Premium
Assessing the generalizability of findings from the Alzheimer’s Disease Neuroimaging Initiative to the Atherosclerosis Risk in Communities study cohort
Author(s) -
Power Melinda C,
Bennett Erin,
Glymour M Maria,
Gianattasio Kan Z,
Couper David,
Mosley Thomas H,
Gottesman Rebecca F,
Griswold Michael E,
Wei Jingkai,
Mehrotra Megha,
Stuart Elizabeth A
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.038874
Subject(s) - generalizability theory , cohort , medicine , atherosclerosis risk in communities , confounding , alzheimer's disease neuroimaging initiative , logistic regression , neuroimaging , biomarker , covariate , cohort study , disease , psychology , cognitive impairment , psychiatry , statistics , developmental psychology , biology , biochemistry , mathematics
Background The Alzheimer’s Disease Neuroimaging Initiative (ADNI) collects and shares high quality imaging, biomarker, genetic, and clinical data. Whether findings from ADNI generalize broadly is unclear, given its highly‐selected, predominantly white and well‐educated participants. We compared associations estimated in ADNI to those estimated in the Atherosclerosis Risk in Communities (ARIC) study, which initially recruited participants randomly from four US communities, to examine the potential generalizability of ADNI findings. Method We identified common risk factor, cognitive, and imaging variables at the ADNI screening/baseline visits and ARIC Visit 5. Data were pooled to estimate associations between risk factors and cognitive or imaging data, and between cognitive and imaging data, using adjusted linear and logistic regression models. Models included a term for cohort and its interaction with the variable of interest, allowing for cohort‐specific estimates and statistical evaluation of cohort differences. We repeated analyses on data subsets defined by race and cognition. Sensitivity analyses accounted for cohort differences in the impact of confounders by including cohort by covariate interactions. Result The proportion of estimated associations that differed significantly by cohort (interaction p‐value<.05) in primary analyses was 42% (range 25‐42% across subset and sensitivity analyses). Many differences were substantively meaningful (e.g. OR for APOE‐4 on amyloid positivity in ARIC: OR=2.75; in ADNI: OR=8.44; OR for any functional limitations on MMSE score < 25: OR = 5.05 (2.87 ‐ 8.90) in ADNI, OR = 1.34 (0.99 ‐ 1.81) in ARIC). Conclusion The proportion of associations that differed significantly between ADNI and ARIC was substantially higher than would be expected by chance. Differences may stem from inherent differences in the populations from which participants were recruited. This has implications for the generalizability of highly selected samples, including deeply phenotyped samples typically used for biomarker research.