Nonignorable nonresponse models for categorical survey data
Nonignorable nonresponse models for categorical survey data
Sample surveys are an important source of information in the social sciences, and qualitative information that can be summarised by variables with categorical outcomes are commonly used when analysing the composition and behaviour of social structures. Nonresponse leads to incomplete data, the analysis of which causes a number of theoretical and practical difficulties. If the reason for nonresponse depends on the missing data values, then the nonresponse is said to be nonignorable. Model-based analysis of data subject to nonignorable nonresponse requires the specification of a model for the nonresponse mechanism. Two nonignorable nonresponse models for the analysis of incomplete contingency tables are considered in this thesis.
The first model considered is for an incomplete multi-way table where one variable is incompletely observed. Previous applications of this model have found that nonignorable models fit the incomplete data well but produce parameter estimates on the boundary of the parameter space. Maximum likelihood estimation is investigated using a geometric interpretation to explain the occurrence of boundary estimates. It is found that the parameters for nonignorable nonresponse are weakly identified and unstable. The consequences of this finding are investigated by simulation studies of interval estimation and model selection. It is found that the method of interval estimation must be carefully chosen, and that the power of the likelihood ratio test can be severely affected by boundary estimates.
A second model is proposed for modelling nonignorable nonresponse from household based surveys. The design of household-based surveys can mean that within household nonresponse behaviour can be highly correlated; for example, if one household member acts as a proxy for the other household members. To control for household correlation, the probability of household nonresponse is modelled as a function of the household members' characteristics. Simulation is used to demonstrate the possible bias from fitting models which ignore household correlation, and then the model is applied to estimating labour force gross flows from the British Labour Force Survey. It is found that unemployed individuals are under-represented in the sample, although the degree of this mis-representation is subject to the instability of nonignorable model estimates.
University of Southampton
1998
Clarke, Paul Simon
(1998)
Nonignorable nonresponse models for categorical survey data.
University of Southampton, Doctoral Thesis.
Record type:
Thesis
(Doctoral)
Abstract
Sample surveys are an important source of information in the social sciences, and qualitative information that can be summarised by variables with categorical outcomes are commonly used when analysing the composition and behaviour of social structures. Nonresponse leads to incomplete data, the analysis of which causes a number of theoretical and practical difficulties. If the reason for nonresponse depends on the missing data values, then the nonresponse is said to be nonignorable. Model-based analysis of data subject to nonignorable nonresponse requires the specification of a model for the nonresponse mechanism. Two nonignorable nonresponse models for the analysis of incomplete contingency tables are considered in this thesis.
The first model considered is for an incomplete multi-way table where one variable is incompletely observed. Previous applications of this model have found that nonignorable models fit the incomplete data well but produce parameter estimates on the boundary of the parameter space. Maximum likelihood estimation is investigated using a geometric interpretation to explain the occurrence of boundary estimates. It is found that the parameters for nonignorable nonresponse are weakly identified and unstable. The consequences of this finding are investigated by simulation studies of interval estimation and model selection. It is found that the method of interval estimation must be carefully chosen, and that the power of the likelihood ratio test can be severely affected by boundary estimates.
A second model is proposed for modelling nonignorable nonresponse from household based surveys. The design of household-based surveys can mean that within household nonresponse behaviour can be highly correlated; for example, if one household member acts as a proxy for the other household members. To control for household correlation, the probability of household nonresponse is modelled as a function of the household members' characteristics. Simulation is used to demonstrate the possible bias from fitting models which ignore household correlation, and then the model is applied to estimating labour force gross flows from the British Labour Force Survey. It is found that unemployed individuals are under-represented in the sample, although the degree of this mis-representation is subject to the instability of nonignorable model estimates.
This record has no associated files available for download.
More information
Published date: 1998
Identifiers
Local EPrints ID: 463392
URI: http://eprints.soton.ac.uk/id/eprint/463392
PURE UUID: dd16d5a0-919a-4a86-aaa3-e067fcedde4d
Catalogue record
Date deposited: 04 Jul 2022 20:51
Last modified: 04 Jul 2022 20:51
Export record
Contributors
Author:
Paul Simon Clarke
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