Statistical analysis

We aggregated across purchases made by households within each quarter and calculated overall mean daily sugar purchased (g) per household member in each quarter. We also estimated this by food group and for healthier/less healthy beverages/foods separately. We applied weights provided by Kantar to account for differences in the frequency of purchase of products across different places of purchase. This ensured that product-specific purchases reported in the panel were representative of the GB market. To estimate changes in sugar purchases over time, we used generalised linear models with a log-link function⁽²⁶⁾ using generalised estimating equations to account for the hierarchical structure of the data at household level (observations nested within households)⁽²⁷⁾. We used unstructured working correlation structures, unless the model failed to converge, in which case we used exchangeable working correlations. We adjusted for age of the main shopper (18–29, 30–39, 40–49, 50–59, 60–69, 70+), number of children (0, 1, 2, 3+) and number of adults (1, 2, 3, 4+) in the household, region of residence (ten regions in Britain) and season (i.e. quarter). We specified interactions between social grade measure (grades A&B, C1, C2, D and E) with year dummies to allow for differences in time trends in sugar purchasing by social grade. These socio-demographic covariates were almost time-invariant (not presented). Based on the estimated coefficients, we then predicted mean amount of sugar (g) purchased by social grade and year at fixed values of the other covariates. These were defined using distributions from the GB population for each of the covariate in order to obtain estimates representative of the GB population average (see online supplementary material, Supplemental Table S2). Inequalities in sugar purchasing were examined for 2014 and 2017 using (i) a test of overall evidence of difference between the five social grades and (ii) estimated differences between the lowest (grade E) and the highest social grades (grade A&B). Data analyses were performed using Stata MP version 15.1⁽²⁸⁾.