Investigate "data points" for trials with no events in any arm

[tommorris168]

In the binary model, when a trial has no events in any of its arms, a warning is displayed on the 1c. Network Plot tab:

Removing the offending trial in the sensitivity analysis options removes the warning from the sensitivity model (but not from the main model). When running the main (Bayesian) analysis and the sensitivity analysis, the model statistics are different, including the number of "data points":

According to the warning, the dropped trial is "not considered in network meta-analysis". In that case I would expect the data points (and the other statistics) to be the same. They are the same in the frequentist model, but not the Bayesian. For the other statistics this can be explained by Bayesian randomness, but that doesn't explain the difference in data points.

The first step to resolving this is to find out how trials with no events in any arm are handled in gemtc.

[nabury]

Could you add the file you've used and the options chosen in the app so that I can reproduce this?

[tommorris168]

Dataset here: Data points and no events problem.csv

• Select Binary on the homepage
• Load the data
• Go to Data Analysis -> 1.Data summary -> 1c.Network Plot to see the warning.
• Remove "Okuyemi 2022" on the left to see that the warning is removed for the sensitivity analysis.
• The Frequentist tab shows that the frequentist results are identical.
• Run both main and sensitivity analysis on the Bayesian tab to compare the results. Data points are 31 in the main and 29 in the sensitivity.

[BalancedmonkeY]

Summary information on tab 1c is created using netconnection() which is part of the {netmeta} package. For conducting an NMA using the {netmeta} package, if both arms of a two-arm study have 0 outcome, the entire study will be removed from the frequentist analysis -> hence the warning from netconnection()

However, the Bayesian analysis does allow studies with zeros in every arm -> hence the inclusion of all the data points.

So the differing information is because of the differences between packages.

Action-> It should be put somewhere on tab 1c that the warning about removed studies is only applicable to the frequentist analysis (or similar)