PNS322 REVIEW OF METHODOLOGICAL APPROACHES TO ANALYSE LONGITUDINAL UTILITY DATA

Abstract

Reliable estimates of the impact of health events on quality of life are essential when estimating the cost-effectiveness of health care interventions. Many cost-effectiveness models describe longitudinal changes in patients’ health utility as their disease progresses through different stages. This review aimed at identifying existing statistical methods to derive utility values from clinical trials or longitudinal observational studies, summing-up their goals, assumptions, pros and cons. A search strategy focusing on keywords related to “utility” and “longitudinal analysis” was implemented in Embase/Medline in April 2019. Six statistical approaches were identified in the literature. First, generalized linear mixed models (GLMM): they are designed to estimate subject-specific effects, do not require strong assumption with regards to missing data pattern, but require large sample sizes. Second, generalized-estimating-equations (GEE): they are designed to estimate population-averaged effects, but make strong assumptions with regards to missing data. Third, two-part-mixed-models (TPM): they investigate both subject-specific and population-specific effects, although modelling movements from part-one (utility=1) to part-two (utility<1) in consecutive surveys can be problematic. Fourth, latent-class models (LCMs): they assume that there is an unobserved variable splitting the population into two, and can be useful when the distribution of observed utilities is bimodal. Fifth and sixth, Tobit and censored least absolute deviations (CLAD) models: both assume that utility can exceed 1 and is censored at 1, but their use is questionable, since utility values cannot exceed 1 in reality. Further research is encouraged, whether this comes from agencies providing guidelines or from individual initiatives presenting case studies and recommendations. Challenges are especially foreseen for small sample sizes, handling of missing data, comorbidities effect and recurrent events.