Back to PLS Help
Brain scores with error bars
Posted on 03/11/09 13:02:34
Number of posts:
Is it reasonable to perform a t-test or ANOVA on the orig_usc to be able to interpret Brain score with CI? I know that PLS has used the statistic information of the whole brain for multivariate analysis,
but there should be a way to give some statistical evidence of brain score with CI to be able to claim that ,let's say one specific condition will not contribute in one direction, compare to the other condition (this could be seen by error bars as well )?That's why I feel that one needs a statistical proof (like a number given by F-test or t-test) for the brain score with CI.
my second question is if one wants to do such a kind of analysis, is it better to perform it on b_scores or on boot_result.orig_usc? if you say by the latter one,then one would be able to provide three value for each condition of interest, one orig_usc and the other two would be ulusc_adj and llusc_adj. So the degree of freedom for performing let's say t-test by using b_scores would be num of subject-1 but with orig_usc and it's upper and lower level border, would be 2. which one is better?
Actually I have read the previous reply by jimmy,but I have not convinced yet.
Thanks
but there should be a way to give some statistical evidence of brain score with CI to be able to claim that ,let's say one specific condition will not contribute in one direction, compare to the other condition (this could be seen by error bars as well )?That's why I feel that one needs a statistical proof (like a number given by F-test or t-test) for the brain score with CI.
my second question is if one wants to do such a kind of analysis, is it better to perform it on b_scores or on boot_result.orig_usc? if you say by the latter one,then one would be able to provide three value for each condition of interest, one orig_usc and the other two would be ulusc_adj and llusc_adj. So the degree of freedom for performing let's say t-test by using b_scores would be num of subject-1 but with orig_usc and it's upper and lower level border, would be 2. which one is better?
Actually I have read the previous reply by jimmy,but I have not convinced yet.
Thanks
Replies:
Untitled Post
I'm Online
Posted on 03/11/09 13:47:34
Number of posts: 291
Sorry that I could not convince you for the questions that you proposed.
For b_scores & boot_result.orig_usc, they should be in proportion. The major difference is: boot_result.orig_usc averaged subjects effect for brain scores. The rest differences are very minor.
For b_scores & boot_result.orig_usc, they should be in proportion. The major difference is: boot_result.orig_usc averaged subjects effect for brain scores. The rest differences are very minor.
Untitled Post
Posted on 03/13/09 13:22:52
Number of posts: 394
Is it reasonable to perform a t-test or ANOVA on the orig_usc to be
able to interpret Brain score with CI? I know that PLS has used the
statistic information of the whole brain for multivariate analysis,
but there should be a way to give some statistical evidence of brain score with CI to be able to claim that ,let's say one specific condition will not contribute in one direction, compare to the other condition (this could be seen by error bars as well )?That's why I feel that one needs a statistical proof (like a number given by F-test or t-test) for the brain score with CI.
The confidence intervals around the scores are all you really need to interpret the differences between conditions. It is really not valid to compare the brainscores with an ANOVA because they are biased towards showing a difference (thats what they are optimized to do). If you have 95% confidence intervals that do not overlap or do not cross zero, then you have your statistical proof.
my second question is if one wants to do such a kind of analysis, is it better to perform it on b_scores or on boot_result.orig_usc? if you say by the latter one,then one would be able to provide three value for each condition of interest, one orig_usc and the other two would be ulusc_adj and llusc_adj. So the degree of freedom for performing let's say t-test by using b_scores would be num of subject-1 but with orig_usc and it's upper and lower level border, would be 2. which one is better?
If you wanted to do that, you would do the analysis on the b_scores since there is one per subject/condition/lv.
but there should be a way to give some statistical evidence of brain score with CI to be able to claim that ,let's say one specific condition will not contribute in one direction, compare to the other condition (this could be seen by error bars as well )?That's why I feel that one needs a statistical proof (like a number given by F-test or t-test) for the brain score with CI.
The confidence intervals around the scores are all you really need to interpret the differences between conditions. It is really not valid to compare the brainscores with an ANOVA because they are biased towards showing a difference (thats what they are optimized to do). If you have 95% confidence intervals that do not overlap or do not cross zero, then you have your statistical proof.
my second question is if one wants to do such a kind of analysis, is it better to perform it on b_scores or on boot_result.orig_usc? if you say by the latter one,then one would be able to provide three value for each condition of interest, one orig_usc and the other two would be ulusc_adj and llusc_adj. So the degree of freedom for performing let's say t-test by using b_scores would be num of subject-1 but with orig_usc and it's upper and lower level border, would be 2. which one is better?
If you wanted to do that, you would do the analysis on the b_scores since there is one per subject/condition/lv.
Brain score with CI
Posted on 03/14/09 10:34:54
Number of posts:
Hello Randy,
Thank you for your clarification.So, If I correctly understood, if one condition have a confidence interval crossing zero,that might not contribute to the pattern. it also sounds clear to me ,but lets consider figure 23 in the PLS manual showing task PLS with CI . For group 1, there is no difficulty to claim that the Blue condition is the dominant condition in positive direction since it neither have overlap with the other three condition nor crossing zero. But in the negative direction, The red condition does not have overlap with the other conditions and does not cross the zero.Can we ignore the effect of the other two condition (green and pink) here? if you say yes , I would say that their confidence interval as you would see in the figure are different from zero,so they could have their effect on the patter.
If you say no, I would say that the confidence interval for the red shows a difference between red condition and both green and pink.so it could be more dominant.
So,Now the question (regarding interpretation of this pattern) is if one can claim that this pattern reflects Blue versus Red or Blue Vs. red,green,pink?
The last question is if we need to take a look at the original confidence interval or adjusted one (the one the is usually used to supress the effect of skew). I believe taken any of them, will lead to different interpretation.
Please answer me because this is the only obsticle that I found to be able to utilize PLS for my data.
Thanks in advance.
Hamed
Thank you for your clarification.So, If I correctly understood, if one condition have a confidence interval crossing zero,that might not contribute to the pattern. it also sounds clear to me ,but lets consider figure 23 in the PLS manual showing task PLS with CI . For group 1, there is no difficulty to claim that the Blue condition is the dominant condition in positive direction since it neither have overlap with the other three condition nor crossing zero. But in the negative direction, The red condition does not have overlap with the other conditions and does not cross the zero.Can we ignore the effect of the other two condition (green and pink) here? if you say yes , I would say that their confidence interval as you would see in the figure are different from zero,so they could have their effect on the patter.
If you say no, I would say that the confidence interval for the red shows a difference between red condition and both green and pink.so it could be more dominant.
So,Now the question (regarding interpretation of this pattern) is if one can claim that this pattern reflects Blue versus Red or Blue Vs. red,green,pink?
The last question is if we need to take a look at the original confidence interval or adjusted one (the one the is usually used to supress the effect of skew). I believe taken any of them, will lead to different interpretation.
Please answer me because this is the only obsticle that I found to be able to utilize PLS for my data.
Thanks in advance.
Hamed
Untitled Post
I'm Online
Posted on 03/14/09 13:16:08
Number of posts: 229
if one
condition have a confidence interval crossing zero,that might not
contribute to the pattern.
Its contribution is not reliable.
figure 23 in the PLS manual
the confidence intervals should be interpreted like any standard post-hoc test: if there is no overlap, there is a stable difference.
For both groups, R1C1 clearly differs from the other conditions. You should not "ignore" the R2C1 and R2C2 conditions because, while they may not differ from R1C2, they are different from the first.
Regarding the differences among the three conditions with negative brainscores, you should check both the "orig" and "adjusted" confidence intervals, since in this case, the CIs are fairly skewed with respect to the observed brainscores, and are close to having overlap. The adjusted confidence intervals may be more conservative, but are informative in the case of highly biased sampling distributions. If the distribution of the resampled results is not skewed, there should be little difference in the CIs.
In the end, it is your decision as to which set of CIs are appropriate for your data.
nancy
Its contribution is not reliable.
figure 23 in the PLS manual
the confidence intervals should be interpreted like any standard post-hoc test: if there is no overlap, there is a stable difference.
For both groups, R1C1 clearly differs from the other conditions. You should not "ignore" the R2C1 and R2C2 conditions because, while they may not differ from R1C2, they are different from the first.
Regarding the differences among the three conditions with negative brainscores, you should check both the "orig" and "adjusted" confidence intervals, since in this case, the CIs are fairly skewed with respect to the observed brainscores, and are close to having overlap. The adjusted confidence intervals may be more conservative, but are informative in the case of highly biased sampling distributions. If the distribution of the resampled results is not skewed, there should be little difference in the CIs.
In the end, it is your decision as to which set of CIs are appropriate for your data.
nancy
Login to reply to this topic.
