|| Checking for direct PDF access through Ovid
Introduction: Continuous Net Reclassification Improvement (NRI) compares risk prediction between 2 models, but the NRI scale is hard to interpret as a difference of differences of correct vs. incorrect reclassification. We developed Discrimination Improvement Analysis (DIA), a mathematically-related alternative. We illustrate DIA in i3C comparing P(adult obese) estimated from childhood either using body mass index (BMI; base model) or BMI percentile (CDC Chart 2000; alternative model).Methods: In both DIA and NRI (described in the Table), reclassification is based on the regression discriminant to assess reclassification up when predicted risk is increased under the alternative model (>0), and down when the opposite is true (≤0). The DIA metric is the “added spread”: observed P(adult obese) in those reclassified up minus in those reclassified down. Using Bayes rule, continuous NRI is a weighted average of the P(adult obese) in those reclassified up vs down. We used first childhood BMI (age 3-19 y) and last adult BMI (age 30-50 y) in i3C. Logistic regression estimated P(adult obese) in the base model and the alternative model.Results: Of 8864 participants (47% men, 89% white, mean child BMI age 11 y (SD 4), mean adult age 38 y (SD 6)), 2436 were obese adults (last adult BMI≥30 kg/m2). Both absolute BMI and BMI percentile were independent predictors (Table). Average P(adult obese) in the base model was 24.1%. The alternative model significantly added spread of 11.2%: observed P(adult obese) 30.2% in those reclassified up vs 19.0% in those reclassified down, P<0.001). Using the continuous NRI yields a result qualitatively similar to DIA: 0.28 (95% CI: 0.25, 0.31).Conclusions: The DIA and NRI agreed that prediction of P(adult obese) was improved using BMI percentile over using absolute BMI, although DIA and NRI may disagree due to different weighting. The DIA method of quantifying added spread in prediction has the advantage of being on the natural scale. The DIA principle may also be applied to a continuous dependent variable.