Activity Number:
|
77
- Data, Linked Data, and Model-Based Analytics in Social Science
|
Type:
|
Contributed
|
Date/Time:
|
Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
|
Sponsor:
|
Social Statistics Section
|
Abstract #323469
|
|
Title:
|
Assessing Residual Diagnostics for Ordinal Response Models
|
Author(s):
|
Altea Lorenzo-Arribas* and Antony M Overstall and Mark J Brewer
|
Companies:
|
University of Southampton / Biomathematics and Statistics Scotland and University of Southampton and Biomathematics and Statistics Scotland
|
Keywords:
|
Cumulative logit models ;
Goodness of fit ;
Residuals
|
Abstract:
|
Model assessment in cumulative logit ordinal response models entails several practical issues. Traditional goodness-of-fit measures available for linear fixed and random-effects models and diagnostic plots of individual residuals are generally unavailable and difficult to implement for these models, and residual diagnostics are generally accepted as underdeveloped. The main problems are associated with the discrete nature of the data. General challenges include: failure to provide only one value per subject; the overall direction of the observed values is not preserved; non-monotonic behaviour with respect to the observed values for those with the same covariates; lack of symmetry with respect to zero; and, different from zero expectation (in contrast to the ideal features proposed by Li and Shepherd). We build on recent progress and assess the behaviour of residuals proposed for ordinal response models. Preliminary results have shown that randomized quantile residuals often miss lack of fit and that probability-scale residuals show a tendency to produce "false negatives".
|
Authors who are presenting talks have a * after their name.