Do Changes in Pension Incenives Affec Reiremen? A Longiudinal Sudy of Subjecive Reiremen Expecaions February 2001 Sewin Chan Rober F. Wagner School of Public Service New York Universiy sewin.chan@nyu.edu Ann Huff Sevens Deparmen of Economics Yale Universiy ann.sevens@yale.edu The auhors hank Sephen Haider, Ron Miller and seminar paricipans a he Universiy of California, Davis, Universiy of Connecicu, Huner College, Universiy of Maryland, New York Universiy and Universiy of Torono for heir commens and discussions, and graefully acknowledge financial suppor from he Naional Science Foundaion, grans 9905275 and 9907824. Excellen research assisance was provided by Min Qi and Mahew Johnson. A previous version of his paper was circulaed under he ile Reiremen Incenives and Expecaions.
Do Changes in Pension Incenives Affec Reiremen? A Longiudinal Sudy of Subjecive Reiremen Expecaions Absrac This paper invesigaes he responsiveness of individuals reiremen decisions o forwardlooking measures of pension accumulaions. In conras o previous research, we are able o use wihin-person variaion in reiremen incenives and conrol for unobserved heerogeneiy in ases for reiremen ha may be correlaed wih reiremen incenives, by sudying a panel of subjecive reiremen expecaions. We find ha he response o pension incenives esimaed from fixed-effecs regressions is a mos half he magniude of ha esimaed from similar OLS regressions. Our resuls sugges poenially large biases from he failure o conrol for unobserved heerogeneiy in previous esimaes of such relaionships. JEL classificaion: J2 1
I. Inroducion Undersanding he deerminans of reiremen from he labor force is crucial for designing reiremen programs and public policies ha affec older individuals. While here has been much research in his area, virually all of he exising empirical sudies ha use longiudinal daa have relied heavily on cross-secional variaion in pension incenives, and have paid lile aenion o he possibiliy of unobserved heerogeneiy in ases for reiremen ha are correlaed wih boh reiremen behavior and pension incenives. This paper makes use of repeaed observaions of individuals reiremen expecaions o invesigae he effecs of reiremen incenives on hose expecaions, and o evaluae he robusness of he resuls o conrols for unobserved heerogeneiy. We find evidence of subsanial biases in esimaes of he effec of forwardlooking reiremen incenives when here are no conrols for unobserved heerogeneiy. Thus, previous research may have subsanially oversaed he effecs of reiremen incenives on reiremen behavior. The reiremen problem does no lend iself direcly o approaches based on wihin-person variaion because acual reiremen (defined as a permanen exi from he labor force) will be observed only once. Thus, repeaed observaions of acual behavior will provide limied informaion abou responses o changes in reiremen incenives and may be subjec o bias from unobserved heerogeneiy in ases for reiremen. Facors ha deermine he naure of pension plans and Social Securiy benefis, such as employer choice and earnings hisory, are very likely o be correlaed wih labor force aachmen, leading o an oversaemen of he reiremen response o incenives. We are able o esimae he effec of incenives on reiremen using wihin-person variaion by examining a differen dependen variable: individuals subjecive probabiliies of 2
coninuing work beyond age 62, or beyond age 65, from a survey quesion in he Healh and Reiremen Sudy (HRS). The advanage of his measure is ha we observe hese subjecive probabiliies a up o hree differen daes. Thus, we can conrol for any fixed individual characerisics and learn much more abou he updaing of reiremen plans in response o new informaion. Mos longiudinal sudies of he effec of pension incenives on reiremen behavior use cross-secional variaion as he main source of idenificaion and ignore he likely correlaion beween pension incenives and ases for reiremen. Even sudies ha are supposedly based on ime-varying incenives are ofen idenified from he cross-secion since here is lile unanicipaed wihin-person variaion over ime. One excepion is Krueger and Pischke (1992) who invesigaed he labor supply effec of Social Securiy using cohor analysis wihin an aggregae panel daa se. Their source of wihinperson variaion comes from a naural experimen: a subsanial and unanicipaed decline in Social Securiy benefis for hose born beween 1917 and 1921 (he so-called noch generaion), compared wih oherwise idenical individuals who were born before 1917. They find an insignifican relaionship beween Social Securiy wealh and labor force paricipaion for he noch generaion and heir resuls sugges ha previous measures of a large negaive Social Securiy wealh effec were overesimaed. Alhough hey were no able o invesigae he role of privae pensions or oher sources of wealh ha affec he reiremen decision, heir general conclusion is consisen wih findings in his paper. Using more recen longiudinal daa from he HRS, Coile and Gruber (2000) have ried o overcome he reliance on cross-secional variaion in reiremen incenives by exploiing nonlineariies in he Social Securiy program rules as heir main source of variaion in Social 3
Securiy wealh accumulaion. This, hey argue, isolaes a source of variaion in reiremen income accumulaion ha is less subjec o bias from omied unobserved characerisics. However, hey do no consider he analogous poenial correlaion arising from privae pensions. If earnings (and hus Social Securiy benefis) are correlaed wih ases for leisure or reiremen, i also seems likely ha privae pension srucures will be similarly correlaed. Individuals wih a preference for early reiremen are more likely o seek ou and remain wih employers offering generous plans for early reiremen. Thus, heir approach is sill poenially subjec o bias from unobserved heerogeneiy in ases for reiremen ha is correlaed wih reiremen incenives. Since we are no using a naural experimen in our analysis, i is imporan o consider wha migh be driving wihin-person variaion in he incenives o reire. Among workers who are employed by he same firm hroughou heir 50s and 60s, he rules of heir pension eligibiliy and benefis should be known o hem from he very beginning of heir job enure. This assumpion is he basis of he opimizing behavior inheren in forward-looking models of reiremen. This also suggess, however, ha unless new informaion is revealed from one year o he nex, wihin-person variaion in expecaions will no be informaive. Below we describe possible sources of wihin-person variaion, including informaion acquisiion, early reiremen incenives and job separaions, and we show ha our main resuls are robus o a sample resricion o individuals who have experienced an idenifiable and arguably exogenous change in reiremen incenives. This paper conribues o our undersanding of he relaionship beween reiremen incenives and reiremen by providing a crucial cross check of esimaes based on sudying acual reiremen. We demonsrae ha ignoring he likely correlaion beween reiremen incenives and ases for reiremen leads o a subsanial oversaemen of he effecs of pension 4
incenives on reiremen decisions. Since we are using a differen dependen variable reiremen expecaions raher han acual reiremen our findings will no be direcly comparable o he previous lieraure. However, our evidence on he sensiiviy of he incenive esimae o unobserved heerogeneiy will have imporan implicaions for any work ha seeks o undersand he iming of reiremen. The nex secion of he paper discusses in more deail he relaionship beween our approach and ha aken in he recen lieraure on reiremen decisions. Secion III presens our daa and economeric sraegy. Secion IV discusses he sources of wihin-person variaion ha is driving our esimaes. Secion V presens our resuls and he final secion concludes. II. Relaion o Previous Lieraure The lieraure on he deerminans of reiremen has emphasized he imporance of fuure reiremen income accumulaion ha comes from Social Securiy benefi formulas and pension srucures. The behavioral assumpion ha underlies he empirical models is ha individuals decide wheher o reire based on an evaluaion of lifeime uiliy associaed wih curren and fuure reiremen daes. 1 The approach ha we ake in his paper also adops his fundamenal idea. Our esimaing framework is mos direcly relaed o he reduced form implemenaions of he Sock and Wise (1990) opion value model of reiremen, firs discussed in Lumsdaine, Sock and Wise (1992). We briefly summarize heir mehodology and subsequen applicaions here. 1 Esimaed srucural models of reiremen, including Rus (1989), Gusman and Seinmeier (1986), Sock and Wise (1990) and Rus and Phelan (1997), all share his same basic feaure. 5
In Sock and Wise (1990), individuals chose heir reiremen dae, R, o maximize a lifeime uiliy funcion V, of he following form: R 1 s s (1) V ( R) = β U ( Y ) + β U ( B ( R ) s= W s S s= R R s where β is a subjecive discoun facor, U w is uiliy during working years, U R is uiliy while reired, Y is income while working, B is reiremen benefis or income while reired, and S is he final period. Individuals choose heir reiremen dae o maximize he presen value of uiliy from he sreams of income while working and afer reiremen. Sock and Wise implemen he opimizaion underaken here wih a simplificaion of he full dynamic programming rule: in every period individuals compare he lifeime uiliy of reiring oday versus reiremen a a fuure opimal dae, he reiremen dae a which lifeime uiliy is maximized. This difference beween he uiliy from immediae reiremen and delayed opimal reiremen is referred o as he opion value of work or delayed reiremen. Reduced form implemenaions of he model involve calculaing he opion value variable using some assumed uiliy funcion parameer values and esimaing is effec on reiremen. Samwick (1998) applies his mehodology using deailed pension plan informaion from he Survey of Consumer Finances and Pension Provider Surveys. His calculaed opion value closely follows he framework of he Sock and Wise model, and is hus a non-linear funcion of earnings, Social Securiy benefi accruals, and privae pension benefi accruals. The opion value variable is found o be highly significan, as are measures ha focus on he accrual of wealh from working jus one addiional year (as opposed o he opion value measures which consider he gain from working unil some fuure opimal reiremen dae). Levels of pension and oher wealh are no found o be major deerminans of reiremen. Samwick s conribuion is 6
imporan since i uses a daase from a naionally based sample, while Sock and Wise s resuls were based on daa from workers a a single firm. Coile and Gruber (2000) raise a concern abou using opion value and similar measures in a reduced-form framework. They argue ha a problem wih he reduced-form approaches (and o some exen he srucural models as well) is a lack of careful aenion o he sources of idenificaion of he reiremen incenive effecs hey esimae. (Page 9). Coile and Gruber focus heir criicism on he fac ha he opion value combines informaion on privae pensions, Social Securiy benefis, and wages. Mos imporanly, hey noe ha opion value depends criically on an individual s wage, boh because he wage eners direcly as he main par of income while working, and because earnings are an imporan deerminan of Social Securiy benefis. Their primary concern is ha wages are likely o be correlaed wih underlying ases for reiremen, and so using cross-secional variaion in opion value may resul in biased esimaes of he effecs of changes in reiremen incenives on reiremen. Coile and Gruber s alernaive is o develop measures ha capure he forward-looking accumulaion of reiremen wealh or income inheren in he opion value, bu o focus on individual variables ha separaely measure accumulaion of Social Securiy benefis and accumulaion of pension wealh, afer conrolling for a hos of wage and earnings measures. Raher han using a uiliy-based measure such as he Sock and Wise opion value, hey calculae a measure hey call peak value ha is equal o he difference beween oal discouned Social Securiy wealh a is maximum expeced value and is value if reiremen occurs immediaely; a similar peak value measure is calculaed including boh pension and Social Securiy accumulaions. Unsurprisingly, peak value is equivalen o he Sock and Wise opion value under a se of uiliy funcion parameer resricions, as explained in Samwick (1999). Because 7
Social Securiy and pensions are no longer combined wih one anoher and wih wages, as in he uiliy-based opion value measures, Coile and Gruber can conrol separaely for wages, and use nonlineariies in he Social Securiy program rules as heir main sources of variaion in Social Securiy wealh accumulaion. This, hey argue, isolaes a source of variaion in reiremen income accumulaion ha is less subjec o bias from omied unobserved characerisics in he reduced-form esimaing equaions. We share he concerns raised by Coile and Gruber abou he source of variaion in reduced-form implemenaions of he opion value model, bu i is likely ha Coile and Gruber s approach does no go far enough in eliminaing variaion in reiremen incenives ha could be correlaed wih unobservables such as ases for reiremen. As menioned above, if wages are likely o be correlaed wih ases for leisure or reiremen, i also seems likely ha privae pension srucures will be similarly correlaed. Thus, simply using variaion in pension wealh accumulaion may resul in similar problems from omied variable bias. In his paper, we develop an idenificaion sraegy ha conrols for unobserved heerogeneiy ha is poenially correlaed wih boh privae pension incenives and reiremen incenives coming hrough Social Securiy benefis. Our esimaing model is a modificaion of he opion value framework ha akes accoun of he differen dependen variable based on reiremen expecaions. 8
III. Daa and Economeric Mehodology To examine he effecs of reiremen incenives on reiremen expecaions, we use daa from he firs hree waves of he Healh and Reiremen Sudy (HRS), conduced in 1992, 1994 and 1996. The HRS inerviewed individuals aged 51 o 60 as of he firs survey wave in 1992, along wih heir spouses, and colleced deailed informaion on demographics characerisics, employmen, employmen hisory, pensions, asses, income, healh, and subjecive expecaions of various fuure evens. Our dependen variable comes from he following quesion asked of all employed individuals age 62 and younger: Thinking abou work generally and no jus your presen job, wha do you hink are he chances ha you will be working full-ime afer you reach age 62? A similar quesion is asked wih respec o he chances of coninuing work afer age 65. For ease of exposiion, we refer o he age 62 quesions, alhough we have performed many of our analyses using he age 65 quesion as well. We begin by presening some simple descripive saisics on our dependen variables, he subjecive probabiliies of working beyond age 62 or beyond age 65, in Table 1. There is subsanial variaion in individual responses o his quesion, wih a large number of individuals who are cerain ha hey eiher will or will no work beyond he given age, leading o responses ha are clusered a 0 and 1. One-hird of women and one-quarer of men repor ha here is no chance hey will be working full-ime afer hey reach he age of 62, and around half repor a zero probabiliy of working afer age 65. Approximaely 17 percen of women and 23 percen of men respond ha hey are cerain o be working afer age 62, while 6 and 9 percen of women and men are cerain ha hey will be working afer age 65. There is also a smaller clusering of responses around 0.5. 9
Because we are using hese reiremen expecaion responses in a conex in which acual reiremen behavior is he primary concern, i is imporan o relae hese subjecive expecaions o some measure of acual behavior. To do his, we ake responses o he reiremen expecaions from he iniial wave of he survey in 1992 and hen form a sample of hose individuals who have reached age 63 by he ime of he hird wave survey, approximaely four years laer. 2 Some simple condiional mean calculaions make clear ha subjecive reiremen expecaions from he wave 1 survey are relaed o employmen saus a wave 3. Among individuals over he age of 62 by wave 3 who are no longer working full ime (a leas 35 hours per week), he mean of he wave 1 expecaion of work beyond age 62 is 0.43, compared o 0.67 for hose who are working full ime a wave 3. Furhermore, he correlaion coefficien beween wave 1 expecaions and wheher individuals (over age 62) are working full ime a wave 3 is 0.35 for men and 0.34 for women, while he analogous correlaion coefficien based on wave 2 expecaions is significanly higher a 0.60 for men and 0.54 for women. The significan increase in predicive power beween wave 1 and wave 2 remains eviden when we esimae linear probabiliy models for wheher an individual (over age 62) is working full ime a wave 3, conrolling for he wave 1 or wave 2 expecaion as well as age, healh, marial saus and educaion. This is consisen wih new informaion arriving beween waves 1 and 2 ha allows individuals o adjus heir expecaions. As individuals grow closer o he relevan reiremen age, heir expecaions are sronger predicors of heir fuure work or reiremen saus. 2 We have performed he same exercises using individuals aged 66 or over by wave 3, and he expecaions of working beyond age 65. The resuls are similar o hose for he sample aged 63 and over, alhough he sample sizes are much smaller wih he higher age requiremen. 10
The exensive informaion on earnings, employmen, pensions and asses colleced a each survey wave allows us o calculae reiremen incenives among older workers in he HRS. Employmen and pension informaion is colleced a he wave 1 survey on earnings and pensions conneced wih boh curren jobs and jobs ha have ended. Pension deails include eligibiliy, srucure, and benefi amouns on up o hree differen pension plans associaed wih each employer. A each subsequen survey wave, respondens are asked wheher heir employmen and/or pension informaion has changed and wha hose changes are. Even for hose who do no repor a change in pension plans, informaion is again colleced a each survey wave on he pension benefis and eligibiliy rules. We are relying primarily on self-repored pension informaion from he hree survey waves of he HRS o consruc our measures of reiremen incenives. Alhough Pension- Provider and Social Securiy Adminisraion daa mached o he HRS files have been used by researchers o forecas pension accumulaion and Social Securiy wealh 3, he mached employer pension plan file provides deails on pensions a wave 1 only. Our ineres in changes in reiremen incenives means ha we mus use he self-repored longiudinal daa. While concerns have been raised regarding he accuracy and compleeness of self-repored pensions in he HRS, for our purposes, he self-repors mach well wih he subjecive reiremen expecaions since i is individuals undersanding of heir own pensions ha should affec heir expecaions. 3 See, for example, Gusman and Seinmeier (1998), Moore and Michell (1997), and Uccello and Perese (1999). 11
An imporan consequence of our use of pension plan informaion is ha we mus delee from our samples hose individuals who claim o have a pension, bu who do no provide any informaion on he benefi or accoun amouns, or on age of eligibiliy. While his is a maer for some concern, we have few alernaives. Noe ha a similar problem arises wih he mached employer pension daa since a subsanial fracion of individuals refused o auhorize he collecion of pension informaion from heir employers and some auhorized employers failed o respond. Moreover, if an individual is unaware of heir pension eligibiliy, no informaion was colleced regarding heir employer s pension plan. To calculae he Social Securiy benefis o which individuals in our sample are eniled, we make use of he complee Social Securiy covered earnings hisories of individuals in he HRS sample. We calculae each individual s benefi enilemen based on years of covered earnings and he level of earnings in each year. This involves calculaing each individual s Annual Indexed Monhly Earnings (AIME) based on heir highes 35 years of covered earnings and using he Social Securiy program rules o conver he AIME ino annual benefi amouns for alernaive reiremen daes. Since he earnings hisories are only available up o 1991, we use self-repored earnings for subsequen years in which respondens are sill working. 4 We also calculae eligibiliy for spousal benefis, and assume ha individuals will receive he maximum of heir own benefi enilemen, or half of heir spouse s, following Social Securiy program rules. 4 We assume ha earnings remain consan beween one earnings repor and he subsequen repor in he nex wave of he HRS. 12
We nex provide some deails on how he pension and oher daa are used o form our main independen variables of ineres. This requires some discussion of he models o be esimaed. The basic esimaion approach ha we use follows closely he reduced-form implemenaions of Lumsdaine, Sock and Wise (1992), Samwick (1998) and Coile and Gruber (2000) discussed above. In hese models, an indicaor for reiremen is he dependen variable, and he main independen variable of ineres is he opion value of delayed reiremen. Since non-uiliy-based measures of opion value (such as Coile and Gruber s peak value) are nesed in he uiliy-based measures (as in he original Sock and Wise specificaion), we will henceforh refer o opion value broadly defined as encompassing eiher case, and noe ha lifeime uiliy refers o he presen value of pension wealh when opion value is defined wihou uiliy parameers. To disinguish beween he wo measures, we refer o he measure using only pension wealh (no uiliy parameers or oher income informaion) as pension gain. In he sandard reduced-form approach o Sock and Wise, he lifeime uiliy V (R) associaed wih each possible reiremen age R is calculaed for each individual. Opion value OV is calculaed by aking he highes possible lifeime uiliy (associaed wih he opimal reiremen dae R* beween ime and he maximum ime S), and subracing from i he lifeime uiliy available if he individual reires immediaely a ime : (2) OV = V (R * ) V () where R* = argmax { R S)} V (R) The likelihood of reiremen is hen esimaed wih OV as an explanaory variable. 13
Our key explanaory variables differ from he previous sudies in wo ways. Firs, our use of reiremen expecaions as he dependen variable necessiaes a sligh modificaion of his opion value approach. In order o derive a measure of reiremen incenives ha focuses on he pre-/pos-age 62 disincion, we calculae he maximum lifeime pension wealh (or uiliy) from reiring afer age 62, and subrac from i he maximum lifeime pension wealh available from reiring prior o or a age 62: (3) OV62 = V (R * >62 ) V (R * 62 ) where R * 62 = argmax { R 62)} V (R) R * >62 = argmax {62 < R S)} V (R) This reains he concep of maximizing behavior inheren in radiional opion value approaches, bu modifies i o fi he expecaion based dependen variable: OV62 is he opion value from delaying reiremen unil afer age 62. We employ hree differen measures of he opion value of delaying reiremen: (i) pension gain, (ii) pension and Social Securiy benefi gain (iii) opion value based on assumed uiliy funcion parameers. The firs wo measures are analogous o he peak value of Coile and Gruber ha does no require explici assumpions abou he uiliy parameers. In he second measure, we look a he presen value of he sum of pension and Social Securiy benefis, and calculae he gain variable as described above. In calculaing he value of pension wealh a each poenial reiremen age, we discoun fuure income and weigh his income by age-specific survival probabiliies. In he resuls presened below, we use a discoun rae of 3 percen, following boh Samwick (1998) and Coile and Gruber (2000). Survival probabiliies are aken from he Social Securiy Adminisraion s Annual Saisical Supplemen, 1997. 14
15 A second deviaion from previous work arises in our measure of uiliy-based opion value. This measure is based on he Sock and Wise framework, bu we use an alernaive form for he uiliy funcion ha allows saving and borrowing across ime periods. The framework is discussed in Chan and Sevens (1999). We assume ha a ime, individuals maximize a lifeime uiliy funcion of he form: (4) ( ) s s S s s l c u V, = = β subjec o a lifeime budge consrain a ime of: (5) ( ) R y i A c i s S s s s S s s = = + + = + 1 1 1 1 where c is consumpion, l is leisure, S is he las possible dae, β is he subjecive discoun facor, i is he real ineres rae, A is he sock of asses a ime and y s is income in period s, which depends on he reiremen dae R. We assume a consan relaive risk aversion uiliy funcion wih risk aversion parameer 1/α. Furher, we assume ha he uiliy of leisure is 0 while working and equal o a consan U L when reired. Individuals chose he reiremen age from he maximizaion of: (6) max ( ) ( ) L S R s s S s s S s s U R Y A R V = = = + + = β α β β α 1 1 The reiremen age maximizes uiliy, which is a funcion of he presen value of lifeime wealh,
including earnings, asses, and reiremen income, as well as he uiliy of leisure. 5 In previous work (Chan and Sevens, 1999) we have experimened wih alernaive values for he risk aversion and uiliy of leisure parameers, and found ha hey make lile difference o he performance of our opion value measure in reduced-form regressions. Thus, in he work below, we follow Samwick (1998) and Coile and Gruber (2000), in assuming values for he risk aversion and uiliy of leisure parameers, bu do no find resuls o be sensiive o he paricular values assumed. The uiliy of leisure parameer is se so ha on average, he value of an 5 This seup is similar in spiri o ha of Sock and Wise. Their funcional form is power uiliy wih parameer γ and he value of leisure expressed as a muliplicaive facor k on reiremen benefis during reiremen. V R 1 s γ s ( R) = β Y + β ( kb) s= S s= R γ If heir income Y and benefis B are replaced wih consumpion C w before reiremen, and consumpion C r afer reiremen, and log uiliy is used (a limiing case of heir power uiliy, and α=1 in our seup), we ge V R 1 s s ( R) = β ln( Cws ) + β ln( kcrs ) s= S s= R Breaking ou he second erm: V R 1 s s s ( R) = β ln( Cws ) + β ln( Crs ) + β ln( k) s= S s= R S s= R Firs order condiions would have ha consumpion is equal in all periods, i.e., C w =C r, and hus we have our equaion [6] where our U L corresponds o heir ln(k). 16
addiional year of leisure is worh 55 o 65 percen of an addiional year of income from working, depending on he age of reiremen. We assume a risk aversion parameer, α, of 2/3. Our key independen variables of ineres, he pension gain, or opion value from coninuing work beyond ages 62 or 65 are summarized in Appendix Table 1. Noe ha only 48 percen of women and 66 percen of men in he sample repor having any employer-provided pension. This means ha a subsanial porion of he sample has zero values for he pension gain variables. Below we invesigae he sensiiviy of he resuls o focusing only on hose individuals wih privae pension plans. In addiion, some workers repor having a pension from a previous employer raher han heir curren employer, and so will have a zero gain o remaining employed wih heir presen employer. Among hose wih pensions, here are large gains and losses associaed wih working beyond a paricular age. For individuals wih a pension eligibiliy age ha is afer age 62, he pension gain from working beyond age 62 can be subsanial. The average gain among women wih posiive pension gain values is $71,179, and for men his average of posiive values is $164,435. For individuals eligible o receive pension benefis a or before age 62, here can also be subsanial financial penalies o remaining employed beyond he wealh-maximizing age. Among women wih negaive pension gain values, he average is $58,577, while for men he corresponding average is $95,543. Our basic esimaing equaion can be expressed as: (7) Yi = X i β + δ OV 62i + α i + ε i where Y is he subjecive expecaion of working beyond age 62, and OV62 is he measure of reiremen incenives, opion value or pension gain. X includes some sandard demographic conrols, such as age, educaion, healh saus and marial saus, and in mos specificaions, we include earnings and asse measures. Our concern is ha he individual fixed-effec, α i, may be 17
correlaed wih he key independen variable OV62 and so OLS esimaion of equaion (7) will lead o biased esimaes of δ. To address his concern we esimae fixed-effecs models ha accoun for his permanen individual componen of he error erm. IV. Wihin-Person Variaion in Reiremen Incenives The fixed-effecs specificaion direcly focuses aenion on how he wihin-person variaion in opion value is generaed. We believe ha one imporan source of variaion in repored pension deails across years for a given individual is he acquisiion of informaion abou pension plans. Gusman and Seinmeier (1999) have found ha pension beneficiaries in he HRS do no know he deails of heir pension plans very well, bu ha his knowledge seems o improve somewha wih age. Because hese changes in self-repored pension eligibiliy and benefis do no represen acual changes in pension plans, i migh be emping o dismiss his source of variaion as noise. However, for our purposes, his is acually an imporan source of informaion since i is he pension incenive an individual believes o be available ha should influence his or her reiremen plans in a given year, no he acual benefis ha may be known only by an employer or he Social Securiy Adminisraion. The purpose of making his disincion is simply ha changes across waves in self-repored pension informaion need no reflec acual changes in pension plans o provide meaningful variaion in our esimaions. Wihin-person variaion in reiremen incenives may also come from acual changes in pension plans. In our esimaions below, we repor resuls ha focus exclusively on hose individuals experiencing some observable evens ha are likely o generae changes in individuals pension incenives, including involunary job changes, early reiremen incenives, 18
and oher discree changes in employer-sponsored pension plans. This is inended o focus on individuals for whom here is a relaively clear source of wihin-person variaion in he pension gain or opion value variables. Addiionally, i seems likely ha acquisiion of informaion abou a pension plan may be endogenous o he reiremen decision. Individuals who have decided o reire in he near fuure may be mos likely o updae heir knowledge of heir pension deails. Focusing on a sample of individuals wih specific changes in heir employmen or pension saus will reduce he reliance on (poenially endogenous) informaion acquisiion as a source of variaion. Specifically, in some of he resuls below, we limi he sample o individuals who mee one of hree crieria: (i) repor having involunarily los a job (due o a plan closing or layoff) beween waves 1 and 3, (ii) repor having received a special early reiremen incenive a one of he survey waves, or (iii) repor ha somehing abou heir pension plan has changed from one survey wave o he nex. Noe ha we do no include individuals who have experienced volunary job changes in his group. These individuals are likely o have significan changes in heir pension incenives, bu his may be a resul of heir planned reiremen behavior. As noed by Ruhm (1990), many older workers leave heir career job for bridge jobs before compleely reiring. For such people, he pension gain variable may be deermined by a prior decision o reire in he near fuure. Job loss will produce a change in pension incenives by, a a minimum, changing a pension from being associaed wih a curren job o one ha is associaed wih a previous job. This means ha a decision o reire from a pos-displacemen job does no have any impac on pension income associaed wih a pre-displacemen job. In previous work (Chan and Sevens, 2001) we show ha job loss among older workers ypically resuls in workers mainaining some 19
pension benefis, and ha hey ofen begin receiving pension benefis immediaely as a form of severance pay, regardless of heir re-employmen saus. Early reiremen incenives may emporarily aler an individual s expeced pension benefis in a number of ways such as reducing he minimum age for pension eligibiliy, or alering he annual benefi amoun. Because we have informaion on when hese early incenive plans were offered, we can disinguish beween periods when he increased pension incenives were and were no available. Finally, individuals who remain wih he same employer across survey waves are asked direcly wheher some aspec of heir pension plan rules has changed since he previous wave. For individuals who answer yes o his quesion, we again can have more confidence ha here is some rue wihin-person variaion in he pension incenives ha hey face. 6 The wihin-person variance of he pension gain measure in his change sample is approximaely 2.5 imes larger han he wihin-person variance for he sample of individuals who do no experience an easily idenifiable shock o heir pension incenives. Esimaion based on his resriced sample produces resuls very similar o esimaion using he full sample. 6 Noe ha, even if individuals do no repor a change in heir pension rules, hey are again asked for informaion regarding benefi amouns, eligibiliy ages, accoun balances, ec.. Thus, here are someimes significan changes in repored pension deails even among individuals who claim ha nohing has changed. Our focus on individuals who repor an explici change is mean o focus on a sample ha we view as more likely o have experienced a well-defined change in heir pension plans across survey waves. 20
V. Resuls We nex begin o examine he relaionship beween workers subjecive reiremen expecaions and he pension gain or opion value. In Table 2, we repor OLS and fixed-effecs esimaes of he expecaion of work beyond age 62 on he pension gain measure, esimaed separaely for men and women. Sandard errors are correced for clusering of observaions from he same individuals, and for possible heeroscedasiciy arising from informaion acquisiion over ime. 7 The variable of ineres here is he pension gain variable. The coefficiens on pension gain in he OLS regressions are posiive and significan for boh men and women. This is consisen wih much of he previous lieraure, showing ha pension incenives significanly affec reiremen decisions. Larger gains o coninuing work beyond age 62 significanly increase individuals saed expecaions of working beyond age 62. To inerpre he magniude of he pension gain coefficien, consider an individual wih he average privae pension plan ha encourages delaying reiremen beyond age 62, and replace ha pension wih anoher ha has a zero pension gain. This scenario reflecs one possible effec of he swich in populariy from defined benefi o defined conribuion pension plans ha has occurred in recen years. The coefficiens from he firs and hird columns of Table 2 imply ha a woman wih he average posiive pension gain ($71,179) would have a 0.06 lower expecaion of working afer age 62 once her pension is changed o having zero pension gain. A man wih he average posiive pension gain ($164,435) is prediced o have an expecaion ha is 7 There is also a poenial problem arising from he clusering of he expecaion responses a zero, 0.5 and one. However, when we esimae he models using a sub-sample of individuals wih responses oher han zero, 0.5 and one, we obain similar resuls. 21
approximaely 0.10 lower once his pension is changed o having zero pension gain. Noe ha he larger effec for men is driven enirely by he larger average value of he pension gain, while he esimaed coefficien on pension gain is smaller for men han for women. These regressions also include conrols for age, healh, educaion, marial saus, race, ehniciy, as well as annual earnings and oal wealh. Mos of he coefficiens from he OLS regressions have he expeced signs. The coefficien on age is negaive and significan, bu he squared erm is posiive. The expecaion of working afer age 62 is iniially declining wih age, bu begins o increase wih age afer age 57 or 58. Poor healh reduces he expecaion of working beyond age 62 for men and women. There are some small differences in reiremen expecaions by educaion level. Black women are significanly less likely o expec work beyond age 62. Being married decreases a woman s expeced probabiliy of coninuing work beyond age 62 by 0.16, bu has no effec for men. This is no surprising since married women ypically have significan sources of labor and reiremen income hrough heir husband s earnings, pensions, and Social Securiy benefis. We have also included annual earnings and oal non-pension asses o accoun for differences among workers ha migh be correlaed wih heir pension wealh and wih reiremen plans. Gruber and Coile (2000) argue ha many opion value measures are highly correlaed wih earnings, and he same may apply o wealh. In he OLS resuls, hese variables have he expeced signs, alhough he asse measures are generally insignifican. Higher earnings increase individuals expecaion of coninuing work ino heir 60s, while increased wealh reduces hese expecaions. To address he problem of omied variable bias, he second and fourh columns of Table 2 repea he analysis using a fixed-effecs esimaor. If here are imporan omied variables such 22
as ases for leisure ha are negaively correlaed wih boh he probabiliy of working beyond age 62 and wih he pension gain variable, we would expec he fixed-effecs esimaes o resul in a smaller coefficien on he pension gain variable. Indeed, here are sharp reducions in he magniude of he pension gain coefficien in he fixed-effecs specificaions. For boh men and women, he pension gain effec esimaed from wihin-person changes is less han half he magniude of he OLS coefficien. The fixed-effecs coefficiens imply ha a woman wih he average posiive pension gain value would have an expecaion of working afer age 62 ha is 0.02 lower if her pension changed o having zero pension gain, and ha a man under he same circumsances would have an expecaion 0.04 lower. Recall ha under he OLS esimaes, hese pension changes would have induced a much larger decrease in expecaion (0.06 for women and 0.10 for men). The difference in expecaion change is consisen wih a srong correlaion beween unobserved fixed-effecs and he pension gain measure and suggess ha cross-secional esimaes of he relaionship beween reiremen and reiremen wealh or accumulaion measures may be subjec o significan bias. A Hausman es allows us o rejec he hypohesis ha he person-specific effecs are independen of he regressors. The second half of he able repeas he OLS and fixed-effecs regressions using he variable for he expecaion of working beyond age 65, and he equivalenly defined pension gain variables. Looking firs a he OLS resuls, he esimaed magniudes of he coefficiens on he pension gain from working pas age 65 are somewha smaller han hose for age 62. The fixedeffecs esimaes using he age 65 variables produce dramaically smaller coefficiens on he pension gain variable compared wih he age 62 resuls. One possible explanaion for his paern is ha relaively few privae pensions have peaks afer age 65. Thus, here is less variaion in he 23
gain o working beyond age 65 han in he gain o working beyond age 62. In mos of he resuls ha follow, we focus on he variables for working beyond age 62, bu noe ha here may be imporan differences in pension incenive effecs depending on he specific reiremen ages being considered. 8 Because he quesions abou working beyond age 62 or 65 do no differeniae beween an individual s desire o coninue working, and heir abiliy o coninue working, we have also esimaed he basic models including some addiional conrols. If individuals anicipae losing a job or suffering diminished healh over he nex few years, hey may have lower expecaions of working in he fuure for reasons ha have lile o do wih financial incenives for reiremen. To es for such a possibiliy, we have used addiional expecaions daa from he HRS surveys. A each wave, individuals were also asked o provide heir expecaion of losing heir curren job, heir expecaions of being able o find a similar job (in he even of a job loss), and heir expecaion ha healh would limi heir abiliy o work in he nex several years. While he expecaion of losing he curren job was never significan, he expecaion of easily finding a similar job was consisenly posiive and significan. This laer variable may be reflecing a high valuaion of one s human capial and aachmen o he labor force. Surprisingly, he coefficien on he expecaion of facing a healh limiaion in he fuure was posiive and significan. Including responses o hese quesions as addiional conrols, however, had lile effec on he esimaed coefficiens on he pension gain variables. 8 A similar poin is made by Coile and Gruber (1999), who noe ha he effecs of Social Securiy and pension incenives are much sronger beween ages 62 and 69 han beween ages 55 and 61. 24
Some addiional specificaions of he basic OLS and fixed-effecs regressions are summarized in Table 3. We repea he analysis using boh he sum of privae pension and Social Securiy wealh and using he uiliy-based opion value variable (described above) ha combines pension, Social Securiy, earnings, and asse informaion, along wih assumed uiliy funcion variables. We believe i is imporan o incorporae Social Securiy measures ino he incenives measures, alhough here are some caveas necessary for hese resuls. Firs, he pension and Social Securiy measures ha are combined in Table 3 are no compleely comparable, since he Social Securiy measures represen acual benefi amouns and eligibiliy, and he pension measures represen individuals undersanding of heir pension benefis and eligibiliy. 9 Second, he sources of wihin-person variaion in he Social Securiy measures are no necessarily he same as for he pension daa. The main reason why individuals Social Securiy enilemen changes across survey waves is ha earnings change across he waves. We conrol separaely for earnings, so he change in Social Securiy benefis reflecs he non-linear way in which earnings ranslae ino benefis. However, if hese earnings changes reflec eiher volunary job changes or measuremen error, he fixed-effecs approach may no produce an unbiased esimae. 10 In shor, he idenificaion sraegy used here is likely o be beer suied o he self-repored pension incenives han o he Social Securiy measures. 9 Alhough he HRS also conains subjecive esimaes of he magniude of fuure Social Securiy benefis a he subjecive expeced dae of firs receip, i was no clear how hese should be adjused for differen reiremen daes. 10 Recall ha we have adminisraive earnings repors wih which o compue Social Securiy benefis only hrough 1991; for subsequen years we use self-repored annual earnings. 25
The coefficiens on he pension plus Social Securiy measures are similar o hose using pension informaion only. We have also esimaed hese models using only he Social Securiy gain measure (and ignoring privae pensions enirely). The coefficiens on he Social Securiy gain variable are wrong-signed and someimes significan in he OLS regressions, and are no saisically significan in he fixed-effecs regressions. This is consisen wih here being oo lile wihin-person variaion in he Social Securiy gain variable for he fixed-effecs esimaes o be informaive. I is worh noing ha Coile and Gruber (1999) found no saisically significan relaionship beween reiremen and he peak value defined using Social Securiy benefis when hey resriced heir sample o hose under 61, as is consisen wih our sample here. This may reflec he fac ha Social Securiy becomes more imporan o reiremen decisions once individuals reach he age of eligibiliy in heir 60s. The opion value measures based on uiliy comparisons are similar o hose based on pension gain. The fixed-effecs esimaes are posiive and saisically significan, and are subsanially below he corresponding OLS esimaes. To provide a meaningful inerpreaion, consider he previous example of replacing an individual s pension wih he average incenive o coninuing work pas age 62 wih a pension ha has zero incenive. The fixed-effecs coefficien on opion value for men suggess ha his individual would have an expecaion ha is 0.01 lower when he no longer has he posiive pension incenive while keeping oher componens of he opion value measure he same. Decreasing an individual s opion value by one sandard deviaion would lower he expecaion of working beyond age 62 by 0.03. In Table 4, we repor resuls from wo alernaive samples of men and women, focusing again on he pension gain. Firs, we use a sample ha is resriced o individuals who repor having a privae, employer-provided pension. (The full sample resuls are repeaed in he firs se 26
of columns for comparison.) This follows Samwick (1998) and ess wheher here is a disincion beween he effec of having no pension versus any pension, and having a more generous or more seeply sloped pension wealh-reiremen age profile. Samwick found ha a sample resriced o individuals wih pensions produced slighly sronger effecs of opion value on reiremen. He inerpreed his finding as evidence ha he pension effec was no being driven enirely by he conras beween hose wih pensions and hose wihou and our resuls confirm his finding. The OLS and fixed-effecs esimaion using he sample of hose wih pensions (mos comparable o Samwick s specificaion) produce esimaes of pension gain effecs ha are very close o hose based on he full sample. The similariy of he wihin-person esimaes across he wo samples is no surprising since very few individuals go from reporing a pension o reporing none in a subsequen wave. The second new se of resuls in Table 4 is based on a sample limied o individuals wih a well-defined change in heir pension srucures or incenives beween survey waves. As menioned above, his is inended o focus on individuals for whom here is a relaively clear source of wihin-person variaion in he pension gain or opion value variables. The regression resuls for his sample are consisen wih resuls from he broader samples. For men, he pension gain coefficiens are similar in magniude o he full sample and show significan reducions in moving from he OLS o he fixed-effecs esimaes. The pension gain coefficien from he fixed-effecs regression is approximaely half he size of he OLS coefficien. For women he OLS and fixed-effecs coefficiens are somewha smaller han in he full sample and he fixed-effecs coefficien on pension gain is no longer saisically significan, perhaps because he sample size is so dramaically reduced. Overall our resuls, paricularly for men, are no very 27
sensiive o hese sample changes which may focus on more exogenous sources of wihin-person variaion. Several previous auhors have invesigaed wheher levels of reiremen wealh (boh pension and non-pension) significanly affec reiremen behavior, afer condiioning on variables capuring he fuure accumulaion of pension wealh such as opion value or pension gain. If individuals are liquidiy consrained or have relaively low wealh holdings, hey may be consrained o coninue working. In Table 5, we es wheher variables for low pension wealh and low non-pension wealh affec reiremen plans. In paricular, we include a variable equal o one if oal asse are less han $10,000, and a variable indicaing ha he value of pension wealh (assuming reiremen a age 62) is $50,000 or less. 11 The resuls in Table 5 sugges ha low levels of wealh may be imporan deerminans of reiremen planning. While he magniude of coefficiens on boh low asses and low levels of reiremen wealh are reduced by including fixed effecs in he regression, hey remain posiive and are usually saisically significan. Having very low asse holdings increases he likelihood of coninuing work beyond age 62 by 0.05. This is slighly larger han he effec of a change from a zero pension incenive o one he average posiive pension incenive. For women, low pension wealh also increases he expecaion of working beyond age 62. Inclusion of hese variables slighly decreases he magniudes of he pension gain coefficiens, bu he differences are no saisically significan. The finding ha levels of pension and non-pension wealh are relaed o reiremen expecaions differs from recen findings by Samwick (1998). Samwick 11 These resuls are no sensiive o changes in he definiion of low asses. Similar resuls are obained if we use alernaive definiions of low asse holdings up o $30,000. 28
noes in his conclusion ha i is only he change in pension wealh (hrough he opion value) and no he level ha affecs he decision o reire. While we do no focus on acual reiremen, bu on reiremen expecaions, our findings sugges ha wealh levels may be imporan. Individuals wih low levels of wealh are more likely o plan o coninue working beyond age 62. Furher, hese effecs remain afer conrolling for individual fixed effecs. While opion value or pension gain variables are imporan, having adequae levels of wealh also appears o be crucial o reiremen iming. Coile and Gruber (2000) also provide evidence ha wealh levels may be imporan, even afer conrolling for reiremen wealh accumulaion. We have esimaed several addiional specificaions as robusness checks on our main resuls. Firs, because many individuals in heir early 50s may no have begun o hink in grea deail abou reiremen, we have resriced he sample o include only hose over he age of 55. Resuls for his group are very similar o he resuls in Table 2. Second, our calculaion of he pension gain and opion value variables use survival probabiliies ha differ only by age and gender. Heerogeneiy across individuals in healh or life expecancy may lead some individuals o more heavily discoun fuure pension reurns han ohers. To accoun for his, we have included an addiional variable in he regressions, indicaing an individual s subjecive expecaion of living beyond age 75, as well as an ineracion beween his variable and he pension gain or opion value measure. None of he ineracion erms are saisically significan, and inclusion of he life expecancy variables does no change he coefficiens on our main variables of ineres. Finally, because single and married women may have very differen plans and resources available for reiremen, we have esimaed he resuls for women separaely by marial saus. The effecs of measured individual reiremen incenives are similar for married and unmarried women. 29