Models and estimation for repeated ordinal responses, with applications to telecommunications experiments
Models and estimation for repeated ordinal responses, with applications to telecommunications experiments
Responses made on scales with ordered categories (ordinal responses) can be analysed using multinomial models which include 'threshold parameters'. These models have become established over the last 20 years, in which time research has focused on models with location terms which allow for a change in location of the threshold parameters. These location terms explain one type of difference between patterns of ordinal responses. Another type of difference can be explained by the inclusion of scaling terms in the model.
Where ordinal responses are observed repeatedly on the same subject the analyst has a challenge to explain the correlation between these intra-subject responses. This thesis presents a new approach of this challenge which involves fitting one of the multinomial models, the cumulative logit model, with subject-specific location and scaling terms. These terms may be fixed effects or random effects and both cases are investigated. The approach is motivated by data from telecommunications experiments, and when used to analyse these data, it is found that the model gives a good explanation of the correlation between intra-subject responses. A new piece of general-purpose software is introduced which allows the fitting, by maximum-likelihood, of cumulative link models with both location and scaling terms.
It is possible to fit the cumulative logit model by using generalized estimating equations (GEE). One particular type of GEE is discussed in detail and referred to as 'independent binomials'. An advantage to the use of this method of fitting the cumulative logit model is its straightforward implementation. Some theoretical comparisons are performed to compare the efficiency of independent-binomials estimation of model parameters with the efficiency of maximum-likelihood estimation. It is concluded that the loss of efficiency in independent-binomials cannot be considered too great to warrant its dismissal as a method of estimation.
University of Southampton
Wolfe, Rory St.John
e8077801-d801-4336-9e12-e936ee9d69eb
1996
Wolfe, Rory St.John
e8077801-d801-4336-9e12-e936ee9d69eb
Wolfe, Rory St.John
(1996)
Models and estimation for repeated ordinal responses, with applications to telecommunications experiments.
University of Southampton, Doctoral Thesis.
Record type:
Thesis
(Doctoral)
Abstract
Responses made on scales with ordered categories (ordinal responses) can be analysed using multinomial models which include 'threshold parameters'. These models have become established over the last 20 years, in which time research has focused on models with location terms which allow for a change in location of the threshold parameters. These location terms explain one type of difference between patterns of ordinal responses. Another type of difference can be explained by the inclusion of scaling terms in the model.
Where ordinal responses are observed repeatedly on the same subject the analyst has a challenge to explain the correlation between these intra-subject responses. This thesis presents a new approach of this challenge which involves fitting one of the multinomial models, the cumulative logit model, with subject-specific location and scaling terms. These terms may be fixed effects or random effects and both cases are investigated. The approach is motivated by data from telecommunications experiments, and when used to analyse these data, it is found that the model gives a good explanation of the correlation between intra-subject responses. A new piece of general-purpose software is introduced which allows the fitting, by maximum-likelihood, of cumulative link models with both location and scaling terms.
It is possible to fit the cumulative logit model by using generalized estimating equations (GEE). One particular type of GEE is discussed in detail and referred to as 'independent binomials'. An advantage to the use of this method of fitting the cumulative logit model is its straightforward implementation. Some theoretical comparisons are performed to compare the efficiency of independent-binomials estimation of model parameters with the efficiency of maximum-likelihood estimation. It is concluded that the loss of efficiency in independent-binomials cannot be considered too great to warrant its dismissal as a method of estimation.
Text
487749.pdf
- Version of Record
More information
Published date: 1996
Identifiers
Local EPrints ID: 462966
URI: http://eprints.soton.ac.uk/id/eprint/462966
PURE UUID: ddd1c567-f632-4c6f-9133-d8f88aa902a0
Catalogue record
Date deposited: 04 Jul 2022 20:31
Last modified: 16 Mar 2024 19:00
Export record
Contributors
Author:
Rory St.John Wolfe
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics