-
Linear Latent Structure Analysis: from Foundations to Algorithms and Applications
Authors:
I. Akushevich,
M. Kovtun,
A. I. Yashin,
K. G. Manton
Abstract:
A new statistical technique for constructing linear latent structure (LLS) models from available data, supported by well established theoretical results and an efficient algorithm, is presented. The method reduces the problem of estimating LLS model parameters to a sequence of linear algebra problems. This assures a low computational complexity and an ability to handle large scale data that invo…
▽ More
A new statistical technique for constructing linear latent structure (LLS) models from available data, supported by well established theoretical results and an efficient algorithm, is presented. The method reduces the problem of estimating LLS model parameters to a sequence of linear algebra problems. This assures a low computational complexity and an ability to handle large scale data that involve thousands of variables. An overall computational scheme and all its components are discussed in detail. Simulation experiments demonstrate the excellent performance of the algorithm in reconstructing model parameters. Step-by-step analysis of a demographic survey is presented as an example.
The technique is useful for the analysis of high-dimensional categorical data (e.g., demographic surveys, gene expression data) where the detection, evaluation, and interpretation of a underlying latent structure are required.
△ Less
Submitted 18 October, 2005; v1 submitted 16 August, 2005;
originally announced August 2005.
-
Linear Latent Structure Analysis: Mixture Distribution Models with Linear Constraints
Authors:
Mikhail Kovtun,
Igor Akushevich,
Kenneth G. Manton,
H. Dennis Tolley
Abstract:
A new method for analyzing high-dimensional categorical data, Linear Latent Structure (LLS) analysis, is presented. LLS models belong to the family of latent structure models, which are mixture distribution models constrained to satisfy the local independence assumption. LLS analysis explicitly considers a family of mixed distributions as a linear space and LLS models are obtained by imposing li…
▽ More
A new method for analyzing high-dimensional categorical data, Linear Latent Structure (LLS) analysis, is presented. LLS models belong to the family of latent structure models, which are mixture distribution models constrained to satisfy the local independence assumption. LLS analysis explicitly considers a family of mixed distributions as a linear space and LLS models are obtained by imposing linear constraints on the mixing distribution. LLS models are identifiable under modest conditions and are consistently estimable. A remarkable feature of LLS analysis is the existence of a high-performance numerical algorithm, which reduces parameter estimation to a sequence of linear algebra problems. Preliminary simulation experiments with a prototype of the algorithm demonstrated a good quality of restoration of model parameters.
△ Less
Submitted 1 July, 2005;
originally announced July 2005.
-
A New Efficient Algorithm for Construction of LLS Models
Authors:
Mikhail Kovtun,
Igor Akushevich,
Kenneth G. Manton,
H. Dennis Tolley
Abstract:
We present a new efficient algortithm for construction of linear latent structure (LLS) models. This algorithm reduces a problem of estimation of model parameters to a sequence of problems of linear algebra, which assures a low computational complexity and ability to handle on desktop computers data that involve up to thousands of variables.
We present a new efficient algortithm for construction of linear latent structure (LLS) models. This algorithm reduces a problem of estimation of model parameters to a sequence of problems of linear algebra, which assures a low computational complexity and ability to handle on desktop computers data that involve up to thousands of variables.
△ Less
Submitted 1 July, 2005;
originally announced July 2005.
-
Grade of Membership Analysis: One Possible Approach to Foundations
Authors:
Mikhail Kovtun,
Igor Akushevich,
Kenneth G. Manton,
H. Dennis Tolley
Abstract:
Grade of membership (GoM) analysis was introduced in 1974 as a means of analyzing multivariate categorical data. Since then, it has been successfully applied to many problems. The primary goal of GoM analysis is to derive properties of individuals based on results of multivariate measurements; such properties are given in the form of the expectations of a hidden random variable (state of an indi…
▽ More
Grade of membership (GoM) analysis was introduced in 1974 as a means of analyzing multivariate categorical data. Since then, it has been successfully applied to many problems. The primary goal of GoM analysis is to derive properties of individuals based on results of multivariate measurements; such properties are given in the form of the expectations of a hidden random variable (state of an individual) conditional on the result of observations.
In this article, we present a new perspective for the GoM model, based on considering distribution laws of observed random variables as realizations of another random variable. It happens that some moments of this new random variable are directly estimable from observations. Our approach allows us to establish a number of important relations between estimable moments and values of interest, which, in turn, provides a basis for a new numerical procedure.
△ Less
Submitted 22 March, 2004;
originally announced March 2004.
