Mobility and Earnings in Ethiopia s Urban Labor Markets: Arne Bigsten Taye Mengistae Abebe Shimeles. Abstract

Transcription

1 Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Mobly and Earnngs n Ethopa s Urban Labor Markets: Arne Bgsten Taye Mengstae Abebe Shmeles Abstract An analyss of panel data on ndvduals n a random selecton of urban households n Ethopa reveals large, sustaned, and unexplaned earnngs gaps between publc and prvate, and formal and nformal sectors over the perod We have no formal evdence whether these gaps reflect segmentaton of the labor market along eher of these dvdes. In other words, we cannot show whether they are at least n part due to mpedments to entry to the hgher wage sector. However, we do have evdence that, f segmentaton explans any part of the observed earnngs gaps, then could only have weakened over the survey decade. We fnd, frst, that the rate of mobly ncreased between the two pars of sectors. Sample transon rates grew across survey waves, whle state dependence n sector choce decreased. Second, the sensvy of sector choce to earnngs gaps ncreased over the same perod. In partcular, the role of comparatve earnngs n selecton nto the nformal sector was evdent through out the survey decade and ncreased n magnude over the second half of the perod. JEL Classfcaton: J42, J62, J31, J45, O17 Key Words: Segmented labor markets, labor market mobly, wage dfferentals, nformal sector, publc sector labor markets World Bank Polcy Research Workng Paper 4168, March 2007 WPS4168 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 ced accordngly. The fndngs, nterpretatons, and conclusons expressed n ths paper are entrely those of the authors. They do not necessarly represent the vew of the World Bank, s Executve Drectors, or the countres they represent. Polcy Research Workng Papers are avalable onlne at Unversy of Gothenburg, Gothenburg, Sweden. World Bank Group, Washngton DC, USA.

2 1 Introducton 1 In ths paper we nvestgate the extent to whch the structure of Ethopa s urban labor market mght have changed snce the early 1990s by lookng at changes n the rate of workers mobly across sectors, n sectoral earnngs gaps, and n the sensvy of mobly rates to earnngs dfferentals. The am s to see f these have changed n ways consstent wh the market gettng more or less segmented n the sense of the exstence of entry barrers to relatvely hgh payng sectors of employment. Our analyss s based on data from the 1994 to 2004 waves of the Ethopa Urban Household Soco Economc Survey of the Adds Ababa Unversy and the Unversy of Gothenburg, Sweden. 2 There are good enough theoretcal reasons and anecdotes that could lead us to suspect that urban labor markets n low-ncome economes lke Ethopa s are segmented along the lnes ndcated n ths paper. At the same tme, there s consensus that the hypothess of segmentaton s dffcult, f not mpossble, to test formally n datasets lke the one avalable to us. 3 We nonetheless thnk that the hypothess provdes a useful framework wh the help of whch one can analyze changes n the structure of labor markets. Although we cannot drectly test s fundamental predcton that there are barrers to entry to certan sectors of the labor market, we can stll nterpret tme seres of mobly rates and wage gaps or the relatonshps thereof as ndcatons of whether labor markets are gettng more or less flexble. In other words, absent the possbly of segmentaton, these same data would not be as nterestng n terms of ther mplcatons for the structure or functonng of the market. Segmentaton s ndcated whenever there are persstent sector wage premums that are unexplaned n the sense that we cannot attrbute them to selectvy or to dfferences n comparatve advantage, unmeasured ably, job attrbutes or taste. 4 One hypotheszed 1 Ths paper s based on an earler verson prepared as part of a World Bank study of Ethopa s urban labor markets (World Bank, 2007). We thank Jen Klugman and Caterna Laderch, for very useful comments on that and a subsequent verson as well as for the support and gudance they provded to our work on the paper. 2 As dscussed n greater detal n secton 2 these data are not statstcally representatve of labor market ndcators for urban areas n Ethopa as a whole. Those can be found, together wh a thorough dscusson of dfferent labor market data sources for Ethopa n World Bank (2007). 3 Although the balance of studes has so far found evdence aganst the segmentaton hypothess, almost everyone rejectng are quck to pont out serous lmatons of data or methods used n the analyss. 4 Dckens and Lang (1985) defne segmentaton n terms of the nterrelated phenomena of sustaned and unexplaned wage premums arsng from endogenous wage rgdes and the non-prce ratonng of prmary-sector or hgh-wage jobs arsng from. Magnac (1991) defnes a segmented labor 2

3 nstance of s between formal sector wage employment and own account work or nformal sector employment. 5 A second s between publc sector employees and prvate sector wage earners. 6 If segmentaton does ndeed exst along eher of these dvdes, then s also possble that some proporton of the unemployed have been ratoned out of pad jobs or out of self-employment by non-prce factors. Exstng formal tests of segmentaton rely on a host of restrctve assumptons ncludng the log normaly of potental earnngs and the absence of state dependence n mobly costs, tastes and sklls. 7 No attempt s therefore made n ths paper to carry out any such tests, or formally establsh f Ethopa s urban labor market s segmented. 8 Instead, we look at changes n varables that defne segmentaton wh a vew to assessng whether these changes would have strengthened or reduced f n fact exsted over the survey perod. Unexplaned and non-compensatng wage dfferentals that cannot be attrbuted to dfferences n comparatve advantage or to selectvy are sustanable only f the market as one n whch a) rewards vary across sectors for equally productve workers and b) wages are rgd at least n one (-presumably the hgh-payng) sector of the labor market as a state n whch equally productve worker are rewarded. 5 The classc story of segmentaton along the formal vs. nformal dchotomy s Felds (1975) extenson of the Harrs -Todaro (1970) model of an underdeveloped economy s labor market quanty-adjustng to shocks through changes n unemployment and rural-to-urban mgraton n the face of wage-rgdy. Stglz (1974), Shapro and Stglz (1984) and Bulow and Summers (1985) provde effcency wage explanatons to the wage-rgdy assumed n the Harrs-Todaro model. The extenson by Felds ntroduces job search costs that rural-to-urban mgrants fnanced by enterng a low wage, low productvy no-barrer-to entry sector upon arrval to towns. Both entry barrers and wage premums are essental features of Felds model of segmentaton. Rauch (1991) provdes an alternatve occupatonal-choce-cum-ndustral-organzaton perspectve to the modellng of the dualy between the formal and the nformal sectors of the economy. Although the dchotomy n Rauch s model s defned n terms of departures from an optmal sze dstrbuton of frms rather than from an optmal dstrbuton of workers and earnngs per se, also arses from (government mposed) prce rgdy n the labor market. 6 The large body of emprcal work comparng pay rates between the two sectors s overwhelmngly focused on advanced economes, and s concerned not so much wh labor market segmentaton as wh the mport of pay gaps to publc fnances or to the ably of government agences to attract sklled workers. There has nonetheless been sgnfcant nterest n publc vs. prvate sector wage gaps n developng economes as well, partly out of smlar polcy concern as n advanced economes (e.g. Stevenson, 1992; Heller and Ta, 1994), and also n the context of testng the segmentaton hypothess (Lndauer, 1991). Examples nclude Lndaur and Sabot (1983) on Tanzana, Van der Gaag and Vjerberg (1988) on Cote d Ivore, Terrel on Ha (1993), and Nelsen and Rosholm (2001) on Zamba. Regardless of ther sgn, unexplaned and non-compensatng publc sector pay prema would be ndcatve of segmentaton, snce mobly must be mpeded n one drecton or the other for them to exst. 7 Heckman and Hotz (1986) and Magnac (1991) dscuss the assumptons n the context of the swchng regresson test proposed n Dckens and Lang (1985). 8 On balance emprcal studes of labor market studes on mddle ncome developng economes reject the segmentaton hypothess. See, for example, Magnac (1991) on Colomba, Maloney (1999) on Mexco, and Pratap and Quntn (2005) on Argentna. 3

4 mobly of workers across sectors s somehow mpeded. An ncrease n the rate of workers mobly nto hgh-wage sectors over the survey perod would therefore ndcate decrease n the degree of segmentaton. In order to see f the rate of workers mobly has ncreased over the survey perod, we compare sample transon matrces across survey waves. We also estmate dynamc bnary sector choce models for four states, namely, unemployment, nformal sector employment, self-employment, and publc sector employment. 9 A decrease or an ncrease n the mobly of workers across sectors over tme should be reflected n a decrease or an ncrease the degree of state dependence n sector choce, the parameter of state dependence beng the coeffcent of the nal state n the sortng equaton. It s mportant n ths context that the parameter of state dependence we estmate measures true state persstence rather than merely reflects unobserved ndvdual heterogeney. Even then true state dependence could be a consequence as much of state dependence n mobly costs, taste or sklls as could be of entry barrers, of whch only the latter defnes segmentaton. 10 Stll, f we can assume that mobly costs and the skll and taste dstrbuton of workers reman the same over tme, we can tell whether or not the market s gettng more segmented by lookng at what happens to the persstence parameter over the same perod. 11 Gven that the rate of workers mobly s usually hghly correlated wh the sze of sectoral earnngs gaps, the change n the earnngs gap self s a second (alternatve) ndcator of change n the degree of segmentaton. A thrd alternatve ndcator s change n the sensvy of sector choce to earnngs gaps. The less sensve s state (or sector) choce to pay gaps, the more lkely s that mobly nto the hgher payng sector has been mpeded. To look at these second set of ndcators, we compute sector earnngs dfferentals based on estmated sector specfc earnngs equatons. The sensvy of 9 For completeness one could also estmate a separate model of bnary choce nto formal prvate sector employment. However, the estmaton would not add nformaton to that the models of choce nto unemployment, nformal sector employment and publc sector employment provde between them. 10 Heckman and Hotz (1985) hghlght the dstncton between mobly costs and entry barrers as an mportant ssues n testng for segmentaton. 11 Maloney (1999) s probably the frst to make use of mobly data n conjuncton wh post transon changes n earnngs to test for segmentaton. Gong and van Soest (2002) buld on ths dea to jontly model earnngs and nter-sectoral mobly and estmate state dependence n sector choce as one of the test parameters of segmentaton. Although we borrow from these two papers the dea of usng mobly data, our goal here s less ambous than thers snce we are not seekng to test for segmentaton drectly. We are n fact sde steppng the task of testng for segmentaton, whle askng whether our data are consstent wh becomng more or less pronounced over the perod of observaton. 4

5 sector choce to earnngs gaps s estmated as a parameter of a dynamc structural sector choce model. 12 Our estmates mply large earnngs dfferentals between the publc and the prvate sectors of wage employment. In the absence of drect evdence on mpedments to mobly, these dfferentals would not necessarly mply segmentaton. Stll, they should make the hypothess of segmentaton more credble n the Ethopan context than would otherwse be. At the same tme, the sensvy of sector choce to earnngs gaps seems to have gone up n more recent waves of the survey, not only between the prvate and the publc sectors of wage employment, but also between formal sector wage employment and the nformal sector. In partcular, the role of relatve earnngs n selecton nto the nformal sector seems to have gone up substantally. These results suggest that, f Ethopa s urban labor market s ndeed segmented along the publc vs. prvate or formal vs. nformal sector dchotomes, has got less and less so snce The raw transon matrces we compute from the survey sample support ths concluson as do the parameters of the dynamc sector choce models that we have estmated on the same dataset. We see from the raw transon matrces that, although there was ltle mobly across sectors pror to the 1997 wave of the survey, mobly rates ncreased consderably between 1997 and Mobly rates of the 2004 wave were also at least as hgh as those of the 2000 wave. The ncrease n the rates of mobly across sectors of employment was accompaned by a small but persstent declne n the sample rate of open unemployment. The ncrease n sample transon rates across survey waves s consstent wh changes n state dependence parameters of the dynamc choce model that we estmate. The parameters are all posve and statstcally sgnfcant n the equaton for publc sector employment, n the equaton for nformal sector employment and n the unemployment equaton. At the same tme, they have become smaller n more recent survey waves n each case. 12 Our approach here s smlar to that of Gong and van Soest (2002), who estmate a dynamc structural sector choce model by usng predcted wages (though not premums) as one of the potental determnants of mobly between the formal and nformal sectors. 5

6 The organzaton of the rest of the paper s as follows. In the next secton we descrbe our data. Secton 3 dscusses patterns and rates of mobly across states n terms of raw transon matrces. Secton 4 descrbes our estmaton and testng framework. We present results of estmaton of (reduced form) dynamc sector choce models n Secton 5, and those of ther structural counterparts n Secton 6. We conclude n Secton 7. 2 Data The frst wave of the Ethopa Urban Household Soco Economc Survey took place n There have been four waves snce then, one n each of the years 1995, 1997, 2000 and Some 9,000 to 10,000 ndvduals n 1500 to 1600 households were covered n each of these waves. A major pont of strength of the dataset generated by the survey s that has a szeable panel component. A sgnfcant proporton of the ndvduals covered n the frst wave were also tracked by all four subsequent waves. An even hgher proporton of them were covered by at least three waves. A weakness s that the survey samples were drawn exclusvely from the country s seven largest urban centers. Whle these account for the bulk of the urban sector of the economy, we would hesate to project detals of our results to the broader (natonal) urban labor market. The smaller towns may dffer n terms of ther labor market structure from the major centers covered by the survey. That sad, we would be surprsed f the more qualatve aspects of our conclusons dd not hold by and large for the urban sector countrywde. Table 1 provdes the breakdown of the survey sample by age groups and labor market states. In we lm the labor force to those aged 15 to 64 years nclusve. Ths leads to a labor market partcpaton rate of about 54%, whch does not seem to have changed much over the decade spanned by the surveys. The table shows the dvson of the economcally actve between the unemployed, on the one hand, and the employed of varous categores. The unemployed are defned as those household members of the age group who were not self-employed, or workng for anyone else and were actvely lookng for pad work. Among the employed, the table dstngushes between the selfemployed and those workng for someone else. Self-employment s understood to nclude own account workers and owner manages of small busnesses. We defne formal sector employment as wage or salary employment n a prvate company, n a state-owned busness or n a government agency. The term formal prvate sector worker refers to wage earners n prvate companes. We have put 6

7 famly workers, domestc workers and casual laborers under the category other prvate sector workers. In the context of ths paper the nformal sector conssts of the selfemployed and these other prvate sector workers. We nclude n the formal sector employees of prvate companes and the publc sector. 13 We provde n table 2, descrptve statstcs for monthly earnngs, annual sales revenue of owned busnesses and selected covarates of these and of labor market status. The covarates nclude age, gender, schoolng, and occupaton. Statstcs are gven by survey wave as well as for the observatons pooled across waves. Wh the excepton of earnngs and busness sales revenue all varables n the table are dchotomous. Schoolng s presented usng two classfcatons of school attanment. One of these s farly standard, needs ltle explanaton, and s used to report the sample proportons of those who attaned a partcular level for the 1994, 1997 and 2000 waves. Unfortunately the same classfcaton was not used n the 2004 wave of the survey, whch dstngushed only between those who had no educaton, those who had attaned grade 10 or lower, those who had completed preparatory school, and those who had some tertary educaton. Other covarates hghlghted n the analyss nclude locaton of employment, ethncy and famly background varables, ncludng parental schoolng and parental occupaton. 3. Aggregate transon rates and patterns As a prelude to our analyss of the determnants of labor mobly, we descrbe n ths secton the patterns of mobly observed n the data n terms of sample transon matrces between employment and unemployment, and across three sectors of employment, namely, the publc sector, formal prvate sector employment, and nformal sector employment. 13 There s no employer sze dmenson to our defnon of nformaly. People who have reported to be employees of a prvate sector organzaton or enterprse have all been classfed here as formal sector workers. It s common to defne nformaly n a way that takes nto account the sze of employers. There are a varety of operatonal defnons of nformaly wh whch our admtedly lose characterzaton s consstent. One defnon dentfes nformaly wh lack of labor regulaton (Pratap and Quntn, 2006). A second s n terms of employer sze (Maloney 1999, Gong, van Soest and Vllagomez, 2004). A thrd s n terms of the knd of work beng done such as casual or pece rate work and also own account work, such as casual work or self employment (Magnac, 1991). A fourth alternatve s to defne frms as nformal f they are not legally regstered (Bgsten, Kmuyu, Lundvall, 2004) 7

8 There was not much transon across these sectors between the 1994 wave and the 1995 wave, or between the 1995 wave and the 1997 wave. On the other hand, transon rates were relatvely hgh between the 1994 wave and the 1997 wave, the 1997 wave and the 2000 wave, and the 2000 wave and 2004 wave, and, as expected, even hgher between any par of waves that were further apart by more than four years. We report the sample transon matrces for these parngs n table 3. Two key patterns seem to emerge across the four waves. Frst, there was a small but statstcally sgnfcant and persstent declne n the sample unemployment rate snce the 1997 wave. Secondly, ths was assocated wh an ncrease n the share of formal sector n employment. Unemployment transons The probably of someone reportng unemployment n the 1994 sample landng a job or shftng to self-employment whn a year was almost zero. However, the probably of someone n that state exng unemployment ncreases as we expand the tme horzon of observaton. Panel a) of table 3 tracks the respondents of the 1994 wave through to the 1997 wave. Although the table regsters some movement out of unemployment, ths occurs at very low rates: less than 7 percent of those who were unemployed n the 1994 sample had some form of employment by A ltle over a thrd of these went nto formal sector wage employment. Thngs do change sgnfcantly over the next three years. By the 2000 wave, more than half of the unemployed of the 1994 sample had changed status (table 3, panel b). About one n eght had left the labor force altogether, presumably, havng stopped lookng for work. About one n fve went nto own account work or nto nformal wage employment. Another 11 percent got jobs n prvate frms. And approxmately 9 percent landed publc sector jobs, about one thrd of these beng n state owned companes. By the 2004 wave, less than a thrd of the unemployed of the 1994 wave were stll avalable for work (panel c). Of the rest, about 23 % left the labor force; some 19 % joned the nformal sector (n self-employment or n nformal wage contracts); 17 % landed jobs n prvate frms; and about 12 % got publc sector jobs. There are some nterestng comparsons of rates of transon out of unemployment between the perods and The rate of absorpton of the unemployed nto the formal sector ncreased from 2 % to 18 %. The rate of absorpton 8

9 by the nformal sector (both self-employment and wage contracts) ncreased from about 5% to 19%. The rate of absorpton by the publc sector ncreased from under 1% to almost 8%. The rate of absorpton by the formal prvate sector ncreased from under 2% to about 10%. The relatvely hgh rates of transon from unemployment durng the perod were also sustaned n the next four years. The probably of someone reportng unemployment n the 2000 sample takng up some form of employment by the tme of the 2004 survey was 41 percent. Ths s sgnfcantly hgher than the correspondng transon rate for the perod , but by an amount that could be explaned entrely by the fact that the second nterval s one year longer. 14 Informal sector/self-employment transons Lke rates of transons out of unemployment, probables of transons between the formal and nformal sectors of employment would seem to be neglgble over any gven year. But, agan, the probably of transon rses sgnfcantly when we expand the tme horzon by three more years (table 3, panel b). Ths s largely on account of movement of workers to the prvate formal sector. The probably that someone reportng selfemployment or nformal employment n the 1994 survey would have joned the publc sector by the 2000 survey was only 5 %. The probably that the same ndvdual would have taken up a job n a prvate frm was about 15 %. These probables do not change much as we expand the horzon further to 10 years (table 3, panel c). 14 Other related transon rates of nterest for the years are the followng: About 24 % of the unemployed of 2000 sample landed formal sector jobs by 2004; another 17 % became selfemployed, or entered nto nformal job contracts; some 13 % joned prvate frms; about 11 % joned the publc sector. To llustrate the orgns of the unemployed, some 17 % of those who were out of the labor force n the 2000 sample joned the unemployment pool by 2004 (table 3, panel e). Gven that the out-of-the-labor force group s by far the largest category of survey respondents n any year, ths suggests that fresh entry nto the labor force must be the largest source of entry nto unemployment n the economes from whch the sample was drawn. The second largest source were prvate sector job losses were the second mportant source of gross addon to employment for the same perod. The probably of a formal prvate sector job loss was 14 %. The probably of an nformal sector job loss was even hgher at 17%. Ths contrasts wh 7-8 % rate of publc sector job losses. Gven that there were more prvate sector employees than there were publc sector workers n the 2000 sample, a larger proporton of the unemployed of 2004 must have come from the prvate sector than dd from the publc sector. 9

10 To see f the rate of mobly from nformal sector employment to formal sector jobs has ncreased over tme, we agan compare panel a, whch captures transons over wh panel d, whch relates to As s the case wh transons from unemployment to employment, there was a drastc change here also: the probably of transon from nformal sector wage employment to formal sector jobs rose from less than 1 percent for to 12 % for the perod , and 15% for the perod Publc sector transons There was a 48% chance that someone who worked for the government (as opposed to a state owned company) n the 1994 sample would have left that partcular sector by We arrve at ths rate by addng the followng probables from panel c: that the person retred or otherwse left the labor force (20.5%); that they were lad off or pensoned off but dd not leave the labor force (10%); and that they joned the prvate sector (17%). More than half of those who moved to the prvate sector would have worked for a formal sector frm. The correspondng probables of transon for a publc enterprse worker were: 35% of leavng the labor force, 10% of beng unemployed, and 16% of jonng the prvate sector. Two-thrds of those jonng the prvate sector would have joned a formal sector frm. These rates reflect more the hgher transon probables of the ntervals between the fourth and ffth ( ) and the thrd and fourth ( ) waves of the survey (panel d and panel e) than the rates of the frst three waves. As can be seen from panel a there was practcally no movement from the publc sector durng the nterval between the frst and the thrd waves ( ). On the other hand, there was a 12 % chance that someone who was a government employee n 2000 would have joned the prvate sector n self-employment (4 %) or as an employee (8%) by The probably of a 15 To llustrate where the nformally employed come from, more than one n ten of people who were out of the labor force n the 2000 sample would have become nformal sector wage worker or selfemployed by And just under 20% of those who were unemployed n the 2000 sample would be self employed or n nformal wage work by These rates are much larger than the probably of transon from the formal sector. Because the combned share of the out-of-the labor force and the unemployed was 60% of the full sample of ndvduals n the 2000 sample, ths means that by far the largest fracton of those who joned the nformal sector by 2004 could only have come from these sources. Indeed, the probably that anyone who had a formal sector job n the 2000 sample would be found n self employment or n nformal pad jobs n 2004 was que low, standng at 6% for a government worker, at 8% for a publc enterprse employee, and at 12 % for an employee of a prvate frm. 10

11 publc enterprse employee n the 2000 sample movng to the prvate sector by 2004 was even hgher, at 19% Estmaton and testng Suppose we would lke to know whether the labor market s gettng more or less segmented between any two sectors over tme: say, formal sector wage employment vs. nformal sector employment. One way of nvestgatng ths s to see f the probably that anyone who starts out n the nformal sector n year t 1 wll have remaned n the same sector n year t has ncreased, decreased, or remaned the same over tme. The more moble are workers across any dvde, the poorer s status at tme t 1 as a predctor of status at tmet. A way of measurng changes n the degree of segmentaton s therefore to use changes over tme n the extent of state dependence n sector choce. In dong so we leave open the possbly that mobly costs are state dependent n the sense, for example, that prospectve entrants to formal sector or publc sector employment mght have to mgrate from elsewhere at sgnfcant transport costs, whle smlar costs could be sunk or entrely absent for those already n those sectors. We also allow for the possbly of sector scarrng effects as defned below as well as for skll formaton and skll deprecaton rates to vary by labor market state. As already noted, any of these factors would affect the rate of nter-sectoral mobly and, consequently, result n true state dependence n sector choce, but s not part of what defnes market segmentaton. In estmatng the degree of state dependence based on our sample, we assume that the current dstrbuton of ndvduals n terms of belongng to a sector or not s generated by the process (1) * y = 1( y > 0) where * y = X β + δy + ε 1, 16 Lookng at transons n the reverse drecton, some 12 per cent of those who were workng for prvate frms n the 2000 sample had joned the publc sector by Some 9.5% of nformal sector employees and 3 % of self-employees n the same sample moved to the publc sector over the same perod. Agan these rates are much hgher than the neglgble transon probables between 1994 and 1997, but are only a ltle hgher than the rates for the perod

12 X s a vector of exogenous observable determnants of such status, ε s a random varable summng up all unobservable nfluences on the same; β and δ are constants; and * y s a latent varable regsterng a threshold value of comparatve earnngs or of utly governng ndvdual s sector choce, or a crcal value of a measure of entry barrers to whch the choce s subject. Followng most of the relevant emprcal lerature on labor market transons, we assume that ε s dstrbuted normal wh zero mean and constant varanceσ. There s state dependence n the process whenδ 0, so that the 2 ε probably Pr( y = 1 X ) E( y X ) s a functon of y 1. In the present context we should expect δ to be posve n the event of s beng dfferent from zero. In that case a larger value of δ would mply greater state dependence n the sense that y 1 would be a stronger predctor of y. We use equaton (1) to descrbe any one of four dstnct processes n Ethopa s labor market: unemployment vs. employment n general, nformal vs. formal sector employment, self-employment vs. wage employment, and publc vs. prvate sector employment. Suppose (1) relates to the unemployment process, where y = 1 f s unemployed at tmet, but y = 0 otherwse. In the lerature on unemployment, a larger value of δ could be a result of longer unemployment leadng to loss of human capal through dsuse, whch would make the unemployed less attractve than the employed as job canddates (Pssardes, 1992). It could also reflect the scarrng effect of unemployment f prospectve employers use unemployment hstory as a rule for screenng out nferor workers (Arulampalam et al. 2002, Phelps, 1972). A thrd possbly s that gettng a new job nvolves a fxed relocaton or transport cost that becomes sunk once ncurred. Whle each of these three scenaros s consstent wh competve labor markets, a hgher δ could also be the result of segmentaton, that s, a result of barrers of entry nto employment on the part of the unemployed, due to employment protecton laws, the exercse of nsder power as n Lndbeck and Snower (1986), or jobs beng ratoned va effcency wage payment (as n Shapro and Stglz, 1984; and Bulow and Summers, 1986). Lkewse, state dependence n the sortng of workers between the nformal and formal sectors could also arse from segmentaton (by entry barrers to the formal sector) or 12

13 from any one of the three competve sources of state dependence lsted above for the case of unemployment. If we let y = 1 when s an nformal sector worker, and y = 0 otherwse, a larger value of δ could reflect the fact hgher barrers of entry to the formal sector arsng from the protecton of formal sector jobs by law or va nsder power, or some combnaton of formal sector effcency wage payment (Stglz, 1974) and search costs of formal sector jobs (Felds, 1975) n a Harrs-Todaro settng (Harrs and Todaro, 1970). Ths would be the segmentaton scenaro. Alternatvely, a large δ mght smply be caused by the fact that stayng n the nformal sector nvolves loss of sklls relevant to formal sector employment or has some knd of scarrng effect smlar to that assocated wh unemployment, or because the nformal sector workers would have to ncur reallocaton costs to formal sector jobs. Smlarly, a hgher value of δ could reflect hgher entry barrers to the publc sector f equaton (1) descrbes the allocaton of workers between that sector and prvate frms. Publc sector jobs are probably more often lfe tme jobs than prvate sector jobs n many low-ncome economes. At the same tme ther avalably for new entrants to the labor force s lmed by government fnances. Gven any of the three bnary sector choce suatons dscussed above, we would have no way of tellng what component of the δ we mght estmate for any of them reflects segmentaton and what part does not. We can nonetheless tell f segmentaton has ncreased, decreased or remaned the same along any of the dchotomes by estmatng δ over the varous segments of the survey decade under an dentfyng assumpton. Specfcally, we assume that the dstrbuton, across sectors, of mobly costs, skll formaton/deprecaton rates, scarrng effects, and other possble competve sources of state dependence across sectors does not change over the perod of nterest, so that the share of δ that we can attrbute to these sources s constant over tme. A fall n the degree of state dependence n sector choce would then ndcate declne n segmentaton whle an ncrease n the same would mean greater segmentaton. However, ths s only one aspect of the problem of dentfcaton that we face n tryng to estmate equaton (1). Segmentaton or not, our estmate of δ as a measure of true state dependence n general would be consstent only f the nal values, y 0, of labor market 13

14 states as observed at the start of the survey are exogenous (Heckman, 1981). 17 If y 0 s endogenous, a posve estmated value of δ would sgnfy wholly or partally what Heckman (1981) charactersed as spurous state dependence, that s, as a case of upward bas that the endogeney gves rse to n the estmaton. The nal condon would be exogenous only f (a) ε are serally ndependent, and (b) the start of the survey concdes wh the onset of the process self. The latter beng unlkely the case, s conventonal to assume that y 0 s endogenous, determned as s by the selecton rules of the survey sample. The assumpton that ε s serally uncorrelated would also be unrealstc snce would rule out ε ncludng persstent ndvdual effects. Yet such effects could well be present snce we are unlkely to observe all potental determnants of sector choce or the all the constrants under whch s made. Unobserved heterogeney of ths knd generate seral correlaton n the error term, whch n turn produces correlaton between the error term and y 1 and y 0. We wll therefore thnk of ε as consstng of two components: a term summng up persstent elements of unobserved heterogeney, c, that s potentally correlated wh y 0, and an..d component, u uncorrelated wh y 0, y 1 and X such that (2) ε = c + u 2 σ c Ths n turn makes ε serally correlated wh correlaton coeffcent ρ =, The 2 2 σ + σ correlaton between c and y 0 n turn bases δ upwards (Stewart, 2005). Heckman (1981), Arulampalam et al. (2000), Wooldrdge (2005), and Stewart (2005), dscuss alternatve ways of avodng ths bas. Heckman s estmator nvolves estmatng (1) jontly wh a prob model for the nal condon (3) y 1( W γ + u 0) where 0 = > W s a vector of exogenous varables ncludng pre-survey characterstcs of the sample not ncluded n X possbly along wh some of the exogenous elements of γ s a vector of constants, and u s a normal error term uncorrelated wh u, but c u X, 17 Consstent estmaton would also be possble wh serally uncorrelated errors f can be assumed that the process s n equlbrum at the start of the survey. However, the later assumpton would be unrealstc as excludes the tme varyng exogenous varables from the process (Heckman, 1981). 14

15 possbly correlated wh c. Heckman s estmator s obtaned by maxmzng the lkelhood of y, y..., y,..., y ) condonal on X, Z ) as gven by (1) to (3), whle ( 0 1, T ( treatng c as random effects. The reason for the random effects formulaton s that MLEs of parameters of (1) would be based for any fne tme seres length, T, of the panel under the assumpton of fxed effects. On the other hand, MLEs of parameters of (1) are consstent as the cross-secton sze, N, tends to nfny for gven seres length f ndvdual effects are random (Heckman, 1981). 18 Treatng c as random effects n turn amounts to assumng that they are uncorrelated wh observable nfluences on can be relaxed by assumng that tme varyng elements of wh mean zero and varance c X y *. Ths, however, s not as restrctve as mght seem. It c s a lnear combnaton of tme means, X,of the X, and a random effects component α dstrbuted normal 2 σ a, but assumed to be uncorrelated wh X such that = λ + α (Wooldrdge 2000). In that case the prmary equaton of Heckman s estmator should be wrten as the followng random effects prob: 19 (4) y 1( X β + δy 1 + X λ + α + u > 0) = Heckman s estmator deals wh the nal condon problem of dentfcaton of true state dependence by nstrumentng the nal condons, that s, by maxmzng the jont lkelhood of (, y 1, y 0..., y,..., y ) X, W. Wooldrdge (2000, 2005), proposes an T alternatve estmator that maxmzes the jont lkelhood of ( y 1,..., y,..., yt ) X, y0, that s, maxmzes the post-nal sub sample condonal on the nal condon and exogenous covarates of y. The underlyng dea s that, f y 0 s endogenous whle u s serally uncorrelated, then c must be correlated wh y 0. Snce there can also be exogenous covarates of y 0 that may produce ndvdual effects ndependently of ther nfluence on nal condons, Wooldrdge assumes that c = π 0 y 0 + Z π + a, where Z are elements of W correlated wh c ndependently of y 0, and a s dstrbuted 18 See Wooldrdge (2000), Hsao (2003) and Stewart (2005) for revews of the lerature on consstent estmaton of state dependence n the context of a prob model wh unobserved heterogeney. 19 Ths s as opposed to the tradonal random effects prob defned by equatons (1) and (2). 15

16 2 normal wh zero mean and varance σ a so that the lkelhood functon s wrten n terms of (5) y ( X β + δy + π y + Z π + a + u 0) = > Agan the lkelhood s maxmzed by treatng a as random effects. Because s easer to mplement than Heckman s estmator, we use Wooldrdge s estmator. It turns out that the results that we obtan are que smlar to those we get by usng a two stage procedure proposed by Arulampalam, Booth and Taylor (2000). 20 The procedure estmates n two steps Heckman s specfcaton (Heckman, 1981) as gven by equatons (1) to (4) above. The frst stage nvolves the estmaton of (3) by maxmum lkelhood as a smple prob. In the second stage one estmates the followng random effects prob by maxmum lkelhood: (6) y ( X β + δy + X λ + b uˆ + v~ + u 0) = > where û s the generalzed resdual from the frst stage prob, replacng the true error component u n the equaton (7) y = 1( X β + δy 1 + X λ + b1u + v + u > 0) and v ~ s such that b uˆ v~ = b u + v = a We can test for the endogeney of nal condons based on Wooldrdge s estmator by testng for the statstcal sgnfcance of π 0 n equaton (5). The corresponds to the test for the statstcal sgnfcance of b 1 n equaton (6) n the context of the two-step estmator. Snce endogeney s controlled for n each case, the estmate of δ we obtan by maxmzng the log lkelhood of the ample accordng to eher equaton should be consstent for true state dependence n choce nto the labor market state of nterest. In order to assess the mer of the two estmators relatve to estmates from maxmum lkelhood estmates of a smple dynamc prob, or a tradonal random effects prob that would assume the exogeney of nal condons, we have estmated equatons (5) and (6) on the full fve-wave panel over the perod The objectve of the analyss beng to gauge possble changes n δ over the same perod, we also estmate each specfcaton over two consecutve segments of the decade, namely, the perod and the perod The panel consstng only of fve waves, we had to choose overlappng perods n order to make sure that the estmaton of δ 20 Henley (2004) s an example of more recent work usng the two-step estmator. 16

17 controls for unobserved effects. As already noted the fve survey years are 1994, 1995, 1997, 2000 and Ths means the two perods overlap over the years The value of δ therefore can change between the two perods only n as far as the value of δ for s dfferent from the value of the same for We would nterpret a hgher value of δ for the observaton than that for the observatons as ndcaton ncrease n the degree of segmentaton of the labor market over the survey decade. An alternatve measure of change n the degree of segmentaton s the correspondng change n the sensvy of mobly to wage gaps that one observes over the perod of nterest. Consderng equaton (3), let earnngs of person n the sector of current employment, and w be the w 0 the person s potental earnngs n the alternatve sector. If workers respond to sector wage gaps n decdng where to work, then should be one of the determnants of w w0 employment status, y. We assume that the relatonshp between pay gaps and sector choce s such that (8) y ( a ( w w ) + Z θ + ξ 0) where a 1 s a constant; = > Z 1 s a vector of exogenous ndvdual characterstcs, θ s a vector of parameters (ncludng a constant term), and ξ s a zero mean d error term orthogonal to pay gaps and to Z 1. The proposon that workers sector choce and mobly decsons depend on pay gaps can be tested by estmatng equaton (8) and testng for the statstcal sgnfcance of a 1. More mportantly n the present context, an ncrease n a 1 over tme would suggest that segmentaton has ncreased n the sense that mobly has become more and more sensve to sectoral earnngs gaps. Conversely, f segmentaton grows over tme n the sense that workers have become less responsve to pay gaps ths should be reflected n a 1 dmnshng over tme. A problem that has to be tackled n tryng to estmate a 1 on the bass of (8) s that, although each worker mght have a reasonable dea of what they would earn n alternatve states of employment, we can observe only one of these, whch s the wages of current employment. A commonly used soluton to ths problem s to replace the unobserved wage by a counterfactual pay rate mpled by an estmated earnngs equaton. 17

18 To obtan the estmates we assume that such that (9) w = X 1Π1 + ζ 1 and (10) w0 = X 1Π 2 + ζ 2 where w and w 0 are lnear n a set of parameters X 1 s a vector of exogenous covarates of earnngs, Π 1 and Π 2 are vectors of parameters (ncludng constant terms), and ζ 1 and ζ 2 are d error terms orthogonal to X 1. The counterfactuals that one would compute based on the OLS estmaton of equatons (9) and (10) wll be based f there are unobservable worker attrbutes that nduce workers to self select nto eher sector and at the same tme enhance or reduce ther earnngs n the same sector. In order to avod ths bas we estmate the two equatons havng added a selectvy term based on the estmaton of equaton (6). Let Λ1( yˆ φ( yˆ ) ) = where Φ( yˆ ) φ s the densy functon of ξ, Φ the correspondng cumulatve dstrbuton functon, and ŷ s predcted value of y that one would obtan from parameter estmates of equaton (2) and Λ OLS to (9b) and (10b) 2 ( yˆ φ( yˆ ) ) = 1 Φ( yˆ w = X 1 Π1 + 1Λ y ) 1. We estmate the earnngs equatons by applyng ) σ ( ˆ + ς σ ( ˆ + ς w0 = X 1Π 2 + 2Λ y ) 2 where σ 1 and σ 2 are constants and ς 1 and ς 2 are d error terms orthogonal to all other rght hand sde varables n ther respectve equatons. The structural sector choce equaton we actually estmate s (8b) y ( a ( wˆ wˆ ) + Z θ + v 0) = > where the hat symbol ndcates estmates based on the estmaton of (9b) and (10b). 18

19 5 State dependence n sector choce The most mportant fndng emergng from our estmates of dynamc bnary sector choce models s that, whle there has always been state dependence along the three dvdes of Ethopa s urban labor market, has grown sgnfcantly smaller over the survey decade. The estmates show that the probably of beng n any one of the four sectors durng a partcular survey wave s hgher for those who were n that state n the precedng wave, but ths effect of last-perod state on current choce has grown weaker n more recent waves. We nterpret ths as evdence that the rate of nter-sectoral mobly has ncreased over the years. Detals of the fndng are reported n tables 4, 5 and 6, where we estmate a dynamc prob for each of the four labor market states. The prob for nformal employment s estmated under alternatve defnons, namely, one n whch we equate nformal sector employment wh self-employment, and a second n whch the category also ncludes those who work for others under nformal wage contracts. In the frst panel of table 4, we estmate a dynamc prob for each of the four labor market states on observatons pulled across all fve waves on the assumpton that there are no unobserved ndvdual effects and that the nal labor market state observed n the frst wave of the survey s exogenous. Snce neher s necessarly true n the lght of the dscusson n secton 4, what we are reportng n the panel are estmates of a baselne model rather than estmates of the true data generatng process. Indeed we do show that neher assumpton s correct n table 5 wh respect to our data. Table 5 s also where we compare the results of addressng the nal condon problem usng the Wooldrdge s estmator and the two step estmator. As part of mplementaton of the latter we estmate a smple prob model of employment status at the tme of the 1994 wave of the survey and report n the second panel of table 4. As a selectvy equaton, ths s dentfed by ncludng famly background varables n havng excluded them from the frst panel and from the specfcatons n tables 5 and 6. The underlyng assumpton s that these varables affect current status only n as far as they nfluence nal condons. It turns out that both paternal occupaton and maternal occupaton are sgnfcant nfluences on nal condons. Interestngly, ndvduals whose fathers owned non-farm busnesses were less lkely to be to be found n the publc sector n 1994, but 19

20 they were also less lkely to be unemployed, as they were more lkely to be self-employed or to work for someone else n nformal contracts. Also, maternal occupaton dd not seem to nfluence the chances of nal publc sector employment. Those who had workng mothers were more lkely to be unemployed as they were less lkely to work n the nformal sector as employees or as own account workers. Surprsngly, parental educaton seems to be only weakly correlated wh nal labor market status. The level of own educaton does nfluences the same status. Notably, those who had secondary educaton as the hghest level of schoolng were more lkely to be unemployed n 1994 than those who had less schoolng than that and those who had some tertary educaton. Ths s n spe of the fact that there was no sgnfcant correlaton between educaton and publc sector employment, and reflects the fact that the more educated were less lkely to be found n the nformal sector or n self-employment at the tme. Nevertheless, the publc sector attracted the more sklled than the prvate sector n 1994 whle nformal sector employment was assocated wh lower sklls, when sklls are measured n terms of broad occupaton groups. Gender and age were also sgnfcant factors n sector choce n In partcular, women were more lkely to be selfemployed. Publc sector employment and self-employment were also posvely correlated wh age, whle the unemployed were predomnantly n the age group. Turnng to table 5, we report Wooldrdge s estmator of each sector s prob n the frst panel, and the correspondng two-step estmates n the second panel. Results of the frst panel are obtaned by estmatng equaton (5) by maxmum lkelhood. Those n the second are also maxmum lkelhood estmates of the specfcaton gven by (6). There s clear evdence of endogeney of nal condons n both panels. Ths can be seen n the fact that all coeffcents of y 1994 are statstcally sgnfcant n the frst panel. Lkewse, the coeffcents of the generalzed resdual of the nal status prob of the frst stage are statstcally sgnfcant. The fact that both sets of coeffcents are posve means that the coeffcents of y t 1 n the frst panel of table 4 would be based upwards for true state dependence due to the endogeney of nal condons. However, what s left when we remove the bas s stll statstcally sgnfcant: the coeffcents of y t 1 are posve and sgnfcantly dfferent from zero n both panels of table 5. It s also clear that there are unobservable nfluences on sector choce beyond the observable factors controlled for n the table. However, the same unobservable varables happen to be hghly correlated wh 20

21 tme varant observables, namely, educaton, age, and occupaton. Indeed, there s no evdence of unobserved effects that are uncorrelated wh those observables. Ths can be seen from our estmates of σ a and σ v, whch are reported n the frst and second panels of the table, respectvely, and are not statstcally sgnfcant. Snce estmaton results are que smlar across the two panels of table 5, we confne ourselves to the use of Wooldrdge s estmator n table 6, where we compare state dependence n sector choce between the perods and As already noted, ths amounts to comparng estmates for wh those for The concluson that state dependence n sector choce has dmnshed over tme can be read from the frst row of the table, where each coeffcent n the frst panel s sgnfcantly lower than the correspondng entry n the second panel. For example, the coeffcent of y t 1 n column 1 s sgnfcantly smaller than the coeffcent of the same n column 5, suggestng that the average unemployment duraton for was smaller than that for Lkewse, the coeffcent of y t 1 n column 4 s sgnfcantly smaller than the correspondng coeffcent n column 8, mplyng that past publc sector employment was a weaker predctor of current publc sector employment durng the perod than was durng The same can be sad of nformal employment n general and self-employment n partcular: a hstory of employment n eher form was a weaker predctor of beng n the same type of employment n than was n On the assumpton that the combned effect of competve sources of true state dependence of sector choce has not changed sgnfcantly over the survey decade n all three cases, the observed fall n the persstence parameters over the same perod can only ndcate weakenng segmentaton f there are ndeed entry barrers to formal sector employment n Ethopa. 6 Earnngs gaps and sector choce We now turn to our second approach to gaugng changes n the degree of segmentaton, whch s estmatng the sensvy of sector choce to sectoral earnngs gaps. In tables 7, 8, and 9 we report results of estmaton of a structural mode (equaton 8b) of the sortng of workers between the publc and prvate sectors of wage employment. In column 4 of 21 At the same tme, there s no evdence that the nfluence of any of the key observable determnants of labor market state has changed between the two perods. Ths ncludes gender, age and schoolng. The only excepton to ths s that age effects n publc sector employment seem to have come down sgnfcantly. 21