3dttest++¶
Gosset (Student) t-test of sets of 3D datasets.¶
[* Also consider program 3dMEMA, which can carry out a *]
[* more sophisticated type of 't-test' that also takes *]
[* into account the variance map of each input dataset. *]
[* When constructing 3dttest++ commands consider using *]
[* gen_group_command.py to build your command FOR you, *]
[* which can simplify the syntax/process. *]
* Usage can be similar (not identical) to the old 3dttest;
for example [SHORT form of dataset input]:
3dttest++ -setA a+tlrc'[3]' b+tlrc'[3]' ...
* OR, usage can be similar to 3dMEMA; for example [LONG form]:
3dttest++ -setA Green sub001 a+tlrc'[3]' \
sub002 b+tlrc'[3]' \
sub003 c+tlrc'[3]' \
... \
-covariates Cfile
* Please note that in the second ('LONG') form of the '-setA' option,
the first value after '-setA' is a label for the set (here, 'Green').
++ After that, pairs of values are given; in each pair, the first
entry is a label for the dataset that is the second entry.
++ This dataset label is used as a key into the covariates file.
++ If you want to have a label for the set, but do not wish (or need)
to have a label for each dataset in the set, then you can use
the SHORT form (first example above), and then provide the overall
label for the set with the '-labelA' option.
++ The set label is used to create sub-brick labels in the output dataset,
to make it simpler for a user to select volumes for display in the
AFNI GUI. Example:
-labelA Nor -label Pat
then the difference between the setA and setB means will get the
label 'Nor-Pat_mean', and the corresponding t-statistic will get
the label 'Nor-Pat_Tstat'.
++ See the section 'STRUCTURE OF THE OUTPUT DATASET' (far below) for
more information on how the results are formatted.
** NOTES on the labels above:
++ The '-setX' label (above: 'Green') will be limited to 12 characters
-- this label is used in the sub-brick labels in the output files,
which are shown in the AFNI GUI 'Define Overlay' buttons for
choosing the volumes (sub-bricks) you want to look at.
++ However, the dataset labels (above: 'sub001', etc) are only limited
to 256 characters. These labels are used to pick values out of the
covariates table.
++ In the 'LONG' form input illustrated above, the set label and the
dataset labels are given explicitly.
++ In the 'SHORT' form input, the set label must be given separately,
using option '-labelA' and/or '-labelB'. The dataset labels are
taken from the dataset input filenames -- to be precise, the 'prefix'
part of the filename, as in:
'Ethel/Fred.nii' -> 'Fred' and 'Lucy/Ricky+tlrc.HEAD' -> 'Lucy'
If you are using covariates and are using the 'SHORT' form of input
(the most common usage), the prefixes of the dataset filename must
be unique within their first 256 characters, or trouble will happen.
++ I added this note [15 Dec 2021] because failing to distinguish between
these labels and their limits was causing some confusion and angst.
* You can input 1 or 2 sets of data (labeled 'A' and 'B' by default).
* With 1 set ('-setA'), the mean across input datasets (usually subjects)
is tested against 0.
* With 2 sets, the difference in means across each set is tested
against 0. The 1 sample results for each set are also provided, since
these are often of interest to the investigator (e.g., YOU).
++ With 2 sets, the default is to produce the difference as setA - setB.
++ You can use the option '-BminusA' to get the signs reversed.
* Covariates can be per-dataset (input=1 number) and/or per-voxel/per-dataset
(input=1 dataset sub-brick).
++ Note that voxel-level covariates will slow the program down, since
the regression matrix for the covariates must be re-inverted for
each voxel separately. For most purposes, the program is so fast
that this slower operation won't be important.
* The new-ish options '-Clustsim' and '-ETAC' will use randomization and
permutation simulation to produce cluster-level threshold values that
can be used to control the false positive rate (FPR) globally. These
options are slow, since they will run 1000s of simulated 3D t-tests in
order to get cluster-level statistics about the 1 actual test.
* You can input plain text files of numbers, provided their filenames end
in the AFNI standard '.1D'. If you have two columns of numbers in files
AA.1D and BB.1D, you could test their means for equality with a command like
3dttest++ -prefix stdout: -no1sam setA AA.1D\' -setB BB.1D\'
Here, the \' at the end of the filename tells the program to transpose
the column files to row files, since AFNI treats a single row of numbers
as the multiple values for a single 'voxel'. The output (on stdout) from
such a command will be one row of numbers: the first value is the
difference in the means between the 2 samples, and the second value is
the t-statistic for this difference. (There will also be a bunch of text
on stderr, with various messages.)
* This program is meant (for most uses) to replace the original 3dttest,
which was written in 1994, "When grass was green and grain was yellow".
++ And when the program's author still had hair on the top of his head /:(
SET INPUT OPTIONS¶
* At least the '-setA' option must be given.
* '-setB' is optional, and if it isn't used, then the mean of the dataset
values from '-setA' is t-tested against 0 (1 sample t-test).
* Two forms for the '-setX' (X='A' or 'B') options are allowed. The first
(short) form is similar to the original 3dttest program, where the option
is just followed by a list of datasets to use.
* The second (long) form is similar to the 3dMEMA program, where you specify
a distinct label for each input dataset sub-brick (a difference between this
option and the version in 3dMEMA is only that you do not give a second
dataset ('T_DSET') with each sample in this program).
***** SHORT FORM *****
-setA BETA_DSET BETA_DSET ...
[-setB]
* In this form of input, you specify the datasets for each set
directly following the '-setX' option.
++ Unlike 3dttest, you can specify multiple sub-bricks in a dataset:
-setA a+tlrc'[1..13(2)]'
which inputs 7 sub-bricks at once (1,3,5,7,9,11,13).
*** See the '-brickwise' option (far below) for more information ***
*** on how multiple sub-brick datasets will be processed herein. ***
++ If multiple sub-bricks are input from a single dataset, then
covariates cannot be used (sorry, Charlie).
++ In the short form input, the 'prefix' for each dataset is its label
if '-covariates' is used. The prefix is the dataset file name with
any leading directory name removed, and everything at and after
'+' or '.nii' cut off:
Zork/Fred.nii -> Fred *OR* Zork/Fred+tlrc.HEAD -> Fred
++ In the long form input (described below), you provide each dataset
with a label on the command line directly.
++ For some limited compatibility with 3dttest, you can use '-set2' in
place of '-setA', and '-set1' in place of '-setB'.
++ [19 Jun 2012, from Beijing Normal University, during AFNI Bootcamp]
For the SHORT FORM only, you can use the wildcards '*' and/or '?' in
the BETA_DSET filenames, along with sub-brick selectors, to make it
easier to create the command line.
To protect the wildcards from the shell, the entire filename should be
inside single ' or double " quote marks. For example:
3dttest++ -setA '*.beta+tlrc.HEAD[Vrel#0_Coef]' \
-setB '*.beta+tlrc.HEAD[Arel#0_Coef]' -prefix VAtest -paired
will do a paired 2-sample test between the symbolically selected sub-bricks
from a collection of single-subject datasets (here, 2 different tasks).
***** LONG FORM *****
-setA SETNAME \
[-setB] LABL_1 BETA_DSET \
LABL_2 BETA_DSET \
... ... \
LABL_N BETA_DSET
* In this form of input, you specify an overall name for the set of datasets,
and a label to be associated with each separate input dataset. (This label
is used with the '-covariates' option, described later.)
SETNAME is the name assigned to the set (used in the output labels).
LABL_K is the label for the Kth input dataset name, whose name follows.
BETA_DSET is the name of the dataset of the beta coefficient or GLT.
++ only 1 sub-brick can be specified here!
** Note that the label 'SETNAME' is limited to 12 characters,
and the dataset labels 'LABL_K' are limited to 256 characters.
-- Any more will be thrown away without warning.
-- This limit also applies to the dataset labels taken
from the dataset filenames in the short form input.
** Only the first 12 characters of the covariate labels can be
used in the sub-brick labels, due to limitations in the AFNI
dataset structure and AFNI GUI. Any covariate labels longer than
this will be truncated when put into the output dataset :(
** The program determines if you are using the short form or long **
** form to specify the input datasets based on the first argument **
** after the '-setX' option. If this argument can be opened as a **
** dataset, the short form is used. If instead, the next argument **
** cannot be opened as a dataset, then the long form is assumed. **
-labelA SETNAME = for the short form of '-setX', this option allows you
[-labelB] to attach a label to the set, which will be used in
the sub-brick labels in the output dataset. If you don't
give a SETNAME, then '-setA' will be named 'SetA', etc.
***** NOTE WELL: The sign of a two sample test is A - B. *****
*** Thus, '-setB' corresponds to '-set1' in 3dttest, ***
*** and '-setA' corresponds to '-set2' in 3dttest. ***
***** This ordering of A and B matches 3dGroupInCorr. *****
*****-------------------------------------------------------------*****
***** ALSO NOTE: You can reverse this sign by using the option *****
*** '-BminusA', in which case the test is B - A. ***
*** The option '-AminusB' can be used to explicitly ***
***** specify the standard subtraction order. *****
------------ Dataset (e.g., Subject) level weights [Mar 2020] ------------
These options let you mark some datasets (that is, some subjects) as
weighing more in the analysis. A larger weight means a subject's
data will count more in the analysis.
-setweightA wname = Name of a file with the weights for the -setA
*and/or* datasets. This is a .1D (numbers as text) file
-setweightB that should have 1 positive value for each
volume being processed.
* A larger weight value means the voxel values for
that volume counts more in the test.
* In the least squares world, these weights would
typically be the reciprocal of that subject's
(or volume's) standard deviation -- in other words,
a measure of the perceived reliability of the data
in that volume.
* For -setweightA, there should be the same number
of weight values in the 'wname' file as there
are volumes in -setA.
++ Fewer weight values cause a fatal ERROR.
++ Extra weight values will print a WARNING
message and then be ignored.
++ Non-positive weight values cause a fatal ERROR.
* You can provide the weights directly on the
the command line with an option of the form
-setweightA '1D: 3 2 1 4 1 2'
when -setA has 6 input volumes.
* You can use -covariates and -setweight together.
--LIMITATIONS-- ** At this time, there is no way to set voxel-wise weights.
** -setweight will turn off -unpooled (if it was used).
** -paired will turn off -setweightB (if used), since
a paired t-test requires equal weights
(and equal covariates) in both samples.
** -singletonA will turn off -setweightA.
** Using -setweight with -rankize is not allowed.
Implementation of weights is by use of the regression method used
for implementing covariates. For convenience in the program, the
provided weights are normalized to average 1, separately for
-setA and -setB (if present). This means that the total weight
actually used for each set is the number of volumes present in that set.
The t-statistic for setA-setB is testing whether the weighted
means of the two samples are equal. Similar remarks apply to
the individual sample means (e.g., weighted mean of setA
tested versus 0).
Dataset weights are conceptually different than dataset covariates:
* Weights measure the reliability of the input dataset values - larger
weight for a dataset means its values are more reliable.
* Covariates are measures that might directly affect the input dataset values.
In a different language, weights are about the variance of the input dataset
values, whereas covariates are about the size of the input dataset values.
As with covariates, where you get the weights from is your business.
Be careful out there, and don't go crazy.
TESTING A SINGLE DATASET VERSUS THE MEAN OF A GROUP OF DATASETS¶
This new [Mar 2015] option allows you to test a single value versus
a group of datasets. To do this, replace the '-setA' option with the
'-singletonA' option described below, and input '-setB' normally
(that is, '-setB' must have more than 1 dataset).
The '-singletonA' option comes in 3 different forms:
-singletonA dataset_A
*OR*
-singletonA LABL_A dataset_A
*OR*
-singletonA FIXED_NUMBER
* In the first form, just give the 1 sub-brick dataset name after the option.
* In the second form, you can provide a dataset 'label' to be used for
covariates extraction. As in the case of the long forms for '-setA' and
'-setB', the 'LABL_A' argument cannot be the name of an existing dataset;
otherwise, the program will assume you are using the first form.
* In the third form, instead of giving a dataset, you give a fixed number
(e.g., '0.5'), to test the -setB collection against this 1 number.
++ In this form, '-singleton_variance_ratio' is set to a very small number,
since you presumably aren't testing against an instance of a random
variable.
++ Also, '-BminusA' is turned on when FIXED_NUMBER is used, to give the
effect of a 1-sample test against a constant. For example,
-singletonA 0.0 -set B x y z
is equivalent to the 1-sample test with '-setA x y z'. The only advantage
of using '-singletonA FIXED_NUMBER' is that you can test against a
nonzero constant this way.
++ You cannot use covariates with this FIXED_NUMBER form of '-singletonA' /:(
* The output dataset will have 2 sub-bricks:
++ The difference (at each voxel) between the dataset_A value and the
mean of the setB dataset values.
++ (In the form where 'dataset_A' is replaced by a fixed)
(number, the output is instead the difference between)
(the mean of the setB values and the fixed number. )
++ The t-statistic corresponding to this difference.
* If covariates are used, at each voxel the slopes of the setB data values with
respect to the covariates are estimated (as usual).
++ These slopes are then used to project the covariates out of the mean of
the setB values, and are also applied similarly to the single value from
the singleton dataset_A (using its respective covariate value).
++ That is, the covariate slopes from setB are applied to the covariate values
for dataset_A in order to subtract the covariate effects from dataset_A,
as well as from the setB mean.
++ Since it impossible to independently estimate the covariate slopes for
dataset_A, this procedure seems (to me) like the only reasonable way to use
covariates with a singleton dataset.
* The t-statistic is computed assuming that the variance of dataset_A is the
same as the variance of the setB datasets.
++ Of course, it is impossible to estimate the variance of dataset_A at each
voxel from its single number!
++ In this way, the t-statistic differs from testing the setB mean against
a (voxel-dependent) constant, which would not have any variance.
++ In particular, the t-statistic will be smaller than in the more usual
'test-against-constant' case, since the test here allows for the variance
of the dataset_A value.
++ As a special case, you can use the option
-singleton_variance_ratio RRR
to set the (assumed) variance of dataset_A to be RRR times the variance
of set B. Here, 'RRR' must be a positive number -- it cannot be zero,
so if you really want to test against a voxel-wise constant, use something
like 0.000001 for RRR (this is the setting automatically made when
'dataset_A' is replaced by a fixed number, in the third form above).
* Statistical inference on a single sample (dataset_A values) isn't really
possible. The purpose of '-singletonA' is to give you some guidance when
a voxel value in dataset_A is markedly different from the distribution of
values in setB.
++ However, a statistician would caution you that when an elephant walks into
the room, it might be a 500,000 standard deviation mouse, so you can't
validly conclude it is a different species until you get some more data.
* At present, '-singletonA' cannot be used with '-brickwise'.
++ Various other options don't make sense with '-singletonA', including
'-paired' and '-center SAME'.
* Note that there is no '-singletonB' option -- the only reason this is labeled
as '-singletonA' is to remind the user (you) that this option replaces the
'-setA' option.
COVARIATES - per dataset and per voxel¶
-covariates COVAR_FILE
* COVAR_FILE is the name of a text file with 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 one sample (dataset). Note
that you can use '-covariates' only ONCE -- the COVAR_FILE should contain
the covariates for ALL input samples from both sets.
* Rows in COVAR_FILE whose first column don't match a dataset label are
ignored (silently).
++ This feature allows you to analyze subsets of data collections while
using the covariates file for a large group of subjects -- some of whom
might not be in a given subset analysis.
* An input dataset label that doesn't match a row in COVAR_FILE, on the other
hand, is a fatal error.
++ The program doesn't know how to get the covariate values for such a
dataset, so it can't continue.
* There is no provision for missing values -- the entire table must be filled!
* The format of COVAR_FILE is similar to the format used in 3dMEMA and
3dGroupInCorr (generalized to allow for voxel-wise covariates):
FIRST LINE --> subject IQ age GMfrac
LATER LINES --> 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]
* The first line of COVAR_FILE contains column headers. The header label
for the first column (#0) isn't used for anything. The later header labels
are used in the sub-brick labels stored in the output dataset.
* The first column contains the dataset labels that must match the dataset
LABL_K labels given in the '-setX' option(s).
* If you used a short form '-setX' option, each dataset label is
the dataset's prefix name (truncated to 12 characters).
++ e.g., Klaatu+tlrc'[3]' ==> Klaatu
++ e.g., Elvis.nii.gz ==> Elvis
* '-covariates' can only be used with the short form '-setX' option
when each input dataset has only 1 sub-brick (so that each label
refers to exactly 1 volume of data).
++ Duplicate labels in the dataset list or in the covariates file
will not work well!
* The later columns in COVAR_FILE contain numbers (e.g., 'IQ' and 'age',
above), OR dataset names. In the latter case, you are specifying a
voxel-wise covariate (e.g., 'GMfrac').
++ Do NOT put the dataset names or labels in this file in quotes.
* A column can contain numbers only, OR datasets names only. But one
column CANNOT contain a mix of numbers and dataset names!
++ In the second line of the file (after the header line), a column entry
that is purely numeric indicates that column will be all numbers.
++ A column entry that is not numeric indicates that column will be
dataset names.
++ You are not required to make the columns and rows line up neatly,
(separating entries in the same row with 1 or more blanks is OK),
but your life will be much nicer if you DO make them well organized.
* 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!
* If you want to omit some columns in COVAR_FILE from the analysis, you
can do so with the standard AFNI column selector '[...]'. However,
you MUST include column #0 first (the dataset labels) and at least
one more column. For example:
-covariates Cov.table'[0,2..4]'
to skip column #1 but keep columns #2, #3, and #4.
* Only the -paired and -pooled options can be used with covariates.
++ If you use -unpooled, it will be changed to -pooled.
++ The same limitation on -unpooled applies to -setweight.
* If you use -paired, then the covariate values for setB will be the
same as those for setA, even if the dataset labels are different!
++ If you want to use different covariates for setA and setB in the
paired test, then you'll have to subtract the setA and setB
datasets (with 3dcalc), and then do a 1-sample test, using the
differences of the original covariates as the covariates for
this 1-sample test.
++ This subtraction technique works because a paired t-test is really
the same as subtracting the paired samples and then doing a
1-sample t-test on these differences.
++ For example, you do FMRI scans on a group of subjects, then
train them on some task for a week, then re-scan them, and
you want to use their behavioral scores on the task, pre- and
post-training, as the covariates.
* See the section 'STRUCTURE OF THE OUTPUT DATASET' for details of
what is calculated and stored by 3dttest++.
* If you are having trouble getting the program to read your covariates
table file, then set the environment variable AFNI_DEBUG_TABLE to YES
and run the program. A lot of progress reports will be printed out,
which may help pinpoint the problem; for example:
3dttest++ -DAFNI_DEBUG_TABLE=YES -covariates cfile.txt |& more
* A maximum of 31 covariates are allowed. If you have more, then
seriously consider the likelihood that you are completely deranged.
* N.B.: The simpler forms of the COVAR_FILE that 3dMEMA allows are
NOT supported here! Only the format described above will work.
* N.B.: IF you are entering multiple sub-bricks from the same dataset in
one of the '-setX' options, AND you are using covariates, then
you must use the 'LONG FORM' of input for the '-setX' option,
and give each sub-brick a distinct label that matches something
in the covariates file. Otherwise, the program will not know
which covariate to use with which input sub-brick, and bad
things will happen.
* N.B.: Please be careful in setting up the covariates file and dataset
labels, as the program only does some simple error checking.
++ If you REALLY want to see the regression matrices
used with covariates, use the '-debug' option.
++ Which you give you a LOT of output (to stderr), so redirect:
3dttest++ .... |& tee debug.out
***** CENTERING (this subject is very important -- read and think!) *******
++ This term refers to how the mean across subjects of a covariate
will be processed. There are 3 possibilities:
-center NONE = Do not remove the mean of any covariate.
-center DIFF = Each set will have the means removed separately.
-center SAME = The means across both sets will be computed and removed.
(This option only applies to a 2-sample test, obviously.)
++ These operations (DIFF or SAME) can be altered slightly by the following:
-cmeth MEAN = When centering, subtract the mean.
-cmeth MEDIAN = When centering, subtract the median.
(Per the request of the Musical Neuroscientist, AKA Steve Gotts.)
++ If you use a voxel-wise (dataset) covariate, then the centering method
is applied to each voxel's collection of covariate values separately.
++ The default operation is '-center DIFF'.
++ '-center NONE' is for the case where you have pre-processed the
covariate values to meet your needs; otherwise, it is not recommended!
++ Centering can be important. For example, suppose that the mean
IQ in setA is significantly higher than in setB, and that the beta
values are positively correlated with IQ IN THE SAME WAY IN THE
TWO GROUPS. Then the mean beta value in setA will be higher than in
setB simply from the IQ effect.
-- To attempt to allow for this type of inter-group mean differences,
in order to detect other difference between the two groups
(e.g., from disease status), you would have to center the two groups
together, rather than separately (i.e., use '-center SAME').
-- However, if the beta values are correlated significantly differently
with IQ in the two groups, then '-center DIFF' would perhaps be
a better choice. Please read on:
++ How to choose between '-center SAME' or '-center DIFF'? You have
to understand what your model is and what effect the covariates
are likely to have on the data. You shouldn't just blindly use
covariates 'just in case'. That way lies statistical madness.
-- If the two samples don't differ much in the mean values of their
covariates, then the results with '-center SAME' and '-center DIFF'
should be nearly the same.
-- For fixed covariates (not those taken from datasets), the program
prints out the results of a t-test of the between-group mean
covariate values. This test is purely informative; no action is
taken if the t-test shows that the two groups are significantly
different in some covariate.
-- If the two samples DO differ much in the mean values of their
covariates, then you should read the next point VERY CAREFULLY.
++ The principal purpose of including covariates in an analysis (ANCOVA)
is to reduce the variance of the beta values due to extraneous causes.
Some investigators also wish to use covariates to 'factor out' significant
differences between groups. However, there are those who argue
(convincingly) that if your two groups differ markedly in their mean
covariate values, then there is NO statistical test that can tell if
their mean beta values (dependent variable) would be the same or
different if their covariate values were all the same instead:
Miller GM and Chapman JP. Misunderstanding analysis of covariance.
J Abnormal Psych 110: 40-48 (2001).
http://dx.doi.org/10.1037/0021-843X.110.1.40
http://psycnet.apa.org/journals/abn/110/1/40.pdf
-- For example, if all your control subjects have high IQs and all your
patient subjects have normal IQs, group differences in activation can
be due to either cause (IQ or disease status) and you can't turn the
results from a set of high IQ controls into the results you would have
gotten from a set of normal IQ controls (so you can compare them to the
patients) just by linear regression and then pretending the IQ issue
goes away.
-- The decision as to whether a mean covariate difference between groups
makes the t-test of the mean beta difference invalid or valid isn't
purely a statistical question; it's also a question of interpretation
of the scientific issues of the study. See the Miller & Chapman paper
(above) for a lengthy discussion of this issue.
-- It is not clear how much difference in covariate levels is acceptable.
You could carry out a t-test on the covariate values between the
2 groups and if the difference in means is not significant at some
level (i.e., if p > 0.05?), then accept the two groups as being
'identical' in that variable. But this is just a suggestion.
(In fact, the program now carries out this t-test for you; cf supra.)
-- Thanks to Andy Mayer for pointing out this article to me.
++ At this time, there is no option to force the SLOPES of the
regression vs. covariate values to be the same in the two-sample
analysis. [Adding this feature would be too much like work.]
OTHER OPTIONS¶
-paired = Specifies the use of a paired-sample t-test to
compare setA and setB. If this option is used,
setA and setB must have the same cardinality (duh).
++ Recall that if '-paired' is used with '-covariates',
the covariates for setB will be the same as for setA.
++ If you don't understand the difference between a
paired and unpaired t-test, I'm not going to teach you
in this help file. But please consult someone or you
will undoubtedly come to grief!
-unpooled = Specifies that the variance estimates for setA and
setB be computed separately (not pooled together).
++ This only makes sense if -paired is NOT given.
++ '-unpooled' cannot be used with '-covariates'.
++ Unpooled variance estimates are supposed to
provide some protection against heteroscedasticty
(significantly different inter-subject variance
between the two different collections of datasets).
++ Our experience is that for most FMRI data, using
'-unpooled' is not needed; the option is here for
those who like to experiment or who are very cautious.
-toz = Convert output t-statistics to z-scores
++ -unpooled implies -toz, since t-statistics won't be
comparable between voxels as the number of degrees
of freedom will vary between voxels.
-->>++ -toz is automatically turned on with the -Clustsim option.
The reason for this is that -Clustsim (and -ETAC) work by
specifying voxel-wise thresholds via p-values -- z-statistics
are simpler to compute in the external clustering programs
(3dClustSim and 3dXClustSim) than t-statistics, since converting
a z=N(0,1) value to a p-value doesn't require knowing any
extra parameters (such as the t DOF).
-- In other words, I did this to make my life simpler.
++ If for some bizarre reason you want to convert a z-statistic
to a t-statistic, you can use 3dcalc with a clumsy expression
of the form
'cdf2stat(stat2cdf(x,5,0,0,0),3,DOF,0,0)'
where 'DOF' is replaced with the number of degrees of freedom.
The following command will show the effect of such a conversion:
1deval -xzero -4 -del 0.01 -num 801 \
-expr 'cdf2stat(stat2cdf(x,5,0,0,0),3,10,0,0)' | \
1dplot -xzero -4 -del 0.01 -stdin -xlabel z -ylabel 't(10)'
-zskip [n]= Do not include voxel values that are zero in the analysis.
++ This option can be used when not all subjects' datasets
overlap perfectly.
++ -zskip implies -toz, since the number of samples per
voxel will now vary, so the number of degrees of
freedom will be spatially variable.
++ If you follow '-zskip' with a positive integer (> 1),
then that is the minimum number of nonzero values (in
each of setA and setB, separately) that must be present
before the t-test is carried out. If you don't give
this value, but DO use '-zskip', then its default is 5
(for no good reason).
++ At this time, you can't use -zskip with -covariates,
because that would require more extensive re-thinking
and then serious re-programming.
++ You CAN use -zskip with -paired, but it works slightly
differently than with a non-paired test [06 May 2021]:
-- In a non-paired test, setA and setB are pruned of
zero values separately; e.g., setA could lose 3
values at a given voxel, while setB loses 5 there.
-- In a paired test, if EITHER setA or setB has a zero
value at a given voxel, both paired values are discarded.
This choice is necessary, since a paired t-test
requires subtracting the setA/setB values pairwise
and if one element of a pair is invalid, then the
other element has nothing to be paired with.
++ You can also put a decimal fraction between 0 and 1 in
place of 'n' (e.g., '0.9', or '90%'). Such a value
indicates that at least 90% (e.g.) of the values in each
set must be nonzero for the t-test to proceed. [08 Nov 2010]
-- In no case will the number of values tested fall below 3!
-- You can use '100%' for 'n', to indicate that all data
values must be nonzero for the test to proceed.
-rankize = Convert the data (and covariates, if any) into ranks before
doing the 2-sample analyses. This option is intended to make
the statistics more 'robust', and is inspired by the paper
WJ Conover and RL Iman.
Analysis of Covariance Using the Rank Transformation,
Biometrics 38: 715-724 (1982).
http://www.jstor.org/stable/2530051
Also see http://www.jstor.org/stable/2683975
++ Using '-rankize' also implies '-no1sam' (infra), since it
doesn't make sense to do 1-sample t-tests on ranks.
++ Don't use this option unless you understand what it does!
The use of ranks herein should be considered very
experimental or speculative!!
-no1sam = When you input two samples (setA and setB), normally the
program outputs the 1-sample test results for each set
(comparing to zero), as well as the 2-sample test results
for differences between the sets. With '-no1sam', these
1-sample test results will NOT be calculated or saved.
-nomeans = You can also turn off output of the 'mean' sub-bricks, OR
-notests = of the 'test' sub-bricks if you want, to reduce the size of
the output dataset. For example, '-nomeans -no1sam' will
result in only getting the t-statistics for the 2-sample
tests. These options are intended for use with '-brickwise',
where the amount of output sub-bricks can become overwhelming.
++ You CANNOT use both '-nomeans' and '-notests', because
then you would be asking for no outputs at all!
-nocov = Do not output the '-covariates' results. This option is
intended only for internal testing, and it's hard to see
why the ordinary user would want it.
-mask mmm = Only compute results for voxels in the specified mask.
++ Voxels not in the mask will be set to 0 in the output.
++ If '-mask' is not used, all voxels will be tested.
-->>++ It is VERY important to use '-mask' when you use '-ClustSim'
or '-ETAC' to computed cluster-level thresholds.
++ NOTE: voxels whose input data is constant (in either set)
will NOT be processed and will get all zero outputs. This
inaction happens because the variance of a constant set of
data is zero, and division by zero is forbidden by the
Deities of Mathematics -- cf., http://www.math.ucla.edu/~tao/
-exblur b = Before doing the t-test, apply some extra blurring to the input
datasets; parameter 'b' is the Gaussian FWHM of the smoothing
kernel (in mm).
++ This option is how '-ETAC_blur' is implemented, so it isn't
usually needed by itself.
++ The blurring is done inside the mask; that is, voxels outside
the mask won't be used in the blurring process. Such blurring
is done the same way as in program 3dBlurInMask (using a
finite difference evolution with Neumann boundary conditions).
++ Gaussian blurring is NOT additive in the FWHM parameter.
If the inputs to 3dttest++ were blurred by FWHM=4 mm
(e.g., via afni_proc.py), then giving an extra blur of
FWHM=6 mm is more-or-less equivalent to applying a single
blur of sqrt(4*4+6*6)=7.2 mm, NOT to 4+6=10 mm!
++ '-exblur' does not work with '-brickwise'.
++ '-exblur' only works with 3D datasets.
++ If any covariates are datasets, you should be aware that the
covariate datasets are NOT blurred by the '-exblur' process.
-brickwise = This option alters the way this program works with input
datasets that have multiple sub-bricks (cf. the SHORT FORM).
++ If you use this option, it must appear BEFORE either '-set'
option (so the program knows how to do the bookkeeping
for the input datasets).
++ WITHOUT '-brickwise', all the input sub-bricks from all
datasets in '-setA' are gathered together to form the setA
sample (similarly for setB, of course). In this case, there
is no requirement that all input datasets have the same
number of sub-bricks.
++ WITH '-brickwise', all input datasets (in both sets)
MUST have the same number of sub-bricks. The t-tests
are then carried out sub-brick by sub-brick; that is,
if you input a collection of datasets with 10 sub-bricks
in each dataset, then you will get 10 t-test results.
++ Each t-test result will be made up of more than 1 sub-brick
in the output dataset. If you are doing a 2-sample test,
you might want to use '-no1sam' to reduce the number of
volumes in the output dataset. In addition, if you are
only interested in the statistical tests and not the means
(or slopes for covariates), then the option '-nomeans'
will reduce the dataset to just the t (or z) statistics
-- e.g., the combination '-no1sam -nomeans' will give you
one statistical sub-brick per input sub-brick.
++ If you input a LOT of sub-bricks, you might want to set
environment variable AFNI_AUTOMATIC_FDR to NO, in order
to suppress the automatic calculation of FDR curves for
each t-statistic sub-brick -- this FDR calculation can
be time consuming when done en masse.
-->>++ The intended application of this option is to make it
easy to take a collection of time-dependent datasets
(e.g., from MEG or from moving-window RS-FMRI analyses),
and get time-dependent t-test results. It is possible to do
the same thing with a scripted loop, but that way is painful.
++ You CAN use '-covariates' with '-brickwise'. You should note
that each t-test will reuse the same covariates -- that is,
there is no provision for time-dependent covariate values --
for that, you'd have to use scripting to run 3dttest++
multiple times.
++ EXAMPLE:
Each input dataset (meg*.nii) has 100 time points; the 'X'
datasets are for one test condition and the 'Y' datasets are
for another. In this example, the subjects are the same in
both conditions, so the '-paired' option makes sense.
3dttest++ -brickwise -prefix megXY.nii -no1sam -paired\
-setA meg01X.nii meg02X.nii meg03X.nii ... \
-setB meg01Y.nii meg02Y.nii meg03Y.nii ...
* The output dataset will have 200 sub-bricks: 100 differences
of the means between 'X' and 'Y', and 100 t-statistics.
* You could extract the output dataset t-statistics (say)
into a single dataset with a command like
3dTcat -prefix megXY_tstat.nii megXY.nii'[1..$(2)]'
(Or you could have used the '-nomeans' option.)
This dataset could then be used to plot the t-statistic
versus time, make a movie, or otherwise do lots of fun things.
* If '-brickwise' were NOT used, the output dataset would just
get 2 sub-bricks, as all the inputs in setA would be lumped
together into one super-sized sample (and similarly for setB).
* Remember that with the SHORT FORM input (needed for option
'-brickwise') you can use wildcards '*' and '?' together with
'[...]' sub-brick selectors.
-prefix p = Gives the name of the output dataset file.
++ For surface-based datasets, use something like:
-prefix p.niml.dset or -prefix p.gii.dset
Otherwise you may end up files containing numbers but
not a full set of header information.
-resid q = Output the residuals into a dataset with prefix 'q'.
++ The residuals are the difference between the data values
and their prediction from the set mean (and set covariates).
++ For use in further analysis of the results (e.g., 3dFWHMx).
++ Cannot be used with '-brickwise' (sorry).
++ If used with '-zskip', values which were skipped in the
analysis will get residuals set to zero.
-ACF = If residuals are saved, also compute the ACF parameters from
them using program 3dFHWMx -- for further use in 3dClustSim
(which must be run separately).
++ HOWEVER, the '-Clustsim' option below provides a resampling
alternative to using the parametric '-ACF' method in
program 3dClustSim.
-dupe_ok = Duplicate dataset labels are OK. Do not generate warnings
for dataset pairs.
** This option must precede the corresponding -setX options.
** Such warnings are issued only when '-covariates' is used
-- when the labels are used to extract covariate values
from the covariate table.
-debug = Prints out information about the analysis, which can
be VERY lengthy -- not for general usage (or even for colonels).
++ Two copies of '-debug' will give even MORE output!
ClustSim Options – for global cluster-level thresholding and FPR control¶
The following options are for using randomization/permutation to simulate
noise-only generated t-tests, and then run those results through the
cluster-size threshold simulation program 3dClustSim. The goal is to
compute cluster-size thresholds that are not based on a fixed model
for the spatial autocorrelation function (ACF) of the noise.
ETAC (infra) and ClustSim are parallelized. The randomized t-test steps are
done by spawning multiple 3dttest++ jobs using the residuals as input.
Then the 3dClustSim program (for -Clustsim) and 3dXClustSim program (for -ETAC)
use multi-threaded processing to carry out their clusterization statistics.
If your computer does NOT have multiple CPU cores, then these options will
run very very slowly.
You can use both -ETAC and -Clustsim in the same run. The main reason for
doing this is to compare the results of the two methods. Using both methods
in one 3dttest++ run will be super slow.
++ In such a dual-use case, and if '-ETAC_blur' is also given, note that
3dClustSim will be run once for each blur level, giving a set of cluster-
size threshold tables for each blur case. This process is necessary since
3dClustSim does not have a multi-blur thresholding capability, unlike
ETAC (via program 3dXClustSim).
++ The resulting 3dClustSim tables are to be applied to each of the auxiliary
t-test files produced, one for each blur case. Unless one of those blur
cases is '0.0', the 3dClustSim tables do NOT apply to the main output
dataset produced by this program.
++ These auxiliary blur case t-test results get names of the form
PREFIX.B8.0.nii
where PREFIX was given in the '-prefix' option, and in this example,
the amount of extra blurring was 8.0 mm. These files are the result
of re-running the commanded t-tests using blurred input datasets.
-Clustsim = With this option, after the commanded t-tests are done, then:
(a) the residuals from '-resid' are used with '-randomsign' to
simulate about 10000 null 3D results, and then
(b) 3dClustSim is run with those to generate cluster-threshold
tables, and then
(c) 3drefit is used to pack those tables into the main output
dataset, and then
(d) the temporary files created in this process are deleted.
The goal is to provide a method for cluster-level statistical
inference in the output dataset, to be used with the AFNI GUI
Clusterize controls.
++ If you want to keep ALL the temporary files, use '-CLUSTSIM'.
They will include the z-scores from all the simulations.
** Normally, the permutation/randomization z-scores are saved
in specially compressed files with suffix '.sdat'. If you
want these files in the '.nii' format, use the options
'-DAFNI_TTEST_NIICSIM=YES -CLUSTSIM'.
** However, if '-ETAC' is also used, the '.sdat' format will
be used instead of the '.nii' format, as the program that
implements ETAC (3dXClustSim) requires that format.
** You can change the number of simulations using an option
such as '-DAFNI_TTEST_NUMCSIM=20000' if you like.
++ Since the simulations are done with '-toz' active, the program
also turns on the '-toz' option for your output dataset. This
means that the output statistics will be z-scores, not t-values.
++ If you have less than 14 datasets total (setA & setB combined),
this option will not work! (There aren't enough random subsets.)
** And it will not work with '-singletonA'.
-->>++ '-Clustsim' runs step (a) in multiple jobs, for speed. By
default, it tries to auto-detect the number of CPUs on the
system and uses that many separate jobs. If you put a positive
integer immediately following the option, as in '-Clustsim 12',
it will instead use that many jobs (e.g., 12). This capability
is to be used when the CPU count is not auto-detected correctly.
** You can also set the number of CPUs to be used via the Unix
environment variable OMP_NUM_THREADS.
** This program does not use OpenMP (OMP), but since many other
AFNI programs do, setting OMP_NUM_THREADS is a common way
to set the amount of parallel computation to use.
-->>++ It is important to use a proper '-mask' option with '-Clustsim'.
Otherwise, the statistics of the clustering will be skewed.
-->>++ You can change the number of simulations from the default 10000
by setting Unix environment variable AFNI_TTEST_NUMCSIM to a
different value (in the range 1000..1000000). Note that the
3dClustSim tables go down to a cluster-corrected false positive
rate of 0.01, so that reducing the number of simulations below
10000 will produce notably less accurate results for such small
FPR (alpha) values.
**-->>++ The primary reason for reducing AFNI_TTEST_NUMCSIM below its
default value is testing '-Clustsim' and/or '-ETAC' more quickly
-->>++ The clever scripter can pick out a particular value from a
particular 3dClustSim output .1D file using the '{row}[col]'
syntax of AFNI, as in the tcsh command
set csize = `1dcat Fred.NN1_1sided.1D"{10}[6]"`
to pick out the number in the #10 row, #6 column (counting
from #0), which is the p=0.010 FPR=0.05 entry in the table.
-->++ Or even *better* now for extracting a table value:
a clever person added command line options to 1d_tool.py
to extract a value from the table having a voxelwise p-value
('-csim_pthr ..') and an FDR alpha level ('-csim_alpha ..').
Be sure to check out those options in 1d_tool.py's help!
**-->>++ NOTE: The default operation of 3dClustSim when used from
3dttest++ is with the '-LOTS' option controlling
the thresholds used for the tabular output.
You can change that to the '-MEGA' option = a larger
table, by setting Unix environment variable
AFNI_CLUSTSIM_MEGA to YES. You can do that in several
ways, including on the command line with the option
'-DAFNI_CLUSTSIM_MEGA=YES'. [15 Dec 2021 - RWCox]
---==>>> PLEASE NOTE: This option has been tested for 1- and 2-sample
---==>>> unpaired and paired tests vs. resting state data -- to see if the
---==>>> false positive rate (FPR) was near the nominal 5% level (it was).
---==>>> The FPR for the covariate effects (as opposed to the main effect)
---==>>> is still somewhat biased away from the 5% level /:(
****** The following options affect both '-Clustsim' and '-ETAC' ******
-prefix_clustsim cc = Use 'cc' for the prefix for the '-Clustsim' temporary
files, rather than a randomly generated prefix.
You might find this useful if scripting.
++ By default, the Clustsim (and ETAC) prefix will
be the same as that given by '-prefix'.
-->>++ If you use option '-Clustsim', then the simulations
keep track of the maximum (in mask) voxelwise
z-statistic, compute the threshold for 5% global FPR,
and write those values (for 1-sided and 2-sided
thresholding) to a file named 'cc'.5percent.txt --
where 'cc' is the prefix given here. Using such a
threshold in the AFNI GUI will (presumably) give you
a map with a 5% chance of false positive WITHOUT
clustering. Of course, these thresholds generally come
with a VERY stringent per-voxel p-value.
** In one analysis, the 5% 2-sided test FPR p-value was
about 7e-6 for a mask of 43000 voxels, which is
bigger (less strict) than the 1.2e-6 one would get
from the Bonferroni correction, but is still very
stringent for many purposes. This threshold value
was also close to the threshold at which the FDR
q=1/43000, which may not be a coincidence.
-->>++ This file has been updated to give the voxel-wise
statistic threshold for global FPRs from 1% to 9%.
However, the name is still '.5percent.txt' for the
sake of nostalgia.
-no5percent = Don't output the 'cc'.5percent.txt file that comes
for free with '-Clustsim' and/or '-ETAC'.
++ But whyyy? Don't you like free things?
-tempdir ttt = Store temporary files for '-Clustsim' in this directory,
rather than in the current working directory.
-->>++ This option is for use when you have access to a fast
local disk (e.g., SSD) compared to general storage
on a rotating disk, RAID, or network storage.
++ Using '-tempdir' can make a significant difference
in '-Clustsim' and '-ETAC' runtime, if you have
a local solid state drive available!
[NOTE: with '-CLUSTSIM', these files aren't deleted!]
-seed X [Y] = This option is used to set the random number seed for
'-randomsign' to the positive integer 'X'. If a second integer
'Y' follows, then that value is used for the random number seed
for '-permute'.
++ The purpose of setting seeds (rather than letting the program
pick them) is for reproducibility. It is not usually needed by
the ordinary user.
++ Option '-seed' is used by the multi-blur analysis possible
with '-ETAC', so that the different blur levels use the same
randomizations, to make their results compatible for multi-
threshold combination.
++ Example: -seed 3217343 1830201
***** These options (below) are not usually directly used, but *****
***** are described here for completeness and for reference. *****
***** They are invoked by options '-Clustsim' and '-ETAC'. *****
-randomsign = Randomize the signs of the datasets. Intended to be used
with the output of '-resid' to generate null hypothesis
statistics in a second run of the program (probably using
'-nomeans' and '-toz'). Cannot be used with '-singletonA'
or with '-brickwise'.
++ You will never get an 'all positive' or 'all negative' sign
flipping case -- each sign will be present at least 15%
of the time.
++ There must be at least 4 samples in each input set to
use this option, and at least a total of 14 samples in
setA and setB combined.
++ If you following '-randomsign' with a number (e.g.,
'-randomsign 1000'), then you will get 1000 iterations
of random sign flipping, so you will get 1000 times the
as many output sub-bricks as usual. This is intended for
for use with simulations such as '3dClustSim -inset'.
-->>++ This option is usually not used directly, but will be
invoked by the use of '-Clustsim' and/or '-ETAC'. It is
documented here for the sake of telling the Galaxy how the
program works.
-permute = With '-randomsign', and when both '-setA' and '-setB' are used,
this option will add inter-set permutation to the randomization.
++ If only '-setA' is used (1-sample test), there is no permutation.
(Neither will there be permutation with '-singletonA'.)
++ If '-randomsign' is NOT given, but '-Clustsim' is used, then
'-permute' will be passed for use with the '-Clustsim' tests
(again, only if '-setA' and '-setB' are both used).
++ If '-randomsign' is given and if the following conditions
are ALL true, then '-permute' is assumed (without the option
needed on the command line):
(a) You have a 2-sample test.
And, you are not using '-singletonA'.
[Permutation is meaningless without 2 samples!]
(b) And, you are not using '-unpooled'.
(c) And, you are not using '-paired'.
-->>++ You only NEED to use '-permute' if you want inter-set
permutation used AND you the '-unpooled' option.
+ Permutation with '-unpooled' is a little weird.
+ Permutation with '-paired' is very weird and is NOT allowed.
+ Permutation with '-covariates' may not work the way you wish.
In the past [pre-March 2020], covariates were NOT permuted along
with their data. Now, covariate ARE permuted along with their data.
This latter method seems more logical to me [RWCox].
++ There is no option to do permutation WITHOUT sign randomization.
-->>++ AGAIN: This option is NOT usually used directly by the user;
it will be invoked by the '-Clustsim' or '-ETAC' operations.
-nopermute = This option is present if you want to turn OFF the automatic
use of inter-set permutation with '-randomsign'.
++ I'm not sure WHY you would want this option, but it is here
for completeness of the Galactic Chronosynclastic Infundibulum.
ETAC Options – [promulgated May 2017]¶
The following options use the ETAC (Equitable Thresholding And Clustering)
method to provide a method for thresholding the results of 3dttest++.
-ETAC uses randomization/permutation to generate null distributions,
as does -Clustsim. The main difference is that ETAC also allows:
* use of multiple per-voxel p-value thresholds simultaneously
* use of cluster-size and/or cluster-square-sum as threshold parameters
* use of multiple amounts of blurring simultaneously
'Equitable' means that each combination of the above choices is treated
to contribute approximately the same to the False Positive Rate (FPR).
Major differences between '-Clustsim' and '-ETAC':
* -Clustsim produces a number: the cluster-size threshold to be used everywhere.
* -ETAC produces a map: the cluster figure of merit (FOM) threshold to be
used as a function of location.
* -ETAC allows use of a FOM that is more general than the cluster-size.
* -ETAC allows the use of multiple per-voxel p-value thresholds simultaneously.
* -ETAC allows the use of multiple blur levels simultaneously.
*** ALSO see the description of the '-prefix_clustsim', '-tempdir', and ***
*** '-seed' options above, since these also affect the operation of ETAC ***
*** The 'goal' of ETAC is a set of thresholds that give a 5% FPR. You ***
*** can modify this goal by setting the 'fpr=' parameter via '-ETAC_opt' ***
* ETAC can use a lot of memory; about 100000 * Ncase * Nmask bytes,
where Ncase = number of blur cases in option '-ETAC_blur' and
Nmask = number of voxels in the mask.
For example, 50000 voxels in the mask and 4 blur cases might use about
50000 * 100000 * 4 = 20 billion bytes of memory.
* Run time depends a lot on the parameters and the computer hardware, but
will typically be 10-100 minutes. Get another cup of tea (or coffee).
*** You should use ETAC only on a computer with ***
*** multiple CPU cores and lots of RAM! ***
*** If 3dXClustSim fails with the message ***
*** 'Killed', this means that the operating ***
*** system stopped the program for trying to ***
*** use too much memory. ***
-ETAC [ncpu] = This option turns ETAC computations on.
++ You can put the maximum number of CPUs to use
after '-ETAC' if you want, but it is usually
not needed -- just let the program choose.
++ The ETAC algorithms are implemented in program
3dXClustSim, which 3dttest++ will run for you.
++ As with '-Clustsim', you can put the number of CPUs
to be used after the '-ETAC' option, or let the
program figure out how many to use.
-ETAC_global = Do the ETAC calculations 'globally' - that is, produce
multi-threshold values to apply to the entire volume
rather than voxelwise.
--->>++ This is the default mode of operation for ETAC.
++ These global calculations are kind of like '-Clustsim'
in that they produce a set of cluster thresholds to
apply everywhere in the brain - a small set of numbers.
The difference from '-Clustsim' is that for a given FPR,
the set of cluster threshold values are intended to
be applied simultaneously.
Output files -->>>++ The output from global ETAC is a binary mask file indicating
binary mask which voxels survived the multi-thresholding process.
(main result) The name of such a file follows the format
{PREFIX}.{NAME}.ETACmask.global.{SIDE}.{FPR}.nii.gz
where {PREFIX} is from'-prefix' or '-prefix_clustsim',
{NAME} is the name given in '-ETAC_opt',
{SIDE} is '1pos' and '1neg' if 1-sided testing
was ordered in '-ETAC_opt',
or is '2sid' if 2-sided testing was ordered.
{FPR} is the false positive rate (e.g., '7perc')
-> It is very possible that this output mask will be all
zero, indicating that nothing survived. A quick way
to see how many voxels made it through the ETAC process:
3dBrickStat -non-zero -count MaskDatasetName.nii.gz
This command will print (to stdout) a single integer
of the count of non-zero voxels in this mask dataset.
Output files --->>++ A similarly named file, with '.ETACmaskALL.' replacing
which tests '.ETACmask.' is also output, which has 1 binary volume for
'passed' in each thresholding sub-test (i.e., number of p-thresholds
each voxel times number of blur levels); this dataset marks each
voxel with the set of tests that were passed there.
Output files --->>++ The actual output thresholds are stored in text files
the thresholds (using an XML format) with a name like
globalETAC.mthresh.{PREFIX}.{NAME}.ETAC.{LEVEL}.{FPR}.niml
where {PREFIX} is from'-prefix' or '-prefix_clustsim',
{NAME} is the name given in '-ETAC_opt',
{LEVEL} is the blur level name (e.g., 'B8.0')
{FPR} is the false positive rate (e.g., '7perc')
The multiple thresholds are available as a column of
numbers in the single XML element in this file.
If multiple blur levels are used, there will be one
such file for each blur level.
-ETAC_mem = This option tells the program to print out the
estimate of how much memory is required by the ETAC
run as ordered, and then stop.
++ No data analysis of any kind will be performed.
++ You have to give all the options (-setA, -ETAC, etc.)
that you would use to run the analysis.
++ The purpose of this option is to help you choose
the computer setup for your run.
-ETAC_blur b1 b2 ... = This option says to use multiple levels of spatial
blurring in the t-tests and ETAC analysis.
++ If you do NOT use -ETAC_blur, then no extra
blurring is used, beyond whatever might have
been used on the inputs to 3dttest++.
++ Note that Gaussian blurring is NOT additive
in the FWHM parameter, but is rather additive in
the square of FWHM. If the inputs to 3dttest++
are blurred by FWHM=4 mm (for example), then giving
an extra blur of FWHM=6 mm is equivalent to a
single blur of sqrt(4*4+6*6)=7.2 mm, NOT to 10 mm!
++ The list of blur FWHM parameters can have up to 5
entries, but I recommend no more than 2 or 3 of them.
3dXClustSim memory usage goes up sharply as the
number of blur cases rises.
++ You can use '0' for one of the blur parameters here,
meaning to not apply any extra blurring for that case.
++ We recommend blurring no more than 3 times the original
EPI voxel dimension.
++ You can only use '-ETAC_blur' once.
++ '-ETAC_blur' is implemented via '-exblur', and the blurring
is done only inside the analysis mask (cf. 3dBlurInMask).
-ETAC_opt params = This option lets you choose the non-blurring parameters
for ETAC. You can use this option more than once, to
have different thresholding cases computed from the
same set of permutations -- since the permutations are
slow, this multiple cases ability is here to help
you speed things up when you are trying out different
possibilities.
The 'params' string is one argument, with different
parts separated by colon ':' characters. The parts are
NN=1 or NN=2 or NN=3 } spatial connectivity for clustering
sid=1 or sid=2 } 1-sided or 2-sided t-tests
pthr=p1,p2,... } list of p-values to use
hpow=h1,h2,... } list of H powers (0, 1, and/or 2)
fpr=value } FPR goal, between 1 and 9 (percent)
} - must be an integer
} - or the word 'ALL' to output
} results for 1, 2, 3, 4, ..., 9.
} - or the word 'MUCHO' to output
} result for 1, 2, ..., 24, 25.
name=Something } a label to distinguish this case
For example:
-ETAC_opt NN=2:sid=2:hpow=0,2:pthr=0.01,0.005,0.002,0.01:name=Fred
++ You can use '-ETAC_opt' more than once, to make
efficient reuse of the randomized/permuted cases.
-->> Just give each use within the same 3dttest++ run a
different label after 'name='.
+ It is important to use distinct names for each
different '-ETAC_opt' case, so that the output
file names will be distinct.
++ There's no built-in upper limit to the number of
'-ETAC_opt' cases you can run.
Each time you use '-ETAC_opt', 3dXClustSim will be run
(using the same set of permutations/randomizations).
++ The H powers ('hpow') allowed are 0, 1, and/or 2;
the clustering figure of merit (FOM) is defined as the
sum over voxels in a cluster of the voxel absolute
z-scores raised to the H power; H=0 is the number of
voxels in a cluster (what 3dClustSim uses).
+ Although ETAC allows you to multi-threshold across
multiple hpow values, there is little reason to
actually do this. My recommendation:
Choose hpow=0 for cluster-size, if you want
to make it easier to explain your methods.
Choose hpow=2 to weight voxelwise z-statistic
more, which will make detection of small
high intensity clusters somewhat more likely.
++ There is no built in upper limit on the number of
'pthr' levels allowed. If you wish to use an
arithmetically evenly spaced set of p thresholds,
you can do something like 'pthr=0.01/0.001/19' to
use 19 thresholds evenly spaced from 0.01 to 0.001,
with step size (0.01-0.001)/(19-1)=0.0005.
+ In the form 'pthr=A/B/N', the count N must be at
least 2, or only the value of A will be used.
+ pthr values must be in the range 0.1 .. 0.0001
(inclusive), or this program will be unhappy.
+ pthr values are interpreted in the context of
1-sided or 2-sided testing when the actual
statistic values for thresholding are computed.
+ Of course, the program gets slower for larger
numbers of pthr levels and will use more memory.
A practical upper bound for the number of pthr
levels is about 20. I have run it (experimentally)
with 91 pthr levels, which made very little
difference in the results from using 10 pthr values,
and it took much longer to run.
++ NN=1 means clustering using the 6 nearest neighbors;
NN=2 means clustering using 18 neighboring voxels
(NN=1 plus the 12 second nearest neighbors);
NN=3 means clustering using 26 neighboring voxels
(NN=2 plus the 8 third nearest neighbors).
++ sid=1 means to carry out voxelwise 1-sided t-tests,
which will result in output masks labeled
'1pos' for the set of voxels that survive the
positive side of the t-tests (at the given
1-sided p-thresholds) plus ETAC clustering, and
'1neg' for the corresponding negative side
of the t-tests.
sid=2 means to carry out voxelwise 2-sided t-tests,
which will result in an output mask labeled
'2sid'.
++ Do not confuse the '1-sided' and '2-sided' testing
choice with the '1-sample' or '2-sample' analysis
being carried out by 3dttest++. Although these
concepts are completely distinct, the naming
with numerals can be a source of distraction.
-->>++ If you do not use '-ETAC_opt' at all, a built-in set
of parameters will be used. These are
NN=2 sid=2 hpow=2 name=default
pthr=0.01/0.001/10
=0.010,0.009,0.008,0.007,0.006,0.005,0.004,0.003,0.002,0.001
fpr=5
-->>++ Note that using 'fpr=ALL' will make the ETAC calculations
slower, as the software has to compute results for 9 different
FPR goals, each of which requires thrashing through all
the pseudo-random simulations at least once.
+ On the other hand, seeing how the results mask varies
as the FPR goal changes can be illuminating.
-ETAC_arg something = This option is used to pass extra options to the
3dXClustSim program (which is what implements ETAC).
There is almost no reason to use this option that I
can think of, except perhaps this example:
-ETAC_arg -verb
which will cause 3dXClustSim to print more fun fun fun
information as it progresses through the ETAC stages.
*** WARNING: ETAC consumes a lot of CPU time, and a lot of memory ***
*** (especially with many -ETAC_blur cases, or 'fpr=ALL')! ***
+++ (: One of these days, I'll expand this section and explain ETAC more :) +++
+++ (: ------------------------------ MAYBE ---------------------------- :) +++
STRUCTURE OF THE OUTPUT DATASET¶
* 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]
#1 SetA-SetB_Tstat
#2 SetA-SetB_IQ = difference of slopes wrt covariate IQ
#3 SetA-SetB_IQ_Tstat
#4 SetA_mean = mean of SetA [covariates removed]
#5 SetA_Tstat
#6 SetA_IQ = slope of SetA wrt covariate IQ
#7 SetA_IQ_Tstat
#8 SetB_mean = mean of SetB [covariates removed]
#9 SetB_Tstat
#10 SetB_IQ = slope of SetB wrt covariate IQ
#11 SetB_IQ_Tstat
* In the above, 'wrt' is standard mathematical shorthand for the
phrase 'with respect to'.
* If option '-BminusA' is given, then the 'SetA-SetB' sub-bricks would
be labeled 'SetB-SetA' instead, of course.
* If option '-toz' is used, the 'Tstat' will be replaced with 'Zscr'
in the statistical sub-brick labels.
* If the long form of '-setA' is used, or '-labelA' is given, then
'SetA' in the sub-brick labels above is replaced with the
corresponding SETNAME. (Mutatis mutandis for 'SetB'.)
* If you produce a NIfTI-1 (.nii) file, then the sub-brick labels are
saved in the AFNI extension in the .nii file. Processing further
in non-AFNI programs will probably cause these labels to be lost
(along with other AFNI niceties, such as the history field).
* If you are doing a 2-sample run and don't want the 1-sample results,
then the '-no1sam' option can be used to eliminate these sub-bricks
from the output, saving space and time and mental energy.
* The largest Tstat that will be output is 99.
* The largest Zscr that will be output is 13.
++ FYI: the 1-sided Gaussian tail probability of z=13 is 6.1e-39.
HOW COVARIATES WORK¶
Covariates work by forming a regression problem for each voxel, to
estimate the mean of the input data and the slopes of the data with
respect to variations in the covariates.
For each input set of sub-bricks, a matrix is assembled. There is one
row for each sub-brick, and one column for each covariate, plus one
more column for the mean. So if there are 5 sub-bricks and 2 covariates,
the matrix would look like so
[ 1 0.3 1.7 ]
[ 1 0.5 2.2 ]
X = [ 1 2.3 3.3 ]
[ 1 5.7 7.9 ]
[ 1 1.2 4.9 ]
The first column is all 1s, and models the mean value of the betas.
The remaining columns are the covariates for each sub-brick. (The
numbers above are values I just made up, obviously.)
The matrix is centered by removing the mean from each column except
the first one. In the above matrix, the mean of column #2 is 2,
and the mean of column #3 is 4, so the centered matrix is
[ 1 -1.7 -2.3 ]
[ 1 -1.5 -1.8 ]
Xc = [ 1 0.3 -0.7 ]
[ 1 3.7 3.9 ]
[ 1 -0.8 0.9 ]
(N.B.: more than one centering option is available; this is the default.)
The set of equations to be solved is [Xc] [b] = [z], where [b] is
the column vector desired (first element = de-covariate-ized mean
of the data values, remaining elements = slopes of data values
with respect to the covariates), and [z] is the column vector of
data values extracted from the input datasets.
This set of equations is solved by forming the pseudo-inverse of the
matrix [Xc]: [Xp] = inverse[Xc'Xc] [Xc'], so that [b] = [Xp] [z].
(Here, ' means transpose.) For the sample matrix above, we have
[ 0.2 0.2 0.2 0.2 0.2 ]
Xp = [ 0.0431649 -0.015954 0.252887 0.166557 -0.446654 ]
[ -0.126519 -0.0590721 -0.231052 0.0219866 0.394657 ]
Because of the centering, the first column of [Xc] is orthgonal to
the other columns, so the first row of [Xp] is all 1/N, where N is
the number of data points (here, N=5).
In reality, the pseudo-inverse [Xp] is computed using the SVD, which
means that even a column of all zero covariates will not cause a
singular matrix problem.
In addition, the matrix [Xi] = inverse[Xc'Xc] is computed. Its diagonal
elements are needed in the t-test computations. In the above example,
[ 0.2 0 0 ]
Xi = [ 0 0.29331 -0.23556 ]
[ 0 -0.23556 0.22912 ]
For a 1-sample t-test, the regression values computed in [b] are the
'_mean' values stored in the output dataset. The t-statistics are
computed by first calculating the regression residual vector
[r] = [Xc][b] - [z] (the mismatch between the data and the model)
and then the estimated variance v of the residuals is given by
i=N
q = sum { r[i]*r[i] } and then v = q / (N-m)
i=1
where N=number of data points and m=number of matrix columns=number of
parameters estimated in the regression model. The t-statistic for the
k-th element of [b] is then given by
t[k] = b[k] / sqrt( v * Xi[k,k] )
Note that for the first element, the factor Xi[1,1] is just 1/N, as
is the case in the simple (no covariates) t-test.
For a 2-sample unpaired t-test, the '_mean' output for the k-th column
of the matrix [X] is bA[k]-bB[k] where 'A' and 'B' refer to the 2 input
collections of datasets. The t-statistic is computed by
vAB = (qA+qB) / (NA+NB-2*m)
t[k] = (bA[k]-bB[k]) / sqrt( vAB * (XiA[k,k]+XiB[k,k]) )
For a 2-sample paired t-test, the t-statistic is a little different:
i=N
q = sum { (rA[i]-rB[i])^2 } and then vAB = q / (N-m)
i=1
and then
t[k] = (bA[k]-bB[k]) / sqrt( vAB * XiA[k,k] )
A paired t-test is basically a 1-sample test with the 'data' being
the difference [zA]-[zB] of the two input samples.
Note the central role of the diagonal elements of the [Xi] matrix.
These numbers are the variances of the estimates of the [b] if the
data [z] is corrupted by additive white noise with variance=1.
(In the case of an all zero column of covariates, the SVD inversion)
(that yields [Xi] will make that diagonal element 0. Division by 0)
(being a not-good thing, in such a case Xi[k,k] is replaced by 1e9.)
For cases with voxel-wise covariates, each voxel gets a different
[X] matrix, and so the matrix inversions are carried out many many
times. If the covariates are fixed values, then only one set of
matrix inversions needs to be carried out.
HOW SINGLETON TESTING WORKS WITH COVARIATES¶
(1) For setB, the standard regression is carried out to give the
covariate slope estimates (at each voxel):
[b] = [Xp] [z]
where [z] = column vector of the setB values
[Xp] = pseudo-inverse of the [X] matrix for the setB covariates
[b] = covariate parameter estimates
Under the usual assumptions, [b] has mean [b_truth] and covariance
matrix sigma^2 [Xi], where sigma^2 = variance of the zB values, and
[Xi] = inverse[X'X]. (Again, ' = transpose.)
(If centering is used, [X] is replaced by [Xc] in all of the above.)
(2) Call the singletonA value (at each voxel) y;
then the statistical model for y is
y = yoff + [c]'[b_truth] + Normal(0,sigma^2)
where the column vector [c] is the transpose of the 1-row matrix [X]
for the singletonA dataset -- that is, the first element of [c] is 1,
and the other elements are the covariate values for this dataset.
(The null hypothesis is that the mean offset yoff is 0.)
The covariate slopes [b] from step (1) are projected out of y now:
y0 = y - [c]'[b]
which under the null hypothesis has mean 0 and variance
sigma^2 ( 1 + [c]'[Xi][c] )
Here, the '1' comes from the variance of y, and the [c]'[Xi][c] comes
from the variance of [b] dotted with [c]. Note that in the trivial
case of no covariates, [X] = 1-column matrix of all 1s and [c] = scalar
value of 1, so [c]'[Xi][c] = 1/N where N = number of datasets in setB.
(3) sigma^2 is as usual estimated by s^2 = sum[ (z_i - mean(z))^2 ] / (N-m-1)
where N = number of datasets in setB and m = number of covariates.
Under the usual assumptions, s^2 is distributed like a random variable
( sigma^2 / (N-m-1) ) * ChiSquared(N-m-1).
(4) Consider the test statistic
tau = y0 / sqrt(s^2)
Under the null hypothesis, this has the distribution of a random variable
Normal(0,(1 + [c]'[Xi][c]) / sqrt( ChiSquared(N-m-1)/(N-m-1) )
So tau is not quite t-distributed, but dividing out the scale factor works:
t = y0 / sqrt( s^2 * (1 + [c]'[Xi][c]) )
and under the null hypothesis, this value t has a Student(N-m-1) distribution.
Again, note that in the case of no covariates, [c]'[Xi][c] = 1/N, so that
t = y / sqrt( s^2 * (1+1/N) )
If we were testing against a constant y, rather than y itself being random,
we'd have
t_con = y / sqrt( s^2 / (N-1) )
which shows that the t statistic for the '-singletonA' test will usually be
smaller than the t statistic for the 'test against constant' case --
because we have to allow for the variance of the singleton dataset value y.
Please note that the singleton dataset is assumed to be statistically
independent of the reference datasets -- if you put the singleton dataset
into the reference collection, then you are violating this assumption --
a different statistic would have to be computed.
A test script that simulates random values and covariates has verified the
distribution of the results in both the null hypothesis (yoff == 0) case and the
alternative hypothesis (yoff !=0) case -- where the value t now takes on the
non-central Student distribution.
Below is a sketch of how a covariate might be useful in singleton tests:
* the 'z' labels are voxel values from setB
* the 'y' label is the voxel value from singletonA
* y is not markedly different from some of the z values
* but for the singleton subject's age, y IS very different
* a test WITHOUT the age covariate would not give a large t-statistic for y
* a test WITH the age covariate will show a larger t-statistic for y
D | z |
a | z |
t | z z z z |
a | z z z z |
| z z z z z |
v | z z z z z |
a | z z z z z |
l | z z z z |
u | z z z y |
e | z z |
| |
| |
| |
Subject age
After linear regression removes the covariate effect (values at smaller
ages are increased and values at larger ages are decreased), the cartoon
graph would look something like this, where the modified y value is
now clearly far away from the cluster of z values:
R D | |
e a | |
g t | z z z |
r a | z zz z z z z z |
e | z z zz |
s v | z z z z z |
s a | z z z z zzz z |
e l | z z z |
d u | z z z |
e | |
| |
| |
| y |
Subject age
A NOTE ABOUT p-VALUES (everyone’s favorite subject :)¶
The 2-sided p-value of a t-statistic value T is the likelihood (probability)
that the absolute value of the t-statistic computation would be bigger than
the absolute value of T, IF the null hypothesis of no difference in the means
(2-sample test) were true. For example, with 30 degrees of freedom, a T-value
of 2.1 has a p-value of 0.0442 -- that is, if the null hypothesis is true
and you repeated the experiment a lot of times, only 4.42% of the time would
the T-value get to be 2.1 or bigger (and -2.1 or more negative).
You can NOT interpret this to mean that the alternative hypothesis (that the
means are different) is 95.58% likely to be true. (After all, this T-value
shows a pretty weak effect size -- difference in the means for a 2-sample
t-test, magnitude of the mean for a 1-sample t-test, scaled by the standard
deviation of the noise in the samples.) A better way to think about it is
to pose the following question:
Assuming that the alternative hypothesis is true, how likely
is it that you would get the p-value of 0.0442, versus how
likely is p=0.0442 when the null hypothesis is true?
This is the question addressed in the paper
Calibration of p Values for Testing Precise Null Hypothese.s
T Sellke, MJ Bayarri, and JO Berger.
The American Statistician v.55:62-71, 2001.
www2.stat.duke.edu/~berger/papers/99-13.ps
The exact interpretation of what the above question means is somewhat
tricky, depending on if you are a Bayesian heretic or a Frequentist
true believer. But in either case, one reasonable answer is given by
the function
alpha(p) = 1 / [ 1 - 1/( e * p * log(p) ) ]
(where 'e' is 2.71828... and 'log' is to the base 'e'). Here,
alpha(p) can be interpreted as the likelihood that the given p-value
was generated by the null hypothesis, versus being from the alternative
hypothesis. For p=0.0442, alpha=0.2726; in non-quantitative words, this
p-value is NOT very strong evidence that the alternative hypothesis is true.
Why is this so -- why isn't saying 'the null hypothesis would only give
a result this big 4.42% of the time' similar to saying 'the alternative
hypothesis is 95.58% likely to be true'? The answer is because it is
only somewhat more likely the t-statistic would be that value when the
alternative hypothesis is true. In this example, the difference in means
cannot be very large, or the t-statistic would almost certainly be larger.
But with a small difference in means (relative to the standard deviation),
the alternative hypothesis (noncentral) t-value distribution isn't that
different than the null hypothesis (central) t-value distribution. It is
true that the alternative hypothesis is more likely to be true than the
null hypothesis (when p < 1/e = 0.36788), but it isn't AS much more likely
to be true than the p-value itself seems to say.
In short, a small p-value says that if the null hypothesis is true, the
experimental results that you have aren't very likely -- but it does NOT
say that the alternative hypothesis is vastly more likely to be correct,
or that the data you have are vastly more likely to have come from the
alternative hypothesis case.
Some values of alpha(p) for those too lazy to calculate just now:
p = 0.0005 alpha = 0.010225
p = 0.001 alpha = 0.018431
p = 0.005 alpha = 0.067174
p = 0.010 alpha = 0.111254
p = 0.015 alpha = 0.146204
p = 0.020 alpha = 0.175380
p = 0.030 alpha = 0.222367
p = 0.040 alpha = 0.259255
p = 0.050 alpha = 0.289350
You can also try this AFNI package command to plot alpha(p) vs. p:
1deval -dx 0.001 -xzero 0.001 -num 99 -expr '1/(1-1/(exp(1)*p*log(p)))' |
1dplot -stdin -dx 0.001 -xzero 0.001 -xlabel 'p' -ylabel '\alpha(p)'
Another example: to reduce the likelihood of the null hypothesis being the
source of your t-statistic to 10%, you have to have p = 0.008593 -- a value
more stringent than usually seen in scientific publications. To get the null
hypothesis likelihood below 5%, you have to get p below 0.003408.
Finally, none of the discussion above is limited to the case of p-values that
come from 2-sided t-tests. The function alpha(p) applies (approximately) to
many other situations. However, it does NOT apply to 1-sided tests (which
are not testing 'Precise Null Hypotheses'). See the paper by Sellke et al.
for a lengthier and more precise discussion. Another paper to peruse is
Revised standards for statistical evidence.
VE Johnson. PNAS v110:19313-19317, 2013.
http://www.pnas.org/content/110/48/19313.long
For the case of 1-sided t-tests, the issue is more complex; the paper below
may be of interest:
Default Bayes Factors for Nonnested Hypthesis Testing.
JO Berger and J Mortera. J Am Stat Assoc v:94:542-554, 1999.
http://www.jstor.org/stable/2670175 [PDF]
https://www2.stat.duke.edu/~berger/papers/mortera.ps [PS preprint]
What I have tried to do herein is outline the p-value interpretation issue
using (mostly) non-technical words.
((***** What does this all mean for FMRI? I'm still thinking about it. *****))
TESTING THIS PROGRAM¶
A simple 2-sample test of this program is given by the script below,
which creates 2 datasets with standard deviation (sigma) of 1; the
first one (ZZ_1) has mean 1 and the second one (ZZ_0) has mean 0;
then the program tests these datasets to see if their means are different,
and finally prints out the average value of the estimated differences
in their means, and the average value of the associated t-statistic:
3dUndump -dimen 128 128 32 -prefix ZZ
3dcalc -a ZZ+orig -b '1D: 14@0' -expr 'gran(1,1)' -prefix ZZ_1.nii -datum float
3dcalc -a ZZ+orig -b '1D: 10@0' -expr 'gran(0,1)' -prefix ZZ_0.nii -datum float
3dttest++ -setA ZZ_1.nii -setB ZZ_0.nii -prefix ZZtest.nii -no1sam
echo '=== mean of mean estimates follows, should be about 1 ==='
3dBrickStat -mean ZZtest.nii'[0]'
echo '=== mean of t-statistics follows, should be about 2.50149 ==='
3dBrickStat -mean ZZtest.nii'[1]'
\rm ZZ*
The expected value of the t-statistic with 14 samples in setA and
10 samples in setB is calculated below:
delta_mean / sigma / sqrt( 1/NA + 1/NB ) / (1 - 3/(4*NA+4*NB-9) )
= 1 / 1 / sqrt( 1/14 + 1/10 ) / (1 - 3/87 ) = 2.50149
where division by (1-3/(4*NA+4*NB-9)) is the correction factor
for the skewness of the non-central t-distribution --
see http://en.wikipedia.org/wiki/Noncentral_t-distribution .
CONVENIENCE NOTE¶
When constructing 3dttest++ commands, consider using gen_group_command.py
to simplify the process!
VARIOUS LINKS OF INTEREST¶
* http://en.wikipedia.org/wiki/T_test
* https://open.umn.edu/opentextbooks/textbooks/459
* http://en.wikipedia.org/wiki/Mutatis_mutandis
AUTHOR -- RW Cox -- don't whine TO me; wine WITH me (e.g., a nice Pinot Noir)
++ Compile date = Dec 17 2024 {AFNI_24.3.10:linux_ubuntu_24_64}
