Are we there yet? Estimating the waves of follow-up required for stable effect estimates in cognitive aging research

Cognitive aging research relies on longitudinal data, but extended follow-up is costly. The extent to which estimates and precision from data with shorter follow-up diverge from estimates based on longer follow-up is unknown. The necessary follow-up period may depend on model specification, i.e., modeling the timescale as time-since-baseline or current age. We used data on adults age 65+ from the 2006-2018 U.S. Health and Retirement Study. For associations of 8 commonly studied dementia risk factors with cognitive decline, we compared coefficients and variance estimates to results from benchmark models (i.e., using 7 waves and/or specifying time-since-baseline). We varied the hypothetical follow-up length (1-7 waves, representing 0-12 years of follow-up) and timescale specification. Among individuals 65-80 years old at baseline, estimates of cognitive change in models with <4 waves of follow-up differed meaningfully in terms of both coefficients and variance from estimates using full follow-up, regardless of timescale specification. Differences by length of follow-up time were less pronounced among those >80 years of age at baseline, in part due to sample attrition. In models assuming equal follow-up duration, estimates of cognitive change specified by current age differed from estimates using time-since-baseline but were more precise, especially with shorter follow-up.