Study Design
Is more always better? Implications of the number of cognitive assessments within a limited follow-up time on bias and precision of estimated effects on cognitive change Mary Thoma* Mary Thoma Jingxuan Wang Elizabeth-Rose Mayeda Charles McCulloch Maria Glymour Eleanor Hayes-Larson Jacqueline Torres
Background: Repeated cognitive performance assessments are necessary for dementia research, but longitudinal data collection is costly. Within a set follow-up period, similar estimates of association with cognitive change may be achievable with fewer assessments. We compared the bias and precision of cognitive change estimates (age effect) across models with varying total assessments across 8 years of follow-up.
Methods: Using data from the U.S. Health and Retirement Study among 65-80 year-olds in 2006 (N=7736), we estimated linear mixed models for associations of baseline diabetes, stroke, depression, and neighborhood safety with change in cognitive performance z-scores via interactions between each risk factor and current age. Reference estimates were derived from a model in a random 50% of the cohort with 5 cognitive assessments completed every 2 years over 8 years of follow-up (2006-2014). Comparison estimates were obtained from the other 50% of the cohort using (a) standard linear regression with cross-sectional data (2006); linear mixed models with (b) 2 assessments separated by 8 years (2006, 2014); and (c) 3 assessments each separated by 4 years (2006, 2010, 2014). Bias of estimates was defined as the absolute difference from reference model estimates. Design effects (DEFF) were calculated as the variance from a given estimate relative to the variance of the reference model estimate.
Results: Estimates of the four risk factors using cross-sectional data had substantial bias (mean across risk factors=0.013; SD=0.009) and DEFF (mean across risk factors=3.49; SD=0.32). Using 2 assessments the mean bias was 0.007 (SD=0.004) with DEFF=1.30 (SD=0.07). Using 3 assessments improved results only slightly to a mean bias of 0.005 (SD=0.004) and DEFF of 1.15 (SD=0.06).
Conclusions: When total follow-up time was limited to 8 years, using 2-3 equally spaced assessments resulted in minimal bias and loss of precision compared to using 5 equally spaced assessments.
