nbr_lm_aov {NBR}

Network-based R-statistics using Linear Model ANOVA

Description

This function computes the specified linear model (LM) ANOVA for each edge in the network, and calculates the family wise error (FWE) p-value for the size of the clusters of connected edges that are individually below the P threshold (thrP), or above the F threshold (thrF). FWE estimation is based on the null distribution of the maximum size of sets of connected edges (defined as above), obtained with nperm permutations of the original data.

Usage

nbr_lm_aov(net, nnodes, idata, mod, diag = FALSE, nperm,
           thrP = 0.05, thrF = NULL, cores = NULL,
           nudist = FALSE, expList = NULL,
           verbose = TRUE, ...)

Arguments

net

3D volume (2D matrices for each observation) or 2D matrix of edges as columns.

nnodes

Number of network nodes.

idata

Matrix or data.frame including independent variables of interest of the model.

mod

Model, specify as a string, e.g., "~Group + Age".

diag

Logical indicating if matrix diagonal is to be included in the analysis (default: FALSE).

nperm

Number of permutations.

thrP

Individual edge p-value threshold (if NULL, thrF should be given).

thrF

Individual edge F-value threshold (if NULL, thrP should be given).

cores

Number of selected cores for parallel computing (default: NULL).

nudist

Logical indicating if null distribution should be returned (default: FALSE).

expList

Character string adding variable names to the varlist of 'clusterExport' (default: NULL).

verbose

Logical indicating if messages should be printed (default: TRUE).

...

Additional arguments to be passed to the low level 'lm' function.

Details

It's VERY IMPORTANT when giving net as a 2D matrix or data.frame, to be completely sure that column distribution fits that of the upper triangle indices of an nnodes * nnodes matrix. This may be verified through the edge indices, e.g., "which(upper.tri(matrix(nrow = nnodes, ncol = nnodes)), arr.ind = T)" (see vignette NBR-LME for more details).

Regarding nperm, I suggest first setting it to small values (5 or 10) in order to test that everything runs fine. After that, set nperm to 1000 or larger number to decrease the margin of error of the FWE p-value (see https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/Theory#Conditional_Monte_Carlo_Permutation_Tests to explore the behavior of FWE p-value as a function of nperm).

Value

List containing the observed statistics and their corresponding FWE p-values, if requested by nudist it will return the null distribution.

1. Observed statistics for every individual edge: corresponding subset of connected nodes and strength for each model term.

2. FWE for components: binary and strength sum, with their corresponding FWE p-value.

3. Null Distribution: maximal component size and strength for each permutation. Only returned if nudist is TRUE.

Examples

data(frontal2D)

ncores <- 2
library(parallel)
if(detectCores() < ncores) ncores <- NULL
nbr_result <- nbr_lm_aov(net = frontal2D[,-(1:3)],
  nnodes = 28, idata = frontal2D[,1:3],
  mod = "~ Group + Sex * Age",
  thrP = 0.01, nperm = 5, cores = ncores)
show(nbr_result)