As predicted, regression-based methods picked up on variability not captured by norms, which may prove helpful for identifying early decline in those who start out as high performers. However, in our relatively young sample, it is still unclear which of these approaches will do the best job of targeting AD-specific change. Additional followup of this sample is needed to determine the overall sensitivity and specificity of these approaches for identifying patients who are on a trajectory to develop MCI and AD. In the meantime, clinicians may want to consider both absolute and relative low performance for a given individual when assessing patient risk of AD. Distribution of classifications: In general, Method 1 identified fewer people as performing worse or better than expected than did either of the other methods. The average percentage (worse, better) across the seven outcomes was (4, 6) for Method 1, (6, 9) for Method 2, and (6, 10) for Method 3. Furthermore, the average performance of those identified by Methods 2 and 3 was less extreme (Table 1). This suggests that Methods 2 and 3 allow one to identify more subtle declines. Model agreement: The classifications made by Models 2 and 3 had good to very good agreement for all outcomes (weighted kappa values .71-.83). However, Model 1 did not agree well with either of the other two (weighted kappa values .08-.38), suggesting that norm-based methods and regression-based methods do not identify the same people as low performers (Fig. 1-4). Overlap of cognitive indicators: Poor performance on one outcome did not generally transfer to poor performance on another outcome. Figure 5 shows the relative proportion of low performers who were classified as “Worse” on multiple composite score outcomes by Methods 1, 2, or 3. TaBleS & figureS