Food Substitution Modeling Approaches: A Methodological Study.

Background: The interpretation of results from observational studies in nutrition depends on how food intake is modeled. When adjusting for total energy or total food intake, substitution effects are estimated. Simulation studies show that the choice of modeling approach for estimating substitution effects strongly influences the results, with some models producing estimates opposite to the true simulated effect. However, few studies have investigated these different approaches using empirical data. Objectives: We compared different substitution models using fermented and nonfermented milk as the exposure and all-cause mortality as the outcome in a Swedish cohort of n = 1063 elderly men. Methods: Two substitution modeling strategies were compared: a nonspecified approach (adjusting for total energy) and a specified approach (adjusting for total energy and other foods). The specified approach was divided into a leave-one-out model (excluding substituted food) and a comprehensive leave-one-out model (excluding substituted food while adjusting for remaining foods). We also examined the influence of unit of choice [grams, kilocalories (kcal), or mixed] across all models. All substitution models ( Results: Nonspecified substitutions for nonfermented milk varied by unit of choice, with the mixed-unit approach (HR: 1.18; 95% CI: 0.95, 1.48) differing from the same-unit (kcal) approach (HR: 1.00; 95% CI: 0.80, 1.25). Fermented milk showed no estimated differences. The specified substitution models provided similar estimates across the unit of choice. Conclusions: In nonspecified substitution modeling, estimates differed for nonfermented milk. Fermented milk showed similar differences in both the nonspecified and specified leave-one-out models. Although estimates were generally consistent across specified substitution models, each estimate represents a distinct research question, and making this explicit is advised for future studies.