Skip to content

[R-meta] Impact of very high tau-square value.

3 messages · jideofor thomas, Gerta Ruecker, Guido Schwarzer

#
Hi all,
I did a random effect meta-analysis using the metafor package and I obtained a tau-squared ( heterogeneity)? value greater than 1.

I was worried because I am nit sure what to make of it; so I did the same analysis under the Bayesian framework. I also obtained a tau-squared value of about 32.8.??
I find it confusing, for one it is quite high and secondly I hadn't come across any discussion on how high the tau-squared value could get.
My questions are:
How high can the tau-squared value? be?
Secondly, under this situation can I go ahead and interpret my result but with a warning on the amount of heterogeneity present in the study?? Or does the high value of the tau-square make non-sense of my analysis?
Thanks in advance for your response.?
Jide Ogbu
#
Dear Jide,

in principle, tau -squared can take every non-negative value. Note that 
tau is the standard deviation of the random effects. That is, it is 
measured on the same scale as the outcome and has the same dimension. 
That is, if you measure blood pressure, it is on the scale of blood 
pressure, mmHg; if you measure body weight, it is on the scale of body 
weight, i.e. kg. There is no other general rule.

If you look at a dimensionless outcome (e.g., (log) odds ratio or 
standardized mean difference), the same holds: tau is on the log odds 
scale or on the normal scale, respectively.

Thus the meaning of your values depends on what you are measuring.

See the attached paper, page 7.

Best,

Gerta

Am 26.06.2019 um 09:32 schrieb jideofor thomas:

  
    
#
Dear Jide Ogbu,

It is not possible to give any specific advice on your dataset as you do 
not provide sufficient information. So, I will make some general comments.

A plethora of methods to estimate tau-squared is available (Veroniki et 
al., 2016) and they can give different estimates; see argument 'method' 
in metafor.

Furthermore, whether a tau-squared value of 32.8 is large depends on the 
outcome measure (and for some outcomes on the underlying metric). This 
value could be judged to be small if you are interested in a mean 
difference - depending on your metric. However, if your outcome measure 
is a standardized mean difference or a log odds ratio a tau-squared 
value of 32.8 would be very extreme.

Best wishes, Guido

Veroniki AA, Jackson D, Viechtbauer W, Bender R, Bowden J, Knapp G, et 
al. Methods to estimate the between-study variance and its uncertainty 
in meta-analysis. Research Synthesis Methods. 2016;7:55?79.