The Place Premium: Michael A. Clemens Claudio E. Montenegro Lant Pritchett

Transcription

1 Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Pol c y Re s e a rc h Wo r k n g Pa p e r 4671 Ba c k g r o u n d Pa p e r t o t h e 2009 Wo r l d De v e l o p m e n t Re p o r t The Place Premum: Wage Dfferences for Identcal Workers across the US Border Mchael A. Clemens Claudo E. Montenegro Lant Prtchett The World Bank Development Economcs Department, World Development Report Team Center for Global Development (CGDEV) Harvard Kennedy School July 2008 WPS4671

2 Polcy Research Workng Paper 4671 Abstract Ths paper compares the wages of workers nsde the Unted States to the wages of observably dentcal workers outsde the Unted States controllng for country of brth, country of educaton, years of educaton, work experence, sex, and rural-urban resdence. Ths s made possble by new and unquely rch mcrodata on the wages of over two mllon ndvdual formal-sector wage-earners n 43 countres. The paper then uses fve ndependent methods to correct these estmates for unobserved dfferences and ntroduces a selecton model to estmate how mgrants wage gans depend on ther poston n the dstrbuton of unobserved wage determnants. Followng all adjustments for selectvty and compensatng dfferentals, the authors estmate that the wages of a Bolvan worker of equal ntrnsc productvty, wllng to move, would be hgher by a factor of 2.7 solely by workng n the Unted States. Whle ths s the medan, ths rato s as hgh as 8.4 (for Ngera). The paper documents that (1) for many countres, the wage gaps caused by barrers to movement across nternatonal borders are among the largest known forms of wage dscrmnaton; (2) these gaps represent one of the largest remanng prce dstortons n any global market; and (3) these gaps mply that smply allowng labor moblty can reduce a gven household s poverty to a much greater degree than most known n stu antpoverty nterventons. Ths paper a product of the World Development Report Team, Development Economcs Department; the Center for Global Development (CGDEV) and the Harvard Kennedy School s part of a larger effort n the department to mprove our understandng of the consequences of mgraton restrctons. Polcy Research Workng Papers are also posted on the Web at The authors may be contacted at: mclemens@cgdev.org, cmontenegro@worldbank. org, or lant_prtchett@harvard.edu. The Polcy Research Workng Paper Seres dssemnates the fndngs of work n progress to encourage the exchange of deas about development ssues. An objectve of the seres s to get the fndngs out quckly, even f the presentatons are less than fully polshed. The papers carry the names of the authors and should be cted accordngly. The fndngs, nterpretatons, and conclusons expressed n ths paper are entrely those of the authors. They do not necessarly represent the vews of the Internatonal Bank for Reconstructon and Development/World Bank and ts afflated organzatons, or those of the Executve Drectors of the World Bank or the governments they represent. Produced by the Research Support Team

3 BACKGROUND PAPER TO THE 2009 WORLD DEVELOPMENT REPORT The Place Premum: Wage Dfferences for Identcal Workers across the US Border Mchael A. Clemens Center for Global Development Claudo E. Montenegro World Bank and Department of Economcs, Unversdad de Chle Lant Prtchett Harvard Kennedy School and Center for Global Development JEL Codes F22, J61, J71, O15. We are grateful to Indermt S. Gll and hs team at the World Development Report of the World Bank who bult and allowed use of the database. We thank Samuel Bazz for excellent research assstance. We receved helpful comments from Chrstopher Blattman, Wllam Easterly, Davd Lndauer, Davd McKenze, Mark Rosenzweg, and semnar partcpants at Yale Unversty and the Center for Global Development. Ths work was partally supported by generous grants from the John D. and Catherne T. MacArthur Foundaton and from AusAID. The paper represents the vews of the authors alone and not necessarly those of ther employers or funders.

4 1 Introducton Three questons have each launched a thousand papers. Frst, how large are the gaps n compensaton caused by varous types of labor market dscrmnaton? Second, how large are the relatve prce dfferentals n global markets caused by nternatonal borders? Thrd, how can publc polces rase the ncomes of poor households? We brng these lteratures together wth an estmaton of the dfferences between the wages of workers n 42 low- and mddle-ncome countres and the wages those same people would earn n the Unted States. Ths calculaton at once documents an enormous form of wage dscrmnaton, measures a massve cross-border prce wedge, and suggests a polcy that could dramatcally rase the earnngs of many low-ncome famles. The frst secton of the paper creates baselne estmates of wage gaps controllng for ndvdual observable trats. It does so wth a unque harmonzed database on the purchasng power prce-adjusted wages and other trats of over two mllon workers n 43 countres, ncludng the Unted States. Ths allows us to predct wages of observably dentcal workers on ether sde of the US border for each of these countres. Crucally, the US data dentfy the ndvduals country of brth and, for the foregn-born, year of arrval n the US. Ths allows our defnton of observably dentcal to go beyond standard covarates such as years of schoolng, age, sex, and rural/urban resdence. We can also compare workers of the same country of brth mplctly controllng for culture, language, and socal networks and same country of schoolng whch adjusts for the qualty and relevance of schoolng. The wage gaps that emerge from these ntal estmates are large. For nstance, n our preferred econometrc specfcaton, 1 a Bolvan-born, Bolvan-educated, 35 year-old urban male formal sector wage worker wth 9-12 years of schoolng earns an average of US$1,831 per month workng n the Unted States but US$460 (at purchasng power party) workng n Bolva. Hence the earnngs rato between these observably dentcal people s We produce estmates of the wage ratos of observably dentcal workers for each of 42 countres we call these ratos R o, where the subscrpt sgnfes observably dentcal. 2 Bolva s rato R o of 3.98 s near the medan, whle the lowest such ratos we observe are from the Domncan Republc, at (dependng on the regresson functonal form) and the hghest are for Ngera, at But wage gaps for observably equvalent workers do not necessarly reflect wage dscrmnaton. The second part of the paper grapples wth the fact that, no matter how many ndvdual trats are controlled for, wage dfferentals for observably equvalent workers do not necessarly consttute evdence of wage dfferentals across workers of 1 We test the senstvty of the results to econometrc specfcatons, n partcular relaxng key assumptons mposed n the standard Mncer functonal form of wage regresson specfcatons that have recently receved mportant crtcsm (e.g. Heckman, Lochner, and Todd (2006) n the US context and Rosenzweg (2007) for nternatonal comparsons). 2 Rosenzweg (2007) estmates country specfc skll prce dfferentals comparng the same ndvduals on both sdes of the Unted States border usng the observatons of wage changes of ndvduals from the US New Immgrant Survey. Those calculatons heavly nfluence ours, but the results reported here are the frst we know of that pool ndvduallevel survey data across countres to estmate the mpact of borders.

5 equal ntrnsc productvty, as foregn-born workers n the US can obvously dffer n unobservable ways from ther observably dentcal counterparts back home. Ths ssue s common to all attempts to measure dscrmnaton. In other words, wage ratos for observably-equvalent workers R o are not the same as wage ratos for workers of equal ntrnsc productvty who would be wllng to move from one country to another; we call ths latter rato R e. One factor that leads to wage gaps between foregn-born workers n the US and observably dentcal workers abroad s selecton on unobservable determnants of productvty selecton both by the mgrants themselves and by mgraton polcy. The effect of selecton on the wage gaps we measure s complex and we explore t below wth a new theoretcal model. The true wage gan to a typcal mgrant depends on two separate but related aspects of selecton: where mgrants come from wthn the source-country dstrbuton of unobservable productvty determnants (selecton), and where they end up wthn the destnaton-country dstrbuton of unobservable productvty determnants (sortng). The hgher s mgrants typcal poston n the orgn-country dstrbuton of unobserved productvty determnants all else equal the lower s the wage gan. But the hgher s mgrants typcal poston n the destnaton-country dstrbuton of unobserved productvty determnants all else equal the hgher s the gan. The model shows that postve selecton on unobservable trats from the orgn country s nether necessary nor suffcent for overestmaton of the wage gan to mgraton. If mgrants come from the upper part of the dstrbuton of unobserved productvty determnants n the orgn and were randomly sorted nto the wage dstrbuton n the destnaton country, then comparng average workers would ndeed overstate the wage gan. But f selected mgrants sort nto the upper end of the dstrbuton n the destnaton country, the comparson of wages for average workers wth gven observed trats can accurately reflect or even understate the gan. That s, f the people who are uncommonly ntellgent, energetc or ambtous n the source country selectvely mgrate and are people who are uncommonly ntellgent, energetc and ambtous n the destnaton country, ther wage gan could be the same as or even larger than the wage gan to less ntellgent and energetc people. Furthermore, among those postvely selected on unobservables from the orgn, those bound for the upper end of the dstrbuton n the destnaton are more lkely to be seen n the data than those bound for the mddle or the bottom the former have more to gan from mgraton than the latter and are thus more lkely to move. We match a theory of selecton (from the source country) and sortng (n the destnaton country) wth data to estmate the bas attrbutable to selecton on unobservables. Another, separate factor that can cause the wages of observably dentcal workers to dffer across countres n the absence of dscrmnaton s what we call natural barrers. Workers mght requre a compensatng dfferental to bear the costs broadly consdered of movng to a new land. These nclude the dffculty of learnng a new language, beng away from one s famly, and enterng new socal networks, as well as the drect cost of travel. Workers mght also be credt-constraned and have dffculty fnancng the move. 2

6 Only a completely exogenous movement of workers across borders would allow estmaton of wage gans wthout selecton and wthout natural forces determnng who s wllng and able to move. We do not present (or desre!) such an experment. Instead, we trangulate usng fve dstnct methods to place estmated bounds on selecton and natural barrers drawng on theory and varous emprcal lteratures. These ndependent calculatons yeld the remarkably consstent result that selecton of mgrants on unobservable wage determnants results n an observed US-to-foregn wage rato for observably equvalent workers (R o ) of around 1.25 tmes the true rato for equalproductvty workers on average across countres, and that the combned effect of selecton and natural barrers produces observed ratos about 1.5 tmes the true rato for equal-productvty workers wllng and able to move (R e ). Even after ths correcton, wage gaps across borders reman extremely large. Gven our medan observed rato R o of about 3.9, the medan rato purged of selecton on unobservable wage determnants and the effects of natural barrers wage ratos of equally productve workers wllng to move (R e ) s roughly 2.6 (=3.9/1.5). Even ths conservatve estmate of R e s above 3 for many countres ncludng Inda, Vetnam, Indonesa, Ghana, Yemen, Egypt, Hat, and Ngera. In other words, a worker from one of these countres can expect at the margn hs or her wages to trple or more, solely due to steppng across the US border. Ths wage gap s a margnal effect n two dstnct senses: It s the effect on the wage of the next person who would arrve after a small relaxaton of the mgraton barrer not the effect on the typcal person n the sendng country and t s the margnal effect gven a small change n current levels of mgraton not the general equlbrum wage under fully open borders. The fnal secton relates our results to the three separate lteratures on wage dscrmnaton, border-nduced prce wedges, and the margnal mpacts of antpoverty polces. Researchers measurng each of these would do well to pay much more attenton to restrctons on mgraton; the wage gaps we measure consttute one of the largest known forms of all three. Emprcal estmates n these other lteratures are comparable to ours because they, too, are measured at the margn. 2 Baselne estmates It s obvous that there are large wage dfferences across countres; the queston s the source of those wage dfferences. Some of the dfference mght be due to sector of employment or occupatonal composton across countres (e.g. more lawyers n the US and more small farmers n Inda). But prevous researchers have documented that wage gaps across countres are enormous even for workers n the same sector, such as manufacturng, or n the same narrowly defned low-skll occupatons, such as carpenters, laborers, or bus drvers (World Bank (1995)). Table 1 gves a samplng of these estmates 3

7 from other sources for the countres also n our sample. The rato of real wages n the US to those n Inda for the same low-skll occupaton s somewhere between 5 and 14. Table 1: Prevous estmates of the rato of wages n the US to those n other countres (PPP adjusted), wthout controllng for ndvdual trats Freeman & Oostendorp 3 Rama & Artecona 4 UBS5 Occupaton Carpenter Laborer Industry Laborer Year Medan N Selected countres Bolva Inda Mexco Ngera Turkey N gves the number of countres n the source that 1) have data for both the country n queston and the US, and 2) are one of the 42 countres studed n ths paper. Blank cells ndcate no data for that country. But even n the same sector or occupaton, workers n the US can dffer from those n Inda wth characterstcs (e.g. educaton) that affect ther ntrnsc productvty. The unanswered queston s how much of these observed gaps reflect dfferences n compensaton between otherwse equally productve workers who would wsh to mgrate. Such gaps are (a) the result of wage dscrmnaton (b) nduced by border restrctons (c) whose relaxaton would create large ncome gans for very poor people. Wage ratos adjusted for worker productvty (and compensatng dfferentals) are our ultmate goal. We begn by layng out the assumptons that underle the baselne 3 Freeman and Oostendorp (2005) calculate average monthly wage rates for male workers, n US dollars at Purchasng Power Party, n Carpenter refers to ILO occupaton code 88 ( constructon carpenter ), and laborer refers to ILO code Rama and Artecona (2002) calculate ndustry wages as: Labor cost per worker n manufacturng n current US dollars per year. Includes male and female workers. Calculated as the rato between total compensaton and the number of workers n the manufacturng sector as a whole. Compensaton ncludes drect wages, salares and other remuneraton pad drectly by the employer; plus all employers contrbutons to socal securty programs on behalf of ther employees. Data on labor costs per worker are from plant-level surveys coverng relatvely large frms, mostly n the formal sector of the economy. Fgures are converted nto US dollars usng the average exchange rate for each year. In countres of the former Sovet Unon, the exchange rate of 1989 s used for prevous years. Government wage s Average wage of employees n the central or general government, n current US dollars per year. Includes male and female employees. Calculated dvdng the government payroll by the total number of employees. Data are from government records. Fgures are converted nto US dollars usng the average exchange rate for each year. Both of these are converted to PPP dollars usng the PPP-to-offcal-exchange-rate rato from World Bank (2007). 5 The UBS estmates (Hoefert and Hofer (2007)) are for urban areas (respectvely: Buenos Ares, New Delh, Seoul, Mexco Cty, Manla, Bangkok, and Istanbul, wth the US represented by New York Cty), and show the hourly wage (assumng 50 workng weeks per year) of a buldng laborer, 25 years old, sngle, unsklled or sem-sklled (p. 41) adjusted for cost of lvng n each cty by the prces of 95 goods and 27 servces (p. 8). 4

8 estmates of wage ratos for observably equvalent workers R o, and calculatng these ratos for 42 countres. 2.1 The estmaton problem Our analyss takes the vantage pont of the mgrant destnaton country, thus h ( home ) s the country of destnaton (US) and f ( foregn ) s the country of mgrant orgn. People move untl wages n foregn equal wages n the US but for a factor δ 0, so that 1 + δ w = w. ( ) f h The wedge ( + δ ) ( 1+ )( 1+ ) 1 δ n δ p s the result of two forces. The wedge δ n represents the effect of natural barrers such as credt constrants, transportaton costs, language dfferences, psychc costs of leavng home, and job-search n an unfamlar settng, whch requres a compensatng dfferental to make a mover ndfferent between movng and stayng n the absence of any polcy-based mpedments. The other element, δ p, represents Becker s (1971) dscrmnaton coeffcent the cumulatve result of all polcy barrers such as prce and quantty restrctons on nternatonal movement enacted by governments. 6 The wage of an ndvdual born n home and resdng n home s denoted w hh, where the frst subscrpt denotes country of brth and the second subscrpt denotes country of resdence. That wage s the product of a functon θ hh of a vector of ndvdually specfc observable trats x (schoolng, age, sex, and rural/urban resdence) and a functon φ hh of a vector of unobservable wage determnants x. Thus w ( ) ( hh θhh x φhh x' ). Smlarly, the wage of a person born n foregn and resdng n foregn s w ( ) ( ff θ ff x φ ff x ). We wsh to compare these wages to w ( ) ( fh θ fh x φ fh x ), the wage of a person wth trats x and x born n foregn who has mgrated to home. Ths s captured smply and ntutvely by the rato of home wages to foregn wages for a person born n foregn, R w / w so that: E fh ff w fh [ R ] = E = ( +δ n )( + δ p ) 1 1. (1) w ff The estmaton problem of (1) s that, at any gven moment, the counterfactual w ff s unobservable for those who have mgrated from foregn (e.g. Bolva) to home (the US) as they are workng n the US. 6 It s certanly possble for governments to affect the psychc costs of leavng home, language usage, and so forth. We abstract away from mnor effects of ths knd and consder the man effect of government acton on wage gaps to arse through restrctons on entry and domestc job-search placed upon people who resde n other countres. 5

9 2.2 The selected mgrant Our startng pont to estmate quantty (1) s a set of dentfyng assumptons that we wll later relax. Ths set of assumptons A 1, whch permt unbased measurement of R o for the margnal selected mgrant, are: fh fh, can be approxmated by the observed wage returns to the observable and unobservable trats of those who have already mgrated from foregn to the US, denoted by θˆ fh() and φˆ fh() ; that the unobserved trats x of the typcal mgrant do not dffer from the unobserved trats of the non-mgrant; that the partal assocaton of wages and unobservable trats n foregn s ndependent of the same assocaton n home that the wage returns to mgrants attrbutes, θ ( ) and φ ( ) ( E [ φ ( x ) φ ( x )] = E[ φ ( x )]/ E[ φ ( x )]) fh / hh fh hh ; and that mgraton s costless to the mgrant. In other words, f we restrct ourselves to consderaton of the margnal mgrant, f we assume that the margnal mgrant s unobserved trats are dentcal to those of nonmgrants, and f we assume that the translaton of unobserved trats nto wages happens n ndependent fashon on each sde of the border, then we can estmate Re 1 + δ p. In ths case, equaton (1) reduces to: R e [ 1 ff [ ] ( x) E ˆ φ fh ( x ) ( x) E[ φ ( x )] ˆ θ fh A1 E R A ] =. (2) θ The rght-hand sde of (2) s observable. The result s an estmate of the rato of wages after and before mgraton from foregn to home of the typcal selected mgrant wth observed trats x and unobserved trats x. Notably, t does not assume that observable trats have the same return n foregn and home. We can test whether or not beng born n Ghana or educated n Ghana, for example, have dfferent wage returns n the US labor market relatve to Ghana s. On the other hand, t has mportant dsadvantages: It makes the strong assumpton that the unobserved trats of emgrants from foregn are the same as the unobserved trats of non-mgrants. In partcular, f there s postve selecton on unobservables f the unobserved trats of mgrants contrbute postvely to ther earnngs relatve to nonmgrants then the estmate (2) wll be based upwards by a factor 1 + δ s ( selecton ). Furthermore, nonzero costs of mgraton wll bas estmates of (2) by the wedge 1 + δ n. To the extent that there s selecton and that mgraton s costly, estmates of (2) wll be wage ratos for observably dentcal workers (R o ) rather than for equal-productvty workers wllng to move (R e ).: ff 6

10 R o ( + δ )( 1+ δ ) R = ( 1+ δ )( 1+ δ )( + δ ) = 1 1, (3) n s e n s p that s, a combnaton of natural barrers, selecton, and polcy-nduced wage dscrmnaton. We wll return to the subject of selecton n detal below. It turns out that the effect of ths assumpton on the estmates s closely related to the assumpton of the φ. ndependence of φ ( ) and ( ) 2.3 Data fh x ff x We estmate equaton (2) usng unusually rch and standardzed collecton of ndvdual level data sets on wage-earners compled by the World Bank 7 plus the US Census Publc Use Mcrodata Sample (PUMS) fve percent fle. A seres of steps brngs us from the raw collecton of data sets to the estmaton sample. Frst, we remove all self-employed people and unpad famly workers from the data, leavng only wage-earners. Ths has the advantage of ncreasng the comparablty and accuracy of the earnngs measures, but has the dsadvantage of elmnatng a large porton (though not all) of the nformal sector from the sample especally many agrcultursts n the poorest countres. Second, we remove all people aged 14 or less and all people aged 66 or greater, as well as all people reportng zero wage earnngs. Thrd, we remove data from twelve transton countres because many of these countres were undergong extraordnary nstablty of prces, wages, and currences at the tme the survey was admnstered. Rather than pckng and choosng among them we just elmnated the transton countres as a block. 8 Fourth, we randomly delete US-born US-resdents from the PUMS to reduce the sze of that group from about 6.13 mllon to about half a mllon, due to bndng memory constrants n the mcrocomputer conductng the statstcal analyss (the resultng group was 8.15% of ts orgnal sze, so the samplng weght of each remanng ndvdual was multpled by ). Ffth, we drop Chad from the sample because the sample of US resdents n the publc-use data does not happen to contan any workng-age wage-earners born n Chad. Fnally, we drop Honduras from the sample for reasons descrbed below. The result s a data set wth 2,015,411 wage-earners resdng n 43 countres. Ths comprses 891,158 ndvduals resdng n 42 developng countres, 623,934 ndvduals born n those same 42 developng countres but resdng n the US, and 500,319 ndvduals born n the US and resdng n the US. Each ndvdual record contans the person s wage n 1999 US dollars at Purchasng Power Party, country of resdence, years of schoolng, age, sex, an ndcator of urban or rural resdence, and ndcator varables for the perodcty of the reported wage (weekly, monthly, etc., wth monthly as the base group). For those resdng n the US, there s addtonal nformaton on country of brth and year of arrval for the foregn-born. A samplng weght s assgned to each 7 The sources for all data are gven n the appendx. The basc database s also descrbed n Montenegro and Hrn (2008). 8 The twelve we remove are: Armena, Azerbajan, Bulgara, Belarus, Croata, Hungary, Latva, Moldova, Republc of Macedona, Russa, Romana, and Slovaka. 7

11 observaton ndcatng the number of ndvduals n the natonal populaton represented and s used n all regressons. The US census data were collected for the year 1999 whle the surveys were n the 1990s and early 2000s (only Inda s survey was carred out n 1999). We convert each wage estmate n current year local currency to current year US dollars at Purchasng Power Party usng factors from the World Bank (2007) and then deflate these dollar amounts to 1999 PPP US dollars usng the PPP factor deflator. 9 To the extent that real wages rose (or fell) relatve to the US between 1999 and the year of a country s survey, the wage ratos for those countres wll be slghtly under (or over) estmated. Convertng to PPP also naturally ntroduces the possblty that errors n any gven country s PPP calculaton could affect the results; note, however, that each of the 42 wage ratos we calculate s ndependent of any data from the 41 other countres. Thus any error n any one country s PPP rate does not propagate to the other estmates. By usng PPP exchange rates we assume that all consumpton of the wage gan occurs n the US, whch substantally understates the gans to overall earnngs for mgrant famles, n two ways. Frst, ths gnores remttances. If a worker s n one country wth nuclear famly members n another, and f we assume a untary household utlty functon, then household consumpton should be deflated n the locaton where consumpton occurs. Ths suggests at the least that all remttances should enter the analyss at sendng country prces (offcal exchange rates), not PPP. Second, mgrants, and especally temporary workers, should optmally have very hgh savngs rates. A smple model of ntertemporal consumpton smoothng would suggest that f a worker had access to a much hgher wage rate for an explctly temporary perod they should optmally smooth these wndfall gans over hs or her lfetme. Alternatvely, temporary mgraton s often modeled as drven by target savers who accumulate savngs for a specfc purpose (e.g. a house, busness, car, weddng/marrage), consumpton that agan would occur n ther country of orgn. Much, perhaps most consumpton of the US earnngs of temporary mgrants would be n ther own country, not the US. For nstance, take our medan country, Bolva. The rato of offcal exchange rate (5.81 LCU/$) to PPP (2.09 LCU/PPP$) n 1999 was Assumng all consumpton occurs n the US produces a wage rato for our base case observably dentcal worker of 4 at PPP. If one assumes that half of consumpton s n Bolva the rato would be 7.5 and f only 20 percent of consumpton were n the US (a combnaton of remttances and very hgh propensty to save) the rato would be 9.7. The mean rato of offcal to PPP exchange rates for countres wth GDP per capta under PPP $10,000 n 1999 was 3.7 (larger than Bolva s). At ths rato of offcal to PPP exchange rates even f only 20 percent of mgrant worker earnngs n the US were consumed (through remttances of savngs) n ther home country ths would mean our estmates were understated by a factor of 1.5 and f as much as 50 percent were consumed n ther home country the PPP comparsons 9 After we carred out our analyss the World Bank announced ntentons to retroactvely adjust the PPP factors we use, but these were unavalable at the tme of wrtng. We note, however, that the most mportant adjustments foreseen are those to Inda s and Chna s PPP factors, both of whch wll tend to lower the PPP dollar-value of non-mgrants earnngs and therefore make the wage ratos reported here tend to underestmate the true ratos. 8

12 understate welfare gans by more than a factor of 2. In the sprt of keepng our results as conservatve as possble the paper wll devote a great deal of attenton to adjustments, such as for mgrant selectvty or compensatng dfferentals that scale down the raw estmates of wage dfferentals. But we hghlght the fact that our assumpton that all consumpton occurs n the destnaton country goes a long way toward ensurng that our estmates understate the gaps n real wages we document. A key queston s the relablty and comparablty of reported earnngs. Research comparng multple sources of ncome data at the ndvdual level suggest that selfreported ncome s an unbased estmator of true ncome, both n rch countres (Bound and Krueger (1991)) and n poor countres (Akee (2007a)). There s less certanty about comparablty. Wage data for the US reflect total earnngs from all jobs, whereas wage data for the 42 developng countres n our sample reflect wages from the respondent s prncpal occupaton. For the vast majorty of formal-sector wage earners n the sample we nevertheless expect wage earnngs from the prncpal occupaton to closely reflect total wage earnngs. Furthermore, wage data for the Unted States reflect gross earnngs before taxes, and we expect that most people respondng to a general queston about ther wages or earnngs would have provded gross wages on most of the country surveys, but for a handful of countres t may be that the responses reflect after-tax wages. 10 If respondents provded net-of-tax nstead of gross wages ths would result n some upward bas to our estmated R o. Ths bas wll be small, however, f t s present at all. Formalsector ncome taxes are on the order of 5% n most developng countres (Easterly and Rebelo (1993)). For the medan rato of 3.92, for example, a 5% underestmaton of the denomnator means that the corrected rato s Method We seek estmates of nternatonal wage ratos for equal-productvty workers (R e ); we begn by estmatng wage ratos for observably-dentcal workers (R o ). Frst we compare workers resdng n one par of countres at a tme the US and another country j J estmatng a separate regresson for each country j. We do ths wth three dfferent regresson specfcatons. We dscuss the standard Mncer specfcaton frst snce t s the easest to descrbe, though we rely on t the least: ln w j δ 0 + β0sj 1 r r r δ j + β j sj Ij = X jζ + l l l + ε j δ j β js j I. (4) + j e e e δ j β j s j I + j 10 In a small number of the countres (such as Yemen) the survey explctly requests after-tax earnngs, and n a few of the others (such as Chle) custom may dctate that formal sector wages refer to after-tax earnngs unless otherwse specfed. The text of the wage queston from each survey s n the Appendx. 9

13 Where w j s the wage of person n country j. In the Mncer specfcaton ζ s a vector of coeffcents to be estmated and X j s age, age squared, an ndcator varable for sex, and an ndcator varable for resdence n a rural area. The number of years of schoolng s denoted s j, and each δ and β s a coeffcent to be estmated. The ntercept and slope term on schoolng are allowed to dffer across four values of the ndcator varables that combne country of brth and country of resdence. I (r for resdent of country j) takes the value 1 f ndvdual resdes n country j, or 0 r j l otherwse; these are people born n foregn, resdng n foregn. I j (l for a late arrver) s 1 f ndvdual was born n country j, now resdes n the US, and arrved n the US at or above age 20. I (e for early arrver) takes the value 1 f ndvdual was born n e j country j, now resdes n the US, and arrved n the US below age 20. As the base group s the Unted States, scalar β 0 s the coeffcent on years of schoolng for US-born USresdents, and β r 0 + β j s the coeffcent on schoolng for resdents of country j. We dstngush between early and late arrvers because we do not wsh to assume that a year of schoolng acqured abroad has the same value n the US labor market as a year of schoolng acqured n the US. We assume that most late arrvers receved most of ther educaton n ther countres of brth. Snce we focus on workers wth educaton of 12 years or less, vrtually all of those wth 9 12 years of educaton who arrved at or after age 20 were ndeed educated n ther country of brth. Heckman, Lochner, and Todd (2006) queston the valdty of assumptons underlyng the tradtonal Mncer functonal form, so we also estmate two alternatve specfcatons. The extended Mncer specfcaton ncludes quadratc and cubc terms n years of schoolng, sex, and age (and ther nteractons). 11 We do not present the results from the extended specfcaton, as the estmates of wage ratos based on ths less restrctve functonal form do not dffer substantally from those of the standard Mncer specfcaton. In our second alternatve, the category specfcaton on whch our analyss focuses the columns of X j contan ndcator varables for nne qunquennal age groups, an ndcator varable for sex, and an ndcator varable for resdence n a rural area. In that specfcaton, s j s a vector of fve ndcator varables for dfferent levels of r l educaton based on years completed (1-4, 5-8, 9-12, 13-16, 17-28). Hence the β 0, β j, β j, e and β j are also 5 1 vectors of coeffcents to be estmated. 12 Asde from defnng age and educaton groups, ths last makes no assumptons about the functonal form of the relatonshp between wages and schoolng or age. Ths s partcularly mportant as t 11 Ths addresses the ambguty of functonal form by Stone-Weerstrass approxmaton of the unknown functon wth hgher-order polynomals. Lettng a = age and g = sex, the extended Mncer specfcaton replaces the constant, sex, age, and schoolng terms above wth λ μ ν βλμν g j aj sj. λ= 0 μ= 0 ν= 0 12 The sx schoolng categores are: 1) zero (base group), 2) 1-4 years, 3) 5-8 years, 4) 9-12 years, 5) years, and 6) years. The ten age categores are 1) (base group), 2) 20-24, 3) 25-29, 4) 30-34, 5) 35-39, 6) 40-44, 7) 45-49, 8) 50-54, 9) 55-59, 10) (ntentonally ncludes 65). 10

14 reduces the nfluence of returns to tertary/hgher educaton on the estmates of returns to earler educaton. The raw regresson results from estmatng equaton (4) for the Mncer specfcaton for each of the 42 countres are gven n Appendx Table A1. The results from the category specfcaton are gven n Appendx Table A2. 13 From these blateral regresson results we estmate (2), the rato of the expected wage of an ndvdual wth s years of educaton, born and educated n country j ( late arrval) but now workng n the US to the expected wage of the observably equvalent person workng n the foregn country of brth j: Rˆ o, j ( ˆl r l r δ j ˆ δ j ) + ( ˆ β ˆ j β j )s e (5) l r l where ˆ δ, ˆ δ, ˆ β, and j j j r βˆ j are emprcal estmates of the correspondng parameters. Regresson (4) allows estmaton of other versons of (5) as well, such as comparson of the wage of an early arrval (many of whch receved educaton n the US) to the wage of a foregn resdent. 2.5 Baselne estmates of wage ratos from blateral regressons Table 2 presents a varety of the wage rato estmates. Columns I-III report the results of the regressons usng educaton and age categores, as these are our preferred estmate as they mpose fewer assumptons about the structure of wage profles than the standard Mncer specfcatons. They also compare ndvduals who are the most smlar: foregnborn, foregn-educated (late-arrval) people on ether sde of the border, allowng educaton acqured abroad to have dfferent returns than educaton acqured n the US. Wth ths specfcaton the medan Rˆ o for 9-12 years of schoolng s 3.9. The hghest s Ngera at 13.6 (earnng $1,625/month n the US versus $120/month n Ngera) whle the lowest s Domncan Republc at 1.4 ($1,553 versus $1,137). The medan dfference n PPP-adjusted annual wage earnngs by observably equvalent workers (35-39 year-old formal-sector urban male wage-earners wth 9-12 years of schoolng, born and educated n each country of orgn) s $15,339, from a hgh of $21,000 n Ghana to a low of roughly $5,000 n Domncan Republc. These ratos are robust to changes n the specfcaton of the underlyng regressons. Comparng the results usng a category of 9-12 years of schoolng to the ratos usng the Mncer specfcaton n column IV, computng a rato for each of the years 9 to 12 and takng the geometrc average produces results wth a cross-country correlaton of 0.99 but whch are lower on average by about We also estmate versons of each specfcaton that assume no country-specfc coeffcents on schoolng, whch s to say that they assume that all coeffcents β except β 0 are equal to zero. The raw results of these, and those usng the extended Mncer specfcaton, are avalable upon request. 11

15 The wage prema tend to be modestly lower at hgher levels of educaton (although ths s n ratos; n absolute terms the gap grows). Ths can be attrbuted mechancally to the fact that the partal assocaton of wages n the US labor market and schoolng acqured abroad (medan 6.1% ncrease n wages per year of schoolng) s typcally substantally lower than the assocaton of US wages and US schoolng (medan 12.3%) or the assocaton of foregn wages and foregn schoolng (medan 8.2%). Fnally, column V shows estmates of wage ratos that do not control for any observable trats at all. Note that these are comparable n order of magntude to the low-skll occupaton-specfc ratos from Table 1 that do not control for observable worker trats besdes occupaton or ndustry. Comparng column V wth the rest of Table 2 reveals that controllng for the observable characterstcs of workers accounts for, on average, about 30-40% percent 14 of the observed cross-natonal dfferences n wage ratos. It s not just cross-natonal dfferences n the quantty and qualty of schoolng, but also dfferences n average age and n rural-urban resdence that account for the explaned porton of the varance. Other authors have pooled mcrodata sets on PPP adjusted ncome (or consumpton expendtures) to compute measures of nternatonal personal nequalty. Mlanovc (2008) shows that country fxed effects explan roughly 60 percent of all ncome nequalty across ndvduals n the world, but ths ncludes nequalty due to dfferental access to captal and dfferent levels of human captal. In contrast, our results are specfc to labor ncome for workers wth the same characterstcs. These estmates of R o use regressons smply as a tool for calculatng condtonal means for data descrpton we are only comparng the dstrbutons of wages of workers n gven categores (born outsde the US, male, urban, aged 35-39, wth 9-12 years of schoolng acqured n the country of orgn) on both sdes of the border. On ths level the ratos R o are just factual summary statstcs (ratos of condtonal means) that happen to be robust to a varety of ways of makng adjustments across categores mpled n varous functonal forms. As wth most emprcal work n economcs all the theory comes n decdng how to nterpret these facts, to whch we now turn. 14 Medan /6.20 = 0.366; average, /7.27 =

16 Table 2: Estmates of the wage ratos of observably equvalent workers (male, urban, 35 years old) comparng late arrvers workng n the US versus ther country of brth at varous levels of schoolng Column I II III IV V VI Specfcaton Category Category Category Mncer Raw wage Annual Schoolng Geom. ratos, no dollar gan n Avg controls column I Average $14,999 Medan $15,339 Correlaton wth col. I Ngera $18,068 Hat $18,459 Egypt $19,788 Yemen $18,344 Ghana $21,053 Serra Leone $19,436 Indonesa $17,971 Cameroon $18,883 Vetnam $16,994 Venezuela $16,337 Pakstan $15,381 Inda $16,827 Bangladesh $15,031 Ethopa $17,308 Ecuador (medan) $14,520 Jordan $17,643 Camboda $18,547 Sr Lanka $17,446 Bolva (medan) $16,458 Uganda $17,925 Phlppnes $16,882 Nepal $14,846 Guyana $16,537 Brazl $15,193 Chle $15,297 Panama $15,084 Jamaca $16,881 Peru $14,248 Thaland $12,992 Turkey $11,814 Uruguay $14,307 Colomba $10,505 Guatemala $10,782 Ncaragua $11,447 Morocco $10,009 Mexco $9,180 South Afrca $13,939 Argentna $12,420 Belze $8,597 Paraguay $14,024 Costa Rca $7,548 Domncan Rep $4,991 Sorted n descendng order by column IV, for comparablty wth Table 3. No controls means that the regresson ncludes only country and wage-unt dummes. Mncer s standard Mncer specfcaton. Category uses dummes for fve educaton levels and nne age levels. Both specfcatons nclude nteracton terms allowng dfferent returns to schoolng acqured abroad. The numerator s the predcted wage for US-born (US-born US-resdents), and the denomnator s predcted wage for Foregn-born, late arrval (foregn-born US-resdents who arrved at age 20+). 13

17 3 Boundng the bases of selecton and natural barrers We wsh to know what porton of the ratos R o s the result of polces that prevent the movement of labor. That s, we wsh to know R e, the wage rato for equally productve workers who wsh to move. R e s the relevant measure because (a) t represents wage dscrmnaton as workers would be wllng to work at those wages but are prevented from dong so by government polcy, (b) t s analogous to prce gaps across dentcal goods and servces nduced by borders, and (c) t represents the potental welfare gan to the margnal mover f the polcy nduced costs were removed. Gong back to equaton (3), we want to know what porton of R o represents () unobservable dfferences n productvty between selected mgrants and non-mgrants ( 1 + δ s ), () a compensatng dfferental for natural barrers such as credt constrants, search costs, dstance, language, and unfamlarty ( 1 + δ n ) to make movers wllng to move, and () what porton represents the pure wedge ntroduced by borders ( 1 + δ p ). No one would clam thatδ p = 0, whch would mply that the entre apparatus of control of the US border from vsas at arports and border crossngs to fences and agents was neffectual wth respect to all countres and workers from all countres are just ndfferent between movng and stayng. Ths s obvously counter-factual as both hgh-skll and low-skll vsas to work n the Unted States are vastly oversubscrbed: the entre stock of hgh-skll temporary H-1 work vsas for all of fscal year 2008 was famously exhausted n a few hours on the frst day n 2007 that they became avalable. The same year, the US dversty vsa lottery drew 6.4 mllon applcatons (representng 10 mllon ndvduals) for 50,000 vsas oversubscrbed by a factor of 200 to 1. Drawng upon economc theory and a range of emprcal results n the lterature, we dscuss fve separate ways to place useful bounds on the relatve contrbutons of selecton, natural barrers, and polcy: Smulaton exercses wth a calbrated structural model of selecton A verson of our regressons that elmnates selecton on unobservables, but must be adjusted for nternatonal dfferences n wage-returns to observables Dscusson of the sole extant expermental analyss allowng precse evaluaton of the wage effect of an exogenous change to δ p n solaton Several mgraton flows n whch polcy barrers δ p are roughly zero, allowng measurement of ( 1 + δ s )( 1+ δ n ) Comparson wth predctons based macroeconomc data, where δ 0 Surprsngly, all of these methods gve very smlar results. They suggest that the wage ratos of observably equvalent workers overestmate wage dscrmnaton and by at most a factor of ( 1+ δ s )( 1+ δ n ) 1. 5, resultng n roughly equal measure from selecton δ s and natural barrers δ n. s 14

18 3.1 A calbrated selecton model The key ssue wth selecton s that, whle we have controlled for observables, t may well be that Bolvans workng n the US would have had hgher wages, had they remaned n Bolva, than observably equvalent Bolvans who stayed. Thus whle we observe that our base-case worker earns a monthly wage of $1,831 n the US and $460 n Bolva, t s possble that had the typcal Bolvan-born worker n the US been n Bolva he or she would have made more than $460. Keep n mnd, however, the magntudes at hand. Abstract away from compensatng dfferentals for a moment and suppose that the mgrant worker s hghly selected on unobservable trats, and would have been 50 percent more productve than the average observably-dentcal non-mgrant. Ths large dfference would only reduce the true wage rato from 3.9 to 2.6. Even f the mgrant worker were twce as productve as an observably-dentcal non-mgrant the true wage rato would stll be 2.0. Fgure 1 gves another way to get a feel for the lmted degree to whch selecton can explan the observed ratos R o. For four countres of orgn Hat, Vetnam, Mexco, and Ghana these kernel-densty plots show the dstrbuton of the unexplaned component of wages n the blateral regressons wth the category specfcaton. Four groups are shown: US-born US-resdents ( ˆ0 δ + ˆ ε ), early arrvers ( e j δ ˆ δ j ˆ ˆ0 + + ε ), late arrvers j l ( δ ˆ r δ j ˆ ˆ0 + + ε j ) and foregn resdents ( δ ˆ δ j ˆ ˆ0 + + ε j ). In order for selecton to fully explan the wage gap between Vetnam resdents and Vetnam-born Vetnam-educated US-resdents, two condtons must smultaneously be met: 1) the typcal mmgrant from Vetnam must be drawn from the top 1% of the dstrbuton of unobserved determnants of earnngs n Vetnam or hgher, and 2) there must be zero or negatve correlaton between that person s unobserved component of wages n Vetnam and n the Unted States that s, ths observed characterstc must have rased ther wages n Vetnam but not n the Unted States. For Hat and Mexco, smlar condtons would need to apply: enormously hgh postve selecton on unobservables (top 3-4% of the foregn dstrbuton) and zero or negatve correlaton of the foregn and home unobserved component of wages. These condtons are extreme. 15

19 Fgure 1: Kernel denstes of the unexplaned component of wages Kernel densty HAITI Kernel densty VIETNAM Constant plus resdual from ln(wage) regresson Constant plus resdual from ln(wage) regresson USA born, USA res, USA educ HTI born, HTI res, HTI educ USA born, USA res, USA educ VNM born, VNM res, VNM educ HTI born, USA res, USA educ HTI born, USA res, HTI educ VNM born, USA res, USA educ VNM born, USA res, VNM educ Kernel densty MEXICO Kernel densty GHANA Constant plus resdual from ln(wage) regresson Constant plus resdual from ln(wage) regresson USA born, USA res, USA educ MEX born, MEX res, MEX educ USA born, USA res, USA educ GHA born, GHA res, GHA educ MEX born, USA res, USA educ MEX born, USA res, MEX educ GHA born, USA res, USA educ GHA born, USA res, GHA educ Economc theory can moreover serve to place bounds on the degree of bas ntroduced by selecton. If people move n order to maxmze wages subject to an ndvdual-specfc cost of movement, we can derve the precse degree to whch any gven relatonshp between the returns to unobservables n foregn and home creates a gap between R o and R e for the new mover n response to a margnal relaxaton of mgraton barrers. We do ths va a smple selecton model rooted n the classc model ntroduced by Roy (1951). Let f ln represent the wages of a gven foregn-born ndvdual n foregn (the w ff x mgrant-orgn country), for a gven vector of observables x, and let h ln represent w fh x the wages of the same ndvdual f he or she resdes n home (the mgrant-destnaton country). Indvdual mgrates from foregn to home f h f + c, where c ln( 1+ δ ) where δ 0 s the cost of movng (broadly consdered), expressed n fractons of the ndvdual s foregn wage. 16

20 Suppose now that each person wth observable trats x has home and counterfactual f, N f h, σ, σ, ρ foregn wages that are jontly normally dstrbuted as ( ) π [( ) ( ) ] where ρ s the correlaton. Any level set of ths dstrbuton ( h) π = Π h, f h f, s an ellpse, as n Fgure 2. Indvduals to the upper left of the lne h = f + c move from foregn to home, and the rest do not. The modal new mgrant due to an nfntesmal decrease n c s represented by the tangent pont ( f m, h m ). Usng the equatons of the ellpse and movement threshold, we can calculate precsely the wage gan to ths margnal mgrant. Suppose that the margnal mgrant s US wage h m s observed, but hs foregn wage f m s not observed and s nstead approxmated by f, the average for workers n foregn wth h m equal observable trats. In ths case f the true wage rato m f R e = e s estmated by R hm f o = e, wth a bas of R o R fm f e = e. There are two cases: No selecton or negatve selecton: Where there s no selecton of mgrants or negatve selecton on unobserved trats n the foregn market ( f m f 0), the estmated wage rato underestmates the true rato. The left-hand panel of Fgure 2 shows one such case. Recent evdence from Fernández-Huertas Moraga (2008) ndcates that Mexcan mgrants to the US are modestly negatvely selected on unobserved trats. f m, as n the rght-hand panel of Fgure 2, the estmated wage rato overstates the true rato by Postve selecton: Where there s postve selecton on unobserved trats ( f > 0) where ( γ ( ρ γ )) ( ργ ) ( hm h )Θ R R = e (5) o e Θ 1 and γ σ f σ h and ρ s the correlaton. 15 To begn to evaluate the magntude of bas (5) note that f γ 1 (as for three of the four countres n Fgure 1) then Θ = 1 (as long as ρ < 1) and ( h h R m ) o Re = e. Atγ = (the average γ across the countres) and wth postve selecton overestmaton occurs f and only f h m < h and ρ < Marks (1982) shows that the equaton of ths ellpse s a fˆ 2 bfh ˆˆ + ah ˆ2 = Z, where fˆ f f / σ hˆ 2 2 h h / σ 1/ 1 ρ b 2ρ / 1 ρ, and Z s some constant. ( ) f, ( ) h, a ( ), ( ) The mplct functon theorem yelds yelds dhˆ / dfˆ ( bhˆ 2afˆ ) ( 2ahˆ bfˆ ) the slope of the tangent lne =, whch equals f, such that that γ σ f σ h at any pont ( m h m ) ( 2a bγ ) ( b aγ ) 2 h h Θ γ 2aγ 2a bγ = ( γ ρ γ ) ( 1 ργ ) hˆ m fˆ m = 2. Ths mples that for the margnal new mgrant, f f = ( )Θ b. m m, where ( ) ( ) ( ) 17