Meta-Analysis Resources

Tools for Those Who Summarize the Evidence Base

Resources and networking for those who conduct or interpret meta-analyses related to any phenomenon that is gauged in multiple studies.

Hello all,

Happy new year. I discovered this forum while searching for meta-analysis resources on google. 

I am currently working on a meta-analysis where one of the outcomes is reduction in pain as measure by different pain scales (mostly 100mm VAS). The challenge we are facing is how does one backtranslate the SMD to something meaningful like raw means and NNT? If we can solve this problem, then the SMD value we have will be clinically meaningful. 

What I am looking for is something that sounds like this:

"Physical activity was statistically significantly more effective than the control at the end of the intervention period (standardized mean difference (SMD) -0.36, 95% confidence interval (CI) -0.62 to -0.10; back translated to mean difference of 14.4 points lower, 95% CI -4.0 to -24.8 on a 100 point scale where a lower score means less fatigue; number needed to treat for an additional beneficial outcome (NNTB) 7, 95% CI 4 to 26) demonstrating a small beneficial effect upon fatigue."

I got this above sentence from a recent Cochrane review:

The cochrane handbook does provide some some help on this topic: ; however what I could not figure out is what the assumed control group risk value is to put in the equation. So, either ways, I am stuck. 

I certainly appreciate any help or guidance. Thanks in advance.

Views: 110

Replies to This Discussion

Happy New Year!

That's an interesting problem. The answer would seem to revolve around knowing what the population standard deviation (SD) is on the critical outcome measure (or in your case, measures plural). The Cochrane review you cite probably did dual meta-analyses, one focused on the SMD and the other focused on the unstandardized meas difference (UMD). If they did not, then they probably needed to make an assumption about the population SD. Doing so might be clear in some cases but much trickier in others. For example, very-well normed scales have a well-established mean and standard deviation, yet even in these cases assumptions are made (e.g., were the prior data collected only from students in universities? How would the scores compare to non-university samples?). One might think of absolute scales as relatively easy to norm, but in fact they can face similar problems. For example, blood pressure (BP) is universally measured using the metric system (a BP of 130/70 is literally millimeters of mercury for systolic and diastolic blood pressure, respectively), which might suggest that BP would be readily comparable. Yet, blood pressure varies from sub-population to sub-population (e.g., race/ethnicity, age) and as means rise or fall, so too do the SDs.

We recently conducted a Monte Carlo simulation that pits analysis of the SMD with that for the UMD. One version of the report is here. (We're working on subsequent articles based on the concept.) We recommend understanding better what makes SDs get bigger or smaller. (And we recommend, generally, using the SMD especially in messy domains.)

Hope that helps!


© 2022   Created by Blair T. Johnson.   Powered by

Badges  |  Report an Issue  |  Terms of Service