University of Michigan Center for Statistical 


Record of Changes to METAL

March 2011

Added LIST option to genomic control option.
Fixed bug in calculation of I^2 statistic.
Fixed bug that led to crashes when strand and heterogeneity analysis were used together.

July 2010

Fix to output of heterogeneity statistics.

June 2010

Added I^2 heterogeneity statistic to output.

April 2010

Bug fix in counter of invalid effect sizes.

Fix to improve handling of "NaN" an"Inf" (non-a-number and infinity 
entries) in the effect size column.

February 2010

Minor fixes to ADDFILTER syntax, so as to avoid unnecessary warnings and error messages.

October 2009

Added the COLUMNCOUNTING option. When this is set to STRICT (the default is LENIENT) Metal 
will require that every line in an input file should have the exact same number of columns.
This stricter checking can help catch defective files.

When the new LOGPVALUE option is set to ON, Metal now reports log(p-value) instead of the
raw p-value for meta-analysis and heterogeneity test results. This makes it easier to 
process, browse and read results with very small p-values.

Added support for custom variables. These are user specified columns whose contents are 
tracked across multiple files. New variables are created with the CUSTOMVARIABLE command. 
For example, the command:


Creates a custom variable called MENDEL_ERRORS. This variable must next be linked to 
specific columns in each input file with the LABEL command. For example:

    LABEL MENDEL_ERRORS AS bad_transmissions
    PROCESS input_file1.txt

    LABEL MENDEL_ERRORS AS non_mendelian_events
    PROCESS input_file2.txt

In the example above, the column "bad_transmissions" in "input_file1.txt" and 
"non_mendelian_events" in "input_file2.txt" would be linked to the MENDEL_ERRORS
column. Metal would report the sum of these two columns in each row of output.

Minor fixes to heterogeneity analysis to improve handling of p-values very close to zero and
of odds ratios recorded in their natural (non-log) scale.

February 2009

Increased precision of p-value calculations so that p-values very close to zero can now be
better distinguished. 

Added ability to track additional user specified variables. This can be useful, for example, 
to enable calculation of the total sample size in a standard error based meta-analysis.
Custom variables are created with the CUSTOMVARIABLE command and their corresponding 
column labels can be customized with the LABEL command.

By default, METAL now requires that exactly the same number of tokens should be present
in every row of each file. This option makes it harder for column misalignments to occur.
The original behavior, which allowed for extra information to be appended to the end of 
each line, can be requested with the COLUMNCOUNTING LENIENT command.

Tweaks to screen output to facilitate readability and debugging.

January 2009

When genomic control modeling is enabled, previous versions of METAL could
discard all or part of the information in previously processed datasets
when an empty result set was processed. METAL now handles empty result sets
more gracefully.

September 2008

This version of Metal can carry out simple heterogeneity analyses. To do this,
issue the command "ANALYZE HETEROGENEITY". Metal will then carry out a standard
meta-analysis (using either effect sizes and standard errors or sample size, 
direction of effect and p-value) and, once this is completed, evaluate evidence
for heterogeneity in effect sizes.

August 18, 2008

This version includes enhancements to GENOMICCONTROL and FILTER commands and 
a new VERBOSE option. Used together, the new VERBOSE option and improvements
to the FILTER command can help retrieve details about interesting results in
a meta-analysis.

* The GENOMICCONTROL option now allows for a specific genomic control parameter
as input. In this case, instead of automatically calculating the GENOMICCONTROL
value based on available data, METAL will apply the correction according to a
user-specified parameter. This option is convenient when full genome scan results
are not available.



The original GENOMICCONTROL ON and GENOMICCONTROL OFF commands are still supported.

* The previous version of METAL allowed filters based on simple comparison operators,
for example:

    ADDFILTER MARKER = rs12345

This new version allows filters that select a set of markers. For example:

    ADDFILTER SNP IN (rs12345,rs98765,rs4567)

* This new version of METAL includes a verbose option, enabled with "VERBOSE ON". When
this option is enabled METAL will print details of each analyzed SNP for each input
file. Unless you use an appropriate filter (see above), this option can easily generate
an unmanageable amount of ouput.

Details for each SNP will be output after aligning alleles and strands to the same strand,
but before any genomic control correction is applied. In sample size based analysis, the
sample size, z-score and p-value will be output. In standard error based analyses, effect
sizes and standard errors output instead. In either case, allele frequency information
will also be output if available.

July 1, 2008

This version fixes a bug in the GENOMICCONTROL command. Starting with this version,
a genomic control correction is only applied to studies where the genomic control
parameter lambda > 1.0. In previous versions, if genomic control was enabled, a 
correction was applied even when the lambda parameter was less than 1.0 (i.e. in
studies where statistics appeared slightly deflated in comparison with null 
expectations, Metal inflated these statistics.)

Another change is that Metal now outputs additional digits when the standard error
option is enabled.

This version also introduces two minor feature, for convenience:

* Effect sizes can be log transformed on the fly. This is convenient when 
meta-analyzing odds ratios. In this case, two studies have the same direction
of effect if their log(odds ratio) statistics are both positive or negative,
corresponding to odds ratios that are both >1.0 or <1.0.

This transformation is activated through the EFFECTLABEL command. For example,
EFFECTLABEL log(ODDSRATIO) instructs metal to log-transform the ODDSRATIO column
prior to analysis.

* Script names can now be given as command line arguments. If multiple scripts 
are named, Metal will proceed, left to right, executing each one.

* The SOURCE command allows one script to include another script. This can be 
convenient when you want to try multiple analyses that include similar filesets.
For example, if the scripts PROCESS-FEMALES.txt and PROCESS-MALES.txt loaded
sex specific result sets, you could use the following script to carry out
three meta-analyses:

          # Carry out a males only meta-analysis

          # Carry out a females only meta-analysis

          # Carry out a combined meta-analysis
          OUT META-ALL .TBL

April 14, 2008

This version introduces several new capabilities:

* Meta-analysis weighted by standard error can now be carried out by issuing the 
"SCHEME STDERR" command before analysis and then using the "STDERRLABEL [LABEL]"
command to specify the appropriate heading for the column containing standard
errors in each file. Effect sizes should be in the column specified with the
"EFFECTLABEL" command.

* Automatic genomic control correction of each scan can now be enabled with the

* On the fly filtering of input files is now possible with the "ADDFILTER ..."
and "REMOVEFILTERS" options. For example, you could use a filter such as
"ADDFILTER MAF > .01" to exclude SNPs with a value of .01 or less in the 
MAF column.

* Use of explicit strand information for alignment of results at "G/C" and "A/T"
SNPs is now possible, using the "STRANDLABEL [LABEL]" and "USESTRAND ON" commands.
Previous versions assume that these SNPs have been pre-aligned to the correct 

* Tracking of allele frequencies can now be enabled using a combination of the 
frequencies will be averaged using the same weights used to combine test
statistics or effect sizes.

January 2, 2008

First release of Metal, supporting basic meta-analysis of genome wide association
scans, weighted by sample size.


University of Michigan | School of Public Health | Abecasis Lab