A new paper argues that foods, in isolation, do not have a definitive effect on health. What matters is what they replace on the plate, and this change can change the way nutritional evidence is interpreted.
Opinion: Is this food healthy? Reframing nutritional evidence through counterfactual comparisons. Image Credit: Anna Puzatykh/Shutterstock
An opinion paper was recently published in the journal clinical nutrition Advocates for reframing nutrition research through counterfactual comparisons.
Despite decades of research, nutrition science continues to produce findings that are often considered context-dependent or inconsistent. Nutrition science often relies on proxy outcomes such as intermediate physiological measures and biomarkers, and their interpretation is dependent on underlying causal paradoxes. Dietary interventions are inherently structural, meaning that increasing the intake of one food requires reducing the intake of another food.
Thus, studies assessing the same food may reflect heterogeneous causal contrasts defined by different contexts and comparisons. Aggregating these contrasting contrasts in meta-analyses will yield estimates that may obscure the relationship between health and diet. In this paper, the authors argue that improving nutrition evidence synthesis warrants meta-analytic reframing through a causal inference lens that integrates the comparative context.
Counterfactual Framework in Nutrition Science
Moving beyond associative or descriptive explanations toward causal inference requires accounting for the counterfactual framework. Modern causal inference highlights that causal effects are defined relative to specific interventions and choices. Thus, meaningful causal interpretation is dependent on exposure specification and counterfactual comparisons.
The stationarity assumption is a key requirement: it states that, for a given risk, an observed outcome reflects the likely outcome associated with that intervention. This requires that the exposure represent a well-defined intervention, such that different treatment versions should not yield systematically different outcomes. If not, the estimated effect will be unclear.
For example, red meat consumption may refer to unprocessed lean meat, processed meat, or meat eaten with refined carbohydrate-rich foods or vegetables. Although these scenarios have the same exposure label, they are different interventions with different health effects and biological mechanisms. Therefore, treating such different versions as interchangeable risks weakens the causal inference.
Diet replacement and relational health effects
Many studies consider foods to have intrinsic effects independent of dietary context; Yet, the structural nature of the diet challenges this assumption. Modifying any one component of the diet does not happen alone; Instead, it matches specific scenarios based on how other dietary components are allowed to change.
This characteristic has implications for the definition of estimates and the interpretation of results. The paper distinguishes between effects that allow widespread dietary change and substitution effects that reflect replacing one food with another under constant intake. The health effects of a given food correspond to specific substitutions rather than an intrinsic property of the food. For example, a randomized controlled trial (RCT) Consumption of dried ham was compared to cooked ham (control).
While the intervention shows favorable changes in metabolic markers compared to controls, the interpretation critically depends on the nature of the replacement tested. If the comparator has less favorable effects, the observed benefit would indicate a relative improvement over the alternative diet rather than the inherent cardioprotective properties of the intervention.
Network meta-analysis and causal inference
Meta-analysis of RCTOften represent the best evidence, provided that the underlying studies examine the same causal question. However, many nutrition meta-analyses analyze overall effect estimates from different dietary contrasts without equivalent counterfactual comparisons. As a result, the pooled estimates lack a clear causal explanation.
These challenges do not mean that the evidence synthesis is intrinsically flawed. Instead, they suggest that traditional meta-analytic methods may be inadequate to account for compositional risks such as diet. In contrast, network meta-analysis (NMA) provides a methodological framework addressing some of these limitations by incorporating multiple comparators. NMA The relational nature of dietary interventions can be preserved by modeling competing choices.
In the examples discussed by the authors, NMA reveal differences specific to the comparators, whereas traditional meta-analyses may report minimal or neutral effects. Remarkably, NMAThis does not eliminate all challenges, as valid causal interpretation requires three key assumptions to be met: consistency (indirect and direct evidence are consistent), variability (studies remain comparable across treatment contrasts), and clinical comparability (interventions are sufficiently homogeneous).
Counterfactual Nutrition Research Implications
Overall, the limitations described here do not indicate a failure of the meta-analysis. Instead, they reflect a mismatch between the causal structure of dietary exposure and conventional evidence synthesis. approach like NMAComparators that preserve structure align better with causal inference.
Yet, methodological tools alone are not sufficient; Therefore, improving nutrition science requires reframing research questions to reflect well-defined counterfactual paradoxes. The authors also call for clearer exposure definitions and more transparent reporting of replacement references and energy balances. “Is the food healthy?” Moving from “What makes this food healthier compared to this one?” The translational relevance, coherence, and interpretability of nutrition science can be improved.
