Meta-analysis using mean difference and standard error

[Kyle Le]

Hello STATA forum members,

I am trying to conduct a meta-analysis incorporating two different continuous outcome measures. In this case, I usually do metan N1 Mean 1 SD1 N2 Mean2 SD2 in order to get SMD.
However, the problem I am facing here is that higher score is better in one outcome measure and lower score is better in the other scale. So if I use the command above, I think the results of meta-analysis will not be accurate because the above command will calculate effect size as Mean 1 - Mean 2.

So instead, I manually calculated mean difference (MD) between the groups and SE for each study (where I have all MD in positive values since the net difference is I am interested in) and hope to run metan MD SE, random. The question I have here is that would the resulting meta-analysis be considered standardized mean difference? or weighted mean difference? Looking at the pooled effect size, it didn't look like SMD.

Is there a better way to approach this problem?

[Tiago Pereira]

Hi, Kyle.

It is unclear whether your outcome is on a bounded scale (e.g., like pain intensity scores which can range between 0 and 10, or quality of life, say, 0-100). So, we can provide you with general suggestions only.

You can coin all data to be on the same direction (e.g., lower scores mean a better outcome). There are several ways to approach this. A simple approach is to convert all data to SMD (or MD) and associated standard error. Then, you multiply by -1 those SMDs (or MDs) that need to be inverted. For MDs, you need to be sure all outcomes are on the same scale.

[Kyle Le]

Thanks for your feedback. Yeah so I will be more specific - one scale is 0-100 with higher score indicating better outcome and then another scale is 0-7 with lower score indicating better outcome.
Given the difference in the outcome measure, I would definitely use SMD, and what I am getting at from your advice is that manually calculate SMD and associated SEs (multiply by -1 for inversion) then run metan SMD SE?
Would you be able to elaborate more on -1 to the SMD that needs to be inverted and are you aware of reference to cite this for my paper?

Thanks again for your help!

[Tiago Pereira]

You can rescale the 0-7 instrument, where lower scores indicate a better outcome, to a 0-100 scale with higher scores indicating a better outcome. If you provide us with some data, we can give more details. The data can be simulated and/or mimic what you have in your original dataset.

You can cite these references:

https://onlinelibrary.wiley.com/doi/10.1002/jrsm.46
https://www.bmj.com/content/bmj/364/bmj.k4817.full.pdf