We trained various classes of
instrument-level models on the normative sample at V5. We
refer to these classes of models as
full, race, edu, adi, wrat, adi+wrat.
Each model-class was used to generate global z-Scores for the full V5 sample.
We followed the full V5 sample to see whether or not they had
ARIC-adjudicated dementia at V6 (cogstatus61 from
derive61).
For each model-class, we fit a logistic regression model on the
full V5 sample. For each model, the outcome was V6 dementia and
the predictor was the V5 global z-Score from a model-class.
(Example Stata command:
logistic v6_dem v5_globz_full.)
We then used each logistic regression model to predict a "risk score": probability for dementia at visit 6. The basic idea is something like "the higher your V5 global z-Score, the lower your risk for dementia at V6."
Suppose that we wanted to place risk scores into "low risk" and "high risk" categories. This categorization would depend on a "risk-cutoff." Based on this cutoff, we could count many participants were below and above the cutoff for a given batch of risk scores.
Then we could compare these counts with the counts corresponding to the batch of risk scores from the full model. This idea is devloped in more detail with these tables.