Palatability of Interventions for Good Self-Nudging Complementarities in Food Environments Using an Experimental Online Supermarket

Prior to week 2, participants will be randomly assigned to one of three treatment arms:
The three treatment arms, applied in week 2, are (and described in chronological order):
1. Effect of MPL (Control): MPL (no action identity cue) – Control.
2. Effect of Cue on MPL and behaviour: Action-Identity Cue + MPL
3. Effect of Cue on behaviour: MPL + Action-Identity Cue.

Two treatment arms of our experiment design will include an “action-identity link” informational cue (Tonke, 2025). The other will not include any identity cue. In all three treatment arms, participants are asked to go give their (incentive-compatible) willingness to pay between two online grocery shop formats. One is the standard layout of searched grocery items, the other format has all goods searched for sorted by their eco scores (Clark, 2022), from most eco to least. Participants then will undertake a typical weekly shop on an experimental online grocery shopping platform, Woods (https://woodssupermarket.co.uk). For each participant the Woods ‘sort by’ layout will be allocated based on the willingness to pay choices they previously declared in the Multiple Price List (same way as Epperson and Gerster, 2024).

Prior to week 2, participants will be randomly assigned to one of three treatment arms:
The three treatment arms, applied in week 2, are (and described in chronological order):
Effect of MPL (Control): MPL (no action identity cue) – Control.
Effect of Cue on MPL and behaviour: Action-Identity Cue + MPL
Effect of Cue on behaviour: MPL + Action-Identity Cue.

Two treatment arms of our experiment design will include an “action-identity link” informational cue (Tonke, 2025). The other will not include any identity cue. In all three treatment arms, participants are asked to go give their (incentive-compatible) willingness to pay between two online grocery shop formats. One is the standard layout of searched grocery items, the other format has all goods searched for sorted by their eco scores (Clark, 2022), from most eco to least. Participants then will undertake a typical weekly shop on an experimental online grocery shopping platform, Woods (https://woodssupermarket.co.uk). For each participant the Woods ‘sort by’ layout will be allocated based on the willingness to pay choices they previously declared in the Multiple Price List (same way as Epperson and Gerster, 2024).

The identity-action link information cue “Last time, you purchased eco items — you’re one of our eco-friendly shoppers! Great, keep going and we will keep providing green options.” Is deigned to create a salient link between a participant’s past eco-purchasing behaviour and an eco-identity. The identity cue was developed based on responses from a pre-test survey with 50UK participants (recruited via Prolific). Participants rated six candidate statements across several dimensions, including perceived personal relevance, emotional resonance, motivational impact, clarity, and autonomy. The final cue was constructed from the two statements that received the highest overall ratings across these dimensions.

This cue is provided in two treatment arms. For one treatment arm it is provided before the multiple price list (MPL) eliciting the participant’s willingness to pay for standard and eco grocery online sorting (and also before a shopping task). For the other treatment arm with the cue, it is provided after the MPL but before a shopping task.

For the control, no action-identity link cue is provided at any point. Participants still have to complete the willingness to pay MPL and the shopping task.

Across all treatment arms, the same products, prices and dual eco-nutri labels are provided. In none of the treatment arms can the participant change the default allocated ‘sort by’ grocery ordering (whether it be standard or eco) once implemented.

These treatment arms lead to our main hypotheses:
Hypothesis 1:
H1_The action-identity link cue will increase willingness to pay for the ‘eco sort by’ ordering option for the Woods supermarket platform, relative to a combined baseline that merges the two experimental arms without the identity-action cue prior to the Multiple Price List (MPL) — specifically, Control and Treatment arm 3., defined above.
This hypothesis is tested using a cross-sectional linear regression with a pseudo unbalanced treatment allocation. The unbalanced allocation arises from combining two control conditions to cost-effectively compare against a slightly larger treatment group for this research question.

Hypothesis 2:
H2_a The action-identity link cue leads to a greater reduction in the average environmental impact (per 100g and per GBP spent) of shopping baskets in the intervention group, compared to the change in the control group (treatment arm 1) from baseline (week 1) to follow-up (week 2).
This hypothesis is tested using a (log)-linear difference-in-difference regression (the difference between week 1 pre- and week 2 post-intervention eco-values of shopping baskets for each participant) between control and treatment arm 3.

H2_b The impact of the action-identity link cue on change on basket eco scores depends on the ‘sort-by’ ranking of the shopping platform.
We use Conditional Average Treated Effect (CATE) analysis to explore H2_a further. We suspect the effect of the action cue is moderated by the shopping platform’s sort-by configuration. This also forms a robustness check on H2_a.
We have no theoretical priors.

Exploratory Analysis:
We will conduct additional exploratory analysis to investigate the following research questions:

Does the higher willingness to pay for eco layout identify who will be most affected by the eco sort by layout?
Does consumer search effort negatively predict the effect of the sort by layout?
Does participants’ green identity exhibit an inverted-U relationship with the positive effect of the identity cue?
Is green identity positively increased between week 1 and week 2, for those receiving the identity cue, compared to those that did not receive the identity cue.

Difference in difference empirical approach (Hypotheses 2_a and 2_b):
To evaluate the average treatment effect of our manipulations we follow (Panzone et al., 2021, 2024) by using a difference-in-difference (DID) estimator (Bertrand et al. 2004; Imbens and Wooldridge 2009; Wing et al. 2018). We remember that participants shop in two separate weeks, called tasks. Task 1 has no intervention, regardless of the participant. For Task 2, individuals are randomly assigned (through stratified randomisation, as previously described) to one of three experimental groups k = 1, 2 and 3. Here, k=1, MPL (no action identity cue), will be the reference basket, control. Within each shopping task t = week 1 and week 2 participants i will make their purchase decisions, choosing food items to fill basket Ci;t which will have a normalised eco and nutri value (per grammes of basket and GBP spent). The average treatment effect (ATE)[we will also explore CATE, as done by Epperson and Gerster, 2024] will then be estimated as the difference between the average changes in basket values observed from week 1 to 2, for treatment arms 1 to 2 (Effect of Cue on MPL and behaviour) and from 1 to 3 (Effect of Cue on behaviour) between individuals in different groups. And testing the complementarity will be a comparison between treatment arms 2 and 3.

E.g. Cue on behaviour: \emptyset_{k;2}=Ek;2-Ek;1]-[E1;2 - E1;1 k=3

Where \left[\bar{E_{k;t}}\right] is the average eco score value of individuals in experimental group k in task t.
For our analysis, we perform a log-linear panel regression for concerns of non-normality and heteroskedasticity of the dependent variable:
ln(E_{i;t})=\ \alpha_{1;i}\ +\ \alpha_{1;i}W_t\ +\ \Pi_{k;t}W_t\ +\ \varepsilon_{i;t}

\alpha_{1;i}: Individual’s fixed effects
W_t: Time dummy capturing the task number t.
\varepsilon_{i;t}: Error term.

To convert the DID interaction term \Pi_{k;t} back to our desired ATE estimate, we use the Puhani (2012) transformation:
\emptyset_{i;k;t}=\ exp\ (ln(E_{i;t}))\ -\ exp(E_{i;t}\ \ -\ \Pi_{k;t}W_t\ )
NOTES: We also note that Panzone et al. (2024) use a 10% significance level for judgements on significance due to inherent noisiness and overestimation of standard errors (Betrand et al., 2004) in DID. To minimise some of the individual-level heterogeneity in shopping choices, our study design incorporates a prescriptive shopping list.

Note on self-selection and randomisation:
With regards to the effect of the cue, as individuals are randomly allocated to the three treatments, we have no concern of self-selection, and consider that the three populations comparable with any preference heterogeneity in one treatment group are likely also present in the others.
For the allocation to the sort by (based on each participant’s willingness to pay, we remind the reader that we have a higher probability to select the extreme MPL choices (which are chosen to be beyond likely willingness to pay for a favoured sort by option over another), as done in Epperson and Gerster, 2024. This means we still have pseudo randomisation and we can use an Inverse Probability Weighting (IPW) to explore CATE analysis, distinguishing effects between those that were more or less interested in the eco sort-by within each treatment arm.

2025-04-24

2025-05-01

Primary Outcomes (end points): This study has two main outcome variables.

First primary variable: The willingness to pay from each participant to have the grocery item menu ‘sorted by’ the standard format (based on order of appearance in the background grocery dataset) or from most to least eco-friendly on an incentivised online platform supermarket (https://woodssupermarket.co.uk). The willingness to pay is elicited through a Multiple Price List following Epperson and Gerster (2024).

Second primary variable: Eco-score of the final shopping baskets from Week 1 (pre-intervention baseline) and Week 2 (post-intervention treatment) based on the final product choices made by all participants in both weeks (in a shopping budget of £25-50). Eco-scores are based on calculations by Clark et al. (2022). The final basket eco scores will be re-balanced per 100g of produce bought and per GBP spent.

First primary variable:
We measure changes in willingness to pay (WTP) for two different online grocery sorting options -‘Standard sort by’ and ‘Eco sort by’- after participants search for key grocery items. This is compared across different treatment arms. The WTP is incentivised and elicited through a multiple price list. One of the binary decisions in the MPL is randomly chosen and applied.

The core research question is whether an identity-action cue, which highlights that a participant’s past behaviours are consistent with those of an eco-minded consumer, can encourage more eco-conscious consumer behaviour. Specifically, we examine whether this cue increases participant’s WTP to re-order their shopping environment to prioritize eco-friendly options (i.e., self-nudging by switching to the ‘Eco sort by’ layout).

This hypothesis surrounding the effect of the action-identity cue draws on the dual self-belief model from Bénabou and Tirole (2006; 2011), where individuals seek consistency with their self-image or past actions.

While prior studies have shown that altering the default sort order of items (e.g., placing healthier or eco-friendly items at the top) can strongly influence consumer choice, our experiment tests whether consumers will actively choose to make such changes themselves, rather than being passively guided by defaults.

Second primary outcome variable:
In both weeks of the shopping task, we collect data on the products each participant selects in the experimental online supermarket. We use this data to compute a basket-level eco score, which reflects the average environmental impact of the chosen items, normalized per 100g and per GBP spent.

This outcome allows us to test whether selecting the ‘Eco sort by’ option leads to meaningfully more eco-friendly purchasing behaviour, as reflected by the grocery basket eco score.

Because of the design of our multiple price list (MPL) mechanism, we are able to observe basket eco scores under both sorting conditions—Eco and Standard—across participants with varying preferences (based on their relative MPL willingness to pay). This enables us to compare eco scores both by preference (e.g., participants with above-median willingness to pay for the eco sort) and against preference (e.g., when participants experience a sort condition they would not have chosen). This facilitates part of heterogeneity analysis and conditional average treatment effect measures.

Our experiment includes two secondary variables.

We capture Green Consumer Identity through a series of statements to which participants indicate their degree of agreement (following Guiidetti et al, 2023). The identity measures are given after the shopping tasks in week 1 and 2.

Our other secondary variable measured in both week 1 and 2, is search effort in the shopping task (based on time spent going through the shopping platform). Through the experimental online supermarket platform and the survey platform, we will get these two separate measures for each participant each week.

In addition to these two key secondary variables, we collect data on socio-demographics (age, income, weekly grocery expenditure, children education, etc). We also collect preference elicitation data for the cognitive reflection test, locus of control, self-control.

The Green Consumer Identity will offer deeper insight on the mechanism driving any effects from the identity-action cue. Specifically, whether treated participants have an increased green identity relative to non-treated participants.

The search effort variable (in week 1) should be a strong moderating factor for the effect of the ‘sort by’ ordering. Specifically, we expect consumers with a shorter search effort to be more affected by a ‘sort by’ ordering, in our case, with a more eco basket when allocated to the green ‘sort by’ ordering, regardless of it was desired or not.

All the other variables are for exploratory purposes. These will be used for mechanism and heterogeneity analysis.

Similar to (Panzone et al., 2021a, 2021b; Zizzo, 2010), we will employ a between and within-subject experimental design, where participants are recruited to shop twice on our simulated grocery store. Week 1 serves as a baseline, where all participants are asked to do a weekly shop on the platform. Participants will then be randomised into one of 3 conditions using a stratified randomisation procedure and will be invited to complete a second survey and shopping task in the following week. Importantly, our experimental design will be incentive-compatible and can be classified as a framed field experiment (List, 2011).
A UK adult (18+) representative (as defined by Prolific) population sample, and pre-screened by Prolific for diet (to avoid participants with strict dietary requirements, such as veganism, Halal or Kosher requirements), online shopper (to select those that interact with online grocery shopping at least a few times a year) and primary household shopper (to get the main grocery shoppers) is initially recruited through the Prolific recruitment platform. We will also track whether individuals have other dietaries (vegetarianism, food allergies and intolerances, whom we might exclude from our main analysis at a later date.

To allocate the participants to one of the 3 treatment arms in week 2, we use a stratified randomisation technique, based on age, gender, income and eco score of participants’ week 1 shopping task baskets.

The unit of randomisation is at the individual level.

No

N/A

We collect 2-week data (from 2 pre- and post-shop surveys and two shopping tasks) from c.900 participants (not recruitment number) for the whole experiment (spread over 3 treatments). This sample size is based on previous papers of a similar nature and power calculation simulations for linear and log-linear difference-in-difference (detail below) and for unbalanced cross-sectional regression analysis (where the treated arm needs 300 people). This does not covers likely non-completion of the survey and experiments, attrition across the two-week study, comprehension and attention check failures, and budgetary constraints.

The recruitment sample size collected for week 1 is 345 participants per treatment arm. This sample size is based on previous papers of a similar nature and power calculations. This accounts for non-completion and attrition in the range of c.15% across the two-week study.

In the absence of prior data, we simulate required sample sizes using Stata 18, targeting 80% power at a 5% significance level for detecting a standardized effect size (Cohen’s d) in the range of 0.15 to 0.20. For the difference-in-difference simulation, we also estimate a 0.8 correlation between a participant’s basket in week 1 and week 2 (based on Panzone et al., 2021). We also compare minimum sample size requirements for linear and log-linear difference-in-difference (DiD) regressions. For a Cohen’s d of 0.15, the simulations suggest a required sample of approximately 280 participants per treatment arm (for both linear and log-linear DiD specifications), whereas for a Cohen’s d of 0.2, only requires a sample size of 165 participants. For our cross-sectional linear regression analysis (hypothesis 1), we simulated the required sample sizes using Stata 18, targeting 80% power at a 5% significance level for detecting a standardized effect size (Cohen’s d) in the range of 0.20. By combining data results from the control and treatment arm 3 (in which the cue is delivered after the MPL), we achieve cost-effective power through an unbalanced allocation. Based on these simulations, we find two viable options: a) Maintaining a 2:1 control-to-treatment allocation, requiring approximately 300 participants in each group (900 total); b) Holding constant the power calculations sample size for our DiD analysis (c.265 participants). This leads to a required 310 participants in the treatment arm of study for sufficient power. While the 2:1 allocation offers more robust power—particularly under potential treatment effect heterogeneity—it comes at higher cost. The second option, while slightly less conservative, provides adequate power and optimizes resource use by leveraging existing recruitment targets. Considering the other experimental risks, e.g. attrition and other screen-outs, we keep the upper bound of the sample size, 345 participants per treatment arm (with an additional 10 participants for the cue then MPL treatment arm, as a buffer for the most power sensitive treatment arm) to ensure sufficient power to detect a small effect size. In total, we will recruit 1,045 people in week 1.