Camila Olarte Parra; Martina Amongero; Ekkehard Glimm

This session is a collection of presentations considering theoretical aspects of the estimands topic.

Hypothetical estimands in clinical trials: implementation of causal inference and missing data methods - Camila Olarte Parra 
The ICH E9 addendum outlines different strategies for handling intercurrent events but does not suggest statistical methods for their estimation. In this talk, we focus on the hypothetical estimand, where the treatment effect is estimated under the hypothetical scenario in which the intercurrent event is prevented. To estimate a hypothetical estimand, we consider methods from causal inference and missing data, establishing that certain ‘causal inference estimators’ are identical to certain ‘missing data estimators’. These links may help those familiar with one set of methods but not the other. We show that hypothetical estimands can be estimated by exploiting data after intercurrent event occurrence, which is typically not used. Moreover, using potential outcome notation allows us to state more clearly the assumptions on which causal inference and missing data methods rely to estimate hypothetical estimands and which (time-varying) variables should be adjusted for. The different estimators will be applied to a clinical trial conducted in patients with type 2 diabetes, where rescue medication needs to be available for ethical reasons. Accounting for rescue medication using the hypothetical strategy in this context, we will illustrate how the different estimators can be implemented, compare their assumptions, and their results. We will also discuss how to simultaneously account for other intercurrent events such as treatment discontinuation with either the treatment policy or hypothetical strategy. 
 
Treatment policy estimands for recurrent event data with missing data: COPD Vaccine case study using IPWC - Martina Amongero 
Under the Treatment Policy, whether an intercurrent event has occurred or not is irrelevant, the data will be collected and analyzed regardless. In presence of missing measurements (informative censoring), data need to be predicted based on plausible assumptions. For example, multiple imputation approach has been used to impute missing data based on similar subjects who remained in the trial.  
In this work, we explore an alternative causal inference approach called Inverse Probability of Censoring Weighting (IPCW). As a motivating example we consider a Phase 2 COPD Vaccine study where the clinical outcome is exacerbation (recurrent event) and the informative censoring is due to withdrawals. 
 
Intercurrent event time-based copy-to-reference method - Ekkehard Glimm 
The definition of intercurrent events (ICEs) for the estimand of interest is always linked to a specific strategy in regards to the analysis of the data. For estimands with multiple ICEs that use different strategies, it can be a challenge to define the imputation method and the analysis model. This becomes even more challenging, if missing data are imputed based on a multiple imputation approach. We will present an analysis method using only one imputation model, to impute data under the missing-at-random (MAR) or missing-not-at-random (MNAR) assumption which covers all different strategies we use for our multiple ICEs.  
 
The method that we propose is flexible, such that imputations can be done either based on the MAR assumption or using a reference based approach (MNAR), both within the same procedure in SAS® PROC MI.  
 
This method has various advantages: The data are imputed in one imputation step and hence the correlation and variability is based on the full set of patients. The method is simple to implement and intuitive, the switch to the reference group (if indicated) at the time of the ICE is implemented at a patient level and hence considers the pre-ICE profile and post-ICE profile of each patient. Furthermore, this method considers the half-life of the active treatment by including previous visits into the imputation model. Finally, it considers different imputation strategies without the need of running separate imputation algorithms.