Real-world nudging, pricing, and mobile physical activity coaching was insufficient to improve lifestyle behaviours and cardiometabolic health: the Supreme Nudge parallel cluster-randomised controlled supermarket trial

We conducted a parallel cluster-randomised controlled trial (RCT)—the Supreme Nudge trial—in 12 Dutch supermarkets from one chain in socially disadvantaged neighbourhoods. Supermarkets were randomised to a control arm (n = 6) or intervention arm receiving nudging and pricing strategies to promote healthy food and beverage purchases (n = 6). Participants were additionally randomised to a mobile physical activity coaching app intervention (‘SNapp’), receiving either a mobile coaching app and a step counter app (intervention arm) or only the step counter app (control arm). The duration of intervention implementation was six (n = 4 supermarket clusters) to 12 months (n = 8) depending on supermarket enrolment date.

More detailed information on the trial design is published in a protocol paper [37], including a correction paper detailing protocol changes due to the COVID-19 pandemic [38]. We also briefly summarise these protocol changes further down. The current paper focusses mainly on the supermarket-related outcomes. In a separate paper, we will describe and evaluate the mobile physical activity coaching app intervention in more detail. This trial was prospectively registered in the Dutch Trial Register (ID NL7064) on 30 May 2018. The reporting this paper presents is in accordance with the 2010 Consort statement for cluster-randomised trials [39].

Eligible supermarkets were regular stores (i.e. no convenience store) located in socially disadvantaged neighbourhoods (below average postal code SEP-scores according to The Netherlands Institute for Social Research) [40] and had implemented an adapted cash-register system to allow for pricing strategies. Of 69 eligible supermarkets, we selected 12 balancing the lowest number of competing supermarkets within a 2-km radius and the highest number of potentially eligible residents living in the neighbourhood. There were no systematic differences between supermarkets at baseline.

Eligible individuals were aged 30–80 years, regular shoppers at a participating supermarket (self-reported to purchase > 50% weekly groceries at a participating supermarket), planning on continuing shopping there for the study period, and could communicate in Dutch. Participants had to provide written informed constant prior to study enrolment. For additional individual-level randomisation to the mobile coaching app, participants had to use a smartphone with a mobile data plan and at least Android 8 or iOS 13, have experience with mobile text messaging, and be able to climb a flight of stairs.

We used a mix of passive and active recruitment strategies prior to the intervention implementation, as described elsewhere [41]. Passive strategies entailed, for example, online news articles in local newspapers, flyers and invitation letters sent to every household in the supermarkets’ neighbourhood, flyers in supermarkets, and e-mail invitations to supermarket customer panels. Active recruitment strategies included word-of-mouth promotion among included participants and in-store recruitment by research staff.

Measurements took place at baseline (T0), after 3 months (T1), after 6 months (T2), and after 12 months (T3). The primary outcome was the change in total diet quality (Dutch Healthy Diet 2015 index) from baseline until 12 months [47], measured with a validated short 40-item food frequency questionnaire (FFQ). The short FFQ asked about the average dietary intake during the past month from 15 components representing the food-based Dutch dietary guidelines of 2015. For each component, a score between 0 and 10 was assigned, resulting in a total score ranging from 0 (low diet quality) to 150 (high diet quality) [48, 49]. Secondary outcomes were as follows: the total percentage of healthier food purchasing (calculated based on grams purchased derived from weekly data collected via loyalty cards), daily step counts (step counter app) [46], cardiometabolic risk markers self-assessed by participants at home (glycated haemoglobin (HbA1c), low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, total cholesterol (TC), TC/HDL-ratio, triglycerides (TG) collected via a finger prick measurement, and waist circumference via a measuring tape, presented and analysed stratified by sex) [37], and questionnaire items relating to overall customer satisfaction, food-decision styles in relation to vegetables and to snacks (reflective, habitual, and impulsive) [50‐54], social cognitive factors related to nudges (heath goals, healthy shopping, perceived social norm, and attractiveness of healthy foods) [55‐60], and acceptance and awareness of nudges after six or 12 months. The purchase data were collected weekly over the complete duration of the intervention and recalculated to an average purchase per blocks of 4 weeks, leading to 13 time points. Step count data was collected daily (Additional file 1: Table S3 [37, 38]). As explorative outcomes, we described medication use, diet quality scores of various food groups, and percentage of healthy purchasing within various food groups per time point. Further details on data collection methods are described elsewhere [37, 41].

At the supermarket-level, we explored whether nudging and pricing strategies changed supermarket-level sales trends to account for potential selective trial participation [61]. We used sales data of the intervention period plus 6 months of pre-intervention data. Sales data comprised the number of products and Euros spent on healthy and unhealthy products per supermarket per week, based on 63 product groups which were pre-defined by the supermarket chain. We primarily investigated changes in the total percentage of healthy product sales of all sales in intervention supermarkets compared to control supermarkets during the intervention period, while taking into account the pre-intervention sales trends. We assigned the 63 product groups into 10 overarching food groups and examined the percentage change in sales from: fruits, vegetables, legumes, and nuts, grain products, milk and yogurt products, cheese, meat products, meat substitutes and eggs, fish, oils, fats and herbs and spices, non-alcoholic beverages, products from all remaining food products (e.g. pre-packaged meals, and baking products), and sweet and savoury snacks of total sales. Lastly, we examined changes in total sales revenue (Euros).

We monitored the implementation fidelity of the nudging and pricing strategies monthly to bi-monthly as the trial progressed. Research staff visited intervention supermarkets and quantitatively scored intervention components separately by their implementation status (none/minimal implementation scored 1, to (almost) optimal implementation scored 5). A summary score of all components was calculated for each supermarket by time point.

The trial required 141 participants in each supermarket arm to detect an average change of 5 points in the diet quality score by the end of the intervention period, assuming a 14.1 point standard deviation (SD) of the average change (power = 80%; two-sided alpha = 0.05). The expected 5-point change was based on results from a virtual supermarket experiment which indicated a 4% increase in healthier purchasing following a combination of nudging and salient pricing strategies [22]. We did not use a design factor since baseline data of diet quality among participants from the first eight enrolled supermarkets showed no correlation for observations between study sites (intra-cluster correlation coefficient (ICC) = 0.00) [37]. To allow for 25% drop out, 176 participants needed to be recruited per arm, resulting in 352 participants.

Twelve supermarkets were block-randomised by the first author to the control arm or the intervention arm, using a web-based random-number generator tool with blocks of four supermarkets. Individual-level randomisation in the mobile physical activity coaching app intervention or only the step counter app was conducted via a cloud-based data-management software application (Castor Electronic Data Capture). We used variable block randomisation alternating block sizes of two, four, and six, with equal weight for both groups. A research assistant conducted the randomisation of participants eligible for the coaching app intervention.

Blinding of participants was not feasible due to the nature of the interventions. Participants were not actively informed about the content of the supermarket interventions and not about the intervention allocation of their supermarket location, not prior to their study enrolment nor during the trial. Those randomised in the coaching app intervention were actively informed since they were requested to install the app(s).

Due to the COVID-19 pandemic and its associated governmental restrictions (e.g. social distancing measures), the main protocol underwent some notable changes [38]. As we were unable to perform physical measurements face-to-face, we abandoned planned blood pressure measurements. Instead, waist circumference and blood markers including HbA1c and lipid profile were measured remotely via a self-assessed finger prick. Furthermore, the lack of active recruitment opportunities required a revision of recruitment goals. Due to a lower expected response rate, we recruited from twelve rather than eight supermarket locations. Due to the fixed project end-date, the additional four stores could only be followed-up for six instead of 12 months. The expected lower response rates also forced us to combine intervention arms, leading us to comparing a combined nudging and pricing arm to a control arm rather than also investigating the independent effect of nudging strategies. Finally, in response to the expected lower response rates and participant burden of self-assessment of cardiometabolic risk markers, we revised our primary outcome to only include changes in diet quality, as was already part of the primary outcome, and to exclude systolic blood pressure, LDL cholesterol, and HbA1c from the primary outcome. These changes required us to recruit 352 participants instead of the originally planned 1485.

Population characteristics described were stratified by trial arm to visually inspect potential baseline differences as mean (SD) for normally distributed continuous variables, median (IQR) for skewed continuous variables, or frequencies (percentage) for categorical variables.

For evaluation of the primary outcome, we used a linear mixed model with group allocation as independent variable and the diet quality score at T1, T2, and T3 as dependent variable. The model included the covariates diet quality at T0, to account for regression to the mean, and time as categorical variable, considering the varying number of missing values per time point. Interaction between time and treatment indicated no time-specific intervention effects so this interaction was left out of the model. All models included random intercepts at the participant-level and the supermarket-level. The addition of random slopes did not improve the model fit and were therefore not included. Residuals were confirmed to be normally distributed via residual plots. Missing data were not imputed since mixed models use the maximum available data, even if participants had incomplete data. We reported the average within- and between group change in diet quality over all follow-up time points with a 95% confidence interval (CI) for the intervention arm, compared to the control arm and compared to the baseline measurement. We also reported the effect sizes on total diet quality at T1 and T2.

Changes in all secondary outcomes with the exception of step counts were analysed as described above for the primary outcome. Analysis of step count data followed a similar approach, but with only a random intercept at the participant-level, since their appeared no clustering of step count data between study sites, and including adjustment for smartphone operating system and sensor type used for step count tracking. Variables related to acceptance of nudges were only measured at T3 and therefore presented descriptively per arm by their mean (SD) or frequency (%).

Participants were analysed according to their initial treatment assignment (intention-to-treat), and we included participants who had a baseline measurement and at least one follow-up measurement of diet quality; this was considered our analytical sample. For secondary outcomes, we used an analytical sample per outcome, including those participants with at least a baseline measurement and one follow-up measurement for that outcome, leading to varying sample sizes by outcome.

For each explorative supermarket-level outcome, we described the average (mean (SD)) pre-intervention and post-intervention sales for both the intervention and the control supermarkets. We conducted controlled interrupted time series analyses based on a linear mixed model [62, 63]. For each outcome, we tested for changes in the average sales trends over time in intervention supermarkets compared to the control supermarkets, treating the pre-intervention period sales trends as fixed effects and including a random intercept for supermarkets. All analyses excluded the first week of the intervention implementation to allow time for the intervention to be implemented. We modelled changes in sales outcomes as an immediate step change, which assumed an immediate and stable effect of the intervention. Moreover, the average implementation fidelity scores per supermarket per time point were graphically presented, and the median score (range) over time is presented for all intervention stores combined. As a sensitivity analysis, the intervention supermarket with the lowest implementation fidelity score over time was excluded from the analyses of the total percentage of healthy product sales of all sales.

We did not account for multiple testing since our primary, secondary, and explorative analyses all reflect different outcome constructs and were planned in our study protocol. We used a two-sided alpha of 0.05 to define statistical significance, with the exception of the interaction tests for which we used an alpha of 0.10. Analyses were conducted using R statistical software (version 4.2.1), except for the step count data (SPSS version 27).

Recruitment ran from mid-January 2021 to November 2021. In total, 783 individuals registered for participation, of whom 602 were eligible and 421 signed informed consent (Fig. 1). Overall, 199 participants in the control arm and 162 in the intervention arm completed the baseline measurement and at least one follow-up measurement of the primary outcome, thereby reaching our target sample size of 360 participants [37]. Across trial arms, 73% of participants were female, 42% were highly educated and the average age was 58 (SD 11) years. The population characteristics were equally balanced between arms (Table 2) and across clusters [41], with the potential exception of prevalence of hyperlipidaemia.

In addition, 233 of the participants also completed the baseline measurement and at least one follow-up measurement of the mobile coaching app intervention, of whom 114 were randomised to the mobile coaching app plus step counter app (intervention) and 119 only to the step counter app (control). Participant characteristics were equally balanced between groups and comparable to the total study population (Additional file 1: Table S4).

Average pre-intervention sales of the total percentage of healthy products were 27.4% (SD 2.5) in control supermarkets and 27.5% (SD 2.1) in intervention supermarkets (Additional file 1: Table S9). Compared to control supermarkets and to pre-intervention sales trends, nudging and pricing strategies did not change the total percentage of healthy products sales (β − 0.0 (95% CI − 0.0 to 0.0)) nor the percentage of healthy products sales within various food groups (Table 4). Nudging and pricing strategies also did not significantly change retailer revenue (β − 76.9 Euros (95% CI − 274.6 to 120.4)). Results were comparable when excluding the intervention supermarket with the lowest implementation fidelity score over time (β_{total percentage of healthy sales} 0.0 (95% CI − 0.0 to 0.1)).

The median implementation fidelity score for all supermarkets over the intervention period was 3.6 (range 2.9–4.1) which translates to an implementation fidelity of 72%. Mean scores for individual supermarkets by time point are presented in Additional file 2: Figure S5.

Our results demonstrate that co-creation of interventional strategies with a retailer led to a predominant focus on the promotion of healthy products in supermarkets, which appeared not to be sufficient to shift purchasing patterns and dietary intake towards healthier patterns. Nevertheless, nudging and pricing strategies can be part of a comprehensive set of ambitious policy measures, while simultaneously focusing on discouraging unhealthier purchasing (e.g. restricting marketing and promotions of unhealthy products, banning of unhealthy products at checkouts, and restricting the availability of unhealthy products) [83, 84]. Such a combination of policy measures would create a level playing field among food retailers which is crucial in overcoming commercial barriers that impede impactful changes in supermarket environments. Furthermore, combined policy measures can lead to a small shift in population diet quality which can have a positive impact on population health outcomes [11]. Implementation fidelity and its effects on purchasing patterns needs to be monitored, and measures should be reformulated based on emerging insights.

There is a need for further studies focusing on the healthiness of purchase patterns or dietary intake patterns as a whole, rather than on single products or product groups, to better estimate the intervention implications for public health. Researchers may consider using more flexible study designs for real-world contexts that aim to generate public health benefits; various types of intervention components can be implemented simultaneously and innovated over a trial period based on real-time evaluations of their effectiveness [85]. Such an adaptive design could increase the chances of developing the most effective intervention(s), with the highest likelihood of achieving sustainable effects.