Meta-Analysis Resources

Tools for Those Who Summarize the Evidence Base

Resources and networking for those who conduct or interpret meta-analyses related to any phenomenon that is gauged in multiple studies.

Hello Mike and others;

I am trying to conduct random-effects TSSEM on a dataset of 61 studies (data set, "hamed.dat", is attached), using metaSEM 0.9.6, Openmx 2.3.1, and R 3.2.2 (2015-08-14). However, I am stuck in stage 1 of the random-effects analysis with error code "6". I even changed the default optimizer in Openmx from "SLSQP" to "NPSOL", but no luck. Regardless of changes I make to my data, it seems that I can not produce error code "0" or "1" in stage 1 of random-effects TSSEM. This is while the eigen values for the pooled correlation matrix are all positive and the matrix is actually positive definite.

I have attached the RScript that I am using for this analysis.

Any help will be greatly appreciated.

Thank you,

/Hamed

Views: 142

Replies to This Discussion

Hi Hamed,

There is only 1 study in many of these cells. It is not possible to estimate the heterogeneity variances.

Regards,
Mike

pattern.na(MASEM$Data, show.na = FALSE)
   x1 x2 x3 x4 x5 x6 x7 x8 x9
x1  4  2  1  1  1  4  4  1  1
x2  2 29 10 14 15  7 15  1  3
x3  1 10 14  5  5  5  7  1  1
x4  1 14  5 27 12  3 16  5  3
x5  1 15  5 12 29  8 23  4  2
x6  4  7  5  3  8 14  9  4  2
x7  4 15  7 16 23  9 33  6  6
x8  1  1  1  5  4  4  6 19 13
x9  1  3  1  3  2  2  6 13 18

Thank you very much, Mike.

So, if I use the resulted pooled correlation matrix in stage 2 of TSSEM, does it mean that for the data points with one study, the correlation will be used as a covariance toward estimating the structural model?

If there is only 1 study, that correlation will be used in fitting the structural model. It is then difficult to call it a "meta-analysis" as there is only 1 study.

Mike

Thank you, Mike! I see the limitation.

Dear Mike and other users,

I am encountering the same problem. I have a smaller sample of 15 studies with a 5-variable matrix (1 binary variable, 4 continuous variables), but all 15 studies contribute data to every cell. Every study matrix is positive definite--is.pd(vector) gives TRUE for every study matrix. I get OpenMx status=6 when I run tssem1(vector, n, method="REM", RE.type="Diag"). When I rerun tssem1, I still get OpenMx status=6. 

1.  Within each study matrix, the correlations are sometimes based on different n due to missing data for certain variables within each study. Is this causing the problem?

2. What should I do to improve the optimization? I have run tssem1 on smaller datasets with 5 studies and similar 5-variable matrix and the first run of tssem1 gives OpenMx status=5 or 6, but after rerunning the OpenMx = 0. 

3. For the 15-study dataset with tssem1 OpenMx status=6, if I run tssem2, OpenMx status=0. Can I interpret this result, and what caveats should I note given that the tssem1 had optimization problems? 

Thank you in advance for your help!
Mei Yi

Dear Mei Yi,

Could you email me the data and R code so that I can check it?

Best,
Mike

RSS

© 2017   Created by Blair T. Johnson.   Powered by

Badges  |  Report an Issue  |  Terms of Service