automixfit: Automatic Fitting of Mixtures of Conjugate Distributions to a...

In RBesT: R Bayesian Evidence Synthesis Tools

Automatic Fitting of Mixtures of Conjugate Distributions to a Sample

Description

Fitting a series of mixtures of conjugate distributions to a sample, using Expectation-Maximization (EM). The number of mixture components is specified by the vector Nc. First a Nc[1] component mixture is fitted, then a Nc[2] component mixture, and so on. The mixture providing the best AIC value is then selected.

Usage

automixfit(sample, Nc = seq(1, 4), k = 6, thresh = -Inf, verbose = FALSE, ...)

Arguments

sample	Sample to be fitted by a mixture distribution.
Nc	Vector of mixture components to try out (default seq(1,4)).
k	Penalty parameter for AIC calculation (default 6)
thresh	The procedure stops if the difference of subsequent AIC values is smaller than this threshold (default -Inf). Setting the threshold to 0 stops automixfit once the AIC becomes worse.
verbose	Enable verbose logging.
...	Further arguments passed to mixfit, including type.

Details

The type argument specifies the distribution of the mixture components, and can be a normal, beta or gamma distribution.

The penalty parameter k is 2 for the standard AIC definition. Collet (2003) suggested to use values in the range from 2 to 6, where larger values of k penalize more complex models. To favor mixtures with fewer components a value of 6 is used as default.

Value

As result the best fitting mixture model is returned, i.e. the model with lowest AIC. All other models are saved in the attribute models.

References

Collet D. Modeling Survival Data in Medical Research. 2003; Chapman and Hall/CRC.

Examples

# random sample of size 1000 from a mixture of 2 beta components
bm <- mixbeta(beta1=c(0.4, 20, 90), beta2=c(0.6, 35, 65))
bmSamp <- rmix(bm, 1000)

# fit with EM mixture models with up to 10 components and stop if
# AIC increases
bmFit <- automixfit(bmSamp, Nc=1:10, thresh=0, type="beta")
bmFit

# advanced usage: find out about all discarded models
bmFitAll <- attr(bmFit, "models")

sapply(bmFitAll, AIC, k=6)

RBesT documentation built on May 29, 2024, 10:40 a.m.