RoBMA Random-effects meta-analysis (with meta-regression)
Greetings,
I would like to use RoBMA, on JASP (preferably) or R, to conduct a Bayesian random-effects meta-analysis and meta-regression. I have 3 questions which I have asked separately elsewhere (JASP YouTube channel and another user's forum post here), but decided I would also put it together here on the forum for myself.
1. How can I work only with a random-effects model i.e., remove all fixed effects models from the 'Inference' and 'Plots' functions and then get the appropriate (model-averaged) estimates/plots? I am intrigued to know because this was a possibility mentioned on P.18 of Bartoš, Maier, and Wagenmakers' paper "simply removing the fixed effects models from RoBMA".
2. I also wish to explore the possibility of conducting a meta-regression to complement the main RoBMA analysis. I suspect it would be possible and sound to use JASP's Bayesian Linear Regression function for this purpose.
According to Cochrane Handbook for Systematic Reviews of Interventions:
"Meta-regressions are similar in essence to simple regressions, in which an outcome variable is predicted according to the values of one or more explanatory variables."
With a meta-analysis, since we are usually working with and reviewing aggregate data, does a regression analysis turn into a 'meta-regression' by virtue of using these data as part of a systematic review and meta-analysis?
In other words, is it OK to use the Bayesian Linear Regression function in JASP to conduct a Bayesian 'meta-regression' to complement the RoBMA analysis? Ideally with continuous and categorical (dummy-coded) potential effect moderators testing 3 models, e.g., participant characteristics, intervention characteristics, and general study characteristics.
Bergh, D.v.d., Clyde, M.A., Gupta, A.R.K.N. et al. A tutorial on Bayesian multi-model linear regression with BAS and JASP. Behav Res (2021). https://doi.org/10.3758/s13428-021-01552-2
3. How can I use (or plug-in) different point/variance estimators e.g., Hedges' g instead of the default Cohen's d (and apply the corresponding name/label to the figures)? For my analysis, I need to work with the standardized mean difference, but should the need arise, it would be great if I could use alternative estimators like Hedges' g to account for potentially small studies. Any advice on how to implement this on JASP/R would be appreciated.
I am new to the field and area of meta-analysis, so pardon me for my superficial understanding and if I do not understand whether there is a genuine difference between a meta-regression and an 'ordinary' regression for Bayesian analysis.
Thank you in advance!
Comments
I'll forward this to our experts
Cheers,
E.J.
Hi jber3175,
1) You can remove any type of models/prior specifications under the
Modelssection. There is a checkbox at the bottom of the section calledSet null priors. Selecting the checkbox will open additional settings that allow you to specify models for the null hypotheses. By clicking on theXbehind the Spike(x) under theHeterogeneity (null)removes the fixed-effect models from the ensemble (as the heterogeneity parameter tau = 0 correspond to fixed-effect models). Nevertheless, I would advise doing this only if you have a strong justification for it. In a recent paper (under review), we showed that fixed-effect models actually out-predict random-effect models in many cases and constitute an interesting hypothesis. Furthermore, if the data are indeed predicted well by the fixed-effect hypothesis, your test for the presence/absence of the effect and publication bias will become weaker.2) Regarding the second question, I will just paste my response from the other thread. Let me know if there was something unclear:
Currently, the Bayesian Meta-Analysis analysis does not support a metaregression. However, you are correct that you can obtain a similar functionality from the Bayesian Linear Regression analysis - it's important to keep two things in mind though:
a) You have to use ``WLS Weights'' argument to pass the weighting of the studies (usually 1/se^2). You would discard information about the precision of the study effect size estimate otherwise.
b) This will result in weighted least squares meta-regression that differs a bit from the fixed/random effects meta-regression models regularly used in psychology. Nevertheless, some authors (e.g., Stanley and Doucouliagos) that WLS meta-regression has better properties than fixed/random effects meta-regressions.
3) You can use Hedge's g and the corresponding SE as input in the module, however, I'm not sure how does the small study correction interplay with the weighted likelihood. Especially whether it keeps the corresponding p-values cutoffs. I will need to do more research into this.
Cheers,
Frantisek
Dear Frantisek,
1. Thank you again for clarifying this. Is it possible to share the pre-print of the paper?
2. Is the "WLS Weights" argument accessible on JASP or only accessible on R?
Cheers,
Josh
Hi Josh,
sorry for the late reply,
1) we unfortunately do not have a pre-print. Since the paper is already under review, we will post-print it. I will ask my coauthors whether I can share it with you though.
2) yes, it's accessible in the JASP Regression module:
Cheers,
Frantisek
Hi Frantisek,
Thanks again for taking the time to share your advice. I really appreciate it :-)
Cheers,
Josh