Skip to contents

GLM LSS Estimation Convenience Function (Single Trial Estimation)

Source: R/fmri_betas.R
glm_lss.Rd

A convenience wrapper around estimate_betas for least squares separate (LSS) estimation. This is primarily designed for single trial estimation, where each individual trial/event gets its own separate beta estimate rather than averaging across trials of the same condition.

Usage

glm_lss(
  dataset,
  model_obj,
  basis_obj,
  basemod = NULL,
  block = ~1,
  use_cpp = TRUE,
  progress = TRUE,
  ...
)

Arguments

dataset

A matrix_dataset object containing the fMRI time series data

model_obj

An event_model object specifying the experimental design

basis_obj

An HRF basis object (e.g., from fmrihrf::HRF_SPMG1, HRF_FIR, etc.)

basemod

A baseline_model instance to regress out of data before beta estimation (default: NULL)

block

A formula specifying the block factor (default: ~ 1 for single block)

use_cpp

Logical; whether to use C++ implementation for speed (default: TRUE)

progress

Logical; show progress bar (default: TRUE)

...

Additional arguments passed to estimate_betas

Value

A list of class "fmri_betas" containing the estimated trial-wise coefficients

Details

Primary Use Case - Single Trial Estimation:

  • Trial-wise beta estimation: Each trial gets its own beta coefficient

  • Single trial analysis: Useful for decoding, representational similarity analysis (RSA)

  • Trial-by-trial variability: Captures individual trial responses rather than condition averages

  • Avoiding trial averaging: Preserves trial-specific information that would be lost in standard GLM

Method Details: LSS (Least Squares Separate) fits a separate model for each trial, where the trial of interest gets its own regressor while all other trials of the same condition are modeled together. This approach avoids the collinearity issues that would arise from including separate regressors for every trial simultaneously.

For standard condition-level estimation (averaging trials within conditions), use glm_ols() instead.

See also

estimate_betas for the underlying estimation function, glm_ols for condition-level estimation

Examples

if (FALSE) { # \dontrun{
# Create event model and data
event_data <- data.frame(
  onset = c(10, 30, 50, 70),
  condition = factor(c("A", "B", "A", "B")),
  run = rep(1, 4)
)
sframe <- fmrihrf::sampling_frame(blocklens = 100, TR = 2)
model_obj <- event_model(onset ~ hrf(condition), 
                        data = event_data, 
                        block = ~ run, 
                        sampling_frame = sframe)

# Create data matrix (100 timepoints, 10 voxels)
Y <- matrix(rnorm(1000), 100, 10)

# Create matrix_dataset with event table
dset <- matrix_dataset(Y, TR = 2, run_length = 100, event_table = event_data)

# Fit with LSS - estimates separate beta for each individual trial
fit <- glm_lss(dset, model_obj, fmrihrf::HRF_SPMG1)
dim(fit$betas_ran)  # 4 trials x 10 voxels (NOT averaged by condition)

# This is useful for:
# - Decoding analysis (predicting condition from single trial patterns)
# - RSA (representational similarity analysis)
# - Studying trial-by-trial variability
} # }