Presentation (CTI)

OBJECTIVES: To quantify the systematic bias introduced when missing health state utility data are ignored in responder analyses, examining how the strength of association between utility values and missingness probability affects health state utility values (HSUVs) used in economic evaluations.
METHODS: A simulation study using 400 hypothetical patients across responder and non-responder health states. Missing data were simulated under a Missing Not At Random (MNAR) mechanism where patients with lower baseline utility values and lower change from baseline values had progressively higher probabilities of missing data for utility or response. We varied missingness rates (10%-50%) and the strength of utility-missingness correlation (0.5-0.9) across 1,000 simulation scenarios. Complete case analysis (CCA), non-responder imputation (NRI), and multiple imputation (MI) approaches to estimate HSUVs were compared against true population HSUVs with a focus on the difference between responder and non-responder HSUVs.
RESULTS: All approaches systematically overestimated utility in responder and non-responder health states compared to the true utility values. This systematic overestimation increased with higher correlation assumed between missingness and utility. CCA and MI overestimated the difference between responder and non-responder HSUVs. In contrast, NRI systematically underestimated this difference, leading to a more conservative estimate in health economic terms.
CONCLUSIONS: Ignoring missing utility data in responder analyses can be seen as excluding patients with the worst outcomes, creating a bias that potentially inflates treatment benefits. This bias can directly impact cost-effectiveness ratios used in health economic evaluations. This simulation study shows that NRI is the most conservative approach to estimating HSUVs to ensure differences between responder and non-responder HSUVs are not overestimated.