A comparison of arm-based and contrast-based models for network meta-analysis

Area of Science:

• Biostatistics
• Medical Informatics

Background:

• Network meta-analysis (NMA) models, specifically arm-based (AB) and contrast-based (CB) approaches, present debated differences in application.
• Existing models include the Lu-Ades (2006) CB model, the Hong et al. (2016) AB model, and intermediate variations.

Purpose of the Study:

• To compare the Lu-Ades CB model, the Hong et al. AB model, and two intermediate models for NMA.
• To elucidate the primary differences stemming from fixed versus random study intercepts.

Main Methods:

• Comparative analysis using hypothetical and real datasets.
• Examination of four key differences related to study intercepts and model assumptions.

Main Results:

• Fixed study intercepts utilize only within-study data, while random intercepts incorporate between-study data, potentially introducing bias.
• Random study intercepts facilitate broader estimands but require external data for optimal risk derivation; fixed intercepts are equally effective with external data.
• The Hong model permits relating treatment effects to study intercepts, unlike the Lu-Ades model.
• The Hong model's relaxed missing data assumption (missing at random arms) does not appear to mitigate bias.

Conclusions:

• Both AB and CB models are appropriate for NMA.
• The use of random study intercepts in NMA models necessitates a robust rationale, such as linking treatment effects to study intercepts.