Abstract All cohorts are conditioned on survival to a study's baseline. The validity of estimates drawn from these cohorts of survivors may be compromised if those who die prior to enrollment have different covariate structures than survivors. In this research note, we used data from the “HAALSI Cohort” (Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa) of older adults in rural South Africa and a “Mortality Cohort” of individuals who would have been eligible for HAALSI but died before they had the opportunity to enroll, drawing on complete population mortality data from the Agincourt Health and Socio-Demographic Surveillance System. We simulated the prevalence of cognitive impairment under different assumptions about the prevalence of such impairment in the Mortality Cohort. We constructed a random forest classification model to predict cognitive impairment in the Mortality Cohort and compared it with observed estimates in the HAALSI Cohort. The prevalence of cognitive impairment was sensitive to assumptions about the prevalence in the Mortality Cohort. The predictive model revealed meaningfully higher predicted probability of cognitive impairment in a counterfactual scenario with no prebaseline deaths. Researchers should consider prebaseline mortality in the interpretation of prevalence estimates, especially when the magnitude of prebaseline deaths is likely large.
Wang et al. (Thu,) studied this question.