1 Introduction

Author

Sebastian Weber - sebastian.weber@novartis.com

Applied modeling can facilitate interpretation of clinical data and is hence a valuable too for drug development. However, clinical data is very diverse in nature (e.g. endpoints) and brings along various statistical challenges (e.g. longitudinal type data, censoring or missingness). As a result building models may require a great deal of flexibility in terms of the statistical model being applied. Generally models should be kept as simple as possible, but they need to be as complex as needed. For this reason it is of great advantage to any applied modeller to use a flexible modeling framework capable of easily expanding simple models towards more complex models whenever that is required.

The R package brms (Bayesian regression models using Stan) provides a powerful and flexible modeling framework. The package is essentially an easy to use frontend for the Stan statistical modeling platform. This allows brms to leverage the full flexibility of the feature rich statistical modeling language Stan. However, users of brms only need to specify R formulas to define their models. By design brms is written in a modular manner and offers many advanced features. Given the modular design most features can be combined arbitrarly.

While brms (and Stan) is intended primarily for Bayesian analyses neither the use of prior information is strictly required nor is it a requirement to use MCMC sampling (Stan supports penalized maximum likelihood based on L-BFGS). However, using MCMC sampling used for Bayesian problems brings a number of advantages for applied modelers:

Easy-to-communicate interpretation via probabilistic statements about estimands of interest (unknown model parameters, future observations, etc.)
Natural measures of uncertainty: all inference flows from the posterior (no need for large-sample approximations, even for complex nonlinear models)
Good small sample properties including when evaluated by frequentist operating characteristics
Ability to synthesize data from multiple sources
Ability to incorporate weakly-informative prior information reflecting physical constraints
Ability to incorporate informative prior information from historical trials or even experts

However, a key obstacle to wide adoption of Bayesian methods is the availability of and knowledge about easy-to-use software to carry out inference. In the following vignettes of case studies, we introduce the powerful package brms, created by Paul Bürkner, which we believe can help break down this obstacle. The package includes a very flexible syntax for model specification (both likelihood and prior) that can be done entirely in R; these specifications are translated into Stan code, and fit using Stan as a powerful backend for MCMC. One major advantage of brms is that its formula syntax is already familiar to most R users from functions like lm and glm, as well as the popular lme4 package for mixed models.

The purpose of the case studies are to illustrate the wide variety of drug-development applications, which can all be solved using a single R package - brms. The example case studies have been selected to demonstrate specific features of brms as summarized in the overview table below. For colleagues who are not familiar with the package or its syntax, the example code presented here may also be very useful.

This document envisioned as a living document. Please reach out to the team in case you find that some material can be improved or added. In case you have used brms in your work and would like to include the material here, then you are very welcome to contribute to the document!

1.1 Public documentation

There are many resources available online for learning about brms. We list a few:

The homepage for the package, https://paul-buerkner.github.io/brms/
Bürkner, P.-C. (2017). brms: An R Package for Bayesian Multilevel Models Using Stan. Journal of Statistical Software, 80(1), 1–28. https://doi.org/10.18637/jss.v080.i01
Paul-Christian Bürkner, Advanced Bayesian Multilevel Modeling with the R Package. The R Journal (2018) 10:1, pages 395-411. https://journal.r-project.org/archive/2018/RJ-2018-017/index.html

An excellent place to ask questions on brms in the public is the Stan discourse forum.

More useful ressources from the Stan community:

FAQ on cross-validation
Stan runtime warnings and convergence problems explained

--- author: - Sebastian Weber - <sebastian.weber@novartis.com> --- # Introduction {#sec-introduction} ```{r, include = FALSE} source(here::here("src/setup.R")) ``` Applied modeling can facilitate interpretation of clinical data and is hence a valuable too for drug development. However, clinical data is very diverse in nature (e.g. endpoints) and brings along various statistical challenges (e.g. longitudinal type data, censoring or missingness). As a result building models may require a great deal of flexibility in terms of the statistical model being applied. Generally models should be kept as simple as possible, but they need to be as complex as needed. For this reason it is of great advantage to any applied modeller to use a flexible modeling framework capable of easily expanding simple models towards more complex models whenever that is required. The R package `brms` (Bayesian regression models using Stan) provides a powerful and flexible modeling framework. The package is essentially an easy to use frontend for the Stan statistical modeling platform. This allows `brms` to leverage the full flexibility of the feature rich statistical modeling language Stan. However, users of `brms` only need to specify R formulas to define their models. By design `brms` is written in a modular manner and offers many advanced features. Given the modular design most features can be combined arbitrarly. While `brms` (and Stan) is intended primarily for Bayesian analyses neither the use of prior information is strictly required nor is it a requirement to use MCMC sampling (Stan supports penalized maximum likelihood based on L-BFGS). However, using MCMC sampling used for Bayesian problems brings a number of advantages for applied modelers: * Easy-to-communicate interpretation via probabilistic statements about estimands of interest (unknown model parameters, future observations, etc.) * Natural measures of uncertainty: all inference flows from the posterior (no need for large-sample approximations, even for complex nonlinear models) * Good small sample properties including when evaluated by frequentist operating characteristics * Ability to synthesize data from multiple sources * Ability to incorporate weakly-informative prior information reflecting physical constraints * Ability to incorporate informative prior information from historical trials or even experts However, a key obstacle to wide adoption of Bayesian methods is the availability of and knowledge about easy-to-use software to carry out inference. In the following vignettes of case studies, we introduce the powerful package `brms`, created by Paul Bürkner, which we believe can help break down this obstacle. The package includes a very flexible syntax for model specification (both likelihood and prior) that can be done entirely in R; these specifications are translated into [Stan](https://mc-stan.org/) code, and fit using Stan as a powerful backend for MCMC. One major advantage of `brms` is that its formula syntax is already familiar to most R users from functions like `lm` and `glm`, as well as the popular `lme4` package for mixed models. The purpose of the case studies are to illustrate the wide variety of drug-development applications, which can all be solved using a single R package - `brms`. The example case studies have been selected to demonstrate specific features of `brms` as summarized in the overview table below. For colleagues who are not familiar with the package or its syntax, the example code presented here may also be very useful. This document envisioned as a living document. Please reach out to the team in case you find that some material can be improved or added. In case you have used `brms` in your work and would like to include the material here, then you are very welcome to contribute to the document! ## Public documentation There are many resources available online for learning about `brms`. We list a few: * The homepage for the package, [https://paul-buerkner.github.io/brms/](https://paul-buerkner.github.io/brms/) * Bürkner, P.-C. (2017). brms: An R Package for Bayesian Multilevel Models Using Stan. Journal of Statistical Software, 80(1), 1–28. [https://doi.org/10.18637/jss.v080.i01](https://doi.org/10.18637/jss.v080.i01) * Paul-Christian Bürkner, Advanced Bayesian Multilevel Modeling with the R Package. The R Journal (2018) 10:1, pages 395-411. [https://journal.r-project.org/archive/2018/RJ-2018-017/index.html](https://journal.r-project.org/archive/2018/RJ-2018-017/index.html) An excellent place to ask questions on `brms` in the public is the [Stan discourse forum](https://discourse.mc-stan.org). More useful ressources from the Stan community: * FAQ on [cross-validation](https://mc-stan.org/loo/articles/online-only/faq.html) * Stan [runtime warnings and convergence problems explained](https://mc-stan.org/misc/warnings.html) ::: {.content-visible when-profile="dummy"} #### Introduction video from the workshop with Paul Bürkner {{< video videos/brms-1-intro.mp4 >}} :::