Effects from TSSEM larger than in original studies

1 reply [Last post]
jbauer's picture
Offline
Joined: 04/14/2014

Hi Mike,

many thanks for the fantastic metaSEM package! I used it to combine models from two original studies.

(1) What astonishes me is that some of the structural parameters estimates from TSSEM Stage 2 are even larger than the ones from the individual studies (e.g. study 1: beta = -.40; study two: beta = -.53; TSSEM2: beta = -.66). Naively I had expected that the comnbined TSSEM2 etimates should be somewhere between the ones of the original studies. I am wondering how this is possible. Can I trust this finding?

(2) Since my model contains two mediation paths: Is it possible to get LB confidence intervals for indirect effects?

Many thanks
Johannes

AttachmentSize
output.txt37.29 KB
Mike Cheung's picture
Offline
Joined: 10/08/2009
Hi Johannes, Thanks. (1) I

Hi Johannes,

Thanks.

(1) I don't have a definite answer here. Were the beta = -.40 and beta = -.53 obtained from the over-identified models, e.g., CFA and SEM? If yes, the average of them can be different from TSSEM. The values beta = -.40 and beta = -.53 are based on the assumption that the fitted models are correct. If they are not correct, the average between -.40 and -.53 is missleadig. It is usually difficult to defend that the proposed model fits well in all studies.

The TSSEM takes other assumptions. If you used the fixed-effects TSSEM, the first stage assumes that the correlation matrices are homogeneous, whereas the stage 2 analysis assumes that the proposed model is correct. Personally, I find it more reasonable to use the TSSEM approach.

(2) Yes, please see the examples in Becker92 and Hunter83.

Cheers,
Mike