Marcel Wolbers, Alessandro Noci 

Standard & reference-based imputation methods based on conditional mean imputation Many RCTs compare a continuous outcome variable that is assessed longitudinally at scheduled visits between subjects assigned to an intervention vs. a control group. Missing outcome measurements may occur because subjects miss an assessment or drop out from the trial. Missing data methods based on MI are increasingly used to align the analysis strategy with the defined estimand.  Imputations may be based on a MAR assumption (with or without the inclusion of time-varying covariates such as ICE indicators) or on a reference-based imputation assumption. We propose & justify deterministic conditional mean imputation (based on maximum likelihood estimation of imputation parameters) combined with the jackknife for inference as an alternative approach. In an application and a simulation study, we demonstrate that our proposal provides unbiased treatment effect estimates & correct frequentist inference with accurate SE estimation & type I error control.

Many randomized controlled clinical trials compare a continuous outcome variable that is assessed longitudinally at scheduled follow-up visits between subjects assigned to a intervention treatment group and those assigned to a control group. Missing outcome measurements may occur because subjects miss an assessment or drop out from the trial altogether. Moreover, intercurrent events (ICEs) such as discontinuations of the assigned treatment or initiations of rescue medications may affect the interpretation or the existence of the outcome measurements associated with the clinical question of interest. The ICH E9(R1) addendum on estimands presents a structured framework to link trial objectives to a precise description of the targeted treatment effect in the presence of ICEs and missing data. Missing data methods based on multiple imputation (MI) are increasingly used to align the analysis strategy with the defined estimand.  Imputations may be based on a missing-at-random assumption (with or without the inclusion of time-varying covariates such as ICE indicators) or on a reference-based imputation assumption. Reference-based methods impute missing data in the intervention treatment group based on observed data from a reference group which is typically defined as the control group of the trial. Typically, imputation is implemented using Bayesian random multiple imputation and Rubin’s rules for pooling results across imputed datasets. However, this approach requires the specification of prior distributions and MCMC sampling. Moreover, it overestimates the frequentist standard error for reference-based imputation. We propose and justify deterministic conditional mean imputation (based on maximum likelihood estimation of imputation parameters) combined with the jackknife for inference as an alternative approach. In an application and a simulation study, we demonstrate that our proposal provides unbiased treatment effect estimates and correct frequentist inference with accurate standard error estimation and type I error control. Additionally, it can result in substantially more efficient treatment effect estimators under reference-based imputation assumptions than the Bayesian approaches. A further advantage of the method is that it does not rely on random sampling and is therefore easily replicable and unaffected by Monte Carlo error. The implementation of the method in the publicly available R package “rbmi” will also be described.