Faculty Research Working Papers Series

Transcription

1 Faculty Research Workng Papers Seres Socal Interactons and Smokng Davd Cutler Department of Economcs Harvard Unversty Edward L. Glaeser Department of Economcs & John F. Kennedy School of Government Harvard Unversty March 2008 RWP The vews expressed n the KSG Faculty Research Workng Paper Seres are those of the author(s) and do not necessarly reflect those of the John F. Kennedy School of Government or of Harvard Unversty. Faculty Research Workng Papers have not undergone formal revew and approval. Such papers are ncluded n ths seres to elct feedback and to encourage debate on mportant publc polcy challenges. Copyrght belongs to the author(s). Papers may be downloaded for personal use only.

2 SOCIAL INTERACTIONS AND SMOKING by Davd Cutler and Edward L. Glaeser Harvard Unversty and NBER August 2007 We are grateful to Alce Chen for research assstance, to Are Kapteyn for comments, and to the Natonal Insttutes on Agng for research support.

3 Abstract Are ndvduals more lkely to smoke when they are surrounded by smokers? In ths paper, we examne the evdence for peer effects n smokng. We address the endogenety of peers by lookng at the mpact of workplace smokng bans on spousal and peer group smokng. Usng these bans as an nstrument, we fnd that ndvduals whose spouses smoke are 40 percent more lkely to smoke themselves. We also fnd evdence for the exstence of a socal multpler n that the mpact of smokng bans and ndvdual ncome becomes stronger at hgher levels of aggregaton. Ths socal multpler could explan the large tme seres drop n smokng among some demographc groups. Davd Cutler Edward Glaeser Department of Economcs Department of Economcs Harvard Unversty Harvard Unversty 1875 Cambrdge Street 1875 Cambrdge Street Cambrdge, MA Cambrdge, MA dcutler@harvard.edu eglaeser@kuznets.harvard.edu and NBER and NBER 2

4 I. Introducton A large and growng lterature suggests that ndvdual choces are nfluenced by the choces of ther frends and neghbors. These peer effects have been found n droppng out, unemployment, crme, pregnancy and many other settngs (Crane, 1991, Case and Katz, 1991, Glaeser et al., 1996, Topa, 2001, Brock and Durlauf, 2001, Kuzemko, 2006). The older work n ths lterature was crtczed because the company you keep s rarely random (Mansk, 1993). Newer work n ths area has documented peer effects n settngs where there s real random assgnment lke college dormtores (Sacerdote, 2001). There are many reasons to thnk that peers matter for health-related behavors. In many cases, health-related behavors are more fun to do when others are dong them too (drnkng, for example). Peers are also a source of nformaton (the benefts of a mammogram) or about what s acceptable n socety (the approbaton accorded smokers). A recent study suggested that a good part of the obesty epdemc n the Unted States s spread from person to person, n a manner remnscent of vral nfectons (Chrstaks and Fowler, 2007). These nterpersonal complementartes can have enormous socal mpact. In addton to helpng us understand how health behavors operate, they magnfy the mpact of polcy nterventons. The exstence of socal nteractons mples that a polcy nterventon has both a drect effect on the mpacted ndvdual and an ndrect as that person s behavor mpacts everyone around. These ndrect effects create a socal multpler where the predcted mpact of nterventons wll be greater when the nterventons occur at large geographc levels than when they occur ndvdually (Glaeser, Sacerdote and Schenkman, 2003). The socal multpler also suggests that parameter estmates from aggregate regressons can mslead us about ndvdual level parameters. In ths paper, we assess the evdence on socal nteractons n one partcularly mportant healthrelated behavor: smokng. There are a number of reasons we mght expect to see socal nteractons n smokng, as we dscuss n Secton II. These nclude drect socal nteractons (where one person s utlty s affected by whether others are dong the same thng); the socal formaton of belefs; and supply-sde nteractons from market creaton n a stuaton n fxed costs.

5 Secton III lays out the emprcal mplcatons of socal nteractons. The most straghtforward mplcaton of socal nteractons s that an exogenous varable that ncreases the costs of a behavor for one person wll decrease the prevalence of that behavor s hs or her peers. Socal nteractons models also predct excess varance n smokng rates across aggregates. Fnally, the exstence of socal nteractons mples that the measured mpact of an exogenous varable on an outcome becomes larger at hgher levels of aggregaton. In Sectons III and IV, we look at these three emprcal predctons. At the ndvdual level, we examne the mpact of workplace smokng bans on spousal smokng. Evans, Farrelly and Montgomery (1999) show that workplace bans have a sgnfcant mpact on the probablty that an ndvdual wll smoke and that these bans survve varous estmaton strateges that address selecton of smokers nto smoke frendly workplaces. We look at whether people are more lkely to smoke f ther spouse smokes, usng workplace smokng bans as an nstrument for spousal smokng. The IV estmate s large: we estmate that an ndvdual whose spouse smokes s 40 percent more lkely to smoke. The nstrumental varables estmate s hgher for men than for women, suggestng that men are more nfluenced by spousal smokng. These effects are also stronger for people wth some college than for people wth college degrees or people who were hgh school dropouts. In Secton IV, we turn to the other emprcal mplcatons of socal nteractons. We frst show that the mpact of smokng bans appears to be greater at the area level than at the ndvdual level. At an ndvdual level, a workplace ban reduces the probablty of smokng by about fve percent. At the metropoltan area level, a ten percent ncrease n the share of workers facng workplace bans reduces the share of people who smoke by more than three percent sx tmes greater than the.5 percent predcted by the ndvdual model. At the state level, the socal multpler rses to more than ten. We also examne the predcton that socal nteractons create excess varance of aggregate smokng rates. We fnd that the standard devaton of smokng rates across metropoltan areas or states are about seven tmes hgher than the rates that would be predcted f there were no socal nteractons and f there were no exogenous varables that dffered across space. Snce there are sgnfcant exogenous varables that dffer across space, we do not put complete stock n these 2

6 numbers. Stll these hgh varances provde suggest evdence supportng the exstence of socal nteractons n smokng. Secton VI turns to the queston of whether socal nteractons can help us make sense of the tme seres of smokng. Socal nteractons predct s-shaped adopton curves and changes are a functon of current levels of smokng. A smple regresson suggests that socal nteractons are not obvous n the natonal dynamcs of cgarette prevalence, but our samples for ths regresson are small. The last secton concludes. II. Sources of Socal Interactons Why should one person s smokng ncrease hs neghbor s tendency to smoke? There are three broad categores of reasons for such socal nteractons: (1) drect socal nteractons, ncludng socal approval and stgma, (2) the socal formaton of belefs and (3) market-medated spllovers that occur because of fxed costs n the provson of healthy or unhealthy behavor. In ths secton, we brefly revew these three possble reasons for nter-personal complementartes n smokng and other health related behavors. The frst reason that one person s smokng, or eatng or exercse, mght postvely nfluence a neghbor s choces s that t s more pleasant to do somethng together than alone. Ths s most obvous n the context of eatng, where t s more pleasurable (most of the tme) to eat wth others rather than eatng alone. Because of the desre to eat together, people are more lkely to go to donut shops, steak houses, or McDonald s, f ther frends are also dong so. Drnkng s also a socal actvty; f one s frends lke to drnk n bars, the returns from gong to bars rses. Smokng and exercse may be somewhat less socal actvtes, but many people lke to exercse or smoke wth frends around. Conversely, smokng around a non-smoker can be much less pleasant because of the dscomfort caused by second-hand smoke to a non-smoker. Whle there may be debate about the health consequences of second hand smoke, there s less dsagreement about whether non-smokers dslke smoke. If a smoker has some degree of altrusm for the uncomfortable non-smoker, or f the non-smoker chooses to recprocate hs dscomfort by scoldng the smoker, then ths wll decrease the returns to smokng around non-smokers. 3

7 A second reason for socal nteractons n health behavors s that belefs may themselves be formed through socal learnng. One type of socal learnng model suggests that people nfer truth from the behavor of others (e.g. Ellson and Fudenberg, 1993). A person may not know whether moderate drnkng s good or bad, but they can get gudance on ths by watchng others they beleve have more nformaton. In these models, the presence of frends and neghbors who smoke, drnk or exercse wll provde evdence about the benefts of these actvtes. Conversely, the absence of smokng wll be taken to mean that there s somethng wrong wth lghtng up. Of course, conversaton also transmts nformaton (e.g. DeMarzo, Vayanos and Zwebel, 2003). If smokng, or any other harmful actvty, ncreases one s belef n the net benefts of that actvty perhaps because of cogntve dssonance then smokers are lkely to artculate the vew that cgarettes are pleasurable or not harmful. These vews wll then be transmtted n conversaton and perhaps persuade some peers that smokng s less harmful. The power of these vews wll depend, of course, on the extent to whch other messages about the benefts or harms of the actvty are beng regularly broadcast. The thrd reason for socal nteractons works through the market. The typcal assumpton about markets s that supply curves slope up: when more people consume a good, the prce of that good rses. Ths creates a negatve socal nteracton; more people smokng wll drve up the prce of cgarettes, and dscourage some margnal smokers from smokng. However, as Waldfogel (2003, 2006) has recently emphaszed, n the presence of fxed costs these negatve market-based socal nteractons can be reversed. Supplers are only lkely to pay the fxed costs to set up f the market sze s suffcently hgh. In that case, the market creates a strong postve socal nteracton. Ths market-based nterpersonal complementarty s more lkely n goods wth fxed costs, such as restaurants, grocery stores, bars or health clubs. Cgarettes producton tself has large fxed costs, but snce transport costs are low, cgarette avalablty does not depend on local market sze. However, several studes have shown that healthy foods are hard to buy n low ncome areas, presumably because of lmted demand. The presence of health clubs and bars also depend on the presence of szable local demand. 4

8 The relatve mportance of these dfferent types of socal nteractons wll dffer across behavors. Drect nteractons and belef formaton seem more mportant for smokng. Market-based nteractons are more lkely to be mportant for exercse and consumpton of healthy food. In the next secton, we wll not dstngush between these dfferent sources of socal nteractons but dscuss more generally the emprcal mplcatons of nterpersonal complementartes n healthrelated behavors. III. Emprcal Tests of Socal Interactons The lterature on socal nteractons has broadly dentfed four dfferent emprcal mplcatons of socal nteractons. Frst, socal nteractons mply that a person s more lkely to undertake an actvty when hs or her peers are also undertakng that actvty. Second, the exstence of socal nteractons mples a socal multpler, where the mpact of some exogenous characterstc on the outcome at an ndvdual level s much smaller than the mpact of that same characterstc on the outcome at an aggregate level. Thrd, socal nteractons mply hgh levels of varance n the actvty across space (Glaeser, Sacerdote and Schenkman, 1996). Fourth, n a dynamc settng, socal nteractons lead to an S-shaped adopton curve. In ths secton, we present a partcularly smple socal nteracton model that llustrates the frst three ponts. In Secton VI, we dscuss a dynamc model. We start wth a smple model of socal nteractons. We assume that ndvdual receves prvate benefts from an actvty, X, of A X, where A dffers across ndvduals. The cost of the actvty s.5x 2. To capture socal nteractons, we assume that benefts ncrease by b tmes that average consumpton of X among person s frends, whch we denote therefore 2 ( A bx ) X X Xˆ. The utlty of ndvdual s + ˆ. When ndvduals set margnal benefts equal to margnal costs, the optmal level of X wll satsfy X = A + bxˆ. Aggregatng ths relatonshp mples that Xˆ ˆ = A /(1 b), where  refers to the average value of A n s peer group. Substtutng ths term n then mples that ndvdual X wll equal 5

9 A + baˆ /(1 b). If b s greater than ½, then the mpact of average A s greater than the mpact of ndvdual A. These calculatons delver the basc emprcal mplcatons of socal nteractons models. Frst, there wll be greater varaton n the outcome across space than would be predcted based on ndvdual dfferences alone. Wthn groups, the varance of the outcome wll be Var(A ) whle 2 the varance of outcomes across groups wll equal Var( Aˆ ) /(1 b). If there are N people n each group who are allocated randomly, then Var( Aˆ ) = Var( A ) N, so n that case, the rato of the / aggregate varance to the ndvdual wthn group varance should equal 1/N(1-b) 2. Hgh group level varance s a sgn that b s hgh. 1 Whle we mplement ths test, we note one obvous dffculty wth t: the rato of across to wthn group varance s lkely to be based upwards because of omtted characterstcs that dffer at the group level. For example, f exogenous tastes for smokng dffer across areas and we cannot control for tastes, we wll attrbute the varaton n smokng rates across areas to socal spllovers rather than tastes. One method of dealng wth ths problem s to control extensvely for observable characterstcs and then to assume that the heterogenety across groups n the unobservable characterstcs s some multple of the heterogenety across groups n observable characterstcs. A second mplcaton of the model s the exstence of a socal multpler. To see ths, assume that A = a + δz where δ s a constant and z s an exogenous characterstc such as ncome or publc polcy regulatons. In ths case, regressng the outcome on z at the ndvdual level wll gve a coeffcent of δ, whle the same regresson at the aggregate level wll gve a coeffcent of δ /(1-b). Thus, the group level relatonshp wll be stronger than ndvdual relatonshp, whch s the defnton of a socal multpler. The most common emprcal approach to socal nteractons has been at the ndvdual level, estmatng a regresson of one person s outcomes on the outcomes of a neghbor. The reflecton problem (Mansk, 1993) means that a drect regresson of ths sort does not recover the parameter 1 We conduct our test usng standard devatons: the rato of the standard devaton at the group level, to the standard devaton at the ndvdual level dvded by the square root of N s an estmate of 1/(1-b). 6

10 b. For example, assume a peer group of two people, and j. Then, person s outcome s A +bx j and person j s outcome s A j +bx. Solvng these two equatons mples that person s outcome equals (A +ba j )/(1-b 2 ) and person j s outcome equals (A j +ba )/(1-b 2 ). Straghtforward analyss shows that a unvarate regresson where person s outcome s regressed on person j s outcome does not yeld the parameter b, but rather 2b/(1+b 2 ). External factors can help us wth ths problem, however. Specfcally, f A = a + δz and z j s used as an nstrument for A j then the nstrumental varables estmate of the socal nteracton (Cov(A,z j ) / Cov(A j,z j )) wll equal b. We wll follow ths approach n our analyss. IV. Socal Interactons n Smokng: Drect Tests Surely a spouse s among the most mportant of all socal nfluences. For all of the reasons dscussed above, we would expect the nfluence of behavors to be partcularly large wthn a famly. In addton, smokng mght be senstve to peers or other people smlarly stuated. In ths secton, we look at the nfluence of one spouse s smokng decsons on the smokng propensty of the other spouse. We also look at the nfluence of smokng rates for people wth smlar demographc characterstcs. Clearly the decson of two marred people or frends to smoke s endogenous. To address the endogenety ssues dscussed above, we follow Evans, Farrelly and Montgomery (1999) and use the presence of workplace smokng bans as an nstrument for the smokng of one spouse. We use the Current Populaton Survey (CPS) tobacco supplement data for nformaton on smokng rates and workplace smokng bans. The CPS asks about smokng and smokng bans n four perods: 1992/93, 1995, 1998 and We sample people between the ages of 15 and 64. The smokng data s asked of everyone. The smokng ban queston s asked only of ndoor workers. We dscuss ths more below. Table 1 shows the means and standard devatons from ths data sources. Between 1992 and 2002, the overall smokng rate declned from 25 percent to 20 percent, a reducton of one-ffth. The declne for ndoor workers, who are those effected by smokng bans, was smlar: 24 percent n 1992/93 to 20 percent n

11 Smokng bans for ndoor workers were spreadng rapdly n the 1990s. Whle the overall share of the sample wth a smokng ban ncreases from 35 percent n 1992/93 to 45 percent n 2002, the share of the ndoor workers wth smokng bans ncreased from 66 percent n 1992/93 to 79 percent ten years later. The current omnpresence of workplace bans represents a remarkable change over 25 years. Evans, Farrelly and Montgomery (1999) report that as late as 1985, only one-quarter of workplaces banned smokng. As Evans, Farrelly and Montgomery (1999) dscuss, the estmated mpact of smokng bans on smokng may be based because of sortng across jobs. Smokers may choose jobs that are partcularly smoke-frendly, and ths wll cause a negatve correlaton between workplace bans and smokng that does not reflect the mpact of the bans. Ther own nstrumentaton strategy suggests that ths selecton (wthn ndoor jobs) s relatvely weak. We have no comparable sources of exogenous varaton. As such, we wll look at the mpact of workplace bans drectly wthout usng nstruments. We start by lookng at the mpact of smokng bans on the smokng rates of people affected by them. To do ths, we estmate a model of smokng rates as a functon of demographcs and the presence of a smokng ban: (1) Smoke = β 0 + β1 Smokng Ban + Z β + ε where denotes ndvduals and Z s the control varables. We nclude a number of standard controls: age and ts square, gender, famly sze, famly ncome, a dummy for mssng ncome, educaton (<hgh school, hgh school, some college, college grad, >college), race/ethncty (whte, black, Hspanc, other race), martal status (marred, dvorced, separated, wdowed, never marred), ndustry dummes, occupaton dummes, a dummy for whether the person s employed, and a dummy for whether the person s an ndoor worker. We also control for metropoltan area and year fxed effects so that our results reflect changes n smokng bans wthn regons over tme. The frst column n Table 2 shows our bass results. Snce the dependent varable s dchotomous, we report margnal effects from a Probt regresson. We estmate that workers who face workplace smokng bans are 4.6 percent less lkely to be smokers. The coeffcent s hghly statstcally sgnfcant. The magntude here s smlar to that found n Evans, Farrelly and 8

12 Montgomery (1999), who estmated that smokng bans reduce workplace smokng by fve percent. We are less concerned wth the other varables, but some are worthy of note. Surprsngly, we do not fnd a sgnfcant effect of cgarette taxes on smokng. The coeffcent s negatve, as expected, but not statstcally sgnfcant. It may be that by the late 1990s, the most prce senstve smokers have already left the market. More educaton s negatvely related to smokng, wth large coeffcents. College graduates are 15 percent less lkely to smoke than hgh school graduates. Blacks and Hspancs are less lkely to smoke than are whtes, and employed people smoke less. We now turn to the models ncludng spllovers. In regresson (2), we show the ordnary least squares regresson when ndvdual smokng s regressed on all of the varables n the frst regresson and on an ndcator varable for whether the spouse smokes. 2 The regresson shows that people whose spouse smokes are 21 percent more lkely to smoke themselves. We would normally expect ths coeffcent to be based upwards both because of the endogenety of spousal smokng and because of selecton of spouses. Regresson (3) looks at the spllovers of smokng n a more general peer group. As s common n the lterature, we defne the peer group as people n the same metropoltan area and cohort group wthn the same metropoltan area and wth the same age (14-30, 31-50, and 51-64) and educaton level (<hgh school, hgh school, come college, college graduate). There s a very hgh correlaton of smokng rates across people n a common reference group. The coeffcent on reference group smokng s 0.8, whch means that as the share of peers that smokes ncreases by 10 percent, the probablty that an ndvdual wll hmself smoke ncreases by eght percent. As n the case of the spousal smokng coeffcent, we expect ths coeffcent to be based upwards because ndvduals nfluence ther peers and because of omtted varables that are correlated across peers. The obvous soluton n each case s nstrumental varables. In the case of spousal smokng, we nstrument wth whether the spouse has a smokng ban at work. In the case of peer group 2 Snce ths s a prelude to the nstrumental varables estmates, we also nclude dummes for whether the spouse s employed, and whether the spouse s an ndoor worker. 9

13 smokng, we nstrument wth the share of the peer group that has a smokng ban at work. Regressons (4) and (5) show these results the former for spousal smokng only, and the latter for spousal and reference group smokng. The nstrumentaton has very dfferent effects on the estmated spouse and reference group coeffcents. When we nstrument usng smokng bans facng one s spouse, we fnd that the estmated mpact of spousal smokng ncreases to.4, so that people whose spouses smoke are 40 percent more lkely to smoke themselves. Whle the magntude of ths coeffcent s not unreasonable, we are somewhat skeptcal about the fact that the estmated coeffcent rses. One nterpretaton of ths mght be that we are not measurng the ntensty of spousal smokng, and workng n a place wthout a ban mght be partcularly correlated wth ntensve smokng. An alternatve nterpretaton s that spouse s workplace smokng bans are correlated wth other characterstcs, lke the pro-smokng atmosphere n one s socal group, that we cannot adequately control for. In regresson (5), we see that the nstrumental varables approach completely elmnates the estmated mpact of peer smokng on an ndvdual s decson to smoke. Whle the standard error s large (29 percent), the coeffcent s very small (5 percent). The coeffcent on spousal smokng, n contrast, s essentally unchanged. One nterpretaton of these results s that spousal smokng does have spllovers, but peer group smokng does not. Another vew s that our nstrumental varables peer group coeffcent s not precsely estmated enough to really say much about the mpact of peers on smokng. One queston commonly speculated about s how spllovers dffer by demographc group. One often hears that less educated groups mght be more response to peer nfluences, though nformaton dssemnaton s perhaps greater n better educated groups. In Table 3, we estmate the spllover effects separately for dfferent populaton subgroups. The regressons are all smlar to those n Table 2, though but we only report the coeffcents on workplace bans, spousal smokng and peer group smokng. The frst row n the table reports our benchmark results from column (5) of Table 2. The next two rows report these results separately for men and women. Workplace smokng bans have a larger mpact on men (5.2 percent) than on women (2.9 percent). Ths may be because 10

14 men are more lkely to work full tme, or because men nfer more from a workplace smokng ban than do women. Men are also more senstve to spousal smokng than are women. The coeffcent on (nstrumented) spousal smokng s 0.50 for men and 0.37 for women. Accordng to these fndngs, wves have a bgger mpact on husbands than husbands have on wves. The reference group smokng rate s nsgnfcant for both genders. The next four rows show the results for four separate educaton groups: hgh school dropouts, hgh school graduates wth no college, people wth some college educaton and people wth college degrees. The mpact on workplace bans s strongest for those ndvduals n the mddle educaton categores. The mpact of spousal smokng s strongest for people wth some college and weakest for people who are hgh school dropouts. The reference group effects dffer substantally across educaton subgroups but are never statstcally sgnfcant. Overall, these fndngs support the dea of a substantal socal nteracton n smokng between spouses. Whle we are not confdent that the rght coeffcent s.4, rather than.2, we are reassured by the fact that the postve socal spllover s robust to our nstrumental varables strategy. The reference group may also be mportant, but the fact that t s not robust to our nstrumental varables strategy makes us less confdent about ts strength. V. Socal Multplers and Excess Varance n Smokng We now turn to other evdence for socal spllovers n smokng: varablty across groups and socal multplers. We start wth non-parametrc evdence: the varablty n smokng rates across groups. At the ndvdual, our estmated smokng rate of 24 percent mples a standard devaton of.43. If there were no omtted varables across metropoltan areas and f there were no socal nteractons, then ths varance should declne substantally wth group sze. Specfcally, the standard devaton of smokng rates across a group of sze N should equal. 42/ N. Our metropoltan area samples have, on average, 3,238 ndvduals, whch mples that the standard devaton of smokng rates across groups should equal approxmately.008. As table 3 shows, ths s approxmately one-sxth of the actual varaton n smokng rates across our metropoltan area samples. At the state level, our average sample sze s 10,684 whch mples 11

15 that the standard devaton of smokng rates across state groups should equal approxmately.004. Agan, the actual standard devaton s almost seven tmes larger than ths amount. Usng the calculatons n Secton II, an aggregate to ndvdual standard devaton of 6 suggests a value of b of.83. Surely, ths estmate s based upwards because of omtted group level characterstcs. Nonetheless, there s a hgh level of varaton at the group level, whch supports the dea that socal nteractons may be mportant n smokng. A thrd test for socal multplers s to look at the mpact of external factors on smokng rates at the ndvdual and group level. As Secton II ponted out, n a stuaton of socal multplers, the aggregate mpact of a partcular factor wll be greater than the ndvdual mpact. We test ths usng the ndvdual, MSA, and state-level samples. The basc approach of these regressons s to regress smokng on the same characterstcs at the ndvdual, metropoltan area and state level. If socal nteractons are mportant then we should expect the mpact of characterstcs to become more mportant at hgher levels of aggregaton (Glaeser, Sacerdote and Schenkman, 2003). In prncple, a socal multpler could show up n any varable, but we would be less nclned to see t n varables that are strongly correlated wth socal groupngs. For example, even though age s correlated wth smokng, we mght not expect to fnd that a large socal multpler n age, because people of smlar age groups tend to sort together. Thus, the presence of a large number of young smokers n a partcular locale would not have a large mpact on the smokng habts of older people. Wth ths n mnd, we focus most heavly on our key varable the presence of smokng bans and look at whether the mpact of ths varable ncreases at hgher levels of aggregaton. We also look at the spllovers assocated wth years of educaton, ncome, and basc demographcs (age and gender). Table 5 shows the results of ths estmaton. The frst column of Table 5 shows our basc ndvdual level specfcaton. The coeffcent s smlar to Table 2, though slghtly larger, reflectng the restrcton to 2001 and the compresson of educaton nto a sngle varable. The second and thrd columns repeat ths specfcaton at the metropoltan area level, and the state level. The coeffcent on the smokng ban varable ncreases across columns. The ndvdual coeffcent of ncreases to at the metropoltan area level and at the state level. 12

16 A socal multpler of four at the metropoltan area level and 12 at the state level gves us another estmate of 1/(1-b) whch s agan compatble wth an estmate of b rangng from Of course, just as the varance estmates can potentally be based by omtted area level characterstcs, the socal multpler numbers are also lkely to be based upwards. Nonetheless, ths provdes suggestve support for sgnfcant socal nteractons n the smokng. Perhaps the other two most natural canddates for varables n whch to look for socal multplers are ncome and educaton. The years of educaton measure shows essentally no socal multpler. The logarthm of ncome shows a much stronger socal multpler of three at the metropoltan level and fve at the state level. Agan, ths s compatble wth hgh levels of socal nteractons, between.67 and.8. Table 6 looks at these socal multplers wthn educaton categores. In ths case, we just look at the socal multpler on the smokng ban varable. We fnd the largest socal multplers for hgh school graduates and the smallest for college graduates. In these regressons, socal nfluence n smokng s more mportant for less educated people. VI. The Smokng Tme Seres In the prevous two sectons, we focused on cross-sectonal mplcatons of socal nteractons. In ths secton, we turn to the dynamc mplcatons of socal nteracton models and ther connecton wth the tme seres of cgarette consumpton. The basc structure of dynamc socal nteractons models s to assume that the rate at whch ndvduals choose a behavor s an ncreasng functon of the share of the populaton that s already selectng that behavor. For example, f the populaton was fxed and nfntely lved, and f people who started smokng never stopped, then a dynamc socal nteracton model mght take the form: (2) S(t+1)-S(t) = (a 0 +a 1 S(t)) (1-S(t)), where S(t) s the share of the populaton that smokes at tme t and a 0 and a 1 are parameters. In ths framework, all non-smokers have some probablty of swtchng to become smokers (a 0 ) and 13

17 ths probablty ncreases wth the share of the populaton that s already smokng. The parameter a 1 determnes the power of the socal nteractons. In ths formulaton, hgher values of S(t) are assocated wth a more S-shaped curve, and t s ths S-shaped curve that s the hallmark of dynamc socal nteracton models. For example, Fgure 1 shows the tme paths mpled by three dfferent values of a 1. In all three cases, we assume that S(0)=.05, and a 0 =.02. We show results for a 1 =.1, a 1 =.2 and a 1 =.3. Hgher values of a 1 mply both a faster convergence to everyone smokng and also a more s-shaped curve. Whle ths one-sded model mght be approprate for a tme perod when smokng was rsng the frst half of the century, for example t seems ll-suted for the last 40 years, when cgarette smokng has been declnng. A more sensble model mght assume that both smokers and nonsmokers have a probablty of transtonng nto the other group. For example, we mght assume that a non-smoker becomes a smoker between tme t and t+1 wth probablty a 0 +a 1 S(t), and a smoker becomes a non-smoker between tme t and tme t+1 wth probablty b 0 +b 1 (1-S(t)). In ths formulaton, both the constant transton probabltes and the socal mpacts of smokng may dffer. A partcularly natural assumpton mght be that a 1 = b 1 so that the socal mpacts of smokng and non-smokng are dentcal, but that the basc transton probabltes (a 0 and b 0 ) dffer. We thnk of changes n belefs about the health consequences of smokng as reflectng changes n those parameters. Wth these assumptons, the new dfference equaton characterzng smokng rates s: (3) S(t+1) - S(t) = (a 0 +a 1 S(t))(1-S(t)) - (b 0 +b 1 (1-S(t))S(t). The change n smokng ncludes non-smokers who become smokers (the frst term on the rght hand sde) and smokers who become non-smokers (the second term on the rght hand sde). Ths equaton can be rewrtten: S(t+1)-S(t)= a 0 + (a 1 -a 0 -b 0 -b 1 )S(t)+(b 1 - a 1 )S(t) 2. In the case that socal nteractons are the same for smokng and non-smokng (a 1 =b 1 ), ths equaton reduces to: (4) S(t+1) - S(t) = a 0 -(a 0 +b 0 )S(t). 14

18 In ths case, the system wll converge to a steady state S = a 0 /(a 0 +b 0 ). Fgure 2 shows two cases where the mpact of socal nteractons on both transtons s the same. 3 In the frst case, S(t) converges to the steady state level from below (startng at a smokng rate of 5 percent) and n the second case t converges from above (startng at a rate of 95 percent). Our modest emprcal mplementaton of ths s to regress changes n the smokng rates snce 1965 on the ntal share of the populaton that smokes and the square of that share. We use data from 1965 because that s when data on adult prevalence are frst avalable from the Natonal Health Intervew Survey (NHIS). Not all years of the NHIS asked about smokng; we use data from all the years that do, and consder adjacent years of the data. When we estmate ths equaton, we fnd: (4) ΔRate = *Lagged Rate +.004*Lagged Rate 2 (5.57) (.38) (.56) Standard errors are n parentheses. There are 18 observatons and the r-squared s essentally zero (2 percent). Changes n the smokng rate over the past 25 years some to be uncorrelated wth the ntal level. Ths seems to suggest that socal nteractons operate weakly at an aggregate level, though clearly the number of observatons makes us cautous of drawng strong conclusons. VII. Concluson Ths paper dscusses the possble reasons why the decson to smoke mght depend on the smokng decsons of one s peers, and the emprcal mplcatons of socal nteractons n smokng. The most obvous mplcaton s that exogenous forces that make one person s smokng less lkely wll decrease the probablty that a peer wll also smoke. Other mplcatons are that socal nteractons wll create hgh levels of varance across aggregates, and that there wll be socal multplers, where exogenous attrbutes matter more at hgher levels of aggregaton. 3 We assume a 0 =.02 and b 0 =.04, so that the steady state smokng rate s 33 percent. 15

19 We found that ndvduals whose spouse faced a workplace smokng ban where less lkely to smoke themselves. The nstrumental varables estmate of the mpact of spousal smokng suggests a 40 percent reducton n the probablty of beng an ndvdual smokng f a spouse quts. These mpacts were greatest for people who modest levels of educaton, although not unformly so. The varance n smokng rates across states and metropoltan areas s about seven tmes hgher than t would be f there were no socal nteractons and f there were no exogenous varables dfferng across space. We also fnd a sgnfcant socal multpler n the mpact of smokng bans. The bans have a much stronger mpact at hgher levels of aggregaton. These results suggest that polcy nterventons that mpact an ndvdual s smokng habt wll have both drect effects and also ndrect effects through on the smokng of peers. Workplace bans seem not only to have reduced worker smokng but also the smokng of the worker s spouse. Our results also suggest that nterventons are lkely to have larger mpacts when they are mposed at hgher levels of aggregaton, although we found lttle evdence suggestng that socal nteractons can explan the shape of the tme seres of smokng rates. 16

20 References Brock, Wllam A., and Steven N. Durlauf, Dscrete Choce wth Socal Interactons, The Revew of Economc Studes, 68(2), Aprl 2001, Case, Anne C., and Lawrence F. Katz, The Company You Keep: The Effects of Famly and Neghborhood on Dsadvantaged Youths, NBER Workng Paper No. 3705, May Chrstaks, Ncholas A., and James H. Fowler, The Spread of Obesty n a Large Socal Network over 32 Years, New England Journal of Medcne, 357(4), July 26, 2007, Crane, Jonathan. "The Epdemc Theory of Ghettos and Neghborhood Effects on Droppng Out and Teenage Chldbearng." Amercan Journal of Socology, 96(5), March 1991, DeMarzo, Peter M., Dmtr Vayanos, and Jeffrey Zwebel, Persuason Bas, Socal Influence, and Undmensonal Opnons, Quarterly Journal of Economcs, 118(3), August 2003, Ellson, Glenn, and Drew Fudenberg, Rules of Thumb for Socal Learnng, Journal of Poltcal Economy, 101(4), August 1993, Evans, Wllam N., Farrelly, Matthew C. and Montgomery, Edward B., "Do Workplace Smokng Bans Reduce Smokng?" Amercan Economc Revew, 89(5), September 1999, George, Lsa and Joel Waldfogel. "Who Affects Whom In Daly Newspaper Markets?," Journal of Poltcal Economy, 111(4), August 2003, George, Lsa and Joel Waldfogel. The New York Tmes and the Market for Local Newspapers, Amercan Economc Revew, 96(1), March 2006, Glaeser, Edward, Bruce Sacerdote, and Jose Schenkman, Crme and Socal Interactons, Quarterly Journal of Economcs, 111(2), May 1996, Glaeser, Edward, Bruce Sacerdote, and Jose Schenkman, The Socal Multpler, Journal of the European Economcs Assocaton, 1(2), Aprl-May 2003, Kuzemko, Ilyana, Is Havng Babes Contagous? Fertlty Peer Effects Between Adult Sblngs, mmeo, Prnceton Unversty, Mansk, Charles, Identfcaton of Endogenous Socal Effects: The Reflecton Problem, Revew of Economc Studes, 60(3), July 1993, Sacerdote, Bruce, Peer Effects wth Random Assgnment: Results for Dartmouth Roommates, Quarterly Journal of Economcs, 116(2), May 2001, Topa, Gorgo. Socal Interactons, Local Spllovers and Unemployment, Revew of Economc Studes, 68 (2), Aprl 2001,

21 Fgure 1: Smulated Smokng Rate wth Intaton Only Percent smokng Year a1=.1 a1=.2 a1=.3 18

22 Fgure 2: Smulated Smokng Rate wth Quttng and Intaton Percent smokng Year 19

23 Table 1: Trends n smokng rates and smokng bans Measure 1992/ Smokng rate, overall 25% 25% 24% 20% Smokng rate, ndoor workers Percent wth smokng ban, overall Percent wth smokng ban, ndoor workers Note: The sample s self-respondents aged from the Current Populaton Survey. Data are weghted usng sample weghts. 20

24 Table 2: Explanng Smokng Decsons Indvdual Ban Only Wth Peer Effects Independent OLS OLS OLS IV IV Varable (1) (2) (3) (4) (5) Smokng Smokng ban (0.005) *** (0.005) *** (0.005) *** (0.005) *** (0.005) *** Spouse smokes (0.005) *** (0.006) *** (0.082) *** (0.084) *** Reference grp smokng rate (0.012) *** (0.285) Cgarette tax (0.009) (0.009) (0.009) (0.010) (0.010) Demographcs Age (0.001) *** (0.001) *** (0.001) *** (0.001) *** (0.004) *** Age (9.4E-6) *** (1.1E-5) *** (1.1E-5) *** (1.2E-5) *** (4.5E-5) *** Female (0.003) *** (0.003) *** (0.003) *** (0.003) *** (0.004) *** Famly Sze (0.001) *** (0.001) *** (0.001) *** (0.001) *** (0.001) *** Ln(famly nc) (0.002) *** (0.002) *** (0.002) *** (0.003) *** (0.004) *** Income mssng (0.024) *** (0.026) *** (0.026) *** (0.030) *** (0.038) *** < Hgh school (0.006) *** (0.006) *** (0.005) *** (0.006) *** (0.006) ** Some college (0.004) *** (0.004) *** (0.004) *** (0.004) *** (0.020) College grad (0.005) *** (0.005) *** (0.005) *** (0.006) *** (0.056) ** > College (0.005) *** (0.005) *** (0.006) ** (0.008) *** (0.055) ** Black (0.005) *** (0.005) *** (0.005) *** (0.005) *** (0.006) *** Hspanc (0.005) *** (0.005) *** (0.005) *** (0.006) *** (0.011) *** Other Race (0.007) *** (0.006) *** (0.006) *** (0.006) *** (0.007) *** 21

25 Table 2 (contnued) Indvdual Ban Only Wth Peer Effects Independent OLS OLS OLS IV IV Varable (1) (2) (3) (4) (5) Dvorced (0.005) *** (0.006) *** (0.006) *** (0.014) *** (0.014) *** Separated (0.010) *** (0.010) *** (0.011) *** (0.017) *** (0.017) *** Wdowed (0.010) *** (0.012) *** (0.011) *** (0.016) *** (0.017) *** Never Marred 0.03 (0.004) *** (0.005) *** (0.005) *** (0.012) *** (0.013) *** Employed (0.008) *** (0.008) *** (0.008) *** (0.008) *** (0.008) *** Indoor worker (0.006) *** (0.006) *** (0.006) *** (0.006) *** (0.007) *** Spouse employed (0.005) ** (0.004) (0.005) ** (0.005) ** Spouse ndoor worker (0.004) ** (0.003) (0.004) *** (0.005) *** Pct reference group employed (0.013) *** (0.030) Pct reference grp ndoor worker (0.011) *** (0.012) ** MSA dummy Yes Yes Yes Yes Yes varables Year dummy Yes Yes Yes Yes Yes varables N 195, , , , ,579 R Notes: Data are from CPS Sept. 1992/May 1993, Sept. 1995, Sept. 1998, and Feb Tobacco Supplement Surveys. Sample composton s of people aged All regressons also nclude major ndustry (21 dummes) and major occupaton (13 dummes) effects, and are weghted by the self-response supplement sample weght. Models for ndvduals and spouses are clustered by famly d. Models ncludng cohort effects are clustered by clustered by the MSA-cohort-educaton level wth cohort ages of 14-30, 31-50, and and educaton levels of less than hgh school, hgh school, some college, and college graduates or hgher. Spouse smokes nstrumented by spouse smokng ban, and reference group smokng rate nstrumented by share of reference group wth a smokng ban. ** ( *** ) ndcates statstcal sgnfcance at the 5% (1%) level. 22

26 Table 3: Examnng the Response to Smokng Bans by Demographc Group Instrumental Varable Estmates Group Smokng ban Spouse smokes Reference group smokng rate N R 2 All , (0.005) *** (0.084) *** (0.285) By Gender Men , (0.008) *** (0.196) ** (0.416) Women , (0.006) *** (0.073) *** (0.628) By Educaton <Hgh School , (0.014) ** (0.525) (2.235) Hgh School , (0.012) *** (0.261) (5.203) Some College , (0.011) *** (0.177) *** (0.668) College (0.008) *** (0.191) (1.148) 52, Note: The reference group s based on the msa-cohort-educaton level. All regressons nclude age, age squared, famly sze, log(famly ncome), mssng ncome dummy, three ndcators for ethncty, four ndcators for martal status, cgarette tax (state + federal), 21 ndustry ndcators, and 13 occupaton ndcators. Regresson for all, men, and women also nclude 4 ndcators for educatonal attanment. Regressons for all and educaton bns nclude ndcator for gender. Spouse smokes nstrumented by spouse smokng ban, and reference group smokng rate nstrumented by share of reference group wth a smokng ban. Regressons weghted by self-response supplement weght. ** ( *** ) ndcates statstcal sgnfcance at the 5% (1%) level. 23

27 Table 4: The Varablty of Smokng Across Areas Average observatons per unt Predcted standard devaton Actual standard devaton Rato: Actual/ Predcted Indvdual MSA 3, State 10, Note: The sample s self-respondents aged from the Current Populaton Survey. Data are weghted usng sample weghts. 24

28 Table 5: The Spllover Effects of Smokng Independent Varable Indvdual MSA State Smokng ban (0.007) *** (0.112) ** (0.312) ** Years of educaton (0.001) *** (0.009) (0.026) Log (Income) (0.003) *** (0.039) *** (0.082) *** N 64, R Note: Data are from CPS June 2001 Tobacco Supplement Survey. Sample composton s respondents 18 years and older. Regressons weghted by self-response supplement weght. Regressons nclude controls for age, gender, employed, ndoor worker, and a dummy for mssng ncome. For years of educaton, frst, second, thrd, and fourth grades were averaged to 2.5 years. Ffth and sxth grades were averaged to 5.5 years, seventh and eghth grades were averaged to 7.5 years, hgh school dploma and GEDs were treated as 12 years, some college and assocates degrees were treated as 14 years, bachelors degrees were treated as 16 years, masters degrees were treated as 18 years, professonal degrees (such as MD s, DD s) were treated as 20 years, and doctorate degrees (such as PhD s or EdD s) were treated as 21 years. For ncome, <$5,000 was coded as $2,500, and >$75,000 was coded as $75,000. All other categores were averaged over the range n the choce. ** ( *** ) ndcates statstcal sgnfcance at the 5% (1%) level. 25

29 Table 6: Spllover Effects by Educaton Impact of Smokng Ban Educaton Group Indvdual MSA State <Hgh School (0.025) (0.202) (0.469) Hgh School Grad (0.013) *** (0.129) * (0.423) *** Some College (0.013) *** (0.123) *** (0.320) College Grad (0.011) ** (0.079) (0.187) Note: Data are from CPS June 2001 Tobacco Supplement Survey. The sample s ndvduals aged 18 and older. Regressons are weghted and control for age, gender, employed, and ndoor workng. 26