Sustainable by Design: Choice Architecture and the Carbon Footprint of Grocery Shopping

Abstract

The increase in global temperatures requires substantial reductions in the greenhouse emissions from consumer choices. The authors use an experimental incentive-compatible online supermarket to analyze the effect of a carbon-based choice architecture, which presents commodities to customers in high, medium, and low carbon footprint groups, in reducing the carbon footprints of grocery baskets. The authors relate this choice architecture to two other policy interventions: (1) a bonus-malus carbon tax on all grocery products and (2) moral goal priming using an online banner noting the moral importance of reducing one’s carbon footprint. Participants shopped from their home in an online store containing 612 existing food products and 39 existing nonfood products for which the authors had carbon footprint data over three successive weeks, with the interventions occurring in the second and third weeks. Choice architecture reduced participants’ carbon footprint significantly in the third week by reducing the proportion of choices made in the high-carbon aisle. The carbon tax reduced carbon footprint in both weeks, primarily by reducing overall spend. The goal-priming banner led to a small reduction in carbon footprint in the second week only. Thus, the design of the marketplace plays an important role in achieving the policy objective of reducing greenhouse gas emissions.

A key challenge facing society is how to avoid unsustainable increases in mean global temperature. Meeting this challenge requires substantial reductions in the emissions of greenhouse gases (GHG) along the supply chain of all goods and services (Metcalf and Weisbach 2009; Vandenbergh and Steinemann 2007; Weidema et al. 2008). Businesses and other organizations have started to reduce GHG emissions along the supply chains of the goods and services they provide, from sourcing sustainable raw materials through production, transport, and storage of their products. This is known as their “carbon footprint.” Yet consumer choices still account for a large share of food-related GHG emissions (Poore and Nemecek 2018), and policy makers’ attention has increasingly focused on the design of interventions targeting consumers (Dietz et al. 2009; Metcalf and Weisbach 2009; Panzone et al. 2018; Panzone, Wossink, and Southerton 2013; Vandenbergh and Steinemann 2007).

However, consumers are often reluctant to switch to more sustainable choices, even when this course of action is beneficial to them (Allcott and Rogers 2014; Gifford 2011; Steg 2016). The difficulty of changing behavior on a large scale arises for three main reasons. First, climate change is (to some) a hypothetical scenario, far in the future and distant in space, and the benefits of any costly (to the decision maker) change in behavior are shared with a large group of individuals, most unknown personally to the individual (Gifford 2011, 2014; Goeschl and Perino 2012; Steg 2016; Weber 2018; White, Habib, and Hardisty 2019). Second, consumers struggle in estimating the damage that their choices cause to the environment (Camilleri et al. 2019; Gifford 2014; Panzone, Lemke, and Petersen 2016; Panzone et al. 2020), and the information (e.g., the carbon footprint) is usually not available when choices are made (Muller, Lacroix, and Ruffieux 2019; Osman and Thornton 2019). Third, supermarkets have large choice sets, and the search for low-carbon alternatives is time intensive (Muller, Lacroix, and Ruffieux 2019; Panzone et al. 2018). Consumers who are strongly concerned with their carbon footprint may end up paying high search costs to lower their carbon footprint, whereas less-motivated consumers may be unwilling to do the same, refusing to change their behavior.

While the reduction in carbon footprint is a public policy problem connected to the management of the environmental public good, retailers can offer viable solutions to address climate change at a large scale by intervening directly where choices are made (Demarque et al. 2015; Muller, Lacroix, and Ruffieux 2019; Panzone et al. 2018). In particular, changes to the architecture of a choice can be used to facilitate large-scale transitions to low-carbon grocery shopping. Choice architecture refers to a family of tools that changes the decisions consumers make by altering the way choices are presented to decision maker, without altering the assortment of the store or the prices of the goods (Johnson et al. 2012; Lamberton and Diehl 2013; Thaler, Sunstein, and Balz 2014; Theotokis and Manganari 2015). In policy practice, choice architecture can be used to reduce externalities (Gravert and Kurz 2019) and may also have a habit formation potential if it remains in place for a sufficiently long time (White, Habib, and Hardisty 2019).

In this article, we study an innovative carbon-based choice architecture to reduce the carbon footprint of shoppers’ grocery baskets using a framed field experiment (Harrison and List 2004). In the experiment, we generated a unique data set of 273 individuals shopping over three weeks in an experimental store with over 600 products, with full incentive compatibility to ensure that choices are real. Methodologically, we advance previous experiments in the same area (Demarque et al. 2015; Panzone et al. 2018; Perino, Panzone, and Swanson 2014; Vanclay et al. 2011) by conducting the research entirely online. An additional novelty is the study of the performance of this choice architecture intervention in the presence of two motivators: (1) a bonus-malus carbon tax ¹ and (2) moral goal priming. The bonus-malus carbon tax, presented here for the first time in a retail experiment, combines a carbon tax that adds the environmental costs associated with the GHG emissions of each product with a flat subsidy to all goods (D’Haultfœuille, Givord, and Boutin 2014; Hilton et al. 2014). The moral goal was primed through a banner displaying the message “Keep Carbon Low” on the webpage participants used to place their orders.

Results indicate that choice architecture leads to a strong reduction in the carbon footprint of food baskets, but only in the final week of the experiment. This reduction is due to consumers purchasing fewer high-carbon goods. The carbon tax reduced carbon footprint in both weeks, reducing overall spend rather than driving within-store product substitution. Finally, the goal-priming banner led to a small reduction in carbon footprint in the first week only.

Retail Interventions to Promote Sustainable Consumption

While public policy makers have an important role in protecting the environmental public good, retailers can play a major role in driving changes in consumer behavior at a local level. The proximity between consumers and retailers in the marketplace helps retailers understand consumers’ needs, therefore providing access to knowledge that can be used to change the choices of a large number of households relatively quickly (Macfadyen et al. 2015). In particular, behavioral public policy interventions (e.g., nudging, choice architecture) can be implemented straightforwardly in retail environments for three reasons (Gravert and Kurz 2019; Sunstein and Reisch 2014): (1) they have low implementation costs, (2) they can be introduced without the need for formal regulation (unlike taxes or bans), and (3) they are not invasive, because they neither change prices nor alter the choice set.

Previous research shows that the provision of information on the carbon footprint of individual food products, or that of the food basket, reduces the GHG of the basket of goods purchased. For instance, carbon labels changed behavior in a small choice set (Vanclay et al. 2011), while traffic-light and kilometric labels (which report GHG in terms of kilometers driven by a car) on individual products successfully reduced the carbon footprint in a small experimental online supermarket (Muller, Lacroix, and Ruffieux 2019). Similarly, the use of color-coded feedback on the carbon footprint of a basket successfully reduces the total GHG purchased (Kanay et al. 2021).

Of particular interest in the promotion of sustainable consumption are interventions that affect the environmental self-concept of the consumer. Using an online supermarket in an initial task, Mazar and Zhong (2010) show that the purchase of “green” goods affects consumers’ moral self-concept, influencing subsequent unrelated prosocial behavior. ² Panzone et al. (2018) show that consumers in a large experimental supermarket who were asked to recall past (successful) environmental behavior prior to entering the shop reduced the carbon footprint of their food shopping. Similarly, knowledge of the consumption norm (how many people purchase environmentally friendly products) increases the purchase of environmentally friendly items (Demarque et al. 2015). The role of the self-concept in prosocial consumption is further supported by research showing that prosocial behavior self-signals prosocial predisposition to the decision maker: a carbon subsidy reduces environmental motivation by removing this self-signal (Perino et al. 2014), and the voluntary acceptance of a prosocial price premium motivates further prosocial behavior by strengthening it (Gneezy et al. 2012).

Changing the Architecture of Choices to Facilitate More Sustainable Choices

Rational consumers base their decisions on beliefs that their choices enable them to attain valued outcomes (e.g., reducing the risks of climate change), but they need information to guide their choices (e.g., the carbon footprints of products). Apart from internal determinants of behavior, consumer choices are also influenced by the way choices are structured and presented to them, something known as the choice architecture (Johnson et al. 2012; Thaler et al. 2014). Early experiments designed the architecture of choices by exploiting decision-making biases to protect the long-term benefits of the decision maker (Johnson et al. 2012; Kamenica 2012; Thaler et al. 2014). More recently, choice architecture has explored how the organization of products in a retail environment drives consumer choices (Diehl, Van Herpen, and Lamberton 2015; Lamberton and Diehl 2013; Sarantopoulos et al. 2019). For instance, complement-based organizations (e.g., presenting trousers jointly with shirts) are more effortful to process than substitute-based organizations (Diehl, Van Herpen, and Lamberton 2015) but increase consumer spending by expanding the consideration set consumers use (Sarantopoulos et al. 2019). Apart from Lamberton and Diehl (2013), these architectures alter only the structure of the choice, providing no additional information on the products in the store (e.g., the damage they may cause to the environment).

There is almost no previous research using choice architecture to motivate sustainable consumption; an exception is Theotokis and Manganari (2015), who study default policies. The architecture of a choice can be designed to provide consumers with information about a perceived benefit (e.g., sustainable consumption) and reduce the costs of finding products that satisfy that benefit. For instance, a benefit-based architecture that groups products into low-, medium-, and high-carbon aisles would reduce considerably the (nonmonetary) search costs of finding low-carbon goods. A benefit-based architecture that organizes products by their abstract environmental benefit is expected to motivate environmentally friendly consumption by activating an abstract mindset ³ : it induces consumers to focus on the abstract reasons for taking an action (why am I doing it?), rather than the concrete means to achieve a goal (how do I do it?) (Lamberton and Diehl 2013; Reczek, Trudel, and White 2018). This leads to our first hypothesis:

H₁: A choice architecture that organizes a product assortment by carbon footprint reduces the carbon footprint of shopping baskets relative to a control group with no choice architecture.

Motivating Low-Carbon Baskets: Goal Priming and Taxes

Proenvironmental behavior requires that an environmental goal is active during the shopping trip (Steg 2016; Steg and Vlek 2009; White, Habib, and Hardisty 2019). A choice architecture that highlights the environmental (carbon) impact of a choice makes the environmental goal salient, and may interact with instruments promoting the same goal. We test two instruments that can influence the effectiveness of choice architecture: a goal-priming banner and a bonus-malus carbon tax.

Moral Goal Priming

Moral goal priming aims to persuade consumers to take the environmental impact of their choices into account by making a sustainability goal salient. For instance, a sustainability goal would remind consumers to purchase low-carbon food shopping baskets to protect the natural environment. The communication of a goal contains a normative element, because it suggests the existence of a social norm (e.g., environmental preservation; White, Habib, and Hardisty 2019). Importantly, providing a message with an environmental frame motivates participants by influencing their self-concept (i.e., how strongly an individual identifies as an environmentally caring person), influencing the motivation to engage in proenvironmental behavior as a way to maintain a desired moral self-image (Bolderdijk et al. 2013). Goal priming has been shown to successfully motivate proenvironmental behavior. For instance, priming consumers with information about waste motivated the intention to purchase loose (vs. packaged) fruits and vegetables (Tate, Stewart, and Daly 2014), while information on carpooling in the United States motivated paper recycling in a sample of U.K. students (Evans et al. 2012). Similarly, priming a dominant normative self-control goal reduces its depletion over time, thereby maintaining a high level of commitment toward the goal primed (Walsh 2014).

The efficacy of an environmental message crucially depends on the structure of the message. Specifically, a message is less effective when it presents an injunctive norm (White and Simpson 2013), when it is too assertive (Zemack-Rugar, Moore, and Fitzsimons 2017), or it is seen as an unrequested interference (Laran, Dalton, and Andrade 2010). In these instances, the message may still lead to a reduction in carbon footprint; however, this reduction will be lower than that of a more persuasive message due to a negative impact on the self-concept of the recipient: an assertive message increases guilt associated with noncompliance (Zemack-Rugar et al. 2017), and injunctive messages weaken proenvironmental attitudes (White and Simpson 2013). With these considerations in mind, our second hypothesis is as follows:

H₂: A morally nuanced environmental goal-priming message produces a lower basket carbon footprint than a control group with no goal priming.

Interactions between choice architecture and moral priming

The moral goal prime makes salient a clear sustainability goal (i.e., low-carbon shopping choices), and carbon-based choice architecture increases the consumer’s ability to achieve this goal. Then, a goal-priming banner that promotes the same low-carbon goal underlying the design of the choice architecture may provide a synergistic effect that further reduces the carbon footprint of a basket. A further argument for a positive interaction between moral goal priming and choice architecture is that a benefit-based architecture activates an abstract construal (Lamberton and Diehl 2013), which increases the consumer’s ability to adhere to the moral goal prime’s call to buy low carbon items. Thus,

H_3a: Goal priming increases the strength of the choice architecture manipulation, so that the joint effect is a larger (in absolute value) reduction in the carbon footprints of baskets than the sum of the individual effects.

However, the moral goal priming may reduce the effectiveness of the choice architecture manipulation. As we noted in H₂, if consumers perceive the goal-priming message as injunctive, it may have a negative effect on their self-concept (White and Simpson 2013; Zemack-Rugar et al. 2017), which can reduce the effectiveness of the goal-priming banner when presented alone. When moral goal priming is implemented along with choice architecture, consumers may view the combination as an excessive attempt to influence their decisions, causing them to place less effort in the search of low-carbon goods. These arguments suggest that H_3a might be rejected in favor of an alternative hypothesis:

• H_3b: The joint effect of choice architecture and goal priming is a smaller (in absolute value) reduction in the carbon footprint of baskets than the sum of the individual effects.

Bonus-Malus Carbon Tax

Economists have long argued that a key policy tool for reducing emissions of pollutants is to impose a tax that reflects the cost of environmental damage caused by the production and consumption of each good. In climate change policy, a carbon tax (Metcalf and Weisbach 2009; Stern 2007) targets a reduction in pollutants known as GHG, measured in grams of carbon-dioxide equivalent (gCO₂e). A tax changes behavior extrinsically by imposing a penalty on undesirable behaviors (Bénabou and Tirole 2003; Perino et al. 2014): a carbon tax reduces GHG by motivating a change from high-carbon products (e.g., beef) to low-carbon substitutes (e.g., chicken), making carbon-intensive goods relatively more expensive and low-carbon substitutes relatively cheaper (e.g., Brännlund and Nordström 2004; Edjabou and Smed 2013; Kehlbacher et al. 2016; Panzone et al. 2018). A tax leads to two (experimentally nonseparable) effects: (1) the substitution effect, because it motivates substitution from high- to low-carbon goods, and (2) the income effect, which reduces consumption by decreasing the (real) disposable income of consumers. Because of the income effect, participants exposed to a tax manipulation have a lower budget than those not exposed to the tax, so that reductions in carbon footprint are partially caused by consumers spending less, rather than changing behavior. ⁴

A practical alternative is a bonus-malus tax (D’Haultfœuille, Givord, and Boutin 2014; Hilton et al. 2014). This is a fiscal system that involves the imposition of a tax in proportion to the carbon content of the good, combined with a uniform pro rata subsidy to the prices of all goods. Previous research has focused on a conventional (malus only) carbon tax (see, e.g., the tax in Panzone et al. [2018]), which linearly adds the social cost of carbon to current prices, thereby increasing all prices. Instead, the bonus-malus tax results in a net subsidy for products that the paternalistic policy maker considers “desirable” (e.g. low-carbon goods) and a net tax for products that are considered “undesirable” (e.g., high-carbon goods). The subsidy is set at a level that makes the bonus-malus tax scheme revenue neutral from the perspective of the government, therefore removing the income effect of the tax, on average. Consequently, the bonus-malus tax motivates a reduction in carbon footprint through substitution from high-carbon to low-carbon products.

H₄: A bonus-malus carbon tax reduces the carbon footprint of shopping baskets by simultaneously subsidizing low-carbon goods and taxing high-carbon goods, relative to a control group with no carbon tax.

Real-life implementation of bonus-malus taxes is rare. For governments, a key barrier to the implementation of a revenue-neutral bonus-malus tax is that calculating the rate of the subsidy requires anticipating what the levels of consumption will be after the tax is introduced. As a result, the policy maker needs to forecast sales accurately to estimate the size of the subsidy; this task is increasingly feasible for policy makers due to the availability of high-frequency data that can be used proficiently for forecasting purposes. As we have noted, a key advantage of the bonus-malus tax is that it redistributes the tax revenues automatically in the experimental store, leading to an efficient redistribution mechanism.

Interactions between choice architecture and bonus-malus carbon tax

The bonus-malus design of the tax reduces the price of products in the low-carbon aisle, while increasing the price of products in the high-carbon aisle. As a result, this tax incentivizes consumers to switch from high-carbon to low-carbon products, and the choice architecture makes it easier for consumers to find low-carbon products. Thus,

H_5a: The carbon tax increases the strength of the choice architecture manipulation, so that the joint effect is a larger (in absolute value) reduction in the carbon footprint of baskets than the sum of the individual effects.

However, the presence of a carbon tax can reduce the effectiveness of the choice architecture manipulation. The carbon tax may lead the consumer to attribute the desire to reduce the carbon footprint of the shopping to a money-saving goal (Bénabou and Tirole 2003; Bolderdijk et al. 2013). This extrinsic goal may be less effective in reducing the carbon footprint of the basket compared with an intrinsic moral desire to do good for the environment. At the same time, the imposition of the bonus-malus carbon tax may suggest to consumers that low-carbon products are “good” and high-carbon products are “bad” (Hilton et al. 2014), therefore providing an injunctive message that is visible only when the choice architecture is present. As with moral goal priming, the consumer may perceive the combination of both a carbon tax and choice architecture as an intrusive attempt to influence the consumer’s choices, reducing the motivation to use choice architecture to find low-carbon products. In support of this negative effect of the tax on prosocial motivation, the pricing of a socially desirable behavior (such as sustainability) has been associated with a reduction in the number of individuals performing the behavior ⁵ (Frey and Jegen 2001; Perino et al. 2014). As a result, H_5a might be rejected in favor of an alternative hypothesis:

• H_5b: The carbon tax reduces the effectiveness of choice architecture by shifting the attention to a money-saving goal, so that the joint effect is a lower (in absolute value) reduction in the carbon footprint of baskets than the sum of the individual effects.

The Effectiveness of Policy Interventions over Time

The literature exploring possible ways to increase environmentally sustainable consumption is mostly based on lab experiments (e.g., Demarque et al. 2015; Osman and Thornton 2019; Panzone et al. 2018; Reczek, Trudel, and White 2018; Tate, Stewart, and Daly 2014; Theotokis and Manganari 2015) and field experiments where individual behavior is difficult to track (e.g., Baca-Motes et al. 2013; Gneezy et al. 2012; Gravert and Kurz 2019). This literature presents limited evidence of the impact of nudges over time. The evidence currently available indicates that nudges can have long-lasting effects: Allcott and Rogers (2014) show that behavioral interventions targeting a reduction in energy consumption remain effective over time if the nudge is in place, but the impact erodes slowly when the nudge is removed. The present study contributes to this literature by trying to gauge the effectiveness of the three interventions over time in a three-week field experiment.

Method

The Online Supermarket

The experiment was run using an online supermarket called Manchester-Newcastle Online Shopping Research Tool, developed by the University of Manchester and Newcastle University (for the first version of the online supermarket, see Panzone et al. [2018]), based on a similar research tool developed by Demarque et al. (2015). The supermarket was stocked with 651 products currently sold by Tesco in the United Kingdom, including 612 commonly purchased food items and 39 nonfood items (detergents and personal hygiene products). All products were sold at the prices charged by Tesco stores in the same shopping week.

During a shopping trip, the store allows participants to monitor their basket in real time. A “current basket” window presented participants with a list of the products purchased while allowing them to remove or change the quantity of products. Below the basket, participants could see the total spend, total carbon footprint, total kilocalories, and total macronutrients of their shopping basket. These metrics would be updated any time a choice was made or amended. To see the characteristics of the products, participants could click on an icon to obtain the carbon footprint and, for food items, click on another icon to obtain information on nutritional facts (kilocalories and sugar, salt, and fat content). The information was only available by clicking (clicks were not recorded). A search engine would also allow consumers to find products containing a specific keyword in their name (e.g., “organic”) or their category (e.g., meat); this information was not recorded. These features were available to all consumers (including the control group).

Participants

We recruited 306 students from the University of Manchester to take part in the main experiments. Before starting the experiment, participants had to register and supply a personal ID and password for use in all three weeks to ensure that the same individual could be uniquely identified over time. Once registered, participants had to complete an online consent form to ensure that they understood the experimental setup and that they agreed with the terms and conditions of the experiment. Anyone failing to complete these tasks was excluded from the experiments. A total of 273 participants completed the three-week experiment.

Experimental Procedure

The three-week experiment started with a first baseline week (week 0), where all participants shopped with no intervention in place. Participants were then randomly allocated to one of six experimental groups, discussed in the “Experimental Design” subsection, which were present in both the second (week 1) and third (week 2) weeks. Participants remained in the same experimental group in both weeks.

The budget

As part of the experiment, all participants were given a budget of £25 each week to shop in the online store from their own computer anywhere, without the presence of an investigator. The figure of £25 was pretested and found to be in line with the grocery expenditures of a typical student in the study area. The £25 could not be saved or transferred across weeks.

The shopping trip

The online experiment ran over three successive weeks. In each week, participants shopped at any time between 9 a.m. on Monday and midnight on Sunday. They had to spend at least £7, in line with the minimum spending required by supermarkets for online shopping, and could spend as much as they wanted from the £25 budget, knowing that any unspent budget would be returned to them jointly with the participation fee. Participants could enter the store as many times as they wished during the experimental week, but they could only check out from the store once. Baskets could not be updated after checking out. To ensure that this was the case, we deactivated the checkout functionality once the transaction was completed and reactivated it at 9 a.m. the following Monday. Consequently, participants could do their shopping in phases or update and revise their shop before checking out, but they could check out only once a week.

Food delivery

To ensure incentive compatibility, we informed participants that at the end of the three weeks we would randomly select one of the baskets they chose, and they would receive the goods they had ordered in that week and any unspent balance of the £25 budget, plus the participation fee. Participants received the products and the unspent budget balance only for the randomly selected week, and these were received at the end of the study. This step was crucial to ensure that participants were aware that their choices had real consequences: there was a high (33%) chance of getting their basket, and manipulating choices in an attempt to accommodate the experimenter was not an optimal strategy. In the terminology of Harrison and List (2004), this is a “framed field experiment,” an experiment that provides field context (real products in a recognizable shopping environment) to participants. Consequently, although the experimental supermarket we use is not a store for people to shop at regularly, the procedure ensures that the decisions are real. This is a key property of this methodology. To avoid more impulsive shopping patterns in the last experimental week, which was closer to the delivery date, we told participants that delivery would happen two weeks after the end of the last shopping week.

Participation fees

Participants outside Manchester received their goods directly at home, with a participation fee of £8. Participants from the Manchester area collected their baskets from their local Tesco store, receiving a participation fee of £14 (the £6 difference reflects the costs of traveling to a physical store). After the last shopping week, participants were informed which week’s basket had been randomly chosen for them as well as how much cash they would receive ⁶ (unspent budget and fees).

Final questionnaire

At the end of the last experimental week, a questionnaire (available in the Web Appendix) collected information on the participants. Demographics were collected at the recruitment stage. Specifically, the environmental identity scale was adapted from Aquino and Reed (2002). Knowledge of carbon footprint was measured through eight questions asking participants to identify high-carbon options within a pair of goods (similar to Panzone et al. [2020]); four pairs referred to products sold in the online store (knowledge: in-store), and four referred to products not sold in the store (knowledge: out-of-store). Following Cornelissen et al. (2008), we measured proenvironmental attitudes (how consumers feel about environmental behaviors) using three questions: one on moral obligation (how strongly consumers feel morally obliged to protect the environment) and two measures of self-perception. The self-perception measures included general self-perception (whether the mentioned person considers their own behavior environmentally responsible) and purchasing self-perception (whether the person considers the environment when shopping). Note that because this questionnaire measures these metrics only at the end of the third shopping week, we compare only the impact of the two-week exposure to the experimental stimuli with nonexposure (the control group).

Limitations of the experimental procedure

The setup of the online grocery store tool incorporates as much of the online shopping experience as possible for an experiment. However, the experiment has some limitations. First, the incentive compatibility is limited: consumers receive only one of their three baskets, with food arriving two weeks after the end of the experiment (i.e., two to four weeks later, depending on the week of the selected basket). This approach ensures that consumers know that the choices they make in all experimental weeks will have consequences, despite the delay. Second, the budget does not come from the traditional leisure/labor trade-off. However, previous research shows that participants who receive unearned income contribute to public goods as much as those who do not (Clark 2002), an indication that this might not be a major limitation of this experiment. Finally, the knowledge that participants will receive the unspent balance of their participation fee might encourage some participants to underspend, particularly as they cannot use it for immediate consumption. Using a simpler, but similar, setup, Spiteri, James, and Belot (2019) show that nutrient purchases correlate with nutrient consumption, an indication that the presence of the unspent balance may not be problematic. Nevertheless, the design of such an intervention is complex, and we leave the study of these problems for future research, recognizing that although the experiment does not perfectly replicate an online supermarket, it mimics it very closely.

Experimental Design

The experiment followed a 2 (choice architecture: present vs. absent) × 3 (motivation: baseline vs. goal priming vs. tax) design. As indicated previously, week 0 had no intervention; the manipulations were then introduced in week 1 and remained the same in week 2.

Choice architecture

In weeks 1 and 2, participants in the choice architecture manipulation were told that products had been rearranged into three aisles on the basis of their carbon footprint per British pound sterling spent (CO₂e/£). This manipulation allocated the bottom third of products to the low-carbon aisle, the top third to the high-carbon aisle, and the remaining third to the mid-carbon aisle. Participants could shop across aisles with no restrictions. The normalization by expenditures maintains the carbon hierarchy of the products: the Pearson correlation between CO₂e and CO₂e/£ is ρ = .77 (p < .001). Similarly, Table 1 shows that the average product carbon footprint increases progressively when moving from the low-carbon to the high-carbon aisle. The use of CO₂e/£ to allocate products in the carbon aisles ensured that consumers could find products in as many categories as possible ⁷ and classifies some expensive high-carbon products as sustainable because they cause an indirect reduction in the carbon footprint of a basket by reducing the income available to purchase other carbon-emitting goods.

Table 1.

Carbon Footprint and Taxes by Carbon Aisle.

Carbon Aisle	Products	CO2	Pretax Price	Tax	Subsidies	BM Tax	% BM Tax
Low	217	685	£1.74	£.05	£.075	−£.03	−2.74%
Medium	217	1,702	£1.65	£.12	£.076	£.04	1.39%
High	217	6,778	£1.68	£.47	£.075	£.40	24.99%
All products	651	3,055	£1.69	£.21	£.076	£.14	7.88%

Notes: The % BM tax is estimated as the value of the tax divided by the final (i.e., posttax) price.

Compared with Lamberton and Diehl (2013), this manipulation estimates the net impact of a benefit-based assortment by presenting it nested onto an attribute-based assortment that is in both the control and choice architecture condition: within each aisle consumers could find the same organization of products (e.g., meat, fish, vegetables) as all other treatments, and most product categories (14 of 26) offered products in all three aisles. ⁸ The details of food categories allocated to each aisle can be found in Table A1 in the Appendix. Figure 1 shows the information given to participants about this intervention (note that low- and high-carbon aisles were randomly placed at either the top or the bottom of the list); a banner reminded consumers of the intervention throughout the shopping trip.

Figure 1.

Graphical representation of the choice architecture treatment.

Motivation

Baseline (control)

Participants in the baseline motivation condition shopped with a standard shop layout, and no additional intervention or information was added to the online supermarket.

Moral goal priming

In weeks 1 and 2, participants in the moral (environmental) goal-priming treatment could see a banner (Figure A1 in the Appendix) displayed above the shopping area that communicated a clear goal—“Keep Carbon Low”—and a brief rationale—“Caring for the environment is an important moral value. So, choose products with a lower carbon footprint.” The message was designed using an absolutist moral message (a moral imperative), as this has been shown to motivate prosocial behavior (Rai and Holyoak 2013).

Bonus-malus carbon tax treatment

In weeks 1 and 2, when entering the online store, and before entering the aisles, participants were notified that prices had been adjusted to include a specific carbon tax ⁹ of £70/ton CO₂e, which raised prices in proportion to the carbon footprint of each good, and an ad-valorem sales subsidy of 6.5% (the average increase in expenditures caused by the tax in Panzone et al. [2018]). The price of the product after the tax, ${\mathrm{p}}^{\tau }$ , was calculated as ${\mathrm{p}}^{\tau }=(\tau \times \mathrm{C}{\mathrm{O}}_2+{\mathrm{p}}^0)\times (1-\sigma )=[{\mathrm{p}}^0(1-\sigma )]+\{\mathrm{tau}(1-\sigma )\times \mathrm{C}{\mathrm{O}}_2,$ where ${\mathrm{p}}^0$ is the initial price, $\mathrm{\tau }$ is the carbon tax rate, CO₂ is the carbon footprint, and $\mathrm{\sigma }$ is the constant subsidy rate. Figure 4 presents the specific information given to participants, which provided a brief explanation of the mechanism of the carbon tax for transparency. A banner above the shopping aisle, visible throughout the shopping trip, reminded participants of the tax (Figure A2 in the Appendix). When shopping, consumers would see only the full price of the product, inclusive of all taxes and subsidies. Table 1 confirms that the tax worked as a net subsidy for low-carbon products and a net tax for high-carbon products. Table A1 in the Appendix presents the average tax per product class.

Figure 4.

Average ln(carbon footprint) of the basket, by experimental treatment and week.

While the simultaneous presence of a tax and a subsidy may induce some consumers to believe that the two offset each other, the increased price difference between high- and low-carbon products gives consumers the incentive to switch to more sustainable alternatives. A limitation of this manipulation is that participants were aware that the bonus-malus tax applied only to the experimental shop, so they had an incentive to buy fewer high-carbon products in the experimental shop (they were cheaper elsewhere) and to buy more low-carbon products in the experimental store (they were cheaper in the experiment than elsewhere).

Econometric Identification of Treatment Effects

In the experiment, all participants i of personal characteristic d_i shopped over three successive weeks, t = 0, 1, 2. In the first week (t = 0), all participants made their purchase decisions in the absence of any intervention (a baseline); in the second and third weeks (t = 1, 2) the treatments k = 0, 1, …, 5 were introduced and continued in week 2, where k = 0 is the control group, and k = 1, …, 5 are the treatments described in the “Experimental Design” section. The key metric to evaluate behavior is the total carbon footprint (in gCO₂e) of the basket of goods chosen by consumer i at time t, C_it .

Changes in basket carbon footprint from week 0 to week 1 or 2 could be driven by external factors unobservable to the experimenter (e.g., information on media, sustainability events in the local community) rather than the treatment (Dhar and Baylis 2011). These unobservable effects can be removed by comparing the change in consumption to the change in consumption over time of the control group (which is not exposed to any experimental stimuli), effectively removing the change that occurred “naturally” over the same period if an individual was not exposed to the stimuli. This is a difference-in-difference (DID) estimator (Bertrand, Duflo, and Mullainathan 2004; Imbens and Wooldridge 2009), generalized for multiple groups and time periods (Wing, Simon, and Bello-Gomez 2018). In this technique, the impact of the intervention, also termed the average treatment effect (ATE), is the difference between the average change observed in the treated individuals and the change in individuals in the control group over the same time (Imbens and Wooldridge 2009), or

$${\varphi }_{\mathrm{kt}}\equiv [{\overline{\mathrm{C}}}_{\mathrm{kt}}-{\overline{\mathrm{C}}}_{\mathrm{k}0}]-[{\overline{\mathrm{C}}}_{0\mathrm{t}}-{\overline{\mathrm{C}}}_{00}]\mathrm{k}=1,\dots ,5;\mathrm{t}=1,2,$$

(1)

where ${\overline{\mathrm{C}}}_{\mathrm{kt}}$ is the average carbon footprint of individuals in experimental group k in week t, compared with the baseline week 0. Crucially, because the treatment is randomly allocated, there is no self-selection into a treatment, ensuring the absence of endogeneity in the DID estimator. In a linear equation, the ATE is estimated directly from the regression

$${\mathrm{C}}_{\mathrm{it}}={\alpha }_{0i}+\sum _{\mathrm{k}=1}^{\mathrm{k}=5}{\alpha }_{1\mathrm{k}}{\mathrm{G}}_{\mathrm{ik}}+{\alpha }_2{\mathrm{W}}_{\mathrm{t}}+\sum _{\mathrm{k}=1}^{\mathrm{k}=5}{\varphi }_{\mathrm{kt}}{\mathrm{W}}_{\mathrm{t}}{\mathrm{G}}_{\mathrm{ik}}+{\nu }_{\mathrm{it}},$$

(2)

where ${\mathrm{\varphi }}_{\mathrm{kt}}$ is the ATE. In Equation 2, α_0i are fixed individual effects (i.e., preferences and characteristics that are constant over time); G_ik= 1 if i belongs to experimental group k = 1, …, 5, and 0 otherwise; and W_t= 1 are time dummies if t = 1, 2, and 0 otherwise.

In our analyses, the dependent variable is log-transformed to mitigate problems of nonnormality and heteroskedasticity (see Manning and Mullahy 2001), and we estimate the nonlinear regression

$${\mathrm{y}}_{\mathrm{it}}=\ln ({\mathrm{C}}_{\mathrm{it}})={\alpha }_{0\mathrm{i}}+\sum _{\mathrm{k}=1}^{\mathrm{k}=5}{\alpha }_{1\mathrm{k}}{\mathrm{G}}_{\mathrm{ik}}+{\alpha }_2{\mathrm{W}}_{\mathrm{t}}+\sum _{\mathrm{k}=1}^5{\theta }_{\mathrm{kt}}{\mathrm{W}}_{\mathrm{t}}{\mathrm{G}}_{\mathrm{ik}}+{\nu }_{\mathrm{it}}.$$

(3)

In a nonlinear equation such as Equation 3, the interaction term θ_kt is not the DID estimator, and ${\mathrm{\varphi }}_{\mathrm{kt}}$ is a function of θ_kt (for more detail, see Athey and Imbens [2006] and Puhani [2012]). In this article, the ATE can then be calculated as

$${\varphi }_{\mathrm{ikt}}=\exp ({\hat{\mathrm{y}}}_{\mathrm{it}})-\exp \left({\hat{\mathrm{y}}}_{\mathrm{it}}-\sum _{\mathrm{k}=1}^5{\theta }_{\mathrm{kt}}{\mathrm{W}}_{\mathrm{t}}{\mathrm{G}}_{\mathrm{ik}}\right).$$

(4)

Results

This section presents the results of the statistical analyses. The 273 participants shopped over three weeks, resulting in 819 baskets containing a total of 10,606 choices (product-consumer-week combinations, with an average of 12.9 choices/person/week), purchasing 14,858 items. ¹⁰ In the analyses, the carbon footprint variable is the total gCO₂e in a participant’s shopping basket. Note that we use a 10% significance level for judgments on significance because DID analyses are known to be noisy and overestimate standard errors (Bertrand et al. 2004).

Descriptive Characteristics of the Sample

Table 2 presents the demographic characteristics and the psychological profile of the sample. Environmental identity consists of two components: internalization, which measures how central environmental identity is to the respondent’s self-concept, and symbolization, which measures how much the respondent believes their actions reflect this identity. These two variables were obtained from the answers to 11 statements, averaging the items in each factor as identified by Aquino and Reed (2002) (Table AO1 in the Web Appendix reports the statements, together with the results of a principal component analysis with varimax rotation, which are similar to those in Aquino and Reed’s article). Environmental attitudes refer to the average of the three attitudinal questions. Self-perception items were kept separate because of a low Cronbach’s α (< .7; see Lance, Butts, and Michels 2006). These constructs are weakly or moderately correlated (see Table OA2 in the Web Appendix). Finally, in-store knowledge and out-of-store knowledge are the sum of the knowledge questions in each scale.

Table 2.

Average Characteristics of the Sample of Participants.

	Total	Control	CA	GP	Tax	CA × GP	CA × Tax	Kruskal–Wallis χ2(5)	Cronbach’s α
Age (years)	23.47	23.77	23.97	23.60	23.27	22.93	23.28	2.54	—
Male (0/1)	.39	.38	.36	.51	.45	.35	.27	6.75	—
Year of study (1–5)	2.72	2.67	2.82	2.57	3.09	2.52	2.63	6.81**	—
British (0/1)	.55	.58	.58	.65	.40	.70	.39	14.54**	—
Env. identity: Internalization (1–7)	5.32	5.45	5.20	5.16	5.36	5.38	5.41	2.40	.7271
Env. identity: Symbolization (1–7)	3.15	3.49	3.41	2.69	3.19	3.17	2.98	11.40**	.8911
CO2e knowledge: Internal (0–4)	1.21	1.16	1.29	1.20	1.21	1.26	1.15	.69	.5486
CO2e knowledge: External (0–4)	1.12	1.09	.98	1.27	1.17	1.00	1.20	3.04	.5918
Environmental attitude (1–7)	5.35	5.52	5.35	5.21	5.52	5.16	5.37	4.86	.8370
Self-perception: General (1–7)	4.60	4.73	4.22	4.78	4.91	4.22	4.76	18.21***	—
Self-perception: Purchasing (1–7)	3.64	3.98	3.41	3.62	3.72	3.39	3.76	5.94	—
Moral obligation (1–7)	5.25	5.58	5.11	5.02	5.45	5.07	5.32	7.49	—
Participants	273	45	45	49	47	46	41

*p < .10.

**p < .05.

***p < .01.

Notes: CA = choice architecture; GP = goal priming.

The student sample contains mainly young participants (84.6% were in the 18- to 25-year age range) with mixed nationality (55% British nationals). Participants in the sample have strong attitudes, internalized environmental identity, and moral obligation; moderate symbolized environmental identity and self-perceptions; and low knowledge. Kruskal–Wallis tests show that experimental groups differ significantly in the year of study and British nationality; moreover, the general self-perception and symbolization components of the environmental identity differ across experimental groups and are lowest under goal priming (see the “Food Delivery” subsection).

The Carbon Footprint and Economic Structure of the Basket

Table 3 indicates that, apart from the control, consumers spent slightly less in weeks 1 and 2 compared with week 0 in all experimental groups (exception made for the joint choice architecture and goal-priming group). The reduction in expenditures is driven primarily by consumers buying fewer items, and, to a lesser extent, paying higher unit prices. The slight (nonsignificant) decrease in consumption in the control group could be due to several reasons, such as the store not selling a favorite brand, which would only be known after week 0, or the participant feeling tired of the experiment.

Table 3.

Changes in Average Basket Metrics over Time.

	Week	Control	CA	GP	Tax	CA × GP	CA × Tax	Kruskal–Wallis χ2(5)
Expenditures	0	21.62	21.16	21.85	21.23	21.88	21.45	1.40
	1	21.65	20.31*	20.94*	19.76**	21.30	20.71*	2.67
	2	20.40	19.98**	20.39**	18.86***	20.55	20.09**	3.47
Items	0	19.53	19.33	19.69	19.66	20.30	17.95	1.91
	1	18.67	17.87***	17.76**	16.85*	19.09	17.59	2.62
	2	18.16	17.49***	16.16***	16.09***	18.04***	16.17	2.72
Price	0	1.23	1.31	1.31	1.28	1.28	1.40	1.20
	1	1.26	1.40*	1.41	1.49***	1.38	1.44	1.20
	2	1.30	1.44	1.52***	1.57*	1.44	1.43	2.01
Entropy	0	1.00	1.00	.95	.98	.97	.99	3.61
	1	.97	.73***	.97	.94	.81***	.82***	23.57***
	2	.95	.70***	.89	.85***	.70***	.77***	21.77***

*p < .10.

**p < .05.

***p < .01.

Notes: Differences in carbon footprint from the baseline (week 0) are based on a Wilcoxon matched-pairs signed-rank test. Differences between experimental groups are tested using a Kruskal–Wallis test. CA = choice architecture; GP = goal priming.

The manipulations reduced the carbon-aisle variety within a food basket. Variety is defined in terms of the proportion of items, known as SKUs, within each carbon aisle. Specifically, this variable is measured by the entropy of each individual basket (Van Herpen and Pieters 2002), calculated as ${\eta }_{\mathrm{it}}=-\sum _{\mathrm{l}=1}^{\mathrm{L}}{\mathrm{s}}_{\mathrm{ilt}}\ln ({\mathrm{s}}_{\mathrm{ilt}})$ , where ${\mathrm{s}}_{\mathrm{il}}$ is the proportion of SKU purchased by individual i in aisle l at time t, not accounting for quantity. This implies that consumers making all choices in a single aisle have zero entropy (no variety), whereas a consumer making exactly one third of its choices in each category has entropy ${\eta }_{\mathrm{it}}=-3[(1\mathrm{\setminus }3)\ln (1\mathrm{\setminus }3)]=1.099$ (maximum variety). Table 3 shows that, apart from the control and the goal-priming manipulation, variety decreased over time in all experimental groups, an indication that consumers concentrated their expenditures in fewer carbon aisles.

Table 4 shows that the GHG of shopping baskets decreased significantly over the three experimental weeks in all five experimental groups, with no significant change in the control group. The tax and the choice architecture groups show the strongest effect size, although the effect of choice architecture is significant only in week 2. Goal priming only shows a weakly significant change in carbon footprint of the basket in both weeks 1 and 2. Kruskal–Wallis tests show that the carbon footprint of baskets differed across the six experimental groups in both weeks 1 (marginally) and 2. Table OA3 in the Web Appendix presents a similar comparison for the nutritional composition of the food baskets. ¹¹

Table 4.

Summary CO₂e Consumption Statistics, by Treatment and Week.

Condition		Week			Trend Test
		0	1	2
Control	Mean	26,221	25,128	23,788	−.86
	SD	14,861	10,718	13,506
	Cohen’s d	—	.0853	.1733
CA	Mean	25,877	26,255	18,313***	−2.89***
	SD	12,806	17,936	11,058
	Cohen’s d	—	−.0245	.6394
GP	Mean	31,846	25,714*	25,185*	−1.69*
	SD	20,812	10,951	14,345
	Cohen’s d	—	.3726	.3765
Tax	Mean	25,593	20,806***	18,304***	−2.72***
	SD	13,944	11,316	9,311
	Cohen’s d	—	.3811	.6214
GP and CA	Mean	28,063	26,163	22,998**	−1.54
	SD	18,259	19,048	13123
	Cohen’s d	—	.1029	.3220
Tax and CA	Mean	26,418	20,863**	19,601***	−2.16**
	SD	14,686	11,543	10,244
	Cohen’s d	—	.4258	.5451
Kruskal–Wallis χ2(5)		2.699	9.371*	12.146**

*p < .10.

**p < .05.

***p < .01.

Notes: Differences in carbon footprint from the baseline (week 0) are based on a Wilcoxon matched-pairs signed-rank test. CA = choice architecture; GP = goal priming.

Manipulation Checks

Choice architecture

Results indicate that choice architecture operates primarily by simplifying the search for low-carbon goods, consequently reducing carbon-aisle variety. Repeated-measure analysis of variance (ANOVA)–style tests (Table A2 in the Appendix) show that consumers who have access to the carbon-based choice architecture are more likely to make a first purchase in the low-carbon aisle and make a larger percentage of choices in the low-carbon aisle (measured in terms of the proportion of SKUs bought per aisle). This behavior translates into a reduction in the entropy of food baskets. The strategy adopted by consumers in the choice architecture manipulations is shown graphically in Figure 2: all consumers who have access to the choice architecture are more likely to make an initial sequence of choices in the low-carbon aisle compared with the other experimental groups. Table 2 and ANOVAs (Tables A3 and A4 in the Appendix) provide no evidence that choice architecture had an impact on the psychological constructs we measured or increased product-specific carbon footprint knowledge.

Figure 2.

Probability of purchasing in the low-carbon aisle in the sequence of purchases, by treatment and week.

Moral goal priming

ANOVA and ANOVA-style tests in Table A3 in the Appendix show that the goal-priming manipulation operated by making participants insecure about their interest in the environment. Specifically, participants primed with a moral goal reported lower general self-perception and lower symbolized environmental identity. Table 2 and the ANOVAs (Tables A3 and A4 in the Appendix) indicate that no other psychological construct was affected.

Bonus-malus carbon tax

Figure 3 shows that the tax increases the price of high-carbon goods and decreases the price of low-carbon goods; the increase in price of high-carbon goods is larger (in absolute value) than the decrease in price of low-carbon goods. Table 2 and the ANOVAs (Tables A3 and A4 in the Appendix) find no evidence that the tax impacted the psychological constructs we measured.

Figure 3.

Final product price and carbon footprint level of products.

DID Analysis

Next, we present the estimated impact of each intervention using the DID estimator described previously. This analysis advances the simple comparison of means presented in the “Manipulation Checks” subsection by accounting for systematic differences in carbon emission levels across individuals (within-respondent variation) and removing unobservable factors (e.g., computer literacy, news in the media) that have influenced all experimental groups within a given week. Crucially, the analysis estimates the impact of the intervention as the reduction in carbon footprint in an experimental group minus the same reduction in the control group, thereby correcting the treatment effect for the (small) natural reduction in carbon footprint.

Trend tests

A key assumption of DID estimators is that any relevant unobservable variable is either a group attribute that is time-invariant (consumer fixed effects in our subsequent analysis) or a time-varying factor that is constant within each experimental group (week dummies in our analysis) (Angrist and Pischke 2009; Wing, Simon, and Bello-Gomez 2018). However, different groups can be characterized by different trends because, for instance, participants in an experimental group were in the process of reducing their carbon footprint before the experiment started. More generally, individuals in both the control and an experimental group may differ in time-varying observable or unobservable characteristics, which cause an omitted variable bias. This is particularly relevant when participation in the experiment is voluntary.

Because our experiment has only one week of data before the experiment started, it is not possible to test for parallel pretreatment trends. However, the randomization of participants into experimental groups precludes the need for testing pretrends (Borusyak and Jaravel 2018; Keller, Geyskens, and Dekimpe 2020; Wing, Simon, and Bello-Gomez 2018) because the assignment of individuals to the different experimental groups is unrelated to environmental preferences (Table 2) and expenditures (Table 3). Table 4 shows that experimental group membership is unrelated to the pretreatment (week 0) level of carbon footprint. Figure 4 represents graphically the change in consumption (log-carbon footprint) across groups and time (Wing, Simon, and Bello-Gomez 2018); nonparametric trend tests in Table 4 also indicate that, apart from the control, all experimental groups had a significant downward sloping trend in carbon footprint.

DID estimates of carbon reductions

This section implements the generalized DID estimator, regressing the logarithm of the total carbon footprint in the basket of consumer i at time t over week-specific dummies (fixed time effects) and the interactions between week and manipulation variables. Regressions also use dummies for the day of checkout. Using a fixed-effects panel estimator, we remove time-invariant individual effects (e.g., psychometric variables) that could be correlated with the error term of Equation 3, and the (fixed) experimental group membership. In all regressions, the control group is the baseline treatment dummy, and experimental stimuli are used as main effects, assigning a dummy variable whenever a specific intervention (e.g., carbon tax) is present, independently of whether it is the only intervention or not. This approach enables a direct test of H_3a/H_3b and H_5a/H_5b. Estimates of the standard errors are obtained through bootstrapping, stratifying by experimental group; to remove serial autocorrelation, the bootstrapping procedure clusters residuals at individual consumer level ¹² (Bertrand et al. 2004; Cameron and Miller 2015). A cluster-robust Hausman test shows no difference between the fixed- and random-effects estimator in all equations; we report the fixed effects specification to align with the existing DID literature.

Table 5 reports the estimated parameters of the DID regressions. In all regressions, the dependent variable is the natural logarithm of total carbon footprint of the basket (gCO₂e). Model A regresses this variable over the experimental manipulations only, Model B adds total expenditures (in natural logarithm) to adjust for basket size, Model C adds entropy to adjust for the different composition of the basket, and Model D adds the total number of units purchased to control for the potential strategy of “buying less” to keep the carbon footprint low. Three additional regressions estimate the impact of the manipulations on entropy (E), in-store expenditures (F), and number of items purchased (G).

Table 5.

DID Estimates of the Impact of Treatments on Carbon Footprint.

	Model A	Model B	Model C	Model D	Model E	Model F	Model G
	ln(CO2e)	ln(CO2e)	ln(CO2e)	ln(CO2e)	Entropy	ln(expenditure)	ln(items)
Intercept	10.100***	7.1326***	7.0686***	7.0493***	.9727***	3.0357***	2.8602***
SE	.0657	.2926	.2750	.2700	.0342	.0357	.0533
Week 1	.0076	.0150	.0301	.0322	−.0282	−.0076	−.0267
SE	.0749	.0735	.0730	.0730	.0230	.0252	.0472
Week 2	−.1277	−.0499	−.0334	−.0299	−.0511*	−.0795***	−.1027**
SE	.0798	.0697	.0660	.0661	.0264	.0258	.0502
CA: Week 1	−.1662	−.1027	.0238	.0254	−.2369***	−.0649	−.0982
SE	.1182	.1097	.1042	.1044	.0614	.0470	.0617
CA: Week 2	−.3132***	−.3006***	−.1588*	−.1576*	−.2485***	−.0129	−.0524
SE	.1181	.1073	.0935	.0941	.0588	.0529	.0715
GP: Week 1	−.1711*	−.1175	−.1504	−.1445	.0413	−.0548	−.0892
SE	.0984	.0960	.0926	.0926	.0354	.0398	.0882
GP: Week 2	−.1047	−.0825	−.0803	−.0670	−.0104	−.0228	−.1270
SE	.1114	.1000	.0991	.1000	.0528	.0470	.0932
Tax: Week 1	−.2139**	−.1210	−.1276	−.1147	−.0157	−.0950**	−.1858**
SE	.0867	.0874	.0869	.0872	.0367	.0468	.0841
Tax: Week 2	−.2149**	−.1265	−.0950	−.0806	−.0802*	−.0904	−.2024**
SE	.1031	.0942	.0868	.0872	.0482	.0579	.0965
CA + GP: Week 1	.2171	.1410	.1153	.1092	.0666	.0778	.1231
SE	.1543	.1388	.1324	.1310	.0931	.0668	.1156
CA + GP: Week 2	.3060*	.2929*	.2702*	.2577*	.0430	.0135	.1170
SE	.1655	.1513	.1421	.1395	.0905	.0748	.1196
CA + Tax: Week 1	.0926	−.0165	−.0650	−.0815	.1153	.1117	.2419**
SE	.1651	.1467	.1324	.1332	.0857	.0847	.1159
CA + Tax: Week 2	.2943	.2161	.1355	.1173	.1619	.0800	.2348*
SE	.1803	.1564	.1298	.1308	.1010	.0922	.1226
ln(expenditure)	—	.9775***	.8131***	.7068***	—	—	—
SE		.0994	.0874	.1109
Entropy	—	—	.5790***	.5632***	—	—	—
SE			.1016	.1015
ln(Number of products)	—	—	—	.1250**	—	—	—
SE				.0607
Day of checkout	Yes	Yes	Yes	Yes	Yes	Yes
Hausman test	10.11	8.01	12.78	18.12	10.91	8.66	8.54
Observations	819	819	819	819	819	819	819
Participants	273	273	273	273	273	273	273
R2	.1403	.3137	.3937	.3998	.1947	.1061	.09948
χ2	87.24***	170.23***	221.54***	241.96***	114.42***	47.27***	57.70***

*p < .10.

**p < .05.

***p < .01.

Notes: CA = choice architecture; GP = goal priming. SEs are based on bootstrapping (200 replications), stratified at the treatment level (six strata), and clustered at the individual consumer level. Model A regresses the carbon footprint (in natural logarithm) over the experimental stimuli only; Model B adds total expenditures (in natural logarithm); Model C adds the entropy of the basket; Model D adds the total number of units purchased; and Models E, F, and G regress entropy, in-store expenditures, and number of items purchased over the experimental stimuli only, respectively.

A first result is that choice architecture successfully reduces the carbon footprint of food baskets by around 31%, but only in week 2, providing partial support for H₁. In week 1, the same manipulation achieves a 17% reduction, not significantly different from zero (p = .150 in Model A) because the response is heterogeneous (the standard error is large). This result suggests that the manipulation required a learning phase that could have long-lasting benefit. The effect of choice architecture disappears when adjusting for entropy (equation E), an indication that this manipulation operated through a reduction in carbon-aisle variety.

Second, moral goal priming shows a 17% reduction in carbon footprint in Model A, a marginally significant effect only in week 1, and a 10% reduction in week 2 (not significant), therefore providing weak support for H₂. Third, the bonus-malus carbon tax reduces the carbon footprint of the basket, and the effect is significant in both experimental weeks, accounting for a stable 21% reduction in carbon footprint in Model A. This result provides evidence in support of H₄. The impact of the tax is mediated by a reduction in in-store expenditures (equation F) and in the number of items in the basket (equation G).

Finally, in week 2 choice architecture interacts with the tax and goal priming. However, interactions increase carbon footprint, somewhat offsetting the effects of choice architecture. Interactions are only marginally significant, weakly supporting H_3b and H_5b. Note, however, that the total effect of the interactions is still negative (i.e., an overall reduction in carbon footprint).

Table 6 reports the marginal effects of the DID coefficients. The marginal effects of the choice architecture and the bonus malus tax manipulations are large, indicating the benefits these interventions can have.

Table 6.

Estimated Marginal Effects from the DID Regression.

	Week	Model A		Model B		Model C		Model D
		Margin	SE	Margin	SE	Margin	SE	Margin	SE
CA	1	−3,387.6	2,321.5	−2,214.1	2,324.4	539.9	2,376.5	579.0	2,390.2
CA	2	−5,333.7***	1,989.7	−5,694.6***	1,995.6	−3,179.2*	1,859.0	−3,170.2*	1,876.8
GP	1	−3,459.3*	1,959.0	−2,508.6	2,026.8	−3,202.6*	1,947.9	−3,092.1	1,956.4
GP	2	−1,896.2	1,997.9	−1,672.0	2,001.9	−1,645.0	2,003.6	−1,386.0	2,045.6
Tax	1	−4,241.3**	1,692.5	−2,574.2	1,839.7	−2,731.4	1,840.5	−2,475.8	1,862.8
Tax	2	−3,723.9**	1,799.0	−2,515.5	1,858.0	−1,932.4	1,758.6	−1,656.2	1,783.7
CA + GP	1	5,168.7	4,095.6	3,353.5	3,565.4	2,734.7	3,348.2	2,592.2	3,303.5
CA + GP	2	6,639.3	4,261.1	7,014.3*	4,252.2	6,469.5	3,940.6	6,162.3	3,837.2
CA + Tax	1	2,083.9	3,888.4	−366.2	3,220.1	−1,420.3	2,793.8	−1,774.2	2,772.4
CA + Tax	2	6,368.5	4,612.0	5,001.8	4,103.7	3,050.7	3,175.9	2,629.1	3,152.2

*p < .10.

**p < .05.

***p < .01.

Notes: Standard errors are calculated using the Delta method. CA = choice architecture; GP = goal priming. Column names refer to the models of Table 5, as follows: Model A regresses the carbon footprint of the basket over the experimental manipulations only, Model B adds total expenditures, Model C adds entropy, and Model D adds the total number of units purchased.

To understand how demographics and psychometrics influence the carbon footprint of shopping baskets, we regress the fixed effects (measured in log-carbon footprint) from the panel regression, estimated as ${\alpha }_{0\mathrm{i}}=\ln ({\mathrm{C}}_{\mathrm{it}})-\sum _{\mathrm{k}=1}^{\mathrm{k}=5}{\alpha }_{1\mathrm{k}}{\mathrm{G}}_{\mathrm{ik}}-{\alpha }_2{\mathrm{W}}_{\mathrm{t}}-\sum _{\mathrm{k}=1}^5{\theta }_{\mathrm{kt}}{\mathrm{W}}_{\mathrm{tG}}\mathrm{ik}-{\nu }_{\mathrm{it}}$ , over time-invariant variables. To limit the impact of multicollinearity, we include only those variables that vary significantly across experimental groups in Table 1. Because the dependent variable (the fixed effects) is estimated rather than observed, standard errors are based on bootstrap estimates. Results (Table 7) indicate that carbon emissions are higher in male and British consumers, and individuals with stronger symbolized environmental identity have baskets with a lower carbon footprint. The inclusion of the five treatment dummies formally tests whether the six experimental groups had the same carbon footprint in week 0, a hypothesis that is rejected only for goal priming in regression C. Results are stable across specification.

Table 7.

Impact of the Fixed Effects on the Carbon Footprint of the Basket.

	Model A		Model B		Model C		Model D
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
Intercept	−.2589	.2094	−.1502	.1405	−.1637	.1410	−.1381	.1371
Male	.1272**	.0595	.1279***	.0431	.1309***	.0426	.1488***	.0424
British	.3702***	.0613	.1490***	.0446	.1497***	.0431	.1306***	.0426
Age	.0058	.0057	.0037	.0043	.0029	.0043	.0025	.0041
Identity: Symbolization	−.0485*	.0248	−.0410**	.0182	−.0329*	.0176	−.0366**	.0175
Self-perception: General	−.0075	.0273	.0020	.0185	.0002	.0182	.0003	.0184
CA	.0491	.1036	.0690	.0700	.0652	.0646	.0662	.0628
GP	.1116	.1009	.1145	.0713	.1507**	.0687	.1472**	.0691
Tax	.0255	.1154	.0054	.0745	.0158	.0708	.0097	.0704
CA + GP	−.1276	.1386	−.1469	.1041	−.1535	.1010	−.1519	.0997
CA + Tax	.0472	.1508	.0082	.1061	.0102	.1017	.0193	.1007
Observations	273		273		273		273
R2	.1613		.1067		.1148		.1182
Adapted R2	.1293		.0726		.0810		.0846
χ2	54.32***		35.96***		38.35***		38.87***

*p < .10.

**p < .05.

***p < .01.

Notes: SEs are based on bootstrapping (200 replications), stratified at treatment level (six strata). CA = choice architecture; GP = goal priming. Column names refer to the models of Table 5, as follows: Model A regresses the carbon footprint of the basket over the experimental manipulations only, Model B adds total expenditures, Model C adds entropy, and Model D adds the total number of units purchased.

DID analysis of budget allocations

The reduction in carbon footprint observed in the previous subsection was achieved through a change in the composition of the basket. This section explores the changes in the share of consumer expenditure in the three carbon aisles, plus the share of savings (the budget kept for consumption outside the experimental store). We estimate the impact of the experimental manipulations using a random-effects panel Tobit estimator to account for the truncation at zero of expenditure shares. All regressions adjust for the day of checkout. Parameters are presented with bootstrapped standard errors (200 replications), stratified at the level of the experimental group, clustered at individual consumer level. Each equation is estimated independently.

Results (Table 8) indicate that the significant drop in carbon footprint observed in the choice architecture treatment in week 2 originates from consumers substituting high-carbon options with low-carbon options, and the more moderate drop in week 1 was characterized by consumers substituting from medium- to low-carbon options. Conversely, the reduction in carbon footprint in the presence of a tax is driven by an increase in the share of unspent budget in week 1 (the week 2 coefficient is of comparable size, but not significant). Goal priming primarily motivated the substitution from the medium- to the low-carbon aisle, with a nonsignificant reduction in carbon footprint. The interaction of the instruments motivates the substitution away from low-carbon items toward high-carbon items.

Table 8.

Impact of the Manipulations on the Expenditure Allocation Decision.

Budget Share	Low Carbon	Medium Carbon	High Carbon	Savings
Intercept	.2177***	.3375***	.2363***	.1885**
SE	.0677	.0577	.0656	.0874
Week 1	−.0337	.0240	.0038	.0068
SE	.0283	.0235	.0220	.0186
Week 2	−.0308	−.0150	−.0123	.0564***
SE	.0297	.0231	.0239	.0180
CA: Week 0	−.0345	−.0094	.0303	.0182
SE	.0359	.0306	.0309	.0489
CA: Week 1	.0384	−.1010***	−.0156	.0269
SE	.0426	.0332	.0441	.0304
CA: Week 2	.1275**	−.0637*	−.1111***	−.0140
SE	.0505	.0371	.0386	.0350
GP: Week 0	−.0240	.0136	−.0143	.0190
SE	.0408	.0347	.0335	.0437
GP: Week 1	.0548	−.0798**	−.0016	.0308
SE	.0389	.0353	.0267	.0258
GP: Week 2	−.0131	−.0197	.0211	.0033
SE	.0405	.0385	.0366	.0298
Tax: Week 0	−.0122	−.0037	.0270	−.0044
SE	.0386	.0299	.0316	.0468
Tax: Week 1	.0075	−.0454	−.0278	.0595**
SE	.0399	.0329	.0305	.0295
Tax: Week 2	.0257	−.0455	−.0481	.0419
SE	.0417	.0362	.0333	.0354
CA + GP: Week 0	.0045	−.0138	.0254	−.0179
SE	.0587	.0445	.0480	.0591
CA + GP: Week 1	−.0426	.1206**	−.0114	−.0368
SE	.0590	.0546	.0548	.0420
CA + GP: Week 2	−.0169	.0052	.0325	.0093
SE	.0683	.0589	.0581	.0483
CA + Tax: Week 0	.0512	.0070	−.0408	−.0340
SE	.0616	.0401	.0443	.0684
CA + Tax: Week 1	.0266	.0764*	−.0030	−.0558
SE	.0645	.0454	.0564	.0510
CA + Tax: Week 2	−.1275*	.0894	.1186**	−.0150
SE	.0697	.0587	.0560	.0563
Male	−.0936***	.0442**	.0458**	−.0005
SE	.0244	.0198	.0216	.0270
British	.0252	.0617***	.0728***	−.1482***
SE	.0247	.0203	.0173	.0280
Age	.0021	.0000	−.0004	−.0009
SE	.0021	.0015	.0014	.0021
Identity: Symbolization	.0228**	−.0194**	−.0130*	.0057
SE	.0097	.0083	.0077	.0112
Self-perception: General	.0029	−.0040	−.0051	.0057
SE	.0105	.0091	.0085	.0124
Day of checkout	Yes	Yes	Yes	Yes
% zeros: Week 0	3.30%	1.83%	4.03%	3.30%
% zeros: Week 1	8.42%	9.16%	8.79%	2.20%
% zeros: Week 2	10.62%	12.82%	17.21%	4.40%
Observations	819	819	819	819
Participants	273	273	273	273
χ2	48.14***	93.22***	65.32***	82.64***

*p < .10.

**p < .05.

***p < .01.

Notes: Random-effects panel Tobit estimation. SEs are based on bootstrapping (200 replications), stratified at treatment level (six strata), and clustered at the individual consumer level. CA = choice architecture; GP = goal priming.

Finally, British consumers purchased more mid- and high-carbon goods and kept smaller shares of the budget, and male consumers purchased more mid- and high-carbon goods and fewer low-carbon goods. This result indicates that, on average, men and British consumers behave differently from the rest. Symbolized proenvironmental identity is linked with a larger budget share spent on low-carbon goods and a lower share on mid- and high-carbon goods.

Discussion

The aim of this research was to explore the efficacy of a choice architecture intervention that organizes products into three “aisles” of low, medium and high carbon footprints to facilitate a reduction in the carbon footprint of grocery shopping baskets. The effectiveness of the intervention was tested in the presence of two novel policy and marketing instruments: a bonus-malus carbon tax (D’Haultfœuille, Givord, and Boutin 2014; Hilton et al. 2014) and a banner priming a sustainability goal. Key research questions are whether these stimuli have a significant role in complex incentive-compatible situations where consumers are exposed to a multiplicity of stimuli and whether these interventions can bring sustained changes in behavior over time. To answer these questions, we designed a unique incentive-compatible experimental supermarket where consumers make real (binding) choices. We advance previous experiments (e.g., Demarque et al. 2015; Panzone et al. 2018) by having an online experiment, with no personal interaction between experimenter and participants. A key feature of the method is the large amount of information that the online retailer generates, allowing for the study of consumer behavior using advanced marketing analytics. This section summarizes and contextualizes these results to understand their implications for public policy and marketing practice.

What Motivates Reductions in Carbon Footprint? A Summary of the Results

Results indicate that choice architecture led to strong and significant reductions in carbon footprints in week 2. This intervention leads to a nonsignificant reduction in carbon footprint (−17%) in week 1, when consumers reduced their expenditures on medium-carbon goods, and to a significant reduction in carbon footprint (−31%) in week 2, when consumers switched from high-carbon to low-carbon goods. The delayed response might be caused by a learning phase in which participants become acquainted with the new structure of the choice set or learn about the benefits of this reorganization. Results suggest that this manipulation operates by reducing search costs, in particular by making environmental information more easily accessible. Consumers who have access to the choice architecture are more likely to start shopping in the low-carbon aisle and to make more choices there. The manipulation has no impact on the psychological variables collected in this experiment; however, previous research indicates that benefit-based choice architecture activates an abstract construal of the choices (Lamberton and Diehl 2013), which in turn motivates proenvironmental behavior (Reczek, Trudel, and White 2018; White, MacDonnell, and Dahl 2011).

The bonus-malus carbon tax resulted in a smaller but still significant reduction in carbon footprint compared with the choice architecture (−21%), an effect that remained stable over time. These results are in line with literature showing that environmental taxes can motivate dietary change (Kehlbacher et al. 2016; Panzone et al. 2018). The bonus-malus tax operates by making low-carbon options cheaper and high-carbon options more expensive. The bonus-malus tax had no impact of on the psychological variables collected in this experiment. In response to the tax, consumers reduced the amount of money spent in the experimental store. The data do not enable us to explore the environmental impact of the unspent budget. However, the tax gave consumers the incentive to buy higher-carbon products at a lower price outside the experimental store while buying the cheaper low-carbon goods in the experimental store (because they were shopping from home, they could check the difference in price).

Finally, moral goal priming is the manipulation with the smallest main effect size (−6% to −17%)—in the right direction but only marginally significant in week 1. This result is in line with literature showing that goal priming motivates virtuous choices (Lindenberg and Steg 2007; Tate, Stewart, and Daly 2014; Walsh 2014). Results indicate that the priming message had a negative impact on the environmental identity and environmental self-perception of consumers, which may have reduced the value they assigned to proenvironmental choices. This weak response may be due to the use of an injunctive message, which has been associated with a reduction in prosocial behavior (Kavvouris, Chrysochou, and Thøgersen 2020; White and Simpson 2013). This result indicates that persuasive message design is a key element of a priming manipulation. A targeted message that consumers could monitor (e.g., “Keep your carbon footprint below X gCO₂e”) might have been easier to monitor, while the more generic “Keep Carbon Low” might be causing the decline in self-concept by instilling doubts in the mind of consumers (e.g., “If I do not know what ‘low-carbon’ is, then I am not an environmentally caring person”).

The interaction of choice architecture with either tax or goal priming partially offsets the reduction of choice architecture alone (marginally significantly only in week 2). In other words, interactions resulted in a reduction in the carbon footprint of baskets, but by less than the sum of the reduction of each individual experimental stimulus. This result is consistent with the literature on reactance (Kavvouris, Chrysochou, and Thøgersen 2020; Ma, Dixon, and Hmielowski 2019; Reich and Robertson 1979; Zemack-Rugar et al. 2017). In the case of the tax, reactance is consistent with research on motivational crowding out (Frey and Jegen 2001; Perino et al. 2014): consumers may use the low-carbon aisle to signal interest in prosocial behavior to the self and to others; the carbon tax motivates consumers to do so to avoid the tax. Conversely, goal priming weakens the effect of choice architecture, perhaps because the priming message has a negative impact on participants’ self-concept (consistent with Laran, Dalton, and Andrade [2010], Ma, Dixon, and Hmielowski [2019], White and Simpson [2013], and Zemack-Rugar et al. [2017]). Overall, these results suggest that choice architecture is driven by intrinsic proenvironmental motivations, which are negatively affected by the presence of an injunctive priming message and by a tax.

The Duration of the Effects of the Interventions

This research indicates that a key component of policy design is time. Most behavioral and experimental literature does not explore the impact of nudges over multiple periods—whether they wear off or grow over time (Allcott and Rogers 2014). The duration of the effects of a (costly) intervention is important for policy makers in business and government, because it affects the total social benefits the intervention generates over time. We aimed to address this question by having interventions that took place over two weeks (after one baseline week). This time horizon is certainly short, but results give a flavor of how policy interventions differ in their ability to generate consistent low-carbon choices. Carbon taxes resulted in an immediate and persistent negative effect on the carbon footprint of the baskets. Choice architecture led to reductions in carbon footprint that grew in magnitude over time. The effect of goal priming reduced over time. Further research should explore ways to extend the time horizon, to determine whether these effects are long-lasting or whether they lead to some sort of behavioral adaptation, maybe using consumer panels over many time periods. The added complexity of analyzing a larger data set over a long time period would be compensated by in-depth learning about the dynamics of compliance to prosocial interventions.

Implications for Public Policy

Consumers often find it difficult to engage with environmental preservation efforts (Gifford 2011; Whitmarsh, Seyfang, and O’Neill 2011), limiting the impact of interventions that require consumers to actively engage to protect the environment. Directly changing the choice set or the shopping infrastructure can be an easy way to facilitate large-scale changes in behavior, nudging consumers into the construction of lower-carbon baskets “by design” (Sunstein and Reisch 2014; Theotokis and Manganari 2015). Results in this article indicate that changes in the architecture of a choice can give stronger social benefits and facilitate behavior change in retail environments compared with taxes and persuasive messages. This type of intervention can be market-led, without the need for additional formal regulation, and thus faster to implement than more traditional policy instruments. However, this choice architecture interacts negatively with the tax and the injunctive message, an indication that the design of interventions consisting of multiple stimuli requires careful planning. Nonetheless, removing from the consumers the task of identifying the low-, medium-, and high-carbon products appears to be a suitable approach to achieve significant reductions in carbon footprint at a large scale.

Also of interest to policy makers is the performance of the bonus-malus carbon tax (D’Haultfœuille, Givord, and Boutin 2014; Hilton et al. 2014), a revenue-neutral mechanism that does not currently have any real-life implementation in a retail environment. Like choice architecture, the tax introduces structural changes into the market, changing all prices to account for the damage they cause to the environment. This tax performs well by itself, and its impact is stable over time. Importantly, government could “market” this tax as a revenue-neutral tool to tackle climate change to try to gain public consensus for the tax measure. Governments might also try to work with retailers in designing this tax, perhaps by using retailers’ sales data or through the collaborative design of field experiments (Gneezy 2017), to learn about which goods to subsidize to best protect real incomes of particular segments of the population. Such collaboration may help offset potential resistance of retailers and industries most affected by the tax.

The results from the choice architecture intervention have implications for other public policies that can improve the sustainability of grocery consumption. For instance, growing attention is given to carbon labeling, as well as the simpler “traffic light” carbon footprint labeling (Muller, Lacroix, and Ruffieux 2019; Osman and Thornton 2019; Panzone et al. 2020). Labeling policies are designed to provide relevant information about all products to consumers and require consumers to search for products with the characteristic they want. The choice architecture presented in this study achieves the same objective as a traffic-light system, classifying products as high, medium, or low carbon. However, the search by carbon content is faster to execute, as it only requires entering a specific aisle. Future research should further investigate the impact of choice architecture on search costs, for instance, observing whether the same or better results could be obtained by replacing choice architecture with traffic light labels or whether additional mechanisms that reduce search costs (e.g., store filters) could create positive synergies with choice architecture.

Managerial Implications for Retail Shelf Management

While there is clear responsibility for consumers to make environmentally sensitive choices, retailers could operate by designing a shopping environment that simplifies the consumer task of being sustainable. The literature primarily understands variety in terms of availability: if low-carbon options are introduced in a choice set, consumers can find them and buy them (e.g., Demarque et al. 2015; Mazar and Zhong 2010; Muller, Lacroix, and Ruffieux 2019; Panzone et al. 2018; Perino et al. 2014; Van Herpen and Pieters 2002). Yet the organization of products has an impact on search costs: currently, online retailers supply thousands of products and require consumers to identify the low-carbon options on their own. The choice architecture presented in this article simplifies this task by providing carbon footprint information for the whole store. This simple change to the presentation of the choice set plays a major role in driving sustainable behavior change. Crucially, retailers need to be aware that this architecture can interact negatively with other stimuli, as discussed previously. Overall, while research on choice architecture in retail environments is limited (e.g., Diehl, Van Herpen, and Lamberton 2015; Lamberton and Diehl 2013; Sarantopoulos et al. 2019), it shows considerable potential for public policy.

In practical terms, the choice architecture presented in this article, which organizes products by carbon footprint first and by product category within a footprint aisle, was conceived for online retailing, where it can be implemented straightforwardly and at low cost. Online shopping is growing in size and importance, and online retailers can very flexibly offer a menu of possible aisles, organizing products by carbon footprint, nutrition, or any other topical environmental or social issue (e.g., plastic content) so that customers can tailor their shopping experience to their needs. The proportion of retail revenues generated online is still a smaller, though growing, proportion of overall retail revenues than in-store (Panzone, Larcom, and She 2021). The same choice architecture is more complex to introduce in brick-and-mortar retailers. However, physical stores could still implement the organization proposed in this store as a “virtual” carbon aisle, ordering products within each aisle by carbon footprint (e.g., classifying all vegetables into low-, medium-, and high-carbon items). This study does not allow us to generalize our findings to the physical environment, which consumers may use to satisfy different priorities, and this question is left for future research.

At the same time, the specific use of this choice architecture in online shops makes it appealing as a subject for research using machine learning, designing an infrastructure that can use the large amount of data that can be collected online (e.g., recording clickstreams to measure information search and attention) to facilitate the transition to low-carbon shopping. This article did not record some information, such as the amount of time spent reading the carbon footprint or the nutritional profile of a product (to capture attention), the keywords used to search for products (to capture search), or the items added/removed from a basket (to capture basket changes). The ability to expand the range of data collected could provide additional information on what motivates consumers in the marketplace, with the ultimate goal of driving large-scale changes in consumption that reduce the total carbon footprint from food and grocery shopping. This advance hinges on understanding the modeling and the data availability and requirements; this article is a first step in this direction and hopefully a useful contribution on this topic to the academic community.

Conclusions

The societal challenge of limiting the increase in mean global temperature requires ensuring that consumers reduce the GHG they emit from consumption. These reductions can be achieved effectively by intervening at the point where consumers make their choices. Overall, the results in this article indicate that structural changes in the layout of the store to facilitate the identification of low-carbon options (choice architecture) can lead to strong reductions in carbon footprint. The effect can be larger than that of a bonus-malus carbon tax, which, however, has a persistent effect on the carbon footprint of grocery baskets. Nudging in the environmental decision-making domain is somewhat understudied, but the results of this research suggest the potential for the design of online stores that motivate low-carbon choices. This research also shows the benefits of using an experimental online supermarket to test a wide range of policy interventions. This tool is particularly valuable for studying consumer behavior over short periods of time (i.e., a few weeks, as customer retention decreases faster over time) and lends itself to the design of local as well as nationwide experiments. The approach can provide a better understanding of what drives consumer behavior when making decisions with environmental implications, a key area of research to ensure a more sustainable future.

Supplemental Material

Supplemental Material, sj-pdf-1-ppo-10.1177_07439156211008898 - Sustainable by Design: Choice Architecture and the Carbon Footprint of Grocery Shopping

Supplemental Material, sj-pdf-1-ppo-10.1177_07439156211008898 for Sustainable by Design: Choice Architecture and the Carbon Footprint of Grocery Shopping by Luca A. Panzone, Alistair Ulph, Denis Hilton, Ilse Gortemaker and Ibrahim Adebisi Tajudeen in Journal of Public Policy & Marketing