Back to PLS HelpPosted on 10/30/14 17:42:51
Number of posts: 9
Dear Randy and Nancy et al.,
I am using structural behavioural PLS to study grey matter abnormalities in a subclinical population. Behaviourally, in my sample age and sex are significantly correlated with some of the personality scales that I am using as behaviour in the PLS analysis. Is there any way I can include those variables as covariates?
This is what I have tried: I did a mean-centered behavioural PLS with only the personality scales first. I then used the resulting weights from the significant LV in a non-rotated behavioural PLS as weights for the personality scales and in the same analysis I included age and sex weighted as zero. The resulting pattern of grey matter density reductions overall look very similar in both analyses, but sex is a significant contributor in the second analysis and the correlation responses in some areas are slightly reduced.
Would you be able to give me any advice on whether that is an appropriate way to include covariates in PLS? I wasn't quite sure what happens computationally when I weight variables as zero.
Thank you so much for your help. Kristina
Replies:
Posted on 10/31/14 08:08:30
Number of posts: 394
Dear Randy and Nancy et al.,
I am using structural behavioural PLS to study grey matter abnormalities in a subclinical population. Behaviourally, in my sample age and sex are significantly correlated with some of the personality scales that I am using as behaviour in the PLS analysis. Is there any way I can include those variables as covariates?
This is what I have tried: I did a mean-centered behavioural PLS with only the personality scales first. I then used the resulting weights from the significant LV in a non-rotated behavioural PLS as weights for the personality scales and in the same analysis I included age and sex weighted as zero. The resulting pattern of grey matter density reductions overall look very similar in both analyses, but sex is a significant contributor in the second analysis and the correlation responses in some areas are slightly reduced.
Would you be able to give me any advice on whether that is an appropriate way to include covariates in PLS? I wasn't quite sure what happens computationally when I weight variables as zero.
Thank you so much for your help. Kristina
Hi Kristina,
I am not sure I completely understand the second analysis, but did you try a behavior PLS where you included both the personality measure, age and sex in the behavior matrix? That would be the most straight forward way to ascertain whether what you are seeing is "just' an age or sex effect. You will get this by looking at the weights on the behavior side (I would focus on the correlation plot with confidence intervals). I suspect you will have one LV where everything loads, and then perhaps other where just the personality measures load.
Randy
Posted on 11/03/14 17:52:34
Number of posts: 9
Hi Kristina,
I am not sure I completely understand the second analysis, but did you try a behavior PLS where you included both the personality measure, age and sex in the behavior matrix? That would be the most straight forward way to ascertain whether what you are seeing is "just' an age or sex effect. You will get this by looking at the weights on the behavior side (I would focus on the correlation plot with confidence intervals). I suspect you will have one LV where everything loads, and then perhaps other where just the personality measures load.
Randy
Hi Randy,
thank you so much for your reply and sorry that I wasn't very clear about my second analysis. Here a second attempt.
My first analysis was a regular behavioural PLS with just the personality measures. I got a pattern of grey matter volume reductions across the brain to which 3 of the 4 personality measures contributed significantly.
In my second analysis I wanted to look at that pattern more closely and check how including age and sex would affect the pattern, i.e. whether parts of the pattern are significantly correlated with age and/or sex. My rationale was that if I then take out the portions of the pattern that are significantly correlated with age and/or sex I will get a pattern that is independent of those variables.
To this end, I used a non-rotated behavioural PLS with the personality measures, as well as age and sex in the behaviour matrix. As weights for the personality measures I took the weights that I got from the first (regular behavioural PLS) analysis (from result.lvcorrs), as I wanted to preserve the individual contributions of the personality measures. I weighted age and sex as zero. I summarised the design data in the table below.
| Personality measure 1 | Personality measure 2 | Personality measure 3 | Personality measure 4 | Age | Sex |
| 0.49 | 0.48 | 0.06 | 0.15 | 0 | 0 |
Long story short, here is my main question: How does PLS treat variables that are weighted as 0 in a non-rotated analysis? I weighted age and sex as zero assuming that this will result in a pattern that is not correlated with age and sex, i.e. to which age and sex are not contributing. However, looking at the resulting correlation plot with confidence intervals makes me think that this assumption was wrong. Sex seems to be contributing significantly (values from result.lvcorrs in the table below).
| Personality measure 1 | Personality measure 2 | Personality measure 3 | Personality measure 4 | Age | Sex |
| 0.49 | 0.48 | 0.06 | 0.15 | -0.04 | 0.63 |
Thank you for your time,
Kristina
Posted on 11/04/14 07:26:53
Number of posts: 394
Hi Kristina,
Thanks for the clarification.
For non-rotated analysis, whether task or behavior, conditions or variables that are weighted zero remain in the analysis and form the "Y" matrix to which "X" is fit - X being the contrast you enter.
The correlations that come out of the analysis give you an idea of how well "Y" actually fit the pattern in "X". In your case, Age was in fact not correlated with the brain data, but sex was moderately.
Does that make sense?
Posted on 11/07/14 17:37:37
Number of posts: 9
Hi Kristina,
Thanks for the clarification.
For non-rotated analysis, whether task or behavior, conditions or variables that are weighted zero remain in the analysis and form the "Y" matrix to which "X" is fit - X being the contrast you enter.
The correlations that come out of the analysis give you an idea of how well "Y" actually fit the pattern in "X". In your case, Age was in fact not correlated with the brain data, but sex was moderately.
Does that make sense?
That makes perfect sense. Thank you so much for your help, Randy.
Kristina
Login to reply to this topic.
