Cluster randomised trials (CRTs) often face issues with missing data and treatment non-adherence. Guidelines such as CONSORT require that the numbers of clusters and individuals randomised, receiving treatment and analysed are reported, while newer guidelines go further (see for example the ICH 9 addendum on estimands), suggesting that an adherence-adjusted estimand is reported alongside an intention-to-treat estimate. For both analyses, missing data in covariates and outcome need to be handled appropriately as failure to do so can introduce bias in treatment effect estimates, leading to invalid inferences.
Multiple imputation has become a popular method to handle missing data, and instrumental variables methods can be used to obtain adherence-adjusted average treatment estimates (CACE). Both of these statistical techniques should reflect the hierarchical nature of CRT data.
In this talk, I will present multilevel multiple imputation for continuous and binary data, and then
present two approaches to obtain CACE for CRTs: (1) cluster-level analysis two-stage least square method, with inferences at the cluster level, and (2) mixture models with random effects for individual level CACE.
I will illustrate these methods by re-analysing a CRT in UK primary health settings. The OPERA study trial, which studied the effect of a physiotherapist-led exercise intervention on depression and physical health in elderly residents of nursing home.