Lesion analysis FAQ
From VoxBoWiki
For frequently asked questions about the VoxBo package in general, see the FAQ.
Contents |
General
What is VLSM?
The term "Voxel-based lesion-symptom mapping" (VLSM) was coined by Bates et al. in their brief but seminal article in 2002 (see suggested readings below). It describes a process that is essentially the analog of SPM for lesion data - mapping the relationship between injury and behavior on a voxel-by-voxel basis. A typical VLSM analysis involves calculating, for each voxel, a t statistic describing the difference in some behavioral measure between patients with and without a lesion.
I'm happy to use the term "VLSM" loosely to refer to a broad variety of lesion analysis methods (VLSM is already a mild misnomer, given that most practitioners are not interested in symptoms per se).
Can you recommend some readings on lesion-behavior mapping?
This 2003 article by Bates et al. started VLSM rolling. It's not that methodologically interesting anymore, but it's also very short.
- Bates, E.; Wilson, S. M.; Saygin, A. P.; Dick, F.; Sereno, M. I.; Knight, R. T. & Dronkers, N. F. (2003), 'Voxel-based lesion-symptom mapping.', Nat Neurosci 6(5), 448-50.
I've written a short practical introduction to VLSM as a methodology whitepaper available here:
Rorden and Karnath wrote a very nice and somewhat higher level discussion of the merits of lesion-behavior mapping
- Rorden, C. & Karnath, H.O. (2004), 'Using human brain lesions to infer function: a relic from a past era in the fMRI age?', Nat Rev Neurosci 5(10), 813-9.
More coming soon!
VoxBo and MRIcro(n)/NPM
I ran the same analysis in VoxBo and NPM/MRIcron and got different results. Why?
To the best of my knowledge, if you really run exactly the same analysis in both packages, you will get the same results. But it's easy to run a slightly different analysis without realizing it. Here are some things to check, assuming you're using VoxBo's vbtmap:
- vbtmap by default produces a t map, while NPM always converts t scores to z scores. The output of NPM might make you think you're looking at t statistics, but you're not. To get z scores in vbtmap, use the -z flag.
- vbtmap produces positive t/z statistics for voxels in which high scores correlate with damage (typically appropriate for error scores or response times). NPM does the opposite (typically appropriate for accuracy measures). To get vbtmap to work like NPM, use the -f flag.
- Make sure you're using the same inclusion criterion for the minimum number of lesions per voxel, and make sure it's at least 2 (VoxBo's minimum). NPM will allow you to use a criterion of 0, which you should never do and which will produce anomalous results.
Statistical issues
What statistical tests does VoxBo use?
vbtmap provides the t test for two independent samples with pooled variance. If you use the -w flag, it will give you Welch's t-test for unequal variances. Note that if you use the regular t-test, each voxel has the same df (n-2), so you're safe doing everything as a t map. If you use the Welch's test, each voxel has different df, so it's best to convert to z scores (using the -z flag).
vbvolregress provides multiple regression, where you can mix and match spatially varying variables (4D volume series with one volume per subject) and spatially invariant variables (1D vectors with one value per subject). It's much slower than vbtmap, and in the simple case where you have a single independent measure for lesion/no-lesion, should give you exactly the same results. Use vbvolregress if you really need multiple regression.
What statistical test should i use?
Sneak preview: it's up to you. More detailed hedging coming soon!
How many patients do i need per voxel?
There is no very principled answer to this question, and perhaps if we used better behaved statistical tests it wouldn't make much difference. But I offer the following informal advice, in case it helps your thinking: use the same criterion you would if you were running a non-VLSM patient study with a single statistical test. If you would feel comfortable reporting a finding based on a comparison between 4 impaired patients and a larger control group, then you should be comfortable using the same criterion in a VLSM study.
Why are values repeated in my permutation distribution and/or in my permutation thresholds?
Coming soon?
What about the Brunner-Munzel test?
The Brunner-Munzel test is a non-parametric rank-based test that has some nice properties for VLSM, described in several articles by Chris Rorden and colleagues.
VoxBo doesn't provide the Brunner-Munzel test at present. If you want to use it, you need to use NPM. But just in case you're interested, here are some facts about the test.
The Brunner-Munzel calculation produces a Brunner-Munzel t statistic, sometimes called BMt, and a df. When you have at least 10 subjects in each group, it behaves a lot like the familiar Student's t statistic, and you can treat it just like a regular t. When you have fewer than about 10 subjects in one or both groups, the BMt is ill-behaved, and you will get anomalous p values if you just look it up in a t table. For VLSM studies, this means you have to be careful if using the Brunner-Munzel while including voxels with fewer than 10 lesions (or with fewer than 10 intact patients).
Note that having a large number of patients does not automatically protect you from this issue. If you have any voxels with fewer than 10 patients, and you use a permutation "maximum" test, those bad voxels can foul up the whole test.
To get around this problem, NPM automatically calculates the p value (and from it a z score) in a different way for voxels with small numbers of lesioned patients. It calculates a permutation test just within that voxel to derive a p value, and from there, a z score. Bear in mind that this is not related to whether or not you derive your significance threshold from a permutation test.
Some versions of MRIcron/NPM released in 2008 had a partial implementation of the Brunner-Munzel, omitting the special treatment of ill-behaved voxels, and could therefore produce anomalous results when voxels with small numbers of lesions were included. This has been fixed in more recent versions (although to be safe, I recommend only running it under Windows).
