econstor Make Your Publications Visible.

Transcription

1 econstor Make Your Publcatons Vsble. A Servce of Wrtschaft Centre zbwlebnz-informatonszentrum Economcs Kuhn, Peter; Skuterud, Mkal Workng Paper Internet Job Search and Unemployment Duratons IZA Dscusson paper seres, No. 613 Provded n Cooperaton wth: IZA Insttute of Labor Economcs Suggested Ctaton: Kuhn, Peter; Skuterud, Mkal (2002) : Internet Job Search and Unemployment Duratons, IZA Dscusson paper seres, No. 613, Insttute for the Study of Labor (IZA), Bonn Ths Verson s avalable at: Standard-Nutzungsbedngungen: De Dokumente auf EconStor dürfen zu egenen wssenschaftlchen Zwecken und zum Prvatgebrauch gespechert und kopert werden. Se dürfen de Dokumente ncht für öffentlche oder kommerzelle Zwecke vervelfältgen, öffentlch ausstellen, öffentlch zugänglch machen, vertreben oder anderwetg nutzen. Sofern de Verfasser de Dokumente unter Open-Content-Lzenzen (nsbesondere CC-Lzenzen) zur Verfügung gestellt haben sollten, gelten abwechend von desen Nutzungsbedngungen de n der dort genannten Lzenz gewährten Nutzungsrechte. Terms of use: Documents n EconStor may be saved and coped for your personal and scholarly purposes. You are not to copy documents for publc or commercal purposes, to exhbt the documents publcly, to make them publcly avalable on the nternet, or to dstrbute or otherwse use the documents n publc. If the documents have been made avalable under an Open Content Lcence (especally Creatve Commons Lcences), you may exercse further usage rghts as specfed n the ndcated lcence.

2 DISCUSSION PAPER SERIES IZA DP No. 613 Internet Job Search and Unemployment Duratons Peter Kuhn Mkal Skuterud October 2002 Forschungsnsttut zur Zukunft der Arbet Insttute for the Study of Labor

3 Internet Job Search and Unemployment Duratons Peter Kuhn Unversty of Calforna, Santa Barbara and IZA Bonn Mkal Skuterud Statstcs Canada, Ottawa Dscusson Paper No. 613 October 2002 IZA P.O. Box 7240 D Bonn Germany Tel.: Fax: Emal: Ths Dscusson Paper s ssued wthn the framework of IZA s research area The Future of Labor. Any opnons expressed here are those of the author(s) and not those of the nsttute. Research dssemnated by IZA may nclude vews on polcy, but the nsttute tself takes no nsttutonal polcy postons. The Insttute for the Study of Labor (IZA) n Bonn s a local and vrtual nternatonal research center and a place of communcaton between scence, poltcs and busness. IZA s an ndependent, nonproft lmted lablty company (Gesellschaft mt beschränkter Haftung) supported by the Deutsche Post AG. The center s assocated wth the Unversty of Bonn and offers a stmulatng research envronment through ts research networks, research support, and vstors and doctoral programs. IZA engages n () orgnal and nternatonally compettve research n all felds of labor economcs, () development of polcy concepts, and () dssemnaton of research results and concepts to the nterested publc. The current research program deals wth (1) moblty and flexblty of labor, (2) nternatonalzaton of labor markets, (3) welfare state and labor market, (4) labor markets n transton countres, (5) the future of labor, (6) evaluaton of labor market polces and projects and (7) general labor economcs. IZA Dscusson Papers often represent prelmnary work and are crculated to encourage dscusson. Ctaton of such a paper should account for ts provsonal character. A revsed verson may be avalable on the IZA webste ( or drectly from the author.

4 IZA Dscusson Paper No. 613 October 2002 ABSTRACT Internet Job Search and Unemployment Duratons After decades of stablty, the technologes used by workers to locate new jobs began to change rapdly wth the dffuson of nternet access n the late 1990 s. Whch types of persons ncorporated the nternet nto ther job search strategy, and dd searchng for work on lne help these workers fnd new jobs faster? We address these questons usng measures of nternet job search derved from the December 1998 and August 2000 CPS Computer and Internet Supplements, matched wth job search outcomes from subsequent CPS fles. We fnd that nternet searchers are postvely selected on observables, but negatvely selected on unobservables. A benefcal (unemployment-duraton reducng) causal effect of nternet job search s consstent wth our estmates only f negatve selecton on unobservables s especally strong, n other words only f the populaton of on-lne resumes s strongly adversely selected. JEL Classfcaton: J64 Keywords: unemployment, duraton, hazard models, nternet, job search Correspondng author: Peter Kuhn Department of Economcs Unversty of Calforna Santa Barbara CA USA Tel.: Fax: Emal: pjkuhn@econ.ucsb.edu

5 Usng CareerBulder to fnd a job s lke drvng n the carpool lane. -half-page ad for an nternet job ste n the Los Angeles Tmes, Frday March 1, (p. C5) Thnk Monster for the best resumes, the best canddates. -Monster.com web ste, Sept. 19, Introducton After decades of stablty, the technologes used by workers to locate new jobs began to change rapdly wth the dffuson of nternet access n the late 1990 s. As early as August 2000, one n four unemployed U.S. jobseekers reported that they regularly used the nternet to look for jobs; one n ten employed persons sad they regularly looked for other jobs on lne. The use of nternet job and recrutng stes s generally free of cost for workers and much cheaper for frms than tradtonal prnt advertsements. In addton, these servces offer frms and workers the promse of nstant access to a much larger number of possble matches than tradtonal channels, as well as the potental for the exchange of much more detaled nformaton about both worker and job attrbutes. 1 Not surprsngly, economsts have begun to speculate on the potental effects of the above developments on labor markets. For example, commentators have argued that the hgher contact rate, lower cost, and greater nformaton content provded by ths technology could lead to lower frctonal unemployment (Mortensen 2000), hgher average match qualty (Krueger 2000), a reducton of noncompettve wage dfferentals (Autor 2001), and an amplfcaton of ablty-related wage dfferentals (Kuhn 2000). If 1 For example, at frms request WebHre wll check the followng worker credentals: socal securty numbers; current and prevous addresses; references; educaton; crmnal, cvl and bankruptcy court records; drvng and credt reports; and workers compensaton clams. Also offered are on-lne sklls and personalty testng. The combnaton of nternet applcaton procedures and tradtonal database management software also dramatcally smplfes the process of searchng through submtted resumes for approprate matches. Fnally, workers can now gan much more nformaton about workng condtons and job requrements from job boards as well as company webstes.

6 2 even some of these clams are correct, the advent of nternet job search wll have mportant mplcatons for both labor- and macroeconomc polcy. 2 Ths artcle has two man goals, the frst of whch s to understand the process by whch jobseekers choose to use nternet methods to look for work. Second, we am to estmate the causal effect, for an ndvdual worker, of ncorporatng the nternet nto hs or her job search strategy. We are of course well aware that, even f nternet search has prvate, ndvdual benefts, t does not follow that the equlbrum effects of ntroducng ths technology on unemployment rates, wages and other outcomes are socally benefcal. 3 However, snce n most equlbrum models, some frst-order, or prvate effects are a necessary condton for any general equlbrum effect to occur, the questons posed n ths paper seem to be the rght ones to ask frst. In order to answer our questons we use measures of nternet search derved from the December 1998 and August 2000 CPS Computer and Internet Use Supplements, matched wth job search outcomes from all subsequent CPS fles that contan some of the same survey respondents. Throughout our analyss we focus on the search methods and outcomes of unemployed persons only. Ths s because the regular monthly CPS does not collect data on non-nternet job search by employed persons. 4 Thus, for those wth jobs, CPS data does not allow one to dstngush nternet job search actvty from the decson to look for work at all. We also restrct our attenton to one partcular outcome of the job search process jobless duraton. In part, ths s drven by data consderatons: n the CPS, job qualty (.e. wage) nformaton s not avalable for a suffcent sample of 2 One potentally relevant aspect of labor market polcy s the ratonale for government-provded job matchng servces such as the states Employment Servces. Macro polcy mplcatons could follow from any change n the natural unemployment rate caused by nternet job search technology. 3 For example, Lang (2000) has suggested an asymmetrc-nformaton model n whch a reducton n the costs of applyng to jobs can be Pareto-worsenng, n part by reducng the average match qualty n every frm s applcant pool.

7 3 jobseekers. 5 For many polcy purposes, however, unemployment duratons are the outcome of most drect nterest, justfyng our focus here. Ths paper contrbutes to an emergng lterature on the effects of nternet technology on product market performance (e.g. Brown and Goolsbee 2002 n lfe nsurance markets; Brynjolfsson and Smth 2000 on book and CD markets, and Carlton and Chevaler 2001 on varous consumer goods); to our knowledge ours s the only study of the effects of nternet technology on the functonng of the labor market. The current paper also contrbutes to an older lterature on the relatve effectveness of dfferent job search methods. For example, Holzer (1987, 1988), Bortnck and Ports (1992), Osberg (1993) and Addson and Portugal (2001) compare the job-fndng rates of unemployed workers usng a varety of search methods. Thomas (1997) focuses specfcally on the effectveness of publc employment agences. Fnally our work also relates to a recent lterature on the dgtal dvde, whch asks whether dfferental access to computer or nternet technology aggravates nequalty along varous dmensons (e.g. Farle 2001); as we dscuss below, at least some of our results regardng selecton nto nternet job search are surprsng n the lght of ths lterature. In our data, smple means ndcate that nternet job searchers are more lkely to be employed one year after ther search methods are observed than are other unemployed workers. However t s also the case that nternet job searchers are better-educated, prevously worked n occupatons wth lower unemployment rates, and had several other characterstcs whch are generally assocated wth shorter unemployment duratons n our sample as well as others. Once these observable dfferences are held constant, we 4 See Skuterud (2001) for a recent analyss of trends n on-the-job search usng the occasonal CPS surveys that do collect ths nformaton.

8 4 fnd no dfference n employment rates twelve months later. Further adjustng our estmates (a) to ncorporate all the avalable nformaton n our sample on unemployment duratons, and (b) for length-based samplng (Lancaster 1979), yelds estmated nternet job search effects that are counterproductve,.e. nternet job search appears to lengthen unemployment spells. Fnally, we develop and add to the above model an nstrumental-varables-type technque to adjust for endogenous selecton nto nternet search on unobservables. To our knowledge ths s the frst applcaton to smultaneously model selecton nto a treatment (nternet search) and the effects of the treatment n a duraton model, allowng for a contnuous jont dstrbuton of the unobservables n the two equatons. 6 Whle ths model cannot unequvocally rule out a benefcal causal effect of nternet job search on unemployment duratons, t does mply --usng a zero effect of nternet search as a lower bound on the true effect-- a statstcally sgnfcant level of negatve selecton on unobservables; n other words, nternet job searchers are less lkely than observatonallydentcal unemployed workers to be re-employed regardless of the search methods they use. As a consequence, a sgnfcant benefcal effect of nternet search would be consstent wth our data only f selecton nto nternet search were even more negatve than ths baselne level. Internet search frms who smultaneously clam to employers that ther applcants are postvely selected (on hard-to-document characterstcs) and to ther applcants that nternet search wll reduce ther search tme are therefore makng clams that are nconsstent wth our evdence. 5 CPS wage nformaton s of course only avalable for persons who fnd new jobs, and who are n the outgong rotaton groups. Further, a credble analyss of re-employment wages also requres controls for pre-unemployment wages, a restrcton whch reduces the sample to non-useful levels. 6 Heckman and Snger s (1984) approach would be to model unobserved heterogenety n both the unemployment and search equatons as a dscrete bvarate dstrbuton. In practce, ths typcally amounts to allowng each of the error terms to take two dstnct values only.

9 5 2. Data and Descrptve Statstcs As noted, our data on nternet job search come from the December 1998 and August 2000 Computer and Internet Use Supplements to the Current Populaton Survey. These supplements ncluded the followng queston: Do(es) (you) (any one) REGULARLY use the Internet... to search for jobs?. As always, the regular monthly CPS survey n these months also asked unemployed ndvduals whch out of a lst of nne tradtonal job search methods they used. Internet job search rates n these two surveys, classfed by labor force status, are shown n Table 1. As already noted, the fracton of unemployed jobseekers 7 lookng for work on lne was 25.5 percent n August 2000, up from 15.0 percent n November 1998, less than two years earler. As Table 1 also shows, much of ths ncrease was assocated wth a large rse n home nternet access among unemployed persons (from 22.3 to 39.4 percent), but nternet use for job search condtonal on nternet access also rose over ths perod. By August 2000, regular nternet job search was also surprsngly common among the employed (around 11 percent) and among labor force nonpartcpants, at least those who were nether retred nor dsabled (around 6 percent). 8 In order to measure the job-fndng success of nternet versus other job searchers, we matched observatons n the December 1998 supplement wth the same persons n the ten subsequent CPS regular monthly surveys (January-March 1999, September 1999 through March 2000) n whch some of the same ndvduals were re-ntervewed. Smlarly the August 2000 survey was matched wth September-November 2000, and 7 All unemployed workers not expectng to be recalled to ther former employer are classfed by the BLS as jobseekers.

10 6 May through November Matchng was done usng establshed methods (see for example Madran and Lefgren 1999); some detals about our procedure are provded n Appendx A. To be n our sample, a person had to be unemployed accordng to the offcal Bureau of Labor Statstcs defnton n a Computer/Internet supplement month (December 1998 or August 2000), yeldng a sample of 4139 persons. 9 To be consdered unemployed, the ndvdual had to be not workng, and ether on layoff from a job to whch he/she expected to be recalled, or searchng for work usng at least one of nne recognzed actve methods. 10 These methods are lsted n Table 2; note that they could themselves nvolve nternet use (for example sendng resumes could nclude sendng resumes va emal). The role of our nternet supplement varable s to dstngush persons who ncorporated the nternet nto ther job search strategy from those who dd not, holdng other dmensons of ths strategy fxed. Sample means of all the varables used n the regresson analyses below are presented n Table 2 separately for unemployed persons who searched for a new job on the nternet and those who dd not. In most cases, unemployed workers who look for jobs on lne have observable characterstcs that are usually assocated wth greater job search success than other unemployed workers. For example, n the Computer/Internet Supplement month, the average unemployed nternet searcher had already been unemployed for 3.44 months, somewhat less than the 3.75-month retrospectve 8 Kuhn and Skuterud (2000) compare these recent rates of on-the-job nternet job search (IJS) to hstorcal measures of on-the-job search (OJS) va any method. They are sgnfcantly hgher, suggestng that the nternet may have contrbuted to an ncrease n total OJS. 9 Ths ncludes a small group of persons who were never matched wth an observaton after those dates. Whle these observatons contrbute no nformaton on unemployment duratons, they do contrbute nformaton on the determnants of nternet search, and are retaned n our analyss for that reason.

11 7 duraton of the non-nternet searchers. Internet searchers resded n states wth somewhat lower unemployment rates than other unemployed workers, and had prevously worked n occupatons wth consderably lower unemployment rates. They were more lkely to have been employed pror to the current unemployment spell, were much better educated, and were more lkely to be n ther prme workng ages (26-55) (versus under 26 or over 55). Internet job searchers were less lkely to be black, Hspanc or mmgrant and more lkely to be homeowners than other unemployed persons. Fnally, on average, unemployed workers who looked for work on lne were more lkely, not less lkely, to use tradtonal job search methods than other unemployed workers. In all, they used an average of 2.17 tradtonal search methods, compared to 1.67 for other unemployed workers, suggestng an overall greater nvestment n search. 11 Table 2 also reports rates of nternet use outsde the home among the members of respondents households. These rates dffer between nternet job searchers and others, wth the spouses and other household members (excludng spouses, parents and chlldren) of nternet job searchers beng more lkely to use the nternet outsde the home. Fnally, Table 2 shows that nternet job searchers lve n states wth hgher mean overall nternet access rates, and where a smaller share of households must make a long-dstance call to access the nternet. There s no sgnfcant dfference n state mean nternet access fees between nternet searchers and other unemployed persons. By constructon, no one n our sample was workng n the month n whch we observe whether or not ther job search strategy ncorporated the nternet (December 1998 or August 2000). The fracton of our sample observed n employment at varous 10 We also conducted some analyses that excluded workers expectng recall, as well as some analyses that ncluded margnally-attached workers (nonpartcpants who engaged n passve job search only). In nether case were the results substantally dfferent.

12 8 ponts after these dates s reported n Table 3. For example, among those ndvduals whose labor market status was observed one month after the Supplement date (.e. n January 1999 or September 2000), 29.1 percent were employed. Two months after the supplement date, 37.5 percent were employed, and a year later 55.9 percent were employed. If we pool all ndvduals who were re-ntervewed at least once after the date n whch we observe ther nternet search actvty, the same share, 55.9 percent, were seen n re-employment at some tme after the Supplement date. Comparng nternet job searchers wth other unemployed workers, essentally no dfference n employment rates s evdent one or two months after an ndvdual s nternet job search actvty s observed. A year later, however, 64.6 percent of unemployed nternet searchers are re-employed, compared to 53.3 percent of other unemployed workers. Ths dfference, lke the dfference n re-employment at any tme after the Supplement date (n row 4 of the Table), s statstcally sgnfcant. On the surface, Table 3 thus seems to suggest that nternet search facltates re-employment, at least f one allows a few months to elapse for ths method to yeld results. 3. Re-employment Probts A frst step n ascertanng whether the dfferences found n Table 3 are truly causal effects of nternet search s to see whether they are smply artfacts of dfferences between nternet searchers and other unemployed persons n observable characterstcs, such as educaton, local labor market condtons, and the use of non-nternet job search methods. To ths end, Table 4 presents probt estmates of the probablty an unemployed ndvdual s re-employed 12 months after we observe ther nternet job search actvty n 11 Ths apparent complementarty between nternet and other job search methods s examned n more detal n Kuhn and Skuterud (2000).

13 9 the CPS Computer/Internet Supplement. We focus on 12 months because ths s where the largest apparent nternet effect was observed n Table Of course, re-employment probabltes n the above probts wll lkely depend on how long an ndvdual had already been unemployed when we observe whether or not he/she uses the nternet for job search,.e. at the Supplement date. As s well known, there are at least two dstnct reasons for ths: duraton dependence (long unemployment spells may have a causal effect on subsequent ext rates from unemployment), and unobserved heterogenety (ndvduals who have been unemployed a long tme are dsproportonately less employable on unobserved dmensons). In Sectons 3-5 of ths paper we handle both these possbltes n a smplstc manner: we smply nclude retrospectve (pre-supplement date) unemployment duratons as a regressor n our models of post-supplement duratons. Sectons 6 and 7 wll handle both these ssues more formally. Throughout Table 3 as well as all the followng tables, we wll present specfcatons of each equaton wth and wthout a control for home nternet access. Whle we do not beleve home nternet access has a causal effect on the job-fndng rate - -what should matter s whether the nternet s used for job search we can thnk of plausble arguments both for and aganst holdng t constant n a comparson of nternet searchers and others. On the for sde, home nternet access may be correlated wth other unobserved characterstcs (for example wealth, whch n turn s correlated wth past employment) that do affect job-fndng rates. On the other hand, home nternet 12 Smlar analyses were performed for re-employment wthn a month, wthn two months, or at any tme after nternet search actvty s observed. (In the latter specfcaton, we added a control for the number of months n whch the ndvdual s observed after the Supplement month). In all cases, the results were smlar to those n Table 4: whenever even a relatvely parsmonous set of demographc controls are used, the nternet search coeffcent s ether nsgnfcant or negatve.

14 10 access s a very powerful predctor of on-lne search among the unemployed, and much of the varaton n home access may be drven by genunely exogenous dfferences n the rate of nternet dffuson across space, tme and ncome groups; n ths case controllng for access could be dscardng a large amount of useful varaton. Thus we present both specfcatons. Column 1 of Table 4 reproduces the sgnfcant dfference n Table 3, where only nternet search s ncluded as a regressor. Addng a control for home nternet access n Column 2 reduces the value of the coeffcent. Thus, part of the apparent effect of nternet use on re-employment rates n Table 3 s n fact a re-employment dfferental between ndvduals who have nternet access and those who do not. Ths should not be surprsng snce these ndvduals may be, as dscussed, more re-employable on unobserved dmensons. A smlar, but stronger message emerges when addtonal controls for observable characterstcs are added n the remanng columns of Table 4. Columns 3 and 4 add controls for labor market condtons local and occupatonal unemployment rates-- and for varous characterstcs of the unemployment spell. The latter nclude how long the spell had been n progress by the Supplement month, whether the ndvdual was on layoff and therefore expectng recall, what actvty (school, publc sector employment, prvate sector employment, self employment, school) preceded the unemployment spell, and the reasons for leavng any prevous job ( lost job and temporary job, wth quts as the omtted category). We also nclude a fxed effect for the 2000 survey to capture any changes n macroeconomc condtons between the surveys. As for columns 1 and 2, we present one specfcaton wth and one wthout a home nternet access control. The

15 11 apparent effect of nternet search on re-employment remans postve, but s agan smaller and becomes statstcally nsgnfcant n the presence of a home access control. Effects of the control varables n columns 3 and 4 are generally n lne wth expectatons. For example, although the coeffcent s not qute sgnfcant at conventonal levels, we see that ndvduals wth hgh retrospectve duratons are less lkely to be re-employed a result that mrrors the common fndng of declnng reemployment hazards n duraton studes. A hgh occupatonal unemployment rate depresses job-fndng rates, and ndvduals who worked or went to school mmedately before the onset of ther current unemployment spell are much more lkely to be reemployed than those who dd nether. Persons whose last job was n the prvate sector fared better n re-employment than those whose last job was n the publc sector or n self-employment, or who dd not work just pror to the current unemployment spell. 13 Columns 5 and 6 add controls for demographc characterstcs. They show, as expected, that younger workers are re-employed more quckly, and less-educated and black workers more slowly. Sngle men are less lkely to be re-employed than sngle women, but marred men are more lkely to be re-employed than marred women, possbly reflectng greater geographcal search constrants among marred women (Crossley, Jones and Kuhn, 1994). The nternet effect on re-employment now becomes hghly nsgnfcant n both specfcatons. The last two columns of Table 4 add controls for the use of other, tradtonal job search methods. Interestngly, we detect sgnfcant postve effects on re-employment for three of these methods: drect employer contact, sent resumes and publc 13 Note that n a substantal number of cases the ndvdual s last job preceded a spell of nonpartcpaton, so that these sector ndcators do not smply subdvde the group who entered unemployment drectly from a job.

16 12 employment agences, whch ncdentally are also the search methods most commonly used by unemployed persons n our data. For the remanng methods, no statstcallysgnfcant effects on the job-fndng rate are found. Lkewse, addng the nternet to one s job search strategy appears not to ncrease re-employment rates. In sum, when we control for observed characterstcs of unemployed workers and ther unemployment spells, nternet job search does not appear to be effectve n reducng unemployment duratons. 4. Econometrc Issues Whle the results n columns 7 and 8 of Table 4 certanly suggest that ncorporatng the nternet nto one s job search strategy s neffectve n reducng jobless duratons, there are at least three reasons why ths concluson may be premature. In ths secton we descrbe these reasons, and outlne our strategy for dealng wth them n the remander of the paper. A frst reason why the re-employment probts summarzed n Table 4 mght fal to reveal a true, benefcal effect of nternet job search s smply an neffcency n the estmaton procedure. In partcular, any probt focusng on a worker s labor force status at only a sngle date (n the above case 12 months after hs/her search actvty s observed) dscards a consderable amount of nformaton on the actual duraton of unemployment. To address ths ssue, n what follows we shall estmate duraton models that ncorporate all the avalable nformaton about a worker s jobless spell followng the Supplement date. Of course, the nformaton avalable to us on duratons n the CPS s hghly dscrete: at best, we only know the month n whch re-employment occurred; n some cases (the gap between the two four-month CPS observaton wndows ), we only know that re-employment occurred durng an eght-month perod. Ths makes

17 13 contnuous-tme duraton models hghly napproprate. For ths reason we develop and estmate a dscrete-tme hazard model that takes nto account the partcular features (potentally large falure wndows whose structure vares across observatons) of CPS duraton data, whle stll allowng for a fully flexble form of the baselne hazard functon. 14 A second reason why the estmates n Table 4 mght dsguse a true, benefcal effect of nternet search on jobless duratons results from the fact that our data s sampled at random from the stock of workers who were unemployed n the month of the Computer and Internet supplement. As a result, the probablty of beng n ths sample s drectly proportonal to the dependent varable.e. the length of an ndvdual s completed unemployment spell-- a property sometmes referred to as length-based samplng. Snce, n the smplest case, such a systematc undersamplng of short spells wll bas our nternet search coeffcents towards zero 15, addressng ths ssue s also essental to rulng out a true, benefcal effect of nternet search on jobless duratons. In what follows, we wll augment our duraton model usng a technque ntroduced by Lancaster (1979) to address ths ssue. Essentally we wll condton each observaton s contrbuton to the lkelhood functon on the fact that t lasted long enough to be observed n our sample. The remanng potental source of bas n Table 4 concerns the endogenety of the nternet job search varable. For example, one mght be concerned that ndvduals who look for work on lne are a postvely-selected sample, n the sense that they are more 14 Exstng dscrete-tme hazard models, such as Meyer s (1990) requre the structure of ntervals to be the same across observatons. 15 Suppose that (asde from a constant term) nternet search was the only regressor n a smple OLS regresson model, and that ts true effect was to reduce unemployment duratons. Then the systematc undersamplng of short duratons nduced by stock samplng wll nduce a postve correlaton between the

18 14 motvated and able to fnd a new job than observatonally equvalent non-nternet searchers. Indeed as already noted, ths clam s sometmes made by nternet job stes n marketng ther servces to employers. Of course, f ths s the case, then the estmates n Table 4 exaggerate the benefts of nternet job search, thus strengthenng the case that nternet job search does not reduce unemployment duratons. But what of the possblty of negatve selecton nto nternet search on unobservables? We can thnk of at least three mechansms that could generate ths. Frst, as suggested by Holzer (1987) n another context, persons who use formal and anonymous job search channels such as the nternet may be dong so because ther nformal contacts and socal networks are poor. 16 Second, and related, s the possblty of prvate nformaton about re-employablty: persons usng a larger number of search methods ncludng the nternet may do so n response to prvate nformaton that ther search prospects are partcularly poor. Fnally, especally condtonng on home nternet access, nternet job search s a very low-cost job search method. Thus the costs of engagng n t are unlkely to screen out ndvduals wth only a very margnal nterest n fndng a new job. Gven these possbltes, n order to complete the case aganst an unemployment-reducng effect of nternet search we would need to rule out negatve selecton as an explanaton of the postve partal correlaton between nternet search and jobless duratons we observe n our data. In order to address the endogenety ssue, we need to do two thngs: one s to dentfy some nstrumental varables that affect nternet use but are unlkely to be correlated wth dosyncratc varaton n ndvdual workers re-employablty. The level of nternet search and the error term. Falng to account for ths wll bas the (negatve) coeffcent on nternet search upward,.e. towards zero.

19 15 second s to develop a means of ncorporatng these nstruments (whch are essentally excluson restrctons) nto a duraton model that both handles the peculartes of CPS duraton data and accounts for the length-based samplng problem dscussed above. Regardng the latter ssue, we shall proceed by jontly modellng the process of selecton nto nternet job search among the unemployed and the duraton of search. By adaptng a technque frst used by Han and Hausman (1990) n another context, we are able to allow the dosyncratc, unobserved determnants of both these outcomes to be correlated, and to estmate the degree of correlaton emprcally. 17 Regardng nstruments, we propose two sets. The frst s a set of ndcators of nternet use by other members of the respondent s household outsde the home. The ratonale s that the presence of such a person n the household should reduce any costs of becomng famlar wth the nternet or wth on-lne job search stes. Further, unlke the total varaton n nternet job search, nternet search that s nduced by the above factors should not be affected by unobserved adverse nformaton about the ndvdual s jobfndng prospects. Also, nternet use assocated wth havng close contact wth nternetusers s unlkely to be a response to poor nformal networks. Our alternatve set of nstruments comprses three varables measurng mean nternet access costs and nternet dffuson at the state level. These are the mean level of access fees pad by nternet users n the respondent s state, the share of households n the respondent s state who need to make a long-dstance telephone call to access the nternet, and smply the state mean home nternet access rate. Because these are state means, they should be purged of any ndvdual dosyncraces n re-employablty. Also, snce (at 16 In partcular, Holzer suggests that mnorty youth dsproportonately use formal and anonymous job search networks n part due to low access to nformal contacts n the world of work, and that ther relance on formal methods n part explans ther lower job-fndng rates.

20 16 least n two of the three cases) they focus on the costs rather than benefts of nternet job search, they should not be contamnated by unobserved prvate nformaton regardng expected unemployment duratons, ether across ndvduals or states. It s probably worth notng that, whle a good case can be made that these nstruments purge our estmates of any spurous negatve correlaton between nternet search and re-employment, they may go too far n that drecton. For example, one mght argue that, despte our controls for state unemployment rates, our state-level nternet access means are correlated wth unobserved labor demand shocks that reduce unemployment duratons. Or one mght argue that, controllng for observables, hghlyreemployable unemployed persons dsproportonately lve n households wth persons who use the nternet at non-home locatons. If anythng, therefore, our nstruments are probably based towards estmatng an unemployment-reducng nternet search effect; thus f we fal to fnd a statstcally sgnfcant benefcal effect even wth these nstruments, we can be qute sure that nternet search s neffectve n shortenng unemployment spells. The analyss n the remander of the paper proceeds as follows. Secton 5 develops and estmates a duraton model that uses all the avalable nformaton on duratons and takes account of the pecular structure of CPS duraton data. Secton 6 then develops an extenson of ths model that ncorporates both length-based samplng and endogenous selecton nto nternet search. Secton 7 dscusses the results of estmatng ths model, whch addresses all three of the above econometrc concerns smultaneously. 17 Han and Hausman (1990) use a bvarate normal dstrbuton n combnaton wth a flexble baselne hazard to model potentally correlated competng rsks.

21 17 5. A Unvarate Duraton Model We begn, as s common, by assumng the hazard rate nto re-employment, λ (τ ), s separable nto a baselne component that depends on elapsed duraton λ ( ), and a component that depends on a lnear combnaton of observed characterstcs X and estmated parameters β: 0 τ λ( τ ) = λ 0 ( τ ) exp( X β ) (1) From assumpton (1) t follows that (see Kefer 1988, pp ): log Λ 0 ( t ) = β + µ (2) X where the random varable Λ ( t ) s the ntegrated baselne hazard up to each 0 observaton s realzed duraton,.e.: 0 0 ) 0 t Λ ( t ) = λ ( τ dτ (3) and where µ follows a type-1 extreme-value dstrbuton. 18 Thus the transformed duraton varable, log Λ 0 ( t ),--whch s monotoncally ncreasng n t -- can be thought of as the dependent varable n a lnear regresson. Suppose now that a partcular unemployment spell s known to have ended between two dates, t a < t b. Defnng δ a log Λ 0 ( t a ) and δ b logλ 0 ( t b ), the lkelhood of such a spell s just: F( δ X β ) F( δ X β ), (4) a b where F()) s the cdf of µ. Duratons known only to have ended after, say, t a (.e. rghtcensored duratons) have a lkelhood of 1 F( δ X β ) ; duratons known to have ended between t=0 and, say, t b, have a lkelhood of F( δ X β ). 19 a b

22 18 In our data, job searchers are observed no more frequently than once per month. Recognzng ths dscreteness, we dvde the set of possble jobless duratons nto dsjont ntervals. 20 Denote the number of such ntervals by T+1; n the results reported n Table 5 (whch focus on post-supplement duratons only), we used eght ntervals: 0-1, 1-2, 2-3, 3-10, 10-11, 11-12, and more than 13 months. For some of our observatons (for example those persons observed as unemployed n one month and employed the next), we know n exactly whch of these ntervals ther unemployment spell ended. Others are rght-censored, due to attrton or rotaton out of the sample. For yet others (ncludng, but not lmted to, persons who were not matched n a perod before they are frst observed n employment) we know only that they became employed at some pont wthn a set of adjacent ntervals. To allow for the latter types of observatons, defne V as a 1xT vector of lower bound dummy varables (thnk of these as applyng, n order, to each of the T+1 ntervals defned above except the hghest one). Set V equal to zero for all ntervals except the one precedng the nterval n whch worker s unemployment spell s known to have ended. 21 Defne V as a 1xT vector of upper bound dummy varables, equal to zero for all ntervals except the one durng whch we knew the unemployment spell ended. 22 Fnally, let δ be a Tx1 coeffcent vector correspondng to the cut ponts between the above ntervals. Because the elements of δ correspond to the log of the 18 The cdf for the extreme-value dstrbuton s gven by F ( µ ) = exp( exp( µ )) 19 Unlke observed duratons whch must be postve, note that the transformed duratons and the error term µ occupy the entre real lne. 20 Appendx B descrbes how we constructed unemployment duratons from the matched CPS fles. 21 If the observaton s rght-censored ths s the nterval before t became rght-censored; f the observaton became re-employed durng the frst nterval V s a vector of zeroes. 22 If the observaton s rght-censored, V s a vector of zeroes.

23 19 ntegrated baselne hazard at the upper end of each nterval, and becauseδ s estmated, ths procedure allows for an unrestrcted baselne hazard functon. Puttng all the above together, the log lkelhood for the entre sample can be expressed as: log L = Cens= L Cens= 0 Cens= R log[ F( V δ X β)] + [ F( V δ X β ) F ( V δ X β) ] log log[ 1 F( V δ X β) ]. + (5) where Cens = L, 0 and R ndcates the observaton s left-censored, not censored, or rghtcensored, respectvely. (Note that we refer to observatons that became re-employed n the frst month of ther unemployment spell as left-censored because the transformed duraton varable, log Λ 0 ( t ), has no lower bound for ths group). Whle the dervaton leadng to (5) reles on F havng an extreme value dstrbuton, n Table 5 we present estmaton results based on a normal dstrbuton for F as well. Ths ordered probt-type specfcaton does not follow drectly from the proportonal-hazards specfcaton n (1), but yelds predcted duratons (both wth and wthout nternet search) that are very smlar to those obtaned from the extreme-value specfcaton. 23 The value of the ordered-probt specfcaton s that t allows us to model correlaton between the dsturbance term ( µ ) n our unemployment duraton equaton (2), and unobserved characterstcs that nduce unemployed ndvduals to look for work on lne n the followng secton. 23 Predcted survvor curves for all four specfcatons n Table 5 are avalable from the authors. The normal and extreme-value based curves are essentally ndstngushable from each other. The lkely reason why functonal form s of so lttle consequence s that our specfcaton allows for an unrestrcted baselne hazard: movng the cut ponts for the two dstbutons gves us a large number of degrees of freedom wth whch to ft observed transton patterns.

24 20 As n Table 4, Table 5 presents specfcatons wth and wthout controls for home nternet access. (Note that because the ndex X β enters equaton (1) negatvely, a postve coeffcent n Table 5 ndcates that the varable n queston reduces the hazard rate,.e. t ncreases expected unemployment duraton). Recallng also that our estmaton framework so far contnues to treat pre-supplement unemployment duraton as an exogenous covarate, Table 5 shows that persons who are far nto ther unemployment spells (.e. wth hgh retrospectve duratons n the Supplement month) have longer remanng unemployment duratons after that date. Duratons were lower n the 2000 Supplement, reflectng the tghter aggregate labor market condtons prevalng around the tme of that survey. Hgh state unemployment rates rase unemployment duratons. As n Table 4, younger workers have shorter unemployment duratons and older workers reman unemployed longer. One nterestng dfference from Table 4 s that persons wth home nternet access now have sgnfcantly shorter jobless duratons. The most surprsng fndng from Table 5, however, s that accordng to the coeffcent estmates, nternet job search now appears to be not smply neffectve, but n fact sgnfcantly counterproductve. In other words, holdng constant observable characterstcs of the person and the prevous duraton of the unemployment spell, persons who searched for work on lne actually entered re-employment more slowly than those who dd not, durng the perod after we observe whether they search on lne. Whle not provng that nternet search s n fact counterproductve, these results certanly present a strong prelmnary case aganst the argument that nternet job search reduces unemployment duratons. Of course, these results are also consstent wth negatve selecton nto nternet search on unobservables, especally n the specfcaton where nternet access s held constant, where the estmated counterproductve effect s the strongest.

25 21 6. Length-Based Samplng and Endogenety of Internet Search: Methods a) Length-Based Samplng We begn agan wth the lnear-regresson representaton of the proportonal hazard model n (2), but now re-nterpret the duraton varable t, as total spell duraton ncludng the retrospectve component, (thus t s known to have lasted at least dates, t a < t b (where both t a and t b must exceed our data s that, condtonal on lastng at least R t. Consder agan a spell observed n our data R t months), known to have ended between two R t ). Thus the nformaton provded by R t months, ths spell lasted between t a and t b months. Once agan defnng δ a log Λ 0 ( t a ) and δ b logλ 0 ( t b ), and now R R δ logλ 0 ( t ), the lkelhood of such a spell s 24 : F( δ a X β ) F( δ b X β ). (6) R 1 F( δ X β ) Next, we dvde up the set of possble total duratons (ncludng the retrospectve porton) nto T+1 ntervals and defne the 1xT lower- and upper- bound vectors V and V as before for these spell duratons. 25 Lastly, defne the 1xT vector R V as equal to zero for all ntervals except the one precedng the supplement month. (Thus, for example, wth month-long ntervals, a worker who became unemployed n September 1998 has the thrd --November element of R V set to one and the rest to zero). Parallel to (6), the log lkelhood for the entre sample, corrected for length-based samplng, can now be wrtten: 24 Ths assumes a constant nflow rate nto unemployment before the Supplement month. 25 In the results reported here we used 22 ntervals for these total duratons. Wth the excepton of months 7 and 8 (whch were combned due to small sample szes) these comprsed ndvdual months up to 16. Beyond that, the categores were 16-22, 23-26, 27, 28, 29-37, and over 38 months.

26 22 log L = Cens= L Cens= 0 Cens= R log [ F( V δ X β) ] F( V δ X β ) F ( V δ log R 1 F( V ) δ Xβ 1 F( V δ X β) log. R 1 F( V δ Xβ) + X β) + (7) Note that because all left-censored spells n ths context are new spells, no lengthbased samplng correcton apples to them. Of course, parameter estmates n (7) have a dfferent nterpretaton than n (5): they refer to the effect of each covarate on total unemployment duratons (ncludng the retrospectve porton) rather than on post-supplement duratons. In the nterpretaton of both (5) and (7), note agan that gven the absence of nternet search data at any dates other than the Supplement month we treat and nterpret nternet search, lke all our other covarates, as a non-tme-varyng characterstc of the unemployment spell. 26 b) Endogenety Fnally, to ncorporate the possble endogenety of nternet job search nto our model, we rewrte (2) as: log Λ 0 ( t ) = X β + y γ + µ (8) where y (prevously ncluded n X ) s the nternet search dummy, wth coeffcent γ. Now defne the dfference between the margnal beneft and margnal cost of nternet job search as the latent varable: y = * W θ + ε. (9) 26 Any other treatment of nternet search would requre data on nternet search actvty at more than one pont durng an unemployment spell, whch s currently not avalable.

27 23 where W s a vector of exogenous, non-tme varyng covarates, X, plus a set of nstrumental varables excluded from X. The latent varable nstead we observe: * y s not observed, but y = * 1 f y > 0 0 otherwse. Our concern s that ε n (9) may be correlated wth µ n (8). Ths leads to bas, e.g., n the estmate of γ n (8) because the endogenous varable, y, s correlated wth the error term µ. Unfortunately, we are aware of no wdely-accepted technque for estmatng hazard models wth an endogenous covarate. The dffculty, essentally, s modelng the jont dstrbuton of ε and µ when µ s non-normal. Our approach s therefore to extend the ordered-probt verson of (7) to the bvarate case, where standard bvarate normal results can be used to model the jont dstrbuton of ε and µ. Wth the excepton of the nterval nature of our duraton measure and the correcton for length-based samplng, our approach s smlar to Greene s (1998) bvarate probt model wth an endogenous dummy varable. 27 To extend the model n (7) to the case where ε and µ have a jont normal dstrbuton wth (potentally) non-zero correlaton, note frst that an observaton wll be n our sample ff: t > t R or, or, log Λ 0 ( t ) > δ R µ > δ X β yγ R

28 24 or, R µ > δ X β I Wθ + ε ) γ (10) ( where the ndcator functon, I (), returns 1 f y * > 0 and 0 otherwse. Thus (for the bvarate normal case) the lkelhood of beng n the sample s gven by: Φ 2 ( ρ z, z, ρ ) + Φ ( z, z, ) (11) where Φ2 s the standard bvarate normal cdf wth correlaton ρ, and: z = 0 W θ z 11 = R V δ X β γ R z10 = V δ X β. By adjustng the denomnator n (7) to reflect ths new condton, and the numerator n (7) to account for the jont dstrbuton of µ and ε, we can obtan an unbased estmate of γ. To express the bvarate lkelhood functon, frst defne: q = 2 y 1 (thus q = 1 when y = 1, and q = -1 when y = 0), The complete lkelhood functon for ths model can then be wrtten: where: L = log log Φ [ Φ ( q z, z, q ρ ) ] Φ2 ( q z0, z3, q ρ ) log Φ2 ( z0, z11, ρ ) ( z 0 0 Φ2 ( q z0, z, z, ρ ) + Φ 11 Φ2 ( q z0, z Φ ( z, z Cens = Cens = R Cens = L z2 = V δ X β yγ, and z = V δ X β y γ , q ρ ) ( z, z 0 2, q ρ), ρ ) 10., ρ ) + (12) 27 See also Greene (2000), pp

29 25 6. Endogenety of Internet Search and Length-Based Samplng: Results Results from maxmzng the lkelhood functon n equaton (12) are presented n Table 6. In all, sx specfcatons are reported, each of whch contans two equatons: a search equaton for the determnants of nternet search, and an unemployment duraton equaton. The frst two columns present estmates of the entre model, but wth the correlaton (ρ) between the search and duraton error terms ( µ and ε ) constraned to equal zero. Essentally, ths amounts to ntroducng the correcton for length-based samplng as well as the dscrete-tme hazard framework descrbed n Sectons 5 and 6(a), but not the correcton for endogenous nternet search n Secton 6(b). As n prevous tables, we present a verson that ncludes controls for home nternet access (n both the search and duraton equatons) and one that does not. To economze on space we report only the nternet search and nternet access coeffcents n the duraton equaton, and the estmated correlaton between the error terms, ρ. 28 The man message of columns 1 and 2 of Table 6 s that length-based samplng alone cannot account for our estmated counterproductve search effects. Whle the nternet search effect on duratons s now nsgnfcant n the absence of an nternet access control (column 1), column 2 shows that ths s due to two offsettng effects: (1) unemployed ndvduals wth home nternet access have shorter unemployment duratons, whether they search on lne or not; and (2) among the unemployed wth access, those who use the nternet to look for work actually have longer unemployment duratons than those who do not. Whether these longer duratons are a perverse causal effect of nternet 28 Complete results for the four specfcatons wth ρ unrestrcted are provded n Appendx C; results for the two remanng specfcatons are very smlar. Note also that when ρ=0, the nstruments used n the search equaton should not, and do not affect our estmates of the duraton equaton. Thus Table 6 does not present separate estmates for the two nstrument sets when ρ=0.

30 26 search or an artfact of selecton on the remanng unobservables cannot be determned from columns 1 and 2; to address that queston we must turn to the estmates n the remander of the Table. As noted, columns 3 through 6 of Table 6 relax the ρ = 0 constrant. Columns 3 and 4 present estmates of the duraton equaton when nstrument set 1 non-home nternet use by other household members s used; columns 5 and 6 use state means of nternet access costs and nternet dffuson as nstruments. Three features are noteworthy: frst, all the postve estmated effects of nternet search on duratons dsappear; thus nternet job search no longer appears to be counterproductve. Instead (pont two), the model prefers to attrbute the postve partal correlaton between nternet search and unemployment duratons to what we have been callng negatve selecton: all our estmates of ρ are postve, ndcatng that unemployed ndvduals who have a hgher unexplaned dsposton to look for work on lne tend to have longer unemployment duratons. Thrd, however, nether the estmate of ρ nor the estmated effect of nternet search on unemployment duratons s sgnfcantly dfferent from zero. Thus, despte farly powerful nstruments (ncludng some t-ratos as hgh as 5 see Appendx C), our model s unable to dstngush selecton on unobservables from a causal effect of nternet search. Can our data, then, when all potental bases are taken nto account, place any restrctons on ether the selecton process nto nternet search or on ts causal effects? A reasonable lower-bound estmate of the causal effect of nternet search s zero: t s hard to magne anyone addng ths element to hs/her search strategy f t makes hm or her worse off. Imposng a causal nternet effect of zero on the models estmated n Table 6 yelds estmates for ρ of.107 (.033) and.106 (.033) respectvely for the household access