Just a quick, slightly critical remark about this: the BIC and the CAIC
are *by definition* pretty much the same, so it shouldn't be interpreted
as some kind of additional confirmation if they point to the same number
of classes, it's rather (more or less) "the same information twice".
Christian
On Tue, 24 May 2011, David Joubert wrote:
I have used PoLCA for this purpose, and not the e1071 package.
You should use a variety of fit indices to choose the number of classes. The
BIC may not always be the best choice, depending on your sample size and
frequency table. In the best case, AIC, CAIC and BIC values agree as to the
optimal number of classes. The Cressie-Read statistic is useful with sparse
tables, but I havent found a way to obtain it in R. If you're a coder there
might be a way to write a function to obtain it.
With poLCA, you can quickly call output for a range a latent classes, and
evaluate solutions. The commands are very simple.
David Joubert
Date: Tue, 24 May 2011 12:30:01 +0100
From: chr...@stats.ucl.ac.uk
To: dan...@umd.edu
CC: r-help@r-project.org
Subject: Re: [R] Latent class analysis, selection of the number of classes
Dear Daniel,
the BIC can be used to estimate the number of classes. This is actually
given out by lca, so you could run lca with several different k and pick
the solution that gives you the best BIC.
Unfortunately I can't tell you whether "large is good" or "small is good"
for the BIC implementation of lca, because there are both versions found
in the literature, BIC with positive and negative sign. (I think that if
there is any standard, then it should rather be "large is good"; you
certainly can check it looking up the values of the loglikelihood and a
definition of the BIC in a book. "Large is good" if the likelihood is
used in the definition with a positive sign.)
With a bit of experimentation it should be able to find out which way
round it is, or you may ask the e1071-maintainer.
Hope this helps (actually I may have missed if somebody responded before),
Christian
On Mon, 23 May 2011, Daniel Malter wrote:
Hi,
I perform latent class analysis on a matrix of dichotomous variables to
create an indicator of class/category membership for each observation. I
would like to know whether there is a function that selects the best fit in
terms of number of classes/categories.
Currently, I am doing this with the lca() function of the e1071 package.
This function requires me to specify the number of classes and to compare
fit statistics for each run of lca. This becomes somewhat cumbersome the
more variables the data matrix contains and, thus, the greater the number of
possible classes is. I was wondering whether there is an alternative
implemented in a different package that does exactly that.
Thanks,
Daniel
--
View this message in context:
http://r.789695.n4.nabble.com/Latent-class-analysis-selection-of-the-number-of-classes-tp3545538p3545538.html
Sent from the R help mailing list archive at Nabble.com.
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
*** --- ***
Christian Hennig
University College London, Department of Statistical Science
Gower St., London WC1E 6BT, phone +44 207 679 1698
chr...@stats.ucl.ac.uk, www.homepages.ucl.ac.uk/~ucakche
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
*** --- ***
Christian Hennig
University College London, Department of Statistical Science
Gower St., London WC1E 6BT, phone +44 207 679 1698
chr...@stats.ucl.ac.uk, www.homepages.ucl.ac.uk/~ucakche
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.