Question about meta-analysis

#1
Hello! I'm new here and I'm not a statistician but I would appreciate your inputs.

Is it possible to make a meta-analysis investigating results from various randomized controlled trials (i.e. one has a population of diabetic patients, another has hypertensive patients, another one has pre-diabetic patients, yet another has breast cancer survivors, etc. --- all of these patients are obese and are adults) as long as all of them have been done with weight loss as an outcome and almost the same intervention applied (behavioral counseling)?

I have been arguing with a research partner about it where she says it can't be done, that I'll have to analyze each group differently while I say that it can be done. Obviously, we're not statisticians so we're just going around in circles.

Would really appreciate everyone's input here. Thanks.
 

hlsmith

Not a robit
#2
Well anything can be done, it just comes down to whether it would be appropriate. Most might flinch at pooling such heterogeneous samples even if you use random effects. The issue may be there is underlying differences in the samples and these differences may effect the effectiveness of the intervention. It would be fine to do for fun, but would get dinged for validity reasons if you tried to draw inferences or disseminate results.
 

hlsmith

Not a robit
#5
Interesting. I knew someone who did this with multilevel data where they had 28 hospitals, and would hold one out each time they ran the model.

Still not sure this addresses the issue, but perhaps you could try it as well as say holding out all diabetes samples, then HTN samples, etc.
 

ondansetron

TS Contributor
#6
Interesting. I knew someone who did this with multilevel data where they had 28 hospitals, and would hold one out each time they ran the model.

Still not sure this addresses the issue, but perhaps you could try it as well as say holding out all diabetes samples, then HTN samples, etc.
I would think a meta regression with dummy variables for the ailment would be okay. Thoughts?
 

hlsmith

Not a robit
#7
Well, as you probably already know - the standard sensitivity analysis for meta-analyses is to rerun the analyses m times, where m = number of studies. So you run it holding out each study once and checking to see if the results change when holding out a single study. If results change enough, you know the results are sensitivity to a study, possibly based on sample size or effect size.


Now running a regression and dummy coding ailment studies would control for effects, but it seems like it could neglect to control for within and between ailment groups covariance structures. I wonder if there would be a possibility of running a meta-analysis controlling for ailments as another level in the model, like multilevel modeling. Though I wonder if the model would be stable given possible small number of studies within ailment clusters.


Over all, I am not sure what would be the best approach. It would be interesting to try a bunch of approaches, but eventually the results would be questionable, since they would be dependent on the organic model building process and methods were not a priori stated.


Best approach may be to comb the literature and see how others address topic. I have a strong feeling they just run a MA for each ailment, like when you see authors run the RCT in one MA and observational studies in another MA.