maximum rank to search.
ridge_lambda: Optional ridge penalty lambda for rrpack::rrs.fit.
...: Extra fields stored on the model spec.

Value

A model spec of class "repmap_model".

Details

Internally, the item-level seed features (X) and ROI patterns (Y) are column-centered prior to reduced-rank regression. Returned voxelwise R-squared values are in-sample and may be negative when the mapping underperforms the mean model; this can be useful as a diagnostic rather than an error.

Examples

ds <- gen_sample_dataset(D = c(4, 4, 4), nobs = 12, blocks = 3, nlevels = 2)
items <- as.character(sort(unique(ds$design$train_design$.rownum)))
seed_feats <- matrix(rnorm(length(items) * 4), nrow = length(items),
                     dimnames = list(items, NULL))
repmap_des <- repmap_design(items = items, seed_features = seed_feats)
model <- repmap_model(ds$dataset, ds$design, repmap_des, key_var = ~ .rownum,
                      rank = 2)
class(model)
#> [1] "repmap_model" "model_spec"   "list"