[R-meta] Zero seTE in network MA

[Carla Gomez Creutzberg]

Dear Guido,

Thank you so much for your prompt reply and kind explanation. Indeed I was
a bit confused because the error message I get states that:
1. The comparisons for the treatment with 0 standard error were not being
considered  in the network meta-analysis
2. As a result of removing those comparisons, the study ends up having the
wrong number of comparisons and, consequently,
3. I should consider removing the study from the analysis.
In your example this doesn't happen because when the rows with a NA  (or 0)
standard error are removed, the study where they are ends up with just one
comparison (between treatments B and C) which can still be included in the
analysis. In my case, when the study with the 0 standard error is removed
the links between the remainder comparisons are lost so the study can't be
included in the analysis (eg. we loose comparison AB and end up with AC and
BC). I was thus wondering if there was any way to work around this or
whether 0 variances (& consequently standard errors) were simply not
allowed in a meta-analysis at all?

For the time being I have excluded the study from the analysis and,
coincidentally,I have also encountered a second problem with the checks for
the netmeta version 0.9 - 6. For another of the studies in my analysis we
have 3 arms and the situation depicted in the network below:

[image: Inline image 1]

Each node of the image above is a treatment in the study with its
respective mean and, in parenthesis, approximate variance. The arrows of
the edges indicate the direction of the comparison such that the arrow
originates at the numerator treatment and points to the denominator
treatment for the log response ratio.The numbers in the edges correspond to
the log response ratios for each treatment comparison and, in parenthesis,
the corresponding variance. You can also find a copy of this data (along
with another dummy study) and the code that generates the error in the
attached folder. In this case the error message I initially get suggests
that I activate the details.chkmultiarm for more details. When I do so, I
get  the following error:


Multi-arm studies with inconsistent treatment effects:



 studlab           treat1            treat2         TE     resid

  S14172    exotic.forest  forest.harvested  0.7929675  0.528645

  S14172    exotic.forest indigenous.forest 16.0182753 -0.528645

  S14172 forest.harvested indigenous.forest 16.8112428  0.528645





Multi-arm studies with zero treatment arm variance:



 studlab             treat  var.treat

  S14172     exotic.forest 0.02237507

  S14172  forest.harvested 0.01261770

  S14172 indigenous.forest 0.00000000



Legend:

 resid - residual deviation (observed minus expected)

 TE    - treatment estimate

 var.treat - treatment arm variance

Error: Problems in multi-arm studies!

  - Study with inconsistent treatment estimates: 'S14172'

  - Study with zero treatment arm variance: 'S14172'

  - Please check original data used as input to netmeta().

  - Argument tol.multiarm in netmeta() can be used to relax consistency

    assumption for multi-arm studies (if appropriate).

I was wondering where I could find out more information about what is
causing the treatments to be inconsistent (is it the 0 variance for one of
the treatments) and what are the consequences of relaxing the consistency
assumption in this case.

Again, thank you so much for your kind attention and help!

With best wishes,

Carla

On Wed, Sep 20, 2017 at 10:57 PM, Guido Schwarzer <sc at imbi.uni-freiburg.de>
wrote:

> Dear Carla,
>
> I think this is a misunderstanding.
>
> R function netmeta() works perfectly with zero standard errors (by
> excluding pairwise comparisons with zero or missing standard errors from
> the network meta-analysis). This is the same behaviour as in "basic"
> meta-analysis of pairwise comparisons (see metagen() in R package *meta*).
> I attached a corresponding fictitious example for netmeta() to this email.
>
> The new checks in *netmeta*, version 0.9-6, you are referring to, related
> for *treatment arm variances in multi-arm studies* which are calculated
> internally. Sometimes, users enter standard errors for pairwise comparisons
> in multi-arm studies that simply do not "add up". Accordingly, treatment
> arm variances can be negative.
>
> Maybe somebody else can also comment on your idea to use a (very) small
> standard error for a pairwise comparison with zero standard error. In my
> view, the problem with such an approach is that this study gets a very
> large weight in the (network) meta-analysis.
> Best wishes,
> Guido
>
> --
> Dr. Guido Schwarzer (sc at imbi.uni-freiburg.de)
> Institute for Medical Biometry and Statistics
> Stefan-Meier-Str. 26, D-79104 Freiburg | Phone: +49 (0)761 203 6668 <+49%20761%202036668>http://www.imbi.uni-freiburg.de        | Fax:   +49 (0)761 203 6680 <+49%20761%202036680>
>
>


-- 
*Carla Gómez Creutzberg*
PhD. Candidate - Tylianakis Lab
<http://www.tylianakislab.org/the-group.html>
University of Canterbury - *Te Whare Wānanga o Waitaha*
Christchurch, New Zealand <http://www.tylianakislab.org/the-group.html>
cgomezcre at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://stat.ethz.ch/pipermail/r-sig-meta-analysis/attachments/20170922/6d5db44c/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Error14172b.jpeg
Type: image/jpeg
Size: 98671 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-meta-analysis/attachments/20170922/6d5db44c/attachment-0001.jpeg>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: inconsistency-ex.zip
Type: application/zip
Size: 1056 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-meta-analysis/attachments/20170922/6d5db44c/attachment-0001.zip>