Plan the sample size of a prospective study for a Bayesian graphical model, given a prior elicited from a previous study.

An informative prior elicited from a previous study carries a quantifiable amount of information, the prior effective sample size. A prospective study should be large enough that its data outweigh that prior. designbgm estimates the prior effective sample size for Bayesian graphical models and plans the prospective sample size accordingly, under a data-to-prior information ratio or a Bayes factor design analysis.

Installation

# install.packages("remotes")
remotes::install_github("Bayesian-Graphical-Modelling-Lab/designbgm")

Usage

Start from the parameters of a previous study: a precision matrix K, its graph G, and the study size nu.

library(designbgm)

params <- ggm_parameters(K, G, nu = 100)
ep     <- elicit_prior(params)

How much information does this elicited prior carry?

prior_ess(ep)

Plan the size of the prospective study. "DPIR" targets a data-to-prior information ratio over the model parameters; "BFDA" targets the Bayes factor power at a representative edge:

plan <- design(ep, method = "DPIR")
plan

Validate the recommendation:

validate(plan)

For simulation studies, simulate_prior_study() generates prior studies from the family likelihood:

st  <- ggm_study(p = 10, nu = 100, structure = "smallworld")
pst <- simulate_prior_study(st, n_studies = 1)

Background

Prior effective sample size (ESS) is a pre-data measure of how much information an elicited prior carries. It matters because it sets a baseline to the planning: a study should collect more than ESS(θ) observations, or the prior determines the posterior. The ESS estimators and the planning methods implemented in this package are described in Arena et al. (2026).

Scope

Currently, the package supports Gaussian graphical models with a Wishart (complete graph) or G-Wishart (sparse graph) prior. The package is structured by family model, and ordinal Markov random fields are planned.

References

• Arena, G., et al. (2026). What is your Prior Worth? Effective Sample Size and Sample Size Planning for Gaussian Graphical Models arXiv. arXiv:2606.22687