This package allows you to compute the singilar Bayesian information criterion as described in Drton and Plummer (2017) for collections of the following model types:
All of these models, excluding Gaussian latent forests, are described in the above paper. For details regardings the use of the sBIC with Gaussian latent forests see Drton et al (2014).
This package makes extensive use of the R.oo
package
(Bengtsson 2003) which allows for the use of some object oriented
principles in R. While not strictly necessary to use this package it may
be helpful to read sections 1 and 2 of Bengstsson (2003) which serve as
an introduction to R.oo.
An important consequence of the use of R.oo
is that
objects in the sBIC
package use call by reference semantics
and are modified by calling their associated methods.
Each collection of models is defined as its own class. As an example for how to use the package we will compute the sBIC for a collection of Gaussian mixture models with at most 8 components:
set.seed(123)
Create an object representing a collection of Gaussian mixture models with at most 8 components in 2 dimensions.
library(sBIC)
gms = GaussianMixtures(maxNumComponents = 8, dim = 2, restarts = 100)
Generate some simulated data, a mixture of 3 bivariate normals.
library(MASS)
n = 175
class = sample(0:2, n, replace = TRUE)
X = (class == 0) * mvrnorm(n, mu = c(0, 0), Sigma = diag(2)) +
(class == 1) * mvrnorm(n, mu = c(2.5, 2.5), Sigma = diag(1.3, 2)) +
(class == 2) * mvrnorm(n, mu = c(-3, 2.5), Sigma = diag(1.2, 2))
Compute the sBIC on the mixture models with the randomly generated data.
sBIC(X, gms)
Notice that the BIC too strongly penalizes the (true) model with 3 components.