
Estimation of the Expected Value of Partial Perfect Information (EVPPI) using a linear regression metamodel approach
Source:R/evppi.R
calc_evppi.Rdevppi is used to estimate the Expected Value of Partial Perfect
Information (EVPPI) using a linear regression metamodel approach from a
probabilistic sensitivity analysis (PSA) dataset.
Arguments
- psa
object of class psa, produced by
make_psa_obj- wtp
willingness-to-pay threshold
- params
A vector of parameter names to be analyzed in terms of EVPPI.
- outcome
either net monetary benefit (
"nmb") or net health benefit ("nhb")- type
either generalized additive models (
"gam") or polynomial models ("poly")- poly.order
order of the polynomial, if
type == "poly"- k
basis dimension, if
type == "gam"- pop
scalar that corresponds to the total population
- progress
TRUEorFALSEfor whether or not function progress should be displayed in console.
Value
A list containing 1) a data.frame with WTP thresholds and corresponding EVPPIs for the selected parameters and 2) a list of metamodels used to estimate EVPPI for each strategy at each willingness to pay threshold.
Details
The expected value of partial perfect information (EVPPI) is the expected
value of perfect information from a subset of parameters of interest,
\(\theta_I\), of a cost-effectiveness analysis (CEA) of \(D\) different
strategies with parameters \(\theta = \{ \theta_I, \theta_C\}\), where
\(\theta_C\) is the set of complementary parameters of the CEA. The
function calc_evppi computes the EVPPI of \(\theta_I\) from a
matrix of net monetary benefits \(B\) of the CEA. Each column of \(B\)
corresponds to the net benefit \(B_d\) of strategy \(d\). The function
calc_evppi computes the EVPPI using a linear regression metamodel
approach following these steps:
Determine the optimal strategy \(d^*\) from the expected net benefits \(\bar{B}\) $$d^* = argmax_{d} \{\bar{B}\}$$
Compute the opportunity loss for each \(d\) strategy, \(L_d\) $$L_d = B_d - B_{d^*}$$
Estimate a linear metamodel for the opportunity loss of each \(d\) strategy, \(L_d\), by regressing them on the spline basis functions of \(\theta_I\), \(f(\theta_I)\) $$L_d = \beta_0 + f(\theta_I) + \epsilon,$$ where \(\epsilon\) is the residual term that captures the complementary parameters \(\theta_C\) and the difference between the original simulation model and the metamodel.
Compute the EVPPI of \(\theta_I\) using the estimated losses for each \(d\) strategy, \(\hat{L}_d\) from the linear regression metamodel and applying the following equation: $$EVPPI_{\theta_I} = \frac{1}{K}\sum_{i=1}^{K}\max_d(\hat{L}_d)$$ The spline model in step 3 is fitted using the `mgcv` package.
References
Jalal H, Alarid-Escudero F. A General Gaussian Approximation Approach for Value of Information Analysis. Med Decis Making. 2018;38(2):174-188.
Strong M, Oakley JE, Brennan A. Estimating Multiparameter Partial Expected Value of Perfect Information from a Probabilistic Sensitivity Analysis Sample: A Nonparametric Regression Approach. Med Decis Making. 2014;34(3):311–26.