Engelhart, Luella. The impact of different estimation procedures on net monetary benefit in clinical trials. Retrieved from https://doi.org/doi:10.7282/T3DF6R5X
DescriptionBackground: Growth in health care spending has led to greater use of cost-effectiveness analysis (CEA) in assessing health technologies. Traditional CEA uses the incremental cost-effectiveness ratio (ICER), a measure with statistical issues and limitations with missing data. Better analytic methods for CEA are needed to inform health care policy decisions. Objectives: The study evaluated estimates of cost-effectiveness from three models using incremental net monetary benefit (INMB) rather than ICER. Estimates were compared under different conditions of missingness. Data were simulated to include missing at random (MAR) and missing not at random (MNAR) nonresponse mechanisms as defined by Little and Rubin (2002). Methods: The parameter of interest was INMB. Models were ANCOVA, mixed effects (ME), and joint mixed effects and log of time-to-dropout (joint ME), a selection model. Because the joint ME model incorporates correlation between time-to-dropout and random effects of the longitudinal model of NMB into one model, the hypothesis was it would produce the best estimate. Simulated treatment effect provided a “true” INMB for model evaluations that included bias (absolute difference from “true”), precision (ratio of variances), and cost-effectiveness acceptability curves with willingness-to-pay (λ) values from $0 to $100k. Base case used a threshold criterion for dropout. Sensitivity analyses assessed impact of higher missingness. Post-hoc analysis used a trajectory criterion for dropout. Results: Base case analyses resulted in ANCOVA and ME models producing the least biased estimates. At λ = $50k, bias was $1.3k, $1.4k, and $2.3k, and precision was 1.27, 0.90, and 1.24 for ME, ANCOVA, and joint ME, respectively. ANCOVA estimates were best in sensitivity analyses although estimates were poor. The joint ME model performed best in the post hoc analysis. Conclusions: The models performed differently under alternative missingness conditions and were sensitive to nonresponse mechanisms. All estimates were poor when missingness was high, therefore, primary prevention of missing data should be a goal of research. MNAR nonresponse mechanisms are more complicated than implied by Little and Rubin’s definitions as shown by results with threshold versus trajectory criteria for dropout. Further research is needed with selection models in CEA and INMB as the measure of cost-effectiveness.