The effects of missing data due to study dropout on longitudinal analysis inference using outcome-dependent sampling.

Background: Existing longitudinal cohort study data and associated biospecimen libraries provide abundant opportunities to efficiently examine new hypotheses through retrospective specimen testing. Outcome-dependent sampling (ODS) methods offer a powerful alternative to random sampling when testing all available specimens is not feasible or biospecimen preservation is desired. For repeated binary outcomes, a common ODS approach is to extend the case-control framework to the longitudinal setting.For ODS designs, we consider the impact of incomplete follow-up when missingness is completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). We further consider sampling from (i) complete cases, in which, in an attempt to maximize power, participants who dropped out before study completion were excluded; and (ii) all individuals, including those with incomplete follow-up.

Methods: Simulation studies based on the Advancing Clinical Therapeutics Globally HIV Infection, Aging, and Immune Function Long-Term Observational Study cohort were used to examine the impact of MCAR, MAR, and MNAR missingness, assuming specimens were sampled from either (i) complete cases; or (ii) all individuals. Three ODS analytical methods were considered.

Results: When longitudinal data are MNAR, ODS methods exhibit bias similar to that seen in random sampling. MNAR and MAR bias is exacerbated when sampling only participants with complete follow-up. Simulations indicate that ODS analyses that include participants with incomplete follow-up are robust to MCAR and less biased by MAR missingness.

Conclusion: Dropout is common in longitudinal cohort studies. Investigators utilizing ODS methods must consider the effect of dropout in both the retrospective sampling design and analysis.