A Minority of the Variance in Single-Recall Nutrient-Density Pattern Scores Is Between-Person: A Reliability and Regression-Calibration Framework

Background: Dietary pattern–mortality studies usually report one estimate from a single 24-hour recall, without asking what share of the variance in the resulting pattern score is between-person rather than within-person variation. Methods: We pooled ten National Health and Nutrition Examination Survey (NHANES) cycles (1999–2018) linked to the 2019 Mortality File, derived nutrient-density patterns from 26 harmonized nutrients by non-negative matrix factorization, and modeled cause-specific mortality with survey-weighted Fine–Gray subdistribution hazards. Results: The cohort comprised 47,257 adults and 7629 deaths over a median of 9.2 years. No reconstruction-error elbow appeared at any rank from k = 2 to 8, so the two retained components are reported as macronutrient-dominant axes, one fat-dominant and one carbohydrate-dominant, not as discovered food-based patterns. Between-person variation accounted for 31.6% and 35.2% of the variance of a single observed day’s score. After Holm–Bonferroni correction no pattern–mortality contrast reached significance. The strongest, the fat axis against other-cause mortality, gave a hazard ratio of 0.92 (95% CI: 0.82–1.04, padj = 1.000). Regression calibration against a second-day recall moved that estimate away from the null (0.80, 95% CI: 0.57–1.10) while widening its interval, quantifying the attenuation a single recall imposes. Conclusions: The reliability coefficients and calibration factors are reusable by other single-recall NHANES analyses.