In ICON-in-R/NMA: Easy Network Meta-Analysis

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Introduction

This document describes how to use the main functions of NMA to run a single network meta-analysis with survival data.

Example

First load the required packages.

library(NMA)
library(dplyr)
library(purrr)

Settings

Define the BUGS parameters for MCMC. This is not necessary, but recommended, because there are default values for these.

bugs_params <-
  list(
    PROG = "openBugs",  # which version of BUGS to use to run the MCMC
    N.BURNIN = 10,#00,  # number of steps to throw away
    N.SIMS = 150,#0,    # total number of simulations
    N.CHAINS = 2,       # number of chains
    N.THIN = 1,         # thinning rate
    PAUSE = TRUE)

Define the scenario we will use for the analysis.

RANDOM <- FALSE             # is this a random effects model?
REFTX <- "X"                # reference treatment
data_type <- c("hr_data", "surv_bin_data", "med_data") # which type of data to use
label_name <- "label_name"

Read in datasets

The trials data consist of up to 3 separate data frames. A main table, subData, and optional tables for median event time and binary data, survDataMed and survDataBin respectively. Lets read in the each data set separately. In another article we will show how to do this in one function call by including a Reference file in the data folder which contains the meta data of how to read in the study data. If there is no binary or median data used in the NMA then the variables survDataBin and survDataMed are assigned NA.

file_name <- here::here(file.path("inst", "extdata", "survdata_"))

survDataHR <-
  read.csv(paste0(file_name, "hr_test.csv"),
           header = TRUE,
           as.is = TRUE)

survDataBin <-
  tryCatch(
    read.csv(paste0(file_name, "bin_test.csv"),
             header = TRUE,
             as.is = TRUE),
    error = function(e) NA)

survDataMed <-
  tryCatch(
    read.csv(paste0(file_name, "med_test.csv"),
             header = TRUE,
             as.is = TRUE) %>% 
      mutate(medR = floor(medR)),
    error = function(e) NA)

The format of the data should be fairly self-explanatory and looks like the following. For the hazard ratio data

knitr::kable(survDataHR)

For the binary data

knitr::kable(survDataBin)

and for the median time data

knitr::kable(survDataMed)

More information about these data is available in the help documentation which can be accessed with e.g. help(subData).

Build model

The package uses the model in @Woods2010 for NMA combining count and hazard ratio statistics in multi-arm trials. For count data, the cumulative count of subjects who have experienced an event in arm $k$ of study $s$

$$ r_{s,k} \sim Bin(F_{s,k}, n_{s,k}) $$

A log cumulative hazard for each trial arm

$$ \ln (H_{s,k}) = \ln ( - \ln (1 - F_{s,k}))) $$

where

$$ \ln (H_{s,k}) = \alpha_s + \beta_k - \beta_b $$

Under an assumption of proportional hazards, the $\beta_k$ coefficient is equal to the log hazard ratio.

The counts and hazard ratio data can then be used together, placing a normal distribution error on the log HR.

A random effect is included in the model as follows.

$$ \ln (H_{s,k}) = \alpha_s + \beta_k - \beta_b + re_{s,k} - re_{s,b} $$

Now we can create the NMA object to use in the modelling. The workflow is to first create this separately to actually doing the fitting. This then means that we can perform modified fits but we don't have to redo any of the preparatory work.

nma_model <-
  new_NMA(survDataHR = survDataHR,
          survDataMed = survDataMed,
          survDataBin = survDataBin,
          bugs_params = bugs_params,
          is_random = RANDOM,
          data_type = data_type,
          refTx = REFTX ,
          effectParam = "beta",
          label = "",
          endpoint = "")

nma_model

We can view the network graph.

library(sna)
plotNetwork(nma_model)

Run MCMC

The NMA MCMC function calls the appropriate BUGS model.

nma_res <- NMA_run(nma_model, save = FALSE)

nma_res

http://127.0.0.1:34255/graphics/plot_zoom_png?width=1121&height=900

Reconfigure model

It is simple to modify an existing analysis without repeating the previous steps. For example, we can run the NMA for a random effects rather than a fixed effects model version of the same model.

nma_model2 <-
  NMA_update(nma_model,
             is_random = TRUE)

nma_res2 <- NMA_run(nma_model2, save = FALSE)

nma_res2

Plot and tables

BUGS plots are available for diagnosing the performance.

diagnostics(nma_res2, save = TRUE)

Different NMA tables can be created. They can provide a record of the analysis.

write_data_to_file(nma_model)
write_results_table(nma_model, nma_res)

pairwiseTable(nma_model, nma_res)

Currently available NMA plots are a treatment effect forest plot of posterior samples and a rank probability grid.

txEffectPlot(nma_model, nma_res)
rankProbPlot(nma_model, nma_res)

It's also possible to create all of the BUGS and output plots and table functions together and write to an analysis folder.

nma_outputs(nma_res2, save = TRUE)

References

ICON-in-R/NMA documentation built on Nov. 14, 2023, 10:54 a.m.