SMALL TRIALS, BIG ASSUMPTIONS: USING MULTILEVEL NETWORK META-REGRESSION TO ADDRESS IMPERFECT RANDOMIZATION IN CLINICAL EVIDENCE NETWORKS
Author(s)
Xingzhu Leng, MSc1, Jan Tuzil, PhD2, Will Sullivan, PhD3, Nathaniel Smith, PhD4.
1Maple Health Group, Antilopespoor, Netherlands, 2Maple Health Group, Prague, Czech Republic, 3Principal, Maple Health Group, Sheffield, United Kingdom, 4Maple Health Group, New York, NY, USA.
1Maple Health Group, Antilopespoor, Netherlands, 2Maple Health Group, Prague, Czech Republic, 3Principal, Maple Health Group, Sheffield, United Kingdom, 4Maple Health Group, New York, NY, USA.
OBJECTIVES: To evaluate whether main-effect covariate adjustments in multilevel network meta-regression (ML-NMR) correct for chance covariate imbalance in a small, randomized trial embedded in a placebo-controlled systemic lupus erythematosus (SLE) evidence network. To our knowledge, this role of main effect has not been previously examined.
METHODS: Six SLE trials; BLISS-52 (belimumab), MUSE (anifrolumab), PAISLEY (deucravacitinib), SLEek (upadacitinib and elsubrutinib), RISE and PHOENYCS GO (dapirolizumab pegol); were combined with a simulated trial, WOBBLE. WOBBLE compared a hypothetical treatment, “randomizumab”, with placebo in 120 subjects randomized 1:1 and was simulated with 0%, 5%, 10%, and 20% imbalances in Systemic Lupus Erythematosus Disease Activity Index (SLEDAI) score. Fixed-effect network meta-analysis (FE NMA) and FE ML-NMR (multinma R package) estimated the odds of Systemic Lupus Erythematosus Responder Index 4 (SRI-4) response, with baseline SLEDAI set to be correlated to corticosteroid use. Covariate effects were assumed shared across treatment arms to ensure identifiability. Mean estimated relative treatment effects (RTE) on the logit scale across 100 iterations were compared to the true RTE of 1.16 for randomizumab.
RESULTS: With 5% imbalance in SLEDAI score, the NMA-estimated randomizumab RTE (0.90, 95% credible interval 0.25-1.60) was biased. Main effect shifted the RTE estimate to 1.06, by a mean of 21% towards the true RTE. The addition of interaction terms did not improve bias correction (RTE=1.00, mean shift 14% vs. NMA). Comparable RTE estimates were obtained for 10% and 20% imbalance, with bias corrected in a dose-dependent manner (29% and 44% mean shift, respectively).
CONCLUSIONS: Main effect use in early phase networks can shift estimated RTE meaningfully by correcting for chance covariate imbalance. It may therefore be justified not only for precision, but also for reducing bias.
METHODS: Six SLE trials; BLISS-52 (belimumab), MUSE (anifrolumab), PAISLEY (deucravacitinib), SLEek (upadacitinib and elsubrutinib), RISE and PHOENYCS GO (dapirolizumab pegol); were combined with a simulated trial, WOBBLE. WOBBLE compared a hypothetical treatment, “randomizumab”, with placebo in 120 subjects randomized 1:1 and was simulated with 0%, 5%, 10%, and 20% imbalances in Systemic Lupus Erythematosus Disease Activity Index (SLEDAI) score. Fixed-effect network meta-analysis (FE NMA) and FE ML-NMR (multinma R package) estimated the odds of Systemic Lupus Erythematosus Responder Index 4 (SRI-4) response, with baseline SLEDAI set to be correlated to corticosteroid use. Covariate effects were assumed shared across treatment arms to ensure identifiability. Mean estimated relative treatment effects (RTE) on the logit scale across 100 iterations were compared to the true RTE of 1.16 for randomizumab.
RESULTS: With 5% imbalance in SLEDAI score, the NMA-estimated randomizumab RTE (0.90, 95% credible interval 0.25-1.60) was biased. Main effect shifted the RTE estimate to 1.06, by a mean of 21% towards the true RTE. The addition of interaction terms did not improve bias correction (RTE=1.00, mean shift 14% vs. NMA). Comparable RTE estimates were obtained for 10% and 20% imbalance, with bias corrected in a dose-dependent manner (29% and 44% mean shift, respectively).
CONCLUSIONS: Main effect use in early phase networks can shift estimated RTE meaningfully by correcting for chance covariate imbalance. It may therefore be justified not only for precision, but also for reducing bias.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR98
Topic
Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)