Conducting a metaSEM on two-arm randomized controlled trial data

We are working on a meta-analysis of RCTs, looking at the efficacy of brain stimulation (active vs sham TMS) in reducing OCD severity as the primary outcome, along with secondary outcomes as reduction in anxiety and depression severity. We have 17 studies included in this analysis.

We estimated the effect sizes as SMD/Hedge's g - for each study, using the pre-post differences in each arm. In the univariate meta-analysis, we found that active TMS is somewhat superior to sham with respect to OCD (ES 0.26) & anxiety (ES 0.24) but only marginally for depression (ES 0.18, LCI 0.004).

We were wondering if a mediational analysis with metaSEM could be done, to look at whether there is a direct relationship of TMS on OCD severity, or is it mediated via anxiety/depression, or vice-versa.

Here is the model that we we tried out:

model <- "YBOCS ~ c*Treatment + b*Anxiety
Anxiety ~ a*Treatment
Treatment ~~ 1*Treatment"

Where YBOCS & Anxiety are the effect sizes of each study on the OCD & anxiety outcomes, respectively. We do not have the correlation/covariance between the outcome measures for most of the studies, and so we planned to do a sensitivity analysis using various values for it.

So we made matrices for each study that looks like this:

$data[[10]]
YBOCS Anxiety Treatment
YBOCS 1.0000000 0.6000000 0.2887533
Anxiety 0.6000000 1.0000000 0.1349759
Treatment 0.2887533 0.1349759 1.0000000

(Only one study that had an effect size of >1, and was excluded as its matrix wasn't positive definite.)

As an alternative, we also estimated point biserial correlations for each study, the same matrix for the above study looks like this

$data[[10]]
YBOCS Anxiety Treatment
YBOCS 1.0000000 0.60000000 0.10156272
Anxiety 0.6000000 1.00000000 0.02320059
Treatment 0.1015627 0.02320059 1.00000000

I tried running the tssem with the codes you'd provided in the article, - code and data are attached. However, I wanted to know if this approach of combining SMDs/point biserial correlations with pearson correlation coefficients within an SEM model is correct?

For both the approaches, the model didn't converge (OpenMx status 5), this is possibly due to lesser sample sizes?

If not, is there any alternative to answer our research question?

Thanks in advance, look forward to hearing back from you.

Srinivas