A cross-sectional exploration of cognitive ability across age via stacked ensembles.

Abstract

Age-related changes in cognitive and biological processes mean that older adults show markedly lower performance on cognitive assessments than younger adults. Characterizing the precise nature of age-related differences in cognitive performance and whether they vary as a function of key demographic characteristics has been challenging due to small effect sizes, underpowered samples, and blunt analysis methods. In the present study, we address these issues by using a massive cross-sectional data set of approximately 750,000 English-speaking participants who completed at least one battery from the NeuroCognitive Performance Test. We employ stacked ensembles, a machine learning approach, to model differences in age-related cognitive performance from 25 to 80 years based on gender and education. We utilize bootstrapping to quantify uncertainties and compare predicted performances across age, gender, education, and subtest while accounting for data variability. We then use clustering techniques to identify cognitive subtests with similar patterns across demographics. Our novel approach reveals several notable trends. For example, tasks reliant on semantic knowledge and fluid reasoning, such as completing patterns or arithmetic word problems, exhibit similar education-dependent variation. On tasks where men outperform women at early ages, men's predicted performance also shows greater decline across the age range, resulting in a narrower or nonexistent gender gap at older ages. We discuss additional age, gender, and education interactions, as well as variations in the magnitude and onset age of change in the predicted slope of performance, most of which appear dependent on the specific cognitive area being evaluated. Implications for theories of aging are discussed. (PsycInfo Database Record (c) 2024 APA, all rights reserved).