I am using the 3dttest++ for comparing two different groups of subjects and would like to control for the effect of gender as a covariate. I referred to the help file of the 3dttest and found the following illustration on the structure of the output:
STRUCTURE OF THE OUTPUT DATASET ~1~
The output dataset is stored in float format; there is no option
to store it in scaled short format
For each covariate, 2 sub-bricks are produced:
++ The estimated slope of the beta values vs covariate
++ The t-statistic of this slope
++ If there are 2 sets of subjects, then each pair of sub-bricks is
produced for the setA-setB, setA, and setB cases, so that you’ll
get 6 sub-bricks per covariate (plus 6 more for the mean, which
is treated as a special covariate whose values are all 1).
++ Thus the number of sub-bricks produced is 6*(m+1) for the two-sample
case and 2*(m+1) for the one-sample case, where m=number of covariates.
For example, if there is one covariate ‘IQ’, and a two sample analysis
is carried out (‘-setA’ and ‘-setB’ both used), then the output
dataset will contain the following 12 (6*2) sub-bricks:
#0 SetA-SetB_mean = difference of means [covariates removed]
#2 SetA-SetB_IQ = difference of slopes wrt covariate IQ
#4 SetA_mean = mean of SetA [covariates removed]
#6 SetA_IQ = slope of SetA wrt covariate IQ
#8 SetB_mean = mean of SetB [covariates removed]
#10 SetB_IQ = slope of SetB wrt covariate IQ
I believe the output I would need to explore is sub brick #1, which states SetA-SetB_Tstat and demonstrates group mean differences. My understanding was that the term ‘removed’ in [covariates removed] means that the covariate is controlled for in this comparison. I would be thankful if you could please confirm.
Thank you so much.
Also, my second question was that because gender is a categorical variable, in my previous attempts with 3dMVM I used to use letters (F/M) in the covariates file. But in the 3dttest++ we are instructed to only use numeric variables as a covariate. Accordingly, I have used 1 for females and 0 for males. I was just concerned that the algorithm may treat the data differently if a categorical covariate is presented in a numeric format. Or is it just fine?
Referring to the instructions provided on the 3dttest++ help file, I think it must be fine. I would be thankful if you could please confirm. I am pasting this section of the help file at the end of this email.
Thank you so much for your time.
- You cannot enter covariates as pure labels (e.g., ‘Male’ and ‘Female’).
To assign such categorical covariates, you must use numeric values.
A column in the covariates file that contains strings rather than
numbers is assumed to be a list of dataset names, not category labels!
the term ‘removed’ in [covariates removed] means that the covariate is controlled for in this comparison.
You’re correct that the description in the 3dttest++ help can be improved.
You seem to have a 2 x 2 data structure. If you decide to use 3dttest++ instead of 3dMVM to handle the situation, there are a couple of subtleties involved. One, it sounds like you’re going to ignore the interaction effect. Two, dummy coding a categorical variable can be tricky. For example, 0/1 coding for sex could cause some unwanted side effect, If you really want to use 3dttest++,
- make sure the assumption of no interaction is reasonable
- code the two sexes with something -0.5 and 0.5 (or -1 and 1). See more details here: https://stats.idre.ucla.edu/r/library/r-library-contrast-coding-systems-for-categorical-variables/
Thank you very much, Gang!