Metan excludes study due to SD equals zero

[YT Wang]

Hi all,

I am trying to include a study where sample size was quite small, and the outcomes have a mean of 6 and SD of 0 in the intervention group, with a mean of 5 and SD of 1.7 in the control group.
Metan excludes this study - I imagine the zero is the issue.
Can I get around this problem in anyway? If not, how can I explain this in my manuscript?

Thanks

[Joseph Coveney]

I imagine the zero is the issue.

One way to find out: make the standard deviation nonzero in the experimental treatment group in that study. See what happens.

I'm not familiar with metan (I used a similarly named user-written command years ago, but it took binomial data), but if you feed your metan summary statistics, just make the standard deviation for the experimental group for that study some small value, say, 0.1.

Can I get around this problem in anyway? If not, how can I explain this in my manuscript?

Same way. You can do a sensitivity exploration by varying the value of the standard deviation and seeing whether that substantially affects your conclusion as well as by determining whether making the standard deviation nonzero provides for a more robust and representative meta-analysis than simply excluding the study.

You can explain to your audience (i) what you did and why, (ii) how (via the sensitivity exercise) you assured that it did not affect your conclusions (presumably), and (iii) in any case, it would be a worst-case analysis by attenuating the difference between experimental treatment group and control treatment group in that study.

[Leonardo Guizzetti]

Joseph has given some excellent pragmatic advice already. I will add that if the one treatment arm really has an SD=0, then it is very likely the arm size is quite small (maybe only 3 people). In which case you should see no practical impact upon excluding the study (which as Joseph pointed out, should be handled in a sensitivity analysis). A second approach could be to derive a pooled estimate of standard deviation using a Satterthwaite approach or similar, and use this pooled SD for each arm.