Random Effects Latent Class Analysis

Fits standard and random effects latent class models. The single level random effects model is described in Qu et al and the two level random effects model in Beath and Heller . Examples are given for their use in diagnostic testing.


News

Changes in randomLCA version 1.0-15

o corrected exports to allow some methods to be called

Changes in randomLCA version 1.0-14

o added method to obtain class with maximum posterior probability

Changes in randomLCA version 1.0-13

o change citation to refer to Journal of Statistical Software article

o added extra examples

Changes in randomLCA version 1.0-12

o add option to obtain outcome probabilities on either the probability or transformed scale

o add extra documentation to postClassProbs method

Changes in randomLCA version 1.0-11

o add vignette of Journal of Statistical Software article (without logos) and code

o added registration for C routines

Changes in randomLCA version 1.0-10

o remove test of best minimum loglikelihood as this was producing too many unnecessary error messages

o incorporate penalised likelihood into EM algorithm

o changes to starting values to improve speed

o added parameter to allow choice of convergence tolerance for EM algorithm

Changes in randomLCA version 1.0-9

o removed reliance on SciencePo package by implementing Dirichlet density directly

o fixed minor bug in calculating penalty

o changed default penalty to 0.01 increasing speed and with only very minor change in estimates

o various modifications to code to improve speed

o avoid error messages due to problems in inverting Hessian in certain cases

Changes in randomLCA version 1.0-8

o Removed spaces after doi: in description file

Changes in randomLCA version 1.0-7

o Improved speed of random effects models

o Fixed bug with 2 level models with constant loadings which caused crashing

Changes in randomLCA version 1.0-6

o Added warning if only one set of starting values produced the maximum value of the maximum likelihood

o Changed likelihood penalty from product of beta to Dirichlet to match the Galindo Garre paper

o Changed default penalty to 0.001, which produces neglible effect except to remove some numerical instabilities

o Changed most of the Depends to Imports which removes the messages on loading

Changes in randomLCA version 1.0-5

o Added argument to refit method to use the estimates from current model for the initial values

o Changed outcomeProbs to use the initial values from the current model when bootstrapping. This avoids problems with label switching.

Changes in randomLCA version 1.0-4

o Added check for rank of Hessian and give warning that model may not be identified or there may be boundary estimates.

o Improve starting values for random effects models

o For refit now totally redetermine starting values rather than using those from model, as this produces more reliable results

Changes in randomLCA version 1.0-3

o When selecting initial starting values only perform the EM algorithm to choose the best starting values, and then use the quasi-Newton to perform the final optimisation to increase speed.

o To avoid possible problems removed "." from some method names. Changed class.probs to classProbs, post.class.probs to postClassProbs, outcome.probs to outcomeProbs, calc.cond.prob to calcCondProb, calc.cond2.prob to calcCond2Prob, calc.marg.prob to calcMargProb

o add AIC3 method to calculate AIC with a penalty of 3, and added this to the summary

o reduced number of S3 methods to only those that are necessary

o added print method for randomLCA objects that just calls summary

o with plots use outcome labels from the data

o improved graphing code to incorporate better labels and allow these to be overriden from the plot command

o added penalty parameter to AIC to conform to stats arguments

Changes in randomLCA version 1.0-2

o fixed bug that caused some non-dentifiable models to be missed

Changes in randomLCA version 1.0-1

o improvements to help documentation

o added check for non-identifiability of model

o include patterns with zero frequency in fitting if supplied - this increases fitting time but gives the predictions for these patterns

o fixed errors in examples vignette

Changes in randomLCA version 1.0-0

o changed default graph to marginal for random effects models

o removed graphtype of both

Changes in randomLCA version 0.9-0

o there have been major changes in this version

o added constload parameter to specify that loadings are constant for each outcome and modified use of blocksize parameter so it now refers only to size of level2 units or repeated outcomes in single level models.

o reduce stepmax in nlm to reduce chance of not finding true maximum

o use routine from fastGHQuad for Gauss-Hermite quadrature

o remove initmodel parameter

o changed fitting algorithm for both random effects models to use generalised EM algorithm.

o tidy up namespace requirements.

o change default method for standard errors with outcome.probs to normal bootstrap with 50 simulations.

o increased the quadrature points in the calculation of marginal probabilities, just to be safe. This should be converted to use adaptive quadrature.

o change name of class.probs to post.class.probs to better reflect what it does, and add class.probs function to return overall probability of each class

Changes in randomLCA version 0.8-7

o moved vignettes to correct folder

o added imports from lattice and boot to fix error message in check

Changes in randomLCA version 0.8-6

o passes all CRAN tests

Changes in randomLCA version 0.8-5

o reduced examples execution time

Changes in randomLCA version 0.8-4

o changed fitting parameters fro examples to fix convergence problem on SPARC machines

Changes in randomLCA version 0.8-3

o remove patterns when supplied frequencies are zero

o fixed documentation of symptoms dataset

Changes in randomLCA version 0.8-2

o output both standard an penalised log likelihood and added penalty parameter

Changes in randomLCA version 0.8-1

o simplified code

o fixed bug which caused stop when using verbose=TRUE

o added parameter to control the number of quasi-Newton steps between each adaptive step, and reduced the default to 5

o restricted the outcome probabilities using a penalty, to prevent numerical problems

Changes in randomLCA version 0.7-5

o improved output formatting in summary

o more information in vignette

Changes in randomLCA version 0.7-4

o fixed problems with import of AIC and BIC

Changes in randomLCA version 0.7-3

o added bibliography to vignette

o changes to documentation to improve readability

o added check for valid patterns must be 0 or 1

Changes in randomLCA version 0.7-2

o changes to documentation to improve readability

Changes in randomLCA version 0.7-1

o added simulate method to produce random data from fitted models

o changed default calculation of standard errors to true

o added bootstrap estimation of confidence intervals for outcome probabilities

o removed most of the warnings about generated NaN during fitting

o checked for excessive number of adaptive iterations

Changes in randomLCA version 0.6-2

o increased speed of fitting for standard latent class models

o added Gender Roles dataset

o updated vignette to allow for label switching

Changes in randomLCA version 0.6-1

o added an accessor for outcome probabilities with confidence intervals

o standard latent class now allows probit (this obviously wont affect the outcome probabilities but will change the confidence intervals)

o allowable quadrature points increased to 190

o changed handling of random seeds for random starting values

Changes in randomLCA version 0.5-3

o added marginal outcome probabilities and link to summary

o added accessors for logLik, AIC, BIC

o changed log.Lik to logLik

Changes in randomLCA version 0.5-2

o split summary to have separate print method

o in summary print names for loadings

Changes in randomLCA version 0.5-1

o added new example to vignette

o improved summary function

o increased speed of multilevel models by recode in C

o fixed bug with returned outcoem probabilities in multilevel models

Changes in randomLCA version 0.3-2

o changes to vignette to allow running on some systems

Changes in randomLCA version 0.3-1

o recode of algorithm for single level random effects in C

o for standard LC always perform Quasi-Newton to guarantee convergence

Changes in randomLCA version 0.2-1

o fixed bug which could cause very occaisonal crash when outcome probabilities approached 0 or 1

o recode of EM algorith for standard LC models in C, so very much faster

Changes in randomLCA version 0.1-4

o fixed bug in adaptive random by class model fit

Changes in randomLCA version 0.1-3

o made vignette creation faster

Changes in randomLCA version 0.1-2

o fixed bug where convergence failure wasn't identified for some models, resulting in infinite loops

o fixed bug causing outcome probabilities of 0 or 1 not to be plotted

o included vignette for examples

First realease version 0.1-1

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.

install.packages("randomLCA")

1.0-16 by Ken Beath, 4 months ago


Browse source code at https://github.com/cran/randomLCA


Authors: Ken Beath [aut, cre]


Documentation:   PDF Manual  


Task views: Cluster Analysis & Finite Mixture Models, Psychometric Models and Methods


GPL (>= 2) license


Imports boot, fastGHQuad, Matrix, Rfast, parallel

Depends on lattice

Suggests R.rsp


See at CRAN