FITTING A MULTIDIMENSIONAL COMPONENT MODEL TO BINARY DATA.
Abstract
A solution to factoring binary data matrices is proposed by weakening the symmetric-least-squares error function of component analysis to requiring only monotonicity over individuals between item responses and reproduction from the analysis. Two classes of measures of monotonicity are considered: one analogous to regression of the observations on the model (class I) and the other analogous to regression of the model on the observations (class II). Several computing procedures were tried: a gradient-oriented procedure which led to local minima and a refactoring procedure which led to general minima for class II functions but not for class I functions. Solutions for the Class II functions indicated existence of undesirable degenerate results. Suggestions are made for further modified computing procedures for class I functions. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1968
- Accession Number
- AD0677294
Entities
People
- Raymond Frederick Koopman Jr
Organizations
- University of Illinois Urbana–Champaign