Handling attrition and non-response in longitudinal data with an application to a study of Australian youth

Abstract

A standard concern with long term longitudinal studies is that of attrition over time. Together with initial non--response this typically leads to biased model estimates unless a suitable form of adjustment is carried out. The standard approach to this has been to compute weights based upon the propensity to respond and to drop out and then carry out weighted analyses to compensate for response bias. In the present paper we argue that this approach is statistically inefficient, because it drops incomplete data records, is inflexible, and in practice gives rise to undue complexity involving a proliferation of weighting systems for different analyses. Instead we set out an alternative approach that makes use of recently developed imputation procedures for handling missing data and show how this can be used to improve the quality of the statistical analysis. An example analysis is given using the Longitudinal Study of Australian Youth (LSAY).

A Bayesian model for measurement and misclassification errors alongside missing data, with an application to higher education participation in Australia

“…The first two columns of results in Table 3 replicate these analyses, pooling the two categories of nonGovernment school (Catholic, Private) that were treated separately by Cumming and Goldstein (2016) but in fact showed only a small and non-significant difference and so have been combined in our analysis.. The level 2 units are the year 9 schools, and Table 3 lists the predictor variables with full details given by Cumming and Goldstein (2016). Note that the scale of the SES measures has been divided by 100 and the test scores divided by 10 to provide more significant figures for the coefficient estimates.…”

“…Data were collected on variables related to education, training, work, financial matters, health, social activities and attitudes as well as background family characteristics such as SES. A description of the variables is given by Cumming and Goldstein (2016). LSAY started in 1995 by sampling Year 9 school students, with an average age 14.5 years, in Australian secondary schools and subsequently is following them up every year on a further 11 occasions (LSAY, 2013a).…”

“…While Cumming et al (2016) did account for attrition and missing data they did not allow for test score unreliability. Such unreliability is a general feature of educational tests and arises from several sources of variation including the choice (sampling) of test questions, conditions of test administration and short term fluctuations within students.…”

“…Cumming and Goldstein (2016) analyse this dataset but considered only the case of missing data and ignored possible measurement errors, using the algorithm described by Goldstein et al (2014) to obtain efficient parameter estimates. The first two columns of results in Table 3 replicate these analyses, pooling the two categories of nonGovernment school (Catholic, Private) that were treated separately by Cumming and Goldstein (2016) but in fact showed only a small and non-significant difference and so have been combined in our analysis.. The level 2 units are the year 9 schools, and Table 3 lists the predictor variables with full details given by Cumming and Goldstein (2016).…”

“…In Cumming and Goldstein (2016) it was concluded that the principal effect of adjusting for missing data was a gain in efficiency, with a small increase in the estimate of the difference between Government and non-Government schools, so that there were no important policy implications.…”