Impact of increasing vegetarian availability on meal selection and sales in cafeterias

Significance

Reducing meat consumption in higher income countries is vital to protect the environment and improve public health. Few studies have tested the real-world performance of different strategies to increase plant-rich diets, and none has examined the impact of altering the availability of vegetarian meal options. In robust observational and experimental studies, we show that doubling the proportion of vegetarian meals offered increases vegetarian sales by between 41% and 79%. Our study assesses the impact of increasing the proportion of plant-based meal options on selection and is based on over 90,000 meal choices. We suggest that our findings have the potential to make a significant contribution to the global ambition for more sustainable diets.

Abstract

Shifting people in higher income countries toward more plant-based diets would protect the natural environment and improve population health. Research in other domains suggests altering the physical environments in which people make decisions (“nudging”) holds promise for achieving socially desirable behavior change. Here, we examine the impact of attempting to nudge meal selection by increasing the proportion of vegetarian meals offered in a year-long large-scale series of observational and experimental field studies. Anonymized individual-level data from 94,644 meals purchased in 2017 were collected from 3 cafeterias at an English university. Doubling the proportion of vegetarian meals available from 25 to 50% (e.g., from 1 in 4 to 2 in 4 options) increased vegetarian meal sales (and decreased meat meal sales) by 14.9 and 14.5 percentage points in the observational study (2 cafeterias) and by 7.8 percentage points in the experimental study (1 cafeteria), equivalent to proportional increases in vegetarian meal sales of 61.8%, 78.8%, and 40.8%, respectively. Linking sales data to participants’ previous meal purchases revealed that the largest effects were found in the quartile of diners with the lowest prior levels of vegetarian meal selection. Moreover, serving more vegetarian options had little impact on overall sales and did not lead to detectable rebound effects: Vegetarian sales were not lower at other mealtimes. These results provide robust evidence to support the potential for simple changes to catering practices to make an important contribution to achieving more sustainable diets at the population level.

High-income countries produce and consume animal-derived food (meat, fish, dairy, and eggs) at levels that are incompatible with meeting greenhouse gas emissions (GHGE) reduction targets (1). Livestock and aquaculture are responsible for 56 to 58% of the global food system’s GHGE and use 83% of farmland despite contributing just 18% of calories and 37% of our protein (2). In particular, meat from ruminants (cows, sheep, and goats) has average GHGE per kilogram 5-fold higher than pork, 7-fold higher than chicken, and 43-fold higher than legumes (3). Shifting toward a more plant-based diet is therefore 1 of the most effective ways of reducing the environmental footprint of food (2, 4). For the United Kingdom, it is estimated that switching from a high-meat diet (>100 g⋅d⁻¹) to an entirely vegetarian diet would reduce the GHGE of a typical person’s food by 47% (5).

Shifting diets to achieve sustainability outcomes is likely to require an array of strategies for changing human behavior (6, 7). Education to bring about behavior change is a popular and uncontroversial method; however, while it can raise awareness, it appears to be largely ineffective at actually changing behavior (8, 9). Models suggest that taxes on the most polluting foods would result in savings of 1 gigatonne of GHGE worldwide (4), but these taxes can be regressive and are politically unpopular given their lack of public support (8). A third group of interventions—changing the physical, economic, and social context (the so-called choice architecture) in which decisions are made—could potentially deliver improved environmental outcomes at a low cost and with little controversy, but this group has received relatively little empirical attention to date (10–13).

As 1 form of nudging, altering the relative availability of different food types has shown promise as a lever for changing dietary behavior to improve population health. Reducing the availability of high-calorie foods is estimated to be the third most effective strategy for combatting obesity after lowering portion size and reformulation, although the evidence for subsequent behavior change is rated as “limited” (14). A Cochrane review (15) found only 5 studies on altering availability that met the inclusion criteria (16–20), with a meta-analysis showing a nonsignificant decrease in consumption and a large significant decrease in selection. Other studies on availability not included in the Cochrane review have found increasing the relative availability of low- and moderate-fat entrées in a US school cafeteria from 33 to 50% increased their selection by 108% and 63%, respectively (21), and in 4 English workplace cafeterias, decreasing the number of high-calorie cooked meals offered to 1 option per lunchtime (while keeping the total number of options offered constant) reduced the mean energy per main meal sold by 26.1% (22).

Turning to reducing meat consumption, a recent review found no studies on the effects of changing the availability of plant-based meals (13). The likely patterns are hard to anticipate: At 1 extreme increasing relative availability might have a directly proportional impact on relative sales; conversely, if people have fixed preferences for meat or vegetarian meals, changing their relative availability might have no impact. It is important in such work that outcomes are assessed over sustained periods, because effects can wane over time (23, 24), and, if possible, that interindividual variation is examined too: An online study altering menu configurations found different responses between those who frequently or infrequently ate vegetarian foods (25). However, we are aware of only 1 study (again focused on health rather than meat consumption) that presents long-term individual-level data on how availability affects food choices (26). There are 2 further considerations: For any intervention to be acceptable to caterers, it is important that total sales and revenue do not substantially drop as a result (24, 27), and to have a genuinely additional environmental effect, it is important there are no sizeable rebound effects (28) whereby meat consumption increases on other occasions. However, almost no studies address rebound effects or effects on total sales (24).

To tackle these research gaps, we conducted 2 studies—1 observational and 1 experimental—in 3 college cafeterias at the University of Cambridge. These studies examined the effect on vegetarian sales of increasing the proportion of vegetarian options available (hereafter “availability”). We tested the hypothesis that meal selection is influenced by availability, such that increasing the availability of vegetarian options increases their selection. In these studies, we take advantage of year-long and anonymized individual-level data to analyze whether increasing vegetarian availability had effects that differed with the prior levels of vegetarian meal consumption of individual diners, affected total sales, or resulted in rebound effects at other mealtimes when vegetarian availability was not altered.

Research Setting

We collected data from 3 University of Cambridge college cafeterias during weekday term-time lunches and dinners (the university’s colleges are broadly equivalent to halls of residence). All colleges already varied the number of total meal options and vegetarian options served at lunch and dinner. Vegetarian options contained no meat or fish, but may have included eggs and dairy products; vegan options were entirely plant-based, and therefore contained no eggs or dairy products. Approximately 30% of the vegetarian options on offer were vegan. Hereafter, vegetarian and vegan options are both referred to as “vegetarian.” Study 1 comprised nonexperimental data of 86,932 hot main meals (hereafter referred to simply as “meals”; salads and sandwiches were not included) from colleges A and B across lunch and dinner during the spring, summer, and autumn terms in the 2017 calendar year (Fig. 1). Study 2 consisted of experimental data of 7,712 meals from college C lunches during the autumn term in 2017, when we experimentally altered the number of vegetarian options on offer at lunchtimes (Fig. 1).

Fig. 1.

Overview of data and levels of analyses in study 1 and study 2. Image credit: Icons from the Noun Project (https://nounproject.com).

We summarized the sales transaction data into 1) aggregate data, summarizing the total vegetarian and meat/fish (hereafter simply “meat”) sales at each lunch and dinner and 2) individual-level data on whether each diner at a meal selected a vegetarian or meat meal. Purchases made with university cards enabled anonymized individual diner-level purchases to be tracked; this is useful in evaluating how diners with different prestudy levels of purchasing vegetarian meals responded to increasing vegetarian availability (Methods). We used the total number of vegetarian and meat meals sold at a mealtime to analyze total sales. Measuring rebound effects (i.e., increased meat purchases at another time) is not possible for study 1 as vegetarian availability varied across lunches and dinners. For study 2, although we cannot completely capture rebound effects as we do not have information on what diners ate outside the cafeteria, as a proxy, we measured vegetarian sales at college C during dinnertimes, which were not included in the experimental intervention. We had originally intended dinners to be included, but this posed too much of an operational burden for the cafeteria (Methods). This created the opportunity to conduct a post hoc analysis of rebound effects that was not part of the original study design.

We estimated the effect of vegetarian availability on vegetarian meal sales and total meal sales, adjusting for other predetermined variables, including day of the week, ambient temperature, and average price difference between vegetarian and meat options (Methods) using linear models (LMs) and binomial generalized linear models (GLMs) for aggregate data. Binomial generalized linear mixed models (GLMMs) were used for the individual-level data, with individual diner fitted as a random effect, which allows each diner to have a different likelihood of selecting a vegetarian meal (29). A 95% confidence level was used to calculate confidence intervals (CIs). Models were evaluated using the Akaike information criterion (AIC), interpretability, and model diagnostics (30).

Study 1: Observational

Aims and Design.

For study 1, we did not experimentally alter the menu (SI Appendix, Tables S1 and S2) but observed the number of vegetarian and meat options available from the sales data. We analyzed long-term data from 269 mealtimes at college A and 266 mealtimes at college B. Excluding the few mealtimes when no vegetarian options were served (SI Appendix, Tables S3 and S4), vegetarian availability ranged from 16.7 to 75% in college A and 12.5 to 66.7% in college B.

Vegetarian Sales: Aggregate Data.

Vegetarian availability alone explained 20.9% and 31.9% of variation in vegetarian sales at college A and college B, respectively (binomial GLMs, McFadden’s pseudo-R²). When controlling for other variables, the best GLMs for colleges A and B explained 26.1% and 39.3%, respectively, of the variability in vegetarian sales (SI Appendix, Tables S5 and S6), with vegetarian availability remaining a highly significant predictor of vegetarian sales for both colleges (college A: n = 51,251 meals, P < 0.001; college B: n = 35,681 meals, P < 0.001). Specifically, the models estimated that doubling vegetarian availability from 25 to 50% increased vegetarian sales by 61.8% in college A (from 24.1% [CI = 22.5%, 25.7%] to 39.0% [CI = 36.7%, 41.3%] of total sales) and by 78.8% in college B (from 18.4% [CI = 16.8%, 20.1%] to 32.9% [CI = 30.6%, 35.4%]) (Fig. 2A and SI Appendix, Tables S5 and S6).

Fig. 2.

Effects of vegetarian availability on vegetarian and total sales for study 1. (A) Raw values (jittered) of vegetarian sales against vegetarian availability. (B and C) Modeled likelihood of selecting a vegetarian meal for individual diners at colleges A and B, with individual diners divided into least vegetarian (LeastVeg) to most vegetarian (MostVeg) quartiles. LessVeg, less vegetarian; MoreVeg, more vegetarian. (D) Raw values (jittered) of total sales against vegetarian availability. Lines of best fit and CIs were generated from the models using conditional regression and the visreg package in R (Methods).

Other variables also correlated with vegetarian sales but often had different effects in the 2 colleges. For example, as the vegetarian option became relatively cheaper compared with the meat options, vegetarian sales increased in college A but decreased in college B, and higher ambient temperatures were associated with higher vegetarian sales in college A but lower vegetarian sales in college B. However, increasing vegetarian availability increased vegetarian sales consistently in a similar way across colleges, indicating a strong and potentially generalizable effect.

Vegetarian Sales: Individual-Level Data.

A total of 1,394 identifiable individual diners at college A and 746 at college B used the cafeteria during the study period; this excludes guests and cash-only diners. Of these, 597 and 222 diners, respectively, purchased ≥10 meals in autumn 2016 (prior to our main study) and were divided into quartiles within each college, based on their level of vegetarian meal consumption during this period (Methods, Fig. 1, and SI Appendix, Tables S7 and S8). In both colleges, diners in every quartile from the most vegetarian to the least vegetarian bought more vegetarian meals as vegetarian availability increased (Fig. 2 B and C). For both colleges A and B, the least vegetarian quartile had the strongest response to increasing vegetarian availability (GLMM; college A: n = 32,687 meals, interaction effect size = 1.012 [CI = 1.004, 1.020], P = 0.004; college B: n = 19,663 meals, interaction effect size = 1.024 [CI = 1.014, 1.034], P < 0.001) (SI Appendix, Tables S9 and S10).

Total Sales.

College A sold an average of 191 main meals at a mealtime, and college B sold an average of 134. When adjusted for other variables, increasing vegetarian availability had no significant effect on total sales in college A and a small negative effect in college B, where the mean total of meals sold decreased from 138 (CI = 129, 147) to 128 (CI = 118, 137) as vegetarian availability increased from 25 to 50% (LM for main meals sold at a mealtime; college A: n = 51,251 meals, availability effect size = 1.001 [CI = 0.997, 1.003], P = 0.707; college B: n = 35,681 meals, availability effect size = 0.998 [CI = 0.997, 0.999], P < 0.001) (Fig. 2D and SI Appendix, Tables S11 and S12). The different quartiles of diners in college A did not respond differently, in terms of number of meals bought at a mealtime, as vegetarian availability increased (LM; n = 33,180 meals, interaction terms P > 0.05). In college B, those in the least vegetarian quartile responded more negatively to increasing vegetarian availability than those in other quartiles, in terms of total number of meals purchased (LM; n = 19,950 meals, interaction effect size = 0.995 [CI = 0.992, 0.998], P < 0.001). This was, however, still a small drop from a mean of 27.4 (CI = 26.2%, 28.6%) meals to 24.7 (CI = 23.2%, 25.9%) meals as vegetarian availability increased from 25 to 50%.

Study 2: Experimental

Aims and Design.

We tested the causality of the association between vegetarian availability and vegetarian sales by running an experiment at college C in the autumn term of 2017 based on fortnightly alternation between 1 (control) and 2 (experiment) vegetarian options at lunchtimes (Methods and SI Appendix, Fig. S1 and Tables S13 and S14). We analyzed data from 44 lunchtimes. Vegetarian availability ranged from 16.7 to 50%, (impacted by differences in the total number of options served, as well as our manipulation; SI Appendix, Table S15).

Vegetarian Sales: Aggregate Data.

Vegetarian availability alone explained only 3.9% of the variation in vegetarian sales (binomial GLM; n = 7,712 meals, McFadden’s pseudo-R² = 0.039, P < 0.001) in a univariate analysis. When controlling for other variables (Methods), 31.8% of the variation was explained (day of the week, week of the term, and the price differential of vegetarian and meat meals were the predictors that explained most of the variation in vegetarian sales), and availability remained a highly significant predictor of vegetarian sales (P < 0.001) (Fig. 3A and SI Appendix Table S16). The model estimated that doubling vegetarian availability from 25 to 50% increases vegetarian sales by 40.8% (from 19.1% [CI = 15.1%, 23.9%] to 26.9% [CI = 21.5%, 33.1%] of total sales) (SI Appendix, Table S16).

Fig. 3.

Effects of vegetarian availability on vegetarian and total sales for college C, study 2. (A) Raw values of vegetarian sales against vegetarian availability. (B) Modeled likelihood of selecting a vegetarian meal for individual diners, divided into least vegetarian (LeastVeg) to most vegetarian (MostVeg) quartiles. LessVeg, less vegetarian; MoreVeg, more vegetarian. (C) Raw values of total sales against vegetarian availability. (D) Raw values of vegetarian sales at dinner during the control and experimental weeks, with model mean estimates and CIs shown in white. Lines of best fit and CIs in A and C and model mean estimate with CIs in D were generated from the models using conditional regression and the visreg package in R (Methods).

Vegetarian Sales: Individual-Level Data.

A total of 121 of the 491 individual diners who bought a main meal during our experiment could be assigned a quartile based on their level of vegetarian meal consumption in the previous term, summer 2017 (Fig. 1 and SI Appendix, Tables S17 and S18). When other variables were controlled for, diners in every quartile (except most vegetarian) bought more vegetarian meals in response to increasing vegetarian availability (SI Appendix, Table S19). Similar to study 1, for college C, the least vegetarian quartile of diners had a significantly stronger response to increasing vegetarian availability than the other quartiles (GLMM; n = 1,585 meals, interaction term effect size = 1.053 [CI = 1.002, 1.106], P = 0.041) (Fig. 3B and SI Appendix, Table S19).

Total Sales and Possible Rebound Effects.

College C sold an average of 175 meals per lunchtime, and increasing vegetarian availability had no effect on total sales (LM for main meals sold at lunchtime: n = 7,712 meals, availability effect size = 1.000 [CI = 0.993, 1.004], P = 0.942) (Fig. 3C and SI Appendix, Table S20). Moreover, the different quartiles of diners responded similar to each other in terms of numbers of meals bought at a mealtime as vegetarian availability increased (LM; n = 3,201 meals, interaction terms P > 0.1). In college C, unlike in study 1, vegetarian sales at dinnertimes could be used to explore possible rebound effects. We analyzed dinner sales for the 71% of autumn term lunchtime diners who also ate at dinner. When adjusted for other variables, they bought similar numbers of vegetarian meals during the experimental weeks (when there were 2 vegetarian options at lunchtimes) as in the control weeks (with 1 vegetarian option) (GLM; control vs. experimental weeks: n = 5,287 meals, experimental weeks effect size = 0.953 [CI = 0.795, 1.141], P = 0.601) (Fig. 3D and SI Appendix, Table S21). Hence, we found no evidence for a rebound effect involving a drop in vegetarian sales at dinnertimes during weeks when there were higher vegetarian sales at lunchtimes.

Discussion

In all 3 participating colleges across study 1 and study 2, increasing the proportion of vegetarian meals offered increased vegetarian sales, with a large effect size that was greatest among those who were less likely to select vegetarian meals prior to the study. From 94,644 meals selected, we found that doubling vegetarian availability from 25 to 50% increased vegetarian sales (and decreased meat sales) by 7.8, 14.9, and 14.5 percentage points, equivalent to 40.8%, 61.8%, and 78.8% increases, respectively. Increasing vegetarian availability had little effect on total sales or vegetarian sales at other mealtimes not involved in experiments, indicating rebound effects were probably small or nonexistent. In 2 of 3 cafeterias, increasing vegetarian availability did not to lead different responses, in terms of number of meals bought, by diners with different prior levels of vegetarian meal selection. In the third college, there was a modest difference (with those previously eating meat responding slightly negatively to increasing vegetarian meal availability); however, together, these results suggest that increasing vegetarian availability did not substantially put off meat eaters.

Although it might seem intuitive that providing proportionally more vegetarian options would increase vegetarian sales, to our knowledge, this is untested. If meal preferences were fixed, changing the availability of vegetarian options would have no effect. If meal selections were random, this would lead to sales tracking the proportion of each meal option available. Our results indicate that meal selection is neither fixed nor random, but rather is partially determined by availability. These results suggest that increasing the proportion of vegetarian options may have a larger effect than many other choice architecture interventions included in a recent systematic review on meat selection and consumption (13): In previous studies, neither restructuring food menus with different meal descriptions nor positioning meat in less prominent positions reduced meat uptake. Providing US and UK participants with meat substitutes, recipes, and educational materials led to large reductions in meat consumption (13): a 40% reduction in red and processed meat (31), a 54% reduction in spending on meat (32), and a 70% reduction in meat consumed (33). These results are impressive but, unlike increasing vegetarian availability, are time- and resource-intensive; thus, they may not be scalable, and their effects can diminish over time (24, 31). One paper found that meat consumption was 60% lower at the end of the intervention than at the baseline but that after 2 mo, the effect had decreased to 40% (31). Reducing the serving size of meat portions reduced meat consumption by 13 to 14% (34, 35); hence, increasing vegetarian availability combined with smaller meat portions could be a powerful combined strategy to reduce the mass of meat served by cafeterias.

Our studies have several strengths. While many recent papers have stressed the importance of reducing meat consumption (1–3, 36), very few studies have tested which interventions might work. For example, a recent systematic review found only 18 studies with 11,290 observations that tested how changing some aspect of choice architecture could reduce meat consumption (13). Our studies have 94,644 observations from months of robust, individual-level data. We collected both observational and experimental data and included analyses on total meal sales. We have shown that increasing vegetarian availability can substantially reduce meat consumption, even for those with low prior levels of vegetarian meal consumption—the most important demographic group to shift to reduce the GHGE of the food system (5).

However, our studies also have several limitations. First, due to the design of the studies, we did not collect data on the nutrition of the cafeteria meals or their palatability to students, which are important considerations for catering managers (12, 37). Second, in keeping with other similar field studies (22), some data were misclassified. Miscoding of a small number of vegetarian meals as meat meals in college C led to a slight underestimation in study 2 of the effect of vegetarian availability on vegetarian sales (Methods); however, this is highly unlikely to change the results in a significant direction.

The current studies suggest opportunities for future research. First, they were conducted in a university setting with students and staff. While this is a good context in which to generate proof-of-concept evidence for the intervention, studies are now needed in other types of food outlets serving other populations, including those in middle- and low-income countries, to estimate the generalizability of the current findings. Second, we were informed by catering managers that ingredient costs were considerably cheaper for vegetarian meals but that labor costs might be higher. Future research could investigate the effects of increasing sales of vegetarian meals on profits. Third, to achieve tangible environmental benefits, any reduction in demand for meat needs to lead to reduced livestock farming, and not simply redirecting livestock products to other countries (38). Shifting both diets and agricultural production toward less meat will require the support of governments and farmers, as well as pressure from citizens (38, 39).

Nevertheless, our results demonstrate the potential of choice architecture for making progress toward improved sustainability. Increasing the availability of vegetarian options in cafeterias is a relatively cheap and easily implemented strategy that generally goes unnoticed: It does not require restructuring the canteen layout, or running meat-free days, the latter of which can prove unpopular (40), and it can save money on ingredients (24). Increasing the availability of plant-based meals will require diversification of vegetarian provision by cafeterias and restaurants, which may, in turn, necessitate changes in the training offered to chefs (37). Interest in reducing meat consumption and in “flexitarianism” is on the rise (41), and our results show that caterers serving more plant-based options are not just responding to but also reshaping customer demand. Further long-term studies—intervening on availability in addition to other aspects of choice environments and conducted in a wider range of settings—might usefully test behavioral interventions that are scalable and offer the potential to significantly mitigate climate change and biodiversity loss.

Methods

This research was approved by the University of Cambridge Psychology Research Ethics Committee (PRE.2016.100). In keeping with research governance for interventions that target environments and not individuals directly, consent was obtained from those who have authority over these environments (i.e., the managers of the college cafeterias). Signed consent forms, approved by the Research Ethics Committee, were obtained from each of the catering managers of the 3 participating colleges.

Study Setting.

Colleges A and B have both undergraduate and postgraduate members. College A has over 1,100 members, and college B has over 500 members. College C is a graduate college with over 600 members. All 3 colleges admit students of any gender identity. Students pay for meals by swiping their university cards, and meals are not included in the tuition or accommodation fees. In colleges A and B, students top up their card with credit throughout the academic year, and in college C, students pay the bill at the end of each term. Meals typically cost between £2.30 (€2.51, $2.45) and £3.70 (€4.04, $4.50). Although many students eat in the college cafeteria, others cook their own meals or eat elsewhere. In the cafeterias, vegetarian and meat meals are available throughout the mealtime; if meat or vegetarian options run out, they are quickly replaced by an option in the same category.

Study Design.

Study 1.

In their normal operations, colleges A and B varied both the total number of options and the number of vegetarian options available. We did not experimentally alter the menus from these colleges but observed how the availability of vegetarian meals related to their relative sales. We used data from lunch and dinner on weekdays (Monday to Friday) during spring (16 January to 17 March), summer (24 April to 30 June), and autumn (2 October to 1 December) terms of 2017.

Study 2.

College C experimentally altered the number of vegetarian meals on its menus. The original experimental design specified that that both lunch and dinner would alternate between 1 and 2 vegetarian options week by week. However, this was too much for the cafeteria to implement within the time frame of the study. Therefore, only lunchtimes alternated between the experimental condition of 1 and 2 vegetarian options, every 2 wk. The number of vegetarian options still sometimes varied from experimental allocation due to cafeteria constraints (SI Appendix, Table S15). Some misclassifications at the checkout occurred, resulting in some vegetarian meals being recorded as meat sales. This meant that vegetarian sales may have been up to 21.5% greater than recorded (E.E.G., personal observation). No meat meals were misclassified as vegetarian. Although unfortunate, this error is conservative and suggests that the true effect of availability at college C could be substantially greater than that reported, and closer to that estimated from the observational work at colleges A and B.

We collected and analyzed the experimental data from weekday lunchtimes from college C to test the effect of vegetarian availability, and also compared this with weekday dinner sales to investigate if increasing vegetarian availability at lunch affected vegetarian sales at dinner. Data were collected across the autumn term and the first 2 wk of the Christmas holidays in 2017 (2 October to 15 December). Unlike colleges A and B, college C is a graduate college, and meals were served to staff and students outside of normal university term times; thus, to increase the sample size, we included the first 2 wk of the Christmas holidays. These 2 wk did have slightly lower total sales than term-time weeks (SI Appendix, Table S19) but did not have significantly different vegetarian sales (SI Appendix, Table S15).

Data Collection.

Sales data were downloaded from the online catering platforms Uniware (42) and Accurate Solutions (43), and identifiable data were stored on a secure online server. All 3 colleges had online menus; however, the options served sometimes varied from this. At colleges A and B, the number of vegetarian options and total number of options could be inferred from how the sales data are coded. At college C, it was not possible to infer the number of vegetarian options and total options from the sales data; therefore, visits were made at lunchtimes to directly observe the options available. When the lunch offer included a pasta bar, this commonly had 2 sauces, often 1 vegetarian and 1 meat; we counted each sauce + pasta as one-half of an option.

Data Preparation.

We summarized the sales data into 1) aggregate data, summarizing the total vegetarian and meat sales at each lunch and dinner, and 2) individual-level data on whether each individual diner at a meal selected a vegetarian or meat meal. Eight mealtimes at college A and 3 at college B served no vegetarian main meals; therefore, vegetarian availability and vegetarian sales were 0. These data were excluded from the analysis to avoid overestimating the effect of availability (SI Appendix, Table S3). In college B, 1 mealtime only served 1 main meal in total, and this was also excluded from the analysis. Only lunchtimes when direct observations were made of the vegetarian and total options available were included in the analysis for college C.

Aggregate data included main meals bought by both college members and guests. Individual-level data only included meals bought by college members on their university cards, as only these meals could be associated with individual diners. An individual diner who bought 1 or more vegetarian meals at a mealtime was coded as 1, and an individual diner who bought 1 or more meat meals was coded as 0. Any individual diners who bought both vegetarian and meat meals at a single mealtime were coded as not applicable (NA), and we excluded those meal choices from the analysis; this removed 1.6% of the individual-level data at college A (699 of 43,751), 1.5% at college B (468 of 31,956), and 4.5% at college C (207 of 4,565).

We wanted to test if the response to increasing vegetarian availability varied with background levels of meat consumption. To calculate this, for individuals who bought ≥10 main meals during the preceding term (autumn 2016 for colleges A and B, summer term 2017 for college C), we calculated the proportion of main meals bought that were vegetarian, and these values were used to divide the individual diners into within-college quartiles: least, less, more, and most vegetarian.

Statistical Approaches.

We carried out analyses in R 3.5.2 (44), using the lme4 (45) package. We used binomial GLMs for the aggregate data, and binomial GLMMs for the individual-level data with each individual diner included as a random effect. Models were evaluated using AIC values and interpretability. We follow the recommendations of Simmons et al. (46), which include citing the effect of vegetarian availability, with and without covariates. Initial analyses showed that relative vegetarian availability (number of vegetarian options/number of total options) was a better predictor of vegetarian sales than number of vegetarian or meat options; therefore, we used this as the predictor variable for vegetarian availability. We estimated the effect of vegetarian availability on vegetarian sales and total sales, adjusting for other predetermined variables (Table 1). After model selection, we used the predict function to generate the predicted values and plotted out lines of best fit, using conditional regressions with 95% CIs using the effects (47) and visreg packages (48).

Table 1.

Variables considered for statistical models

Model	Variable	Description
All models	Vegetarian availability	Number of vegetarian options/total options available
	Total options available	Number of different meal options offered at a mealtime
	Total main meals sold	Number of main meals sold at a mealtime
	Vegetarian price differential (£)	Difference between mean cost of the meat options and vegetarian options
	Ambient temperature (°C)	Mean temperature over 24 h each day in Cambridge (°C) (49)
	Day	Monday, Tuesday, Wednesday, Thursday, Friday
	Week of term	1 to 11
For study 1 only (no variation in study 2)	Meal	Lunch or dinner
	Term	Spring, summer, autumn
For individual-level models only	Individual diner as a random effect
For individual-level models and models of total sales considering diner background	Prior level of vegetarian meal consumption	Individual diners at each college were divided into least, less, more, and most vegetarian quartiles, and we tested for any interaction effects with vegetarian availability
For study 2 rebound model	Week condition	Control or experimental week