Hostname: page-component-7b7567c8c-2n9c9 Total loading time: 0 Render date: 2026-11-15T14:26:16.968Z Has data issue: false hasContentIssue false

A Copula-Based Joint Model for Speed and Accuracy

Published online by Cambridge University Press:  16 October 2026

Sunbeom Kwon
Affiliation:
Psychology, University of Illinois Urbana-Champaign , USA
Susu Zhang*
Affiliation:
Psychology; Statistics, University of Illinois Urbana-Champaign , Champaign, USA
*
Corresponding author: Susu Zhang; Email: szhan105@illinois.edu
Rights & Permissions [Opens in a new window]

Abstract

This study proposes a copula-based model that relaxes the multivariate normality assumption among latent traits in the joint modeling of multimodal data, such as item response and response times. In the structural model, the joint distribution of latent traits is defined via a copula with tail dependence. An expectation–maximization algorithm with stochastic approximation is implemented for parameter estimation. A simulation study evaluates the proposed method’s accuracy in ability estimation under various conditions, comparing it to traditional approaches. By better capturing the dependence structure between latent ability and speed, the proposed method outperforms the traditional approach with multivariate normality assumption. Applied to a variety of empirical datasets, the proposed method’s flexibility and the improvement in model fit are illustrated. The copula-based latent variable modeling can be generalized to broader applications in psychological testing involving multivariate latent traits.

Information

Type
Theory and Methods
Creative Commons
Creative Common License - CCCreative Common License - BYCreative Common License - NCCreative Common License - SA
This is an Open Access article, distributed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike licence (https://creativecommons.org/licenses/by-nc-sa/4.0), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the same Creative Commons licence is used to distribute the re-used or adapted article and the original article is properly cited. The written permission of Cambridge University Press or the rights holder(s) must be obtained prior to any commercial use.
Copyright
© The Author(s), 2026. Published by Cambridge University Press on behalf of Psychometric Society
Figure 0

Figure 1Bivariate distributions constructed with copula functions and standard normal margins.Note: X and Y were simulated using a bivariate copula with standard normal marginal distributions. The red contours represent the bivariate density, and the dots indicate the randomly simulated data points.Figure 1 long description.

Figure 1

Table 1Examples of copula functionsTable 1 long description.

Figure 2

Figure 2Bivariate distributions of latent traits under rotated copulas.Note: The marginal distributions of θ$\theta $ and τ$\tau $ are standard normal, and the joint distributions are defined using rotated copulas.Figure 2 long description.

Figure 3

Figure 3Bivariate distributions of latent ability and speed under different dependence conditions.Figure 3 long description.

Figure 4

Table 2True item parameter values in the simulationTable 2 long description.

Figure 5

Table 3Marginal likelihood-based model selection resultTable 3 long description.

Figure 6

Table 4Parameter recovery under the Gumbel copula dependence conditionTable 4 long description.

Figure 7

Table 5Accuracy of ability estimates under the Gumbel copula dependence conditionTable 5 long description.

Figure 8

Figure 4Conditional RMSE of ability estimates under the Gumbel copula dependence condition.Figure 4 long description.

Figure 9

Figure 5Conditional bias of ability estimates under the Gumbel copula dependence condition.Figure 5 long description.

Figure 10

Table 6Accuracy of ability estimates under the MVN dependence conditionTable 6 long description.

Figure 11

Figure 6Conditional RMSE of ability estimates under the MVN dependence condition.Figure 6 long description.

Figure 12

Table 7Descriptive statistics of the empirical datasetsTable 7 long description.

Figure 13

Table 8Model selection results with log-likelihood values from the empirical data analysesTable 8 long description.

Figure 14

Figure 7Scatter plots of θ$\theta $ and τ$\tau $ estimates from empirical data analyses.Note: Each row displays latent trait estimates obtained from the MVN model, the 90∘$90^\circ $-rotated Joe copula, and the 270∘$270^\circ $-rotated Gumbel copula.Figure 7 long description.

Figure 15

Figure 8θ$\theta $ and τ$\tau $ estimates of MVN and 90∘$90^\circ $-rotated Joe from the 2018 PISA Science Data, Cluster 1, USA.Figure 8 long description.

Figure 16

Figure 9Squared and raw differences in θ$\theta $ estimates between MVN and 90∘$90^\circ $-rotated Joe from the 2018 PISA Science Data, Cluster 1, USA.Figure 9 long description.

Figure 17

Figure 10θ$\theta $ and τ$\tau $ estimates of MVN and 90∘$90^\circ $-rotated Joe from the 2018 PISA Math Data, Cluster 3, JPN.Figure 10 long description.

Figure 18

Figure 11IMV values from the 2018 PISA Math and Science Data.Figure 11 long description.

Figure 19

Figure 12Person-level predicted accuracy versus observed accuracy from the 2018 PISA Math and Science Data.Figure 12 long description.

Figure 20

Figure 13Person-level predicted median RT versus observed median RT from the 2018 PISA Math and Science Data.Figure 13 long description.

Figure 21

Table 9Ability estimation accuracy of the Gumbel copula and conditional dependence modelsTable 9 long description.

Figure 22

Table 10Model selection results under the Gumbel copula and the CD modelTable 10 long description.

Figure 23

Table D1Average log marginal likelihood in the simulation studyTable D1 long description.

Figure 24

Table D2Parameter recovery under the Joe copula dependence conditionTable D2 long description.

Figure 25

Table D3Accuracy of ability estimates under the Joe copula dependence conditionTable D3 long description.

Figure 26

Table D4Accuracy of speed estimates under the Gumbel copula dependence conditionTable D4 long description.

Figure 27

Table D5Item parameter recovery under Gumbel copula condition (N=500$ N = 500$, I=18$I = 18$, and δ=1.43$\delta = 1.43$)Table D5 long description.

Figure 28

Table D6Item parameter recovery under MVN dependence condition (N=500$ N = 500$, I=18$I = 18$, and ρ=0.45$\rho = 0.45$)Table D6 long description.

Figure 29

Figure D1Conditional RMSE of ability estimates under the Joe copula dependence condition.Figure D1 long description.

Figure 30

Figure D2Conditional RMSE of speed estimates under the Gumbel dependence condition.Figure D2 long description.

Figure 31

Table E1Model selection result under the 270∘$270^\circ $-rotated Gumbel copulaTable E1 long description.

Figure 32

Table E2Model selection result under the 270∘$270^\circ $-rotated Joe copulaTable E2 long description.

Figure 33

Table E3Parameter recovery under the 270∘$270^\circ $-rotated Gumbel copulaTable E3 long description.

Figure 34

Table E4Parameter recovery under the 270∘$270^\circ $-rotated Joe copulaTable E4 long description.

Figure 35

Figure E1Conditional RMSE of ability estimates under the 270∘$270^\circ $-rotated Gumbel copula.Figure E1 long description.

Figure 36

Figure E2Conditional RMSE of ability estimates under the 270∘$270^\circ $-rotated Joe copula.Figure E2 long description.

Figure 37

Table F1Dependence parameter estimates from the empirical data analysesTable F1 long description.

Figure 38

Table F2Cross-validation results from the empirical data analysesTable F2 long description.

Figure 39

Table F3Model selection results with BIC values from the empirical data analysesTable F3 long description.

Figure 40

Table F4Model selection results with AIC values from the empirical data analysesTable F4 long description.

Figure 41

Figure F1Item-level predicted accuracy versus observed accuracy from the 2018 PISA Math and Science Data.Figure F1 long description.

Figure 42

Figure F2Item-level predicted median RT versus observed median RT from the 2018 PISA Math and Science Data.Figure F2 long description.