Network Meta-analysis in R part III

Previously, i have shown how to structure the data, use the netmeta() function to implement the Drucker et al. method and plot the direct evidence network and the network effects forest plot. As i mentioned before the Network evidence is composed of both the direct and indirect effects. In this tutorial i will show how to separate the direct and indirect effects using the netsplit() function and the plot using different setting on the forest plot. Finally i will show how to interpret all the evidence, indirect, direct and network effects, togheader.

We will be using the same dataset as before as a continuation from the previous tutorial https://github.com/DarkoMedin/Network-Meta-Analysis-II/blob/main/nmadataset2.csv

We will also use the same first part of the code :

Next step is very important. I will use the netsplit() function which will split the network into two types of evidence estimates, the indirect and direct estimates. Why is this important? We can step up the interpretation and use more detailed plots for final interpretation and see exactly where out model added only the indirect estimates and where we had both the direct and indirect estimates. Finally we can compare indirect and direct estimates, this part of the interpretation is essential in NMA, ideally one would want to make sure that both types of evidence agreement / disagreement of effects is interpreted well. Now to the coding part...

As you can see i used the netsplit() function to create an nsplit object which now contains the split indirect and direct evidence. The a simple forest() function can be used to plot the separate indirect and direct evidence, but make sure to use spacing = 0.9 or higher to avoid overlaps of the effects as they are abundant on the plot. You also might need to export the plot using the .pdf format and set the height to 14.5 and width to 6 to be able to see the whole plot vertically. Here is the resulting plot.

Network Meta-analysis forest plots are the most important segment of Network Meta-analysis data visualizations. While many consider the Network graphs evaluated in the previous tutorials as the main topic, its actually the forest plots in NMA that provide the most detailed esitmates and should be used dominantly as a one of the most useful inference approaches in NMA. In the plot above we can see all the details in terms of agreement between indirect and direct evidence, in terms of risk ratios and their confidence intervals and the network effects too. We can see that treatment C and A are more effective versus placebo, but for treatment C the agreement between indirect and direct evidence is better. On the other hand we have more direct evidence for treatment A and the direct evidence being slightly more trustworthy also shows that Treatment C is better. So its a complex situation, but i intentionally chose it to show how complicated the interpretation might be. We may use the network to compare any pair of treatments or them again the rest of the network. It would be safe to say that treatments A and C are better then the rest in this analysis with quite high level of confidence and this would be a simpler part of the interpretation.

Now to an even more specific part. When you observe the network, you will see that there are no comparisons for those combinations of treatments for which there was no direct comparison initially. For example there is no A : C comparison on the network. Lets get back for a moment to the direct evidence network.

In the graph we can see that there are not direct comparisons between A and C or between B and G and so on. The power of Network Meta-analysis is that we can use the network structure to infer the indirect evidence even if no direct comparisons were made between different treatments. I can dot this bu using the show =' ' argument.

You can see that the comparisons which were not there by default such as A : C are now there. And the result actually tells the story i interpeted before, A and C are very similar, around 1 and the confidence interval both ways bellow 1 and above 1. Remember this comparison was not done directly in a study or studies but infered using the indirect inference using the network created (please note that all comparisons here are not from real data but from practice data). If you scroll down, you may notice some other comparisons which were also not in the direct evidence network, but i was able to infer them using the network effects (indirect effects specifically in this case).

That's it for tutorial NMA in R part III. In the next one we will discuss the publication ready styling of the forest plots. This is a very important segment in any NMA as the publications originating from them must have standardized and high quality plots. While the plots made so fare have enough quality for most journals, in the next tutorial we will discuss this topic in even more detail.

by Darko Medin