Tools for Those Who Summarize the Evidence Base
There are a few different path analysis models that I want to test, each model may use some of the same variables and some different variables. For example, model 1 includes mediator A and outcome at pre and post-treatment, model 2 includes mediator B and outcome at pre and post-treatment, model 3 includes mediators A and B and outcome at pre, mid, and post-treatment. When I extracted data from each study, I extracted correlations and sample sizes for every measure at every timepoint (against every other measures at every other timepoint) to create a large correlation matrix for each study, but most studies have some measures or timepoints missing.
When doing the two-stage TSSEM, which approach should I take?
1. Pool the large correlation matrix in stage 1, and conduct separate stage 2 analysis for each model that I want to test using the same large pooled correlation matrix.
2. Trim the large correlation matrix for each study into multiple smaller correlation matrices that each includes only the measures/timepoints that I want to test for each specific model, then create multiple pooled correlation matrices, one for each model, in stage 1. Then run each stage 2 analysis on a different pooled correlation matrix specifically created for that model.
It seems to me that the first approach might be less work, but may cause more errors because there are more likely to be missing values or very different sample sizes in each cell of the large pooled correlation matrix. But I don't understand what difference there may be statistically, and would be grateful if you could enlighten me.
Thank you and best wishes,
Hi Mei Yi,
Both approaches sound reasonable.
It is slightly more tricky to apply the first approach because you have to select the correspondent elements in the pooled correlation matrix and in the asymptotic sampling covariance matrix. It is easier to apply the second approach.
Thank you Mike, I did use the second approach and it seems to work well.