Data types and sources

Volume sales data was sourced from Euromonitor and accessed through the Oxford University Library. The top 10 UK food and soft drink manufacturers and their brands were identified based on global company names using 2018 sales data from Euromonitor [5]. A company is defined by Euromonitor as: “the legal entity that produces or distributes an individual or group of brands in the UK”. All of the brands manufactured by these companies between 2015 and 2018 were identified, including those that dropped in or out of the market. Brands were defined as a set of products that have the same generic name and are manufactured by one company.

The composition data were provided by Edge by Ascential (previously Brand View), a private analytics company that collects product information, including nutrient composition data, by scraping the websites of the UK’s three leading retailers: Asda, Sainsbury’s and Tesco. These data were scraped from these three websites on the same date (13^th December) for four consecutive years (2015, 2016, 2017 and 2018). The sales data and nutrition composition data were automatically matched in Python based on three identifier variables that were present in both databases: brand name, category and year. A 10% random sample of brands was checked manually for any errors. Of the 20 brands checked, 4 brands were identified as pairing with the correct brand name but incorrect category. All 4 of these errors were brands that appeared in more than one category (e.g. Cadbury is present in five categories, including baked goods and confectionery). The matching code was adjusted so that it first paired based on matching categories, and then brand names, and no errors were identified after further checks.

Applying the FSA/Ofcom nutrient profile model

The FSA/Ofcom nutrient profile model was applied to the individual product composition data. The appropriate points were awarded based on each product’s energy, saturated fat, total sugar and sodium content (“A-points”) and fibre, protein and fruit, nut and vegetable (FNV) content (“C-points”) per 100g, as set out by FSA/Ofcom’s technical guidance [9]. This system was developed for the purposes of restricting advertising of food to children, but here we have used it to classify products as healthy and unhealthy. A food is classified as ‘less healthy’ if it scores four points or more. A drink is classified as ‘less healthy’ if it scores 1 point or more. For the purpose of comparing companies’ entire product portfolios, we converted the nutrient profile score to a 1–100 scale (-2(original score) +70), so that a higher score indicates healthier products. In order to directly compare drink scores with food scores, we also applied a linear adjustment to the distribution of the soft drinks scores (11x – 704, where x is the score for drinks on the 1–100 scale). The linear adjustment was selected so that the 33^rd percentile and 66^th percentile of both foods and drinks received the same score (44 and 66, respectively). After the scale conversion and linear adjustment, the thresholds for products to be considered healthy according to the FSA/Ofcom nutrient profile model were 62 or more for foods and 66 or more for drinks.

If the nutrient content for a product was missing, then data was imputed by calculating a brand average for foods in the same category, and if this was not possible, an overall category average. FNV content was estimated based on the ingredients list to categorise ingredients into ‘fruit’, ‘nut’, ‘vegetable’ and ‘other’. The percentage composition of ingredients was identified if this information was provided in the ingredients list. For the products where percentage of ingredients were not given, values were imputed based on a brand and category average, or if this was not possible, a category average.

Variables calculated

The total value (£ millions) and volume of food and soft drinks (tonnes) and the sales weighted mean nutrient profiling score (referred to in figure labels as sales-weighted score) were calculated in R for each company and brand, both overall and by category. When one brand had multiple product variants, a simple mean was used. While all brands were included in the analysis, only the top five for each company (n = 50) were presented for the brand-level analyses (Fig 3) for clarity. Bubble and chewing gum and milk formulas for infants, toddlers and children were excluded.

Statistical analysis

Chi-squared tests were performed in R to test if there were any significant changes in the number of brands and products each company manufactured over time (2015–2018). ANOVA tests were used to test for differences over time in the nutrient profiling scores overall and for each company, category and brand.

Strengths and limitations

By pairing composition data with sales data and applying a nutrient profile model, both the relative healthiness of individual foods and drinks available, and the relative healthiness of what is sold have been assessed, and how this has changed over time. This gives an idea of how companies are responding to voluntary reformulation targets to improve the nutritional quality of their products overall, rather than in relation to a single nutrient.

Only 10 companies, based on global company name, were included in the analysis, which represented 24% of total value sales in the UK in 2018 [5]. These companies were selected based on their value sales, although they are also the top 10 companies in terms of volume sales. By selecting companies based on their global, rather than national, names, UK retailers were excluded from the analysis. This is a major limitation given that own-label brands from the top 3 UK retailers (Tesco, Sainsbury’s and Asda) represented a total market share of 21% in 2018 [5]. While this study sets out a useful and important method for ranking companies in terms of healthiness of product portfolios, future studies should include retailers and a wider range of companies. This would give a more comprehensive picture of how food and drink companies and retailers in the UK are changing their products to meet public health targets. There are a number of data-driven limitations. The first is in relation to missing and imputed data. The values for seven nutrients (energy, saturated fat, total sugars, sodium, fibre, protein and FNV content) are needed to calculate the FSA/Ofcom nutrient profile score of a product. 32% of the 13,371 products included in this study had missing values for fibre, and 67% products had insufficient ingredients information and composition detail to be able to calculate %FNV accurately. There was no difference in the proportion of missing values over time. Missing values were imputed with either a category and/or brand average. The high proportion of missing/imputed fibre and FNV was to be expected as the labelling of fibre on foods is not mandatory (unlike other macronutrients) [12] and the percentages for individual ingredients (i.e. FNV ingredients) only have to be stated when the product title includes an ingredient name, or when a claim about the amount of an ingredient has been made on the label [13].

To test what impact the imputed fibre data had on the results, a sensitivity analysis was conducted. 31% (n = 4186) of all products in the original dataset had imputed fibre values, and these were evenly distributed across the four years. For our sensitivity analysis, we adjusted the fibre content for these products to 0.0g/100g, with the FSA/Ofcom points awarded for fibre also then given 0, the lowest score possible. The number of products that were classified as healthy fell from 47% to 46% in 2018, and there were negligible changes in the total sales-weighted nutrient profiling score for 2018, which fell from 51 to 50.

For fruit, nut and vegetable (FNV) content, 8896 (67%) of included products had imputed values, although three-quarters of these (n = 6624) fell into categories that you would not expect to contain enough FNV to score one point: baked goods, confectionery, dairy, ice cream, savoury snacks, soft drinks, spreads and staples. To test what impact the imputed FNV data may have had on the results, the remaining 2272 products (baby food, breakfast cereals, ready meals, and sauces, dressings and condiments) had their %FNV adjusted to 0%. After this adjustment, 25% (n = 564) of the 2272 products saw a change in their final Ofcom score. The overall proportion of products classified as healthy in 2018 fell from 47% to 46%. The results were the same as those found with the fibre sensitivity analysis, with a similar group of products being affected by the lack of fibre and FNV values. These results suggest that while the missing fibre and FNV values is a weakness in the dataset, the interpretation of the data was unchanged, and it has not affected the overall results.

Data restrictions meant that time period covered changes between 2015 and 2018. Previous reformulation efforts made before 2015, for example as part of the salt reduction programme that began in 2006, will have been excluded. Using a wider historic time period may show that some companies who started reformulation efforts promptly have made more signficant changes than recorded here. Applying this method to datasets in multiple countries may offer insight into how companies are responding in countries with varying public health nutrition policies, for example voluntary reformulation targets in the UK compared to taxes on energy dense foods in Mexico [14] and mandatory warning labels in Chile [15].

The FSA/Ofcom nutrient profile model was used because it is designed for and used in the UK market and has been widely validated in terms of how its use may impact on dietary choices [16]. However, its original purpose was for the assessment of whether or not a product should be advertised to children, rather than to assess the nutritonal quality of a company’s product portfolio and classifying products as healthy and unhealthy, as it was used here. It would be possible to conduct similar analyses using other nutrient profiling models such as Health Star Rating [17] and Nutri Score [18], though since all rely on changes in the underlying nutrient composition differences between scoring systems are likley to be modest.

We combined the distributions of food and drink products by using a linear transformation that matched the distributions at two points–the 33^rd and 66^th percentile. The selection of the two matching points was arbitrary. Matching at different points (e.g. the 25^th and 75^th percentiles) would have produced a different linear transformation and hence different scores for drinks. This is an inevitable limitation associated with combining scores for companies with both food and drink profiles.

Implications of research

This study shines a spotlight on the very small changes over time in the nutritional quality of food and drink products from the UKs largest food and beverage companies. While the proportion of volume sales increased from 44% to 53% over time, this change was entirely down to increased volume sales of bottled water, low/no calorie drinks and high-scoring fruit juices. The brands that saw the biggest changes to their scores over time were soft drinks. Once soft drinks were removed, the total volume sales of foods classified as healthy dropped to just 6% in 2018, down from 7% in 2015. This strongly suggests that PHE’s reformulation targets for sugar, salt and calories have not had a substantive impact on the nutritional quality of foods.

This method of ranking food and drink companies based on the nutritional quality of their product portfolios could be used to benchmark companies as a tool for ‘healthier’ impact investment. There is an increasing interest by investment banks and other financial organisations to assess what impact food companies are having on public health and how responsible their business practices are (known as impact investment) [25]. This has already been done in part by ATNI in collaboration with Shared Action [26] and INFORMAS [22].

Transparent monitoring of this kind also allows for greater consumer understanding of the work that is, or is not, being undertaken by companies. There is some evidence that pressure from the social environment is a factor influencing corporate behaviour [27], and public benchmarking exercises may increase pressure on companies to make meaningful change.