key issue: the covariance and correlations of BISR do not align in their meaning, but not always
MPlus website reports that this issue is fairly common. Muthens advise to report one or the other
this doesn't work in our case because we want to compare the effect across studies and compute a meta-analytic index of the average effect
possible solution: report variances with se and pvals, but also provide just the correlations (without ci ). The correlations would be reported for the cross-study comparison, with the sidenote to see the significance in the raw variances
the problem with this solution: if we want to compute the average effect, then the correlation that are high but non-significant will make the results murky: they will influence the average but should not be trusted. conundrum
Homework for @annierobi :
[ ] send the list of models that exhibit differences between covariances and correlations
[ ] investigate the differences between MLR and MLF estimation methods to see if they can resolve the inconsistencies
[ ] as investigation unfolds, collect the outputs with informative errors The list of the errors to look out for will be incorporated into the extraction script for flagging the problematic model solutions
Homework for @andkov :
[ ] incorporate the estimator selection into the extraction scripts, so that we can explore the differences systematically
[ ] using the initial list of problematic models (covariances and correlations disagree) create a way to detect the inconsitencies automatically.
next meeting is scheduled for Friday 2016-10-07-12:00 EDT, but the next two weeks we communicate via email.
Homework for @annierobi :
Homework for @andkov :
next meeting is scheduled for Friday 2016-10-07-12:00 EDT, but the next two weeks we communicate via email.