Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Identifiability of Diagnostic Classification Models

Guanhua Fang, Jingchen Liu, Zhiliang Ying

TL;DR

The paper develops a general identifiability theory for diagnostic classification models within a latent class framework, addressing identifiability of item parameters, attribute distributions, and the item-specific partial information structure encoded by the Q-matrix. It leverages Kruskal-type arguments and T-matrix rank conditions to derive sufficient identifiability results and introduces a Bayesian estimation framework with a Dirichlet process–style stick-breaking prior to recover an unbounded latent class structure; partial information is inferred via clustering of item-response probabilities. Through extensive simulations across NIDA, NC-RUM, and LCDM settings and a real NESARC Social Phobia analysis, the authors demonstrate consistent estimation and successful reconstruction of qualitative item-attribute relationships, supporting reliable diagnostic inference. The work provides a unified, model-parameter-agnostic foundation for identifiability in DCMs and offers practical estimation tools for uncovering the underlying attribute structure without pre-specifying the number of latent classes.

Abstract

This paper establishes fundamental results for statistical inference of diagnostic classification models (DCM). The results are developed at a high level of generality, applicable to essentially all diagnostic classification models. In particular, we establish identifiability results of various modeling parameters, notably item response probabilities, attribute distribution, and Q-matrix-induced partial information structure. Consistent estimators are constructed. Simulation results show that these estimators perform well under various modeling settings. We also use a real example to illustrate the new method. The results are stated under the setting of general latent class models. For DCM with a specific parameterization, the conditions may be adapted accordingly.

On the Identifiability of Diagnostic Classification Models

TL;DR

The paper develops a general identifiability theory for diagnostic classification models within a latent class framework, addressing identifiability of item parameters, attribute distributions, and the item-specific partial information structure encoded by the Q-matrix. It leverages Kruskal-type arguments and T-matrix rank conditions to derive sufficient identifiability results and introduces a Bayesian estimation framework with a Dirichlet process–style stick-breaking prior to recover an unbounded latent class structure; partial information is inferred via clustering of item-response probabilities. Through extensive simulations across NIDA, NC-RUM, and LCDM settings and a real NESARC Social Phobia analysis, the authors demonstrate consistent estimation and successful reconstruction of qualitative item-attribute relationships, supporting reliable diagnostic inference. The work provides a unified, model-parameter-agnostic foundation for identifiability in DCMs and offers practical estimation tools for uncovering the underlying attribute structure without pre-specifying the number of latent classes.

Abstract

This paper establishes fundamental results for statistical inference of diagnostic classification models (DCM). The results are developed at a high level of generality, applicable to essentially all diagnostic classification models. In particular, we establish identifiability results of various modeling parameters, notably item response probabilities, attribute distribution, and Q-matrix-induced partial information structure. Consistent estimators are constructed. Simulation results show that these estimators perform well under various modeling settings. We also use a real example to illustrate the new method. The results are stated under the setting of general latent class models. For DCM with a specific parameterization, the conditions may be adapted accordingly.

Paper Structure

This paper contains 17 sections, 7 theorems, 47 equations, 17 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Latent class models and diagnostic classification models
  4. Examples of diagnostic classification models
  5. Identifiability of model parameters
  6. On the identifiability of diagnostic classification models
  7. Identifiability of item parameters and the attribute distribution
  8. Identifiability of partial information structure
  9. Estimation via a latent class model with Dirichlet allocation
  10. Simulation study
  11. NIDA model simulation
  12. Reduced NC-RUM model simualtion
  13. LCDM model simulation
  14. Real data analysis
  15. Discussion
  16. ...and 2 more sections

Key Result

Theorem 1

We consider the general setting of a latent class model with $M>2$ latent classes. The responses are binary and take values in $\{0,1\}$. For each item $j$, let $p_{j\boldsymbol \alpha} = P(Y_j= 1 | \boldsymbol \alpha)$. Let $\pi_{\boldsymbol \alpha}$ be the probability of each latent class. Suppose Then, the item parameters $p_{j\boldsymbol \alpha}$ and the latent class population $\pi_{\boldsymb

Theorems & Definitions (13)

  • Example 1: DINA model, Junker
  • Example 2: NIDA model
  • Example 3: Reduced NC-RUM model
  • Example 4: DINO model
  • Example 5: C-RUM model
  • Definition 1
  • Theorem 1
  • Corollary 1
  • Theorem 2
  • Theorem 3
  • ...and 3 more