Can Missing Data Boundary Estimates Cast Doubt On Previous Conclusions Reevaluating A 10 Year Cost-Effectiveness Analysis (CEA) Of The Diabetes Prevention Program Outcomes Study (DPP/DPPOS)?

Abstract

In a previous reanalysis of this CEA of placebo, metformin, individual and group lifestyle (ILS and GLS) interventions, we showed, contrary to the original study’s claim, that there appeared to be no place for metformin in diabetes prevention. Here we explore whether unknown (and ignored) results from missing study participants, could have had an effect on the results and to explore the implications for decision uncertainty via probabilistic sensitivity analysis (PSA). Manski (1989) suggested using boundary estimates to estimate potential effects of missing data without making potentially unreasonable assumptions about such data (e.g. ignorability). We implemented this by assuming all missing participants were alternatively diabetic and non-diabetic for each intervention, and assigning mean QALYs and costs conditional on diabetes status. To test the ability to contest our previous conclusion on metformin not having a place in diabetes prevention, we focused on the lifestyle pessimistic case (all missing had diabetes) vs. the metformin optimistic case (all missing remained pre-diabetes). GLS remained cost-effective relative to the other three interventions (pessimistic ILS dominated). The incremental cost-effectiveness ratio (ICER) between GLS pessimistic (highest QALYs) and metformin optimistic was approximately $32,000/QALY. CEAC/Fs showed that GLS pessimistic probability of being cost-effective was the highest of the interventions at a WTP approximately equal to that ICER. If GLS was not available, the ICER between ILS pessimistic and metformin optimistic was approximately $55,700. Even using pessimistic GLS estimates, GLS remained the cost-effective choice. If for some reason, GLS is not available, pessimistic ILS is cost-effective compared to optimistic metformin at conventional WTP threshold values. Based on these data, even when using boundary estimates to the maximal advantage of metformin, there appears to be no place for metformin in diabetes prevention. Other rationales would need to be employed (e.g. heterogeneity of treatment effect, unexplored here).