Hi Max,

Yes, you are exactly right, both approaches are perfectly fine (and using a reference site can make the specification of your contrast vector/matrix simpler in may cases).

Just to elaborate a bit, for example assuming that want to estimate patients vs control differences using data from three different sites it is equivalent to test a model of the form:

  Subject effects: Patients, Controls, Site1, Site2, Site3

  Between-subjects contrast: [1, -1, 0, 0, 0]

and a second model of the form:

  Subject effects: Patients, Controls, Site1, Site2

  Between-subjects contrast: [1, -1, 0, 0]

(where "Site1" is a variable with 1's for subjects in site 1 and 0's for all other subjects). The two models are equivalent because both design matrices span the same column space (only the first model design matrix contains a "redundant" column as the sum Patients+Controls is equal to the sum Site1+Site2+Site3), and both contrasts define the same dimension within that column space.

Even though both forms are exactly equivalent (and they will produce the exact same results/statistics),  what makes the second form desirable in many cases is that it simplifies the interpretation and specification of some common post-hoc contrasts. For example, if you knew that you wanted to use site3 as reference and would like to evaluate the adjusted means in Patients and Controls at the level of this reference site, in the first analysis you would do so by specifying a contrast of the form:

  Between-subjects contrast: [1, 0, 0, 0, 1] (patients adjusted means at level of site3)

  Between-subjects contrast: [0, 1, 0, 0, 1] (controls adjusted means at level of site3)

while in the second analysis you would do so by specifying a contrast of the form:

  Between-subjects contrast: [1, 0, 0, 0] (patients adjusted means at level of site3)

  Between-subjects contrast: [0, 1, 0, 0] (controls adjusted means at level of site3)

with the latter being simpler as the model regressor coefficients directly inform you of the adjusted means within each group at the desired control level. 

Hope this helps

Alfonso

Originally posted by max345:

Hi Alfonso,

I have a question about this issue of study sites as covariates in CONN (see your comment below). What you described sounds to me like a full dummy coding with n-covariates for n-scanning sites and each participant scanned at these sites is coded with 1 and the others with 0. Is it possible in CONN to dummy code the scanning sites with a reference group (see figure below) with n-1 covariates?

source: https://de.mathworks.com/help/stats/dumm...

Or is it more common to code the scanning sites as full dummy variables rather than dummy variables with a reference group?

Thank you so much for your help!

 

Best,

Max

 

Originally posted by Alfonso Nieto-Castanon:

Hi Till, Exactly, simply define a set of site-specific covariates (e.g. SITE_1, SITE_2, etc.) and include these as covariates-of-no-interest in your second-level analysis (e.g. image attached).  For example, if you have three sites and have already imported a 2nd-level SITE covariate with values 1 to 3 indicating the site of each subject, you could:  1) discretize that variable: in the Setup.Covariates.2nd-level tab select your covariate SITE, and then click on 'Covariate tools. Discretize selected covariate'. That will create three new covariates named SITE_1, SITE_2 and SITE_3 2) (optionally / rarely-necessary) center those new site-specific covariates to your desired control-level (e.g. average across all subjects):  in the same tab select jointly SITE_1, SITE_2 and SITE_3, then click on 'Covariate tools. Orthogonalize selected covariates', and select the variable 'AllSubjects' as your only orthogonal factor 3) add those site-covariates as controls in your second-level analysis. For example, in the Results (2nd-level) tab, after defining your desired analysis (e.g. a two-sample t-test comparing PLACEBO and TREATMENT subjects), simply select the option that reads 'add/remove SITE_1 as control covariate' (and repeat for SITE_2 and SITE_3). You should be seeing something like in the example attached, which will estimate the differences between groups while controlling for site effects.  Hope this helps Alfonso   Originally posted by Till Langhammer:

Originally posted by Zahra Mor:

Dear Ali, conn users and experts, I wonder if you have found an answer to your question regarding how to correct the multi site (scanner) effect, probably through defining a categorical second level covariate? Thanks in advance!

Hey People! I have the same problem!!!! Anybody found a solution? best wishes Till