Methods: After rank ordering 8 health states, 22 patients post-temporal lobe surgery assessed their utilities using the standard gamble. Rank order was identical for 6 of the 8 health states. We explored 4 approaches: 1) logit for the best health state and separate independent normal distributions for mean differences versus next best health state (BASE); 2) a single sample from a normal distribution was used and applied to each mean difference and standard deviation (ONE); 3) to preserve the rank ordering, the maximum utility of less desirable health states was that of the next best health state (MAX) except for the 2 health states; and 4) a combination of ONE and MAX a single sample and utilities for less desirable health states could not exceed that of the next best health state (ONE MAX). Using our Monte Carlo simulation comparing surgery versus continued medical management for patients with pharmacoresistant temporal lobe epilepsy, we estimated the gain in quality-adjusted life expectancy (QALYs) and the frequency with which surgery was preferred.
The table presents the results of the 4 different methods used to estimate utilities in probabilistic sensitivity analyses using Monte Carlo simulation. As seen in the table, the standard BASE approach has the lowest QALY benefit from surgery and a lower frequency with which surgery was preferred.
|QALY Benefit (95% CI)||Surgery Preferred (%)|
|ONE||8.4 ( 1.0-21.5)||99%|
|ONE MAX||8.7 (-0.5-23.0)||96%|
Methods that account for rank ordering of health state utilities during Monte Carlo simulation result in different outcomes than that under the standard approach. The current method that ignores rank ordering of utilities may underestimate treatment benefit. Using a single sample from normal distribution leads to higher treatment benefit estimates.