This study was approved by The University of New South Wales Human Research Ethics Committee (approval number HC200244).
Study design and population
This study used household purchase data from the NielsenIQ Homescan Consumer Panel, a dataset that contains household-level food and beverage purchase data from a panel of approximately 10,000 Australian households. These households are recruited to be broadly representative of the demographic composition and geographic location of Australian households [23, 24]. Participating households are provided with a handheld electronic scanner and are asked to scan the barcode of all foods and beverages brought into the home from all retail outlets including supermarkets, grocers, and convenience stores. Data on non-barcoded items such as deli meats, fresh bakery items and fresh unpackaged fruits and vegetables are also collected through use of a scanning guide booklet provided to the households by NielsenIQ [23, 24]. No data are collected on food purchased and consumed outside of the home, such as take-away and restaurant foods. For the present analyses, both barcoded and non-barcoded items were included.
Sociodemographic information is also collected from households, including information about ethnicity, education level of the main shopper in the household, household income, lifestage (e.g., adult households and young families), and age and sex of all household members. To capture annual changes in purchasing behaviours from 2015 to 2019, all household purchases were aggregated for each calendar year over the five-year period (January 1st – December 31st).
Household eligibility
To exclude households with potentially unreliable data, we applied criteria set by NielsenIQ. We removed households who: (i) were not on the panel for the entire 52-week time frame; (ii) did not report purchase data (at least one barcode per week) for at least 50% of the weeks within a 12-month timeframe; (iii) were missing any demographic information; (iv) did not meet the minimum spend criteria (≥ $5 on average for each week for all purchases). As previously described, to further reduce the potential impact of under-reporting, we excluded households with the lowest annual food and beverage expenditure (< 2.5th percentile defined separately for single-member households and multi-member households) [23,24,25,26,27].
Nutrition information
To determine the energy content of each food and beverage product at the time of purchase, barcoded products in the Homescan dataset were linked with corresponding nutrition information from the FoodSwitch nutrition composition database [28]. This database contains nutrition information for more than 80,000 packaged foods and beverages that have been available for sale in Australia since January 2013. Most of the data (~ 60% of all products) are captured by trained data collectors through in-store surveys at five large Australian supermarkets owned by Aldi, Coles, Harris Farm, Independent Grocers of Australia (IGA) and Woolworths in the Sydney metropolitan area [29]. Images of the pack of each food and beverage product are captured (front of pack, nutrient declaration, ingredients list, manufacturer details), using a bespoke smartphone application. The product name, brand name, package size (g) and nutrient content per 100 g/mL and per serve are then extracted [23, 30]. The database also contains data that are (i) crowdsourced using the FoodSwitch smartphone application (~ 30%) and (ii) provided directly by the food industry (~ 10%) [29].
As FoodSwitch contains only nutrition information for packaged foods that carry a nutrition information panel, we extracted energy content information for unpackaged, unbarcoded products reported by households (i.e. scanning guide items) from similar food or beverage subcategories in Australian Food and Nutrient database (AUSNUT) 2011–2013 [31]. AUSNUT is a food nutrient database containing nutrient values for 5740 generic foods and beverages with reported consumption in the 2011–2013 Australian Health Survey [31]. Where multiple relevant products were available in the AUSNUT dataset, we used the average energy content of all relevant products.
All products were assigned to food categories based on the categorisation system developed by the Global Food Monitoring Group, which classifies all products into a hierarchical category tree to allow for comparison of nutritionally similar foods [28]. This system classifies each product into a food group (e.g., bread and bakery), category (e.g., bread), subcategory (e.g., flat bread) and minor category (e.g., regular wraps).
Merging NielsenIQ and FoodSwitch datasets
The first step in merging the two datasets was to exclude products not relevant for the analyses. This included the removal of non-food and beverage products from the NielsenIQ Homescan database, such as medicinal items and cleaning products. We also excluded alcoholic beverages, vitamins and supplements from both databases.
The remaining food and beverage products in the NielsenIQ dataset were then linked with their corresponding nutrient information from FoodSwitch to obtain the energy content and NOVA classification. Initial matching of NielsenIQ Homescan to FoodSwitch was carried out using the unique barcode associated with each product followed by additional steps to further improve the coverage of products purchased by households [23, 32]. This included linking products by product name only, then by product name after removal of irrelevant descriptors (e.g., shape and size information). For unpackaged foods and beverages (i.e., those without a barcode), the energy content was matched to information from AUSNUT. After these additional steps were applied, the match rate across the NielsenIQ and FoodSwitch datasets was approximately 96.6% according to the total volume of products purchased over the five-year period. There was similar coverage across each of the five years (2015 = 96.1%, 2016 = 96.6%, 2017 = 96.8%, 2018 = 97.0% and 2019 = 96.6%).
Level of processing classification
The NOVA system categorises products into four categories based on the extent and purpose of industrial food processing. These include Group 1: Unprocessed or minimally processed foods (e.g., rice, meat, fish, milk, eggs, fruit, vegetables, nuts, and seeds); Group 2: Processed culinary ingredients (e.g., sugar, oils, butter); Group 3: Processed foods (e.g., canned fruit, canned fish, freshly baked bread, some cheeses); and Group 4: Ultra-processed foods (e.g., mass produced packaged breads, cookies/pastries, confectionery, savoury snacks, reconstituted meat products and sugar sweetened beverages) [3].
We categorised products matched across the NielsenIQ and FoodSwitch datasets (96.6% of all product units) into two groups based on level of processing: (1) ultra-processed (NOVA Group 4) and (2) non-ultra-processed (NOVA Group 1 to 3). Using previously described methods, products with ingredient list information (~ 93% of all product units) were classified as ultra-processed if they contained ultra-processed ingredients i.e. ingredients that are never or rarely used in household kitchens or additives that function to make foods more palatable and/or appealing, including flavours, emulsifiers, modified starches, vegetable gums, stabilisers and artificial sweeteners [7, 33, 34]. A full list of ingredients used to identify ultra-processed foods is provided in Supplementary Table 1. For products missing ingredient list information (~ 3% of all product units), we applied the NOVA system using food category information [3]. For example, any eggs, legumes, herbs, unprocessed and unflavoured meat, poultry and seafood products were categorised under Group 1: unprocessed or minimally processed foods, whereas sugar sweetened beverages, sweet and savoury snack foods, chocolate, ice-cream, breakfast cereals were categorised under Group 4: Ultra-processed foods.
Socio-economic status
The SES of participating households was assessed based on their postcode using the Index of Relative Social Advantage and Disadvantage (IRSAD), which is a Socioeconomic Index for Areas (SEIFA) as defined by the Australian Bureau of Statistics (ABS) [35]. IRSAD ranks geographic areas according to relative socio-economic advantage and disadvantage using a range of indicators including education, income, occupation and housing [35]. Using this index, households were divided into quintiles according to SES (Quintile 1: lowest SES; quintile 5: highest SES).
Statistical analysis
We assessed Australian household purchases of ultra-processed foods in 2019 using two outcome measures, 1) mean per capita purchases (grams/day); the amount of ultra-processed foods purchased daily per person and 2) contribution to total daily energy purchases (% energy); amount of energy purchased from ultra-processed foods as a proportion of total energy purchased from all grocery purchases. The major food categories contributing to total purchases of ultra-processed foods across Australian households were identified and ranked according to their relative contribution (%) to total daily energy purchases. We also explored differences in household purchases of ultra-processed foods across quintiles of SES. Differences in mean per capita purchases of ultra-processed foods across each of the SES quintiles were assessed using survey-weighted linear regression.
We also assessed changes in purchases of ultra-processed foods by SES between 2015 and 2019. Weighted linear mixed models were fit with the household as a random effect nested within the region with the year as a fixed effect. Purchases of ultra-processed foods and the contribution of ultra-processed foods to total purchases of energy were treated as dependent variables in the models. The model included the number of children in the household, the number of adults in the household, and life stage of the household (young singles & couples, young families, mixed families, older families, older singles & couples, adult households) as a fixed effect and were adjusted for in the model as these factors are likely to impact purchasing behaviours. To examine differences over time between SES groups, weighted linear mixed models were fit with an interaction term between SES and year. Using the parameters from the models, predicted means using a household of two adults and one child in the purchases of ultra-processed foods (g/d per capita) and contribution of ultra-processed food to total purchases of energy (% energy) among households of different socio-economic status over time were calculated. Survey weights (provided by Nielsen IQ) were applied throughout all analyses to ensure annual purchases were representative of the SES, demographic, and geographic composition of the Australian population [23].
All statistical analyses were performed using R Studio (version 1.4.1106) & R (version 4.1.0). Packages survey and lme4 were used for this analysis. A two-sided p-value of < 0.05 was considered statistically significant.