RM1 - SENSITIVITY ANALYSIS FOR NOT-AT-RANDOM MISSING DATA IN TRIAL-BASED COST-EFFECTIVENESS ANALYSIS USING MULTIPLE IMPUTATION

Abstract

Multiple imputation is being increasingly used to address missing data in cost-effectiveness analysis (CEA) of randomised trials, assuming data are missing-at-random (MAR). However, in many CEA settings the missing data is related to unobserved values, i.e. data are missing-not-at-random (MNAR). For example, if patients in poorer health are less likely to complete health questionnaires, given the observed data. Guidelines recommend sensitivity analyses under plausible MNAR assumptions, but this is rarely done in practice. We aim to illustrate an accessible framework to conduct sensitivity analysis for departure from MAR in trial-based CEA, using the Ten Top Tips (10TT) trial evaluating a weight management intervention. We illustrate the implementation of pattern-mixture models using multiple imputation, modifying the imputed data to reflect a plausible departure from the MAR assumption. Sensitivity analyses are conducted under a range of plausible values (deviations from MAR), to assess the robustness of the study conclusions. We applied this framework to the 10TT CEA, to address the concern that participants who dropped out of the study (about 42%) could be those less successful at losing weight (MNAR). Under the base-case MAR assumption using multiple imputation, 10TT resulted in 0.004 fewer quality-adjusted life years (95%CI -0.074 to 0.066) and £35 lower costs (-504 to 434) compared to the control arm, with a 48% probability of being cost-effective at £20,000 per QALY. The sensitivity analysis illustrated that the study conclusions were sensitive to small departures from MAR (e.g. assuming missing quality-of-life was 10% lower than observed), with the probability of 10TT being cost-effective ranging from 16% to 75%. MNAR is an important concern in trial-based CEA and sensitivity analyses are recommended to assess whether the study conclusions are robust to departures from the missing-at-random assumption. We illustrated an accessible framework to conduct and report these sensitivity analyses.