trialr and escalation

Summary

This post sets out how two of my R packages, trialr and escalation, will work together and grow in the future.

Background

trialr (Brock 2020b; Brock 2019) is a package I wrote to fit complex Bayesian clinical trial models using Stan. Because I generally work on early phase trials, and because early phase trials are generally predisposed to Bayesian methods, the package ended up containing quite a few dose-finding methods like CRM (O’Quigley, Pepe, and Fisher 1990) and EffTox (Thall and Cook 2004). I initially added code to this package that would aid practical implementation of these methods, like conducting simulation studies and calculating future dose-paths.

Later I introduced the escalation (Brock 2020a) package, containing a more general range of dose-finding methods (i.e. not just those that require MCMC methods for model-fitting). The beauty of escalation is that formalises the interface for dose-finding designs. For instance, all methods recommend a dose through the recommended_dose() function and distinguish between the need to stop or continue via the continue() function. This standard interface allows the elements of designs to become modules that can be arbitrarily combined.

For example, escalation allows you to start with a CRM model, add behaviour to stop when a certain sample size is reached, but ensure that the recommended dose is evaluated in at least \(n\) patients before stopping is permitted, all the while monitoring for excess toxicity at the lowest dose. Because all of the modules obey the same interface, they can be combined in arbitrary ways to tailor the behaviour of designs. The module that stops a trial when a target sample size is reached works whether it is preceded by a CRM design or BOIN design, for instance.

Do we need both?

By the time escalation was released, there was quite a bit of overlap between trialr and escalation. So do we need both? Yes, and I will explain why! Is trialr now redundant? Not at all, but it will become more focused.

escalation went far further than trialr in making dose-finding designs hot-swappable and customisable. It makes sense that inferential tools like running simulations are offered in escalation so that comparing designs is easy.

trialr still has its place, however. It takes care of fitting models to data using MCMC via Stan. This is a reliable way of fitting complex (e.g. non-linear, multi-parameter) models when other methods struggle. Henceforth, trialr will focus only on Stan-based MCMC model-fitting.

escalation will now use trialr alongside dose-finding packages from other authors like dfcrm (Cheung 2013) and BOIN (Yuan and Liu 2018) to support the broadest possible array of dose-escalation methods.

How might this look in future?

One of my near term goals is to add dose-response models to trialr like Emax. Dose-response modelling is a topic close to dose-finding trials but quite distinct from it. When I add the Emax model, the Stan implementation will go in trialr. If the Emax model stands to be useful as a dose-selection tool, then it can be added to support the usual methods in the escalation package.

The situation now

I recently updated both packages on CRAN:

• trialr is now at v0.1.5 at https://CRAN.R-project.org/package=trialr
• escalation is now at v0.1.4 at https://CRAN.R-project.org/package=escalation

escalation now depends on trialr to deliver Stan-MCMC implementations of CRM, EffTox, and the phase I method of Neuenschwander, Branson, and Gsponer (2008).