Data and Study Design
This observational field study utilises data from 142 worksite cafeterias run by a single commercial partner in the UK. Data was available from cafeteria sites the commercial partner runs across Great Britain. The worksites were located in England, Wales, and Scotland. Although the legislation did not make labelling mandatory in Scotland or Wales, the catering provider implemented calorie labels uniformly across their cafeteria sites. The catering provider operates cafeterias across a variety of worksites, with approximately 40% of these being manufacturing or distribution centres, ~40% at administrative or managerial offices, and the remaining ~20% classified as a mix of these two. Each cafeteria site has a base menu provided by the central office, from which they can select the foods which will be served in their specific cafeteria, meaning that while most cafeterias will offer the same basic items (e.g. burgers), they are able to choose which burgers they order and offer. This base menu changes every 12 weeks, and cafeterias run menus on a 4-week cycle (i.e. every four weeks the cafeteria cycles back through the same food items). Cafeterias select the foods they will serve during the 12-week time period from the base menu (but do not typically serve all items from this menu), and repeat this in three 4-week menu cycles.
The primary analysis covers a 3-month period before and 6-month period after the effective date of the new law to compare purchasing activity before and after the implementation, covering January to October 2022. Due to logistical constraints and the legal requirement to universally implement calorie labelling in the UK as of 6 April, it was not possible to randomise study sites or control the rollout of the implementation. This means all worksite cafeterias in the study sample introduced calorie labels at the same time and there could be no control group, given the legal requirement to implement labels.
We analysed data six weeks and six months following the implementation of the policy to assess any changes over time. Additionally, a post-hoc analysis used data from one year before labels were implemented to six months after, to provide more information on seasonal changes.
Data sources
Sales data from worksite cafeterias were obtained from the commercial partner. Data included product-level sales data (i.e., number of units sold of each item) for selected cafeterias (those who sold a mean of at least 50 meals per day before labels were introduced in April 2022). Individual customer transactions were not available.
Energy labelling was implemented on all products made on site. Products prepared on site were grouped into one of 10 categories for analysis, matching categories within the catering provider system. These categories were: Breakfast; Cakes, Pastries, Biscuits, and Discretionary; Fruits and Vegetables; Jacket Potatoes; Meals; Miscellaneous and Condiments; Salads and Cold Snacks; Sandwiches; Savoury Snacks; and Starters. Energy content of these items during this study period was provided by the commercial partner, as was the cost of each item. Energy data for pre-packaged items (e.g., bottled drinks, pre-packaged crisps and confectionery) were not available from the commercial partner – however, the provision of calorie information for these items did not change as a result of this legislation.
Study period: primary analysis
The overall study period for primary analysis is from 10 January 2022 to 10 October 2022, with the calorie labels implemented in sites on 4 April 2022. Following the pre-registered analysis plan (https://osf.io/t5jbh/), interrupted time series models were used to assess the potential impact of calorie label implementation. For the interrupted time series assessing level and slope change, the intervention point was at 4 April 2022, with a slope change after this point. For the interrupted time series assessing level changes at the time of catering provider menu changes, time was a fixed effect and menu changes were included as dummy variables. One of the menu changes occurred on 4 April 2022. All models were linear regression models, and for both outcomes, four models were run (using data from six weeks post-labelling, data from six months post-labelling, applying a secondary model, and a post-hoc analysis model). All models are centred at zero (i.e. time 0 represents the start of the intervention), and they are outlined in the below sections and Supplementary File A. Due to evidence of autocorrelation, all ITS and regression models applied Newey-West robust standard errors, with a lag of 28 days for the monthly menu cycles. Data cleaning and statistical analyses were conducted in R version 4.3.2.
Statistical analysis
Impact of calorie labelling implementation on energy purchased
Interrupted time series analysis with level and slope change was used to evaluate the impact of the implementation on energy (kcal) purchased from items made on site before and after the calorie labelling intervention. The primary outcome variable was mean daily calories purchased per food item from aggregated data. Weekends and bank holidays were removed, given the likelihood of unusual trading on these dates. Annual seasonality was difficult to account for, given the study period from January to October, however, a sensitivity analysis was run considering seasonal menu changes as covariates (which occurred approximately every 12 weeks). The primary model can be found in Supplementary File A.
Secondary analyses repeated the same analysis for different food categories (e.g. main meals, sandwiches), to examine the potential for different impacts by category.
Sales data for retail items (i.e. pre-packaged crisps, chocolates or snacks) and drinks were also analysed in secondary analyses for changes in quantity sales, assessing any potential knock-on effects of implementing the calorie labels. In these analyses, the outcome variable was number of items sold.
Impact of calorie labelling implementation on menu composition
Interrupted time series analysis with only level change was used to evaluate the impact on mean energy (kcal) of prepared menu items (e.g. hot meals, jacket potato toppings, etc.) offered before and after the implementation of calorie labels. The level-only ITS analysis considered the difference in energy offered on menus, using time as a fixed effect in the linear regression model and including when the catering provider made menu changes during the study period as dummy variables. This model was selected a priori, since it was known prior to the analysis that the menu change would coincide with the implementation of calorie labels on menus, and that another menu change would follow this 12 weeks later, providing the next notable opportunity for a reformulation of the menu. The model used to assess the impact of calorie labelling implementation on menu composition can be found in Supplementary File A.
Secondary analyses repeated the analysis by food category, and a sensitivity analysis considered the overall trend in energy offered, applying an interrupted time series model with level and slope changes. This model was considered, given the possibility that individual catering managers may select different foods from the base menu during the 12-week menu cycle, and this model could better assess a gradual shift over time. The model for the secondary analyses was the same as the primary analysis, and the model for the sensitivity analysis was the same as the energy purchased primary analysis.
Post-hoc analysis
Following primary analysis, a post-hoc analysis was run, where weekly-level product sales data from March 2021 to October 2022 were obtained. This allowed for a longer period of analysis, with an entire year of data prior to the implementation of calorie labels, so that each week post-implementation had a corresponding week in the pre-implementation stage to better account for seasonal changes. This period of time involved site closures, so there are a variable number of sites open in any given week for the aggregation and analysis, however, to consider this, a sensitivity analysis was also run including only those sites that were in the dataset from the beginning.
An interrupted time series analysis with level and slope change was run for both energy purchased and energy offered. Seasonality was included as a variable for every four-week period, equating to thirteen dummy variables in a fifty-two week-long year. No other additional variables were included (i.e. seasonal menu changes that are included in the primary analysis, since the seasonal week variable accounts for seasonality). Weeks with holidays or with incomplete data were excluded from analysis. The model used for the post-hoc analysis can be found in Supplementary File A.
The same method for robust standard errors was used as in other analyses, applying Newey-West standard errors with a lag of 4 weeks for the monthly menu cycle.
Sensitivity analysis
Not all sites were in the data set from the beginning to the end of the available data. Sensitivity analyses were run both for the primary analysis and post-hoc analysis models, including only those sites that appeared in the dataset at the first available data point. This included 137 of 142 sites that were in the dataset from the beginning for the primary analysis, and 97 of 142 sites for the post-hoc analysis.
Deviations from protocol
This analysis was originally intended to be on a dataset from January to September 2022. Additional data was used for a post-hoc analysis in light of concerns about the potential impacts of seasonality on the primary outcome. Secondary analyses on retail products were added in as well, to assess any potential knock-on effects to food items not affected by the change in calorie labelling. The catering provider predominantly offers two different sets of menus, with the menu type often associated with the type of worksite (e.g. manufacturing/distribution vs. office-based); however, information on type of menu offered was not available, so planned analyses exploring whether any impact differed by menu type could not be run. The sensitivity analysis with a reduced number of sites, only including those that were in the dataset from the beginning, was also not originally included in the analysis plan.
Add Comment