Introduction
The global food production system is one of the largest contributors to climate change, estimated to be responsible for 21–37% of anthropogenic greenhouse gas (GhG) emissions (Poore and Nemecek, Reference Poore and Nemecek2018; Willett et al., Reference Willett2019; Mbow and Rosenzweig, Reference Mbow and Rosenzweig2022). Population growth and the global shift towards resource-intensive ‘Western diets’ (OECD-FAO, 2023) are expected to intensify the environmental impact of food production, highlighting the urgent need to find more sustainable ways to grow and consume food (Juniper, Reference Juniper2020).
Dietary changes, particularly a shift to lower meat consumption, have the potential to significantly reduce GhG emissions (Committee on Climate Change, 2020; IPCC, 2023). Meat and animal products are emissions-intensive, generating an estimated 60% of food-related total releases (Xu et al., Reference Xu2021). Large differences exist in GhG emissions (kg CO2e/kg) between different protein sources, such as beef (37), pork and chicken (10–12), or tofu and beans (0.5–3), so shifting to diets with a lower meat content can significantly reduce emissions (Poore and Nemecek, Reference Poore and Nemecek2018).
This study focuses on how meal choices can be influenced to encourage more sustainable consumption in canteens. Canteens (alternatively referred to as cafeterias) serve over five million meals per day in the UK across workplace and institutional venuesFootnote 1 (Compass Group, 2023; IBISWorld, 2024) and form an important part of the out-of-home food consumption landscape. Among the primary strategies to encourage more sustainable diets in canteens are: (i) adjusting the availability of different meal options, (ii) altering relative prices, (iii) modifying the choice architecture (often referred to as ‘nudges’) and (iv) providing information (Wirsenius et al., Reference Wirsenius, Hedenus and Mohlin2011; Godfray et al., Reference Godfray2018). This study specifically assesses the impact of carbon labelling, an intervention that combines elements of choice architecture and information provision. Carbon labelling is often considered a practical, entry-level approach for influencing consumer behaviour, as it is relatively unobtrusive, highly visible and involves the consumer in the carbon reduction journey (Maier and Fesenfeld, Reference Maier and Fesenfeld2024).
Carbon labelling seeks to address the lack of consumer understanding of the externalities of food production and enable consumers to make socially responsible choices (Moran, Reference Moran2021). Camilleri et al. (Reference Camilleri2019) found that consumers consistently underestimated the environmental impact of their diets, as well as the differences between high- and low-carbon food types. Their research showed that carbon labels can reduce these consumer misperceptions and enable a more informed choice. Moreover, carbon labels often employ traffic-light colour schemes to increase the salience of ‘good’ and ‘bad’ choices, with the aim of simplifying decision-making and guiding consumers towards more climate-friendly options (Reisch et al., Reference Reisch, Sunstein and Kaiser2021; Schulze Tilling, Reference Schulze Tilling2023).
There have been several field experiments testing the impact of carbon labels in canteens in Europe and the UK, which have observed small, positive changes in meal choices and emissions (Spaargaren et al., Reference Spaargaren2013; Brunner et al., Reference Brunner2018; Slapø and Karevold, Reference Slapø and Karevold2019; Lohmann et al., Reference Lohmann2022; Beyer et al., Reference Beyer2023). These studies have focused almost entirely on university canteens, although these make up only a small portion of catered meals and tend to serve a young, well-educated demographic, which is unusually pro-environmental, pro-vegetarian and open to change messaging (Chang et al., Reference Chang2023). In contrast, workplace canteens serve half of all catered meals but only two studies have been conducted in these settings, neither of which found any impact from labelling (Pechey et al., Reference Pechey2022; Bschaden et al., Reference Bschaden2024).
Pechey et al. (Reference Pechey2022) researched the impact of eco-labels (presented as letters A–E) on 111,837 meal choices in 28 manufacturing and distribution workplace canteens in the UK, and found no evidence that labelling improved the sustainability of meal purchases. The authors did, however, note several limitations in the trial’s design and execution, including concerns over label salience and the meal banding methodology, as well as constraints associated with conducting the trial during the COVID-19 pandemic (Pechey et al., Reference Pechey2022). Similarly, Bschaden et al. (Reference Bschaden2024) found no significant effects of carbon footprint information (presented using cloud symbols) in three company canteens. However, the latter experiment employed simple before-and-after comparisons, which may be confounded by unobserved time-varying factors. There is therefore a need for further research, employing causal experimental designs in workplace catering venues, to understand whether carbon labels offer widespread potential for change (Pechey et al., Reference Pechey2022).
To contribute to this research, the objective of this study was to assess whether the presence of carbon labels had an impact on the sustainability of food choices in a workplace canteen setting. We conducted a natural field experiment in collaboration with four worksite cafeterias catering to call centre employees in the North of England. We utilised a difference-in-differences (DID) methodology to estimate the causal effect of carbon labelling. Three hypotheses were tested:
H1: Carbon labels will stimulate cafeteria diners to choose fewer meat dishes and more vegetarian dishes.
H2: Carbon labels will stimulate cafeteria diners to choose fewer high-carbon dishes and more low- or medium-carbon dishes.
H3: Cafeterias where diners are provided with carbon labels will have a lower average carbon footprint in the intervention period than those where diners are not provided with carbon labels.
Method
Ethics approval was granted for the research on 3 October 2023 by the Cambridge Institute for Sustainability Leadership, University of Cambridge. The study was pre-registered prospectively on the Open Science Framework on 13 November 2023.Footnote 2
A natural field experiment was conducted in four workplace cafeterias where one intervention, the addition of carbon labelling to the main hot meal lunch options, was evaluated. Two cafeterias received the intervention and the remaining two acted as a control (see Appendix Section 3 for details of assignment of sites to treatment or control groups). The experiment was run between October and December 2023. The sites were cafeterias in call centres, and the majority of diners were call centre employees who ate lunch in the cafeterias regularly. The cafeterias operated a standardised central menu, so the same hot meal options (one meat, one vegetarian) were served each day across all four sites. Employees paid for their meals, and the meat and vegetarian options were priced the same.
The carbon footprint values were calculated based on ingredient-level data supplied by the catering company for each hot main meal. Three Life-cycle Assessment Databases (Clune et al., Reference Clune, Crossin and Verghese2017; Hilborn et al., Reference Hilborn2018; Poore and Nemecek, Reference Poore and Nemecek2018) were used, with the value most relevant to UK food procurement selected. Emissions were expressed in kg CO2e per meal and included emissions from ‘cradle to retail’ as described in Poore and Nemecek (Reference Poore and Nemecek2018) dataset. Calculations were made for 48 unique cafeteria main meals, including an analysis of 1,300 ingredient lines.
Labels were placed on the menus as well as on stands beside the dishes at the counter for maximum salience. The labels were designed following best practices (Thøgersen and Nielsen, Reference Thøgersen and Nielsen2016; Taufique et al., Reference Taufique2022), using a traffic-light colour scheme (red, amber and green) and footprint logo. To account for colour-blind diners, the words HIGH, MEDIUM or LOW were included, while the words CARBON FOOTPRINT were incorporated to clarify the concept, as some diners in testing thought the footprint related to running or health. See Figure 1 for the design of the labels, and Figure 2 for a picture of the labels deployed in one of the treatment cafeterias.
Label design deployed in this study.
Labels deployed in the study, displayed on menus and stands beside the meals. Menu image anonymised for confidentiality purposes.
The intervention was advertised as a new cafeteria sustainability initiative and was accompanied by an information campaign. Information was displayed in the canteen, on the digital screens, on the workplace intranet, and on the app, explaining why the labels were being introduced (for images, see Appendix 5). Further context was provided on how meal choices can impact the environment, through comparisons with emissions from a car journey or phone charge. Participants were not made aware that the intervention was being evaluated in a research study.
Fidelity to protocol was monitored through the daily sharing of counter photos for all sites, via WhatsApp – these included images of the labels for intervention sites. There was regular communication between the researcher and the sites. The small number of sites and the strong motivation of canteen managers enabled strong fidelity to protocol, and the labels were applied correctly every day.
Daily numbers of each meal sold were obtained from the electronic sales registers. The final sample size was determined by logistical constraints, as well as pre-defined exclusion criteria. Post hoc power calculations are provided in Appendix Section 2. Data on pre-intervention meal choices were collected over a baseline period of 5 weeks, labels were then applied daily in two of the sites, and data were collected for the 6-week intervention period. Certain data exclusions were determined from the outset – Fridays were excluded as menu variations meant the menus were not standardised. Additionally, further exclusions were made after the data was collected, based on pre-determined criteria. Any days were excluded, across all sites, if the standardised menu was not served. This ensured that like-for-like meal and label distribution was maintained across all sites over the entire study period.
The trial data were analysed using Stata (Stata Statistical Software: Release 18.0). A difference in differences approach was used, which assessed (1) the difference over time in the treatment group (before and after the labels), (2) the difference over time in the control group and (3) the difference between these two differences, which denoted the intervention effect. Formal regression models are presented in Appendix Section 1. We additionally estimated two-way fixed effects (TWFE) regressions with unit (site) and time (date) fixed effects (see equation 2), providing our main results. Identification of a causal effect relies on the assumption of parallel trends, meaning that treatment and control groups would have followed similar trends in the absence of the intervention. Although this assumption is inherently untestable, sites were carefully selected to increase the likelihood that it would hold: all sites were required to be on a standardised menu cycle and to serve the same population of employees (i.e. call-centre workers) in the North of England. Moreover, we find no evidence for diverging trends between treatment and control prior to the intervention period (see Appendix Figure A1).
Models were estimated by ordinary least squares (OLS) and standard errors were clustered by site, to account for the dependence of observations within sites (repeated daily observations of the same individuals with similar habits and experience over time, aggregated at the site level). We use OLS as our primary modelling approach to facilitate the interpretation of the interaction term in our DID framework (Puhani, Reference Puhani2012) and account for the small number of clusters using the wild cluster bootstrap-t procedure (Cameron et al., Reference Cameron, Gelbach and Miller2008). For robustness, we estimate zero and one-inflated beta regressions, which better account for the observed distributions of our proportional meal choice outcomes (see Appendix Figure A3 and Table A3).
Taken together, the key trial design elements (standardised menus, DID design, rigorous carbon footprint calculations and daily fidelity checks) enabled us to conduct a field experiment with high internal validity, isolating the behavioural effects of the labels and producing causal estimates with strong ecological validity.
Results
The results included 4,068 meals, with an average of 29 meals per cafeteria per day. The total number of observations in the trial was 140; this represented the number of individual cafeteria days, the unit of analysis in the trial. Table 1 outlines the average meal shares and carbon emissions across the full trial period. Distributions of the outcomes are shown in Appendix Figure A3. On average, the uptake of vegetarian meals was 8% across the trial period, with 92% of meals sold being meat-based. This was in the context of 50% vegetarian and 50% meat dish availability, with one of each type served each day. The uptake of high-labelled meals was 33%, medium-labelled meals was 61% and low-labelled meals was 6%. The average daily emission per meal was 1.68 kg CO2e per meal, with a wide range between 0.76 kg CO2e and 4.96 kg CO2e .
Descriptive statistics
Notes: Table reports average sales shares (columns 1–5) and average emissions (columns 6–7) across the full trial period. Distributions of the outcome variables are shown in Appendix Figure A3.
The results of the trial are presented in Table 2, organised by the three hypotheses: Columns 1 and 2 relate to Hypothesis 1, focusing on the uptake of meat and vegetarian meals; Columns 3–5 relate to Hypothesis 2, comparing uptake across high-, medium- and low-carbon impact labels; and Columns 5 and 6 examine Hypothesis 3, focusing on canteen emissions. We present DID estimates obtained from linear regression analysis, with standard errors clustered at the site-level using the wild-bootstrap procedure. Panel A presents estimates of the canonical DID model, with indicators for treatment group and intervention period, as well as their interaction. Panel B presents estimates from TWFE regressions, our primary specification. These results are described in the following three sections with reference to our main pre-registered hypotheses. Estimates from zero- and one-inflated beta regressions are reported in Appendix Table A3 and are largely consistent with the results presented here.
Main statistical results
Notes: Statistical significance denoted as
* < 0.1, ** < 0.05, *** < 0.01. p-values in parentheses estimated using wild cluster bootstrap procedure. The outcome variables in columns (1)–(5) are continuous site-by-day sales shares. The outcomes in columns (6) and (7) represent emissions quantities, measured in kilograms of CO₂-equivalent (kg CO₂e). DID refers to the difference in differences, which is the intervention effect of the labels. Panel A presents estimates of the standard DID model without controls (see equation 1 in Appendix Section 1). Panel B presents estimates of the TWFE model (equation 2), including site and date fixed effects, as well as controls for the number of sales. Full regression output reported in Appendix Tables A4 and A5. Columns (3)–(5) exclude days on which no low-, medium- or high-carbon emitting options were available on the menu, respectively, as these options could not be selected on those days. This is reflected in differences in the observation numbers, and in DID Estimators, which across columns (3)–(5) do not sum to zero. Estimates from zero- and one-inflated beta regressions are reported in Appendix Table A3.
Hypothesis 1 asserted that carbon labels would stimulate cafeteria diners to choose fewer meat dishes and more vegetarian dishes. This hypothesis was confirmed with the results showing that following the application of labels, the uptake of vegetarian meals was relatively 1.5%pts higher in the treatment sites compared to the control sites. As shown in Table 2, the difference is statistically significant (p = 0.034) and robust to the inclusion of additional control variables and fixed effects (Panel B). In Appendix Figure A1, we explore dynamic treatment effects grouped by calendar week. The results suggest that the overall increase is primarily driven by a rise in vegetarian meal uptake during the first week of the labelling intervention (b = 0.05, p = 0.042).
Hypothesis 2 asserted that carbon labels would stimulate cafeteria diners to choose fewer high-carbon dishes and more low- or medium-carbon dishes. This hypothesis was not confirmed. The results in Table 2 indicate that, following the application of labels, the uptake of high meals was 2.0%pts lower (p = 0.492) in the treatment sites compared to the control sites, while medium labelled dishes remained similar (0.3%pts lower, p = 0.786) and the uptake of low labelled dishes rose (1.1%pts, p = 0.378). While this pattern aligns with the observed shift from meat to vegetarian dishes, the estimates were not statistically significant. This study, therefore, could not confirm any effect on the uptake of high-carbon (vs medium or low) meals as a result of labelling.
Hypothesis 3 asserted that cafeterias providing carbon labels would have a lower average carbon footprint during the intervention period compared with those without labels. This hypothesis was not confirmed. The results in Panel B of Table 2 show that, following the application of labels, emissions in the canteens where labels were applied did not materially change (0.008 kg CO2e or 0.5% increase) relative to those that did not receive the intervention. While this result contrasts with the observed increase in vegetarian and low-labelled dish choices, once controlling for site and date fixed effects (such as daily menu-specific effects), the estimate is not statistically significant (p = 0.338). In contrast to average canteen emissions, total cafeteria emissions decreased in line with the observed shift from meat to vegetarian; however, the estimates do not reach statistical significance (p = 0.166). Overall, it was not possible to demonstrate any significant impact on canteen emissions following the label intervention.
Discussion
This study sought to test whether the benefits of carbon labelling evidenced in previous university studies could be replicated among the wider workplace population. It showed that carbon labels, accompanied by an explanatory information campaign, did encourage consumers to make more sustainable food choices, but that the behavioural shift evidenced in these worksite settings was small. During the intervention period, there was a 1.5%pt shift from meat to vegetarian meal uptake in the sites where labels were introduced compared to the sites without labels. This effect was almost entirely driven by a 5%pt rise in vegetarian sales during the first week of the intervention, likely due to its novelty, with the effect fading by the second week. There were no consistent, statistically significant changes in the uptake of high-labelled meals, or importantly, in overall cafeteria emissions.
This trial is the first in a workplace cafeteria to evidence any behavioural impact of carbon labels. However, compared to previous university studies, the results were less pronounced, with workplace diners responding to the labels only during the first week of implementation (Appendix Figure A1). Trials in university canteens observed shifts from meat to vegetarian uptake of 1.7–4.6%pts, shifts from high to medium/low labelled dishes of 2.4–2.7%pts and reductions in cafeteria emissions of 1.5%–5.0% (Spaargaren et al., Reference Spaargaren2013; Brunner et al., Reference Brunner2018; Lohmann et al., Reference Lohmann2022; Beyer et al., Reference Beyer2023).
The results must be assessed in the context of dramatically different baseline consumption patterns that exist in call-centre workplace canteens compared to universities. The most readily available measure that demonstrates these differences is the vegetarian meal uptake. Diners in this workplace trial had a baseline vegetarian meal uptake of 7%, compared to between 15% and 66% in the university studies where this measure was reported (Brunner et al., Reference Brunner2018; Slapø and Karevold, Reference Slapø and Karevold2019; Lohmann et al., Reference Lohmann2022; Beyer et al., Reference Beyer2023). Where canteens start from a low baseline vegetarian uptake, it appears more challenging to encourage substantial percentage point shifts. The baseline should therefore be taken into account in expectation-setting and assessment when these interventions are implemented.
The differences in baseline behaviours and also diners’ responsiveness to labelling are likely to reflect the different demographics of students vs the non-student workplace population. The call centre workforce tends to be older, have a lower average formal education level, and potentially lower disposable income than the student population, once accounting for family obligations. Systematic studies on carbon labelling in food retail and hospitality suggest that these factors are associated with reduced responsiveness to carbon labelling (Potter et al., Reference Potter2021; Rondoni and Grasso, Reference Rondoni and Grasso2021; Hargreaves et al., Reference Hargreaves2022; Majer et al., Reference Majer2022; Chang et al., Reference Chang2023).
The trial results indicate that where worksite cafeterias have ambitious carbon reduction targets (potentially as part of Scope 3 GHG reduction targets), carbon labelling alone is unlikely to be sufficient to meet those targets. Where transformative change is required, and where cafeterias want to implement carbon labelling, a multi-modal approach, combining carbon labelling with one or more other interventions may hold potential (Fesenfeld et al., Reference Fesenfeld2020; Van Der Linden et al., Reference Van Der Linden, Pearson and Van Boven2021). There is some evidence that combining interventions which seek to affect deliberate decision-making processes (e.g. through information and labelling) with those impacting implicit decision-making (e.g. through choice architecture) can be synergistic, with the result greater than the sum of the individual approaches (Banerjee et al., Reference Banerjee2023; Chang et al., Reference Chang2023). Interventions that could be used alongside carbon labelling include increasing the relative availability of vegetarian or low-carbon dishes compared to meat or high-carbon dishes (Garnett et al., Reference Garnett2019), introducing plant-based default menus (Meier et al., Reference Meier2022), reducing meat portion sizes or reformulation of recipes to reduce carbon (e.g. mixing meat with plant-based ingredients) (Robertson et al., Reference Robertson, Andersson and Lunn2023), or the creation of pricing differences to encourage sustainable choices (Garnett et al., Reference Garnett2021).
Taking a wider perspective, several broader policy implications arise from this research. The lack of an established carbon labelling scheme meant that multiple decisions needed to be taken during the trial on how to implement the labels. Inaccuracies and missing values in the system that was intended to automatically generate the LCA (life cycle analysis) carbon footprints meant implementing labelling became a labour-intensive, manual process. Furthermore, the lack of familiarity with, or trust of, the labels among the diner population may have limited the understanding and effectiveness of them in this trial, despite the information campaign. If carbon/eco-labelling is to become widely established, a governmental (or pan-governmental) labelling scheme with consistent label design and corresponding LCA dataset would be of benefit to facilitate implementation and ensure consumer trust in the information provided (Gorton et al., Reference Gorton2021).
The main limitation of this experiment was the small number of observations, which was determined by the timeframe and number of selected sites, as well as the relatively high number of exclusions needed to ensure data consistency. There is thus a continued need for more real-world studies to test the effectiveness of carbon labelling in workplace settings over longer periods of time or including more sites. Moreover, future research could test multiple combinations of interventions (Alt et al., Reference Alt2024), as well as employ qualitative methods to understand the motivation of consumers and the wider impact of labelling on their decision-making outside the canteens.
There is an urgent need to reduce the impact that food production is having on the environment. Carbon labelling is one of the few interventions that educates and involves the consumer, and as such, it can play a valuable role in the carbon reduction journey. Indeed, labelling may enhance the effectiveness of other canteen interventions and potentially make more intrusive policies more acceptable (Taufique et al., Reference Taufique2022). However, on its own, labelling is not a panacea and is insufficient to drive transformational change, so where catering companies and their clients have the ambition to materially reduce cafeteria emissions, labelling should be applied alongside other interventions.
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/bpp.2025.10023.
Competing Interests
The authors declare no competing interests.
Open access
