Aging
Estimating Effects of Exposures on Memory Trajectories With Death and Dropout: Comparing Modeling Frameworks and Prediction Strategies L. Paloma Rojas-Saunero* L. Paloma Rojas-Saunero Rojas-Saunero Rojas-Saunero Rojas-Saunero Department of Epidemiology, Fielding School of Public Health, University of California Los Angeles
Observational studies on the effect of time-fixed exposures on memory decline in older adults commonly use linear mixed-effects models (LMMs) and summarize results by evaluating predicted memory trajectories at reference or mean covariate values, which differ from marginal estimates obtained from generalized estimating equations (GEE). We used data from the Health and Retirement Study to estimate the effect of completing high school versus not on memory decline over 24 years of follow-up. We included participants with a memory score assessment in 1998 (n=19208, mean baseline age=66.3 years) and constructed a binary indicator for high school educational attainment (yes/no). We estimated mean differences in memory function over time using a g-formula approach based on GEE with inverse probability censoring weights (IPCW) for attrition due to death and dropout. We also estimated mean differences in memory function over time from a single LMM with random intercepts, estimated using three prediction strategies: the g-formula, prediction at reference covariate values, and prediction at mean covariate values. For both GEE and LMM, we included exposure, practice effects, time with natural splines, interaction between time and exposure, baseline age, sex/gender, southern birth, parental educational attainment, and interaction between covariates and time. For the IPCW, we additionally included race/ethnicity and time-varying health-related variables. We calculated 95% confidence intervals with 500 bootstraps. High school attainment was associated with higher memory scores, with mean differences increasing over time. Mean differences estimated using GEE were slightly smaller than those based on LMM, and all three LMM-prediction approaches yielded identical estimates (Figure). Overall, mean differences over time varied across modeling approaches, not across prediction strategies; GEE with IPCW is a more explicit strategy for accounting for potentially informative attrition.
