Binary Data Factor Analysis and Multidimensional Latent Trait/Item Response Theory (IRT) Models


Overview

Several advanced methods are available for factor analysis of binary data, including:

  1. Full-information maximum-likelihood estimation of a normal-ogive (Gaussian) multidimensional latent trait/IRT model (Bock, Gibbons & Muraki, 1988).

  2. Factor analysis of the tetrachoric correlations between all item pairs (Knol & Berger, 1991).

  3. The LISCOMP method (Muthen, 1978).

  4. Nonlinear factor analysis (McDonald, 1982).

(These do not include methods based on logistic-ogive or Rasch models, with which I am less familiar.)

Methods 1--3 are theoretically similar; all assume (a) the dichotomous manifest variables are discretized versions of latent continuous variables; and (b) the underlying continuous variables have a multivariate normal distribution. I don't know much about Method 4, but it appears related to the other three methods and, if so, might be expected to produce similar results.

Knol and Berger (1991; also see Parry & Mcardle, 1991) compared methods and basically found that factoring tetrachoric correlations worked as well as other methods. This is helpful since commonly available software such as PRELIS (distributed with LISREL) can be used to calculate a matrix of tetrachoric correlations, and, say, SAS PROC FACTOR can be used to factor the matrix. For a full explanation of this method, including examples, click here.

Besides the methods above, Uebersax (1993) described another approach: first one performs a latent class analysis of the data; then one locates the latent classes in a multidimensional space. This is potentially useful when (a) the assumption of latent multivariate normality is inappropriate; or (b) one wishes to consider the group (latent class) structure of cases as well as data dimensionality.

Recommended Readings

The book by Bartholomew is very helpful; it devotes two chapters to the subject and is perhaps the best summary available. The Knol and Berger (and the Parry & McArdle paper, which is similar) gives a good empirical comparison of different methods. The Takane and de Leeuw paper--more technical and not for every reader--rigorously examines the relationships between different approaches.

Software

Following are programs I know of for factor analysis of binary data and/or multidimensional latent trait modeling.


Software distributor contact information:

Bibliography


Go to the Latent Trait Analysis pages
Go to the Latent Class Analysis pages
Go to the Statistical Methods for Rater Agreement pages

This page maintained by

John Uebersax
jsuebersax@yahoo.com

Revised: 8 July 2000


Copyright (c) by John S. Uebersax. All rights reserved.