Modeling frameworks in nutritional epidemiology matter: comparing isotemporal and time-lagged Bayesian and frequentist approaches of carbohydrate intake and adiposity

Background: Understanding how different modeling strategies affect associations in nutritional epidemiology is critical, especially given the temporal complexity of dietary and health data. Objective: To compare how different modeling frameworks—including isotemporal versus time-lagged designs and frequentist versus Bayesian inference—affect estimated associations between carbohydrate subtypes and adiposity. Methods: Longitudinal data of 415 adults from the NoHoW Study were used to investigate associations between four carbohydrate predictors (free sugars, intrinsic sugars, starch, and dietary fiber) and three indices of adiposity (body fat percentage, BMI, and waist circumference) as outcomes. Four statistical approaches were used contrasting frequentist and Bayesian methods across both isotemporal (concurrent measurement) and time-lagged (6-month temporal shift) frameworks. To specifically evaluate change in adiposity outcomes over time, we implemented additional baseline-adjusted longitudinal models. Results: Isotemporal and time-lagged models showed directional agreement for nearly all associations; in all but one case, the models either aligned in the direction of the association or differed only in relation to the null. However, time-lagged models identified statistically significant associations and produced larger effect sizes for body fat outcomes and for starch and fiber predictors. Other associations, including intrinsic and free sugars, were weaker and varied with model specification, losing statistical support under time-lagged models. Frequentist models exhibited greater variation across temporal frameworks, including one directional shift among significant associations. Effect estimates were substantially attenuated after adjustment for baseline adiposity. Discussion: Time-lagged modeling shifted associations between carbohydrate intake and anthropometric outcomes, with increased effect sizes and additional significant associations for starch and fiber, and fewer statistically significant associations for intrinsic and extrinsic sugars. In contrast to frequentist models, Bayesian models yielded more stable and consistent estimates across time-lagged and isotemporal frameworks, showing no differences in the directions of associations across temporal frameworks. Models unadjusted for baseline adiposity overstate dietary impacts; including baseline adiposity is essential to isolate true diet-change effects from initial weight. Conclusion: Our findings suggest that incorporating temporal structure, especially through Bayesian models, can uncover relevant relationships that concurrent models may overlook. This study demonstrates that model specification, both in temporal framework and statistical approach, meaningfully influences both the detection and interpretations of associations in nutritional epidemiology.