Missing continuous outcomes under covariate dependent missingness in cluster randomised trials

TL;DR

This paper evaluates methods for analyzing cluster randomized trials with missing continuous outcomes due to covariate-dependent missingness, comparing bias, standard error, and coverage of different approaches.

Contribution

It provides a comprehensive comparison of unadjusted, covariate-adjusted, and mixed model analyses, highlighting conditions for unbiased estimates under missing data scenarios.

Findings

Unadjusted and covariate-adjusted cluster-level analyses are unbiased only if missingness mechanisms and covariate effects are the same across groups.

Linear mixed models and multiple imputation are unbiased under all scenarios if interaction terms are included when covariate effects differ.

Multiple imputation slightly overestimates standard errors, reducing statistical power.

Abstract

Attrition is a common occurrence in cluster randomised trials (CRTs) which leads to missing outcome data. Two approaches for analysing such trials are cluster-level analysis and individual-level analysis. This paper compares the performance of unadjusted cluster-level analysis, baseline covariate adjusted cluster-level analysis and linear mixed model (LMM) analysis, under baseline covariate dependent missingness (CDM) in continuous outcomes, in terms of bias, average estimated standard error and coverage probability. The methods of complete case analysis (CCA) and multiple imputation (MI) are used to handle the missing outcome data. Four possible scenarios are considered depending on whether the missingness mechanisms and covariate effects on outcome are the same or different in the two intervention groups. We show that both unadjusted cluster-level analysis and baseline covariate adjusted cluster-level analysis give unbiased estimates of the intervention effect only if both intervention groups have the same missingness mechanisms and the same covariate effects, which is arguably unlikely to hold in practice. LMM and MI give unbiased estimates under all four considered scenarios, provided that an interaction of intervention indicator and covariate is included in the model when the covariate effects are different in the two intervention groups. MI gives slightly overestimation of average standard error, which leads to a decrease in power.