Informal Employment in Bolivia: A Lost Proposition?

Transcription

1 Informal Employment n Bolva: A Lost Proposton? Mara Tannur-Panto Donald M. Panto Omar Aras Ths verson: March 15 th, 2004 Abstract We study partcpaton and relatve earnngs n the formal, nformal, and self-employed sectors n Bolva. We estmate quantle earnngs equatons corrected for self-selectvty to address potental bases n the estmates of relatve earnngs gaps due to the endogenety of sector partcpaton. Selectvty s sgnfcant n all three sectors for all three years studed. The benefts of beng more formal lke at low quantles of the nformal sector vansh from 1997 to 2002 as the avalablty of formal jobs decreases. The human captal model s very well ft for 1993 and In 2002 t s best ft for the formal sector where educaton and experence explan much of a worker's earnngs, and worst ft for the self-employed sector where educaton does not play a role and experence s only mportant at hgh quantles. We explot the sem-parametrc nature of quantle regresson to lnk the condtonal returns to worker characterstcs, obtaned from the quantle regressons, wth the poverty status of households to determne the extent to whch unobserved earnngs determnants nteract wth observed characterstcs to penalze non-formal workers n poor households. We fnd that females n non-formal employment suffer the largest penaltes. In unreported results (avalable from the authors upon request) we perform a counterfactual analyss of condtonal earnngs by sector, decomposng the earnngs gaps nto dfferences n endowments of sklls and dfferences n returns to sklls. The results suggest segmentaton between the formal and nformal sector at the lowest condtonal quantles, whle hgher productvty workers seem to have a choce of whch sector to work n. KEYWORDS: earnngs gaps, sample selecton, quantle regresson, multple-choce models. JEL CODES: J23, J24, J31, C14. Mara Tannur-Panto s a professor at the Unversty of Brasla, Brazl, D. Panto s a doctorate student at the Unversty of Pernambuco, Brazl, and Omar Aras s a Senor Economst at the Latn Amerca and Carbbean Department at the World Bank. Correspondng author and presenter: Mara Tannur-Panto. E- mal: tannur@unb.br. Address: SQN 205 Bl. C Apto. 606, Brasíla-DF, , Brazl. We are grateful to Wllam Maloney for valuable dscusson of deas. The opnons and results do not reflect that of the nsttutons to whch the authors are afflated.

2 1 Introducton Bolva has one of the hghest levels of nformal employment n Latn Amercan and the Carbbean. The nformal economy (the salared mcro-enterprse and self-employed sectors) comprsed 68% of the remunerated urban employment n 2002 and over three fourth of jobs for households n the two poorest quntles. Snce the poor derve most of ther ncome from labor assets the nablty of the Bolvan economy to generate suffcent formal sector jobs s often cted as a central factor underlyng the hgh and persstent poverty n the country. In the tradtonal vew the nformal sector s seen as the repostory of jobs for less-advantage workers ratoned out of a superor formal sector, manly as a result of an overly regulated labor market. However, mountng research n the regon has questoned ths vew (see Maloney, 2003). Workers n the nformal economy exhbt great heterogenety. An mportant fracton of nformal jobs may reflect the voluntary choce of workers gven ther preferences, sklls, competng earnngs prospects and job characterstcs. In Bolva, about 44% of employment n the top earnngs quntle s nformal salared or self-employed. For many of these the vew of an ncpent entrepreneur sector wth potental for productvty growth may be more conformng. Ths artcle analyzes the profle of nformal salared and self-employed urban workers n Bolva as well as the factors underlyng dfferences n ther labor market performance compared to formal sector workers located throughout the entre earnngs dstrbuton. We look for characterstcs that can cause nformal and self-employed workers to perform better than ther formal counterparts. We examne the role played by the unobserved (unmeasured) heterogenety of households and ndvduals n determnng labor market performance. To ths am, we estmate sem-parametrc quantle Mncer models corrected for potental selectvty n the sortng of workers across sectors. We capture heterogenety n the returns to characterstcs throughout condtonal quantles of the earnngs dstrbuton and relate t to the poverty status of workers (ther poston n the uncondtonal per capta household ncome dstrbuton). Ths allows us to dentfy the ndvdual characterstcs or specfc determnants of the wage structure of the dfferent sectors whch most contrbute to lower earnngs for the poor. Our focus s on better understandng, among workers partcpatng n urban labor markets, the dfferences between those who fnd jobs n growng sectors and those laggng behnd, payng partcular attenton to the role played by ther unobserved attrbutes. Ths knowledge s mportant to desgn publc polces to better ntegrate nformal workers to the manstream economy and enhance socal protecton schemes. Our results suggest that smplstc segmented market stores are generally at odds wth the observed varety of wage patterns n Bolva. Our fndngs conform to two ters of nformalty. The lower seemngly conssts of workers wth a szable wage dsadvantage, and the upper comprses those wth more dynamc earnngs, some tmes hgher (condtonally) than formal earnngs. The unobserved characterstcs that gve an earnngs advantage appear to be based aganst poor households, partcularly among the female self-employed. Although the dsadvantaged nformal salared stll lag behnd, often they and the self-

3 employed appear at the front lne of earnngs gans. The flow of workers to nformal salared jobs, partcularly those who end up n the lower wage nformal jobs, may be vewed as consstent wth the relatvely better prospects of wage growth n the sector. The paper s organzed as follows. Secton 2 dscusses defntons and the data for the three sectors and years studed. In Secton 3 we descrbe the econometrc methods used n our analyss. Secton 4 contans the emprcal results, and Secton 5 concludes. 2 The Informal Economy n Bolva 2.1 Data and Defntons We use data from the Bolvan 2002 household survey (MECOVI), the 1997 labor force survey and the 1993 household survey (pror to MECOVI) conducted by the Natonal Insttute of Statstcs (INE). Our analyss focuses on workers 15 to 65 years of age, lvng n metro areas and recevng cash remuneratons. Snce we are prmarly concerned wth the ablty of the nformal sector to create good qualty jobs, we focus on frm sze and dstngush three groups of workers: nformal salared workers (workers n establshments wth 1-4 employees plus domestc employees), formal workers (whte-collar and blue-collar workers n establshments wth 5 or more employees), and the self-employed (those who self classfy as cuenta propa, employers of mcro enterprses, and cooperatve workers). Although the self-employed are typcally consdered part of the nformal economy, we treat them as a separate group snce they are not subject to the same job relatons as salared workers (they have no boss, no predetermned workng hours, etc.). TABLE 1 reports average characterstcs for the three groups n 1993, 1997 and The nformal salared sector mostly comprses workers n small shops (servces and commerce), wth no formal contractual arrangement and lackng coverage of benefts such as pensons and health nsurance. On average, they are four to nne years younger than formal workers and the self-employed, have 8 years of schoolng (3 to 4 fewer than formal workers and about the same as the self-employed) and shorter occupaton tenures, and are more lkely to stll lve wth ther parents. As n most of LAC, the Bolvan self-employed comprse a very heterogeneous sector ncludng from street vendors to small artsans under subcontractng producton arrangements. They tend to have hgher potental experence 1, the longest tenure n occupaton, and a hgher fracton wantng to work more hours. Note the smlartes but also some mportant dfferences between the nformal salared and self-employed. Both groups work longer hours per week, although the dfference decreases for the nformal and vanshes for the self-employed n Also, both groups are dsproportonally represented by females (about half) and the ndgenous populaton (44% of the nformal and 61% of the self-employed n 2002). However, despte ther smlar average schoolng, the educaton dstrbuton tends to be more dsperse for the self-employed. Whle a hgher fracton of the self-employed have no completed basc educaton (23% vs. 16% n 1 To avod overstatng the potental experence of the low educated ndvduals, we defne potental experence as mn {age-educaton-6, age-14}.

4 2002) the share wth unversty educaton s twce as hgh than among the nformal salared (8% and 3% n 2002). Medan hourly earnngs n 2002 were 70 to 90 percent hgher n the formal sector than for the self-employed and nformal salared. 2 Note that total earnngs are lower for the nformal salared than for the self-employed, despte workng an average of 5 hours more per week. Thus, t s relatvely harder to escape poverty through nformal employment. Durng the perod of economc boom (1993 to 1997) the self-employed were better off than the nformal salared, but ths pattern reverted durng the years of recesson (1998 to 2002). However, as stressed n the recent lterature, these uncondtonal average hourly earnngs cannot be used as evdence of the superorty of formal sector jobs. Frst, t cannot be clamed that the lower nformal earnngs are due to the characterstcs of nformal jobs rather than to the productve attrbutes of the workers (both observed and unobserved). One should compare earnngs of Bolvan workers wth smlar characterstcs, both observed and unobserved. Second, gaps n average earnngs can hardly characterze the stuaton of workers at all ponts of the earnngs scale. Average earnngs gaps may mask the dfferental stuaton faced by Bolvans whose unobserved characterstcs place them below or above the condtonal mean wage functon. Thrd, and more mportantly, monetary earnngs gaps do not fully capture dfferences n the qualty of jobs across sectors n so far characterstcs such as flexble work schedules, the degree of protecton and non-monetary benefts (e.g., health nsurance, tranng) are also valued (dfferently) by ndvduals. These need to be factored n as part of the cost-beneft choce calculaton of workers. We attempt to address these ssues below by gong beyond narrow average, cross-sectonal earnngs comparsons and a more careful sem-parametrc, multvarate analyss of earnngs equatons corrected for selectvty. We next compare the entre uncondtonal earnngs dstrbutons for formal, nformal salared workers and the self-employed (FIGURE 1A). It s clear that medan earnngs mask substantal dspartes between workers at dfferent ponts of the earnngs scale. The dstrbuton for formal sector workers s further to the rght of the nformal salared reflectng ther wage advantage at any wage level, and n fact, the dstrbutons separate further at the rght tal-- earnngs gaps are larger between workers at jobs wth hgher pay. Meanwhle the formal salared dstrbuton jons the self-employed dstrbuton at the rght tal. A worker at the 0.10 quantle of the dstrbuton for the formal salared (whose wage places hm/her above 10% of formal workers) earns about 122 percent more than a worker at the 0.10 quantle of the dstrbuton of the self-employed n The earnngs gap s then reduced to 33 percent for the self-employed at the 0.90 quantle and goes to zero as we move to hgher quantles (the hghest pad self-employed earn wages smlar to the hghest pad formal employees). These earnngs gaps n part arse from dfferences n productvty-related characterstcs of workers. We estmate (but do not report) the gaps for workers at dfferent ponts of the sector-specfc wage dstrbutons condtonng on observed factors and adjustng for dfferences n sector partcpaton, that s, the earnngs dstrbutons that would result f workers had the same set of measured characterstcs and worked n a dfferent sector. 2 Log of hourly earnngs are calculated from all reported labor earnngs from the man job (net salary plus ndrect salares and benefts) and the number of hours worked n the prevous week.

5 2.2 A Prmer on Regresson We use quantle regressons (Koenker and Bassett, 1978) to estmate earnngs and return gaps between formal and nformal workers at dfferent ponts of the condtonal earnngs dstrbuton. Just as least squares models the mean of the dstrbuton of the dependent varable Y condtonal on the regressors Z, quantle regressons gve models for dfferent percentles of ths dstrbuton. The τ-th quantle of Y condtonal on Z s gven by: Q τ ( Y Z ) = Z í φ() τ where the coeffcent φ(τ) s the slope of the quantle lne gvng the effect of changes n Z on the τ-th condtonal quantle of Y. Estmaton for dfferent quantles (τ from 0 to 1) yelds regresson lnes for varous percentles of the condtonal dstrbuton of Y such that at least a τ proporton of regresson resduals are below the estmated regresson lne and approxmately a (1-τ) fracton are above t. For nstance, medan regresson (τ = 0.5) splts the sample n half (half of the resduals above and below the regresson lne) and gves the same results as ordnary least squares when the dstrbuton s symmetrc. FIGURE 1B captures the basc ntuton of our approach. We compute the dfference n the ntercepts and educaton coeffcents from the estmated quantle Mncer functons for formal and nformal workers located at the same quantle of the condtonal dstrbuton of each sector. Thus, we examne: Q ln w e, X Q lnw e, X = α τ α τ + β τ β τ e + θ τ θ τ τ ( ) ( ) ( () ()) () () f τ f ( ) ( () ()) X coeffcents have the usual regresson nterpretaton. For example, takng τ = 0.9, α ( 0.9) α ( 0.9) f (the dstance A-A ) gves the sectoral gap n the level of wages for uneducated workers at the 90 th quantle of the condtonal wage dstrbuton of each group, that s, the dfference between the wage floor of the best pad 10% of uneducated formal and the wage floor of the top 10% of uneducated nformal (for any gven X). Smlarly, α.1 α 0.1 measures the adjusted wage gap at the 10 th quantle of the condtonal ( 0 ) ( ) f dstrbutons (dstance C-C ). Meanwhle β ( 0.9) f s the slope of the Mncer regresson lne ftted through the 90 th condtonal quantles, and as s conventonal refers to the return to educaton at ths quantle. It gves the percentage change n the wage floor of the best-pad 10% of formal workers (wthn each observed skll level) from an addtonal year of schoolng. Thus, β ( 0.9) β ( 0.9) gves the gap n the returns to educaton between formal f and nformal workers at ths quantle. In the case of dummy varables, each coeffcent measures the log earnngs dfference between a worker wth the partcular characterstc (e.g., secondary educaton) and an otherwse smlar worker wth the excluded category (e.g., basc educaton) at the same condtonal quantle. The ant-log of the coeffcents (mnus 1) gve the relatve (adjusted) earnngs percentage gap of hgh school workers wth respect to those wth only basc educaton at each gven quantle.

6 For example, the college educaton dummy for the 10 th percentle (τ = 0.1) gves the ncome per capta gap between households wth less than basc and college educated workers located at the 10 th percentle of the condtonal ncome dstrbuton, that s, the dfference between the earnngs floor of the bottom 10 percent of college educated workers and the floor of the bottom 10 percent of workers wth less than basc educaton (controllng for other explanatory characterstcs). Smlarly, the coeffcent at the 50 th percentle measures the college earnngs premum at the medan earnngs of the two condtonal dstrbutons. In the case of contnuous regressors, the coeffcent measures the conventonal slope of the regresson lne ftted through a gven condtonal quantle. It s mportant to stress that ths nterpretaton pertans to a condtonal analyss where confoundng effects on ncome per capta arsng from the correlaton of the varous household characterstcs are beng solated. We can thnk of bottom condtonal quantles as pertanng to workers wth wages lower than granted by ther educaton, experence level and other measured wage determnants, and the upper quantles to workers wth wages hgher than predcted by observed sklls. The relatve postonng of workers n the condtonal wage dstrbuton can be related to dfferences n "ablty", whch may nclude a worker's labor market connectons, famly human captal, school qualty, and/or work ethcs (Aras, Hallock and Sosa, 2001). The nterplay of ths unobserved heterogenety wth each regressor results n regresson coeffcents that vary across quantles 3 Econometrc Approach Mncer earnngs equatons (Koenker and Bassett, 1978) and multnomal choce models are estmated by a two-stage procedure (Ftzenberger, 2003 and Buchnsky, 1998). We use a smlar approach to the one proposed by Ftzenberger (2003) for multple choce sample selecton models and quantle regresson. In the frst stage we determne the probablty of partcpaton n the formal/nformal/self-employed sectors. In the second stage, we correct sector-specfc earnngs equatons for selecton bas, caused by unobservable characterstcs that cause partcpants to jon a sector even when they have a low probablty of beng n that sector. 3.1 Selectvty n Regresson There are few approaches avalable n the lterature for selectvty correcton n quantle regresson. The method developed by Buchnsky (1998) allows for frst-stage semparametrc estmaton of the partcpaton equaton usng Ichmura s (1993) SLS estmator. He derves the small sample propertes of the second-stage quantle regresson estmator for the case when one accounts for sample selecton by ncludng a polynomal expanson n the nverse Mll s rato n the quantle models. Ths methodology does not rely on parametrc assumptons about the resduals nether n the frst nor n the second-stage estmatons. However, only bnary choce models can be estmated n the frst-stage. Also adoptng a two-stage approach, Ftzenberger (2003) makes use of a methodology that accommodates polychotomous choce problems n the frst-stage partcpaton equaton. He estmates lnear probablty models usng seemngly unrelated regressons (SUR) for a three choce model (full-tme/part-tme/non-employment). In the second-stage quantle wage

7 equatons, he uses a second order polynomal expanson n the estmated probabltes to correct for selectvty. Fnally, lackng the analytcal dervaton of the covarance matrx of the coeffcents, he bootstraps the standard errors. We employ a two-stage method n whch we estmate the frst-stage partcpaton decson wth a multnomal logt. Ths does not free us from makng parametrc assumptons on the frst-stage resduals; however, t does allow us to consder smultanety n the choce of sector. We rely on n -consstency of the frst-stage estmator and on dentfcaton assumptons as n Newey (1988). In the second-stage quantle Mncer equatons, we nclude a polynomal expanson n the multnomal equvalent of the nverse Mll s rato and bootstrap the standard errors Why Correct for Selectvty Bas? Selectvty bas plays an mportant role n the estmaton of wage equatons n the formal, nformal salared, and self-employed sectors. Wthout correcton, based estmates of mportant varable such as educaton may be obtaned. For example, a worker wth characterstcs typcal of a formal worker (a hgh level of educaton, perhaps) has a low probablty of workng n the nformal sector. However, f such a worker does choose to work n an nformal job t s lkely because an excellent offer was made (remember, small frms may stll be hghly productve). Wthout correctng for selectvty, the estmates of the return to ths formal-lke workers level of educaton would be based upward, n the nformal sector. Snce we correct for selectvty, ths effect would appear as a postve coeffcent for the term λ.f (as observed n 1993 and 1997) and the educaton coeffcent would reman unbased. 3.2 Two-stage Estmaton The varable that ndcates the sector n whch worker s employed, I, takes the values: 1 for the formal, 2 for the nformal, and 3 for the self-employed sector. For ths multnomal choce model we have the followng equatons for the latent ndces: * (1) Is = z γ s + ηs, where s = 1,2, 3 specfes the sector and = 1,2, K, N specfes the observaton. The employment sector s determned by: (2) I = s ff z ( γ s γ j ) > η j ηs for all j s. We assume that η s are ndependently and dentcally dstrbuted wthn the sector (but not between sectors), wth the type I extreme-value dstrbuton. Ths assumpton allows the estmaton of γ s va maxmum lkelhood multnomal logt estmaton. To estmate the multnomal logt, we must choose the base category for whch the coeffcents are set to zero. We choose to exclude the formal sector, whch mples γ = 1 0. Followng Maddala (1983, Secton 9.5), f one wrtes a mean wage equaton for sector 1 as y1 = x β 1 + u1, the selecton bas nduced by the fact that y 1 s only observed when I = 1 can be corrected. Ths correcton requres assumptons about the covarance between the

8 resduals from the multnomal logt and the resduals from the wage equaton, whch we do not dscuss here (see Maddala, 1983). Frst defneωj η j η whch have the multvarate logstc dstrbuton, and tj z ( γ γ j ) such that the nequaltes n equaton (2) becomeω < t. The expectaton of the wage equaton can then be wrtten as: sj sj (3) E( y < = 1 ω 1 j t1 j) x β1 + λ j f j ( t12, t13), 3 j = 2 where f ( ) can be calculated explctly (see Maddala, 1983). j Here, we consder quantle wage equatons for the varous sectors, s : * * (4) Quantτ ( ys x = x ) = x β sτ + usτ = x βsτ + hsτ ( I2, I3) + ε sτ, where x s a subset of z, and Quant ( u x = x, I = s) 0, leadng to based estmates of τ sτ β s τ f we do not correct for selectvty. However, the ncluson of the term * * (5) hs τ ( I2, I3) Quantτ ( us τ z = z, I = s) n the wage equaton allows unbased estmates of β s τ snce Quantτ ( ε sτ z = z, I = s) = 0, by defnton. Ths term ncludes nformaton about the unobservable characterstcs of workers that affect ther choce of sector to work n (note the condtonalty on z and not x ). * * Unfortunately, we have no closed form for hs τ ( I2, I3) and we must resort to approxmatons. We follow Ftzenberger (2003) and Buchnsky (1998) and approxmate * * hs τ ( I2, I3 ) by a power seres whose coeffcents are to be estmated by the regresson. We choose to expand n lnear terms of the functons f j ( ) n equaton (3). In partcular, (6) h sτ ( tsj, tsk ) λ sτ + λs τj fs, j + λs τk fs, k where s j, s k, j < k. Our restrcton to lnear terms s made based on cross-valdaton technques whch demonstrate the worsenng of the model s ft as hgher order terms are ncluded, or f no selectvty s ncluded at all (Newey, Powell and Walker, 1990; and Newey 1988). Newey (1988) and Buchnsky (1998) show that the second stage estmates are consstent f the frst stage estmates are. For clarty we dscuss these functons for the formal sector. The functon f s the 1, 2 multnomal equvalent of the nverse Mll's rato. It ncreases monotoncally as the probablty of beng nformal ncreases. It therefore has very small values when the probablty of beng self-employed or formal s hgh. Meanwhle f represents the same for 1, 3 ncreasng probabltes of self-employment. If λ1, 2 s postve t ndcates that ndvduals wth more nformal lke characterstcs that stll jon the formal sector receve a premum for jonng. The coeffcent λ 1, 3 has a smlar nterpretaton for ndvduals wth more selfemployed lke characterstcs that stll are part of the formal sector. Fnally, we obtan the standard errors for by bootstrappng. β s τ

9 3.3 Lnkng Poverty and Condtonal s As dscussed above, quantle regresson estmates address heterogenety n the condtonal earnngs dstrbuton. We would lke to lnk these condtonal returns to workers' postons n the uncondtonal per capta household ncome dstrbuton. To do ths we frst dentfy the condtonal quantle of each worker. We perform quantle regressons at the 0.05, 0.15,..., 0.95 quantles and then dentfy at whch quantle each worker had the smallest resdual. Mathematcally, the condtonal quantle of worker n sector s, gven by θ., s determned by θ = argmn( ε ) where τ = 0.05,0.15,0.25, K,0.95. Ths allows us to dentfy the returns to characterstcs receved by each ndvdual, β sθ, as well as the returns they would receve at the same condtonal quantle n another sector, j, ( β jθ ) and therefore the penalty (or beneft) that they receve for each of ther characterstcs by not workng n the formal sector, P = β. β τ sτ β sθ 1θ Gven the sector penalty, P β, and the uncondtonal quntle of per capta household ncome, C, for each worker, we regress the penalty on ndcator varables of the uncondtonal quntles. The regressons are weghted by the nverse of the square of the standard errors from the quantle regresson estmatons. The regresson sample s composed of non-formal workers only, snce formal workers have penaltes equal to zero. Ths exercse allows us to determne f poorer households are more severely hurt by partcpaton n the nformal or self-employed sectors and whch characterstcs are most damagng. 4 Emprcal Fndngs Because the man focus of ths verson of the paper s on the econometrc methods, such as selectvty correcton for multnomal choce models n quantle regresson and the lnk between condtonal returns and uncondtonal earnngs through quantle regresson, we only present the most mportant results. We have estmatons for 1993, 1997 and 2002 and n some cases we may show results just for one year, beng the others beng avalable upon request. 4.1 Probablty Model We model the partcpaton decson of ndvduals to engage n the formal, nformal or self-employment sector usng a multnomal logt model. We choose selecton (dentfyng) varables that may affect the decson to partcpate n a gven sector and not affect the earnngs receved n that sector. Both the probablty model and the earnngs equaton model nclude: potental experence (and squared); experence n occupaton (and squared); ndcator varables for levels of educaton basc, prmary, secondary, techncal and unversty 3 ; dummes for ethncty and female; an ndcator for ndvduals who work less than 20 hours per week; dummes for sector of actvty (wth commerce as the excluded category) and department (La Paz s the excluded category). The probablty model s further dentfed by 3 The category less than basc represents years of educaton 4; basc represents 5 years of educaton 7; and prmary ndcates 8 years of educaton 11. The defnton of secondary, techncal and unversty depends on the type of educaton attanment.

10 the ncluson of the followng varables: ndcators for martal status and school attendance, an ndcator for head of household and nteracton wth female; number of kds less than 15; number of kds less than 6 nteracted wth female; number of elderly n the household nteracted wth female; an ndcator for lvng wth workng parents; other famly per capta monthly ncome (dvded by 1000); an ndcator for other household member workng n the formal sector; and ndcators for wantng (and not) and beng able (and not) to work more, wth the excluded category beng not wantng and not beng able to work more. Ths model specfcaton vares a lttle throughout the analyzed years due to avalablty of the varables. The results for the multnomal model for 1993 are presented n TABLE 2C. We use the formal sector as the comparson group and report relatve rsk ratos. Experence enters the model both lnearly and quadratcally and therefore t s a lttle trcky to nterpret. When t s sgnfcant, potental experence 4 decreases the probablty of beng nformal relatvely to formal, but t ncreases the probablty of beng self-employed. The effect of occupaton specfc experence (years n occupaton) also depends on the evaluaton pont 5. Years n occupaton have a weaker effect on the probabltes of beng nformal (negatve) and selfemployed (postve). Educaton attanment decreases the probablty of beng both nformal and self-employed. In 1993, havng prmary or hgher levels of educaton (compared to the excluded category less than basc) would decrease substantally the probablty of nformalty 6. In 1993, ether basc and prmary or techncal and unversty educaton would reduce the probablty of selfemployment. The educaton varables have a stronger negatve mpact on the partcpaton n the nformal salared sector than on the self-employed sector for all years analyzed. Belongng to an ethnc group ncreases substantally the chances of self-employment n 1993, 1997, and 2002, although t s only sgnfcant (barely) and negatve for the nformal salared n 1993 and Beng a female has a strong, postve and sgnfcant effect on the probabltes of nformalty and self-employment. Beng marred, however, decreases the chances of nformalty for all years, whle ncreasng the chances of self-employment (only n 1993). Even though the number of kds less than 15 years old n the famly s only sgnfcant for the nformal n 1993, ncreasng the probablty of nformalty, havng young kds (less than 6 years old) ncreases the probablty of women beng self-employed n It shows that self-employment mght be an opton for ndvduals who need flexble workng hours. Beng the head of household or havng another member of the household that s formal ncrease the probablty of formalty. Fnally, the surveys ask questons about the wllngness and avalablty to work more hours and they turn to be very mportant selecton varables. If an ndvdual wants and s able to work more hours t ncreases hs chance of beng self-employed for all three years, relatvely to the excluded category (not wantng and not beng able to work more). A smlar mpact s also found for the nformal sector n 1993 and We also fnd that the support structure of famles, measured by lvng wth parents and other famly per capta ncome can affect the 4 Potental experence s only barely sgnfcant for the self-employed n 2002 and not sgnfcant at all for the nformal n Ths varable does not exst n the 1997 survey. 6 In 1997 and 2002, only educatonal levels hgher than secondary have sgnfcant coeffcents.

11 partcpaton decson. Other famly per capta ncome reduces the chances of nformalty n 1993 and Lvng wth parents defntely decreases the probablty of nformalty for 1993 and 1997 and of self-employment n Earnngs Equatons We proceed wth the second-stage estmatons, ncorporatng selectvty polynomals (expansons n the multnomal equvalent of the Inverse Mll s rato) n the quantle earnngs models. TABLE 3C presents the results for the formal, nformal and self-employed sectors (at the 10 th, 50 th and 90 th condtonal quantles) n 1993, but an overvew of the quantle results for all three years are presented on FIGURES 2A-2C, 3A-3C and 4A-4C. The frst thng to observe s that all selectvty polynomals are sgnfcant as a group for all quantles of the formal sector n 1993, the 10 th and the 50 th n 1997 and the 10 th and the 90 th n 2002, ndcatng that unobservable characterstcs are relevant n determnng earnngs n ths sector. Beyond ths, formal specfcaton tests (Newey, Powell, and..., 1990) demonstrate that the coeffcents from the quantle regressons would be based f one dd not correct for selectvty (.e. the null hypothess of no change n coeffcents between the models wth and wthout selectvty correcton s rejected). The selectvty polynomal s not sgnfcant at the 10 th quantle of the nformal sector n 1997 and 2002, nor at the 90 th quantle n It s also not sgnfcant at any quantle n 2002, nor at the 90 th quantle n 1993 for the selfemployed sector Analyss of Selectvty In the formal sector, the ndvdual coeffcents of the selectvty polynomal can be nterpreted as follows. The coeffcents of lambda.f. (coeffcent of λ 1, 2 ) are negatve and sgnfcant and lambda.f.se (coeffcent of λ 1, 3 ) are postve and sgnfcant for all quantles of 1993 and the 90 th quantle of Ths ndcates that as a worker's probablty of beng formal decreases (whle stll beng formal), unobservables wll cause her to earn less than expected based on her observed level of human captal (wage equaton characterstcs) f she has more nformal lke characterstcs (negatve selecton). However, she wll make more than expected f she has more self-employed lke characterstcs (postve selecton). Ths suggests that the opportunty cost for a worker wth nformal lke characterstcs to work n the nformal sector nstead of the formal sector s reduced by the effect of unobservables n 1993, whereas the opposte holds for a worker wth self-employed lke characterstcs. However, n 2002, ths trend s nearly reversed at the 10 th quantle as lambda.f. becomes large and postve whle lambda.f.se becomes large and negatve. The mantaned sgn at hgh quantles and the reversal of sgn at the 10 th quantle for the lambda.f.se term suggests that workers wth self-employed lke characterstcs nclude both hgh ablty entrepreneurs and, ncreasngly snce the start of the perod of economc stagnaton, low ablty workers. The change n sgn of the lambda.f. term at the 10 th quantle from negatve to postve and nearly sgnfcant suggests that the lowest productvty formal workers wth nformal lke characterstcs left (or were forced to leave) the sector, thereby causng the postve selecton bas at low condtonal quantles. For the self-employed, the frst reported coeffcent s lambda.se.f (coeffcent of λ 3, 1) and t s always postve when sgnfcant as the second lambda.se. (coeffcent of λ 3, 2 ) s always negatve when sgnfcant (1993 and 1997). Ths means that t s advantageous to look more

12 formal lke than nformal lke n the self-employed sector. Ths corroborates the dea that the self-employed sector s made up of well performng rsk takng entrepreneurs who are postvely selected nto the sector (those wth formal-lke characterstcs) and those who are unable to fnd any other job and are forced to set up shops for themselves on the sde of the road (those wth nformal-lke characterstcs). The coeffcents become nsgnfcant n 2002, although the sgns reman the same wth only slghtly smaller magntude, suggestng that the shfts n sector composton from 1997 to 2002 placed some formal-lke workers wthout the rsk takng entrepreneur sprt n the self-employed sector. One strkng result from our analyss of the nformal sector s that lambda..f (coeffcent λ ) goes from beng postve and sgnfcant n 1993 and 1997 to beng negatve (though of 2, 1 nsgnfcant) at the 10 th and 90 th quantles n Ths may ndcate that durng the boom from 1993 to 1997 when formal jobs were plentful, workers wth formal-lke characterstcs only accepted those nformal jobs whch offered the best opportuntes, causng postve selecton nto the sector (remember that frm sze s only a proxy for frm productvty). Ths agrees wth the nterpretaton of lambda.f. gven above. However, the reducton of the formal sector from 1997 to 2002 may be drvng the loss of ths postve self selecton as lower productvty workers from the formal sector lose ther jobs and move to the nformal salared sector. The term lambda..se (coeffcent of λ 2, 3 ) s negatve and sgnfcant only at the 50 th and 90 th quantles n 1993, suggestng a negatve selecton of self-employed lke workers nto the nformal sector durng those years. Ths coeffcent remans negatve (though nsgnfcant) at the 10 th quantle throughout all the years, whle becomng large and postve (though nsgnfcant) at the 90 th quantle n 2002 suggestng that hgh ablty workers wth selfemployed lke characterstcs are enterng the nformal sector Analyss of the Earnngs Equatons FIGURES 1A, 2A and 3A show plots of some selected quantle regresson coeffcents for the formal, nformal and self-employed sectors n Returns to potental experence decrease wth quantle n all three sectors n In 1997, returns to experence are steady through quantles for the formal and self-employed sectors, and t s consderably hgher at the 90th quantle of the nformal sector. Potental experence s not sgnfcant at the 90th quantle of the formal sector n 1993 and 2002, at the 90th quantle of the self-employed n 1993 and 1997, and t s not sgnfcant at all for the self-employed n Ths may be a result of the many older uneducated workers wth exaggeratedly large levels of potental experence n the self-employed sector. Returns to tenure at the occupaton play an mportant role to the self-employed n 1993 and to the nformal salared n 2002, wth the largest returns comng n the nformal sector and the smallest n the self-employed sector. In 1993 and 1997, educaton s a sgnfcant determnant of earnngs n all three sectors, however, n 2002 t only plays a role for the formal sector. The great excepton s the returns to unversty n 1993 and 1997, whch ncrease wth quantle and present smlar magntude n all three sectors. Returns to techncal educaton are never sgnfcant for the self-employed sector 7, ndcatng that nvestments n ths type of educaton would only beneft formal and nformal salared workers. 7 It s negatve and sgnfcant for the 90th quantle of the self-employed n 1993.

13 One's beng ethnc (ndgenous) severely hurts n the formal sector, but t does not affect earnngs n the nformal sector. For the self-employed, penalty for ethnc background s only sgnfcant at medan and hgh quantles n 1993 and low and medan quantles n Women are severely penalzed throughout the self-employed dstrbuton wth the worst penaltes at the lowest quantles. In 1993, women also receve lower earnngs at the formal and nformal sectors, but the penaltes are hgher at the top of the condtonal earnngs dstrbuton. In 2002, the human captal model s well ft for the formal sector and poorly ft for the other sectors. FIGURES 1B, 2B and 3B present the coeffcents for the formal, nformal and self-employed sectors n Returns to potental experence vary sgnfcantly across quantles only n the self-employed sector as t becomes nsgnfcant at hgher quantles. The formal and nformal sectors receve the hghest returns. In the formal sector, returns to basc, prmary, and techncal educaton decrease wth quantle, whle all other returns reman constant. Perhaps one's earnngs are capped wthout a unversty educaton or some forms of educaton substtute for ablty, rather than proxy for t. In the nformal sector, for all but basc and prmary educaton, the premum s larger at hgh decles, becomng larger than the formal premums for techncal and unversty educaton. We observe a smlar behavor n the selfemployed sector for secondary educaton, though the premums reman smaller than the formal sector's. In contrast to 2002, educaton and potental experence are sgnfcant determnants of earnngs for all sectors, not just the formal one. In 1997, ethncty was not penalzed n the nformal sector. It hurts throughout the formal sector, worsenng wth ncreasng quantle, and s penalzed at the low quantles of the selfemployed sector. In 2002, the ethnc penaltes were not sgnfcant at any quantle of the selfemployed sector, though they were large and nearly sgnfcant at the 90 th percentle. Beng female s only penalzed n the self-employed sector. Ths penalty does not sgnfcantly vary across quantle (unlke 2002 where the penalty decreased wth ncreasng quantle). In the formal sector, there s no female earnngs dsadvantage n 1997, whereas n 2002 the hghest quantles suffered ths penalty. The human captal models were well specfed for all sectors n 1997, rather than just the formal as n The plots for 1993 are n FIGURES 1C, 2C and 3C. Returns to potental experence vary sgnfcantly across quantles only n the self-employed sector as t becomes nsgnfcant at hgher quantles, as n Returns to tenure at occupaton decrease wth quantle n the formal sector and are not sgnfcantly dfferent across quantles n the other sectors. Returns to educaton, when sgnfcant, are constant throughout quantles for the formal sector, except for unversty for whch the returns ncrease wth quantle. In the nformal sector any sgnfcant returns to educaton are constant throughout quantles. For the self-employed sector, returns to educaton ncrease wth quantle for techncal, secondary, and unversty educaton. Ethncty s penalzed: steadly throughout the formal sector (exceptng a slght ncrease at the 90 th quantle); ncreasngly wth quantle n the self-employed sector; and not at all n the nformal sector. The female earnngs dsadvantage s steady throughout quantles of all sectors (wth perhaps a slght ncrease at the 90 th quantle of the formal sector) and s largest for the self-employed as n 1997 and 2002.

14 4.3 Lnkng Poverty and Condtonal s TABLES 4 through 10 contan the results from the analyss where we lnk condtonal returns and household poverty status. The dfferences are measured relatvely to the formal sector. TABLE 4 shows the penalty for beng a non-formal female s much larger n the poorest households throughout all years ( (16%) n 2002, obtaned by summng the constant and the result for quntle 1). TABLE 5 shows the dfference n returns to ethncty (ndgenous) n non-formal sectors. In fact, there s a premum for ethnc workers n the nformal sector and the poorest households beneft the most. Ths premum ncreased from 1993 to The remanng fgures (7 to 11) treat the dfferences assocated wth returns to educaton. In 2002, among the workers wth basc educaton (21% of nformal and 23% of self-employed) only the poorest ones receved a weak penalty (around -0.03, see TABLE 6). If we break down the results for the nformal salared and self-employed wth basc educaton, the former receve a premum, whle the latter a penalty. Ths may reflect the effects of the recesson n shortenng the salared jobs n The 1993 and 1997 regressons (TABLE 6) showed a premum for workers wth basc educaton n the nformal sector, but the poorest ones beneft the least. For workers wth prmary educaton (30% of nformal and 23% of selfemployed n 2002) there s a small premum for the rchest self-employed (0.04, see TABLE 7) and a penalty (-0.16) for the rchest nformal salared. In both cases, the poorest ones are more benefted and less hurt, respectvely. Small penaltes for the poorest households are consstent wth the results for 1997 and 1993, however n these years wealther households receved slghtly hgher relatve returns to a prmary educaton n non-formal employment (see TABLE 7). In 2002, 26% of nformal workers and 19% of self-employed workers had secondary educaton. Havng secondary educaton helps the poorest non-formal workers, but hurts the rchest ones n 2002 (see TABLE 8). Ths s n contrast to 1997 and 1993 when poorer households receved larger penaltes for returns to secondary educaton n nonformal employment (see TABLE 8). The remanng educatonal categores (techncal and unversty) account for only 7% and 12% of the nformal and self-employed workers, respectvely and tend to be concentrated n the wealther households. The regressons for 2002 and 1997 show qute large penaltes for non-formal techncally educated workers (around -0.6 n 2002 and n 1997 for the rchest ones, see TABLE 9). In 1993 and 1997 the poorest non-formal workers wth techncal educaton are worse off than the rchest ones (-0.33 and -0.55). Fnally, there are huge penaltes for unversty educated workers n the nformal and self-employed sectors relatve to ther formal counterparts n 1997 and 2002 (around -0.8 (55%) n 2002, see TABLE 10). The penaltes were a bt smaller for the poorest self-employed n 1997, but the poorest nformal salared wth unversty educaton were severely hurt n Ths may be related to the economc stagnaton from 1997 to 2002, whch faled, to provde a suffcent number of formal jobs for unversty educated workers. In summary, ethncty s the only characterstc whch yelds hgher returns n the nformal sector for poor workers for all years studed. Among the educaton categores, only prmary educaton benefted the workers n the poorest households n 2002, whle n 1997 t s basc educaton that benefts the poorest non-formal workers. In 1993, basc, prmary and secondary educaton levels provde premums for the poorest non-formal workers, relatve

15 to returns n the formal sector. It s mportant to pont out that those penaltes/premums are calculated based on returns to characterstcs, wthout takng nto account the magntude of the models' ntercepts, whch could make the condtonal earnngs gaps behave n a dfferent, sometmes contradctory way. 5 Concluson Labor partcpaton and earnngs follow dfferent patterns n each of the formal, nformal and self-employment sectors n urban Bolva. The choce of self-employment s certanly of some utlty for ndvduals wth restrcted workng hours, such as women wth young chldren, as more than 70% of self-employed workers who work less than 20 hours per week are female. Ths may explan the lower earnngs they are wllng to accept for selfemployed work. Formal employment s often found by heads of households, ndvduals n households wth other formal employed, and ndvduals wth hgh levels of other famly per capta ncome, ndcatng the strong role played by networkng n fndng a formal job. Selectvty plays an mportant role n determnng earnngs for the three sectors. Our models ndcate that f one does not possess typcal formal characterstcs and stll works n the sector; he wll earn less than expected based on hs observed characterstcs. Ths suggests that there are some specfc abltes, whch are not beng captured by our human captal models that are valued by the formal sector. A noteworthy fndng from our estmatons s that the selectvty patterns changed wth tme. In 1993 and 1997, nformal workers wth formal-lke characterstcs were postvely selected nto the nformal sector. In ths tme perod, formal jobs were plentful and workers wth formal-lke characterstcs only accepted those nformal jobs whch offered excellent opportuntes. By 2002, ths postve selecton had eroded away and workers wth formal-lke characterstcs were negatvely (though nsgnfcantly) selected nto the 10 th quantle of the nformal sector. The reducton n the sze of the formal sector from 1997 to 2002 may have been responsble by causng lower productvty workers from the formal sector lose ther jobs and move to the nformal salared sector. The selectvty estmates from the formal sector confrm ths hypothess. Workers wth nformal-lke characterstcs go from beng negatvely to postvely (and nearly sgnfcant) selected nto the formal sector from 1993 to Agan, ths suggests that the lowest productvty formal workers wth nformal-lke characterstcs left (or were forced to leave) the sector, thereby causng the postve selecton bas at low condtonal quantles. The formal sector, as expected, s the sector for whch the human captal models are best adjusted, specally n Returns to educaton are consderably hgher n the formal and nformal sectors than n the self-employed. Gettng an educaton degree ncreases yours chances of becomng formal and ncreases your earnngs once you are part of the sector. Educaton s rather unmportant for earnngs determnaton n the nformal and selfemployed sectors n 2002, whereas t was much more mportant n 1993 and The worst penaltes for poor households from non-formal work are assocated wth beng female for all years. There are also some bg penaltes for the poor from techncal educaton, but only for 1993 and Ethncty provdes a small premum for the poor non-formal workers, relatve to formal returns, throughout the years. In 2002, the educatonal penaltes

16 are small for basc and prmary educaton and ncrease for secondary and above, leadng us to conclude that the low educated (male) non-formal worker does not receve large penaltes from not partcpatng n the formal sector. For those wth secondary educaton the penaltes are larger n 2002, but decreasng wth ncome levels, and for those wth hgher educaton there are very few observatons n poor households, although the penalty for non-formal work s qute hgh.

17 6 References Buchnsky, M. (1998). The dynamcs of changes n the female wage dstrbuton n the USA: A quantle regresson approach. Journal of Appled Econometrcs 13, Ftzenberger, B. (2003). Gender wage dfferences across quantles, accountng for sample selecton, mmeo, Unversty of Mannhem Ichmura, H., (1993). Semparametrc least-squares (SLS) and weghted SLS estmaton of sngle-ndex models. Journal of Econometrcs, Vol. 58, Koenker, R. and Bassett, G. (1978). Regresson s, Econometrca, Vol. 46 No.1, pp Lay, J. (2001). Segmentaton and nformalty n urban labour markets: evdence from Bolva and mplcatons for poverty reducton. Kel Insttute of World Economcs. Machado, J. and Koenker, R. (1999). Goodness of ft and related nference processes for quantle regresson, Journal of the Amercan Statstcal Assocaton, Vol. 94, pp Maddala, G. S. (1983). Lmted-dependent and qualtatve varables n econometrcs, Cambrdge Unversty Press, New York. Maloney, W. (2003). Informalty revsted, World Bank Polcy Research Workng Paper n Newey, W. (1988) Two step seres estmaton of sample selecton models, mmeo, Department of Economcs, MIT. Newey, W., J. Powell. and J. Walker (1990) Semparametrc estmaton of selecton models: some emprcal results, The Amercan Economc Revew, vol. 80, no. 2, Tannur-Panto, M. and D. Panto (2002). Informal employment n Brazl a choce at the top and segmentaton at the bottom: a quantle regresson approach, Workng paper n 236, Department of Economcs, Unversty of Brasla. Wodon, Q. (2003). Internal Mgraton, Urbanzaton and Poverty: case studes for Latn Amerca, The World Bank, draft.

18 TABLE 1- Summary Statstcs, 1993, 1997, Weghted sample averages. Formal Informal Self-Employed Varables Log hourly earnngs Labor earnngs month Hours week Age (years) Pot. Exp. (years) Years Occupaton Educaton (years) Educ. less than basc Educ. basc Educ. prmary Educ. secondary Educ. techncal Educ. unversty Ethnc (language) Female Marred Less 20 hours week Industry & Agrculture Transport & Utltes Constructon Commerce Government & Educ Servces La Paz (captal) Chuqusaca La Paz Cochabamba Oruro Potos Tarja Santa Cruz Ben & Pando Attend school Head HH Head HH * Female N. kds less 15 years N. kds less 6 * Female N. elderly HH * Female Lve wth parents Want/Able work more Want/Not able Not want/able Oth.Fam.Inc.(pc/1000) Other HH Formal Sample Sze 2,342 1, , ,382 1,877 1,154 Note: Medan values for Log hourly earnngs and Labor earnngs month.