Skip to content

[R-meta] meta-analysis of 0 events in one or both arms

4 messages · Bighelli, Irene, Gerta Ruecker, Nicky Welton

Original
Bighelli, Irene
# 
Dear all,

I am dealing with meta-analysis and network meta-analysis of dichotomous data where the number of event in one or both arms is = 0. I would like to find a way to include these studies in the analysis, without applying a continuity correction. Is it possible?

I know there are different ways to deal with this,

in metabin: I understand that a continuity correction is applied by default (https://cran.r-project.org/web/packages/meta/meta.pdf page 60-62<https://cran.r-project.org/web/packages/meta/meta.pdf%20page%2060-62>): "For studies with a zero cell count, by default, 0.5 is added to all cell frequencies of these studies; if incr is "TACC" a treatment arm continuity correction is used instead (Sweeting et al., 2004; Diamond et al., 2007). For odds ratio and risk ratio, treatment estimates and standard errors are only calculated for studies with zero or all events in both groups if allstudies is TRUE." Does it mean that with allstudies=TRUE only studies with zero events are counted, and the ones with some events excluded?

in netmeta: I read in the manual of the pairwise function (https://cran.r-project.org/web/packages/netmeta/netmeta.pdf), where the argument "allstudies=TRUE" allows to include in the calculations such studies ("A logical indicating if studies with zero or all events in two treatment arms are to be included in the meta-analysis"). However, I was not able to find a description of how this works, and whether it applies a continuity correction.


Thanks a lot in advance for your help,

Irene

________________
Irene Bighelli  PhD

Technical University of Munich | School of Medicine | Klinikum rechts der Isar

Department of Psychiatry and Psychotherapy

Section for Evidence Based Medicine in Psychiatry

Ismaningerstr. 22, 81675 M?nchen, Germany

Tel: +4908941404243

Mail: irene.bighelli at tum.de<mailto:irene.bighelli at tum.de>
Gerta Ruecker
# 
Dear Irene,

Double-zero studies can be safely ignored for relative measures such as 
RR and OR, because they do not contribute to the likelihood (see, for 
example, https://www.ncbi.nlm.nih.gov/pubmed/32065224 ).

Studies with a zero in only one arm (single-zero studies) are included 
with the Mantel-Haenszel method, without need of a continuity 
correction. In the meta package, to avoid a continuity correction, one 
has to set incr = 0. The Peto method includes one-zero studies always 
without a continuity correction, but likewise ignores double-zero studies.

The sentence you cited from the meta help page means that in order to 
include studies with zero events in both groups you have to set 
allstudies=TRUE. You need a continuity correction. If not, they are 
excluded.

In netmeta, it is similar: If you set allstudies=TRUE, all studies are 
included, but with an increment, by default incr=0.5.

For network meta-analysis, instead of using the netmeta() function, you 
may consider using function netmetabin(), which uses one-stage methods 
like the Mantel-Haenszel method. See example(netmetabin).

Summarizing this,

  * it is no problem to include single-zero studies without needing a
    continuity correction when using the Mantel-Haenszel estimator (or
    other one-stage methods).
  * Double-zero studies can only be included by using an increment
  * But this is not necessary/recommended because they are not
    informative for relative measures.

Best,

Gerta

Am 03.04.2020 um 10:15 schrieb Bighelli, Irene:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Nicky Welton
    

    

      
      #
    

  


  

      
Dear Irene,

If there are 0's on both arms then that study should be excluded from the analysis in my opinion. The study is not large enough to measure the outcome in question, and the results are equally consistent with an OR of 0.5, 1, or 2, etc. 
For the studies with a 0 on one arm. The only way to incorporate these without a continuity correction is to use a Bayesian analysis where exact binomial likelihoods are given. Even then depending on your evidence structure, it may still be necessary to add a continuity correction to obtain convergence. 

Best wishes,

Nicky


?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Nicky J. Welton
Professor of Statistical and Health Economic Modelling
Director of the NICE Technical Support Unit
?
Population Health Sciences, Bristol Medical School
University of Bristol
Canynge Hall, 39 Whatley Road
Bristol. BS8 2PS
UK
?
Tel: +44(0)117 3313918
Email: Nicky.Welton at bristol.ac.uk
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
A top 10 UK university and top 50 world university (QS Rankings 2020)
A top 2 UK university with leading employers (High Fliers 2019)
A top 5 UK university for research (THE analysis of REF 2014)
A top 6 European university for teaching (THE 2018) 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~



-----Original Message-----
From: R-sig-meta-analysis <r-sig-meta-analysis-bounces at r-project.org> On Behalf Of Bighelli, Irene
Sent: 03 April 2020 09:16
To: r-sig-meta-analysis at r-project.org
Subject: [R-meta] meta-analysis of 0 events in one or both arms

Dear all,

I am dealing with meta-analysis and network meta-analysis of dichotomous data where the number of event in one or both arms is = 0. I would like to find a way to include these studies in the analysis, without applying a continuity correction. Is it possible?

I know there are different ways to deal with this,

in metabin: I understand that a continuity correction is applied by default (https://cran.r-project.org/web/packages/meta/meta.pdf page 60-62<https://cran.r-project.org/web/packages/meta/meta.pdf%20page%2060-62>): "For studies with a zero cell count, by default, 0.5 is added to all cell frequencies of these studies; if incr is "TACC" a treatment arm continuity correction is used instead (Sweeting et al., 2004; Diamond et al., 2007). For odds ratio and risk ratio, treatment estimates and standard errors are only calculated for studies with zero or all events in both groups if allstudies is TRUE." Does it mean that with allstudies=TRUE only studies with zero events are counted, and the ones with some events excluded?

in netmeta: I read in the manual of the pairwise function (https://cran.r-project.org/web/packages/netmeta/netmeta.pdf), where the argument "allstudies=TRUE" allows to include in the calculations such studies ("A logical indicating if studies with zero or all events in two treatment arms are to be included in the meta-analysis"). However, I was not able to find a description of how this works, and whether it applies a continuity correction.


Thanks a lot in advance for your help,

Irene

________________
Irene Bighelli  PhD

Technical University of Munich | School of Medicine | Klinikum rechts der Isar

Department of Psychiatry and Psychotherapy

Section for Evidence Based Medicine in Psychiatry

Ismaningerstr. 22, 81675 M?nchen, Germany

Tel: +4908941404243

Mail: irene.bighelli at tum.de<mailto:irene.bighelli at tum.de>

  
  


  


      

    

    

      
      

        


  

        

      Gerta Ruecker
    

    

      
      #
    

  


  

      
Dear Nicky, dear Irene,

As outlined in my first post, it is not true that single-zero studies 
can be included without continuity methods only with Bayesian methods.

The Mantel-Haenszel method is a very simple and also intuitive way to 
include them. If you denote the four entries of a 2x2 table by a,b,c,d, 
the formula (here OR) is

OR.pooled = sum(a*d)/sum(b*c)

If, for example, one of the a's is zero, but not the corresponding c, 
this summand becomes zero in the numerator, but not in the denominator. 
Thus the study is included, and the estimator is well defined, except 
all denominators are zero (no events at all in the second group).

A double-zero is excluded (and does not contribute to the likelihood).

Best,

Gerta

Am 03.04.2020 um 11:35 schrieb Nicky Welton: