Dear all,
I am trying to conduct a metaSEM on four data sets which contain missing values. I have the codes I've used to calculate the estimates and their confidence intervals with the full data sets (missing values imputed using ML), and which worked perfectly.
However, when I am repeating the same analysis with the raw data sets with no imputations, I get error messages at the fixed2 stage, and for calculating the confidence intervals, I get NA for the Upper and Lower bounds.
I was also wondering whether it is possible to get the R2 in these analysis.
I have attached the codes I have been using and the four data sets I am using.
Any help is highly appreciated.
Best regards,
Arin
Dear Arin,
I have a couple of suggestions.
1) Since the raw data are available, it is easier and better to use multiple-group SEM.
2) From the attached R code and data, I can only find the correlation matrices based on the listwise deletion (or the complete case analysis). I do not find any data based on the missing values imputed using ML (multiple imputation?).
3) If the variables are standardized, R2 on y is the same as 1-Error variance on y.
Mike
Dear Mike,
Thank you very much for your prompt reply.
1. We had decided to use Meta-SEM as it would allow us to summarise the results across the four contexts, while taking each of the contexts into consideration. We did not chose multiple-group SEM as we are not interested in comparing the coefficients across the different contexts.
Is there any article showing that multiple-group SEM is better than MetaSEM when we want to summarise coefficients across several contexts?
2. I did not attach the data with imputed data as I already have the results from these data sets using metaSEM. I just need to repeat the same analysis with listwise deletion. I thought using the same code I used for the imputed data sets (Expectation-Maximization) would work out, but the code is giving me errors.
3. The variables are not standardized.
Best,
Arin
Dear Arin,
Mike
Dear Mike,
Thank you very much for your help. I highly appreciate it.
In regards to points 1 and 3, everything is clear now.
As for point 2, the model should be theoretically sound as all of our hypotheses are based on strong social psychological theoreis. However, in the pdf document you've sent, the Likelihood based 95% confidence intervals are all NA. How can we get these confidence intervals for the estimates coefficients for the direct and indirect effects (as I am basing my conclusions on the CI rather than p value)? When I used the code in the full data sets (e.g., missing data imputed), the code did give me the intervals.
Also, is MetaSEM gives any kind of indices of variance for the coefficents (direct and indirect), that would show how much the estimate varies across the four contexts.
Best regards,
Arin
Dear Arin,
You may try rerun() and see if you can get the CIs. It may take some time to compute them.
If you want to get the heterogeneity (variances) of the direct and indirect effects, you may need to apply a different model. Cheung and Cheung (2016) discusses this topic.
Cheung, M. W.-L., & Cheung, S. F. (2016). Random-effects models for meta-analytic structural equation modeling: Review, issues, and illustrations. Research Synthesis Methods, 7(2), 140–155. https://doi.org/10.1002/jrsm.1166
https://dl.dropboxusercontent.com/u/25182759/Random%20effects%20models%20for%20meta%20analytic%20structural%20equation%20modeling%20Review%20issues%20and%20illustrations.pdf
Mike
Thank you for the reference.
I have tried to re-run the code several times. I do get the CI now, but I do not get the SE and p-values, and most of the times I get the OpenMx status 6.
Is there any possibility I can get all the information (e.g., CI, p values and SE)?, as well as the openMx Status 0 or 1?
Below is the revised code where I had to add few indirect effects that I had misspecified in the previous code.
I would highly appreciate your help.
Best regards,
Arin
I deliberately disable SEs and p values when the users call the likelihood-based (LB) confidence interval. The LBCI may not match with the SEs and p values. If you want to get the SEs and p values, you may use intervals.type="z".
The attached R code works fine for me. I recommend you to update OpenMx to the latest version before running your analyses.
Mike