Large cohort studies are subject to the problem of attrition. The most prominent types of attrition include those participants who have died during the follow-up period, those who cannot be located because of (e)migration, and those who do not respond to the follow-up survey (i.e., non-responders) [1, 2]. Although some causes cannot be influenced by the researcher, study design and efforts to contact the study population can modify the degree of attrition .
High rates of non-participation to a follow-up survey can lead to selection bias, when the persons who drop-out differ significantly from the participants in characteristics that are related to the outcome being studied [3, 4]. This loss of a selective group can reduce the external validity as well as the generalizability of the research findings [1, 2, 5, 6]. The success of any longitudinal study, therefore, depends upon its participants remaining in the study . Assessment of information on initial participation and retention rates helps to evaluate potential selection bias when non-participation during follow-up is not random [1, 7, 8]. Furthermore, assessment of determinants of attrition may identify characteristics of participants who are most unlikely to respond to the follow-up survey . This may aid in management strategies to target specifically individuals with such characteristics and thus leading to reducing non-response [1, 4, 6, 9]. Various population-based longitudinal cohort studies have shown that non-responders often differ from those who respond to a follow-up survey with respect to demographic, socioeconomic and health characteristics. Many factors have been investigated, though not all factors are consistently found to be significantly associated with non-response [10, 11]. However, in most studies, non-responders are more likely to be among the youngest [1–3, 12] or oldest participants [6, 8, 9], to live alone [1–4, 6, 9, 13], to be less educated [1, 4, 6, 8, 11–14], unemployed [2, 5, 9, 14] and to have a low income [5, 6, 11]. Non-responders are more likely to have an unhealthy lifestyle, especially being a smoker [2–4, 7, 8, 11, 13]. The general health profile of non-responders tends to be worse than that of responders [1, 4, 8, 9, 11, 13, 15, 16] and a higher prevalence of obesity is observed [3, 8, 12].
To date, studies on determinants of non-response have been mainly conducted in single population-based cohorts where all participants were followed for the same time period. The present study, however, is based on data from almost 500.000 participants from 10 European countries, as part of the EPIC-PANACEA (European Prospective Investigation into Cancer and Nutrition-Physical Activity, Nutrition, Alcohol, Cessation of Smoking, Eating out of home And obesity) study. EPIC-PANACEA aims to investigate the determinants of obesity and body weight changes in Europe. For the purpose of EPIC-PANACEA data from a second assessment of body weight collected several years after baseline were centralized and combined with the EPIC baseline dataset. The length of follow-up as well as the method of contacting participants (i.e. by postal surveys, directly by phone or by a request to visit a study center for physical examination) differed between the collaborating centers. This allows insight in whether non-response differs with various methods of contacting participants and diverse durations of follow-up.
The purpose of the present study was twofold. First, we investigated whether baseline demographic, socio-economic, health variables, length of follow-up and method of contacting the participants predicted non-response to an invitation for a second assessment of lifestyle factors and body weight excluding those who were not (yet) contacted, and those who either died or emigrated during follow-up. This provides insight in important determinants of non-response that can be used to enhance cohort maintenance in future studies. Second, we compared all baseline participants for whom a second body weight assessment was missing (including non-responders, (e)migrated, deceased or not yet contacted participants) with responders, to evaluate whether the population lost to follow-up formed a selective group causing potential selection bias in future analyses.