Predictors of two forms of attrition in a longitudinal health study involving ageing participants: an analysis based on the Whitehall II study

Abstract

Background: Longitudinal studies are crucial providers of information about the needs of an ageing population, but their external validity is affected if partipants drop out. Previous research has identified older age, impaired cognitive function, lower educational level, living alone, fewer social activities, and lower socio-economic status as predictors of attrition.

Methods: This project examined attrition in participants of the Whitehall II study aged between 51-71 years, using data from questionnaires participants have completed biennially since 1985 when the study began. We examine the possibility of two distinct forms of attrition--non-response and formally requesting to withdraw--and whether they have different predictors. Potential predictors were age, gender, marital status, occupational grade, retirement, home ownership, presence of longstanding illness, SF-36 quality of life scores, social participation and educational level comparing participants and those who had withdrawn from the study.

Results: The two forms of attrition share many predictors and are associated but remain distinct. Being older, male, having a lower job grade, not being a home owner, not having a long standing illness, having higher levels of education, and not having retired, were all associated with a greater probability of non-response; being married was associated with higher probability in women and lower in men. Being older, male, having a lower job grade, not being a home owner, having lower SF-36 scores, taking part in fewer social activities, and not having a long standing illness, were all associated with greater probability of withdrawal.

Conclusions: The results suggest a strong gender effect on both routes not previously considered in analyses of attrition. Investigators of longitudinal studies should take measures to retain older participants and lower level socio-economic participants, who are more likely to cease participating. Recognition should be given to the tendency for people with health problems to be more diligent participants in studies with a medical screening aspect, and for those with lower socio-economic status (including home ownership), quality of life and social participation, to be more likely to request withdrawal. Without taking these features into account, bias and loss of power could affect statistical analyses.

Figure 1

Two plausible causal models linking non-response and withdrawal.