ISSN 1440-771X Australa Department of Econometrcs and Busness Statstcs http://www.buseco.monash.edu.au/depts/ebs/pubs/wpapers/ Language and Labour n South Afrca A new approach for a new South Afrca Katy Cornwell February 2006 Workng Paper 5/06
Language and Labour n South Afrca A new approach for a new South Afrca January 2006 Katy Cornwell Department of Econometrcs and Busness Statstcs Monash Unversty Clayton VIC 3800 Australa Phone +61 3 9905 2453 Fax +61 3 9905 5474 katy.cornwell@buseco.monash.edu Abstract Ths paper consders the role of language n labour earnngs n South Afrca over the perod 1996 to 1998. Our pooled cross-secton comprses of over 160,000 workng age adults, and the analyss consders the decson to partcpate n the labour force, employment prospects and labour earnngs. Models nclude varables for ndvdual mother tongue n addton to populaton group. After condtonng on a number of soco-economc and demographc factors, we fnd that havng Englsh as one s mother tongue s one of the pvotal determnants of labour earnngs. These results are robust across two models of sample selecton. Such fndngs shed lght on the economc consequences of South Afrca s natonal polcy of lngustc heterogenety. Keywords: Unemployment, Income, South Afrca, Language Polcy, Race JEL Codes: J64, J21, J23, J19, J31 The author would lke to thank Brett Inder and Ana Deumert for useful dscussons.
1. Introducton Unemployment n South Afrca has been metaphorcally descrbed as an untamed beast (Kngdon and Knght, 2004). Indeed, South Afrca s charactersed by unemployment rates amongst the hghest n the world, wth the most ferocous rates of up to 45% amongst black South Afrcans (see Table 1). Moreover, hardshp s not overcome once pad employment s found, partcularly for the majorty black South Afrcan populaton. As a resdue from the aparthed era, substantal dfferences n earnngs between racal groups reman. Table 1 demonstrates that on average whte South Afrcans are earnng almost four tmes as much as blacks. Hgh unemployment rates coupled wth dsparate labour earnngs have lead to numerous studes on the determnants of ncome, mostly focussng on the returns to educaton. The latest studes nclude the work of Keswell and Poswell (2002) and Serumaga-Zake and Naude (2003). The former work questons the emprcal relevance of the standard human captal theory of dmnshng margnal returns to educaton, and provdes a thorough overvew of the vast South Afrcan returns to educaton lterature. Serumaga-Zake and Naude (2003) utlse double hurdle and Heckman sample selecton models n examnng the prvate returns to educaton of black South Afrcan males and females. We extend ths work by consderng multple years of data from the South Afrcan October Household Surveys and ncorporatng addtonal varables. Aparthed dctated that populaton group was the prmary determnant of educatonal and occupatonal opportunty. Wth well-documented evdence of the effect of educaton on earnngs, t was natural for the lterature to lnk populaton group, educaton and earnngs n South Afrca. The breakdown of aparthed saw the formaton of the new South Afrca where black economc empowerment s recognsed as fundamental to redressng past mbalances and enablng the country to move on to acheve sustanable development and prosperty (southafrcanfo, 2004). Great efforts have been made to eradcate racal dscrmnaton and undo the njustces of the past, and the Ranbow Naton looks forward to the day when they can say wth confdence that populaton group no longer determnes one s fate. Dsappontngly, studes contnue to fnd populaton group dummes strongly sgnfcant n ncome and employment equatons. We would argue that a new South Afrca calls for a new approach to modellng the South Afrcan labour market: an approach whch looks much further than populaton group n dentfyng the determnants of earnngs n the multlngual new South Afrca. We begn to explore ths noton by examnng whether mother tongue language provdes a better nsght than populaton group nto what s of mportance to an ndvdual s relatve success n the labour market. Is t populaton group per se that leads to hgher unemployment rates for black South Afrcans, or s t that 1
Englsh s not ther natural mother tongue language, creatng a barrer of entry to employment and an mpedment to earnngs? It s ths aspect of the labour market whch we seek to address n ths paper. Our nterest s prmarly on ntroducng language as a potental determnant of labour earnngs. However, before the ndvdual s able to report earnngs, they must overcome two hurdles: the ndvdual must frst choose to partcpate n the labour force, and then from ths labour force pool the ndvdual must also be selected for employment. Recognsng a propensty for sample selecton bas, we model ncome usng two models of sample selecton: a verson of Cragg s (1971) double hurdle and Heckman s (1979) sample selecton model, wth mother tongue ncluded n addton to populaton group and other soco-economc and demographc varables at each of the partcpaton, employment and ncome stages of the models. Gven that ths avenue s a new drecton for the labour earnngs lterature, our next secton s devoted to dscusson of the South Afrcan labour market n the context of language. Secton 3 follows wth a descrpton of the methodology, whle secton 4 ntroduces the data. Results a presented n secton 5, and dscusson follows n secton 6. 2. Language as the new drecton Embracng lngustc pluralsm n ts consttuton, the new South Afrca recognses and guarantees equal status to each of ts eleven offcal languages 1. However, hstorcal whte domnance n government and commerce s reflected n Englsh and Afrkaans beng the most commonly used languages n offcal and commercal publc lfe, despte the Afrcan languages of Xhosa and Zulu beng the more common languages spoken at home (see Table 2). In partcular, Englsh s eghth on the lst of mother tongues ranked accordng to frequency for the respondents n our sample. Lterature on the economcs of language n the labour market s lmted. The majorty nvolve consderaton of the role of language n labour market nteractons and earnngs for mmgrants and Hspancs n the Unted States. Dscusson tends to fnd consensus n favour of lngustc homogenety. The theoretcal bass for the mmgrant/hspanc lterature s generally pnned to the noton of language as the facltator of communcaton. In ths sense, language can be seen as the medum for communcaton exchange, whereby lngustc heterogenety ncreases the transacton costs of ths exchange and consequently n the absence of blngualsm, less exchange wll take place between those speakng dfferent languages. 2
Consder the mplcatons of ths for the ndvdual job seeker n the labour market. Informaton about jobs flows through open channels of communcaton. The ndvdual s prvy to these channels of nformaton dependng on her ablty to communcate wth the people n these channels. If Englsh s the domnant language used n the work envronment, the Englsh-speakng ndvdual can tap nto nformaton drectly from the pool of the employed and also drectly to the employer. Consequently, the Englsh-speaker holds an advantage on the employment front over the non-englsh speaker. Ths suggests that job search may not only be facltated by language channel, but also lmted by t, and hence lngustc dsadvantage would present tself n both the partcpaton decson and the employment outcome. The lterature advocates that earnngs may also be ted to language knowledge. Where access to occupaton s determned by language channel, a worker may fnd themselves n low pad occupatons relatve to skll. Kossoudj (1988) suggests that there could also be some element of ndvdual choce to be among peers of the same language background, thereby maxmsng ndvdual utlty rather than ncome. Indeed, ths could be one explanaton for why large ncome dfferences are observed by language group (see Table 2). From the employer s vewpont, communcatve ablty s a form of human captal n that t enhances productvty, and productvty s lnked to earnngs. The communcator s better able to convey ther comparatve skll advantage, from whch the employer s able to realse any productvty gans from specalsaton. McManus (1985) compares a group of employees sharng a lngua franca wth a group who are unable to communcate. He suggests that n the latter case, the dvson of tasks would be accordng to average characterstcs of the group wth no allowance for personal varaton, whereas the former would be more productve n the sense that productvty gans from specalsaton could be realsed. Hence, those wth hgh levels of communcaton would fnd themselves deemed more productve and awarded accordngly wth hgher ncome. Furthermore, by acceleratng the absorpton of nformaton, communcaton mproves the return to educaton (McManus et al 1983), such that employers may choose to tran workers wth hgh language sklls n new technology more readly than those wth lmted language sklls, enablng the employee wth hgh language sklls to clmb further up the promotonal and therefore ncome ladder. Emprcal applcaton corroborates ths theoretcal dscusson, revealng that language attrbutes play an mportant role n earnngs for mmgrants and Hspancs. Grener (1984) s able to use language to explan up to one thrd of the relatve wage dfference between Anglo and Hspanc men. Kossoudj (1988) concurs n her selecton bas corrected specfcaton of a random utlty model for occupaton and earnngs. Other studes have been able to ncorporate Englsh profcency. For nstance, Rvera-Batz (1990) uses test-based 1 The 11 Offcal languages of the Republc of South Afrca are Seped, Sesotho, Setswana, sswat, Tshvenda, Xtsonga, Afrkaans, Englsh, sndebele, sxhosa and szulu. 3
Englsh profcency measures to examne the mpact on earnngs, fndng t to be a major factor. Mora (2003) models a standard Mnceran earnngs functon wth Englsh fluency, geographcal regon and ethncty as condtonng varables, fndng there to be a great deal of nteracton between educaton, experence and schoolng. Interestngly, for males wth no educaton and no experence, Mora (2003) suggests that those who speak Englsh earn sgnfcantly less than those who do not speak Englsh, yet ths result s reversed at hgher levels of schoolng: for a male wth 12 years of educaton, Englsh language profcency adds an earnngs premum of 30%. Small pockets of studes have also looked at the effect of mnorty languages on educatonal outcomes n developng economes, however, whle nterestng, these are manly descrptve and ther emphass on the mplcatons for blngual educaton dstracts from the focus of our study concernng labour market outcomes. Moreover, lke the work on mmgrants, these studes are concerned wth the mplcatons of a mnorty populaton group beng unable to converse n one offcal and domnant language. The case of South Afrca, however, s unque n that through a hstory of poltcal dscrmnaton, the languages of the mnorty populaton group domnate commerce and offcal lfe, yet t s the majorty and also the poorest populaton group who speak the languages whch could well be the mnor n the labour market context. We turn now to our own analyss n the hope of sheddng lght on ths stuaton. 3. Methodology Our analyss s prmarly concerned wth the factors whch contrbute to monthly earnngs for all South Afrcans. Pror to reportng ncome, the ndvdual must frst choose to partcpate n the labour force. Of course, n South Afrca partcpaton n the labour force does not guarantee employment, and so a further decson on behalf of the employer must be made to draw the ndvdual from the labour force pool. Only once the ndvdual s employed do they report ncome. Accordngly, only a subsample of all South Afrcans are employed and able to report earnngs. It s lkely that the soco-economc characterstcs of the employed are dfferent to those who are not, and lkewse, the characterstcs of labour force partcpants are dfferent from non-partcpants. In partcular, unobservable characterstcs affectng the decson to work would be correlated wth the unobservable characterstcs affectng ncome. Selectvty bas would arse, therefore, f we were to make statements about the determnants of earnngs for all South Afrcans based on the observed earnngs of the subset whom are employed. The approprate model must be one whch copes wth sample selecton at each stage of partcpaton and employment. 4
We specfy a sample selecton model wth the prmary dependent varable of nterest of the form I I I I y = x β + u, I where x s the vector of soco-economc and demographc explanatory varables, β I the vector of unknown I coeffcents and u the error term. There are two latent decson functons: (1) The partcpaton decson: I P * P P = x β +u P, wth ndcator varable I P R S P f I > = T 1 * 0. 0 otherwse Such that the ndvdual partcpates n the labour force f I P* > 0. And (2) The employment decson: I E * E E = x β +u E, such that the ndvdual s selected for employment f I E* > 0 and hence I E R S E f I > = T 1 * 0. 0 otherwse 5
These choces are partally observed: we do not observe the employment outcome for the non-partcpant nor the ncome for unemployed and non-partcpants. Correspondngly, the partcpaton equaton s defned over the whole South Afrcan workng age populaton, the employment equaton over labour force partcpants and the ncome equaton over those who are employed. We utlse 2 models of ncome determnaton whch ncorporate sample selecton correcton factors: the double hurdle model and Heckman s sample selecton model wth two sample selecton mechansms. Double hurdle model The double hurdle lterature s dvded about whether the partcpaton and employment decsons are jont or sequental that s, does one s perceptons of ther employment prospects nfluence ther decson to partcpate n the labour force, or s the partcpaton decson taken frst, ndependently of employment? Maddala (1993) embraces the jont versus sequental debate by pontng out that the dstncton rests n the covarance of error terms n the partcpaton and employment equatons. As extensons to the Heckman-Lee two stage estmaton methods, Maddala (1993) outlnes a sequental double hurdle model wth uncorrelated errors, and then relaxes the zero correlaton assumpton to descrbe a jont double hurdle model relyng on bvarate probt estmaton. Smth (2002) renders jont double hurdle approaches superfluous, argung that the specfcaton of Cragg s orgnal model s already a jont model. Confuson as to whether decsons are jont or sequental, he purports, stem from the ease wth whch Cragg s model lends tself to sequental nterpretaton. Such unresolved ambguty ultmately leaves the choce between jont and sequental models n the hands of the researcher. To add a further complexty to the ssue, South Afrca has long debated on the partcular defnton of the unemployed: under the offcal (narrow) defnton, workng-age ndvduals who are not employed must be actvely searchng for work and be able to start work wthn 1 week n order to be deemed unemployed. An alternatve defnton (the so-called broad or expanded defnton) relaxes the need for the ndvdual to be actvely searchng, and to some extent allows ncluson of dscouraged workers. The dfference that ths crteron makes for the unemployment rate s qute substantal (see Table 1). For comparablty and completeness, our analyss s undertaken usng both defntons. The ssue of correlaton between employment and partcpaton s somewhat ntertwned wth these defntons of unemployment. The dfference n results owng to choce between Cragg and Maddala, or jont and sequental s lkely to be mnmal, hence we opt for the computatonal smplcty of Maddala s (1993) sequental model. 6
Assumng normalty of the error terms, Maddala s (1993) double hurdle model nvolves estmatng two separate probt models for partcpaton and employment. From these estmated models we obtan two correcton factors P λ = e j P P φ x β P P x Φe β j and E λ = e j E E φ x β, E E x Φe β j where φ. bg and Φ bg. are, respectvely, the probablty densty and cumulatve dstrbuton functons of the standard normal dstrbuton. Restrctng the sample to those employed, ncome s regressed on a number of soco-economc and demographc varables (outlned n secton 4) as well as both the obtaned correcton factors. Heckman s sample selecton model Heckman s model dffers from Maddala s (1993) sequental double hurdle model n ts ncluson of the two correcton factors n estmaton. The Heckman partcpaton probt and ts correspondng correcton term are dentcal to those of the double hurdle. In modellng the employment probt, however, the correcton factor from the partcpaton equaton s ncluded as an addtonal varable for the Heckman model. The second correcton factor s then obtaned as E λ = e j E E φ x β, E E x Φe β j where x E now ncludes λ P as an addtonal varable. The second correcton factor alone s then ncluded n the ncome equaton as an addtonal regressor. 7
4. Data Data for ths study s extracted from the South Afrcan October Household Surveys of 1996 through to 1998. Ths provdes a mult-stage cluster sample of some 166,534 workng age adults. Our double hurdle and Heckman models characterse ncome as the prmary contnuous dependent varable of nterest. Partcpaton n the labour market and employment are two hurdles whch must be overcome before an ndvdual s observed as recordng ncome. Adoptng a Mnceran form to our dependent varable, ncome s taken as the natural logarthm of real monthly ncome and deduced from waged and/or selfemployed sources. We defne success n the labour market as one havng a full-tme job and earnng a consstent wage, and that anyone workng less than full-tme throughout the month s dong so due to lack of work rather than by choce. Gven ths defnton, we do not adjust monthly ncome accordng to hours worked, snce total ncome earned durng the month would capture relatve success. Monthly ncome s deflated accordng to temporal, provncal and regonal (rural/urban) prce dfferences. Such a prce deflator does not exst for South Afrca, and nether can t be constructed from the offcal Stats SA CPI publcatons: CPI seres are avalable at the provncal level, but for each provnce the seres s constructed to have a base of 100 n 1998, renderng the seres ncomparable across provnces. Moreover, the prces used to construct these ndces were collected at only the urban level, and do not provde any nformaton on prces n the rural areas. We thus construct our own prce deflator by usng cluster (vllage) level prce data n the World Bank s 1993 LSMS dataset n addton to the offcal provncal urban seres. 2 The frst stage of constructng the prce ndces nvolves usng 1996 offcal Stats SA CPI weghts to calculate a food prce ndex for 1993 by (new) provnce and regon, relatve to a base of 100, representng the urban overall country average. The offcal urban provncal level CPI s were then rebased to ther 1993 food prce ndex value, and re-projected out to 2005, creatng an urban provncal level seres relatng prces back to the 1993 urban country average. The rural counterpart was then constructed at the provncal level accordng to the wthn provnce rural-urban food prce dfference observed n the 1993 LSMS data. Ths constructon provdes a deflator seres for the whole country, mappng tme, provnce and regon relatve to one base: average 1993 urban food prce. 3 Its relablty wll depend on the rural-urban prce dfferences that prevaled n 1993 remanng stable over tme. Ths may not be the case, but gven the lack of alternatve data, ths represents the best avalable approxmaton t s certanly better than gnorng dfferences n prces between rural and urban areas. 2 A smlar method was used by Kngdon and Knght (1999) to construct spatal prce ndces for South Afrca n 1993. 8
Partcpaton and employment are the secondary bnary dependent varables of nterest. In the partcpaton equaton, the dependent varable takes a value of 1 where the ndvdual partcpates n the labour force and 0 otherwse. Smlarly, the dependent varable n the employment equaton takes a value of 1 where the person s employed. For ths second bnary varable, the sample s restrcted to labour force partcpants. The October Household Survey for 1996 ntroduced a new queston regardng the mother tongue of the respondent. Gven that South Afrca embraces eleven offcal languages, t s of partcular nterest whether mother tongue language nfluences labour market outcomes, partcularly employment and ncome whch are determned by the employer. We also condton on a number of soco-economc and demographc varables ncludng populaton group, gender, household head, rural resdency, age and tme dummes. Lnear regresson splnes for hghest level of educaton attaned are used to allow for dfferng slopes across prmary, secondary, and tertary educaton. For the ncome equaton, we also nclude dummy varables for employment n the nformal sector and self employment. The martal status dummy s nteracted wth ndvdual gender n the partcpaton equaton, but omtted from the employment and ncome equatons for dentfcaton. A more comprehensve descrpton of the varables s provded n Table 3. In each of the double hurdle and Heckman models, we consder unemployment under both the offcal and expanded defntons, wth lttle varaton n results. 5. Results 5.1 Labour force partcpaton Table 4 presents results for the partcpaton decson under both the offcal and broad defntons of unemployment. Defned over the entre workng-age populaton, the estmated double hurdle and Heckman models are dentcal n ths frst stage of partcpaton. The models n columns (1) and (2) follow the conventon n the lterature for studes on the South Afrcan labour market: ncluson of dchotomous ndcator varables for aparthed-era racal classfcaton of the ndvdual, alongsde other soco-economc characterstcs such as status n the household, gender, age, educaton etc. As Table 4 reveals, our results corroborate the common fndngs of such lterature, wth all populaton group dummes strongly sgnfcant. Under the narrow defnton of unemployed (column (1)), we fnd that Asan, coloured and whte South Afrcans are all more lkely to partcpate n the labour market than are black South Afrcans. 3 The prce ndex data s avalable from the author on request. 9
Males and those regarded as head of the household are more lkely to enter the labour force than females and those not regarded as head of ther household respectvely. Where the male s also marred, there s an addtonal lkelhood that they wll partcpate, yet ths s reversed for marred females. Moreover, age has a quadratc effect on labour force partcpaton, fallng as one approaches old age. These results hghlght a strong patrarchal culture for South Afrca and the mportance of the household as an economc and cultural unt: the decson to partcpate n the labour force s made n conjuncton wth the cultural oblgatons of the (often extended) famly. Educatonal attanment s mportant for labour force partcpaton. Under the narrow defnton, attanment of prmary school level educaton has no effect on entry nto the labour force over no educaton. As one progresses through secondary school and tertary studes, entry nto the labour force becomes ncreasngly more common. The slope of the lnear splne eases off for those wth other levels of educaton, yet remans steeper than that for secondary educaton. Indeed, hgher schoolng may mprove access to nformaton about employment, enhance one s percepton of ther employablty and lead to a hgher rate of partcpaton n the labour force. Evdence suggests a hgher labour force partcpaton rate amongst urban dwellers, reflectng the lack of opton for subsstence farmng that the rural area provdes. Fnally, there s some evdence of labour force expanson for South Afrca as a whole over the three years, wth some provncal varatons. The move from the offcal to broad defnton means that around 18,000 ndvduals move from beng nonpartcpants to unemployed partcpants. These people could perhaps be called dsheartened workers as opposed to dscouraged workers: they would work gven the opportunty, but are not actvely seekng work. The results, n column (2), do change under ths alternatve defnton. In most cases the change s only n sgnfcance, but for two of the populaton group and one of the provncal dummes, the estmated coeffcents change sgn. For those offcally classfed as Asan or whte, a sgnfcant postve estmate under the narrow defnton becomes a sgnfcant negatve under the broad defnton. Indeed, 92% of the so-called dsheartened workers are black South Afrcans, hence the reclassfcaton boosts the partcpaton rate amongst black South Afrcans, ceters parbus. Moreover, the now negatve coeffcent on the whte dummy reflects the old money syndrome: typcally whte South Afrcans have hstorcal wealth for use as a safety net, enablng them to drop out of the labour force much more easly than the nvarably poorer black South Afrcan. Interestngly, the coeffcent on prmary schoolng becomes stronger under the broad defnton such that those wth prmary school attanment are less lkely to partcpate n the labour market than those wth no schoolng. Ths could mply that a greater proporton of the dsheartened have attaned only prmary school 10
level educaton. Alternatvely, wth the model defned over the entre workng age populaton (15-65), and school attanment not necessarly mplyng the ndvdual has ceased school, we may be nadvertently capturng those currently n school and hence not partcpatng n the labour force at ths stage. Columns (3) and (4) present results for our alternatve varable specfcaton: ncluson of mother tongue language. We group the offcal Afrcan languages together to form the base language category, and use dummy varables to dstngush Englsh, Afrkaans and other (unoffcal) languages. We also nteract Englsh and Afrkaans wth populaton group to allow for a dfferng effect of the commerce-domnant languages accordng to populaton group. The results are practcally dentcal for all varables except the populaton group dummes. In partcular, we see a drastc change n the mportance of populaton group for labour force partcpaton once ndvdual mother tongue language s taken nto account. These results mply a new mndset for a new South Afrca: racal dscrmnaton no longer domnates, rather, the ablty of the ndvdual to tap nto the world of the employed. Lookng more closely at the results for language, t seems that whle coloured South Afrcans generally have hgh labour force partcpaton rates whch reman unchanged by language, those who have Englsh as ther mother tongue are much less lkely to partcpate. Dfferental partcpaton rates accordng to language are seen most strongly for the base populaton group: black South Afrcans. In partcular, under the narrow defnton of unemployment dsplayed n column (3), havng Englsh as a mother tongue mproves the chances of the average black South Afrcan beng economcally actve by around 14%. For completeness, we also present results based on black South Afrcan workng age adults alone n Table 5. By concentratng on only the black populaton, we are able to further dsaggregate the offcal Afrcan languages. We construct a new set of 7 dummy varables for language based on prevalence: 6 dummes representng the 6 most common languages to be spoken n the home, wth the remanng languages grouped together n the 7 th dummy as mnorty languages. Ths tme, we use Englsh as the base n order to determne just how much dfference there s between the wdely spoken languages and that whch domnates commerce. Under the narrow defnton, all the Afrcan and mnorty languages are assocated wth lower labour force partcpaton rates. An Afrkaans mother tongue, however, s not statstcally dfferent to an Englsh mother tongue for partcpaton rates amongst black South Afrcans. In contrast, the broad defnton provdes much weaker results, wth no sgnfcant dfference n partcpaton accordng to language. Indeed, the lnk between language and partcpaton may not be as strong as that between language and employment or language and earnngs, snce partcpaton s more a reflecton of ndvdual motvaton rather than success. 11
The remanng estmaton results mrror those of the entre workng age populaton drven by the fact that 80% of the South Afrcan populaton s black. One mportant dfference n results, however, s observed n prmary school attanment under the narrow defnton, the estmated coeffcent s now sgnfcantly postve, ndcatng that for black South Afrcans, any educaton mproves partcpaton. As wth our prevous fndngs, we also see a negatve change n slope upon use of the broad unemployment defnton, yet ths s nsuffcent to deem the effect sgnfcantly dfferent from zero. 5.2 Employment outcomes Results for employment outcomes are presented n Table 6. Usng the Heckman and double hurdle models, under both the offcal and broad defntons of unemployment, we next consder the mportance of language for employment outcomes of the entre South Afrcan workng age populaton. Columns (1) - (4) provde estmaton results for the conventonal models, whle columns (5) (8) allow for mother tongue dfferences. We use the orgnal language dummes: offcal Afrcan languages form the base, to be compared to Englsh, Afrkaans and other (unoffcal) languages. We agan nteract Englsh and Afrkaans wth populaton group to allow for a dfferng effect of language. Indvdual martal status s omtted under the presumpton that employers are ndfferent to the martal status of ther employees, yet we retan the household head dentfer as a proxy for ndvdual leadershp qualtes. Results are equvalent across both Heckman and double hurdle methodologes, and any dfferences are found only between offcal and broad employment defntons. The nverse mlls rato for the Heckman model s sgnfcant under the narrow defnton, but nsgnfcant under the broad defnton. Ths suggests that any sample selecton bas s ntroduced by omttng the dsheartened workers, and calls for the use of broad over offcal defnton of unemployed. The populaton group dummes are postve across all 8 sets of results. Asan, coloured and whte South Afrcans all appear more lkely to be employed than black South Afrcans under all model specfcatons. However, once language s taken nto account, t becomes clear that the determnants of employment go much deeper than populaton group alone, wth changes n both the margnal effects and the standard errors. The coeffcents on the Asan dummes become slghtly larger, yet smaller on the coloured and whte dummes. Interestngly, the margnal effects now suggest more of a dvde n employment propensty of Asan and black than between whte and black South Afrcans, at the base level of all other characterstcs. The margnal effects for Englsh are qute nterestng: Englsh mproves the employment prospects of black, coloured and whte, however, the margnal effect s greatest for black South Afrcans. In net effect, 12
Afrkaans- or Englsh-speakng whtes reman better off for employment than any other group defned on the bass of populaton group and language, whle the worst off depends on the unemployment defnton: under the narrow defnton, Asans who speak Afrkaans or Englsh at home are worst off, whle under the broad defnton, Afrcan-speakng blacks fall just short of Afrkaans- or Englsh-speakng Asans. Males and household heads are more lkely to be employed than females and those other than the household head. Age as a proxy for potental experence has a postve, slowly dmnshng advantage for employment. Those ndvduals who have attaned prmary school level educaton appear to have a lower propensty for employment than those wth no educaton. Presumably, those who have never been to school would have nstead pursued or been forced nto work, and hence by default bult up vocatonal skll, leadng to employment but not necessarly hgh levels of ncome. Returns to secondary and tertary educaton are ncreasngly hgh n terms of employment outcomes. Lookng to regonal dspartes, the results provde evdence for rural-urban employment dfferences, although the drecton shfts dependng on the defnton of the unemployed. Under the offcal defnton, rural dwellers are more lkely to be employed than urban dwellers, yet ths s reversed under the broad defnton. Ths mples that there are more dsheartened workers n the rural areas n the urban areas there may be greater opportunty to engage n nformal self employment n the absence of fndng waged work, whereas n the rural areas, the employment opportuntes are more restrctve. Fnally, all provnces seem to have much hgher unemployment rates than the Western Cape, and the underlyng country unemployment rate has contnued to rse over the three year perod. We next turn to the results for black South Afrcans n Table 7. Snce the majorty of black South Afrcans record Afrcan mother tongues, we agan defne dummes accordng to prevalence, usng Englsh as the base for comparatve purposes. In both selecton models and under both unemployment defntons, Englsh mother tongue affords the ndvdual much greater success n employment outcomes than any other language. The models suggest that those who speak Zulu n the home are around 37.3% less lkely to be employed under the broad defnton than an Englsh mother tongued black South Afrcan wth the same soco-economc characterstcs. And ths s the estmated outcome even after controllng for the level of educaton of the ndvdual. Afrkaans ranks less than Englsh, but greater than other languages for employment. Ths mples that beng able to speak the most wdespread language (and even extendng ths to havng more networkng potental) s not what s mportant for employment, but rather, t s one s ablty to speak the language of the workplace, ths n most cases beng Englsh. 13
5.3 Earnngs The two sample selecton models provde smlar estmates for earnngs, as shown n Table 8. Wth only a dfference n how the selectvty correcton terms are constructed, there s mnmal dfference n results, as antcpated. There s some evdence of sample selecton bas, and ths bas seems to be reduced upon use of the broad defnton. The Heckman versons all fnd sgnfcant upward bas n the ncome equaton, whle the double hurdle versons are more mxed, tendng to attrbute bas to the unobservablty of employment outcomes for non-partcpants. Most relevant to ths paper are the results on populaton group and language. Agan, we provde the conventonal models n columns (1) (4) and our language-ncluded alternatves n columns (5) (8). The results agan hghlght model msspecfcaton when language s omtted. The conventonal models suggest Asan, coloured and whte South Afrcans earn more on average than black South Afrcans wth the same soco-economc characterstcs. Ths populaton group-based ncome premum ranges from 10% for coloured and 50% for whte South Afrcans, however, the results change dramatcally once language s taken nto account, wth the premum for beng whte slashed to almost half that found under the conventonal models. Furthermore, our approach deems all other populaton group dummes nsgnfcant. Such evdence suggests not only that much of the conventonal populaton group effects are largely language effects, but also that, apart from some ncome dfferences for the whte populaton, there are no other ncome dfferences on the bass of populaton group. Ths mples that whle some populaton group-based earnngs dfferentals do stll exst, South Afrca may be well on ts way to breakng down the racal dscrmnaton legacy of ts past. Lookng more closely at language partculars, we fnd those wth an Englsh mother tongue to earn an ncome premum above all other languages, and unlke employment, ths premum does not dffer accordng to populaton group. Despte ths populaton group-constant Englsh language premum, the proportonal ncome mprovement for Englsh-speakng whte South Afrcans s mnute compared to that experenced by Englsh-speakng black South Afrcans. Afrkaans s only benefcal for earnngs f the ndvdual s Asan, yet the seemngly large margnal effect s not partcularly strong and largely netted out by the (albet nsgnfcant) negatve coeffcent on Asan. We agan fnd postve estmated coeffcents on male and household head. The models suggest that any educaton mproves earnngs, and furthermore, acceleratng returns to educaton. Ths corroborates the fndngs of Keswell and Poswell (2002). The Mnceran proxy for experence, age, combned wth age 2, has a postve but slowly dmnshng effect on earnngs. 14
Occupatons of an agrcultural, artsan or operatng nature do not seem to yeld hgher monthly earnngs than elementary and domestc occupatons, whle sklled, and to greater extent manageral/professonal occupatons, earn hgher ncome. Gven that most agrcultural and mnng operatons are set n the rural areas, there are drastcally low ncomes for workers n ths sector, even for the more sklled workers. Monthly earnngs n the trade, transport and fnance ndustres are smlar to those of the domestc servces ndustry, whle manufacturng, utltes and constructon sectors do better. Informal sector workers do qute poorly, partcularly f they are employees rather than self-employed persons. Whle ths nformal dsadvantage s largely netted off for the self employed, most nformal sector workers would stll suffer the low ncome experenced by workers n elementary occupatons. In addton to lower rural wages, there are also ncome dspartes across provnces and tme, even after takng nto account regonal prce dfferences. Results are robust when modellng black South Afrcans alone. Dsaggregaton of the more prevalent Afrcan languages reveal Afrkaans to not only rank below Englsh for earnngs, but also below some of the Afrcan and unoffcal languages. Interestngly, the mother tongue language yeldng earnngs closest to Englsh s Seped, the language wth the hghest unemployment rate. The most negatve partal elastcty s found among those wth Sesotho mother tongue, the domnant language n the Free State and Lmpopo, and suggests ncomes 30.1% lower than those whose mother tongue s Englsh, even after controllng for occupaton and educaton. 6. Dscusson Ths paper has examned the mportance of language for labour force partcpaton, employment and earnngs n South Afrca. The estmated models suggest that Englsh mother tongue language s mportant for success n the labour market, even after condtonng on populaton group and level of educatonal attanment. It s recognsed that ths study has a number of shortcomngs and that the sgnfcance of the mother tongue varables should be taken wth cauton. Frstly, nformaton on language profcency s a mssng yet mportant pece of the puzzle. The ndvdual possessng hgher profcency and thus potentally better communcatve sklls would fnd themselves n a better barganng poston for jobs than those who are less well off n ther ablty to communcate va language. Unfortunately, the October Household Surveys only provde nformaton on the language spoken at home, whch we term the mother tongue. No ndcaton s gven of ablty to speak other languages, nor of ther profcency. Indeed, for a black South Afrcan to speak 15
Englsh at home could mply that the household has had an hstorcally more prvleged exstence under aparthed than other black South Afrcans. In ths case, Englsh mother tongue could be an ndcaton of class. Data on multple language profcency would allow some dstncton between a class effect and the degree to whch language ablty matters n the labour market. Moreover, n attackng the 2001 census, Donnelly (2003) labels responses to a queston seekng to dentfy a sngle mother tongue from a generalsed lst as unrevealng. We also must be careful to dstngush the extent to whch language determnes earnngs drectly, as opposed to language determnng occupaton type, whch n turn determnes earnngs. Further analyss could nclude nteracton effects between combnatons of populaton group, tme, educaton and language. Despte the fact that the South Afrcan government spends a large proporton of ts budget on schools, t may be that educatonal attanment matters consderably more for those wth Afrcan mother tongues. The ncorporaton of data over a longer tme perod and populaton group/tme nteracton effects may reveal some nterestng results concernng populaton group as a determnant of employment and earnngs snce aparthed. Despte the lmtatons, these tentatve results are qute marked: the results suggest that a black South Afrcan who speaks Englsh at home s more lkely to be employed and to earn a hgher ncome than, for example, a Xhosa speakng black South Afrcan wth otherwse dentcal characterstcs. Such a result has mportant mplcatons for ndvduals whch must not be confused wth those for polcy. Whle n the current South Afrcan context, the ndvdual black South Afrcan lookng to enhance ther employment prospects should prortse learnng the major labour force language, Englsh, ths advce does not necessarly extend to the government by suggestng adopton of Englsh as the unversal offcal language of South Afrca. Results n the lterature based on Hspanc mmgrants to the USA would suggest an abolton of multlngualsm n favour of one offcal language. However, South Afrca s stuaton s unquely dfferent from that n North Amerca n that the mmgrants generally represent a small mnorty group who also speak a mnorty language or one that s largely non-exstent n the labour force. Yet n South Afrca, the domnant (most populous) racal group speak a number of seemngly mnor languages n the labour force context, whle the less populous racal groups speak the major labour force language as a result of mbalanced hstorcal factors. When employees do not share a common lngua franca wth ther employer and/or fellow employees, the mmgraton lterature screams neffcency, wth uncommuncated comparatve skll advantage condemnng producton to run at an overall average skll level. However, the key s that such comparatve skll advantage s uncommuncated, not uncommuncable. Technologcal progress need not be thwarted and producton need not be neffcent f there was, for example, adequate provson for translators. 16
Moreover, the racal structure of the South Afrcan populaton s such that effcency gans can be realsed by the mnorty racal groups optng to embrace the more populous Afrcan languages, rather than many Afrcans strvng to learn one language to communcate wth the mnorty populaton. The beauty of the new South Afrca s that t s all about freedom and equty: the South Afrcan consttuton embraces freedom of the people through allowng and facltatng each populaton group to communcate n ther own language. It would therefore go aganst the sprt of the consttuton to revert to a sngle offcal language, partcularly f t were the language of the least populous group. Hence, rather than Englsh as the domnant language n commerce beng pushed upon the non-englsh speakng populaton, commerce tself could be adapted to embrace the many Afrcan languages and subcultures. As a consequence, commerce would then concde wth South Afrca s poltcal agenda. Perhaps also, regonal economc development wll pave the way for emergence of domnant languages of commerce other than Englsh. And these other domnant languages need not be all 11 offcal languages: one should not gnore the smlartes of the Afrcan languages, n that profcency n one Afrcan language does not lmt the ndvdual to be ncommuncable wth those of other Afrcan tongues. Ths may even allow for blngualsm, for nstance, profcency n Englsh and (any) one Afrcan language for all South Afrcans. Is ths not the approach that s already taken n schools for language educaton? Indeed, mono- vs. blngual educaton s complex and hghly senstve, and our results cannot adequately address the ssue. Nonetheless, our fndngs certanly call for a new way of lookng at the South Afrcan stuaton: a new drecton for a new, free and equtable South Afrca. 17
References Cragg, J., 1971, Some Statstcal Models for Lmted Dependent Varables wth Applcaton to the Demand for Durable Goods, Econometrca, 39, 5, 829-844. Donnelly, S., 2003, Language and the Census, Mal & Guardan, 14/8/03. Grener, G., 1984, The Effect of Language Characterstcs on the Wages of Hspanc Amercan Males, Journal of Human Resources, 19, 25-52. Heckman, J., 1979, Sample Selecton Bas as a Specfcaton Error, Econometrca, 47, 1, 153-162. Kngdon, G. and J. Knght, 2004, Unemployment n South Afrca: The Nature of the Beast, World Development, 32, 3, 391-408. Kngdon, G. and J. Knght, 1999, Unemployment and Wages n South Afrca: A Spatal Approach, Unversty of Oxford Workng Paper WPS/99-12, Oxford. Keswell, M. and L. Poswell, 2002, How Important s Educaton for Gettng Ahead n South Afrca?, CSSR Workng Paper No. 22, Cape Town. Kossoudj, S., 1988, Englsh Language Ablty and the Labor Market Opportuntes of Hspanc and East Asan Immgrant Men, Journal of Labor Economcs, 6, 2, 205-228. Maddala, 1983, Lmted Dependent and Qualtatve Varables n Econometrcs, Cambrdge Unversty Press. McManus, W., 1985, Labour Market Costs of Language Dsparty. An Interpretaton of Hspanc Earnngs Dfferences, Amercan Economc Revew, 75, 4, 818-827. McManus, W., W. Gould and F. Welch, 1983, Earnngs of Hspanc Men: The Role of Englsh Language Profcency, Journal of Labour Economcs, 1, 2, 101-130. Mora, M., 2003, An Overvew of the Economcs of Language n the U.S. Labor Market: Presentaton Notes, Amercan Economc Summer Mnorty Program Presentaton Notes, Unversty of Colorado, Denver. 18
Rvera-Batz, F., 1990, Englsh Language Profcency and the Economc Progress of Immgrnts, Economcs Letters, 34, 295-300. Serumaga-Zake, P. and W. Naude, 2003, Prvate Rates of Return to Educaton of Afrcans n South Afrca for 1995: a Double Hurdle Model, Development Southern Afrca, 20, 4, 515-528. Southafrcanfo, 2004, http://www.southafrca.nfo/ess_nfo/sa_glance/demographcs/language.htm. Smth, 2002, On Specfyng Double Hurdle Models, n Handbook of Appled Econometrcs and Statstcal Inference, A. Ullah, A. T. K. Wan and A. Chaturved, (eds), Marcel Dekker, New York. 19
Table 1 Labour market outcomes by offcal populaton group classfcaton Offcal Unemployment Rate Expanded Average Real Monthly Earnngs Labour Force Partcpants Income Earners Afrcan 28.3 45.4 4.0 11.22 Asan 11.7 15.5 12.2 23.97 Coloured 14.1 21.9 12.6 12.60 Whte 3.8 5.7 25.3 39.73 Overall 22.6 37.4 6.4 15.70 Source: 1996-1998 October Household Surveys. Table 2 Labour market outcomes by mother tongue Mother Tongue Language Number of Labour Force Partcpants Afrcan Asan Coloured Whte Total Average Real Monthly Earnngs Labour Force Partcpants Income Earners Afrkaans 908 93 13618 6817 21436 1140 1935 Englsh 350 3795 1794 3834 9773 2018 3359 Seped 10968 0 17 3 10988 372 1233 Sesotho 12435 2 33 13 12483 402 971 Setswana 12960 0 80 4 13044 444 1119 Xhosa 21455 3 73 35 21566 311 1072 Zulu 27286 12 35 8 27341 390 1162 Other language 13701 147 103 217 14168 431 1222 Total/Overall 100063 4052 15753 10931 130799 631 1570 Source: 1996-1998 October Household Surveys.
Table 3 Varable defntons Varable Descrpton Asan Colour Whte Dummy varables for populaton group, takng a value of 1 where the respondent s offcally classfed as Asan, Coloured and Whte respectvely. Base: Afrcan. Male Gender dummy takng the value of 1 where the respondent s male. Ths varable s also nteracted wth MARRIED to allow a dfferng effect on labour market partcpaton f the respondent s a husband. Household head Dummy varable takng the value of 1 where the respondent s regarded as head of the household. Prmary school Secondary school Dploma Degree Other educaton Hghest level of educatonal attanment. Lnear regresson splnes were used to allow dfferng slopes across prmary school, hgh school, tertary and other levels of educatonal attanment. Base s no educaton. Rural Dummy varable takng the value of 1 where the respondent resdes n a rural area. Marred Age Age 2 Eastern Cape Northern Cape Free State Kwa-Zulu Natal North West Gauteng MpumaLanga Northern Dummy varable takng the value of 1 where the respondent s marred. Used n partcpaton and ncome equatons only. Ths varable s also nteracted wth Male to allow a dfferng effect on labour market partcpaton f the respondent s a husband. Age and Age 2 to allow for a nonlnear effect of age on earnngs. Ths would also capture Mnceran potental experence. Dummy varables takng a value of 1 f the ndvdual resdes n the named provnce. Base: Western Cape. Afrkaans Offcal Afrcan language Unoffcal language Offcal language dummes. Varables take a value of 1 where the language spoken at home s Afrkaans, one of the Afrcan languages deemed an Offcal language of South Afrca and a language not deemed an Offcal language of South Afrca. Ths set of dummes comprse an alternatve to the next set of language dummes. Base: Englsh. Afrkaans Seped Sesotho Setswana Xhosa Zulu Other language Dfferentated Afrcan language dummes takng the value of 1 where the language spoken at home s Afrkaans, Seped/Northern Sotho, Setswana/Tswana, Isxhosa/Xhosa, Iszulu/Szulu/Zulu,and a language whch s not Englsh and not one of the lsted 6. Englsh s the language of partcular nterest, rankng 8 th most common out of all possble responses. Hence, 8 categores were chosen. Ths set of dummes comprses an alternatve to the prevous set of language dummes. Base: Englsh. Manageral occupaton Clercal occupaton Agrcultural occupaton Dummy varable takng the value of 1 f the occupaton of the ndvdual s of a manageral, professonal or sem-professonal; clercal, sales or sklled servce; agrcultural, artsan or operatng nature. Base: Elementary or domestc occupatons. Agrcultural ndustry Constructon ndustry Trade ndustry Dummy varable takng the value of 1 f the occupaton of the ndvdual s n the ndustry of agrculture or mnng; manufacturng, utltes or constructon; trade, transport or fnance. Base: domestc servces ndustry. Informal A dummy takng the value of 1 for employment n the nformal sector. Derved from the man category of occupaton and/or, for self-employed persons, an absence of regstraton of the busness for VAT or wth the regster of companes, the Commssoner of unemployment nsurance or the Commssoner of workmen s compensaton. Self employed A dummy takng the value of 1 f the respondent earned ncome through a busness or other actvty of ther own. Respondents who also earned ncome from an employer would also report a 1 for ths varable. 1
Table 4 Labour Force Partcpaton Double Hurdle and Heckman Sample Selecton Models All Workng Age Adults (1) (2) (3) (4) Defnton of Unemployed Narrow Broad Narrow Broad Asan 0.040** -0.089** -0.042-0.153* (4.06) (9.12) (0.69) (2.49) Coloured 0.160** 0.086** 0.116** 0.091** (23.10) (12.96) (3.28) (2.70) Whte 0.064** -0.034** 0.090 0.004 (9.55) (5.04) (1.94) (0.09) Afrkaans 0.127** 0.070** (6.29) (3.61) Asan x Afrkaans 0.011 0.063 (0.12) (0.70) Coloured x Afrkaans -0.063-0.063 (1.57) (1.55) Whte x Afrkaans -0.163** -0.132** (3.36) (2.58) Englsh 0.140** 0.047 (4.00) (1.38) Asan x Englsh -0.053 0.017 (0.75) (0.24) Coloured x Englsh -0.140** -0.125* (2.86) (2.40) Whte x Englsh -0.121* -0.050 (2.12) (0.85) Unoffcal language -0.017-0.037 (0.36) (0.82) Male 0.135** 0.117** 0.136** 0.117** (38.96) (35.71) (38.97) (35.74) Household head 0.170** 0.120** 0.170** 0.120** (43.25) (30.01) (43.25) (30.04) Marred -0.091** -0.143** -0.091** -0.142** (20.14) (31.27) (20.09) (31.13) Marred x Male 0.272** 0.306** 0.272** 0.306** (40.65) (47.01) (40.64) (47.00) Age 0.085** 0.103** 0.085** 0.103** (111.82) (126.78) (111.84) (126.83) Age2-0.001** -0.001** -0.001** -0.001** (98.68) (111.96) (98.70) (112.03) Prmary school -0.001-0.005** -0.001-0.005** (1.46) (5.44) (1.40) (5.41) Secondary school 0.028** 0.019** 0.028** 0.020** (27.70) (19.61) (27.80) (19.87) Dploma/certfcate 0.239** 0.171** 0.238** 0.171** (28.05) (20.93) (27.97) (20.91) Degree 0.009 0.020 0.005 0.017 (0.55) (1.16) (0.29) (0.98) Other educaton -0.163** -0.148** -0.161** -0.147** (11.14) (9.85) (11.02) (9.77) Rural -0.083** -0.062** -0.082** -0.062** (19.56) (15.41) (19.31) (15.35) Eastern Cape -0.150** -0.093** -0.145** -0.090** (19.45) (12.02) (18.67) (11.61) Northern Cape -0.033** -0.012-0.043** -0.020* (3.63) (1.31) (4.67) (2.16) Free State 0.010 0.015 0.018* 0.019* (1.18) (1.77) (2.00) (2.26) Kwa-Zulu Natal -0.053** -0.001-0.046** 0.003 (6.67) (0.10) (5.77) (0.40) North West -0.042 0.020* -0.035** 0.025** (4.90)** (2.46) (4.00) (2.99) Gauteng 0.029 0.069** 0.035** 0.073** (3.74)** (9.39) (4.52) (9.80) Mpumalanga 0.004-0.011 0.012-0.006 (0.47) (1.31) (1.39) (0.73) Northern -0.120-0.098** -0.113** -0.094** (13.89)** (11.40) (12.94) (10.82) y1997-0.001 0.009-0.001 0.009* (0.13) (2.30)* (0.23) (2.18) y1998 0.049 0.039 0.049** 0.039** (10.16)** (8.52)** (10.14) (8.51) Observatons 164243 164243 164243 164243 Table reports coeffcent estmates; Robust z statstcs n parentheses; * sgnfcant at 5%; ** sgnfcant at 1% 2