VoI book code: Case study and guidelines for estimating EVPPI using regression in R"
In voi: Expected Value of Information

Change this path to reflect where you downloaded the Chemotherapy_Book repository.

chemo_dir <- system.file("Chemotherapy_Book",package="voi")
# setwd(chemo_dir) # only if running interactively

This block of code is only needed if you want to compile this Rmd source file into a document.

library(knitr)
library(voi)
opts_knit$set(root.dir = chemo_dir)
options(digits=3, scipen = 1e+05)

Draw a sample from probabilistic analysis of the model.

source("04_analysis/01_model_run.R")
nb <- m_net_benefit[ , , wtp_seq == 20000]

Calculating single-parameter EVPPI.

Calculate the single-parameter EVPPI for all the parameters in the model. This returns a data frame containing the EVPPI estimates.

library(voi)
pars_all <- as.list(names(m_params))
ev_single <- evppi(outputs=nb, inputs=m_params, pars=pars_all)
ev_single

Dot-plot of the estimates, sorted with the highest values at the top.

plot(ev_single, order = TRUE)

Single-parameter EVPPIs for a specified subset of parameters.

evppi(outputs=nb, inputs=m_params, 
      pars=list("logor_side_effects", "p_side_effects_t1", "u_hospital"))

Multi-parameter EVPPI for four groups of parameters: those associated with side effects, transition probabilites, costs and utilities respectively.

par_groups <- list(
  "side_effects" = c("p_side_effects_t1","logor_side_effects"),
  "trans_probs" = c("p_hospitalised_total","p_died",
                    "lambda_home","lambda_hosp"),
  "costs" = c("c_home_care","c_hospital","c_death"),
  "utilities" = c("u_recovery","u_home_care","u_hospital")
)
ev_grouped <- evppi(outputs=nb, inputs=m_params, pars=par_groups)
ev_grouped

In this example, it is clear that the parameters associated with the risk of side effects have the greatest EVPPI.

Checking regression models for EVPPI calculation

Figure shown in the book

ev_single <- evppi(outputs=nb, inputs=m_params, pars=pars_all, check=TRUE)
check_regression(ev_single, pars = "logor_side_effects")

Additional analysis with standard errors:

evppi(outputs=nb, inputs=m_params,
      pars = list("p_side_effects_t2","u_hospital"), se=TRUE)

Alternative regression models:

evppi(outputs=nb, inputs=m_params,
      pars = list("p_side_effects_t2","u_hospital"))
evppi(outputs=nb, inputs=m_params,
      pars = list("p_side_effects_t2","u_hospital"),
      method="earth")

Comparing different regression specifications: single-parameter EVPPI

(e1 <- evppi(outputs=nb, inputs=m_params, pars=par_groups[1], check=TRUE))
(e2 <- evppi(outputs=nb, inputs=m_params, pars=par_groups[1], 
             gam_formula="s(p_side_effects_t1) + s(logor_side_effects)", check=TRUE))
check_regression(e1, plot=FALSE)
check_regression(e2, plot=FALSE)

earth models with two-way versus three-way interactions

(e1 <- evppi(outputs=nb, inputs=m_params, pars=par_groups[1], method = "earth", check=TRUE))
(e2 <- evppi(outputs=nb, inputs=m_params, pars=par_groups[1], method = "earth", 
             degree=3, check=TRUE))
check_regression(e1,plot=FALSE)
check_regression(e2,plot=FALSE)

Comparing different regression specifications: multi-parameter EVPPI

costs_utilities <- c(par_groups$costs, par_groups$utilities)
ev1 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="gam", 
             gam_formula = "s(c_home_care) + s(c_hospital) + s(c_death) + 
                            s(u_recovery) + s(u_home_care) + s(u_hospital)",
             check=TRUE)
ev2 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="gam", 
             gam_formula = all_interactions(costs_utilities, 2), check=TRUE) 
ev3 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="earth", check=TRUE)
ev4 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="earth", 
             degree=2, check=TRUE)
ev5 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="gp") 
ev6 <- evppi(outputs=nb, inputs=m_params, pars=costs_utilities, method="inla")