I developed a meta-analysis correlation matrix which contains 15 constructs, N (i.e. harmonic mean) = 193 and SD = 1. I have seven correlations >.80 (i.e. 80-82). I am unable to run my model in Amos as I keep receiving an error message "sample moment is not positive definite". I managed to run the model but after dropping many of my constructs which reduces my model significantly. Searching online indicates this is an expected issue with meta-analysis correlation matrix due to aggregate pairwise correlation. Can anyone suggest a solution to this issue without dropping too many variables?
My suggestion would be to abandon AMOS and use MetaSEM for R. It should be able to deal with matrices that are not positive definite. And, yes, if you aggregate data for each cell, and cells include data from different studies, you will often end up with a matrix that is 'unbalanced'.
https://cran.r-project.org/web/packages/metaSEM/metaSEM.pdf
I Agree with Jan, metaSEM is better able to handle this type of data. A second benefit is that by using metaSEM, you can actually use a proper weight for each cell based on the sample size for each cell, rather than the harmonious mean which under/overrepresents the amount of data you have for any individual cell. In other words, metaSEM allows you to extract both a correlation matric and a variance-covariance matrix from your data. The problem with using AMOS (or other SEM) is that it assumes that the correlation matrix you insert is in fact a covariance matrix.
Check: https://courses.nus.edu.sg/course/psycwlm/internet/ and https://www.amazon.com/dp/1119993431
Thank you very much all. I really appreciate your help and responses. I will try to learn R.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis