Using Pre-Post data in a meta-analysis

#1
Hello,

I am trying to figure out the best way to calculate an effect size for a single-group design when I do not know the pair-wise correlation or standard deviation for the difference scores. I have received mixed messages as to whether or not I can use the formula for independent measures and would appreciate any insights that others may have on the matter.

Thank you for your consideration.
 

hlsmith

Omega Contributor
#2
So did all of your studies of interest have the same study design and report differences? Or do you just have this one study that you are trying to fit into the meta-analysis with the others?
 
#3
It's a combination of the two. Some are RCTs, so post-intervention scores are compared to a control group and others are a single-group design with pre-post tests.
 

hlsmith

Omega Contributor
#4
Sounds pretty ugly. Are there at least a sufficient amount of each type of results? You use mixed models to control for heterogeneity between studies, including design issues. Though, you need to make sure you can get them all into comparable standard units and most importantly that each original study is asking and framing the same question similarly.


So back to your question, one study had a pre/post design and reports differences. do they report a one sample ttest against 0 or a Wilcoxon sign rank test. List what they did and how they reported their numbers. I am guessing it may be difficult to make comparisons between these types of studies overall.
 
#5
I have 6 RCTs and 10 single groups with pre-post scores. The results from most of those studies are derived from paired t-tests and one uses a Mann-Witney U test. All of the studies have the same general question. Although they each have slightly different populations, they use one of the three self-esteem questionnaires to measure the impact of a cognitive-behavioural intervention.

In each of the studies, they list the pre-scores and post-scores, but they don't list the correlations for the pre-post pairs. I have been informed that if I assume a correlation of 0.5, then I can calculate the standard error and generate a standardised mean difference, which can then be compared to the independent groups used in the RCTs. Does that sound right?

Sounds pretty ugly. Are there at least a sufficient amount of each type of results? You use mixed models to control for heterogeneity between studies, including design issues. Though, you need to make sure you can get them all into comparable standard units and most importantly that each original study is asking and framing the same question similarly.


So back to your question, one study had a pre/post design and reports differences. do they report a one sample ttest against 0 or a Wilcoxon sign rank test. List what they did and how they reported their numbers. I am guessing it may be difficult to make comparisons between these types of studies overall.