Design and setting
This was a prospectively registered 2 × 2 factorial randomised controlled trial conducted in a custom-made simulated online supermarket platform (www.woodssupermarket.co.uk) developed by Cauldron, UK (http://cauldron.sc/clients#woods). The supermarket was developed to emulate a real online supermarket website as previously described [12]. It contains a food database with ~ 11,000 products, downloaded from a real UK grocery retailer (Tesco.com API, February 2012), which includes standard UK branded products. Nutrient composition information per 100 g was supplemented by manual linkages with food labels at online supermarket websites and with data provided by Kantar WorldPanel and the Medical Research Council Human Nutrition Research food and nutrient database [13]. Data were collected and managed using the supermarket platform and the REDCap (Research Electronic Data Capture) electronic data capture tools hosted at the University of Oxford [14]. REDCap is a secure, web-based application designed to support data capture for research studies, providing: 1) an intuitive interface for validated data entry; 2) audit trails for tracking data manipulation and export procedures; 3) automated export procedures for seamless data downloads to common statistical packages; and 4) procedures for importing data from external sources. The protocol was implemented without changes except minor prospective revisions, made prior to analysis which are noted in the analysis plan (Additional file 1).
Participants
Participants were recruited between March and July 2018 through an online research agency [15]. Invitations were sent to a random subsample of a pool of 6968 who had all been pre-screened as eligible. Due to a technical limitation, the invitation and response rates were not recorded. Participants were eligible if they lived in the UK, were aged 18 years or over, were the main (or shared) grocery shopper for their household, were able to read English, had access to a computer and Internet connection (by virtue of being part of the Prolific participant pool), and were willing and able to provide informed consent. People were not eligible if they were following any restricted diet such as a vegetarian, vegan, dairy-free, sugar-free, or gluten-free diet. Following online screening for these criteria, participants provided consent electronically. Following consent, they answered standard demographic questions as well as a few additional questions on shopping habits and health status at baseline (Additional file 1: Appendix B).
Randomization
The statistician generated the randomization sequence using the R package ‘blockrand’ [16] and the lead researcher uploaded the sequence in REDCap. Following the baseline questionnaire, participants were allocated to trial groups by REDCap via computerised random number generation on a 1:1:1:1 basis with random block sizes. Allocation concealment was achieved, as participants were recruited from Prolific independently of the research team and automatically randomised without human involvement.
Shopping task
Following randomization, participants were redirected to the supermarket website that introduced the shopping task. The website explained how to complete the task. As with real online supermarkets, participants could find items by browsing the supermarket departments and shelves or using a search function. They were asked to select 10 ‘everyday’ foods from a pre-specified shopping list. They were instructed to imagine they were doing their own grocery shopping and to choose foods that they and their household would want to eat. The 10 foods were major sources of SFA in the UK, within food categories where lower SFA options are also available. Participants were not prevented from selecting as many items as they wished, in unlimited quantities, but the instructions requested selection of only a single item per category from the shopping list below. The list comprised:
-
Milk for everyday use
-
Butter or margarine for everyday use
-
Cheese for use in sandwich or light meal
-
Ready-to-eat savoury entree item (e.g. cured meats, samosas)
-
Ready-to-eat individual chilled dessert
-
Meat/fish/vegetarian alternative to cook for 4 people
-
Dessert for a meal of 4 people
-
Something to eat with a hot drink
-
A sweet snack item to eat now
-
A savoury snack item to eat now
In the consent process, all participants were informed in one sentence that: “This study aims to investigate if two different ways of making healthier choices when shopping online are acceptable to shoppers and effective in reducing the saturated fat in the foods in their basket.” However, this sentence was among many in the participant information pages and no further reference was made to SFA afterwards except in the swap condition as a necessary part of that intervention (see below). Following completion of this task, participants were redirected to REDCap to complete a survey assessing the intervention acceptability, their usual shopping behaviour, and an open-ended box for comments. Upon completion, participants were reimbursed with £5 for their participation.
Interventions
Participants were randomly allocated to one of the following groups:
Swaps (individual-level intervention)
Participants were offered an explicit swap with less SFA. Swaps were offered if an alternative product within the same category had at least 2 percentage points less SFA (i.e. 2 g less SFA per 100 g of product), was between 60 and 140% of the weight of the original product, and was between 0 and 200% of the price of the original product. Swaps were offered at the point of selection, immediately after selecting an item to be added to their shopping basket. If multiple swaps were available, the one of the same brand as the base product was offered and if no products of the same brand were available a swap that met the criteria was randomly offered. An example swap is shown in Fig. 1a. Before commencing the task, participants in this group were advised that they might be offered a swap with lower SFA. They were advised to choose the food with lower SFA only if they would choose this food if offered in their normal shop, choose it if they and their family would eat it gladly, or are prepared to eat it to lower their SFA intake. They were advised not to choose the product if they wouldn’t be willing to eat it. If they refused the swap, no further swaps for that food were offered.
Altering the default order (environmental-level intervention)
When searching or browsing foods, participants viewed a list of products in ascending order of SFA content (i.e. the products with the lowest SFA content appeared at the top of the screen) but this order was not made explicit to participants. Moreover, the SFA content of the food was not displayed in the product list but, in common with all UK online supermarkets, the SFA content along with other nutrients from the nutrient facts panel was shown if the participant clicked on a product in search of more information. The SFA order was applied to each list of foods offered to participants when searching for products. An example of the intervention is shown in Fig. 1b.
Combination of individual- and environmental-level interventions
Participants allocated to this arm received both interventions as described above. Participants were exposed to the environmental-level intervention while searching for items and the individual-level intervention after selecting specific items. Accordingly, the environmental-level intervention was viewed before the individual-level one.
Control
Participants in this arm shopped using the default version of the website with a random order of the foods displayed in response to searching or browsing with no swaps offered.
Blinding
Investigators were not blinded to intervention allocation, but they were not able to manipulate any study parameter following the initial study set up, as all study procedures took place in the online platform. The outcome assessment was blinded, as it happened automatically in the online platform. The statistician was blinded to intervention allocation. Participants were necessarily unblinded and were aware of the study aims.
Primary outcome
The primary outcome was the difference in the SFA content of the basket measured as the difference in the percentage of total energy between each of the four trial groups. We adjusted for total energy, because it places the focus on the nutritional composition of the foods selected and not the total amount of food purchased, as it would be the case if we used an absolute amount of saturated fat. Furthermore, it makes the outcome comparable to the nutritional recommendations for saturated fat, which are expressed as percentage of energy.
Secondary outcomes
We examined the differences between the baskets in each group:
-
i.
the proportion of products with less than 1.5 g of SFA per 100 g of product (1.5% SFA) [17]
-
ii.
cost of the basket, expressed as £/100 g
-
iii.
total energy (kcal), energy density (kcal/100 g), sugar (percentage of total energy), and salt (g/100 g).
We also examined the differences between offering swaps alone and the combined intervention arm in:
-
i.
the percentage of energy from SFA per swap accepted
-
ii.
the proportion of swaps accepted out of those offered
-
iii.
the proportion of swaps accepted out of those offered by median observed change in SFA (high SFA change vs low SFA change)
-
iv.
the proportion of swaps accepted out of those offered for (a) butter, margarine, and spreads, (b) cheese, (c) milk, (d) meat and meat products, and (e) sweets and desserts, including chocolates, sweets, ice cream, cakes, pies, biscuits, and sweet items from the bakery
-
v.
the proportion of accepted swaps out of the number of items selected.
Sample size
The relationship between SFA intake and cardiovascular outcomes is linear and so at a population-level even very small reductions in SFA intake will be of public health significance. We made a pragmatic decision to power this study based on a 2% reduction in energy from SFA. This magnitude of reduction is estimated to be associated with an 11% lower risk of cardiovascular disease mortality [18]. Assuming 7% standard deviation in the total basket between any of the 4 groups and using intention to treat analyses with 90% power and two-sided α = 0.05, we required 258 participants per group (total n = 1032). A final sample of 1240 participants would account for 20% attrition of participants not completing the shopping task.
Statistical analysis
We followed a pre-specified statistical plan (Additional file 1) published in advance of the analysis in the ISRCTN registry (ISRCTN13729526). An independent trial statistician analysed the primary outcome and secondary outcomes i-iii using two-way analysis of variance. As the comparisons had been pre-specified, we did not correct for multiple testing [19]. We also tested for interaction between the two interventions and the main outcome by introducing an interaction term in the regression model [20]. Secondary outcomes iv-viii are presented as medians with interquartile range (IQR). We analysed data from participants who bought at least one product from at least 5 out of 10 categories of the shopping list and, when participants bought more than the 10 items requested, we included all items bought. We performed pre-specified subgroup analyses by sex, age (below or above the median), ethnic group (white vs non-white), obesity, education (none/secondary vs higher), and household income (low/middle vs higher). Estimates of comparative effectiveness for all outcomes are reported as mean differences with 95% confidence intervals (CI). Two researchers analysed the open-ended comments using manifest content analysis counting the frequency and grouping specific content evident in the comments [21]. All statistical analyses were conducted in R (version 3.5.0, Vienna, Austria).