scores_MAP: Decompose factor scores by cluster

In Keefe-Murphy/IMIFA: Infinite Mixtures of Infinite Factor Analysers and Related Models

Decompose factor scores by cluster

Description

Takes posterior summaries of the overall factor scores matrix and returns lists of sub-matrices corresponding to the G-cluster MAP partition.

Usage

scores_MAP(res,
           dropQ = FALSE)

Arguments

res

An object of class "Results_IMIFA" generated by get_IMIFA_results.

dropQ

A logical indicating whether columns of the factor scores matrix should be dropped such that the number of columns in each sub-matrix corresponds to the cluster-specific number of factors (if the number of factors is indeed cluster-specific). When FALSE (the default), the number of columns instead remains common to all sub-matrices - given by the largest of the cluster-specific numbers of latent factors. Note that this argument is irrelevant (i.e. always FALSE) for the finite factor methods ("FA", "MFA", "OMFA", and "IMFA").

Details

Under the models in the IMIFA family, there exists only one factor scores matrix. For the finite factor methods, this has dimensions N * Q.

For the infinite factor methods ("IFA", "MIFA", "OMIFA", and "IMIFA"), the factor scores matrix has dimensions N * Qmax, where Qmax is the largest of the cluster-specific numbers of latent factors q_1,\ldots,q_g. Entries of this matrix thus may have been padded out with zero entries, as appropriate, prior to the Procrustes rotation-based correction applied within get_IMIFA_results (thus now these entries will be near-zero).

In partitioning rows of the factor scores matrix into the same clusters the corresponding observations themselves belong to according to the MAP clustering, the number of columns may vary according to the cluster-specific numbers of latent factors (depending on the value of dropQ and the method employed).

Value

For models which achieve clustering, a list of lists (say x) decomposing the posterior mean scores (x$post.eta), the associated variance estimates (x$var.eta) and credible intervals (x$ci.eta), and the last valid sample of the scores (x$last.eta) into lists of length G, corresponding to the MAP clustering, with varying or common numbers of cluster-specific factors (depending on the value of dropQ and the method employed).

For models with only one component, or the "FA" and "IFA" methods, scores cannot be decomposed, and posterior summaries of the scores will be returned unchanged.

Author(s)

Keefe Murphy - <[e-mail]>

See Also

get_IMIFA_results

Examples

data(coffee)
sim <- mcmc_IMIFA(coffee, n.iters=1000)
res <- get_IMIFA_results(sim)

# Examine the single posterior mean scores matrix
res$Scores$post.eta

# Decompose into G matrices, common numbers of columns
eta <- scores_MAP(res)
eta$post.eta

# Allow the number of columns be cluster-specific
scores_MAP(res, dropQ=TRUE)$post.eta

Keefe-Murphy/IMIFA documentation built on Jan. 31, 2024, 2:15 p.m.