The first version found should be 1.1.4 or later. If instead the new version is lower down, you need to remove the older version(s). See help adoupdate and help ado - in particular, try ado uninstall mvmeta.

2. You may not have the latest version of mvmeta. Follow the instructions above (but clicking on mvmeta instead of network in step 3).

I get an error message that my mvmeta needs updating, but I've updated it

If you are running a version of network dated between 3jun2014 (v0.6) and 12mar2015 (v1.0) then you may see an error message like

network requires mvmeta version 2.10 or later

This is a bug that was corrected in network version 1.1, so the solution is to update network.

I'm getting an unexpected error message

Please send me the log file errorlog.txt created by running the following code:

log using errorlog.txt

network which

set trace on

set tracedepth 2

<your command>

log close

How should I cite the package?

The preferred citation is

White IR (2015) Network meta-analysis. Stata Journal 15: 1-34.

Other possible citations are

White, I. R. (2011). Multivariate random-effects meta-regression: Updates to mvmeta. Stata Journal, 11, 255-270.
White, I. R., Barrett, J. K., Jackson, D., & Higgins, J. P. T. (2012). Consistency and inconsistency in network meta-analysis: model estimation using multivariate meta-regression. Research Synthesis Methods, 3, 111-125.

Using the network suite

How do I compare coefficients using lincom?

At present you have to do this by writing commands like

lincom [_y_C]_cons - [_y_B]_cons

I have drafted a command -networkcompare- to automate this.

How does the network suite select the reference treatment?

If treatment is numeric, it uses the numerically first treatment; otherwise it uses the alphabetically first treatment. You can change this using the ref() option on network setup.

Is the parameterisation of the inconsistency model arbitrary?

Yes, it's an arbitrary parameterisation. However, different parameterisations give the same overall model: in particular the test statistic for inconsistency is the same.

I think of this as being like regression with a categorical variable: depending on which level you take as the reference level, you will get different parameter estimates, but the model is the same and the overall test for differences between levels is the same.

Is the network model a fixed or random effects model?

The term "fixed effects" is very confusing.

In meta-analysis it has come to mean "no heterogeneity between studies", although for this meaning it should really be "fixed effect" or better still "common effect" (Higgins et al. A re-evaluation of random-effects meta-analysis. JRSSA 2009;172:137-159).
In the rest of statistics it means that a set of parameters are to be estimated entirely separately, rather than being assumed to come from a particular distribution. We might better call this "fixed parameters" or "separate parameters" as opposed to "random parameters".

So my network package:

allows heterogeneity between studies by default in all network meta-analysis (though the fixed option fits a homogeneity or "fixed-effect" model);
when it allows for inconsistency, it does so using fixed parameters not random parameters.

Is the confidence interval of the network estimate always at least as narrow as that of the direct estimate or the indirect estimate?

Yes under a fixed-effect model.

No under a random-effects model, since heterogeneity estimation across the network can have strange results.

Do I need the commonparm option to perform meta-regression in NMA?

The commonparm option is a technical option of mvmeta; in the context of using mvmeta to do NMA, it is primarily used for analysing NMA data in the standard format.

Meta-regression is easier to do in augmented format. In fact if you look at the network meta help file you will see the option:

regress(varlist) Specify covariates for network meta-regression. Every treatment contrast is allowed to depend on the covariate(s) listed. This option is currently only allowed in augmented format.

For example, regress(gender) allows every treatment contrast to depend - in a different way - on gender. So for example gender might modify the A-B contrast but not the C-D contrast. [Gender is a poor example since we should be talking about study-level covariates - perhaps imagine all studies in our network were single-gender studies.]

It's important to consider whether you want every treatment contrast to depend in a different way on gender. Dias et al propose three different models: (1) Unrelated Treatment-Specific Interactions, (2) Exchangeable and Related Treatment-Specific Interactions, (3) Same Interaction Effect for All Treatments. They favour model (3), which implies that if A is the reference treatment, all contrasts with A are modified by gender, but all other contrasts (B-C, D-E etc.) are NOT modified by gender. I prefer model (1), which handles all treatments symmetrically.

Contents

How do I use the latest version?