Session (CTI)

PURPOSE: Individual clinical trials are rarely powered to detect treatment effect modifiers, leaving subgroup analyses underpowered and prone to spurious findings. Combining information across trials makes identifying such subgroups more feasible. The canonical example is tamoxifen in breast cancer: individual trials were too small to determine whether the drug worked differently by oestrogen receptor status, but the EBCTCG patient-level meta-analysis demonstrated this convincingly, reshaping clinical practice. Identifying which patients benefit most is central to coverage and reimbursement decisions, yet available data vary widely — from full individual participant data (IPD) to aggregate subgroup summaries only. Suitable methods require careful adjustment to the setting at hand. DESCRIPTION: Dr. Remiro-Azócar will frame the problem: why subgroup-specific estimates matter for HTA, why subgroup analyses from individual trials are surprisingly unreliable, and why evidence synthesis offers a path forward (7 min). Dr. Jansen will present the network meta-analysis framework for subgroup analysis, including shrinkage estimation to stabilize uncertain subgroup-specific estimates, network meta-regression with aggregate and individual participant data, and how models can be set up to minimize ecological bias when patient-level effect modifiers are only available as study-level summaries (13 min). Renato Panaro will address subgroup-specific meta-analysis using aggregate subgroup data, showing how standard subgroup meta-analysis can disagree with interaction meta-analysis when subgroup contributions vary across studies, and presenting contribution-adjusted methods that resolve this inconsistency (15 min). Dr. Campbell will introduce new methods for leveraging subgroup-level summary data — frequently reported but typically ignored — within population-adjusted network meta-analysis (15 min). The session will conclude with a scenario involving a treatment network with varying data availability. The audience will be asked, via live polling, to identify appropriate methods, required assumptions, and to interpret results considering key sources of bias (10 min).