You are on page 1of 16

Available online at www.sciencedirect.

com
Journal of Health Economics 27 (2008) 249264
Asymmetric peer effects in the analysis of cigarette smoking
among young people in the United States, 19921999
Jeffrey E. Harris
a,
, Beatriz Gonz alez L opez-Valc arcel
b
a
Department of Economics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
b
Department of Quantitative Methods for Economics and Management, University of Las Palmas de Gran Canaria, Spain
Received 11 September 2006; received in revised form 20 July 2007; accepted 20 July 2007
Available online 4 December 2007
Abstract
We extend the recent literature on peer effects to test the possible role of asymmetric social inuences in the determination of
youth smoking. We analyzed cigarette smoking among people aged 1524 in approximately 90,000 households in the 19921999
U.S. Current Population Surveys. The presence of additional smoking sibling in a household, we estimated, raised a young persons
probability of smoking by 7.6%, while each non-smoking sibling lowered the probability by an estimated 3.5%. Moreover, the
overall deterrent effect of an increase in cigarette price on the probability of smoking was approximately 60% greater than the
estimated effect when peer inuences were held constant. The concept of asymmetric social inuence may have applications in
other elds, including labor economics, education, crime prevention, and group dynamics.
2007 Elsevier B.V. All rights reserved.
JEL classication: A120; C250; C720; D120; D620; D830; I12
Keywords: Social interaction; Tobacco; Price; Non-cooperative games
1. Introduction
Three recent, path-breaking papers in this Journal have focused on the role of social interactions or peer effects in
the economic analysis of cigarette smoking behavior. Powell et al. (2005), in a study of over 17,000 students sampled
from 202 high schools in the U.S. in 1996, found that the prevalence of smoking at the students school was an
independent predictor of the students own smoking participation. Lundborg (2006), in a study of over 2600 students
aged 1218 years who attended one of 149 school classes in a medium-sized town in Sweden in 20012002, found
that the prevalence of smoking in the respondents class was likewise an independent and signicant predictor of the
respondents smoking participation. Clark and Etil e (2006), in a study of nine successive annual waves of data on a
panel of 10,000 spouses in 5000 households in Great Britain during 19911999, found a strong correlation between an
individuals current smoking participation and his or her partners smoking participation in the prior year. The authors
concluded, however, that the principal explanation for the observed inter-marriage correlation in smoking rates was
spouse self-selection or positive assortative matching in the marriage market.

Corresponding author.
E-mail addresses: jeffrey@mit.edu (J.E. Harris), bvalcarcel@dmc.ulpgc.es (B. Gonz alez L opez-Valc arcel).
0167-6296/$ see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhealeco.2007.07.005
250 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
In the present paper, we go one step further, explicitly emphasizing the importance of asymmetric peer effects in the
analysis of smoking behavior. In a pooled cross-sectional analysis of smoking participation among more than 122,000
respondents aged 1524 years who belonged to one of 90,000 households surveyed in the U.S. during 19921999, we
nd that pro-smoking effect of a siblings smoking was signicantly larger in absolute terms than the deterrent effect
of a siblings not smoking.
By contrast, the recent work of Powell et al. (2005) and Lundborg (2006), as well as earlier papers by Norton et
al. (1998) and Gaviria and Raphael (2001), did not explicitly consider asymmetric peer effects. Instead, they assumed
that an individuals smoking participation was a function of the overall smoking prevalence in his or her peer group.
To the extent that an individuals probability of smoking is a nonlinear function of the peer group prevalence (as,
for example, in a probit or logit specication), these studies implicitly incorporated a component of asymmetry. In
Clark and Etil e (2006), the relevant peer group was the individuals spouse, so that the peer group prevalence was no
more than the spouses smoking participation. Clark and Etil e (2006) did distinguish the effect of a wifes behavior
on the husbands smoking participation from the effect of a husbands behavior on the wifes smoking participation,
and they also considered whether peer inuences depended on the duration of the marriage. However, the possibility
that a spouses smoking has a quantitatively different inuence than a spouses not smoking was only implicit in the
nonlinear specications in their econometric models.
The consequences of asymmetric peer effects for public policy toward youth smoking are substantial. If the pro-
smoking social inuence of a smoker exceeds the deterrent inuence of her non-smoking counterpart, then policies to
segregate smokers will reduce overall smoking rates. Moreover, as Powell et al. (2005) and others have noted, such
public policies as anti-smoking informational campaigns, restrictions on youth access and cigarette excise tax increases
could have both a direct effect on the individual and an additional indirect effect through peer inuences. But in the
presence of asymmetric peer inuences, such anti-smoking policies might likewise have asymmetric peer effects. Thus,
a decrease in cigarette price may enhance smoking prevalence among young people more than an equivalent increase
in price reduces smoking. If so, then the cigarette price war in the early 1990s in the highly concentrated U.S. cigarette
market may have enhanced youth smoking, dollar for dollar, more than the price increases effectuated by excise tax
hikes or settlement payments later in the same decade (Gruber, 2001). What is more, if peer effects are asymmetric,
then an informational campaign with a constant intensity of anti-smoking messages will have a larger deterrent effect
on youth smoking than an on-and-off campaign with the same total number of messages.
In Section 2, we provide an overviewof our database. We briey note three well-known problems in the interpretation
of evidence on peer effects, namely, self-selection, unobserved common shocks, and the phenomenon of reection.
In Section 3, we present a simple model of asymmetric peer inuence. In a related Appendix, we show how our
simple model can be interpreted in the context of learning, where a young person learns that smoking is safe, socially
acceptable and non-addictive in short, that smoking is cool when she observes one of her peers smoking. In
this learning context, the asymmetry in social inuence arises because a peers smoking is a specic but not a very
sensitive test that smoking is cool. Such a learning interpretation draws upon the contagion of opinion literature
in nancial markets (e.g., Bikhchandani et al., 1998), where investors rely on the buying and selling patterns of fellow
traders in order to form expectations about future securities prices.
In Section 4, we consider the consequences of our model of asymmetric peer inuence for the joint determination
of smoking participation in a household with more than one young person. Our inquiry leads us to analyze a class of
multi-player non-cooperative binary games similar to the two-player game studied by Clark and Etil e (2006). As noted
by Clark and Etil e (2006) and others, the simultaneous-move versions of such games admit the possibility of multiple
Nash equilibria in pure strategies. To avoid this indeterminacy, we rely on a sequential-move solution, in which older
players move rst. Our sequential-move approach resembles the Stackelberg-type solutions explored in other contexts
by Kooreman (1994), Bjorn and Vuong (1997), Hiedemann (1998), and Chao (2002).
In Section 5, we describe the use of multidimensional probit models to estimate the parameters of the resulting
multi-player family smoking game. Our use of a specic random-effects error structure permits us to address the
problem of common shocks, identify the parameters and thus solve the problem of reection as well.
Section 6 shows our empirical results. At the sample means of our independent variables, each additional smoking
sibling raises the probability of smoking participation by an estimated 7.6%, while each non-smoking sibling lowers
the probability of smoking participation by an estimated 3.5%. We conrm that if smoking is the cool thing to do,
then there is a 37% probability that ones peer will smoke, but if smoking were uncool, there is a 74% probability
that ones peer will not smoke. In the nal Section 7, we discuss the limitations of our results.
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 251
2. Overview of the data and statement of the problem
Table 1 shows the prevalence of current cigarette smoking among persons aged 1524 years in relation to the smoking
status of other young people, also aged 1524, who reside in the same household. The cross-tabulation is based upon
122,010 individual responses to the Tobacco Use Supplements (TUS) that were appended to the U.S. Current Population
Surveys during three waves: September 1992May 1993; September 1995May 1996; and September 1998May 1999
(see U.S. Bureau of the Census, 1996; U.S. Bureau of the Census, 1998; and the Data Appendix in Harris and Gonz alez
L opez-Valc arcel, 2004).
Although the pattern is not completely consistent in the largest households, where the sample sizes are small, Table 1
shows that the inclusion of non-smoking housemates appears to decrease a young persons probability of smoking,
while the addition of more smokers appears to increase her probability of smoking. What is more, the quantitative effect
of including one more smoking housemate appears to exceed the effect of including one more non-smoking housemate.
For example, if a young person (call her Bea) started out with no young housemates, then adding two non-smokers
would reduce Beas log-odds of smoking by 0.99, while adding two smokers would increase her log-odds by 1.98.
Although the data in Table 1 suggest the presence of social inuences within the household, we need to be wary
of three well-known, critical problems of interpretation, namely, self-selection, common shocks, and reection. Under
self-selection, the apparent correlation of smoking behaviors within the household can arise simply because Bea chose
to cohabit with housemates who had compatible smoking habits. Clark and Etil e (2006) found such a phenomenon
to be a particularly important explanation for the observed long-term correlation of smoking habits among spouses.
Similarly, both Powell et al. (2005) and Lundborg (2006) explicitly acknowledged that when parents choose which
school district to live in, they may indirectly choose their childrens peers. In our TUS sample, however, only 4% of
young persons residing in the same household were unrelated. In all likelihood, most young people had little choice
but to live with their parents and other siblings. For convenience, in what follows, we shall use the terms sibling and
young housemate interchangeably.
Table 1
Prevalence of current cigarette smoking among young people, aged 1524, in relation to current smoking status of other young household members,
aged 1524
Number of siblings, aged 1524, who smoke Number of siblings, aged 1524, who do not smoke
0 1 2 3
0 19% 12% 8% 5%
62,817 37,561 7,646 1,724
62,817 18,781 2,549 431
1.45 1.99 2.44 2.94
1 50% 34% 28% 17%
8,859 1,746 366 72
4,430 582 92 14
0.00 0.66 0.94 1.59
2 63% 49% 54% 0%
808 201 39 16
269 50 8 3
0.53 0.04 0.16
3 73% 53% 100%
121 30 4
30 6 1
0.99 0.12
Source: Combined Tobacco Use Supplements of the CPS during September 1992May 1993, September 1995May 1996, and September 1998May
1999 (U.S. Bureau of the Census, 1996; U.S. Bureau of the Census, 1998). The four entries in each cell are, respectively: the percentage of young
people, aged 1524, who are current smokers; the total number of young people in the cell; the total number of households in the cell; and the log
odds of the prevalence of current smoking, that is, ln(p/(1 p)), where p is smoking prevalence. A current smoker is a respondent who answered
some days or everyday to the question, Do you smoke cigarettes some days, everyday, or not at all?. Among 126,352 respondents with known
smoking status, we excluded 3809 who resided in families where the smoking status of at least one other sibling was unknown. Not shown are the
data for another 533 respondents who resided in families with more than three non-smoking siblings or more than three smoking siblings.
252 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
In empirical studies of peer effects, investigators routinely attempt to control for observable factors that members of
the peer group may have in common. In the case of smoking within a household, such observable factors would include
family income and cigarette price (Harris and Chan, 1999). It is also possible, however, that unobserved common
inuences, or common shocks, within the same household underlie the patterns observed in Table 1. Thus, while the
TUS explicitly asked respondents whether smoking was permitted in the home, there may be an unstated family rule
that nobody should smoke in the presence of an inrm, elderly relative. Or perhaps underage siblings in the same
household obtain cigarettes from the same source in the schoolyard. What is critical about these common shocks is
that they are unobserved by the analyst, but they are common information among the economic agents under study.
Quite apart from biases due to self-selection and unobserved common shocks, there remains the conundrum of
two-way causation, what Manski (1993) has called the phenomenon of reection. Suppose that 18-year-old Bea has a
19-year-old brother Jeff who lives in the same household and does not smoke cigarettes. We observe that Bea likewise
does not smoke. Now consider an otherwise identical two-sibling household, endowed with the same income, facing
the same retail cigarette prices, caring for the same sick grandmother. In this clone household, 18-year-old Bea has a
19-year-old brother Jeff who does smoke cigarettes, and we observe that Bea likewise smokes. Do we conclude that
Jeffs smoking practices inuenced Beas? Or is it the other way around?
If we could perform such an experiment, then we could at least measure the two-way social interaction between
Beas and Jeffs smoking. Beyond that, we would have to impose strong restrictions to infer one-way causation. We
could posit, for example, that if Jeff began to smoke before Bea, then Jeff caused Bea to smoke, but not vice versa.
Such a restriction is dubious because Beas smoking, even if she began after Jeff started, may also be keeping Jeff from
quitting.
When we observe households with three or more young people, however, we can impose restrictions that follow
naturally from the organization of the data in Table 1. Suppose that we observed another household, with Bea and Jeff
as above, which differed fromthe rst clone in that there was a third sibling, 20-year-old Pepe. The combined inuence
of Beas two housemates on Beas smoking, we posit, depends only on the number of smoking siblings, and not their
identity. In that case, all other things equal, Beas probability of smoking when Jeff does smoke and Pepe does not
is the same as her probability of smoking when Jeff doesnt and Pepe does. When we impose such a restriction, and
when we use the data on both the two-sibling and three-sibling households together, then as we will see below, we can
identify the one-way causal effect of one young persons smoking on all the remaining sibs.
3. Basic model
A household consists of n 1 young people, whom we call siblings. Each sibling i has a binary choice to smoke
or not to smoke. Her smoking behavior is represented by the binary variable y
i
, where y
i
=1 means that she smokes,
while y
i
=0 means that she does not. For each i =1, . . ., n, we specify the net utility of smoking (that is, the difference
between the utility of smoking and the utility of not smoking) as
U
i
= u
i
+
n

j=i
y
j
+
n

j=i
(1 y
j
) (1)
where u
i
can be regarded as the baseline net utility of smoking absent peer inuence, and where 0 represents the
pro-smoking inuence of those siblings who do smoke, while 0 represents the anti-smoking inuence of those
siblings who do not. Asymmetric peer inuence means that +=0. Specically, if the pro-smoking inuence of a
smoking sibling exceeds the anti-smoking inuence of a non-smoking sibling, then +>0.
If we condition on the smoking behavior of all other siblings j =i, then the decision rule for sibling i is
y
i
= 1 if u
i
+
n

j=i
y
j
+
n

j=i
(1 y
j
) 0; and y
i
= 0 otherwise (2)
In Appendix A, we present a more detailed model that specically interprets the quantities {u
i
, , } in terms of
learning. In this learning model, each sibling is uncertain whether smoking cigarettes is, on the one hand, a safe, non-
addicting and socially acceptable thing to do, or on the other hand, a dangerous, addicting and socially unacceptable
practice. Such a young person draws Bayesian inferences from the smoking practices of other young household
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 253
members to resolve this uncertainty. In particular, let and , respectively, denote the sensitivity and specicity of
peer smoking. From Beas point of view, the quantity is the probability that Jeff would smoke if it were cool (that
is, safe, non-addicting and socially acceptable), while the quantity is the probably that Jeff would not smoke if it
were uncool (that is, unsafe, addicting and socially unacceptable). In Appendix A, we specically derive Eq. (2),
interpreting =ln(/(1 )) and =ln((1 )/) as log-likelihood ratios. We showthat if peer smoking is informative,
that is, +>1, then the strict inequalities >0 and <0 hold and, moreover, the condition for asymmetric inuence
+>0 implies that >, that is, the specicity of Jeffs smoking exceeds its sensitivity.
Letting = , we rewrite (1) as
U
i
= u
i
+(n 1) +
n

j=i
y
j
(1)
for all i =1, . . ., n. Since 0 and 0, we therefore have 0. The coefcient measures the net effect of a young
persons switching from a non-smoker to a smoker in a household with a xed number n of young people.
For purposes of empirical analysis, we specify the linear link function:
u
i
= X
i
+Z +
i
(3)
for all i =1, . . ., n, where x
i
is a vector of observable personal characteristics (such as age and gender) that vary from
one young person to the next within the household, Z is a vector of observable characteristics (such as household
income and cigarette price) that vary between households, (, ) is a parameter vector, and
i
is a disturbance term
that captures unobservable factors. For now, we assume that these unobserved factors are random variables (
1
, . . .,

n
) with a common mean of zero and a joint cumulative distribution function F, and that they are independent of the
exogenous variables (X
1
, . . ., X
n
, Z). We impose parametric assumptions on the joint distribution of (
1
, . . .,
n
) in
Section 5 below. We further assume that each young person in the household can observe all of the random factors (
1
,
. . .,
n
), while we, as economic analysts, cannot. This assumption distinguishes our analysis from other studies where
each
i
is regarded as private information known only by player i (see, for example, Bajari and Krainer, 2004).
We thus have the specication:
U
i
= X
i
+Z +(n 1) +
n

j=i
y
j
+
i
(5)
for all i =1, . . ., n. It will be helpful to abbreviate W
i
=X
i
+Z +(n 1) for all i =1, . . ., n, so that the net utility in
(5) becomes
U
i
= W
i
+
n

j=i
y
j
+
i
(6)
In Eq. (1), the termcapturing pro-smoking peer inuence (that is,

n
j=i
y
j
) depends only on the number of smoking
siblings, while the term capturing anti-smoking peer inuence (that is,

n
j=i
(1 y
j
)) depends only on the number of
non-smoking siblings. In the learning model described in Appendix A, these restrictions imply that each young person
views the smoking practices of each of her siblings as independent information about the coolness of smoking. In
what follows, we assume that the parameters and in Eq. (1), and therefore the parameter in Eqs. (3), (5) and (6)
are constant and independent of the total number of siblings n in the family. However, we shall relax this assumption
in our econometric analyses later on.
4. The family smoking game
When n >1, we can think of the system(6) as delineating the payoffs of each player in a binary, static, non-cooperative
game of complete information. In this family smoking game, each young persons decision to smoke (that is, y
j
=1)
enhances the propensity of the others to follow suit (that is, y
i
=1). Our family smoking game differs from the market
entry games (Bresnahan and Reiss, 1990; Berry, 1992) where the entry of a rival rm j into the market (that is, y
j
=1)
makes it less protable for rm i to be in the market (that is, y
i
=1). It also differs from the binary game studied in
detail by Tamer (2003), in which y
j
=1 makes it more likely that y
i
=1, but y
i
=1 makes it less likely that y
j
=1. In its
254 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
two-player version, our family smoking game is analogous to the spousal bargaining (or battle of the sexes) game
studied by Clark and Etil e (2006), in which each spouses smoking imposed a negative health cost on the other spouse
in the partnership.
A critical feature of our family smoking game is the order in which the players move. Solving (2) as a system of
simultaneous equations would be equivalent to specifying a simultaneous-move game. As we shall see below, such
an approach can lead to indeterminacy in the model, with multiple Nash equilibria. However, given that the young
people in the household are of different ages and that cigarette smoking is addictive, we posit that the older players
can commit to move before the younger ones. This sequential-move specication solves the indeterminacy problem
that arises in a simultaneous-move game.
4.1. Sequential-move games
We employ the two-sibling case to elucidate the differences between the solution concepts to our family smoking
game. Appendix B shows the corresponding analysis in the three-player case. For now, our household has two young
people, where Bea is assumed to be the older sibling #1, while her brother Jeff is the younger sibling #2.
We rst consider the two-player simultaneous-move game. The Nash equilibria in pure strategies represent the
solutions to the system of simultaneous equations:
y
1
= 1 if W
1
+y
2
+
1
0; y
1
= 0 otherwise
y
2
= 1 if W
2
+y
1
+
2
0; y
2
= 0 otherwise
(7)
Fig. 1 shows the possible Nash equilibria in pure strategies for various values of the error terms in the (
1
,
2
) plane.
When
1
W
1
, smoking cigarettes is a dominant strategy for Bea, that is, she will smoke whether or not Jeff does. Con-
versely, when
1
<W
1
, abstaining from smoking is a dominant strategy for her. But when W
1
>
1
W
1
,
Bea will choose to smoke only if Jeff does, too. In the central rectangular region R of the (
1
,
2
) plane where
W
1
>
1
W
1
and W
2
>
2
W
2
, there will be two Nash equilibria: (y
1
, y
2
) =(0, 0); and (y
1
, y
2
) =(1,
1). Our Fig. 1 is analogous to Fig. 1 in Clark and Etil e (2006), but differs from Fig. 3 in Bresnahan and Reiss (1990)
and Fig. 1 in Tamer (2003).
We now consider a sequential-move game in which both siblings are present in the household, but Bea (#1), as
the older sibling, can commit to move before Jeff (#2). In that case, the indeterminacy encountered in the foregoing
simultaneous-move game is eliminated. Thus, inthe regionwhere
2
W
2
, Bea infers that smokingwill be a dominant
strategy for Jeff. That is, y
2
=1 independent of Beas initial move. In that case, the only rationalizable strategy is for
Bea to smoke when
1
W
1
and not to smoke when
1
<W
1
. These outcomes are concordant with the
equilibria shown in the upper portion of Fig. 1 above the level
2
=W
2
. Proceeding with the same logic, we nd
that the only difference between the sequential game and the simultaneous-move game is that in the center region R
Fig. 1. Solutions in the (
1
,
2
) plane in the two-sibling smoking game. In the sequential-move game, the entire gray-shaded area, including the
rectangular region R, yields the solution (1, 1). In the simultaneous-move game, the rectangle R admits two Nash equilibria in pure strategies: (0, 0)
and (1,1).
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 255
Fig. 2. Equilibria in the (
1
,
2
) plane in the restricted sequential-move two-sibling smoking game when Bea (#1) moves before the entry of Jeff
(#2).
of Fig. 1 where W
1
>
1
W
1
and W
2
>
2
W
2
, Bea will elect to smoke and we have the determinate
outcome (y
1
, y
2
) =(1, 1), rather than multiple equilibria.
We also consider a restricted sequential-move game, in which sibling #1 must permanently decide whether or not
to smoke before sibling #2 is old enough to become a regular cigarette smoker. For example, if Bea were 5 years older
than Jeff, then she might decide to smoke at age 14 and Jeff, still only 9 years old, was not yet a player in the game.
In essence, Bea moves rst under the assumption that Jeffs only move is y
2
=0. After Bea commits to her rst move,
and once Jeff is old enough, he responds to Beas initial move. The outcomes for this restricted sequential-move game
are shown in Fig. 2.
In our empirical analysis below, we focus principally on the sequential-move game. We test the sequential-move
hypothesis against the alternative simultaneous-move and restricted sequential-move games.
4.2. Other solution concepts
Our sequential-move approaches resemble the Stackelberg-type solutions explored in other contexts by Kooreman
(1994), Bjorn and Vuong (1997), Hiedemann (1998), and Chao (2002). In the present context, the sequential-move
equilibriumturns out to be equivalent to the cooperative solution, which has been studied by Clark and Etil e (2006). Our
sequential-move solution concepts have a distinct advantage over the Nash equilibrium in a simultaneous-move game,
which suffers from indeterminacy in the sense that the binary decision variables {y
i
} are not necessarily uniquely
determined by the values of {W
i
,
i
, }. Thus, in the two-sibling simultaneous-move family smoking game, in the
central region R in Fig. 1, there can either be zero or two smokers, and in the three-sibling family smoking game, as
shown in Appendix B, there are regions of (
1
,
2
,
3
)-space where there can be zero or three smokers. To be sure,
in some simultaneous-move market entry games, the quantity

n
i=1
y
i
, which corresponds to the number of entrants,
is uniquely determined even when there are multiple equilibria (Bresnahan and Reiss, 1990; Berry, 1992), but this is
clearly not the case in the simultaneous-move family smoking game.
To resolve the indeterminacy, one could rely upon mixed strategies. For example, within the region R in Fig. 1, the
mixed strategy equilibrium entails Beas smoking with probability (W
2
+
2
)/ and Jeffs smoking with probability
(W
1
+
1
)/. When n >2 and no player has a dominant strategy, the mixed strategy equilibrium entails sibling i
smoking with probability (1/n)

(W
i
+
i
)

n
j=i
(W
j
+
j
)

. However, we are not sure what interpretation to attach


to such mixed strategies in the smoking context.
Alternatively, one could attempt to address the problem of multiple equilibria in the simultaneous-move game by
the ad hoc inclusion of incidental parameters. In those regions of the (
1
,
2
, . . .,
n
) space where multiple pure-strategy
Nash equilibria prevail, we could assume that the observed equilibrium is the outcome of a random trial governed by
mixing probabilities. These mixing probabilities, which become incidental parameters, are assumed to be independent
of the exogenous variables (X, Z). The main drawback of such an approach to equilibrium selection is that the number
of incidental parameters increases nonlinearly with family size. For example, in the three-sibling family partially
256 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
described in Appendix B, we would in principle need one distinct parameter for each of the four regions in which there
are two equilibria Harris and Gonz alez L opez-Valc arcel (2004).
5. Econometric methods
5.1. Multivariate normal disturbances
We specify a multivariate normal error structure that specically incorporates the possibility of unobserved common
shocks. We denote the smoking behavior of young person i in household h, which contains n
h
young people, by the
binary variable y
ih
. Expanding on the specication in Eq. (5), we write the net utility of smoking by sibling i in
household h as
U
ih
= X
ih
+Z
h
+(n
h
1) +Y
ih
+
h
+
ih
(8)
where Y
ih
=

n
h
j=i
y
jh
denotes the number of siblings of young person i who currently smoke; each common shock
component
h
is independently identically normally distributed with mean 0 and variance [0, 1]; each
ih
is inde-
pendently identically normally distributed with mean 0 and variance 1 ; and (, , , , ) are unknown parameters.
Under this random-effects error structure, each combined disturbance term
ih
=
h
+
ih
is normally distributed with
zero mean and unit variance. Moreover, E[
ih

jh
] = for i =j, that is, the disturbances {
ih
} may be correlated within
the household, even if they are independent across households.
The assumption of multivariate normality permits us to estimate the parameters of our model by maximumlikelihood
methods. The two-player family smoking game illustrates the point. In that case, the disturbances (
1h
,
2h
) are jointly
normally distributed with zero means, unit variances, and correlation coefcient . For the solutions to the sequential-
move game in Fig. 1, the log-likelihood of observing a household h in which sibling #1 does not smoke while sibling
#2 smokes (that is, (y
1h
, y
2h
) =(0, 1)) would be
(0, 1) = ln[(W
1h
, W
2h
, )] (9)
where (u, , ) is the bivariate cumulative normal cumulative distribution function, where W
ih
=X
ih
+Z
h
for i =1,
2, and where the parameter is suppressed as it cannot be identied solely from data on two-sibling households. The
log-likelihood that only the rst sibling smokes (that is, (y
1h
, y
2h
) =(1, 0)) would be
(1, 0) = ln[(W
1h
, W
2h
, )] (10)
and the log-likelihood that both siblings smoke (that is, (y
1h
, y
2h
) =(1, 1)) would be
(1, 1) = ln (W
1h
+, W
2h
+, ) (11)
Finally, the log-likelihood of observing two non-smoking siblings in the household (that is, (y
1h
, y
2h
) =(0, 0)) would
be
(0, 0) = ln[(W
1h
, W
2h
, ) +(W
1h
, W
2h
, ) (W
1h
, W
2h
, )] (12)
Our assumption that the disturbances are jointly normally distributed permits us to take advantage of the global
concavity of the multivariate probit log-likelihood function to establish global identication of all the parameters.
Again, specic reference to the two-person case helps to clarify the identication strategy. To that end, we note that
observations in which (y
1h
, y
2h
) =(0,1), (y
1h
, y
2h
) =(1, 0), or (y
1h
, y
2h
) =(1, 1) for different households h, which
correspond to Eqs. (9)(11), respectively, are all bivariate probit log-likelihoods. But for the presence of the peer effect
parameter , we could simply invoke the well-known properties of the bivariate probit model to identify {, , }. Now
consider only those observations in which (y
1h
, y
2h
) =(0, 1). So long as we can observe data {X
1h
, X
2h
, Z
h
} where X
1h
varies across households h while {X
2h
, Z
h
} remain constant, we can identify the peer effect parameter as well. Since
we have imposed the restriction that both X
1h
and X
1h
have the same coefcient , varying X
1h
across households in
which sibling #1 does not smoke varies the probability (0, 1) in Eq. (9) only with respect to . The same identifying
strategy applies to households with discordant smoking in which (y
1h
, y
2h
) =(1, 0). We can identify the parameter so
long as we can observe data {X
1h
, X
2h
, Z
h
} where X
2h
varies across households h while {X
1h
, Z
h
} remain constant.
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 257
Although we do not present a formal argument for identication of the parameters in households with more than
two siblings, the basic argument is a generalization of the foregoing. Moreover, it should be clear from Eq. (8) that if
we study households of different sizes n, then we can identify the parameter as well.
5.2. Estimation and testing strategy
Our estimation and testing strategy proceeded fromthe particular to the general. We rst estimated the model for the
subset of two-sibling families, and tested the sequential game solution against the restricted sequential-game solution.
As it turned out, the restricted sequential-game solution concept, in which the older sibling initially assumes that
the younger sibling is a non-smoker, was strongly rejected. With xed n
h
=2, we obtained a precise estimate of the
structural parameter , but not the parameter , which is indistinguishable from the constant term implicit in the vector
(X
ih
, Z
h
).
Second, we estimated the sequential-move model for the subsets of families of xed size. In the three-sibling,
four-sibling and ve-sibling analyses, the likelihood functions are analogous to that shown for the two-sibling family
in Eq. (9)(12), and the relevant multivariate probit integrals were computed by standard quadrature methods. Again,
with xed n
h
, these analyses provided us with a precise estimate of the parameter , but not the parameter .
Third, in pooled analyses of all households with one to four siblings, we were able to identify not only the parameter
, but also the parameter . From this information, we could also recover the parameter = +. Since >0 and <0,
the null hypothesis of symmetric social inuence would mean that +=0 or, equivalently, +2=0. To test this
null hypothesis against the two-sided alternative, we computed the ratio of the likelihood for the constrained version
of model, where +2=0, divided by that for the unconstrained version.
Fourth, in our pooled sample of households with four or fewer siblings, we formally tested the hypotheses that that
the peer inuence parameters and depended on family size. In particular, we assumed the linear link functions:
=
0
+
1
(n
h
1) and =
0
+
1
(n
h
1), so that the net utility (8) becomes
U
ih
= X
ih
+Z
h
+
0
(n
h
1) +
1
(n
h
1)
2
+
0
Y
ih
+
1
(n
h
1)Y
ih
+
h
+
ih
(13)
From the estimated parameters in (13), we could then recover the dependence of the parameter on family size
as well, that is, =(
0
+
0
) +(
1
+
1
)(n
h
1). To test the null hypothesis of symmetric social inuence against the
two-sided alternative, we computed the ratio of the likelihood for the constrained model, where both
0
+2
0
=0 and

1
+2
1
=0, divided by that for the unconstrained version.
Fifth, in our pooled sample of households with four or fewer siblings, we estimated the reduced-form model in
which we assumed that both =0 and =0. This case is equivalent to a random-effects probit model of the form:
y
ih
= 1 if X
ih
+Z
h
+
h
+
ih
0 and y
ih
= 0 otherwise (14)
Comparison of the estimated parameters (, ) from the reduced-form model in Eq. (14) with those from the
structural model (8) permitted us to assess the social multiplier associated with within-family peer inuence. In this
context, we focused specically on the contrast between the structural effect and the reduced-form effect of an increase
in the real price of cigarettes.
Finally, we performed a number of sensitivity analyses. We analyzed a pooled sample of families with two to four
siblings, excluding households with only one young person. We studied an alternative model to (8) where the peer
inuence parameter depended on a young persons age. We also estimated the model (8) under the dubious assumption
that the variable Y
ih
, the number of siblings of young person i who currently smoke, was exogenous.
6. Empirical results
6.1. Parameter estimates for subsamples of households of xed size
In our analysis of two-sibling households, we strongly rejected the restricted sequential-move solution (Fig. 2) in
favor of our basic sequential-move solution concept (Fig. 1). In 12,632 two-sibling households, the sequential-move
solution (Fig. 1) gave an estimate

= 0.521 with a 95%condence interval of [0.488, 0.554]. The restricted sequential-
move solution (Fig. 2) gave

= 0.362 with a 95%condence interval of [0.248, 0.476]. However, restricting the model
so that which the older sibling must decide whether to smoke while the younger sibling remains a non-smoker reduced
258 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
Table 2
Estimates of the peer inuence parameters and the coefcient of cigarette price in a pooled sample of 49,898 households with 14 siblings
Parameter Description Social interaction No social interaction
Eq. (8) Eq. (13) Eq. (14)
Coefcient of price 0.016 [0.023, 0.010] 0.016 [0.023, 0.010] 0.026 [0.036, 0.016]
Coefcient of Y
h
(# smoking
sibs in household)
0.479 [0.455, 0.503]
Coefcient of n
h
1 (# sibs in
household)
0.150 [0.170, 0.130]

0
Intercept of =
0
+
1
(n
h
1) 0.576 [0.517, 0.635]

1
Slope of =
0
+
1
(n
h
1) 0.063 [0.098, 0.027]

0
Intercept of =
0
+
1
(n
h
1) 0.114 [0.161, 0.067]

1
Slope of =
0
+
1
(n
h
1) 0.013 [0.037, 0.011]
+ 0.329 [0.308, 0.350]

0

0
+
0
0.462 [0.410, 0.514]

1

1
+
1
0.076 [0.104, 0.048]
LLR
a
test for
symmetric
inuence
LLR=52.75
2
=105.5
(1 d.f.) P<10
6
LLR=61.92,
2
=123.84
(2 d.f.) P<10
6
a
LLR=log-likelihood ratio.
the log-likelihood by 44.3. In our subsequent empirical analysis, we used only our basic sequential-move solution
concept.
Separate analyses of subsets of xed-sized households, based on the sequential-move solution concept (Fig. 1)
yielded positive and statistically signicant estimates of the peer inuence parameter . In 1786 three-sibling households
alone, we observed

= 0.445 with a 95% condence interval of [0.277, 0.613]. For 308 four-sibling families alone,
we obtained

= 0.396 with a 95% condence interval of [0.318, 0.473], while in 50 ve-sibling households alone,
we observed

= 0.241 with a 95% condence interval of [0.080, 0.402].
6.2. Pooled analysis of households with one to four young people
The column identied as Eq. (8) in Table 2 shows our estimates of the parameters and in Eq. (8), as well as the
estimated coefcient for cigarette price, in a pooled sample of 49,898 households with one to four siblings. Given
maximum likelihood estimates of

= 0.479 and = 0.150, we computed =

+ = 0.329. As Table 2 further
shows, the null hypothesis of symmetric peer inuence, namely, +=0, was strongly rejected (P<10
6
). At the
sample means of the right-hand side variables, each additional smoking sibling raised the probability of smoking by an
estimated 7.6%, while each non-smoking sibling lowered the probability by an estimated 3.5%. Thus, the pro-smoking
inuence of a sibling who smokes is more than twice the deterrent effect of a non-smoking sibling.
The results for Eq. (8) in Table 2 permitted us to recover the parameters in our learning model of Appendix A.
In particular, the sensitivity of a peers smoking as a test of its coolness was an estimated 0.366 (approximate
95% condence interval, 0.3350.397), while the specicity was an estimated 0.737 (approximate 95% condence
interval, 0.7010.772). Put differently, if smoking is uncool, then Bea believes that there is about a 74% probability
that her brother Jeff will not smoke. But if smoking is cool, then there is only about a 37% probability that Jeff will
smoke.
Inour analysis of subsets of households of different size, we foundthat the estimatedvalues of were inverselyrelated
to the number of young people. We therefore investigated a model, specied in Eq. (13), wherein both the peer inuence
parameters and depended on family size through the linear link functions =
0
+
1
(n
h
1) and =
0
+
1
(n
h
1).
The parameter estimates for the resulting model, based upon the pooled sample of 49,898 households with one to four
siblings are also shown in Table 2. We found that while the parameter declined with the number of siblings, the
parameter did not. Thus, as the number of siblings in the household increased, the pro-smoking inuence of any one
smoking sibling was diluted, while the anti-smoking inuence of a non-smoking sibling was not clearly dependent on
family size. In particular, in a two-sibling family (that is, n
h
=2), the estimated values of and were 0.513 and 0.127,
respectively, and the corresponding values of the sensitivity and specicity were 0.297 and 0.798, respectively.
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 259
In a four-sibling family, the estimated values of and were 0.387 and 0.153, respectively, and the corresponding
values of the sensitivity and specicity were 0.442 and 0.650, respectively. As in Eq. (8), our likelihood ratio test
strongly rejected the null hypothesis of symmetric peer inuence in the pooled sample of families all sizes. Moreover,
the values of in models where the parameters were estimated separately for subsets of households of different sizes
were indistinguishable from the values estimated from Eq. (13) in the pooled sample of households of all sizes (results
not shown).
By way of sensitivity analysis, we found that the inuence parameter did not vary signicantly with each siblings
age (results not shown). When we used standard random-effects probit methods to estimate Eq. (14) under the dubious
assumption that the number of smoking siblings Y
ih
was exogenous, we obtained estimates of 0.90 and 0.22,
values that (in absolute terms) signicantly overstated those shown in Table 2 (detailed results not shown).
The right-most column of Table 2 shows the estimate of the coefcient for cigarette price in the reduced-formmodel
(14). The estimated coefcient of price in the reduced-form model was 1.6 times the value derived from the structural
models of (8). At the sample means of the right-hand side variables, each $1-increase in the real price of a carton of
cigarettes (or 10-cent increase in the price per pack) reduced the probability of smoking in the reduced-formmodel (14)
by 0.0060, while it reduced the probability of smoking in the structural model (8) by only 0.0038. Thus, we found that
the marginal deterrent effect of a price increase on the probability of smoking, as derived fromthe reduced-formmodel,
was likewise about 60% greater than the marginal effect of price derived from the structural model. This comparison
gives us an estimate of the social multiplier effect of peer inuence within the household.
The full set of parameter estimates for Eqs. (8), (13) and (14) in Table 2 are shown in the Results Appendix of
Harris and Gonz alez L opez-Valc arcel (2004). In particular, we found that the effect of a household adults smoking is
only about two-thirds that of a younger peers smoking. In separate sensitivity analyses, we rejected the possibility that
older siblings had more inuence than younger ones (results not shown). We also re-estimated our pooled analyses
with a subset of 14,781 families with two, three, or four siblings, but the results were not markedly different from those
reported here (results also not shown).
7. Discussion and conclusions
In our analyses of smoking practices among young people aged 1524 in a large nationwide sample of U.S.
households polled during the 1990s, we found that the pro-smoking inuence of a sibling who smokes is more than
twice the deterrent effect of a non-smoking sibling. While this asymmetry was somewhat reduced in larger households,
the pro-smoking inuence of a smoking peer continued to dominant the deterrent effect of a non-smoker peer even in
households with ve young residents.
Althoughit is not essential toour analysis, we caninterpret our results interms of a model of informationexternalities,
in which each young person employs the information contained in each of her peers smoking practices to update her
prior probability that smoking is safe, non-addictive, and the socially acceptable, in short, the cool thing to do
(Appendix A). We estimated that, on average, the sensitivity of a peers smoking was 37%, while the specicity was
74%. Put differently, young Bea believes that if smoking were in fact uncool, then there would be only about a
one-in-four chance that her brother Jeff would smoke anyway. But if smoking were in fact cool, then she believes
that there would be a nearly two-thirds chance that Jeff could still be a non-smoker.
The partial dilution of a smokers inuence in large families can likewise be expressed in terms of sensitivity and
specicity. If Jeff were Beas only sibling, then we estimated that Bea would attach about a 30% sensitivity and an
80% specicity to his smoking. By contrast, if Jeff were just one of Beas three siblings, then the sensitivity of Jeffs
smoking would rise to about a 44%, while the specicity would fall to about 65%. Our ndings concerning diluted
inuence suggest that young people in fact understand the interdependence of their peers smoking practices. Thus, if
Pepe already smokes, Jeffs smoking is less inuential for Bea because she recognizes that Jeff is smoking, at least in
part, because Pepe is.
Comparing the results from a reduced-form model to our structural estimates, we estimated that the aggregate
deterrent effect of an increase in cigarette price on youth current smoking prevalence was approximately 60% greater
than the sumof the deterrent effects on each individual separately. Equivalently, we estimated a social multiplier of 1.60.
Powell et al. (2005) estimated the overall participation elasticity in a sample of high school students to be 0.50, but only
0.32 when they controlled for peer effects. This implies a social multiplier of approximately (0.50)/(0.32) =1.56,
which is consistent with our results.
260 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
Although the static equilibriummodel employed in this paper (the family smoking game) allowed for asymmetric
peer inuence, it did not by itself generate asymmetric social multipliers. While we might consider more complex
static models, or even dynamic models of youth smoking, we leave the modeling and estimation of such asymmetric
social multipliers to later work. Nonetheless, asymmetric social multipliers may be an important source of asymmetric
responses of demandtoincreases andreductions incigarette taxes andother measures intendedtoreduce youthsmoking.
In order to avoid problems of multiple equilibria and model indeterminacy, we relied on a sequential-move solution
to our multi-player, non-cooperative family smoking game, in which older players can commit to move before younger
ones. Our sequential-move solution accommodates the facts that young people generally do not become regular smokers
until early adolescence and that regular cigarette smoking, once established, is addictive. We rejected a more restricted
version of the sequential-move model, in which each player must decide whether or not to smoke before her younger
siblings are old enough to play the family smoking game. The sequential-move solution concept permits the rst-
moving older player to take account her inuence on her young siblings, as well as their reciprocal inuence upon
her. However, above and beyond their rst-mover advantage, we did not nd older siblings to have a larger inuence
parameter than their younger counterparts. Our results indicate that the peer inuence of younger smokers can affect
the maintenance of smoking among older ones. This nding is consistent with evidence that addiction to cigarettes
develops gradually along a continuum during youth (Harris and Chan, 1999).
Our analysis has a number of limitations. First, we may have relied too heavily upon the assumption of a multivariate
normal error structure (Eq. (8)) to distinguish peer effects from household-specic common shocks. While we have
not addressed the question of identication at the most general level, nonetheless we suspect that identication of
the peer-effect parameter does not necessarily require such a restricted error structure. In Eq. (1), we dened two
peer-effect parameters: , the pro-smoking effect of other siblings who smoke; and , the anti-smoking effect of other
siblings who do not smoke. In Eq. (3), we then dened = . Given this transformation of the parameters, we
concluded that the identication of required only that we have data on households of variable size n. In order to
identify , we relied on our assumption that the person-specic characteristics (that is, X
1h
and X
2h
in the two-sibling
case) had the same structural coefcient . In households with discordant smoking (for example, (y
1h
, y
2h
) =(0, 1) in the
two-sibling case), we could identify the peer effect parameter so long as we observed variations in the person-specic
characteristics of the non-smoking siblings (that is, in the two-sibling case, so long as we had data {X
1h
, X
2h
, Z
h
}
where X
1h
varies across households h while {X
2h
, Z
h
} remain constant).
Second, the peer group in our study was restricted to only young people within the household. While we observed
the smoking practices of older adults in the household, we treated such data as household-specic exogenous variables
(that is, Z
h
in Eq. (8)). When it comes to smoking decisions, we assumed that adults inuenced young people, and that
young people inuenced each other, but that young people did not inuence adults. Given that smoking initiation is
rare after age 25 years, and that the annual probability of successful quitting in an established adult smoker is on the
order of 12% (Harris, 1980), we do not regard our treatment of adults as problematic. Unfortunately, the Tobacco Use
Supplements to the U.S. Current Population Survey did not provide information on the smoking practices of friends
and classmates outside the household. If Bea and Jeff had a close friend in common named John, who did not reside
within the household, then Johns smoking practices would constitute an unobserved, household-specic common
shock. However, we are not sure how the omission of data on other peers smoking practices biases our nding that,
within the household, peer effects are asymmetric.
While we have conned our study to youth smoking, concepts of asymmetric social interaction have wider appli-
cability. Glaeser et al. (1996), for example, observed that rates of larceny and auto theft in a neighborhood inuence
the propensity of young residents to commit such crimes. In this context, asymmetric social interaction would mean
that the deterrent effect of a decrease in the neighborhood crime rate could be absolutely larger or smaller than the
crime-enhancing effect of an increase in the crime rate. Clark and Oswald (1996) found that workers job satisfaction
was signicantly negatively correlated with their co-workers income. In this context, asymmetric social interaction
would mean that the disutility of a high-income co-worker could be greater or smaller than the utility of a low-income
colleague. In the context of the contagion of opinion in nancial markets, where investors follow the behavior of
other fellowtraders to formexpectations about future prices (Bikhchandani et al., 1998; Lux, 1998; Bajari and Krainer,
2004), asymmetric social interaction would mean that investors might make different inferences when they see their
peers buy, rather than sell, securities.
In the educational and labor contexts, if the positive effect of a high-performing peer exceeds the negative effect of
a low performer, then mixing good and bad workers or students might enhance overall group performance (Sacerdote,
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 261
2001; Hoxby, 2002; Winston and Zimmerman, 2003). Recent study of an educational policy experiment in South
Korea in the 1970s revealed that mixing students by ability, rather than tracking theminto homogeneous ability groups,
resulted in overall higher adulthood earnings. That is, the positive effect of mixing on low-ability students exceeded
the negative effect on high-ability students (Kang et al., 2007).
Asymmetric peer effects may also afford a natural way to predict group polarization without ad hoc nonlinearity
assumptions. Thus, in an experimental study of jury deliberation on punitive damages, Schkade et al. (2000) found that
post-deliberation damage award was systematically higher than the average pre-deliberation judgment of the members.
This phenomenon, which the authors labeled a severity shift, is consistent with asymmetric peer effects.
Acknowledgements
This paper is an abridged version of Harris and Gonz alez L opez-Valc arcel (2004). We acknowledge the helpful
comments of Patrick Bajari, Victor Chernozhukov, and Cass Sunstein on earlier drafts. Prof. Gonz alez L opez-Valc arcel
was supported in part by grant PR2002-0134 from the Ministerio de Educaci on Cultura y Deporte, Espa na. This work
is the authors sole responsibility, and does not necessarily represent the opinions of the Massachusetts Institute of
Technology, the University of Las Palmas de Gran Canaria, or any other organization.
Appendix A. A learning model
We think of each young person as uncertain whether smoking cigarettes is, on the one hand, a safe, non-addicting
and socially acceptable thing to do, or on the other hand, a dangerous, addicting and socially unacceptable practice.
Such a young person draws Bayesian inferences from her peers smoking, especially the smoking practices of other
young household members, to resolve this uncertainty.
A.1. Expected utility
Consider the decision of one such young person, whomwe will call Bea. Let denote Beas point prior subjective
probability that smoking is cool (that is, safe, non-addicting and socially acceptable) rather than uncool (that is, a
dangerous, addicting and socially unacceptable). We refrain from any rational-expectations assumptions that relate
to the actual prevalence of smoking in Beas larger peer group.
Let V
1
denote Beas utility of smoking when it turns to be cool, while V
2
is the utility of smoking if it turns out
to be uncool. Analogously, V
3
is her utility of refraining from smoking when it turns to be cool, while V
4
is her
utility of not smoking when it turns out to be uncool. Since our focus is on social interactions, we sidestep questions
as to whether the quantities V
1
, . . ., V
4
are short- or long-run lifetime utilities.
Bea smokes so long as the expected utility of smoking exceeds the expected utility of not smoking, that is
V
1
+(1 )V
2
V
3
+(1 )V
4
(15)
Let B=V
1
V
3
denote the benet fromsmoking when it is cool to smoke, while C=V
4
V
2
is the cost of smoking
when it is uncool to do so. We assume that B, C>0. We further denote Beas smoking behavior by the binary variable
y
1
, where y
1
=1 if Bea smokes and y
1
=0 if not. Then, equivalently, Bea smokes so long as her subjective probability
is no less than a threshold value = C/(B +C), that is
y
1
= 1 if and y
1
= 0 otherwise (16)
Denoting =ln(/(1 )) and = ln( /(1 )) = ln C ln B as the log-odds transformations of and ,
respectively, we can rewrite the condition (17) as
y
1
= 1 if and y
1
= 0 otherwise (17)
A.2. Sensitivity and specicity of peer smoking
We assume that Bea updates her point prior distribution on when she learns whether or not a member of her
peer group (whom we call Jeff) smokes cigarettes. Since our data are derived from young people who responded
262 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
to a household-based survey, and since nearly all such household members were brothers or sisters, we will refer
to the members of Beas peer group as her siblings. Let the sensitivity denote the probability that Jeff would
smoke if it were cool to do so, and let specicity denote the probability that Jeff would refrain from smoking if
it were uncool. When the sensitivity is high, Bea thinks, My brother Jeff would denitely smoke if it were the
cool thing to do. When the specicity is high, the Bea thinks, My brother Jeff wouldnt be caught dead with a
cigarette if it werent cool. When +>1, Jeffs smoking is informative about its coolness. By contrast, when
+=1, Jeffs smoking is entirely uninformative, and knowledge of Jeffs smoking would not lead Bea to update her
prior .
If Jeff in fact smokes, then by Bayes rule, the log-odds of Beas posterior probability that smoking is cool becomes
+ln ln(1 ) (18)
If Jeff does not smoke, then the log-odds of Beas posterior probability that smoking is cool becomes
+ln (1 ) ln (19)
We similarly denote Jeffs smoking behavior by the binary variable y
2
, where y
2
=1 if Jeff smokes and y
2
=0 if not.
Then, in light of the information about Jeffs smoking, Beas decision to smoke can be described by the rule:
y
1
= 1 if +ln

y
2
+ln

(1 y
2
) 0 and y
1
= 0 otherwise (20)
When Jeffs smoking is informative, that is, +>1, the coefcient ln(/(1 )) will be strictly positive, which
means that his smoking (y
2
=1) will reinforce Beas decision to smoke. Similarly, when Jeffs smoking is informative,
the coefcient ln((1 )/) will be strictly negative, which means that Jeffs abstaining (y
2
=0) will reinforce Beas
decision to refrain.
Jeffs inuence on Bea, however, has a built-in asymmetry. That is, the reinforcing effect of Jeffs smoking,
ln(/(1 )), is not necessarily equal to and of opposite sign to the dissuading effect of Jeffs abstaining, ln((1 )/).
When the former is stronger than the latter, we have ln(/(1 )) +ln((1 )/) >0. So long as Jeffs smoking is infor-
mative ( +>1), this condition is equivalent to >, that is, the specicity of Jeffs smoking exceeds its sensitivity.
Conversely, the dissuading inuence of Jeffs abstaining exceeds the reinforcing inuence of his smoking when <.
Only when = would symmetric social inuence prevail.
A.3. Each siblings smoking is independently informative
Generalizing across an arbitrary number n >1 of young persons in a household, we now let the pairs (
i
,
i
) denote
the respective prior beliefs and utilities of young person i, while we let y
i
be the binary variable representing that
persons smoking practices. We assume that each young person views the smoking practices of each of her siblings
as independent information about the coolness of smoking. Thus, if Bea had another brother, Pepe, she would not
dismiss the fact of Pepes smoking as uninformative simply because he might copy his brother Jeff.
Once young person i observes the smoking habits of all of her siblings {y
j
; j =i}, her posterior probability that
smoking is cool is (in log-odds form):

i
+{ln ln(1 )}
n

j=i
y
j
+{ln(1 ) ln }
n

j=i
(1 y
j
) (21)
We write =ln(/(1 )) >0 and =ln((1 )/) >0 for compactness of notation. (The inverse mapping is:
=(exp() exp( +))/(exp() exp()) and =(exp() 1)/(exp() exp()).) We also denote u
i
= (
i

i
).
Then each young persons smoking is described by the rule:
y
i
= 1 if u
i
+
n

j=i
y
j
+
n

j=i
(1 y
j
) 0; and y
i
= 0 otherwise (22)
J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264 263
Eq. (23) reproduces Eq. (1) in the main text. Letting = , we can rewrite this rule as
y
i
= 1 if u
i
+(n 1) +
n

j=i
y
j
0; and y
i
= 0 otherwise (23)
for all i =1, . . ., n. This reproduces Eq. (2) in the main text.
In Eqs. (23) and (24), the term u
i
= (
i

i
) captures each young persons prior beliefs (
i
) about the coolness
of smoking, as well as the perceived net benets of smoking (
i
= ln B
i
ln C
i
). The former can depend on other
social inuences, such as the smoking practices of adults and peers outside the household, or any other information
about the perceived dangers or acceptability of smoking. The latter can depend on the price of cigarettes, income, age,
gender, and other personal or household characteristics.
So long as we assume that each young person views the smoking practices of each of her siblings as independent
information about the coolness of smoking, the parameters and in Eq. (24) will remain constant and independent
of the total number of siblings n in the family. However, if Bea recognizes the interdependence of the smoking practices
of her brothers Jeff and Pepe, then the additional observation that Pepe smokes may be more or less informative than
the initial observation that Jeff smoked. In that case, we expect the parameter to vary with n.
Appendix B. Solution concepts in the n-player game
The three-sibling family smoking game, involving Bea, Jeff, and Pepe (sibling #3), entails the payoffs:
U
1
= W
1
+(y
2
+y
3
) +
1
U
2
= W
2
+(y
1
+y
3
) +
2
U
3
= W
3
+(y
1
+y
2
) +
3
(24)
Fig. A1 shows the solutions in the (
2
,
3
) subspace where W
1
>
1
W
1
. In the lower right rectangle
in the gure, where the disturbances (
1
,
2
,
3
) satisfy {W
1
>
1
W
1
2;
2
W
2
; W
3
2 >
3
},
smoking is a dominant strategy for Jeff (sibling #2), while abstinence is a dominant strategy for Pepe (sibling #3).
Through iterative exclusion of dominated strategies, we conclude that Beas (sibling #1s) only rationalizable strategy
is to abstain, too, and therefore the only solution is (y
1
, y
2
, y
3
) =(0, 1, 0). This conclusion holds whether Bea moves
rst or the players move simultaneously.
In the rectangular area T, however, where (
1
,
2
,
3
) instead satisfy {W
1
>
1
W
1
2,
2
W
2
;
W
3
>
3
W
3
2}, the outcome depends on the solution concept. First consider a simultaneous-move game.
Although smoking is Jeffs dominant strategy, Bea will smoke only if Pepe does, too, and vice versa. Thus, in a
Fig. A1. Equilibria in the (
2
,
3
) plane in the three-player smoking game when W
1
>
1
W
1
2. In a simultaneous-move game, the
regions R, S and T each have two Nash equilibria in pure strategies: (0, 0, 0) or (1, 1, 1) in R; (0, 0, 1) or (1, 1, 1) in S; and (0, 1, 0) or (1, 1, 1) in T.
In a sequential-move game where player #1 moves rst, the unique equilibrium is (1, 1, 1) in R, S, and T.
264 J.E. Harris, B. Gonz alez L opez-Valc arcel / Journal of Health Economics 27 (2008) 249264
simultaneous-move game, there are two Nash equilibria in pure strategies: (y
1
, y
2
, y
3
) =(0, 1, 0) and (1, 1, 1). However,
if Bea moves rst, she will choose to smoke and the determinate outcome is (y
1
, y
2
, y
3
) =(1, 1, 1). The same logic
applies to the other labeled rectangular areas R and S in the gure. For example, in the central rectangle R, no player
has a dominant strategy. In a simultaneous-move game, there are two Nash equilibria are: (y
1
, y
2
, y
3
) =(0, 0, 0) and (1,
1, 1). But when Bea moves rst and Jeff second, the only rationalizable solution is (y
1
, y
2
, y
3
) =(1, 1, 1).
The indeterminacy in the simultaneous-move solution gets worse in higher dimensions. Consider a four-
sibling smoking game in which both Bea (#1) and Jeff (#2) will smoke if just one sibling does, too (that is,
W
i
>
i
W
i
; i =1, 2), while Pepe (#3) and Silvia (#4) will each smoke only if at least three siblings do,
too (that is, W
i
2 >
i
W
i
3; i =3, 4). Under the simultaneous-move solution concept, there are three pure-
strategy equilibria: (y
1
, y
2
, y
3
, y
4
) =(0, 0, 0, 0); (1, 1, 0, 0); and (1, 1, 1, 1). In our sequential-move solution concept,
the determinate equilibrium is (y
1
, y
2
, y
3
, y
4
) =(1, 1, 1, 1).
References
Bajari, P., Krainer, J., 2004. An Empirical Model of Stock Analysts Recommendations: Market Fundamentals, Conicts of Interest, and Peer
Effects. Working Paper 10665. National Bureau of Economic Research, Cambridge MA, http://www.nber.org/papers/w10665.pdf.
Berry, S.T., 1992. Estimation of a model of entry in the airline industry. Econometrica 60, 889917.
Bikhchandani, S., Hirshleifer, D., et al., 1998. Learning from the behavior of others: conformity, fads, and informational cascades. Journal of
Economic Perspectives 12 (3), 151170.
Bjorn, P.A., Vuong, Q.H., 1997. Modeles d equations simultanees pour variables endogenes ctives: Une formulation par la theorie des jeux avec
application a la participation au marche du travail. LActualite Economique 73, 161205.
Bresnahan, T.F., Reiss, P.C., 1990. Entry in monopoly markets. Review of Economic Studies LVII, 531553.
Chao, L.-W., 2002. A comparison of consensus and nonconsensus approaches to modeling contraceptive choice behavior. Health Economics 11,
599622.
Clark, A.E., Etil e, F., 2006. Dont give up on me baby: spousal correlation in smoking behaviour. Journal of Health Economics 25 (5), 958978.
Clark, A.E., Oswald, A.J., 1996. Satisfaction and comparison income. Journal of Public Economics 61 (3), 359381.
Gaviria, A., Raphael, S., 2001. School-based peer effects and juvenile behavior. Review of Economics and Statistics LXXXIII, 257268.
Glaeser, E.L., Sacerdote, B., et al., 1996. Crime and social interactions. Quarterly Journal of Economics CXI, 507548.
Gruber, J., 2001. Risky behavior among youths: an economic analysis, introduction. In: J. Gruber (Ed.), Risky Behavior Among Youths: An Economic
Analysis. University of Chicago Press, Chicago, pp. 128.
Harris, J.E., 1980. Patterns of Cigarette Smoking. The Health Consequences of Smoking for Women, A Report of the Surgeon General. US
Department of Health and Human Services, Washington, DC, pp. 1542.
Harris, J.E., Chan, S.W., 1999. The continuum-of-addiction: cigarette smoking in relation to price among Americans aged 1529. Health Economics
8 (1), 8186.
Harris, J.E., Gonz alez L opez-Valc arcel, B., 2004. Asymmetric Social Interaction in Economics: Cigarette Smoking among Young People in the
United States, 19921999. Working Paper No. 10409, March, 2004. National Bureau of Economic Research, Cambridge, MA.
Hiedemann, B., 1998. Stackelberg model of social security acceptance decisions in dual career households. Journal of Economic Behavior and
Organization 34, 263278.
Hoxby, C., 2002. The power of peers: how does the makeup of a classroom inuence achievement. Education Next 2 (2), 5663.
Kang, C., Park, C., et al., 2007. Effects of ability mixing in high school on adulthood earnings: quasiexperimental evidence from South Korea.
Journal of Population Economics 20, 269297.
Kooreman, P., 1994. Estimation of econometric models of some discrete games. Journal of Applied Econometrics 9 (3), 255268.
Lundborg, P., 2006. Having the wrong friends? Peer effects in adolescent substance use. Journal of Health Economics 25 (2), 214233.
Lux, T., 1998. The socio-economic dynamics of speculative markets: interacting agents, chaos, and the fat tails of return distributions. Journal of
Economic Behavior & Organization 33 (2), 143165.
Manski, C.F., 1993. Identication of endogenous social effects: the reection problem. Review of Economic Studies LX, 531542.
Norton, E.C., Lindrooth, R.C., et al., 1998. Controlling for the endogeneity of peer substance use on adolescent alcohol and tobacco use. Health
Economics 7 (5), 439453.
Powell, L.M., Tauras, J.A., et al., 2005. The importance of peer effects, cigarette prices and tobacco control policies for youth smoking behavior.
Journal of Health Economics 24 (5), 950968.
Sacerdote, B., 2001. Peer effects with random assignment: results for dartmouth roommates. Quarterly Journal of Economics 106, 681704.
Schkade, D., Sunstein, C.R., et al., 2000. Deliberating about dollars: the severity shift. Columbia Law Review 100 (4), 11391176.
Tamer, E., 2003. Incomplete simultaneous discrete response model with multiple equilibria. Review of Economic Studies LXX, 147166.
U.S. Bureau of the Census, 1996. Current Population Survey, January 1996: Tobacco Use Supplement Technical Documentation. Bureau of the
Census, Washington, DC.
Usbc. U.S. Bureau of the Census, 1998. Current Population Survey, September 1998: Tobacco Use Supplement Technical Documentation. Bureau
of the Census, Washington DC.
Winston, G.C., Zimmerman, D.J., 2003. In: Hoxby, C. (Ed.), Peer Effects in Higher Education. College Decisions: How Students Actually Make
Them and How They Could. University of Chicago Press & National Bureau of Economic Research, Chicago.

You might also like