Bias in population oral health research: longitudinal studies.

Abstract

Bias in longitudinal studies have been well described and the longer the follow-up, the higher the proportion of drop-outs. Here, I present some key issues related to selection bias, time-varying confounders, solutions to bias and challenges in longitudinal studies in dental research. Selection bias creates distortions in measures of disease frequency or association due to losses of follow-up or use of specific population groups. It is shown that even if losses are not associated with baseline values, measures such as odds ratios may be seriously distorted. Such problems can be understood by directed acyclic graphs, identifying the collider bias, or by missing data theory. Time-varying confounding occurs when an exposure varies over time and is affected by past exposure of other time-varying covariates, creating a complex scenario to adjustment in multiple regression. Under some assumptions, missing information may be informed by other variables in the dataset, and techniques such as multiple imputation or inverse probability weighting can be helpful, but the best solution is to prevent losses of follow-up as much as possible. Finally, I present challenges for longitudinal studies that use electronic health records and the need to incorporate area-based contextual measures. The first allows linkage of dental records with other information systems to create longitudinal (big) data. The second allows evaluation longitudinally of the effect of contextual factors, including social and health policies, on oral health.