Counterfactuals and models for mean change in longitudinal outcomes in studies with high mortality

Mixed effects models are often used to evaluate changes in repeated measurements of longitudinal outcomes obtained during the follow-up period of randomized clinical trials. Often, a high proportion of the planned measurements are missing due to factors such as intermittent missed visits, dropout, or death. If the probability of missingness does not depend on the unobserved measurements, the data are said to be missing at random (MAR), and standard likelihood-based inference can be conducted without explicitly modeling the probability of missing data. If the MAR assumption is not satisfied,“ informative censoring” models may be used which jointly model both the longitudinal outcome and the process leading to missing data. However, when attrition due to death is high, these approaches have been criticized because the analyses estimate hypothetical complete-data marginal means. For example, an estimate of the treatment effect at 3 years follow-up refers to the difference in the mean of the outcome variable at 3 years between the treatment and control groups in all patients, including those who died prior to 3 years. For those patients who did in fact die shortly after randomization, one has to consider obscure hypothetical quantities that would have been observed had the patients not died.

In this talk I will apply the framework of the Rubin Causal Model to suggest a modification of pattern-mixture informative censoring models in which the estimated parameters refer to subsets of patients who would have survived sufficiently long for the parameters to be meaningful if the patients had been randomized to the control group. For patients randomized to the treatment group, this approach requires the interpretation of counterfactual variables. While the dependence of these models on counterfactual quantities leads to certain difficulties, the parameters estimated under this approach nonetheless appear to have a more clinically meaningful interpretation than the hypothetical complete-data marginal means.� 

Concepts from the talk will be illustrated using data from a recently completed randomized trial conducted in patients undergoing hemodialysis.