Causal Analysis of Multidimensional Dietary Data to Assess Effects on All-Cause Mortality

Background: Methods applied under explicit causal assumptions can provide estimates that support potential causal interpretations of the effects of dietary factors on health outcomes. However, the high dimensionality inherent in dietary data presents a challenge. Objectives: Using multivariate analysis methods under causal assumptions, we identified dietary patterns and estimated their associations with all-cause mortality, as well as the effects of a 100 g/day increase in individual components. Methods: Data from 12,635 individuals aged 20 years and above from the National Health and Nutrition Examination Survey (NHANES), United States, were analyzed. K-means clustering was first used to identify dietary patterns, and then their associations with mortality risk were estimated both with and without inverse probability weighting (IPW). Second, the multivariate generalized propensity score (mvGPS) method was employed to estimate the average effects of dietary components on all-cause mortality under causal assumptions. Third, mutually adjusted models (non-mvGPS) were utilized to determine the effects of each dietary component. Relative risks (RR) and 95% confidence intervals (CI) were computed using fully adjusted Poisson generalized linear models. Results: In a 15-year follow-up period, 400 (3.2%) deaths were recorded. ‘Healthy’, ‘unhealthy,’ and ‘mixed’ dietary patterns were identified. Variations in estimates of ‘healthy’ and ‘unhealthy’ patterns with mortality were observed in non-IPW (RR = 0.96; 95% CI: 0.67–1.13 and RR = 0.79; 0.56–1.11) and IPW models (RR = 0.75; 0.55–1.04 and RR = 0.92; 0.63–1.36, respectively) compared to the ‘mixed’ pattern. In the mvGPS model, added sugar (RR = 1.21; 1.06–1.36), processed meat (RR = 1.20; 0.96–1.48), and legumes (RR = 0.82; 0.73–0.90) showed the strongest effects. Only whole grains (RR = 0.68; 0.46–0.98) had an effect in the non-mvGPS model. Conclusions: Applying mvGPS to multidimensional dietary data may help improve covariate balance across measured confounders and support more interpretable analysis of correlated dietary components. However, findings from this observational study should be interpreted as estimates under explicit causal assumptions, rather than definitive causal effects. Future studies should apply and further evaluate these approaches in larger and more diverse populations.