Purpose: To highlight the potential for misinterpreting the expected value of perfect information (EVPI) in the presence of structural uncertainty.
Method: The analyses were conducted using a model evaluating group cognitive behavioural therapy (gCBT) for treatment of postnatal depression within a UK setting. Efficacy and utility data had assigned statistical distribution, however two key parameters: the duration of the comparative advantage of gCBT, and the cost per woman of treatment were each assumed to be one of 3 values (12 months, 18 months and 24 months; £750, £1,000 and £1,500 respectively). As is common in economic evaluations, the nine scenarios were analysed using probabilistic sensitivity analyses with one (18 months and £1,000) being designated the base case. EVPI analyses were conducted for each scenario. The EVPI for the base case was compared with alternate EVPI estimates: The weighted average (assuming equal likelihood of each scenario) of the EVPIs calculated independently for each scenario using each associated adoption decision; with the weighted average assuming the adoption decision associated with the base case and with that when it were assumed the the values for costs and duration were the 95% confidence intervals for a lognormal and a normal distribution respectively.
Result: The EVPI per woman varied considerably. The base case value was £1; the weighted average of the individual EVPIs was £8; the weighted average assuming the base case adoption decision was £92 and the value using statistical distributions was also £92. Given the high prevalence of PND (14%) the discrepancies in the maximum cost of undertaking research estimated by each method will be large.
Conclusion: Given structural uncertainty, EVPI values can change markedly dependent on the method of calculation and care must be taken in interpreting these values. The base case alone, and the average EVPI considering each scenario in isolation significantly underestimates the EVPI compared with either averaging the values for each decision using the basecase adoption decision, or when synthesising the data into statistical distributions.
Candidate for the Lee B. Lusted Student Prize Competition