Quick start guide

Estimate TFCE

Use any existing 2nd-level Model from a previous (parametric) SPM12 analysis of volume or surface data.

Select SPM.mat
Select the SPM.mat file that contains the design specification from a previous (parametric) estimation, where all required contrasts are already specified.
Select an additional mask
Select an additional mask image or surface to restrict your analysis. As default the mask in the analysis folder is used. Here you can select a mask to additionally restrict the analysis to regions of interest (i.e. small volume/surface correction).
Results title
Heading on results page - determined automatically if left empty
Contrast index
Index(es) of contrast according to the contrast manager: Each contrast in SPM is indicated by a sequential number that is displayed in the first column of the contrast manager. You can enter one or more contrasts. If only one number is entered, and this number is "Inf", you can select one or more contrasts interactively using the contrast manager. Do not define here the contrast itself. This should be done in the contrast manager, that is automatically called if "Inf" is kept as entry.
Number of permutations
In order to obtain reliable estimates you need about 5000-10000 permutations. If the number of maximal possible permutations is smaller, then this number is used resulting in an exact permutation test. Thanks to the tail approximations used for the p-values, the smallest reportable p-value is no longer tied to this number, as it was in earlier versions.
Permutation method to deal with nuisance variables
A number of methods are available to obtain parameter estimates and construct a reference distribution in the presence of nuisance variables. Smith permutation method is used if any nuisance variables exist and is selected by default. If no nuisance variables were found in the model then Draper-Stoneman method is automatically used. Freedman-Lane is another permutation method to deal with nuisance parameters, but it should be applied carefully. It's only necessary to change the permutation method if a large discrepancy between parametric and non-parametric statistic was found, which is indicated at the Matlab command line. See also Winkler et al. 2014 for more information.
Stop early if nothing can become significant
Stop the permutations as soon as it is certain that nothing in the image can become significant, instead of always running the full number of permutations. The largest value in the image is watched, because no other element rests on fewer exceedances than it does: an element with a smaller value is exceeded at least as often by the permutation maxima. Once the largest value has been exceeded often enough, and its corrected p-value is at least three standard errors above the largest alpha you asked about, no element can still become significant and no corrected p-value can still move appreciably. The permutations stop there and the results are saved as usual.

An image with nothing in it reaches that almost at once, because its largest value is an ordinary draw from the very distribution it is being compared against. An image with a real effect never reaches it, nor does an image whose corrected p-value sits anywhere near alpha, and both run the full number of permutations. So permutations are only ever saved where they could not have changed the answer. At least 500 permutations are always run, which the tail approximations need.

This is off by default: it changes how many permutations a given analysis runs, and analyses run with a fixed number of permutations are easier to compare with one another. Switch it on when you are screening many contrasts or designs and most of them are expected to be null.

Exchangeability blocks
For longitudinal and repeated-measures designs, so that data are only ever permuted within the blocks they may be permuted within. A wrong block structure invalidates the whole test, so the permutations are verified against the blocks before any of them is used.
TBSS data
Use 2D optimization (e.g. for TBSS DTI data) with internal TFCE parameters H=2, E=1.

A note on the TFCE parameters E and H

The idea of the TFCE approach is to combine focal effects with large voxel height as well as broad effects. How the two are weighted against one another is set by the parameters E (extent) and H (height). These are not user settings. They are fixed by the kind of data: E=0.5 and H=2 for 3D volumes, and E=1 and H=2 for surfaces and for TBSS data.

Smith and Nichols (Neuroimage 2009) established E=0.5 and H=2 empirically as giving good statistical power for volume data, and the surface and TBSS values follow the same reasoning for data of lower effective dimension. Earlier versions of the toolbox exposed E as a setting. It no longer is, because changing it changes what "significant" means in a way that is not comparable across studies and is easy to misuse.

< TFCE