Hello,
I am currently conducting a Meta-SEM, with different concepts like satisfaction, trust or loyalty. Among the articles I selected for this meta-analysis, the authors are using different definitions. For example, for the loyalty concept, some authors are using only the behavioral approach, while others are using the attitudinal approach, or even a mixt approach. For my meta-analysis, I then coded the type of definition used (Loy-be, Loy-att and Loy-mix). This way I will be able to run a moderator analysis and see if there are differences amongst the effect sizes, depending on the type of definition used.
First, do you think it is a good approach?
Plus, I have another issue, based on this problem. If the articles are using different definitions under the same term “loyalty”, some are defining these dimensions as strict different concepts. For example, I have one article with two different concepts in the model: attitudinal loyalty and behavioral loyalty. How can I treat these articles? Indeed, there is an issue, because since there is 2 concepts, there will be 2 correlation coefficients for the relationship “satisfaction-loyalty”. But in the correlation matrix, I can only have one correlation coefficient for this relationship.
Then, I was thinking of two options:
- Removing these articles from the analysis
- Choosing only one variable: to analyze either “attitudinal loyalty” or “behavioral loyalty” but not both
- Building two correlation matrixes for the articles with the two types of loyalty: it seems to be a bad idea because there will be a repetition of the other correlation coefficients.
What do you think?
Thanks in advance for your help,
Kathleen
Yes, what you are doing is good.
Yes, build two correlation matrices.
The other options all lose potentially valuable information. Your coding will record that the two matrices come from the same study, so you can check back and remove one later if you want to check for sensitivity.
You are on the right track and asking important questions. I'm wondering whether you plan to conduct seperate (bivariate) meta-analyses first, construct the meta-analytical cor matrix then, and run SEM | or if you will conduct TSSEM (Cheung, 2016) by constructing correlational matrices for each study and then, in two stages, optimizing the model . This makes all the difference in the world for answering your question correctly, but I'm assuming the latter here.
First, you may be interested in the general relationships. In that case, it doesn't matter how they defined the construct and you include all relevant measures. If a single study uses multiple, you can average the correlations. You construct a single correlation matrix for each study.
Second, you may be interested in the relationship for a specific construct definition, such as the behavioral approach to loyalty. In that case, you select only correlation coefficients for your MA that reflect this approach, i.e., only loy-be constructs. You can construct seperate correlation matrices from each study for each distinct construct definition.
Third, you may be interested in whether there are (significant) differences between these approaches to define loyalty. In that case, your suggested approach of coding the type of construct at the study level works. In for example metafor, you should be able to cluster matrices from the same studies using the inner|outer formula for random coefficients (with something like ~Construct|ID, [construct = loyalty definition; ID=study identifier]). However, Im guessing from your description that you will use metaSEM (Cheung, 2016), and I have not seen any approaches to including this type of moderators or clusters in random coefficients there. You might reach out to Mike, he's very responsive in my experience.
A wonderfully comprehensive review, Leonidas.
Thank you for posting.
I shall be sharing this with my colleagues.
Thank you very much for your clear answers!
Robert Suurmond, to answer your question, I was thinking about conducting the TSSEM of Mike Cheung :)
For my MetaSEM, I am interested, at the end, in the general relationships so the avergage can be a good idea, and I'll then code the definition of loyalty as a mixt definition (average between attitudinal and behavioral definitions so a mixt definition)
Thanks again for your precious help !
Best regards,
Kathleen Desveaud
Dear Robert Suurmond
As you suggested I did the average of the correlations coefficients when an author is separating one concept into two different concepts. For example, let's say I am interested in the relationship Satisfaction - Loyalty. In one article I have a model with satisfaction, behavioral loyalty and attitudinal loyalty. Here, the concept of brand loyalty is divided in two different concepts: the attitudinal and the behavioral loyalty.
To code the correlation coefficient of my relationship of interest for my meta-analysis, I compute the average of the correlation of satisfaction with attitudinal loyalty and the correlation of satisfaction with behavioral loyalty. Of course, in my coding, I also coded the type of definition used, which is here the composite approach (behavior + attitude).
This way I have a single coefficient correlation for the relationship Satisfaction-Loyalty.
However, I still have issue to code other relevant metrics. What should I do to code the mean, SD, alpha coefficient, CR and AVE for the concept of loyalty, since I have two different values (one for the behavioral loyalty and another for the attitudinal loyalty) ? Should I also compute the average?
Thanks in advance for your precious help,
Best regards,
Kathleen
Kathleen Desveaud
I don't have a clear answer to that question, but what do you intend to do with that information? If you already have the (average) correlation coefficient and the sample size, you have enough to run the meta-analysis.
If you intend to correct for reliability, I'd assume the average Alpha works.
I'm not sure why you'd need the mean, SD, etc for the actual analysis. Please let me know.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis