This is the authors’ accepted version of the manuscript for personal use. The definitive version is published in
PNAS, www.pnas.org/cgi/doi/10.1073/pnas.1907207116 . Supporting information online at
SOCIAL SCIENCES: Sustainability Science
Impact of increasing vegetarian availability on meal selection and sales in cafeterias
Emma E. Garnetta,1
Mark A. Pillingc
Theresa M. Marteauc
a Department of Zoology, University of Cambridge, CB2 3EJ, UK; b Department of Geography,
University of Cambridge, CB2 1QB, UK; c Behaviour and Health Research Unit, University of
Cambridge, Institute of Public Health, CB2 0SR, UK
1 To whom correspondence may be addressed. Email: email@example.com
Diet; behaviour change; meat; vegetarian; choice architecture; sustainability; climate
The aggregate data and summaries of the individual-level data can be found at
Shifting people in higher-income countries towards 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 behaviour 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. Anonymised individual-level data from 94,644 meals purchased in 2017 were
collected from three 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 (two cafeterias) and by 7.8 percentage points in the experimental
study (one 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 novel and 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.
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 is the first study to assess the impact of increasing the
proportion of plant-based meal options on selection, and is based on over 90,000 meal
choices. We suggest our findings have potential to make a significant contribution to the
global ambition for more sustainable diets.
High-income countries produce and consume animal-derived food – meat, fish, dairy, eggs –
at levels that are incompatible with meeting greenhouse gas emissions (GHGE) reduction
targets (1). Livestock and aquaculture are responsible for 56-58% 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, goats) has average GHGE per
kg five times higher than pork, seven times higher than chicken and 43 times higher than
legumes (3). Shifting towards a more plant-based diet is therefore one the most effective
ways of reducing the environmental footprint of food (2, 4). For the UK it is estimated that
switching from a high meat (>100g/day) 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 behaviour (6, 7). Education to bring about behaviour change is a popular
and uncontroversial method but – while it can raise awareness – it appear to be largely
ineffective at actually changing behaviour (8, 9). Models suggest that taxes on the most
polluting foods would result in savings of 1Gt 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 so far has received
relatively little empirical attention (10–13).
As one form of nudging, altering the relative availability of different food types has shown
promise as a lever for changing dietary behaviour 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 behaviour change is rated as “limited” (14). A Cochrane review (15) found
only five studies on altering availability that met the inclusion criteria (16–20), with a meta-
analysis showing a non-significant 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 USA school
cafeteria from 33% to 50% increased their selection by 108% and 63% respectively (21); and
in four English workplace cafeterias, decreasing the number of high-calorie cooked meals
offered to one 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 one 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 inter-individual 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 one study (again
focused on health rather than meat consumption) which presents long-term individual-level
data on how availability affects food choices (26). There are two 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 two studies – one observational and one
experimental – in three college cafeterias in 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 anonymised
individual-level data to analyse whether increasing vegetarian availability had effects which
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.
We collected data from three 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 non-experimental 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 spring, summer and autumn terms in the
2017 calendar year (Figure 1). Study 2 consisted of experimental data of 7712 meals from
College C lunches during autumn term 2017, when we experimentally altered the number of
vegetarian options on offer at lunchtimes (Figure 1).
We summarised the sales transaction data into a) aggregate data, summarising the total
vegetarian and meat/fish (hereafter simply “meat”) sales at each lunch and dinner and b)
individual-level data on whether each diner at a meal selected a vegetarian or meat meal.
Purchases made with university cards enabled anonymised individual diner-level purchases
to be tracked; this is useful in evaluating how diners with different pre-study 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 analyse 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 dinner times, 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 pre-determined variables including day of the week, ambient
temperature, average price difference between vegetarian and meat options (Methods)
using Linear Models (LMs) and binomial Generalised Linear Models (GLMs) for aggregate
data. Binomial Generalised 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).
Figure 1: Overview of data and levels of analyses in Study 1 and Study 2. Credit: icons from
Study 1: observational
Aims and design
For Study 1 we did not experimentally alter the menu (Supporting Information (SI)
Appendix, Tables S1 and S2) but observed the number of vegetarian and meat options
available from the sales data. We analysed long-term data from 269 mealtimes at College A
and 266 mealtimes at College B. Excluding the few mealtimes where 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 College 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%), Figure 2a and SI Appendix Tables S5 and S6).
Other variables also correlated with vegetarian sales but often had different effects in the
two colleges. For example, as the vegetarian option became relatively cheaper compared to
the meat options, vegetarian sales increased in College A but decreased in College B; 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
1394 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 (Figure 1, Methods and SI Appendix Tables S7 and S8). In
both colleges every quartile from the Most Vegetarian to the Least Vegetarian bought more
vegetarian meals as vegetarian availability increased (Figure 2b&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).
College A sold an average of 191 main meals at a mealtime, and College B, 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 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)(Figure 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%) as vegetarian availability
increased from 25% to 50%.
Figure 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): Modelled likelihood of selecting a vegetarian meal for
individual diners at Colleges A and B, with individual diners divided into Least Vegetarian to Most Vegetarian
quartiles; d) Raw values (jittered) of total sales against vegetarian availability. Lines of best fit and confidence
intervals generated from the models using conditional regression and the visreg package in R (Methods).
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 autumn term 2017 based on fortnightly
alternation between one (control) and two (experiment) vegetarian options at lunchtimes
(Methods, SI Appendix Tables S13 and S14 and Figure S1). We analysed 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=7712 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 term and the price differential of vegetarian and meat
meals were the predictors which explained most of the variation in vegetarian sales), and
availability remained a highly significant predictor of vegetarian sales (p<0.001, Figure 3a
and SI Appendix, 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).
Vegetarian sales: individual-level data
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 (Figure 1, 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). Similarly 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=1585
meals, interaction term effect size= 1.053 (CI= 1.002, 1.106), p=0.041, Figure 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=7712 meals, availability
effect size= 1.000 (CI= 0.993, 1.004), p=0.942; Figure 3c and SI Appendix Table S20).
Moreover the different quartiles of diners responded similarly to each other in terms of
numbers of meals bought at a mealtime as vegetarian availability increased (LM, n=3201
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 analysed 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 two vegetarian options at lunchtimes) as in the control weeks (with one
vegetarian option)(GLM, control v experimental weeks, n=5287 meals, experimental weeks
effect size= 0.953 (CI= 0.795, 1.141), p=0.601, Figure 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.
Figure 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) Modelled likelihood of selecting a vegetarian meal for
individual diners, divided into Least Vegetarian to Most Vegetarian quartiles; 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 confidence intervals in white. Lines of best fit and confidence intervals
in a) and c) and model mean estimate with confidence intervals in d) generated from the models using
conditional regression and the visreg package in R (Methods).
In all three participating colleges across Study 1 and Study 2 increasing the proportion of
vegetarian meals offered increased vegetarian sales, with a large effect size which was
greatest amongst those who prior to the study were less likely to select vegetarian meals.
To our knowledge this is the first year-long study on how altering availability affects
sustainable food choices. 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.
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
non-existent. In two out of three 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) but 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
(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 – so may not be scalable – and their effects can diminish over time (24,
31): one paper found that at the end of the intervention meat consumption was 60% lower
than at the baseline but after two months the effect had decreased to 40% (31). Reducing
the serving size of meat portions reduced meat consumption by 13-14% (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 underestimate 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 generalisability of the current findings. Second, we were informed by
catering managers that ingredients costs were considerably cheaper for vegetarian meals,
but that labour 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 towards 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 towards improved sustainability. Increasing the availability of vegetarian options in
cafeterias is a relatively cheap and easily-implemented strategy which generally goes
unnoticed: it does not require restructuring the canteen layout, or running meat-free days
that 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 re-shaping 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 behavioural interventions that are scalable and offer
the potential to significantly mitigate climate change and biodiversity loss.
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 three participating colleges.
Colleges A and B have both undergraduate and postgraduate members. College A has over
1100 members, and College B over 500. College C is a graduate college with over 600
members. All three colleges admit students of any gender identity. Students pay for meals
by swiping their university cards, 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, 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.
Colleges A and B in their normal operations 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
(16th January to 17th March), summer (24th April to 30th June) and autumn terms (2nd
October to 1st December) 2017.
College C experimentally altered the number of vegetarian meals on their menus. The
original experimental design specified that that both lunch and dinner would alternate
between one and two vegetarian options week by week. However, this was too much for
the cafeteria to implement within the timeframe of the study. Therefore, only lunchtimes
alternated between the experimental condition of one and two vegetarian options, every
two weeks. 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 (EG,
pers. obs.). No meat meals were misclassified as vegetarian. Though 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 analysed 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 autumn term and the first two weeks of the Christmas
holidays 2017 (2nd October to 15th December). Unlike College A and B, College C is a
graduate college and meals were served to staff and students outside of normal university
term-times, so to increase the sample size we included the first two weeks of the Christmas
holidays. These two weeks did have slightly lower total sales than term time weeks (SI
Appendix Table S19) but did not have significantly different vegetarian sales (SI Appendix
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 three 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 two sauces, often one vegetarian and one meat; we counted each
sauce+pasta as half an option.
We summarised the sales data into a) aggregate data, summarising the total vegetarian and
meat sales at each lunch and dinner and b) individual-level data on whether each individual
diner at a meal selected a vegetarian or meat meal. Eight mealtimes at College A and three
at College B served no vegetarian main meals, and therefore vegetarian availability and
vegetarian sales were zero. These data were excluded from the analysis to avoid
overestimating the effect of availability (SI Appendix Table S3). In College B one mealtime
only served one 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 one or more vegetarian meals at a mealtime was coded as 1; an individual diner
who bought one or more meat meals was coded as 0. Any individual diners who bought
both vegetarian and meat meals at one meal time were coded as NA and we excluded those
meal choices from the analysis; this removed 1.6% of the individual-level data at College A
(699/43,751), 1.5% at College B (468/31,956) and 4.5% at College C (207/4,565).
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.
We carried out analyses in R 3.5 (44), using the lme4 (45) packages. We used Binomial
Generalised Linear Models for the aggregate data, and Binomial Generalised Linear Mixed
Models 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 includes 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 and 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 pre-determined 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% confidence
intervals using the effects (47) and visreg packages (48).
Table 1: Variables considered for statistical models.
Number of vegetarian options/ total options available
Number of different meal options offered at a
Total main meals
Number of main meals sold at a mealtime
The difference between the mean cost of the meat
options and the vegetarian options
Mean temperature over 24 hours each day in
Monday, Tuesday, Wednesday, Thursday, Friday
Week of term
For Study 1 only (no variation
in Study 2)
Lunch or dinner
Spring, summer, autumn
For individual-level models
Individual diner as a
For individual-level models
and models of total sales
considering diner background
Prior level of
Individual diners at each college were divided into
Least, Less, More and Most Vegetarian quartiles and
we tested for any interaction effects with vegetarian
For Study 2 rebound model
Control or experimental week
We thank Ivan, Paul, Rob, Alex, Gary and Christine for generously contributing their time
and data, and Benno Simmons for help with coding in R. E.G is funded by a NERC
studentship, grant number NE/L002507/1. A.B. is supported by a Royal Society Wolfson
Research Merit award.
E.G., A.B., C.S. and T.M. designed research; E.G. performed research; E.G. and M.P. analysed
data; E.G., A.B., C.S. and T.M. wrote the paper.
Conflict of interest
The authors declare no conflict of interest.
1. Bryngelsson D, Wirsenius S, Hedenus F, Sonesson U (2016) How can the EU climate targets be met? A
combined analysis of technological and demand-side changes in food and agriculture. Food Policy
2. Poore J, Nemecek T (2018) Reducing food’s environmental impacts through producers and consumers.
Science (80- ) 992(6392):987–992.
3. Clune S, Crossin E, Verghese K (2017) Systematic review of greenhouse gas emissions for different
fresh food categories. J Clean Prod 140:766–783.
4. Springmann M, et al. (2017) Mitigation potential and global health impacts from emissions pricing of
food commodities. Nat Clim Chang (7):69–74.
5. Scarborough P, et al. (2014) Dietary greenhouse gas emissions of meat-eaters, fish-eaters, vegetarians
and vegans in the UK. Clim Change 125(2):179–192.
6. Reddy SMW, et al. (2016) Advancing Conservation by Understanding and Influencing Human Behavior.
Conserv Lett 00(May):1–9.
7. Marteau TM (2017) Towards environmentally sustainable human behaviour: targeting non-conscious
and conscious processes for effective and acceptable policies. Philos Trans R Soc A Math Phys Eng Sci
8. Diepeveen S, Ling T, Suhrcke M, Roland M, Marteau TM (2013) Public acceptability of government
intervention to change health-related behaviours: a systematic review and narrative synthesis. BMC
Public Health 13(1):756.
9. Bianchi F, Dorsel C, Garnett E, Aveyard P, Jebb SA (2018) Interventions targeting conscious
determinants of human behaviour to reduce the demand for meat: A systematic review with
qualitative comparative analysis. Int J Behav Nutr Phys Act 15(102):1–25.
10. Bucher T, et al. (2016) Nudging consumers towards healthier choices: a systematic review of positional
influences on food choice. Br J Nutr (9):1–12.
11. Cadario R, Chandon P (2017) Which Healthy Eating Nudges Work Best? A Meta-Analysis of Behavioral
Interventions in Field Experiments.
12. Campbell-Arvai V, Arvai J, Kalof L (2014) Motivating sustainable food choices: The role of nudges, value
orientation, and information provision. Environ Behav 46(4):453–475.
13. Bianchi F, Garnett E, Dorsel C, Aveyard P, Jebb SA (2018) Restructuring physical micro-environments to
reduce the demand for meat: A systematic review with qualitative comparative analysis. Lancet Planet
14. McKinsey Global Institute (2014) Overcoming obesity: An initial economic analysis.
15. Hollands GJ, et al. Altering the availability or proximity of food, alcohol and tobacco products to change
their selection and consumption. Cochrane database Syst Rev (3).
16. Fiske A, Cullen KW (2004) Effects of promotional materials on vending sales of low-fat items in
teachers’ lounges. J Am Diet Assoc 104(1):90–93.
17. Foster GD, et al. (2014) Placement and promotion strategies to increase sales of healthier products in
supermarkets in low-income, ethnically diverse neighborhoods: A randomized controlled trial. Am J
Clin Nutr 99(6):1359–1368.
18. Roe LS, Meengs JS, Birch LL, Rolls BJ (2013) Serving a variety of vegetables and fruit as a snack
increased intake in preschool children. Am J Clin Nutr 98(3):693–699.
19. Stubbs RJ, Johnstone AM, Mazlan N, Mbaiwa SE, Ferris S (2001) Effect of altering the variety of
sensorially distinct foods, of the same macronutrient content, on food intake and body weight in men.
Eur J Clin Nutr 55(1):19–28.
20. Kocken PL, et al. (2012) Promoting the Purchase of Low-Calorie Foods From School Vending Machines:
A Cluster-Randomized Controlled Study. J Sch Health 82(3):115–122.
21. Bartholomew JB, Jowers EM (2006) Increasing frequency of lower-fat entrees offered at school lunch:
An environmental change strategy to increase healthful selections. J Am Diet Assoc 106(2):248–252.
22. Pechey R, et al. (2019) Impact of increasing the proportion of healthier foods available on energy
purchased in worksite cafeterias: A stepped wedge randomized controlled pilot trial. Appetite
23. Clark M (2017) Chronic effects of replacing red and processed meat with non/reduced meat
alternatives. Dissertation (University of Cambridge).
24. Gravert C, Kurz V (2017) Nudging À La Carte – A Field Experiment on Food Choice
25. Bacon L, Krpan D (2018) (Not) Eating for the environment: The impact of restaurant menu design on
vegetarian food choice. Appetite 125:190–200.
26. Whitaker RC, Wright JA, Finch AJ, Psaty BM (1993) An environmental intervention to reduce dietary fat
in school lunches. Pediatrics 91(6):1107–1111.
27. Grech A, Allman-Farinelli M (2015) A systematic literature review of nutrition interventions in vending
machines that encourage consumers to make healthier choices. Obes Rev 16(12):1030–1041.
28. O’Reilly GA, et al. (2017) Sugar Restriction Leads to Increased Ad Libitum Sugar Intake by Overweight
Adolescents in an Experimental Test Meal Setting. J Acad Nutr Diet 117(7):1041–1048.
29. McCulloch CE, Searle SR, Neuhaus JM (2008) Generalized, Linear, and Mixed Models (Wiley-Blackwell).
30. Hosmer DW, Lemeshow S, Sturdivant RX (2013) Applied Logistic Regression. Applied Logistic
Regression (Wiley), pp 89–152. Third Edit.
31. Ali M, Simpson EJ, Clark M, Razak A, Salter A (2017) The impact of dietary meat intake reduction on
haematological parameters in healthy adults. Proc Nutr Soc 76(OCE3):E70.
32. Flynn MM, Reinert S, Schiff AR (2013) A Six-Week Cooking Program of Plant-Based Recipes Improves
Food Security, Body Weight, and Food Purchases for Food Pantry Clients. J Hunger Environ Nutr
33. Holloway T, Salter AM, Mccullough FS (2012) Chapter 4: Dietary intervention to reduce meat intake by
50 % in University students –a pilot study. Proc Nutr Soc 71(OCE2):119–151.
34. Reinders MJ, Huitink M, Dijkstra SC, Maaskant AJ, Heijnen J (2017) Menu-engineering in restaurants -
adapting portion sizes on plates to enhance vegetable consumption: a real-life experiment. Int J Behav
Nutr Phys Act 14(1):41.
35. Rolls BJ, Roe LS, Meengs JS (2010) Portion size can be used strategically to increase vegetable
consumption in adults. Am J Clin Nutr 91(4):913–922.
36. Godfray HCJ, et al. (2018) Meat consumption, health, and the environment. Science (80- ) 243(July).
37. Volkhardt I, et al. (2016) Checklist for a vegan lunch menu in public catering. Ernaehrungs Umschau
38. Buckwell A, Nadeu E (2018) What is the Safe Operating Space for EU livestock? (Brussels) Available at:
39. Committee on Climate Change (2018) Land use: Reducing emissions and preparing for climate change.
40. Lombardini C, Lankoski L (2013) Forced Choice Restriction in Promoting Sustainable Food
Consumption: Intended and Unintended Effects of the Mandatory Vegetarian Day in Helsinki Schools. J
Consum Policy 36(2):159–178.
41. Eating Better (2017) The future of eating is flexitarian: Companies leading the way.
42. Uniware Available at: http://www.uniware.co.uk/ [Accessed January 8, 2019].
43. Accurate Solutions Available at: http://www.accurate-solutions.co.uk/ [Accessed January 8, 2019].
44. R Core Team (2018) R: A language and environment for statistical computing. Available at:
45. Bates D, Bolker BM, Maechler M, Walker SC (2015) Fitting Linear Mixed Effects Models using lme4. J
Stat Softw 67(1):251–264.
46. Simmons JP, Nelson LD, Simonsohn U (2011) False-Positive Psychology: Undisclosed Flexibility in Data
Collection and Analysis Allows Presenting Anything as Significant. Psychol Sci 22(11):1359–1366.
47. Fox J, Weisberg S (2018) Visualizing Fit and Lack of Fit in Complex Regression Models with Predictor
Effect Plots and Partial Residuals. J Stat Softw 87(9):1–27.
48. Breheny P, Burchett W (2016) Visualization of Regression Models Using visreg. R J 9(December):56–71.
49. Cambridge Daily Weather Graphs (2018) Available at: