Font Size:
Methods for handling attrition: an overview
Damon Stanley Berridge, Roger Norman Penn
Building: Law Building
Room: Breakout 7 - Law Building, Room 028
Date: 2012-07-12 11:00 AM – 12:30 PM
Last modified: 2012-06-19
Abstract
This paper presents a general overview of the development of techniques for the statistical modelling of longitudinal data subject to attrition. It will also explore the various computational solutions to fitting such models currently.
Attrition occurs at random when each respondent in a panel survey is equally likely to drop out at any given wave. When attrition is correlated with the social process under scrutiny, attrition is non-random and may be informative about that process. The most efficient and effective way of handling attrition is to model simultaneously the response and attrition processes. Joint models for response and attrition can be classified into two broad types: selection models; and pattern-mixture models. The difference between the two types lies in the manner in which the joint distribution of the response-attrition mechanism is factored.
A major problem encountered in the development of such models is that results are often sensitive to violations of assumptions about the nature of attrition. Heckman (1979) developed a two-stage estimator and outlined a number of semi- and non-parametric approaches to estimating selection models that relied on weaker assumptions about the nature of the attrition process. Little (1995) suggested that sensitivity analyses should be performed in order to assess the effect on inferences of alternative assumptions about the attrition process.
Follmann & Wu (1995) proposed separate models for continuous responses with binary dropout processes which were linked by a common random parameter. An approximation to the response model was conditioned on attrition thereby precluding the need to specify an explicit attrition model. This was extended by Ten Have et al. (1998) to handle longitudinal binary responses subject to informative dropout. They made use of a shared parameter model with logistic link.
The shared parameter model is equivalent to a joint response-attrition model in which the random effects in each component (that is, the response and attrition models) are correlated with each other. The shared parameter is equivalent to the association parameter in this correlated random effects model. We will demonstrate how this correlated random effects model may be fitted within the software package SABRE (http://sabre.lancs.ac.uk ).
Attrition occurs at random when each respondent in a panel survey is equally likely to drop out at any given wave. When attrition is correlated with the social process under scrutiny, attrition is non-random and may be informative about that process. The most efficient and effective way of handling attrition is to model simultaneously the response and attrition processes. Joint models for response and attrition can be classified into two broad types: selection models; and pattern-mixture models. The difference between the two types lies in the manner in which the joint distribution of the response-attrition mechanism is factored.
A major problem encountered in the development of such models is that results are often sensitive to violations of assumptions about the nature of attrition. Heckman (1979) developed a two-stage estimator and outlined a number of semi- and non-parametric approaches to estimating selection models that relied on weaker assumptions about the nature of the attrition process. Little (1995) suggested that sensitivity analyses should be performed in order to assess the effect on inferences of alternative assumptions about the attrition process.
Follmann & Wu (1995) proposed separate models for continuous responses with binary dropout processes which were linked by a common random parameter. An approximation to the response model was conditioned on attrition thereby precluding the need to specify an explicit attrition model. This was extended by Ten Have et al. (1998) to handle longitudinal binary responses subject to informative dropout. They made use of a shared parameter model with logistic link.
The shared parameter model is equivalent to a joint response-attrition model in which the random effects in each component (that is, the response and attrition models) are correlated with each other. The shared parameter is equivalent to the association parameter in this correlated random effects model. We will demonstrate how this correlated random effects model may be fitted within the software package SABRE (http://sabre.lancs.ac.uk ).