Back to PLS HelpI'm Online
Posted on 12/12/14 09:52:47
Number of posts: 11
Dear all,
I'm trying to implement in the command line what is discussed here:
http://www.rotman-baycrest.on.ca/index.php?action=view_thread&id=869&module=bbmodule&src=%40random45c35fcb17881
In summary, I've created a datamatrix with (Nsubjects*Nconditions) x (Nchannels*Nfrequency_bins), and a behavior matrix with (Nsubjects*Nconditions)x 1 behavior measure. In the function pls_analysis.m I don't see an option to define the duration of the epoch nor the sampling rate. I'd like to see the effect of the conditions for the particular frequencies and channels, however when I run my behavior PLS analysis I obtain matrices (s, v, usc, vsc) similar to the ones I obtain if I had used just one single frequency: (Nsubjects*Nconditions) x (Nchannels for a specific frequency). The matrices result.u and result.boot_result.compare_u do reflect the size (Nchannels*Nfrequency_bins) x Nconditions.
My question is: should I focus on these two matrices to see the effect and the BSR? If the order is mantained, then could I plot the saliences individually for each frequency band? That is: first 128 rows correspond to saliences of frequency bin 1, the next 128 rows correspond to salience of frequency bin 2 and so on?
Thank you for any comment or suggestion!
Cheers,
Laura
Replies:
Posted on 12/12/14 17:09:11
Number of posts: 394
Hi Laura,
If I understand your question, yes you can extract the effects related to each frequency band by extracting the BSRs from the 'compare_u' matrix. The order will be identical to the ordering you used to create the data matrix. We created the command-line version of PLS to be as flexible as possible to be used with any data set, so the user inputs are minimal, but that also require a bit more of the user to sift through the results.
Randy
I'm Online
Posted on 12/16/14 08:05:03
Number of posts: 11
Dear Randy,
Thank you for your help. I've plotted the saliences per frequency band like you suggested.
Now I have a more specific question about the BSR and the contrasts. It would be very helpful if you could shed some light into them.
1) When doing the PLS analysis on a single frequency band, I know that I can get the "stable" electrodes using the percentiles 95 and 5 from result.boot_result.compare_u as thresholds, and I can do this per latent variable as shown: http://www.rotman-baycrest.on.ca/index.php?action=view_thread&id=1540&module=bbmodule&src=%40random45c35fcb17881.
When doing the PLS analysis on multiple frequency bins, the matrix from result.boot_result.compare_u is of size (Nchannels*Nfrequency_bins) x Nlatent_variables. How do I define then the thresholds for getting the stable electrodes? Does it make sense to get the percentiles 95 and 5 from this matrix per frequency bin (i.e. obtaining upper and lower thresholds per frequency bin)? If that's the case, then how do I check which frequency bin is actually the one mostly correlated to the effect I'm studying?
2) I'm only getting familiar with the design matrices. For a non-rotated Task PLS analysis (option.method = 2), do the contrasts need to sum up to 0? In my case I have 3 conditions, so for example: option.stacked_designdata = [2; -1; -1]. What options do I have if I need to check an increasing or decreasing trend, for example [3; 2; 1]? Is it akin to contrasts in parametric designs in which a decreasing trend should be [1; 0; -1]?
Cheers,
Laura
Posted on 12/16/14 08:42:21
Number of posts: 394
HI Laura:
>When doing the PLS analysis on multiple frequency bins, the matrix from result.boot_result.
You can use the same threshold, but you can't really check explicitly whether one frequency is most correlated with the effect unless you compare the saliences to one another using their estimated standard error (i.e, do they overlap)? How you do this will depend on how hard you want the inferenece to be. If just a description of the effect, then you don't have to compare, if you want to say one band does not show an effect, then you will need to show the standard errors don't overlap
>2) I'm only getting familiar with the design matrices. For a non-rotated Task PLS analysis (option.method = 2), do the contrasts need to sum up to >0? In my case I have 3 conditions, so for example: option.stacked_designdata = [2; -1; -1]. What options do I have if I need to check an increasing >or decreasing trend, for example [3; 2; 1]? Is it akin to contrasts in parametric designs in which a decreasing trend should be [1; 0; -1]?
They should sum to zero for your own peace of mind, though the code mean centres the contrast so [3,2,1] will get converted to [1,0,-1]
Login to reply to this topic.
