O4–07–05: Correcting bias in estimates of differential rates of cognitive decline caused by informative missing data through the use of telephone cognitive evaluations

In longitudinal studies of cognitive aging, missing data may bias both overall estimates of cognitive decline, and the assessment of the effects of risk factors. In this study we use a telephone memory assessment instrument to help correct for such biases in the estimation of cognitive decline in APOE-ε4 carriers (ε4 +) and non-carriers (ε4 -). An initially healthy sample of community residing adults without dementia, age 70+ were followed in the Einstein Aging Study using a cognitive assessment battery which included the free and cued selective reminding test. All subjects also received a telephonic assessment of cognitive status even when they could not complete in-person assessments. We jointly model in-person and telephone cognitive data using linear mixed models in order to assess the degree to which differences in rates of memory decline between APOE-ε4 positive (ε4+) and APOE-ε4 negative (ε4-) participants may be affected by informative missing data. We also simulated missing data scenarios to evaluate the likely effectiveness of this approach in correcting for bias due to missing data. Among 640 EAS participants, 473 were e4- and 147 were e4+. EAS participants contributed 2,434 in-person assessments over 2,085 person-years of follow-up. Compared to approaches that ignore missing data, the joint model showed rates of memory decline that were 19% greater in ε4 negative participants and 26% greater in ε4 positive participants. The estimated effect of APOE-ε4 on memory decline was 37% larger using joint modeling with the telephone assessment instrument compared with traditional models. Simulations showed that the correction for the bias due to missing data is only partial. Informative loss to follow-up leads to underestimates of rates of cognitive decline and may attenuate the estimation of the effect of risk factors. Joint modeling with auxiliary data partially corrects these biases, and may facilitate the discovery of genetic and environmental risk factors, and more accurate characterization of their effects.