Elementary decision tree (Evans 1997)
In rdecision: Decision Analytic Modelling in Health Economics

#| purl = FALSE,
#| include = FALSE
# read vignette source chunks from corresponding testthat script
knitr::read_chunk(
  file.path("..", "tests", "testthat", "test-model-Sumatriptan.R")
)
# read vignette build utility functions
knitr::read_chunk(file.path("vutils.R"))

#| gbp,
#| echo = FALSE

#| include = FALSE,
#| purl = FALSE
knitr::opts_chunk$set(
  echo = FALSE,
  collapse = TRUE,
  comment = "#>"
)

#| purl = FALSE
#nolint start

library(rdecision)

#| purl = FALSE
#nolint end

Introduction

This vignette is an example of modelling a decision tree using the rdecision package. It is based on the example given by Briggs [-@briggs2006] (Box 2.3) which itself is based on a decision tree which compared oral Sumatriptan versus oral caffeine/Ergotamine for migraine [@evans1997]. In this vignette, we consider the problem from the perspective of a provincial health department.

Creating the model

Model variables

The following code defines the variables for cost, utility and effect that will be used in the model. There are 14 variables in total; 4 costs, 4 utilities and 6 probabilities.

#| modvars,
#| echo = TRUE

Constructing the tree

The following code constructs the decision tree. In the formulation used by rdecision, a decision tree is a form of arborescence, a directed graph of nodes and edges, with a single root and a unique path from the root to each leaf node. Decision trees comprise three types of node: decision, chance and leaf nodes and two types of edge: actions (whose sources are decision nodes) and reactions (whose sources are chance nodes), see Figure 1.

#| model,
#| echo = TRUE

#| results = "hide",
#| fig.keep = "all",
#| fig.align = "center",
#| fig.cap = "Figure 1. Decision tree for the Sumatriptan model"
DT$draw(border = TRUE)

Running the model

The method evaluate of decision tree objects computes the probability, cost and utility of each strategy for the model. A strategy is a unanimous prescription of the actions at each decision node. In this example there is a single decision node with two actions, and the strategies are simply the two forms of treatment to be compared. More complex decision trees are also possible.

The paths traversed in each strategy can be evaluated individually using the method evaluate(by = "path"). In rdecision a strategy is defined as a set of action edges with one action edge per decision node. It is necessary to use the option by = "path" only if information about each pathway is required; normally it is sufficient to call evaluate which will automatically aggregate the evaluation by strategy.

Model results

Base case

The evaluation of each pathway, for each strategy, is done as follows:

#| eval-by-path,
#| echo = TRUE

and yields the following table:

#| echo = FALSE
knitr::kable(
  ep[, c("Leaf", "Probability", "Cost", "Utility")],
  align = "lrrr",
  digits = c(2L, 4L, 2L, 5L),
  format.args = list(scientific = FALSE)
)

There are, as expected, ten pathways (5 per strategy). The expected cost, utility and QALY (utility multiplied by the time horizon of the model) for each choice can be calculated from the table above, or by invoking the evaluate method of a decision tree object with the default parameter by = "strategy".

#| eval-by-strategy,
#| echo = TRUE

This gives the following result, consistent with that reported by Evans et al [-@evans1997].

#| echo = FALSE
knitr::kable(
  es[, c("d1", "Cost", "Utility", "QALY")],
  align = "lrrr",
  digits = c(2L, 2L, 4L, 4L),
  format.args = list(scientific = FALSE)
)

#| icer-basecase,
#| echo = FALSE

The incremental cost was $Can r gbp(x = delta_c, p = TRUE) (r gbp(x = cost_s, p = TRUE) - r gbp(x = cost_c, p = TRUE)) and the incremental utility was r round(delta_u, 2L) (r round(utility_s, 2L) - r round(utility_c, 2L)). Because the time horizon of the model was 1 day, the incremental QALYs was the incremental annual utility divided by 365, and the ICER was therefore equal to r gbp(icer) \$Can/QALY, within 5% of the published estimate (29,366 \$Can/QALY).

Univariate sensitivity analysis

Evans et al [-@evans1997] reported the ICER for various alternative values of input variables. For example (their Table VIII), they reported that the ICER was 60,839 $Can/QALY for a relative increase in effectiveness of 9.1% (i.e., when the relief from Sumatriptan was 9.1 percentage points greater than that of Caffeine-Ergotamine) and 18,950 $Can/QALY for a relative increase in effectiveness of 26.8% (these being the lower and upper confidence intervals of the estimate of effectiveness from meta-analysis).

To calculate these ICERs, we set the value of the model variable p_sumatriptan_relief, and re-evaluate the model. The lower range of ICER (with the greater relative increase in effectiveness) is calculated as follows:

#| relief-threshold-upper,
#| echo = TRUE

#| icer-upper,
#| echo = FALSE

This yields the following table, from which the ICER is calculated as r gbp(icer_upper) \$Can/QALY, close to the published estimate of 18,950 \$Can/QALY.

#| echo = FALSE
knitr::kable(
  es[, c("d1", "Cost", "Utility", "QALY")],
  align = "lrrr",
  digits = c(2L, 2L, 4L, 4L),
  format.args = list(scientific = FALSE)
)

The upper range of ICER (with the smaller relative increase in effectiveness) is calculated as follows:

#| relief-threshold-lower,
#| echo = TRUE

#| icer-lower,
#| echo = FALSE

This yields the following table, from which the ICER is calculated as r gbp(icer_lower) \$Can/QALY, close to the published estimate of 60,839 \$Can/QALY.

#| echo = FALSE
knitr::kable(
  es[, c("d1", "Cost", "Utility", "QALY")],
  align = "lrrr",
  digits = c(2L, 2L, 4L, 4L),
  format.args = list(scientific = FALSE)
)