2
$\begingroup$

I am intending to perform a meta-analysis to analyze efficacy and safety profile of particular intervention. However, out of 9 studies included, some of them consists of a single group (experimental only), and the other consists of 2 groups (experimental and control).

I was wondering, can I perform meta-analysis with these data or do I have to exclude some of them? If I can, is there any guideline or example of how it's done? Another question will be, what software can I utilize to perform the meta-analysis?

Thank you very much

$\endgroup$
  • $\begingroup$ Very general questions -> very general answers. Be more specific. $\endgroup$ – user2974951 Jan 17 at 7:31
  • $\begingroup$ A perfect choice would be mvmeta, which can encompass missingness. $\endgroup$ – Joe_74 Jan 19 at 17:52
2
$\begingroup$

Firstly, remember all the limitations of uncontrolled trials (see e.g. ICH E10 Choice of Control Group in clinical trials). So, let's assume they have some value in your case (the good thing is that to some extent good statistical methods will discount such trials a lot, if some assumptions are not meet).

Secondly, one of the more obvious approaches is to use a generalized linear mixed effects model (GLMM). The very simplest form would be $$g(E Y_{ij}) = \beta_i + \gamma_j + \delta_{ij},$$ where $Y_{ij}$ is the outcome in arm $j$ of trial $i$, $g()$ is the link function, $\beta_i$ is the main effect for trial $i$, $\gamma_j$ is the (fixed) main effect for treatment $j$ and $\delta_{ij}$ is the deviation from effect off treatment $j$ in trial $i$ (treatment by trial interaction). Crucially, $\beta_i$ and $\delta_{ij}$ are random effects (which makes this model identifiable) that e.g. follow $N(0, \tau_\beta)$ and $N(0, \tau_\delta)$ distributions. If you do a frequentist analysis this does not matter, but I believe one would usually use a sum to zero constraint on the treatment main effects for a Bayesian analysis.

This type of model can be implemented in most statistical packages, e.g. the author of the article I mention provides SAS code. Since you are effectively doing an arm-based network meta analysis, looking at the software used by the practioners in that field might be useful. Stan (can be used in R via the rstan package) could be another very flexible choice that gives you complete freedom of specification for both a Bayesian and frequentist analysis.

$\endgroup$
  • $\begingroup$ (+1) and if the OP prefers a frequentist approach netmeta is a very fully featured package from a respected team. $\endgroup$ – mdewey Jan 17 at 13:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.