ShinyEQBayes

Introduction

ShinyEQBayes (or EQBayes) is an R Shiny app for combining respondent EQ5D utility values and EQVAS scores to get a joint utility measure using Markov Chain Monte Carlo (MCMC). The modelling in the app is based on this Value in Health paper. It uses the nimble package to turn WinBUGS code, which the original paper used, into C++ code, making computation much faster.

Why use EQBayes?

There are two main reasons for using this methodology to combine EQ5D utility and EQVAS.

1. There are prior beliefs that the sample population is somewhat different to the population norms being used to translate EQ5D answers into utility values. If responses from the EQVAS are notably different to the derived utility values, this may be the case.
2. To reduce interval widths of the health utility variable.

Running the app offline

In some cases, for example if you’re experiencing crashing or want to avoid uploading information to the server, it may be easier to run the app on your local machine. To do this, simply open RStudio and install the necessary packages if not already installed, then use the shiny package to load the app from GitHub.

# Install required packages
install.packages(c("shiny", "shinydashboard", "shinycssloaders", "nimble", "tidybayes", "dplyr", "ggplot2", "here", "renv"))

# Run locally from Git
shiny::runGitHub("robinblythe/ShinyEQBayes")

How to use the app

Access the hosted app at https://aushsi.shinyapps.io/ShinyEQBayes/, or run it on your local machine. There is a brief introduction on the first panel. You will need a .csv file with at least two columns: the EQ5D health utility values (usually ranging from -1 to 1) and the EQVAS values (from 0 to 100). If you have a binary variable of interest that you want to explore, for example treatment or respondent sex, you can use this to stratify your results.

1. Upload your .csv file in the “Upload” sidebar tab.
   • Select the columns denoting your utility values, VAS, and binary variable, if using.
   • Check that the file upload details are correct, and inspect the dataset at the bottom of the tab.
2. Inspect your uploaded data in the “Explore” sidebar tab.
   • Check that the histograms look appropriate, and note the spread of the variables.
   • Use the scatterplot of paired EQ5D utility/VAS values at the bottom to visually inspect correlation. The notches in the axes are a rug plot, representing the marginal distributions of the variables. These variables will be used to construct the joint distribution of the adjusted utility value.
3. Select your MCMC model inputs. A sensible default based on the paper above has been selected, but this can be modified.
   • Number of posterior draws per chain refers to how many samples of the posterior distribution you want for the final model. Extra samples are comparably inexpensive to run with C++, but don’t go overboard. Between 5000-10000 per chain is probably sufficient, but we can come back to this later.
   • Chains represent how many times you want to re-run the model. Because MCMC is a random process, it is recommended to use at least 2 Chainz as they may explore different parts of the posterior distribution. A maximum of 4 chains has been set for this app. Each chain is relatively expensive to run so don’t go overboard.
   • Thinning refers to how many iterations to skip during the sampling process before a posterior draw is taken. This smooths out some variability from the sample, but there is debate around whether this is desirable. The default is 3.
   • Number of iterations to burn for warmup is an important input, because the samples from MCMC can take a while to explore the distribution space. Using a burn-in period reduces this unnecessary variation and is highly recommended.
   • Setting your seed guarantees reproducibility. It allows other researchers to check on the validity of your work and know whether your results are just a result of re-running your code until you get favourable results. Select a consistent number each proejct when using the app for research.
4. Hit run model without (intercept only) or (with X variable) and wait for the results to appear on the “Analyse” tab. You may need to switch panes. There should be a progress bar in the bottom right indicating how far along you are. It is unlikely that the app will freeze while in use, so please be patient. You can’t do anything except view the already-rendered plots while the models are running.
5. You can see your results, including the density plots of the posterior samples of your new joint distribution compared to the old ones, in the “Analyse” pane. A table underneath the plots shows the summary details of your results - pay particular attention to the ‘Width’ part of the panel to see the reduction in uncertainty from the joint distribution. If you want to see trace plots, the “Diagnostics” pane shows these for each parameter.
   • This is where the extra iterations come in. If the chains look densely populated around one Y value without much wandering, this shows that the algorithm has probably converged. If not, you probably need to increase the number of iterations.
   • For more complex models, there is a tradeoff between the stability of your results and computation time, so more complex models usually should be run for longer to avoid variability in your results. In this app, however, the models are quite simple and should all converge quickly.

Happy hunting!