Methods for observed-cluster inference when cluster size is informative: a review and clarifications

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Clustered data are common in epidemiology, often with missing outcome data.
• Missing outcomes can arise from non-existence (e.g., death) or other reasons.
• The distribution of outcomes in complete clusters may differ from observed clusters.

Purpose of the Study:

• To evaluate statistical methods for clustered data with missing outcomes.
• To determine if existing methods provide inference for observed or complete clusters.
• To identify conditions where shared random-effects models yield valid observed-cluster inference.

Main Methods:

• Review of weighted and doubly weighted generalized estimating equations.
• Analysis of shared random-effects models for clustered data.
• Investigation of conditions for observed-cluster inference.

Main Results:

• Proposed methods for observed-cluster inference may actually provide inference for complete clusters.
• This holds true even when observed clusters are internally complete.
• Conditions are identified for shared random-effects models to accurately reflect observed data.

Conclusions:

• Misinterpretation of estimates from shared random-effects models is a significant risk.
• Careful consideration of complete vs. observed cluster inference is crucial.
• Psoriatic arthritis data illustrate the potential for misleading results.