The Question
Population-level cognitive decline is well-characterized: roughly 0.5–1 point per year on standardized tests after age 65, modulated by years of education. But individual trajectories diverge enormously. Some 80-year-olds stay sharp; some 60-year-olds decline rapidly. Can ML identify the individual-level risk factors — depression, cardiovascular disease, physical inactivity — that separate these trajectories?
ADM Prediction (Made Before Running Models)
Predicted winner: ML. Age-education curves predict the population average well (Schaie’s Seattle Longitudinal Study showed ~0.5 SD decline per decade after 65). But individual decline rates vary enormously — depression accelerates cognitive loss (Dotson et al., 2008), cardiovascular disease damages brain vasculature (Gorelick et al., 2011), and physical inactivity compounds both. ML should capture these interactions.
Results
RMSE by Prediction Horizon
Feature Importance (Top 8)
Multi-Model Comparison
Accuracy by Trajectory Type
Cognitive decline rates differ by age group, sex, and education level. Does the ML advantage hold across all subgroups, or does it matter most for specific populations?
The ADM Insight
Cognitive decline is the outcome where individual trajectories diverge most from population averages. A 70-year-old with depression, diabetes, and physical inactivity follows a fundamentally different trajectory than an active, healthy 70-year-old with the same education. The ML model’s advantage grows with prediction horizon — precisely because individual divergence accumulates over time.
Cohort: HRS RAND respondents aged 50–90 with COGTOT (total cognitive score, 0–35) in 5+ waves. The COGTOT score combines immediate and delayed word recall, serial 7s, and backward counting — a well-validated telephone-administered cognitive battery.
Domain baseline: Age-education curve with published decline rates. Cognition declines approximately 0.5–1 point per year after age 65, modulated by years of education (each year of education offsets roughly 0.3 points of age-related decline).
ML model: GradientBoostingRegressor (200 trees, max depth 4, learning rate 0.1). Features include baseline cognitive scores, demographics (age, sex, education, race), chronic conditions (diabetes, heart disease, stroke, lung disease), depression (CES-D), BMI, smoking status, physical activity, and early cognitive trajectory (slope across first 2–3 waves).
Evaluation: Train-test temporal split (train on earlier waves, predict later waves). Bootstrap 95% CIs from 1,000 resamples. RMSE evaluated at 4, 6, 8, and 10-year prediction horizons.
Self-administered tests: COGTOT is based on the TICS telephone interview, not a clinical neuropsychological battery. Sensitivity to mild cognitive impairment is limited compared to in-person assessments like the MoCA or full neuropsych testing.
No biomarkers: HRS does not include ApoE4 genotype, amyloid PET, or tau levels. These biomarkers would significantly improve prediction, especially for Alzheimer’s-related decline.
Practice effects: Repeated administration of the same cognitive tests may inflate scores in later waves, as participants learn the test format. This can mask true decline, particularly in the early stages.
Survivorship bias: People who experience the most severe cognitive decline are more likely to drop out of the study (due to institutionalization, inability to complete the interview, or death). This means our sample underrepresents the most rapid decliners.