MCMCbasedBayesianMS.jl

This package provides a sampling-based implementation of the Bayesian model-selection method of Stephan et al. (2009) in NeuroImage. Earlier versions of this code were used for the analyses in Modirshanechi et al. (2025a) and Modirshanechi et al. (2025b).

Theory

The theoretical details behind the implementation of the package are provided in theory.pdf.

Installation

using Pkg; Pkg.add(url = "https://github.com/modirshanechi/MCMCbasedBayesianMS.jl", rev="main")

Usage

using MCMCbasedBayesianMS

# generating a synthetic data set
#   20 subjects * 2 models
#   LMatrix[n,m] = log-prob of subject n under model m
LMatrix = (rand(20,2) .- 1) * 10;
#   assigning the first 15 to model 1 and the rest to model 2
LMatrix[1:15,1] .+= 10; LMatrix[16:20,2] .+= 10; 

# Bayesian Model Selection (BMS) object with 10 chains and 100K samples per chain
BMS = BMSObj(LMatrix, 
        bms_basechain = BMSBase(n_chains = 10, n_samples = 100000),
        bms_borsamp = BMSBorSampNaive(n_samples = 100000))

# performing inference
InfResults = doBMSinference(BMS)

# plotting the statistics
plotstats(InfResults)

Expected outcome:

# plotting the density of model frequency for different chains
plotrsamples(InfResults,figsize=(500,300))

Expected outcome: