bayprior

A toolkit for constructing, validating, and justifying Bayesian priors in clinical trial settings. Implements expert elicitation via quantile matching, the roulette method, and moment matching across six distribution families, linear and logarithmic expert pooling, meta-analytic-predictive (MAP) priors derived from historical trials, prior-data conflict diagnostics including the Box p-value (computed from the exact prior predictive distribution where one has a closed form), an exploratory surprisal (S-value) companion to it, surprise index, information divergence, and Mahalanobis distance, sensitivity analyses with tornado and influence heatmap plots, sceptical, robust, and power priors, and automated prior justification reports. Includes a fully modular 'Shiny' application for interactive use. Methods based on O'Hagan et al. (2006, ISBN:9780470029886), Box (1980) <doi:10.2307/2982063>, Greenland (2023) <doi:10.1111/sjos.12625>, Oakley and O'Hagan (2010) <https://tonyohagan.co.uk/shelf/>, Schmidli et al. (2014) <doi:10.1111/biom.12242>, Neuenschwander et al. (2010) <doi:10.1177/1740774509356002>, Roever et al. (2021) <doi:10.1002/jrsm.1475>, Ibrahim and Chen (2000) <doi:10.1214/ss/1009212673>, Spiegelhalter et al. (1994) <doi:10.2307/2983527>.

Releases

5

Needs compilation

No

Needs R

>= 4.1.0

Config/Needs/vignettes

knitr, rmarkdown, rstanarm, brms

Config/testthat/edition

3

Language

en-GB

RoxygenNote

7.3.3

SystemRequirements

Quarto CLI (>= 1.5, https://quarto.org)

• Clinical trial design

Since March 2026

• cli
• config
• dplyr
• DT
• ggplot2
• glue
• golem
• plotly
• purrr
• rlang
• shiny
• shinycssloaders
• shinydashboard
• shinyjs
• shinyWidgets
• stats

Installing it adds

77 packages, 73.8 MB of source

Of those, need compilation

38

• brms
• htmlwidgets
• knitr
• metafor
• patchwork
• quarto
• rmarkdown
• rstanarm
• shinytest2
• spelling
• testthat
• vdiffr
• withr

Lines by kind (18,062)

Lines by language

Languages are counted from file extensions.

Files

138

R functions

124

Object systems

S3: 8 methods

Median

3

Mean

5.77

Highest

36

Median lines per function

29

R functions

124

Calls

172

Unconnected

9

Groups

4

Lines counted

5,945

Lines the tests ran

2,035

Tests passed

Yes

Measured with

covr 3.6.5 on R 4.6.1

Has tests

Yes

Test frameworks

testthat

Test cases

330

Snapshot tests

11

Lines of tests

4,059

Test lines per line of code

0.52

• Introduction to bayprior
• Prior Elicitation Methods
• Prior-Data Conflict Diagnostics
• Regulatory Reporting
• Robust, Sceptical, and Power Priors
• Sensitivity Analysis
• Using bayprior Priors with rstanarm and brms

README

Yes, 1,597 words

Release notes file

Yes

Help pages

36

Lines of documentation

2,054

Commits

303

Open pull requests

0

Tagged releases

16

Latest tagged release

v0.5.0, 21 Oct 2026

In development

Default branch at 0.5.0.9000, ahead of CRAN 0.5.0 (checked 2026-10-25)

Issues closed in

1 day (median)

Pull requests closed in

0 days (median)

Pull requests merged

98%

Created

24 Apr 2026

Licence file reads as

–

Method, repository practices

Across ndohpenngit

Changes to default behaviour

• prior_conflict() now computes box_pvalue (and the s_value derived from it) from the exact prior predictive distribution by default wherever a closed form exists: Beta-Binomial (Beta prior, type = "binary") and Gamma-Poisson / Negative Binomial (Gamma prior, type = "poisson" or "survival"). The exact argument now defaults to NULL (exact where available, silent fallback to the Normal approximation otherwise); exact = TRUE additionally messages on fallback and exact = FALSE reproduces the Normal-approximation results of 0.4.0 and earlier. The exact check requires whole-number counts (and, for binary data, a whole-number sample size); otherwise the Normal approximation is used. $box_pvalue_method reports which was used. Reason: a simulation study (see validation/) found that the Normal approximation, whose standard error is estimated from the observed proportion or count, can reject far more often than its nominal level when there is no prior-data conflict, for informative priors and small samples. For example, for a Beta(5, 45) prior with n = 10 and no conflict, the approximation flagged conflict in about 39% of simulated datasets at the 5% level, because zero observed events (probable under the prior) yield a zero standard error and a small approximate p-value. The exact version's rejection rate stayed at or below the nominal level in all simulated settings. Results for moderate samples and weakly informative priors are essentially unchanged.
• The prior predictive p-value now also covers mixture priors (e.g. robust_prior(), and linear pooling from aggregate_experts()): the predictive is the weighted mixture of the component predictives, exact for Beta (binary) or Gamma (Poisson/survival) components. Other proper priors (Normal, lognormal, Weibull, exponential, or a Normal vague component) are handled by equal-probability quadrature over the prior, truncated to the parameter space; $box_pvalue_method is then "numerical". The definition of the p-value is the same throughout (total predictive probability of outcomes no more probable than the observed one). In the simulation (validation/sim_mixture.R), pooled-consensus mixtures with the Normal approximation rejected up to 43% of the time at the 5% level under no conflict; the prior predictive version never exceeded 5%. For lognormal, Weibull and exponential priors on a rate (validation/ sim_numerical.R, 84 cells) the numerical p-value never exceeded the nominal level either (the approximation reached 63%). At equal size, the prior predictive p-value has the same average power as the old approximation, but less power against downward shifts for skewed priors (validation/sim_power.R; see REPORT_boxp.md for the full comparison).
• New field $box_pvalue_caution on prior_conflict() results: set when the Normal approximation produced the p-value for discrete data with zero or all events, or fewer than 10 events or non-events. Shown in print() and in the app.
• conflict_severity is now classified on the reported box_pvalue: "severe" below 2 * pnorm(-3) (about 0.0027), "mild" below alpha. This equals the earlier rule (surprise_index >= 3) whenever the Normal approximation is used, and avoids mixing an exact p-value with an approximate distance otherwise.
• Shiny app (Conflict module): the "exact prior predictive" option is now on by default; unchecking it uses the Normal approximation throughout.

New features

• plot_sensitivity() and plot_tornado() gain an interactive argument. The default (FALSE) returns a ggplot object exactly as before; interactive = TRUE returns a plotly object with readable tooltips (rounded values and parameter labels instead of raw column names) and, for the tornado plot, an explicit title and subtitle.

Shiny app

• Sensitivity module: plots are interactive with formatted tooltips, the range sliders no longer show tick marks, a note is shown when the prior is a mixture (the analysis uses a moment-matched working prior), and the "Analysis type" and outcome buttons clearly distinguish the selected option.
• Dark mode: plotly plots are now recoloured natively instead of through a CSS colour-inversion filter, so plot colours and hover labels are correct and plots no longer flash light after an analysis is run.