This is the manual for the GeMTC user interface for network meta-analysis. It starts with a brief introduction to network meta-analysis in the Bayesian framework, including issues such as model fit and convergence. This is followed by a guide to the GeMTC user interface itself. The GeMTC user interface uses question mark icon to provide help for specific terms, which when clicked will show a brief explanation as well as a link to further information in this manual.
Although this manual contains a brief introduction to the terminology and methodology used in GeMTC, it is not a complete guide to network meta-analysis. For further background, we recommend the following excellent open access publications:
- The Medical Decision Making series on network meta-analysis based on the NICE Decision Support Unit series on evidence synthesis: Table of Contents, MDM 2, MDM 3, MDM 4
- The ISPOR guidelines on network meta-analysis: ISPOR 0, ISPOR 1, ISPOR 2
- Evaluating the quality of evidence from a network meta-analysis: Salanti et al. (2014)
GeMTC is available as a stand-alone package at gemtc.drugis.org, and is also part of the ADDIS decision support system for evidence based medicine, developed at drugis.org. The functionality in the GeMTC user interface is supported by the GeMTC R package. All drugis.org software is open source, and source code is available on github.com/drugis.
Preparing your dataset
After signing in to GeMTC, you will be redirected to your personal home page. It contains a list of your previously created analyses (which will be empty until you create one), and a button to create a new analysis. Clicking this button will open the “New analysis” dialog, where you choose a title for your analysis and outcome, and upload a dataset file:
Datasets can be uploaded in CSV format. The CSV file should contain one row per arm, and needs to contain “study” and “treatment” columns, which can contain names or identifiers for the study and treatment. The data columns may be one of:
- Continuous data: “mean” and “std.err”
- Continuous data: “mean”, “std.dev”, and “sampleSize”
- Dichotomous data: “responders” and “sampleSize”
- Survival data: “responders” and “exposure” (exposure in person-time)
The following is an example CSV file for a Parkinson’s disease dataset:
"study","treatment","mean","std.dev","sampleSize" "1","A",-1.22,3.7,54 "1","C",-1.53,4.28,95 "2","A",-0.7,3.7,172 "2","B",-2.4,3.4,173 "3","A",-0.3,4.4,76 "3","B",-2.6,4.3,71 "3","D",-1.2,4.3,81 "4","C",-0.24,3,128 "4","D",-0.59,3,72 "5","C",-0.73,3,80 "5","D",-0.18,3,46 "6","D",-2.2,2.31,137 "6","E",-2.5,2.18,131 "7","D",-1.8,2.48,154 "7","E",-2.1,2.99,143
Including covariates in the dataset
To include covariates in a dataset, simply add additional columns to the CSV file, but do not to use any of the column names already recognized by GeMTC (study, treatment, mean, etc.). Covariates are currently only considered at the study level. Therefore, you can include these data in one of two different ways:
- Specify the covariate value for only one of the study arms, and leave it as missing or "NA" for the others.
- Specify the covariate value for all arms of the study, making sure that each arm within a study has the same value as the other arms.
Including contrast-based data
Contrast based data (e.g. log odds ratios) can also be analyzed in GeMTC. Datasets can contain both contrast-based and arm-based data, but any single study must be either completely contrast-based or completely arm-based. Contrast based data are effect estimates and their standard errors on a specific scale (e.g. mean difference or log-odds ratio); that scale must be specified when uploading contrast-based data.
Contrast-based data are also specified using the one arm per row format. For every arm except the base arm, specify the effect estimate and its standard error in the columns "re.diff" and "re.diff.se". For the base arm, these columns must be left as missing or "NA".
In multi-arm (that is, more than 2 arms) trials, the treatment contrasts are correlated. This must be accounted for in the likelihood. To do this, the standard error of the (absolute) effect in the base arm must be specified in the column "re.base.se". For example, if the outcome scale is the log-odds ratio, this is the standard error of the absolute log-odds in the base arm. By a convenient statistical coincidence, the standard error of the base arm squared is equal to the covariance between the treatment contrasts. Thus, if the covariance is known, set "re.base.se" to the square root of the covariance.
There are a number of caveats when using contrast-based data. First, as the number of participants is not known for this data, the size of nodes in the evidence graph does not reflect the number of participants. Second, the scale at which the relative effects were computed will restrict the available likelihoods. Third, model fit statistics for contrast-based data are not computed at the arm level, but at the study level.