Heterogeneity Analysis of Associations Involving the Large-Scale Online MindCrowd Survey Memory Test

Alzheimer's disease and related disorders (ADRDs), as well as general age-related cognitive decline, are known to be multifactorial with heterogeneous etiologies. Identifying and accommodating heterogeneity in any one ADRD-related data set can be pursued using different analytical techniques, each with different assumptions or purposes. For example, whereas a great deal of research has explored clustering individuals or variables that exhibit greater similarity in some way, little research has explored evidence for heterogeneity in the relationships between relevant outcomes, such as performance on a memory test, and risk factors such as environmental exposures, behaviors, or genetic factors among individuals.

We explored evidence of heterogeneity in the relationships between ability on a memory test, specifically the paired associate learning (PAL) test, and multiple social and demographic risk factors using the large MindCrowd study database (n > 90,000 individuals). We focused on mixtures of regression models but compared models assuming many interaction effects among independent variables as well as random effects.

We ultimately find substantial evidence for heterogeneity and offer an intuitive explanation for it involving individual motivation for participating in the MindCrowd study. Basically, we argue that our mixture of regression model analysis results suggest that a smaller group of individuals (∼16%) likely participated in the MindCrowd study out of a concern for their cognitive abilities as they exhibit stronger and statistically significant negative associations between age, number of medications they are on, some ancestries, and the number correct on the PAL test. They also exhibit stronger positive associations between education and PAL test results in a dose-dependent manner suggesting that a "cognitive reserve" associated with greater education could benefit them. Analysis models assuming interaction terms and random effects suggested that other forms of heterogeneity in the relationships between variables exist in the data set, but their results do not carry with them the same intuitive explanation that the results of the mixture model analyses do.

We find evidence for heterogeneity in the relationships between social and demographic variables and PAL test results in the large MindCrowd study database. This heterogeneity is likely due to individuals with and without concerns for their cognitive abilities participating in the study. We also find other types of evidence in the data set. Our results should motivate caution in the use of large epidemiological study or survey-oriented data sets to build predictive models of clinical or subclinical pathologies without exploring or accommodating heterogeneity. Our results also suggest that one should include questions about motivation to participate in large epidemiological studies since different motivations may impact important relationships between independent and dependent variables.