Developing appropriate methods for cost-effectiveness analysis of cluster randomized trials.

Area of Science:

• Health Economics
• Biostatistics
• Clinical Trials

Background:

• Cost-effectiveness analyses (CEAs) often utilize data from cluster randomized trials (CRTs).
• Many existing analytical methods for CEAs of CRTs do not account for the clustered nature of the data.
• This oversight can lead to inaccurate estimates of uncertainty and the value of further research.

Purpose of the Study:

• To compare the performance of different statistical methods for conducting CEAs of CRTs.
• To identify appropriate methods for accommodating clustering in CEAs of CRTs.
• To evaluate methods under various scenarios, including different numbers of clusters and cost distributions.

Main Methods:

• A simulation study compared five statistical methods: seemingly unrelated regression (SUR) ignoring clustering, SUR with robust SE, generalized estimating equations (GEEs) with robust SE, a 2-stage nonparametric bootstrap (TSB) with shrinkage, and a multilevel model (MLM).
• Performance was assessed using bias, root mean squared error (rMSE), and confidence interval (CI) coverage for incremental net benefits (INBs).
• Scenarios included balanced/imbalanced clusters, few clusters, and skewed costs, alongside a case study.

Main Results:

• All methods showed low bias. SUR without robust SE had poor CI coverage (0.89 vs. 0.95).
• MLM and TSB demonstrated robust performance across all scenarios (CI coverage 0.92-0.95).
• GEE and SUR with robust SE had CI coverage below 0.90 with few clusters. Ignoring clustering underestimated uncertainty in the case study.

Conclusions:

• Multilevel models (MLMs) and the 2-stage bootstrap (TSB) are suitable for CEAs of CRTs.
• Seemingly unrelated regression (SUR) and generalized estimating equations (GEEs) are not recommended when CRTs have few clusters.
• Accurate statistical methods are crucial for reliable CEAs of CRTs to avoid underestimating uncertainty.