A team of researchers has developed and tested a method for assigning weights to risk impact indicators across a food system, combining structured expert elicitation with probabilistic analysis before the weights are fed into a multi-risk assessment. Reporting in Food Safety and Risk, the study used the French lentil food system as its case study and elicited judgements from eleven experts on sixteen indicators spanning agriculture, environment, economics, nutrition and food safety. The work addresses one of the most stubborn problems in multi-risk assessment: aggregating indicators that cannot be compared directly requires weights, and the weighting step has repeatedly been criticised as subjective and opaque.
The rationale for the exercise lies in the structure of food system evaluation. Assessing a system from farm to fork means tackling risks in several domains at once, and those risks can conflict. The authors note, for example, that controlling microbiological hazards through more stringent thermal treatments raises energy demand and therefore environmental risk, and may even increase chemical risk by generating neo-formed contaminants during processing. Within the food safety domain alone, multiple hazards must be considered, from environmental contaminants such as heavy metals and dioxins to pesticides from agricultural practice, furan formed during cooking, and substances like bisphenol A migrating from packaging. Scoring each of these indicators demands data and models, and doing so across several scenarios multiplies the workload.
When the goal is to rank alternative scenarios rather than simply describe them, the indicators must be aggregated. The researchers point to Multi-Criteria Decision Analysis based on outranking methods as a suitable tool for handling multiple, often conflicting criteria of a semi-quantitative and ordered nature. Such methods have been advocated for food system comparisons and were demonstrated by the FAO in evaluating interventions ranging from pre-harvest options to dietary modification. But outranking MCDA requires a weight for every indicator, and it is precisely this step that the literature has flagged as insufficiently transparent. The new study set out to make that step objective, transparent and robust.
The chosen vehicle was expert elicitation embedded within a broader Living Lab experience. Living Labs are open innovation ecosystems that bring together researchers, experts, policy-makers and other stakeholders in structured, participatory and iterative co-creation. The authors argue that situating elicitation in such a setting turns weighting into an explicit, collective and reflexive activity, in which assumptions, trade-offs and knowledge limitations can be openly discussed and documented, rather than a purely technical task performed behind closed doors.
The Living Lab ran for four years through exploration, experimentation and evaluation phases, and the elicitation described here took place as a participatory workshop in the experimentation phase. Eleven participants were purposively selected for expertise in food systems resilience, agricultural production, climate change adaptation, food policy, consumer behaviour and risk assessment. They came from both EU and non-EU countries, including the United States, Spain, France, Italy, the Netherlands, Denmark, Portugal and Greece, with five women and six men. All gave informed consent through a dedicated online form covering the research purpose, voluntary participation, GDPR-compliant data processing and a non-disclosure declaration. Although the case study focused on the French lentil system, the aim was not to capture local stakeholder preferences but interdisciplinary expertise relevant to farm-to-fork risk assessment.
The workshop unfolded in three parts. After an introduction to the lentil supply chain, its impact indicators and its scenarios, participants completed an Excel-based questionnaire individually, distributing 100 points across the sixteen indicators according to their importance. The question asked which indicator should be prioritised with climate change in 2050 in dried lentils, and it was answered under four scenarios: the current situation; a 2050 lentil system under climate change with increased consumption; a 2050 system with increased importation of fried lentils from Canada; and a 2050 system with increased local consumption. Facilitators answered clarification questions while limiting undue influence among participants. A group discussion of the scores followed, and individual scores were then submitted after the workshop for analysis in R.
The sixteen indicators covered the full continuum. They included cadmium as a representative heavy metal; the nutrients iron, selenium and zinc; the decrease in ischemic heart disease as a nutrition outcome; the pathogens Bacillus cereus and Listeria monocytogenes; economic measures comprising price to farmers, market price paid by consumers, willingness to pay, and the cost of illness; and environmental measures including global warming potential over 100 years based on IPCC 2021, land use, eutrophication potential, ozone layer depletion and acidification potential. After a pre-analysis check for completeness, one expert was discarded for missing responses, leaving ten. These experts collectively represented agriculture, financial cost economy, food science, food safety and environmental science, with several holding expertise in more than one domain.
The raw scores already revealed patterns. With sixteen indicators, an even split would give each a score of 6.25, but the distributions differed markedly. Price to consumer and land use attracted the highest weights, each with a 75th percentile above 10, while zinc received far less attention, with a 75th percentile of 5. The widest dispersion appeared for global warming potential, where raw weight scores ranged from zero to 50, reflecting sharply divergent views among experts from different backgrounds.
The in-depth analysis then introduced two refinements designed to increase robustness. First, rather than using raw points, a PERT distribution was constructed for each indicator from the 25th percentile, median and 75th percentile of the ten responses, deliberately excluding outliers so that isolated individual opinions would not dominate. Second, each expert’s self-rated confidence in their knowledge of each dimension, scored from 0 for limited knowledge to 100 for expert knowledge, was used to build an uncertainty interval around their weight through a linear function with an adjustable constant set at 500: the higher the confidence, the narrower the interval. Individual uncertainty was represented as a uniform distribution over that interval, while variability between participants was captured by the PERT distribution. A second-order Monte Carlo procedure, run with 100 iterations in the uncertainty dimension and 10,000 in the variability dimension using the mc2d package in R, then combined the two dimensions, and the final weights were normalised to sum to 100.
The resulting distributions separated variability from uncertainty, a distinction the authors stress is highly recommended in risk assessment and management. For price to consumer, the median weight was estimated at 7.18 with a 95% uncertainty interval of 7.15 to 7.21, ranging from a 2.5th percentile of 5.52 to a 97.5th percentile of 9.72, confirming the relatively high priority experts gave this indicator. Decrease in ischemic heart disease received a median weight of 5.02 with a 97.5th percentile of 6.17, below the even-split average of 6.25. Global warming potential, despite its raw dispersion, settled into a much narrower distribution with a median of 7.18, spanning a 2.5th percentile of 5.74 to a 97.5th percentile of 9.25, precisely because outliers were excluded from the final construction. The authors present the approach as a proof of concept: once obtained, the weights can be plugged into an outranking MCDA to rank alternative scenarios in a decision process, and if the method transfers successfully to other food systems, it could remove a key obstacle to multi-risk aggregation, namely the subjectivity of choosing each indicator’s weight. They also acknowledge practical limits, noting that weighting more than sixteen indicators risks expert fatigue, and that data availability for the underlying scores remains a separate scientific challenge.
Subject of Research: Expert elicitation and probabilistic weighting of risk impact indicators for a lentil food system multi-risk assessment
Article Title: Weighting risk impact indicators in a food system: expert elicitation and analysis prior to inclusion in a multi-risk assessment
Article References: Membré, J.-M., Guiziou, L., Altamore, S., Careccia, A., & Feliciano, R. J. (2026). Weighting risk impact indicators in a food system: expert elicitation and analysis prior to inclusion in a multi-risk assessment. Food Safety and Risk, 13(1), Article 32. https://doi.org/10.1186/s40550-026-00157-x
Image Credits: AI Generated
DOI: 10.1186/s40550-026-00157-x
Keywords: food safety, multi-risk assessment, expert elicitation, Living Lab, Monte Carlo, lentil food system, MCDA, risk indicators, uncertainty analysis, food systems, climate change scenarios, probabilistic weighting
News Source: Alan Morgan. (October 29, 2026). Experts Weigh 16 Risk Indicators for Lentils in Multi-Risk Food Assessment. Scienmag.
