THEROLE OFINFORMATION ANDSOCIAL INTERACTIONS INRETIREMENT PLANDECISIONS: EVIDENCE FROM ARANDOMIZED EXPERIMENT*

ESTHER DUFLO AND EMMANUEL SAEZ

Thispaper analyzes a randomizedexperiment to shed light on the role of informationand social interactions in employees’ decisions to enroll in a Tax DeferredAccount (TDA) retirementplan within a largeuniversity. The experi- mentencouraged a randomsample of employees in a subsetof departments to attenda beneŽts information fair organized by the university, by promising a monetaryreward forattendance. The experiment multiplied by more than Ž vethe attendancerate of these treated individuals (relative to controls), and tripled that ofuntreated individuals within departments where some individuals were treated.TDA enrollmentŽ veand eleven months after the fair was signiŽ cantly higherin departmentswhere some individuals were treated than in departments wherenobody was treated. However, the effect on TDA enrollmentis almost as largefor individuals in treated departments who did not receive the encourage- mentas for those who did. We provide three interpretations— differential treat- menteffects, social network effects,and motivational reward effects—to account forthese results.

I. INTRODUCTION Thereis growingconcern in theUnited States aboutlow levelsof savings forretirement. For most U. S.families,employ- ers’pensions are the main source of cash income during retire- ment,over and aboveSocial Security beneŽ ts (see,e.g., Poterba, Venti,and Wise[1996]). However,over the last 25 years,tradi- tionalDeŽ ned BeneŽ ts and DeŽned Contribution employer pen- sionplans whereemployee participation is mandatoryhave been partly replacedwith TaxDeferred Account (TDA) retirement plans suchas 401(k)s whereemployees choose whether to partici- pate and howmuch to save for their retirement (see Poterba, Venti,and Wise[2001]). As aresult,most U. S.workersnow have tomake a decisionabout how much to save for their retirement,

*Wethank , George Akerlof, Orley Ashenfelter,Joshua Angrist,David Autor, Abhijit Banerjee, David Card, Edward Glaeser,Lorenz Goette,Jonathan Gruber, Guido Imbens, Lawrence Katz, JeffreyKling, Botond Koszegi,Michael Kremer, Alan Krueger, David Laibson, Brigitte Madrian, Sen- dhilMullainathan, Thomas Piketty, , two anonymousreferees, andnumerous seminar participants for very helpfulcomments and discussions. Wegratefully acknowledge Ž nancialsupport from the National Science Founda- tion(SES-0078535). We areespecially grateful to all the members of the BeneŽ ts OfŽce of theUniversity for their help and support in organizing the experiment. Thispaper does not re ect theviews of the University or its BeneŽ ts OfŽ ce.

© 2003by thePresident and Fellowsof Harvard Collegeand theMassachusetts Institute of Technology. The Quarterly Journal ofEconomics, August 2003

815 816 QUARTERLY JOURNALOF instead ofbeingpassive participants in theiremployer’ s pension plan. Thismakes it veryimportant to understand howretirement savings decisionsare made. Deciding howmuch to save for retirement and howto invest requiressolving a complicatedintertemporal optimization prob- lem,and having informationabout the rules governing different instruments.In sucha context,one could expect that information mayhave a largeimpact onsavings behavior.As aresult,Ž nan- cial educationis considereda potentially importantavenue to improvethe quality ofŽ nancial decisionmaking, both by policy makers[Summers 2000] and by companies.A telephonesurvey weconducted with all Fortune500 companiesrevealed that 71 percentof thesecompanies systematically hold Žnancial informa- tionsessions. A further10 percentconducts them occasionally. Bernheimand Garrett[1996], Bayer,Bernheim, and Scholz [1996], and Bernheim,Garrett, and Maki [2001], amongothers, presentevidence that participation in aŽrm’s savings plan is higherwhen Ž rmsoffer Ž nancial education.However, they rec- ognizethat an employer’s decisionto provide this information mightbe endogenous, which complicates the interpretation of thesedifferences. Becausehow much and howto save is adifŽcult decision, it is alsolikely that individuals’decisions are affected by thedeci- sionsof others in theirpeer group. First, they may obtain infor- mationabout the employer retirement plan fromdiscussions with theircolleagues, or make inferences based onobserving their decisions.Second, consumption and savings behaviormay be subjectto peer pressure and socialnorms, leading toconformity in behavior.As aresult,social network effects within thework- place mightplay an importantrole in thedecision to contribute to 401(k) retirementplans. Thispaper analyzes theevidence from a randomizedexperi- ment,designed toshed light onboth the role of information and socialinteractions on the employees’ decision to enroll in the employer-sponsoredTDA plan ofa largeuniversity. This allows usto overcome some of the very difŽ cult identiŽ cation problems in thepresence of peer effects, described notably in Manski [1993, 1995]. Each year,the university organizes a beneŽts fair and invites all ofits employeesto the fair in orderto provide information on beneŽts. In particular, astated goalof thefair is toincrease the enrollmentrate in TDAwhichis relativelylow (around 35 per- THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 817 cent).Obviously, comparing the TDA enrollmentdecisions of fair attendeeswith thosewho did notattend thefair would notpro- vide convincingevidence of a causal effectof fair attendanceon TDAenrollment,because the decision to attend thefair is endoge- nous.1 Tocircumvent this selectionproblem, we have imple- mentedthe following experiment. We selected a randomsample ofemployees not yet enrolled in theTDA and sentthem an invitationletter promising a $20 rewardfor attending thefair. Thistype ofexperiment is aclassical encouragementdesign, oftenused in medicalscience, where treatments are offered to a randomgroup of patients whothen decide whether or not to take thetreatment. Furthermore, we designed ourexperiment such that weare able toestimate social interaction effects. Namely, “treated”individuals whowere sent the invitation letterwere selectedonly within arandomsubset of departmen ts (the “treated”departments). TheŽ rststage of ourstudy analyzes theeffect of the invita- tionletter on fair attendance.Treated individuals aremore than Žvetimes as likelyto attend thefair as controlindividuals. Interestingly,nontreated individuals in treateddepartments are threetimes as likelyto attend thefair as controlindividuals in nontreateddepartments, despite thefact that onlyoriginal letter recipientscould claim the $20 reward.This shows that theinvi- tationletters not only increased the fair attendancerate for individuals whoreceived them but alsohad aspilloversocial effecton their colleagues within departments. Thesecond stage of the study triesto estimate the causal effectof fair attendanceand socialeffects on thedecision to enroll in theTDA. We showthat, Ž veand elevenmonths after the fair, individuals in treateddepartments are signiŽ cantly morelikely tohave started contributingto the TDA than controlindividuals. Thisshows that ourexperiment, and hencethe fair, was success- ful in increasingTDA enrollment.However, there is nosigniŽ - cant differencein TDAenrollmentbetween those who actually receivedour encouragement letter and thosein thesame depart- mentswho did not.We propose three different interpretations, notnecessarily mutually exclusive, to account for these facts. First,this couldbe explained by socialeffects at thedepartment

1.For example, individuals who had already decided to enroll, but are not sureexactly how much they want to contribute, may be more likely to attend the fair.See Madrian and Shea [2002] for evidence of selection in the decision to attendinformation sessions within a largeŽ rm. 818 QUARTERLY JOURNALOF ECONOMICS level.Fair attendees might be able tospread informationob- tained fromthe fair in theirdepartments. Second, our results couldalso be explained by differential treatmenteffects. Employ- eeswho come to thefair onlybecause of theŽ nancial rewardare different fromthose who decide to come to the fair becauseof theircolleagues, and it is plausible tothink that thetreatment effectis largerfor the latter group than forthe former. Finally, ourresults might also be explained by motivationalreward ef- fects.Paying individuals toattend thefair mightaffect their subjectivemotivation and thereforethe perceived value or quality ofthe information they obtain at thefair. Ourexperiment does notallow usto separately identify thesethree effects, but it allows us toconclude that theimportant decision about how much tosave for retirement can be affected by smallshocks such as a verysmall Ž nancial rewardor the in uence of peers, and thus doesnot seem to be the consequence of an elaboratedecision process. Theremainder of thepaper is organizedas follows.Section II presentsa briefdiscussion of themechanisms by whichŽ nancial educationand socialinteractions can affect retirementsavings decisions.Section III describesthe beneŽ ts fair and thedesign of ourexperiment. Section IV discussesthe reduced-form evidence. SectionV developsa simplemodel to interpret our results. Fi- nally,Section VI offersa briefconclusion.

II. INFORMATION AND SOCIAL INTERACTIONS IN SAVINGS DECISIONS Anumberof recent studies haveemphasized theimportant roleof factors others than Žnancial incentivesin thedecision to enrollin TDAplans. Madrian and Shea[2001] and Choi,Laibson, Madrian, and Metrick[2001a, 2001b] showthat default rules havean enormousimpact onemployees’ participation, contribu- tions,and assetallocations. When employees are enrolled by default in aTDA,very few opt out,and mostemployees do not changethe default contributionrate or the default allocationof assets.This evidence could be interpreted in twoways: either individuals lackinformation, and interpretthe default optionas asignal,or they do notthink very much about their retirement savings,and canbe in uenced by verysmall changes in their environment.Distinguishing thesetwo mechanisms is important, sincethey have very different policyimplications. If lackof infor- mationis important,this suggestsa potentially importantrole for THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 819

Žnancial educationsessions, through which individuals canob- tain generalinformation about retirement plan featuresas well as beguided throughtheir intertemporal maximization decision. However,if Žnancial educationhas onlya modestimpact on retirementplan decisions,this suggeststhat thesecond hypothe- sis is true. Theliterature on social interactions suggests that in both cases,social interactions are likely to affect retirementdecisions. First,individuals maylearn from their coworkers, either through discussionsor by makinginferences from their actions. The lit- eratureon informational cascades [Bikhchandani, Hirshleifer, and Welch1992; Banerjee1992; Ellison and Fudenberg1993] providereasons why information(correct or not) obtained from coworkersmay be an importantfactor in deciding whetherto participate and howto invest— giving riseto peer effects. Second, savings decisionsmay be in uenced by socialnorms or beliefs aboutsocial norms. By observingcoworkers, people can learn aboutthe proper behavior of theirsocial group, as emphasized by modelsof conformity [Bernheim 1994]. Individuals maywant to maintain thesame consumption level as what is commonin their socialgroup. In bothcases, there is a“socialmultiplier” effect: the aggregateimpact ofan interventionon a groupis largerthan the sumof its effectson each individual’ s decision.As discussed in Glaeser,Sacerdote, and Scheinkman[2002], it is oftenimportant forpolicy purposes to separate direct individual effectsfrom socialmultiplier effects. Thereis agrowingempirical literature which shows evidence ofsocial interaction effects in anumberof areas.Some empirical papers havefocused on information transmission, 2 whileothers havefocused on peer pressure. 3 Mostof these studies areobser- vational and hencesubject to difŽ cult identiŽ cation problems [Manski 1993, 1995]. Glaeser,Sacerdote, and Scheinkman[1996, 2002] propose twomain avenues to obtain suggestiveevidence on the presence ofsocial interaction effects. First, in thepresence of positive spillovers,the variance across social groups will belarger than what would bepredicted byrandomdraws. Second,there will be

2.See Besley and Case [1994] and Foster and Rosenzweig [1995] on technol- ogy adoptionin developingcountries. 3.See, for example, Evans, Oates, and Schwab [1992]on teenagers’ behavior, Bertrand,Mullainathan, and Luttner [2000] on welfareparticipation, and Munshi [2000]on contraception. 820 QUARTERLY JOURNALOF ECONOMICS acorrelationbetween individual behaviorand apredictionof aggregatebehavior based ondemographic characteristics in the group.Du o and Saez [2002a] showthat, in theuniversity stud- ied in this paper, botheffects are present: there is littlevariance ofparticipation within departments,compared with thevariance in participation rates across departments,and individual partici- pation ratesare correlated with predicted participation in their peergroups. While this evidenceis suggestive,it mightbe con- taminated by omittedvariables, correlated within thegroup and correlatedwith theobserved variables usedto predict aggregate participation rates.To address this problem,we set up arandom- ized experiment,where we affect theincentives of asubsetof the peergroup in somerandomly selected groups, and evaluate whetherthe impact ofthis interventionextends beyond the tar- getedgroup, which would bedirect evidence of asocialmultiplier effect.4

III. CONTEXT AND EXPERIMENT DESIGN

III.A. BeneŽts and theBeneŽ ts Fair Theuniversity we study has approximately12,500 employ- ees.About aquarterof the employees are faculty members.Our study was limitedto nonfaculty employees only. 5 Theuniversity providesretirement beneŽ ts toits employeesthrough a tradi- tionalmandatory pension plan, but employeescan also voluntar- ily contributeto a complementaryTax Deferred Account (TDA) 403(b) plan. Everyemployee can contribute to the 403(b) plan any percentageof theirsalary uptothe IRS limit($10,500 peryear for eachindividual in 2001). Theuniversity does not match contri- butions.Employees can choose how to invest their contributions in any combinationof fourdifferent vendors. Each year,the university organizes a beneŽts fair whereall

4.Two recent studies have used experimental or quasi-experimental situa- tionsto study socialinteraction effects. Katz, Kling,and Liebman [2001] evaluate arandomlyassigned housing voucher program whereby households living in high poverty publichousing projects were given the opportunity to move out of the project.Sacerdote [2001] analyzes peer effects among Ž rst-year studentsat Dart- mouthCollege randomly assigned to dorms. Both studies have found evidence of spillovers. 5.Du o andSaez [2002a] present suggestive evidence that staff employees’ TDA choicesare not in uenced by faculty choices and vice versa. Furthermore, staffemployees may be morerepresentative of averageU. S.workers thanfaculty members. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 821 employeesare invited tocome and learnabout the different kinds ofbeneŽ ts (such as healthbeneŽ ts, retirement beneŽ ts, etc.) provided by theuniversity. The fair is held ontwo consecutive days in earlyNovember in twodifferent locations,each one close tothe two separate main university campuses. About oneweek beforethe fair, everyemployee receives a letterthrough the universitymail system inviting herto attend thefair. Thisletter alsoprovides a briefdescription of the event. At thesame time, underseparate cover, every employee receives a packetdescrib- ing in detail universitybeneŽ ts alongwith enrollmentforms. Novemberis “openenrollment” month during whicheach em- ployeemay change her beneŽ ts choicesby submittingthe enroll- mentform. If theemployee does not send back the form, her beneŽts choicesare automatically carried over from the previous year.However, employees are free to enroll in theTDA orchange theircontribution level or investment decision at any time throughoutthe year. In bothlocations, the fair is held in alargehotel reception room.There are a largenumber of stands representingthe uni- versityBeneŽ ts OfŽce, and thevarious health and retirement beneŽts serviceproviders. The university BeneŽ ts OfŽce offers informationon all beneŽts throughdirect conversation with Bene- Žts OfŽce staff presentat thefair, and througha numberof informationpamphlets freelyavailable at theirstand. Thebene- Žts ofŽce alsoprovides information on howthe other stands at the fair areorganized. These other stands arerun by eachof the specialized serviceproviders. For example, each of the mutual fund vendorshas astand at whichthey provide information about theTDA plan and thespeciŽ c servicesthey offer within that plan. Thefair alsooffers individuals thechance to use a specially designed computerprogram to analyze theirspeciŽ c situation. Employeesare free to come any timeduring thethree and ahalf hoursduring whichthe fair is held,and visit any numberof stands theywant. III.B. ExperimentDesign Theuniversity organizes the annual fair inorderto dissemi- nateinformation about beneŽ ts and tohelp its employeesmake betterdecisions. The beneŽ ts ofŽce ofthe university realizes that theparticipation rateamong staff (34 percent)is toolow com- pared with that at otheruniversities, and suspectsthat this may bedue to lack of information. 822 QUARTERLY JOURNALOF ECONOMICS

In orderto identify thecausal effectof fair attendanceon TDAenrollment,we setup an “encouragementdesign,” by prom- ising arandomsubset of employeesa small amountof moneyfor attending thefair. In orderto shed light onsocial effects within departments,we sent those letters only within arandomsubset of departments.There are thus two distinct treatmentsin ourex- periment:receiving the letter, and beingin thesame department as someonewho receives a letter. Weused a crosssection of administrative data provided by theuniversity on all its employeesas ofAugust 2000. Were- strictedthe sample to staff employees(i.e., nonfaculty employees) aged lessthan 65 and eligibleto participate in theTDA. Of the 9700 employeesmeeting these criteria, around 3500 wereen- rolledin theTDA as ofAugust 2000. Fromnow on, we refer to theseindividuals as thepre-enrolled individuals. Theremaining 6200 individuals werenot enrolled in theTDA by August 2000. As veryfew employeesstop contributingto theTDA oncethey are enrolled,we focus on the decision to start participating in the TDA.Thus, the sample of 6200 nonenrolledindividuals is our sampleof primaryinterest. 6 TheUniversity is divided into330 departments.Depart- mentsinclude each of theacademic departments such as Econom- icsor Cell Biology,etc. In addition toacademic departments, thereare many administrative departments.For example, each library is aseparatedepartment. Each ofthedining halls isalso adepartment.In mostcases, each department has asinglegeo- graphical location.Departments sometimes share a samebuild- ing or oorwithin abuilding, but evenin thosecases, work interactionswithin departmentsare much more intense than acrossdepartments. 7 Of course,there is communicationacross departments,but it ismainlyconcentrated among higher ranked employeeswithin departments. 8 Theaverage number of staff employeesper department is 30, but themedian size is much smaller,around 15. Therefore,except in afewlarge departments,

6.Only 80of the 3500 employees enrolled in the TDA stoppedcontributing duringthe one-year period we examine. More than Ž vetimes as manyemployees startedcontributing to the TDA duringthe same period. We havenot found any signiŽcant differencesin the decision to stop contributing to the TDA across treatedand nontreated departments. 7.Academics know very wellthat, even when departments are close geo- graphically,interactions across departments are always minimal. To a large extent,the same is true for staff in administrative departments. 8.For example, administrative managers from different departments partici- patein many meetings with the central administration. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 823 wewould expecteach employee to know most of hercolleagues in thedepartment. The fact that departmentsmay not correspond exactlyto social units should lead toan attenuationbias in our socialnetwork effects estimates. In theŽ rststep, we randomly selected two-thirds ofthe departmentsof theuniversity (220 outof atotalof 330) as follows. In orderto maximizethe power of theexperiment (in acontextin whichwe know there are strong department effects), we Ž rst matcheddepartments according to their size (i.e., number of employees)and participation ratein theTDA beforethe fair. We separated departmentsinto deciles of participation ratesamong thestaff. Each decilecontains 33 departments.We then ranked themby sizewithin eachdecile, and formedgroups of three departmentsby putting threeconsecutive departments on these lists in thesame triplet. Within each of these triplets, we ran- domlyselected two departments to be part ofthegroup of treated departments.From now on, we denoteby thedummy variable D thetreatment status ofdepartments. We have D 5 1 in treated departmentsand D 5 0in controldepartments. In thesecond step, within eachof the treated departments, any individual notenrolled as ofAugust 2000, was selectedwith aprobability ofone-half. 9 Thistreatment group is composedof 2039 individuals. Fromnow on, we denote by thedummy variable L theselection status ofindividuals. Wereferto this groupas the Treatedindividuals and denotethem by 11 ( D 5 1forTreated departmentand L 5 1forbeing selected). The group formed by theemployees in thetreated departments who were not selected contains2129 individuals and is denotedby 10 ( D 5 1forTreated departmentand L 5 0fornot being selected). In total,there are 4168 individuals in thetreated departments. The control group is formedby employeesin thecontrol departments where no treat- mentswere selected; it contains2043 individuals and isdenoted by 00 (D 5 0 and L 5 0). Oneweek before the fair, wesent a lettervia universitymail tothe 2039 employeesin thetreatment group 11. Theletter remindedthem of the fair and informedthem that theywould receivea checkfor $20 fromus if theywere to come to the fair and registerat ourdesk. This letter is reproducedin theAppendix. At thefair, weset up astand forthe employees who received ourinvitation letter to register their name. Unfortunately, the

9.This selection probability is independent across individuals. 824 QUARTERLY JOURNALOF ECONOMICS

BeneŽts OfŽce did notauthorize us torecord the names of thefair participants whodid notreceive our letter. However, we recorded theirtotal number: a student stoodat thefair entranceand distributed acouponto each person who entered the hall. The couponshad different colorsaccording to the status ofthepartici- pant (active orretired), which allowed us to count the number of activeemployees who attended thefair. Everybodyhad topass throughthe narrow entrance to enter the fair, and thefew people whorefused the coupon were carefully counted. We are thus conŽdent that weaccuratelyrecorded the number of participants. In orderto collectinformation on theTDA status and thedepart- mentafŽ liation ofall thefair participants, weorganized a rafe. Thecoupons that weredistributed at theentrance of thefair had twoparts, with anumberwritten twice. Each fair attendant who wanted toparticipate in theraf e gaveus half ofthecoupon. We askedall theraf e participants theirdepartment afŽ liation and whetherthey were currently enrolled in theTDA. The raf e was held every30 minutes,and theprize was a$50 Macy’s gift certiŽcate. A totalof 1617 activeemployees attended thefair, and 573 ofthem had receivedour letter. Out ofthe remaining 1044 employees,766 (i.e.,about three-quarters) came to play theraf e and registeredtheir department afŽ liation and TDAenrollment status.An importantissue that arisesis whetherthere was selectionby D 5 1 versus D 5 0departmentsin whodecided to play theraf e (and henceprovide their department afŽ liation and TDAstatus). Wedo not believe this was thecase. Most of thosewho refused to play theraf e did sobecause they visited our stand justafter the previous raf e had beenplayed, and did not want tostay at thefair longenough to wait forthe next raf e. Therefore,we assume that fair attendants whodid notregister theirdepartment afŽ liation are distributed between D 5 1 and D 5 0departmentsas thosewho did register.Therefore, in what follows,we scaleup theattendance recorded in eachdepartment by afactorof 1044/ 766. 10 In orderto assessthe effects of theexperiment and thefair on TDAparticipation, theuniversity provided uswith threewaves of data. TheŽ rstwave was obtained inSeptember2000, justbefore thefair. Thesecond wave was fromMarch 2001 (4.5 monthsafter thefair), and thethird wavefrom October 2001 (11 monthsafter the fair).

10.We willdiscuss how modifying this assumption would affect our results. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 825

IV. RESULTS: SUMMARY STATISTICSAND REDUCED-FORM DIFFERENCES In thepresence of socialinteractions, employees who work in departmentswhere some people received the letter can be af- fectedby theexperiment even if theydid notreceive the letter themselves.They may be more likely to come to the fair them- selves,because they are reminded by othersof the event, or becauseemployees come to the fair in groups. 11 Theymay also be morelikely to enroll in theTDA evenif theydo notcome to the fair themselves,either because they are directly in uenced by the actionof thosewho went to the fair, orbecausethese individuals sharethe information they gathered at thefair. Thus,employees arepotentially subjectedto two kinds oftreatments: they can receivethe invitation letterthemselves (group 11), ortheycan be in adepartmentwhere some employees received the letter (group 10 and group11). Thosewho receive the letter are, obviously, subjectto both treatments. Thesummary statistics aredisplayed in TableI, broken down intofour groups. In columns(1) to(3) wepresent the statistics forindividuals whobelong to treateddepartments. Col- umn(1) has thestatistics forthe entire group (group D 5 1), column(2) has thestatistics forthe group of treatedindividuals (group 11), and column(3) has thestatistics forthe untreated individuals in treateddepartments (group 10). In column(4) we presentthe statistics forindividuals whobelong to the untreated departments(group 00). 12 PanelA presentsbackground characteristics. In theŽ rst wave(in September2000, beforethe fair), averysmall proportion ofemployees started contributingto the TDA (theŽ rstwave is fromSeptember 2000, but weused data fromAugust 2000, to constructthe randomization), but thereis noapparent difference acrossgroups in theseproportions. Since we are interested in changescaused by thefair, wefocus in theremainder of the analysis onindividuals whowere still notenrolled in theŽ rst wave(i.e., by September2000). Becausethe groups were chosen randomly,the mean of observable characteristics such as sex, yearsof service, annual salary,and age,are very similar across groups,and noneof the differences are signiŽ cant.

11.This is something we observed at thefair. 12.It is important to note that all these statistics (except the Ž rst row of PanelA andthe second row ofPanel B) focusonly on individuals not enrolled in theTDA inSeptember 2000, before the fair. 826 QUARTERLY JOURNALOF ECONOMICS

TABLE I DESCRIPTIVE STATISTICS, BY GROUPS

Treateddepartments

Treated Untreated Untreated All (group (group departments (group D 5 1, D 5 1, (group D 5 1) L 5 1) L 5 0) D 5 0)

(1) (2) (3) (4)

PANEL A:BACKGROUND CHARACTERISTICS TDA participationbefore 0.010 0.009 0.011 0.012 thefair (Sept. 2000) (.0015)(.0021) (.0022) (.0024) Observations 4168 2039 2129 2043 Sex (fractionmale) 0.398 0.400 0.396 0.418 (.0076)(.0109) (.0107) (.011) Yearsof service 5.898 5.864 5.930 6.008 (.114)(.161) (.16) (.157) Annualsalary 38,54738,807 38,297 38,213 (304) (438) (422) (416) Age 38.3 38.4 38.2 38.7 (.17) (.24) (.24) (.24) Observations 4126 2020 2106 2018 PANEL B:FAIRATTENDANCE (REGISTRATION DATA) Fairattendance rate among 0.214 0.280 0.151 0.049 non-TDAenrollees (.0064)(.01) (.0078) (.0048) Observations 4126 2020 2106 2018 Fairattendance rate for all 0.192 0.063 staffemployees (.0132) (.0103) Observations 6687 3311 PANEL C:TDA PARTICIPATION(ADMINISTRATIVE DATA) TDA participationrate after 0.049 0.045 0.053 0.040 4.5 months (.0035)(.0049) (.0051) (.0045) Observations 3726 1832 1894 1861 TDA participationrate after 0.088 0.089 0.088 0.075 11 months (.005)(.0071) (.007) (.0065) Observations 3246 1608 1638 1633

a. Standarderrors are in parentheses. b. TheŽ rst partof Panel B includesall individualsnot enrolled in the TDA by September 2000. The secondpart includes all employees(enrolled or notin the TDA). c.The average fair participation in the nontreated departments was obtainedfrom the registration informationcollected at thefair. Since only 75 percentof the participants registered,the participation was adjustedby a proportionalityfactor. d.Demographic information and TDA participation are all obtainedfrom administrative data.

In PanelB wecan see that ourinducement strategy had a dramatic effecton theprobability ofattending thefair: in treated departments,as manyas 21.4 percentof individuals attended the THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 827 fair. In controldepartments, fewer than 5percentof individuals attended thefair. Comparingtreated individuals versuscontrols in thetreated departments in columns(2) and (3) showsthat socialeffects account for a largefraction of the effect of our experimenton fair attendance.The fair attendancerate of those whoreceived our letter is 28 percent,and is 15.1 percentfor those in thetreated departments who did notreceive the letter. Thus, thedifference in theattendance rate between group 10 and group 00 (which is solelydue to social effects) is over10 percentage points.13 In PanelC welook at TDAparticipation. After 4.5 months relativelyfew peoplehave enrolled. However, employees in treateddepartments are already signiŽcantly morelikely to be enrolledthan employeesin controldepartments (4.9 percentver- sus4 percent).However, individuals in group11 arenot more likelyto be enrolledthan individuals ingroup10. Thedifference betweengroups 10 and 00 is relativelylarge at 1.3 percentage points.Eleven months after the fair, enrollmentis higherstill, and thedifference between treated departments and controlde- partmentsis 1.4 percentagepoints. The difference between groups11 and 10 is nowpositive, but still verysmall and insig- niŽcant. The difference between group D 5 1 and group D 5 0 remainsequal to 1.3 percentagepoints. In orderto analyze thedifferences, we consider simple re- duced-formregression speciŽ cations. Denote, respectively, by fi j and yi j thefair attendanceand theTDA enrollmentdecisions of individual i in department j.Similarly, Li j is thedummy for receivingthe inducement letter, and Dj thetreatment status of thedepartment. The average effects on fair attendanceand TDA enrollmentof being in atreateddepartment ( D 5 1) versus a controldepartment ( D 5 0)(irrespectiveof individual treatment status L)arecaptured by thefollowing speciŽ cations:

(1) fij 5 a1 1 b1Dj 1 eij, and

13.This result is, of course, sensitive to the assumption we made about departmentafŽ liation of fair attendants who did not register at our desk. If we makethe extreme assumption that all nonregistered individuals come from D 5 0departments,the fair participation rate for group 10would drop to 11 percent butstill be higher than for group 00(which would go up to 9 percent).In addition, weshow below that the increase in fair attendance in group 10isparalleled by an increasein theirTDA participation. 828 QUARTERLY JOURNALOF ECONOMICS

TABLE II REDUCED-FORM ESTIMATES (OLS)

Dependentvariable

TDAenrollmentafter Fair attendance 4.5 months 11 months (1) (2) (3)

PANEL A:Averageeffect of departmenttreatment Treated 0.166 0.0093 0.0125 Departmentdummy D (.013) (.0043) (.0065) Observations 6144 5587 4879 PANEL B:Effect ofletter and department treatment Letterdummy L 0.129 20.0066 0.0005 (.0226) (.0061) (.0102) Treated 0.102 0.0125 0.0123 Departmentdummy D (.0139) (.0054) (.0086) Observations 6144 5587 4879

a. Dependentvariables are individual fair participation (column (1)), TDA enrollment 4.5 monthsand 11 monthsafter the fair (columns (2) and (3)). b. Independentvariable in Panel A is thedepartment treatment dummy D. c.Independent variables in Panel B arethe individual letter dummy L andthe department treatment dummy D. d.All regressionscontrol for the triplet of thedepartment, gender, year of service, age, and salary. e.Standard errors (in parentheses) are corrected for clustering at thedepartment level.

(2) yij 5 a2 1 b2Dj 1 hij.

Theestimates for b1 and b2 arereported on Panel A ofTable IIforfair attendance,(column (1)), and TDAenrollmentafter 4.5 months(column (2)) and 11 months(column (3)). Theseestimates correspondto the difference in fair attendanceand TDAenroll- mentbetween treated and untreateddepartments reported in columns(1) and (4) ofTable I, respectively.The regressions also includeŽ xedeffects for the stratiŽ cation triplet (seeSection III), as wellas controlsfor background variables— gender, year of service,age, and salary.All standard errorsare corrected stan- dard errorsfor clustering at thedepartment level. 14 Being in a treateddepartment increases the probability ofattending thefair by 16.6 percentagepoints. It alsoincreases signiŽ cantly theTDA

14.Adding thetriplet dummies reduces the standard errors, by absorbing someunexplained differences across departments of similar size and prefair TDA enrollmentrates. Baseline covariates are also included to improve the precision of ourestimates. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 829 enrollmentrate by 0.93 and 1.25 percentagepoints (after 4.5 and 11 months). Obtaining signiŽcant differencesbetween these randomly chosengroups means that ourexperiment did havean impact on TDAenrollment.This impact is largein relativeterms (an in- creaseof 24percentand 19 percentin thelikelihood of enrollment after4.5 and 11 months).However, because people update their TDAstatus veryinfrequently, it is smallin absoluteterms (an increaseof only1.25 percentagepoints ofenrollment,on abaseof 34 percent).This effect is tiny comparedwith interventionsthat changethe default rulesfor TDA enrollment(such asin Madrian and Shea[2001] and Choiet al. [2001a, 2001b]) orthat offer individuals theoption of automatically allocating futurepay raisesto TDA contributions[Thaler and Benartzi 2001]. In orderto estimate separately the effect of receiving the letterpersonally and that ofjust being in adepartmentwhere somecolleagues received the letter, we run the following reduced- formregressions:

(3) fij 5 a1 1 m1Lij 1 d1Dj 1 eij, and

(4) yij 5 a2 1 m2Lij 1 d2Dj 1 hij. Theresults of these regressions are reported in PanelB of

TableII. Theparameters m1 and m2 capturethe difference in fair attendanceand TDAenrollmentbetween groups 11 and 10 (col- umns(2) and (3) ofTableI). Theparameters d1 and d2 capture the differencein fair attendanceand TDAenrollmentbetween groups 10 and 00 (columns(3) and (4) ofTable I). Consistentwith the resultsfrom Table I, beingin atreateddepartment increases the probability ofattendanceby 10.2 percentagepoints, and receiving theletter increases it furtherby 12.9 percentagepoints. These resultssuggest that thepromise of the $20 rewarddid havea strongimpact onthe decision to attend thefair. Moreover,the fact that colleaguesreceived the letter also increased one’ s prob- ability ofattending. These peer effects can be explained in two ways.First, an employeewho sees colleagues receiving the in- ducementletter might be reminded of thefair and beled tothink that this is an importantevent (worth rewarding employees for attending) and thusmight decide to attend herself.Second, indi- viduals whoreceive the letter and decideto go to the fair might asktheir colleagues to join them. Our experiment does not allow 830 QUARTERLY JOURNALOF ECONOMICS usto separate these two effects but doesallow usto conclude that socialinteractions play an importantrole in thedecision to attend the fair. Columns(2) and (3) ofTableII showthat receivingthe letter doesnot increase the probability ofenrolling in theTDA (the effectis slightly negativebut insigniŽcant after4.5 monthsand slightly positivebut insigniŽcant as wellafter 11 months),while beingin atreateddepartment does increase the probability of TDAenrollment(by 1.25 and 1.23 percentagepoints after4.5 and 11 months).15 Thenext section presents simple models to inter- pretthese results.

V. ESTIMATINGTHE EFFECTSOF THE EXPERIMENT

V.A.The Model Weposit thefollowing simple speciŽ cation to explain the effectof the experiment on TDA enrollment:

(5) yij 5 a 1 gi fij 1 G z Dj 1 uij. Thisequation states that an individual’s decisionto partici- pate in theTDA is potentially inuenced by theirown attendance at thefair as wellas by whethersome colleagues received induce- mentletters (treatment department dummy D).Theeffect of beingin atreateddepartment could be direct (when many people goto the fair, theircolleagues feel compelled to go to the fair as well,and toenroll in theTDA), channeledthrough conventional peereffects (higher fair attendancein adepartmentleads to higherTDA participation, whichin turnin uences the participa- tionof others), or resulting from the diffusion oftheinformation obtained at thefair. Hereagain, theseeffects cannot be sepa- ratelyidentiŽ ed, and wewill makeno attemptto separate them.

Theindividual fair effect gi mayvary across individuals in oursample, for at leasttwo reasons. First, the effect of attending thefair onTDA participation couldvary across individuals. In particular, ourexperiment induced twodistinct groupsof indi- viduals toattend thefair. Thosewho were in treateddepartments (D 5 1),and thosewho in addition tobeing in atreateddepart- ment,received the inducement letter themselves ( D 5 1, L 5 1).

15.The estimate after 4.5 months is signiŽcant atthe 5 percentlevel while thecoefŽ cient after eleven months has a t-statisticof 1.45. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 831

As wediscuss below,the effect of the fair maybe different for thesetwo groups. Second,it is conceivablethat, even for an individual who would havecome to thefair with noexternalinducement, receiv- ing theletter offering the $20 rewardaffects thefair effective- ness.Because the individual is nowpaid toattend thefair, she mightconvince herself that sheis comingjust for the $20 and thusthat sheis notreally interested in thecontent of the fair. Thistype ofeffectis notstandard ineconomicmodels, but there is substantial evidencein thepsychology literature on the moti- vational consequencesof rewards.This literature is summarized in Rossand Nisbett[1991, pp. 65–67] and morerecently in Frey and Jegen[2001]. Thismotivational reward effect can be captured by assuming that thetreatment effect gi ispotentially (negatively) correlated with theletter treatment Lij.In orderto simplify thepresenta- tion,let us assume that gi takesthe following simple form:

S (6) gi 5 gi 2 nLij,

S where gi (thestandard treatmenteffect component) is indepen- dent of Lij, and n representsthe motivational reward effect. Assuming that nomotivational reward effect amounts to simply assumingthat n 5 0and thus that gi is independentof Lij. Each individual belongsto one of thegroups 11, 10, or00. In orderto deŽ ne treatment effects of fair attendanceon TDA en- rollment,it is usefulto introduce the notion of potential outcomes forfair attendance.For each individual, wedenote by fij(11), fij(10), and fij(00)thefair attendancedecision of individual i, if hehad beenin group11, 10, or00. Obviously,for each individual ij,weobserve only one of the three potential outcomes for fair attendance.As theliterature on differential treatmenteffects has recognized[Imbens and Angrist 1994], in orderto be able to identify parametersof interest, we need to make the following assumption.

ASSUMPTION 1.Monotonicity assumption. For each individual i,

fij(11) $ fij(10) $ fij(00).

Thisassumption states that receivingthe letter can only encouragean individual toattend thefair (and in nocase deter them),and that having one’s colleaguesreceive the letter can also onlyencourage an individual toattend thefair. Thisassumption 832 QUARTERLY JOURNALOF ECONOMICS soundsvery plausible in thesituation we analyze. The Monoto- nicityassumption implies that thepopulation canbe partitioned intofour different types.

First, the nevertakers areindividuals suchthat fij(11) 5 fij(10) 5 fij(00) 5 0.Theseindividuals would notattend regard- lessof the group to which they belong. Second, we deŽ ne the

Žnancial rewardcompliers type as individuals suchthat fij(11) 5 1 . fij(10) 5 fij(00) 5 0.Theseindividuals attend thefair only if theyreceive the letter with theŽ nancial rewardpromise. Third, we deŽne the social interactioncompliers as individuals suchthat fij(11) 5 fij(10) 5 1 . fij(00) 5 0.Theseindividuals would not attend thefair if nobodyin theirdepartment receives the letter, but would attend thefair if theyare in atreateddepartment (whetheror not they themselves receive the letter). Finally, we deŽne the always takers as individuals suchthat fij(11) 5 fij(10) 5 fij(00) 5 1.Theseindividuals attend thefair regardless ofthe group to which they belong. Wemake the following additional assumption.

ASSUMPTION 2.Exclusion restriction assumption. uij is indepen- dent of Lij and Dj.

Theassumption that theerror term uij is independentof the letterassignment status Lij meansthat theletter inviting the employeeto the fair has nodirect effect on TDA participation decisionsof thosewho do not attend thefair (beyond its effecton individual and departmentalfair attendance).Likewise, the fact that otherpeople received the letter is assumedto haveno effect onTDA participation. Toensurethe validity ofAssumption 2,we did notmention TDA in theletters, and theletter did notcontain any mentionof the employee’ s TDAstatus (theletter is repro- duced in theAppendix). 16 It is nowapparent that thereare four parameters of interest in themodel: the average treatment effect for Ž nancial reward compliers E[giufij(11) 2 fij(10) 5 1], theaverage treatment effectfor social interaction compliers E[giufij(10) 2 fij(00) 5 1], thesocial network effect parameter G,and themotivational re- ward effect n.However,our experiment provides us with onlytwo instruments Lij and Dij,makingit impossibleto identify all four

16.A follow-upquestionnaire which contained precise questions about sav- ingsand the TDA didnot have an effect on TDA enrollment(see Du o andSaez [2002b]for details). As a result,it is highly unlikely that the inducement letter couldhave had an effect on TDA enrollment. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 833 parameters.Only if wemake additional assumptionsabout two of thesefour parameters can we estimate the remaining two parame- ters.In thenext subsection we discuss alternativesets of assump- tionsunder which the remaining parameters of the model could beestimated. Our goal is notto claim that any particular set ofassumptions is correct,but ratherto explore the implica- tionsof each assumption, and toprovide bounds tothe different effects.

V.B.Interpretation under Alternative IdentiŽ cation Assumptions If weassume that thereis nomotivationalreward effect ( n 5

0) and gi is equal to g forall individuals, equation(5) reducesto

(7) yij 5 a 1 gfij 1 G z Dj 1 uij. Thisis astandard InstrumentalVariables setup,and bothpa- rameters g and G areidentiŽ ed. They can be obtained by an IV estimationof equation (7), using Dj and Lij as instruments.These estimatesare presented in column(1) in TableIII. Theresults show,as weexpectedfrom Section IV, that thedirect effect of fair attendanceis zerowhile the social effect of being in atreated departmentis positive(and signiŽcant after4.5 months).Being in atreateddepartment increases the probability ofenrollment by 1.8 and 1.2 percentagepoints (after 4.5 and 11 months,respec- tively).Under this setof assumptions, all theeffects of the experi- mentare channeled indirectly through the social effect. If weassume away socialnetwork effects, the parameter G is equalto zero, and equation(5) thenreduces to

(8) yij 5 a 1 gi fij 1 uij. If weassume Ž rstthat thereare no motivational reward effects

(n 5 0), thenan IVregressionof equation (8) using Lij as an instrumentfor fij forthe subsample oftreateddepartments ( Dj 5 1)providesan estimateof theaverage treatment effect of Žnan- 17 cial incentivecompliers, E[giufij(11) 2 fij(10) 5 1]. The esti- matesare reported in column(2) ofTableIII. As weexpected, the averagetreatment effect for Ž nancial incentivescompliers is zero and notsigniŽ cant. Since it is reasonableto assume that thefair doesnot have a negative effecton any individual’s participation

17.Note that the presence of social effects would not bias this estimate as the socialeffect is assumed to be constant within departments in equation(5). 834 QUARTERLY JOURNALOF ECONOMICS

TABLE III IV ESTIMATES OF FAIR ATTENDANCEAND DEPARTMENT EFFECTS ON TDA ENROLLMENT

Assumingno socialeffects

Assuming Effect on Effect on constant Žnancial social treatment incentive interaction effect compliers compliers OLS Na¨•ve IV

(1) (2) (3) (4) (5)

PANEL A:Dependentvariable: TDA participationafter 4.5 months Fairattendance 20.046 20.050 0.117 0.016 20.002 (.0431) (.0429)(.0465) (.0109) (.0255) Treateddepartment 0.018 (.0092) Observations 5587 3726 37551832 5587 PANEL B:Dependentvariable: TDA participationafter 11 months Fairattendance 0.003 0.005 0.1310.049 0.032 (.0681) (.0685)(.0826) (.018) (.0397) Treateddepartment 0.012 (.0147) Observations 4879 3246 32711608 4879 Sample Entire Treated No letter Letter Entire sample departments only only sample

a. Dependentvariables are individual enrollment in the TDA 4.5 monthsand 11 monthsafter the fair. b. Independentvariable are individual fair attendance and department treatment dummy D in column (1). c.Independent variable is individualfair attendance in columns (2) to (5). d.All regressionscontrol for the triplet of thedepartment, gender, year of service, age, and salary. e.Standard errors (in parentheses) are corrected for clustering at thedepartment level. decision,the very small coefŽ cient in column(2) (evenslightly negativeafter 4.5 months)would imply that thetreatment effect is veryclose to zero for all Žnancial rewardcompliers, which seemsunrealistic. This suggests that therewas verylikely a motivationalreward effect associated with receivingthe letter. Theaverage treatment effect for social interaction compliers

E[giufij(10) 2 fij(00) 5 1]canbe obtained byan IVregressionof (8) using Dj as an instrumentfor fij forthe subsample ofindivid- uals with noletter ( Lij 5 0). Column(3) in TableIII presents theseIV estimates,for TDA enrollment4.5 monthsand 11 monthsafter the fair. Theestimates are positive and signiŽcant showingthat attending thefair increasesthe probability ofen- rollingby 11.7 and 13.1 percentagepoints after4.5 and 11 monthsin this sample.The social interaction compliers are THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 835 clearlynot affected by themotivational reward, but maybe sub- jectto peer effects. Therefore, the IV estimatesis an upper bound ofthe direct effect of thefair. Theseeffects are of comparablesize (slightly higher)than thoseestimated by Madrian and Shea [2002] in anonexperimentalsetup. Therefore, the IV estimates suggesta positivetreatment effect on social interaction compli- ers,and noeffecton Ž nancial rewardcompliers. This differential treatmenteffect is plausible. Thosewho attend becauseof the rewardmay be less interested in thefair than thosewho decide to attend becauseof their colleagues. If weassumethat thereare motivational reward effects, then theestimates in column(2) givethe average treatment effect for Žnancial rewardcompliers less the motivational reward effect. Wecannot obtain estimatesof the motivational reward effects unlesswe assume that, in theabsence of motivational reward effects,the treatment effect would beconstant for both groups of compliers.In that case,all thedifference between columns (3) and (2) canbe attributed tomotivational reward effects. Under these assumptions,straightforward computations(see Du o and Saez [2000b]) showthat receivingthe letter reduces the treatment effectof the fair by 63 percentfor TDA participation after4.5 months,and 41 percentfor TDA participation after11 months. It is usefulto compare the effects of fair attendanceon TDA enrollmentof columns (2) and (3) with theOLS effectobtained by regressingTDA enrollmenton fair attendance.The OLS esti- matesare reported in column(4) forthe sample of individuals whoreceived the letter. 18 TheOLS coefŽcient after 11 monthsis positiveand signiŽcant, and would lead theresearcher to con- cludethat thefair increasedparticipation by 4.9 percentage points forthose who attended it.This coefŽ cient, as expected,is biased upward by selectionbias. In column(5) wepresent the “ naive”IV estimatethat uses theletter dummy as an instrument,in thecomplete sample, withouttaking socialeffect into account. This estimate lies be- tweenthe estimates of column (2) and column(3). Thenaive estimatewould underestimatethe overall effect of thefair (since part ofthe “ control”group is actually treated)and overestimate

18.That is the only group wherewe have actual individual fair attendance information.Computing the OLS estimatein the sample of controlswould have beenmore interesting but is unfortunately unfeasible. 836 QUARTERLY JOURNALOF ECONOMICS thedirect effect on those who received the letter. This shows the potentialbias in randomizedtrials that ignoresexternalities. Thedistinction betweendifferential treatmenteffects, social networkeffects, and motivationalreward effects is clearconcep- tually but ourexperiment does not allow usto tell them apart. Thus,it is usefulto describe what type ofalternativeexperimen- tal designs would beneeded to separatethese effects. Differential treatmenteffects arise in oursetting because there is aŽrststage in ourexperiment where individuals decidewhether or not to attend thefair. As aresult,only a self-selectedfraction of indi- viduals attends thefair. Motivational rewardeffects arise be- causeindividuals receivea monetarypayment for attending the fair. Socialnetwork effects could be identiŽ ed with thefollowing experiment.Within a subsample of“treated”departments, a sub- sampleof employees would all automaticallyattend an informa- tionsession targeted to them only (and nottheir colleagues). This couldbe done by makingattendance a jobrequirement for these employees.One could then test whether the TDA participation of nonattendeesin treateddepartments rises relative to that of individuals in untreateddepartments. Motivational rewardef- fectscould be estimated by paying peoplefor attending aninfor- mationsession in asituationwhere everybody is supposed to attend.For example, in manyŽ rms,new hires are often invited to attend informationsessions about beneŽ ts. In somedepartments, this informationsession could be presented as anormalprocess throughwhich all newemployees go. In otherdepartments, at- tending this informationsession could be presented as voluntary, but aŽnancial rewardcould be offered for attendance (large enoughto induce virtually everybodyto attend). If everybody attends in bothcases, the average treatment effect would be expectedto be the same in bothgroups in theabsence of a motivationalreward effect. Evidence of differential treatment effectscould potentially beobtained by using nonmonetaryincen- tivesof various intensity to attend thefair. Forexample, some employeescould be sent a lettersimply remindingthem of the beneŽts fair. Otherscould be sent a morepointed lettertelling themthat importantinformation can be obtained at thefair. One couldalso use emails, personal phone calls, or evenremind them in personto attend thefair. Thesedifferent encouragementde- signs areassociated with different groupsof compliers and may thus allow estimationof differential fair treatmenteffects. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 837

V.C.Additional Evidence In orderto cast furtherlight onourresults, we have divided oursample by sizeof department, pre-experiment TDA partici- pation rate,gender, salary, and yearsof service.These results are reportedon Table IV. Column(1) reportsfair attendanceamong thosewho received the letter (we know the identity ofthosewho attended thefair onlyfor this group).Fair participation was largerin smalldepartments than in largedepartments, and for womenthan formen. In columns(2) and (3) weshow the differ- encein TDAenrollmentbetween treated and controldepartments after4.5 and 11 months,respectively. PanelA showsthat theTDA enrollmenteffects are about the samein largeand in small departments.Panel B showsthat effectsappear tobe stronger in smallerdepartments than in largerones after 11 months.Panel C showsthat theeffects are thesame for men and women.After 4.5 monthsthe treatment effectseems somewhat larger in departmentswhere average salariesare high (PanelD), orfor employees with moreyears of service(Panel E). However,after 11 months,this differenceis actually reversed(in PanelsD and E). Thissuggests that it takes moretime for those in departmentswith lowersalaries or senior- ity toadjust theirTDA participation. Overall,there is noevidence that treatmenteffects are widely different acrossgroups deŽ ned by observablesafter eleven months. However, most of thecoefŽ - cientsin TableIV areimprecisely estimated, and cautionshould betaken in theinterpretation of results. Asmentionedabove,following our experime nt,we sentout afollow-up questionnaireto 917 employeesdesigned tomea- surethe employee s’knowledg eoftheretireme nt beneŽts sys- temin theuniversit y,as wellas questionsto elicit alternativ e retirement savings optionsavailable toemployee sand tomea- surethe extent of procrastination.All theresults are described in detail in Duo and Saez [2002b]. Interestingly,we found that satisfactionwith thefair was signiŽcantly higherfor group11 than forgroup 10. Thisis consistentbothwith dif- ferentialtreatmen teffectsand motivational reward effects. Thequestionn aireresults also show that individuals in group11 donot seem to have a betterknowledge of retirement beneŽ ts than thosein group10, eventhough they are less likely to think that they sufferfrom a lackof information. 838 QUARTERLY JOURNALOF ECONOMICS

TABLE IV FAIR ATTENDANCEAND TREATMENT EFFECT IN DIFFERENT GROUPS

Differencegroup D 5 1 2 group D 5 0

Fairattendance TDA TDA amongletter participation participation recipients after 4.5 after 11 (L 5 1) months months

(1) (2) (3)

PANEL A:DEPARTMENT SIZE Belowmedian (81) 0.328 0.008 0.007 (.015) (.007) (.0104) Observations 985 2797 2403 Abovemedian (81) 0.235 0.007 0.012 (.0132) (.0047) (.0087) Observations 1035 2790 2476 PANEL B:DEPARTMENT AVERAGE PARTICIPATIONIN THETDA BEFORE THEEXPERIMENT Belowmedian (34%) 0.259 0.009 0.018 (.0134) (.0064) (.0098) Observations 1062 2929 2523 Abovemedian (34%) 0.304 0.010 0.008 (.0149) (.0063) (.0089) Observations 958 2658 2356 PANEL C:GENDER Women 0.320 0.011 0.012 (.0134) (.0072) (.0117) Observations 1213 3298 2843 Men 0.221 0.008 0.010 (.0146) (.007) (.0085) Observations 807 2289 2036 PANEL D:SALARY Belowmedian ($34,021) 0.269 0.003 0.018 (.0141) (.0065) (.0093) Observations 983 2745 2291 Abovemedian ($34,021) 0.291 0.015 0.010 (.0141) (.0063) (.0104) Observations 1037 2842 2588 PANEL E:YEARS OFSERVICE Belowmedian (2.84 years) 0.312 0.005 0.015 (.0145) (.0071) (.0115) Observations 1027 2706 2196 Abovemedian (2.84 years) 0.248 0.013 0.010 (.0137) (.0054) (.0083) Observations 993 2881 2683

a. Thesample in column (1) is composedof individuals in group 11 only. b. Incolumns(2) and (3), dependent variables are TDA enrollment 4.5 monthsand 11 monthsafter the fair.Independent variable is departmenttreatment dummy D. c.Regressions in columns (2) and (3) control for the triplet of the department, gender, year of service, age, and salary. d.Standard errors (reported in parentheses below the coefŽ cient) are corrected for clustering at the departmentlevel. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 839

VI. CONCLUSION Small Žnancial incentiveshave successfully induced treated employees,as wellas membersof their peer groups, to attend a beneŽts fair. Moreover,individuals affected by theexperiment, whetherdirectly or indirectly, are more likely to enroll in the TDAafterthe fair. Interestingly,the direct effect is nolarger than theindirect effect: in treateddepartments, those who re- ceivedthe letter and thosewho did notare about as likelyto subsequentlyenroll in theTDA. Weproposed three different interpretationsof this Žnding: differential treatmenteffects, so- cial networkeffects, and motivationalreward effects. Our experi- mentdoes not allow us tounambiguously distinguish thesein- terpretations,which illustrate how the analysis ofa simpleex- perimentin asocialand economiccontext may be substantially morecomplicated than expected. Thesethree different explanations,however, have a common feature.They suggest that an individual’s decisionto participate in theTDA isaffected by small changesin theenvironment, and notonly by theinformation content of thefair. Thestrong social effectsobtained in thefair attendancedecision are not of primary economicinterest per se. However, they are important in an indirectway, as theylead tosigniŽcant changesin thedecision to enrollin theTDA, which is averyimportant economic decision. Thus,social network effects deŽ nitely caused somepeople to take steps whichultimately led themto change their TDA participa- tiondecision. Theincrease in TDAcontributiongenerated by this experi- mentwas muchlarger than its costs.Our program increased participation afterone year by about1.25 percentagepoints for the4000 nonenrolledemployees in treateddepartments relative tocontrol departments. Hence our experiment induced 50 extra employeesto start contributingto the TDA. Assuming that such employeescontributed about $3500 peryear (the average contri- butionof newly enrolled employees), 19 theextra TDA savings generatedby ourexperiment is about$175,000 peryear. If the treatmenteffect persists for many years, the total extra TDA savings couldbe many times that amount.Therefore, the extra

19.After elevenmonths the average yearly contribution of newcontributors is$3500, and that Ž gureis almost identical across groups 11, 10, and 00. New contributorsalso contribute $3500 on average after 4.5 months suggesting that employeesrarely update their contribution levels after they enroll in the plan. 840 QUARTERLY JOURNALOF ECONOMICS savings obtained is withoutdoubt verylarge relative to the in- ducementcost (the rewards distributed amountedto about $12,000).20 However,these effects remain very small compared with changingdefault enrollmentrules [Madrian and Shea2001] or offeringdelayed enrollment,as in the“ SaveMore Tomorrow” program[Thaler and Benartzi 2001]. Thelarge effect of a small rewardon fair attendance,ampliŽ ed by socialeffects, also sug- geststhat individuals donot optimally seekout and process available informationon their own. This suggests that individu- als maynot be giving muchthought to their retirement savings decisions.This has extremelyimportant policy implications for theoptimal design ofretirement plans.

APPENDIX October31, 2000 Name Line 1 Line 2 City statezip Dear Name: Youhave just received your Open Enrollmentpacket from theBeneŽ ts ServicesGroup, inviting youto the BeneŽ ts Fair 2001. TheFair will beheld in twolocations: November7, 11am— 2:30pm ADDRESS ERASED November8, 11am— 2:30pm ADDRESS ERASED Thisyear, as part ofastudy (conductedjointly by theBene- Žts ServicesGroup and economicsresearchers) to better under- stand theimpact ofthe Fair on beneŽts choices,we areoffering a reward of $20 to2,000 employees,just for attending theFair. Funding forthese rewards was contributedfrom a research grant.We selectedthose employees by asimplelottery, and your namewas amongthose drawn. In orderto receive this $20 reward,all youhave to do is to

20.Moreover, this does not take into account potential effects on other beneŽts decisions, which are likely to be impacted by the fair as well. Unfortu- nately,we do not have access to the data on other beneŽ ts to study theseother effects. THE ROLE OFINFORMATIONAND SOCIALINTERACTIONS 841 cometo the Fair with this letter,and giveyour name at the registrationtable that will belocated in themain hall. You will receivea checkwithin thetwo weeks following the Fair. Wehope that youwill Žnd theFair helpful in makingyour beneŽts choices.However, we want toemphasize that thereward is completelyindependent of your beneŽ ts decisions. Makea noteof these dates (November7 orNovember 8) in yourcalendar, and welook forward toseeing you there. Sincerelyyours, Nameof theBeneŽ ts OfŽce AssociateDirector

MASSACHUSETTS INSTITUTE OF TECHNOLOGYAND NBER UNIVERSITY OF CALIFORNIAAT BERKELEY AND NBER

REFERENCES Banerjee,Abhijit, “ ASimpleModel of Herd Behavior,” Quarterly Journalof Economics, CVII(1992), 797– 817. Bayer,Patrick, DouglasBernheim, and Karl Scholz, “ TheEffects of Financial Educationin the Workplace: Evidence from a Survey ofEmployers,”National Bureauof EconomicResearch Working Paper No. 5655, 1996. Bernheim,Douglas, “ ATheoryof Conformity,” Journalof Political Economy, CII (1994),841– 877. Bernheim,Douglas, and Daniel Garrett, “TheDeterminants and Consequences of FinancialEducation in the Workplace: Evidence from a Survey ofHouse- holds,”National Bureau of Economic Research Working Paper No. 5667, 1996, Journalof Public Economics, forthcoming. Bernheim,Douglas, Daniel Garrett, andDean Maki, “ Educationand Saving: The Long-TermEffects of HighSchool Financial Curriculum Mandates,” Journal ofPublicEconomics, LXXXIII(2001), 435– 465. Bertrand,Marianne, , and Erzo Luttmer,“ Network Effects andWelfare Cultures,” Quarterly Journalof Economics, CXV (2000), 1019–1057. Besley,Timothy, and Anne Case, “ Diffusionas aLearningProcess: Evidence from HYVCotton,”Discussion Paper No. 174, RPDS, PrincetonUniversity, 1994. Bikhchandani,Sushil, David Hirshleifer, and Ivo Welch, “ ATheoryof Fads, Fashion,Custom, and Cultural Change as Informational Cascades,” Journal ofPolitical Economy, C(1992),992– 1026. Choi,James, David Laibson, Brigitte Madrian, and Andrew Metrick, “ForBetter orforWorse: Default Effect and401(k) Savings Behavior,” National Bureau ofEconomicResearch Working Paper No. 8651, 2001a. Choi,James, David Laibson, Brigitte Madrian, and Andrew Metrick, “DeŽned ContributionsPensions: Plan Rules, Participants Decisions, and the Path of LeastResistance,” National Bureau of EconomicResearch Working Paper No. 8655,2001b. Duo, Esther,and Emmanuel Saez, “ Participationand Investment Decisions in a RetirementPlan: The In uence of Colleagues’ Choices,” Journalof Public Economics, LXXXV(2002a), 121– 148. Duo, Esther,and Emmanuel Saez, “ TheRole of Information and Social Interac- tionsin Retirement Plan Decisions: Evidence from a RandomizedExperi- ment,”National Bureau of Economic Research Working Paper No. 8885, 2002b. Ellison,Glenn, and Drew Fudenberg,“ Rulesof Thumbs for Social Learning,” Journalof Political Economy, CI(1993),612– 643. 842 QUARTERLY JOURNALOF ECONOMICS

Evans,William, William Oates, and Robert Schwab, “ MeasuringPeer Group Effects:A Modelof Teenage Behavior,” Journalof Political Economy, C (1992),966 – 991. Foster,Andrew, andMark Rosenzweig,“ Learningby Doing and Learning from Others:Human Capital and Technical Change in Agriculture,” Journal of Political Economy, CIII(1995), 1176 –1209. Frey,Bruno, and Reto Jegen, “ MotivationCrowding Theory:A Survey ofEmpiri- calEvidence,” Journalof EconomicSurveys, XV(2001),589 – 611. Glaeser,Edward, Bruce Sacerdote,and Jose ´Scheinkman,“ Crimeand Social Interactions,” Quarterly Journalof Economics, CXI(1996), 507– 548. Glaeser,Edward, Bruce Sacerdote,and Jose ´Scheinkman,“ TheSocial Multiplier,” NationalBureau of EconomicResearch Working Paper No. 9153, 2002. Imbens,Guido, and Joshua Angrist, “ IdentiŽcation and Estimation of Local Av- erageTreatment Effects,” Econometrica, LXII(1994), 467– 476. Katz, LawrenceF., Jeffrey R. Kling,and Jeffrey B. Liebman,“ Movingto Oppor- tunityin Boston:Early Resultsof aRandomizedMobility Experiment,” Quar- terly Journalof Economics, CXVI(2001), 607– 654. Madrian,Brigitte, and Dennis F. Shea,“ ThePower of Suggestion: Inertia in 401(k)Participation and Savings Behavior,” Quarterly Journalof Economics, CXVI(2001), 1149 –1187. Madrian,Brigitte, and Dennis F. Shea,“ Preachingto the Converted and Convert- ingthose Taught: Financial Education in theWorkplace,” Graduate School of Business,University of Chicago,2002. Manski,Charles, “ IdentiŽcation of ExogenousSocial Effects: The Re ection Prob- lem,” Review ofEconomicStudies, LX(1993),531– 542. ——, IdentiŽcation Problemsin the SocialSciences (Cambridge:Harvard Univer- sityPress, 1995). Munshi,Kaivan, “ SocialNorms and Individual Decisions During aPeriodof Change:An Applicationto the Demographic Transition,” University of Penn- sylvania,2000. Poterba,James, Steven Venti, and David Wise, “ HowRetirement Saving Pro- gramsIncrease Saving,” Journalof Economic Perspectives, X(1996),91– 112. Poterba,James, Steven Venti, and David Wise, “ TheTransition to Personal Accountsand Increasing Retirement Wealth: Macro and Micro Evidence,” NationalBureau of EconomicResearch Working Paper No. 8610, 2001. Ross,Lee, and Richard Nisbett, The Person andthe Situation:Perspectives of SocialPsychology (Philadelphia:Temple University Press, 1991). Sacerdote,Bruce, “PeerEffects with Random Assignment: Results for Dartmouth Roommates,” Quarterly Journalof Economics, CXVI(2001), 681– 704. Summers,Lawrence, “ HelpingAmerica to Save More. Remarks by Treasury Secretary LawrenceH. Summers,”Treasury News Press Release LS-524, April4, 2000. Thaler,Richard, and Shlomo Benartzi, “ SaveMore Tomorrow: Using Behavioral Economicsto Increase Employee Saving,” Graduate School of Business, Uni- versityof Chicago,2001.