MBA¶
Welcome to MBA¶
Matrix-Based Analysis Program through Bayesian Multilevel Modeling
#+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Version 1.0.7, Apr 17, 2023
Author: Gang Chen (gangchen@mail.nih.gov)
Website - https://afni.nimh.nih.gov/gangchen_homepage
SSCC/NIMH, National Institutes of Health, Bethesda MD 20892
#+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Usage:¶
MBA performs matrix-based analysis (MBA) as theoretically elaborated in the
manuscript: https://www.biorxiv.org/content/10.1101/459545v1
MBA is conducted with a shell script (as shown in the examples below). The
input data should be formulated in a pure-text table that codes the regions
and variables. The response variable is usually correlation values (with or
without Fisher-transformation) or white-matter properties (e.g., fractional
anisotropy, mean diffusivity, radial diffusivity, axial diffusivity, etc.),
but it can also be any values from a symmetric matrix (e.g., coherence,
mutual information, entropy). In other words, the effects are assumed to be
non-directional or non-causal. Diagonals can be included in the input if
sensible.
Thanks to Zhihao Li for motivating me to start the MBA work, and to
Paul-Christian Bürkner and the Stan/R communities for their strong support.
Citation:¶
If you want to cite the approach for MBA, consider the following¶
Chen, G., Bürkner, P.-C., Taylor, P.A., Li, Z., Yin, L., Glen, D.R., Kinnison, J.,
Cox, R.W., Pessoa, L., 2019. An Integrative Approach to Matrix-Based Analyses in
Neuroimaging. Human Brain Mapping. In press. https://doi.org/10.1101/459545
Read the following carefully!!!
A data table in pure text format is needed as input for an MBA script. The
data table should contain at least 4 columns that specify the information
about subjects, region pairs and the response variable values with the
following fixed header. The header labels are case-sensitive, and their order
does not matter.
Subj ROI1 ROI2 Y Age
S1 Amyg SMA 0.2643 11
S2 BNST MPFC 0.3762 16
0) You are performing Bayesian analysis!!! So, you will directly obtain
the probability of an effect being positive or negative with your data,
instead of witch hunt - hunting the straw man of p-value (weirdness of your
data when pretending that absolutely nothing exists).
1) Avoid using pure numbers to code the labels for categorical variables. The
column order does not matter. . You can specify those column names as you
prefer, but it saves a little bit scripting if you adopt the default naming
for subjects ('Subj'), regions ('ROI1' and 'ROI2') and response variable ('Y').
The column labels ROI1 and ROI2 are meant to indicate the two regions
associated with each response value, and they do not mean any sequence or
directionality.
2) Only provide half of the off-diagonals in the table (no duplicates allowed).
Missing data are fine (e.g., white-matter property deemed nonexistent).
3) Simple analysis can be done in a few minutes, but computational cost can be
very high (e.g., weeks or even months) when the number of regions or subjects
is large or when a few explanatory variables are involved. Be patient: there
is hope in the near future that further parallelization can be implemented.
4) Add more columns if explanatory variables are considered in the model. Currently
only between-subjects variables (e.g., sex, patients vs. controls, age) are
allowed. Each label in a between-subjects factor (categorical variable)
should be coded with at least 1 character (labeling with pure numbers is fine
but not recommended). If preferred, you can quantitatively code the levels of a
factor yourself by creating k-1 columns for a factor with k levels. However, be
careful with your coding strategy because it would impact how to interpret the
results. Here is a good reference about factor coding strategies:
https://stats.idre.ucla.edu/r/library/r-library-contrast-coding-systems-for-categorical-variables/
5) It is strongly suggested that a quantitative explanatory variable be
standardized with option -stdz; that is, remove the mean and scale by
the standard deviation. This will improve the chance of convergence
with each Markov chain. If a between-subjects factor (e.g., sex) is
involved, it may be better to standardize a quantitative variable
within each group in terms of interpretability if the mean value differs
substantially. However, do not standardize a between-subjects factor if
you quantitatively code it. And do not standardize the response variable
if the intercept is of interest!
6) With within-subject variables, try to formulate the data as a contrast
between two factor levels or as a linear combination of multiple levels.
7) The results from MBA are effect estimates for each region pair and at each
region. They can be slightly different across different runs or different
computers and R package versions due to the nature of randomness involved
in Monte Carlo simulations.
8) The range in matrix plot may vary across different effects within an analysis.
It is possible to force the same range for all plots through fine-tuning
within R using the output of .RData. The criteria of color coding for the
strength of evidence in matrix plots in the output is as follows:
Green - two-tailed 95% compatible/uncertainty interval (or probability of effect
being positive >= 0.975 or <= 0.025)
Yellow - one-tailed 95% compatible/uncertainty interval (or probability of effect
being positive >= 0.95 or <= 0.05)
Gray - one-tailed 90% compatible/uncertainty interval (or probability of effect
being positive >= 0.90 or <= 0.10)
white - anything else
Installation requirements:¶
In addition to R installation, the R package "brms" is required for MBA. Make
sure you have the most recent version of R. To install "brms", run the following
command at the terminal:
rPkgsInstall -pkgs "brms" -site http://cran.us.r-project.org"
Alternatively you may install them in R:
install.packages("brms")
*** To take full advantage of parallelization, install both 'cmdstan' and
'cmdstanr' and use the option -WCP in MBA. However, extra stpes are required:
both 'cmdstan' and 'cmdstanr' have to be installed. To install 'cmdstanir',
execute the following command in R:
install.packages('cmdstanr', repos = c('https://mc-stan.org/r-packages/', getOption('repos')))
Then install 'cmdstan' using the following command in R:
cmdstanr::install_cmdstan(cores = 2)
# Follow the instruction here for the installation of 'cmdstan':
# https://mc-stan.org/cmdstanr/articles/cmdstanr.html
# If 'cmdstan' is installed in a directory other than home, use option -StanPath
# to specify the path (e.g., -StanPath '~/my/stan/path').
Running:¶
Once the MBA command script is constructed, it can be run by copying and
pasting to the terminal. Alternatively (and probably better) you save the
script as a text file, for example, called myMBA.txt, and execute it with the
following (assuming on tcsh shell),
nohup tcsh -x myMBA.txt > diary.txt &
nohup tcsh -x myMBA.txt |& tee diary.txt &
The advantage of the commands above is that the progression is saved into
the text file diary.txt and, if anything goes awry, can be examined later.
The 'nohup' command allows the analysis running in the background even if
the terminal is killed.
Examples:¶
Example 1 --- Simplest scenario. Values from region pairs are the input from
each subject. No explanatory variables are considered. Research
interest is about the population effect at each region pair plus
the relative strength of each region.
MBA -prefix output -r2z -dataTable myData.txt \
The above script is equivalent to
MBA -prefix myWonderfulResult -chains 4 -iterations 1000 -model 1 -EOI 'Intercept' \
-r2z -dataTable myData.txt \
The 2nd version is recommended because of its explicit specifications.
If a computer is equipped with as many CPUs as a factor 4 (e.g., 8, 16, 24,
...), a speedup feature can be adopted through within-chain parallelization
with the option -WCP. For example, the script assumes a computer with 24 CPUs
(6 CPUs per chain):
MBA -prefix myWonderfulResult -chains 4 -WCP 6 \
-iterations 1000 -model 1 -EOI 'Intercept' -r2z -dataTable myData.txt \
The input file 'myData.txt' is a data table in pure text format as below:
Subj ROI1 ROI2 Y
S01 lFFA lAmygdala 0.162
S02 lFFA lAmygdala -0.598
S03 lFFA lAmygdala 0.249
S04 lFFA lAmygdala 0.568
If the data is skewed or has outliers, consider using the Student t-distribution
through the option -distY:
MBA -prefix myWonderfulResult -chains 4 -WCP 6 \
-iterations 1000 -model 1 -EOI 'Intercept' -distY 'student' -dataTable myData.txt \
If t-statistic (or standard error) values corresponding to the response variable
Y are available, add the t-statistic (or standard error) values as a column in the input
data table so that they can be incorporated into the BML model using the option -tstat
or -se with the following script (assuming the tstat column is named as 'tvalue),
MBA -prefix myWonderfulResult -chains 4 -WCP 6 \
-iterations 1000 -model 1 -EOI 'Intercept' -tstat tvalue -dataTable myData.txt \
or (assuming the se column is named as 'SE'),
MBA -prefix myWonderfulResult -chains 4 -WCP 6 \
-iterations 1000 -model 1 -EOI 'Intercept' -se SE -dataTable myData.txt \
Example 2 — 2 between-subjects factors (sex and group):¶
MBA -prefix output -Subj subject -ROI1 region1 -ROI2 region2 -Y zscore\
-chains 4 -iterations 1000 -model '1+sex+group' \
-cVars 'sex,group' -r2z -EOI 'Intercept,sex,group' \
-dataTable myData.txt
If a computer is equipped with as many CPUs as a factor 4 (e.g., 8, 16, 24,
...), a speedup feature can be adopted through within-chain parallelization
with the option -WCP. For example, For example, consider
adding '-WCP 6' on a computer with 24 CPUs.
The input file 'myData.txt' is formatted as below:
subject region1 region2 zscore sex group
S1 DMNLAG DMNLHC 0.274 F patient
S1 DMNLAG DMNPCC 0.443 F patient
S2 DMNLAG DMNRAG 0.455 M control
S2 DMNLAG DMNRHC 0.265 M control
Notice that the interaction between 'sex' and 'group' is not modeled in this case.
Example 3 --- one between-subjects factor (sex), one within-subject factor (two
conditions), and one quantitative variable:¶
MBA -prefix result -chains 4 -iterations 1000 -model '1+sex+age+SA' \
-qVars 'sex,age,SA' -r2z -EOI 'Intercept,sex,age,SA' \
-dataTable myData.txt
If a computer is equipped with as many CPUs as a factor 4 (e.g., 8, 16, 24,
...), a speedup feature can be adopted through within-chain parallelization
with the option -WCP. For example, For example, consider adding
'-WCP 6' on a computer with 24 CPUs.
The input file 'myData.txt' is formatted as below:
Subj ROI1 ROI2 Y sex age SA
S1 DMNLAG DMNLHC 0.274 1 1.73 1.73
S1 DMNLAG DMNPCC 0.443 1 1.73 1.73
S2 DMNLAG DMNRAG 0.455 -1 -0.52 -0.52
S2 DMNLAG DMNRHC 0.265 -1 -0.52 -0.52
Notice
1) the 'Y' column is the contrast between the two conditions.
2) since we want to model the interaction between 'sex' and 'age', 'sex' is
coded through deviation coding.
3) 'age' has already been standardized within each sex due to large age
difference between the two sexes.
4) the 'SA' column codes for the interaction between 'sex' and 'age', which
is the product of the two respective columns.
Options:¶
Options in alphabetical order:
-chains N: Specify the number of Markov chains. Make sure there are enough
processors available on the computer. Most of the time 4 cores are good
enough. However, a larger number of chains (e.g., 8, 12) may help achieve
higher accuracy for posterior distribution. Choose 1 for a single-processor
computer, which is only practical only for simple models.
-cVars variable_list: Identify categorical (qualitive) variables (or
factors) with this option. The list with more than one variable
has to be separated with comma (,) without any other characters such
as spaces and should be surrounded within (single or double) quotes.
For example, -cVars "sex,site"
-dataTable TABLE: List the data structure in a table of long format (cf. wide
format) in R with a header as the first line.
NOTE:
1) There should have at least four columns in the table. These minimum
four columns can be in any order but with fixed and reserved with labels:
'Subj', 'ROI1', 'ROI2', and 'Y'. The two columns 'ROI1' and 'ROI2' are
meant to code the two regions that are associated with each value under the
column Y, and they do not connotate any indication of directionality other
than you may want to keep track of a consistent order, for example, in the
correlation matrix. More columns can be added in the table for explanatory
variables (e.g., groups, age, site) if applicable. Only subject-level
(or between-subjects) explanatory variables are allowed at the moment. The
columns of 'Subj', 'ROI1' and 'ROI2' code each subject and the two regions
associated with each region pair, and these labels that can be any identifiable
characters including numbers. The column 'Y' can be correlation value,
Fisher-transformed correlation value, or white-matter property between
grey-matter regions (e.g., mean diffusivity, fractional anisotropy, radial
diffusivity and axial diffusivity).
2) Each row is associated with one and only one 'Y' value, which is the
response variable in the table of long format (cf. wide format) as
defined in R. In the case of correlation matrix or white-matter property
matrix, provide only half of the off-diagonals. With n regions, there
should have at least n(n-1)/2 rows per subject, assuming no missing data.
3) It is fine to have variables (or columns) in the table that are
not used in the current analysis.
4) The context of the table can be saved as a separate file, e.g., called
table.txt. In the script specify the data with '-dataTable table.txt'.
This option is useful when: (a) there are many rows in the table so that
the program complains with an 'Arg list too long' error; (b) you want to
try different models with the same dataset.
-dbgArgs: This option will enable R to save the parameters in a file called
.MBA.dbg.AFNI.args in the current directory so that debugging can be
performed.
-distY distr_name: Use this option to specify the distribution for the response
variable. The default is Gaussian when this option is not invoked. When
skewness or outliers occur in the data, consider adopting the Student's
t-distribution or exGaussian by using this option with 'student' or
'exgaussian'.
-EOI variable_list: Identify effects of interest in the output by specifying the
variable names separated with comma (,). For example, -EOI "sex,age".
By default the Intercept is considered to be an effect of interest.
Currently only variables, not their interactions, can be directly
requested for output. However, most interaction effects can be obtained by
either properly coding the variables (see example 3) or post processing.
-fullRes: Use the option to indicate that a full set of results is shown in the
the report. When option is not invoked (default), only those region pairs
whose effect reaches at least 90% quantile are shown.
-help: this help message
-iterations N: Specify the number of iterations per Markov chain. Choose 1000 (default)
for simple models (e.g., one or no explanatory variables). If convergence
problem occurs as indicated by Rhat being great than 1.1, increase the number of
iterations (e.g., 2000) for complex models, which will lengthen the runtime.
Unfortunately there is no way to predict the optimum iterations ahead of time.
-MD: This option indicates that there are missing data in the input. With n
regions, at least n(n-1)/2 values are assumed from each subject in the
input with no missing data (default). When missing data are present,
invoke this option so that the program will handle it properly.
-model FORMULA: This option specifies the effects associated with explanatory
variables. By default (without user input) the model is specified as
1 (Intercept). Currently only between-subjects factors (e.g., sex,
patients vs. controls) and quantitative variables (e.g., age) are
allowed. When no between-subject factors are present, simply put 1
(default) for FORMULA. The expression FORMULA with more than one
variable has to be surrounded within (single or double) quotes (e.g.,
'1+sex', '1+sex+age'. Variable names in the formula should be consistent
with the ones used in the header of data table. A+B represents the
additive effects of A and B, A:B is the interaction between A
and B, and A*B = A+B+A:B. Subject as a variable should not occur in
the model specification here.
-prefix PREFIX: Prefix is used to specify output file names. The main output is
a text with prefix appended with .txt and stores inference information
for effects of interest in a tabulated format depending on selected
options. The prefix will also be used for other output files such as
visualization plots such as matrix plot, and saved R data in binary
mode. The .RData can be used for post hoc processing such as customized
processing and plotting. Remove the .RData file to save disk space once
you deem such a file is no longer useful.
-qContr contrast_list: Identify comparisons of interest between quantitative
variables in the output separated with comma (,). It only allows for
pair-wise comparisons between two quantitative variables. For example,
-qContr "age vs IQ, age vs weight, IQ vs weight", where V1, V2, and V3 are three
quantitative variables and three comparisons, V1 - V2, V1 - V3 and V2 - V3
will be provided in the output. Make sure that such comparisons are
meaningful (e.g., with the same scale and unit. This can be used to
formulate comparisons among factor levels if the user quantitatively
codes the factor levels.
-qVars variable_list: Identify quantitative variables (or covariates) with
this option. The list with more than one variable has to be
separated with comma (,) without any other characters such as
spaces and should be surrounded within (single or double) quotes.
For example, -qVars "Age,IQ"
-r2z: This option performs Fisher transformation on the response variable
(column Y) if it is correlation coefficient. Do not invoke the option
if the transformation has already been applied or the variable is
not correlation coefficient.
-ROI1 var_name: var_name is used to specify the column name that is designated as
as the region variable for the first set of each region pair. The default
(when this option is not invoked) is 'ROI1'.
-ROI2 var_name: var_name is used to specify the column name that is designated as
as the region variable for the second set of each region pair. The default
(when this option is not invoked) is 'ROI2'.
-ROIlist file: List all the regions in a text file with one column in an order
preferred in the the plots. When the option is not invoked, the region
order in the plots may not be in a preferred order.
-se: This option indicates that standard error for the response variable is
available as input, and a column is designated for the standard error
in the data table. If effect estimates and their t-statistics are the
output from preceding analysis, standard errors can be obtained by
dividing the effect estimatrs ('betas') by their t-statistics. The
default assumes that standard error is not part of the input.
-show_allowed_options: list of allowed options
-stdz variable_list: Identify quantitative variables (or covariates) to be
standardized. To obtain meaningful and interpretable results and to
achieve better convergence of Markov chains with reasonable iterations,
it is recommended that all quantitative variables be standardized
except for the response variable and indicator variables that code for
factors. For example, -stdz "Age,IQ". If the mean of a quantitative
variables varies substantially between groups, it may make sense to
standardize the variable within each group before plugging the values
into the data table. Currently MBA does not offer the option to perform
within-group standardization.
-Subj var_name: var_name is used to specify the column name that is designated as
as the measuring unit variable (usually subject). The default (when this
option is not invoked) is 'Subj'.
-tstat var_name: var_name is used to specify the column name that lists
the t-statistic values, if available, for the response variable 'Y'.
In the case where standard errors are available for the effect
estiamtes of 'Y', use the option -se.
-verb VERB: Specify verbose level.
-WCP k: This option will invoke within-chain parallelization to speed up runtime.
To take advantage of this feature, you need the following: 1) at least 8
or more CPUs; 2) install 'cmdstan'; 3) install 'cmdstanr'. The value 'k'
is the number of thread per chain that is requested. For example, with 4
chains on a computer with 24 CPUs, you can set 'k' to 6 so that each
chain will be assigned with 6 threads.
-Y var_name: var_name is used to specify the column name that is designated as
as the response/outcome variable. The default (when this option is not
invoked) is 'Y'.
