Jing You 1. November 2010

Transcription

1 Evaluatng poverty duraton and transton: A spell-approach to rural Chna Jng You 1 1 The Unversty of Manchester, Manchester, UK ng.you@manchester.ac.uk Brooks World overty Insttute ISBN : November 2010 BWI Workng aper 134 Creatng and sharng knowledge to help end poverty

2 Abstract Ths artcle uses a dscrete-tme multvarate duraton model to study poverty transton n rural Chna between 1989 and The analyss dentfes nonlnear negatve duraton-dependence for both ext and re-entry rates of poverty. There s sgnfcant dfference n hazard rates of ext and reentry assocated wth geographc locaton and educatonal level of households, but less related to gender, occupaton or ethnc background of household head. The factors facltatng households endng a poverty spell are found to be educaton, land ownershp, asset accumulaton, health nsurance and out-mgraton, whle larger famly sze and dependence rato may reduce the chance of ext. Ths artcle s forthcomng n Appled Economcs Letters, Keywords: duraton analyss, hazard model, persstent poverty, rural Chna Jng You s hd Student n Economcs, School of Socal Scences, The Unversty of Manchester, and Doctoral Student Assocate, Brooks World overty Insttute, The Unversty of Manchester, UK. 2

3 1. Introducton overty dynamcs n rural Chna have been well examned from the perspectves of ther transent and chronc components (Jalan and Ravallon, 1998a, b, 2000) and the probablty of becomng poor (McCulloch and Calandrno, 2003; Zhang and Wan, 2006). Whle useful for understandng the changes of households poverty status wthn a gven perod, they have weak explanatory power for the persstent poverty whch has been emergng snce the late 1990s. Chen and Ravallon (2008) fnd that, although the ncdence of poverty dropped sharply by 68 percent between 1981 and 2005, 47 percent of ths reducton had happened before The mssng explanatory factor may be tme-varyng and ndvdual-specfc determnants of households poverty transtons (Bgsten and Shmeles, 2008). If so, the spell-approach s more nsghtful, as t reveals ndvdual households traectores of sldng n and out of poverty spells and the determnants of these repeated shfts. Ths approach has been wdely appled to poverty transtons n developed countres (e.g., n the UK by Devcent, 2002, 2010; n Italy by Devcent and Gualter, 2007) and a few developng economes (e.g., Ethopa by Bgsten and Shmeles, 2008). However, lttle has been known for rural Chna. As far as we are aware, only Glauben et al. (2006) use duraton analyss to measure to what extent and how households ndvdual past (non-)poverty experence affects ther probabltes of sufferng or escapng poverty n future. Ther study shows that past exposure to poverty may be less decsve, because both ext and entry rates of poverty tend to ncrease at longer duraton. Nevertheless, ths may lack representatveness for rural Chna, especally the poor areas, as ther samples were selected from Zheang provnce only, whch s coastal and one of the rchest provnces. Moreover, ther hazard model s based on the presumpton of underlyng contnuous data wthout unobserved heterogenety, whch may be an over-smplfcaton and overestmate (underestmate) negatve (postve) duraton-dependence and the proportonate response of the hazard to an estmated negatve (postve) coeffcent (Jenkns, 2005). Ths artcle offers new evdence on the shape and correlates of transton n and out of poverty for rural Chna, by usng a hghly representatve panel and a dscrete-tme hazard model controllng for unobserved heterogenety. The next secton sets up the model. Secton 3 descrbes the data and dscusses the results. Secton 4 concludes. Analytcal framework Modellng poverty duraton There are two states, poverty and non-poverty, between whch households shft over tme. 1 Followng Bgsten and Shmeles (2008), the (dscrete) survval tme s ndexed by t, t 2,K t, Kt wth equal ntervals for smplcty. The rates of ext pertan to households who ust started a poverty 1 k 1 As Bgsten and Shmeles (2008), there s presumably no correlaton between repeated spells for the same household over tme,.e. ndependence between multple spells. In fact, we splt households nto subects wth sngle-spells and then pooled them for estmaton. 3

4 spell. 2 Among them, d households end ther poverty spells at t. n households stay poor n at least waves and are at rsk of movng out of poverty at t + 1. The survval functon s therefore defned by d Sˆ ( t ) = 1 (1) t t n Correspondngly, the hazard rates for endng a poverty spell at t are calculated by h ( t ) Sˆ ( t ) ( t ) ˆ S( t 1) Sˆ ( t ) 1 f = 1 = ˆ S (2) f > 1 By the same token, the poverty re-entry rates refer to those who ust started a non-poverty spell. The hazard rates of endng non-poverty spells are calculated analogously. Nevertheless, there has been growng concern over spurous transton between the two states. A cause may be the measurement errors n consumpton data. A household mght be msclassfed as poor smply because ts consumpton seems to fall below a certan poverty lne, but ths may be a measurement error rather than evdence of adverse events. Followng Devcent (2002), ths problem could be addressed by adustng the poverty lne so that households are deemed to be poor (non-poor) only f ther per capta consumpton falls below (surpasses) 90 (110) percent of the unadusted poverty lne at US$1.25/day. Another cause s the constructon of survval and hazard functons tself. Equatons 1 and 2 are essentally aggregate measures of transton nto and out of poverty for the full sample, whle some households sharng certan characterstcs mght reman poor/non-poor for a long tme. These characterstcs can be ether observed or unobserved, such as the lack of endowments and ntrnsc ncapabltes. It s hence necessary to nvestgate whether the revealed shape of poverty transton s a common feature. In ths artcle, ths s done n two ways. Non-parametrc estmates of survval and hazard functon are replcated for varous sub-groups. We also mplement a multvarate analyss to explore the correlates of ext from and re-entry nto poverty. Explanng the correlates of poverty transton For the household n the tme nterval, a standard dscrete-tme hazard model takes the followng specfcaton: h ( t ) = ( T = t T t ) r (3) 2 The concept employed here s n lne wth Devcent (2002, 2010) and Bgsten and Shmeles (2008). A household ust startng a (non-)poverty spell at t means t was n (non-)poverty at t-1 and shfts out of ths state at t. Our sample contans seven waves of the surveys. Therefore, the frst (non-)poverty spell starts at the second wave and the maxmum duraton s fve. 4

5 where T s the tme a (non-)poverty spell ends. Emprcally, a complementary log-log hazard functon s used to model poverty ext and re-entry rates separately. Followng Devcent and Gualter (2007), the probablty that household escapes from poverty at duraton d at tme t, gven t has stayed n poverty spells up to t, s expressed by: ( d X υ ) = 1 exp[ exp( f ( d ) + X β u )] e, + (4) where the vector X contans household-specfc tme-varyng characterstcs; f ( d ) s a functon explctly modellng how ext rates depend on the duraton that households have spent n poverty spells; u log( υ ) denotes the unobserved heterogenety whch s tme-nvarant and common across s all poverty spells. 3 Smlarly, the probablty that the household re-enters poverty at duraton d at tme t, gven t has been non-poor up to t, s wrtten by: N N N ( d X υ ) = 1 exp[ exp( f ( d ) + X β u )] N r, + (5) In order to ntegrate out the unobservables n estmatng the hazard models, normal dstrbutons are assumed for u and f d and f N ( d ) u. 4 To make models more flexble, the baselne hazards ( ) N take a fully non-parametrc form motvated by Devcent (2002; 2010): a set of duratonnterval-specfc dummes, at whch households are at rsk of shftng out of (non-)poverty spells. Data and emprcal results Data A balanced panel contanng 1,429 rural households s extracted from seven rounds of Chna Health and Nutrton Surveys (CHNS) n 1989, 1991, 1993, 1997, 2000, 2004 and The samples are bascally equally dstrbuted n seven provnces, from coastal to nland Chna. 5 Table 1 summarses the varables used n estmaton. 3 Households ntal (non-)poverty status s assumed to be exogenous to ther characterstcs. Devcent s (2010) model controls for endogenety of ntal condtons, whch may lead to our future research. 4 We also expermented wth Gamma and Heckman and Snger s (1984) mxed mass-pont dstrbutons, but maxmsaton procedures faled to converge to a soluton. 5 Coastal provnces are Jangsu and Shandong. Central provnces are Henan, Hube and Hunan. Western provnces are Guangx and Guzhou. 5

6 Table 1 Descrptve statstcs Varables Mean SD Mean SD hh per capta consumpton hh sze age of hh head dependence rato % male adults % complete prmary edu. wthn the hh % complete at least secondary edu. wthn the hh ln(farm land) ln(value of agrcultural assets) % local off-farm employment wthn the hh % vllage out-mgraton (outmgraton networks) % havng health nsurance wthn the hh % sown land affected by natural dsasters wthn the provnce Note: All monetary varables are n 2006 prces. relmnary exploraton of transton probabltes suggests co-exstence of persstence and transton of poverty n rural Chna. The upper panel of Table 2 shows that percent of households had experenced at least one perod of poverty wthn the sample tme span. Among those who were poor at the begnnng of the surveys, percent ended up n poverty agan. The degree of ths persstent hardshp s even greater (64.09 percent) f measured aganst the adusted poverty lne. In comparson, however, percent of the ntally non-poor were lkely to retan ther lvelhood poston at the end of the surveys. As one mght predct, usng the adusted poverty lne makes t harder to reman non-poor (78.04 percent). Meanwhle, there s also evdent poverty transton. Of the ntally poor households, percent successfully moved out of deprvaton, whle only percent of those who were non-poor slpped back nto poverty. Table 2 overty transton matrx (%), overty Non-poverty Total Unadusted poverty lne overty Non-poverty Total Adusted poverty lne overty Non-poverty Total Estmated survval and hazard functons The estmated survval and hazard functons n Table 3 ndcate strong negatve duratondependence assocated wth the rates of poverty re-entry. Ths mples a good chance for 6

7 households to escape from poverty n the long term. For those who ust started a non-poverty spell, 65.7 percent successfully remaned above the unadusted poverty lne, after spendng fve perods n non-poverty. Ther re-entry rates quckly approach to zero. In the case of unadusted poverty lne, f a household has survved for fve perods, t has only a 1.6 percent lkelhood of sldng nto poverty n the next perod. Table 3 Survval and hazard functons of ns and outs of poverty Tme snce the start of overty ext Unadusted Adusted spell Sur.(s.e.) Ext (s.e.) Sur. (s.e.) Ext (s.e.) 1 1 (.). (.) 1 (.). (.) (0.009) (0.011) (0.009) (0.011) (0.012) (0.014) (0.011) (0.013) (0.013) (0.017) (0.012) (0.016) (0.014) (0.034) (0.013) (0.030) (0.013) (0.044) (0.013) (0.039) overty re-entry Tme snce the start of Unadusted Adusted spell Sur. (s.e.) Re-ent. (s.e.) Sur. (s.e.) Re-ent. (s.e.) 1 1 (.). (.) 1 (.). (.) (0.013) (0.017) (0.012) (0.015) (0.015) (0.014) (0.014) (0.010) (0.016) (0.010) (0.014) (0.006) (0.016) (0.007) (0.014) (0.005) (0.017) (0.007) (0.015) (0.005) Note: Kaplan-Meer estmates. The ext rates are also negatvely assocated wth duraton n the frst three perods n poverty for those who ust started a poverty spell. In other words, the longer the tme spent n poverty, the lower the probablty of escape for these households s becomng. The average length of a poverty spell s 2.55 perods, whch are equvalent to 5.1 years f countng the real gap of years between surveys. Meanwhle, t s also worth notng that after four perods n poverty, ext rates tend to ncrease, sgnallng an opportunty for the poor to escape at longer duraton. Ths seemngly mxed duraton-dependence for ext wll be examned more carefully by the multvarate analyss n the next sub-secton. As aforementoned, adusted poverty lnes tend to brng about more dffcultes for households sldng nto and out of poverty. Ths s demonstrated by estmates n Table 3. The hazard rates of poverty ext (re-entry) are hgher (lower) n the case of adusted poverty lnes, relatve to the unadusted one. In order to best accommodate measurement errors n consumpton data, from here, ths artcle keeps usng the adusted poverty lnes to splt households poverty/non-poverty epsodes n the analyss. As noted n Secton 2, the hazard rates n Table 3 are estmated, based on the assumpton of homogenous populaton. We further consder whether poverty ext and re-entry dverge for categores defned by households geographc locaton, natonalty, household heads educatonal level, gender and occupaton. For each of the sub-groups, the dfferences of hazard rates between sub-categores are examned by log-rank and Wlcoxon tests (Table 4). Wth respect to the lkelhood of extng poverty, dstncton exsts across dfferent educaton levels and regons at one percent sgnfcance level, whle for the rsk of re-enterng poverty, varaton s only found across regons at 10 percent sgnfcance level. As Glauben et al. (2006), we also observe frst a decreasng and then an ncreasng relatonshp between ext rates and the duraton of poverty 7

8 spells n coastal provnces, but consstently decreasng ext rates n western provnces. Households resdng n less developed regons are more lkely to be trapped n persstent poverty. Ths supports our argument that Glauben et al. s (2006) concluson does not represent the general stuaton n rural Chna. Table 4 Heterogenety n hazard rates (adusted poverty lne) Ext Re-entry Wlcoxon Log-rank test test Category Sub-group Log-rank test Regon Gender Occupaton Educaton costal, central, western male, female hh heads farmer, sklled worker, non-sklled worker, professonal llterate, prmary, secondary, tertary ethnc maorty, Natonalty ethnc mnortes Note: p-values are n parentheses (0.00) 1.13 (0.29) 3.21 (0.36) (0.00) 0.59 (0.44) 3.10 (0.21) 0.43 (0.51) 1.80 (0.62) (0.01) 0.06 (0.81) 5.16 (0.08) 1.87 (0.17) 5.67 (0.13) 0.13 (0.73) 0.47 (0.49) Wlcoxon test 2.62 (0.27) 1.89 (0.17) 4.63 (0.20) 0.92 (0.82) 0.68 (0.41) Correlates of poverty transton The LR test (Table 5) shows that unobserved heterogenety matters n poverty ext, but seems to be less of a problem n re-entry regressons. Negatve duraton-dependence can be confrmed n cases of both poverty re-entry and ext. However, t would dsappear after four perods n nonpoverty for the former and two perods n poverty for the latter. The multvarate analyss seems not to support the ncreasng hazard rates of ext at longer duraton revealed by the non-parametrc examnaton. Moreover, the magntude of estmates suggests that the negatve relatonshp between spell duraton and hazard rates s non-lnear. Among varous demographc characterstcs, larger famly sze and hgher dependence rate are maor mpedments to poverty ext and drvers of poverty re-entry. rmary educaton can reduce the rsk of re-entry, whle secondary and tertary educaton s more helpful to chances of escape. Gender and ethnc background of a household head appear not to exert much nfluence on poverty transtons, whle occupaton may play a role. Households led by non-farmer heads are more lkely to move out of poverty. As expected, more asset accumulaton, land ownershp, out-mgraton and health nsurance are conducve to shftng out of entrenched deprvaton. When researchng the mpact of aggregate shocks on poverty ext, weather rsk features. Compared wth coastal provnces, lvng n less developed western and central regons may also hamper prosperty. 8

9 Table 5 Covarates of hazard rates of poverty ext and re-entry overty ext regresson overty re-entry regresson Indep. varable Wthout heterogenety Wth normal heterogenety Wthout heterogenety Wth normal heterogenety Duraton dependence D (0.080) *** (0.081) *** (0.156) *** (0.156) *** D (0.098) *** (0.099) *** (0.287) *** (0.287) *** D (0.094) (0.095) (0.387) *** (0.387) *** D (0.132) (0.133) (0.460) *** (0.460) *** D (0.141) (0.143) ( ) ( ) Household characterstcs hh sze (0.021) *** (0.021) *** (0.036) *** (0.036) *** hh head s age (0.003) *** (0.003) *** (0.005) *** (0.005) *** % completng prmary edu. (0.111) (0.113) (0.230) * (0.231) * % completng at least sec. edu. (0.178) *** (0.182) *** (0.451) (0.454) % male adults wthn hh (0.100) (0.102) (0.192) (0.193) gender of hh head (male=1) (0.110) (0.112) (0.235) (0.237) dependency rato (0.107) *** (0.108) *** (0.225) (0.226) ethnc mnortes of hh head (=1) (0.121) (0.123) (0.268) (0.270) hh head s occup.: farmer (0.093) (0.095) (0.196) (0.197) hh head s occup.: unsklled labour (0.119) ** (0.120) ** (0.279) (0.281) Wealth ln(farm land) (0.019) ** (0.019) ** (0.036) (0.036) ln(value of agr assets) (0.008) *** (0.008) *** (0.017) ** (0.017) ** rasng lvestock

10 (yes=1) (0.069) * (0.070) * (0.136) (0.137) Access to labour market % local off-farm empl. wthn hh (0.100) *** (0.100) *** (0.196) *** (0.198) *** vllage out-mg networks (0.661) *** (0.678) *** (1.631) (1.649) Socal protecton % havng health nsur. wthn hh (0.076) *** (0.077) *** (0.230) *** (0.231) *** Aggregate shocks prov. % land n natural dsasters (0.662) *** (0.667) *** (1.062) (1.068) Geographc locatons lvng n central provnces (yes=1) (0.082) *** (0.084) *** (0.163) (0.164) lvng n western provnces (yes=1) (0.087) *** (0.089) *** (0.201) (0.203) Log-lkelhood LR test of 2 2 ρ = σ u /(1 + σ u ) = 0 (p-value) Note: *, **, *** denote 10%, 5% and 1% sgnfcance levels. Standard errors are n parentheses. Conclusons The analyss dentfes negatve duraton dependence for poverty ext and re-entry n rural Chna n the perod Ths ndcates that poverty tends to become a persstent state for those who started out wth a poverty spell. olces amng to end current poverty may also facltate households movng out of poverty n the future. The catalyst for poverty ext and mpedments to poverty re-entry nclude educaton, asset accumulaton, health nsurance and out-mgraton. Lvng n less developed regons, larger famly sze and dependence rate reduce the possblty of escape from poverty. 10

11 References Bgsten, A. and Shmeles, A. (2008). overty transton and persstence n Ethopa: World Development 36, Chen, S. and Ravallon, M. (2008). Chna s poorer than we thought, but no less successful n the fght aganst poverty. World Bank olcy Research Workng aper, No Washngton, DC: The World Bank. Devcent, F. (2002). overty persstence n Brtan: a multvarate analyss usng BHS, Journal of Economcs Supp 9, Devcent, F. (2010). Estmatng poverty persstence n Brtan. Emprcal Economcs forthcomng [avalable at DOI: /s ]. Devcent, F. and Gualter, V. (2007). The dynamcs and persstence of poverty: evdence from Italy. LABOR Workng aper, No. 63. Turn, Italy: Collego Carlo Alberto. Glauben, T., Herzfeld, T. and Wang, X. (2006). The persstence of poverty n rural Chna: applyng an ordered probt and a hazard approach. aper presented at the Internatonal Assocaton of Agrcultural Economsts Conference, Gold Coast, Australa. Heckman, J. J. and Snger B. (1984). A method for mnmzng the mpact of dstrbutonal assumptons n econometrc models for duraton data. Econometrca 52, Jalan, J. and Ravallon, M. (1998a). Transent poverty n postreform rural Chna. Journal of Comparatve Economcs 26, Jalan, J. and Ravallon, M. (1998b). Determnants of transent and chronc poverty: Evdence from rural Chna. World Bank olcy Research Workng aper, No Washnton, DC: The World Bank. Jalan, J. and Ravallon, M. (2000). Is transent poverty dfferent? Evdence for rural Chna. Journal of Development Studes 36, Jenkns, S. (2005). Survval analyss. Unpublshed manuscrpt. Insttute for Socal and Economc Research, Unversty of Essex, Colchester, UK. Onlne resource avalable at: (accessed 23 November 2010). McCulloch, N. and Calandrno, M. (2003). Vulnerablty and chronc poverty n rural Schuan. World Development, 31, Zhang, Y. and Wan, G. (2006). An emprcal analyss of household vulnerablty n rural Chna. Journal of Asa acfc Economy 11,

12 Executve Drector rofessor Davd Hulme Research Drector rofessor Armando Barrentos The Brooks World overty Insttute (BWI) creates and shares knowledge to help end global poverty. BWI s multdscplnary, researchng poverty n both the rch and poor worlds. Contact: Brooks World overty Insttute The Unversty of Manchester Humantes Brdgeford Street Buldng Oxford Road Manchester M13 9L Unted Kngdom Our am s to better understand why people are poor, what keeps them trapped n poverty and how they can be helped - drawng upon the very best nternatonal practce n research and polcy makng. The Brooks World overty Insttute s chared by Nobel Laureate, rofessor Joseph E. Stgltz. Emal: bwp@manchester.ac.uk