Network meta-analysis
Comparing three or more treatments, including ones never tested against each other
Use this module when you have several treatments and not all of them were compared directly. The network uses the direct comparisons to estimate the indirect ones: if A was compared with B, and B with C, then A against C can be estimated.
Follow the three-step flow, but both the input and the output are different here.
Step 1, the data
The format is long, one row per arm rather than per study. There are no .e and .c columns: studies repeat, once for each arm.
study |
treatment |
responders |
sampleSize |
|---|---|---|---|
| Study A | Drug A | 12 | 120 |
| Study A | Drug B | 20 | 118 |
| Study B | Drug A | 15 | 130 |
| Study B | Placebo | 23 | 128 |
There are three submodules:
| Submodule | Required columns |
|---|---|
| Binary | study, treatment, responders, sampleSize |
| Continuous | study, treatment, mean, std.dev, sampleSize |
| Pre-calculated | study, treat1, treat2, TE, lower, upper |
In the pre-calculated submodule a row is a comparison, not an arm, which is why treat1 and treat2 exist.
Drug A, drug A and Drug-A become three different nodes in the network. It is the most common error here, and the network comes out disconnected with nothing pointing at the reason. Standardize the names in the spreadsheet before loading.
Each study needs at least two arms. Studies with three or more are accepted, and the Data check reports how many there are, because they require special statistical handling, which netmeta already does.
Step 2, the parameters
Outcome settings
Outcome name goes into titles and file names.
Summary measure (sm) depends on the submodule: RR, OR and RD in the binary one, MD and SMD in the continuous one, and HR, RR, OR, MD, SMD in the pre-calculated one. Each behaves as in the paired modules.
Reference treatment (optional) sets which treatment the others are compared against in the forest plot. Left blank, the app picks one on its own. This does not change the network, only the point of view of the figure.
For ranking, smaller values are (default Good) tells the app whether smaller values are good or bad. It sets the direction of the P-score ranking. For an outcome such as mortality, where smaller values are desirable, choose Good; for one such as treatment response, where larger values are desirable, choose Bad.
Getting it wrong raises no error and no warning. The P-score comes out computed, neat, and in reverse order. Check it before reporting any ranking.
Statistical settings (advanced)
Analysis model (default Random-effects) and Tau method (default REML) work as in the binary module. In a network the random-effects model is almost always the right choice, because the network gathers studies that are even more diverse than in a pairwise comparison.
There is no Prediction interval, Random CI method or Prediction method in this module.
Step 3, export and results
This is the part that differs most from the other modules. There are eight cards.
Forest plot compares every treatment against the chosen reference.
Network graph draws the network: treatments are the nodes and direct comparisons are the edges, with thickness proportional to the amount of direct evidence. It is worth being the first thing you look at, because it shows at once whether the network is connected and where the evidence is thin.
League table puts every pairwise comparison into a single table, exported to Excel. It is the table that usually goes into supplementary material.
Between-study heterogeneity opens the design-based decomposition and the Q statistics, which separate heterogeneity within designs from inconsistency between them.
Node splitting compares direct and indirect evidence for each comparison where both exist. It is the local test of inconsistency.
P-score ranking orders the treatments. The P-score runs from 0 to 1 and summarizes how likely a treatment is to be better than the others.
Comparison-adjusted funnel plot examines small-study effects across the network comparisons.
Pairwise object downloads the pairwise object used to build the network, useful for checking and reproducibility.