Methods: Data were obtained from a large randomized trial called the Atrial Fibrillation Follow-up Investigation of Rhythm Management (AFFIRM) study. The AFFIRM investigators compared rhythm-control to rate-control for treatment of atrial fibrillation and reported a non-significant mean survival gain (0.08 years) in the rate-control arm. There was considerable censoring in the patient level data, with only 16% of patients experiencing the primary endpoint (death) during follow-up. We compared the results of an economic evaluation accounting for the censoring to differing extents. Firstly, with no adjustment for censoring in either cost or survival, then with partial adjustment accounting for censoring of survival but not of cost, and finally, with adjustment for censoring in both cost and survival. In each case, the expected cost and expected survival duration were obtained via bootstrapping, and the cost-effective strategy was identified based on the expected values. The uncertainty surrounding the estimates of expected cost and survival duration were presented using cost-effectiveness planes and cost-effectiveness acceptability curves. The expected value of perfect information (EVPI) was calculated to determine the potential worth of further research to reduce uncertainty.

Results: The strategy identified as cost-effective was the same (rate-control) irrespective of the extent of the adjustment made for censoring. However, the extent of the adjustment for censoring impacted on the level of uncertainty surrounding the decision regarding cost-effective provision and consequently the potential worth of further research, through the EVPI. For a maximum willingness-to-pay for survival of $50,000 per life year gained, the uncertainty surrounding the decision to adopt rate-control varied between 98% (full adjustment) and 100% (partial adjustment). For a patient population of 1.5 million over a period of 5 years, the EVPI varied between $234,000 (partial adjustment) and $44 million (full adjustment).

Conclusions: Cost-effectiveness analyses employing patient level data should fully adjust for censoring in both costs and effects in order to appropriately measure the extent of the uncertainty surrounding the decisions regarding cost-effective provision and the worth of further research.