Meta-regression detected associations between heterogeneous treatment effects and study-level, but not patient-level, factors.

OBJECTIVE: Two investigations evaluate Bayesian meta-regression for detecting treatment interactions. STUDY DESIGN AND
SETTING: The first compares analyses of aggregate and individual patient data on 1,860 subjects from 11 trials testing angiotensin converting enzyme (ACE) inhibitors for nondiabetic kidney disease. The second explores meta-regression for detecting treatment interaction on 671 covariates, including the baseline risk, from 232 meta-analyses of binary outcomes compiled from the Cochrane Collaboration and the medical literature.
RESULTS: In the ACE inhibitor study, treatment effects were homogeneous so meta-regression identified no interactions. Analysis of individual patient data using a multilevel model, however, discovered that treatment reduced glomerular filtration rate (GFR) more among patients with higher baseline proteinuria. The second investigation found meta-regression most effective for detecting treatment interactions with study-level factors in meta-analyses with >10 studies, heterogeneous treatment effects, or significant overall treatment effects. Under all three conditions, 46% of meta-regressions produced strong interactions (posterior probability >0.995) compared with 6% otherwise. Baseline risk was associated with the odds ratio in 6% of meta-analyses, half the rate found using maximum likelihood.
CONCLUSION: Meta-regression can detect interactions of treatment with study-level factors when treatment effects are heterogeneous. Individual patient data are needed for patient-level factors and homogeneous effects.