Panel Presentation at the Royal Statistical Society Conference 2025
OBJECTIVES: Health Technology Assessment (HTA) guidelines in the UK mention the importance of uncertainty in cost-effectiveness analysis (CEA). The presence of missing data within data sets used to provide inputs for CEA can be a source of uncertainty. One setting in which missing data may be prevalent is for utility scores and these are often used in CEA. The intention of this analysis is to formalise recommendations for dealing with missingness and outline a minimum accepted reporting standard for missing utility data.
METHODS: A simulated patient level dataset (SIPD) was created in an oncology setting. The SIPD included utility scores for hypothetical individuals and treatments. Key prognostic variables were also simulated. Missing data was generated for utility for 10% and 30% of the observations, for missing completely at random (MCAR), at random (MAR) and not at random (MNAR) missingness mechanisms. Four methods for addressing missing data were analysed, complete case analysis (CCA), mean score estimation (MSE), multiple imputation via chained equations (MICE) and linear mixed modelling (LMM). The outcome of interest was the mean difference between the true utility value and the estimated utility value from each method.
RESULTS: MICE tended to result in the lowest mean difference across levels and mechanism. For the 10% level and MCAR mechanism, CCA resulted in negligible differences from the true utility value. MNAR data tended to result in substantial differences from the true utility value regardless of the level of missingness and method used. LMMs suppressed the standard deviation of utility.
CONCLUSION: Differing results across levels and mechanisms for each method used to deal with missingness were observed here. This suggest that the level and mechanism of missingness should be investigated and reported as a minimum standard for all data sets that inform economic models for HTA in the UK.