Abstract
Objective: To assess the impact of attrition in longitudinal studies on bias and efficiency.
Background: The Dunedin Cohort study is world renowned for its high retention rate. Information on difficulty of contact is used to assess the likely impact of attrition on the estimation of the prevalence of risk factors and associations of these risk factors with poor health outcomes.
Methods: 972 (96% of living) Dunedin Study members who presented at phase 32 were scored for difficulty of contact. Bias in the prevalence of risk factors and associations with poor health outcomes when those who were most difficult to contact ('hard-to-get') are excluded (quasi attrition) are assessed by simulation (using R). Where associations are not biased, we estimate the increment in sample size which would be needed without the hard-to-get, in order to achieve the same efficiency as the sample with the hard-to-get. The relative value (RV) of a study member who is hard-to-get is defined as the ratio of the additional sample size needed to the number of the hard-to-get recruited.
Results: Work in progress has demonstrated that Dunedin Study members who were hard-to-get were more likely than the easy-to-get group to have higher rates of substance use and dependence, and to have behavioural disorders. Hence the prevalence of risk factors for poor health for the sample without the 'hard-to-get' were negatively biased. Associations between exposures and poor health outcomes may not be biased, for example that between cannabis consumption and periodontal disease. However, excluding those who are hard-to-get, a greater than expected loss of efficiency is observed (RV>1). Generalised results for the dependence of the RV on prevalence of exposure amongst the hard-to-get are derived.
Conclusions: Because those who are hard-to-get tend to be at the extremes of distributions of risk factors for poor health outcomes, or to have greater prevalence of conditions that affect only small proportions of the population, they contribute more information about associations between such predictors and conditions than do those who generally participate in studies. Efforts in retention can therefore represent greater value for money than increases in sample size.