Mind the Gap: Addressing Missing Person Time When Estimating Outcome Incidence in Longitudinal Data

Longitudinal data often include gaps in observation when outcomes (and other variables) are unmeasured due to missed study visits or dropout. We explore the fundamentals of data gaps and use simulation to compare approaches for handling data gaps when estimating outcome incidence.

We generated a simulation of 1000 individuals across 10 study visits. We used 4 data-generating mechanisms: (1) missingness was independent of the outcome; (2) there was a baseline common cause of missingness and the outcome; (3) there was a time-varying common cause; and (4) the outcome directly affected future missingness. We estimated the risk and rate of the first outcome occurrence (generated as a transient, repeated, and permanent outcome), using crude and adjusted approaches, across 1000 iterations, and compared bias and empirical standard error.

Under Scenario 1, in crude analyses, results were unbiased when censoring before a data gap but not when allowing participants to return. Under scenarios 2-4, all crude approaches were biased. Inverse probability of censoring weights and multiple imputation were relatively unbiased across scenarios and outcome types; multiple imputation was more precise. Inverse probability of observation weights was biased when the outcome was permanent and was less precise than either of the other two approaches.

Crude approaches allowing participants to return following a data gap are not recommended because they can be biased even when missingness and the outcome are independent. Instead, one should either censor or handle the data gap using multiple imputation.