Penalized regression in network meta-analysis: A new approach for analyzing networks of interventions with rare events

Tags: Oral
Evrenoglou T1, Mavridis D2, Chaimani A1
1Inserm Research Centre in Epidemiology and Statistics (CRESS-U1153), Université de Paris, 2Department of Primary Education, University of Ioannina

Background: Network meta-analysis (NMA) allows comparing simultaneously multiple interventions and, under certain conditions, provides the highest possible level of evidence for the development of clinical guidelines. However, conventional NMA models for dichotomous outcomes can provide biased and imprecise results when the available studies are small or there are few events. Cochrane suggests analyzing at least one efficacy and one safety outcome and the latter typically includes few events . Methods that allow properly analyzing individual studies with low event rates do exist but have never been considered in the context of NMA.

Objectives: To allow more accurate and less biased conclusions from NMAs evaluating rare safety outcomes by extending appropriate statistical methods used in the analysis of individual studies into full networks of interventions.

Methods: We developed a new statistical approach for low event rate binary data forming a full network of interventions. Following ideas previously suggested in the literature for the analysis of a single study, we reduce the bias of NMAs with rare events by modifying the likelihood function of the model. We evaluate the performance of our approach using various simulation scenarios and two real datasets: a network comparing the safety of different drugs for chronic plaque psoriasis and a network comparing interventions for decreasing blood loss and blood transfusion requirements during liver transplantation. To facilitate the use of our method, we have implemented it in R.

Results: In comparison to three alternative NMA models that have been suggested for handling rare events, our method gave on average more precise and less biased relative effect estimates in most simulated scenarios. In the real examples, our model led to much smaller confidence intervals than the other methods particularly for comparisons involving only one or two studies. This is explained partly because the method gives more precise results for trials with small event rates and partly because our approach allows the inclusion of all available studies no matter the number of events per arm. It this way, we also avoid the risk of ending up with disconnected networks due to the exclusion of studies with zero events without making arbitrary continuity corrections.

Conclusions: When we have studies with small event rates we should employ appropriately tailored methods to synthesize them. The suggested methodology offers a reliable and more informative alternative to existing approaches for the analysis of networks of interventions with rare events. Nevertheless, in the presence of small event rates we should always place less confidence in our effect estimates.

Patient or healthcare consumer involvement: None.