Figure 12.2 shows an example of a bivariate, or two-way, sensitivity analysis, also taken from Suárez et al.4 In this particular case, the percentage of cross-protection for some non-vaccine serotypes was investigated in conjunction with discounting outcomes against Polish PE guideline values versus undiscounted outcomes. It can be seen from the lines that the undiscounted results are quite insensitive to inclusion of assumed cross-protection, whereas the discounted results do show some relevant sensitivity. Two-way sensitivity analysis is typically represented by different lines, as in Figure 12.2. Alternatively, a 3-dimensional graph can be constructed, as was done, for example, by Hubben et al. to depict the dependencies on discount rates for health and costs separately for infant pneumococcal vaccination in the Netherlands (Figure 12.3).5,6Rozenbaum et al. typically present a best-case and a worst-case cost-effectiveness for their analysis on antenatal HIV testing in the Netherlands.7 Taxonomy in best- and worst-case analyses can sometimes be a bit counter-intuitive as, for example, in this specific publication, a higher prevalence of HIV among pregnant women con- tributes to an improved cost-effectiveness. Yet, a higher prevalence is difficult to be envisaged as “best” in many other respects. In particular, the authors estimated that antenatal HIV testing would cost €6495 per life-year gained in the best case (maxi- mum cost-effectiveness ratio), whereas antenatal testing would be cost saving in the worst case.12.3.2 probabilisTic saProbabilistic SA concerns the assignment of formal probability distributions or den- sity functions to specific parameters in the model. Probabilistic SA is sometimes referred to as stochastic SA. This type of analysis was first suggested by Doubilet et al.8 Generally, these probability distributions are designed for the mean values of the selected parameters (second-order SA), rather than for the sample data from which the estimated mean is derived (first-order). Using these distributions, typically 1000 or more simulations are done using random draws from the defined distributions in each simulation. Each individual one (often referred to as “replicate”) from these multiple simulations translates into an estimate of the incremental cost-effectiveness ratio. Again from the Suárez et al. paper,4 Figure 12.4 shows a scatter plot of 10,000 replicates around the base-case estimate of cost-effectiveness for vaccinating teen- age girls against HPV in Ireland. Both nondiscounted, as well as discounted, out- comes using a 3.5% discount rate (according to the UK PE guideline), are shown.Probabilistic SA (or PSA) is often further represented in a cost-effectiveness acceptability curve (CEAC). The CEAC shows, for a range of acceptability or will- ingness-to-pay (often denoted with λ), the proportions in the scatter plot that are below each individual λ. Figure 12.5 shows the corresponding CEAC to the scatter plot in Figure 12.4. Additionally, a CEAC with 2% discounting is included in the fig- ure, possibly better reflecting the Irish underlying time preference (see Chapter 10 on discounting). For example, it can be read that with a discount rate of 3.5%, approxi- mately 80% of replicates correspond to a cost-effectiveness ratio below €50,000 per QALY. Also, 95% of replicates, or more, provides an acceptable cost-effectiveness if λ is chosen at €40,000 or more, using a discount rate of 2%.Of course, the major issue in probabilistic SA concerns the exact choice and specification of the probability distributions for the mean parameter values. In the absence of adequate information, often uniform or triangle distributions are taken over plausible ranges with the base-case parameter values as midpoints or expected values. In particular, for both of these types of distributions, a minimum
and maximum are defined, with equal probabilities for each value in between for the uniform distribution and increasing probabilities from the minimum or maxi- mum if moving to the predefined top of the triangle. Also, referring to the central limit theorem, normal distributions are often considered. Indeed, Suárez et al.4 used uniform distributions for parameters such as unit costs and screening coverage, and normal distributions for vaccine effectiveness and sensitivity of screening. De Vries et al.9 used normal and triangle distributions for transition probabilities in the decision tree reflecting progression to aspergillosis and candidosis underlying their analysis of cost-effectiveness of itraconazole prophylaxis against invasive infec- tions for neutropenic cancer patients. Also, Postma et al.10 used normal and uniform distributions for average length of stay, antibiotic prescriptions, and indirect costs of production losses in their analysis of the cost-effectiveness of treatment with oseltamivir for influenza patients.
đang được dịch, vui lòng đợi..