Why causal effects of ultra-processed foods cannot be identified: a systematic review of subtypes of ultra-processed foods and risk of type 2 diabetes

BackgroundOne of the key assumptions underlying causal inference is the assumption of consistency; that is, the absence of relevant multiple versions of the same treatment. Although ultra-processed food (UPF) as a dietary construct has been associated with a greater risk of type 2 diabetes (T2D), subtypes of UPFs may exert different strengths and directions of associations on T2D. The main objective of this study was to systematically evaluate the current literature on subtypes of UPFs and their associations with T2D and to contextualize these findings within the framework of counterfactual theory, with particular emphasis on the consistency assumption. MethodsWe searched PubMed and Embase for relevant articles. Cohort studies of adult men and women without T2D and with the exposure (subtypes of UPFs) and outcome (incident T2D) of interest were included. Two independent researchers screened the articles and evaluated risk of bias using the ROBINS-E tool. UPFs were categorized into eight subtypes. No meta-analyses were conducted. A post-hoc analysis of n=52,201 in the Danish Diet, Cancer and Health (DCH) cohort was performed to illustrate the implications of causal inconsistency. We formed a fictive food group (the terrible five) combining sugar sweetened beverages, processed meat, red meat, refined grains and vegetables and investigated the associations with T2D. ResultsOut of n=222 articles screened for eligibility, n=6 cohort studies with n=635,332 participants (average follow-up of 6-26.1 years) were finally included. Most studies on UPF breads and cereals; packaged snacks; and sauces, spreads and condiments showed no associations with T2D. Most studies on UPF ready-to-eat-dishes; meat-based products; and artificially and sugar-sweetened beverages were positively associated with T2D. UPF dairy products; and sweets, snacks and confectionaries were generally inversely associated with T2D, although with some variation. Overall risk of bias was considered moderate to high in all included studies. In the DCH cohort, a positive association between the "terrible five" and T2D was observed (HR: 1.14, 95% CI: 1.05, 1.23). Intake of vegetables was inversely associated with T2D (HR: 0.96, 95% CI: 0.95, 0.98). ConclusionThe systematic review showed different strengths and directions of associations on the risk of T2D across the different subtypes of UPFs. Aggregating these sources into a coarse food category such as UPF therefore violates the assumption of causal consistency and, as illustrated in our post hoc analysis, can mask individual associations. This study illustrates potential unintended consequences of broad implementation of UPF-related policies without further improvement of the definition or mechanistic understanding. RegistrationThe protocol was registered in PROSPERO (ID: CRD420251058458).