Dear CONN users,

I am a bit confused how to enter my between-subjects contrasts.
I have 2 linear measures (Iron and Amyloid) but not all subjects were acquired on the same MRI system (8 channel vs 32 channel coil) so I should correct for that. Therefore I created a covariate ScannerType with '1' or '2' depending on which system they were scanned on.

1) How do I investigate if there is an effect of Iron*Amyloid while correcting for ScannerType?

2) Some subjects are special (very high age), how would I do separate analyses for this group?
Do I need to create separate IronOld and AmyloidOld covariates?

Dear Jiri

Regarding (1), if by "Iron * Amyloid" you are referring to the interaction between those two measures (e.g. you hypothesized that connectivity may be weaker than expected when both measures are low) then, in addition to your "Iron" and "Amyloid" covariates you would first also define another second-level covariate as "IronAmyloid" and enter in the 'values' field the expression "Iron .* Amyloid" (without the quotes) so that this new covariate contains the product of the other two covariate values. Then, in your second-level model you would select AllSubjects, Iron, Amyloid, and IronAmyloid and enter a [0 0 0 1] between-subjects contrast. If, in addition, you would like to control for potential differences between MRI systems, then simply select instead AllSubjects, Iron, Amyloid, IronAmyloid, and ScannerType and enter a [0 0 0 1 0] contrast.

Regarding (2), yes, if you want to perform separate analyses that include only a subset of subjects, there are several exactly-equivalent ways to do this. The first, and the one we typically recommend, is by creating a new set of second-level covariates that only have values for these subjects and contain 0 values for all other subjects. When all of the subject-effect variables selected in your second-level model contain zeros for some subjects those subjects are automatically eliminated from the analysis (you can check that this is happening by looking at the design matrix). In your case, for example, you would perform the analyses above by selecting 'SubjectsOld', 'IronOld', 'AmyloidOld', and 'IronAmyloidOld' and entering a [0 0 0 1] contrast (where all of these covariates contain some values for the older subjects, and zero's for the younger subjects). This is the way we typically recommend performing these analyses, but there is an alternative that is simpler in many cases (but in can be a bit confusing to new users), which is to use the "missing-values" capabilities in CONN. Basically, CONN can automatically eliminate subjects from some second-level analyses if some of the entered subject-effects in your second-level model contain "missing-values" (which are indicated by NaN values). This can be used to your advantage when you want to perform separate analyses only over a subset of subjects by creating a single new covariate, e.g. named GroupOldOnly, that includes 1's for the subjects that you want to include, and "NaN" (without the quotes), for the subjects that you do not want to include. Then, when you define your second-level analysis that includes the GroupOldOnly, Iron, Amyloid, and IronAmyloid effects (contrast [0 0 0 1]), the younger subjects will be automatically eliminated from these analyses without having to create new IronOld, AmyloidOld, etc. variables.  

Hope this helps
Alfonso 

Originally posted by Jiri van Bergen:

Thank you! This helps a lot!

Just for my understanding: 

Why do we write a 0 for Iron and Amyloid?
From the documentation of CONN it never became truly clear to me what a '0' really means.
I interpreted it as: We include this covariate in the design matrix in order to let it explain variances in the data, but it's not part of the statistical test for significance.

Dear Conn users,

Thankyou Alfonso et al.,  for providing such a wonderful software and regularly customising it on the basis of user requirements.

I have come across different opinions regarding putting categorical variable as covariate in the design. What would be everybody's opinion regarding Question 1 in this thread on  the below statements

1. Putting '0' and '1' or 8 channel and 32 channel in this case for the variable 'scanner Type'
or
2. Putting '1' and '2' for 8 channel and 32 channel in this case for the variable 'scanner Type'
or
3. Putting '1' for for the subjects scanned with 8 channel and '0' for the subjects scanned with 32 channel for the variable 'Eight channel' and similarly  putting '1' for for the subjects scanned with 32 channel and '0' for the subjects scanned with 8 channel for the variable 'Thity-two channel'

and then how would be the correction procedure in second level model for all the these three cases. I feel the contrast to be entered like 

for case 1 AllSubjects, Iron, Amyloid, IronAmyloid, and ScannerType as [0 0 0 1 0]

for case 2 AllSubjects, Iron, Amyloid, IronAmyloid, and ScannerType as [0 0 0 1 0]

for case 3 AllSubjects, Iron, Amyloid, IronAmyloid, Eight channel and Thirty-two channel as [0 0 0 1 0 0]

Which of the three option is recommended for analysis. Your suggestions are valuable 

Regards
Himanshu Joshi

Dear Conn experts,
It would be very helpful to have your comments on my previous mail
Regards
Himanshu Joshi
Originally posted by Himanshu Joshi:

Dear Joshi,

This is a very good question. If I am understanding correctly all three options will give you exactly the same results, and as such there is no 'better' or 'correct' choice. Strictly speaking, this is only true because you are testing the association with a covariate (IronAmyloid interaction), and that association will be the same irrespective of the level of ScannerType factor chosen (since IronAmyloid by ScannerType interactions have not been entered into your model). If you try to test other contrast in those same models, for example [1 0 0 0 0] in (1), [1 0 0 0 0] in (2), or [1 0 0 0 0 0] in (3), or if you include the ScannerType.*IronAmyloid interactions into your model, then those three options that you mention will result in different statistics/results, and in that case the 'better' choice would be the one where the zero-level of your ScannerType covariate corresponds to the level where you want to evaluate your effect of interest (e.g. IronAmyloid interaction). Often that 'better' choice would be something like:

4. Putting '-1' and '1' for the 8-channel and 32-channel subjects for the variable 'ScannerType' (in this case the zero-level of this covariate represents the average effect across both scanner types)

Hope this helps
Alfonso

Originally posted by Himanshu Joshi:

Hi Jiri,

On a related note, I am also looking into combining the data acquired from different scanner models (Philips + Siemens) and of course the protocols are not exactly the same. Can the multi-site 2nd-level covariate correction method be applied to the data with different scanner brands and MRI protocol? 

Thank you!

Regards,
Dilip Kumar