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.
I'm trying to run a model similar to Tang & Cheung's (2016) model where you have two reflective construct that each have observed measures and then these two constructs in turn predict another…Continue
Started by Sergio Canavati. Last reply by Sergio Canavati Dec 1, 2018.
Hi there,I am using A/Prof Mike Cheung's TSSEM (See:…Continue
Started by Ryan Tang. Last reply by Jisoo Ock Oct 7, 2018.
Dear All,I am new to MASEM and I am trying to test mediator effect in meta analysis using MASEM. When I run TSSEM 2 I get a saturated model with df =0 for the target model. I want to see how to…Continue
Started by saneesh edacherian. Last reply by saneesh edacherian Oct 5, 2018.
Dear all, (sorry if this request is cross-posted) one aspect in meta-analyis that is not entirely clear to is why some meta-analytical studies use the standardized slopes of one (or more)…Continue
Started by Empi. Last reply by Mike Cheung Oct 4, 2018.
Loading feed
Comment
Hi Sergio,
I have two clarifications.
The approach by Wilson, Polanin & Lipsey (2016) is not the same as the multi-level multivariate model in Chapter 6 of my 2015 book.
My tssem2() or wls() were developed for fitting SEM with correlation/covariance matrices. I don't know whether it is appropriate to use them to fit multi-endpoint studies.
Regards,
Mike
Hi Mike,
Good point. I'm using the approach by Wilson, Polanin & Lipsey (2016) which you cite in Chapter 6 of your 2015 book in the discussion of multi-level multivariate models. You also wrote an introduction to the special issue in which it was published in RSM. Basically, in this approach I use the metafor rma.mv function (indicting random effects at the effect size and study levels [that way I can use multi-endpoint studies]) to estimate no-intercept estimates that I can use to populate my correlation table and generate the asymptotic covariance matrix.
Then, in the second stage I enter the aCov from metafor into the wls function in metaSEM. My n = sum(sample sizes of all studies).
Is there another way I can perform MASEM using multi-endpoint studies without having to average effect sizes from multi-endpoint studies?
Thanks,
Sergio
Hi Sergio,
I am confused. I thought that you were fitting SEM with the TSSEM. But you have just mentioned multi-endpoint studies which are not related to the TSSEM.
Regards,
Mike
Hi Mike,
I tried OSMASEM for a study I'm conducting but the results changed because it required that I average effect sizes in studies that provide more than one effect size for a specific construct relationship. Since I have an average of 2 effect sizes for each construct relationship in each sample study, I was using this method (https://doi.org/10.1002/jrsm.1199) to estimate the aCov and the used the WLS function to fit the SEM in the second stage. In your recent PsychFrontiers article, you show that the estimates in multi-endpoint studies are biased in multi-level multivariate models when we assume homogeneity of variance. The solution you propose there also doesn't help me because the studies are using provide multiple end-points for the same construct relationships. How do you suggest I proceed?
Dear Sergio,
If you were referring to the regression coefficients or factor loadings are U shape, I don't think that it can be easily done given that this is a meta-analysis. If you were referring to the relationship between a moderator and the regression coefficients is U shape, it might be possible under the One-Stage MASEM. https://psyarxiv.com/ce85j But it won't be an easy task.
Best,
Mike
Dear Mike,
How can I test inverse u-shaped variable relationships (quadratic terms) in TSSEM?
Sergio
Dear Empi,
As I don't use STATA, I am not sure how difficult it is to implement the WLS fit function in STATA.
I have a draft on this topic (as a co-author). If you would like to try the STATA code, please drop me an email.
Best,
Mike
Dear all,
my question is whether there is an established minimum number of studies necessary to conduct a random effects meta analysis?
In the neighboring field of multilevel analaysis, the typical advice is to use at least 20-30 groups for random intercept models. However, my impression is that random effects meta-analytical models - which seem to resemble the logic of twolevel analyses - are often based on relatively few studies, say approx. k=10-15. I couldn't find any simulation study, hence I would be very grateful for suggestions...
Best,
Empi
Dear Empi,
#1: I think so though I have not tested it yet. The standard error with the “stdyx” is based on the delta method, which should be the same as the one with the SEM approach.
#2 Yes.
#3 Regarding synthesizing interaction effects, I would suggest using the raw data rather than the summary statistics. If the scales of the variables are comparable across studies, you may compare whether the slopes of the interaction term are equal across samples. It is a three-way interaction.
Best,
Mike
© 2019 Created by Blair T. Johnson. Powered by
You need to be a member of Meta-analytic structural equation modeling to add comments!