# 3dMEMA¶

```
Usage:
------
3dMEMA is a program for performing Mixed Effects Meta Analysis at group level
that models both within- and across- subjects variability, thereby requiring
both regression coefficients, or general linear contrasts among them, and the
corresponding t-statistics from each subject as input. To get accurate
t-statistics, 3dREMLfit should be used for the linear regression (a GLS
regression program using an ARMA(1,1) model for the noise), rather than
3dDeconvolve.
It's required to install R (https://www.r-project.org/), plus 'snow' package
if parallel computing is desirable. Version 1.0.1, Dec 21, 2016. If you want to
cite the analysis approach, use the following at this moment:
Chen et al., 2012. FMRI Group Analysis Combining Effect Estimates
and Their Variances. NeuroImage. NeuroImage 60: 747-765.
The basic usage of 3dMEMA is to derive group effects of a condition, contrast,
or linear combination (GLT) of multiple conditions. It can be used to analyze
data from one, two, or multiple groups. However, if there are more than two
groups or more than one subject-grouping variables (e.g., sex, adolescent/adults,
genotypes, etc.) involved in the analysis, dummy coding (zeros and ones) the
variables as covariates is required, and extremely caution should be exercised
in doing so because different coding strategy may lead to different
interpretation. In addition, covariates (quantiative variables) can be
incorporated in the model, but centering and potential interactions with other
effects in the model should be considered.
Basically, 3dMEMA can run one-sample, two-sample, and all types of BETWEEN-SUBJECTS
ANOVA and ANCOVA. Within-subject variables mostly cannot be modeled, but there are
a few exceptions. For instance, paired-test can be performed through feeding the
contrast of the two conditons as input. Multi-way ANOVA can be analyzed under the
following two scnearios: 1) all factors have only two levels (e.g., 2 X 2 repeated-
measures ANOVA) can be analyzed; or 1) there is only one within-subject (or
repeated-measures) factor and it contains two levels only. See more details at
https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/statistics/mema.html
Notice: When comparing two groups, option "-groups groupA groupB" has to be
present, and the output includes the difference of groupB - groupA, which is
consistent with most AFNI convention except for 3dttest++ where groupA - groupB is
rendered.
Convenience note: Please consider using the AFNI program gen_group_command.py
to build your 3dMEMA command. It should greatly help simplify the syntax of
constructing the command.
Example 1 --- One-sample type (one regression coefficient or general linear
contrast from each subject in a group):
--------------------------------
3dMEMA -prefix ex1 \
-jobs 4 \
-set happy \
ac ac+tlrc'[14]' ac+tlrc'[15]' \
ejk ejk+tlrc'[14]' ejk+tlrc'[15]' \
...
ss ss+tlrc'[14]' ss+tlrc'[15]' \
-max_zeros 4 \
-model_outliers \
-residual_Z
3dMEMA -prefix ex1 \
-jobs 4 \
-set happy \
ac ac+tlrc'[happy#0_Coef]' ac+tlrc'[happy#0_Tstat]' \
ejk ejk+tlrc'[happy#0_Coef]' ejk+tlrc'[happy#0_Tstat]' \
...
ss ss+tlrc'[happy#0_Coef]' ss+tlrc'[happy#0_Tstat]' \
-missing_data 0 \
-HKtest \
-model_outliers \
-residual_Z
Example 2 --- Two-sample type (one regression coefficient or general linear
contrast from each subject in two groups with the constrast being the 2nd group
subtracing the 1st one), heteroskedasticity (different cross-subjects variability
between the two groups), outlier modeling, covariates centering, no payment no
interest till Memorial Day next year. Notice that option -groups has to be
present in this case, and the output includes the difference of the second group
versus the first one.
-------------------------------------------------------------------------
3dMEMA -prefix ex3 \
-jobs 4 \
-groups horses goats \
-set healthy_horses \
ac ac_sad_B+tlrc.BRIK ac_sad_T+tlrc.BRIK \
ejk ejk_sad_B+tlrc.BRIK ejk_sad_T+tlrc.BRIK \
...
ss ss_sad_B+tlrc.BRIK ss_sad_T+tlrc.BRIK \
-set healthy_goats \
jp jp_sad_B+tlrc.BRIK jp_sad_T+tlrc.BRIK \
mb mb_sad_B+tlrc.BRIK mb_sad_T+tlrc.BRIK \
...
trr trr_sad_B+tlrc.BRIK trr_sad_T+tlrc.BRIK \
-n_nonzero 18 \
-HKtest \
-model_outliers \
-unequal_variance \
-residual_Z \
-covariates CovFile.txt \
-covariates_center age = 25 13 weight = 100 150 \
-covariates_model center=different slope=same
where file CovFile.txt looks something like this:
name age weight
ejk 93 117
jcp 3 34
ss 12 200
ac 12 130
jp 65 130
mb 25 630
trr 18 187
delb 9 67
tony 12 4000
Example 3 --- Paired type (difference of two regression coefficients or
general linear contrasts from each subject in a group). One scenario of
general linear combinations is to test linear or higher order trend at
individual level, and then take the trend information to group level.
---------------------------------
3dMEMA -prefix ex2 \
-jobs 4 \
-missing_data happyMiss+tlrc sadMiss+tlrc \
-set happy-sad \
ac ac_hap-sad_B+tlrc ac_hap-sad_T+tlrc \
ejk ejk_hap-sad_B+tlrc ejk_hap-sad_T+tlrc \
...
ss ss_hap-sad_B+tlrc ss_hap-sad_T+tlrc \
Options in alphabetical order:
------------------------------
-cio: Use AFNI's C io functions
-conditions COND1 [COND2]: Name of 1 or 2 conditions, tasks, or GLTs.
Default is one condition named 'c1'
-contrast_name: (no help available)
-covariates COVAR_FILE: Specify the name of a text file containing
a table for the covariate(s). Each column in the
file is treated as a separate covariate, and each
row contains the values of these covariates for
each subject. Option -unequal_variance may not be
used in the presence of covariates with two groups.
To avoid confusion, it is best you format COVAR_FILE in this manner
with BOTH row and column names:
subj age weight
Jane 25 300
Joe 22 313
... .. ...
This way, there is no amiguity as to which values are attributed to
which subject, nor to the label of the covariate(s). The word 'subj'
must be the first word of the first row. You can still get at the
values of the columns of such a file with AFNI's 1dcat -ok_text,
which will treat the first row, and first column, as all 0s.
Alternate, but less recommended ways to specify the covariates:
(column names only)
age weight
25 300
22 313
.. ...
or
(no row and column names)
25 300
22 313
.. ...
-covariates_center COV_1=CEN_1 [COV_2=CEN_2 ... ]: (for 1 group)
-covariates_center COV_1=CEN_1.A CEN_1.B [COV_2=CEN_2.A CEN_2.B ... ]:
(for 2 groups)
where COV_K is the name assigned to the K-th covariate,
either from the header of the covariates file, or from the option
-covariates_name. This makes clear which center belongs to which
covariate. When two groups are used, you need to specify a center for
each of the groups (CEN_K.A, CEN_K.B).
Example: If you had covariates age, and weight, you would use:
-covariates_center age = 78 55 weight = 165 198
If you want all covariates centered about their own mean,
just use -covariates_center mean. Be alert: Default is mean centering!
If no centering is desired (e.g.,the covariate values have been
pre-centered), set the center value as 0 with -covariates_center.
-covariates_model center=different/same slope=different/same:
Specify whether to use the same or different intercepts
for each of the covariates. Similarly for the slope.
-covariates_name COV_1 [... COV_N]: Specify the name of each of the N
covariates. This is only needed if the covariates' file
has no header. The default is to name the covariates
cov1, cov2, ...
-dbgArgs: This option will enable R to save the parameters in a
file called .3dMEMA.dbg.AFNI.args in the current directory
so that debugging can be performed.
-equal_variance: Assume same cross-subjects variability between GROUP1
and GROUP2 (homoskedasticity). (Default)
-groups GROUP1 [GROUP2]: Name of 1 or 2 groups. This option must be used
when comparing two groups. Default is one group
named 'G1'. The labels here are used to name
the sub-bricks in the output. When there are
two groups, the 1st and 2nd labels here are
associated with the 1st and 2nd datasets
specified respectively through option -set,
and their group difference is the second group
minus the first one, similar to 3dttest but
different from 3dttest++.
-help: this help message
-HKtest: Perform Hartung-Knapp adjustment for the output t-statistic.
This approach is more robust when the number of subjects
is small, and is generally preferred. -KHtest is the default
with t-statistic output.
-jobs NJOBS: On a multi-processor machine, parallel computing will speed
up the program significantly.
Choose 1 for a single-processor computer.
-mask MASK: Process voxels inside this mask only.
Default is no masking.
-max_zeros MM: Do not compute statistics at any voxel that has
more than MM zero beta coefficients or GLTs. Voxels around
the edges of the group brain will not have data from
some of the subjects. Therefore, some of their beta's or
GLTs and t-stats are masked with 0. 3dMEMA can handle
missing data at those voxels but obviously too much
missing data is not good. Setting -max_zeros to 0.25
means process data only at voxels where no more than 1/4
of the data is missing. The default value is 0 (no
missing values allowed). MM can be a positive integer
less than the number of subjects, or a fraction
between 0 and 1. Alternatively option -missing_data
can be used to handle missing data.
-missing_data: This option corrects for inflated statistics for the voxels where
some subjects do not have any data available due to imperfect
spatial alignment or other reasons. The absence of this option
means no missing data will be assumed. Two formats of option
setting exist as shown below.
-missing_data 0: With this format the zero value at a voxel of each subject
will be interpreted as missing data.
-missing_data File1 [File2]: Information about missing data is specified
with file of 1 or 2 groups (the number 1 or 2
and file order should be consistent with those
in option -groups). The voxel value of each file
indicates the number of sujects with missing data
in that group.
-model_outliers: Model outlier betas with a Laplace distribution of
of subject-specific error.
Default is -no_model_outliers
-n_nonzero NN: Do not compute statistics at any voxel that has
less than NN non-zero beta values. This options is
complimentary to -max_zeroes, and matches an option in
the interactive 3dMEMA mode. NN is basically (number of
unique subjects - MM). Alternatively option -missing_data
can be used to handle missing data.
-no_HKtest: Do not make the Hartung-Knapp adjustment. -KHtest is
the default with t-statistic output.
-no_model_outliers: No modeling of outlier betas/GLTs (Default).
-no_residual_Z: Do not output residuals and their Z values (Default).
-prefix PREFIX: Output prefix (just prefix, no view+suffix needed)
-residual_Z: Output residuals and their Z values used in identifying
outliers at voxel level.
Default is -no_residual_Z
-Rio: Use R's io functions
-set SETNAME \
SUBJ_1 BETA_DSET T_DSET \
SUBJ_2 BETA_DSET T_DSET \
... ... ... \
SUBJ_N BETA_DSET T_DSET \
Specify the data for one of two test variables (either group,
contrast/GLTs) A & B.
SETNAME is the name assigned to the set, which is only for the
user's information, and not used by the program. When
there are two groups, the 1st and 2nd datasets are
associated with the 1st and 2nd labels specified
through option -set, and the group difference is
the second group minus the first one, similar to
3dttest but different from 3dttest++.
SUBJ_K is the label for the subject K whose datasets will be
listed next
BETA_DSET is the name of the dataset of the beta coefficient or GLT.
T_DSET is the name of the dataset containing the Tstat
corresponding to BETA_DSET.
To specify BETA_DSET, and T_DSET, you can use the standard AFNI
notation, which, in addition to sub-brick indices, now allows for
the use of sub-brick labels as selectors
e.g: -set Placebo Jane pb05.Jane.Regression+tlrc'[face#0_Beta]' \
pb05.Jane.Regression+tlrc'[face#0_Tstat]' \
-show_allowed_options: list of allowed options
-unequal_variance: Model cross-subjects variability difference between
GROUP1 and GROUP2 (heteroskedasticity). This option
may NOT be invoked when covariate is present in the
model. Default is -equal_variance (homoskedasticity).
This option may not be useded when covariates are
involved in the model.
-verb VERB: VERB is an integer specifying verbosity level.
0 for quiet (Default). 1 or more: talkative.
#######################################################################
Please consider citing the following if this program is useful for you:
Chen et al., 2012. FMRI Group Analysis Combining Effect Estimates
and Their Variances. NeuroImage. NeuroImage 60: 747-765.
https://afni.nimh.nih.gov/sscc/gangc/MEMA.html
```