Understanding output of 3dLME for ICC (Intraclass correlation)

I’ve been playing with the 3dLME subroutine for evaluating intraclass correlation coefficients and trying to understand the output.

I have an imaging metric that I have calculated for 4 repeated scans in many subjects (using HCP data), and looking to calculate ICC for this metric (Results are quite similar whether I use the bayesian ICCb option or standard REML ICC option.

I followed the syntax in Example 4 for 3dLME:

3dLME -prefix LMEOutput.nii -jobs 12 -model “1” -ranEff ‘Scan+Subj’ -ICC -dataTable
Subj Scan InputFile
s0001 scan0001 Sub_0001_Scan_1.nii
s0001 scan0002 Sub_0001_Scan_2.nii
s0001 scan0003 Sub_0001_Scan_3.nii
s0001 scan0004 Sub_0001_Scan_4.nii
s0002 scan0001 Sub_0002_Scan_1.nii
s0002 scan0002 Sub_0002_Scan_2.nii
s0002 scan0003 Sub_0002_Scan_3.nii
s0002 scan0004 Sub_0002_Scan_4.nii
s0003 scan0001 Sub_0003_Scan_1.nii
s0003 scan0002 Sub_0003_Scan_2.nii
s0003 scan0003 Sub_0003_Scan_3.nii
s0003 scan0004 Sub_0003_Scan_4.nii

The output gives ICC values that are much lower than I get with other ways of calculating ICC. For example, I’ve included a graph that shows this metric across many voxels compared to calculating the ICC for the same data using Matlab’s ICC.m function ICC(data,‘1-k’) option, which is ANOVA-based.

It looks like the 3dLME output is approximately squared compared to the Matlab output (I graphed square root of 3dLME as well). Any help understanding the discrepancy?


A couple of comments:

  1. Make sure that you’re using the 3dLME ICC output with the label with subject, not scan, in your case.
  2. 3dLME computes ICC(2,1) in the conventional notation, which should be the same as the standard definition except that it converts to 0 in case the conventional ICC formula renders a negative value.


I was comparing to ICC(2,k). Now everything makes sense.