Social Assistance and Poverty Reduction in Moldova,

Transcription

1 Publc Dsclosure Authorzed Pol c y Re s e a rc h Wo r k n g Pa p e r 4658 WPS4658 Publc Dsclosure Authorzed Publc Dsclosure Authorzed Socal Assstance and Poverty Reducton n Moldova, An Impact Evaluaton Paolo Verme Publc Dsclosure Authorzed The World Bank Europe and Central Asa Regon Poverty Reducton and Economc Management Sector June 2008

2 Polcy Research Workng Paper 4658 Abstract Ths paper assesses the mpact of socal assstance benefts on household welfare n Moldova. Ignorng standard ssues of mpact evaluatons such as selecton bas, behavoral responses, unobserved heterogenety and endogenety, an ncdence analyss suggests that ncreased spendng on socal assstance enhances the probablty of movng out of poverty and reduces the probablty of movng nto poverty. However, double dfference estmates (based on a mmcked randomzed experment) and parametrc estmates (based on panel data) ndcate that socal benefts have not contrbuted to mprove household welfare or reduce poverty. Double dfference estmates pont to a negatve mpact on welfare. Parametrc estmates do not yeld any consstent sgnfcant mpact on welfare or poverty. The author concludes that the growth n populaton coverage and expendture on cash benefts that characterzed socal assstance polces n recent years has not resulted n a sgnfcant mprovement n welfare, all other factors beng equal. Ths paper a product of the Poverty Reducton and Economc Management Sector, Europe and Central Asa Regon s part of a larger effort n the department to understand the economc transton n former centrally planned economes. Polcy Research Workng Papers are also posted on the Web at The author may be contacted at paolo.verme@unto.t.@worldbank.org. The Polcy Research Workng Paper Seres dssemnates the fndngs of work n progress to encourage the exchange of deas about development ssues. An objectve of the seres s to get the fndngs out quckly, even f the presentatons are less than fully polshed. The papers carry the names of the authors and should be cted accordngly. The fndngs, nterpretatons, and conclusons expressed n ths paper are entrely those of the authors. They do not necessarly represent the vews of the Internatonal Bank for Reconstructon and Development/World Bank and ts afflated organzatons, or those of the Executve Drectors of the World Bank or the governments they represent. Produced by the Research Support Team

3 Socal Assstance and Poverty Reducton n Moldova, An Impact Evaluaton Paolo Verme 1 JEL: H53; I32; I38; P35 Keywords: Socal assstance, Poverty, Evaluaton, Moldova 1 Department of Economcs "S. Cognett de Marts", Unversty of Torno and School of Management SDA-Boccon, Mlan. The work developed out of my contrbuton to the World Bank Report No MD "Improvng Publc Expendture Effcency for Growth and Poverty Reducton: A Publc Expendture Revew for the Republc of Moldova", 2007.

4 1. Introducton The frst decade of the transton to a market economy has been very dffcult for the countres of the Former Sovet Unon (FSU). All the 15 Republcs that consttuted the Unon experenced a deep recesson between 1990 and 1995 wth an average fall n GDP of about 40%. As they were startng to recover n 1996 and 1997, Russa defaulted on ts debt and unleashed a fnancal crss that reached all FSU republcs wth a subsequent new fall n GDP of several percentage ponts. Hdden and open unemployment and poverty ncreased consstently durng the decade leadng to a severe declne n lvng standards (WB 2005). Understandably, the populatons of these countres have lved the 1990s as a panful experence that overshadowed the ntal enthusasm for reforms. Ths s also the story of Moldova, a small FSU republc that suffered severe economc consequences followng the desegregaton of the Sovet Unon, transtonal reforms and the 1998 Russan fnancal crss. By 1999, Moldova had lost almost 60% of ts GDP 2 and had an estmated poverty rate (headcount ndex) of 71% (WB 2004). As for other countres n the regon, the economc stuaton started to mprove after the begnnng of the new mllennum. Between 2001 and 2004, GDP grew, on average, by 7% a year (GoM 2005) whle t s estmated that poverty (headcount ndex) declned from 54.6% to 26.5% (GoM 2004). The begnnng of the new growth perod also concded wth the electon of a new government n 2001 whch ncreased publc spendng on socal assstance very sgnfcantly durng ts frst mandate ( ). Can poverty reducton be attrbuted to specfc pro-poor government polces or s more generally the product of economc growth? Ths paper addresses ths queston by focusng on one polcy nstrument whch s specfcally desgned to support the poor and the vulnerable: socal assstance benefts n cash (henceforth 'socal benefts'). We carry out a welfare mpact evaluaton of socal benefts usng household panel data that cover the perod The purpose s to try to solate from other effects the mpact of socal benefts (and ther ncrease) on the welfare of Moldovan households and assess n ths way the merts of socal assstance polces. The paper s organzed as follows. Secton 2 provdes a bref revew of the lterature on poverty and socal assstance n transton. Secton 3 looks more n detal at the Moldovan case. Secton 4 dscusses some of the key ssues we are confronted wth n makng a retrospectve mpact evaluaton. Secton 5 outlnes the evaluaton strategy we opted for. Secton 6 llustrates data and varables used and secton 7 dscusses the results. Secton 8 concludes. 2. Poverty and socal assstance n transton The socal protecton system n the FSU countres went through mportant changes durng the 1990s. The Sovet system dd not really make a dstncton between socal nsurance and socal 2 See 2

5 assstance. As all adults were expected to work and provded wth the same means, socal protecton was desgned around the needs of the workforce or confned to a set of 'pensons' granted to ctzens who were not able to work such as the dsabled. The transtonal recesson and the growth n unemployment and poverty changed ths scenaro drastcally and governments found themselves unequpped to deal wth these new challenges. Wth the support of nternatonal nsttutons, FSU countres embarked on a program of socal protecton reforms. The dstncton between socal nsurance and socal assstance became more mportant as the penson funds were made extra-budgetary and the nformal sectors were expandng quckly leavng large sectons of the populatons unprotected. It became also necessary to shft from a categorcal type of cash benefts to a means-tested system n order to focus attenton on the real poor. In many FSU republcs, socal protecton reforms are stll largely ncomplete and the evaluaton of such reforms s n an early stage. In partcular, reforms have largely focused on socal nsurance relegatng socal assstance to a backstage role whch explans the scarce attenton of researchers to such nstruments. Ths s also explaned by the fact that the household budget surveys whch are necessary to evaluate publc transfers are stll few and dscontnuous. There are, however, some studes that evaluated publc transfers n transtonal economes combnng socal nsurance and socal assstance transfers. Mlanovc (2000) looked at socal protecton transfers n Latva usng one cross-secton survey. He fnds a weak pro-poor role of socal protecton benefts but does not dstngush between socal nsurance and socal assstance benefts and provdes an ncdence rather than an mpact evaluaton. Lokshn and Ravallon (2000) analyzed the role of the socal safety net n protectng the poor from the 1998 Russan fnancal crss. They use two spells of the Russan Longtudnal Montorng Survey (RLMS) and a rather broad defnton of government transfers that ncludes socal assstance and socal nsurance. They conclude that the socal safety net n place was largely nsuffcent to protect the poor from the Russan crss. Welfare mpact evaluatons of socal assstance benefts n other countres n transton outsde the FSU are also scarce. Ravallon et al. (1995) looked at the early years of the transton n Hungary and focused on the role of the socal safety net n protectng people from poverty and promotng ther escape from poverty. Combnng socal nsurance and socal assstance benefts, they found that the safety net was able to protect effectvely from poverty but dd not play an mportant role n lftng people out of poverty. Van de Walle (2003) followed n the steps of ths last paper to test the publc safety net n Vetnam and found a very margnal role of the socal safety net n protectng people from poverty or promotng an ext from poverty. Okrasa (1999a and 1999b) has used a four years panel survey from Poland to assess the mpact on poverty of the publc safety net usng a survval analyss. The transfers consdered are famly allowances and unemployment benefts and the author fnds a general postve mpact on redstrbuton, a postve but moderate mpact on reducng the poverty spell and a postve mpact on extng poverty wth all 3

6 these effects beng dfferent dependng on the household prototype consdered. The present study may be the frst to focus exclusvely on the welfare mpact of socal assstance benefts n a transtonal economy. The dstncton between socal assstance and socal nsurance s not a trval one for welfare evaluatons. Socal nsurance relates to the world of work and s desgned to protect past and present workers from rsk (old age, dsablty, work njury, unemployment, maternty). Socal nsurance benefts are granted based on contrbutons and rrespectve of welfare status. By contrast, socal assstance benefts are specfcally desgned to assst the poor and the vulnerable rrespectve of whether they have contrbuted to socal nsurance schemes or not. Only socal assstance benefts are specfcally desgned to allevate poverty and for ths reason they are a better canddate for evaluatng pro-poor government polces. 3. The case of Moldova Moldova s one example of the changes occurred n the FSU durng the 1990s wth a deep recesson durng the early part of the decade, two years of stablty n 1996 and 1997 and a new recesson caused by the 1998 Russan fnancal crss. It s an extreme example n that the fall n GDP (-60%) and the peak poverty level (71%) reached by 1999 are remarkable fgures even wthn the FSU scenaro. Indeed, the case of Moldova can be lsted as one of the worst recessons of the twenteth century on record. But what can socal assstance do to help the poor and the vulnerable under such crcumstances? Despte the growng needs, the socal assstance system lagged behnd n terms of reforms. In the Sovet Unon, the system dd not dstngush between socal nsurance and socal assstance and welfare benefts were characterzed by an emphass on specal categores of ctzens such as the war veterans, chldren and the dsabled. Durng the 1990s, the transtonal recesson and attempted reforms led to contrastng effects. On the one hand, reforms amed at two major objectves ncludng the separaton of socal nsurance from socal assstance and the transton of socal assstance from a categorcal to a means-tested system. The frst objectve was prncpally desgned to make socal nsurance ndependent and extra-budgetary. The second objectve amed at focusng scarce resources on the poor. On the other hand, fnancal resources were quckly shrnkng because of the recesson preventng these same reforms from beng mplemented. By the end of the 1990s, the government of Moldova had substantally reduced the number and sze of socal assstance benefts and had also accumulated a certan amount of arrears n payments wthout really mplementng structural changes. Economc recovery started n the year 2000 and GDP growth contnued at a sustaned level between 2000 and 2005 wth annual growth rates n between 5% and 8%. By 2004, poverty had sgnfcantly recovered to a level below the pre-1998 Russan fnancal crss and estmated at 26.5% of the populaton (GoM 2004). The begnnng of the growth perod almost concded wth the electon of the communst government n 2001 and the concomtance of growth and a conservatve 4

7 agenda contrbuted to change the fortunes of the socal assstance system. Between 2000 and 2005, all socal assstance benefts ncreased n number, sze or coverage wth growth rates n real expendture rangng from 6% to 335%. Benefts contnued to be manly targeted to the war veterans, dsabled and chldren and these three categores of benefcares together ncreased n sze between 2001 and 2004 to reach about 90% of all benefcares and total expendture. By 2005, the socal assstance benefts system ncluded nne dfferent types of benefts and had a total expendture estmated at 1.7% of GDP and 4.6% of publc expendture (WB 2007). The stated am of socal assstance benefts n Moldova s to assst the poor and the vulnerable. Accordng to the Socal Report 2001 prepared by the former Mnstry of Labor and Socal Protecton: (1) Socal assstance benefts and facltes are provded to persons that fnd themselves at rsk or are unable to make a lvng. (2) The objectve of socal assstance s to support persons who temporarly or contnuously face dffcultes whch mpede them obtanng necessary condtons for a decent lfe, helpng them to develop ther own capactes and competences for a relevant socal functonng. (GoM 2001, p.9, Englsh verson) Yet, the practce of socal assstance has contnued to be based n favor of selected categores of benefcares rrespectve of wealth status and, to date, none of the benefts are really means tested. 3 The government of Moldova had plans to experment wth means-tested benefts and a few categorcal benefts were complemented wth means-tests but - de facto - all socal assstance benefts contnued to be granted accordng to categorcal crtera. The government of Moldova may have been successful n targetng ts poltcal consttuences. But has t really addressed ts stated ntenton of mprovng welfare? 4. Key ssues We are confronted wth a retrospectve evaluaton. There was no desgn to evaluate socal assstance benefts before the program was launched and we cannot rely on a randomzed experment and/or household surveys whch were specfcally desgned for evaluatng socal assstance benefts. Ths s the rule rather than the excepton n transtonal economes. These countres face fnancal and sklls constrants whch lmt the funds and capactes to undertake a proper evaluaton desgn. The socal program we are evaluatng s expected to be affected by most problems whch are typcal of evaluatons of ths knd ncludng selecton bas and lack of a proper comparson group, unobserved heterogenety, model endogenety and measurement error. Selecton bas accrues from the fact that the group of households that receved socal benefts - the treated group - was not randomly selected by the government. Ths group may be very dfferent n terms of ts characterstcs from the group of households that dd not receve socal benefts - the untreated 3 Benefts for chldren n age 0-3 and materal assstance are supposed to be accompaned by means-tests but n realty the system preselects categores of benefcares makng the means-test an almost redundant exercse. 5

8 group - whch s therefore a poor choce of counterfactual. Several factors whch determne program selecton are not observed n our data and some of the observed factors are lkely to be plagued by endogenety and measurement error. Partcpaton to the program depends on the applcaton decson on the part of the household and on the selecton decson on the part of the government. On the household decson, we do not have any nformaton about whether the household has appled for benefts or not. On the government decson, we know that benefts are assgned to predetermned categores of ctzens such as chldren, the dsabled and the war veterans but we cannot dentfy from the survey the exact categores that correspond to each beneft. For example, we do not know the category of dsablty or the war veteran status. Nor, we can dstngush those varables that determne the household decson from those varables that determne the government decson. For example, havng many chldren may be a factor that nduces households to apply for benefts because poor households tend to have many chldren but s also a categorcal crtera used by the government to assgn benefts. The socal benefts program also suffers from msmanagement and not all benefts are assgned to those who are supposed to be targeted. Dsablty status s a gatekeeper for a whole range of benefts and t s wdely recognzed n Moldova that the commsson that grants dsablty status s not suffcently transparent n ts decsonal process. Endogenety may arse due to the fact that some of the unobserved crtera used for program selecton may be correlated wth observed varables. Also, the proxy-means tests n place for two of the exstng benefts are only subsequent to the categorcal crtera and result n sgnfcant targetng error type I (non coverage of poor people) and targetng error type II (coverage of rch people). Agan, these are not exceptonal crcumstances for an evaluaton of government transfers n transtonal economes but the rule. In substance, program selecton s not entrely determnstc but contans several sources of stochastcty. The socal assstance scheme has also expanded durng the perod consdered and ths has occurred n concomtance wth output growth and poverty reducton whch means that we also need to dsentangle the mpact of socal assstance transfers from the mpact of growth. 5. Evaluaton strategy Our objectve s to evaluate the mpact of socal assstance benefts on household welfare and poverty. We use four consecutve rounds of a household budget survey ( ) that covers ncome and consumpton and whch measures the socal assstance benefts receved by households (more on the survey n the data and varables secton). We also know that the socal benefts program was already on-gong n our perod one and expanded throughout the four year perod consdered. Let be our unt of nterest - the household - wth = 1,2,... n ; t an ndcator of tme wth t =1, 2, 3, 4; P a bnary varable that descrbes whether households partcpate ( P = 1) or 6

9 do not partcpate ( P = 0) to the socal benefts program; B a varable that measures the ntensty of total benefts receved by each household wth B = 0 f the household does not receve benefts and 0 < B f the household receves benefts; and Y a varable B max representng household welfare wth 0 < Y Y max. We wll use the notaton T to dentfy the group treated wth socal benefts and C to dentfy the comparson group. 4 Wth t =1, 2, we can dentfy four groups of households ( A, B, C, D) accordng to partcpaton P as follows: Group P P 1 2 A 0 0 B 0 1 C 1 0 D 1 1 We call group A 'Stayouts', group B 'Joners', group C 'Leavers' and group D 'Stayns'. Our prmary nterest s the comparson of groups A and B over the perod gven that we are tryng to assess the mpact of accrued benefts. However, comparng C and D can also provde a useful counterfactual. For example, Ravallon et Al. (2005) study the mpact on welfare of fallng out of the Trabajar workfare program n Argentna therefore focusng on groups C and D. We propose a three steps evaluaton process. The frst step nvolves estmatng the transton probabltes between dfferent welfare groups over tme. The advantage of ths approach s that s very versatle and provdes nformaton whch can be used for at least four dfferent exercses: 1) An assessment of the upward and downward moblty of households across welfare groups and tme; 2) An assessment of the role of socal benefts n fosterng or hamperng such moblty; 3) Two separate tests that assess the ablty of socal benefts to protect the non-poor from fallng nto poverty (PROT) and to promote the poor to ext poverty (PROM); 4) A poverty ncdence evaluaton. For the frst exercse t s suffcent to read the transton probablty matrx between two tme perods n the presence of benefts. Ths provdes nformaton on household moblty across welfare groups. For the second exercse and f we estmate the transton probablty matrx also n the absence of socal benefts, we can use the percentage of households n each cell of the transton 4 Note that the comparson group s not necessarly the group of non treated households observed n the data. The latter can be one of the possble choces of comparson groups and - as already explaned - not a good choce for our specfc case. 7

10 matrces to measure changes n probabltes wth and wthout benefts. A smlar procedure can be used to calculate the PROT and PROM tests. These tests have been developed by Ravallon et Al. (1995) and used n other works snce (see for example van de Walle 2003). Smplfyng the exposton, let F t ( z) be the share of the poor at tme t n the presence of socal benefts and gven a poverty lne z and let G t ( z) be the correspondng share n the absence of socal benefts. Let also F ( z, z) and G ( z, z) be the shares of those who stay poor between the tme perods consdered. The protecton (PROT) and promoton (PROM) tests are defned as: ( z, z) F ( z) F( z z) ( z, z) F( z z) PROT z) = G ( z) G, (1) ( PROM ( z) = G, (2) Postve values of these measures wll ndcate that socal benefts have been able to protect the non-poor from poverty and promote the poor out of poverty. Usng the same two transton matrces (n the presence and absence of socal benefts) we can estmate the ncdence of socal benefts on poverty and on dfferent welfare groups. To estmate the ncdence on poverty, t s suffcent to read the columns totals and compare these fgures n the estmates wth and wthout benefts. These are useful nformaton but we constraned the analyss to the sphere of ncdence evaluatons where the comparson group s smply the populaton wthout socal benefts. We have not really constructed yet a proper comparson group able to delver an mpact evaluaton. For ths purpose, the second step wll be to mmc a randomzed experment and then use a double dfference ( DD) or 'dfference n dfference' method to estmate the mpact of benefts on welfare. In our two years scheme, groups A and B dd not partcpate n the program n tme 1 whle n tme 2 only group B partcpates. If group A and B were dentcal n all respects n tme 1 then the socal benefts program could be consdered as randomly assgned. In our data, group A happens to be very large whle groups B, C and D are smaller (the coverage of benefts s low). Wth propensty score matchng we can extract from group A n tme 1 the nearest match to group B n tme 1. We can then compare changes n welfare of the two groups durng the perod wth the double dfference method as follows: D AB T T C C = E( Y ) E( Y )] [ E( Y ) E( Y )] (3) [ where E( Y ) s the expected value (the mean) of welfare. Postve values of D AB wll ndcate that socal benefts had a sgnfcant mpact on welfare. The DD estmator s generally used to address the problem of endogenous placement that arses from sngle dfference estmators. The am s to compare partcpants and non partcpants pre and post nterventon. 5 Such procedure generally entals a program start n perod 2 and a 5 DD evaluatons are very popular n all scences. A frequently cted reference of a DD evaluaton for developng countres s Duflo (2001) whle Lokshn and Yemstov (2003) provde an example of DD applcatons to transtonal economes. 8

11 selecton bas whch s tme-nvarant. If ths s the case, the DD estmator s unbased and the program mpact s correctly estmated (Ravallon 2008). In our case, we can relax the tme-nvarant assumpton gven that we do not use group A as a comparson group but a match of group B extracted from group A. One problem may nstead arse from the fact that we need to use the panel component of the surveys rather than the two full cross-secton surveys. 6 We need to trace the same households between perod 1 and 2. Ths mples a certan attrton resultng n a loss of observatons and a possble loss of sample representatveness. Ths potental loss wll be assessed by testng means equalty of the balanced and unbalanced panels. The thrd step wll be to explot the avalablty of a four years perod for a longtudnal parametrc analyss. We regress the ntensty of benefts (condtonal on a number of other varables) on a measure of welfare and the socal benefts status (condtonal on a number of varables) on the poverty status as follows: y t = α + βb + γx + ε (4) t t t Poor t = α + βp + γx + ε (5) t t t where α, β, γ are parameters, ε t s the error term, Poor t s a bnary varable wth 1 = Poor and 0 = Non _ Poor, y t s a contnuous measure of welfare (descrbed n the next secton), B t s a contnuous measure of benefts (descrbed n the next secton), P t represents program partcpaton and and t stand respectvely for unts of observaton and tme. We call [4] the welfare equaton and [5] the poverty equaton. The advantage of ths approach s that we can explot the ablty of longtudnal models to treat non-observed tme-varant and tme-nvarant factors under varous assumptons and address the problems of unobserved heterogenety and endogenety. The choce of panel models estmators s extremely rch. We opted to use Fxed effects (FE), Between Effects (BE) and Random Effects (RE) models addressng separately the queston of consstency of results across specfcatons wth dfferent sets of regressors and the queston of consstency of results across these three models (more on the varables used n the next secton). 7 To check for results consstency across models specfed wth dfferent sets of regressors we use 6 Note that the standard DD approach does not requre a panel structure (Ravallon 2008). 7 We excluded the pooled OLS because ths estmator would provde bas estmates gven that we expect sgnfcant endogenety due to unobserved heterogenety. We also excluded dummy varable regressons (LSDV) because ths s practcal only when N s small and Instrumental Varables (IV) models because they rely on assumptons about the nstrument that cannot be tested. Dynamc panel models were excluded because they would requre IV specfcatons to be unbased. A condtonal fxed-effects logt model was also excluded because t drops observatons where the treatment varable does not change over tme. In our data, these are groups A and D whch would mean wastng most of the dataset. Among error components models we opted to use both FE and RE models. A RE model assumes no covarance between the explanatory varables and the error term that represents person specfc and tme nvarant unobserved heterogenety. Ths would be a rather unrealstc assumpton. In a country lke Moldova, unobserved factors such as ndvdual contacts matter a great deal for fndng a job or accessng cash benefts. However, ths model has the advantage of accomodatng all types of regressors ncludng tme nvarant and cross-secton nvarant regressors. Ths s mportant because we want to test the model condtonal on varables such as rural locaton or the season when the ntervew took place whch would be dropped n a FE model. 9

12 the RE model as ths s the only one that can accommodate all types of varables ncludng tmenvarant and unt nvarant varables. Equaton [4] s estmated wth a GLS estmator whle equaton [5] s estmated wth a probt. The GLS specfcaton s as follows: t ( 1 θ ) α + ( x θ x ) β + [( θ ) ν + ( ε θε )] RE : ( y θy ) = 1 (6) t t where y s the welfare varable and y s ts mean, x s a vector of explanatory varables as 2 2 descrbed n equatons [4] and [5], θ s an arbtrary functon of and σ, α s the ntercept, β s the slope parameter of the regressors whch s assumed to be fxed across observatons, ν s the unt-specfc resdual and ε s the error term wth and t standng for unts of observatons and tme respectvely. In a second stage we addressed the queston of results consstency across dfferent estmaton models by comparng the three models FE, BE and RE. The wthn and between effects models are specfed as follows: FE : yt y + y = + ( xt x + x ) + ( t + + (7) σ v α β ε ε ν ) ε ε BE wth: 2 εt IID( 0, σ ε ); N, t T : y = α + x β + ν + ε (8) y T = t = 1 yt / T ; y = t y t / nt ; x T = t= 1 xt / T ; x = t x t / nt ; T ε ε / T ; = t = 1 t ε = ε / nt t t Wth a fxed (wthn) effects model the problem of tme constant unobserved heterogenety s addressed but the model uses nformaton only on the 'movers' wastng nformaton on the 'stayers'. The FE model also assumes that the ndependent varables and the dosyncratc error term are ndependent. In realty and n our case such endogenety may arse from systematc shocks due to macroeconomc factors such as growth (perod effect), omtted varables such as those varables that explan program selecton and are not observed, shocks that can also affect regressors such as changes n growth that may lead to changes n labor market condtons (smultanety) and measurement error whch we know we have. Autocorrelaton n the dsturbance term may also be an ssue as people who have already successfully appled for benefts may apply agan or smply contnue to be granted benefts wth no further enqury 8. We can address perod effects by ntroducng tme dummes n the fxed (wthn) effects model and ths partly addresses smultanety and we can also estmate a wthn effects model wth autoregressve dsturbance term. We opted nstead to use macroeconomc varables whch dentfy the 8 As already mentoned, some of the factors that determne the decson to apply are lkely to be non observed n the data. 10

13 macroeconomc shocks that occurred durng the perod consdered. We know that between 2001 and 2004 Moldova enjoyed sustaned growth and that employment has mproved durng the perod. We can use the growth and the employment rates as regressors n place of tme dummes and check the mpact of these shocks on the socal benefts parameter. In concluson, the RE model s an oblged choce for comparng models wth all forms of regressors. The FE model wth the macroeconomc varables s probably the best choce to address ssues of unobserved heterogenety, perod effects and smultanety but t drops tme nvarant regressors. The BE model s less relevant but s a useful counterfactual to the FE model whch s a wthn effects model and focuses on tme nvarant effects. 6. Data and varables Data are drawn from the Moldova Household Budget Survey (MHBS). The survey covers approxmately 6,240 households every year ntervewed n monthly blocks of about 520 households each. It ncludes a panel component wth an elaborate rotaton scheme and a maxmum tenure of each household of four years. 9 The survey s a mult-stage samplng and multpurpose survey and ncludes sectons on ncome and consumpton as well as sectons on labor, health and others. In the data used for ths paper, a balanced panel ncludes 2,469 households over two years and 866 over the four years perod. A balanced panel ncludes only observatons ncluded n the panel by desgn. As a measure of welfare we use real monthly consumpton per capta dvded by the poverty lne ( y). Ths s a standard approach n smlar studes (Ravallon et Al. 1995, van de Walle 2003). Consumpton s expected to be a better measure of welfare than ncome gven that ncome s underreported and s more senstve than consumpton to seasonal varatons. We use real consumpton measured at 2001 prces. The poverty lne s the one adopted by the Government of Moldova n 2001 whch was 195 Le (GoM 2004). The poverty lne was calculated wth a cost of basc needs approach based on a food basked of 2,100 calores/day and an extra amount calculated for non food tems. The Purchasng Power Party (PPP) equvalent value of the 2001 poverty lne was approxmately 2.5 USD/day whch s what the World Bank consders as an approprate poverty lne for transtonal economes (WB 2005). We opted to gnore equvalence scales. Ths s the approach followed by the World Bank n ts study on poverty n Moldova (WB 2004) and s also justfed by the arbtrary nature of equvalence scales and by the fact that results 10 are often very senstve to the type of equvalence scale used. Nne types of benefts are consdered: utltes compensatons, chld allowances, war veteran allowances, socal allowances, death grants, Chernobyl compensatons, care-taker allowances for 9 Sgnoret, J.E. (2003) provdes full detals on the panel structure and rotaton mechansms. 10 For example, n a study on poverty (GoM 2004), the Government of Moldova uses an equvalence scale of 1 for the frst adult n the household, 0.7 for the other adults and 0.5 for chldren below the age of 16. The poverty estmatons provded by ths study are very dfferent from the estmatons provded by the World Bank (2004) for the same years. 11

14 the dsabled, transport compensatons for the dsabled and materal assstance. Utltes compensatons, chld benefts, war veteran allowances and socal allowances account for the quas-totalty of benefts captured by the surveys. These benefts are all allocated monthly and the aggregated varable ( B) has been transformed n real and per capta terms and relatve to the poverty lne so as to be comparable wth the welfare ndcator ( y ). We use four groups of condtonal varables to whch we attach dfferent economc nterpretatons. The frst group of varables ( a ) s the characterstcs of the household head ncludng age, sex and educaton. These are standard varables whch are generally known to be relevant for household welfare. The second group ( b ) s varables ndcatng dfferent levels of household dependency from actve ndvduals ncludng the number of chldren and the number of household members relatvely to the number of earners n the household. These varables are expected to be correlated wth welfare and also wth the probablty of recevng benefts but they are not dropped by the software for multcollnearty and the correlaton coeffcents wth the socal benefts varables are rather low. The thrd set of varables ( c ) ncludes those varables that are tme-nvarant but may be relevant for welfare ncludng rural locaton and whether the household has been ntervewed n the summer months or not. Rural areas are known to be poorer and welfare s generally subject to seasonal varatons. The fourth group of varables ncludes two macroeconomc varables whch are expected to capture macroeconomc shocks. One s the growth rate whch s cross-secton nvarant and one s the dstrct employment rate whch s defned as the average dstrct number of earners dvded by the number of adults n the dstrct as captured by the survey. Ths last varable vares over tme and across dstrcts. Means of all varables used n the regressons are reported n Table A1 n the Annex for the four years and for the balanced and unbalanced panels. ( d ) 7. Results Socal assstance benefts n Moldova prvlege the poor, meanng that the greatest share of these benefts s allocated to households wth monthly consumpton per capta below the poverty lne. However, between 2001 and 2004 the share of socal benefts receved by the frst quntle (the poorest 20% of households) has declned sgnfcantly from 57.6% of total benefts to 49.2%. Redstrbuton has occurred n favor of the second and thrd quntles whle the top two quntles mantaned a flat and non nsgnfcant share of around 10% of total expendture each (Fgure 1). Lvng condtons have mproved very sgnfcantly between 2001 and Average consumpton relatvely to the poverty lne ( y ) has ncreased from 1.05 n 2001 to 1.53 n 2004 and the poverty headcount rato has declned from 62.6% to 34.1%. 11 The average beneft relatve to 11 The poverty trends ndcated here dffer from those reported n poverty studes on Moldova (GoM 2005 and World Bank 2004). Ths s explaned by the adopton of dfferent deflators for the consumpton fgures and by the use or non use of equvalence scales. 12

15 B) the poverty lne ( for benefts recpents has nstead declned between 2001 and 2004 from 0.16 to However, ths declne has been accompaned by a remarkable growth n coverage (from 5.7% to 19%) whch explans the growth of average benefts per capta n the populaton (Table 1). The government clearly opted to ncrease expendture and coverage rather than focusng on proper targetng based on means tests and the fne tunng of ndvdual transfers. The transton probablty matrces would suggest a postve ncdence of socal benefts on poverty. Table 2 reports the transton probabltes between poor and non poor. The frst two columns n Table 3 report the dfference between these probabltes estmated wth and wthout benefts. Postve values ndcate an ncreased probablty of movng across poverty groups n the presence of socal benefts and vce-versa for negatve values. To check for consstency, we repeated the estmatons usng the balanced and unbalanced four years panels and also the balanced panels for the three perods , and We can see that the values for the transton from poor to non-poor (P-NP) are all postve whle the values for the transton from non poor to poor (NP-P) are all negatve. The PROT and PROM tests proposed by Ravallon (1995) confrm these fndngs wth postve values for both PROT and PROM for all estmates (Table 3, columns 3 and 4). 12 The poverty headcount ndex n the presence of socal benefts s also lower than n the absence of benefts for all estmates (Table 3, column 5). In substance, the presence of socal benefts enhances the probablty of movng from poor to non-poor and reduces the probablty of movng from non-poor to poor. As already mentoned, the problem wth ths procedure s that we gnore the ssue of the reference group usng the untreated ndvduals as comparson group uncondtonal on household characterstcs. The double dfference estmates wth matched samples provde a rather dfferent pcture. Table 4 reports the relevant statstcs for each group and the respectve match over each of the three perods consdered. When we compare those who joned the socal benefts scheme (group B) wth a matched sample extracted from those who dd not jon (group A) we fnd that all DD estmates are negatve. Ths ndcates that jonng the scheme dd not mprove welfare for the treated everythng else beng equal. As a counterfactual, we estmated the DD measures comparng those who left the scheme (group C) wth a matched sample extracted from those who contnued to stay n the scheme (group D). Leavng the scheme seems to mprove welfare everythng else beng equal. From these estmates, socal benefts seem to have a negatve effect on welfare. The dsadvantage of the double dfference approach as we desgned t s that we constran the sample to a rather small number. Ths s due to the fact that coverage s ntally low and that the treated group s small. On the other hand, f the matched sample s vald, estmates should be 12 Note that columns 1 and 4 n table 3 report the same values. That s because the PROM test s reduced to the dfference of the transton probabltes between estmates wth and wthout benefts. 13

16 consstent and close to a randomzed experment. Also, double dfference estmates and a randomzed experment can mss on some of the behavoral effects that may arse from unobserved heterogenety such as dfferences n ndvdual abltes. Durng a perod of sharp growth such as the one we observe, those who proft from and contrbute to growth are probably the most sklled, wth better contacts and socal abltes, all attrbutes whch are dffcult to measure n household surveys. Ths motvates the thrd step n our analyss where we use longtudnal data to control for unobserved heterogenety. Tables 5a and 5b report random effects estmates usng all 16 possble specfcatons and the balanced and unbalanced panels. 13 Table 5a llustrates results for the welfare equatons and 5b for the poverty equatons. For smplcty of exposton, we report only the coeffcents and the z statstcs for the varables of nterest B and P. Wth a balanced panel (top panel n Table 5a), B s sgnfcant wth a postve sgn n equatons 2, 6, 9 and 13. In all these equatons, the macroeconomc varables (dstrct employment rate and annual growth - d) are omtted whch would suggest that once we control for growth socal benefts do not nfluence welfare. Ths same statement s not true f we consder the unbalanced panel. In 12 of the 16 welfare equatons B s sgnfcant, ncludng four equatons where the macroeconomc varables are present (Table 5a). Smlar results are found wth the poverty equatons (Table 5b). Wth the balanced panel, P s sgnfcant n equatons 2, 6, 9, 12, 13, 14 and 16. In all these equatons the macroeconomc varables are omtted. When we control for the economc cycle, the sgnfcance of socal benefts dsappears. However, wth the unbalanced panel, n eght of the 16 poverty equatons P s sgnfcant ncludng four equatons where the macroeconomc varables are present (Table 5b). Therefore, the unbalanced panel would suggest that the economc cycle s not suffcent to explan mprovements n welfare and that socal benefts may play a role. Tables 5a and 5b used a random effects model. In order to test dfferent estmaton models, we have to exclude the tme-nvarant varables (rural and summer) whch are not supported by the fxed effects model. Table 6a and 6b report the results for the balanced and unbalanced panel estmatons and nclude the three estmaton models consdered (FE, RE and BE). As before, the exercse s repeated for the welfare and poverty equatons. Wth the balanced panel estmatons (Table 6a), none of the socal assstance varables s sgnfcant n ether the welfare or poverty equatons. Wth the unbalanced panel (Table 6b) the socal assstance varables are sgnfcant n equatons 2, 3 and 4. Agan, the dfference s between balanced and unbalanced panel wth the balanced panel suggestng no effect of socal benefts and the unbalanced panel suggestng some effect. 13 By 'all possble specfcatons' we mean nsertng and omttng all groups of varables a,b,c and d one at the tme. 14

17 The balanced panel s more approprate to assess the mpact of benefts provded that the attrton rate, whch s very large over the four years, does not bas the sample. Table A1 n the Annex reports the means for all varables and the t-test for means dfferences between balanced and unbalanced panels. The dfference n means s generally rather small and below 10% across varables wth most varables showng a gap below 20%. The t-test shows a sgnfcant dfference between the balanced and unbalanced panels only for the rural and summer varables. In partcular, for the varables of man nterest (y, poor, B and P) the dfference n means s never very sgnfcant. Moreover, the unbalanced panel largely 'pollutes' tme effects wth ncomplete spells whch s a serous shortcomng f annual macroeconomc shocks are responsble for explanng most of the varaton n welfare whch seems to be our case. The Hausman tests reported on the bottom of Tables 6a and 6b speak also n favor of the FE model rather than the RE model and n the FE model used wth the unbalanced panel the ntensty B) of benefts varable ( s not sgnfcant. In substance, when we use the most approprate model (FE), the most approprate sample (balanced) and when we control for macroeconomc trends socal assstance does not seem to have any postve and sgnfcant effect on welfare or poverty. 8. Concluson Between 2001 and 2004, Moldova experenced sustaned growth and poverty reducton. The system of socal assstance contnued to focus on selected categores of benefcares ncludng manly chldren, the dsabled and the war veterans rrespectve of welfare status. Relatve expendture on the bottom quntle and expendture per benefcary decreased durng the perod n favor of ncreased overall expendture and a sharp rse n populaton coverage. Ths approach does not seem to have mproved household welfare, all other factors beng equal. If we gnore the questons of selecton bas, behavoral responses, unobserved heterogenety, endogenety and measurement error, an ncdence analyss would suggest that socal benefts enhance the probablty of movng out of poverty and reduce the probablty of movng nto poverty. However, double dfference estmates and regresson analyses whch are able to control for at least some of these factors would suggest that socal benefts have not contrbuted to mprove welfare overall. Double dfference estmates show a negatve mpact on welfare. The man dfference between the ncdence and the double dfference analyses as we structured them n ths paper s that the ncdence analyss smply compares the benefcares and non benefcares gnorng the queston of comparablty of the two groups whereas we used a matchng procedure to carry out the double dfference estmatons. Comparng two homogeneous groups wth the double dfference procedure allowed us to observe that the non benefcares of socal assstance benefts have performed better than the benefcares. There may be both demand and supply sde reasons whch could explan the negatve values 15

18 we fnd wth the double dfference estmates. On the supply sde, dfferent dstortons of the socal assstance system such as a lack of transparency n the selecton mechansm of benefcares or the use of categorcal selecton crtera resulted n dstorted targetng. Targetng the non poor, n turn, may have had negatve behavoral effects n terms of labor supply whch could explan foregone ncome opportuntes for those non poor who partcpated to the socal assstance program. These problems may also have ncreased n weght durng the perod consdered gven that the volume of benefts ncreased wthout real restructurng of the targetng system. A second set of explanatons may relate to the demand sde through the mechansm of self-selecton of benefcares. Wth the double dfference procedure we compared two smlar groups selected on the bass of observed characterstcs. However, unobserved characterstcs such as ndvdual abltes may play an mportant role. It may be that the group of non benefcares s also the group wth better ndvdual abltes and whch was able to better explot the emergng growth opportuntes. Ths problem s not really addressed by the matchng procedure and the double dfference estmates and motvated the panel approach that followed. Panel survey estmates dd not evdence any consstent sgnfcance of socal assstance benefts n explanng welfare or poverty and ponted nstead toward a strong growth effect. Therefore, even when we take nto account unobserved heterogenety, socal assstance benefts do not seem to explan poverty reducton. We can conclude that the real engne of poverty reducton n Moldova between 2001 and 2004 has been growth and that socal assstance has had ether a negatve effect through a dsplacement effect (socal assstance benefcares who would have done better wthout benefts) or a non sgnfcant role. References Abade, Alberto, Drukker, Davd, Leber Herr, Jane and Imbens, Gudo, Implementng Matchng Estmators for Average Treatment Effects n Stata. The Stata Journal, 1, pp Abade, Alberto and Imbens, Gudo, Smple and Bas-Corrected Matchng Estmators for Average Treatment Effects. NBER Techncal Workng Paper No. 283 Abade, Alberto and Imbens, Gudo, Large Sample Propertes of Matchng Estmators for Average Treatment Effects. Econometrca 74, 1, Duflo, Esther, Schoolng and labor market consequences of school constructon n Indonesa: Evdence from an unusual polcy experment. The Amercan Economc Revew 91, 4, Government of Moldova, Socal Report Mnstry of Labor and Socal Protecton. Mmeo. 16

19 Government of Moldova, Polcy and Poverty Impact Report. Mnstry of Economy and Trade. Mmeo. Government of Moldova, Medum Term Expendture Framework Mnstry of Fnance. Mmeo. Ima, Kosuke, Kng, Gary, Stuart, Elzabeth, The Balance Test Fallacy n Matchng Methods for Causal Inference. Wred at: polmeth.wustl.edu/ Lokshn, Mchael, and Ravallon, Martn, Welfare mpacts of the 1998 fnancal crss n Russa and the response of the publc safety net. The Economcs of Transton 8, 2, Leuven, Edwn and Sanes, Barbara, PSMATCH2: Stata module to perform full Mahalanobs and propensty score matchng, common support graphng, and covarate mbalance testng. Wred at: Lokshn, Mchael, and Yemtsov, Ruslan, Evaluatng the mpact of nfrastructure rehabltaton projects on household welfare n rural Georga. World Bank Polcy Research Workng Paper Mlanovc, Branko, Socal transfers and socal assstance: An emprcal analyss usng Latvan household survey data. World Bank Polcy Research Workng Paper Okrasa,Wlodzmerz, 1999a. The dynamcs of poverty and the effectveness of Poland s safety net ( ). World Bank Polcy Research Workng Paper Okrasa, Wlodzmerz, 1999b. Who avods and who escapes from poverty durng the transton? Evdence from polsh panel data, World Bank Polcy Research Workng Paper Ravallon, Martn, Galasso, Emanuela, Lazo, Teodoro, and Phlpp, Ernesto, What can expartcpants reveal about a program s mpact? Journal of Human Resources 40, 1, Ravallon, Martn, van de Walle, Domnque, and Gautam, Madhur, Testng a socal safety net. Journal of Publc Economcs 57, Sgnoret, Josè E., The Moldova hbs: A note on the panel sample. Mmeo. 17

20 van de Walle, Domnque, Testng Vetnam s publc safety net. Journal of Comparatve Economcs 32, 4, World Bank, Growth, Poverty and Inequalty: Eastern Europe and the Former Sovet Unon. The World Bank. World Bank, Recesson, Recovery and Poverty n Moldova. The World Bank; Report No World Bank, Improvng Publc Expendture Effcency for Growth and Poverty Reducton: A Publc Expendture Revew for the Republc of Moldova, World Bank Report No MD

21 Fgure 1 Structure of Socal Benefts by Consumpton Quntle Note: Consumpton quntles are based on consumpton net of socal benefts. 19

22 Table 1 Household Benefts and Socal Welfare Average consumpton relatve to PL (y) Obs Popul. 3,630,500 3,623,991 3,553,774 3,603,873 Mean Std. Dev Average benefts relatve to PL for benefts recepents (B) Obs Popul. 269, , , ,745 Mean Std. Dev Average benefts relatve to PL for all (B) Obs Popul. 3,632,407 3,623,991 3,553,774 3,603,873 Mean Std. Dev Coverage and poverty Benefts coverage Poverty headcount

23 Table 2 Transton Probabltes Matrces Wth benefts Wthout benefts Unbalanced panel 4 years Poor Non poor Total Poor Non poor Total Poor Poor Non poor Non poor Total Total Balanced panel 4 years Poor Non poor Total Poor Non poor Total Poor Poor Non poor Non poor Total Total Balanced panel Poor Non poor Total Poor Non poor Total Poor Poor Non poor Non poor Total Total Balanced panel Poor Non poor Total Poor Non poor Total Poor Poor Non poor Non poor Total Total Balanced panel Poor Non poor Total Poor Non poor Total Poor Poor Non poor Non poor Total Total