Philadelphia’s Excise Tax on Sugar-Sweetened and Artificially Sweetened Beverages and Supplemental Nutrition Assistance Program Benefit Redemption

Abstract

Objectives. To assess the effect of a 2017 excise tax on sugar and artificially sweetened beverages in Philadelphia, Pennsylvania, on the shopping patterns of low-income populations using Supplemental Nutrition Assistance Program (SNAP) data.

Methods. I used a synthetic controls approach to estimate the effect of the tax on Philadelphia and neighboring Pennsylvania counties (Bucks, Delaware, and Montgomery) as measured by total SNAP sales (“SNAP redemption”) and SNAP redemption per SNAP participant. I assembled biannual data (2005–2019) from all US counties for SNAP redemption and relevant predictors. I performed placebo tests to estimate statistically significant effects and conducted robustness checks.

Results. Detectable increases in SNAP spending occurred in all 3 Philadelphia neighboring counties. Per-participant SNAP spending increased in 2 of the neighboring counties and decreased in Philadelphia. These effects were robust across multiple specifications and placebo tests.

Conclusions. The tax contributed to increased SNAP shopping in Philadelphia’s neighboring counties across both outcome measures, and decreased spending in Philadelphia (at least by 1 measure). This raises questions about retailer behavior, the effectiveness of the tax’s public health aim of reducing sugar-sweetened beverage consumption, and policy aims of investing in low-income communities. (Am J Public Health. 2021;111(11):1986–1996. https://doi.org/10.2105/AJPH.2021.306464)

Policies aimed at reducing population-level intake of sugar-sweetened beverages (SSBs) via excise taxes have now been adopted by several cities in the United States (e.g., Berkeley, Oakland, and San Francisco, California).¹ Recent evaluations of a $.015-per-ounce excise tax in Philadelphia, Pennsylvania, a broader tax on both sugar-sweetened and artificially sweetened beverages, have documented significantly higher prices and lower sales of taxed items following the policy’s implementation on January 2017.² Though other research has not found any significant changes in unemployment claims across potentially affected industries, retailers and industry groups contend that recent store closures are directly related to the Philadelphia tax.³ Tax opponents often cite the regressive nature of the tax (i.e., on average, lower-income individuals spend a higher proportion of their incomes on SSBs than higher-income individuals), and supporters highlight the tax’s progressivity via its revenue-raising utility for antipoverty initiatives (e.g., universal prekindergarten schooling), a deliberate political feature of the Philadelphia tax identified by qualitative researchers.^4,5 Still, to my knowledge, no research has investigated how low-income shoppers responded to the beverage tax.

The socioeconomic context of food shopping is critical in Philadelphia. While the city’s unemployment rate has declined from a 2012 peak of 10.9% to 6.2% in 2017 (vs national rates of 9.0% and 4.4%, respectively), food insecurity has risen over recent years, and the poverty rate has remained relatively flat at about 26%, the highest among the nation’s largest 10 cities.^6,7 In 2017, 22.1% of Philadelphia households participated in the federal government’s largest and most widespread effort to reduce food insecurity, the Supplemental Nutrition Assistance Program (SNAP), nearly double the national rate of participation (11.2%).⁸ However, the average monthly amount of SNAP benefits redeemed at eligible retailers in Philadelphia also declined by $5.4 million between 2016 and 2017, mirroring, at least partially, national declines in SNAP participation after a period of expansion following the 2008 economic crisis as part of the American Recovery and Reinvestment Act. During the same time, however, SNAP spending increased in the counties immediately neighboring Philadelphia, contrary to state and national trends. A key aim of this study was to identify approximately how much of these changes in SNAP spending, if any, is attributable to the tax.

Generally, SNAP is allocated monthly to participating households via electronic benefit transfers in amounts based on characteristics such as income, age, and the presence of children; nationwide, the average SNAP participant received about $127 per month in 2018. The program is “cash-like,” in that participants can redeem benefits on any food item (excluding prepared foods, alcohol, medicines, or vitamins) at approved retailers. For many low-income households, SNAP benefits comprise a significant portion of monthly food budgets and are fully exhausted by the end of the month, though unused benefits are carried forward.^9,10 While the program helps guard against experiences of deep poverty, food insecurity, and hospitalization,^11,12 the nutritional content of SNAP purchases has also been the subject of scrutiny. Proposals from the political left and right call for restricting or excluding certain food items from SNAP,^13,14 responding in part to studies that identify socioeconomic gradients in diet quality, with poorer households purchasing and consuming less healthy, often lower-cost foods.^15,16 Cyclical patterns are commonly observed in analyses of benefit use, with end-of-month periods characterized by decreased food intake and lower nutritional quality of foods consumed.^9,17 The program’s relationship with neighborhood food environment is also notable: previous research has shown associations between SNAP retailer availability and participant enrollment,^18,19 a multiplier effect of SNAP spending that stimulates broader economic activities,²⁰ and the monthly cycle in SNAP spending cited by retailers as a barrier to operating stores in lower-income communities.²¹ Thus, SNAP spending is key to our understanding of how policies might affect the health and everyday lives of socioeconomically disadvantaged groups.

METHODS

Monthly totals (2005–2019) for the dollar value of all SNAP benefits spent at eligible retailers in a given county (subsequently called “SNAP redemption”) were obtained from the US Department of Agriculture (USDA) for all counties in the United States from 2005 to 2019. Monthly counts (2005–2019) of stores where SNAP could be redeemed were also included. Redemption amounts from counties with 4 or fewer SNAP-eligible retailers are redacted by the USDA to protect retailer identities, though this typically only applies to very small counties.

The count of SNAP participants and amounts of SNAP benefits distributed were also obtained for every county from USDA; these data are released biannually, with month estimates provided for every January and July. State-level agencies that administer SNAP locally collect this information and provide it to USDA. Monthly county-level unemployment rate estimates were obtained from the US Bureau of Labor Statistics. County-level population estimates from the 2012–2016 5-year American Community Survey were also included as covariates. Table A (available as a supplement to the online version of this article at http://www.ajph.org) provides a summary of data sources.

Study Dates

To allow this analysis to incorporate government stimulus‒related SNAP spending patterns, the starting month was selected as January 2005, several years before the onset of the American Recovery and Reinvestment Act. To temporally match data sets, only observations from January and July were included in this analysis, as this was the frequency of USDA reporting for county-level SNAP-participating individuals and SNAP benefits distributed. January 2019 observations were omitted from this analysis, as an anomalous spike appears in SNAP redemption across all counties. July 2019 was selected as the end month, and the resulting monthly data set included 24 pretax observations and 5 posttax observations for each county.

Outcome Variables

The primary outcome variable was total county-level monthly SNAP redemption. As a secondary outcome, I calculated the value of SNAP redemption in a county per SNAP participant (hereafter called “SNAP per participant”) by dividing SNAP redemption by the count of SNAP participants in a county for a given month. I repeated the analyses subsequently described separately for each outcome.

Statistical Analyses

I employed a synthetic control approach to model the effect of the excise tax. I fit 4 separate synthetic control models (Philadelphia and its immediate neighbors, Bucks, Delaware, and Montgomery counties) for each outcome. Philadelphia is a consolidated city‒county, so references to “Philadelphia,” “Philadelphia County,” or “city of Philadelphia” are interchangeable. I did not assess neighboring counties in New Jersey, given the additional physical and economic obstacles to travel (e.g., Delaware River with only toll-bridge crossings; see Figure A, available as a supplement to the online version of this article at http://www.ajph.org). This synthetic control method, pioneered by the work of Abadie, Gardeazabal, Diamond, and Hainmueller,^22–24 has been used in previous social science research and is built upon interactive fixed-effect models, which generate hypothetical or “synthetic” controls for each treated unit, rather than a more traditional matching approaches often used in evaluation research.^24–26 The synthetic control approach leverages time-series outcomes across many potential donor units to identify optimal weights for estimating a counterfactual unit and can also integrate and weight the influence of covariates. Synthetic control studies typically employ a test statistic proposed by Abadie et al.—ratio of the root mean squared prediction errors (RMSPEs)—to compare models and calculate a treatment P value from a placebo test, which refits a synthetic control for each of the donor units.²⁴

Selection of donor counties

As described by Bouttell et al., synthetic control methods rest on several assumptions: (1) the similarity of treated and potential control units, (2) no spillover effects of intervention to potential controls, and (3) no external shocks.^27(p676) To satisfy the first assumption, following the guidance of McClelland and Gault,²⁸ the pretreatment trends for SNAP redemption were also used to narrow the pool of potential donor counties. I calculated simple linear regressions for each potential donor county for 2 periods, 2008–2013 and 2014–2016, and extracted a slope to approximate the trend in SNAP redemption for the county. I chose these periods to reflect the expansion in SNAP under the American Recovery and Reinvestment Act, with peak SNAP enrollments occurring in 2013. Nationally, SNAP enrollment was declining in the second pretreatment period (2014–2016). At least 20 possible donor counties were selected based on their pretreatment predictor values and slopes.

This analysis directly addressed the second assumption, as it attempts to measure the spillover of the Philadelphia policy into its most geographically accessible neighbors by fitting separate models for each neighboring county. These neighboring counties were excluded from each other’s potential donor pools (e.g., Montgomery County was excluded as a potential donor for synthetic Bucks and Delaware counties). Furthermore, any counties with similar local beverage taxes during the study period (San Francisco County and Alameda County in California)²⁹ and their immediate neighbors were also excluded from the data set. I used visual inspection of time-series outcome variable plots to identify potential external shocks (assumption 3).²⁶ I included only counties with complete data during the study period in the final data set.

Synthetic control estimation and specification searching

Recent methodological literature provides further guidance for systematic application of synthetic controls, especially as to the use of pretreatment lagged outcome observations in developing the synthetic unit.³⁰ The guidance provided by Ferman et al. is meant to offer a systematic way of specifying, choosing among, and testing different synthetic control specifications (Table B, available as a supplement to the online version of this article at http://www.ajph.org) for significant treatment effects, as a means of guarding against specification searching for statistical significance, and over- and underrejection of the null hypothesis.³⁰ In short, I used the following steps to identify synthetic control specifications and assess importance of the results: (1) visual inspection of pre-treatment gaps, which should be relatively small for well-fitting synthetic controls; (2) a P value less than .10 from the placebo test described by Abadie et al.;²⁴ (3) consistency of treatment effect observed across 95% confidence set calculations that combine specifications not rejected by steps 1 and 2; and (4) for the specification selected by a MSPE criterion proposed by Dube and Zipperer,³¹ consistency of the treatment effect observed in different covariate configurations, and “leave-one-out” and “in-time” placebo tests. These steps are described in greater detail in the Appendix (available as a supplement to the online version of this article). The assembled data set and code used for conducting the analyses are available on the Oxford University Research Archive (https://doi.org/10.5287/bodleian:0oqGkDBdy).

RESULTS

The average monthly SNAP benefits redeemed in Philadelphia County in 2016 was $61.78 million (SD = $.8 million), and $7.85 million across the 3 bordering counties in Pennsylvania, all higher than the monthly average for Pennsylvania counties ($2.22 million) or the national average ($1.8 million). In 2017, SNAP redemption fell by 6.5% in Philadelphia (to $57.75 million), 1.9% in other Pennsylvania counties (to $2.18 million), and 2.1% nationwide (to $1.76 million). All of Philadelphia’s neighboring Pennsylvania counties saw a rise in SNAP redemption: 3.1% in Bucks (to $5.25 million), 6.0% in Montgomery (to $9.77 million), and 8.9% in Delaware County (to $10.06 million). Figure 1 illustrates the trend for SNAP outcomes in Philadelphia and its immediate neighbors.

FIGURE 1—

Monthly SNAP Outcomes in Philadelphia and Neighboring Pennsylvania Counties by (a) Log SNAP Redemption and (b) SNAP Redemption per Participant: 2005‒2019, Indexed to January 2005

Note. SNAP = Supplemental Nutrition Assistance Program. Dashed vertical line indicates the implementation of the Philadelphia tax in January 2017. Neighboring counties include Bucks, Delaware, and Montgomery counties in Pennsylvania. An alternate version of this figure is provided as Figure B (available as a supplement to the online version of this article at http://www.ajph.org) and illustrates trends for each county separately.

Over the same period, SNAP participation decreased nationally (‒2.8%), with a smaller decrease in Pennsylvania counties (‒0.7%). Philadelphia saw a decrease in SNAP participation of 1.4%, while 2 of its neighbors saw steeper decreases (‒3.9% in Montgomery and −2.2% in Bucks), and there was an increase of less than half a percent (.2%) in Delaware County. From 2016 to 2017, SNAP benefit distributions decreased by 4.4% nationwide, 1.4% in Pennsylvania counties, and 2.1% in Philadelphia. Among neighboring counties, decreases were also observed: −2.7% in Bucks, −0.5% in Delaware, and −5.3% in Montgomery.

SNAP Redemption Outcome

For SNAP redemption in Philadelphia, no specifications appeared to have a good fit, indicating that a synthetic control estimation may not be appropriate. Visual inspection of gap plots illustrated that the differences observed after treatment were not much larger than gaps seen between the treated and synthetic Philadelphia data in pretreatment periods. This may be reflective of the need to relax selection criteria to achieve a donor pool size larger than 20, indicating the difficulty in establishing a reasonable synthetic unit given the predictors used.

For SNAP redemption in Bucks County, all specifications indicated a positive long-run trend; of these, specification 5 (all odd pretreatment periods) had the lowest RMSPE (Table 1) and was also the selected specification per the criterion outlined by Dube and Zipperer.³¹ Figure C (available as a supplement to the online version of this article at http://www.ajph.org) illustrates relatively minor differences between specifications. Two covariates, total population and the number of authorized retailers, had marginal weights in the selected model (.022 and .001, respectively) (Table 2). Placebo testing yielded a P value of .087 (Figure D, available as a supplement to the online version of this article at http://www.ajph.org). Confidence sets for the effect functions of all specifications indicated a positive effect in the posttreatment period, with the earliest posttreatment periods crossing zero (Figure E, available as a supplement to the online version of this article at http://www.ajph.org). Donor county weights are reported for all specifications in Table C (available as a supplement to the online version of this article at http://www.ajph.org). Figure 2 summarizes the estimated gaps between treated and synthetic counties across the study period.

TABLE 1—

Fit Criteria Used in Specification Searching for Synthetic Controls: Philadelphia, PA, and Neighboring Pennsylvania Counties, 2005‒2019

Specification	Philadelphia County			Bucks County			Delaware County			Montgomery County
RMSPE	MSPE	P	RMSPE	MSPE	P	RMSPE	MSPE	P	RMSPE	MSPE	P
Log SNAP redemption
All lags	0.849	0.002	.96	6.803	0.021	.09	6.456	0.032	.04	7.738	0.062	.045
First three quarters	1.275	0.006	.84	6.345	0.019	.09	6.458	0.034	.04	7.644	0.069	.045
First half	1.374	0.007	.84	6.643	0.020	.043	6.189	0.034	.04	6.584	0.052	.045
Odd lags	0.849	0.002	.96	6.814	0.021	.09	6.442	0.035	.04	7.502	0.058	.045
Even lags	0.849	0.002	.96	6.034	0.018	.09	6.222	0.048	.04	7.735	0.062	.045
SNAP redemption per participant
All lags	6.520	33.51	.04	2.313	58.18	.13	9.485	288.24	.029	5.243	1048.11	.054
First three quarters	6.500	33.36	.038	2.331	59.12	.2	9.673	317.78	.029	5.244	1048.49	.054
First half	1.922	18.17	.46	2.952	112.15	.2	7.415	293.81	.029	5.242	1047.69	.054
Odd lags	6.510	33.38	.038	2.357	60.52	.24	9.530	292.06	.029	5.244	1048.24	.054
Even lags	6.462	33.02	.038	2.148	51.77	.28	9.222	295.56	.029	5.075	991.37	.054

Note. MSPE =  mean squared prediction error; RMSPE = root mean squared prediction error; SNAP = Supplemental Nutrition Assistance Program. P values were calculated via placebo test method that reflects the rank of posttreatment gaps in the county under observation compared with placebo counties, divided by the number of placebo counties plus 1.

Source. Used approach described by Ferman et al.³⁰

TABLE 2—

Treated and Synthetic Neighboring County Predictor Means and Weights for Log SNAP Redemption Outcome: Pennsylvania Counties Neighboring Philadelphia, 2005‒2019

	Bucks County			Delaware County			Montgomery County
Treated Mean	Synthetic Mean	Weight	Treated Mean	Synthetic Mean	Weight	Treated Mean	Synthetic Mean	Weight
Predictor
Populationa	13.35	13.28	0.022	13.24	13.03	0.000	13.61	13.53	0.018
SNAP benefitsa	14.95	14.86	0.000	15.65	15.79	0.004	15.2	15.45	0.001
SNAP participantsa	10.22	10.17	0.000	10.89	11.04	0.009	10.46	10.66	0.017
SNAP retailersa	5.38	5.29	0.001	5.81	5.81	0.157	5.66	5.6	0.001
Unemployment rate	6	5.92	0.000	6.35	7.28	0.000	5.4	6.17	0.004
Redemption amount—lagged outcomesa
Jan 2005	NA	NA	NA	15.11	15.13	0.066	14.77	14.74	0.066
Jul 2005	14.33	14.33	0.000	15.14	15.16	0.060	14.79	14.79	0.016
Jan 2006	NA	NA	NA	15.19	15.23	0.031	14.84	14.88	0.046
Jul 2006	14.37	14.37	0.216	15.19	15.18	0.059	14.83	14.85	0.189
Jan 2007	NA	NA	NA	15.23	15.2	0.097	14.9	14.88	0.006
Jul 2007	14.47	14.44	0.166	15.25	15.2	0.086	14.95	14.93	0.028
Jan 2008	NA	NA	NA	15.310	15.27	0.085	15.06	15.03	0.219
Jul 2008	14.59	14.57	0.062	15.31	15.31	0.065	15.11	15.09	0.230
Jan 2009	NA	NA	NA	15.47	15.49	0.049	15.31	15.33	0.002
Jul 2009	15.06	15.1	0.027	15.73	15.73	0.048	15.57	15.6	0.003
Jan 2010	NA	NA	NA	15.79	15.8	0.074	15.66	15.7	0.113
Jul 2010	15.3	15.31	0.197	15.85	15.89	0.110	15.73	15.78	0.040
Jan 2011	NA	NA	NA	NA	NA	NA	NA	NA	NA
Jul 2011	15.48	15.45	0.000	NA	NA	NA	NA	NA	NA
Jan 2012	NA	NA	NA	NA	NA	NA	NA	NA	NA
Jul 2012	15.53	15.51	0.027	NA	NA	NA	NA	NA	NA
Jul 2013	15.5	15.52	0.09	NA	NA	NA	NA	NA	NA
Jul 2014	15.45	15.44	0.03	NA	NA	NA	NA	NA	NA
Jul 2015	15.44	15.43	0.007	NA	NA	NA	NA	NA	NA
Jul 2016	15.43	15.44	0.154	NA	NA	NA	NA	NA	NA

Note. NA = specifications that do not include all observations; SNAP = Supplemental Nutrition Assistance Program.

^a Log-transformed.

FIGURE 2—

Gaps in Synthetic and Treated Counties for (a) Total SNAP Redemption and (b) SNAP Redemption per Participant: Philadelphia, PA, and Neighboring Pennsylvania Counties, 2005–2019

Note. SNAP = Supplemental Nutrition Assistance Program. Dashed vertical line indicates the implementation of the Philadelphia tax in January 2017. Gap values illustrate the estimated difference between treated and synthetic counties. SNAP redemption (panel A) synthetic controls were fit using log-transformed outcome values, which were back-transformed to dollar values here. No appropriate synthetic control could be fit for log SNAP redemption in Philadelphia County (panel A), and no detectable treatment effect was found for Bucks County (panel B), so they are not presented here.

For Delaware County, all specifications also indicated a positive long-run effect, with specification 2 (first three quarters of pretreatment periods) selected based on the lowest RMSPE, though all specifications had similar RMSPE values; specification 5 was selected by the Dube and Zipperer criterion,³¹ though its outcomes and weightings were comparable to specification 2. Three covariates—the number of authorized stores, total SNAP benefits distributed, and number of SNAP participants—were marginally weighted in the model. Placebo testing produced a P value of .04, and confidence set plots also illustrated positive effects in all posttreatment periods.

For Montgomery County, all specifications also indicated a positive effect, and specification 3 (first half of pretreatment periods) was selected for the lowest RMSPE; specification 5 was selected by the Dube and Zipperer criterion,³¹ and produced similar though slightly larger gaps in posttreatment periods than specification 3. All covariates contributed at least marginal weights to the synthetic unit. Placebo testing produced a P value of .045, and confidence set plots showed a positive effect in all posttreatment periods.

Redemption Per Participant Outcome

For SNAP redemption amount per participant in Philadelphia, all specifications showed a negative long-run trend (Figure F, available as a supplement to the online version of this article at http://www.ajph.org), and 4 of 5 produced P values less than .05 via placebo testing (Figure G, available as a supplement to the online version of this article at http://www.ajph.org). Specification 5 had the smallest RMSPE (Table 1), with only the population covariate receiving a small weighting, reflected in the similarities of synthetic controls estimations among all 5 specifications (Table D, available as a supplement to the online version of this article at http://www.ajph.org). Specification 4 (odd pretreatment periods) was indicated by the selection criteria of Dube and Zipperer,³¹ though outcomes were similar to specification 5 (Figure F). Confidence sets for significant specifications illustrated a negative effect for most posttreatment periods (Figure H, available as a supplement to the online version of this article at http://www.ajph.org). Donor county weightings are reported in Table E (available as a supplement to the online version of this article at http://www.ajph.org).

All specifications for neighboring counties demonstrated a positive posttreatment trend. Specifications for Delaware County were deemed significant via placebo testing (P = .029 for all specifications in Delaware County). Specifications in Montgomery and Bucks counties did not yield significant P values through placebo testing (P = .054 for all specifications in Montgomery; specifications ranged from P = .13 to .28 in Bucks). In Delaware County, specification 3 had the lowest RMSPE, with SNAP benefit allocations, participant counts, and retailer counts included as covariates. In Montgomery County, specification 5 had the lowest RMSPE, with all covariates receiving at least a marginal weighting. Dube and Zipperer criteria³¹ indicated specification 5 for both counties. Confidence sets calculated for both Delaware and Montgomery counties illustrated a consistently positive trend across posttreatment periods.

DISCUSSION

Even amid broader macroeconomic changes, such as rates of SNAP participation, unemployment, and benefit allocations, implementation of the Philadelphia excise tax appears to have resulted in a geographical shift in SNAP spending. Though benefits cannot be directly traced from one county to another, synthetic control analyses for the 3 neighboring counties provides a high-level perspective on where these funds may have moved. Each neighbor saw detectable increases in SNAP redemption, with the largest increases observed in Montgomery County, which shares the longest border with Philadelphia of the 3 adjacent Pennsylvania counties. Similarly, per-participant SNAP redemption was shown to decrease in Philadelphia and increase in both Delaware and Montgomery counties. Here, it is reasonable to surmise that the observed changes in SNAP spending in Philadelphia’s neighboring counties is attributable to the implementation of the SSB tax. The mechanisms explaining these changes must be carefully considered.

Probably More Than Sugar-Sweetened Beverages

Earlier research has shown that decreases in Philadelphia SSB sales following the beverage tax were partially offset by increased sales in bordering counties, tempering expectations of broader public health benefits that would result from reduced SSB consumption.² This is consistent with other research from Oakland, California, which finds similar cross-border shopping in response to a SSB tax.³² The present study places these findings in a larger context. While media reports following the tax’s implementation have sensationalized cross-boundary “soda trips,” the magnitude of changes observed in this analysis suggest that SNAP participants may have migrated entire grocery shopping trips (rather than purchases of taxed goods alone) outside city boundaries. Furthermore, USDA estimates that, on average, SNAP households make 9.4 transactions per month, with an average transaction amount of $27.36.¹⁰ Existing research on food shopping among low-income Philadelphians adds to the plausibility of entire trips being migrated by suggesting that retailer proximity, while important, is far from the only factor that influences where to shop, which can include broader economic, logistical, and social considerations.^33,34

With an excise tax, applied at the distributor level, important questions exist about how much of the tax is passed through to customers via product prices, if any, and on which products. One could imagine an excise tax directly passed through via proportionately higher prices on taxed goods, or more indirectly spread among taxed and untaxed products. A recent study of tax implementation in Philadelphia supermarkets found the mean price per ounce of taxed beverages increased by .83 cents from 2016 to 2017, and sales of taxed beverages decreased in Philadelphia and increased in neighboring areas, suggesting that price pass-through did occur.² These higher prices, in turn, could have motivated price-sensitive shoppers to shop outside Philadelphia, especially in neighborhoods near the county border. SNAP is predominantly spent at supermarket retailers, typically with the largest shopping trip occurring at the beginning of the month as benefits are renewed, and involving some form of private vehicle use, even by those who are not vehicle owners, but who might share rides with friends or family.^15,33,34 Thus, if the price differences resulting from the tax were enough to motivate SNAP shoppers to shift these larger, vehicle-oriented trips across city borders, a sizeable effect might be expected.

Looking More Closely at Retailer Behavior

While higher prices on SSBs may have motivated shopping outside the city among low-income individuals, more explanation is needed. For instance, what role did retailer pricing and promotion strategies (on both taxed and untaxed items) play? Evidence from New York State reveals significant increases in in-store marketing of SSBs during periods of SNAP benefit issuance compared with other times of the month, with no attendant increases in marketing of low-calorie or unsweetened beverages,³⁵ illustrating how retailers can target SNAP shopping. Although additional research is needed to unpack the pricing and promotion dynamics that may have motivated SNAP shoppers to move entire shopping trips across city boundaries, it is likely that retailers signaled to consumers beyond pricing alone.

Public Health Implications

In light of these and other findings that suggest shoppers may have moved rather than reduced purchases of taxed beverages, alternative strategies for achieving public health benefits should be considered. At the consumer level, a direct consumer tax, such as sales taxes more commonly associated with alcohol or tobacco, could be set uniformly across city retailers, rather than the general increase in prices that result from distributor and retailer pass-through of an excise tax. Still, this does little to solve the problem of cross-border shopping to avoid the tax, and other research suggests that sales taxes are less salient to consumers than excise taxes.³⁶ Notably, purchases made with SNAP are exempt from sales taxes, so these efforts may do little to reduce SSB consumption among SNAP shoppers.

In terms of retailer-level alternatives, policies intended to reduce consumption of certain goods are unlikely to win allies from the business community. Complementary policies or programs that encourage substitution of goods (e.g., water or unsweetened beverages) or additional purchasing (e.g., “double-up” bonuses for fruit and vegetable purchases) are perhaps more palatable to retailers, and also create both push and pull dynamics to help encourage healthy behavior changes at the population level.

Strengths and Limitations

The synthetic control approach offers a data-driven method that, when carefully applied, helps rule out alternative explanations by matching and weighting donor units based on underlying relationships between variables in a data set. This study merges several discrete county-level, time-series data sets related to the allocation and spending of SNAP benefits. Thus, we can more accurately consider the supply-and-demand dynamics related to SNAP spending.

This study also had several limitations. Here, it was assumed that households did not choose to stop participating in SNAP on the basis of the implementation of the beverage tax. It is also critical to note that this study did not examine items purchased by SNAP shoppers. While we can make comparisons to existing literature on purchasing patterns among SNAP participants, we cannot conclude that the price of SSBs specifically was the cause of the observed shift in SNAP spending, nor do we estimate individual-level changes in spending. More detailed consumer panel data may reveal if particular items motivated the broader shifts in spending.

CONCLUSIONS

The 2017 implementation of an SSB excise tax in Philadelphia coincided with nationwide decreases in SNAP enrollment. I used a synthetic control approach to estimate the average treatment effect of the tax on Philadelphia as well as its immediate neighbors and observed significant treatment effects on SNAP shopping in Philadelphia’s neighboring Pennsylvania counties, potentially indicating a shift in SNAP purchasing away from Philadelphia. However, the magnitude of these effects suggests that they are not likely to reflect SSB sales alone. More research is needed to understand the mechanisms involving both retailer and consumer behavior that might explain these responses to the excise tax.