LACEA WORKING PAPER SERES. No....

THE EFFECTIVENESS OF SIN FOOD TAXES: EVIDENCE FROM MEXICO

Arturo Aguilar

Emilio Gutierrez

Enrique Seira

LATIN AMERICAN AND THE CARIBBEAN ECONOMIC ASSOCIATION

July 2018

The views expressed herein are those of the authors and do not necessarily reflect the views of the

Latin American and the Caribbean Economic Association. Research published in this series may

include views on policy, but LACEA takes no institutional policy positions.

LACEA working papers are circulated for discussion and comment purposes. Citation of such a

paper should account for its provisional character. A revised version may be available directly from

the author.

© 2018 by Arturo Aguilar, Emilio Gutierrez and Enrique Seira. All rights reserved. Short sections

of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full

credit, including © notice, is given to the source.

LACEA WORKING PAPER SERES. No. 0010

LACEA WORKING PAPER SERIES No. 0010 July 2018

The Effectiveness of Sin Food Taxes: Evidence from Mexico

Arturo Aguilar

Department of Economics, ITAM and IZA

[email protected]

Emilio Gutierrez

Department of Economics, ITAM

[email protected]

Enrique Seira

CIE, ITAM

[email protected]

ABSTRACT

We measure the effects of two federal taxes in Mexico aimed at reducing obesity by taxing sugary

drinks (SDs) and high caloric foods (HCFs). We use a weekly scanner panel dataset with more than

58,721 product barcodes. We find 6 percent lower SD consumption but no decrease in HCFs

consumption. We find substantial substitution towards non-taxed goods, smaller unit sizes, and in

the case of HCFs to higher calorie per gram foods. Overall, sugar consumption decreased 2 percent,

saturated fat and cholesterol increased 2 and 6 percent, but total calories remained unchanged.

Available evidence shows no effect on BMI.

JEL Classification: H20, I18, H31

Keywords: Obesity, Sin Taxes, High Caloric Foods.

ACKNOWLEDGEMENTS AND FINANCIAL DISCLOSURE

Kin Gutierrez and Marc Brooks provided outstanding research assistance. We thank Anna Aizer,

Jay Battacharya, John Cawley, Aranxta Colchero, Janet Currie, Sebastian Galiani, Paul Gertler,

Tadeja Gracner, Jonathan Gruber, Fernanda Marquez, Emily Oster, and seminar participants at

Banco de Mexico, ITAM and NEUDC for their very helpful comments. We are grateful to Kantar

World Panel for providing the data used in this paper and to ConMexico for their support to obtain

the data. ConMexico hired the authors of this paper to write a report on the impact of the taxes. The

authors’ right to exploit the dataset also for academic purposes was stipulated in that report’s

contract. ConMexico had no voice or veto regarding the content of this paper. All errors are ours.

2

I. Introduction

The growing concern that obesity represents a public problem has been followed by the de-

sign and implementation of different policies aimed at combatting it. Among these policies,

the introduction of special taxes on high caloric products has received considerable atten-

tion from policy makers and the academic community alike.1 The implementation of special

taxes on high caloric foods and beverages in Mexico in January 2014 is a remarkable case to

study for at least three reasons. First, the taxes were introduced at a national level. Second,

the magnitude of the imposed taxes was considerably larger than that of those introduced in

other contexts. And third, Mexico did not only focus on taxing sugary drinks, but also levied

a special tax on packaged foods with high caloric content. In this paper, we take advantage

of the sudden nature of the price change induced by these measures to study how households

respond to the new incentives and explore their substitution patterns in dimensions such as

product package size, quality and towards untaxed goods. We also study the taxes nutri-

tional implications. We show that despite the taxes achieving some of their objectives (e.g.

reducing soda consumption), they were ultimately unsuccessful at reducing overall caloric

consumption, at least in the short run.

Taxes to fight obesity are controversial and difficult to design, not only because of the

complexity introduced by substitution possibilities across different products, but also be-

cause food is a physiological need, and only overconsumption should be discouraged. These

difficulties and the shortage of empirical evidence may explain the wide disagreement among

economists about the effectiveness of food specific taxes to combat obesity; the evidence so

far is inconclusive. Cawley (2015), an expert on the economics of obesity, recently con-

cluded that: “The results of [existing price] studies are inconsistent; overall, there is weak

1Most saliently, the UK recently announced a tax on SDs. Some US states have imposed taxes on sodas.In 2011, Denmark implemented a tax on all foods with saturated fat content above 2.3 percent, only to revokeit a year later. Finland has imposed extra taxes on candies, ice cream, and soft drinks since 2011. In that sameyear, Hungary levied a tax on a range of products such as soft drinks, energy drinks, pre-packed sweetenedproducts, salty snacks, and condiments. In France, a tax on all beverages with added sugar was introduced in2012. Ireland, Japan, and the US have implemented or are considering implementing such taxes. The city ofBerkeley recently introduced similar taxes.

3

evidence that cheaper food (in terms of money price or time costs) contributes to obesity.”2

In the face of this uncertainty Gruber (2010) recommends starting cautiously and trying out

a sugary drink tax like the one we study.

Whether taxes to fight obesity work is a first order question for economists, not only be-

cause tens of thousands of people are dying yearly of its complications and because its care

costs billions of dollars per year, but also because their introduction is about corrective tax-

ation and consumer choice. Why then do we know so little about it and related policies?

According to Cawley (2015) “this literature is limited by a scarcity of powerful and valid

instruments for prices, time costs, and generally missing data on the intermediate step of

food consumption.” Our paper contributes with both: valid instrument for prices and very

detailed data on food consumption. We also have the benefit of using large tax-induced vari-

ation in prices for a broad base of foods; the implied tax rates introduced in Mexico are 2

to 3 times higher than those of U.S. States. To our understanding, this is the first paper to

measure the causal effect of a large and country-wide tax on both sugar-added drinks (SDs)

and high caloric dense foods (HCFs) on calorie consumption, and the substitution patters

they implied.

Our empirical strategy uses an event study design –which could also be thought of as a

regression discontinuity design– to exploit the sharp change in prices induced by the start

of the tax implementation in January 2014. The policy involved adding a one peso per liter

tax on SDs and an 8 percent VAT on HCFs.3 For our analysis, we rely on an extraordinarily

detailed proprietary home scanner panel dataset from Kantar World Panel, which contains

2Policy-makers promoting food-specific taxes cite correlations between food prices, soda consumption,and weight (e.g., Duffy et al. (2010), Malik et al. (2006)) as supporting evidence for these taxes. Philipsonand Posner (2008) argue that “Probably any feasible tax response would cost more to enforce than it would beworth in reducing the social costs of obesity” and Gruber (2010) argues that “a simple tax on calories coulddo more harm than good by deterring low-income people from getting enough nutrition.” On the other hand,Brownell and Frieden (2009) argue for an excise tax and claim that it would not only be effective, but alsothat “it is difficult to imagine producing behavior change of [the magnitude of a tax] through education alone”.Finally, Griffith and O’Connell (2010) discuss the complexity of such taxes but do not take a strong position foror against them due to the lack of empirical evidence. Other academics such as Bhattacharya and Sood (2011)question whether the very existence of negative externalities caused by obesity, and suggest there may not be aneed for corrective taxation at all.

3This tax has attracted an enormous amount of international attention. For instance, a Google query of“obesity tax Mexico” in February 2016 resulted in 490,000 hits.

4

information on the weekly purchase of 58,721 product-barcodes for 9,953 households in 92

cities across Mexico to study the impact of the taxes on prices, purchases, nutritional content

of households’ consumption bundles, and Body-Mass Index (BMI). In addition, we explore

how the design of these taxes may contribute to explaining their differential effectiveness.

By the nature of an event-study design, our analysis will focus on short run effects.

We find pass-through to consumer prices of 100 percent for the SD tax and 66 percent for

the HCF tax. While the tax was aimed at more obese households, we find that price incidence

almost does not vary for households with different BMIs or socioeconomic status (SES). We

then explore how these price changes translate into changes in the consumption of the taxed

goods. We find a 6 percent decrease in liters of SDs purchased, but no change in kilos

of HCFs purchased. We find that calories from SDs decreased by the same 6 percent, but

calories from taxed HCFs increased slightly. Calorie consumption out of non-taxed drinks

and foods increased as well. Regarding total nutrient intake, which is what the tax was

finally hoping to improve, we find the following: total calories consumed from all products

captured in our dataset did not change. We document a decrease in the consumption of 2%

sugar, but an increase of 1% in fat, 5% in sodium, and 6% in cholesterol, indicating a change

in the composition of the food basket away from sugar and towards fat.4 The results are

robust to a series of checks: accounting for the potential increase in prices associated with

the start of the new year, testing for the potential bias from household stockpiling behavior;

and estimating the effects on prices and quantities with aggregate industry level production

data instead of consumption the household level.

Finally, we document some mechanisms behind these effects, shedding light on the rela-

tive effectiveness of per unit versus ad-valorem taxes. We show that consumers substituted

towards untaxed goods and smaller package sizes of SDs, which experienced a lower per-

centage increase in price. In contrast, for HCFs, consumers substituted towards relatively

cheaper products with higher caloric content, as these experienced a favorable relative price

adjustment.

4Using a difference-in-differences strategy that compares Mexico with four countries in Central Americain the Appendix, we find no change in BMI as a result of the taxes two years after their implementation.

5

What is our contribution to the academic literature? Most studies estimate the demand

elasticity for SDs, not the effect of actual taxes. These studies can be classified into those

that use randomized experiments and those that use naturally occurring variation. The non-

experimental literature has the big problem of using raw price variation across markets to

estimate elasticities, conflating supply and demand. This may explain the wide range of

estimates. For instance, Andreyeva et al. (2010) survey more than 180 studies and find price

elasticities of sodas ranging from -0.33 to -1.24 and price elasticities of foods between -0.27

and -0.81. Other non-experimental studies estimate the direct relationship between weight

vs soda and food prices, and find small correlations (Powell et al. (2009) and Lakdawalla

and Zheng (2011)). The experimental strand has better identification. Their main problem

is that they only observe consumption at the experimental site (e.g. a hospital cafeteria) and

cannot take into account substitution across other sites. These studies also find wide ranging

and sometimes conflicting results.5 The closest studies to ours are Fletcher et al. (2010) and

Colchero et al. (2016). The former exploits across-state variation in soda taxes in the U.S.

to estimate a difference-in-differences model of state level taxes on liters of soda consumed,

calories consumed, and weight. They find that a 1 pp increase in the tax implies a reduction

of only 6 soda-calories per day. Colchero et al. (2016) analyze the SD tax introduced in

Mexico in 2014 (not the HCF tax) using a different data set and a different methodology

from ours. Like us, they find a 6% decrease in SDs consumption. They do not explore

substitution patterns nor the effect on nutritional content of household consumption bundles,

and they do not claim to estimate a causal effect as they rely on extrapolation.

The paper proceeds as follows. Section II reviews the obesity outlook in Mexico and

provides details on the taxes we evaluate. Section III describes our data sources. Section IV

outlines our empirical strategies. Section V presents the main results, while sections VI and

VII discuss mechanisms and robustness checks. The last section concludes.

5In an extensive meta-study of 23 papers on experimental results, Epstein et al. (2012) show that the priceelasticity of foods oscillates between -0.5 and -3.8 depending on the product and setting. Giesen et al. (2011)finds, as we report in our paper, that although changes in prices cause changes in consumption, this is notreflected in changes in calories. In contrast, Yang and Chiou (2010) study price changes in SDs and findprice-calories elasticities of -0.6.

6

II. Context

Obesity has been growing at an alarming rate in several countries (see our online appendix)

and Mexico leads this trend, with one of the fastest growth rates in its overweight population:

the percentage increased from 30 to 70 percent over the past 25 years Gracner (2015). This

has large health implications: obesity is a major cause of chronic diseases like diabetes and

hypertension (Kahn et al. (2006), Eckel et al. (2011)), which account for more than one third

of the costs of health care in Mexico.6 The cost is also high in terms of lives: the mortality

rate due to type 2 diabetes increased from 77.9 to 89.2 per 100,000 inhabitants between 2000

and 2007, making it the number one cause of death in the country (Sanchez-Barriga, 2010)

(see Olshansky et al. (2005) and Bhattacharya and Sood (2011) for US evidence).

As a response to this problem, on January 1st 2014, the Mexican government introduced

two taxes aimed at fighting obesity as part of Mexico’s Special Tax on Production and Ser-

vices (IEPS for its acronym in Spanish). One tax for SDs and another for HCFs.7 The 2014

IEPS reform established that SDs were to be taxed at one peso (6 cents USD) per liter of

SDs, which amounts to approximately 12 percent of their average price. SDs include sodas,

some nectars, concentrates with added sugar, and powdered drink mixes, among other drinks

described in detail in the appendix.8 Exemptions to taxed SDs include alcoholic beverages,

which are subject to another tax rate also defined in the IEPS law9 and dairy products. Drinks

sweetened with non-caloric sugar substitutes are also excluded. The reform also stipulates a

6According to Barquera et al. (2010), the direct and indirect costs of obesity in Mexico increased from35 billion pesos in 2000 to 67 billion in 2008, and are projected to be 150 billion by 2017. According to theMexican Ministry of Health, 3 out of 4 hospital beds are devoted to patients with obesity-related diseases.Similarly, Finkelstein et al. (2009) note that “The estimated annual medical cost of obesity in the US was $147billion in 2008; the medical costs for people who are obese were $1,429 higher than those of normal weight.”Also, the UN (2011) predicts that the cost of diabetes, hypertension, and related chronic diseases in low-andmiddle-income countries will surpass $7 trillion by 2030.

7IEPS was established in 1980, and it imposed specific tax rates on the acquisition or import of productssuch as alcoholic drinks, tobacco, and fuel and on activities such as lotteries and gambling. In fact, the 1980tax included SDs and was later modified in 1987 to exclude them. Before 2014, the government only taxedSDs as part of a 16 percent VAT, while HCFs were VAT exempt. The tax instituted in 2014 applies in additionto the VAT for SDs.

8See Section A3 in the appendix. The law establishes that sugars include monosaccharides, disaccharides,and polysaccharides employed as sugar substitutes with caloric content. Partial liters would be proportionatelytaxed. For example, a 355ml soda would be charged a tax of 35.5 Mexican cents.

9If a soda is mixed with an alcoholic beverage to form a cocktail and sold, then it is considered an alcoholicbeverage (and is taxed as such), and the SD tax does not apply.

7

separate tax on HCFs, which consist of an 8 percent tax on products with a caloric content

greater than or equal to 275 kilocalories per 100 grams, with some exceptions that were con-

sidered part of the basic Mexican diet (such as tortillas and cooking oil). Products subject to

this tax include snacks, candies, chocolate, pudding, marmalade, peanut butter, cereals, and

others. The caloric density is determined using the information on the product’s nutritional

content information label. If a product lacks this information it is automatically taxed. The

taxes are relatively large (about 3 times the average state-level soda tax in the US) and have

a fairly broad base: they apply to 54.9 percent of food products and 50.2 percent of drinks

in our data. Taxed products account for about 30 percent of expenditures captured in our

data, 56 percent of total calories from packaged goods, and 13.9 percent of total food expen-

ditures, according to our estimates using the National Expenditure Survey (ENIGH, 2014).

The introduction of these taxes resulted in $31,945 million pesos of tax revenues in 2014,

$18,280 from SDs and $13,666 from HCFs. This represents a significant fraction of IEPS.10

III. Data Sources

We exploit two main data sources. A very detailed proprietary dataset on household con-

sumption collected by Kantar World Panel (KWP), and data on caloric and nutritional con-

tent for each product from KWP’s dataset gathered by our own research team.

A. Household Barcode-level Consumption

Measuring consumption is difficult for many reasons. Comprehensive measures require long

questionnaires, which lead to respondent fatigue and non-compliance (Beegle et al., 2012);

they typically measure unit values instead of prices, and they rarely measure consumption at

a fine level of disaggregation both in terms of products and frequency of purchase. The ideal

data to measure the effect of taxes on consumption would be a household-level panel with

10In 2014, IEPS revenue was $124,061 million pesos, an increase of 51.7 percent with respect to2013. Source: http://finanzaspublicas.hacienda.gob.mx/es/Finanzas_Publicas/Estadisticas_Oportunas_de_Finanzas_Publicas, last consulted on March 25, 2016.

http://finanzaspublicas.hacienda.gob.mx/es/Finanzas_Publicas/Estadisticas_Oportunas_de_Finanzas_Publicas

http://finanzaspublicas.hacienda.gob.mx/es/Finanzas_Publicas/Estadisticas_Oportunas_de_Finanzas_Publicas

8

high frequency consumption data, with scanned product barcode-level information on quan-

tities and prices, where prices are gathered from ticket receipts to eliminate reporting errors.

This data will ideally include measures of BMI and other socioeconomic characteristics to

study effect heterogeneity. We have such a dataset from KWP.

KWP is an international company that operates in over 50 countries and has over 60 years

of experience. They specialize in the collection of high quality household consumption data

that they process to give advice and information to producers and retailers. In order to en-

sure a high quality panel, KWP recruits and gives incentives to households to participate,

selecting them to be a representative sample of households in the country. Our data covers

93 cities. Households participate in the panel for up to 4 years, during which they are re-

quired to keep ticket receipts for all purchases made in formal and informal establishments.

They are visited by surveyors weekly and all tickets collected are scanned and saved. For

each purchase, the scanned information allows KWP to record the week of the transaction,

barcode of the product, number of units purchased, price per unit, and the type and name of

the store in which it was bought. The barcode itself contains information about the type of

product (such as carbonated flavored beverages, coffee, snacks, soup, yogurt, etc.), brand,

package size (e.g., a 2 liter bottle or 355 ml. can of soda), flavor, regular/light, and main

ingredient (such as milk and non-milk based).11 In our main results, we exploit the infor-

mation in this dataset for all weeks in 2013 and 2014, restricting our inquiry to households

that were interviewed at least once in 2013, which amounts to 9,953 households in 93 cities

and 762,706 household-week observations.12 Our data has 9,017 SD barcodes and 13,464

barcodes on food. KWP also implements a yearly questionnaire that captures a set of so-

cioeconomic and demographic characteristics that include household assets (which we use

to classify households by SES) and the male and female household head’s age, weight, and

height, which we use to calculate BMI. 13. For SES categories, we group households into

11Figures A8 and A9 in the online appendix show examples of the level of detail available in the data.12Of all households in our sample, 11.4 percent report their purchases every week in the 2013-2014 period,

27 percent report for all except three weeks at most, and 71.4 percent report purchases on more than half of allweeks in 2013 and 2014.

13BMI measures are self-reported, which may impose limitations on our analysis. Nonetheless, data checkssuggest they are good measures of individuals’ body weight. Table I shows that BMI as measured in the KWP

9

three groups (ABC+, CD+, and DE), where ABC+ refers to the highest SES level. Using

BMI we classify households as normal weight if the reported BMI of the female head is

lower than 25, overweight if it lies between 25 and 30, and obese if it is above 30.14

Table I shows that KWP matches important moments of Mexico’s 2014 Income and Ex-

penditure Survey and the 2012 Health and Nutrition Survey (ENIGH and ENSANUT for

their acronyms in Spanish), which are the main consumption and health surveys in the coun-

try. The average age of household heads is 46 years and 79 percent are male. It also matches

closely asset holdings, from which KWP constructs SES categories. Finally, BMIs are simi-

lar even though the ENSANUT BMIs are measured rather than self-reported.

While extraordinarily detailed and well-suited for the study of the impact of the taxes in

the Mexican context, the data has some limitations. First, it measures purchases, which do

not necessarily correspond perfectly to consumption. However, we show later that there is no

indication of changes in households’ inventories. Since our identification strategy measures

changes in quantities, as long as waste is unaffected by the tax, our estimates should approx-

imate changes in consumption well. Second, our data has information only on processed

packaged foods (it does not include information on the purchase of fresh foods like fruits

and meat). Table I shows that, according to ENIGH, the average Mexican household’s food

expenditures amount to 3,256 pesos, while, on average, KWP only captures expenditures in

food products of 1,417 pesos. However, ENIGH captures fewer expenditures on the goods

included in KWP. Importantly, taxes are levied on packaged products (included in KWP), not

on fresh food. As long as the degree of substitutability and complementarity between pack-

aged and fresh foods is low, our estimates of the change in consumption of nutrients from

the set of products included in KWP will not be biased. In addition, given that the existing

literature argues that the availability of packaged foods and drinks is likely one of the most

survey is very close in levels with the Mexican Health and Nutrition Survey (ENSANUT), which measuresBMI in situ by professional nurses. The online appendix provides further evidence that BMI as measured byKWP is consistent with the rankings of BMI across countries and shows reasonable densities, time trends,and relationships with structural variables such as age. Self-reported BMI is the norm more than the exception.More than half of the countries that report BMI measures to the World Health Organization rely on self-reportedBMIs.

14We find a large within household correlation above 0.5 for BMI of male and female heads of the samehousehold. Because strong positive correlations of BMI across members of a household are typical, we willtake the liberty of referring to household BMI even if we are using the female household head BMI only.

10

important forces behind the increase in obesity (Cutler et al. (2003), Brownell and Frieden

(2009)), our focus on these goods is not out of place. Finally, KWP records only in-home

consumption. According to ENIGH 2014, 20 percent of all food expenditures correspond to

out-of-home consumption in Mexico. Importantly, to explore if our results are biased due

to the exclusion of out-of-home consumption from our data, we compare our estimates of

the impact of the tax on SD with those obtained using total soda production from national

statistics and find similar results.15

B. Data on Caloric and Nutrient Content

For the taxes to be an effective tool to combat obesity they need to affect caloric intake.

Our main analysis explores if this was the case and provides information on the change in

the purchase of other nutrients, namely sugar, carbohydrates, fat, saturated fat, cholesterol,

and sodium. However, KWP does not collect information about the nutritional content of

products. An additional contribution of this paper is the systematic collection of nutrient

information for a large share of barcodes in KWP. This was possible because the majority

of products in our data are packaged and display nutritional content labels as mandated by

law. Our enumerator team collected information for barcodes representing 71.7 percent of

all purchase events, which corresponds to 68 percent of expenditures in our data. Because

coding the nutritional content required direct observation of products’ label in supermarkets

and convenience stores, the information for products only available in specific regions had

to be imputed based on a careful procedure described in detail in the online appendix. We

are aware that our imputation procedure may introduce measurement error in one of our

main outcomes of interest. We have performed several tests of the imputation quality and

concluded that this problem is small.16 The fact that we find statistically significant effects

15Dubois et al. (2014) is probably the closest paper in terms of data and faces the same limitations. SectionA11.C in the online appendix explores in greater detail the potential bias in our estimates driven by partialobservation of consumption and concludes that we are not underestimating the tax impact. Table A9 of theonline appendix presents additional descriptive statistics for the KWP dataset: 68.9 percent of total expendituresin KWP are devoted to food and drinks, while 14.2 and 13.5 percent of total expenditures correspond to taxedHCF and SD respectively.

16One such check randomly deletes the nutritional information for 10 percent of the products captured anduses the remaining 90 percent of captured products to perform our imputation procedure. We compare the real

11

for some nutrients suggests that noise is not too large.

IV. Empirical Strategy

Identifying the effects of the introduction of a tax reform like the one studied in this paper is

not an easy task for three reasons. The first is that the taxes were not a universal tax on food

and sugar, but were instead aimed at thousands of specific high calorie products. To analyze

the taxes’ impact, we first have to identify which of the barcodes in KWP were subject to

which of the taxes. For this, we relied on the content of the IEPS law and detailed statements

provided by Mexico’s tax authority.

A second challenge is that the impact of the tax may differ across products, and complex

substitution patterns may operate across to untaxed but also within taxed barcodes. We

are not so much concerned with the product-to-product substitution but with the aggregate

implications for consumption and nutrition. Substitution is important as it could undo the

intended effects of the tax. We focus on substitution across broad categories of foods, taxed

and untaxed. For quantities we focus on total calories of taxed SDs and HCFs, as well as

calories from taxed and the untaxed groups. For prices we constructed a Laspeyres-type price

index. Conducting the analysis at this level of agregation has several advantages. First, it

makes the size of the dataset more manageable. Second, it automatically weights the results

by the appropriate consumption quantities, making our results representative for Mexico.

Third, it allows us to focus on salient substitutions, namely between taxed and untaxed goods

in a parsimonious way.17 Fourth, we don’t throw away all interesting heterogeneity, we

present results separately by BMI and socio economic status of the households to look at

quantity effectiveness and price incidence. Finally, we are ultimately concerned about total

calorie consumption and are therefore forced to aggregate quantities. We did run regressions

values to the imputed values and show that the imputation error is tiny. The interquartile range of the imputationerror of calories for SDs is 6.3cal/100gr, with a mean close to zero (see Appendix). This compares favorablywith the interquartile range of calories 40.2cal/100gr and a mean of 97.4cal/100gr of the raw data for SDs.

17We could look for substitutions across all products in our sample but this quickly runs into a dimensionalityproblem as we have thousands of products; it also leads to a statistical problem of multiple testing. Anotheravenue would be to choose a particular subset of products like milk and water, but this would be ad-hoc.

12

at the barcode level with barcode fixed effects. This works well for prices and results are

similar, but note that we have to use weights proportional to household consumption to get

meaningful results. For quantities we are forced to aggregate across barcodes anyways. We

prefer to be transparent about our weighting rather than relying on OLS to do it for us.

A third challenge is to isolate exogenous variation in taxes. Most of the literature uses

comparisons across markets or across persons, conflating supply and demand. Furthermore,

the literature uses variation in prices, not in taxes, and therefore cannot measure the pass-

through of taxes to prices or the effect of tax reforms for that matter.18 In order to identify

causal effects, our main empirical strategy exploits the sharp price variation at the moment

at which the taxes were enacted together with high-frequency data in an event study design.

A. Construction of the Main Variables

To study the broad substitution patterns we decided to aggregate the information on individ-

ual barcodes at the household-week level, and then agregate at taxed and non taxed categories

for SDs and HCFs separately. We had to be careful to identify which barcodes were taxed

and which weights should be used in the aggregation.

Tax variable. To identify the barcodes in the KWP dataset that were subject to each tax,

we rely on the IEPS law and detailed statements by Mexico’s tax authority.19 In general,

drinks that have added sugar are classified as subject to the SD tax (1 peso per liter) and food

products not classified as basic food staples that have more than 275 kilocalories per 100

grams are classified as subject to the HCF tax (8 percent ad-valorem). Ad valorem taxes are

proportional to price so, all else constant (like elasticities), one may expect higher absolute

price increases for more expensive (higher quality) goods within the HCF category.

Quantity variables. We aggregate detailed weekly household and barcode level informa-

tion in the KWP data as follows. For each household i in week t of year y, we calculate

18For expositional purposes, and to show that our data is consistent with this literature, Figure A7 confirmstwo of the stylized (correlational) facts in the existing literature: the cross-sectional relationship between SDprices and liters consumed and the cross-sectional relationship between SD price and BMI are negative. This,however, is not necessarily a causal relationship.

19See section A4 of the online appendix for details on the construction of the tax variable.

13

a measure of calories of taxed SD consumed (CSDity ) by first multiplying the units of bar-

code b consumed, uSDbity, by their calories, cb, and then summing across b for all products

and package sizes in the SD category (identified as the elements in the set SSD). Formally,

CSDity =

∑b∈SSD uSDbitycb. We apply the analogous aggregation for liters (LSD

ity ). We do this

separately for non-SDs, HCF and for all food products consumption.

Price variables. Creating an aggregate measure of price per liter requires a bit more care.

We cannot use current expenditure per liter, as expenditure involves quantities purchased that

the 2014 taxes may have affected. We therefore construct a quantity-weighted average price

index using weights from 2013 consumption that are then independent of any tax effect. We

use a Laspeyres-type index keeping the average 2013 basket fixed for each household. That

is, each household has a weight for each barcode that is equal to the quantity (in liters or

kilos) it bought a year before the tax. This is precisely the weighting we want as it represents

the quantity consumed in the economy. We compute the cost of these baskets each week

with quantity weights fixed but varying prices. Formally, the price index for SDs for each

household i in week t year y (P SDity ) is calculated as follows:

(1) P SDity =

∑b∈SSD

(LSDbit,2013

LSDi,2013

)pSDbity

Where LSDi,2013 =

∑52t=1 L

SDit,2013 measures the total liters of SD consumed by household

i in 2013, and pSDbity is the price of barcode b that household i incurs in week t of year

y = {2013, 2014} . Note that pSDbity is not always observed, as household i may not have

purchased that particular item in that particular week. In such cases, prices are imputed

following a sequential procedure. We first compute the average price for each item in the re-

spective week, city, and store (i.e., cell). When barcode-level prices are still not observed for

a given barcode in a given week, city, and store, we proceed to a higher level of aggregation.

The details are provided in section A5 of the online appendix.20 We follow an analogous

20Several facts make us confident that these imputations are a good approximation of true prices. First, ifwe estimate a barcode-level regression similar to equation 2 but at the barcode level with week, household,and barcode fixed effects, it turns out that the estimated increase in SD and HCF prices are 12 and 6.5 percent,

14

procedure to calculate an index of price per kilogram of HCF and a price per calorie for all

products.

Although not the main focus of the paper, we will estimate the price effects of the tax.

One could do this by using regressions at the barcode level with barcode fixed effects. To

be representative of the economy, we would need to weight each observation by the quantity

consumed of each barcode. We would also need to decide the level of error clustering.

When we do this, regressions give the approximate same answer as what we get from using

the index described here.

B. Identification Strategy

Our empirical strategy exploits the sharp change in prices observed on the date when the

taxes came into effect. It can be viewed as an event study or as a regression discontinuity

design that uses time as a running variable, with the date after which the taxes were intro-

duced (January 1st, 2014) as the treatment of interest. The main identification assumption

is that absent the tax change, prices and quantities of food consumed would have followed a

smooth trend after controlling for seasonality, and that any sharp change in prices in January

2014 can be attributed to the taxes. Many recent papers have profitably used this empirical

design (e.g. Byker (2016), Chetty et al. (2014a), Chetty et al. (2014b)), but it has a long

tradition in economics and finance (MacKinlay (1997)).

Figure I illustrates that these new taxes affected the final prices of the taxed goods but

not those of untaxed goods. Since quantities are kept constant at their 2013 level, the jump

observed in January 2014 is only attributable to changes in prices, not changes in consumer

behavior. Panel A of Figure I displays a very sharp increase in the cost of the taxable con-

sumption bundle just when the tax came into effect. Panel B, on the contrary, shows no such

increase for the cost of the non-taxable consumption bundle, which suggests that the discon-

respectively, which are very close to the price estimates presented in Section V. Second, in each step of theprocedure, imputation-cell dummies explain more than 95 percent of the price variation, which means thatthe procedure of imputing cell means captures almost all variation in prices. Finally, since our strategy is tocompare across time and we follow the same methodology for each year, the imputation should not have animportant influence on our estimates.

15

tinuous jump we observe in Panel A is due to the tax. Panels C and D distinguish between

taxed SDs and HCFs. It plots the price per liter of SDs and the price per kilogram from HCFs

as defined in equation 1 for each week of 2013 and 2014. It is evident that the price per liter

of SDs is very stable, around 8 pesos, for most of 2013. There is a small increase in July

2013, which is later dwarfed by the increase to 9 pesos in the first week of January 2014,

which is strikingly close to the size of the tax change. There is also a discontinuous jump for

HCFs.21 The sharp jumps observed and the fact that they are only present for taxable goods,

strongly suggest that the taxes caused the price changes.

Two additional pieces of evidence support a causal interpretation of this price jump. First,

the appendix (Table A10) shows that the price only increased in Mexico and not in neigh-

boring Central American countries, which suggest it is not driven by international cost of

commodities like sugar or others. Second, Figure A14 and Table A11 in the appendix show

that for HCF the 2014 price jumped at exactly the 275 calories per 100 grams threshold.

Our main specification is laid out in equation 2, where i, t, and y denote household, week

of the year, and year, respectively, I(y ≥ 2014) is an indicator for all observations in 2014

and afterwards.

(2) ln(P SDity ) = αit + βI(y ≥ 2014) + f(ty) + εity

We include household-week of the year fixed effects (αit) as well as a quadratic time

trend f(ty). Thus, our estimation controls for seasonality and relies on within-household

variation in prices close to January 1st, 2014 to identify the effect of the taxes. For ease of

interpretation, we transform all dependent variables x by taking the log of 1 + x, although

our results hold for the untransformed variables as well. The coefficient of interest is β,

which can be interpreted approximately as the percent change in the outcome due to the

introduction of the taxes. We cluster standard errors at the week level.

21Section A11 of the online appendix shows the analogous graphs for a broader set of price indices, includingan extra year of data, in the analysis. No sharp price change is observed in 2014 for the price per liter of non-SDs and price per calorie for non-taxed-foods.

16

V. The Impact of the Taxes

In this section, we first explore the incidence of taxes on prices. Then we focus on quantities

of SDs, HCFs, total calories, and other nutrients. We present results separately by BMI and

SES categories because tax incidence is important. We would like the tax to target more

obese consumers, and to not fall hard on the lower income segments.

A. Sugary Drinks

Table II shows the effect on SDs. Panel A uses the log of our price index as the dependent

variable, and Panel B uses the log of calories as the outcome of interest. Column 1 presents

the estimate for the full sample, while columns 2 through 7 restrict the sample according to

our BMI and SES categories. Panel A shows a precise average price increases of 14 percent.

The results are in line with those in Grogger (2015), who measures the impact of the SD

tax on prices using public data obtained from INEGI and finds slightly more than full tax

pass-through of about $1.2 to $1.5 pesos. Columns 2 to 7 estimate regressions separately by

BMI and SES. Interestingly, SD price changes do not vary considerably across groups. This

means that the actual burden of the SD tax falls on normal weight and low SES people to the

same extent as the obese and high SES.

Panel B shows that calories from SDs decreased by 6.3 percent for the full sample. This

decrease is stronger for richer segments of the population (7.8 percent) and slightly larger

(6.8 percent) for households with obesity. The appendix shows that decreases in liters of

SDs are about the same as decreases in calories. Overall results for prices and quantities

of SDs are similar to other recent papers that have looked at this with different data and

different methods (e.g. Colchero et al. (2016)). Pointing this out is important to note since

KWP data is not public and replication is not easy, and since the tax has been politically very

controversial and ideologically charged, pinning industry against government health centers

and tax authorities.

17

B. High Caloric Foods

Table III presents the results for HCF. Panel A uses the log of the price per kilogram of HCF

as the dependent variable, while Panel B uses the log of calories from HCF purchased as

outcomes of interest (the appendix also shows results for Kilos of HCF). All columns present

the results for the same subsamples as those described in Table II. For the full sample, we

find an increase of 5.3 percent in HCF prices. Given that the tax was an 8 percent ad-valorem

tax, this constitutes about a 66 percent pass-through. Interestingly, the price increase is 0.8

percentage points larger for low income households, while there is no differential effect by

BMI. In spite of the increase in price, we find no change in calories (or kilograms) consumed

on average for the full sample.

Interestingly, the tax on HCFs was less effective in decreasing calories than the SD tax.

As we discuss later, this has to do with the fact that the HCF ad-valorem tax induced coun-

terproductive relative price-per-calorie changes. Package sizes that contain more calories

exhibited lower price increases.

C. Total Calories and Other Nutrients

Taxes like the one we study here aim at decreasing caloric intake and obesity. The event

study methodology we use is not well suited to study long term outcomes like obesity, but

can be used to measure changes in calorie consumption in the shorter run. We only know of

one study that has attempted to measure caloric consumption and taxes. Fletcher et al. (2010)

finds that soda taxes in the U.S. led to a decrease in daily soda consumption of 18 grams,

associated with 6 fewer calories per day, but no decrease in calories once other substitute

foods are taken into account. There are several differences in the taxes we study. First, the

taxes we study are about 3 times larger, and since ours is national and applies to a much

broader set of goods, one may expect less substitution. We study some specific substitution

patterns in the next section. An advantage of focusing on total calories from all foods and

drinks is that it parsimoniously takes into account potentially complex substitution patterns,

and that it is the actual intermediate outcome we care most about.

18

Table IV presents our estimates of the total change in consumption of calories and other

nutrients attributable to the introduction of the taxes. Panel A focuses on total calories, and

panel B shows the impact on other nutrients. All estimations in panel B are run with the

full sample, but each column uses the log of a different nutrient as the dependent variable:

sugar, carbohydrates, fat, saturated fat, cholesterol, and sodium. Panel A shows that there

was a small positive but statistically insignificant increase of 1.6 percent on total calories

purchased for the full sample. Furthermore, for those households with the lowest level of

BMI and income level, we actually find positive and statistically significant increases of 4.2

and 2.4 percent, respectively. We do not find that the obese are consuming less calories.

Panel B suggests that the tax caused a change in food basket composition from sugars to

fats. We find a 2 percent statistically significant reduction in total sugar and increases of 6.4,

4.9, and 2 percent in cholesterol, sodium, and saturated fat, respectively.

Could measurement error explain why we do not detect a decrease in calories? We don’t

think so. The fact that we were able to detect changes in some nutrients discredits measure-

ment error as the explanation for the null effect on total calories. Note also that we did find

calorie reductions for SDs alone of the same magnitude as the decrease in liters consumed,

but this decrease is overturned once we consider untaxed drinks and taxed and untaxed food.

Could it be that calories do decrease but only for products outside the KWP dataset in a

way that we do not capture? We think this is unlikely. For starters, our data contains most

goods to which the new taxes applied. For substitution patterns to overturn the null effect,

households would have had to decrease their consumption of untaxed goods not included in

our dataset in a significant way. We explore this possibility in Section A11.C of the online

appendix by ordering households based on the (estimated) share of total food consumption

we observe. Based on this analysis, we conclude that, if anything, we may be biased against

finding a zero or negative effect on calories, since we find that the more expenditure we cap-

ture the larger is the positive effect we estimate. Secondly, as we will show below we find

similar results when we use production instead of consumption as the dependent variable.22

22Since international trade in sodas is negligible, we think national production approximates total nationalconsumption well, even if it is outside the household.

19

A third and final piece of evidence for the null decrease in total calories comes from estimat-

ing a zero change in BMI. Section A9 of the online appendix uses data from other Central

American countries where the tax was not implemented as a control group and performs a

difference-in-differences estimation in BMI. We find no evidence of a decrease in BMI or

weight, neither in the short run (2014) nor in the medium run (2015).23

We find the following simple back of the envelope calculation persuasive; it also provides

a benchmark to think about the potential effects and gives credibility to our estimates. Stern

et al. (2014) are one of the main proponents of the SDs tax and we use their numbers.

Using the main national health survey that covers in-home and out-of-home consumption,

they find that children (1 to 4 years old) consume only 20 calories from soda per day, while

they consume close to 100 from milk plus 75 from flavored milk and 25 from juice drinks.

Adults consume less than 100 daily calories from sodas. If we take the 6 percent reduction in

soda consumption found by Mexico’s National Health Institute team (Colchero et al. (2016))

and us, and 100 daily calories as baseline, we would conclude that the SD tax achieved a

reduction of only 6 calories per day, and that is before taking substitution into account!24

So far, our results provide strong evidence that the taxes did not decrease consumption

of calories from HCF, and although they did decrease consumption of SDs, the result on

total calories suggests that this was compensated by substitutions that led to an increase in

caloric intake from other products. In the absence of a reduction in caloric intake, the taxes

introduced are unlikely to have the desired effect of reducing obesity. Hill et al. (2003)

estimate that to prevent weight gains among 90 percent of the population, caloric intake has

to decrease by 100 calories per day, while Butte and Ellis (2003) estimate that such decreases

should be 200 calories per day at least. A word of caution is in order. Even if calories are

unchanged, the taxes could have an impact on health as there is a shift from sugar to fats.

Medical research would have to say if such a shift and its magnitude is the desired one.

23We put less emphasis on this result in the paper as one may argue that more than 2 years are needed forchanges in weight to materialize, and that we have too few observational units (countries).

24Another way to benchmark this is again to cite Stern et al. (2014). They say that “From 1999 to 2012,total daily energy from beverages increased among children aged 5-11 (+45.3 kcal), females aged 12-19 (+57.3kcal), and adult females aged 20-49 (+96.4 kcal).” The before-substitution effect of the tax is about 10% of therecent increase in caloric intake of beverages.

20

VI. Mechanisms and Unexpected Consequences

This far, we have shown that the introduction of the taxes did not have the desired impact

on caloric intake. In this section, we explore in detail the potential mechanisms behind this

result. Specifically, we will focus on substitution.

A. Substitution across product types

In the previous section, we described two main direct results: the effect of the two taxes

on the consumption of taxed products, and the effect of the tax on calorie consumption. The

latter effect captures the net effect of all mechanisms, like complementarities and substitution

beteween products. Here we focus on the task of understanding some of those indirect effects

and mechanisms in a parsimonious way. We group the data into four categories: taxed

and untaxed food and drinks. We use specification 2 to estimate separately the effects on

quantities for each of these groups.

Both groups of solid food products show increases in caloric intake. Table V shows that

calories Non Taxed Food (NTFs) exhibit significant increases of 249 calories per household

per week. Meanwhile, calories from taxed HCFs exhibit non-significant increases of 20

calories per household per week. As for drinks, calories from SDs are reduced in the order of

147 calories per household per week. However, untaxed drinks counterweight this reduction

by displaying a 141 calories per household per week increase. On aggregate, these changes

imply an increase of 263 calories per household per week. Substitution towards non-taxed

products, but also within HCFs, limits the effectiveness of these taxes. For completeness

Figure II shows event study graphs of calories for the different groups.

B. Substitution within products

Why was the tax on sugary drinks (moderately) effective, while the tax on high caloric

foods was not? It turns out that the SD and HCF taxes had opposite effects on the relative

21

price of calories.25 First, we found out that the price changes induced by the taxes varied

considerably across barcodes in both the SD and HCF categories (see Figure A19). The

percent quantity variation across years is also very heterogeneous across barcodes26 Second,

we found that while SDs barcodes27 with more calories experimented the larger percent

increases in prices, while the opposite happened for HCFs, as shown in Panels (a) and (b) of

Figure III. This meant that SD barcodes with more calories experienced the largest decline in

consumption, Panel (c); while within HCFs, consumption shifted toward barcodes of larger

caloric content, Panel (d). We believe these patterns of relative price adjustments contribute

to the explanation of the differential effectiveness of the SD vs the HCF taxes.28

Table VI presents evidence exploring the substitution patterns within categories in order

to shed further light. Column 1 studies if consumers switch to cheaper calories. We use

specification equation 2 with the log of the 2013 price per liter of the SDs purchased by each

household each week as the dependent variable; that is, we compute how much the household

would have paid for its current 2014 SD bundle if barcode prices were at their average 2013

price. Column 2 uses as dependent variable the average liters per unit purchased as a measure

of package size. Column 3 simply uses the log of total units purchased as the dependent

variable.29 Analogously, Column 4 uses the log of the 2013 price per 100 calories given

each household’s current consumption basket of HCF in each week. Columns 5 and 6 use

the log of total calories per unit and the log of total units purchased for HCFs as dependent

variables, respectively.

The results show that the decrease in liters of SD bought can be decomposed into a 2.4

25We understand that it is difficult to isolate the effect of the tax on sodas from that of the tax on caloricfoods because they were enacted at the same time and because of substitutability and complementarity betweenthe two sets of goods. This section proceeds under the assumption that the direct effect of the tax dominatesthe indirect effect that the other tax might have.

26As expected with negatively sloped demands and exogenous variation in prices, there is a negative corre-lation for SD and HCFs between the change in price and the change in quantities across barcodes.

27Recall that among the variables that define barcodes are the sizes of products. In Figure A22 we documentthat larger soda package sizes experienced larger price increases.

28The appendix studies in more detail substitution within the SDs using and AIDS model, and shows thatconsumers switched towards buying smaller package sizes. It also estimates structural elasticites and simulatesalternative taxes, showing that lower tax rates could achieve the same revenue by taxing sodas with of largervolumes of liters.

29By unit we mean, for instance, a 255 ml can of soda or a 1L bottle of soda. Multiplying units by liters perunit we get total liters, so columns 2 and 3 can be viewed as a decomposition of liters.

22

percent decrease in units bought and a 4.7 percent decrease in the liters per unit purchased,

i.e. more than half of the decrease in the consumption of liters of sugary drinks is explained

by households demand shift to smaller package sizes. This switch towards smaller package

sizes is also reflected in the increase of 1.3 percent in the counterfactual expense of Column

1, since smaller units are pricier per liter. In contrast, for high caloric foods, individuals con-

sumption shifted towards ex-ante relatively cheaper products, with a higher caloric content:

there is a 2 percent reduction in units bought, but the units bought have almost 1 percent

more calories per unit. This may happen, as consumers are buying what previously would

have been 1.3 percent cheaper calories.30 The HCF tax may have been counterproductive.

It distorted purchase decisions by pushing consumers toward cheaper and probably lower

quality goods while promoting the acquisition of higher caloric content units. These results

show how complex it is to design taxes when the objective is to reduce the intake of specific

nutrients. Taxing nutrients directly may be more effective.

The appendix contains a more involved analysis estimating structural elasticities between

SDs of different sizes with an AIDS model. Using the model we simulate revenue equivalent

taxes that achieve larger decreases in SD liters consumed. It also documents that there is

a positive correlation between BMI and SD unit size, suggesting that taxing larger package

sizes may achieve better tax targeting.

VII. Robustness

A. Anticipatory effects

A crucial assumption of our methodology is that consumers and producers reacted to the

taxes on the date at which they came into effect. In our estimation, we assume that inventories

stayed constant before/after the tax. If consumers anticipated higher prices in 2013, then

we would see inventories increasing right before the tax was imposed. In such a case, our

previous estimation we would be overestimating the effect of the tax, as we would observe

30Table A14 shows that this holds within product categories as well. In particular, the price per calorie ofthe 2013 basket of Snacks, Bread, Cookies, Crackers, Cereal, Spreadbles decreases by 11.5, 3.3, 7.5, 4.5, 0.4,and 0.8 percent, respectively.

23

higher than normal purchases before the implementation of the tax and fewer purchases in

the short term after its implementation, i.e. a larger quantity decline. Note that if this is the

case, it would go against our argument that changes in calories were minuscule.

We perform three checks on whether this channel is operational. First, following the lit-

erature in marketing, Section A11.A of the online appendix explores if (a) there were more

purchase expense of SDs and HCFs in December 2013 vis-a-vis December 2012 and Decem-

ber 2014; (b) whether the probability of purchasing SDs or HCFs was lower in January 2014

than in January 2012 or 2013; and (c) how many weeks a household went without purchases

of SD or HCF in January 2014 and if this number is greater than that of the corresponding

January in 2012 and 2013. We cannot reject the null of no difference in each case, which

strongly suggest that inventories behaved similarly in 2014 and other years.

Second, we redo our estimates by excluding one, two, and three weeks before and after

the introduction of the tax. The results in Table VII show the results of these estimations. As

can be seen, none of the estimated changes in dependent variables vary significantly when

excluding one, two, or three weeks around the introduction of the new tax, suggesting that

our estimates are not being driven by households’ short-term stockpiling behavior strategies.

In a third exercise we conduct a placebo test following the logic of a Fisher exact test in

the Appendix. We estimate our main regressions with our main specification, pooling the

years 2012, 2013, and 2014, except that we construct different placebo tax dummies taking

the value of one on each of the 34 months between February 2012 and November 2014 and

estimate the “treatment effects” of these placebos. We then rank the estimated β′ts and find

out where our true tax effect lies in the distribution of these placebo estimates. We find that

January 2014 is an atypical month for price increases and for the decrease in liters of SD, but

not for decreases of HCF kilos or total calories, consistent with our previous results. We do

not find that months previous to the tax are atypical, which suggests that anticipatory effects

are not important.

24

B. More robustness tests

Our main identification assumption has been that once we control for week dummies and

smooth time trends, the sharp increase in prices in January is due to the new taxes. We

argued that (a) the sharpness of the increase, (b) the fact that the magnitude was very close

to the tax and (c) present only for taxed foods, together with the (d) RD evidence across

the caloric density threshold provide strong evidence for the validity of this assumption. We

now run two additional checks: comparing our estimates with those obtained from four other

Central American countries of the KWP, and comparing two different periods in Mexico.

As placebo checks, Table A10 shows the results of our main specification estimated sepa-

rately for Costa Rica, Guatemala, Panama, and El Salvador. Because we do not have nutri-

tional information for products in these countries, we only use the log of the price per liter of

SDs (Panel A) and the log of the cost of households’ consumption baskets in 2013 (Panel B)

as dependent variables. Compared to the coefficient for Mexico, the coefficient on price is

negative and 20 times smaller for Guatemala, positive and 4 times smaller for Panama, and

negative and 7 times smaller for El Salvador.

Seasonality or beginning of the year effects may be different in Mexico than in Central

America however. Instead of comparing Mexico to other countries, one could compare

Mexico from 2013 to 2014 vs. Mexico in 2012 to 2013. Regression 3 augments our main

specification adding data from 2012 and by including variable Y eary which increases by 1

unit each year. This variable captures the average beginning of year change in price each

year, while our tax variable I(y ≥ 2014) measures the change in price only in 2014. By

including them together, I(y ≥ 2014) captures the extra increase in prices at the beginning

of 2014 above and beyond the beginning of year effect of 2013. Specifically, we run the

following specification:

(3) ln(P SDity ) = αit + βI(y ≥ 2014) + δY eary + f(ty) + εity

Table VIII presents the estimates of δ and β. Panel A focuses on our price indices of liters

25

of SD (Column 1), kilograms of HCF (Column 2), and the price per 100 calories from all

barcodes (Column 3). Panel B uses total liters of SD (Column 4), total kilograms of HCF

(Column 5), and total calories (Column 6) as dependent variables. Results are very close

in magnitude to those presented in Section V, meaning that end of the year effects are not

driving our results.

C. Industry level production

A final potential caveat we consider is that of in-home vs. out-of-home consumption. As

described in Section III, KWP only measures the former, so our results apply only to home

consumption. How problematic is this? According to the ENIGH, about 20 percent of

food expenditure corresponds to out-of-home consumption. By this measure the out-of-

home consumption elasticity would have to be big for us to reach different conclusions. We

already have some evidence suggesting that this is not the case: our estimates indicate that

BMI did not change significantly as a result of the tax, implying that calories must not have

been reduced much. This section provides more evidence.

We analyze changes in total production of sodas in Mexico as a function of the tax. Given

that less than 2 percent of sodas sold in Mexico are imported/exported, that neither exports

nor imports of sodas changed with the introduction of the taxes31, and that inventories should

not vary considerably in periods covering several months; total domestic production should

be a good proxy of total domestic consumption. Detecting a drop in production of a similar

magnitude as that reported above would provide strong evidence that we have not reached

erroneous conclusions for total consumption.

To conduct this test we follow an empirical strategy that is as close as possible to that of

equation 2, except that we are forced to use aggregate time series information on national

prices and production. For prices we use Mexico’s consumer price index (INPC), and for

quantities we use the monthly manufacturing survey (EMIM); both are collected by Mexico’s

INEGI. INPC is gathered from more than 16,000 stores in 46 cities countrywide, while

31See figure A27(a) in the online appendix.

26

EMIM originates from more than 1,200 food and beverage manufacturing plants. EMIM

survey is designed to cover, on average, more than 80 percent of production in every product

category, but for the specific case of sodas all producing plants are surveyed. Soda production

is measured in thousands of liters and is classified into five packaging and flavor categories.

For the analysis, we sum these to get total production. An added advantage of this analysis

is that since information is public it is easy to replicate.

We run the following specification:

(4) log(Yty) = αt + θI(y ≥ 2014) + f(ty) + εty

Where Yty refers to prices and volume produced, f(.) is a second order polynomial in time,

αt are monthly dummies, and I(y ≥ 2014) is our variable of interest. We have monthly data

from 2007 to 2015 on soda prices and quantities.32

Columns 1 and 2 of Table IX present the estimated θ coefficient when the dependent vari-

able is the soda price index and thousands of liters respectively, while Column 3 presents an

elasticity estimate using the tax as an instrument for price. The main takeaway is that the

results are quantitatively close to our previous results: we find an increase of 12 percent in

price of sodas and a decrease of 6.9 percent in liters of sodas produced. This strongly sug-

gests that we are not reaching erroneous aggregate conclusions by not observing out-of-home

consumption in the KWP data.

To assess whether these results are spurious, we also estimate placebo regressions, where

instead of using January 2014 as the start of the tax regime we use January 2012 and 2013.

Columns 4 and 5 present placebo results for 2013 for price and quantity, while columns 6

and 7 do the same for 2012. Placebos show no changes in quantities and small decreases in

price. The online appendix presents graphical evidence.

32Figure A27(b) in the online appendix plots the log of the soda price index and the log of seasonallyadjusted liters produced and reveals two important facts: there is a sharp jump in the price of sodas and aslowdown in the growth of soda production, thus qualitatively replicating our main finding using KWP data.

27

VIII. Conclusions

Several governments and other organizations have recently advocated for the introduction

of special taxes on food in order to combat obesity. Unfortunately, there is little evidence

to date of the effectiveness of such taxes. This paper contributes to the current debate by

exploiting a uniquely detailed and high quality home-scanner dataset to estimate consumers’

(and producers’) responses to a very ambitious, large, and national fiscal policy aimed at

combating obesity through taxation.

The richness of the data allows us to estimate the pass-through of the taxes to consumer

prices and their effect on the consumption of total calories and other nutrients, and to explore

consumers’ substitutions across and within product categories. We find that the SD tax is

moderately effective in decreasing consumption, while the HCF tax is ineffective. The fact

that several foods and drinks were exempt from the tax encouraged substantial substitution

with untaxed foods. The design of the taxes caused changes in the relative price of calories

even within taxed products, which implied a substitution toward smaller SD package sizes

and cheaper higher caloric HCFs. We estimate a statistically significant but small decrease

in the consumption of sugar and an increase in the consumption of saturated fat and sodium,

but no change in total calories. Finally, we found imperfect targeting as tax incidence and

quantity reductions were not concentrated on the obese.

Our results provide useful lessons for those who wish to achieve a reduction in obesity

through special taxes. Taking substitution into account in the design of the tax base is hugely

important. This is especially important in countries with large informal markets in which

sales taxes are largely avoided.33 Taxing calories directly through taxing liters rather than

price was more effective. Package size can be used to have larger consumption effects and

target the tax towards higher BMI consumers (see Appendix A10.C). We abstracted from

discussing of there is a rationale for government interventionat all or the regressive nature of

the taxes introduced. The social costs of obesity and the distribution of the burden imposed

33Stern et al. (2014) show that in the Mexican context, “aguas frescas” usually sold informally, provide asmany calories as sodas for children, for instance. In these cases, a tax on sugar at the source may be moreeffective.

28

by taxes as those introduced in Mexico should be explored in the future.

29

IX. REFERENCES

Andreyeva, T., M. Long, and K. Brownell, “The impact of food prices on consumption:a systematic review of research on price elasticity of demand for food,” American Journalof Public Health, 2010, 100 (2), 216–222.

Barquera, Cervera Simon, Ismael Campos-Nonato, Rosalba Rojas, Juan Rivera, andSecretarıa de Salud Instituto Nacional de Salud Publica, “Obesidad en Mexico: epi-demiologıa y polıticas de salud para su control y prevencion,” Gaceta Medica de Mexico,2010, (146), 397–407.

Beegle, Kathleen, Joachim De Weerdt, Jed Friedman, and John Gibson, “Methods ofhousehold consumption measurement through surveys: Experimental results from Tanza-nia,” Journal of Development Economics, 2012, 98 (1), 3–18.

Bhattacharya, Jay and Neeraj Sood, “Who Pays for Obesity?,” Journal of EconomicPerspectives, 2011, 25 (1), 139–159.

Brownell, Kelly D. and Thomas R. Frieden, “Ounces of Prevention The Public PolicyCase for Taxes on Sugared Beverages,” The New England Journal of Medicine, 2009, 360(18), 1805–1808.

Butte, Nancy F. and Kenneth J. Ellis, “Comment on Obesity and the Environment: WhereDo We Go from Here?,” Science, 2003, 301 (5633), 598.

Byker, T., “Paid Parental Leave Laws in the United States: Does Short-Duration LeaveAffect Women’s Labor-Force Attachment?,” American Economic review, 2016, 106 (5).

Cawley, John, “An economy of scales: A selective review of obesity’s economic causes,consequences, and solutions,” Journal of health economics, 2015, 43 (C), 244–268.

Chetty, R., J. Friedman, and J. Rockoff, “Measuring the Impacts of Teachers I: Eval-uating Bias in Teacher Value-Added Estimates,” American Economic Review, 2014, 104(9).

, , S. Leth-Petersen, T. Nielsen, and T. Olsen, “Active vs. Passive Decisions and Crowd-Out in Retirement Savings Accounts: Evidence from Denmark,” Quarterly Journal of Eco-nomics, 2014, 129 (3).

Colchero, A., B. Popkin, J. Rivera, and S. Wen Ng, “Beverage purchases from stores inMexico under the excise tax on sugar sweetened beverages: observational study,” The BMJ,2016, (352:h6704).

Cutler, D., E. Glaeser, and J. Shapiro, “Why have Americans become more Obese?,”Journal of Economic Perspectives, 2003, 17 (3,Summer), 93–118.

Dubois, Pierre, Rachel Griffith, and Aviv Nevo, “Do Prices and Attributes Explain In-ternational Differences in Food Purchases?,” American Economic Review, 2014, 104 (3),832–867.

30

Duffy, K.J., P. Gordon-Larson, J.M. Shikany, D. Guilkey, D.R. Jacobs, and B.M. Pop-kin, “Food Price and Diet and Health Outcomes,” Archives of Internal Medicine, 2010, 170(5), 420–426.

Eckel, Robert H., Steven E. Kahn, Ele Ferrannini, Allison B. Goldfine, David M.Nathan, Michael W. Schwartz, Robert J. Smith, and Steven R. Smith, “Obesity andType 2 Diabetes: What Can Be Unified and What Needs to Be Individualized?,” The Jour-nal of Clinical Endocrinology and Metabolism, 2011, 96 (6), 1654–1663.

Epstein, L., N. Jankowiak, C. Nederkoorn, H. Raynor, S. French, and E. Finkelstein,“Experimental research on the relation between food price changes and food-purchasingpatterns: a targeted review,” Journal of Clinical Nutrition, 2012, 95 (4), 789–809.

Finkelstein, Eric A., Justin G. Trogdon, Joel W. Cohen, and William Dietz C, “Es-timates Annual Medical Spending Attributable To Obesity: Payer-And Service-Specific,”Health Affairs, 2009, 28 (5), 822–831.

Fletcher, J., D. Frisvold, and N. Tefft, “The effects of soft drink taxes on child and adoles-cent consumption and weight outcomes,” Journal of Public Economics, 2010, 94 (11-12),967–974.

Giesen, J., C. Payne, R. Havermans, and A. Jansen, “Exploring how calorie informationand taxes on high-calorie foods influence lunch decisions,” American Journal of ClinicalNutrition, 2011, 93 (4), 689–694.

Gracner, T., “Bittersweet: How Prices of Sugar-Rich Foods Contribute to the Diet-RelatedDisease Epidemic in Mexico,” Job Market Paper, 2015.

Griffith, R. and M. O’Connell, “Public Policy towards Food Consumption,” Fiscal Stud-ies, 2010, 31 (4), 481–507.

Grogger, J., “Soda Taxes and the Prices of Sodas and Other Drinks: Evidence from Mex-ico,” NBER, 2015, (21197).

Gruber, J., “Taxing Sin to Modify Behavior and Raise Revenue,” Expert Voices NIHCM,2010, (55).

Hill, J. O., H. R. Wyatt, G. W. Reed, and J. C. Peters, “Obesity and the Environment:Where Do We Go from Here?,” Science, 2003, (299), 853.

Kahn, S.E., R. L. Hull, and K. M. Utzschneider, “Mechanisms linking obesity to insulinresistance and type 2 diabetes,” Nature, 2006, (444), 840–846.

Lakdawalla, D. and Y. Zheng, “Food prices, income and body weight,” The Oxford hand-book of the social science of obesity, 2011, pp. 463–479.

MacKinlay, C., “Event Studies in Economics and Finance,” Journal of Economic Litera-ture, 1997, 35 (1).

31

Malik, V. S., M. B. Schulze, and F. B. Hu, “Intake of sugar-sweetened beverages andweight gain : a systematic review,” American Journal of Clinical Nutrition, 2006, 2 (84),274–288.

Olshansky, Passaro, Hershow, Layden, Carnes, Brody, and Ludwig., “A potential de-cline in life expectancy in the United States in the 21st century,” The New England Journalof Medicine, 2005, (352), 1138–1145.

Philipson, T. and R. Posner, “Is the obesity Epidemic a Public Health Problem? A Reviewof Zoltan J. Acs and Alan Lyles’s Obesity Business and Public Policy,” Journal of EconomicLiterature, 2008, 4 (46), 974–982.

Powell, L., J. Chriqui, and F. Chaloupka, “Associations between state-level soda taxesand adolescent body mass index,” Journal of Adolescent Health, 2009, 45 (3), 57–63.

Sanchez-Barriga, J. J., “Mortality trends from diabetes mellitus in the seven socioeco-nomic regions of Mexico,” Revista Panamericana de Salud Publica, 2010, (28), 368–375.

Stern, A., C. Piernas, A. Barquera, J. Rivera, and B. Popkin, “Caloric Beverages WereMajor Sources of Enery among Children and Adults in Mexico 1999-2012,” Journal ofNutrition, 2014.

Yang, C. and W. Chiou, “Substitution of healthy for unhealthy beverages among collegestudents: a health-concerns and behavioral-economics perspective,” Appetite, 2010, 54 (3),512–516.

32

FIGURE I. WEEKLY COST OF HOUSEHOLDS’ 2013 TAXABLE, NON-TAXABLE, SD, AND HCF

(a) Taxable

(b) Non-Taxable

7.5

88.

59

9.5

Pric

e pe

r Lite

r for

Sug

ary

Drin

ksFo

r Tot

al P

opul

atio

n (F

ixed

Bas

ket,

Nom

inal

Pes

os)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

(c) Price per Lt, SD

5051

5253

5455

Pric

e pe

r Kilo

gram

for T

axed

Foo

dFo

r Tot

al P

opul

atio

n (F

ixed

Bas

ket,

Nom

inal

Pes

os)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

(d) Price per Kg, HCFNote: Average consumption bundles keep the 2013 basket constant for each household for each week, so thatonly changes in prices drive the changes in cost. To calculate the cost for the consumption bundles, we multiplythe number of units of each barcode in each households’ 2013 by their current price. The assignment of currentprices to barcodes is explained in detail in section A5 of the online appendix. In panel (c) and (d) we plot pricesdefined in equation (1)

33

TABLE I—SUMMARY STATISTICS

Variable KWP ENIGH ENSANUT

Household’s Head Characteristics

Age46.44 48 49.39

[13.93] [15.2] [33.49]

Male (percentage)0.79 0.72 0.75[0.4] [0.45] [0.43]

Schooling Level

Primary (percentage)0.56 0.49 0.67[0.5] [0.5] [0.47]

Secondary (percentage)0.2 0.16 0.11

[0.4] [0.37] [0.31]

More than Secondary (percentage)0.24 0.35 0.22

[0.42] [0.48] [0.42]

Body Mass Index (BMI)27.6

-28.34

[4.56] [5.49]

Overweight (percentage)0.36

-0.39

[0.48] [0.48]

Obese (percentage)0.38

-0.33

[0.42] [0.33]

Household Assets

TV (percentage)0.95 0.96 0.93

[0.23] [0.19] [0.26]

Computer (percentage)0.39 0.39 0.24

[0.49] [0.49] [0.42]

Car (percentage)0.43 0.49 0.41

[0.49] [0.5] [0.49]

Expenditure

Total Expenditure -10155.45

-[10574.96]

Food -3256.87

-[2274.17]

KWP goodsa1560.26 1490.71

-[652.36] [1063.07]

Sugary Drinks234.14 219.99

-[162.7] [232.1]

Snacks28.34 12.35

-[33.81] [43.82]

Note: This table presents means and standard deviations for households in the Mexican chapter of the KantarWorld Panel used in our empirical analysis, along with descriptive statistics from the 2014 Mexican NationalExpenditure Survey (ENIGH), and the 2012 National Health and Nutrition Survey (ENSANUT). ENIGH pro-vides the national reference values for household demographics, expenditures and income. ENSANUT pro-vides the national reference values for health and nutrition statistics. In particular, ENSANUT measures heightand weight in situ. We apply the same filters in all the samples for comparative purposes. Total expenditures aremeasured in Mexican pesos per week. a KWP goods were constructed for comparison purposes with ENIGH.In this case, only consumption of packaged goods found in the KWP data were considered in the average shownfor ENIGH.

34

TABLE II—IMPACT OF THE TAX ON SUGARY DRINKS

Panel A. Dependent Variable: Log of Household-level Price per liter of SD Index

By Body Mass Index By Socio-Economic StatusTotal BMI Overweight Obese ABC+ CD+ DE

Tax 0.146 0.141 0.145 0.148 0.141 0.146 0.147[0.002] [0.002] [0.002] [0.002] [0.002] [0.002] [0.002]

Mean 7.98 8.09 8.01 7.90 8.24 7.93 7.92Observations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.983 0.985 0.982 0.982 0.980 0.983 0.983

Panel B. Dependent Variable: Log of the Purchase of Calories from Sugary Drinks


Tax -0.063 -0.058 -0.063 -0.068 -0.078 -0.059 -0.061[0.006] [0.007] [0.007] [0.006] [0.007] [0.006] [0.006]

MeanObservations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.777 0.779 0.772 0.776 0.789 0.773 0.774

The tax variable indicates if the observation is captured after January 1st 2014.All estimations include household-week of the year fixed effects and a quadratictime trend. Each column shows the result for a different subsample. Data em-ployed for this estimation comes from the weekly KWP Mexico chapter for 2013and 2014. The unit of observation is the household-week pair. Standard errorsclustered at the week level in brackets.

TABLE III—IMPACT OF THE TAX ON HIGH CALORIC DENSE FOODS

Panel A. Dependent Variable: Log of Household-level Price per kilogram of HCF Index


Tax 0.053 0.055 0.053 0.053 0.043 0.052 0.061[0.001] [0.001] [0.001] [0.001] [0.001] [0.001] [0.001]

Mean 50.9 51.2 51.1 50.5 51.5 50.9 50.6Observations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.983 0.984 0.983 0.982 0.979 0.982 0.986

Panel B. Dependent Variable: Log of the Purchase of Calories from HCF


Tax 0.005 0.019 0.003 0.000 -0.054 0.025 0.010[0.021] [0.024] [0.022] [0.022] [0.029] [0.023] [0.017]

MeanObservations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.650 0.664 0.649 0.644 0.658 0.639 0.659

The tax variable indicates if the observation is captured after January 1st 2014.All estimations include household-week of the year fixed effects and a quadratictime trend. Each column shows the result for a different subsample. Data em-ployed for this estimation comes from the weekly KWP Mexico chapter for 2013and 2014. The unit of observation is the household-week pair. Standard errorsclustered at the week level in brackets.

35

TABLE IV—IMPACT OF THE TAX ON CALORIES AND OTHER NUTRIENTS

Panel A. Dependent Variable: Log of 1 plus household level Total Calories

By Body Mass Index By Socio-Economic StatusTotal BMI 6 25 Overweight Obese ABC+ CD+ DE

Tax 0.016 0.042 0.014 0.006 -0.007 0.020 0.024[0.012] [0.012] [0.014] [0.013] [0.015] [0.013] [0.012]

Mean 16852 15686 16676 17642 17876 17225 15782Observations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.669 0.692 0.666 0.661 0.675 0.659 0.678

Panel B. Dependent Variable: Log of 1 plus household level Total Nutrients

Calories Sugar Carbohydrates Fat Sat. Fat Cholesterol Sodium

Tax 0.016 -0.020 -0.004 0.012 0.020 0.064 0.049[0.012] [0.010] [0.011] [0.014] [0.011] [0.012] [0.012]

Mean 16852 1371 1958 909 231 0.47 25.06Observations 762,706 762,706 762,706 762,706 762,706 762,706 762,706R-squared 0.669 0.705 0.691 0.656 0.675 0.669 0.668

The tax variable indicates if the observation is captured after January 1st 2014. All es-timations include household-week of the year fixed effects and a quadratic time trend.Each column in Panel A shows the results for a different subsample. Each column inPanel B shows the impact of the tax for different nutrients’ consumption. Data em-ployed for this estimation comes from the weekly KWP Mexico chapter for 2013 and2014. The unit of observation is the household-week pair. Standard errors clustered atthe week level in brackets.

36

TABLE V—EFFECTS ON TAXED AND UNTAXED FOODS AND DRINKS

Impact on Calories

Calories of Foods Calories of Drinks

NTF HCFs NTDs SDs

Tax 249.8 20.8 141.1 -147.7[45.3] [26.9] [22.3] [15.5]

Mean 6765 4398 3100 2723Observations 762,706 762,706 762,706 762,706R-squared 0.63 0.66 0.72 0.79

Note: The tax variable indicates if the observation is captured after January 1st 2014. All estimations includehousehold-week of the year fixed effects and a quadratic time trend. In each column the unit of observation isthe caloric consumption per household per week that originates from the consumption of the goods classifiedas follows: column 1 is non-taxed foods (NTF), which are solid products with a density below 275 kcal per 100grams and above but exempt from the tax; column 2 are HCFs, which is the group of solid foods subject to the8 percent VAT; column 3 refers to non-taxed drinks (NTDs), that is, drinks that fall under a tax exemption (e.g.milk) or that do no have added sugar (e.g. water); and column 4 SDs, which is the group with a one peso perliter tax. Data employed for this estimation comes from the weekly KWP Mexico chapter for 2013 and 2014.The unit of observation is the household-week pair. Standard errors clustered at the week level in brackets.

37

FIGURE II. EVENT STUDY PLOTS

(a) Log Taxed Drinks (SD) (b) Log Non-Taxed Drinks (NTD)

(c) Log Taxed Foods (HCF) (d) Log Non-Taxed Foods (NTF)

(e) Log Total CaloriesNote: Each dot represents the bi-weekly average level in logarithm terms after partialling out for householdand week fixed effects. The line corresponds to a cubic parametric estimation with different slopes before andafter the tax reform. The shadowed area represents the 95% confidence interval. The vertical line indicates themoment in which the tax reform began to be implemented (January 1, 2014).

38

FIGURE III. PRICE AND QUANTITY CHANGES BY CALORIES OF BARCODE

(a) %∆Price vs Calories, SD (b) %∆Price vs Calories, HCF

(c) %∆Q vs Calories, SD (d) %∆Q vs Calories, HCFNote: These graphs represent non-parametric regressions of the the caloric content of a barcode against eitherthe change in price of that barcode fron 2013 to 2014, or the change in the quantity. This is done separately forSDs in Panels (a) and (c) or HCFs in Panels (b) and (d).

39

TABLE VI—SUBSTITUTION WITHIN SDS AND HCFS

Units, Sizes and Calories

Sugary Drinks Taxed FoodLog 2013 PL Log liters per Unit Log Units Log 2013 PC Log calories per Unit Log Units

Tax 0.013 -0.047 -0.024 -0.013 0.008 -0.020[0.001] [0.002] [0.006] [0.001] [0.002] [0.006]

Observations 762,706 762,706 762,706 762,706 762,706 762,706R-squared 0.678 0.707 0.757 0.642 0.679 0.694

The tax variable indicates if the observation is captured after January 1st 2014. All estimations includehousehold-week of the year fixed effects and a quadratic time trend. The dependent variable for eachcolumn is as follows: columns 1 and 4 use the logarithm of households consumption of SDs andcalories from HCFs, respectively, but keep 2013 prices constant; columns 2 and 5 use log liters ofSDs and calories of HCFs per unit (a unit is a packaged good sold individually that is identifiedby a barcode); finally, columns 3 and 6 use the logarithm of units of SDs and HCFs bought. Dataemployed for this estimation comes from the weekly KWP Mexico chapter for 2013 and 2014. Theunit of observation is the household-week pair. Standard errors clustered at the week level in brackets.

TABLE VII—ROBUSTNESS TO INVENTORY

Robustness: Anticipatory Behavior with SDs and HCFs

Log Price per Liter (SDs) Log Price per Kilogram (HCFs)Ommited weeks None One Two Three None One Two Three

Tax 0.146 0.147 0.147 0.148 0.053 0.054 0.053 0.053[0.002] [0.002] [0.002] [0.001] [0.001] [0.001] [0.001] [0.001]

Observations 762,706 734,512 705,742 676,610 762,706 734,512 705,742 676,610R-squared 0.983 0.983 0.983 0.983 0.983 0.983 0.983 0.983

Log(1+Liters) (SDs) Log(1+Kilograms) (HCFs)None One Two Three None One Two Three

Tax -0.067 -0.068 -0.069 -0.069 0.003 0.004 0.004 0.004[0.006] [0.006] [0.006] [0.007] [0.003] [0.003] [0.003] [0.003]

Observations 762,706 734,512 705,742 676,610 762,706 734,512 705,742 676,610R-squared 0.777 0.777 0.777 0.777 0.677 0.676 0.676 0.676

The tax variable indicates if the observation is captured after January 1st 2014. All esti-mations include household-week of the year fixed effects and a quadratic time trend. Thecolumn classification None, One, Two and Three refers to the number of weeks that wereomitted before and after the introduction of the taxes (i.e. January 1st 2014). Data employedfor this estimation comes from the weekly KWP Mexico chapter for 2013 and 2014. Theunit of observation is the household-week pair. Standard errors clustered at the week level inbrackets.

40

TABLE VIII—DIFFERENCES IN DIFFERENCES ESTIMATES: USING MEXICO 2013AS A CONTROL

All Dependent Variables in Logs

Panel A: Price per Panel B: QuantitiesLiter Kilogram Calories Liters Kilograms Calories

Year (2013 effect) 0.028 0.017 -0.000 -0.027 0.009 -0.001[0.003] [0.002] [0.006] [0.015] [0.010] [0.029]

Tax 0.145 0.054 0.041 -0.062 0.005 0.023[0.002] [0.001] [0.003] [0.006] [0.003] [0.013]

Observations 1,071,550 1,071,550 1,071,550 1,071,550 1,071,550 1,071,550R-squared 0.979 0.974 0.776 0.682 0.552 0.525

The Y ear variable contains the values 2012, 2013 or 2014 and measures the begin-ning of the year effect for 2013. The tax coefficient indicates if the observation iscaptured after January 1st 2014. In this specification the tax coefficient will capturethe extra beginning of the year effect for 2014 over 2013. All estimations includehousehold-week of the year fixed effects and a quadratic time trend. Data employedfor this estimation comes from the weekly KWP Mexico chapter for 2012, 2013 and2014. The unit of observation is the household-week pair. Standard errors clusteredat the week level in brackets.

41

TABLE IX—EFFECT OF TAX ON INDUSTRY PRODUCTION AND NATIONAL SODA PRICES

Effects and Placebos

2014 2013 2012Price Liters Elasticity Price Liters Price Liters

Tax 0.12 -0.069 -0.57 -0.04 -0.01 -0.03 -0.02[0.006] [0.026] [0.20] [0.01] [0.02] [0.015] [0.015]

Month FE yes yes yes yes yes yes yesDeg. Polynomial Quadratic Quadratic Quadratic Quadratic Quadratic Quadratic QuadraticObservations 107 107 107 107 107 107 107R-squared 0.99 0.89 0.87 0.98 0.91 0.98 0.92

This table presents the results using EMIM industry data. All regressions use month fixed effectsand a quadratic trend. The tax variable reports the θ coefficient from equation 4. Columns 1, 4and 7 use the log of the price index as dependent variable. Columns 2, 5 and 8 use the log ofliters consumed. Column 3 shows the resulting elasticity that results from the columns 1 and 2estimates. Robust standard errors clustered at the date level in brackets.

Online AppendixTaxing to Reduce Obesity

By ARTURO AGUILAR, EMILIO GUTIERREZ, ENRIQUE SEIRA

Draft: July 27, 2017

A1. BMI evolution and measure in our data

The taxes being analyzed in this paper were established as a government response to a growing trendin obesity levels in Mexico. Thus, a variable of interest in this study is the body mass index (BMI),which we use to classify households based on its anthropometric characteristics. Following WHOstandards, we classify households as normal, overweight or obese according to its household’s headbody mass index (BMI) and explore if the effects analyzed in the paper differ for each group. In partof the analysis BMI is also used as a dependent variable itself.

Our BMI data comes from yearly surveys that the Kantar World Panel (KWP) implements to house-holds from our main dataset. As part of the survey, KWP gathers self-reported information of weightand height from the male and female household heads.1 The use of self-reported data might raisesome concerns. We performed some checks with strongly suggest that the BMI measure has highquality. First, it respects the ranking Mexico vis-a-vis the other countries we have in our data. Sec-ond, it is consistent with time trends in BMI as reported to the WHO. Third, and very importantly, it isvery close to data from the ENSANUT survey which is the most important health survey in Mexico,is representative of the country and measures height and weight in situ by trained nurses. Fourth, isdisplays tight relationships with demographics in an intuitive way. Finally, we note that although itwould be better to have measured weight, more than half of the countries reporting to WHO use selfreported statistics like the one in this study. This evidence provides support to the use of this variablein the main text.

First, we show how Mexico compares to other countries in terms of BMI, including a subset ofcountries for which we have KWP data available. Figure A1 confirms that Mexico is one of thecountries with highest average BMI. This map, taken from the WHO website maps, indicates that theaverage BMI in Mexico for individuals aged 18 or older is 27.5.2 Compared with the rest of CentralAmerican countries for which we have KWP data available, Mexico ranks as the highest BMI country.According to KWP data, mean values for Costa Rica, Guatemala, Panama and El Salvador are 26.9,26.3, 27.1 and 26.6, respectively. Mean values according to the WHO are very similar: Costa Rica,26.7; Guatemala, 25.8; Panama, 26.4; and El Salvador, 26.8. The only difference between the KWPand WHO rankings is that KWP reports a larger mean BMI for Panama than for the rest of CentralAmerican countries. Nonetheless, Panama is also the country for which the sample of householdsincluded in KWP is smallest. KWP reports information for 4108, 4468, 3600, and 4008 householdsin Costa Rica, Guatemala, Panama and El Salvador, respectively. Figure A2 gives a first insight ofhow the distribution of BMI compares between Mexico and the Central American countries using the

1More information about KWP and data gathered are available in section A2 of this online appendix.2http://gamapserver.who.int/gho/interactive_charts/ncd/risk_factors/bmi/atlas.html

1

http://gamapserver.who.int/gho/interactive_charts/ncd/risk_factors/bmi/atlas.html

2

FIGURE A1. WORLD OVERWEIGHT STATUS

Note: Source: http://gamapserver.who.int/gho/interactive_charts/ncd/risk_factors/bmi/atlas

female head information from KWP. It shows that Mexico whole distribution is to the right of theother four countries for which we have data.

Part of the interest in obesity fighting taxes is not only the level, but the rate of growth of obesity andBMI. Figure A3 shows the striking growth for selected countries. Figure A16 mirrors this increasingtrend with the five countries for which we have data from Kantar. This is important. One concernwith self reported data is that respondents do not update their weight information. We are witnessinga yearly increase in BMI in our data for all the countries, implying that households are updating theirweight.

Finding that our data respects rankings and trends is comforting. But perhaps the most comfortingevidence was to find out that our self-reported BMI measures coincide substantially with those col-lected in-situ by the 2012 Mexican Health and Nutrition Survey (ENSANUT). Figure A4 plots BMIdensities according to both data sources. The distributions have a very similar mean (see Table 1 inthe paper), although ENSANUT reports a slightly larger variance.

One other consistency check one might implement is to explore correlations of BMI against vari-ables which we know should correlate with BMI. Figure A5 shows the that BMI in our data is closelyrelated to age. Finally, we are not very concerned about using the female household head as a measureof obesity (see Figure A6).

Together, these checks give us confidence that BMI measures reported in KWP are good approxi-mations one of our main variables of interest.

http://gamapserver.who.int/gho/interactive_charts/ncd/risk_factors/bmi/atlas

3

FIGURE A2. BMI DISTRIBUTION, CENTRAL AMERICA AND MEXICO

0.0

2.0

4.0

6.0

8.1

Den

sity

10 20 30 40 50BMI

Mexico Costa RicaGuatemala PanamaSalvador

Note: Kernel density estimates of BMI for our five Latin America countries as reported in KWP.

FIGURE A3. OVERWEIGHT IN CONTEXT

2030

4050

6070

Sha

re O

verw

eigh

t

1980 1990 2000 2010 2020

Australia Canada France Italy Korea

Mexico Spain Switzerland United Kingdom United States

Note: Source: OECD Obesity Update. 2012.

4

FIGURE A4. BMI DISTRIBUTIONS KWP AND ENSANUT

0.0

2.0

4.0

6.0

8.1

Den

sity

0 20 40 60BMI of Household's head

KWP ENSANUTkernel = epanechnikov, bandwidth = 0.7540

Note: Kernel density estimates of BMI according to KWP and ENSANUT.

FIGURE A5. BMI-AGE PROFILES

2526

2728

29B

MI F

emal

e H

ouse

head

20 40 60 80Age

95% CI lpoly smoothkernel = epanechnikov, degree = 0, bandwidth = 2.08, pwidth = 3.12

(a) Female Profile

2526

2728

29B

MI M

ale

Hou

sehe

ad

20 40 60 80Age

95% CI lpoly smoothkernel = epanechnikov, degree = 0, bandwidth = 2.88, pwidth = 4.33

(b) Male Profile

Note: Non-parametric estimations (local polynomials). KWP data.

FIGURE A6. RELATIONSHIP WEIGHT AND BMI OF MALE AND FEMALE HOUSEHOLD HEAD

6065

7075

80W

eigh

t Fem

ale

Hou

sehe

ad

40 60 80 100 120Weight Male Househead

95% CI lpoly smoothtrimmed at>45kg and <110

(a) Weight

2628

3032

BM

I Fem

ale

Hou

sehe

ad

20 25 30 35 40BMI Male Househead

95% CI lpoly smoothtrimmed at>20 BMI and <40

(b) BMI

Note: Non-parametric estimations (local polynomials). KWP data.

5

A. Cross-sectional relationship between BMI and SD Prices and Purchases

FIGURE A7. CROSS-SECTIONAL RELATIONSHIP BETWEEN BMI AND SD PRICES AND PURCHASES

8.5

99.

510

10.5

Avg

. Pric

e pe

r Lite

r

0 100 200 300Yearly Liters per capita

Confidence Interval 95% Local Polynomial

(a) Price-Liters

2727

.528

28.5

BM

I

6 8 10 12 14Price per Liter

Confidence Interval 95% Local PolynomialYearly Liters per capita

(b) BMI-Price

Note: The Figure uses data from the Mexican chapter of the Kantar World Panel. The figures plot local polynomial regressions betweenaverage price per liter and yearly liters per capita, i.e. a naıve demand function (left) and BMI of female head vs average price per liter(right).

6

A2. Kantar World Panel: some examples

Section III in the main paper describes our data sources. Our main source comes from Kantar WorldPanel (KWP). KWP is a company with more than 50 years of experience measuring consumptionof households in dozens of countries. Together with Nielsen they are the most important companiesworldwide.

One remarkable characteristic from the KWP data is its high frequency (weekly) and the the highlevel of disaggregation, which allows us to make a detailed statistical analysis from different perspec-tives. KWP’s product catalogue comprises a total of 58, 721 barcodes (SKUs); 26,101 for foods anddrinks. It classifies barcodes along ten variables. The categories used to classify products include:basic ingredient, flavor, size, brand, product’s characteristics (e.g. light, lactose-free, caffeine-free,etc.), among others. Table A1 below shows an example of how five products are classified along thesedimensions.

TABLE A1—PRODUCT CLASSIFICATION EXAMPLE

id product subproduct brand clas01 clas02 clas03 clas04 clas05 content3510 Cereals No subproduct Kelloggs Corn Flakes . Apple w cinnamon . . 750 gr105944 Cookies Sweet Marinela Principe . Strawberry-filled Chocolate . 44 gr43379 Carbonated drinks Regular Lift Red apple . . . . 1000 ml72523 Milk Pasterurized Alpura . Regular Non lactose-free . . 1000 ml133285 Mayonnaise Regular Kraft Mayo Regular . . Light 443 gr

We used the above information to detect which barcodes were subject to the tax and which wereexempt and to impute nutritional content as we detail in this Appendix. Before proceeding withthis detail however, we would like to exemplify the richness of our data by plotting prices for someselected products. Figure A8 shows the price per liter through time for three Coca-Cola packagesizes: the 355 ml. can, the 600 ml. bottle and the 2 liter bottle. We could potentially plot these graphsby storebrand, by city or by type of household.

Figure A9 gives another example. It plots the price per unit (liter or kilogram, depending on theproduct) for the following products: (i) a 40 gram bag of Lays chips (a taxable food product), (ii) a640 gram of Bimbo bread (a not taxable food), (iii) a 2 liter bottle of Citrus Punch juice (a taxabledrink), and (iv) a 20 liter bottle of Ciel natural water (a non-taxable drink).

Along with the information on purchases, KWP also administers a yearly questionnaire that cap-tures a set of socio-economic characteristics from which SES categories are derived using standardmethodologies. SES categories are derived from different measures of household assets: number ofrooms, type of floor, number of bathrooms, whether the dwelling has a gas stove, number of lightbulbs, number of cars, and household head’s education. Following the procedure, about 21, 52 and27 percent of households in the data are classified in the A/B/C+, C/D+ and D/E SES categories,respectively. Finally, each year KWP gathers also self-reported information on the age, weight andheight of the household’s male and female heads, which allows us to calculate BMI. Using BMI, weproduce obesity and overweight indicators following WHO standards. The SES and BMI categoriesare employed in the analysis to study heterogenous effects.

7

FIGURE A8. PRICE PER LITER OF COCA COLA BY PACKAGE SIZE

1920

2122

2324

1314

1516

177

89

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

Can Coke 600ml Coke 2L Coke

Pric

e of

Fla

g Ite

ms

Sod

as (P

rice

per L

iter,

Nom

inal

Pes

os)

Note: Source: KWP data

FIGURE A9. FOUR PRODUCTS EXAMPLE, PRICES

1401

6018

0200

2202

40Ch

ips (4

0gr L

ays)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

Taxed Items

3638

4042

Loaf

Brea

d (64

0gr B

imbo

)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

Untaxed Items

66.5

77.5

8Ju

ice (2

L Citru

s Pun

ch)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

1.21.2

51.3

1.35

1.4Bo

ttled W

ater (

20L C

iel)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

Price

of F

lag Ite

ms(P

rice p

er Li

ter/K

g, No

mina

l Pes

os)

Note: Source: KWP data

8

A3. Nutritional Content Data and Imputation

A. Collection of Nutritional Data

We mentioned above that we start with 26,101 barcodes for foods and drinks. However some of theseare exactly the same product in the sense they have exactly the same substance in it, and just varyin the quantity per unit. For example a can of 355ml coke has the same substance as a 500ml coke.Once we remove the size differences we end up with 15, 622 barcodes. Our goal in this section isto describe how we recorded or imputed caloric and nutritional content for each of these given thatKantar does not do it.

The first step consisted of recruiting a team of more than 20 enumerators and train them to search forthe items in supermarkets (physically and through their internet website3), manufacturer’s website andlocal grocery stores. This work involved several months of work. Enumerators took pictures of thenutritional label and then coded such label on a predetermined format. This enabled us to do doublecoding for a sample of barcodes and verify that the enumerators actually went to the supermarket andfound the appropriate product. Each enumerator was given a list of products by the researchers ana format to fill out and enter the nutritional content available in the product’s label, mainly: servingsize, calories, sugar, fat, saturated fat, iron, carbohydrates, cholesterol, and sodium. We prioritizedbarcodes that covered a large percentage of purchase events. Finally we performed quality check ona random sample of 5 percent of the barcodes. We found less than 1 percent of imputing errors.

Using this process we managed to collect direct information for 1,953 barcodes which cover 72percent of purchase events and 68 percent of household’s expenditure of the weekly scanner data for2013 and 2014. Table A2 shows the proportion of events and expenditure that were collected for eachproduct-category that covers all of KWP data.4

We could find information of more than 95 percent of the products we selected, and therefore wethink there is no systematic selection problem in the data. However, we excluded some products thatwere very local and that comprised a very low fraction of observations.

B. Caloric and nutrient imputations

After collecting the nutritional information for a very large fraction of purchase events, we had toimpute the nutritional characteristics of the remaining products that appear in KWP’s dataset. Thenext steps were followed to perform such imputation:

1) Convert sizes to the same units: We begin by using serving sizes along with calories and nutri-ents of the products that we gathered to generate a variable that shows the amount of calories ornutrients per 100 grams (or per 100 ml. for drinks). We will refer to this variable as our densitymeasure.

2) Run an exploratory analysis to measure the variance of caloric density for different aggre-gations, using the variables {product,subproduct,brand,clas01, clas02,clas03, clas04,clas05}exemplified in Table A1. We call these the matching variables. The highest level of disaggre-gation is when we use all these 8 variables to create cells. At these level which we call L8 wefind tiny variances with cells. This just means for instance that that Alpura Milk has the samelevel of calories across its package sizes. If we go up one level to L7, there is some variancebut is tiny. This means for instance that Alpura chocolate milk has similar caloric density thanAlpura strawberry milk.

3Mainly Walmart’s website was used: [www.walmart.com.mx]4Product category is a variable in KWP data.

www.walmart.com.mx

9

TABLE A2—INFORMATION COLLECTED BY TYPE OF PRODUCT

ProductExpenditure

(percent)Purchase events

(percent)Cooking oil 83 81Bottled water 38 38Sparkling water 87 85Baby food 45 55Canned tuna 85 82Cereal bar 90 89Powdered drinks 90 90Energy drink 95 95Carbonated soft drink 98 98Sports drink 89 89Non-carbonated soft drink 74 73Snacks 96 96Coffee 85 87Condensed milk candy 45 30Seasonings and broths 78 81Cereal 94 91Chocolate 94 93Tomato puree 88 88Liquid seasoning 56 53Creamer/Substitutes of cream 63 76Sour cream 60 60Cream Spreads 69 60Breadcrumbs 96 93Cookies 80 84Cornflour drink (atole) 73 66Ice creams and popsicles 57 70Vegetable Juice 94 95Condensed Milk 96 96Powdered Milk 35 35Evaporated Milk 95 95Milk 60 60Flavored Milk 85 83Margarine 70 71Mayonnaise 83 84Jams 73 71Honey 53 56Flavored Milk Powder 90 92Industrialized bread 92 90Pastas 63 62Refrigerated Dessert 99 100Powdered Desserts 66 65Tomato puree 87 87Ketchup 86 89Snack sauce 69 71Homemade bottled sauce 84 85Pasta Sauce 53 35Instant Soups 91 90Iced-tea 37 35Yogurt 80 81

10

3) We explored by hand each product-group5 follows a different classification (which is reflectedin the variables shown in Table A1), and within each product-group we imputed the nutri-ent/caloric density to the closest category to which it belonged.

4) In other words, for each article with missing caloric and nutrient information, we search for themost similar product with information available based on the matching variables. We seek toestablish such match with the highest level of disaggregation possible. If an exact match was notpossible for a given level of disaggregation, we used a lower level of disaggregation, that is, weuse a subset of the matching variables. The following table exemplifies this procedure. In thiscase, subproduct and clas01 were chosen as the matching variables. As can be seen, there are4 products with subproduct = REGULAR6 and clas01 = COLA with missing informationfor calories. In this case, these products are matched to those with the same subproduct andclas01 for which information is non missing. The average value of the density measure ofthose products with available information is imputed to those with missing information. Thesame process is followed with the two missing values for the subproduct = REGULAR andclas01 = LEMON LIME products. However, we can see that we have in the first line acherry-flavored drink which cannot be matched to other drink with the same subproduct andclas01 values. In this case, a lower level of disaggregation is used and that product is imputedwith the average density measure value from all drinks that have subproduct = REGULAR.

5) For barcodes for which we were not reasonably sure about the quality of the imputation wesend enumerators to get the information on the supermarket.

To assess the quality of our imputation procedure, we performed the following exercise:

1) We randomly drop the observed values for 10 percent of the products for which we have avail-able caloric and nutrient information.

2) We implement the imputation algorithm described above for the product whose informationwas deleted in the previous step. We compare the imputed versus the observed values.

3) We repeated this exercise 5 times.

4) Figures A10 to A11 plot the density of difference between the observed and imputed values forcalories, sugar and fat. Similar graphs for the rest of the nutrients are available upon request.As the graphs display, there is a big concentration in zero which speaks of the adequacy of theprocess implemented to input the missing information. The imputation errors for calories arerelatively small: for liquids we obtain from the out of sample exercise that the absolute valueof the imputation error is, on average, 8.2 calories per 100 ml, while on average caloric densityis 104.7 calories per 100ml. For solid foods, the corresponding average value of the imputationerror is 61.6 calories per 100 gr, while on average caloric density is 298.4 calories per 100 gr.

5The different types of products-groups available are: cooking oil, bottled water, sparkling water, baby food, canned tuna, cereal bars,powdered drinks, energy drinks, carbonated soft drinks, sport drinks, non-carbonated soft drinks, snacks, coffee, condensed milk candy(cajeta), seasonings and broths, cereal, beer, chocolate, tomato puree, liquid seasoning, creamer, sour cream, sweet spreads, breadcrumbs,cookies, cornflour drink (atole), ice creams and popsicles, vegetable juice, condensed milk, powdered milk, evaporated milk, regular milk,milk-based flavored drinks, margarine, mayonnaise, jams, honey, milk-based flavored powder, industrialized bread, pastas, refrigerateddessert, powdered desserts (e.g. jello), ketchup, salsas, homemade bottled salsas, pasta sauce, instant soups, iced-tea, yogurt

6DIET/LIGHT is the other possible value for subproduct

11

TABLE A3—IMPUTATION EXAMPLE

num product subproduct brand clas01 kcal 100ml124 CARB. DRINK REGULAR GARCI CRESPO CHERRY48 CARB. DRINK REGULAR CABALLITOS VANILLA / ROOT BEER 292 CARB. DRINK REGULAR PEPSI COLA 45.37 CARB. DRINK REGULAR BIG COLA COLA 424 CARB. DRINK REGULAR RED COLA COLA90 CARB. DRINK REGULAR CHIVA COLA133 CARB. DRINK REGULAR SMART COLA39 CARB. DRINK REGULAR BIG COLA DOBLE COLA 431 CARB. DRINK REGULAR COCA COLA COLA 4252 CARB. DRINK REGULAR AURRERA COLA3 CARB. DRINK REGULAR BIG COLA MEGA COLA 42115 CARB. DRINK REGULAR COCA COLA LIFE COLA 1811 CARB. DRINK REGULAR 7 UP LEMON-LIME 348 CARB. DRINK REGULAR SPRITE LEMON-LIME 3681 CARB. DRINK REGULAR BIG FRESH LEMON-LIME97 CARB. DRINK REGULAR AURRERA LEMON-LIME

FIGURE A10. OUTSAMPLE IMPUTATION ERRORS FOR CALORIES

(a) Liquids (b) Solids

Note: Gaussian kernel density estimations shown

12

FIGURE A11. OUTSAMPLE IMPUTATION ERRORS FAT AND SUGAR

(a) Fat (b) Sugar

Note: Gaussian kernel density estimations shown

A4. Tax variable

As described in the main text, two different tax designs were implemented beginning on January 1st,2014: (i) one peso per liter of sugar added drinks, and (ii) an 8 percent ad-valorem tax for non-basicfood items with a caloric density above 275 kilocalories per 100 grams.

To have an accurate identification of the products in KWP’s dataset that would be subject to thesugary drink tax, we started by identifying the tax exempt products. For drinks these include bottledwater, sparkling water, beer, milk, powdered milk, evaporated milk, atole, etc. We then classified astaxable drinks those that fell into powdered drinks, energy drinks, carbonated and non-carbonated softdrinks, sports drinks, vegetable juice, flavored milk and iced-tea. For the products that are consideredfor the tax, the only exception considered is when they are sub-classified as “diet, light” or “non-sugaradded” in KWP’s categories. Then we went inside these broad clarifications to find cases of barcodesthat had no added sugar and therefore should be exempted. We did this barcode by barcode, and forthose where we were not sure we asked industry experts.

As for the ad-valorem tax, non-basic food products with higher caloric density than 275 kilocalo-ries per 100 grams with the following product classification are classified as tax eligible: cerealbars, snacks, condensed milk candy (cajeta), seasonings and broths, cereal, chocolate, sweet spreads(e.g. nutella, peanut butter), breadcrumbs, cookies, ice creams and popsicles, margarine, mayonnaise,jams, honey, industrialized bread, pastas, refrigerated dessert, powdered desserts (e.g. jello), ketchup,salsas, homemade bottled salsas, pasta sauce, and instant soups, etc.

Tables A4 and A5 show, by product-group, the percentage of items classified as being subject toeither the sugary drink (tax beverage)or the high caloric-dense food tax (tax food).

13

TABLE A4—TAX VARIABLE: SUGARY DRINKS

Product typeBarcodes with tax

(percent)Bottled water 0Sparkling water 0Powdered drinks 74Energy drink 94Carbonated soft drink 92Sports drink 100Non-carbonated soft drink 70Coffee 10Beer 0Cornflour drink (atole) 0Vegetable Juice 100Powdered Milk 0Evaporated Milk 0Drinking Milk 0Flavored Milk 100Iced-tea 86

14

TABLE A5—TAX VARIABLE: CALORIC DENSE FOOD

Product typeBarcodes with tax

(percent)Cooking oil 0Baby food 0Canned tuna 0Cereal bar 100Snacks 100Condensed milk candy 98Seasonings and broths 5Cereal 98Chocolate 96Tomato puree 0Liquid seasoning 17Creamer/Substitutes of cream 0Sour cream 0Sweet spreads 100Breadcrumbs 100Cookies 100Ice creams and popsicles 0Condensed Milk 0Margarine 99Mayonnaise 99Jams 0Honey 76Flavored Milk Powder 97Industrialized bread 63Pastas 100Refrigerated Dessert 10Powdered Desserts 6Tomato puree 0Ketchup 4Snack sauce 0Homemade bottled sauce 0Pasta Sauce 0Instant Soups 36Yogurt 0

15

A5. Price Imputation

Barcode selection: We first look at barcodes that exist (i.e. were sold to someone, in some store, insome city in the Kantar data) in all years (i.e. at least sold in one week of 2012, 2013, 2014). Weonly study those barcodes.

The problem of missing values: we are interested in the prices that would be paid for the same2013 basket of consumption for each household7. We keep the basket fixed at 2013 since we wantto isolate any changes on quantities in 2014 derived from higher taxes. We therefore need to pricethe 2013 basket for each week of 2014 and 2013 (and for some exercises for 2012 as well), but it isvery common that a household does not buy the same barcode every week, and therefore in that weekwe have to impute the price that that household would face for that barcode in the that storebrand. Itturns out that for 2012, 2013 and 2014 we have to impute 52,096,512 values of price.

Creation of “cells” for imputation: Next we create (the most dissagregated) cells as a product ofcombinations of city-storebrand-barcode-time to be able to impute mean cell prices for the case theprice is missing for a barcode b which household i purchased in 2013 in a given storebrand j8 in itscity c, in a week t. We could create a full city-store-barcode-week Cartesian product, but that wouldgenerate 6,462,260,532 cells for SDs and 13,854,472,632 cells for HCFs.9 Many of these cells wouldnot be useful however; for example if barcode A is only sold in one city, it does not help to definecells with that barcode for other cities. Similarly, if only 7 Eleven stores sell barcode X it is not usefulto generate a cell of what this barcode would cost at Wallmart as that cell would always be empty. Forparsimony we first define an event-unit (EU) as a triplet (city, store, barcode). Some cities then mayhave more EUs since they may have more storebands or more barcodes. Then we expand EUs to the156 weeks of 2012-2014. This is our finest level of disaggregation. We do the imputation separatelyfor sugary drinks (see Table A6) and for High Caloric Foods (Table A7).

Lets us explain Table A6 for SDs. We begin with the EUs × 156 which generates 24,326,328cells.10 Many of these are empty of course. In fact, only 4.2 percent are non-empty; i.e. 1,021,706cells. This might seem like a lot of missing information, however these 1 million plus non-emptycells are sufficient to impute 33.1 percent of the all the missing household ibjt prices however. Oneway to measure the imputation “precision” is to have an idea of the within cell price variation, i.e. thevariation not explained by the 1,021,706 cell dummies. It turns out that regressing actual prices (i.e.before imputation) on these dummies returns an R2 of 0.98.

To impute prices for the remaining 76.9 percentof missing observations we are forced to go to ahigher level of aggregation. We could for example aggregate across cities but keep separate cellsfor storebrand-barcode-week, or alternatively we could first aggregate across storebrands and keepseparate cells for city-barcode-week. It turned out that even if the second aggregation generates morecells11, which of these two we used first in the sequential imputation was not very important.12 Usingthe storebrand-barcode-week aggregation meant that we could fill-out an additional 29.6 percent ofthe original 24,326,328 cells.13 Now turning to household level imputation, which is what we careabout, this second step imputes an additional 30.8 percent of the original missing values. This means

7A basket of consumption for household i is a duple D:=(barcode, storebrand), i.e. it not only involves what good the household buysbut also where it buys it. A duple D is in the 2013 basket for household i if in any week of 2013 the household bought that barcode is astorebrand j. For example if I only buy coke 500ml in Wallmart and coke 500ml in Costco my basket only has two elements. Differentiatingacross storebrands eliminates price changes due to pure substitution across stores

8A storebrand is for example Wallmart, 7 eleven, Oxxo, Superama, etc.9There are 91 cities, 2863 SD barcodes, 6138 HCF barcodes, 159 storebrands, and 156 weeks.

10Note that we are note that cells do not directly have household in their ‘cartesian’ product.118,092,188 vs 6,485,700.12In fact they have the same R2 of 0.94 in a regression of unimputed price against their cells.13So at this point we have filled out 4.2 percent + 29.6 percent of the original cells.

16

TABLE A6—SEQUENTIAL PRICE IMPUTATION FOR SDS

Imputation of Prices for SDs

Prices Panel Fixed Baskets Panel

Step Aggregation Level R2 CellsFill. Cells(percent)

Imp. Observ.(percent)

1 City, store, week, article 0.98 243263281 4.2 33.12 Store, week, article 0.94 6485700 29.6 30.83 City, week, article 0.94 8092188 16.6 10.34 Week, article 0.85 303264 36.1 17.65 Month, article 0.84 69984 8.0 4.76 Quarter, article 0.83 23328 2.9 1.67 Year, article 0.82 5832 2.5 1.5

Both panels are balanced in the sense that for each price or fixed basket we have 156 weeks ofobservations. 1 Imputed cells of the first aggregation level.

that until this point we have imputed 63.9 percent of the original missing values.We proceed in an analogous fashion to higher and higher levels of aggregation, but note that we

do not need to aggregate much to get almost all of the imputations: by the 3rd step we have 3/4 ofthe imputation and by the 4th step we have more than 9/10. Note actually that even the coarser celldummies have substantial explanatory power: theR2 in step 7 is 82 percent. This means that barcodesare very good predictors of prices and that prices at the barcode level for SDs do not vary too muchby city, time and store.

TABLE A7—SEQUENTIAL PRICE IMPUTATION FOR HCFS

Imputation of Prices for HCFs

Prices Panel Fixed Baskets Panel

Step Aggregation Level R2 CellsFill. Cells(percent)

Imp. Observ.(percent)

1 City, store, week, article 0.97 540012721 4.3 33.32 Store, week, article 0.72 14202864 28.0 29.73 City, week, article 0.89 17483232 15.2 10.34 Week, article 0.41 641472 35.7 17.55 Month, article 0.3 148032 9.2 4.86 Quarter, article 0.27 49344 3.8 1.97 Year, article 0.19 12336 3.8 2.5

Both panels are balanced in the sense that for each price or fixed basket we have 156 weeks ofobservations. 1 Imputed cells of the first aggregation level.

The procedure is analogous for HCFs –but we use HCF barcodes instead of SD barcodes and theTable is interpreted analogously so we wont explain it verbally here. Let us just say that beyond thehigh R2s we have reported, three extra facts make us confident that these imputations are reasonable.First, if we estimate a barcode-level regression with week, household and barcode fixed effects (TableA8), the estimated increase in SD and HCF prices are 14.3 and 5.5 percent, respectively, which arevery close to our estimates presented in the main text. This means that our imputation+aggregationis not distorting the effect of the tax on price. The second is that the distribution of prices before andafter imputing is very similar, as shown in Figure A12. Third, note that we follow the same imputationprotocol for all years. Since our strategy is to compare across years, the imputation should not haveimportant influence on our estimates.

17

FIGURE A12. PRICE DISTRIBUTIONS BEFORE AND AFTER IMPUTATIONS

0.0

2.0

4.0

6.0

8.1

Den

sity

0 20 40 60 80 100Price per Liter

Original Imputedkernel = epanechnikov, bandwidth = 0.6111

(a) SDs

0.0

05.0

1.0

15.0

2D

ensi

ty

0 100 200 300Price per Kg

Original Imputedkernel = epanechnikov, bandwidth = 3.5394

(b) HCFs

Note: Price densities in KWP before and after imputation.

TABLE A8—PRICE EFFECTS AT THE BARCODE LEVEL

Robustness: Impact of the Tax on Prices

Price per Liter Price per Kg Price per 100 Cal(1) (2) (1) (2) (1) (2)

Tax 0.143 0.139 0.055 0.046 0.044 0.046[0.001] [0.002] [0.001] [0.002] [0.001] [0.001]

Barcode×Storebrand×Week FE Yes No Yes No Yes NoBarcode×Storebrand FE No Yes No Yes No Yes

Observations 2,651,936 2,651,936 3,471,284 3,471,284 3,471,110 3,471,110R-squared 0.616 0.657 0.567 0.621 0.795 0.825

All regressions include a quadratic time trend. This table reports RD regression results using an specification analogousto equation (2), except that an observation is a purchase event (i.e. it is dissagregated by barcodes, etc.). We use weightsproportional to liters or proportional to kilograms to have a meaningful comparison with the regressions in the paper. Robuststandard errors clustered at the date level in brackets.

18

A6. Additional Descriptive Statistics

In the main text we show how the main descriptive statistics compare between KWP and other na-tional representative surveys. In this section, we add some descriptive statistics that may be usefulto better understand the composition of households’ consumption bundles in KWP and look at thisstatistics by SES and BMI groups. Table A9 presents additional descriptive statistics regarding thecomposition of households’ consumption bundle according to KWP. Food and drinks comprise 68.9percent of total expenditures in KWP, while 14.2 and 13.5 percent of total expenditures correspondto HCF and SD, respectively. In terms of taxing calories, 42.6 percent of the calories consumed onaverage come from a food or drink product that is now taxed (25.8 percent correspond to HCFs and16.8 percent to SDs). Interestingly, neither the fraction of total expenditures devoted to HCF or SDnor the fraction of total calories purchased from HCF and SD vary considerably across SES and BMIcategories.

TABLE A9—DESCRIPTIVE STATISTICS

Weekly Exp Share of Total Expenditures(2013 mxn pesos) Food HCFs SDs

TotalTotal 392.2 68.9 14.2 13.5

Socio-Economic StatusABC+ 441.2 68.9 14.1 12.8CD+ 402.6 68.7 14.2 13.5DE 350.9 69.3 14.3 14.0

Body Mass IndexSlim/Ideal 364.5 68.4 14.7 12.6Overweight 407.0 69.4 13.9 14.3Obese 391.5 68.6 14.3 13.2

Mean Total Share of Total CaloriesCalories HCFs SDs

TotalTotal 16912.8 25.8 16.8

Socio-Economic StatusABC+ 17285.3 26.8 16.5CD+ 15839.2 25.8 17.0DE 15756.0 25.3 16.8

Body Mass IndexSlim/Ideal 15756.0 26.4 15.5Overweight 17697.7 25.3 17.8Obese 16737.6 26.1 16.4

Table based on our classification of barcodes into taxable and non-taxable, as well as theimputation of calories described in section A2.

19

A7. More on identification

A. Is the Change in Prices Attributable to the Taxes?

Figure A13 presents additional evidence in support of our identifying assumption: that the pricechanges observed in January 2014 are attributable to the taxes introduced. The figures and evidencepresented here are akin to that presented in the paper but including an extra year to the analysis. Thishelps us to corroborate that the changes in prices observed in January 2014 is not common to otherJanuaries.

FIGURE A13. EVOLUTION OF PRICES PER UNIT OF QUANTITY 2012-2014

7.5

88.

59

9.5

Pric

e pe

r Lite

r for

Sug

ary

Drin

ksFo

r Tot

al P

opul

atio

n (F

ixed

Bas

ket,

Nom

inal

Pes

os)

01/01/12 01/01/13 01/01/14 01/01/15

(a) Price per Lt48

5052

5456

Pric

e pe

r Kilo

gram

for T

axed

Foo

dFo

r Tot

al P

opul

atio

n (F

ixed

Bas

ket,

Nom

inal

Pes

os)

01/01/12 01/01/13 01/01/14 01/01/15

(b) Price per Kg

Note: Average consumption bundles are defined from all observed purchases during 2013 for each household, and fixed. For calculatingtheir cost, we multiply the number of units of each barcode in each households’ bundle by their current price. The assignment of currentprices to barcodes is explained in detail in section A5.

20

TABLE A10—JANUARY 2014 PRICE CHANGES IN CENTRAL AMERICA

Log of Price per liter of SDMexico Costa Rica Guatemala Panama El Salvador

Tax 0.146 0.001 -0.007 0.038 -0.019[0.002] [0.005] [0.003] [0.005] [0.004]

Observations 762,706 45,044 48,832 37,785 48,123R-squared 0.983 0.853 0.981 0.958 0.961

Log of Cost of the 2013 Consumption BundleMexico Costa Rica Guatemala Panama El Salvador

Tax 0.051 -0.003 -0.014 0.019 -0.045[0.009] [0.005] [0.003] [0.003] [0.004]

Observations 762,706 45,044 48,832 37,785 48,123R-squared 0.694 0.992 0.988 0.988 0.975

The tax variable indicates if the observation is captured after January 1st 2014,which is the date when the tax began to be implemented in Mexico only. Allestimations include household-week of the year fixed effects and a quadratic timetrend. Data employed for this estimation comes from the weekly KWP Mexicoand Central American chapter for 2013 and 2014. Each column indicates to whichcountry the KWP data corresponds. The unit of observation is the household-weekpair. Standard errors clustered at the week level in brackets.

21

B. RD analysis using calories as a running variable

TABLE A11—RD FOR HCF PRICES

Regression Discontinuity for HCF Prices

NP (bw=5) Linear

Density>275 0.087 0.077[0.037] [0.035]

Observations 8,966 3,322R-squared - 0.008

The first column presents the results of anon parametric (NP) RD estimate. The sec-ond presents a parametric estimate adjust-ing a linear trend on each side of the dis-continuity. Each article is weighted accord-ing to the acts of buying registered in ourdataset. For the linear specification, we re-strict the estimation to the domain of [180,400] kcal per 100 grs.

FIGURE A14. GRAPHICAL ANALYSIS: DIFFERENCE IN BARCODES’ LOG PRICES BY CALORIC DENSITY

-.10

.1D

iffer

ence

of P

rice

per K

g

0 200 400 600Calories per 100 gr

Non Taxed Taxed

Note: We show a non parametric regression at the barcode level, where the vertical axis measures the difference in log price per kilogrambetween 2013 and 2014 and the x-axis orders all barcodes according to their caloric density (kilocalories per 100 grams). The red verticalline indicates the caloric density threshold after which the HCF tax applied (i.e. 275 calories per 100 grams). Shaded areas show 95 percentconfidence intervals.

22

A8. Main Results

The main findings in the paper indicate that the tax caused a 14.6 and 5.3 percent price increase inSDs and HCFs, respectively. Also, the quantities consumed were somewhat affected in the case ofSDs where the consumption decreased 6.7 percent, while HCFs had a non significant 0.3 percentchange in terms of quantities.

Tables A12 and A13 show the results of regressions analogous to those that produce the previousresults, this time using the log of the price per 100 calories and the log of calories purchased fromSD and HCF, respectively. All the estimated coefficients are very similar in magnitude to those wepresent when using liters of SD or kilograms of HCF as our unit measure.14 The results obtained fromthese specifications find a 14.5 and 4 percent price increase for SDs and HCFs, respectively. Mean-while, quantities of SDs deacrease 6.3 percent and those of HCFs have a non-significant increase of0.5 percent. In summary, the changes found are very similar when using calories instead of liters andkilograms.

TABLE A12—RESULTS USING LITERS AND PRICE PER CALORIE FOR SD

Impact of the Tax on the Price per 100 Calories of Sugary Drinks


Tax 0.145 0.141 0.145 0.148 0.141 0.146 0.147[0.002] [0.002] [0.002] [0.002] [0.002] [0.002] [0.002]

Observations 762,706 143,255 329,988 289,277 148,408 348,672 265,440R-squared 0.985 0.987 0.985 0.985 0.984 0.986 0.986

Impact of the Tax on the Purchase of Liters of Sugary Drinks


Tax -0.067 -0.062 -0.065 -0.072 -0.081 -0.063 -0.064[0.0006] [0.007] [0.007] [0.006] [0.007] [0.006] [0.006]


The tax variable indicates if the observation is captured after January 1st 2014. All estimationsinclude household-week of the year fixed effects and a quadratic time trend. Each column shows theresult for a different subsample. Data employed for this estimation comes from the weekly KWPMexico chapter for 2013 and 2014. The unit of observation is the household-week pair. Standarderrors clustered at the week level in brackets.

Also, Table A11 shows the result from using a regression discontinuity design as the one illus-trated in Figure A14 in the main text. The estimates shown in the table correspond to estimating thefollowing specification:

(1) ln (PSDb ) = β0 + β1 I(Denb > 275) + f(Denb) + εb

where b denotes barcode and Denb is the caloric density of barcode b. In the estimates presented,we use a non-parametric estimation in the case of column 1 and a linear trend with different slopes

14Given that magnitudes of liters and kilograms are different, the results presented here normalize the caloric content to that of oneaverage liter and kilogram of SD and HCF to make the coefficients easily comparable across the tables.

23

TABLE A13—RESULTS USING KILOGRAMS AND PRICE PER CALORIE FOR HCF

Impact of the Tax on the Price per 100 Calories of HCF


Tax 0.040 0.039 0.041 0.038 0.030 0.040 0.045[0.002] [0.002] [0.002] [0.002] [0.002] [0.002] [0.002]


Impact of the Tax on the Purchase of Kilograms of HCF


Tax 0.003 0.007 0.004 0.002 -0.008 0.006 0.007[0.003] [0.003] [0.003] [0.003] [0.004] [0.003] [0.003]


The tax variable indicates if the observation is captured after January 1st 2014. All estimationsinclude household-week of the year fixed effects and a quadratic time trend. Each column shows theresult for a different subsample. Data employed for this estimation comes from the weekly KWPMexico chapter for 2013 and 2014. The unit of observation is the household-week pair. Standarderrors clustered at the week level in brackets.

before and after the cutoff in column 2. To make the results comparable to those presented with thehousehold-week as unit of observation, each observation used (which corresponds to a barcode) isweighted by the number of transactions done with that barcode.15

The results found with this specification give a 7.7 to 8.7 percent change in prices, which is aroundthe 8 percent tax rate, meaning something close to a 100 percent pass through. The results of usingthis specification are slightly larger than the 5.3 percent estimated in the main regressions. It shouldbe indicated though that as a traditional RD result, the estimate is valid for the value of density at thecutoff. Also, it is important to recall that in the paper, no evidence of strategic decisions around thediscontinuity is found (see Figure A28).

15The purpose of this is to represent that a product with very high demand should capture more importance that a product that is seldomconsumed.

24

A9. Estimates of effects on BMI

The taxes introduced were aimed at combatting obesity. For this reason, individuals’ BMI may be themost relevant outcome to analyze. If the intake of total calories did not change as a result of the taxes,we should not observe significant changes in BMI. Before presenting a differences in differencesanalysis, Figure A15 just plots kernel densities of BMI for several years in Mexico, from two yearsbefore and two years after the tax. The resemblance is striking. There was no change in BMI in anypart of the distribution.

FIGURE A15. OVERWEIGHT IN CONTEXT0

.02

.04

.06

.08

Den

sity

10 20 30 40 50BMI

2015 20142013 2012

Note: Source: Kernel density estimates of BMI as reported in KWP.

Figure KWP A15 of course does not contain a counterfactual comparison. We would need a controlgroup for which there was no tax in 2014. Our empirical strategy next is to compare Mexico vs itsLatin American neighbors. This is reasonable since (a) we find the same pre-trends in BMI, (b) thetend to consume similar baskets of goods, (c) they were not subject to the tax and we were able to getBMI data at the household level from the same source (KWP) using the same methodology.

We use yearly self-reported information on the weight and height of female household heads forparticipating households in Costa Rica, Guatemala, El Salvador and Panama as a control group16 forMexican households. Our identification assumption is that in the absence of the taxes households inMexico would have experienced a similar change in BMI as households in the these four countries.As shown below this assumption is reasonable; these countries have similar pre-trends in BMI thanMexico.

Figure A16 shows that, consistent with WHO’s data 17, (average across households) BMI has in-creased both in Mexico and the rest of countries in Central America during the period of analysis. Thisadds confidence that the BMI data is not noise with the same mean each year. Moreover, althoughBMI is higher in Mexico than in the rest of Central America, the time trends showed by both Mexicoand the rest of Central America are relatively similar before the enactment of the tax, supporting theidentification assumption of parallel pre-treatment trends.

16KWP collects BMI for a subset of 89 percent of male households heads in Mexico, but unfortunately this information is onlyavailable for 2 years in Central America. We therefore focus on BMI of the female head. We have 54,143, 4,442, 4,918, 4,026 and 4,315household-year observations from Mexico, Costa Rica, Guatemala, El Salvador and Panama, respectively on the female heads BMI.

17See section A1 of the online appendix.

25

FIGURE A16. BMI TRENDS IN MEXICO VS CENTRAL AMERICAN COUNTRIES

26.5

2727

.528

Ave

rage

BM

I of H

ouse

wife

2010 2011 2012 2013 2014 2015Year

Mexico Central America

Note: The graph simply plots BMI for our sample of female household heads in Mexico, and for the rest of Central American countries,pooling all samples together.

Graphical evidence suggest that there was not change in the evolution of BMI over time for Mexicoafter the taxes were introduced. In order to provide statistical evidence we estimate the followingdifference-in-difference (DD) specification:

(2) BMIijy = αj + θy +

2015∑t=2011

βtMexicoj ∗ I(y = t) + εijy

Where i, j and y denote individual, country and year. The equation includes dummies for 5 years2011-2015 (θy), dummies for each Central American country (αj) to allow for country-specific BMIlevels, and an interaction of the Mexico (treatment) dummy with the year dummies. We clusterstandard errors at the country level. The βt’s are the coefficients of interest, which measure thedifferential change in BMI (respect 2010) experienced in Mexico versus other Central Americancountries each year.

FIGURE A17. BMI AND OBESITY DIFFERENCES IN DIFFERENCES ANALYSIS

(a) BMI (b) Obesity

Note: This Figure plots the coefficient estimates from equation 2. Results from years 2011, 2012, and 2013 show that the parallel trendsassumption is satisfied. Results from 2014 and 2015 show that there is no effect of the tax.

26

Because the panels are unbalanced, we present regression results with and without female house-hold head fixed effects. Figure A17 plots the estimated βt’s along with their 95 percent confidenceintervals. The joint F-test for βj = 0 ∀ j ∈ {2011, 2012, 2013} is not rejected (p-value=0.17),suggesting that the parallel trends assumption holds. The coefficients for 2014 and 2015 show noeffect of the tax on BMI for Mexico. The effect of the tax on weight may take time, but the fact thatβ2015 is larger than β2014 is discouraging. Panel B of figure A17 uses the fraction of obese popula-tion as a dependent variable and reaches similar conclusions. Given the strong relationships betweenhousehold members BMIs found in the literature, we believe our results may apply more broadly toother household members.18

FIGURE A18. POWER GRAPH

0.2

.4.6

.81

Pow

er

-.9 -.8 -.7 -.6 -.5Treatment Effect on BMI

Note: Figure A18 displays a power simulation for BMI using our exact same specification 2 and our same data base.

Figure A18 displays a power simulation for BMI using our exact same specification 2 and our samedata base. This gives us an exact power calculation that takes into account the whole structure of ourdata and specification. The simulation proceeds as follows: we take 2013 for Mexico as our placebotax year and define the DID dummy = 1 for a household if the country is Mexico and the year is ≥2013. For each obervation for which DID=1, we add a placebo effect on the BMI dependent variable.The placebo effect is drawn randomly from a normal distribution with mean µ and standard deviationσ. The standard deviation is taken from the empirical distrbution of the female head BMI year-to-yearchange, while µ runs from 0 to -1 BMI points, in steps of 0.02.

For each µwe run 400 simulations. In each simulation for each household we draw a placebo effectfrom the normal distribution described above. Then we estimate the regresion in equation 2, givingus 400 estimated effects. Finally we count the fraction of the placebo effects which are statisticallydifferent from zero at 95 percent confidence. This fraction is our measure of power for each µ.

Figure A18 shows that we can detect an effect of about 0.75 BMI points with above 80 percentpower. For a person with 1.60 meters in height, this is equivalent to about 2kg. We think this isa reasonable effect to detect even on average in 2 years, to decrease obesity people would need todecrease oftentimes more than 10kg. Not finding effects of 2kg is informative indeed.

18We actually estimate a coefficient of 0.46 (t-stat=95) in a regression of male and female household heads BMI’s. Section A1 of thisonline appendix reports quality checks on the weight and BMI measures. Height in the data does not change –as expected– but weight doeschange by reasonable magnitudes with mean changes between 1 and 2 kilos. In power simulations that use our exact empirical specificationwe obtain that we are able to detect effect sizes of about 0.75 BMI points and 1.3 kilos with 90 percent power and 95 percent confidence.

27

A10. Mechanisms

A. Changes in Prices and Quantities by Barcode

FIGURE A19. PRICE AND QUANTITY PERCENTAGE CHANGES: HETEROGENEITY ACROSS BARCODES

(a) (b)

(c) (d)

Note: These graphs histograms of changes in price or quantity for each barcode of SD or HCFs from 2013 to 2014.

B. Change in Consumption Decisions

The results presented in section VI of the paper show interesting patterns in the consumers decisionmaking process. In particular, the substitution towards cheaper goods within the high caloric densityfoods calls for further analysis and attention. Given that this substitution happens within the HCFgroup it can mean that consumers are switching towards cheaper and lower quality products within anhomogeneous category (e.g. choosing a cheaper brand of the same product) or changing consumptionamong HCF categories toward a cheaper group (e.g. changing peanuts or pistachios for chips). Tofurther investigate this, in Table A14 we show the change in the 2013 value of the chosen consumptionbasket for different categories of product. This means that the different estimates will only capturewithin product category substitution. The evidence from the table shows that within product categorysubstitution is an important driver of the results presented in section V. The case of snacks and cookiesis notable for its large decrease. The rest of the categories included in this table show lower decreases

28

FIGURE A20. DIFFERENTIATED IMPACT OF THE TAX ON SDS AND HCFS PRICES

1.8

22.

22.

42.

62.

8Lo

g of

Pric

es p

er L

iter

0 1 2 3 4Liters per Unit

2013 2014

(a) SDs

.88

.9.9

2.9

4.9

6.9

8Lo

g of

Pric

es p

er H

undr

ed C

alor

ies

0 500 1000 1500 2000Calories per Unit

2013 2014

(b) HCFs

Note: The left hand side graph plots the average of the log price of SD barcodes in 2013 and 2014 against the content in liters of eachbarcode. The right hand side graph plots the average of the log price per 100 calories of HCF barcodes in 2013 and 2014 against thecontent in calories of each barcode.

FIGURE A21. PRICE CHANGE IN CALORIES PER UNIT: SDS VS HCFS TAX

.03

.035

.04

.045

.05

.055

.06

Diff

eren

ce o

f Log

s of

the

Pric

es p

er H

undr

ed C

alor

ies

for T

axab

le F

ood

.135

.138

.141

Diff

eren

ce o

f Log

s of

the

Pric

es p

er L

iter f

or S

ugar

y D

rinks

0 500 1000 1500 2000 2500Calories per Unit

Sugary Drinks Taxable Food

Note: The figure orders all SD and HCF barcodes in the Kantar World Panel according to the amount of calories per unit. A unit is apackaged good sold individually that is identified by a barcode. The vertical axis measures the difference in log prices between 2013 and2014. The blue solid line illustrates the case of sugary drinks (SDs) as defined by the tax and in our estimations. The red dotted line showsthe result for high caloric dense foods (HCFs).

29

that range from 4.5 to 0.4 percent for categories such as industrialized bread, crackers, cereals andspreadable creams.

TABLE A14—SUBSTITUTION WITHIN A PRODUCT CATEGORY

Substitution Patterns Within Food Categories

Log 2013 PCSnacks Bread Cookies Crackers Cereal Spreadables

Tax -0.115 -0.033 -0.075 -0.045 -0.004 -0.008[0.007] [0.002] [0.002] [0.002] [0.002] [0.001]

Observations 762,706 762,706 762,706 762,706 762,706 762,706R-squared 0.669 0.644 0.626 0.638 0.640 0.647

The tax variable indicates if the observation is captured after January 1st 2014. All esti-mations include household-week of the year fixed effects and a quadratic time trend. Eachcolumn shows the result for baskets of different categories of products. Data employedfor this estimation comes from the weekly KWP Mexico chapter for 2013 and 2014. Theunit of observation is the household-week pair. Standard errors clustered at the week levelin brackets.

30

C. Substitution Towards Smaller Package Sizes

In 2013 the City of New York issued the Sugary Drinks Portion Cap Rule to prohibit the sale ofseveral sweetened drinks of more than 16 ounces (0.5 liters) in volume.19 This rule was based ontwo motivations. The first was the trend in sizes of drinks; between 1977 and 1996 the averageenergy intake and portion size of the average soft drink consumed in the US increased from 144 to193 kcal (Nielsen and Popkin (2003)). The second is psychological evidence showing that largerportion sizes increase consumption (Wansink B., 2005). This later motivation may be compoundedfor large volume units of sodas, since soda cannot be stored once opened without quickly losing somecarbonation.

One may ask if a price disincentive toward larger package sizes in the form of a tax makes sensein terms of both reducing consumption and targeting this reduction towards those who are moreobese. This is a reasonable question. Figure A22 shows that indeed higher BMI households tend toconsumer larger package sizes. In this section we answer affirmatively to both issues. In the previoussubsection we analyzed the effect of taxes from a reduced form perspective, here we turn to a anempirical strategy with more structure that allows us to estimate own and cross prices elasticitiesacross unit sizes and to simulate alternative tax schedules and their effect on consumption. We focuson SDs since they experienced the larger effects and products are relatively homogeneous acrosssizes, which allows us to focus on size.

FIGURE A22. PACKAGE SIZES VS BMI AND PRICES

2728

2930

BM

I

500 1000 1500 2000 2500 3000Unit Volume

Confidence Interval 95% Local PolynomialYearly Liters per capita

(a) BMI-Prices

11.

051.

11.

151.

2

Pric

e pe

r Lite

r for

Sug

ary

Drin

ks (F

ixed

Bas

ket,

Inde

x 20

13=1

00)

01/01/13 25/06/13 01/01/14 25/06/14 01/01/15

(0,355] (355,600] (600, 1500] (1500,NaN)

(b) Price per Package Size

Note: The left hand side graph shows the relationship between self reported BMI in the Mexican chapter of the Kantar World Panel andthe average package size of SD barcodes bought. Package size is measured with unit volume and expressed in mililiters. Higher BMIconsumers consume units with larger package sizes. The right hand side graph illustrates the evolution of the price index for differentpackage sizes of SD.

Products containing more liters of sugary drinks increased prices in a higher proportion with respectto smaller packages within the same category. Conversely, the price of barcodes in the HCF categorythat contain more calories per unit increased relatively less with respect to barcodes with less caloriccontent within the same category. This finding is illustrated in Figure A21. Figure A20 gives moredetail about this contrasting fact by plotting in two graphs the changes in prices, one for SDs andanother for HCFs. Figure A20 shows that both, in the case of SDs and HCFs, products with moreliters (or calories) per unit have a lower price per liter (or calorie). This is shown by the negative slopein the graphs that relate price per liter and liters per unit (or price per calorie and calories per unit).

19The regulation was struck down by the Court of Appeals before it came into effect.

31

The contrasting effect of the tax between SDs and HCFs is evidenced by the fact that this negativerelation shifted to the right in both cases as a result of taxes. However, this shift is more pronouncedat higher values of liters per unit for SDs and contrarily, in lower values of calories per unit for HCFs.Figure A22 shows that prices increased more for SD with larger unit sizes (panel b).

D. An AIDS model of package sizes

One can think of different package sizes as being different products with their own demands.20 Weassume this is the case and estimate an Almost Ideal Demand System (AIDS), where consumersallocate different shares of expenditure to different package sizes of SD. AIDS has the well-knownadvantages of exact aggregation and consistency with utility maximizing behavior while allowing fora flexible specification of own and cross price elasticities. Since there are more than a dozen sizesof soda package sizes, in order to have a parsimonious model we aggregate package sizes into fourgroups: (0,355],(355,600], (600,1500],(1500,NaN).21 We also have a fifth group to which we allocateall other food expenditures, which allows for substitution between SD and other goods.

Following the AIDS literature, we assume weak separability between the consumption of SDs andall other expenditures. The AIDS expenditure function is displayed in equation 322

log (C(u, pi; Θd)) = α0 +∑g

θgd log (pgi) +1

2

∑g

∑g′

γ∗gg′ log (pgi) log (pg′i)+

+ uβ0∏g

pβggi

(3)

where d indexes demographic group, g indexes product size group, pi is a vector of prices in-curred by household i, Θd is a vector of tastes that are identical for households with demographicsd. Shepard’s Lemma is applied to this function to generate equations 4 of budget shares Sgi as afunction of package sizes and demographics and log prices for each package size, as well as the logreal household expenditures mi/Pi.

(4) Sgi = θgd +∑g′

γgg′ log (pg′i) + βg log

(mi

Pi

)

Following Deaton and Muellbauer (1980), we approximate the price index Pi by a Stone index,lnP ∗

i =∑

g Sgd log (pgi) making the system linear and easily estimable. We use monthly-householdlevel information for all months in 2013.23 The main identification assumption is that prices withindemographic groups are uncorrelated with idiosyncratic tastes. This is a strong assumption that is

20Below we provide some evidence that different types of households buy different package sizes.21The respective market share in terms of items purchased are 17.6, 17.4, 14.2, and 50.8 respectively. When we calculate the market

shares in terms of liters they are, of course, even more concentrated in large packages: 80.8 percent of liters in 2013 were sold in the morethan 1,500 ml category. This concentration is striking in itself.

22This can be generated by a two-stage budgeting process in which households first decide on the allocation of income across broadproduct categories, SD being assigned m pesos, and then allocate their budget within SD size z categories and choose an expenditurefunction given in equation 3.

23We tried also several generalizations with similar results: (a) using a QUAIDS model (Banks et al. (1997)), which allows for richerincome effects (we forgo the use of IV in this more complex environment); (b) not imposing a Stone index and instead using a fully non-linear estimation; (c) using data at the week or year level; (d) estimating the equations with and without week dummies to take advantageof different types of variation in prices; (e) estimating the share equations simultaneously.

32

typically imposed in the AIDS literature. Following Hausman et al. (1994) and Atkin (2014), weestimate the equations using prices in neighboring cities as instruments for the prices. This instru-ment satisfies the exclusion restriction if cost shifts, unrelated to aggregate SD demand in the city ofinterest, are causing price changes in other cities. We impose the restrictions on parameters impliedby utility maximization (adding up, symmetry).24

AIDS model fit

In this section of the appendix we provide some evidence of the quality of the estimated AIDS modelto fit the data and to predict the changes in expenditure out of sample. This is important since theAIDS models have several assumptions which may bias their predictions if they turn out to be wrong.The paper provides one measure of fit by predicting the 2013 liters consumed, and a measure ofpredictive accuracy by projecting the liters consumed out of sample for the year 2014. The resultsshow a reasonable fit. Here Figure A23 shows that the actual 2013 and predicted 2014 share ofexpenditures for the four categories of SDs are very similar. Given that the AIDS estimation wasdone with information from 2013 finding a good 2013 fit is not very surprising. Nonetheless, themodel is then employed to forecast the levels of expenditure shares that would result from the taxeffects and they closely match the actual 2014 shares in the data. The evidence from the graph givesa robust support to our estimates since the observed and predicted levels look very similar despite thefact that this is an out of sample estimation.

FIGURE A23. AIDS IN SAMPLE AND OUT OF SAMPLE FIT

0.0

2.0

4.0

6.0

8.1

Sha

re o

f Exp

endi

ture

2013 2014

(0,3

55]

(355

, 600

]

(600

, 150

0]

(150

0,NaN

)

(0,3

55]

(355

, 600

]

(600

, 150

0]

(150

0,NaN

)

by category of Sugary DrinksAverage Shares and Predicted Shares

Original Predicted

Note: This Figure uses the AIDS model estimated on 2013 data and shows the model predicted “in-sample” shares of 2013. It also showshow the model fares in predicting 2014 shares. This later prediction is an “out-of-sample” prediction as the model was not estimated with2014.

24The AIDS model and the instrumental variables we use rely on mildly strong assumptions. Failure of these assumptions will biascoefficient estimates and predictions. One way to assess the accuracy of the assumptions is to see how the model performs out-of-sample,which is what we do. We estimate the model using 2013 data only and then use the model to predict expenditure shares for 2014. Thisis a stringent test, as we know that soda prices changed in 2014 due to the tax and they changed differentially across package sizes. Theappendix shows that out-of-sample predictions are close to the actual 2014 data, providing some validation of the model.

33

AIDS Results

Table A15 presents the matrix of own and cross-price uncompensated elasticities. Confidence inter-vals are calculated by bootstrapping 1,000 times the elasticity estimates and reporting the 5th and 95thpercentiles.25 Note that all own price elasticities are negative and statistically different from zero. Theelasticity differs across package sizes substantially. The coefficients next to the diagonal are all pos-itive and smaller than those in the diagonal, indicating that similar package sizes are substitutes andthat cross responses are smaller than own-price responses.

Table A16 presents a series of exercises using the estimates of the four tax-price incidence pa-rameters Ij , j = 1, .., 4 displayed in column 2 and the 16 own and cross price elasticity parametersεij , i, j = 1, 2, 3, 4 of Table A15. The first column reports the average 2013 price per liter for SDproducts for each package size. Column 2 reports the estimated percent increase in prices that eachcategory experienced with the introduction of the tax on SDs. Column 3 shows liters consumed in2013 per household per week on average. It also measures liters and tax revenue at the week perhousehold level. Column 4 reports the actual tax structure of one peso per liter across all packagesizes, liters consumed, and tax revenue in 2014. Column 5 is a check on model fit and shows the litersand revenue that our AIDS model would predict using the 2014 tax structure. The fit is strikinglyclose, $6.62 pesos in the data vs. $6.51 for our model. This is a not trivial, since we are using 20estimated parameters and the 2014 prices are substantially different from those of 2013.

Columns 6 and 7 perform counterfactual exercises.26 Scenario S1 in Column 6 simulates whatwould happen with a tax of 1 peso per liter only on the largest packages only. It shows that taking intoaccount all own and cross price elasticities, revenue would be similar to the current 2014 revenue.

One might also ask whether there are tax profiles that achieve more reduction in liters consumedwhile keeping revenue constant. There are several trade-offs to take into account when consideringtilting the tax towards larger or smaller packages. Larger packages command the largest market share;they also showed the greatest price pass through. These two reasons suggest that taxing large packagesmore heavily may achieve larger reductions in liters, all else constant. Note however that they do nothave the largest own-price elasticities, and that we have to take into account substitution patternsrepresented by the cross price elasticities as well if we want to calculate the counterfactual correctly.Problem 5 below takes these four considerations into account, and finds which is the optimal taxprofile (τ1, τ2, τ3, τ4), in the sense of minimizing liter consumption subject revenue staying the sameas in 2014.

(5)

Minimizeτ

4∑i=1

Li(τ) ≡4∑j=1

(

4∑i=1

(L0i εijIj))τj

subject to 0 ≤ Li(τ), i = 1, 2, 3, 4.

0 ≤ τi ≤ 1, i = 1, 2, 3, 4.

4∑i=1

Li(τ)τi = R

Column 7 reports the result. A tax of $1.32 pesos per liter only for the largest packages and a

25To mimic the city-cluster structure of the data in each iteration we draw a random sample with replacement of all the householdswithin each city. When using monthly data, this meant sampling the entire history of a household.

26All these counterfactual analysis assume that incidence and elasticities are constant with respect to the tax rate. This seems reasonablefor the range of price variation we use.

34

zero tax on all the other package sizes is optimal. So we do not have to invoke behavioral biasesor problems of self-control to make an argument for taxing larger packages. There is an additionalargument however. There is a strong and tight positive relationship between BMI and package sizeof SDs. The lefthand panel of Figure A22 shows that this relationship in our data. So, a tax onlarger packages targets more overweight population, which may mean that the tax is more focused onover-consumption and is therefore only distorts consumption where it should.

TABLE A15—UNCOMPENSATED PRICE ELASTICITIES

Uncompensated Price Elasticity Matrix Estimates IV

With Respect the Price ofCommodity Group (ml) (0,355] (355,600] (600,1500] (1500,NaN)

(0,355]-0.99 0.3 0.38 -0.38

[-1.13, -0.85] [0.15, 0.46] [0.23, 0.56] [-0.64, -0.14]

(355,600]0.16 -0.92 0.52 0.25

[0.08, 0.25] [-1.09, -0.71] [0.39, 0.65] [0.03, 0.47]

(600,1500]0.16 0.4 -1.47 0.14

[0.09, 0.24] [0.3, 0.5] [-1.6, -1.36] [-0.01, 0.3]

(1500,NaN)-0.03 0.02 0.02 -0.79

[-0.06, -0.01] [-0.01, 0.05] [-0.01, 0.05] [-0.88, -0.7]

This table shows own and cross price elasticities estimated with the AIDS model. Rows in-dicate product package sizes while columns indicate the price. For instance, the bottom leftcoefficient of -0.03 indicates the percent change in the quantity of the largest package (morethan 1.5 liters) when the price of the smallest package (less than 355 mililiters) changes by1 percent. Prices in a city were instrumented with the prices in its nearest city. Confidenceintervals and significance levels were obtained from 1000 bootstraps.

35

TABLE A16—COUNTERFACTUAL TAX ANALYSIS

Estimated Impact of Alternative Tax Scenarios

CounterfactualScenarios

Package sizecategory (ml)

(1) Price2013

Observed

(2) PriceincidenceEstimated

(3) Liters2013

Observed

(4) Tax-Data2014

Observed

(5) ModelFit

2014(6) S1 (7) S2

(0,355] 16.56 5.6 0.18 1 1 0 0(355,600] 12.57 11.30 0.39 1 1 0 0(600,1500] 9.68 9.90 0.74 1 1 0 0(1500,Nan) 6.71 16.30 5.58 1 1 1 1.32

Total Liters a - - 6.91 6.62 6.51 6.58 6.35Revenue b - - - 6.62 6.51 5.25 6.62Price Incr.c - - - 14.6 15.0 16.3 21.5Change in Liters d - - - -6.7 -5.79 -4.78 -8.10

This table presents simulated scenarios for different tax profiles. The first column reports the average 2013 priceper liter for SD products for each package size. Column 2 reports the estimated percent increase in prices thateach category experienced with the introduction of the tax on SD. Column 3 shows liters consumed in 2013 perhouseholds per week on average. Column 4 reports the actual tax structure of one peso per liter across all packagesizes, liters consumed and tax revenue in 2014. Column 5 is a check on model fit and shows the liters and revenuethat our AIDS model would predict by using the 2014 tax structure. Column 6 simulates what would happenif a tax of 1 peso per liter was applied only on the largest packages. Finally, Column 7 shows what tax profileminimizes liters consumed while keeping tax revenue constant. Data employed for this table includes the 2013and 2014 household consumption from the Mexico chapter of the KWP. The units of measure are a total liters perhousehold and week,b pesos spent per household and week,c weighted average percentage, d total percentage.

36

A11. Robustness checks

A. Anticipation

An alternative behavioral response to taxes is to keep large stocks as the tax implementation ap-proaches. With this response, household can substitute consumption just after the tax implementa-tion, particularly with non perishable products. We perform a robustness test in section VII.A of themain text, where we keep out of the sample a window of time around the beginning of the tax imple-mentation and obtain similar results. Following the marketing literature, in this section we performtwo additional exercises to support that dynamic considerations are not biasing our estimates. Weexplore if (a) there were more purchases of SDs and HCFs in December 2013 vis-a-vis December2012 to accumulate inventories; (b) whether the probability of buying was lower in January 2014than in January 2012 or 2013; and (c) if households spend more weeks without making purchasesof SD and HCF in January 2014 versus January 2012 and 2013. The results of this exercise are pre-sented in Table A17. No evidence of stockpiling behavior from the households is found. The onlystatistically significant coefficient is an increase of 0.1 weeks in time to buy after January 1st 2014.This is economically very small.

TABLE A17—EVIDENCE OF INVENTORY BEHAVIOR

Robustness: Storage of SDs and HCFs

SDs HCFsLog Liters Buy Weeks Log Kgs Buy Weeks

Dec 2012 0.066 0.031[0.105] [0.041]

Jan 2012 0.004 -0.105 -0.007 -0.100[0.009] [0.033] [0.012] [0.034]

Jan 2013 -0.000 -0.055 -0.008 -0.036[0.003] [0.021] [0.004] [0.031]

Mean 1.68 0.86 1.31 0.65 0.90 1.18SD [0.96] [0.34] [1.26] [0.43] [0.3] [0.75]

Observations 708,461 1,071,550 1,053,177 708,461 1,071,550 1,057,024R-squared 0.473 0.204 0.282 0.223 0.097 0.090

This table estimates regressions for different measures of SD and HCF purchases: quantities pur-chased, the likelihood of purchase, and the number of weeks with no purchase. Columns 1 and 4examine whether there were more purchases of SDs and HCFs in December 2013 vis-a-vis Decem-ber 2014 and 2012. Columns 2 and 5 estimate if the probability of buying was lower in January2014 vis-a-vis January 2013 and 2012. Columns 3 and 6 estimate if the household spends a largerspell without purchases of SDs and HCFs in January 2014 vis-a-vis January 2013 and 2012.All regressions include household and month of the year fixed effects and a quadratic time trend.Clustered standard errors at the week level in brackets.

B. Falsification tests

Our second exercise is a set of falsification tests following the logic and output format of a Fisherexact test27. Our aim is to show that the effect captured with the tax variable in our main estimationsis indeed a quite an unusual event in terms of size. This test consists in replicating our main specifi-cation (equation 2) several times and using in each replication a different ‘placebo’ date to generate

27We do not call it a Fisher exact test as that test requires sampling from the same distribution and here we are sampling from differentmonths.

37

the tax variable. In total 34 replications were done, and in each replication a different month be-tween February 2012 and November 2014 was employed to create the tax variable. In particular, thereplication that uses January 2014 is exactly our main specification. The other replications establishartificially the tax at dates where it did not happen and as such are placebo estimates.

If say end of quarter months were particularly good at SDs sales, this would show up in theirrespective month effects. So that these falsification tests can potentially detect if our results are dueto seasonality for example.

We proceeded to estimate the 34 regressions and collected the resulting {β1, . . . , β34} from ourestimations. We do this for each dependent variable separately. If our equation is correctly specifiedand if in reality the tax in January 2014 did have an effect, we would expect all betas other than theone for January 2014 to be statistically zero, except perhaps for sampling variance. With 90 percentconfidence we should therefore observe that the coefficient for January 2014 is either in the top 10percent of the distribution when the effect is positive (e.g. price) or the lowest 10 percent when theeffect is negative (e.g. quantities), and to be outside this range when there is no effect of the tax (e.g.calories). This test is more precise when we have more coefficients/years. Unfortunately we just have3 years of data. We still think they are informative though.

Figure A24 uses the set of 34 coefficients and plots an empirical cumulative distribution function(CDF) for each different dependent variable used. We add a vertical red line to indicate the positionof the January 2014 coefficient (the true tax). We find that for the price of SDs our true tax effect is inthe top 90 percent of coefficients as expected from our previous results, whereas for HCF it is abovethe 80th percentile. For kilos of food it is in around the 30th percentile (strongly indicating a zeroeffect), while for SD the reduction is below the bottom 20th percentile. For total calories the increaseis close to the 70th percentile. The results of these tests are consistent wit the results of statisticalsignificance we got in the paper.

38

FIGURE A24. FISHER EXACT PLACEBO TEST

(a) Price per Lt (b) Price per Kg

(c) Liters (d) Kgs

(e) Calories

Note: This figure estimates 34 regressions with placebo months to mimic the tax month. It then ranks the estimated coefficient and plotsthe cumulative density of these, i.e. the fraction that fall below a particular value. The vertical red line shows where the coefficient for theJanuary 2014 lies. We expect true effects to lie at the extremes of the distribution. This is akin to a Fisher exact test.

39

C. Fraction of Total Expenditures in KWP

One potential criticism of our results is that we capture only a subset of food expenses. In orderto explore how the estimates presented vary with respect to the fraction of total purchases capturedin our dataset, we perform two exercises. First, we categorize households into quintiles of the totalconsumption of calories. Second, we classify households by a proxy for the expenditure that KWPcaptures. We then estimate the main regressions of the paper by these quintiles and show that resultsdo not vary by quintile. This lends credence to our estimates not being biased according to the shareof expenditure we capture.

Before we go to the results, let us briefly describe how we calculate the proxy for the fraction oftotal food expenditure captured by Kantar. First we noted that KWP contains a set of socio-economicvariables also captured in the Mexican Income and Expenditure Survey (ENIGH) which has totalexpenditure in food. Second we estimated a regression of total expenditures against these socioeco-nomic variables on the ENIGH sample.28 We then impute total expenditures to each household in ourKWP dataset using the predicted values from the estimated OLS regression, given the coefficients ob-tained in the regression using ENIGH and the socio-economic variables captured in KWP. By doingthis, we then obtain a (noisy) estimate of total expenditures for each household in our dataset, andare then able to compute, for each household, what fraction of total expenditures is captured in theKWP dataset. We then categorize households into quintiles of the fraction of total expenditures thatare captured in KWP and run our main specification for each of these subsamples, using the price perliter of SD, the price per 100 calories of HCF, total liters of SDs, total calories from HCFs, and totalcalories purchased.

Results are presented graphically in Figures A25 and A26, which focus on quintiles of total caloriesand total expenditures captured in KWP, respectively. As can be seen, the estimated increase in theprice of sugary drinks is (slightly) increasing with the fraction of expenditures and calories capturedin our dataset, while the drop in consumption of liters of sugary drinks is decreasing in magnitude. Asfor calories from taxed foods, the log of price is also increasing with the fraction of total expenditurescaptured, while the change in purchases of calories from these products is negative for the lowestquintile, and positive for the highest. In sum, households for which we capture a larger fraction oftheir total food expenditures exhibit a smaller decrease in calories from SDs and even a slightly largerincrease of calories from HCFs. This is reflected on the fact that total calories increase slightly morefor this group compared to our results (i.e. the mean), although this increase is also not statisticallysignificant. A similar result is obtained if quintiles are based on total calories consumed. Taken atface value, this means that if anything the decrease (increase) in calories would be less (more) if wecould capture more of the food expenditures.29

28We also tried with includying some variables of food expenditures that were in KWP and ENIGH and got similar results on theimputations.

29A caveat from this exercise is that the classification used to form the quintiles might not necessarily reflect groups for which a largerportion of the expenditure is captured, but rather households that have a diet more based on packaged processed foods, which is the bulkof the dataset. In this case the positive relation between effect on consumption and proportion of expenditures captured might only reflectthe fact that those with a diet more based on packaged products are less likely to decrease their consumption and thus will see total caloriesless affected. However, we do not feel this alternative explanation is a threat to our results for the following reasons: (i) the positive relationbetween the effect on consumption and proportion of expenditures is non monotonic; (ii) even though we only capture a portion of the totalcalories consumed, given that packaged goods are the ones affected by the tax and, as such, the main source of calories, the effect on thecalories not captured would need to be substantial to overturn the total calories result.

40

FIGURE A25. MAIN REGRESSIONS WITH CALORIC CONSUMPTION INTERACTION

.14

.145

.15

.155

Coe

ficci

ent o

f P

rice

per L

iter

For

Qui

ntile

s of

Dai

ly P

er C

apita

Cal

orie

s

Quintile 1(338 kcal)





Coefficient Confidence Interval 95%

(a) Price per Lt

.048

.05

.052

.054

.056

.058

Coe

ficci

ent o

f Pric

e pe

r Kilo

gram

F

or Q

uint

iles

of D

aily

Per

Cap

ita C

alor

ies







(b) Price per Kg

-.12

-.1-.0

8-.0

6-.0

4

Coe

ficci

ent o

f Li

ters

F

or Q

uint

iles

of D

aily

Per

Cap

ita C

alor

ies







(c) Liters

-.01

0.0

1.0

2

Coe

ficci

ent o

f K

ilogr

ams

For

Qui

ntile

s of

Dai

ly P

er C

apita

Cal

orie

s







(d) Kgs

-.04

-.02

0.0

2.0

4.0

6

Coe

ficci

ent o

f To

tal C

alor

ies

F

or Q

uint

iles

of D

aily

Per

Cap

ita C

alor

ies







(e) Calories

Note: These graphs plot the coefficient β in equation 2, for different subsamples. In this case, we classify households by quintiles oftheir consumption of total calories. All regressions include household-week of the year dummies and a quadratic time trend as controls.Each point in the graph shows the results of a different regression for the subsample specified. 95 percent confidence intervals from robuststandard errors clustered at the week level are also shown.

41

FIGURE A26. MAIN REGRESSIONS WITH EXPENDITURE PERCENTAGE INTERACTION

.13

.135

.14

.145

.15

.155

Coe

ficci

ent o

f P

rice

per L

iter

For

Qui

ntile

s of

Cap

ture

d E

xpen

ditu

re

Quintile 1(10 %)

Quintile 2(16 %)

Quintile 3(23 %)

Quintile 4(31 %)

Quintile 5(50 %)


(a) Price per Lt

.046

.048

.05

.052

.054

.056

Coe

ficci

ent o

f Pric

e pe

r Kilo

gram

F

or Q

uint

iles

of C

aptu

red

Exp

endi

ture

Quintile 1(10 %)

Quintile 2(16 %)

Quintile 3(23 %)

Quintile 4(31 %)

Quintile 5(50 %)


(b) Price per Kg

-.15

-.1-.0

50

.05

Coe

ficci

ent o

f Li

ters

F

or Q

uint

iles

of C

aptu

red

Exp

endi

ture

Quintile 1(10 %)

Quintile 2(16 %)

Quintile 3(23 %)

Quintile 4(31 %)

Quintile 5(50 %)


(c) Liters

-.04

-.03

-.02

-.01

0.0

1

Coe

ficci

ent o

f K

ilogr

ams

For

Qui

ntile

s of

Cap

ture

d E

xpen

ditu

re

Quintile 1(10 %)

Quintile 2(16 %)

Quintile 3(23 %)

Quintile 4(31 %)

Quintile 5(50 %)


(d) Kg

-.10

.1.2

Coe

ficci

ent o

f To

tal C

alor

ies

F

or Q

uint

iles

of C

aptu

red

Exp

endi

ture

Quintile 1(10 %)

Quintile 2(16 %)

Quintile 3(23 %)

Quintile 4(31 %)

Quintile 5(50 %)


(e) Calories

Note: These graphs plot the the coefficients β in equation 2, for different subsamples. In this case, we classify households by quintilesof the predicted fraction of expenditures in ENIGH that KWP captures, according to their socio-economic characteristics. All regressionsinclude household-week of the year dummies and a quadratic time trend as controls. Each point in the graph shows the results of a differentregression for the subsample specified. 95 percent confidence intervals from robust standard errors clustered at the week level are alsoshown.

42

D. Industry level production

To strengthen the evidence from section VII.C, here we present two additional plots. Figure A27 givesevidence that soda exports and imports are not a concern in terms of validity of our estimates giventhat the proportion of sodas imported and exports is negligible and did not exhibited an importantchange after the tax was introduced. It also gives graphical evidence of the evolution of prices andquantities of sodas nationally with the data used in Table IX. As the figure shows, prices suddenlyincreased their level at the date in which the tax began to be implemented and consumption exhibiteda slight decrease.

FIGURE A27. COMMERCE AND NATIONAL PRODUCTION OF SODA

0.2

5.5

.75

SD

Impo

rts a

s P

erce

ntag

e o

f Tot

al P

rodu

ctio

n

11.

52

2.5

SD

Exp

orts

as

Per

cent

age

of T

otal

Pro

duct

ion

2009 2010 2011 2012 2013 2014Year

Exports Imports

(a) International Commerce

-.2-.1

0.1

.2Lo

g vo

lum

e (1

000'

s lt)

, s.a

.

4.4

4.5

4.6

4.7

4.8

4.9

Log

soda

pric

e in

dex

01/01/6 01/01/8 01/01/10 01/01/12 01/01/14 01/01/16Date

Log Soda Price log Liters

(b) National Production

Note: Source: INEGI

E. Changes in nutritional content

An interesting question that is widely unexplored in the literature is whether taxes of this kind haveeffects, not through changes in the kilograms of food or liters of sugary drinks purchased by con-sumers, but in changes in products’ nutritional content. This supply side effect could potentially bean important channel to reducing caloric intake and obesity above and beyond price-induced changesin demand. Since the tax on SD is per liter, there are few incentives for producers to reformulate anddecrease caloric content unless they drive added sugar to zero; but the tax on HCFs implies incentivesto manipulate products’ nutritional content, as the tax applies only to foods with caloric densitiesabove the 275cal/100gr threshold.

A priori, it is hard to predict if a tax of this magnitude is enough to generate changes in products’caloric content, since this depends on the cost for producers to change product contents and howthese changes may affect demand. We can test if producers changed the caloric content of barcodes.The main challenge is that we only have a one-time cross-section of barcode nutritional informationfrom about 1.5 years after the tax came into effect, so we cannot conduct a before/after comparisonat the barcode level. We therefore rely on an excess mass test. Under the assumption that the cost ofdecreasing calories is smoothly increasing in the size of the caloric decrease, because the tax incentiveis a threshold incentive we should expect excess mass on the left side of the 275 calories per 100 gramsthreshold.

Figure A28 plots the histogram of caloric densities for all food barcodes in our dataset. No excessmass is observed to the left of the tax threshold. This is confirmed by a McCrary test. No significantreformulations seem to have taken place 1.5 years after the tax on HCFs came into effect.

43

FIGURE A28. SUPPLY SIDE RESPONSE ON CALORIC DENSITY

0.0

01.0

02.0

03.0

04D

ensi

ty

0 200 400 600 800 1000Calories per 100 grams, all food items

Note: This figure plots the frequency of caloric densities of all barcodes in the Kantar World Panel. Caloric densities are measured inkilocalories per 100 grams since this is the unit used in the tax definition for HCFs. The red vertical line indicates the caloric densitythreshold (275 kilocalories per 100 grams) that was used for the HCF tax definition.

A12. REFERENCES

Atkin, David, “Trade, Tastes, and Nutrition in India.,” American Economic Review, 2014, 103 (5),1629–1663.

Banks, James, Richard Blundell, and Arthur Lewbel, “Quadratic Engel Curves And ConsumerDemand,” The Review of Economics and Statistics, 1997, 79 (4), 527–539.

Deaton, Angus and John Muellbauer, “An Almost Ideal Demand System,” American EconomicReview, 1980, 70 (3), 312–26.

Hausman, Jerry, Gregory Leonard, and J. Douglas Zona., “Competitive Analysis with Differen-ciated Products,” Annales d’conomie et de Statistique, 1994, (34), 159–180.

JE, North J. Wansink B. Painter, “Bottomless bowls: Why visual cues of portion size may influ-ence intake,” Obesity Research, 2005, 1 (13), 93–100.

Nielsen, S. J. and B. M. Popkin, “Patterns and trends in food portion sizes,” Journal of the AmericanMedical Association, 2003, 289 (4), 450–453.

LACEA WORKING PAPER SERES. No....

Documents

Transcript of LACEA WORKING PAPER SERES. No....