Back to PLS HelpCommand-Line PLS and Bootsrap results
I'm Online
Posted on 06/16/09 14:18:33
Number of posts: 8
Hi All,
I hesitate to post this, as I'm sure that mostly everyone is at HBM. But nonetheless, you never know ...
I am using command line PLS (behavioral) and running into problems interpreting some of the bootstrapping fields generated. I *have* read the documentation, but am still unclear about a few things. So, for example ...
(1) I want to generate a barchart of my behavioral LV values *with* error bars. I see a number of useful fields, such as:
(a) v_se (bootstrapped se?)
(b) compare_v (bootstrap ratios?)
Normally, I would just use the v_se, but my behavior LVs range between [-0.6, +0.6] (approx), whereas the v.se values range between [+3, +18]. So, either we have some sneaky rescaling happening, or v_se is not the bootstrapped se. Ideas?
(2) I would like to create the same sort of display for the correlations between brain scores and behavior. Similarly, I don't know what to use for the error bars. We have a number of candidates:
(a) orig_corr (this is just a copy of lvcorr)
(b) ulcorr / llcorr (upper and lower limits around orig_corr. Are this se, or something else??)
(c) ulcorr_adj / llcorr_adj (OK, adjusted versions of the above. How? Which to use in my plot?)
(3) Question: The compare_u and compare_v are just bootstrap ratios, right? So I could use compare_u to threshold my singular images? Yes?
By the way, have the primary PLS support individuals every considered changing this forum to a listserv? Your forum is full of useful information, which sadly, is not searchable from the outside. I, myself, receive a number of daily digests, and find them useful since (1) they keep me informed and up to speed, and (2) community members are able to help field some of the easier (or previously answered) questions, and (3) it fosters a sense of community. With forums, people generally only sign in if they have a problem (or if they are support people). Does this sound useful?
Thanks, and enjoy HBM (I'm going to ICAD in Vienna :) )
-Jim
I hesitate to post this, as I'm sure that mostly everyone is at HBM. But nonetheless, you never know ...
I am using command line PLS (behavioral) and running into problems interpreting some of the bootstrapping fields generated. I *have* read the documentation, but am still unclear about a few things. So, for example ...
(1) I want to generate a barchart of my behavioral LV values *with* error bars. I see a number of useful fields, such as:
(a) v_se (bootstrapped se?)
(b) compare_v (bootstrap ratios?)
Normally, I would just use the v_se, but my behavior LVs range between [-0.6, +0.6] (approx), whereas the v.se values range between [+3, +18]. So, either we have some sneaky rescaling happening, or v_se is not the bootstrapped se. Ideas?
(2) I would like to create the same sort of display for the correlations between brain scores and behavior. Similarly, I don't know what to use for the error bars. We have a number of candidates:
(a) orig_corr (this is just a copy of lvcorr)
(b) ulcorr / llcorr (upper and lower limits around orig_corr. Are this se, or something else??)
(c) ulcorr_adj / llcorr_adj (OK, adjusted versions of the above. How? Which to use in my plot?)
(3) Question: The compare_u and compare_v are just bootstrap ratios, right? So I could use compare_u to threshold my singular images? Yes?
By the way, have the primary PLS support individuals every considered changing this forum to a listserv? Your forum is full of useful information, which sadly, is not searchable from the outside. I, myself, receive a number of daily digests, and find them useful since (1) they keep me informed and up to speed, and (2) community members are able to help field some of the easier (or previously answered) questions, and (3) it fosters a sense of community. With forums, people generally only sign in if they have a problem (or if they are support people). Does this sound useful?
Thanks, and enjoy HBM (I'm going to ICAD in Vienna :) )
-Jim
Replies:
Untitled Post
I'm Online
Posted on 06/16/09 14:37:58
Number of posts: 291
Please type: help pls_analysis and you will find explanation of those variables:
% compare_v: compared designlv or compared behavlv
% v_se: standard error of designlv or behavlv
You should use "compare_u" for bootstrap ratio.
There is no sneaky rescaling happening to variable "v_se".
% lvcorrs: correlation of behavior data with usc,
% orig_corr: same as lvcorrs
% ulcorr: upper boundary of orig_corr
% llcorr: lower boundary of orig_corr
Don't worry about "ulcorr_adj / llcorr_adj, which is not used normally.
You can simply use the 3 variables above to plot the brain correlation.
Use "compare_u" instead of "compare_v" for bootstrap ratios.
I (jimmy@rotman-baycrest.on.ca), am the primary support of PLS software on MATLAB. However, I do not support PLS forum, which was developed by our webmaster (webmaster@rotman-baycrest.on.ca). Would you mind to put your good suggestions into a separate email to them? Thanks.
% compare_v: compared designlv or compared behavlv
% v_se: standard error of designlv or behavlv
You should use "compare_u" for bootstrap ratio.
There is no sneaky rescaling happening to variable "v_se".
% lvcorrs: correlation of behavior data with usc,
% orig_corr: same as lvcorrs
% ulcorr: upper boundary of orig_corr
% llcorr: lower boundary of orig_corr
Don't worry about "ulcorr_adj / llcorr_adj, which is not used normally.
You can simply use the 3 variables above to plot the brain correlation.
Use "compare_u" instead of "compare_v" for bootstrap ratios.
I (jimmy@rotman-baycrest.on.ca), am the primary support of PLS software on MATLAB. However, I do not support PLS forum, which was developed by our webmaster (webmaster@rotman-baycrest.on.ca). Would you mind to put your good suggestions into a separate email to them? Thanks.
Command-Line PLS and Bootsrap results
I'm Online
Posted on 06/16/09 15:19:57
Number of posts: 8
Hi Jimmy,
Thanks for referring me to the documentation. Actually, I had already read it, I just found some of it to be terse and largely unhelpful. May I ask a couple of follow-up points?
(1) You say that there is no rescaling of the v_se, yet you do not comment on why the values do not match the behavioral LV values (that is, the "v" component of the SVD). Do those values look reasonable to you?
(2) With regard to ulcorr/llcorr, you refer to them as the upper and lower boundaries, yet do not not explain how these boundaries are computed. Are they based on bootstrapped se for the correlations? I've also noticed that they are aymmetrical (thus requiring and upper and a lower). How should I interpret these bounds?
(3) With regard to my suggestion for transitioning from a forum format to a listserv, I just wanted to propose the option to the PLS community as a whole, as I have noticed others encountering the same sorts of problems with the forum. I shall leave it up to the PLS powers-that-be to act on the suggestion (if they find merit in it).
Thanks again,
-Jim
Thanks for referring me to the documentation. Actually, I had already read it, I just found some of it to be terse and largely unhelpful. May I ask a couple of follow-up points?
(1) You say that there is no rescaling of the v_se, yet you do not comment on why the values do not match the behavioral LV values (that is, the "v" component of the SVD). Do those values look reasonable to you?
(2) With regard to ulcorr/llcorr, you refer to them as the upper and lower boundaries, yet do not not explain how these boundaries are computed. Are they based on bootstrapped se for the correlations? I've also noticed that they are aymmetrical (thus requiring and upper and a lower). How should I interpret these bounds?
(3) With regard to my suggestion for transitioning from a forum format to a listserv, I just wanted to propose the option to the PLS community as a whole, as I have noticed others encountering the same sorts of problems with the forum. I shall leave it up to the PLS powers-that-be to act on the suggestion (if they find merit in it).
Thanks again,
-Jim
Untitled Post
I'm Online
Posted on 06/16/09 16:00:05
Number of posts: 291
Q1: You say that there is no rescaling of the v_se, yet you do not comment on why the values do not match the behavioral LV values (that is, the "v" component of the SVD). Do those values look reasonable to you?
A1: I just tested and the value matches behavioral LV values (i.e. the v component of SVD). You can follow the steps below to duplicate it:
result = pls_analysis(newdata_lst, num_subj_lst, k)
x=result.u*diag(result.s)*result.v';
y=result.datamatcorrs_lst{1}';
you can compare x & y, the difference is less than 1e-6
v_se is calculated based on v and bootstrap_v.
Q2: With regard to ulcorr/llcorr, you refer to them as the upper and lower boundaries, yet do not not explain how these boundaries are computed. Are they based on bootstrapped se for the correlations? I've also noticed that they are aymmetrical (thus requiring and upper and a lower). How should I interpret these bounds?
A2: The small help on top of pls_analysis.m will not explain HOW they are computed. Even the PLSgui User's Guide will not do. Please find relevant references for that answer (including the interpretion).
Untitled Post
Posted on 06/16/09 16:06:20
Number of posts: 394
Jim,
the standard errors for both ''v' and 'u' (designlv and brainlv) are computed based on the values in 'u' and 'v' rescaled by their respective singular values.
the upper and lower bounds for the confidence interval (ulcorr, llcorr) are the computed percentilles of the bootstrap distribution. Since a correlation distribution is bounded and not symmetric usually, the confidence intervals are also seldom symmetric.
the standard errors for both ''v' and 'u' (designlv and brainlv) are computed based on the values in 'u' and 'v' rescaled by their respective singular values.
the upper and lower bounds for the confidence interval (ulcorr, llcorr) are the computed percentilles of the bootstrap distribution. Since a correlation distribution is bounded and not symmetric usually, the confidence intervals are also seldom symmetric.
Command-Line PLS and Bootsrap results
I'm Online
Posted on 06/17/09 13:41:21
Number of posts: 8
Randy & Jimmy,
Thank-you very much for your help.
Randy: Ah-hah! Yes, I very much expected "sneaky rescaling". It all makes sense, and my plots look happy and reasonable. Also, thanks for the elucidation on the upper and lower limits, as this is the first question that anyone (reviewer or colleague) is going to ask about when they see the error bars.
Jimmy: Thanks for the help. I'm sure what you meant by your last comment regarding the upper and lower limits, however, I think we've now got it figured out. BTW, for future reference, I'm not matlab naive and am quite able to go through your code if necessary (I've already spent quite a bit of time checking things).
Cheers,
-Jim
Thank-you very much for your help.
Randy: Ah-hah! Yes, I very much expected "sneaky rescaling". It all makes sense, and my plots look happy and reasonable. Also, thanks for the elucidation on the upper and lower limits, as this is the first question that anyone (reviewer or colleague) is going to ask about when they see the error bars.
Jimmy: Thanks for the help. I'm sure what you meant by your last comment regarding the upper and lower limits, however, I think we've now got it figured out. BTW, for future reference, I'm not matlab naive and am quite able to go through your code if necessary (I've already spent quite a bit of time checking things).
Cheers,
-Jim
Login to reply to this topic.
