Dear all!
I have another question and I was wondering if you could enlighten me.
I want to run an analysis that has 2 independent variables, 3 mediators, and 1 outcome variable.
One of the issues that I am having is that, because of my inclusion criteria, I do not have studies that included correlations between two mediators. This means, that I cannot estimate the pooled correlation matrix hence it is impossible to run the full model.
Is it possible to somehow impute this correlation or use the correlation that was found in another study? I do have sufficient data in all other correlations, it's only this one.
An alternative would be to basically include as part of sample of the study other studies that have these correlations, but then I feel I would need to do the same for all other variables, and the process would never end...
I hope I was able to make myself clear?
Cheers!
Max
Printer-friendly version
Hi Max,
I do not think that we can "impute" something in a meta-analysis when there is no data at all.
It is more like a conceptual issue than a statistical issue here. You may use correlations from other studies or meta-analyses. If you include studies excluded by your inclusion criteria, you may need to defend your choice.
Best,
Mike
Perhaps a realistic option here would be to plug in several different plausible values for the missing correlation and consider it to be a sensitivity test that answers the question "what if that correlation was value x?" You could choose the values of x to bracket values reported in the literature. For the most part, fixing the value of the expected correlation for each of the chosen values should do the trick.
Thank you very much!
Indeed, I understand that the issue here is more conceptual than anything else...could it be considered a meta-analysis if you are not using existing correlation tables? On the other hand, if you are using, let say three variables: leadership, autonomy, and well-being. Should the inclusion criteria be studies that only included these relationships? Why would we expect that these studies have relationships that systematically vary from other studies that only observed these variables in pairs? (e.g., studies that only looked at leadership and autonomy; studies that only looked at leadership and well-being; studies that only looked at autonomy and well-being).
But this posits the challenge of when do you stop...much to ponder. Are you aware of any piece of work that ponders on these issues?
Cheers!
Max
It is part of the inclusion and exclusion criteria that researchers have to justify their choices.
I think that most researchers use the bivariate relationship in the inclusion criteria.