Hi all,
I am currently trying to get into bayesian NMAs. I have defined the network I want to analyze and we were now discussing if to include an additional node that would split one of the already existing nodes (A) into two (A and B). Older studies only used treatment A, but after a while a very similar but newer treatment (B) was developed which was though still kept being called A because at first it seemed so similar. After a while though, it was found that A and B are actually more distinct than initially thought. However, usually only more recent studies clearly labeled if A or B was used.
In our case we assume that older studies before a certain publication year could have only used A and that most recent studies clearly make the distinction of A and B. But studies that were published in the time in between won't always have reported specifically if they were using treatment B. Hence, some of them would end up in A although they have technically used B. On one hand I assume that the treatment arms of nodes A and B would therefore have increased correlation. Though, -opposed to a frequentist approach- in a bayesian framework this is for f. ex. including multiarm studies not a problem, when assuming that the effects stem from a multivariate normal distribution. Could I do something similar here? On the other hand, I am not sure in which other problems I might run due to this overlap in the nodes.
Leaving out this distinction of A and B in the network and keeping it all under the umbrella of A is an option but implies potentially loosing valuable information about B -which in fact is available in newer studies. And only including studies that make the distinction of A and B would mean 1. many studies that are older and have meaningful input had to be excluded and 2. this would in turn lead to a lot of information loss on A.
What advise is there regarding the issue?
What kind of problems -specifically analysis wise- do I need to consider if I (don't) split A and B?
Unfortunately I didn't find much on this issue. Maybe someone can give me a hint if it has occurred before and how it is usually dealt with.
I highly appreciate all kind of input
What you are referring to is a common issue in network meta-analysis. Determination of node levels generally depends on the clinical relevance of the distinction i.e. whether clinically relevant information would be lost. There are some statistical approaches to deal with this. The choice between them primarily depends on whether there is a hierarchal relationship (B is a subset of A).
One approach would be to use a network architecture that accounts for multicomponent interventions. This is most commonly done using an additive model, but the assumption in that scenario is that if A and B are combined to form a combined object C, that C is the sum of A and B. If this is true, an additive model can be built to account for this relationship (called component network meta-analysis). Most statistical packages for NMA have this capability. For Bayesian network meta-analysis, it is possible using gemtc for R.
If this is not applicable, a good starting point would be to start by including only the studies that clearly delineate A and B. You could build an additive model with A and B combined and a standard model with them separated and then compare. First, you should build the additive model using the studies that delineate A and B only, so that the same studies are included in each model. You can then assess the goodness of fit of the two models and the inconsistency. If no significant inconsistency is introduced by combining the treatments and the model fit is comparable, it is probably reasonable to pool them and include the other studies. If significant inconsistency is introduced by the combination, they should be left separated and the older studies excluded.
Jack Henry thanks a lot for the elaborate and quick answer, it's really helpful. I'll read into component NMA and will check out the function in gemtc - I decided to use that package anyhow-