(2015) "Local versus Foreign: A Microeconomic Analysis of Cultural ...

WPS 15-05-1 Working Paper Series

Local versus Foreign: A Microeconomic Analysis of Cultural Preferences

Maria Masood

May 2015

Local versus Foreign: A Microeconomic Analysis of Cultural

Preferences

Maria Masood∗

University of Geneva

Geneva School of Economics and Management - Global Studies Institute

This version: August 2015

Abstract: Relying on an augmented social interaction model, the present study investigates the

influence of two, usually separated, determinants of preference across varieties of an addictive good:

a liking by consuming capital and the preferences of other individuals through social interactions.

I use a novel dataset from a survey about cultural preferences of the French population to elicit the

influence of these two determinants on taste for movies, differentiated by their country of origin.

Thereby, this paper sheds light on an important debate about the impact of trade on cultural

diversity: in spite of the growing evidence about the bipolarization of cinema markets around

national and American movies, little is known about the formation of preferences for a given origin

at the individual level. The results reveal that prior exposure to the different varieties is the

crucial ingredient when it comes to movie preferences: through its direct impact but also because

it dampens the social interaction effect.

Keywords: Addictive goods, Social Interactions, Determinants of preferences, Cultural diver-

sity

JEL Classification Numbers: D12, F61, Z13 .

∗e-mail: [email protected]. I thank the Centre Maurice Hawlbachs for providing me with the data. I amgrateful to Celine Carrere for her continued guidance. I thank all the participants of the PhD workshop in Geneva andof the European Workshop on Applied Cultural Economics (EWACE). For their comments and helpful suggestions, Ialso thank Francoise Benhamou, Jeremy Lucchetti, Michele Pellizzari and Frederic Robert-Nicoud. Remaining errorsare my own.

1

1 Introduction

How much we value the consumption of some goods may depend on our prior consumption of

that good (Stigler and Becker, 1977) as well as the preferences of our peers. This is theoretically

the case of most addictive goods such as cigarettes, drugs but also cultural items. Indeed, as Becker

(1992) puts it, “addictive behaviour is generally more subject to pressure from peers than other

behaviour”. To some extend, this can also be applied to experience goods: when an individual does

not know the quality of a given item prior to its consumption, he relies on the information he gets

from his peers (Moretti, 2011) as well as his personal experience with that variety (Israel, 2005).

Hence, the analysis of the preferences for these types of goods differs from the traditional analysis

and requires an adapted framework.

Apart from Reif (2014), there exists to our knowledge, no comprehensive study taking simulta-

neously into account the influence of both determinants on preferences for addictive goods:1 social

interactions and the prior consumption, that is referred here as the liking by consuming capital. In

this paper, I test a new integrating framework that assembles these two, usually separated, theoret-

ical contributions to the understanding of the preferences for the different varieties of an addictive

good, namely movies. I focus on movies as cultural items were taken as a relevant illustration of the

rational addiction process in the seminal contribution of Stigler and Becker (1977). Furthermore,

cultural economics have always emphasized the role of past exposure on current cultural demand.

This postulate has been empirically confirmed in Disdier et al. (2010) and Schulze (1999) about

the geographical pattern of cultural trade: the larger the stock of past consumption from a given

origin, the larger the current imports from that origin2. According to these authors, this result

implies that the existence of an addictive behaviour “tend to reinforce strong and long-established

market positions in cultural exports” (Disdier et al., 2010 p585).

The contribution of the present study is twofold: first, I propose a new integrating framework

to test the influence of past exposure (addiction process) and social interaction simultaneously

1Though Reif (2014) provides a comprehensive theoretical and empirical framework, his work is distinct fromours as he does consider a linear model of social interaction where the addiction process serves to solve the reflectionproblem, while in this paper the use of a discrete choice model eliminates the identification issue and I simply integratethe addiction process as a component of the utility function.

2Furthermore, the focus on the preferences across movies’ country of origin is appealing as it is the less arbitraryclassification of movies (as opposed to genres or quality ranking) and it has been emphasized as a relevant criteriafor movie preferences (Gazley et al., 2011).

2

on movie preferences; second, I shed some light on the debate about the influence of trade on

cultural diversity through the eliciting of the determinant of movie preferences at the individual

level. Usually, consumer behaviour is apprehended only through revealed preferences (consumption

choice) but in this paper, the novel dataset I use allows to focus on preferences themselves. An

original microeconomic approach is proposed, based on the answers to the following question: “what

is the geographical origin of your preferred movies?” from a French household survey conducted in

2008 that provides a representative picture of the French population (see section 6.2.1). Focusing

solely on preferences, instead of revealed preferences, should allow a better understanding of one

of the transmission channels, from the utility perspective, explaining the concentration in global

cinema markets3. Indeed, in spite of the claim about the bipolarization of cinema markets around

national and American productions (Cohen and al., 2008) and the number of theoretical models

predicting the negative influence of trade in cultural goods on cultural diversity (Bala and Van

Long, 2005; Janeba, 2007; Rauch and Trindade, 2009; Maystre, Olivier, Thoenig and Verdier, 2014),

little is known about the determinants of preferences for a given origin of a movie at the individual

level although significant progress has been made notably through the inclusion of consumption

externalities (Bala and Van Long, 2005; Janeba, 2007, Rauch and Trindade, 2009). It is then

relevant to go one step further and identify empirically at the individual level the role of preferences,

and this is the second contribution of this paper: why a consumer might prefer watching a foreign

movie rather than a local one, supposedly closer to his sociocultural context?4

This paper relies on the most recent contributions in the literature to confront the challenges

related to the estimation of a social interaction model. The use of a non linear framework allows to

avoid the reflection problem (Manski, 1993), then I address the potential endogeneity issue through

different instrumentation strategies, I also test the existence of unobservable effects, the robustness

to omitted variable bias and finally assess the extend of the Angrist spurious correlation problem

(2014). The results highlight the importance of two determinants of movie preferences: liking by

consuming capital and the choice of other individuals through social interactions. More specifically,

3Considering preferences instead of consumption choices allows us to focus only on the utility yielded from thedifferent varieties excluding the (direct) influence of other factors, such as prices, availabilities, etc

4Existing analysis of cultural good preferences based on individual data, consist essentially in the empirical testingof the existence of a social stratification of musical tastes (Peterson and Kern, 1996; Van Eijck 1997; Coulangeon, 2005;Coulangeon and Lemel, 2009). Moretti (2011) also tested the impact of social interaction on the movie consumptiondecisions of individuals when quality is ex ante uncertain. For that purpose, he does not use individual data butrather approximate this channel through the evolution of the number of screens devoted to a given movie.

3

prior exposure to the different varieties is the crucial ingredient when it comes to movie preferences

through its direct impact on preferences but also because it dampens the social interaction effect.

The paper is organized as follows: the theoretical framework is exposed in section 2. Section

3 describes the empirical methodology while section 4 reports the results of the estimation and

section 5 concludes.

2 The theoretical framework

2.1 Literature review on the relevant determinants for cultural preferences

The objective of this study is to assess the influence of prior exposure and social interaction

on the preferences among different origins of movies, specifically local and foreign varieties. In

this section, I present a brief literature review of these two determinants applied to the movie sector.

As put by Watts (2007), what individuals look for in their cultural experience is “not so much

to experience the best of everything as it is to experience the same things as other people and

thereby also experience the benefits of sharing”, therefore, tastes should be strongly influenced by

an individual’s peers. Bernheim (1994) was one of the first economists to develop a theoretical

framework generalizing the modelling of consumption with social interactions. He introduces in

the utility function a parameter according to which individuals care about their social status,

relying on the way he is, or his preferences are, perceived by others. This perception appears to be

normative in his model: the group prefers some inclinations and despises others. If the individual

highly values the benefits of having a status compared to his intrinsic preferences, then individuals

tend not to deviate from the norm, complying with the group’s rules. This joins the conclusion

of Bourdieu’s work (1979) who evidenced the role of cultural tastes as a signal, a faire-valoir, for

the individual in a more general framework of inherited cultural capital. In this perspective, an

individual consumes a certain good that helps him in affirming his belonging to a certain social

category, in order to maximise its “self-image”. There exists an important strand in the economic

literature that emphasizes this aspect of social behavior. Theoretical progress have been notably

made by Becker and Murphy (1988), Akerlof and Kranton (2000), Brock and Durlauf (2001, 2007),

De Giorgi and Pellizzari (2013) and Angrist (2014) in identifying the influence of social interactions

4

on individual’s behavior.

The cultural economics literature claims the inadequacy of the standard preference theory:

the preference of an individual for a variety of a cultural good depends upon his skills developed

in appreciating that given variety (Stigler and Becker, 1977, Levy-Garboua and Montmarquette,

1996, 2002). I rely on the seminal theoretical contribution of Stigler and Becker (1977) that

describes the increasing marginal utility yielded by some consumption goods, explaining the

existence of addictive behavior (relevant for cultural goods but also for less healthy items such as

cigarettes, drugs, etc.). The authors made an important contribution to the literature through

the inclusion of prior exposure to a given cultural item as a determinant of present appreciation,

emphasizing the existence of what I term here a “liking by consuming” process. The higher the

prior exposure to a given variety, the higher the ability to produce enjoyment from its consumption.

This determinant is more detailed in the next section.

These two identified set of arguments: social interactions and the liking by consuming capital,

are unified and empirically tested for the first time within the binary choice model with social

interactions developed by Brock and Durlauf (2007) and later exposed in the Handbook of Blume,

Brock, Durlauf, Ioannides (2011). Indeed, this model is widely recognized as a sound workhorse for

studying binary decision process integrating social interactions (as it solves the reflection problem)

and is also well suited for taking into account the specificity of cultural goods preferences as exposed

below.

2.2 A binary choice model with social interactions

I rely on the well established Blume, Brock, Durlauf, Ioannides (2011) framework, henceforth

BBDI, where individual i’s choice wi ∈ (−1, 1) is the result of the comparison of some payoff

function Vi. In this framework, individual choices are coded such that Vi(1), (Vi(−1)) represents

the utility yielded from preferring the local (foreign) variety of movies. Accordingly, individual

choices are influenced by:

• individual specific characteristics: a vector of observables Xi and unobservable εi

• a vector of group specific characteristics: yg

5

• subjective expectation of individual i about the average choice wg,−i of the social group to

which he belongs : mei,g.

In line with this framework, the difference in the utility function is additive in these different

elements (I adopt the same notation as in BBDI, 2011):

Vi(1)− Vi(−1) = k + cXi + dyg + Jmei,g − εi (1)

I assume that the random term εi is independently and identically distributed across individuals,

e.g. it does not depend on individual and/or group characteristics5.

Fεi|Xi,Yg = Fε(εi)

The derivation of the basic structure of the model is provided in appendix. From this, the

probability of preferring the local variety is equal to one if (1) is strictly positive: (Vi(1)−Vi(−1)>0):

Pr(Vi(1)− Vi(−1)>0) = Pr(εi<k + cXi + dyg + Jmei,g)

= Fε(k + cXi + dyg + Jmei,g)

Note that the condition is a strict superiority, implying that the “indifferent” option (for which

both utility levels are strictly equal, i.e. : Vi(1) = Vi(−1)) falls into the foreign, or the non-local,

preference, therefore the “indifferent” choice is also explained by the model. As in BBDI (2011),

I assume that subjective individual expectation about the group choice mei,g “coincides with mg

the mathematical expectation of the average choice in group g given yg”, this is known as the

self-consistency property, then mg is defined as:

mei,g = mg = 2

∫Fε(k + cX + dyg + Jmg)dFX|g − 1 (2)

where FX|g is the within group distribution of Xi. mg does not have any closed form solution and

will be approximated as the proportion of individuals in group g that prefers the local variety in the

empirical section6. In the rest of the section, I describe individually each component of equation

5In order to reduce the risk of violation of that assumption, I add in a robustness check as much individualvariables of control as possible to minimize the correlation across people who belong to the same group.

6As in BBDI (2010), the model exhibits multiple equilibria only when the social interaction term is “very strong”(Lee et al.,2014), see also (BBDI, 2010 - Theorem A.3.). Yet, subsequent studies (Gaviria and Raphael, 2001;Harris and Lopez-Valcacel, 2004; Krauth, 2006; Kooreman and Soatevent, 2007 and Lee et al.,2014) show that theestimated social interaction effect is not strong enough to generate multiple equilibria. Accordingly, the multiplicityof the equilibria is not considered as a concern in this study.

6

(1).

2.3 Individual characteristics: Liking by consuming capital

Following Stigler and Becker (1977), individual characteristics can be apprehended through,

what I term here, the individual liking by consuming capital7. The utility derived from a given

variety of a movie k is a function of the skills developed in appreciating it, a liking by consuming

capital (Xikt) that depends on the accumulation of earlier enjoyment of that variety (∑τ=t−1

τ=0 Mikτ )

in addition to education and other human capital that can have an effect on the inclination of the

individual toward certain variety of movies (Ai):

Xikt = h(∑τ=t−1

τ=0 Mikτ , Ai)

Where,

Mikτ = Mikτ (eikτ , Xikτ )

eikτ being the time allocated to that variety at the time period τε[0, t − 1]. Accordingly, the

enjoyment of variety k (Mikτ ) is a function of the time spent on it as well as the appreciation skills

(or liking by consuming capital) developed till that moment. In turn, the liking by consuming

capital is produced through an endogenous process with the accumulation of earlier experience of

that variety, but depends also on the characteristics of the individual defined in Stigler-Becker’s

model as “education and other human capital”. The basic assumption in the Stigler-Becker model

is that tastes are stable but the skills of the consumer to appreciate art increases with his experience

8. Accordingly, there exists a beneficial addictive process if:

∂Xikt∂Mikτ

>0

And if, as hypothesized by Stigler and Becker, ∂Mikτ∂eikτ

>0, then :

7In their study, Stigler and Becker explain the level of utility yielded from music consumption by the level ofmusic human capital that I term here liking by consuming capital.

8Levy-Garboua and Montmarquette (1996) developed an alternative model of learning by doing process in culturaltastes where there exists an uncertainty. In their model, consumers are not aware of their true preferences that arerevealed through the accumulation of consumption experiences. The consumer might experience positive or negativesurprise and adjust its preferences accordingly. Though the assumption about the underlying process is differentfrom Becker and Stigler model (stochastic instead of deterministic), their approach also reckons the influence ofaccumulated experience on present tastes.

7

∂Xikt∂eikτ

>0

From these expressions, the liking by consuming capital of a given variety (Xikt) relies on

an accumulation of time allocated to that variety in the past (prior exposure). Moreover, the

authors assume that education and other human capital component, Ai, do play a role in the liking

by consuming capital, though the authors do not specify the components of this “other human

capital”. Consistent with the stable preference assumption posited by Stigler and Becker, there is

no discount rate affecting the accumulation of liking by consuming capital: what is relevant is the

experience itself and not the moment when it was experienced9. Following these assumptions, I

can write:

Xikt = h(∑τ=t−1

τ=0 eikτ , Ai) = h(Eikt, Ai)

Conveniently,∑τ=t−1

τ=0 eikτ , henceforth noted Eik can be approximated through the accumulated

amount of time spent watching a given variety of movie. In the framework of the binary choice

problem, the probability of selecting a given variety depends upon the comparison of related liking

by consuming capital. An increase in ∆Ei = Ei,local−Ei,foreign10 reflects a better relative knowledge

of local movies, indicating a longer experience with the local variety that should translate into

a higher probability of selecting it in line with the addictive hypothesis. Regarding individual

characteristics Ai, they are of course invariant across varieties, yet they can influence differently

the tastes for the different varieties. For instance, the place of birth (an individual characteristics)

can influence the taste for a given origin of movie in different ways: a Japanese born individual is

more likely to prefer Japanese movies than a French born. Hence, assuming as in BBDI (2011),

that the difference in the utility levels is additive in the different components, I can explicit Xi in

(1) as follows:

Xi = Xi,local −Xi,foreign = α(Ei,local − Ei,foreign) + (βlocal − βforeign)Ai

And so,

Xi = α∆Ei + βAi (3)

9Indeed, according to Becker and Stigler “a consistent application of the assumption of stable preferences impliesthat the discount rate is zero; that is, the absence of time preference”. Hence, I do not assume the existence of adiscount rate affecting the effects of past exposure. This approach is convenient here as the data does not permit todistinguish the time period of each prior exposure with the different varieties.

10From now on, I omit the time subscript as the analysis is limited to a one period survey.

8

2.4 Group characteristics or the contextual effect

An individual preference might be affected by the group to which he belongs not because of the

behavior of that group (a peer effect) but simply because of shared characteristics. Indeed, the peer

effect implies that individual’s tastes are influenced by the prevalence of these tastes in his social

group while the contextual effect (the characteristics specific to the group that might affect the

choice of the individual) indicates that the preference of an individual depends on the exogenous

characteristics of his group, such as the average education level of his social mates. These two

social interaction processes generate very different implications in terms of policy intervention: the

first effect implies an amplified response to exogenous shocks (multiplier effect) unlike the latter.

For that reason, I need to distinguish these two types of social interactions.

In accordance with the identified determinants of individual preferences above (i.e. the liking by

consuming capital), relevant group specific characteristics can be approximated through the group

average liking by consuming capital. Indeed, it is standard (Manski,1993, Gaviria and Raphael,

2001, Brock and Durlauf, 2007, BBDI, 2011) to consider the contextual effect yg equivalent to the

average individual characteristics of the group, i.e. yg,−i ≡ Xg,−i. In other words:

yg,−i ≡ Xg,−i = α∆Eg,−i + βAg,−i

With Eg,−i (Ag,−i) being the average value of Ei (Ai) for individuals belonging to group g

excluding i11.

2.5 Social interactions

As explained in section 2.1, I assume that the tastes of an individual are in part driven by his

aspiration for social recognition, such that his preferences are influenced by the preferences of his

social group justifying the inclusion of the choice of others in (1) identified as wg,−i:

wg,−i = 1ng−1

∑j 6=iwjg ∈ [0, 1]

with ng the size of the group g and wjg the preference choice of the individual j in group

g. Yet, how to properly identify the social interaction term J? In an influential paper, Manski

11The fact that yg and mg are linearly dependent could pose an identification problem, yet I show in the nextsection that in the case of a non linear model this problem does not arise.

9

(1993) underlines the correlation between the average choice of the group and yg that implies

an impossibility to disentangle the contextual effect (yg) from the peer effect. This identification

problem is termed the reflection problem in Manski (1993), yet, Brock and Durlauf (2001, 2007)

showed that this identification problem “does not arise in the binary choice case” as, by definition,

the binary choice model reveals the nonlinear relationship between wg,−i and yg justifying our use

of the BBDI framework12.

2.6 Empirical specification

According to the previous discussion, the choice of an individual about his movie preferences

can be approximated through the following latent utility function:

Vi(1)− Vi(−1) = k + c(α∆Ei + βAi)︸ ︷︷ ︸individual liking by consuming capital

+ d(α∆Eg,−i + βAg,−i)︸ ︷︷ ︸contextual effect

+ J ¯wg,−i︸ ︷︷ ︸social interactions

−εi

that can be rewritten:

Vi(1)− Vi(−1) = k + γ∆Ei + ζAi + κ∆Eg,−i + θAg,−i + Jwg,−i − εi (4)

The utility function relies on: the individual (group) relative prior exposure to the different

varieties ∆Ei ( ∆Eg,−i), a vector of individual (group) characteristics: human capital Ai (Ag,−i)

and the fraction of the social peers that prefers the local variety wg,−i.

Following Stigler-Becker’s addictive hypothesis, I expect a positive γ (as an increase in ∆Ei

reflects a better relative knowledge of local movies) indicating that a longer experience with the local

variety translates into a higher probability of selecting it as the preferred one. Yet, an endogeneity

12Blume, Brock, Durlauf and Ioannides (2011, p67) state in the following theorem, the sufficient, not necessary,conditions, satisfied in discrete choice models, that allows for the identification of the parameters:

1. conditional on (xi, yg), the random payoff terms εi are i.i.d according to Fε;Fε(0) = 0.5.

2. Fε is absolutely continuous with density dFε; dFε is positive almost everywhere on the interval (L,U), whichmay be (−∞,∞).

3. for at least one group g, conditional on yg, each element of the vector xi varies continuously over all R andsupp(xi) is not contained in a proper linear subspace of RR

4. yg does not include a constant; each element of yg varies continuously over all R : at least one element of d isnon-zero; and supp(yg) is not contained in a proper linear subspace of RS

Then, k, c, d, J and Fε are identified up to scale.Intrinsically to the binary choice model, mg (2) cannot be a linear function of the other regressors in the equation.

This is valid here where there is a sufficient variation in Xi and yg, or when these vectors have sufficiently widesupports (as stated in the third assumption of the theorem).

10

problem may arise as the relative prior exposure might be higher because the individual prefers the

local movies. Therefore, the covariance between the explanatory variable and the residuals would be

different from zero. To confront this potential bias, I implement an instrumentation strategy using

the time spent watching TV and the number of DVD bought as instrumental variables that should

reflect the influence of the environment incentives on individual consumption (a more detailed

description is provided in the next section). The influence of human capital, Ai, on the preference

selection is less clear as the theory does not provide any insight on the direction of the effect of

these variables on individual preferences.

Following the social interaction argument exposed above, I expect a positive J signaling the

compliance of the individual with the preference prescription of his social group. The size of

the effect should provide information on the strength of these social prescriptions: the larger the

coefficient, the larger the identity payoffs yielded from preferring the local variety relative to the

cost of deviating from the group’s tastes13.

3 Data and econometric specification

3.1 Data description

The analysis is based on the “Survey about cultural practices of the French population” con-

ducted by the French Ministry of Culture and Communication (DEPS, 2008) in 2008. This survey

was conducted on a representative sample of 5,004 individuals randomly selected in the French

metropolitan population (aged 15 years and over), stratified by region and agglomeration using

quota method and questioning face to face at the home of the interviewee. To make sure this

database is a representative sample of the French population in 2008, I compare in appendix (Sec-

tion 6.2.1) the distribution of the population in the sample and in the national statistics along

different dimensions. The questionnaire includes a section about socio-economic characteristics as

well as the cultural habits of the interviewees. More central to the analysis, the survey comprises

13In spite of its potential impact on the preference selection, the quality of the producing country in filmmakingis not explicitly integrated in the utility function for two main reasons. First, it is difficult to assess and compare thequality of filmmaking of a (whole) national industry since it is a highly subjective matter and might be highly variablewithin a given country. Second, characteristics of the movies do not vary across individuals, but it is rather theirreaction, depending on their own characteristics, that exhibit some variation. Nevertheless, this (relative) qualityargument is most likely incorporated in the constant term k.

11

a question about the preferences of individuals in terms of geographical origin of cultural goods,

such that: In general, the movies you prefer are (. . . ):

• French;

• American;

• Neither French, nor American but from another country, which one? ;

• I do not care.

5003 (out of 5004) individuals provided an answer to this question that will be used as the

dependent variable in the analysis. Domestic movie is the first category preferred in the sample

(40,6%), followed by American movies (28,1%). A similar share of respondents (30,7%) declares to

be indifferent to the geographical origin of the movie. Only 1% of the sample declares preferring a

movie that is neither French nor American, but rather Asian (11 individuals), British (9), German

(6), Italian (2), Turkish (1), Arabic (1) or Other European (2)14. As already mentioned in the

introduction, the preferences of the French population exhibit a bipolarization between local and

American movies, that altogether represents 70% of overall preferences. Hereafter, foreign movies

refers to American and other varieties altogether.

3.2 A binary response model

3.2.1 Presentation of the model

Assuming that the error terms εi in (4) follows a normal distribution, the choice of the individual

can be modelled using a binomial choice model such as the probit model:

Prob[preferred local=1]=φ(V (1)− V (−1) > 0),

where φ is the standard cumulative normal probability distribution.

14One could worry about a misreporting bias as individuals may feel some pressure to declare preferring Frenchmovies for instance. Yet, it seems to be a minor concern here. Indeed, the descriptive statistics (Table 5 in appendix)provide reassuring information about that concern: individuals declaring preferring local movies are on average thosewho have the least experience with foreign movies (39% of foreign movie watched as opposed to 65% for those whoprefer American movies). Furthermore, figures about relative prior exposure are consistent with movie watchinghistory: those that prefer domestic movies are those that have the highest level of relative prior exposure.

12

3.2.2 Variables description

In this section, I provide a thorough description of the variables that are used in the empirical

section.

∆Ei refers to the difference in terms of exposure of the individual to the different alternatives

(the different origins of movies). Conveniently, the survey includes a question about the prior

experience of the individual with French and foreign movies: a list of movie titles is submitted

to the interviewee who has to select the ones he has already watched15. From the answer to this

question, I compute a ratio of watched movies for each corresponding alternatives (local: priori,local

and foreign: priori,foreign ) indicating the level of familiarity of the individual with these different

varieties. If the percentage for French movies is equal to 100%: the individual is considered to be an

experienced cinephile for French movies, while a 0% indicates that the interviewee’s prior exposure

to the given variety is not significant. Accordingly, the higher ∆Ei, as ∆Ei ≡ Ei,local − Ei,foreign,

the superior is his prior exposure to local movies compared to foreign ones. However, though the list

of movies submitted to the interviewee is exogenous to his preferences (the list is not adapted to the

individual’s response to the preference question), there might remain a reverse causality problem:

because the individual prefers local movies, ∆Ei is higher. In order to confront this issue, I instru-

ment ∆Ei with a continuous variable indicating the amount of time spent watching movies. This

variable should capture the influence of the environment16 on the exposure to the different varieties.

For instance, if foreign movies are largely more displayed than local ones on TV, in DVD stores

and on downloading platforms, then watching a lot of movies should negatively impact the relative

prior exposure of the individual ∆Ei, and this is eventually suggested from the first stage results.

In addition, this variable satisfies the exclusion restriction since its impact on the probability of

preferring a given variety only goes through the relative experience with the different alternatives17.

15The question in the survey is the following : “Dans la liste suivante, quels sont les films que vous avez deja vusque ce soit au cinema, a la television ou en DVD”. Note that this list is submitted to each interviewee whatever hispreferences are. Note that I compute ratio instead of nominal values as the number of local and foreign movies inthe list are not the same.

16Apart from the social group influence which is already captured through wg,−i.17One can think about two limitations about this instrument: first, due to the existence of a general equilibrium

effect, the content broadcasted on TV reflects the preference in the population; second, because the preferred varietyis displayed on TV, the individual watch more TV. The first issue is likely minor in our case, as the distribution ofpreferences in the whole population is almost evenly spread (see Table 5) and the existence of a large number of freechannels in France should allow individuals to watch the variety they want (I can thus consider the TV supply, interms of content, exogenous). The second issue refers to the potential violation of the exclusion restriction, yet in thesample 72% of the individual declare watching TV without knowing the program a priori. To confirm the robustness

13

In the model, vector Ai represents both the education and the human capital of the individual.

Accordingly, Ai comprises a variable indicating the latest educational degree obtained but also

relevant characteristics of human capital that should impact movie preferences between local and

foreign varieties: the geographical origin of the individual (his birthplace), his travel history (how

many times has the individual travelled abroad over the past 12 months), and his cultural habits

depicted through the number of books read during the year.

Note that these two sets of argument, namely ∆Ei and Ai, are also included in a group-wise

form in order to capture the contextual effect. This set of variables, notably ∆Eg, is not instru-

mented in the empirical part as the risk of reverse causality is null and the potential existence

of omitted variable is addressed through the inclusion of large number of individual covariates

(section 6.6). I assume here that there is no unobservable group effects, yet I discuss and test that

possibility in appendix (section 6.5).

wg,−i allows to test to what extend the social group of the individual influences his tastes

(social interaction effect). Recall that wg,−i is defined as the proportion of individuals in the group

g that prefers local versus foreign movies. Yet, how to define relevant social categories? how to

delimitate the social groups acting as prescriptors of tastes? Following the literature, I adopt the

homophily assumption for the reference group such that a social group is identified as a category of

individuals belonging to the same age-gender-occupation group. For instance, men aged between

15-24 and having a blue-collar occupation are considered to belong to the same social group 18.

This combination of characteristics is a relevant proxy for identifying influential social groups in

a randomized population sample assuming that an individual shares his ideas, opinions with the

persons he spends most of the time with and to whom he resembles the most19. In appendix,

of the instrumentation strategy, I perform the same regressions on that restricted sample of individuals that for suredo not violate the exclusion restriction (Table 7).

18More precise identification of social groups is also provided where I apply a more disaggregated definition of ageand occupation group, but as a consequence, the groups that satisfy the necessary number of members to edict tasteprescription falls: because of this tradeoff between precision and sample, I emphasize the latter and prefers the lessprecise identification of social groups. See in appendix - section 6.3 for more details about the identification of socialgroups.

19Of course, the data does not allow me to capture small peer group interactions, yet I believe that the size of thesample allows me to get a broader picture: an average of the small group incentives over the whole population.

14

Figures 6 and 7 provides the size and the fraction of the individuals within each social group that

declares preferring the local variety.

All variables are summarized in Table 1 and summary statistics are provided in appendix

(Table 4 and Table 5).

Table 1: Description of the variablesIndividual characteristics

Birth Place of birth –dummy equal to one if the interviewee is born in FranceTravel Dummy equal to one if has travelled over the past 12 monthsBooks # Books read in the past 12 monthsDegree Latest Educational degree obtainedRegion 21 French administrative regions (metropolitan)

Social category

SPC Socio-Professional Category –Categorical: Inactive/blue/white-collar/RetiredFemale Dummy equal to one if a womanAge Categorical variable : 15-24 / 25-49 / 50+

Prior exposure

Elocal % of movie watched from the list of local moviesEforeign % of movie watched from the list of foreign movies

4 Results

I present in Table 2 the results of the estimation of the probit model. The first column (I) is the

benchmark regression. I perform in columns (II) to (IV) the instrumentation strategy. In column

(III), I exclude from the sample the individuals who declared to be indifferent to the geographical

origin of movies.

4.1 Interpretation of the average marginal effects

The results presented in Table 2 suggests a significant influence of the liking by consuming

capital and the social interactions term. The coefficients reported are the average marginal effects

of a one percentage point increase on the probability of the positive outcome, that is preferring the

local variety over foreign.

Regarding the impact of human capital, it appears that an individual born in France is more

likely to declare preferring local movies. The level of education and the number of books read are

both positively and significantly correlated with the probability of selecting French movies as the

15

Table 2: Estimation resultsDependent variable: preferring local movies

Technique: Probit IV Probit IV Probit IV Probit

(I) (II) (III) (IV)

∆Ei 0.00223*** 0.00937*** 0.0102*** 0.0107***(0.000258) (0.00147) (0.00101) (0.000339)

birthplacei -0.0552* 0.0792** 0.112*** 0.102***(0.0284) (0.0356) (0.0316) (0.0186)

traveli -0.0241** 0.0229 0.0261 0.0345**(0.0111) (0.0151) (0.0177) (0.0161)

booksi -0.00461 0.0139*** 0.0179*** 0.0173***(0.00396) (0.00501) (0.00354) (0.00295)

degreei 0.00103 0.0149*** 0.0161*** 0.0176***(0.00420) (0.00409) (0.00447) (0.00296)

∆Eg,−i -0.000429 -0.00250 -0.00704*** -0.00413**(0.00267) (0.00240) (0.00243) (0.00177)

birthg,−i 0.186 0.213 0.0911 0.114(0.241) (0.182) (0.204) (0.140)

travelg,−i -0.0289 0.0987 0.0590 0.092(0.0959) (0.104) (0.147) (0.0796)

booksg,−i 0.0316* -0.00101 -0.0265 -0.0134(0.0190) (0.0249) (0.0270) (0.0142)

degreeg,−i -0.00448 -0.00770 0.00429 -0.0140

(0.0108) (0.00993) (0.0110) (0.0086)wg,−i 0.00865*** 0.00383** 0.00527* 0.0015

(0.000359) (0.00168) (0.00269) (0.00103)

First stage

TV timei -0.319** -0.328**(0.124) (0.151)

DVDi -1.287***(0.374)

Observations 4,955 3,701 2,514 3,834DWH test 4.58** 7.42*** 33.32***F statistic - 1st stage 25.41 17.31 25.41

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei in columns (II) and (III) is the individual time spent watching movies, in column (IV)

is the number of DVD bought.*** p<0.01, ** p<0.05, * p<0.1

16

preferred ones. Furthermore, I observe a relatively high impact of education and reading habits

(relative to the other determinants) on movie preferences. These results can be interpreted as

evidencing a positive relationship between the level of human capital and the taste for local movies.

As expected, a higher exposure with the local variety, compared to the foreign, is positively cor-

related with a preference for the related type of movie. However, because the prior exposure variable

might be endogenous, which is confirmed by the Durbin-Wu-Haussman test, the corresponding co-

efficient is likely biased. Therefore, I focus on the results of the instrumentation strategy presented

in column (II). The prior exposure variable is instrumented with a continuous variable indicating

the amount of time spent watching movies20. In the first stage regression, the instrument is signifi-

cantly (at the 5% level) and negatively correlated with the relative prior exposure of the individual,

suggesting that the environment of the individual tends to favor the consumption of foreign varieties

over local, at least in France21. As discussed in the variables description section, the coefficient of

the relative exposure is consequently downward biased. The estimation of the instrumented coeffi-

cient suggests that a one percentage point increase in the relative prior exposure ∆Ei is associated

with a 1 percentage point increase in the probability that the individual prefers the French movies.

As expected from the model, the higher the prior exposure to local movies relative to foreign ones,

the higher the probability of preferring the former. Yet, there exists a risk that because foreign

varieties are more exposed on TV, there might be a direct reverse correlation between the depen-

dent variable and the instrument. In order to confront that potential problem (and because of the

constraints in terms of available instruments in the survey), I estimate the same instrumentation

strategy on a restricted sample of individuals who declare being always unaware about the programs

displayed on TV before watching it and thus explicitly excluding those individuals who spend time

watching TV because they prefer certain types of programs. The corresponding results (section 6.4

- Table 7) confirm the robustness of the results regarding our instrumentation strategy. Using an

alternative instrumental variable: the number of DVD bought (column IV) yields the same value

20The instrument used is characterized by a high number of missing values, thus restricting the sample. Yet, theresults from the baseline regression between the total and the reduced sample are similar suggesting the absence ofa sample selection problem. The corresponding results are available upon request.

21It is interesting to note that this first stage result corroborates the observations made by a recent report of theFrench Cinematographic Agency which revealed that “American films are distributed in smaller numbers but moreintensively. During the first week of release they are screened on average in twice as many points of sale (269 in 2011)as French films (131 in 2011). For one French film out of two, less than 50 prints circulate. Less than a quarter ofAmerican films are in this situation” (Paris, 2014).

17

for the coefficient of relative prior exposure. These results are important as it is the first time that

such a relationship is established for the audiovisual sector, and has crucial, though controversial,

meaning. Indeed, these results suggest that the implementation of protectionist measures, such as

local content requirement, might be justified as a way to promote cultural diversity22, or at least

the survival of the (taste for) domestic production.

Turning to the impact of social interactions on preference selection, the corresponding results

indicate that a 10 percentage point increase in the relative fraction of the social groups that prefers

the local variety increases the probability of also preferring local movies by 3.2 percentage points.

Yet, the significance of this variable is sensitive to the instrumentation strategy, as it is no more

significant when using an alternative instrument (the number of DVD bought in column 5). In

appendix (section 6.3 - Table 6), I apply the same strategy using a much looser, or less restrictive,

definition of social groups that eventually yields the same results. In a recent article, Angrist (2014)

evidenced the high risk of obtaining a spurious social interaction term due to a simple mechanical

relationship between group behavior and individual choice. Hence, the social interaction term could

be inflated by the existence of a common influence that can not be interpreted as a peer effect. I

approximate the size of this bias in section 6.7 - Table 10 and evidence that only 1.7% of the social

interaction marginal effect found in Table 2 can be attributed to a spurious correlation.

The sample comprises the individuals declaring to be indifferent to the geographical origin of

the movie they prefer, and following the definition of the model I included those individuals in the

group who prefers the “non-local” movies (as the preference choice results from a strict difference

in utility level - see section 2.2). To make sure the results are not driven by this category, I perform

in column (III) the same strategy as in column (II) but excluding those indifferent individuals. The

results obtained are eventually similar. The similarity of the results is unsurprising as the model

predicts the indifferent choice through the same process: those individuals that are indifferent

should have a relative prior exposure and a social interaction term in between those preferring local

and those preferring foreign, as confirmed in Table 5.

Overall, it appears that the contextual effect is not significant in our model, as all related group-

wise variables (apart from the average relative prior exposure) are not significant probably due to

22Intentionally, I am careful not to mention welfare here as the model does not aim at delivering any insights orconclusions in terms of welfare.

18

the collinearity with the individual-wise variables. In addition, I provide in appendix a robustness

check testing the existence of unobservable characteristics (Section 6.5 - Table 8) and the existence

of an omitted variable bias (Section 6.5 - Table 9). The results obtained eventually suggest that

unobservable group characteristics are not a major problem in our model.

4.2 Going a bit further: exploiting the nonlinearity of the model

So far, I have considered the average marginal effects for the variables of interest meaning

that I have considered the average of marginal effects obtained for individuals having different

characteristics. Yet, it is interesting to see how these effects may vary according to the value of the

other characteristics (variables) of the individual23. More interestingly, I graph the average marginal

effect of relative prior exposure on the probability of preferring the local variety for different values

of the social interactions variable (Figure 1) and vice-versa (Figure 2)24.

Figure 1: Average Marginal Effect of Relative Prior Exposure differentiated by Social Interactionlevels

Computed from the specification (II) in Table 2Confidence interval set at 90% level

Quite intuitively, it appears that the influence of the relative prior exposure on the local pref-

23As in the probit procedure, the marginal effect of a given variable is conditioned to the value of the othervariables. In the linear probability model, however, the partial effect is constant.

24Implementing the same procedure with other variables, namely education level, number of books read, travelhistory, does not yield insightful results: the marginal effects hardly varies along the different values of these variables.

19

erence outcome is reinforced when social interactions goes in the same direction (i.e. when wg,−i

increases). Even so, the marginal effect of relative prior exposure, and its significance, hardly varies

along the different values of the social interaction term.

Figure 2: Average Marginal Effect of Relative Social Interactions differentiated by Prior exposurelevels

Computed from the specification (II) in Table 2Confidence interval set at 90% level

However, the picture is different when looking at the reverse relationship. Indeed, the impact of

social interactions is null, e.g. not significant, if the value of the relative prior exposure variable is

below −20 or above 20, implying that if the individual has watched too much foreign (local) movies

compared to local (foreign) ones and has thus built a larger liking by consuming capital for a given

variety, then, an increase in the taste of his group for local movies has no significant impact on his

preferences. On the other hand, a one percentage point increase in the fraction of the members of

the social group that prefers the local variety has a significant and the highest impact when the

individual is already more familiar with the local alternative (when the relative prior exposure is

around 20). In addition, implementing a t-student test reveals that the average marginal effect

of wg,−i when ∆Ei = 2025 is significantly superior to the average marginal effect of wg,−i when

∆Ei is below a certain value, i.e -60, (Table 3). Having watched far more foreign movies than

25The benchmark value of relative prior exposure is set at 20 for the t-test, since the highest marginal impact ofwg,−i is obtained for this value from Figure 2.

20

local ones does influence the magnitude of peer effect on individual preferences. This result clearly

confirms the influence of prior exposure on the impact of social interactions. It appears that the

relative prior exposure is the crucial ingredient when it comes to movie preferences. As the social

interaction term is not only lower but its significance depends on the prior exposure argument.

Table 3: T-test resultsHypothesis Statistic Result

J |∆E=−80< J |∆E=20 2,64 non rejectionJ |∆E=−60< J |∆E=20 2,55 non rejectionJ |∆E=−40< J |∆E=20 2,18 non rejectionJ |∆E=0< J |∆E=20 0,52 rejectionJ |∆E=40< J |∆E=20 0,16 rejectionJ |∆E=60< J |∆E=20 0,76 rejectionJ |∆E=80< J |∆E=20 1,62 rejection

This test is computed from the results in Table 2 - column (II)The test is computed at the 10% level threshold

5 Concluding remarks

The economic literature has extensively described the dominance of a very few exporters in

world cultural markets using traditional trade arguments (monopolistic competition features) that

focus essentially on the supply side. However, very little is known about the demand side, though

the literature is shifting toward consumption arguments (consumption externalities, hysteresis effect

and non-homotheticity of demand) at the aggregate level. This study contributes to this corpus

through the investigation of the determinants of movie preferences, differentiated by the country

of origin, at the individual utility level. To that extend, I tested a new integrating framework that

assembles two, usually separated, theoretical contributions to the understanding of cultural tastes:

namely social interactions and, what I term here, the liking by consuming capital (that is a function

of human capital and prior exposure to the different varieties).

The empirical results highlight the importance of these two determinants of movie preferences.

Corroborating the addictive theory of cultural good consumption, higher familiarity (prior expo-

sure) with the national variety compared to the foreign one is correlated with a higher probability

of preferring local movies. This result provides evidence for the relevance of cultural policies aimed

at promoting the consumption of local cultural goods at the expense of foreign ones. The impact of

21

social interactions on individual preferences appears to be conditioned on the value of relative prior

exposure: if an individual has accumulated too much experience with the foreign variety, then the

social interaction term has no impact at all. In addition, the social interaction effect is highly sen-

sitive to the specification applied. Relying on a non-linear social interactions model, I confirm the

robustness of the results to the many perils confronted in related analysis through the estimation of

different instrumentation strategies and alternative definitions of social group but also to the test-

ing of the existence of unobservable characteristics and of the Angrist issue. Hence, prior exposure

to the different varieties appears to be the crucial ingredient when it comes to movie preferences,

as the social interaction term effect is not only lower but its significance highly depends on the

prior exposure. These results suggest that the implementation of protectionist measures, though

controversial, such as local content requirement, can be useful in promoting cultural diversity, or

at least the survival of the (taste for) domestic production.

References

Akerlof, G. A. & Kranton, R. E. (2000), ‘Economics and identity’, The Quarterly Journal of

Economics 115(3), 715–753.

Angrist, J. D. (2014), ‘The perils of peer effects’, Labour Economics 30(C), 98–108.

Bala, V. & Van Long, N. (2005), ‘International trade and cultural diversity with preference selec-

tion’, European Journal of Political Economy 21(1), 143–162.

Becker, G. S. (1992), ‘Habits, addictions, and traditions’, Kyklos 45(3), 327–345.

Becker, G. S. & Murphy, K. M. (1988), ‘A theory of rational addiction’, The Journal of Political

Economy 96(4), 675–700.

Bernheim, B. D. (1994), ‘A theory of conformity’, Journal of Political Economy 102(5), 841–877.

Blume, L. E., Vienna, I., Durlauf, S. N., Brock, W. A. & Ioannides, Y. M. (2011), ‘Identification

of social interactions’, Handbook of Social Economics pp. 853–964.

Bourdieu, P. (1979), Distinction: A social critique of the judgement of taste, Routledge.

22

Brock, W. A. & Durlauf, S. N. (2001), ‘Discrete choice with social interactions’, The Review of

Economic Studies 68(2), 235–260.

Brock, W. A. & Durlauf, S. N. (2007), ‘Identification of binary choice models with social interac-

tions’, Journal of Econometrics 140(1), 52–75.

Clark, A. E. & Loheac, Y. (2007), ‘it wasnt me, it was them! social influence in risky behavior by

adolescents’, Journal of Health Economics 26(4), 763–784.

Cohen, D., Verdier, T., Debonneuil, M., Lorenzi, J.-H. & Benghozi, P.-J. (2008), La mondialisation

immaterielle, Documentation francaise.

Coulangeon, P. (2005), ‘Social stratification of musical tastes: questioning the cultural legitimacy

model’, Revue Francaise de Sociologie 46(5), 123–154.

Coulangeon, P. & Lemel, Y. (2009), ‘Les pratiques culturelles et sportives des francais: arbitrage,

diversite et cumul’, Economie et Statistique 423(1), 3–30.

De Giorgi, G. & Pellizzari, M. (2013), ‘Understanding social interactions: Evidence from the class-

room’, The Economic Journal 124(579), 917–953.

DEPS (2008), Pratiques culturelles des francais 2008, DEPS [producteur] Centre Maurice Halb-

wachs (CMH) [diffuseur].

Disdier, A.-C., Tai, S. H., Fontagne, L. & Mayer, T. (2010), ‘Bilateral trade of cultural goods’,

Review of World Economics 145(4), 575–595.

Feld, J. & Zlitz, U. (2014), ‘Understanding peer effects: On the nature, estimation and channels of

peer effects’, SWOPEC Working Paper No. 596 .

Gaviria, A. & Raphael, S. (2001), ‘School-based peer effects and juvenile behavior’, Review of

Economics and Statistics 83(2), 257–268.

Gazley, A., Clark, G. & Sinha, A. (2011), ‘Understanding preferences for motion pictures’, Journal

of Business Research 64(8), 854–861.

Harris, J. E. & Lopez-Valcarcel, B. (2004), ‘Asymmetric social interaction in economics: cigarette

smoking among young people in the united states, 1992-1999’, NBER working paper N 10409 .

23

INSEE (2012), Parts de marche selon la nationalite des films en 2011, Tableaux de l’economie

francaise.

Israel, M. (2005), ‘Services as experience goods: An empirical examination of consumer learning in

automobile insurance’, The American Economic Review 95(5), 1444–1463.

Janeba, E. (2007), ‘International trade and consumption network externalities’, European Economic

Review 51(4), 781–803.

Krauth, B. V. (2006), ‘Simulation-based estimation of peer effects’, Journal of Econometrics

133(1), 243–271.

Lee, L.-f., Li, J. & Lin, X. (2014), ‘Binary choice models with social network under heterogeneous

rational expectations’, Review of Economics and Statistics 96(3), 402–417.

Levy-Garboua, L. & Montmarquette, C. (1996), ‘A microeconometric study of theatre demand’,

Journal of Cultural Economics 20(1), 25–50.

Levy-Garboua, L. & Montmarquette, C. (2002), ‘The demand for the arts’, CIRANO .

Manski, C. F. (1993), ‘Identification of endogenous social effects: The reflection problem’, The

Review of Economic Studies 60(3), 531–542.

Maystre, N., Olivier, J., Thoenig, M. & Verdier, T. (2014), ‘Product-based cultural change: is the

village global?’, Journal of International Economics 92(2), 212–230.

Moretti, E. (2011), ‘Social learning and peer effects in consumption: Evidence from movie sales’,

The Review of Economic Studies 78(1), 356–393.

Paris, T. (2014), New approaches for greater diversity of cinema in europe, Technical report, Eu-

ropean Commission.

Peterson, R. A. & Kern, R. M. (1996), ‘Changing highbrow taste: from snob to omnivore’, American

Sociological Review 61, 900–907.

Rauch, J. E. & Trindade, V. (2009), ‘Neckties in the tropics: a model of international trade and

cultural diversity’, Canadian Journal of Economics/Revue Canadienne d’economie 42(3), 809–

843.

24

Reif, J. (2014), ‘Addiction and social interactions: Theory and evidence’, Unpublished working

paper .

Schulze, G. G. (1999), ‘International trade in art’, Journal of Cultural Economics 23(1-2), 109–136.

Soetevent, A. R. & Kooreman, P. (2007), ‘A discrete-choice model with social interactions: with

an application to high school teen behavior’, Journal of Applied Econometrics 22(3), 599–624.

Stigler, G. J. & Becker, G. S. (1977), ‘De gustibus non est disputandum’, The American Economic

Review 67(2), 76–90.

Van Eijck, K. (1997), ‘The impact of family background and educational attainment on cultural

consumption: A sibling analysis’, Poetics 25(4), 195–224.

Watts, D. J. (2007), Is Justin Timberlake a product of cumulative advantage?, New York Times

Magazine 15 April 2007.

25

6 Appendix

6.1 Basic structure of the binary choice model with social interactions

In this section, I can write the underlying structure of the model developed in section 2 for a

single group g. Following strictly Blume, Brock, Durlauf and Ioannides (2011), choices are coded

wi ∈ {−1, 1} and hi = k + cxi + dyg, so the utility level yielded by a given variety can be written :

Vi(wig) = hiwig −J

2E((wig − wg,−i)2) + ηi(wig) (5)

with wig =∑j 6=i wjgI−1 and ηi(wig) is a variety specific random term. As in Brock and Durlauf

(2001), the individual faces a negative payoff when he deviates from other’s mean choice. As

w2ig = 1,

−J2 (wig − wg,−i)2 = Jwigwg,−i − J

2 (1 + w2−ig)

Replacing it in (5) where the last term is independent from wig, I obtain

Vi(wig) = hiwig + Jwigmeig + ηi(wig)

with meig =

∑j 6=i Ewig |FiI−1 . Accordingly, I obtain

Vi(1)− Vi(−1) = 2hi + 2Jmeig − εi (6)

with εi = ηi(−1)− ηi(1). This explains equation (1).

26

6.2 Descriptive statistics

6.2.1 Representativity of the sample

The validity of the analysis relies on the assumption that the sample is a representative picture

of the French population in 2008. To confirm this, I provide a comparison of actual data about

age, regions and occupation groups provided by the French national statistics institute (INSEE)

with our survey data.

These graphs allow us to conclude positively about the representativeness of the survey data, as

all (age/region/occupation) groups are similarly distributed as the actual French population was

in 2008.

Figure 3: Representativeness of age groups in actual and survey data

Source: INSEE, 2013 and survey dataIn addition, implementing a t-test reveals that the difference in the mean of the two samples is not significant at any

conventional confidence level.

27

Figure 4: Representativeness of French regions in actual and survey data

Source: INSEE, 2013 and survey dataIn addition, implementing a t-test reveals that the difference in the mean of the two samples is not significant at any

conventional confidence level.

Figure 5: Representativeness of French occupation groups in actual and survey data

Source: INSEE, 2013 and survey data

In addition, implementing a t-test reveals that the difference in the mean of the two samples is not significant at any

conventional confidence level.

28

6.2.2 Distribution of preferences

In this section, I present a selection of descriptive statistics in Table 4, and then according to

preference outcomes in Table 5.

Table 4: Sample meansVariables Mean Std err Minimum Maximum Missing valuePersonal characteristicsAge 46.33 19.019 15 95 0Birth abroad 0.1001 0.3002 0 1 2Travel past 12 months 0.2768 0.4475 0 1 0Books past 12 months 2.7773 1.5887 1 6 45Female 0.5216 0.4996 0 1 0EducationLess than Baccalaureate 0.5269 0.4993 0 1 670Baccalaureate to Bachelor 0.3461 0.4757 0 1 670More than Bachelor 0.1269 0.3329 0 1 670OccupationInactive 0.1862 0.3893 0 1 0Blue Collar 0.3185 0.466 0 1 0White Collar 0.2304 0.4211 0 1 0Retired 0.2648 0.4413 0 1 0Prior exposureLocal movies watched 67.89% 29.667 0 100% 0Foreign movies watched 52.16% 30.016 0 100% 0Social interactionw−i,g 40.56 17.44 0 67.27 2Viewing characteristicsTV time 3.25 4.32 0 102 1274DVD 42.07 78.25 0 998 518Channel bouquet 0.81 0.39 0 1 167Number of TV 2.79 0.93 1 5 0Number of computer 2.01 0.89 1 4 0Agglomeration size (1000s)Rural 0.24 0.43 0 1 0<20 inhabitants 0.17 0.37 0 1 0<100 inhabitants 0.13 0.34 0 1 0<200 inhabitants 0.06 0.23 0 1 0>200 inhabitants 0.23 0.42 0 1 0Paris 0.17 0.37 0 1 0

Author’s calculations.Note: The “books” variable is categorical with value 1: 0 books, value 2: 1 to 4 books, value 3: 5 to 9, value 4: 10

to 19, value 5: 20 to 49 and value 6: more than 50.

29

Table 5: Descriptive statistics by preference outcomePrefers French Prefers American Prefers Other Is Indifferent

# Mean sd # Mean sd # Mean sd # Mean sdEducation<Baccalaureate 1748 0.58 0.49 1230 0.52 0.5 31 0.39 0.5 1324 0.47 0.5<Bachelor 1748 0.3 0.46 1230 0.39 0.49 31 0.16 0.37 1324 0.37 0.48>Bachelor 1748 0.12 0.33 1230 0.09 0.29 31 0.45 0.51 1324 0.16 0.36RegionParis 2029 0.15 0.36 1405 0.19 0.4 33 0.36 0.49 1536 0.24 0.43Champagne 2029 0.02 0.12 1405 0.03 0.16 33 0 0 1536 0.03 0.18Picardie 2029 0.03 0.17 1405 0.04 0.19 33 0.03 0.17 1536 0.02 0.16Hte-Normandie 2029 0.04 0.19 1405 0.04 0.19 33 0.03 0.17 1536 0.02 0.13Centre 2029 0.05 0.22 1405 0.03 0.17 33 0 0 1536 0.04 0.19Bse-Normandie 2029 0.02 0.15 1405 0.02 0.16 33 0.03 0.17 1536 0.02 0.15Bourgogne 2029 0.03 0.18 1405 0.02 0.13 33 0 0 1536 0.04 0.18Nord 2029 0.06 0.23 1405 0.08 0.27 33 0.06 0.24 1536 0.08 0.27Lorraine 2029 0.03 0.18 1405 0.05 0.23 33 0.03 0.17 1536 0.03 0.18Alsace 2029 0.03 0.18 1405 0.04 0.19 33 0.18 0.39 1536 0.02 0.14Fche-Comte 2029 0.03 0.16 1405 0.02 0.13 33 0 0 1536 0.01 0.1Pays de Loire 2029 0.06 0.24 1405 0.05 0.21 33 0 0 1536 0.06 0.23Bretagne 2029 0.07 0.25 1405 0.05 0.22 33 0.06 0.24 1536 0.03 0.16Poitou-Cte 2029 0.04 0.19 1405 0.03 0.16 33 0.03 0.17 1536 0.02 0.15Aquitaine 2029 0.06 0.23 1405 0.05 0.22 33 0.06 0.24 1536 0.04 0.19Midi-Pyrenees 2029 0.04 0.2 1405 0.03 0.18 33 0 0 1536 0.05 0.22Limousin 2029 0.02 0.12 1405 0.01 0.09 33 0 0 1536 0.01 0.1Rhone-Alpes 2029 0.09 0.29 1405 0.09 0.28 33 0.06 0.24 1536 0.11 0.31Auvergne 2029 0.03 0.17 1405 0.01 0.11 33 0.03 0.17 1536 0.03 0.16Languedoc 2029 0.03 0.18 1405 0.03 0.18 33 0 0 1536 0.05 0.22Pvce-CteAzur 2029 0.08 0.27 1405 0.1 0.3 33 0.03 0.17 1536 0.06 0.23Individual charactBirth abroad 2028 0.08 0.28 1405 0.11 0.31 33 0.15 0.36 1535 0.11 0.32Travel 2029 0.25 0.43 1405 0.28 0.45 33 0.55 0.51 1536 0.3 0.46Books 2029 2.75 1.61 1405 2.66 1.52 33 3.35 1.89 1536 2.9 1.61Female 2029 0.56 0.5 1405 0.5 0.5 33 0.52 0.51 1536 0.49 0.5PriorForeign 2029 39.23 28.37 1405 65.37 25.77 33 56.28 30.92 1536 57.04 28.97PriorLocal 2029 59.72 29.88 1405 76.65 26.49 33 53.94 36.57 1536 70.96 29.09Age group15-24 2029 0.08 0.26 1405 0.28 0.45 33 0.15 0.36 1536 0.13 0.3325-49 2029 0.3 0.46 1405 0.52 0.5 33 0.48 0.51 1536 0.5 0.550 + 2029 0.62 0.48 1405 0.2 0.4 33 0.36 0.49 1536 0.38 0.49SPC groupInactive 2029 0.15 0.36 1405 0.26 0.44 33 0.21 0.42 1536 0.17 0.37Blue Collar 2029 0.25 0.43 1405 0.42 0.49 33 0.27 0.45 1536 0.32 0.47White Collar 2029 0.18 0.39 1405 0.22 0.41 33 0.39 0.5 1536 0.3 0.46Retired 2029 0.42 0.49 1405 0.1 0.3 33 0.12 0.33 1536 0.21 0.41Social interactionw−i,g 2029 47.86 17.03 1404 32.1 14.28 33 37.11 16.05 1536 38.74 16.34Viewing characteristicsTV time 1308 2.77 4.4 1195 3.77 4.84 26 4.92 5.28 1200 3.23 3.57DVD 1702 30.94 60.38 1347 52.49 115.45 29 51.9 115.45 1407 45.39 88.21Channel bouquet 1973 1.85 0.36 1379 1.78 0.41 27 1.85 0.36 1457 1.79 0.41Number of TV 2029 2.73 0.88 1405 2.96 0.97 33 2.15 0.83 1536 2.72 0.95Number of computer 2029 1.81 0.86 1405 2.18 0.86 33 2.3 0.95 1536 2.11 0.89Agglomeration size (1000s)Rural 2029 0.26 0.44 1405 0.21 0.41 33 0.18 0.39 1536 0.25 0.43<20 inhabitants 2029 0.2 0.4 1405 0.15 0.36 33 0.03 0.17 1536 0.15 0.36<100 inhabitants 2029 0.12 0.32 1405 0.14 0.35 33 0.09 0.29 1536 0.14 0.35<200 inhabitants 2029 0.06 0.24 1405 0.06 0.24 33 0.03 0.17 1536 0.05 0.21>200 inhabitants 2029 0.22 0.41 1405 0.27 0.44 33 0.3 0.47 1536 0.21 0.41Paris 2029 0.13 0.34 1405 0.17 0.38 33 0.36 0.49 1536 0.21 0.40Total observations 2029 1405 33 1536Weighted fraction 40.60% 28.10% 0.66% 30.70%

Author’s calculations.

30

6.3 Identification of the social groups

In order to test the influence of the peers on the tastes of the individual, I have defined the

peers, or the social group, as the individuals belonging to the same gender-age-occupation group.

I can use two alternative definition of these social groups: a restricted and a larger one. For the

larger definition, I focus on 4 subgroupings of the occupation variable (i.e.: inactive/blue collar /

white collar/ retired), and on 3 subgroupings of the age variable (i.e.: 15-24/25-49/50+). In the

restricted definition, I obtain a larger number of groups, and thus of prescription(see Figure 6 and

7). The restricted definition has the advantage of providing a more precise identification of social

groups, and tastes but is more likely to be less reliable as the sample size for some groups might not

be sufficient to identify the actual taste prescriptions. In addition, because there are social groups

for which the number of individuals is not sufficient, this procedure leads to a reduction of the total

sample and potentially a less precise identification of groups taste. As a consequence, I essentially

use the larger definition to test the social interaction variable in the main regressions (Table 2),

and use the restricted one as a robustness check (see Table 6). Figure 6 informs us about the size

of the different group identified26 and Figure 7 provides the taste prescription of each social group.

Figure 6: Identification of the social group samples

The larger definition corresponds to the grey categorization. NA is for non available.The number represents the observations for men (women) in the sample that declares preferring either local or

foreign movies.

For instance, Figure 7 indicates that 25,3% of the male student aged between 18 and 24 declares

preferring local movies while 32,5% of female students of the same age prefers domestic movies.

In Table 6, I perform the same analysis as in Table 2 with a more restrictive definition of social

groups than the one used in the main regressions. It appears that the impact of the social group’s

choice on individual preferences is similar to the one obtained in Table 2 - column 2 confirming the

robustness of the results with the use of an alternative definition of social groups.

26I identify the social groups as those having more than 3 individuals.

31

Figure 7: Prescriptions of the social group

The larger definition corresponds to the grey categorization.The number represents the fraction in the male (female in parenthesis) group that declares preferring the local

variety (as opposed to the foreign one).

Table 6: Using alternative definition of social groups

Dependent variable: preferring local movies

Technique: Probit IV Probit

(I) (II)

∆Ei 0.00232*** 0.00951***(0.000254) (0.00145)

birthplacei -0.0567** 0.0798**(0.0276) (0.0352)

traveli -0.0195* 0.0241(0.0102) (0.0154)

booksi -0.00448 0.0139***(0.00384) (0.00495)

degreei 0.00567 0.0152***(0.00498) (0.00425)

∆Eg,−i -0.00139 -0.00335**(0.00127) (0.00168)

birthplaceg,−i 0.140 0.146

(0.160) (0.124)travelg,−i 0.00644 0.0584

(0.0925) (0.110)booksg,−i 0.0354* 0.00335

(0.0185) (0.0186)degreeg,−i -0.0191 -0.0101

(0.0118) (0.00975)w−i,g,restricted 0.802*** 0.367**

(0.0382) (0.174)

Observations 4,949 3,698

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei is the individual time spent watching movies.

*** p<0.01, ** p<0.05, * p<0.1

32

6.4 Testing the robustness of the instrumentation strategy

In the empirical section, I applied an instrumentation strategy to confront the bias that could

arise due to a possible reverse causality between the relative prior exposure and the preference

selection. I use the time spent watching TV as an instrument with the intuition that an individual

that spends the whole day watching TV will be more sensitive to the incentives to watch a given

variety that would be relatively more broadcast compared to an individual that hardly watch TV.

Accordingly, the instrument should reflect the external incentives to watch a given variety, and thus

impact his relative prior exposure.

Table 7: Testing the robustness of the IV

Dependent variable: preferring local movies

Technique: IV Probit

∆Ei 0.00944***(0.00222)

birthplacei 0.111**(0.0485)

traveli 0.0124(0.0161)

booksi 0.0167***(0.00410)

degreei 0.0135*(0.00737)

∆Eg,−i -0.00316(0.00318)

birthplaceg,−i 0.254

(0.326)travelg,−i -0.0275

(0.112)booksg,−i 0.00343

(0.0288)degreeg,−i -0.000671

(0.0118)wg,−i 0.00334

(0.00219)

Observations 2,808F statistic - 1st stage 21.21

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei is the individual time spent watching movies.

*** p<0.01, ** p<0.05, * p<0.1

Though the existence of a great number of TV channels in France should imply the possibility of

33

watching any variety, one may question the absence of a direct correlation between the dependent

variable and the instrument: the individuals preferring a variety more available on TV could also

spend more time watching TV. To limit the potential risk of using that instrument (and because

the limits imposed by the survey on the choice of instruments), I perform the same instrumentation

as in Table 2 - column 2 on a restricted sample of individuals: excluding those who never watch TV

without knowing the program they are going to watch27 and thus explicitly excluding those who

because they love watching a certain type of program decide to spend more time in front of the

TV. The corresponding results, displayed in Table 7, confirm the significance and the magnitude of

the relative prior exposure impact on the preference outcome and tend to defend the assumption

that the instrument and the dependent variable are not directly correlated. Yet, again, it appears

that the impact of the social interaction variable is highly sensitive to the specification applied.

6.5 Robustness check: accounting for unobservable group characteristics

So far, I have assumed the absence of group specific unobservables, yet they constitute an

important obstacle in the identification strategy of many related analysis (BBDI, 2011; Brock and

Durlauf, 2007). Indeed, including a group specific fixed effect would prevent the identification of

the other group related parameters and most importantly the social interaction effect J . In the

framework of the analysis, the presence of group unobservables would imply that there are either:

• unobservable variables explaining why individuals would prefer a given type of movies and

belong to a given age-gender-occupation group. I add an occupation fixed effect28 to remove

a potential self-selection bias in Table 8 - column (I). The results shows that the inclusion of

an occupation fixed effect cancels out the impact of the social interaction variable suggesting

that there might exist unobservable characteristics explaining why an individual belong to a

given occupation category and declares preferring a given type of movie.

• common environmental incentives/factors causing group members to behave (prefer a given

variety) similarly even in the absence of social effect. From the discussion about the de-

terminants of movie preferences, the environmental factors should goes through the “prior

27The original question in the survey is as follows: “Vous arrive-t-il d’allumer la television en rentrant chez vous,sans connaıtre le programme?”. We excluded in Table 7 those who declared “Never”.

28I use a disagregated occupation variable with the following categories:Inactive-Student-Farmer-Worker-Employee-Intermediate occupation-Artisan-Executive-Retired.

34

exposure” argument and thus be controlled for using the “average prior exposure of the

group”. Accordingly, adding variables approximating external incentives such as the “aver-

age time spent watching TV of the group” should not remove the significance of the average

choice of the group variable. Columns (II) and (III) in Table 8 show that the impact of

environment incentives of the group, approximated through the time spent watching TV, in-

fluence individual choice via the average choice of the group. As the significance of the Hours

TV variable is cancelled out when the group preferences wg,−i is included (Table 8 - column

III)29. Therefore, the results suggest that the external incentives common to the group are

already captured.

29Note that average hours TV of the group is negatively correlated with the local preference evidencing thepreponderance of incentives favoring foreign varieties.

35

Table 8: Testing for the existence of unobservables

Dependent variable: preferring local movies

Technique: IV Probit

(I) (II) (III)

∆Ei 0.00903*** 0.00969*** 0.00940***(0.00185) (0.00113) (0.00143)

birthplacei 0.0749* 0.0852*** 0.0796**(0.0399) (0.0310) (0.0350)

traveli 0.0183 0.0253* 0.0230(0.0167) (0.0137) (0.0149)

booksi 0.0134** 0.0147*** 0.0139***(0.00523) (0.00470) (0.00501)

degreei 0.0138*** 0.0157*** 0.0150***(0.00445) (0.00367) (0.00401)

∆Eg,−i 0.00447 -0.000694 -0.00221(0.00417) (0.00284) (0.00252)

birthg,−i 1.034*** 0.122 0.173(0.357) (0.172) (0.177)

travelg,−i -0.0724 -7.88e-05 0.0656(0.138) (0.105) (0.104)

booksg,−i 0.0639 0.00729 -0.00373(0.0417) (0.0242) (0.0238)

degreeg,−i -0.0870*** -0.0244* -0.0105

(0.0256) (0.0125) (0.0109)wg,−i -0.000730 0.00290**

(0.00151) (0.00134)hoursTVg,−i -0.0651** -0.0216

(0.0291) (0.0183)

Observations 3,701 3,701 3,701

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei is the individual time spent watching movies. In column (I), occupation fixed effects

are included*** p<0.01, ** p<0.05, * p<0.1

36

6.6 Robustness check: adding more individual covariates

In the preceding sections, I assume the absence of omitted variables, yet one can worry that this

problem exists and may result in biased results. I address this question in this section (Table 9)

through the inclusion of the following additional individual covariates: in column (I): subscription

to private TV channels and the number of TV owned in the household (to reflect the quality of

TV subscription), the number of computer owned; in column (II): the size of agglomeration (in

terms of inhabitants); in column (III): the preferred French TV channels (with a dummy equal to

1 for each 7 main channels that should reflect the viewing taste of the individual) and in column

(IV): the level of income. All these variables are included in Table 9 as they might be interpreted

as omitted variables that could be correlated across individuals belonging to the same social group

and therefore lead to the violation of the iid assumption for the error term.

However, the results provided in Table 9 suggest the omitted variable bias is a minor concern.

Indeed, the inclusion of variables indicating the quality of TV subscription (that might be correlated

with consumption history and the social category of the individual) in column (I) hardly modifies the

magnitude and the significance of the variables of interest though it appears that having subscribed

to a private channel bouquet (possessing a large number of computer) is correlated with a positive

(negative) probability of preferring the local variety. The size of the agglomeration, in terms of

inhabitants, does not have any influence on the movie tastes. In column (3), information about

preferred local TV channels is integrated as it is likely to be correlated across individuals within

social groups and may influence at the same time the consumption history. Yet, the coefficient

associated to relative prior consumption hardly varies while the social interaction term remain

positive but less statistically significant. Finally, I add the income level (as income may affect

movie preference, and that there is a correlation between income, quality of TV subscription, and

TV viewing) in the last column though only half of the sample decided to answer to that question

making the results less reliable, again the coefficients of interest are still positive and significant.

37

Table 9: Results with additional individual covariatesDependent variable: preferring local moviesTechnique: IV Probit

(I) (II) (III) (IV)

∆Ei 0.00950*** 0.00948*** 0.00904*** 0.00755*(0.00149) (0.00153) (0.00210) (0.00408)

birthplacei 0.0737** 0.0732** 0.0707** 0.0581(0.0324) (0.0293) (0.0343) (0.0615)

traveli 0.0223* 0.0222* 0.0179 0.0163(0.0134) (0.0122) (0.0118) (0.0197)

booksi 0.0110** 0.0110** 0.00666 0.00255(0.00493) (0.00478) (0.00547) (0.00777)

degreei 0.0133*** 0.0133*** 0.0104*** 0.0103*(0.00402) (0.00414) (0.00387) (0.00524)

∆Eg,−i -0.00238 -0.00237 -0.000251 0.00191(0.00238) (0.00245) (0.00316) (0.00466)

birthg,−i 0.183 0.186 0.172 0.0947(0.193) (0.195) (0.202) (0.262)

travelg,−i 0.127 0.128 0.0866 -0.0190(0.116) (0.119) (0.123) (0.112)

booksg,−i 0.00733 0.00746 0.0261 0.0272(0.0241) (0.0246) (0.0301) (0.0322)

degreeg,−i -0.0153* -0.0152 -0.0192* -0.00455(0.00923) (0.00962) (0.0101) (0.0124)

wg,−i 0.00357** 0.00360** 0.00265* 0.00373*(0.00170) (0.00170) (0.00154) (0.00224)

Channel bouqueti 0.0481** 0.0485** 0.0440* 0.0532*(0.0231) (0.0241) (0.0253) (0.0299)

Nb of TVi -0.0343*** -0.0341*** -0.0288*** -0.0281**(0.0107) (0.0101) (0.0107) (0.0143)

Nb of computeri -0.00258 -0.00270 -0.00367 -0.00576(0.0107) (0.0104) (0.0100) (0.0137)

Agglomeration sizei -0.000101 -0.000490 -0.00369(0.00686) (0.00714) (0.00792)

TF1 vieweri -0.0212 -0.00668(0.0200) (0.0284)

FR2 vieweri 0.0349 0.0602(0.0276) (0.0402)

FR3 vieweri 0.0463 0.0819(0.0378) (0.0583)

Canalplus vieweri -0.00815 -0.0129(0.0257) (0.0246)

FR5 vieweri 0.0529** 0.0619**(0.0241) (0.0263)

ARTE vieweri 0.0653*** 0.0509(0.0215) (0.0332)

M6 vieweri -0.0275 -0.0408(0.0184) (0.0296)

Incomei -0.00105(0.00214)

Observations 3,584 3,584 3,584 2,962

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei is the individual time spent watching movies

*** p<0.01, ** p<0.05, * p<0.138

6.7 Robustness check: accounting for additional perils of peer effect estimation

Angrist (2014) evidenced another “peril” for the identification of the social interaction term:

there exists a risk of erroneously interpreting a purely mechanical relationship between the group

decision and the individual outcome as a true peer effect. Indeed, even in the absence of a social

effect there might exist a significant relationship between the average choice of the group and

individual’s choice simply reflecting the presence of a “shared influence” common to the group.

Accordingly, the social interaction term would capture the peer effect plus a kind of a group

random effect. The ideal solution to that problem would consist in a randomized experiment that

manipulate peer characteristics independently from individual characteristics to prove the causality

of the peer effect. In my setting, i.e. a non-experimental study, this option is unfeasible. Yet, instead

of solving this issue, I can approximate the size of the bias as in Feld and Zolitz (2014) and thus

infer the true value of the social interaction parameter. To do so, I perform the same regressions as

in Table 2 using “placebo” interaction variable: I randomly assign individuals to irrelevant social

groups (through the shuffling of wg,−i). Hence, the estimation of the social interaction term would

reveal the fraction that can not be interpreted as a true peer effect.

The corresponding results are provided in Table 10: the placebo interaction term wg,−irandom is

significant but its magnitude is very low: suggesting that the social interaction term estimated in

Table 2 is affected by the “Angrist mechanics” but marginally. Indeed, only 1.7% of the marginal

impact found in Table 2 (column 2) can be attributed to the Angrist mechanics rather than a true

social interaction effect.

39

Table 10: Robustness check: placebo peer effect

Dependent variable: preferring local movies

Technique: Probit IV Probit IV Probit IV Probit

(I) (II) (III) (IV)

∆Ei 0.00231*** 0.0102*** 0.0108*** 0.0109***(0.000299) (0.000654) (0.000322) (0.000192)

birthplacei -0.0536 0.0976*** 0.129*** 0.111***(0.0335) (0.0247) (0.0211) (0.0168)

traveli -0.0239** 0.0313*** 0.0359** 0.0346**(0.0110) (0.0118) (0.0144) (0.0152)

booksi -0.00343 0.0166*** 0.0194*** 0.0183***(0.00445) (0.00407) (0.00367) (0.00298)

degreei 0.000645 0.0171*** 0.0183*** 0.0183***(0.00415) (0.00362) (0.00449) (0.00295)

∆Eg,−i 0.0141*** 0.00110 -0.000624 -0.00268(0.00255) (0.00310) (0.00306) (0.00210)

birthplaceg,−i 0.509** 0.356** 0.393 0.182*

(0.242) (0.170) (0.242) (0.110)travelg,−i -0.126 0.0963 0.0837 0.119

(0.0987) (0.101) (0.124) (0.0776)booksg,−i 0.265*** 0.0546 0.0457 0.00211

(0.0198) (0.0374) (0.0432) (0.0193)degreeg,−i -0.106*** -0.0362** -0.0319** -0.0242**

(0.00910) (0.0155) (0.0154) (0.0105)wg,−irandom 1.87e-05 0.000638** 0.000620* 0.000628***

(0.000348) (0.000268) (0.000330) (0.000219)

Observations 4,953 3,699 2,512 3,896

Average marginal effects are reported –Robust standard errors in parentheses –Clustered at the region levelThe instrument used for ∆Ei in columns (II) and (III) is the individual time spent watching movies, in column (IV)

is the number of DVD bought.** p<0.01, ** p<0.05, * p<0.1

40