Association of a Sweetened Beverage Tax With Purchases of Beverages and High-Sugar Foods at Independent Stores in Philadelphia

Key Points Question How has the 2017 Philadelphia beverage tax factored into longer-term changes in beverage prices and purchases in independent food retail stores based on observational data? Findings This cross-sectional study found that, 2 years after tax implementation, price audits of stores showed 137% of the tax was passed through to prices and bag checks indicated a 42% decline in volume of taxed beverages purchased in Philadelphia compared with Baltimore. Total calories purchased from beverages and high-sugar foods declined, suggesting food substitution did not offset beverage declines. Meaning These findings suggest a city-level beverage excise tax was associated with persistent declines in purchases of sweetened drinks and calories from sugar in independent stores.


Beverage classification
Philadelphia beverage tax criteria: Beverages subject to Philadelphia's excise tax include soda, fruit drinks (not including 100% juice), sports drinks, flavored waters, energy drinks, pre-sweetened coffee or tea, and non-alcoholic beverages intended to be mixed into an alcoholic drink. Beverages that are not subject to Philadelphia's excise tax include: unsweetened drinks including those to which a purchaser can add sugar or request the addition of sugar (e.g., black coffee purchased at an independent store), baby formulas, beverages that meet the definition of medical food, any product for which more than 50% of its volume is milk or fresh fruit, vegetable, or a combination, and any syrup or other concentrate that a customer combines with other ingredients to create a beverage (e.g., powdered drink mixes like Kool-Aid).

Nutrient and serving size coding
We conducted online searches to identify the nutrient content and serving size for each the items in the high-sugar food categories: candy (n=265 at baseline, n=563 post-tax), sweet snacks (n=343 baseline, 830 post-tax) and pure sugar (n=7 baseline, n=11 post-tax). We searched for specific brand and flavor combinations identified by data collectors (e.g., Brand: Tropical Fantasy, Flavor: Guava) and prioritized nutrition information provided on brand websites (e.g., pepsicobeveragefacts.com). If brand websites did not provide nutrition information, we searched online retailers (e.g., Walmart.com, Amazon.com), and recorded nutrition information only if the retail listing included a photograph of the nutrition label. Next, we searched the USDA FoodData Central database (fdc.nal.usda.gov) for brand and flavor-specific nutrition information. If nutrition data for the brand/flavor combination were not available from any of the above resources, we used the USDA FoodData Central database to find the closes approximation of the food or beverage based on the level of detail provided by data collectors. For example, we were unable to find nutrition information for Lady Linda Honey Buns from our first three sources and instead selected a generic "honey bun" from the FoodData Central database. We were able to identify the nutrient content of foods and beverages for 3,112 (52%) items from brand websites, 1,624 (27%) items from online retailers, 488 (8%) items from the USDA FoodData Central, and 764 (13%) items closely approximated from the USDA FoodData Central.
For beverages, we recorded serving size in fluid ounces; if fluid ounces were not available, we recorded serving size in milliliters. For foods, we recorded serving size in grams. For all items, we recorded the kilocalories and grams of total sugar per serving to create a per-unit value. Using the number of servings per item (recorded by data collectors during purchase assessments), we were able to calculate the total calories and total grams of sugar per item.

Beverage price methods: Store recruitment
Sample size and exclusions. We collected data from 161 independent stores at baseline (53 in Philadelphia, 60 in Baltimore, 48 in neighboring counties). During the study, 36 stores closed or refused to continue participating and were replaced with similar stores near the original ones. Additionally, ten stores changed ownership, but remained in the same physical location. After excluding those stores that changed ownership or did not continuously provide data, the final analytic sample sizes for complete case analysis were 35 stores in Philadelphia (21 low-income, 14 other-income), 43 stores in Baltimore (16 lowincome, 27 other-income), and 38 stores in PA counties neighboring Philadelphia (16 low-income, 22 other-income) bringing the total number of stores to 116 (see Figure 1.3.a). Census data were used to examine the comparability of the neighborhoods where price data were collected in Philadelphia and Baltimore (Table 1.3.b). Data showed neighborhood sociodemographic characteristics were similar. Among all independent stores at baseline and 24 months, we collected data from 103 brand-size combinations including 66 (64.08%) sugar-sweetened beverages, 12 (11.65%) artificially-sweetened beverages, and 25 (24.27%) unsweetened beverages. Energy drinks (n=278 prices from 5 brand-size combinations collected at 65 stores) were excluded from price analyses due to their higher mean price per fluid ounce (energy drinks mean = 23.64 cents per fluid ounce across baseline and post-tax) compared to other taxed beverages (means ranged from 6.08 -6.25 cents per fluid ounce). (See eAppendix   Notes: Tract level estimates were averaged for each income and city group based on 2016 5-year American Community Survey data. A household consists of all people who occupy a housing unit, including for example, roommates living together, a person living alone, or a family. A family household consists of a group of two or more people who are related residing together. SOURCE: Analysis of price data from 5 beverage-size combinations in 65 small, independent stores, 31 in Philadelphia and 34 in Baltimore, before and two years after implementation of the tax. NOTES: a "Taxed" refers to beverages covered under Philadelphia's 1.5 cent per fluid ounce beverage tax on sugar-and artificially-sweetened beverages implemented Jan 1, 2017. b Percent change is calculated by dividing the difference-in-differences coefficient by the sum of the intercept and coefficient for Philadelphia. The numerator represents the change in price or price per fluid ounce 24 months post-tax using Baltimore as a control, and the denominator is the mean price or price per fluid ounce in Philadelphia at baseline. c Percent of taxed passed through to customer is calculated by dividing the difference-in-differences estimate by 1.5¢/fl oz. d P values and confidence intervals were Bonferroni corrected using 2 corrections.

Model selection and sensitivity analyses
We identified covariates that we hypothesized would influence the standard error of our model including: gender, race, ethnicity, education, age, who the purchase was for, frequency visiting the store, city residency, and total amount spent. The table below presents results adjusted for these covariates along with the unadjusted model that we present in the main paper and a complete case model, which excludes observations missing any of the covariates in our adjusted model (n=622, 13.1%). Our plan was to conduct unadjusted analyses because of the difference-in-differences design. We did, however, identify covariates a priori that were likely associated with our outcome to see if there were gains in efficiency. The adjustments did not meaningfully change the parameter estimates, standard errors, or our conclusions, so we report the unadjusted models in the main paper.  SOURCE: Analysis of customer purchase assessments (N=9,483) from 130 small, independent stores. NOTES: Significance for continuous measures is calculated for the within-city standardized mean difference from baseline using a t-test. Significance for independent distribution of categories within cities from baseline is calculated with Chi-Square Test of Independence. *p < .05, **p < .01, ***p < .001, Values may not sum to 100% due to missing values or rounding. The cut points for the store frequency variable (how often do you visit the store) are based on the distribution of the data and differ from the cut points used in our prior paper looking at the association of a sweetened beverage tax with changes in beverage prices and purchases at independent stores one year after tax implementation. 14 a Sugar buyers refers to someone who purchased a sweetened beverage or a high-sugar food.

Stores in the customer purchase assessment sample by city and wave
We conducted customer purchase assessments at a total of 58 independent stores in Philadelphia and 63 independent stores in Baltimore. The stores where customer purchase assessments were conducted were largely similar over time, though there were some differences due to store closures or differences in store traffic. The proportion of stores within categories of store type or income remained consistent.

Price elasticity of demand for taxed beverages
Elasticity was calculated based on the following equation: Elasticity=% change in volume/%change in price. The change in price was observed from our price audits (which do not include every beverage a store sells) and the change in volume was based on the customer purchase assessment data from purchases within Philadelphia and Baltimore.
These elasticity estimates do not account for possible tax avoidance behavior in PA counties neighboring Philadelphia, which would likely reduce the volume change estimates. Customer purchase assessments were not collected in neighboring PA counties due to funding constraints and a short pre-tax data collection window.   Overall, there was no significant change in the total calories of high-sugar foods purchased in Philadelphia compared to Baltimore post tax. There was a significant decrease in total calories purchased from sweetened beverages, and from high-sugar foods and sweetened beverages combined. The difference-in-differences estimates for calories purchased were not moderated by neighborhood income or customer education. SOURCE: Analysis of Philadelphia and Baltimore included 4,631 customer purchase assessments at 119 small, independent stores. NOTES: Baseline data are from 2016. a Income based on census-tract-level data from 2014 5-year American Community Survey estimates. Census tracts with 30% or more of the population living in poverty are considered "low-income" and the rest are "other-income". Of the total 4,631 customer purchase assessments, 2,117 were collected at small, independent stores located in lowincome census tracts and 2,514 were collected at stores in other-income census tracts. b Based on self-report of highest level of education. "Lower education" includes those with a high school degree, GED, or less, while "higher education" includes those with some college or more. Of the total 4,631 customer purchase assessments, 2,574 were collected among customers reporting "lower education" and 1,988 were collected among customers reporting "higher education". 79 customers missing values for education were dropped from education-stratified analyses. All models include a random intercept for store location.

.a Change in sugar purchased from high-sugar foods and beverages 12 months after a Philadelphia beverage tax
Overall, there was no significant change in the total grams of sugar of high-sugar foods purchased in Philadelphia compared to Baltimore post tax. There was a significant decrease in total grams of sugar from sweetened beverages and high-sugar foods and sweetened beverages combined. The difference-in-differences estimates for grams of sugar purchased were not moderated by neighborhood income or customer education.

.b Change in sugar purchased from high-sugar foods and beverages 24 months after a Philadelphia beverage tax
Overall, there was no significant change in the total grams of sugar of high-sugar foods purchased in Philadelphia compared to Baltimore post tax. There was a significant decrease in total grams of sugar from sweetened beverages and high-sugar foods and sweetened beverages combined. Education level moderated grams of sugar purchased from high-sugar food (interaction estimate=+12.9; 95% CI 2.8 to 23.0, p=.01) such that customers with lower education levels increased their grams of sugar purchased from high-sugar foods more than those with higher education levels, but it did not moderate the grams of sugar purchased from sweetened beverages or from high-sugar food and sweetened beverages combined. The difference-in-differences estimates for grams of sugar purchased was not moderated by neighborhood income. SOURCE: Authors' analysis of Philadelphia and Baltimore included 4,738 customer purchase assessments at 121 small, independent stores. NOTES: Baseline data are from 2016. All models include a random intercept for store location. a Income based on census-tract-level data from 2014 5-year American Community Survey estimates. Census tracts with 30% or more of the population living in poverty are considered "low-income" and the rest are "other-income". Of the total 4,738 customer purchase assessments, 1,948 were collected at small, independent stores located in lowincome census tracts and 2,790 were collected at stores in other-income census tracts. b Based on self-report of highest level of education. "Lower education" includes those with a high school degree, GED, or less, while "higher education" includes those with some college or more. Of the total 4,738 customer purchase assessments, 2,459 were collected among customers reporting "lower education" and 2,185 were collected among customers reporting "higher education". 94 customers missing values for education were dropped from education-stratified analyses. c Percent change is calculated by dividing the difference-in-differences coefficient by the sum of the intercept and coefficient for Philadelphia. The numerator represents the change in grams of sugar from high-sugar food and SSBs purchased 24 months post-tax using Baltimore as a control, and the denominator is the mean grams of sugar purchased in Philadelphia at baseline.

Change in total spent per customer purchase assessment 24 months after a Philadelphia beverage tax
Overall, there was no significant change in the total amount spent per intercept for all items purchased in Philadelphia compared to Baltimore post tax. The difference-in-differences estimate for total reported spending was not moderated by neighborhood income or customer education. SOURCE: Analysis of Philadelphia and Baltimore included 4,738 customer purchase assessments at 121 small, independent stores. NOTES: Baseline data are from 2016. a Total spending as reported by the customer. b Income based on census-tract-level data from 2014 5-year American Community Survey estimates. Census tracts with 30% or more of the population living in poverty are considered "low-income" and the rest are "other-income". Of the total 4,738 customer purchase assessments, 1,948 were collected at small, independent stores located in low-income census tracts and 2,790 were collected at stores in other-income census tracts. c Based on self-report of highest level of education. "Lower education" includes those with a high school degree, GED, or less, while "higher education" includes those with some college or more. Of the total 4,738 customer purchase assessments, 2,459 were collected among customers reporting "lower education" and 2,185 were collected among customers reporting "higher education". 94 customers missing values for education were dropped from education-stratified analyses. d Percent change is calculated by dividing the difference-in-differences coefficient by the sum of the intercept and coefficient for Philadelphia. The numerator represents the change in total spent per customer 24 months post-tax using Baltimore as a control, and the denominator is the mean total spent in Philadelphia at baseline. e p values and confidence intervals were Bonferroni corrected using 2 corrections each for store location neighborhood income and customer education. *p<0.05, **p<0.01, ***p<0.001