**Participants**
This study uses longitudinal data on household food and beverage purchases from January 1, 2015 to December 31, 2017 from the consumer panel study, Kantar WorldPanel Chile. Information on the Kantar WorldPanel Chile dataset has been published previously.[37] In brief, the response rate is 95%, and participants are excluded if they do not meet minimum purchasing standards (e.g. purchases from a “basic basket” of goods) or work for an industry that might influence their recording of their purchases (e.g., owner of a food company).
The dataset used in this study included packaged food and beverage purchases from a panel of 2,000 households located in cities (defined by Kantar as having a population >20,000) [41]. With replacement, our analytic sample had 2,380 unique households, with an average (median) follow-up of 29 (35) months, providing 69,760 household-month observations. Data on each purchase are collected by scanning product barcodes or using a codebook to assign barcodes for products without them as well as through weekly receipt review and household pantry inventories conducted weekly by study enumerators. Data on each purchase includes volume (ml) or weight (g), barcode, price per unit, retail outlet, brand, package size, and date. The dataset also contains information about each household in the panel, including household size, composition (e.g., age and gender of each member), household assets (number of rooms, bathrooms and cars), socioeconomic status (as defined by the market and public opinion research association Asociación de Investigadores de Mercado y Opinión Pública de Chile), geographic region, as well as the age of the main shopper and educational level of the household head.
**The Chilean regulation**
The Chilean Law of Food Labeling and Advertising covered both foods and beverages was first implemented in late June of 2016. Products subject to the regulation were required to carry FOP warning labels (e.g., black octagon on the front of the package, with the words “high in” sugar, sodium, saturated fat, or calories), were subject to child-directed marketing restrictions, and were banned from sales and promotion in schools and nurseries. The regulation was designed to become increasingly stringent over time, with more restrictive nutrient thresholds implemented in 2018 (phase 2) and 2019 (phase 3). In phase 2 of the regulation, an additional marketing regulation restricted marketing of regulated foods on any television program from 6am to 10pm, regardless of whether the audience or advertising content was directed towards children. More detail on the regulation has been previously published [31].
**Nutrition facts panel data, food categorization, and categorization by regulation status**
We obtained nutrition facts panel (NFP) data from photographs of products collected by a team of Chilean nutrition research assistants in Santiago food stores during the first quarters of 2015, 2016, and 2017[42]. Data on the NFP was linked at the product level to household food and beverage purchases using a similar process as in previous household purchase evaluations.[37-39] For the pre-regulation period, we linked purchases to NFP data collected in 2015 and 2016 (i.e., data reflecting the nutritional profiles of products available prior to the regulation). For the post-regulation period, we linked purchases to NFP data collected in 2017. If there was no direct 2017 link, we linked to the 2015–2016 NFP data. Linkages were based on barcode, brand name, and product description. Of total food and beverage purchases, 94% and were linked to collected NFP data. If no collected NFP data was available for a purchased product, it was linked to Mintel Latin America (5.9%) or other NFP data resources (0.1%).
After the linkage, a team of Spanish-speaking research assistants reviewed each product for nutritional details and categorization. Each food or beverage was then categorized as to whether it should be subject to the Chilean regulation according to the first phase (2016) nutrient profile model, which specifies different cut-offs for foods and beverages. Specifically, products were categorized as “high-in” (and thus subject to regulation) if they contained added sugar, added sodium, or added saturated fat and exceeded the nutrient thresholds (i.e., for beverages, >100 calories, >100 mg sodium, >6 g sugar, or 3 g saturated fat per 100 ml of product in its as-consumed form; for foods, >350 calories, 800 mg sodium, 22.5 g total sugar, or 6 g saturated fat per 100 g of product in its as-consumed form. “High in” products are required to carry the FOP warning label and are subject to the marketing and school sales restrictions. Products were considered “not high-in” if they did not meet these nutritional criteria.
For the foods analyses, we also classified foods into groups, including cereal-based foods, breakfast cereals, snacks and grain-based desserts, sweets and non-grain-based desserts, meat, poultry, and meat substitutes, dairy products and dairy substitutes, and yogurt (analyzed separately from other dairy because of non-substitutability), fruits and vegetables, condiments and sauces, spices and seasoning, and soups (S1 Table).
Some subgroups were excluded from analysis because data were not collected consistently over the entire time period (e.g., Kantar did not collect or record data on this food group). These subgroups are: wheat grains (within the cereal-based foods group), savory snacks (e.g., tortilla chips, potato chips, popcorn) (within the snacks and grain-based desserts group), and chocolate and candy (within the sweets and non-grain based dessert group)].
To understand how well each food group in Kantar represented food group sales in Chile, we used Euromonitor International sales data from 2015, 2016 and 2017 to compare brands represented in each dataset. For most groups, Kantar had between 78% and 100% of the brands available in Euromonitor. Only fish and seafood (67%) and fruits and vegetables (36%) had less, but these groups have among the smallest purchase volumes in our analytical dataset (due in part to the fact that bulk products could not be included in our analyses) (S2 Table). Further, all Kantar categories had four or five of the five brands with the most sales in Euromonitor (except for fruits and vegetables and food groups with less than five brands in Euromonitor, i.e. fish and seafood, salt and seasoning, and traditional mixed dishes). In addition, when comparing the relative contribution of food groups by volume to the total volume of foods purchased (in Kantar) or sold (in Euromonitor), most food groups were similar in share of contribution, suggesting that the overall profile of purchases in Kantar is similar to that being sold in the market during this time period (S3 Table).
Purchases were considered to have occurred in the pre-regulation period if they were purchased between January 2015 and June of 2016 and in the post-regulation period if they were purchased between July 2016 and December 2017.
**Outcomes**
All data were aggregated at the monthly level and analyzed in terms of calories (kcal), sugar (kcal), saturated fat (kcal), and sodium (mg) per capita per day. Because one of the major goals of the regulation was to prevent further increases in obesity, and excess calories has been linked to increased weight gain, the main outcome was calories per capita per day for total foods and total foods and beverages.
**Covariates**
Covariates included the head of household's educational level (less than high school, high school, more than high school), the age of the main shopper (under 30, 30 or over and under 60, and 60 or over), the household's socioeconomic status (low, middle-low, middle-high, and high), and household composition (a set of discrete variables treated as continuous variables, each with the number of people in the following age categories: children 0-1y, children 2-5y, children 6-13y, adolescents 14-18y, female adults >18y, and male adults >18y), and the monthly regional unemployment rate (https://webanterior.ine.cl/estadisticas/laborales/ene). We also included indicator variables for each calendar month (1-12) to adjust for seasonality and a continuous month variable (1-36) to account for any linear time trend.
**Statistical analyses**
All statistical analyses were conducted using Stata 16 (College Station, TX, USA).
For food-group specific analyses, we did not analyze groups with mean purchase volume wasof less than of 1 g per capita per day overall.
*Unadjusted analyses: descriptive statistics*
We examined household sociodemographic characteristics of households participating in Kantar WorldPanel Chile in 2015, 2016, and 2017. We also examined unadjusted proportions of households that purchased foods from each food group in a given month (i.e., purchases >0g) and unadjusted means of household purchases (calories, sugar, saturated fat, and sodium), overall, by regulation status (high in vs. not high-in), and by food group.
*Adjusted analyses: fixed effects models*
Consistent with our evaluation of changes in beverage purchases after Chile’s Law of Food Labeling and Advertising, [37] we ran models using 18-month windows before and after the regulation. The pre-policy period was considered as the 18 months from January 1, 2015 to June 30, 2016 and the post-policy period was considered as the 18 months from July 1, 2016 to December 31, 2017.
For our main analyses, we used a pre-post quasi-experimental modeling approach to examine changes in household food purchases before and after policy implementation. Similar to previous evaluations of national-level food policies [39-42], we compared predicted values in the post-policy period to a counterfactual, or what would have been expected in the post-policy period based on pre-policy trends projected into the post-policy period. To construct this counterfactual, we included in each model a binary variable for the policy period (pre vs. post) and its interaction between with a count variable for time (e.g., month, from 1 to 36) to account for a potential break in the trend associated with the policy change. Using Stata’s margins command, we calculated the difference in the predicted values in the post-policy period keeping the policy variable fixed at the pre-policy value (i.e., the “counterfactual) as well as at the post-policy value (i.e., post-policy “observed”).
For each outcome (calories, sugar, saturated fat, and sodium per capita per day) overall, high-in, and not-high-in food purchases, we calculated the average absolute and relative differences between the observed post-policy trend and the counterfactual based on predicted values from the model. We used fixed effects models accounting for non-time-varying unobserved household characteristics, and controlled for the aforementioned covariates. Using the dydx option of Stata’s margins command, we then conducted statistical tests to compare the differences in the absolute mean amounts purchased in the post-policy period between the counterfactual and observed, and derived the relative percent difference from this.
To test whether there were differences in response to the regulation by education and socio-economic status, we conducted the same models as above but including an interaction of education or socio-economic status with every variable in the model. The variable of interest was the triple interaction of education or socio-economic status with the policy period and time trend; statistical significance of this joint interaction was tested with an F test. We also predicted the differences between observed and counterfactual for each educational or socio-economic status level using the aforementioned model.
For our analyses at the food group level, we focused our attention on consumers due to high levels of non-consumers at the monthly level in each food group. We first examined changes in the likelihood to purchase high-in products within a food group (i.e., purchasing >0 g of the food group in a given month) between the pre- and post-policy periods using a random-effects probit model, controlling for the same set of covariates as before. Subsequent food-group level analyses focused on analyzing mean amounts of purchases of calories, sugar, saturated fat, and sodium (overall and by regulation status) amongst consumers-only. To account for any change in the selectivity of being a consumer of a particular food group over time, we applied stabilized inverse probability weights (SIPWs) at the household-month level in our fixed-effects models.[43-45] To construct these SIPWs, we used the predicted the probabilities of being a consumer of each food group in a given month from the same random-effects probit model. More specifically, we calculated the SIPWs as:
w_(i,j,k)={■(P_(j,k)/p_(i,j,k) &for consumers@〖1-P〗_(j,k)/(1-p_(i,j,k) )&otherwise)┤
for household i, food group j and month k, where Pj,k is the proportion of consumers of food group j in month k and pi,j,k is household i’s predicted probability of being a consumer of food group j in month k. We then truncated these weights at the 0.5th and 99.5th percentiles, calculated by food group.
Finally, although our data on beverages alone has been previously published,[37] we aggregated data on calories, sugar, saturated fat, and sodium from total food and beverage purchases (overall and by high-in status) in order to be able to see changes in nutrients coming into the household across the total range of products. For these analyses, we again conducted fixed effects models controlling for covariates, comparing the post-policy trends to the expected post-policy trends based on predictions from the pre-policy period.
Sensitivity analyses
For the main outcome (calories of overall, high-in and not-high in food purchases), we conducted several sensitivity analyses to test the robustness of our model to various specifications. First, we wanted to test the robustness of our results to our choice of linking purchases in the third and fourth quarters of 2016 to 2017 nutritional data, by linking all 2016 purchases to pre-regulation nutritional data. While the magnitude of changes by regulation status was smaller (-24.8 kcal and 16.1 kcal for high-in and not-high-in products, respectively, compared to -34.0 kcal and 24.6 kcal originally), the overall change was similar (-8.7 kcal, compared to -9.4 kcal originally).
Further, for our main analyses (calories, sugar, saturated fat and sodium per capita per day overall and by regulation status) we wanted to test the assumption of an absolute change effect implied by our linear models against that of a relative change effect using fixed-effects Poisson regressions, and obtained similar results with regards to differences between the observed and counterfactual.
Finally, we conducted two more sensitivity analyses with respect to inverse probability weighting.
First, we compared the random-effects pooled probit model to a fixed-effects linear probability model to explore the validity of the random-effects assumption. Second, we repeated the analysis by food group with non-truncated weights and weights truncated at the 1st and 99th percentiles and also found similar results. In both cases, we obtained virtually identical mean predicted probabilities.