Approximations of the information matrix for a panel mixed logit model

Wei Zhang, Abhyuday Mandal, John Stufken

Research output: Contribution to journalArticle

Abstract

Information matrices play a key role in identifying optimal designs. Panel mixed logit models are more flexible than multinomial logit models for discrete choice experiments. For panel mixed logit models, the information matrix does not have a closed-form expression and is difficult to evaluate. We propose three methods to approximate the information matrix, namely, importance sampling, Laplace approximation, and joint sampling. The three methods are compared through simulations. Since our ultimate goal is to find optimal designs, the three methods are compared on whether they rank designs similarly, not on how accurate the approximations are. Although the Laplace approximation is not as accurate as the other two methods, it can still be used to rank designs accurately and it is much faster than the other two methods. When an optimal design search using an exchange algorithm takes days to run, the Laplace approximation may be the only viable choice to use in practice.

Original languageEnglish (US)
Pages (from-to)1-27
Number of pages27
JournalJournal of Statistical Theory and Practice
DOIs
StateAccepted/In press - Sep 6 2016

Fingerprint

Logit Model
Information Matrix
Mixed Model
Laplace Approximation
Approximation
Discrete Choice
Multinomial Logit Model
Exchange Algorithm
Importance Sampling
Closed-form
Evaluate
Experiment
Simulation

Keywords

  • A-optimality
  • D-optimality
  • Discrete choice experiments
  • importance sampling
  • joint sampling
  • Laplace’s method
  • optimal designs

ASJC Scopus subject areas

  • Statistics and Probability

Cite this

Approximations of the information matrix for a panel mixed logit model. / Zhang, Wei; Mandal, Abhyuday; Stufken, John.

In: Journal of Statistical Theory and Practice, 06.09.2016, p. 1-27.

Research output: Contribution to journalArticle

@article{9bd7b01cab2540299e77968eb7c6c5a4,
title = "Approximations of the information matrix for a panel mixed logit model",
abstract = "Information matrices play a key role in identifying optimal designs. Panel mixed logit models are more flexible than multinomial logit models for discrete choice experiments. For panel mixed logit models, the information matrix does not have a closed-form expression and is difficult to evaluate. We propose three methods to approximate the information matrix, namely, importance sampling, Laplace approximation, and joint sampling. The three methods are compared through simulations. Since our ultimate goal is to find optimal designs, the three methods are compared on whether they rank designs similarly, not on how accurate the approximations are. Although the Laplace approximation is not as accurate as the other two methods, it can still be used to rank designs accurately and it is much faster than the other two methods. When an optimal design search using an exchange algorithm takes days to run, the Laplace approximation may be the only viable choice to use in practice.",
keywords = "A-optimality, D-optimality, Discrete choice experiments, importance sampling, joint sampling, Laplace’s method, optimal designs",
author = "Wei Zhang and Abhyuday Mandal and John Stufken",
year = "2016",
month = "9",
day = "6",
doi = "10.1080/15598608.2016.1219288",
language = "English (US)",
pages = "1--27",
journal = "Journal of Statistical Theory and Practice",
issn = "1559-8608",
publisher = "Taylor and Francis",

}

TY - JOUR

T1 - Approximations of the information matrix for a panel mixed logit model

AU - Zhang, Wei

AU - Mandal, Abhyuday

AU - Stufken, John

PY - 2016/9/6

Y1 - 2016/9/6

N2 - Information matrices play a key role in identifying optimal designs. Panel mixed logit models are more flexible than multinomial logit models for discrete choice experiments. For panel mixed logit models, the information matrix does not have a closed-form expression and is difficult to evaluate. We propose three methods to approximate the information matrix, namely, importance sampling, Laplace approximation, and joint sampling. The three methods are compared through simulations. Since our ultimate goal is to find optimal designs, the three methods are compared on whether they rank designs similarly, not on how accurate the approximations are. Although the Laplace approximation is not as accurate as the other two methods, it can still be used to rank designs accurately and it is much faster than the other two methods. When an optimal design search using an exchange algorithm takes days to run, the Laplace approximation may be the only viable choice to use in practice.

AB - Information matrices play a key role in identifying optimal designs. Panel mixed logit models are more flexible than multinomial logit models for discrete choice experiments. For panel mixed logit models, the information matrix does not have a closed-form expression and is difficult to evaluate. We propose three methods to approximate the information matrix, namely, importance sampling, Laplace approximation, and joint sampling. The three methods are compared through simulations. Since our ultimate goal is to find optimal designs, the three methods are compared on whether they rank designs similarly, not on how accurate the approximations are. Although the Laplace approximation is not as accurate as the other two methods, it can still be used to rank designs accurately and it is much faster than the other two methods. When an optimal design search using an exchange algorithm takes days to run, the Laplace approximation may be the only viable choice to use in practice.

KW - A-optimality

KW - D-optimality

KW - Discrete choice experiments

KW - importance sampling

KW - joint sampling

KW - Laplace’s method

KW - optimal designs

UR - http://www.scopus.com/inward/record.url?scp=84986188692&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84986188692&partnerID=8YFLogxK

U2 - 10.1080/15598608.2016.1219288

DO - 10.1080/15598608.2016.1219288

M3 - Article

AN - SCOPUS:84986188692

SP - 1

EP - 27

JO - Journal of Statistical Theory and Practice

JF - Journal of Statistical Theory and Practice

SN - 1559-8608

ER -