EVPPI for a subset of parameters

Values perfect information about a subset phi of the parameters, with the remaining parameters psi still uncertain. The inner expectation averages net benefit over psi given phi, the maximum picks the best option for that value of phi, the outer expectation averages over phi, and the value of the current decision is subtracted. The expected value of sample information replaces knowledge of phi with the data from a proposed study.

Signature

EVPPI_phi = E_phi[max_d E_(psi|phi)[NB_d(phi,psi)]] - max_d E_(phi,psi)[NB_d(phi,psi)]
Inputs
InputsDefinitionUnit
NB_dNet benefit of option d as a function of phi and psi, listed across draws and optionscurrency per person
phiSubset of model parameters, such as a group of treatment effects, whose values would be learntunits of the parameters concerned
psiAll other model parameters, which remain uncertain after phi is learntunits of the parameters concerned
Output
EVPPI_phiExpected value of perfect information about the parameters phi alonecurrency per person

Function

Value of information function

Maps the joint distribution of net benefit across the options in a decision to the expected gain from resolving some or all of the uncertainty before choosing, compared with choosing the option with the highest expected net benefit on current information.

Implementations

  • Excel

    EVPPI for a discrete parameter

    With draws in rows, the value of phi in the named range PhiGroup and the net benefits of options A and B in NB_A and NB_B, the formula in I2, filled down beside the distinct phi values in column H, returns the best conditional mean. For phi values with equal numbers of draws, EVPPI is then =AVERAGE(I2:I3)-MAX(AVERAGE(NB_A),AVERAGE(NB_B)).

    =MAX(AVERAGEIFS(NB_A,PhiGroup,H2),AVERAGEIFS(NB_B,PhiGroup,H2))

Assumptions

  • Conditional expectation estimated adequately

    The inner expectation is estimated by nested simulation or by regression of net benefit on phi, as in the generalised additive model and Gaussian process methods. With few inner draws the maximisation biases a nested estimate upwards.

  • Ordering with EVPI

    EVPPI lies between zero and the EVPI for all parameters, apart from small violations caused by numerical error. Parameters are grouped by the research question that could inform them.

Worked examples

  • Four equally likely draws with two parameters

    Parameter phi takes values 1 or 2 and psi takes values 1 or 2, independently and with equal probability. Options A and B have net benefits of £1,000 and £600 at phi 1 and psi 1, £600 and £800 at phi 1 and psi 2, £200 and £800 at phi 2 and psi 1, and £200 and £600 at phi 2 and psi 2. Knowing phi, A is chosen at phi 1 with a mean of £800 and B at phi 2 with a mean of £700, an average of £750. On current information B is chosen with £700, so EVPPI for phi is £50, half the EVPI of £100 for both parameters.

    phi = [1,1,2,2]; psi = [1,2,1,2]; NB_d = [[1000,600],[600,800],[200,800],[200,600]]; EVPPI_phi = 50

Common errors

  • Maximising before averaging over psi

    Taking the best option in each draw before averaging over psi values perfect information about all parameters, so the result is EVPI rather than EVPPI for phi. In the worked example this gives £100 instead of £50.

Sources

  • Partial EVPI from a probabilistic sample

    Strong M, Oakley JE, Brennan A. Estimating multiparameter partial expected value of perfect information from a probabilistic sensitivity analysis sample: a nonparametric regression approach. Medical Decision Making. 2014;34(3):311-326.

    View source →

  • ISPOR methods for EVPPI and sample information

    Rothery C, Strong M, Koffijberg HE, Basu A, Ghabri S, Knies S, Murray JF, Sanders Schmidler GD, Steuten L, Fenwick E. Value of information analytical methods. Report 2 of the ISPOR Value of Information Analysis Emerging Good Practices Task Force. Value in Health. 2020;23(3):277-286.

    View source →

Canonical Identity

Stable URI · Machine-readable · Resolvable · CC BY 4.0