clusterrandanalysis

[Eric Maris]

Dear Claudia,


> i just began learning fieldtrip and am momentarilly busy with
> clusterrandanalysis. perhaps someone can answer some of my questions:
> which hypothesis is actually tested by the t-tests concerning the
> channel-time-frequency-triplets?

Clusterrandanalysis does not test hypothesis at the level of
channel-time-frequency triplets (also called "samples"). Instead, it tests
the null hypothesis that the data in the two (or more) experimental
conditions come from the same probability distribution. The sample-specific
t-statistics are used to construct clusters, namely by thresholding them,
followed by clustering on the basis of spatial, temporal, and spectral
adjacency. From a theoretical perspective, the t-statistic is not important
at all. One can also use the Honolulu-statistic (which does not exist) if
there is a sensible way of thresholding it. Clusterrandanalysis controls the
false alarm rate by calculating a randomization p-value for the maximum
cluster-level statistic. No parametric sampling distribution (such as the
T-distribution) is required to control the false alarm rate.

and why are´nt all triplets included using
> an alphtresh-value of 1.0?

If you set alphathresh at 1.0, you want all samples to be joined in a single
cluster, regardless of the size of the effect at the sample-level. This does
not make sense. Leave alphathresh at its default value 0.05. If you are just
starting with clusterrandanalysis, you will not gain anything by playing
with it.


> does neighbourdist mean: MAXIMAL distance from neighbours?

Yes, neighbourdist is the maximal distance between two sensors that are
considered as neighbours.


> how do i know for which triplet one specific
> p-value/cluster-level-statistic
> etc. stands?

p-values are computed for clusters, not for samples (unless you specify
cfg.makeclusters='no').

> are only significant effects plotted by topoplotTFR?

The effects you plot by topoplotTFR depend on what you ask topoplotTFR to
plot for you. This is specified by the cfg.zparam field. In the
clusterrandanalysis tutorial on the Fieldtrip homepage, I used masking to
select the significant clusters.

> for a two-sided test: which direction do the negative or the positive
> clusters have? is the power of the first ore the second data bigger or
> smaller?

A positive cluster means that the samples in this cluster have a larger
value in condition 1 than in condition 2. For negative clusters, the reverse
holds.

> how are more than two conditions compared (i have two independent
> variables,
> each having two levels, and two groups who are sensitized at the left and
> right side of the body, repectively. so i´m comparing up to eight
> channel-time-frequency-arrays.)

Clusterrandanalysis can only be used to test main effects.
Clusterrandanalysis can perform sample-specific comparisons for more than
two levels (eight, in your case) by means of an F-statistic. However, I
strongly advise you to first analyze your data for every independent
variable separately, each time ignoring the other two independent variables.
This amounts to using an independent samples t-statistic (if the independent
variable is between-subjects) or a paired samples t-statistic (if the
independent variable is within-subjects).


good luck,

Eric Maris