OBJECTIVES: Network meta-analysis (NMA) models are extended to adjust for potential effect modifiers by writing the study level effects as a linear function of the study level covariate. The most commonly used parameterisation of such network meta-regression models is to employ a single common study-level covariate effect versus the baseline treatment. However, connected evidence networks are not always of ‘star’ shape geometry and may not have an obvious choice for the reference treatment node. The aim of this study was to explore approaches to network meta-regression in evidence networks where there is more than one potential candidate for the reference treatment node. METHODS: A previously published NMA comparing exenatide once weekly with other glucagon-like peptide-1 receptor agonists for the treatment of type 2 diabetes mellitus was identified that included an evidence network geometry with two potential candidate reference treatments. A series of exploratory network meta-regressions were performed to predict the mean change in haemoglobin A1c (HbA) from baseline adjusting for the levels of HbA at baseline. These models explored alternative reference treatments and setting two treatments as the reference. RESULTS: The interaction coefficients of the models explored were consistent and setting two treatments as the reference allowed for a gain in statistical power in the regression model compared with the standard regression approach (single reference node). CONCLUSIONS: A potential approach to meta-regression is presented to address evidence networks which have a more complex shape than a star. The potential application of this type of approach is dependent upon the geometry of the evidence network, the clinical plausibility and validity of the assumption that the covariate of interest modifies treatment effect in the same way relative to the two reference treatments.