Representativeness, losses to follow-up and validity in cohort studies

Abstract

We appreciate the comments by Dr. Sheikh on our paper assessing selection bias due to losses to follow-up in the Seguimiento Universidad de Navarra (SUN) cohort [1]. Dr. Sheikh raises three major concerns regarding the validity of our results and conclusions: first, that the non-representative character of our study population could create selection bias; second, that use of inverse probability weighting (IPW) could not adjust for selection bias in cases where information is missing not at random; finally, he affirms that our results are contradictory because, on the one hand, we observed differences among responders and nonresponders to the follow-up questionnaire but, on the other hand, censoring adjustment using IPW did not change the association estimates. Certainly, our study sample is not a representative sample of the whole Spanish population, the Spanish university graduates, or even the graduates from the University of Navarra. Individuals willing to participate in most of the better known epidemiologic cohorts usually are more health-conscious and have healthier lifestyles than the general population. But, in our study, this is not an important limitation. In fact, it could be an advantage [2]. The SUN cohort, as many of the other well-known cohort studies, such as the European Prospective Investigation into Cancer and Nutrition (EPIC), does not aim to define the prevalence or incidence of cardiovascular disorders in the general population. Population registries with full coverage are the most useful tools for this aim. The SUN cohort, however, intends to study the association between some lifestyles and cardiovascular disorders and other outcomes. To increase the internal validity of the results, we have preferred to include only individuals with high educational level, whose information both for exposures and outcomes would probably be more accurate. Once we assume internal validity, we can decide whether the observed results can be generalized to the population or not. As Rothman states, generalization of results in the field of scientific inquiry does not depend on statistical representativeness of the study population but on the mechanisms underlying the observed association and the biological laws governing the process under study [3, 4]. Truly, in some situations, selection of a particular group could introduce bias because of the relationship between factors associated to participation in the study, the exposure of interest and the outcome under study. A mere association between a particular factor and attrition does not necessarily induce selection bias (for example, when that factor does not predict the outcome or when other factors with opposite associations compensate for it). In any case, as Hernan et al. suggest, adjusting in the analysis for those variables could resolve the problem [5]. In our study, we adjusted for the more important risk factors for hypertension and, then, we have controlled in part for this potential selection bias. Another issue is whether using IPW to adjust for censoring really removes the selection bias produced by losses to follow-up. As we affirm in our paper, two assumptions that are required for the validity of the IPW method are the absence of unmeasured confounding and that unmeasured informative censoring may not exist. Applying the usual terminology for the analysis of missing data, we should say A. Alonso (&) Department of Epidemiology, Harvard School of Public Health, 677 Huntington Avenue, Boston, MA 02155, USA e-mail: aalogut@alumni.unav.es