[FieldTrip] Clustering algorithms, large and long clusters, and watershed?
Lam, N.H.L. (Nietzsche)
n.lam at donders.ru.nl
Mon Feb 16 14:34:45 CET 2015
Hi S. Baris Demiral,
I have answers some of your questions, see below. Please note that it was difficult to answer all your questions because you didn't provide provide the actual code you used. Although your description is helpful, being able to see the actual parameters you implemented, and the specific function (e.g, did you use ft_freqstatistics, and did you use ft_clusterplot?) make it easier for anyone in the community attempting to answer your questions. Please see the FAQ for more details: http://fieldtrip.fcdonders.nl/faq/how_to_ask_good_questions_to_the_communityhttp://fieldtrip.fcdonders.nl/faq/how_to_ask_good_questions_to_the_community.
I'd like to point out that you can make good use of the search function (both inside FT - on the top right corner, and just on google), and reading the documentation for the functions that you are using, as many of your answers can be found there.
Finally, this FAQ should be of interest to you: http://fieldtrip.fcdonders.nl/faq/how_not_to_interpret_results_from_a_cluster-based_permutation_test
From: fieldtrip-bounces at science.ru.nl [fieldtrip-bounces at science.ru.nl] on behalf of Baris Demiral [demiral.007 at gmail.com]
Sent: 14 February 2015 17:58
To: FieldTrip discussion list
Subject: [FieldTrip] Clustering algorithms, large and long clusters, and watershed?
I am testing clustering based correction algorithms on a TF power data in a predefined frequent band; theta. I have four conditions. I used F statistic. I defined neighbors moderately so that the number of neighbors is not very small or extremely large. In some analyses I used pairwise t-test statistic to compare between conditions as well. I have a-priory expectations, such that some conditions will increase the centro-frontal theta, and some will increase the posterior theta. I use maxsum and wcm approaches.
I heave the following questions:
-Why do I observe that very distant electrodes are clustered together? I noticed that FCZ is clustered with occipital electrodes and belong to the same cluster written as in stat.cfg.posclusterlabel (label 1).
==> This could be due to the way your defined your neighbourhood structure. However, I can't make any conclusion from your defnition of "not very small or extremely large". Usually, when the neighbours are defined it specifies the neighbourhood size in the matlab workspace, and it would also help to know what you specific for cfg.method, when calling ft_prepare_neighbours.
==> It is important to note that even if there was a watershed method that it wouldn't answer the question of whether the centro-frontal theta is distinct from the occipital theta. More generally, the use of clustering won't answer this question either. It is better to use a feature in the data e.g., power change, to determine whether the theta differs between (groups of) sensors.
In some ways I can understand that because my task produces highly posteriorized theta power. The centro-frontal power is weaker. This leads to my next question: "Is there a watershed type of algorithm to separate these activities?"
- Are the electrodes I see in the plotting (marked by *,x,+) the peak electrodes in the clusters, or do these electrodes form the significant clusters (with smaller p values < .01, .05 etc)?
==> I assume you are using ft_clusterplot, and in the documentation of this function it states that the "(default ['*','x','+','o','.'] for p < [0.01 0.05 0.1 0.2 0.3])"
==> Electrodes marked with the same symbol belong the the same cluster (whether they are significant depends on the symbol, or the way you've assigned what the symbols mean).
Because, if the cluster is formed between distant electrodes as mentioned above, I would expect to see the intermediate electrodes (such as CZ etc.) in the cluster electrode list as well.
-Can you implement in plotting function where color can represent the cluster number? The *,+,x signs represent thresholds, but I cannot see which electrode belongs to which cluster. If you color code electrodes, it will be very helpful.
==> The elements in stat.poscluster/stat.negclusters are sorted according their p-values such that the cluster with the smallest p-value is first.
==> Part of this tutorial also applies to TF data, it should help you with differentiating clusters (and not just using the symbols):
The section of interest begins following text "We now briefly discuss the configuration fields that are not specific for ft_timelockstatistics<http://fieldtrip.fcdonders.nl/reference/ft_timelockstatistics>:".
==> I cannot implement this feature, however, if you would like to contribute to FT by adding this functionality, you're welcome to do so, see http://fieldtrip.fcdonders.nl/contribute
-Is there a range of weight values for the weighted cluster mass (wcm) approach? I looked at the paper, and seems like 0.45-.055 seems to be the weight parameter. Is this correct?
==> As a user, you need to determine and define a weight that is suitable for your data. The parameter to specify the weight is, cfg.wcm_weight.
S. Baris Demiral
10 Center Drive
Building 10, 5C410
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fieldtrip