POVERTY DYNAMICS IN NAIROBI S SLUMS, TESTING FOR STATE DEPENDENCE AND HETEROGENEITY

Transcription

1 POVERTY DYNAMCS N NAROB S SLUMS, TESTNG FOR STATE DEPENDENCE AND HETEROGENETY Ousmane FAYE APHRC - Narob, Kenya Ths verson: August, 2008 Prelmnary and ncomplete draft; Do not quote, do not crculate: for comments only. Abstract: n ths paper, we have been able to offer some nsghts n the dynamcs of poverty n Narob s slums. A very nterestng result of ths paper s that there s substantal state of dependence n poverty after controllng for nal poverty status and for panel retenton. Keywords: Poverty dynamcs, state dependence, unobserved heterogeney, attron, smulated maxmum lkelhood, urban poverty. JEL Classfcaton: C5, C35, 32, O8, R23 Correspondence: Afrcan Populaton and Health Research Centre (APHRC), Shelter Afrque Centre - Longonot Road, Upper Hll Narob, Kenya. P.O. Box GPO Tel: //2, Emal: ofaye@aphrc.org

2 . ntroducton What are the factors assocated wh becomng or remanng poor? Who are the ndvduals at rsk of enterng or exng poverty? s the same ndvduals who are stuck n poverty? n other words, does poverty experenced n one perod has a causal effect on future poverty? Do ndvduals who are poor have partcular characterstcs makng them prone to poverty? Answers to these questons are central to understand poverty and then nform publc polces amed at fxng. Furthermore, when poverty perssts over tme, polcy makers have good reasons for concern over the mpact of such long lastng deprvaton. Also, snce publc resources are lmed, s mportant to understand the dynamc of poverty for a better targetng of the poverty allevaton polces. Ths paper explores poverty persstence and the determnants of transon nto and out of poverty usng panel data collected n two Narob s demographc survellance ses. There are two man processes that may generate poverty persstence. Frst, the fact of experencng poverty n a specfc tme perod, mght n self ncrease the probably of beng poor n subsequent perods (through human capal deteroraton, decreasng self-esteem, etc.). Such a process s sad to exhb state dependence (Heckman, 978). Second, dfferences n characterstcs that make ndvduals prone to poverty mght ncrease the chance of fallng nto poverty and persstent over tme. n that case, ndvduals who experence poverty at tme t because of these (possbly unobserved) characterstcs wll also be lkely to be poor n any other perod because of the very same characterstcs. Ths process s referred as unobserved heterogeney effects. Dstngushng between the two processes s mportant, snce the polcy mplcatons are very dfferent. f poverty persstence s due to state dependence effects, then a polcy amed at fghtng poverty va monetary transfers makes sense. Such a polcy wll help not only wpng out current poverty but also preventng future poverty. n contrast, f poverty persstence s due to unobserved heterogeney effects, a polcy of monetary transfers wll not be the most effectve opton. However, f understandng these two sources of poverty persstence s crucal for desgnng effectve and successful poverty allevaton polces, s worth notng that few studes n Afrca have nvestgated these ssues, despe the prory gven to fghtng poverty n most the Afrcan countres. The reason s that tacklng these ssues requres accurate and comprehensve soco-economc data collected regularly on the same ndvduals or households over tme. Alas, such data are not oftenly collected n the regon.

3 Ths paper takes advantage of the unquely rch dataset from the Narob Urban Health and Demographc Survellance System (NUHDSS). The NUHDSS s run by the Afrcan Populaton and Health Research Center (APHRC). t was set up n 2000 to provde a platform for nvestgatng changng lnkages between urbanzaton, poverty and health and to evaluate the mpact of nterventons amed at mprovng the wellbeng of slum resdents. t covers about 60,000 people lvng n 5,000 households n two slum settlements n Narob Cy, Korogocho and Vwandan. The survellance nvolves vss to all the households once every four months to update nformaton on pregnances and pregnancy outcomes, mgraton, epsodes of morbdy, health seekng behavor, mortaly and causes of death, vaccnaton coverage, maral status, school attendance, lvelhood sources, possessons, shocks, and vulnerables, ncludng copng strateges of households and ndvduals. The paper s structured as follows. Secton 2 revews the related lerature. The estmaton strategy s outlned n the secton 3. Secton 4 descrbes the data and dscusses the explanatory varables. Dscusson of the results follows n secton 5, whle secton 6 concludes. 2. Related lerature Snce Heckman s works (978, 98), the queston arses whether persstence n economc phenomena s due to dfferences n ndvdual characterstcs or due to causal effects of past on future outcomes. Examples range from unemployment ssues (Heckman, 978, 98; Arulampalan et al., 2000), persstence n low pay (Stewart and Swaffeld, 999; Cappellar and Jenkns, 2004) and analyss of poverty persstence (Bewen, 2004; Cappellar and Jenkns, 2002). Dfferent approaches have been used to study the dynamcs and persstence of these economc phenomena. A semnal work by Lllard and Wlls (978) uses the estmaton of components-of-varance models to study poverty n relaton to the evoluton of earnngs or ncome over tme n a sample of male household heads. Usng ther estmates of the permanent and transory varance components of male earnngs, Lllard and Wlls derved probables of varous tme sequences of poverty or low-earnngs status. Bane and Ellwood (986) use a hazard rate approach to measure poverty persstence. They study ndvdual spells of poverty and estmate the probably of endng these poverty spells, allowng for duraton dependence n the hazard rate. But a major drawback of Bane and Ellwood s that focuses on sngle spells whle many ndvduals n ther sample experence more than one poverty spell n the observed tme frame. Usng the hazard rate approach to study ndvdual poverty persstence over lfe tme, Stevens (999) addresses ths ssue. She accounts for multple spells of poverty and ncorporates spell duraton, ndvdual and household

4 characterstcs, and unobserved heterogeney. Her fndngs hghlght the mportance of consderng multple spells n an analyss of poverty persstence, wh half of those who end poverty spells returnng to poverty whn four years. What s common n these prevous studes s the effort to capture the effects of current on future poverty. However, except Stevens (999), they do not clearly dstngush between the potental sources of poverty persstence. Recent studes explore the causes of poverty persstence usng dynamc dscrete choce models whch control for state dependence and unobserved heterogeney. Notceable studes nclude Stewart and Swaffeld (999), Cappellar and Jenkns (2002, 2004), Devcent (2002), Pogg (2003). Most of these studes, assume a frst-order statonary Markov chan for state dependence, and combne wh ndvdual fxed-effect or random-effects models to fx the unobserved heterogeney ssue. But s worth notng that the model proposed by Cappellar and Jenkns (2004) goes further, allowng accountng for multple endogenous selecton mechansms wh panel data such as attron, nal condons, etc. 3. Estmaton strategy State dependence effects are usually analyzed usng dynamc dscrete models wh unobserved heterogeney. f tme t = K,, T, denotes poverty status of ndvdual = K,, N at { α + ϕ z + δ + c} = ε < () { L} s an ndcator functon descrbng the evoluton of poverty condonal on s poverty status at the prevous perod, a vector of exogenous varable z, and two unobserved characterstcs δ and ε. The ndvdual-specfc term δ stands for all unobserved determnants of poverty that are tme-nvarant for a gven ndvdual. The resdual term ε s dosyncratc and s assumed to follow a normal dstrbuton wh zero mean, un varance: ε ℵ( 0,). The bnary varable s equal to f the dsposable ncome s below the threshold c referred as the poverty lne, and 0 otherwse. The value of α determnes how takes n state dependence. fα > 0, experencng poverty at tme t-( = ) ncreases the chance of beng poor at tme t ( = ): (, δ ) > Pr( 0, δ ) Pr = = The dsposable ncome s specfed here as a lnear functon of ndvdual poverty status at tme t-, personal characterstcs and a normally dstrbuted error term.

5 Pr ( ε < δ α ) > Pr( ε < δ ) However, s noteworthy that the specfcaton above does not properly control for ndvdual observed or unobserved heterogeney. n the presence of ndvduals characterstcs, such as ably, motvaton, or ntellgence, that make them prone to poverty, we wll also observe (, δ ) > Pr( 0, δ ) = = gven that these characterstcs persst over Pr tme and even thoughα = 0. Therefore, for testng of genune state dependence, s crucal to correctly control for ndvdual heterogeney. A strategy to address ths ssue conssts of mposng a dstrbuton structure toδ and nterpretng equaton () as a random-effect model. Then one can obtan a lkelhood functon for α by ntegratng out the unobserved term δ (Arellano and Honoré, 200; Cappellar and Jenkns, 2002, 2004; Bewen, 2007). But the queston arses whether results depend on the mposed functon forms and dstrbuton assumptons. Another ssue s that poverty status n the nal perod may also be correlated wh the factors captured byδ. Ths ssue s usually referred as the nal condon problem. gnorng can lead to dstorted estmates, partcularly n short panels (Arulampalan et al., 2000; Heckman, 98). The nal condons problem can be solved n dfferent ways. One way to deal wh, suggested by Wooldrdge (2005/2002), s to let the nal condons be random by usng the jont dstrbuton of all outcomes of the endogenous varables condonal on observed and unobserved heterogeney. n ths paper, we nvestgate the state dependence effects whle accountng for these heterogeney ssues usng Cappellar and Jenkns model (2002, 2004). Cappellar and Jenkns buld on Stewart and Swaffeld (999) and develop a model of transon probables that accounts for nal condons problem but also for panel attron between tmes t- and t. The nterest n Cappellar and Jenkns model s that allows accountng smultaneously for multple endogenous selecton ssues (e.g. nal condons, attron, etc.) and testng for gnorably of one or more of these selecton mechansms. Moreover, the settng proposed by Cappellar and Jenkns crcumvents the assumpton of no feedback effect from the dependent varable on future value of the explanatory varables, unlke most of the models nvestgatng poverty persstence (see Bewen, 2004, 2007; Aassve et al., 2004). as follows: n Cappellar and Jenkns model, equaton () s re-specfed as a dynamc prob model

6 (, R = ) = Φ( α + z + δ + ε ), ( υ = δ + ε ) ( 0,) = ℵ Pr ϕ (2) Where and R represent nal poverty status and attron status respectvely. The model allows a smple test for state dependence based on whether α > 0 ; f true, poverty state at tme t depends on beng poor at tme t-. The nal condon for poverty status s mplemented by a prob model as follows: ( ) = Φ( + λ + µ ), ( θ = λ + µ ) ( 0,) Pr = β x ℵ (3) The retenton status (.e. the probably of not sufferng from attron between t- and t) s also gven as a prob model: Pr ( R ) = Φ( + η + ξ ), ( ψ = η + ξ ) ( 0,) = w ℵ χ (4) Where: x and w are vectors of explanatory varables of nal poverty status and retenton status equatons. The jont estmaton of the three equatons nvolves the evaluaton of the log-lkelhood over =,,N based on a jont trvarate probably. The contrbuton f each ndvdual to the log-lkelhood s as follows: LogL = R + Log [ Φ ( κ ( α + ϕ z ), κ χ w, κ β x ; κ κ ρ, κ κ ρ, κ κ ρ )] 3 ( -R ) Log[ Φ ( κ χ w, κ β x ; κ κ ρ )] Where: T 2 R R R T R 3 T 2 R (5) κ T = R 2 ; κ = 2R ; κ = 2 κ are the correspondng sgn varables that to or - dependng on whether the T, R, observed bnary outcome equals or 0. The estmaton assumes that the error terms of the three equatons (2), (3), and (4) are multvarate normal dstrbuted wh zero mean, un varances, and a covarance matrx Σ. We allow, however, for correlated dsturbances: ρ = corr ρ = corr 2 ρ = corr 3 ( θ, ψ ) = cov( λ, η ) ( θ, υ ) = cov( λ, δ ) ( ψ, υ ) = cov( η, δ ) (6) These three correlaton coeffcents (ρ, ρ2, ρ3) wll be estmated and wll represent the extent to whch unobserved covarates jontly determne the outcomes of nterest. t s worth mentonng that proceedng ths way, the coeffcent estmates from the trvarate prob model

7 wll account for unobserved correlaton among the outcomes and wll be therefore less based and more effcent than those produced by three ndependent models. The estmaton of (5) requres the computaton of dervatves of thrd order ntegrals for whch no general solutons exst. However, the problem can be addressed by recently developed smulaton technques. The method of smulated maxmum lkelhood allows the estmaton of a trvarate prob model by usng the GHK (Geweke-Hajvasslou-Keane) smooth recursve estmator (see Greene, 2003). The GHK smooth recursve estmator decomposes the orgnal three-dmensonally correlated error terms nto a lnear combnaton of uncorrelated one dmensonal standard normal varables. The trvarate dstrbuton s thus transformed nto three sequentally condoned unvarate dstrbutons. n order to evaluate the resultng ntegral, D Halton draws of these standard normal varables are taken from truncated normal dstrbutons, and a sample average of the smulated lkelhoods s used to estmate the probably that enters the lkelhood functon. 4. Data and descrptve statstcs The analyss s based on data from the Narob Urban Health and Demographc Survellance System (NUHDSS) run by APHRC snce The dataset contans nformaton on the well-beng of the approxmately 3,000-5,000 households (60,000 ndvduals) that lve at any one pont n tme n two of Narob s man nformal settlements Vwandan and Korogocho; n partcular the 3 rd and 3 th round of the DSS undertaken n 2003 and Data on lvelhood, possessons, shocks, and vulnerables, ncludng copng strateges of households and ndvduals have been collected durng these two rounds. The focus of ths paper s on household expendure as a measure of welfare. An ndvdual s defned as poor f her household equvalent expendure s less than a gven poverty lne. The latter corresponds to Narob offcal poverty lne from Kenya Natonal Bureau of Statstcs (KNBS). The Narob poverty lnes are 2640 and 293 Kenya Shllngs per month per person (n adult equvalent terms) n 2003 and 2006 respectvely. The expendure varable consdered s the equvalent household expendure, obtaned after addng up all expenses of the household comprsng food, non-food and durable ems, and then dvdng the total by the number of equvalent adults (consderng a chld as half of an adult). Our un of analyss s the ndvdual. We assume an equal sharng of resources whn the household; each member receves the same value of the equvalent household expendure. Table shows the raw poverty transon matrx constructed from the dataset. The frst part of the table ndcates poverty transon for all ndvduals present n the two rounds.

8 Table No : Poverty nflow and outflow between rounds 3 and 3 (row %) wh and whout mssng Poverty status round 3 Poverty status round 3 Non poor Poor Mssng. Sample (mssng at round 3 excluded ) Not Poor Poor Total All ndvduals Not Poor Poor Total Results ndcate that one quarter of those who were not poor n round 3 have become poor n round 3 whle only 5 % of those who were poor n 2003 were no longer poor n n round 3, the poverty rate among those prevously poor s 60 percentage ponts hgher than the poverty rate among those non-poor n round 2. Yet, ths clearly ndcate that the poverty status n a gven perod depend substantally on past poverty status. There s nerta n the dynamc of the poverty status that suggests a state dependence effect. The second secton of table shows of the poverty transon matrx takng nto account the hgh mobly observed n the slums. Results n ths secton confrm what have been already n the frst part of the table. But what s worth hghlghtng s the dfference n the retenton status regardng the prevous poverty status. The proporton of those who have left the slums s hgher among those not poor n round 3. Ths suggests that the slums just constute a trans platform for the non-poor. Then, f ths s case, the retenton n the panel s non-random. To get consstent estmates, we should therefore specfy an equaton characterzng the retenton mechansm and jontly estmate wh the poverty transon equaton. The varables we use n our estmatons are: household characterstc (household sze, household composon, gender of the head of household, age of the head of household, occupatonal status of the head of household, number of workers whn the household), ndvduals characterstc (gender, age and age square, occupatonal status, occupatonal sector usng 4 categores) and dummy varables for housng tenure, lvng areas, and ethnc groups. All these varables are measured usng ther value n round 3, and assumed exogenous. These varables are ncluded n each of the vectors x, w and z. The retenton and nal poverty equatons nclude a number of addonal varables excluded from the poverty transon equaton for model dentfcaton. For the retenton equaton, we consder two excluded varables. The frst s a dummy varable ndcatng whether the household or one of s members have experenced a shock such theft, rape, fre, muggng or demolon n the year precedng the survey. The second excluded varable corresponds to a bnary varable whch ndcates whether

9 the ndvdual was enumerated when the DSS started n 2002 or whether she has joned latter. For the nal poverty equaton the excluded varables correspond to the shock dummy mentoned above and a bnary varable ndcatng f one household member has experenced severe llness durng the year precedng the survey. Table A provdes descrptve statstcs of selected varables. 5. Estmaton results Table 2, 3 and 4 present the results from the trvarate prob regresson. Estmates of the cross-equatons correlatons between the unobserved characterstcs provde nsghts about the endogenous selecton processes. Results n table 3 ndcate that the correlatons assocated wh the sample retenton are not sgnfcant. Ths suggests that the sample retenton mechansm could be gnored. n contrast the correlaton assocatng nal condon and poverty transon equatons cannot be frmly rejected event the correlaton s not strongly sgnfcant. The test of gnorably confrms the rejecton of the endogeney of panel retenton. The test also rejects the null hypothess that nal poverty status s exogenous for poverty transon. Further, the test for the jont sgnfcance of the three correlaton coeffcents suggests that they are jontly sgnfcant at 5 percent. Table No 2: Estmated correlaton coeffcents of unobservable and tests of gnorably Correlatons of unobservable Coeffcents Std. Errors ρ = cov, : nal poverty status, retenton (0.008) ( λ η ) ( λ δ ) ( ) ρ cov, 2 = : nal poverty status, poverty transon (0.06)* = : retenton, poverty transon (0.068) ρ 3 cov η, δ Wald tests of gnorably Ch-2 P-Value Exogeney of panel retenton : 3 ρ = ρ ρ = Exogeney of nal condon : 2 Jont exogeney : ρ = ρ2 = ρ ρ Table 3 shows the mpacts explanatory varables on poverty transon. Evdence clearly ndcates a sgnfcant and szeable state dependence effect. Beng poor n the past ncreases the chance of experencng future poverty once the heterogeney s controlled for. Ths confrms our descrptve fndngs n table. Further, there are few explanatory varables wh sgnfcant effect. ndvduals lvng n households wh many members or households wh hgh proporton of chldren are assocated wh a hgh probably of beng poor. t also appears that ndvduals lvng n ther house whn the slums are very prone to poverty. n contrast, havng a male as a

10 head of household s assocated wh a lower probably of becomng or remanng poor. t s worth hghlghtng that the covarates related to employment have no sgnfcant effects. Table No 3: Poverty transon Explanatory varables Coeffcents Std. Error nal poverty.859 (0.097) *** Household Characterstcs Housng tenure (0.026) *** Household sze (0.008) *** Number of workers (0.009) Household composon (reference: adults [35-49]) chldren [0-5} (0.085) *** ndvdual s characterstcs Head of household Characterstcs Chldren [6-2] (0.084) *** Chldren [3-7] (0.089) *** Adults [8-34] 0.0 (0.054) * Old [50-59] (0.079) Old [60 +[ (0.8) Sex (Male) (0.08) *** Age (0.002) * Age square (0.000) Employee nformal sector (0.03) Self-employee nformal sector (0.029) Other sector (0.087) Kkuyu (0.032) ** Kamba -0.7 (0.038) *** Luo (0.035) *** Luhya -0.2 (0.039) *** Ks (0.059) Somal (0.055) * Age (0.005) Age square (0.000) Sex (Male) (0.022) Occupatonal status (0.093) Korogocho (0.04) Constant *** Log-lkelhood Model ch-square (d. f. = 28) (p < 0.000) Number of observatons 57627

11 Table No 4: nal poverty and retenton estmates Household Characterstcs nal poverty Retenton Coeff. St. Error Coeff. St. Error Housng tenure -0.9 (0.08) *** (0.07) *** Household sze 0.75 (0.005) *** (0.004) *** Number of workers 0.50 (0.006) *** (0.005) Household composon (reference: adults [35-49]) chldren [0-5} (0.052) *** (0.059) *** Chldren [6-2] (0.056) *** (0.05) *** Chldren [3-7] (0.063) *** (0.059) *** Adults [8-34] (0.036) *** (0.033) *** Old [50-59] (0.059) (0.055) Old [60 +[ (0.089) *** (0.087) * ndvdual s characterstcs Sex (Male) (0.02) (0.0) *** Age (0.00) (0.00) ** Age square (0.000) (0.000) *** Employee nformal sector (0.09) *** (0.08) *** Self-employee nformal sector 0.26 (0.020) *** (0.09) * Other sector (0.068) *** 0.2 (0.062) * Kkuyu (0.023) *** (0.002) * Kamba (0.025) *** (0.024) *** Luo (0.025) *** (0.023) *** Luhya (0.026) *** (0.024) *** Ks 0.45 (0.037) *** 0.52 (0.036) *** Somal (0.04) *** (0.039) Head of household Characterstcs Age (0.003) *** 0.06 (0.003) *** Age square (0.000) *** (0.000) *** Sex (Male) (0.06) *** (0.05) *** Occupatonal status (0.07) *** 0.6 (0.066) ** Korogocho (0.04) *** (0.03) *** Shock -04 (0.06) *** (0.05) *** Severe llness (0.022) *** Mgrant (0.02) *** Constant -.26 (0.5) *** (0.07) *** Log-lkelhood Model ch-square (d. f. = 28) (p < 0.000) Number of observatons Concluson n ths paper, we have been able to offer some nsghts n the dynamcs of poverty n Narob s slums. A very nterestng result of ths paper s that there s substantal state of dependence n poverty after controllng for nal poverty status and for panel retenton. ndeed, our results show some heterogeney effects but few covarates n the poverty transon equaton are statstcally sgnfcant.

12 7. References. Arellano, M. and B. Honore (200): Panel data models: some recent developments, n: Heckman, J.J. and E. Leamer (eds.), Handbook of Econometrcs, Vol. 5, Amsterdam: Elsever Scence B.V., pp Arulampalam, W., Booth, A.L. and Taylor, M.P. (2000): Unemployment persstence, Oxford Economc Papers 52, pp Bane, M.J. and Ellwood, D.T. (986), Slppng nto and out of poverty: the dynamcs of spells, Journal of Human Resources, 2(), pp Cappellar, L. and Jenkns, S.P. (2002), Who stays poor? Who becomes poor? Evdence from the Brsh household panel survey, Economc Journal, Cappellar, L. and Jenkns, S.P. (2004), Modellng low ncome transons, Journal of Appled Econometrcs, 9: Heckman, J. (98): Statstcal models for dscrete panel data, n: Mansk, C., D. McFadden (eds.), Structural Analyss of Dscrete Data wh Econometrc Applcatons, Cambrdge, MA: MT Press, pp Lllard, L.A. and Wlls, R.J. (978), Dynamc aspects of earnngs mobly, Econometrca 46: Stevens, A.H. (999), Clmbng out of poverty, Fallng back n. (Measurng the Persstence of Poverty over multple Spells), Journal of Human Resources, 34, Stewart, M.B. and J.K. Swaffeld (999): Low Pay Dynamcs and Transon Probables, Economca 66, pp Wooldrdge (2002b): Econometrc Analyss of Cross Secton and Panel Data, Cambrdge, MA: MT-Press.