JournalofMarketingResearchArticlePostprint

2011,AmericanMarketingAssociation
Allrightsreserved.Cannotbereprintedwithouttheexpress
permissionoftheAmericanMarketingAssociation.

Fast Food Consumption and the Ban on Advertising

Targeting Children: The Qubec Experience1
Tirtha Dhar
Assistant Professor, Division of Marketing, Sauder School of Business
2053 Main Mall, University of British Columbia
Vancouver, BC, Canada, V6T1Z2
email: tirtha.dhar@sauder.ubc.ca
Kathy Baylis
Assistant Professor, Department of Agriculture and Consumer Economics
302B 1301 W. Gregory Dr., University of Illinois
Urbana, IL 61801-3605
email: baylis@illinois.edu
April 8, 2011

1 The

authors would like to thank Sohini Paul and Elizabeth France for their excellent research
assistance on this project. We are grateful to Charles Weinberg, Sonia Angell, Don Fullerton,
Je Perlo, Jon Skinner, and Ellen Goddard for comments. Thanks also go to the UBC-Hampton
Grant for providing funds to conduct this research. Principal authorship is equally shared.

Abstract
Amid growing concerns about childhood obesity and the associated health risks, several
countries are considering banning fast food advertising targeting children. In this paper the
authors study the eect of such a ban in the Canadian province of Quebec. Using household
expenditure survey data from 1984 to 1992, authors examine whether expenditure on fast
food is lower in those groups aected by the ban than in those who are not. The authors use
a novel triple dierence-in-dierence methodology by appropriately dening treatment and
control groups and nd that the bans eectiveness is due not to the decrease in fast food
expenditures per week but to the decrease in purchase propensity by 13 percent per week.
Overall, the authors estimate that the ban reduced fast food consumption by US$88 million
per year. The study suggests that advertising bans can be eective provided media-markets
do not overlap.
Keywords: Advertising regulation, fast food, obesity, dierence in dierence estimator.

Title: Fast Food Consumption and the Ban on Advertising

Targeting Children: The Qubec Experience
Childhood obesity is a growing problem, and governments in dierent countries are considering a variety of policy solutions, including banning advertisements on so-called junk
food. Obesity puts children and adolescents at risk for a range of health problems such as
cardiovascular disease, diabetes, and depression (Krebs and Jacobson, 2003), making obesity
second only to smoking as a cause of preventable death (McGinnis and Foege, 1993; Allison
et al., 1999). Fast food is identied as one of the key drivers of this problem by obesity
researchers, as it signicantly increases caloric consumption per meal (Satia et al., 2005;
Bowman and Vineyard, 2004; Niemeyer et al., 2006; Paeratakul et al., 2003). For example,
one study of 11- to 18-year-olds found that regular consumption of fast food is associated
with ingesting an extra 800 calories per week for boys and 660 extra calories per week for girls
(French et al., 2001). These extra calories translate into a possible weight gain of 10 pounds
or more per year. Further, a study of Mexican children in San Diego found that 4 to 7 year
olds who ate at fast-food restaurants were twice as likely to be obese as those who did not
(Duerksen et al., 2007). In fact, from 1977 to 1996, calorie intake from fast-food restaurants
doubled as a percentage of energy intake for Americans over the age of two (Nielson et al.,
2003).
Fast food is also one of the most heavily advertised product categories targeting children
and according to recent studies such advertising is eective in changing behavior (Institute
of Medicine, 2006; Connor, 2006). For example, Taveras et al. (2006) show that in the
United States, children who view fast food television advertisements are about 50 percent
more likely to eat fast food. Thus, advertising plays a critical role in a households decision to consume fast food and thereby eects health outcomes. Advertising can inuence
obesity in two ways: by encouraging the consumption of unhealthy food and by increasing
advertising-supported TV programming and an associated sedentary lifestyle.1 As a result,
1

For more details, see Crespo et al. (2001); Dietz and Gortmaker (1985); Gortmaker et al., (1996);

a number of countries have responded by either implementing or proposing sweeping restrictions on fast food advertising targeting children. In February 2007, the U.K. banned
junk-food advertising to children, and in 2008, in a report to Congress, the U.S. Federal
Trade Commission recommended that companies restrict advertising to children to healthier
food products (Federal Trade Commission, 2008). Besides the recent initiatives in the U.K.
and the U.S., various forms of advertising bans already exist in some jurisdictions, such as in
the province of Qubec in Canada, along with similar bans in Sweden, Norway and Greece.
Advertising lobby groups argue that despite the ban, children in Qubec are no less obese
than children in other parts of Canada (The Times, May 31, 2004); proponents of advertising
bans, on the other hand, note behavioral studies show that "kiduence" can aect household consumption, and that advertising targeting children is eective in altering consumption
choices (Institute of Medicine, 2006).2 In this paper, we use a quasi-experimental setup and
household-level data to examine whether the Qubec advertising ban, in force since 1980,
has had an eect on consumption of fast food. A better understanding of this connection is
the rst step in comprehending the complex linkage between advertising, consumption and
health-related problems.
Although the Qubec law is widely referenced by both opponents and proponents of advertising bans, very little research has been conducted on the eect of the ban. Goldberg
(1990) published the rst paper analyzing the impact of the ban using a quasi-experiment. He
uses language spoken by children at home to identify the eect of the ban on consumption
behavior, noting that English-speaking (hereafter Anglophone or AP) children in Qubec
have more access than their French-speaking (hereafter Francophone or FP) counterparts to
media from outside Qubec, and are therefore, less likely to be aected by the ban. Interviewing children in Qubec, he nds that AP children have stronger toy-brand recognition
than FP children and, further, that AP children with access to television from the United
States could correctly identify more toys and have a larger number of child-targeted cereal
Boynton-Jarrett et al. (2003); Giammattei et al. (2003); Halford et al. (2004) and You and Nayga (2005).
2
For more details on the global regulatory environment please refer to Hawkes (2007).

brands in their homes. Goldberg concludes that the law is successful in reducing childrens
exposure to cereals and toys and, therefore, in reducing the pressure from children on their
parents to buy them. However, he does not look at the eect of the ban on actual consumption patterns. Further, he compares only FP and AP children within Qubec; as a result,
if the dierence in brand recognition is due to unobserved cultural factors, then the eect
may not be correctly identied. We overcome this problem by comparing household-level
consumption behavior in Qubec with comparable households in the neighboring province of
Ontario. In our natural-experimental setup, we use survey data on expenditure to analyze
the eect of the ban, using the fact that the ban is applicable only in Qubec and not in
Ontario.
In terms of approaches to studying advertising regulations targeting children, our article deviates from the existing studies with respect to both sources of data and estimation
methods. In fact, most proponents of advertising bans refer to the literature in marketing
and child psychology.3 The literature in this area, mainly based on laboratory experiments,
generally nds strong evidence that product promotion to children encourages the consumption of unhealthy food. One weakness of behavioral research in this context, however, is that
controlled behavior in laboratories may not be representative of behavior in the real world.
Thus, one might be concerned about the external validity of the research. In contrast to
the existing studies on advertising regulation targeting children, our goal in this paper is
to study the impact of an advertising ban using household-level eld data. Specically, we
choose fast food as the product category to measure the impact of this ban.
Because we are examining the eectiveness of the advertising ban on expenditure, our
study is also a study of the eectiveness of advertising. Behavioral researchers tend to focus
on the impacts of advertising on the consumers decision process, while most empirical quantitative studies tend to focus primarily on the eectiveness of brand-level advertising.4 Early
quantitative studies on advertising eectiveness mainly use highly aggregated data, either at
3
4

For comprehensive reviews see Institute of Medicine, 2006; Hastings et al., 2003.
For a good review of the behavioral process literature, see Vakratsas and Ambler (1999).

the brand or at the product level, using OLS or simultaneous-equations estimation methods.
Assmus, Farley, and Lehmann (1984) provide an excellent meta-analysis of the early studies.
They nd that these early studies have imperfect "quasi-experimental designs" and do not
isolate the eect of advertising. Lodish et al. (1995) overcome some of the shortcomings
mentioned by Assmus et al. by using proprietary BehaviorSan

matched household data

to estimate the eectiveness of advertising on brand level consumption behavior. In our

study, we control for household-level demographic characteristics as part of the estimation
process and use three key sets demographic characteristics (location, language spoken in the
household, and household composition) to identify the eect of advertising. Unlike Lodish
et al., we also decompose the consumption decision into stages: decision to purchase and
amount spent. Some of the recent research on advertising concentrates more on the strategic
implications of brand-level advertising, or advertising awareness, than on its eectiveness per
se (e.g., Jedidi et al., 1999; Steenkamp et al., 2005; Naik et al., 1998). As a category-level
study, our research complements the recent studies on the eectiveness of advertising at the
brand level.
To measure the eect of the ban, we estimate the dierence in fast food expenditure
between our "treatment" and "control" groups. In the program-evaluation literature, this
approach of estimating the eect of a program, or "treatment", is known as a dierence-indierence estimator. The use of dierence-in-dierence estimators has a long history in labor
economics literature (see, e.g., Ashenfelter and Card, 1985; Lalonde, 1986; Card and Krueger,
1994). In this case, the advertising ban is the treatment, and we identify a treatment group
(i.e., those aected by the ban) and control groups (i.e., those not aected by the ban).
Following Goldberg (1990), we use language as a primary variable to distinguish treatment
and control groups, but rather than simply measuring the eect within Qubec, as Goldberg
did, we compare households across dierent geographic locations and compositions (e.g.,
households with and without children). We believe this is one of the rst applications of
the dierence-in-dierence method to estimate advertising eectiveness within the marketing

literature.
With respect to estimation, the key challenge is the fact that household-level consumption
data do not exist for the period before the ban. To overcome this limitation, we rst identify
the groups highly likely to be unaected by the ban, both within and outside Qubec, and
then compare their consumption behavior with that of the group most likely to be aected.
Specically, we rst test whether fast food expenditure is signicantly dierent for FP and
AP households within Qubec than for FP and AP households in Ontario, then compare
households with children to those without children.5 We also consider whether the ban
continues to aect the consumption patterns of young adults who grew up under the ban
and who are now exposed to advertising. After controlling for individual-level dierences,
our results imply a signicant eect of the ban in terms of lower fast food consumption.
Interestingly, the main source of the eect is in terms of the number of purchase occasions,
not in terms of the amount spent per week. In other words, aected households spend less
on fast food per week because they go out for fast food less often, not because they spend
less on each occasion.
To estimate the eect, we use Statistics Canadas detailed household-level expenditure
survey data over four years. This approach is a distinct departure from existing studies
measuring the eect of advertising regulations, which are primarily based on cross-sectional
surveys or experiments or use country-level data.6 Combining data from the Canadian food
expenditure survey (Foodex) and the household expenditure survey (FamEx) from 1984,
1986, 1990 and 1992, we ask whether consumption of fast food changed as a result of the
ban. To the best of our knowledge, this is the rst formal study to explore the impact of
the Qubec law on household level expenditure using population-representative household
consumption data. Our research is also one of the rst to study the regulation of advertising
5

A similar triple dierence-in-dierence (DDD) approach has been used by Gruber (1994) to study the
eect of maternity benets. For an interesting exposition of DDD please refer to Hamermesh and Trejo
(2000).
6
In economics and marketing, tobacco advertising bans have been extensively studied. For a comprehensive review of this literature, please refer to Saer and Chaloupka (2000).

targeting children using eld-level data.

Background
Advertising targeting children has always been a contentious social policy issue. Social
psychologists have argued that advertising can have a harmful inuence on childrens consumption decisions (Singer and Singer, 2001), leading the American Psychology Association
to support a policy in favor of restrictions on advertising targeting children under the age
of eight.7 Similar concerns that children are not able to process advertising rationally and
this led the Qubec government to introduce the Qubec Consumer Protection Act, which
bans advertising targeting children under the age of thirteen in 1978. The law came into
eect two years later on April 30, 1980. Under this law, products and programs are rated
according to their appeal to children, and products such as toys and fast food for children
cannot be advertised during childrens programs on television or in newspapers, magazines,
or in any other media targeting children. The law applies to both print and electronic media,
although debate and enforcement of the law have mainly focused on television. Articles 248
and 249 of the Act explain the criteria used to determine whether an advertisement targets
children, which include:
[a] the nature and intended purpose of the products advertised (for example, are the
products consumed primarily by children?)
[b] the advertisement itself (does it use fantasy, magic, or children-specic adventures?)
[c] the time and place the advertisement is shown.
During television programs for which children constitute more than 15 percent of the
audience, advertisements targeting either children or a mix of children and adults, are not
permitted. During programs for which children constitute more than 5 percent of viewers,
advertisements directed specically at children cannot be broadcast. Note that the law does
not ban all advertising to children; advertisements can still be broadcast during childrens
7

For details please refer to http://www.apa.org/releases/childrenads.html

programs, as long as they are non-commercial (e.g., public-service announcements) or are for
products and services targeting adults. Thus, ads for cars or detergents are allowed during
childrens shows. Similarly, advertising exclusively targeting children can still be broadcast,
but only during programs that are primarily watched by adults (Caron, 1994). Thus, fastfood chains such as McDonalds can still advertise during late-night shows, but not during
afternoon cartoons. Data on viewing levels and audience composition are compiled by the
Bureau of Broadcast Measurement (BBM Canada) and are used to determine the nature
of the program. Advertisements must be cleared with Qubecs O ce de la Protection du
Consommateur (Consumer Protection O ce and hereafter OPC) before being broadcast.8
How well the ban has been enforced can certainly be questioned. Anecdotal evidence
does suggest active enforcement of the law. For example, in 1984, McDonalds proposed a
commercial in which Ronald McDonald would explain to children the importance of wearing
seat belts; this commercial was refused by OPC because it involved a known and well-liked
childrens character (Government of Canada and Gouvernement du Qubec, 1985). In another case, in early 2007, Saputo, one of the largest baked- and dairy-goods producers in
Canada, sent promotional material to Qubec day-care centers featuring Igor the Gorilla, the
brand mascot of their baked mu n products; consumer advocates argued that the material
was in violation of the advertising ban (Kucharsky, 2007) and as a result, OPC successfully
sued Saputo Inc. (National Post, 2009). Similarly, Burger King was recently sued in connection with its campaign targeting children, and in the end, agreed to a ne and stopped
the campaign (The Gazette, 2009).
Depending on the source, the net loss to Qubecs advertising market from TV advertising
is estimated to be between $3.9 million and $8.2 million per year (Caron, 1994). Anecdotal
8

Like the current proposed bans, the Quebec law was not without controversy. In one of the most famous
free speech cases in Canada, the advertising ban was challenged in Irwin Toy v. Quebec; the Supreme Court
of Canada upheld the ban in 1989, arguing that if children cannot eectively analyze advertisements, then
advertisements to children are inherently misleading. Debate around the ban continues. In the August 15,
2005, issue of Marketing Magazine, an article titled "Time to End It?" discusses the perceived drawbacks of
the ban for Quebec TV programming (Kucharsky, 2005). These concerns were echoed at Canadian Federal
Standing Committee hearings in the Senate earlier in 2005, resulting in calls for the law to be revised or
repealed.

information suggests that in some cases, rms with products targeting children stopped
developing advertising, although there is no evidence of rms exiting Qubec as a result of
this ban.9

One of the legislations weaknesses is that it applies only to media originating

inside Qubec, and thus the advertising ban does not apply to signals originating from
the neighboring Canadian province of Ontario or from the United States. We exploit this
weakness to identify the eect of the ban.

Database
We use data from Statistics Canadas food expenditure survey (Foodex) which provides
detailed information on the bi-weekly food-purchasing behavior of households. Respondents
participating in the survey keep a detailed diary of all food expenditures. In the FoodEx
survey, households use a daily food expenditure diary for two weeks including the number
and type of meals consumed, and the amount spent on these meals. Statistics Canada then
makes aggregated weekly les available for research. In this paper, our focus is the FoodEx
expenditure category meals at fast-food restaurants.10 Finally, Statistics Canada has derived
a set of household weights for use with the publicly available FoodEx data le that take into
account the survey design and the non-response rate. When weighted, the sample is generally
representative of the Canadian population. All results presented in this paper incorporate
these weights. The equivalent survey in the United States is the Consumer Expenditure
Survey conducted by the Bureau of Labor Statistics.
For the purposes of this research, both pre- and post-ban expenditure data would have
been ideal. Unfortunately, Foodex survey data prior to 1984 are not available for research,
meaning that we have no observations of consumption behavior before the ban. Instead, we
use data from the 1984, 1986, 1990, and 1992 surveys for the purpose of our research and
9

For further details please refer to Rapport du Comit Fdral-Provincial sur la Publicit Destin aux
enfants, (Government of Canada and Gouvernement du Qubec, 1985).
10
Statistics Canada broadly denes as fast-food restaurants as places where there is no table service, only
self service, and food is provided in a minimal amount of time. In the survey questionnaire subjects were
specically asked about their expenses at the "fast-food restaurants".

cross-sectionally compare households in dened treatment and control areas. Note that one
of the determinants of the treatment group, mother tongue, is recorded in surveys only up
to 1996, so we cannot use data from more recent surveys for our analysis. We chose only
the neighboring province of Ontario as a control for Qubec because the two provinces have
similar economic and socio-demographic characteristics and because Ontario has a relatively
large Francophone population. Since most information on television consumption behavior
during the period of our study comes from urban areas, we focus our study on large urban
areas in both provinces (i.e., those with a population of more than 100,000).11 Another
reason for dropping small cities and rural areas is that in more remote areas there may be
pockets of FP or AP neighborhoods where distance and lack of transportation make fast
food less available, constraining access to supply.
To keep the sample representative, we delete ve households with fast food expenditures
of more than $150/per week, which is 40 times the weekly average household expenditure.12
We also drop 94 households with no food expenditure. Our nal sample consists of 9177
households (5024 in Ontario and 4153 in Qubec). The four years of data year are stacked,
creating a pooled data set. Population-weighted summary statistics for all the households in
the sample are presented in Tables 1A and 1B. Across most demographic characteristics,
including age, household composition and occupation, the comparable household groups
(e.g., FP and AP households with children in Qubec) are quite similar.
In terms of mother tongue, we classify the households into four types: Francophone (FP)
households, where both spouses are French speaking, Anglophone (AP) households, where
both spouses are English speaking, Allophone households (OP), where both spouses speak
neither English nor French, and mixed households, where spouses have dierent mother
tongues. We drop OP and mixed households, as previous research and anecdotal evidence
11

We are also restricted by the fact that the Foodex and Famex surveys only include urban households in
1984 and 1990. So, we drop households from rural areas included in the 1986 and 1992 surveys. In terms of
population distribution, in 1991, 78 and 82 percent of the population lived in the included cities in Quebec
and Ontario respectively.
12
Retaining these households in our sample did not substantively change our results.

do not provide any guidance in terms of their media and fast food consumption behavior.
Over our time frame, unilingual AP and FP households comprise 69 and 5 percent of all
households in Ontario and 8 and 80 percent of the households in Qubec respectively. By
restricting ourselves to households in which both spouses have the same mother tongue, we
lose about 4 percent of households in Qubec and 6 percent of households in Ontario.

13

Note that in terms of media, one of the key assumptions underlying our identication
strategy is that FP children do not spend a signicant amount of time watching English
television channels compared to AP children. Currently, data on viewership during the period
of study is not publicly available but based on studies of Canadian television consumption
behavior, we believe there is strong evidence that this is the case. In one of the most
comprehensive studies of the impact of the ban on media, Caron states that before the
imposition of the ban, FP children spent only 6 to 7 percent of their viewing time on
English-language programming and that this proportion remained the same after the ban.
Caron also notes that AP children spent a large amount of time watching English broadcasts
that largely originated from the United States. In another study using data from 1987 and
1993 data from Montreal the largest city in Qubec, with 21 percent of the provinces
total population De La Garde (1996) notes that in 1987, FP households viewed French
programming 88 percent of the time and that close to 100 percent of it was supplied by
Qubec-based television stations. This proportion increased to 92 percent in 1993. On the
other hand, English-speaking households spent more than 90 percent of their viewing time
on English programming, which mainly originated from outside the province.
Recent data continue to support this argument. In 2007, Canadian-produced programs
dominated the list of popular prime-time drama/comedy programs in Qubec holding six
of the top ten positions, including all top three spots. All these programs are in French, and
13

When we ran the regression including bilingual households, our results were qualitatively unchanged,
although the magnitudes of the eect of the ban were slightly smaller. When we included OP households,
we found that they tended to have similar consumption patterns to the dominant-language group in each
province. Thus OP households in Ontario had similar consumption patterns as their AP counterparts, while
OP households in Quebec largely acted like their FP neighbours.

10

even the non-Canadian programs are dubbed in French, and transmitted by Qubec-based
television stations. Note that the bias in favor of Qubec-based French television holds not
only for childrens programming but also for adult programming (Variety, 2009). In fact,
dissimilarity in media consumption combined with similarity in brands and other product
consumption by Qubecs French-speaking consumers is turning Qubec into one of North
Americas most ideal geographic locations for test marketing (Advertising Age, 2009).

Estimation Strategy
As mentioned, because we do not have data from before the imposition of the ban, we
compare households cross-sectionally by carefully dening the treatment and control groups.
In the program-evaluation literature, Madrian (1994) uses a similar cross-sectional approach
to identify treatment eect when considering the link between job mobility and health care
benets. We dene the treatment and control conditions in the following three dimensions
to estimate the eect of the ban:
[1] By Language: Because FP households primarily consume media from Qubec-based
French-language media sources, while AP households tend to consume English-language
media from outside the province, we would expect FP households to consume signicantly
less fast food than AP households in Qubec if the ban is eective. In terms of a generic
regression model,14

(1)

Y = f (F r)

where Y is the dependent variable (this can be either the households decision to purchase
fast food or total fast food expenditure conditioned on decision to purchase), and F r is
the dummy variable for FP households (F r = 1 if FP households, 0 if AP households).
Following the literature in program evaluation, the eect of the ban can be estimated as
the rst dierenced estimator as follows:
14

Fr

= YbF r

YbAn , where YbF r and YbAn are the

To focus on the intuition in the exposition we avoid household and province subscripts.

11

estimated average purchase incidence or fast food expenditure of FP and AP households

respectively. Note that this is the underlying model that Goldberg (1990) used to measure
the eect of the ban. One problem with this approach is that if there are persistent intrinsic
dierences between AP and FP households that cause the dierences in expenditures, then
this approach will bias the estimated eect of the ban. As noted above, one approach to
overcome this problem would be to use data from both before and after the ban was imposed.
Because we do not have pre-ban data, however, we use an alternative approach to control for
these potentially intrinsic dierences and classify consumers over another two dimensions:
province and family composition.
[2] By Province: The ban applies to the province of Qubec but not to the neighboring
province of Ontario. This implies that neither FP nor AP households in Ontario come under
the purview of the ban. We expect that if the media-consumption habits of AP households in
Qubec and Ontario are similar, there should not be a signicant dierence in expenditures
between AP households in the two provinces. Similarly, FP households in Ontario are
exposed mainly to media originating in Ontario, and, therefore, their expenditures will be
closer to those of AP households and dierent from those of FP households in Qubec. Using
this approach to measure the eect we respecify the generic regression model as:

(2)

Y = f (F r; Q)

Here Q is the dummy variable for Qubec (Q = 1 if Qubec, 0 if Ontario). In this case,
the dierence-in-dierence (DD) estimator can be expressed as:
YbF r;ON

YbAn;ON , where YbF r;Q

YbAn;Q

and YbF r;ON

F r;Q

YbAn;ON

= YbF r;Q

YbAn;Q

are the estimated av-

erage dierences between FP and AP households in purchase incidence or expenditure in

Quebec and Ontario respectively. One limitation of our data is that we can observe only
household-level expenditure, not the expenditure specic to children. If we consider differences only in households with children, we would not be able to discern whether these
dierences were due to the dierent consumption levels of the children or to the dierent
12

consumption levels of the adults in those households. Therefore, we add another criterion
to classify the households in the sample.
[3] By Children. Note that the nature of the ban implies that households with children
will be aected more than households without children. Specically, if the ban is eective,
FP households with children in Qubec will be the most aected. Thus, to control for
potential dierences in the consumption of adults, we add a dummy variable for households
with children (C = 1 for households with children, 0 for households without a child):
(3)

Y = f (F r; Q; C)

In this case the triple dierence in dierence (DDD) estimator can be expressed as follows:
h
i
b
b
b
b
YF r;Q;C YA;Q;C
YF r;ON;C YA;ON;C
h
i
(4)
F r;Q;C =
YbF r;Q;N C YbA;Q;N C
YbF r;ON;N C YbA;ON;N C

where

YbF r;Q;C

YbA;Q;C ,

YbF r;ON;C

YbA;ON;C

are the estimated average dierences be-

tween FP and AP households with children in purchase incidence or expenditure in Quebec

and Ontario respectively; similarly YbF r;Q;N C

YbA;Q;N C , YbF r;ON;N C

YbA;ON;N C

are the

estimated average dierences in case of households without children. So, in terms of experimental design, we ultimately create eight groups (two provinces

two languages

two

types of households). As a result, if the ban is eective, we should nd that the dierence
between FP and AP households with children in Qubec will be both larger than the equivalent dierence in households in Ontario, and larger than the dierence between FP and AP
households without children.15 Below we provide descriptive statistics of the dierences in
expenditures among these eight groups.
Figure 1A shows fast food expenditures per week by the eight household groups. Note
that across the groups, households in Qubec spend less than households in Ontario. We
observe the largest dierence between our key comparison groups, FP and AP households
15

Note that in the case of linear regression, the parameter associated with a three-way interaction term
can capture this triple-dierenced eect. If the function is linear, such that Y = + F r F r + Q Q + C C +
then F r;Q;C = F r;Q;C .
F r;Q F r Q + F r;C F r C + Q;C Q C + F r;Q;C F r Q C +

13

with children in Qubec; FP households spend CDN$2.19 less per week than their AP
counterparts. This result can be thought of as the rst dierenced estimate. To control
for unobserved cultural eects, we need to adjust this estimate to take into account the
dierence between similar language groups in Ontario. In Ontario, the dierence between
FP and AP households with children is -CDN$0.75. Note that in this case, FP households
with children are not under the purview of the ban, so this dierence may be due to cultural
or other inherent dierences between FP and AP households with children. After adjusting
for this dierence, the dierence-in-dierence (DD) estimate reduces to -CDN$1.44. Still,
this dierence cannot be attributed entirely to the advertising ban, because it may result from
inherent dierences in expenditures on the part of adults in these households. Therefore, we
rst estimate similar DD estimates for the FP and AP households without children and then
use this estimate to adjust the estimate for the households with children. The DD estimate
for the households without children is CDN$0.32. Based on these simple weighted average
expenditures, the DDD estimate will be -CDN$1.76.
Next, Figure 1B shows the average fast food expenditures after excluding households
that do not purchase fast food during the period of the surveys. Note that in this case, the
dierences between comparison groups decrease except in the case of dierence between FP
and AP households with children in Qubec. The pattern of the dierences in all other cases
suggests that once households decide to purchase fast food, there are no large dierences in
the levels of expenditure. In the case of FP households with children in Qubec, this pattern
suggests that it is possible that conditional on their decision to purchase fast food, they still
spend less than AP households with children. Figure 2 plots the percentage of households
that bought fast food at least once during the period of the surveys. Again, we observe
a similar pattern: a smaller number of FP households than AP households purchased fast
food, with the largest observable dierence between FP and AP households with children in
Qubec. Based on these population-weighted averages, the simple average DDD implies that
FP households with children in Qubec have an 8.59 percent lower propensity to purchase fast

14

food in any given week. Note that the dierences presented in Figures 1 and 2 do not control
for key demographic and seasonal dierences. We therefore add numerous demographic
covariates to control for such dierences in our regression analysis.
Next, to simplify the exposition, let
dene our treatment and control groups:

be the vector of dummies and interaction terms that

Fr Q C Fr

Q Fr

C Q

C Fr

We can now express our dependent variable Y as

Y =f(
where

+Z )

(5)

is the vector parameters associated with , and Z is the vector of control variables

and vector

is the associated parameters. As mentioned earlier we estimate the eect of the

ban at two stages, beginning with the level of decision to purchase. We model this rst stage
as a probit model where the dependent variable is the decision to purchase fast food within
a week. Conditional on the decision to purchase, in the second stage, we model the amount
spent per week on fast food. We estimate these two stages simultaneously after taking into
account any correlation in the errors of the two stages of the model. Note that our model
is a variation of the model proposed by Heckman (1976, 1979). The Heckman model is,
in fact, a generalization of the widely used but more restrictive Tobit censored regression
model (Amemiya, 1985). The model in equation (5) above will be a Tobit model if we restrict
the coe cients and regressors to be the same for both purchase and expenditure decisions.
Further, the Heckman model facilitates the use of dierent covariates in the two stages and
provides consistent parameter estimates in the presence of heteroskedasticity (Amemiya,
1985). To check for robustness, we also estimate the eects using OLS.
We use the following set of covariates (Z) to estimate the model. In terms of demographic
variables, we include number of children in the household, household income, number of
household members, occupational category for the head of household and his/her spouse
(blue-collar or manufacturing occupation, pink-collar or service-sector occupation, no occupation; we use white-collar or professional occupations as the excluded category), home
ownership (1 if household owns the home and 0 otherwise), social assistance (1 if household
15

C .

receives social assistance and 0 otherwise), level of education of the male and female heads
of the households (1=less than 9 years education, 2=some secondary education, 3=some
post-secondary education, 4=post-secondary certicate or diploma, 5=university degree),
dual-income households (1 if both female and male household heads are income earners and
0 otherwise), immigration status of spouses (1 if immigrants and 0 otherwise), and age of
heads of the household. Besides these demographic variables, we also use yearly dummy
variables, using 1984 as the base year, and quarterly dummies, with the fourth quarter as
the base, to control for year- specic and seasonal eects on the outcome measures.
In the case of two other covariates (i.e., price and cable TV ownership), we use information
from the existing databases to create two new variables. In terms of price, FoodEx data les
only provide information on the amount spent and number of occasions per week by meal
type (i.e., breakfast, lunch and dinner). So, for the households with a purchase history
during the period of the survey, we divide their total fast food expenditure by the number
of meals consumed within a week to calculate the price. For those households that did not
purchase fast food we use an imputed price. To impute the price of a fast-food meal, we
rst identify the households that purchase only one type of fast food meal (i.e., breakfast,
lunch or dinner) during the week, which allows us to observe the specic price they paid for
that type of meal. Using these households, we estimate the median price of the meal for
the given meal type by region and by year. To estimate a weighted average price for a fast
food meal, we calculate the proportion of breakfasts, lunches, and dinners purchased in the
region that year, and use those proportions as weights. Using these weights multiplied by
the median meal prices, we calculate the weighted average price of a fast food meal for each
province in each year. We use this as the imputed price for those households that do not
report any fast food consumption. As a robustness check, we also use the estimated average
breakfast, lunch, or dinner price per trip by province, week, and year. Qualitatively, the
results presented here are robust to the change in price measurement. Further, we note that
when considering the recorded prices paid for fast food, we nd no signicant dierences

16

between our key treatment and control groups. For example, in 1992, the average fast food
meal cost a FP household in Qubec was CDN $4.64, while their AP neighbors spent an
average of $4.58. The median price is the same for both groups. To make prices across years
comparable, we deate all food prices using the CPI for restaurant food to put prices in 1992
Canadian dollars.16
As previously mentioned, one factor that can aect consumption is exposure to media, particularly television. None of the available databases have detailed information on
household-level television viewing patterns. Of the available databases, only the biannual
Canadian Household Expenditure Survey (i.e., FamEx) provides information on detailed
yearly expenditure on cable and satellite TV subscriptions, but does not contain explicit
information on the amount of time spent in front of the television, or the type of programming viewed. Another estimation challenge is that since the FoodEx and FamEx surveys
are conducted on dierent samples, we cannot directly observe access to television by the
households in the FoodEx. Instead, we project television ownership from Famex to Foodex
data using the following approach: by year and by province, we estimate a probit model
of access to cable television as a function of household characteristics and sampling weights
using the Famex data and then using the same set of characteristics and sampling weights
in the Foodex les we predict the probability (i.e. the propensity score) of cable television
subscription. In 1992, 65 percent of Canadian households in our database paid for cable
television. By relying on cable fees to proxy for television ownership, our proxy for television
ownership will most likely be an underestimate.17
Note that the rst stage of the Heckman model, which is a probit, is highly non-linear in
its parameters. As a result, the estimated three-way interaction parameter will not, in and of
itself, capture the eect of the ban. The eect of the ban is instead given by the dierences
in the probability of purchase between treated and control groups (i.e., the dierences in the
16

Further details on the methods to construct price variable is available from the authors.
Our data also includes whether a renter received cable for free. These houesholds are included as having
access to cable. Further details on the imputation of the cable TV is available from the authors.
17

17

CDFs under treatment and control conditions). Similarly, given that we use the natural log
of expenditure in the second stage of the Heckman, a similar approach is used to estimate
the second-stage dierence-in-dierence eect on fast food expenditures.

Model Estimates
We use the full information maximum likelihood method to estimate the Heckman model.
Note that the model can be identied under either exclusion restrictions or parametric assumptions (Wooldridge, 2002). To check the robustness of our results, we try both. Qualitatively, we obtain similar results with and without exclusion restrictions. Since we do not
have strong empirical or theoretical reasons to exclude variable(s) in either stage of the estimated model, here we present only the model without exclusion restrictions.18 Because
the log of non-zero expenditure data closely approximates a normal distribution, we use the
natural log of expenditure in the second stage. We also use population probability weights
and cluster errors by region and by year to correct for the sampling procedures used in the
survey.19
Table 2 presents the parameter estimates using three regression techniques: OLS, Tobit
and Heckman. In the case of OLS, we estimate the decision to purchase (i.e., Purchase
Decision in Table 2) and the expenditure (i.e., ln(Expenditure)) equation seperately. In the
case of the Tobit model, the parameters for the rst and second stages are assumed to be the
same, so the estimated results are presented as a single equation. In the case of Heckman
model, the two stages are estimated simultaneously (i.e., the decision to purchase and the
amount spent) but relaxing the Tobit assumption of equality of the parameter estimates in
both stages.
18

Regression results based on exclusion restrictions are available from the authors upon request.
Our dataset contains 14,867 observations, which, when weighted, represent almost 20 million households.
This large sample size certainly helps us to avoid problems related multicollinearity in a model with a
comprehensive set of covariates. None of the estimated standard errors is unusually large to become a cause
for concern. For interesting discussions on sample size and multicollinearity please refer to Goldberger (1991),
Wooldridge (2009), and Hansen (2010).
19

18

In terms of the vector of treatment variables ( ), for the rst stage (i.e., selection), in the
case of OLS and the Heckman approach, we nd a negative and signicant eect of the FP
dummy and three way interaction of FP, households with children and Qubec dummies (i.e.,
p < :05 and :01 respectively). Interestingly, in the case of OLS, the Qubec dummy is also
negative and signicant (p < :01), whereas in the Heckman, it is the interaction between the
households with children and the Qubec dummies that is signicantly dierent than zero
(p < :1). In the case of the expenditure equation none of the dummies or interactions are
signicant under either the OLS or Heckman. Note that in the Heckman, the second stage is
conditioned on decision to purchase, implying that once the decision to purchase is taken into
account, none of the treatment dummies and their interactions have a signicant eect on
expenditures. On the other hand, in the case of OLS, expenditure equation is estimated only
for the consumers who purchased fast food. Compared to the OLS and Heckman results, in
the case of Tobit model we nd FP and Qubec, and the three way interaction dummies to
be negative and signicant (p < :01, :01, and :05 respectively).
As further evidence that media exposure matters, we nd that access to cable television
signicantly impacts the outcomes in all three models (in the case of Heckman model it is
at p < 0:01). Although we cannot decisively link television viewing to greater consumption
of fast food, this result does indicate that households with access to cable television also
purchased more fast food, even controlling for income and other demographic characteristics.
Next we explore which estimation model is more appropriate. For the rst stage of the
estimation, given it models a discrete choice, a Probit is more appropriate then OLS. We
also nd that the error terms in the rst and second stage regressions are not independent.
Specically, the correlation between the error terms is signicantly dierent from zero (p <
0:01), implying that we need to take the correlation among the error terms into account when
estimating the two stages. After ruling out the appropriateness of using OLS, we check for the
validity of the parametric restrictions in the Tobit model. Note that in the Heckman model,
some of the parameter estimates on the same characteristics hold dierent signs in the rst
19

and second stage estimates, where both estimates are signicantly dierent from zero. For
example, the number of children signicantly decreases the probability of purchasing fast food
but signicantly increases the amount spent. Other examples of alternating eects are the
coe cient on the dummy variable for blue-collar male occupation, female education, and fast
food price and price squared. The fact that the coe cient estimates are so strikingly dierent
between these two estimated equations leads us to believe that Heckman is the appropriate
approach, rather than the more restrictive Tobit, which assumes these coe cients are the
same (Greene 2003). Finally, note that in the Heckman, model fast-food price has a quadratic
eect on consumption. In the second stage, the positive sign on the linear portion of the
price eect does indicate a possibility of price endogeneity; however, the eect of pricing
across conditions should cancel out in DDD estimates, and we, therefore, do not feel that
this potential endogeneity biases our estimated eect of the ban.20
As noted, because all reported estimates come from non-linear models, we must calculate
the dierences in probabilities and expenditures for the treatment and control groups. We
present the marginal eect of the ban for the OLS, Tobit and Heckman models in Table 3.
In the Tobit case, the parameters for the rst and second stage are forced to be the same.
So, we present the result only for the second stage. For the OLS and Heckman models, we
present estimates for both stages.
Using the Heckman model, we nd signicant dierences in the propensity to consume
fast food: FP households have a 13:4 percent (p < :01) lower propensity to consume fast
food than AP households (row[1]). Note that this comparison is similar to the comparison
used by Goldberg (1990), which indicates that our results echo his ndings in terms of fast
20

Our argument on the negligible eect of probable price endogeneity in DDD estimates suggests that we
should not observe signicant changes if we drop price measures from the model. Indeed, this proves to
be the case; signicance level and directions of the estimated parameters do not change, although we do
observe small changes in estimated eects (e.g., the DDD estimate in the rst stage changed from 13 percent
to 11 percent with no change in the p-value). Also note that had we dropped the price variables from the
model we would have created the potential for other biased parameter estimates. We decided, therefore, to
present results based on a model with price variables. As a robustness check, we also estimated a model
with breakfast, lunch and dinner specic pricing. In that case, the DDD estimate turns out to be 12.75%
with no change in p-value.

20

food consumption. On the other hand, we see no signicant dierence between FP and AP
households in Ontario (row [2]). By subtracting row[2] from row[1], we get the rst dierencein-dierence (DD) estimator which controls for any intrinsic dierences between FP and AP
households with children in Qubec. In this case, the DD estimate (row [3]) implies the
ban led to a decrease in purchase propensity by 10.2 percent (p < 0:05). Next, we compare
households without children. In this case, we nd marginally signicant (p < 0:1) dierences
in the propensity to consume fast food between FP and AP households both in Qubec
and in Ontario. But the estimated dierence-in-dierence (row[6]=row[4]-row[5]) turns out
to be insignicant in this case. Row[7] provides the triple dierence-in-dierence (DDD)
estimates. Here we nd that FP households with children in Qubec have a signicantly
lower probability (13 percent with p < :05) than AP households of consuming fast food
in a given week after taking into account the dierences in the rest of the control groups.
Note that this dierence is, in fact, close to the dierence estimated between FP and AP
households in Qubec (row[1]). In terms of the second-stage decision on how much to spend,
we do not nd any of the dierence estimates to be signicant. This nding implies that
once households make the decision to consume fast food, there are no signicant dierences
in the amount spent on fast food across our comparison groups. This result may be driven
by the fact that most fast food restaurants are nation-wide chains whose menus tend not to
show large product or price variations; as a result, when consumers decide to purchase fast
food, the amount spent does not vary signicantly.
Note that for the rst stage, the OLS and Heckman models provide similar patterns in
estimated eects; more importantly, the crucial DDD estimates are close (11 and 13 percent
respectively) and signicantly dierent from zero at the ve percent level. For the second
stage, OLS and Heckman results are similar except in the case of dierence between FP and
AP households with children in Qubec. We nd marginally signicant DDD estimates with
the Tobit models (p < 0:1) but no signicant eect in the case of the OLS and Heckman
models. Nonetheless, note that in all three cases, the eect is negative. In terms of the DD

21

estimates in the second stage, we nd no signicant dierences in the case of households with
children in all three models (row[10]). But in the case of households without children, Tobit
results provide evidence of marginally signicant dierence (row[13]). Under all three models
in the second stage, the crucial DDD estimates imply a negative eect of the ban on fast food
consumption for FP households with children in Quebec. In terms of signicance, the Tobit
and rst stage OLS and Heckman show similar patterns. But the Heckman approach helps us
to decompose the eect into the two stages of the decision process after relaxing restrictive
assumptions of Tobit and allows us to observe precise estimates of both the decision to
purchase and the amount spent.
As a further robustness test, we use the matching estimator developed by Abadie and
Imbens (2006) to match households across Qubec and Ontario by language and the other
characteristics used in the Heckman regression above. While the DDD approach gives us
more information about the characteristics associated with fast food expenditures and allows
us to disentangle the propensity to purchase from the amount spent, the matching estimators
only inform us about the eect of treatment,in our case, the eect of the ban. However, because the matching estimator does not rely on specic assumptions about functional form, it
serves as a good robustness check for our earlier results. Using this non-parametric approach,
we obtain very similar results, indicating that the ban has a signicant eect, primarily on
the propensity to consume fast food.21
To check whether we are simply observing an overall eect of total food expenditures, we
run a similar regression for total household food expenditures. The DDD estimate in this
case is not signicantly dierent from zero, which indicates that the eect we estimate is not
an artifact of dierences in overall food expenditure patterns.22

Persistence
21
Results based on the matching estimation techniques of Abadie and Imbens are available from the
authors upon request.
22
One dierence in the specications is that because all households do spend on some form of food, we do
not control for selection. Detailed regression results are available from the authors.

22

One of the particular concerns about advertising targeting children is that it may not only
inuence concurrent consumption, but also shape future consumption behavior. We therefore
consider whether the ban aects purchasing patterns as children in Qubec age and are
exposed to advertisements. Here we encounter a further set of data constraints. First,
because data on mother tongue are not collected after 1992, we can consider only those
households that were aected by the ban when it was rst imposed. Second, because we do
not have specic data on the age of children in the household, we cannot consider households
with teenagers over 15 separately from households with multiple adults. Therefore, we
examine fast food expenditures of households composed of individuals under 25 years of
age in 1992 (i.e., consumers targeted by the ban when it was rst imposed) and compare
them to households composed of individuals 35 years of age and above (i.e., consumers not
targeted by the ban in 1980). Third, we have the specic ages only for the individuals who
answer the questionnaire and their spouses; therefore, we limit ourselves here to households
of one or two members, in which the young-adult respondent is less likely to be responding
for his or her aging parents or extended family. Since the overall results using the Heckman
approach suggest that the ban primarily decreases FP households propensity to purchase
fast food, we focus on the persistence of the bans eect on the decision to purchase fast food
for FP households under the age of 25 (i.e., young adults). Fourth, we do not have specic
information on where households previously lived, so our analysis here implicitly assumes
that the majority of the young population remain in the same province. We compare the
probability of purchase between young FP and AP households living in Qubec and Ontario
to their older counterparts.
Results are presented in Table 4. Our results suggest that young FP adults are more likely
to purchase fast food if they live in Ontario than if they live in Qubec, whereas the reverse
is true for young AP adults. Within our data set, an FP young adult is 38 percent less likely
to purchase fast food in a given week if (s)he lived in Qubec than if s(he) lived in Ontario,
while a similar AP young adult is 24 percent more likely to purchase fast food if s(he) lived

23

in Qubec. The resulting dierence-in-dierence is 63 percent, and is signicantly dierent

from zero at the 5 percent level. Comparing the younger adults to their older counterparts,
we nd that both groups of older adults are more likely to purchase fast food if they live
in Ontario than in if they lived in Qubec, and that this dierence is smaller than that for
the young adults. Therefore, we see a large DDD eect, which, although not signicantly
dierent from zero, does provide an indication that the ban on advertising targeting children
may continue to aect purchasing behavior as those children become adults.

Estimated Eects
Given that the ban has a statistically signicant eect on fast food consumption at the
household level, here we estimate the economic eect of the ban. In this case, we use the
signicant DDD estimates from the rst stage of the Heckman model and information from
the existing literature on obesity and nutrition to infer the impact of the ban. As noted
above in Table 3, we estimate that the ban reduces the probability of fast food purchase
incidence by 13 percent per week. Based on the number of FP households with children
in Qubec cities in 1992 (i.e., 310,617 households) this reduction suggests that 40,691 fewer
households purchased fast food in any given week. We can extrapolate to lost annual sales,
noting that FP households purchasing fast food spent an average of CDN$13.09 per week,
which suggests lost revenue from the Qubec urban market of CDN$27.6 million in 1992
dollars. If we assume that the eect is similar for households in small cities and rural areas
then the estimate will be CDN $65.4 million. This amount is equivalent to US$88 million in
2010.
What do these results mean in terms of calories consumed? The amounts above translate
into 7.1 million fewer meals sold in the urban areas and if we extrapolate to all FP households
in Qubec, 16.8 million fewer meals sold overall. With 800-1100 calories per fast food meal23 ,
that means that urban households in Qubec consume between 5.6 billion and 7.8 billion
23

These numbers are based on calorie calculations for regular extra value meals from McDonalds
(http://nutrition.mcdonalds.com/nutritionexchange/nutritionfacts.pdf) .

24

fewer fast food calories per year as a result of the ban. Similarly, if we extrapolate to all
FP households with children in Qubec, the estimate will be between 13.4 billion and 18.4
billion fewer fast food calories. We cannot explicitly estimate the net calories, since these
consumers presumably ate something else when they decided not to purchase fast food.
That said, in two separate studies, consumers eating a fast food meal added an extra 200
calories for that meal (Paeratakul et al., 2003; Bowman and Vinyard, 2004). Recognizing
that these studies were done on U.S. adults, if we assume that Qubec consumption patterns
are otherwise similar to those in the United States, the ban would have reduced net calorie
consumption 1.4 billion calories per year in urban areas, or 3.4 billion calories for the entire
province of Qubec. Pereira et al. (2005) estimate, after controlling for all other probable
factors, that frequent fast food consumption can lead to 4.5 kg (or 9.9 lb) of weight gain
over 15 years. Assuming that the increase in probability of purchase moves a household to
being a "frequent fast food consumer", by 1995, the ban may have resulted in a 0.6 kg (1.3
lb) lower body weight of Qubec FP household members. Perhaps more importantly, lower
fast food consumption is also associated with lower rates of disease. Over the same time
period, Pereira et al. (2005) found that frequent fast food consumption was associated with
a twofold increase in insulin resistance, a key precursor of type II diabetes.

Summary and Concluding Remarks

Advertising targeting children has become a major cause for concern for policy makers in
a number of countries, primarily because of the belief that advertising has increased fast
food consumption and is thus related to the exponential increase in obesity among children.
Several countries are responding by considering banning the advertisement of unhealthy food
to children. One jurisdiction that has a ban on all child-specic advertising is the Canadian
province of Qubec. In this article, we report our analysis of the eect of this ban on fast
food expenditures by households in Qubec.
We identify the eect of the ban by noting that given the nature of Qubecs media mar25

ket and demographic composition, the ban will disproportionately aect French-speaking
(as opposed to English-speaking) households in Qubec and will not aect similar households in Ontario, or households without children in either province. We nd that during
our study period, French-speaking households with children are signicantly less likely to
purchase fast food if they live in Qubec than if they live in Ontario. To address the concern that inherent cultural dierences may aect preferences for fast food and thus may be
responsible for the dierence between Francophone and Anglophone consumption, we use
a dierence-in-dierence-in-dierence estimator, comparing between French and Englishspeaking households without children, and nd a much smaller, and insignicant, dierence
in terms of both the likelihood of purchasing fast food, and the amount spent. Thus, the
result that we observe aects only French-speaking households with children in Qubec
not their English-speaking neighbors, their French-speaking counterparts in Ontario or their
French-speaking neighbors without children. Our estimated Heckman model implies that
the ban signicantly decreased propensity to consume fast food by 13 percent for the effected households. This estimate is robust to alternative estimation methods. Further, we
believe that, if anything, our ndings underestimate the eect of the ban. For example, if
French-speaking adults are aected by the reduction in fast food ads during shows targeted
at both adults and children then our results will be biased downward. In short, we believe
that the present analysis provides evidence that the advertising ban does aect consumption.
Finally, we nd tentative evidence that the eect of the ban persists as the aected children
become young adults.
In terms of policy implications, the present study provides evidence that a ban on advertising targeting children can be eective in lowering or moderating consumption, and
estimates of the eect in expenditures suggest the social-welfare impact of such a ban can
be signicant. One may ask, given this nding, whether other jurisdictions in Canada or
other countries should implement similar bans. Here our results suggest caution. We nd
that it is primarily French-speaking children who are aected by the Qubec ban, while

26

English-speaking children who have greater access to media from the neighboring U.S.
states and Canadian provinces are less aected. This nding indicates that media spillover
can blunt the eect of an advertising ban, which suggest that a ban imposed by a single state
or province may not be eective if there is substantial media overlap, and that advertising
regulations are likely to be more eective if several jurisdictions can coordinate their eorts.
Further, given the rapid changes in information technology that have led to children spending more and more time on video games and computers, any attempt to impose a similar
ban will be challenging. Notably, consumer advocates in Qubec are currently using the ban
to pursue internet advertising.24 It will be interesting to see whether they can expand the
eective scope of the existing ban to address the challenges posed by these new media.
Although at the time of its implementation the law did not have strong support from
majority of consumers in Qubec (Caron, 1994), a survey of Qubec residents in 2007 found
that 60 percent wanted the provinces advertising ban to be applied more strictly. Thus, it
seems that Qubec consumers consider this regulation to be benecial. Note that it is di cult
to assess the eect of the ban on health outcomes without further knowledge of the detailed
food and lifestyle habits (such as frequency and type of physical activity). But, interestingly,
Qubec has one of the lowest childhood obesity rates in Canada, although its children have
one of the most sedentary lifestyles (Statistics Canada 2005). More importantly, the 2004
Canadian Community Health Survey (CCHS) shows that the combined overweight/obesity
rate among 2- to 17-year-olds in Qubec is signicantly below the national level.25
In terms of regulations of advertising targeting children, an alternative to an outright
ban is voluntary industry-led regulations as is currently in practice in the United States
on a limited scale.26 The eectiveness of such voluntary regulations needs to be studied.
24

Specically, the Coalition qubcoise sur la problematique do poids successfully challenged Lucky
Charms cereal for advertising internet games on their food packaging.
25
The CCHS collected information from over 35,000 respondents between January and December 2004
and directly measured most respondents height and weight rather than relying on self-reports, giving the
most accurate picture of rates in 25 years. Regrettably, they did not collect data on mother tongue, so we
cannot replicate our analysis using these data. See www.statcan.ca/Daily/English/050706/d050706a.htm.
26
For further details, please refer to: http://www.bbb.org/us/children-food-beverage-advertising-initiative

27

Interestingly, it is usually argued that as a result of battles for brand market shares, some
products targeting children are advertised excessively. A case in point is the carbonated
soft drinks category, in which in 2006, advertisers spent nearly $20 per American teenager
on targeted advertising.27 If these dollars were largely spent in battles over market share,
then any regulation, whether voluntary or publicly mandated, could turn out to be Pareto
improving for a society.
To the best of our knowledge, the present study is the rst application of the dierencein-dierence approach to estimate advertising eectiveness in a natural experiment. Given
the messiness and incompleteness of eld-level data, we believe that a similar approach can
be used to determine the eectiveness of advertising in other similar contexts. For our data
set, we use household consumption data available from Statistics Canada; similar publicly
available data exist in the United States and many other developed nations. One advantage
of using these data is that they tend to be rich in information on demographic characteristics
and to have wider coverage across geographic regions. In addition, by using the sampling
weights in such databases, it is possible to generalize the outcomes at the population level
with greater condence.
A limitation of the present study is that we look only at the eect the advertising ban
on expenditures on a single food category. Finally, because of data limitations, we were not
able to build and estimate models linking fast food purchase decisions, expenditures and
health outcomes. Also, long term impacts of such bans need to be explored. In the future,
we plan to extend and explore these links further.

References:
1. Abadie, Alberto and Guido W. Imbens (2006), Large Sample Properties of Matching
Estimators for Average Treatment Eects,Econometrica, 74 (1), 235-267.
2. Allison, David B., Kevin R. Fontaine, JoAnn E. Manson, June Stevens, and Teodore B.
27

See : http://www.ftc.gov/speeches/leibowitz/080729foodmarketingtochildren.pdf.

28

VanItallie (1999), "Annual death attributable to obesity in the United States." Journal
of American Medical Association (JAMA), 282 (16), 1530-1538.
3. Amemia, Takeshi (1985), Advanced Econometrics. Harvard University Press, Cambridge, MA.
4. Ashenfelter, Orley and David Card (1985), "Using the Longitudinal Structure of Earnings to Estimate the Eect of Training Programs," Review of Economics and Statistics,
67, 648-60.
5. Assmus, Gert, John U. Ferley and Donald R. Lehman (1984), "How Advertising Aects
Sales: Meta Analysis of Econometric Results." Journal of Marketing Research, 21 (1),
65-74.
6. Bowman, Shanthy and Brian T. Vinyard (2004), "Fast food consumption of US adults:
impact on energy and nutrient intakes and overweight status", Journal of the American
College of Nutrition 23 (2), 163-168.
7. Boynton-Jarrett, Rene, Tracy N. Thomas, Karen E. Peterson, Jean Wiecha, Arthur
M. Sobol and Steven L. Gortmaker (2003), Impact of Television Viewing Patterns on
Fruit and Vegetable Consumption Among Adolescents,Pediatrics 112, 1321-6.
8. Card, David and Alan B. Krueger (1994), "Minimum Wages and Employment: A Case
Study of the Fast Food Industry in New Jersey and Pennsylvania", American Economic
Review, 84, 772-793.
9. Caron, Andre (1994), Children, Advertising and Television Choices in a New Media
Environmentin Children and Advertising: A Fair Game? Stephen Frith and Barbara
Biggins eds, New College Institute for Values Research: University of New South Wales,
94-111.

29

10. Connor, Susan M. (2006), "Food-related advertising on preschool television: building

brand recognition in young viewers." Pediatrics. 118 (4), 1478-85.
11. Crespo, Carlos E., Ellen Smit, Richard Trojano, Susan Bartlett, Caroline Macera and
Ross Anderse (2001), Television Watching, Energy Intake and Obesity in U.S. ChildrenArchives of Pediatrics & Adolescent Medicine 155, 360-365.
12. De la Garde, Roger (1996), "There goes the neighborhood: Montreals television market and free trade." in E.M. McAnany and K.T. Wilkinson edited Mass Media and
Free Trade, University of Texas Press, First edition.
13. Dietz, William H. and Steven L. Gortmaker (1985), Do We Fatten Our Children at
the Television Set? Obesity and Television Viewing in Children and Adolescents,
Pediatrics 75, 807-12.
14. Duerksen Susan C., John P. Elder , Elva M. Arredondo, Guadalupe X. Ayala, Donald J.
Slymen, Nadia R. Campbell, and Barbara Baquero (2007), "Family restaurant choices
are associated with child and adult overweight status in Mexican-American families",
Journal of the American Dietetic Association 107 (5), 849-853.
15. Federal Trade Commission (2008), Marketing Food to Children and Adolescents: A
Review of Industry Expenditures, Activities and Self-Regulation. Report to Congress
(July) http://www.ftc.gov/os/2008/07/P054504foodmktingsreport.pdf
16. French Susan A., Mary Story, Dianne Neumark-Sztainer, Jayne A. Fulkerson and Peter
J. Hannan (2001), "Fast food restaurant use among adolescents: associations with
nutrient intake, food choices and behavioral and psychosocial variables," International
Journal of Obesity, 25, 1823-33.
17. The Gazette (2009), "Burger King must pay $12,000 for advertising to children in
Quebec" May 9, 2009.

30

18. Giammattei, Joyce, Glenn Blix, Helen H. Marshak, Alison O. Wollitzer and David J.
Pettitt (2003), Television Viewing as a Cause of Increasing Obesity Among Children
in the United States.Archives of Pediatrics & Adolescent Medicine 157, 882-6.
19. Goldberg, Marvin E. (1990), A Quasi-Experiment Assessing the Eectiveness of TV
Advertising Directed to ChildrenJournal of Marketing Research 27(4), 445-54.
20. Goldberger, Arthur S. (1991), A Course in Econometrics. Cambridge: Harvard University Press.
21. Gortmaker, Steven L., Aviva Must, Arthur M. Sobol, Karen Peterson, Graham A.
Colditz and William H. Dietz. (1996), Television Viewing as a Cause of Increasing
Obesity Among Children in the United States. Archives of Pediatrics & Adolescent
Medicine 150, 356-62.
22. Government of Canada and the Gouvernement du Qubec (1985), Les eects de la loi
qubcoise interdisant la publicit destine aux enfants. Rapport du Comit fdralprovincial sur la publicit destine aux enfants. Septembre.
23. Greene, William (2003), Econometric Analysis. Fifth Edition, Prentice Hall, NJ.
24. Gruber, Jonathan (1994), "The incidence of mandated maternity benets." American
Economic Review, 84 (3), 622-641.
25. Halford, Jason C.G., Jane Gillispie, Victoria Brown, Eleanor E. Pontin and Terence M.
Dovey (2004), Eect of Television Advertisements for Foods on Food Consumption
in ChildrenAppetite 42, 221-5.
26. Hansen, Bruce E. (2010), Econometrics, available online at http://www.ssc.wisc.edu/~bhansen/
econometrics/Econometrics.pdf.
27. Hastings, Gerard, Martine Stead, Laura McCermott, Alasdair Forsyth, Anne M. MacKintosh, Mike Rayner, Christine Godfrey, Martin Caraher and Kathryn Angus (2003),
31

Review of Research on the Eects of Food Promotion to Children Final Report. University of Strathclyde, Centre for Social Marketing, Glasgow.
28. Hawkes, Corinna (2007), "Marketing Food to Children: Changes in the Global Regulatory Environment 2004-2006. World Health Organization Report.
29. Heckman, James J. (1976), The Common Structure of Statistical Models of Truncation,
Sample Selection and Limited Dependent Variables and a Simple Estimator for Such
Models. Annals of Economic Social Measurement, 5, 4, 475-492.
30. Heckman, James J. 1979. "Sample selection bias as a specication error." Econometrica
47, 153-161.
31. Hamermesh, Daniel S. and Stephen J. Trejo (2000), "The demand for hours of labor:
direct evidence from California." Review of Economics and Statistics, 82 (1), 38-47.
32. Institute of Medicine of the National Academies (2006), Food Marketing to Children
and Youth: Treat or Opportunity. Food and Nutrition Board, McGinnis, J.M. an
J.A.Gootman, V.I. Kraak eds. Available at: www.nap.edu.
33. Jedidi, Kamel, Carl F. Mela and Sunil Gupta (1999), "Managing advertising and promotion for long-run protability." Marketing Science, 18 (1), 1-22.
34. Kelly, Brendan (2009), "Qubec TV Auds glued to the tube", Variety Oct. 3rd, 2009.
35. Krebs, Nancy F. and M.S. Jacobson (2003), "American Academy of Pediatrics Committee on Nutrition: Prevention of pediatric overweight and obesity." Pediatrics, (112),
424-430.
36. Kucharsky, Danny (2007), "Consumer groups demand end to Vachon gorilla campaign"
Marketing Magazine February 20th http://www.marketingmag.ca/daily/20070220/topstory.html.
37. Kucharsky, Danny (2005), "Time to end it?" Marketing Magazine August 15th, (110)
27, 2-23.
32

38. Lalonde, Robert J. (1986), "Evaluating the Econometric Evaluations of Training Programs with Experimental Data," American Economic Review, 76, 604-620.
39. Lodish, Leonard M., Magid M. Abraham, Stuart Kalmenson, Jeanne Livelsberger,
Beth Lubetkin, Bruce Richardson and Mary E. Stevens (1995), How TV Advertising
Works: A Meta-Analysis of 389 Real World Split Cable T.V. Advertising Experiments,
Journal of Marketing Research, 32(2), 125-129.
40. Madrian, Brigitte C. (1994), "Employment-based health insurance and job mobility:
is there evidence of job lock?" Quarterly Journal of Economics, 109 (1), 27-54.
41. McGinnis, J.M. and William H. Foege (1993), "Actual Causes of Death in the United
States." Journal of American Medical Association, 270 (18), 2207-2218.
42. Mullman Jeremy (2009), "The Next Hot Spot for U.S. Test Market is in Canada",
Advertising Age, Nov. 09. 2009.
43. National Post (2009), "Qubec scores sweet victory in junk food wars." G. Hamilton:
Jan 26, 2009.
44. Naik, Prasad A., Murali K. Mantrala and Alan G. Sawyer (1998), "Planning media
schedules in the presence of dynamic advertising quality." Marketing Science, 17 (3),
214-235.
45. Nielsen, Samara, Anna M. Siega-Riz , and Barry Popkin (2003), "Trends in energy
intake in the U.S. between 1977 and 1996: similar shifts seen across age groups,"
Obesity Research 10 (5), 370-378.
46. Niemeier, Heather M., Hollie A. Raynor, Elizabeth E. Lloyd-Richardson, Michelle
L. Rogers, and Rena R.Wing (2006), "Fast food consumption and breakfast skipping:predictors of weight gain from adolescence to adulthood in a nationally representative sample", Journal of Adolescent Health 39, 842-849.
33

47. Paeratakul, Sahasporn, Daphne P. Perdinand, Catherine M. Champagne, Donna H.

Ryan, and George A. Bray (2003), "fast food consumption among US adults and
children: dietary and nutrient intake prole," Journal of American Dietetic Association
103 (10), 1332-1338.
48. Pereira, Mark A., Alex I. Kartashov, Cara B. Ebbeling, Linda Van Horn, Martha L.
Slattery, David R. Jacobs jr. and David S. Ludwig (2005), "fast food Habits, weight
gain, and insulin resistance (the CARDIA study): 15-year prospective analysis." Lancet
365, 36-42.
49. Saer, Henry, and Frank Chaloupka (2000), "The eect of tobacco advertisising bans
on tobacco consumption." Journal of Health Economics, 19, 1117-1137.
50. Satia Jessie A., Joseph A. Galanko, and Anna M. Siega-Riz (2005), "Eating at fast food
restaurants is associated with dietary intake, demographic, psychosocial and behavioral
and behavioral factors among African Americans in North Carolina," Public Health
Nutrition: 7 (8), 1089-1096.
51. Singer, Dorothy G. and Jerome L. Singer (2001), Handbook of Children and the Media,
Thousand Oaks, CA: Sage Publications, 640-643.
52. Statistics Canada (2010), Canadian Community Health Survey: Obesity among children and adultshttp://www.statcan.gc.ca/daily-quotidien/050706/dq050706a-eng.htm.
53. Steenkamp, Jan-Benedict E.M., Vincent R. Nijs, Dominique M. Hanssens and Marnik
G. Dekimpe (2005), "Competitive reactions to advertising and promotion attacks."
Marketing Science, vol. 24(1), 35-54.
54. Taveras Elsie M., Thomas J. Sandora , Mei-Chiung Shih , Dennis Ross-Degnan , Donald
A. Goldmann , and Mathew W. Gillman (2006), "The association of television and
video viewing with fast food intake by preschool-age children." Obesity. 14(11):203441.
34

55. The Times (2004), "Labour will ban junk-food adverts on TV." May 31, 2004.
56. Vakratsas, Demetrios and Tim Ambler (1999), "How Advertising Works: What do we
really know?" Journal of Marketing, 63 (1), 26-43.
57. Wooldridge, Jerey M. (2002), Econometric Analysis of Cross Section and Panel Data,
MIT Press, Cambridge.
58. Wooldridge, Jerey M. (2009), Introductory Econometrics: A Modern Approach (4th
Edition), Cincinnati, Ohio: South-Western College Publishing.
59. You, Wen and Rodolfo M. Nayga (2005) "Fast Food, Television Viewing and Childrens
Dietary Quality," Journal of Agricultural and Resource Economics 30 (2), 302-14.

35

Web Appendices for the paper:

Fast Food Consumption and the Ban on Advertising
Targeting Children: The Qubec Experience
By
Tirtha Dhar and Kathy Baylis

Web Appendix A: Derivation of the price variable

Web Appendix B: Imputation of probability of Cable TV ownership
Web Appendix C: Matched Estimates
Web Appendix D: Regression analysis with exclusion restrictions
Tables:
A1: Estimated effects using matched estimation method of Abadie and Imbens (2006)
A2: Regression with Exclusion Restrictions
A3: Difference-in-Difference Estimates under exclusion restrictions
A4: Total Food Expenditure Regression
A5: Difference-in-difference Expenditure in Total Food
A6: Heckman Regression Results with Standard Errors
A7: Pair-wise correlation of Variables used in regression analysis (highlighter numbers are
significant at 5% level)

0.1

Web Appendix A: Notes on Price Variable

For those who purchase fast food, we use the total expenditure by the
household divided by the number of fast food meals as the price for fast
food. However, we do not observe the price faced by those households
who do not purchase fast food during the two-week period of the survey.
For these households, we impute a median price of fast food per meal
using the information from consumers living in the same region, who
do purchase fast food. To estimate a weighted median price of fast
food, we rst establish a median price paid for each province and each
year. In the data, we only observe the total amount of expenditure
per household per week on fast food and the number of fast-food meals
purchased. For a household purchasing both fast food suppers and
breakfasts, we cannot extract a price paid for each meal type separately.
Thus we begin by restricting ourselves to only those households who
happened to purchase a single meal type in a given week. For example,
we use household A who purchases four breakfast meals during the week
to generate an imputed price for a breakfast fast food meal, and we use
household B who only purchases 3 supper meals as a source of the price
of fast food suppers. We take the median price of fast food breakfasts
from all household like A in a given province and year as our median
price of breakfasts. Similarly, we take all households like B in a province
and a given year who only purchased suppers, to generate a median price
for suppers. So, for example, we might nd that the median imputed
price for breakfast was $2, the median imputed price spent for lunch was
$3 and the median imputed price for supper was $4 per meal. We use
the median price for each meal to avoid biases induced by one or two
particularly high or low imputed prices.
To generate a weighted price for a typical fast food meal for the survey
period during each year, we take the number of all fast food breakfasts,
lunches and suppers purchased in the province in that yearly survey,
and calculate the percent of purchased meals that fall in each category.
Thus, we obtain the percent of fast food meals that are breakfasts, the
percent that are lunches and the percent that are suppers. We use each
of these fractions to weight the median prices for breakfasts, lunches
and dinners to obtain a weighted price for the average meal. Thus, if 20
percent of all meals are breakfasts, 30 percent are lunches and 50 percent
are dinners, our weighted average price would be 0.2*$2+0.3*$3+0.5*$4
= $3.30. This imputed price is then used as the price for consumers
where the household did not purchase fast-food during the period of the
surveys.

0.2

Web Appendix B: Notes on Cable TV Variable

As noted in the main text, one factor that can aect consumption is exposure to media, particularly television. None of the available databases
have detailed information on household level television viewing patterns.
Of the available databases, only the biannual Canadian household expenditure survey (i e. the Famex) provides information on detailed yearly
expenditure on cable and satellite TV subscriptions, but does not contain
explicit information on the amount of time spent in front of the television, or the type of programming watched. Another estimation challenge
is that since the Foodex and Famex surveys are conducted on dierent
samples, we cannot directly infer access to television by the households
in the Foodex data. Instead, we project the television ownership from
Famex to Foodex data by using the following approach. We estimate
a probit model of positive cable television expenditure as a function of
household characteristics and sampling weights using the Famex data
for each province and for each year (results are available from the authors). Then using the same set of characteristics and sampling weights
in the Foodex les we predict the probability or propensity score of any
household having a cable television subscription.
A further challenge is that before 1990, the Famex only contains information on the broader category of entertainment expenditure,which
includes cable TV subscriptions. To address this constraint, using 1990
data, we separately predict cable expenditure and other entertainment
expenditure when the household does not have cable. We then predict the probability of entertainment expenditure given the 1984 and
1986 data. Using the parameter estimates from 1990 and the demographic characteristics from the earlier years, we predict the probability
of a household having entertainment expenditure but without any cable.
We then substract this probability of non-cable entertainment expenditure from the probability of entertainment expenditure to obtain the
probability that each household has cable. In 1992, 65% of Canadian
households in our database paid for cable television.1 By relying on
cable fees to proxy for television ownership, our proxy for television
ownership will most likely be an underestimate.

0.3

Web Appendix C: Matched Estimates

To check for the robustness of the estimates and our key ndings, we
also estimate the eect of the ban using a matching technique proposed
in Abadie and Imbens (2006). Under this method, the treatment eect
1

Our data also includes whether a renter received cable for free. These houesholds
are included as having access to cable.

ii

for a single unit, i is dened as the dierence between the outcome of a

household that is in the treatment group minus the outcome if that same
family is untreated yi (1) yi (0). In our case, the treatment eect is the
dierence in fast food expenditure between a Quebec household aected
by the ban and that of the same household if that household were in
Ontario. If we could observe the treatment and non-treatment outcome
for the same household - i.e. both yi (1) and yi (0), then the eect of the
treatment on household i would simply be the dierence. The problem is
that for each household, we only observe either the outcome of treatment
or not. Thus, following Imbens et al.:

yi = yi (Ti ) =

yi (0) if Ti = 0
yi (1) if Ti = 1

(1)

where Ti represents whether the treatment is received by household

i (Ti = 1) or not (Ti = 0). Given that our interest is in measuring
the eectiveness of the ban
P we focus on the sample average treatment
yi (0)). Note that in the present
eect estimates: = N1 N
i=1 (yi (1)
study, only one of the potential outcomes yi (0) or yi (1) is observed for
each household and the other is unobserved or missing. We impute the
missing potential outcome by using average outcomes for individuals
with similar covariate values. Essentially, the procedure creates a
counterfactual for the expenditure of a household in Quebec by using
expenditure of similar households in Ontario. Let,
^

y i (0) =

1
M (i)

yi if Ti = 0;
X
yl if Ti = 1

(2)

l2M (i)
^

y i (1) =

1
M (i)

yi if Ti = 1;
X
yl if Ti = 0

(3)

l2M (i)

where M (i) is the number of matched households for household i.

This algorithm is also known as nearest neighborhood matching method.
The algorithm rst estimates the distance from a treatment households
to the households in control groups based on a predetermined set of covariates. In the present case we use the Mohalnobis distance measure.
Once the distance is estimated, the algorithm estimates the counterfactual by taking the weighted average outcome of the nearest households.
The treatment eect is then the dierence between the weighted average
outcome of the controls and observed outcome of the treated households.
We match each treatment (control) household to four households in the
iii

control (treatment) group. We can restate the matching treatment eect

estimator as:
N
1 X
(b
yi (1) ybi (0))
(4)
M =
N i=1

Intuitively, if the treated and control units systematically dier in

their characteristics, then we cannot disentangle the eect of the characteristics from the eect of the treatment. Such bias is of paramount
concern in non-experimental studies. So, to rene the estimation process
we also adjust the estimates for biases. With bias corrections the treatment eect becomes:
M

where: yei (0) =

and yei (1) =

1
M (i)

N
1 X
(e
yi (1)
=
N i=1

l2M (i)
1
M (i)

l2M (i)

yi if Ti =0;

yei (0))

yl +(b
p0 (Zi ) pb0 (Zl )) if Ti =1
yi if Ti =1;

(5)

yl +(b
p1 (Zi ) pb1 (Zl )) if Ti =0

where pb0 is the prediction function based linear regression of covariates on group outcomes. Finally, the averages are adjusted for population weights. Given the potential problem of heterogeneity, we use
robust standard errors to estimate signicance.
To keep the matching estimates comparable to the estimates using
the Heckman approach, we use the same set of covariates from the Heckman to match households. For some specications, we also add fast-food
purchase decision as one of the matching variables.2 Further, we use the
explanatory dummy variables from the Heckman analysis for exactly
matching the households and for bias corrections. Matching results are
presented in Table A1 and A2. Since using matched estimation approach
we can only estimate the dierence between two groups at a time, we rst
compare households across provinces (i.e. Table A1i) and then within
provinces across language (i.e. Table A1ii).3
Comparing across provinces (Table A1i), if we take into account
both purchasing and non-purchasing households, then Quebec house2

To address the concern about matching over an outcome variable, we also try
matching over the mills ratio from the Heckman equation, to control for possible
selection bias. Results from this matching estimate are essentially unchanged from
those presented here.
3
We also estimate matched dierences using purchase incidence as one of the exact
matched criteria. Qualitatively results remain the same but, as one would expect
from the results from the second stage of the Heckman estimation, the magnitude is
lower than the results presented.

iv

holds spend $3.93 (p < 0:01) less per week on fast-food than Ontario
households. Next if we consider only the households with positive expenditures then we nd Quebec households spend $0.61 less but this amount
is insignicant. Comparison of these two estimates suggest that the key
dierence in expenditure is in terms of purchase occasions. Next, we
estimate the dierences for households with and without children, for
all households. Again in both cases we nd signicant dierences: in
the case of households with children Quebec households spend $5.48
(p < 0:01) less and in the case of households without children they
spend $2:52 (p < 0:01) less. Interestingly if we consider households with
positive fast-food expenditures, we only nd a signicant dierence for
households with children. This result suggests that Quebec households
with children not only purchase fast-food less often but spend less when
they do.
Next, we compare households within provinces (Table A1ii). In the
case of Quebec, if we consider all households, we nd signicant dierences between AP and FP households; and FP households spend $1:08
(p < 0:01) less than the AP households. Again, if we only consider households with positive fast-food expenditures, then the dierence becomes
insignicant. Next we estimate the dierences between FP and AP
households with children in Quebec. In this case, FP households spend
$3:13 (p < 0:01) less than AP households and if we consider households
with positive expenditures, this dierence becomes $6:37 (p < 0:01). On
the other hand, if we consider households without children, then the difference decreases to $0:47 and is insignicant (p > 0:1). Interestingly, if
we consider households without children but with positive expenditures
then the dierence is signicant but positive with an average dierence
of $1:63 (p < 0:01). This estimate suggests FP households without
children, contrary to all other comparisons conducted, spend more than
comparable AP households. In general, we nd that FP households
with children in Quebec not only went out for fast-food less often, but
spent less than other comparison groups. Finally, we compare groups
within Ontario. In this case not only are the dierences between FP
and AP households small, but they are also insignicant. This nding
clearly conforms to the argument that in a market where the ban was
not imposed the households will show no dierence in expenditures.
Overall, we nd the dierences in expenditures similar to the Heckman results. In most cases the dierence is due to dierences in purchase
incidence across provinces and across households. We nd signicant differences in expenditures between FP and AP households with children in
Quebec. This dierence becomes insignicant if we compare households
in Ontario, suggesting that FP and AP households do not show any difv

ference in expenditures if they are unaected by the ban. We also nd

that compared to any other group, FP households with children in Quebec not only went out less for fast-food but also spent less. In the case of
households without children, we do not nd any signicant dierence in
Quebec but interestingly the dierence become signicant if we consider
only households with positive expenditures. On the other hand, we do
not nd any signicant dierences between households without children
in Ontario.

vi

Table A1i: Across Province Comparison

Quebec vs. Ontario
Quebec vs. Ontario with Children
Quebec vs. Ontario without Children
With Positive Fastfood Expenditure
Quebec vs. Ontario
Quebec vs. Ontario with Children
Quebec vs. Ontario without Children

Estimates
-3.93
0.00
-5.48
0.00
-2.52
0.00
0.61
0.09
-2.21
0.00
0.38
0.36

Table A1ii: Within Province Comparison

Estimates
Within Quebec
Within Ontario
-1.08
-0.49
0.00
0.09
-3.13
-0.47
0.00
0.45
-0.47
-1.45
0.10
0.09

FPHH vs. APHH

FPHH vs. APHH with Children
FPHH vs. APHH without Children
Within Quebec with Positive FF Exp.
FPHH vs. APHH
FPHH vs. APHH with Children
FPHH vs. APHH without Children

vii

-0.99
0.16
-6.37
0.00
1.63
0.01

-0.73
0.34
0.17
0.92
0.24
0.77

Web Appendix D: Regression with Exclusion Restrictions

Under the exclusion restrictions, the first stage probit model is identified due to exclusion of a variable
that affects selection but does not affect the choice of expenditure after the decision to purchase has
been made. Without the exclusion restrictions the model is identified due to non-linearity. Here we
present the results for the model without exclusion restrictions. In the model with exclusion restrictions
we exclude quarterly dummies and dual immigrant household dummy in the Probit model to identify
the model. Quarterly dummies are excluded under the assumption that households decision to
purchase fast-food depends on seasonal variations in weather but the amount spent does not vary as a
result. We expect that under the cold Canadian winter households will generally eat out less, and thus
consume fast-food less often during the fourth quarters compared to other quarters of the year.
Similarly we exclude dual immigrant households in the second stage under the assumption that recent
immigrants are less likely to purchase fast-food purely due to different preferences, and less familiarity
with fast-food. Neither excluded variable significantly affects fast-food expenditure in the second stage
of the model.

viii

Table A2: Regression with Exclusion Restrictions

Estimation Method:

Heckman

Selection
log(Expenditure)
Equation
Estimates
Estimates
Dummy: Children
-0.032
0.001
Dummy: Francophone
-0.265 ***
0.084
Dummy: Quebec
-0.066
0.063
Dummy: Children X Francophone
0.177
-0.103
Dummy: Children X Quebec
0.188 *
-0.021
Dummy: Francophone X Quebec
0.077
-0.097
Dummy: Children X Francophone X Quebec
-0.359 **
0.012
Number of Children
-0.063 **
0.088 ***
Probability of Cable TV access
0.470 ***
0.310 ***
Fast Food Price
-8.599 ***
0.332 ***
Fast Food Price2
1.038 ***
-0.008 ***
Household Income
5.120 ***
0.799
2
Household Income
-12.752
7.522
Number of Household Members
0.016 **
0.026 ***
Dummy: Male Blue Collar Occupation
-0.090 **
0.117 **
Dummy: Male Pink Collar Occupation
-0.138 ***
0.065
Dummy: No Male Occupation
-0.067 *
0.011
Dummy: Female Blue Collar Occupation
-0.360 ***
0.110
Dummy: Female Pink Collar Occupation
0.076 **
0.019
Dummy: No Female Occupation
-0.066 **
0.044
Dummy: Home Ownership
0.018
-0.056 *
Dummy: Recipient of Social Assistance
-0.233 ***
-0.056
Male Education
0.050 ***
0.022
Female Education
0.017 *
-0.036 ***
Dummy: Age of Household Head
-0.014 ***
-0.002
Dummy: Dual Income Household
0.019 ***
0.015 *
Dummy: Both Spouses Immigrants
0.178
Dummy: Year 1986
0.442 ***
0.056
Dummy: Year 1990
0.508 ***
0.064
Dummy: Year 1992
0.496 ***
0.162 ***
Dummy: 1st Quarter
0.122 ***
Dummy: 2nd Quarter
0.121 ***
Dummy: 3rd Quarter
0.157 ***
Intercept
16.665 ***
0.787 ***
Model Fit Measure
0.106
Psuedo R2
Note: *** Significant at 1% level; ** Significant at 5% level and * Significant at 10% level

ix

Table A3: Difference-in-Difference Estimates under exclusion restrictions

Heckman
Decision to Consume Fast-Food (First Stage)
Difference between FP and AP Households with Children in Quebec
-0.134 ***
Difference between FP and AP Households with Children in Ontario
-0.032
Difference in Difference: [1] - [2]
-0.102 **
Difference between FP and AP Households without Children in Quebec
-0.067 *
Difference between FP and AP Households without Children in Ontario
-0.095 *
Difference in Difference: [3] - [4]
0.028
Triple Difference in Difference: [3] - [6]
-0.131 **
Amount Spent (Second Stage) [$/week]
-1.07
[8] Difference between FP and AP Households with Children in Quebec
-0.19
[9] Difference between FP and AP Households with Children in Ontario
-0.88
[10] Difference in Difference: [8] [9]
-0.15
[11] Difference between FP and AP Households without Children in Quebec
0.92
[12] Difference between FP and AP Households without Children in Ontario
-1.06
[13] Difference in Difference: [11] [12]
0.18
[14] Triple Difference in Difference: [10] [13]
Note: *** Significant at 1% level; ** Significant at 5% level and * Significant at 10% level
[1]
[2]
[3]
[4]
[5]
[6]
[7]

Table A4: Total Food Expenditure Regression

Dummy: Children
Dummy: Francophone
Dummy: Quebec
Dummy: Children X Francophone
Dummy: Children X Quebec
Dummy: Francophone X Quebec
Dummy: Children X Francophone X
Quebec
Number of Children
Probability of Cable TV access
Household Income
Household Income2
Number of Household Members
Dummy: Male Blue Collar Occupation
Dummy: Male Pink Collar Occupation
Dummy: No Male Occupation
Dummy: Female Blue Collar Occupation
Dummy: Female Pink Collar Occupation
Dummy: No Female Occupation
Dummy: Home Ownership
Dummy: Recipient of Social Assistance
Male Education
Female Education
Dummy: Age of Household Head
Dummy: Dual Income Household
Dummy: Both Spouses Immigrants
Dummy: Year 1986
Dummy: Year 1990
Dummy: Year 1992
Dummy: 1st Quarter
Dummy: 2nd Quarter
Dummy: 3rd Quarter
Intercept
R-squared
= 0.2860

Estimates
-0.06595 **
0.103681 ***
0.095917 ***
-0.01662
0.121509 **
-0.11948 **
-0.03149
0.077471
0.31892
10.96209
-29.8775
0.065378
0.105398
0.035453
0.053403
0.091643
0.109057
0.08861
0.016561
-0.19545
0.078844
0.036822
0.001456
0.062419
-1.10974
0.166276
0.225544
0.234697
0.022009
0.006512
0.035923
2.95617

Note: *** Significant at 1% level; ** Significant at 5% level

and * Significant at 10% level

xi

***
***
***
***
***
***
*
***
**
***
***
***
***
***
***
*
**
***
***
***

*
***

Table A5: Difference-in-difference Expenditure in Total Food

Total Food
Expenditure
Amount Spent ($/week)
-4.94
[1] Difference between FP and AP Households with Children in Quebec
6.72
[2] Difference between FP and AP Households with Children in Ontario
-11.66 **
[3] Difference in Difference: [1] [2]
0.00
[4] Difference between FP and AP Households without Children in Quebec
10.60 **
[5] Difference between FP and AP Households without Children in Ontario
-10.60 **
[6] Difference in Difference: [4] [5]
-1.06
[7] Triple Difference in Difference: [3] [6]
Note: *** Significant at 1% level; ** Significant at 5% level and * Significant at 10% level

xii

Table A6: Heckman Results with Standard Errors

Estimation Method:

Dummy: Children
Dummy: Francophone
Dummy: Quebec
Dummy: Children X Francophone
Dummy: Children X Quebec
Dummy: Francophone X Quebec
Dummy: Children X Francophone X
Quebec
Number of Children
Probability of Cable TV access
Fast Food Price
Fast Food Price2
Household Income
Household Income2
Number of Household Members
Dummy: Male Blue Collar Occupation
Dummy: Male Pink Collar Occupation
Dummy: No Male Occupation
Dummy: Female Blue Collar Occupation
Dummy: Female Pink Collar Occupation
Dummy: No Female Occupation
Dummy: Home Ownership
Dummy: Recipient of Social Assistance
Male Education
Female Education
Age of Household Head
Dummy: Dual Income Household
Dummy: Both Spouses Immigrants
Dummy: Year 1986
Dummy: Year 1990
Dummy: Year 1992
Dummy: 1st Quarter
Dummy: 2nd Quarter
Dummy: 3rd Quarter
Intercept
Model Fit Measure

Heckman
Purchase Decision
Std.
Estimates
Error
-0.032
0.060
-0.265
0.075
-0.064
0.074
0.178
0.123
0.188
0.102
0.077
0.096
-0.358
0.161
-0.063
0.028
0.468
0.112
-8.608
0.640
1.039
0.073
5.157
1.400
-12.512
7.921
0.016
0.007
-0.091
0.046
-0.139
0.039
-0.067
0.037
-0.362
0.101
0.075
0.032
-0.067
0.033
0.018
0.032
-0.233
0.055
0.05
0.009
0.017
0.009
-0.014
0.001
-0.01
0.060
-0.166
0.034
0.442
0.057
0.509
0.085
0.497
0.079
0.126
0.034
0.118
0.034
0.151
0.035
16.687
1.335
0.106
Psuedo R2

ln(Expenditure)
Std.
Estimates
Error
-0.004
0.057
0.082
0.072
0.062
0.060
-0.092
0.123
-0.013
0.099
-0.098
0.095
-0.011
0.160
0.091
0.319
0.336
-0.008
0.649
4.964
0.028
0.123
0.068
0.011
0.128
0.029
0.05
-0.057
-0.062
0.024
-0.035
-0.002
0.029
-0.022
-0.03
0.031
0.05
0.061
0.064
0.162
0.752

0.028
0.108
0.029
0.001
1.386
7.607
0.007
0.046
0.039
0.037
0.107
0.031
0.034
0.032
0.061
0.009
0.009
0.001
0.066
0.031
0.036
0.040
0.037
0.035
0.035
0.035
0.141

Note: Dummy variables, where White collar is defined as managerial, professional or

teaching; Blue collar is defined as Farming, Fishing, Forestry, Mining, Processing,
Manufacturing or Construction. Pink collar is defined as Clerical, Sales or Service. The
education variable is defined as follows: 1 = less than 9 years; 2 = some secondary; 3 =
some post-secondary; 4 = post-secondary certificate; 5 = university degree. Those with
missing education were dropped.

xiii

Table A7: Pair-wise correlation of Variables used in regression analysis

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

ln(Fast Food)
Dummy: Children
Dummy: Francophone
Dummy: Quebec
Dummy: Children X Francophone
Dummy: Children X Quebec
Dummy: Francophone X Quebec
Dummy: Children X Francophone X Quebec
Number of Children
Probability of Cable TV access
Fast Food Price
Household Income
Number of Household Members
Dummy: Male Blue Collar Occupation
Dummy: Male Pink Collar Occupation
Dummy: No Male Occupation
Dummy: Female Blue Collar Occupation
Dummy: Female Pink Collar Occupation
Dummy: No Female Occupation
Dummy: Home Ownership
Dummy: Recipient of Social Assistance
Male Education
Female Education
Dummy: Age of Household Head
Dummy: Dual Income Household
Dummy: Dual Income Household X Income
Dummy: Both Spopuses Immigrants
Dummy: Year 1986
Dummy: Year 1990
Dummy: Year 1992
Dummy: 1st Quarter
Dummy: 2nd Quarter
Dummy: 3rd Quarter

10

11

12

13

14

15

16

1
0.09
-0.09
-0.09
0.01
0.02
-0.09
0.00
0.09
0.13
0.11
0.20
0.05
0.02
0.02
-0.05
-0.02
0.09
-0.07
0.07
-0.07
0.16
0.10
-0.17
0.11
0.13
0.00
-0.01
0.05
0.04
0.00
0.01
0.02

1
0.00
-0.01
0.58
0.60
0.00
0.55
0.88
0.43
-0.02
0.15
0.28
0.11
0.09
-0.07
0.04
0.13
0.07
0.18
0.05
0.19
0.22
-0.35
0.18
0.14
0.09
0.01
0.01
-0.04
0.02
0.01
-0.01

1
0.77
0.47
0.36
0.91
0.44
-0.01
-0.16
0.34
-0.11
-0.03
-0.01
0.04
0.03
0.00
0.01
0.06
-0.07
0.06
-0.05
-0.09
-0.02
0.00
-0.04
-0.13
0.01
-0.01
0.01
0.00
0.00
0.00

1
0.36
0.45
0.86
0.41
-0.02
-0.21
0.44
-0.14
-0.03
-0.01
0.02
0.03
0.01
-0.01
0.07
-0.07
0.06
-0.05
-0.09
0.01
-0.01
-0.05
-0.11
0.02
-0.03
0.03
0.00
0.00
-0.01

1
0.86
0.42
0.94
0.49
0.19
0.16
0.05
0.15
0.06
0.06
-0.03
0.03
0.08
0.06
0.10
0.05
0.09
0.07
-0.20
0.10
0.06
-0.05
0.02
0.00
-0.01
0.00
0.00
0.01

1
0.41
0.92
0.52
0.19
0.20
0.04
0.15
0.06
0.05
-0.03
0.03
0.07
0.08
0.11
0.05
0.10
0.09
-0.20
0.11
0.07
-0.03
0.03
-0.01
-0.01
0.01
0.00
0.00

1
0.48
-0.01
-0.17
0.39
-0.12
-0.04
0.00
0.04
0.03
0.01
0.01
0.06
-0.06
0.05
-0.05
-0.10
-0.01
0.00
-0.04
-0.13
0.02
-0.01
0.02
0.00
0.00
-0.01

1
0.46
0.17
0.19
0.04
0.14
0.05
0.05
-0.02
0.03
0.07
0.06
0.10
0.04
0.08
0.07
-0.19
0.10
0.06
-0.05
0.03
-0.01
-0.01
0.00
0.00
0.01

1
0.44
-0.02
0.13
0.22
0.11
0.09
-0.06
0.04
0.10
0.10
0.18
0.03
0.20
0.20
-0.32
0.16
0.12
0.09
0.01
0.01
-0.03
0.02
0.01
-0.02

1
-0.14
0.48
0.24
0.07
0.26
0.04
0.06
0.16
0.12
0.51
-0.16
0.19
0.20
-0.14
0.31
0.30
0.11
-0.05
0.00
-0.05
-0.02
0.01
0.00

1
0.06
-0.09
-0.03
0.00
0.06
0.00
0.01
0.05
0.01
0.02
0.02
0.00
0.01
-0.01
0.01
-0.06
-0.08
0.29
0.28
-0.02
-0.01
0.02

1
0.09
0.04
0.09
-0.09
0.01
0.18
-0.10
0.37
-0.21
0.46
0.35
-0.12
0.38
0.60
0.09
-0.10
0.12
0.16
-0.01
0.01
0.02

1
0.00
-0.02
-0.08
0.00
-0.01
-0.05
0.06
0.17
-0.09
-0.01
-0.12
-0.04
-0.03
0.01
-0.03
-0.03
-0.09
0.00
0.01
-0.01

1
-0.12
-0.14
0.08
0.12
0.05
0.07
-0.05
0.03
-0.01
-0.12
0.13
0.06
0.06
-0.01
-0.02
-0.01
0.02
-0.01
0.01

1
-0.18
0.04
0.19
0.00
0.05
-0.07
0.11
0.05
-0.12
0.22
0.15
0.03
0.00
-0.03
0.00
0.01
0.01
-0.01

1
0.01
-0.02
0.31
0.06
0.04
0.03
-0.09
0.24
-0.11
-0.11
0.01
-0.04
0.02
0.07
0.00
0.00
0.00

xiv

Table A6: Pair-wise correlation of Variables used in regression analysis (highlighter numbers are significant at 5% level)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

ln(Fast Food)
Dummy: Children
Dummy: Francophone
Dummy: Quebec
Dummy: Children X Francophone
Dummy: Children X Quebec
Dummy: Francophone X Quebec
Dummy: Children X Francophone X Quebec
Number of Children
Probability of Cable TV access
Fast Food Price
Household Income
Number of Household Members
Dummy: Male Blue Collar Occupation
Dummy: Male Pink Collar Occupation
Dummy: No Male Occupation
Dummy: Female Blue Collar Occupation
Dummy: Female Pink Collar Occupation
Dummy: No Female Occupation
Dummy: Home Ownership
Dummy: Recipient of Social Assistance
Male Education
Female Education
Dummy: Age of Household Head
Dummy: Dual Income Household
Dummy: Dual Income Household X Income
Dummy: Both Spopuses Immigrants
Dummy: Year 1986
Dummy: Year 1990
Dummy: Year 1992
Dummy: 1st Quarter
Dummy: 2nd Quarter
Dummy: 3rd Quarter

17

18

19

20

21

22

23

24

25

26

27

28

29

1.00
-0.07
-0.08
0.01
-0.03
0.00
-0.02
-0.04
0.05
0.03
0.07
0.02
-0.02
0.00
0.00
0.02
-0.01

1.00
-0.34
0.09
-0.09
0.13
0.09
-0.19
0.29
0.23
0.04
-0.04
-0.02
0.06
0.06
-0.01
-0.01

1.00
0.15
0.03
0.02
-0.02
0.26
-0.12
-0.16
0.04
-0.03
-0.05
0.09
-0.02
0.02
-0.01

1.00
-0.20
0.26
0.18
0.18
0.22
0.25
0.06
-0.02
0.06
0.00
-0.02
0.01
0.01

1.00
-0.19
-0.13
-0.03
-0.14
-0.12
-0.05
0.02
-0.03
0.03
0.03
-0.03
-0.01

1.00
0.17
-0.25
0.32
0.34
0.13
-0.01
0.06
0.00
0.01
0.00
0.00

1.00
-0.21
0.25
0.29
0.08
-0.02
0.04
0.04
0.00
0.02
0.00

1.00
-0.19
-0.14
0.02
-0.01
0.02
0.03
-0.02
0.00
0.01

1.00
0.86
0.08
0.25
-0.05
-0.10
0.00
0.00
0.01

1.00
0.08
0.13
0.00
-0.01
-0.01
0.00
0.02

1.00
0.00
-0.31
-0.38
0.00
0.01
0.00

1.00
-0.02
-0.30
0.02
0.00
-0.02

1.00
0.00
-0.01
-0.01
0.02

xv

30

31

32

1.00
-0.01 1.00
0.00 -0.34 1.00
0.01 -0.34 -0.33