AFNI Notes: 10/11/2019
Good AFNI summary (pdf)

Test	Description	Comment
3dttest++	a relatively flexible voxel based t-test	really good option for obvious contrasts
3dttest	as above without new features	largely decommissioned
3dMEMA	like 3dttest++ but uses t-stat for scaling	needs REML, conceptually cool, not sure of benefit
3dMannWhitney	non parametric 2 group ttest non-paired	maybe if the data is just that noisy
3dKruskalWallis	as above for >2 groups	or multiple noisy datasets
3dWilcoxon	paired 2 group non parametric	etc.
3dFriedman	as above for >2 groups	etc.
3dLME	a nice general LM frame work	good option for interactions etc.
3dMVM	like 3dLME with a few less bells	if your comfortable with LME just stay with that
3dANOVAx	Does x way ANOVAs	as it needs equal groups and other reasons don't use
3dRegAna	Does multiple regression	as it does calculations in short and other reasons don't use

3dttest++ Student T test with 1 or 2 sets of data (labeled 'A' and 'B').
-prefix p p = output
-mask mmm mmm = specified mask.
-setA/-setB Use this to specify group membership
1 group just give setA to contrast with base 0 [or setB versus -singletonA 10/perfect10+tlrc'[1]']
2 group will give setA-setB, setA(v. null), and setB(v. null) [-BminusA flips contrast]
Short form: -labelA Group1 -setA mydatasub1+tlrc'[1]' mydatasub2+tlrc'[1]' … etc.
Long form: -setA Group1 sub1 mydatasub1+tlrc'[1]' sub2 mydatasub2+tlrc'[1]' … etc.

-covariates COVAR_FILE a text file with a table for numeric covariate(s)[columns] for each subj[rows]

Rows must match dataset label example:

subject IQ age GMfrac

Elvis 143 42 Elvis_GM+tlrc[8]

Fred 85 59 Fred_GM+tlrc[8]

Ethel 109 49 Ethel_GM+tlrc[8]

Lucy 133 32 Lucy_GM+tlrc[8]

Ricky 121 37 Ricky_GM+tlrc[8]

can only be used with the short form '-setX' option if each input dataset has only 1 sub-brick
-center NONE/DIFF/SAME do you remove mean of covariates? Group or set mean? [default DIFF]
-paired is the data paired. If paired setA will be ignored
-pooled/unpooled for un/pooled variance (blocks covariates)
-toz Convert output t-statistics to z-scores
-rankize Convert the data (and covariates, if any)
-brickwise run ttest on each subbrick for each brick provided
-resid q q = Output residuals to look at with 3dFWHMx
-randomsign makes better residuals by flipping directionality
-clustsim will run 10000 iteration clustsim and apply results to file
-Clustsim keep .1D files

Understanding ClustSim:

Permutations

Understanding ETAC: Article

3dLME Multivariate Linear Model

-jobs NJOBS processors
-prefix Output file
-mask MASK Process voxels inside this mask only.
-model FORMULA Test of model like '~A*B+C+A:B" + = factors; : = interaction * = both (A*B = A+B+A:B)
-SS_type NUMBER Specify the type for sums of squares: 1=sequential 3=marginal
-qVars variable_list comma separated list of quantitative variables
-qVarCenters VALUES comma separated list of means of quantitative variables [default = group mean]
-vVars variable single voxel-wise covariate, i.e. subXTime1+tlrc column in dataTable
-vVarCenters VALUE centering values for voxel-wise covariates [default=mean of all subjects]
-ranEff FORMULA Specify the random effects. For just Subjects = "~1" but could be"~1+RT"
-num_glf/num_glt NG number of general linear F-tests/t-tests
-glfCode/gltCode k CODING Specify the k-th general linear F-test/t-test
Coding example: 'Condition : 1*A -1*B & 1*A -1*C Emotion : 1:pos'

• Weights don't need to sum to 0
• If covariate is not given the mean is used
• If a categorical is not given the mean of all levels I used

-glfLabel/gltLabel k label Specify the label for the k-th GLF/GLT
-dataTable List the data structure with a header as the first line.

• Last item command
• Header on first line must start with Subj and end with InputFile
• Can be a file '-dataTable @table.txt'

Other interesting:

-corStr FORMULA Correlation structure of the residuals (like AR or ARMA) if timeseries entered.

-cutoff threshold: Specify the cutoff value to obtain voxel-wise accuracy [default is 0]

-ICC intra-class correlation [for variables in -ranEff]

-LOGIT Do logistic regression [no random effects]