Bayesian Power Analysis Using 'brms' and 'INLA'.

powerbrmsINLA

Overview

powerbrmsINLA provides tools for Bayesian power analysis and assurance calculations using the statistical frameworks of brms and INLA.

It includes simulation-based approaches, support for multiple decision rules (direction, threshold, ROPE, Bayes factors, precision), sequential and two-stage adaptive designs, and a comprehensive suite of visualisation functions.

What's New in 1.3.0

• Sequential Bayesian analysis module: sequential_design() for prespecifying a sequential analysis (with an MD5 fingerprint of all decision-relevant fields for preregistration), sequential_analysis() for interim monitoring of real accumulating data with an auditable decision trail, plot_sequential_monitor() for trajectory plots, and brms_inla_sequential_trial() for simulating a sequential design's operating characteristics (stopping probabilities, expected sample size, early-stop exaggeration).
• Validation hardening: brms_inla_power() now raises an error when effect_name does not match a formula-level fixed-effect term and the built-in data generator is in use (previously such a name was silently ignored).
• decide_sample_size() in conditional mode now requires at least one decision target, and no longer mistakes the per-cell SD-moment summary columns for effect-grid columns.
• Breaking change: in brms_inla_power_sequential() summaries, the column previously named assurance is now conditional_power (the old name was statistically misleading).
• Bug fix: brms_inla_power_two_stage() no longer errors when called with default error_sd / obs_per_group.
• All engines now fail early with an informative message when INLA is absent.

What's New in 1.2.0

• Unconditional Bayesian assurance via compute_assurance() — averages conditional power over a design prior on the effect size (O'Hagan & Stevens, 2001).
• assurance_prior_weights() for constructing normalised design-prior weights (normal, uniform, beta) over an effect grid.
• decide_sample_size() with both assurance mode (design prior) and conditional mode for recommending sample sizes from simulation output.
• validate_inla_vs_brms() for spot-checking INLA posterior estimates against brms/Stan.
• brms-to-INLA prior translation with full audit trail — specify analysis priors using brms::prior() syntax.
• Variance uncertainty integration — error_sd and group_sd now accept distributional specifications (halfnormal, lognormal, uniform) so that power is integrated over variance uncertainty.
• Marginal-likelihood Bayes factors (bf_method = "marglik") alongside the existing Savage-Dickey method.
• Automatic INLA thread detection when inla_num_threads = NULL.
• 15 new plotting functions for assurance, Bayes factor, decision-rule, precision, and multi-effect visualisation.
• Print methods for brms_inla_power, powerbrmsINLA_assurance, and powerbrmsINLA_sample_size objects.

See NEWS.md for the full changelog.

Installation

Install from CRAN:

install.packages("powerbrmsINLA")

INLA is listed under Suggests and must be installed separately:

if (!requireNamespace("INLA", quietly = TRUE)) {
  install.packages(
    "INLA",
    repos = c(getOption("repos"),
              INLA = "https://inla.r-inla-download.org/R/stable"),
    dep = TRUE
  )
}

To install the development version from GitHub:

# install.packages("remotes")
remotes::install_github("Tony-Myers/powerbrmsINLA")

Quick Example

library(powerbrmsINLA)

# Step 1: Conditional power simulation
results <- brms_inla_power(
  formula      = y ~ treatment,
  effect_name  = "treatment",
  effect_grid  = c(0.2, 0.5, 0.8),
  sample_sizes = c(50, 100),
  nsims        = 50,
  seed         = 123
)
results$summary

# Step 2: Unconditional assurance (new in 1.2.0)
assurance <- compute_assurance(
  results,
  prior_weights = list(dist = "normal", mean = 0.5, sd = 0.2),
  metric = "direction"
)
print(assurance)

# Step 3: Sample size recommendation
decide_sample_size(
  results,
  direction = 0.80,
  prior_weights = list(dist = "normal", mean = 0.5, sd = 0.2)
)

Model Complexity Considerations

For optimal performance:

• Simple to moderate models: All sample sizes supported.
• Complex random effects (e.g., (1 + time | subject)): Recommend n >= 50 subjects.
• Large effect grids: Consider starting with fewer simulations (nsims = 50-100) for initial exploration, or use the sequential/two-stage engines.

Citation

If you use powerbrmsINLA in published work, please cite:

Myers, T. (2026). powerbrmsINLA: Bayesian Power Analysis Using 'brms' and 'INLA'. R package version 1.2.0. https://cran.r-project.org/package=powerbrmsINLA

License

This package is released under the MIT License. See the LICENSE file for details.