Statistical Analysis Beyond P-Values in the Artificial Intelligence Era of Food Science

Food science is entering an analytical era shaped by foodomics, spectroscopy, imaging, sensory science, and artificial intelligence, yet statistical practice still relies on low-dimensional workflows and significance testing. This review argues that such approaches are insufficient for the complexity and decision relevance of modern data. It revisits classical foundations, including experimental design, tests, assumptions, and reporting, and explains why p-value – centred interpretation is increasingly inadequate, especially when statistical significance is mistaken for chemical, technological, nutritional, or sensory relevance. Key challenges in food datasets are examined, including variability, small samples, batch effects, repeated measures, multivariate and compositional structure, sensory heterogeneity, and missingness. The review outlines modern inferential frameworks that improve evidence generation, such as effect sizes, interval-based inference, multiplicity control, resampling, Bayesian methods, equivalence testing, diagnostics, and uncertainty quantification. It also highlights multivariate models used in food chemistry and foodomics, including principal component analysis, partial least squares, clustering, and mixed-effects models, while assessing risks from machine learning. Ultimately, it emphasizes that robust, reproducible, and transparent statistical integration is essential for reliable, decision-oriented food science.