DETERMINANTS OF POVERTY IN KENYA: A HOUSEHOLD LEVEL ANALYSIS * Alemayehu Geda Institute of Social Studies, KIPPRA and Addis Ababa University

Transcription

1 DETERMINANTS OF POVERTY IN KENYA: A HOUSEHOLD LEVEL ANALYSIS * Alemayehu Geda Insttute of Socal Studes, KIPPRA and Adds Ababa Unversty Nek de Jong Insttute of Socal Studes, The Hague Germano Mwabu Unversty of Narob and KIPPRA Mwang S. Kmeny KIPPRA August 2001 * Ths paper s the outcome of collaboratve research between the Insttute of Socal Studes and the Kenya Insttute for Publc Polcy Research and Analyss (KIPPRA).

2 The Insttute of Socal Studes s Europe's longest-establshed centre of hgher educaton and research n development studes. Post-graduate teachng programmes range from sx-week dploma courses to the PhD programme. Research at ISS s fundamental n the sense of layng a scentfc bass for the formulaton of approprate development polces. The academc work of ISS s dssemnated n the form of books, journal artcles, teachng texts, monographs and workng papers. The Workng Paper seres provdes a forum for work n progress whch seeks to elct comments and generate dscusson. The seres ncludes the research of staff, PhD partcpants and vstng fellows, and outstandng research papers by graduate students. For further nformaton contact: ORPAS - Insttute of Socal Studes - P.O. Box LT The Hague - The Netherlands - FAX: E-mal: workngpapers@ss.nl ISSN Comments are welcome and should be addressed to the author:

3 ABSTRACT Strateges amed at poverty reducton need to dentfy factors that are strongly assocated wth poverty and amenable to modfcaton by polcy. Ths paper uses household level data collected n 1994 to examne probable determnants of poverty status, employng both bnomal and polychotomous logt models. The study shows that poverty status s strongly assocated wth the level of educaton, household sze and engagement n agrcultural actvty. In general, those factors that are closely assocated wth overall poverty accordng to the bnomal model are also mportant n the orderedlogt model, but they appear to be even more mportant n tacklng extreme poverty. Keywords: Poverty, Kenya, Afrca, Probablty Models JEL Classfcaton: I320 2

4 CONTENTS 1. INTRODUCTION PREVIOUS POVERTY STUDIES IN KENYA BINOMIAL AND POLYCHOTOMOUS MODELS OF POVERTY DATA ESTIMATION RESULTS Poverty Status: Natonal Sample Poverty Status: Rural and Urban Sub-Samples Ordered Poverty Status: Natonal and Urban-Rural Sub-Samples CONCLUDING REMARKS...11 REFERENCES...12 APPENDIX...14 TABLES

5

6 1. INTRODUCTION Poverty n Kenya s pervasve. Table 1 provdes a general pcture of poverty n Kenya as of Usng a per-adult equvalent measure, the headcount (P 0 ), the poverty gap (P 1 ) and severty (P 2 ) of consumpton-poverty ndces were 48, 19 and 10 per cent n The comparable fgures for 1997, the latest avalable, are 52.9, 19.3 and 9.2 per cent (Government of Kenya, 2000). The fgures reported n Table 1 are n general larger than smlar ndces for Kenya estmated by the Mnstry of Fnance and Plannng (see Government of Kenya, 1998, 2000). The table also shows that poverty s concentrated n rural areas. The pervasve nature of poverty s one of the reasons for the recent focus on poverty-allevaton polces. The Government of Kenya has prepared a poverty reducton strategy paper (PRSP) to gude the poverty reducton effort. One major weakness n the government s PRSP s lack of n-depth nformaton for mplementng and montorng the strategy (see Government of Kenya 2001, Alemayehu et al. 2001). Ths paper should help the government to realse ts poverty reducton goals, by layng the foundaton for analytcal work amed at an n-depth understandng of poverty, and by establshng benchmark condtons for poverty montorng. The remander of the paper s organsed as follows. Secton 2 revews avalable poverty studes n Kenya. Secton 3 presents the models. Secton 4 descrbes the data and Secton 5 dscusses the estmaton results. Fnally, some concludng remarks are made n Secton PREVIOUS POVERTY STUDIES IN KENYA Analytcal work on determnants of poverty n Kenya s at best scanty. Most of the avalable studes are descrptve and focus manly on measurement ssues. Earler poverty studes have focused on a dscusson of nequalty and welfare based on lmted household level data (see Bgsten 1981, Hazlewood 1981, House and Kllck 1981). One recent comprehensve study on the subject s that of Mwabu et al. (2000), whch deals wth measurement, profle and determnants of poverty. The study employs a household welfare functon, approxmated by household expendture per adult equvalent. The authors run two categores of regressons, usng overall expendtures and food expendtures as dependent varables. In each of the two cases, three equatons are estmated whch dffer by type of dependent varable. These dependent varables are: total household expendture, total household expendture gap (the dfference between the 1

7 absolute poverty lne and the actual expendture) and the square of the latter. A smlar set of dependent varables s used for food expendture, wth the explanatory varables beng dentcal n all cases. Mwabu et al. (2000) justfed ther choce of ths approach (compared to a logt/probt model) as follows. Frst, the two approaches (dscrete and contnuous choce-based regressons) yeld bascally smlar results (see below, however); second, the logt/probt model nvolves unnecessary loss of nformaton n transformng household expendture nto bnary varables. Although ther specfcaton s smple and easy to follow, t has certan nherent weaknesses. One obvous weakness s that, unlke the logt/probt model, the levels regresson does not drectly yeld a probablstc statement about poverty. Second, the major assumpton of the welfare functon approach s that consumpton expendtures are negatvely assocated wth absolute poverty at all expendture levels. Thus, factors that ncrease consumpton expendture reduce poverty. However, ths basc assumpton needs to be taken cautously. For nstance, though ncreasng welfare, rasng the level of consumpton expendture of households that are already above the poverty lne does not affect the poverty level (as for example measured by the headcount rato). Notwthstandng such weaknesses, the approach s wdely used and the Mwabu et al. (2000) study dentfed the followng as mportant determnants of poverty: unobserved regon-specfc factors, mean age, sze of household, place of resdence (rural versus urban), level of schoolng, lvestock holdng and santary condtons. The mportance of these varables does not change whether the total expendture, the expendture gap or the square of the gap s taken as the dependent varable. The only notceable change s that the szes of the estmated coeffcents are enormously reduced n the expendture gap and n the square of the expendture gap specfcatons. Moreover, except for mnor changes n the relatve mportance of some of the varables, the pattern of coeffcents agan fundamentally remans unchanged when the regressons are run wth food expendture as dependent varable. Another recent study on the determnants of poverty n Kenya s Oyug (2000), whch s an extenson to earler work by Greer and Thorbecke (1986a,b). The latter study used household calore consumpton as the dependent varable and a lmted number of household characterstcs as explanatory varables. Oyug (2000) uses both dscrete and contnuous ndcators of poverty as dependent varables and employs a much larger set of household characterstcs as explanatory varables. An mportant as- 2

8 pect of Oyug s study s that t analyses poverty both at mcro (household) and meso (dstrct) level, wth the meso level analyss beng the nnovatve component of the study. Oyug (2000) estmates a probt model usng data of the 1994 Welfare Montorng Survey data. The explanatory varables (household characterstcs) nclude: holdng area, lvestock unt, the proporton of household members able to read and wrte, household sze, sector of economc actvty (agrculture, manufacturng/ndustral sector or wholesale/retal trade), source of water for household use, and off-farm employment. The results of the probt analyss show that almost all varables used are mportant determnants of poverty n rural areas and at the natonal level, but that there are mportant exceptons for urban areas (Oyug, 2000). These results are consstent wth those obtaned from the meso-level regresson analyss. It s nterestng to compare the mplcatons of the levels (Mwabu et al. 2000) and probt (Oyug, 2000) regresson approaches. From the levels regressons, age, household sze, resdence, readng and wrtng and level of schoolng are the top fve mportant determnants of poverty at the natonal level. In the probt model, however, n order of mportance the key determnants of poverty are: beng able to read and wrte, employment n off-farm actvtes, beng engaged n agrculture, havng a sde-busness n the servce sector, source of water and household sze. Regon of resdence appears to be equally mportant n determnng poverty status n the two approaches. Although the two approaches dd not employ the same explanatory varables, ths comparson ponts to the possblty of arrvng at dfferent polcy conclusons from the two approaches. 3. BINOMIAL AND POLYCHOTOMOUS MODELS OF POVERTY The approach we follow ntends to explan why some populaton groups are non-poor, poor, or extremely poor. We dentfy dfferent populaton sub-groups n several stages. In the frst stage, we dentfy the poor and non-poor. In the second stage, we examne the probablty of beng n hard-core poverty condtonal on beng dentfed as poor. That s, we also compute the probablty of beng what we term as extremely poor. Ths poverty dentfcaton process s dsplayed n Fgure 1. 3

9 FIGURE 1 A NESTED STRUCTURE OF POVERTY STATUS Total Sample Non-poor Poor Non hard-core poor (Moderately poor) Hard-core poor (Extremely poor) We assumed that the probablty of beng n a partcular poverty category s determned by an underlyng response varable that captures the true economc status of an ndvdual. In the case of a bnary poverty status (.e. beng poor or non-poor), let the underlyng response varable y* be defned by the regresson relatonshp: y * = x ' β + u (1) ' where = β, β... β ] and x ' = [1, x, x... x ] β [ 1 2 k 2 3 k. In equaton (1), y* s not observable, as t s a latent varable. What s observable s an event represented by a dummy varable y defned by: y =1 f y* > 0, and y =0 otherwse (2) From equatons (1) and (2) we can derve the followng expresson: Pr ob ( y = 1) = Pr ob( u > x' β ) = 1 F( x ' β ) (3) where F s the cumulatve dstrbuton functon for u, and Pr ob( y 0 β, x ) = F( x ' β ). = 4

10 The observed values of y are the realsaton of the bnomal wth probabltes gven by equaton (3), whch vares wth X. Thus, the lkelhood functon can be gven by: L = [ F( x ' )] [ 1 ( ' ) ] β F x β y= 0 y= 1 (4a) whch can be wrtten as: L = y = 1 1 [ ( )] y F x ' β [ 1 F( x ' β )] y (4b) The functonal form mposed on F n equaton (4) 1 depends on the assumptons made about u n equaton (1). 2 The cumulatve normal and logstc dstrbutons are very close to each other. Thus, usng one or the other wll bascally lead to the same result (Maddala 1983). Moreover, followng Amemya (1981), t s possble to derve the would-be estmates of a probt model once we have parameters derved from the logt model. Thus, the logt model s used n ths study. We have specfed the logt model for ths study by assumng a logstc cumulatve dstrbuton of u n F (n equatons (4a) and (4b)). The relevant logstc expressons are: 1 F ( x ' β ) = e 1+ e x ' x β ' β (5a) F ( x ' β ) e = 1+ e x x ' β ' β = 1+ e 1 x ' β (5b) As before, X are the characterstcs of the households/ndvduals, and β the coeffcents for the respectve varables n the logt regresson. Havng estmated equaton (4) wth maxmum lkelhood (ML) technque, equaton (5a) bascally gves us the probablty of beng poor (Prob(y =1)) and equatons (5b) the probablty of beng non-poor (Prob(y =0)). 1 The log lkelhood functon for expresson [4a] and [4b] can be wrtten as: n l( β ) = log L( β ) = y log( 1 F( x ' β) ) + (1 y )log F( x ' β) = 0 2 Ths bascally forms the dstncton between logt and probt (normt) models. 5

11 After modelng the process that generates the poor or non-poor status, we focus attenton on the hard-core poor versus the moderately poor and non-poor. Ths can be handled by a polychotomous model, more n partcular an ordered probt or logt model. Ths approach s justfable, because we explctly make the orderng of the populaton sub-samples, usng total and food poverty lnes as cut-off ponts n a cumulatve dstrbuton of expendture. 3 Snce these categores have a natural order, the ordered logt s the approprate model to be employed n the estmaton of relevant probabltes (see Maddala 1983, Amemya 1985, Greene 1993). 4 Assumng three categores (1, 2 and 3 and assocated probabltes P 1, P 2 and P 3 ), an ndvdual would fall n category 3 f u < β x, n category 2 f β x < u β x + α; and n category 1 f u β x + α, where α > 0 and u s the error term n the underlnng response model (see Equaton 1). These relatonshps may be gven by: P = F( â'x ) P = F( â'x + α) F( â'x ) P = 1 F( â'x + α) (6) where the dstrbuton F s logstc n the ordered logt model. Ths can easly be generalsed for m categores (see Maddala 1983). Assumng the underlyng response model s gven by: y = â' x + u (7) we can defne a set of ordnal varables as: Z j =1 f y falls n the j th category Z j =0 otherwse (=1,2,..,n; j=1,2,,m) Pr ob( Z j = 1) = Φ( α j β ' x ) Φ( α j 1 β ' x ) (8) 3 The method used for computng the poverty lnes s gven n the Appendx. For lack of a better term we have used the term moderately poor to desgnate those who are poor but not hard-core (or extremely) poor. 4 Gven the nested nature of the categores n our model, nested model seems also a relevant approach. However, such models are relevant n the context when agents make choces and there s dependence among choces. Snce our categores do not refer to choces beng made, we have opted for the ordered logt model [see Maddala, 1983: 70]. 6

12 where Φ s the cumulatve logstc dstrbuton and the α j s are the equvalents of the α s n equaton (6). The lkelhood and log-lkelhood functons for the model can be gven by equatons (9) and (10) respectvely, as: L n m Z [ ] j Φ ( α j β' x Φ ) ( α j 1 β' x ) = 1 j= 1 = (9) L * [( α β ' x ) Φ( α β ' x )] n k = j j j = 1 j= 1 log L = Z log Φ 1 (10) Equaton (10) can be maxmsed n the usual way, and can be solved teratvely by numercal methods, to yeld maxmum lkelhood estmates of the model (see Maddala 1983). 4. DATA The data used are based on the 1994 Welfare Montorng Survey (Government of Kenya 1998, 2000). These data were collected for the whole country and covered nearly ten thousand households, comprsng about sxty thousand ndvduals (see Mwabu et al., 2000). The fundamental ratonale behnd the choce of a household as a unt of analyss s the assumpton of sharng of resources among households. Although the qualty of the data we use s n general relatvely hgh, two factors need to be borne n mnd n usng the results derved from them. Frst, the results mght be affected by the seasonal effect on household expendture, snce seasonalty was not controlled for whle collectng the data. Second, some dstrcts, especally those from Northeastern provnce, are underrepresented n the sample. We used a comprehensve lst of explanatory varables whch may be grouped nto the followng categores: property-related, such as land and lvestock holdng; household characterstcs, such as status of employment, age, gender, educatonal level, household sze; and others, such as tme spent to fetch water and to obtan energy, place of resdence of the household whether n rural or urban or n a partcular provnce (see Table 2). The estmaton was made after nflatng the number of households n the sample (about 10,000) to that n the total populaton (nearly 26 mllon n 1994), usng expanson factors. The expanson factors are however adjusted downwards for chldren n 7

13 case of adult equvalent-based estmatons. The household characterstcs are assumed to affect (adult-equvalent) members of the household equally ESTIMATION RESULTS 5.1 Poverty Status: Natonal Sample Accordng to the estmaton results, male-headed households are less lkely to be poor. Smlarly, the lkelhood of beng poor s smaller n urban areas than n rural areas. Probably to some extent related to ths, people lvng n households manly engaged n agrcultural actvtes are more lkely to be poor, compared to households n manufacturng actvtes. In all models the most mportant determnant of poverty status s the level of educaton. The effects of ths varable are smlar across the four models. The coeffcent for household sze s almost twce as hgh n the consumpton-based as ncome-based models ones, whle the mpacts of the sector of employment, as well as the number of anmals owned s nsgnfcant n the consumpton-based models. Total holdng of land does not seem to be mportant n any of the specfcatons. An explanaton for ths may le on the mportance of the qualty of land and/or lack of complementary agrcultural nputs (see Alemayehu et al. 2001). Table 3 shows the estmated model and the margnal effects of each explanatory varable on the probablty of beng poor, based on models n whch per adult equvalent consumpton s used to estmate poverty. Estmaton results usng per capta ncome and consumpton are reported n Alemayehu et al. (2001). 5.2 Poverty Status: Rural and Urban Sub-Samples Followng the fndng that place of resdence s assocated wth level of poverty, we have ftted the model to data for rural and urban areas separately. The estmaton results and the margnal effects are gven n Table 4. Agan the detaled results are gven n Alemayehu et al. (2001). In general, the results show that the factors strongly assocated wth poverty (level of educaton, household sze, engagement n agrcultural actvtes) are the same n both rural and urban areas. However, the sze of the coeffcents assocated wth these regressors s larger n rural areas. Moreover, polygamous marrage seems to worsen poverty n urban as opposed to rural areas. Ths may pont at 5 To save space, we have reported only those results derved from estmates based on poverty defned on the bass of consumpton per adult-equvalent. The nterested reader s referred to Alemayehu et al. (2001) for per capta and ncome-based estmates and related detals. 8

14 the larger mportance of labour nput n rural rather than n urban economc actvtes. In rural areas all the members of the extended household do often work n agrculture, whle n urban areas there may be less scope for all the members of the extended household to be meanngfully engaged. Ths result does not seem to hold n the consumpton-based estmaton, however. Gven the relablty problem wth ncome data and the fact that even the consumpton based estmates are not statstcally sgnfcant at conventonal levels, ths result may be taken as nconclusve. The consumpton-based estmaton yeld farly smlar results about determnants of poverty, partcularly wth regard to educatonal attanment. The coeffcents obtaned n the latter model are relatvely smaller, however. Moreover, factors such as age, sze of land holdng (albet wth very small coeffcents) are found to be statstcally sgnfcant n ths verson of the model. Regonal dummes for Western and Eastern provnces that are vrtually nsgnfcant n the ncome-based model are found to be statstcally sgnfcant n the consumpton-based verson of the model for rural areas. Moreover, workng n the urban modern sector seems to reduce the lkelhood of beng poor. 5.3 Ordered Poverty Status: Natonal and Urban-Rural Sub-Samples Followng the dscusson n Secton 3, we have ordered the sample nto three mutually exclusve categores: non-poor (category 1), moderately poor (category 2) and hard-core or extremely poor (category 3), wth households n category 3 beng most affected by poverty. Ths classfcaton s based on the poverty and food poverty lnes computed from the 1994 Welfare Montorng Survey (see Appendx). The estmated model and the margnal effects of the regressors for the consumpton-based models are gven n Table 5. We noted that the consumpton-based model s farly dfferent from the ncome-based model. It exhbts regressors wth statstcally sgnfcant coeffcents as well as weaker explanatory effects n the case of category 1 (non-poor) and category 2 (poor), respectvely (see Alemayehu et al for detals). 6 In general, t s nterestng to note that those factors that are mportant n the bnomal model are stll mportant n the ordered-logt model. More mportantly, by com- 6 The margnal coeffcents for category 3 (hard-core poor) are not reported as they could be derved from the sum of the three, whch should add to zero. Ths s because the probabltes of fallng n ether one of the three categores adds up to one. 9

15 parng the margnal effects for categores 2 and 3, we note that these varables are much more mportant n tacklng hard-core poverty than moderate poverty. The ordered logt model s estmated for rural and urban sub-samples too (not reported here, but avalable on request). Bascally the results are smlar to those obtaned for the natonal sample. However, the followng nterestng dfferences are observed. Frst, although secondary and unversty level educaton are mportant both n rural and urban areas, prmary educaton s found to be extremely mportant n rural areas. Second, agrculture as man occupaton s more closely assocated wth poverty n urban areas than n rural areas. Ths ndcates that agrculture beng the man occupaton s a factor that more strongly dfferentates between beng poor or non-poor n urban areas. Thrd, the negatve mpact of agng s stronger n urban than rural areas. Ths may reflect the collapse of the extended famly network n urban areas, whch normally serves as a tradtonal nsurance scheme n Afrca. Fnally, urban poverty s worst n Western and Northeastern provnces (see Alemayehu et al. 2001). The ordered-logt estmaton of ncome-based models shows that at the natonal level the predcted probablty of fallng n the non-poor category and nto moderately and extremely poor categores are 42, 13 and 45 percent, respectvely. The correspondng fgures for rural areas are smlar, whle for urban areas they are 58, 19 and 23 percent respectvely. Ths bascally shows that for a poor Kenyan resdng n rural areas the probablty of fallng n extreme poverty s much greater than for hs/her urban counterpart. A smlar pattern s observed when the ordered logt model s estmated usng consumpton-based data. However, the probablty for the frst category n general declnes whle that for the thrd category rses. Ths nformaton s summarsed n Table 6. The detals are gven n Alemayehu et al. (2001). The ordered-logt model results show clearly that determnants of poverty have dfferent mpacts across the poverty categores defned. For nstance, f we take the most mportant determnant of poverty status n Kenya,.e. the level of educaton, Table 4 shows that the margnal effect of havng a prmary level of educaton are 0.10, and for non-poor, moderately poor and hard-core poor categores, respectvely. The comparable margnal effects for secondary level educaton are 0.25, and -0.16; and for unversty level educaton 0.36, and -0.22, respectvely. Ths shows that, n general, educaton s more mportant for the hard-core poor than for the moderately poor. The relatve dfference s largest n the case of prmary educaton. 10

16 6. CONCLUDING REMARKS In ths paper an attempt has been made to explore the determnants of poverty n Kenya. We have employed both bnomal and polychotomous logt models usng the 1994 Welfare Montorng Survey data. Although a number of specfc polcy conclusons could be drawn from the estmaton results, the followng polcy mplcatons of the study stand out: Frst, as expected, we have found that poverty s concentrated n rural areas n general, and n the agrcultural sector n partcular. Beng employed n the agrcultural sector accounts for a good part of the probablty of beng poor. Thus, nvestng n the agrcultural sector to reduce poverty should be a matter of great prorty. Moreover, the fndng that the sze of land holdng s not a determnant of poverty status may suggest the mportance n poverty reducton not only of mprovng the qualty of land, but also of provdng complementary nputs that may enhance productvty. Second, the educatonal attanment of the head of the household (n partcular hgh school and unversty educaton) s found to be the most mportant factor that s assocated wth poverty. Lack of educaton s a factor that accounts for a hgher probablty of beng poor. Thus, promoton of educaton s central n addressng problems of moderate and extreme poverty. Specfcally, prmary educaton s found to be of paramount mportance n reducng extreme poverty n, partcularly, rural areas. Thrd, and related to the second pont above, the mportance of female educaton n poverty reducton should be noted. We have found that female-headed households are more lkely to be poor than households of whch the head s a men and that female educaton plays a key role n reducng poverty. Thus, promotng female educaton should be an mportant element of poverty reducton polces. Because there s evdence that female educaton and fertlty are negatvely correlated, such a polcy could also have an mpact on household sze, whch s another mportant determnant of poverty n Kenya. Moreover, gven the mportance of female labour n rural Kenya and elsewhere n Afrca, nvestng n female educaton should be productvty enhancng. Fnally, n lne wth the three strateges that are outlned n the PRSP and drectly related to ssues of poverty (economc growth and macro stablty, rasng ncome opportunty of the poor, and mprovng qualty of lfe), the fndngs n ths study pont to the mportance of focusng on educaton n general and prmary educaton n rural areas n partcular. The study also hghlghts the hgher lkelhood of beng poor of those who are engaged n the agrcultural sector. Thus, the PRSP s strategy of rasng 11

17 ncome opportuntes of the poor should focus on nvestng n agrculture. Snce the macroeconomc envronment s mportant n determnng the productvty of such nvestment, macroeconomc and poltcal stablty are a pre-requste for addressng poverty. REFERENCES Alemayehu, Geda, N. de Jong, M. Kmeny and G. Mwabu, Determnants of Poverty n Kenya: Household Level Analyss, KIPPRA Dscusson Paper (July 2001). Amemya, Takesh, 1981, Qualtatve Response Model: A Survey, Journal of Economc Lterature, Vol. 19, No. 4. Amemya, Takesh, 1985, Advanced Econometrcs, Cambrdge: Cambrdge Unversty Press. Bgsten, Arne, 1981, Regonal Inequalty n Kenya n Tony Kllck. Papers on the Kenyan Economy: Performance, Problems and Polces. Narob: Henemann Educatonal Books Ltd. Government of Kenya, 1998, Frst Report on Poverty n Kenya: Incdence and Depth of Poverty, Volume I. Narob: Mnstry of Fnance and Plannng. Government of Kenya, 2000, Second Report on Poverty n Kenya: Incdence and Depth of Poverty, Volume I. Narob: Mnstry of Fnance and Plannng. Government of Kenya, 2001, Poverty Reducton Strategy Paper (PRSP) for the Perod Narob: The Government Prnter. Greer, J. and E. Thorbecke, 1986a, A Methodology for Measurng Food Poverty Appled to Kenya, Journal Development Economcs, Vol. 24, No. 1. Greer, J. and E. Thorbecke, 1986b, Food Poverty Profle Appled to Kenyan Smallholders, Economc Development and Cultural Change, Vol. 35, No. 1. Hazlewood, Arthur, 1981, Income Dstrbuton and Poverty: An Unfashonable Vew n Tony Kllck. Papers on the Kenyan Economy: Performance, Problems and Polces. Narob: Henemann Educatonal Books Ltd. House, Wllam and Tony Kllck, 1981, Inequalty and Poverty n Rural Economy and the Influence of Some Aspects of Polcy n Tony Kllck. Papers on the Kenyan Economy: Performance, Problems and Polces. Narob: Henemann Educatonal Books Ltd. Maddala, G.S., 1983, Lmted Dependent and Qualtatve Varables n Econometrcs. Cambrdge: Cambrdge Unversty Press. 12

18 Manda, Kulundu, Mwang Kmeny and Germano Mwabu, 2001, A Revew of Poverty and Antpoverty Intatves n Kenya, KIPPRA Workng Paper (Forthcomng). Mukherjee, C., H. Whte and M. Wuyts, 1998, Econometrcs and Data Analyss for Developng Countres. London: Routledge. Mwabu, Germano, Wafula Masa, Rachel Gesam, Jane Krm, Godfrey Ndeng e, Tabtha Krt, Francs Munene, Margaret Chemngch and Jane Marara, 2000, Poverty n Kenya: Profle and Determnants, Narob: Unversty of Narob and Mnstry of Fnance and Plannng. Oyug, Lneth Nyaboke, 2000, The Determnants of Poverty n Kenya (Unpublshed MA Thess, Department of Economcs, Unversty of Narob). 13

19 APPENDIX Computatons of poverty lnes and ndces There are a number of studes on the condton of poverty n Kenya, the most mportant of whch beng the seres of reports publshed by the Mnstry of Fnance and Plannng. In ths paper, we have attempted to follow the method of poverty lne determnaton used by the Mnstry of Fnance and Plannng. Ths s amed at allowng for comparson wth the results of those studes. The frst step we took s to value the monthly food consumpton requred to satsfy the 2250 calores that defnes the bologcal mnmum requred per adult per day. Ths food poverty lne s computed by the Mnstry of Fnance and Plannng for 1994 to be Kshs for urban areas and Kshs for rural areas per adult per month. If, for llustraton purposes, we take the urban areas, the procedure we adopted s as follows. Frst we ranked the households accordng to per adult-equvalent expendture on food and dentfed the household that approxmately spent Kshs per adult equvalent on food tems. Then we computed non-food consumpton per adult equvalent, by takng the mean non-food consumpton per adult equvalent of those households n the neghbourhood of ths partcular household (.e. households wth food per adult-equvalent food expendture n a band of +10% and 20% of the food poverty lne). Addng ths mean non-food consumpton, Kshs , to the Kshs gves the poverty lne per adult equvalent of Kshs per adult per month. A smlar procedure s followed to compute the per capta poverty lne. We have used the same Kshs for urban and Kshs for rural food requrement per month per person as the startng pont. 7 Takng the same range of households as ndcated above, we computed per capta non-food consumpton (Kshs and for urban and rural areas, respectvely). Addng these mean non-food consumpton levels to the Kshs and Kshs gves the per capta poverty lne of Kshs and per month for urban and rural areas, respectvely (see Table A.1 for detals). 7 Notce the assumpton of usng adult-equvalent requrements for each person n the household. Ths mght be a lmtng assumpton but s often made due to lack of an alternatve. 14

20 TABLES TABLE A.1 POVERTY LINES ADJUSTED FOR PRICE CHANGES (IN KSHS. PER MONTH) Per capta Urban Rural Per adult equvalent Urban Rural Deflators used (1986=100)* * CPI of December for 1992 and that of June for 1994 and 1997 TABLE 1 POVERTY IN 1994 (ESTIMATES BY GOVERNMENT OF KENYA IN BRACKETS) Rural Urban Natonal Consumpton Based Income Based Consumpton Based Income Based Consumpton Based Income Based Per capta ncome or consumpton-based measures General poverty Headcount rato 0.64 [0.42] [0.29] [0.40] 0.68 Poverty gap Poverty severty Extreme poverty Headcount rato 0.52 [0.25] [0.10] [0.22] 0.56 Poverty gap Poverty severty Per adult equvalent ncome or consumpton-based measures General poverty Headcount rato 0.50 [0.42] [0.28] [0.44]* 0.58 Poverty gap 0.20 [0.15] [0.09] [0.14] 0.28 Poverty severty 0.10 [0.08] [0.04] [0.07] 0.18 Extreme poverty Headcount rato 0.36 [0.25] [0.10] [0.22] 0.45 Poverty gap 0.13 [0.08] [0.02] [0.07] 0.21 Poverty severty 0.06 [0.04] [0.01] [0.03] 0.13 Source: Authors calculatons based on Welfare Montorng Survey 1994 (see Appendx for the method used) * The 0.40 fgure n the 1998 Government of Kenya report s adjusted to 0.44 n the 2000 verson. 15

21 TABLE 2 DEFINITION OF VARIABLES USED IN THE ESTIMATED EQUATIONS Varables Dependent varable Poverty Defnton P=1 f poor, 0 otherwse Poverty estmate based on consumpton per adult equvalent Symbol n the Estmated Equaton P0_CPAE n bnomal logt model; PM_CPAE n ordered logt model Mean Std dev. Explanatory varables Sex Sex = 1 f male, 0 female SEXD Age and Age square years AGE & AGE Member can read and wrte Martal Status =1 f marred & Monogamy, 0 otherwse =1 f marred & polygamy, 0 otherwse = 1 f yes and 0 otherwse CANREWTE MARYMONO MARYPOLY Employment Sector =1 f formal/publc and 0 otherwse EMPSECD Man occupaton of =1 f n Agrculture (Commercal OCCp member farmer, subsstence farmer and Hghest level attaned (three categores: Prmary, Secondary and Unversty) pastoralsts), 0 otherwse =1 f n Prmary (Standard 1-8 and KCPE) and 0 Otherwse. =1 f n Secondary and certfcate (Form 1-4, KCE/KCSE/KAC, Trade test cert I-III and Other Post Secondary cert) and 0 otherwse =1 f n Unversty degree and 0 otherwse PRIMARD SECONDD UNIVDD Area of Resdence = 1 f n Rural and 0 otherwse URBRUR Total holdng of land n acres TOHOLNOW Number of anmals owned lvestock unts ANIMANOW Provncal Dummes: COAST for Coast Provnce; RIFTV for Rft Valley; WESTERN for Western; EASTERN for Eastern; NEAST for North Eastern, NYANZA for Nyanza and CENTRAL for Central provnce

22 TABLE 3 BINOMIAL LOGIT ESTIMATES FOR CONSUMPTION PER ADULT EQUIVALENT MODEL: NATIONAL SAMPLE Varables Estmated Coeffcents Margnal Effects β' s Z-values Dy/dx Z-values SEXD* MARYMONO* MARYPOLY* OCCPD* * EMPSECD* PRIMARD* * * SECONDD* * * UNIVDD* * * HHSIZE * * ANIMANOW TOHOLNOW * * URBRUR AGE * * AGE ** ** COAST* RIFTV* WESTERN* EASTERN* NEAST* ^ ^ NYANZA* CENTRAL* Constant * Rato of Predcted to actual: 61%; Log Lkelhood= (*) dy/dx s for dscrete change of dummy varable from 0 to 1 *, **, ^ sgnfcant at 1, 5 and 10 per cent level. 17

23 TABLE 4 BINOMIAL LOGIT ESTIMATES FOR CONSUMPTION PER ADULT EQUIVALENT MODEL BY REGION Rural Urban Estmated Coeffcents Margnal Effects Estmated Coeffcents Margnal Effects Varable β Z-values dy/dx Z-values β Z-values dy/dx Z-values SEXD* ^ ** MARYMONO* MARYPOLY* * OCCPD* * * * * EMPSECD* ^ ** PRIMARD* * * SECONDD* * * * * UNIVDD* * * * * HHSIZE * * * * ANIMANOW * ** TOHOLNOW ** ** AGE * * AGE ^ * COAST* RIFTV* WESTERN* * * EASTERN* * NEAST * * NYANZA* CENTRAL* Constant * * (*) dy/dx s for dscrete change of dummy varable *, **, ^ sgnfcant at 1, 5 and 10 per cent level Rural: Number of observatons 9063, Log lkelhood Urban: Number of observatons 1645; Log lkelhood

24 TABLE 5 ORDERED LOGIT ESTIMATES USING CONSUMPTION PER ADULT EQUIVALENT: NATIONAL SAMPLE The Model Probablty of beng Non-poor Probablty of beng Moderately Poor Estmated Coeffcents Margnal Effects Margnal Effects Varable β Z-values dy/dx Z-values dy/dx Z-values SEXD* MARYMONO* MARYPOLY* OCCPD* * * * EMPSECD* PRIMARD* * * * SECONDD* * * * UNIVDD* * * * HHSIZE * * * ANIMANOW TOHOLNOW * * * URBRU ** ** ** AGE * * * AGE * * * COAST* RIFTV* WESTERN* EASTERN* NEAST ^ ** NYANZA* CENTRAL* _CUT _CUT No. of Observatons Log Lkelhood= Pm_cpae= 1 Pr( xb+u<_cut1) Pr(_cut1<xb+u<_cut2) Pr(_cut2<xb+u) 0.33 (*) dy/dx s for dscrete change of dummy varable from 0 to 1 *, **, ^ sgnfcant at 1, 5 and 10 per cent level 19

25 TABLE 6 PREDICTED PROBABILITIES OF BEING NON-POOR, MODERATELY POOR OR EXTREMELY POOR* Income-based Model Consumpton-based Model Sample Probablty of beng Probablty of beng Non-Poor Poor Extremely Poor Non-Poor Poor Extremely Poor Natonal Rural Urban * Fgures may not add to 1 due to roundng up (see Alemayehu et al. 2001). 20