I have been involved in a project with my colleague Gill Mein and others involved in the Whitehall II cohort study to look into predictors of attrition. The paper has just come out in BMC Medical Research Methodology. This is something that has been looked into in other cohorts but our novel approach focussed on the difference between non-response (which may be transitory) and formally requesting withdrawal (which is terminal). We found these shared some predictors and yet had some differences.
One noteworthy feature was the unexpected link between self-reported long-standing illness and decreased attrition, even after adjusting for their SF-36 health-related quality of life. Whitehall involves a periodic health screen examination and it seems that people who regard themselves as ill make more of an effort to stay in, perhaps because they perceive a benefit to the examination. Nowadays we would be cautious about making any claim on the benefits of such a generic examination, which raises an unexpected ethical issue: is it appropriate to offer screening as a way of keeping ill people in longitudinal studies?
I am now working on various Bayesian joint models that try to discriminate between alternative causal models for the attrition data. Watch this space.