Bayesian Exposure-Response Analysis Tools • BayesERtools

BayesERtools provides a suite of tools that facilitate exposure-response analysis using Bayesian methods.

Tutorial (BayesERbook): https://genentech.github.io/BayesERbook/
Package documentation: https://genentech.github.io/BayesERtools/
GitHub repo of the package: https://github.com/genentech/BayesERtools/

Installation

You can install the BayesERtools with:

# install.packages('BayesERtools') # Once on CRAN
devtools::install_github("genentech/BayesERtools") # development version

You also need latest version of rstanemax (>= 0.1.7) to use E_max model.

install.packages('rstanemax', repos = c(ppm = 'https://packagemanager.posit.co/cran/latest'))

Supported model types

	Binary endpoint		Continuous endpoint
	Linear (logit)	Emax (logit)	Linear	Emax
backend	`rstanarm`	`rstanemax`	`rstanarm`	`rstanemax`
reference	🔗	🔗	🔗	🔗
develop model	✅	✅	✅	✅
simulate & plot ER	✅	✅	✅	✅
exposure metrics selection	✅	✅	✅	✅
covariate selection	✅	❌	✅	❌
covariate forest plot	✅	❌	🟡	❌
✅ Available, 🟡 In plan/under development, ❌ Not in a current plan

Quick guide

Here is a quick demo on how to use this package for E-R analysis. See vignette("basic_workflow_bin") for more thorough walk through of a basic workflow.

# Load package and data
library(dplyr)
library(BayesERtools)
ggplot2::theme_set(ggplot2::theme_bw(base_size = 12))

data(d_sim_binom_cov)

# Hyperglycemia Grade 2+ (hgly2) data
df_er_ae_hgly2 <-
  d_sim_binom_cov |>
  filter(AETYPE == "hgly2") |>
  # Re-scale AUCss, baseline age
  mutate(
    AUCss_1000 = AUCss / 1000, BAGE_10 = BAGE / 10,
    Dose = paste(Dose_mg, "mg")
  )

var_resp <- "AEFLAG"

Simple univariable model for binary endpoint

set.seed(1234)
ermod_bin <- dev_ermod_bin(
  data = df_er_ae_hgly2,
  var_resp = var_resp,
  var_exposure = "AUCss_1000"
)
ermod_bin
#> 
#> ── Binary ER model ─────────────────────────────────────────────────────────────
#> ℹ Use `plot_er()` to visualize ER curve
#> 
#> ── Developed model ──
#> 
#> stan_glm
#>  family:       binomial [logit]
#>  formula:      AEFLAG ~ AUCss_1000
#>  observations: 500
#>  predictors:   2
#> ------
#>             Median MAD_SD
#> (Intercept) -2.04   0.23 
#> AUCss_1000   0.41   0.08 
#> ------
#> * For help interpreting the printed output see ?print.stanreg
#> * For info on the priors used see ?prior_summary.stanreg

# Using `*` instead of `+` so that scale can be
# applied for both panels (main plot and boxplot)
plot_er_gof(ermod_bin, var_group = "Dose", show_coef_exp = TRUE) *
  xgxr::xgx_scale_x_log10()

Covariate selection

BGLUC (baseline glucose) is selected while other two covariates are not.

set.seed(1234)
ermod_bin_cov_sel <-
  dev_ermod_bin_cov_sel(
    data = df_er_ae_hgly2,
    var_resp = var_resp,
    var_exposure = "AUCss_1000",
    var_cov_candidate = c("BAGE_10", "RACE", "BGLUC")
  )
#> 
#> ── Step 1: Full reference model fit ──
#> 
#> ── Step 2: Variable selection ──
#> 
#> ℹ The variables selected were: AUCss_1000, BGLUC
#> 
#> ── Step 3: Final model fit ──
#> 
#> ── Cov mod dev complete ──
#> 
ermod_bin_cov_sel
#> ── Binary ER model & covariate selection ───────────────────────────────────────
#> ℹ Use `plot_submod_performance()` to see variable selection performance
#> ℹ Use `plot_er()` with `marginal = TRUE` to visualize marginal ER curve
#> 
#> ── Selected model ──
#> 
#> stan_glm
#>  family:       binomial [logit]
#>  formula:      AEFLAG ~ AUCss_1000 + BGLUC
#>  observations: 500
#>  predictors:   3
#> ------
#>             Median MAD_SD
#> (Intercept) -7.59   0.90 
#> AUCss_1000   0.46   0.08 
#> BGLUC        0.87   0.13 
#> ------
#> * For help interpreting the printed output see ?print.stanreg
#> * For info on the priors used see ?prior_summary.stanreg
plot_submod_performance(ermod_bin_cov_sel)

coveffsim <- sim_coveff(ermod_bin_cov_sel)
plot_coveff(coveffsim)