Can an Employment Guarantee Alleviate Poverty? Evidence from India s National Rural Employment Guarantee Act

Transcription

1 Can an Employment Guarantee Alleviate Poverty? Evidence from India s National Rural Employment Guarantee Act Stefan Klonner and Christian Oldiges University of Heidelberg Draft, January 15, 2013 Abstract This paper examines the e ects of India s recent Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA) on consumption and poverty in rural India. We combine data from India s National Sample Survey (NSS) on household consumption with information on the district-wise roll-out of the MGN- REGA. We perform both di erence in di erences and regression discontinuity estimations. While we do not nd robust program e ects for the sample of all rural households, we nd signi cant poverty-reducing e ects for the sub-sample of households belonging to scheduled castes and scheduled tribes (SC/ST). For this group of marginalized households we nd an increase of average consumption of about 15 percent and a decrease in various poverty measures between one fth and one half. In addition, for the same group of households, we nd that the MGNREGA reduces the exposure to seasonal shocks as the Act s e ects are concentrated on the agricultural lean season. A cost-bene t analysis suggests a reasonable degree of cost-e ectiveness of the Act as the additional consumption enjoyed by SC/ST households gures at about one third of the Act s total expenditures. 1

2 1 Introduction Despite annual growth rates of about 8 per cent India still has a large number of people living in poverty, with estimates ranging between 20 and 50 per cent of India s population of about 1.2 billion. 1 Ever since independence, e orts to reduce poverty have been plenty and diverse. The National Rural Employment Guarantee Act - or as it was later baptized Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA) - is the latest and most comprehensive initiative. The preamble to the Act speci es its main goal, "ensuring inclusive growth in rural India", and channels through which this is to be achieved, most prominently through its "impact on social protection, [and] livelihood security." 2 Enacted in 2005, the workfare program guarantees 100 days of employment to every rural household whose members are willing to do unskilled manual labour at the statutory minimum wage. In the nancial year of alone, more than 50 million households were employed across the Indian subcontinent. Per workday, workers are paid at least Rs. 100 (about $2). The Indian government spends about $10 billion each year on this program, which has amounted to about 4 per cent of its total expenditures in Since its rst year of implementation researchers, activists and civil society organizations have been keen to nd out whether the MGNREGA has been meeting its goals. Issues of interest have been whether funds reach the workers, whether and how works are being taken up, who the bene ciaries are and to what extent participating households bene t. In an overview and compilation of the major studies since the birth of the Act, Khera (2011) describes the process of implementing the Act "in letter and spirit" and the "realization of workers entitlements" as "the battle for employment guarantee". Regarding the Act s e ect on labor market outcomes, four recent econometric studies nd that the Act has increased rural wages signi cantly (Azam, 2012; Berg, Bhattacharyya, Durgam, and Ramachandra, 2012; Imbert and Papp, 2012; Zimmermann, 2012). According to these papers, which are largely based on National Sample Survey (NSS) data and di erence in di erence estimations, the Act resulted in an increase of rural wages between 4 and 8 per cent. Female workers and marginalized groups belonging to scheduled castes and scheduled tribes (SC/ST) are among the main bene ciaries. These studies also show that demand for MGNREGA employment is highly seasonal and that the Act serves as a safety net during the lean season when agricultural work opportunities are scarce. The just-mentioned studies imply large aggregate labor market e ects given that the program is India-wide and the rural work force comprises about 1 According to the Indian Government poverty rates are between 20 and 30 percent depending on the poverty lines, whereas according to Multidimensional Poverty Index estimates about 53 per cent of households are deprived in at least 3 out of 10 indicators (Alkire and Santos, 2010). Also see: 2 See MGNREGA Guidelines Operational_guidelines_4thEdition_eng_2013.pdf 3 For the Indian Budget of , visit ndia.indiatimes.com/home/union-budget- 2011/Highlights-of-Union-Budget /articleshow/ cms. 2

3 300 million people. While agricultural wages and individual consumption might be positively correlated, especially for the rural poor, it is not clear from the just-cited studies to what extent the Act has had an e ect on rural households welfare. For example, has individual consumption increased due to higher wages? Does the Act succeed in reducing households risk to seasonal labor demand shocks and help smooth consumption across agricultural seasons? Have the rural poor, the main target group of the program, bene ted disproportionately from the MGNREGA? Moreover, estimates of money-metric impacts of the Act on individual welfare allow an assessment of the e ectiveness of each Rupee spent for the program. Without such knowledge one cannot argue for or against the statement - often heard by opponents of the Act -, that simple cash hand-outs to the poor would be a much easier, and more e ective alternative (Economist, 2008). In this paper we combine data from several waves of India s nationally representative National Sample Survey (NSS) on household consumption with information on the district-wise roll-out of the MGNREGA to estimate causal e ects of the MGNREGA on rural households consumption expenditures and consumption-based poverty measures. We make use of the phase-wise roll-out of the Act according to which the MGNREGA was implemented rst in 200 districts in the scal year (Phase 1), and in another 130 districts during the following year (Phase 2). We use a district-level panel with NSS and program coverage data for the years to and employ two distinct empirical approaches. First, we conduct di erence-in-di erences (DID) estimations, where we consider consumption levels and consumption-poverty at the household-level as dependent variables and concurrent program availability in the household s district of residence as the explanatory variable of interest to obtain estimates of the intent-totreat e ect of the MGNREGA on household welfare. In our second empirical approach, we implement a regression discontinuity design (RDD) similar to Zimmermann (2012). Toward this we use the o cial district-wise poverty ranking of India s National Planning Commission from 2003, which has served as the basis for program allocation to districts during the years and In this process, the declared intention of policy makers has been to prioritize poorer districts by granting earlier access to the program. Following Zimmermann (2012), we predict actual program status of a district in by whether it is among the 350 poorest districts according to the 2003 Planning Commission ranking. To be precise, we employ a fuzzy RDD by regressing the outcomes of interest on the Planning Commission s district poverty score, which is a continuous variable, and the MGNREGA program status thus instrumented. In our empirical analysis we focus on the second phase of program implementation because the e ects of the rst phase may be blurred by other programs that were introduced earlier and targeted at many of the MGNREGA phase 1 districts. For the sample of all rural households, we nd no statistically signi cant e ect of the program on average household consumption. Regarding poverty, our ndings are inconclusive as we obtain opposite results in our two empirical approaches, poverty-reducing according to DID and poverty-increasing e ects according to RDD. 3

4 Our main interest, however, is the MGNREGA s potential for providing a safety net for the most disadvantaged and marginalized groups of the rural population. One such group that can be immediately identi ed from the NSS data are SC/ST households. According to our sample means, SC/ST households risk of being in extreme poverty is almost three times that of non-sc/st households. In addition, a vast body of literature points out the deprivations and discriminations that these formerly untouchable individuals face, particularly in rural areas (e.g. Sundaram and Tendulkar (2003)). For our subsample of SC/ST households, we nd strong bene cial e ects in both of our empirical approaches. According to our DID and RDD point estimates, the Act has increased SC/ST consumption by 8.8% and 15.0% on average, respectively. At the same time, the Act has decreased measures of both extreme as well as moderate poverty dramatically for this group, between one fth and one half. Taken together, our ndings imply greater average bene ts of the program for SC/STs than for the entire rural population. A second important objective of the Act has been to reduce the exposure to risk and provide income security for poor rural households. Given that unemployment and poverty rates are signi cantly higher during the agricultural slack season, we focus on the seasonal impact of the Act. According to our RDD results, we nd that the overall e ect of MGNREGA for SC/ST households is driven entirely by improvements in consumption and poverty during the (dry) spring season. Estimates of the program e ects allow a rough cost-bene t analysis of the program. For example, our results for mean consumption among the SC/ST population imply additional consumption expenditures per SC/ST household of about Rs. year, which compares to program expenditures of Rs. 1,200 per 1,111 per rural household or roughly Rs. 4,500 per SC/ST household. By and large we conclude that, according to our ndings, the Act has had economically signi cant e ects on rural welfare. Moreover, the Act appears to have contributed successfully to what has been o cially stated as its major goal "inclusive growth in rural India". Our results con rm that it has in fact improved "social protection [and] livelihood security" for a particularly disadvantaged group of the rural population. To our knowledge, there are no existing studies which attempt to estimate aggregate causal e ects of the Act on immediate measures of household welfare - rather than labor market outcomes. Ravi and Engler (2009) use a panel of 320 households in villages of Andhra Pradesh and show that both wages and consumption expenditure increase in response to the Act. Under the assumption that the manual labor o ered under the MGNREGA is a burden for many workers, Lagrange and Ravallion (2012) propose to account for the disutility in a novel utility function. Acknowledging the disutility of work they estimate disutility-adjusted poverty measures of the Foster Greer Thorbeke (FGT) family 4 for a sample of MGNREGA workers in Bihar, using their own survey and NSS data. 5 4 For the FGT poverty measures, which include the headcount ratio, poverty gap, and squared poverty gap, see Foster, Greer, and Thorbecke (1984). 5 Also see Dutta, Murgai, Ravallion, and Van de Walle (2012) for further results of the Bihar survey 4

5 The structure of the paper is as follows. In Section 2, we give a brief overview of the poverty situation and poverty measurement in rural India and then highlight the key features of the MGNREGA. Section 3 describes our data and Section 4 explains our two distinct econometric approaches. We present our main results welfare measures and seasonal impacts including Placebo tests in Section 5. In Section 6. we conclude with some back of the envelope calculations. 2 Background on Poverty and the MGNREGA Poverty in India has been declining over the last few decades. At the same time, however, economic progress came along with greater inequality. Deaton and Drèze (2002) explain that the decline in poverty is evident, but since poverty is still unavoidable for many while others enjoy income gains inequality is higher than before. The decline in poverty is largely uncontested, but its magnitude is a matter of debate. The Tendulkar Committee (T.C.), 2009 recommends reforms to poverty measurement in India. Following the Planning Commission s procedure of applying MPCE based poverty lines the T.C. calls for a re-adjustment of these poverty lines, not just via the CPI-AL but by applying Fisher price indices taking into account rural and urban price di erences (see Tendulkar Committee and others (2009)). According to the T.C. s now widely accepted poverty lines the negative trend in poverty rates over the decades is similar, albeit at di erent levels. Poverty headcount ratios for rural India fell from 40 to 30 per cent and not from 30 to 20 per cent between and Hereon we refer to the poverty gures de ned by the P.C. as extreme poverty and those de ned by the T.C. as moderate poverty. A more holistic approach to measuring poverty or the quality of life is to apply the Alkire and Foster method (Alkire and Foster, 2011a) and calculate the number of households who are deprived in multiple dimensions of well-being. The multidimensional poverty index (MPI) of 2010 is one such attempt (Alkire and Santos, 2010). According to the MPI, 53 per cent of India s rural population is deprived in at least three out of ten indicators of health, education, and living standard. 6. Given this context of widespread poverty (irrespective of the method of measurement), the MGNREGA is a major social sector program. It was preceded by the National Food for Work Programme (NFFWP), which lasted from to and can be viewed as a kind of trial for the MGNREGA. 7 8 Initially implemented by the United Progressive Alliance (UPA) government in the on targeting and rationing of poor households. 6 For a discussion of this technique, refer to Alkire and Foster (2011b); Ravallion (2011). And for further discussions on Indian poverty measurements refer for example to Deaton and Kozel (2005); Sen and Himanshu (2004a,b); Himanshu (2007). 7 For more information regarding the NFFWP, consult the Right to Food Campaign s website, The related identi cation problems shall be discussed in the sections below. 8 Also, workfare programmes are not new in Indian history as the Maharashtra Employment Guarantee Scheme (MEGS) from the 1970s is a famous and well researched example of its kind (see for example (Basu, 1981; Drèze, 1990; Ravallion, Datt, and Chaudhuri, 1993)). 5

6 200 poorest districts of India, it is designed as a safety net for rural households. 9 As presented in Table 2, under the MGNREGA about 20 million households were employed in its rst year, amounting for a total government expenditure of Rs. 9,000 crores. Over the following years, the MGNREGA was rolled out in two additional phases, so that by additional districts were covered (Phase 2) and by all remaining districts of India (Phase 3) were covered. Under the Act every rural household is entitled to 100 days of work at the statutory minimum wage which is set by each state government. It is important to note, that the Act provides for universal entitlement as any rural resident who is willing to volunteer for work, irrespective of gender, caste, or religion, is entitled to the right to work within 14 days of application. According to the Act, any non-compliance would grant unemployment allowance. This, however, along with several other provisions of the Act such as su cient worksite facilities for young mothers and children, a recommended employment quota of 33 per cent for women, timely wage payments and full transparency are often not met across all districts (Drèze and Khera, 2009). Despite the shortfalls all provisions of the Act in letter and spirit,,over 50 per cent of all MGNREGA workers belong to SC/ST households and more than a third of all workers are female in From Table 2, the MGNREGA di ers across the two phases, Phase 1 and 2. Person-days per rural household, for instance, increased from 17 to 20 between and , whereas the number of person-days generated in Phase 2.districts - 11 per rural household - are much lower right from the beginning. For Phase 1 districts we also observe an increase in total expenditure and wages between the two years of interest. Looking at Table 1 we observe that the targeting of poor districts was successful in as much as the average household of a Phase 1 district is indeed poorer than households from districts chosen for the latter phases. For instance, according to our extreme measure of the poverty headcount ratio in , about 30 per cent of all rural households in Phase 1 districts are poor, about 20 per cent in Phase 2 districts, and 15 per cent in Phase 3 districts. Similarly, we see such a trend for our measure of moderate poverty, albeit at higher levels, and for the poverty gap measure. At the same time, with NSS data of the rounds 61, 62, 63, and 64 for the years between and we see a more or less linear trend in falling poverty levels for both Phase 1 and Phase 3 districts. For Phase 2 districts, on the other hand, we notice a hump shaped trend, where poverty rst increases between , and , and then drastically plummets by more than a 30 per cent in From Tables 1 and 2 we conclude that there has been a negative trend in both moderate and the extreme poverty between and , during the time MGNREGA started and grew in intensity. These rather broad observations do not allow us to infer any causal e ects of the MGNREGA. In the following sections we therefore undertake a more complex but standard approach (di erences-in-di erences) to answer whether the poor, the target group, bene t from the Act disproportionately more than then 9 And for a comprehensive account of the Act s history consult the book by Khera (2011). 6

7 non-poor households. 3 Data and Measures of Poverty 3.1 Data In the subsequent analysis we make use of the following NSS rounds: 61, 62, 63, and 64 for the years , , , and , respectively. We use NSS sampling weights to estimates the welfare e ects across all rural households. We do not aggregate at the district-level. Similarly, we calculate all household controls regarding household size and social group of household from these data sets. We de ate all prices to prices using the Consumer Price Index for Agricultural Labourers (CPI-AL), based on Throughout, we rely on MPCE collected via the mixed recall period, and we measure real MPCE in absolute as well as in logarithms (multiplied by a factor of 100). Our sample consists of 504 districts of all major Indian states, including Jammu and Kashmir and Assam, excluding all other Northeastern states and Union territories. From the 504 districts 188 are Phase 1 districts, 103 Phase 2 districts and the remaining 213 are Phase 3 districts. Sample Summary Statistics are given in Table 3 and Phase wise di erences in means are given in Table 4. To control for district-wise heterogeneous time trends we employ a score-interaction term. This score is based on a district ranking by the Planning Commission (2003) which ranks 445 districts on a score of "backwardness" and is calculated as the mean of three subindices: "percentage of SC/ST population", "Agricultural Output per worker", and "Agricultural Wages". The nal composite score ranks districts from a score of (most backward) to a score of (least backward). 3.2 Measures of Poverty As mentioned in Section 2 there are at least two consumption based poverty measures in India. First, there is the what we call extreme headcount ratio which is based on the poverty lines as set by the Planning Commission (P.C.). And second, there is the what we call moderate headcount ratio of poverty as per poverty lines set by the Tendulkar Committee (T.C.). State-wise poverty lines as per the P.C. are extreme in the sense that they are much lower than those set by the T.C.. Therefore, according to the T.C. s poverty lines poverty is more prevalent than under poverty lines of the P.C. In our subsequent analyses we use both the measures to test the e ects on moderate and extreme poverty. Regarding poverty across phases, Table 4 highlights that on average households in Phase 1 are poorer than in Phase 2 districts, which in turn are substantially poorer than households in Phase 3 districts. This speaks for a targeted roll-out of the MGNREGA and underlines the planner s reliance on a poverty ranking. 10 Source: India Budget

8 From Table 3 it is apparent that the subsample of SC/ST households is much poorer than rural households from the entire sample. According to all poverty measures (HCR and PGR), extreme and moderate, the SC/ST population is disadvantaged. The HCR (P.C.) for the SC/ST sample is 50 per cent higher than the corresponding one for the whole sample. 4 Approaches 4.1 Approach 1: Di erence in Di erences (DID) In this section we explain our rst estimation technique. As the MGNREGA was rolled out in phases over the course of three years a di erence in di erences (DID) estimation model is standard and straight forward. For instance, it has been applied by aforementioned studies on labor market e ects (Azam, 2012; Berg, Bhattacharyya, Durgam, and Ramachandra, 2012; Imbert and Papp, 2012) Model for two years of data Our model for two years of data is as follows: Y idt = d + D07 t + T reat d D07 t + X idt + idt ; (1) where Y idt represents the dependent variable for household i in district d and time t:we use data from the years and in our main speci cations. 11 X idt captures control variables for household i, in district d and year t. d is a xed e ect unique to district d and idt is a random error term capturing all disturbances per household, district and year. The binary variable D07 is a dummy equal to one for observations from the year and zero otherwise. The variable T reat is a dummy equal to one for districts in which the MGNREGA was active in and zero otherwise. Our coe cient of interest is. The following is the same speci cation but now identi cation relies only on intrastate variation across districts over time: Y idst = ds + s D07 t + T reat ds D07 t + X idst + idst ; (2) where the subscript s denotes the state in which district d is located. So instead of one dummy for the year , there are S state-speci c year dummies, where S denotes the number of states in our sample. 11 One robustness check uses the year as the reference year. 8

9 A variation of the previous model is Y idt = d + D07 t + 1 P hase1 d D07 t + 2 P hase2 d D07 t + X idt + idt ; (3) which allows to estimate separate treatment e ects for Phase 1 and Phase 2 districts. Notice that in Phase 1 districts are in their second year of implementing the Act while Phase 2 districts are in their rst year. If the MGNREGA s e ect involves a lag of one year, the corresponding coe cient restriction is 2 = 0: Model for more years of data Here we present variations of the speci cations described above which include data of several years. Regression equation with data for three years ( , and ): Y idt = d + 1 D05 t + 2 D06 t + P hase1 d D06 t + X idt + idt ; (4) Regression with data for four years ( , , , and ): Y idt = d + 1 D05 t + 2 D06 t + 3 D07 t + 1 P hase1 d D06 t + 2 P hase1 d D07 t + 3 P hase2 d D07 t + X idt + idt ; (5) Same speci cation but now identi cation relies on only intra-state variation across districts over time: Y idst = ds + 1s D05 t + 2s D06 t + 3s D07 t + 1 P hase1 ds D06 t + 2 P hase1 ds D07 t + 3 P hase2 ds D07 t + X idst + idst ; (6) Interaction terms A concern is the lack of randomness of program placement. For example, Phase 1 districts may have evolved more slowly in the absence of the program. We exploit the Planning Commission s program placement rule according to which 150 of the 200 districts of Phase 1 were chosen from the ranking of "Backwardness". This ranking assigns to each district a score which is calculated as the mean value of three subindices which are separate indices on SC/ST population, agricultural wages and agriculture output per worker. The score then ranks all districts from the most backward to the least backward (from to 2.159). At the same time MGNREGA s e ect on welfare may substantially depend on its level of implementation, which we call intensity. One way to measure the intensity is to calculate person-days per rural household, as practiced in this paper. The following estimating equation holds for the data scenario "Data for 3 years, to ": 9

10 Y idt = d + 1 D05 t + 2 D06 t + 1 P hase1 d D06 t + 2 I d D06 t + 1 I d D06 t + X idt + idt ; (7) where I d is the interaction term of choice. 4.2 Approach 2: Regression Discontinuity Design (RDD) RDD Motivation As highlighted above a concern is the lack of randomness of program placement and it may well be that a DID framework fails to measure the true e ect of the MGNREGA with the main reason being selection bias. To counter this problem, we exploit the Planning Commission s (2003) ranking as a program placement rule and nd a discontinuity therein, which provides an instrument. Doing so we are able to estimate a local average treatment e ect (LATE) at the discontinuity, thereby eleminating the problem of selection bias RDD Idea and Implementation In her study on labor market e ects of the MGNREGA Zimmermann (2012) uses a regression discontinuity framework to identify the program placement rule of the MGN- REGA. Following her technique, we rank all districts in each state in ascending order of the Planning Commission s score (2003), including all Phase 2 and Phase 3 districts. 12 Since we now (ex-post) know the number of districts chosen for Phase 2 in each state, we count from top (poorest district) to button (richest district) in each state until the number of nal Phase 2 districts is reached, so that the last district eligible for Phase 2 receives a zero as its normalized score. Doing so for every state we are able to calculate the probability of being selected for Phase 2 conditional on the normalized score. In Figure 1 it is clearly visible that there is a jump at the normalized score (N-Score) equal to zero. Hence, we assume that the constructed cut-o is a good predictor for the placement rule of the MGNREGA RDD Panel Model We instrument exploiting the discontinuity in the assignment rule of the program. Instrument for Program Placement Phase 2 (First Stage): 12 Phase 1 districts are excluded here, since they were already implementing the MGNREGA. Although we would like to nd a placement rule for Phase 1 districts, too, the Planning Commission s ranking does not help much here. 10

11 P hase2 d = c + 1 Nscore d + 2 Nscore 2 d + 3 Nscore 3 d + 1 fnscore d < 1g + u d : (8) 1 fnscore d < 1g = 1 if Nscore < 1: (9) RDD Panel Regression Equation (Second Stage): Y idt = d + 1 D07 t + P hase2d d D07 t + 1 Nscore d D07 t + 2 Nscore 2 d D07 t + 3 Nscore 3 d D07 t + X idt + idt ; (10) The identifying instrument for the regressor P hase2d d D07 t is 1 fnscore d < 1gD07 t : Hence is our causal e ect of interest. This approach allows for the possibility of di erent evolutions according to the poverty score absent the program. The identifying assumption is that the change in the pattern of the evolution is smooth in the poverty score while the probability of assignment of the program is discontinuous at Nscore = 0. Nscore might have to be transformed monotonically in a way that allows the cubic polynomial a su cient t for explaining the probability of being a phase 2 district RDD Cross-Section Model The speci cation for just the case of a cross-section, i.e. for data for a single year, looks very similar. Speci cation for Cross-Section: Instrument for Program Placement Phase 2 (First Stage): P hase2 d = c + 1 Nscore d + 2 Nscore 2 d + 3 Nscore 3 d + 1 fnscore d < 1g + u d : (11) RDD Cross-Sectional Regression Equation (Second Stage): Y id = 1 D07 + P hase2d d D Nscore d D07+ 2 Nscore 2 d D Nscore 3 d D07 + X id + id ; (12) 11

12 5 Results As the MGNREGA is designed as a poverty alleviating program by providing guaranteed employment at the statutory minimum wage, and as many studies estimate positive labor market e ects, we hypothesize that the MGNREGA also has a positive welfare e ect by raising households consumption expenditure (MPCE). And since for the Indian context conventional money-metric poverty measures as the poverty headcount ratio and poverty gap ratio are based on MPCE we expect that the MGNREGA is eventually lowering poverty according to these two measures. In this section, we present our main results of MGNREGA treatment e ects on welfare. For poverty we are using two measures, the headcount ratio (HCR) and the poverty gap ratio (PGR), and we employ both the extreme poverty lines as per the Planning Commission (P.C.) and moderate poverty lines as per the Tendulkar Committee (T.C.). Delving deeper into an analysis of the poorest and most vulnerable households we separately examine consumption levels solely for rural SC/ST households. As pointed out in the sample statistics (Table 3) and by various studies (Sundaram and Tendulkar (2003) for example) it is a well known fact that SC/ST households are considerably poorer than the average rural household. In our sample, every fth rural household is poor (P.C.), whereas in the much smaller SC/ST sample it is every third. Therefore, it is not too far fetched to wonder whether the MGNREGA does have any impact on the welfare of these impoverished households. After all, the majority of workers employed under the MGNREGA belong to SC/ST communities: 62 per cent in and 56 per cent in In this paper we do not present any results for Phase 1 districts, for the simple reason that we do not have any accurate identi cation strategy as yet. Placebo experiments make us believe that in Phase 1 districts of the MGNREGA it is not only employment guarantee at play with an impact on our measures of welfare but also several other programmes which a simple DID approach would identify as MGNREGA e ects. Hence we treat every DID estimates for Phase 1 estimates with caution and refrain from them for the meantime. Similarly, we have not yet found any accurate instrument for the program placement of Phase 1 district, which prevents us from applying the RDD approach for Phase 1 districts. 5.1 DID Results for all rural households and rural SC/ST households Table 5 and Table 6 show our estimation results using di erences-in-di erences (DID) for the years and In these speci cations, households in Phase 2 districts during the year comprise the treatment group whereas those in Phase 3 districts 13 See Drèze and Oldiges (2009) for an analysis of the rst two years. 12

13 are the control group. The year serves as a reference year, in which the latter Phase 2 districts had not yet implemented the MGNREGA yet. We include heterogenous treatment e ects to control for unobservable di erences at the district level. It is important to note that many of the key decisions important for the implementation of the Act are made at the district level. We try to capture such unobservables by interacting the district speci c index score (Planning Commission s Ranking (2003)) with the treatment dummy. We also include state-year-dummies to control for unobservables at the state level, as regional di erences can be substantial in India. Each column from 1 to 5 in Table 5 and Table 6 presents regression results for the ve di erent measures of welfare which are our dependent variables. In the list of independent variables the rst row includes the variable of interest, our treatment dummy for Phase 2 in the year (Y 07 P 2), followed by a number of controls. With regard to the DID estimation for all rural households (Table 5) we neither nd signi cant e ects for a change in the log of real MPCE nor for the two headcount ratios. However, we do estimate negative e ects for the two poverty gap ratios, which are signi cant at the one percent level. Using the extreme poverty lines as set by the Planning Commission the poverty gap is estimated to fall by almost 50 percent in MGNREGA districts, whereas it is about 27 percent for the moderate poverty lines. Moving to our estimations for rural SC/ST households in Table 6 we estimate highly signi cant results for the poverty gap ratios, for real MPCE and the two headcount ratios. For SC/ST households living in rural areas where Phase 2 was implemented in average real MPCE rose by 8.8 percent. At the same time, the probability to live below the extreme (moderate) poverty line fell for SC/ST households on average by 23 (17 percent) in Phase 2 districts. The two measures of the poverty gap ratio also see signi cant declines, indicating that the poorest of the most vulnerable households are moving towards the poverty line. As a falsi cation test of our analysis we run "placebo" - regressions for a year in which Phase 2 had not started. Table 18 portrays our results for an estimation of hypothetical Phase 2 e ects for the year with as the reference year. We do not nd any signi cant e ects here. This supports the results presented in Table 5 and Table 6 and veri es that our DID estimation solely captures the MGNREGA e ect in Phase 2 districts and nothing else which may be at work simultaneously. To sum up, we do see large treatment e ects for the poorest and most vulnerable households - SC/ST households - in rural India, but hardly for all rural households as a whole. 13

14 5.2 RDD Results for all rural households and rural SC/ST households Table 7 and Table 8 present our estimation results using the RDD approach for the years and As for the DID estimations households in Phase 2 districts comprise the treatment group whereas those of Phase 3 districts are the control group. Again, the year serves as a reference year when MGNREGA was not yet implemented in the latter Phase 2 districts. The tables are structured similar to the DID results tables, where each column presents one separate regression for each dependent variable (columns 1-6). As the RDD is a particular IV-approach, we include the rst stage results in column 1. In the rst stage we regress a dummy variable for Phase 2 program placement in on the constructed intrument for the program placement rule and all the other controls (see equation 8). The second stage results are then reported in the remaining columns (2-6) and the predicted value of Y 07 P 2; which is Y 07 P c 2, indicates the coe ent of interest. For the estimations regarding all rural households (Table 7) we do not nd many signi cant MGNREGA e ects on our measurs of welfare. Instead we nd rather surprising results, as we see all coe cients for our poverty measures being positive and the one for the headcount ratio (P.C.) being highly signi cant, too. In accordance with the theory of RDDs, however, these results are similar to the results obtained from using only cross-sectional data for the treatment year , (see Table 21). Regarding our RDD estimates for rural SC/ST households we nd similar results to the previous DID speci cations for SC/ST households. Average consumption as measured by real MPCE increases by 15.5 percent for SC/ST households living in Phase 2 districts. At the same time we calculate declines in all the four measures of poverty, albeit at low signi cance levels. Hence, on average moderate and extreme levels of poverty decrease for SC/ST households in Phase 2 districts. Our cross-sectional estimations for SC/ST households (Table 22), however, are somewhat puzzling for the time being. 5.3 Brief Discussion of Results Summarizing our results we nd that the two distinct methodologies can point to the same direction. In the case of SC/ST households we can conclude that both the panel estimations of DID and RDD point towards declines in the poverty measures and increases in real MPCE. For all rural households as a whole, however, we are not able to come to a satisfying conclusion. Our DID estimates indicate a poverty alleviating e ect for all rural households in Phase 2 districts, but the RDD estimates do not con rm this result. 14

15 On another note, the RDD cross-sectional results for all rural households are in accordance with the RDD panel results. This is not the case when we compare the two RDD estimations of the much smaller sample of SC/ST households. 5.4 Seasonality Casual labour employment and therefore to a large extent also consumption in rural India follow a seasonal pattern. Figure 3 depicts the rising and falling trend of consumption expenditure during the four NSS data collection sub-rounds. Here the peaks are visible for sub-rounds 2 (October to December) and 4 (April to June) as these are the harvest season and sowing season, respectively. Interestingly, MGNREGA employment also follows a seasonal pattern as workers avail of the public work opportunities when the demand for eld labourer is low. As shown by Ravi and Engler (2009) (see Figure 2) MGNREGA employment peaks in the hot Indian summer months of April and May and then steadily declines until it starts to rise again in December after the harvest season. In order to measure a seasonal speci c impact of the MGNREGA we divide our sample into two seasons, where season 1 will include NSS sub-rounds 1 and 2 (July to December) and season 2 the remaining sub-rounds 3 and 4 (January to June). We run the same regressions, DID and RDD, for both the entire sample of all rural households and the small sample of just the rural SC/ST households. In Tables 9 to 16 we present our results in a similar way as before. The only di erence is that we include an additional estimation for real MPCE as a dependent variable. As before, we do not nd poverty alleviating e ects for all rural households neither via the DID nor via the RDD method. Similar to our earlier estimations, however, we estimate such e ects largely for the smaller sample of rural SC/ST households. Using the DID we estimate signi cant declines in the poverty gap ratio for SC/ST households during the lean season. Employing the RDD method we observe that only the coe cients for poverty alleviating e ects in season 2 are statistically signifciant. Hence, we believe that the poverty alleviating e ect of the MGNREGA for the subgroup of SC/ST households in rural India is rather at work when it is most needed. 6 Concluding Remarks We started our analysis of welfare impacts of the MGNREGA arguing that a pure labor market perspective is certainly important in its own right but not a su cient basis to judge the MGNREGA s e ect on rural households quality of life. Considering that higher wages are only a means to an end we explored whether the MGNREGA does translate into higher levels of living. 15

16 From survey reports, anecdotal evidence, and personal experience we know that many workers employed under the MGNREGA do use their public works wages for goods or services which used to be beyond imagination or just una ordable, like bicycles or the education of their children (see Khera (2011)). In this paper, we employ two distinct approaches, a di erences-in-di erences framework (DID) and a regression discontinuity design (RDD) to examine causal e ects of the MGNREGA on the welfare of rural households living in Phase 2 districts during the scal year While we do not nd robust program e ects for the sample of all rural households, we nd signi cant poverty-reducing e ects for the sub-sample of households belonging to scheduled castes and scheduled tribes (SC/ST). For this group of particularly vulnerable households we nd an increase of average consumption of about 15 percent and a decrease in various poverty measures between one fth and one half. In addition, our estimations of the seasonal welfare impact for SC/ST households let us conclude that the MGNREGA does help to reduce the exposure to risk for the most vulnerable households in India. Our results show that the welfare improving e ect is happen during the (dry) spring rather than during the (wet) autumn season. On another note, we do acknowledge that money metric welfare measures may not be su cient in capturing the entire e ect of a workfare programme like the MGNREGA. Being employed under the MGNREGA may not only increase wages or consumption but may also improve the status of women and marginalised groups within a community, and improve commonly-used infrastructure. Although the burden of hard manual labour needs to be accounted for, one can also think of accounting for positive factors besides earnings that enter the utility function of participating households. In fact, it is well established that the workers entitlements provided under the Act have the potential to empower marginalised groups in many ways, which a mere money-metric welfare measure is unlikely to capture. Finally, this paper only examines the rst two years of the Act. Since then, the MGNREGA has grown not only in scale but also in its design, for instance bank payments have become the norm, laborers have become more aware of their entitlements, and administrative processes have become a routine. Hence, a thorough welfare analysis of the latter years impact is certainly worth further research, albeit methodologically di cult. References Alkire, S., and J. Foster (2011a): Counting and multidimensional poverty measurement, Journal of Public Economics, 95(7), (2011b): Understandings and misunderstandings of multidimensional poverty measurement, Journal of Economic Inequality, 9(2),

17 Alkire, S., and M. Santos (2010): Acute Multidimensional Poverty: A New Index for Developing Countries, OPHI Working Paper 38, Oxford Poverty and Human Development Initiative, University of Oxford. Azam, M. (2012): The Impact of Indian Job Guarantee Scheme on Labor Market Outcomes: Evidence from a Natural Experiment, IZA Discussion Paper No Basu, K. (1981): Food for work programmes: Beyond roads that get washed away, Economic and Political Weekly, 16(1), Berg, E., S. Bhattacharyya, R. Durgam, and M. Ramachandra (2012): Can Rural Public Works A ect Agricultural Wages? Evidence from India, CSAE Working Paper WPS/ Deaton, A., and J. Drèze (2002): Poverty and Inequality in India: A Re- Examination, Economic and Political Weekly, 37(36), Deaton, A., and V. Kozel (2005): Data and dogma: the great Indian poverty debate, The World Bank Research Observer, 20(2), 177. Drèze, J. (1990): Famine prevention in India, in The political economy of hunger. Oxford University Press. Drèze, J., and R. Khera (2009): The battle for employment guarantee, Frontline, Vol 26, Issue 1, 3-16 January(1). Drèze, J., and C. Oldiges (2009): How is NREGA doing?, Frontline, Vol 26, Issue 4, February. Dutta, P., R. Murgai, M. Ravallion, and D. Van de Walle (2012): Does India s employment guarantee scheme guarantee employment?, World Bank Policy Research Working Paper, (6003). Economist, T. (2008): Shovelling for their supper, The Economist, August 24. Foster, J., J. Greer, and E. Thorbecke (1984): A class of decomposable poverty measures, Econometrica: Journal of the Econometric Society, pp Himanshu (2007): Recent trends in poverty and inequality: some preliminary results, Economic and Political Weekly, pp Imbert, C., and J. Papp (2012): Labor Market E ects of Social Programs: Evidence from India s Employment Guarantee, Working Paper. Khera, R. (2011): The Battle for Employment Guarantee. Oxford University Press. Lagrange, A., and M. Ravallion (2012): Evaluating Workfare When the Work is Unpleasant: Evidence for India s National Rural Employment Guarantee Scheme, World Bank Policy Research Working Paper, (6272). 17

18 Planning Commission (2003): Report of the Task Force: Identi cation of Districts for Wage and Self Employment Programmes, Discussion paper, Government of India. Ravallion, M. (2011): On multidimensional indices of poverty, The Journal of economic inequality, 9(2), Ravallion, M., G. Datt, and S. Chaudhuri (1993): Does Maharashtra s Employment Guarantee Scheme guarantee employment? E ects of the 1988 wage increase, Economic development and cultural change, 41(2), Ravi, S., and M. Engler (2009): Workfare in Low income Countries: An e ective Way to Fight Poverty? the Case of NREGS in India, Indian School of Business working paper. Hyderabad: Indian School of Business. Sen, A., and Himanshu (2004a): Poverty and Inequality in India: I, Economic and Political Weekly, pp (2004b): Poverty and Inequality in India: II: Widening Disparities during the 1990s, Economic and Political Weekly, pp Sundaram, K., and S. Tendulkar (2003): Poverty in India in the 1990s: Revised results for all-india and 15 major states for , Economic and Political Weekly, pp Tendulkar Committee and others (2009): Report of the Expert Group to review the Methodology for Estimation of Poverty, Planning Commission, India. Zimmermann, L. (2012): Labor Market Impacts of a Large-Scale Public Works Program: Evidence from the Indian Employment Guarantee Scheme, IZA Discussion Paper No

19 A Appendix A.1 Figures Figure 1: Discontinuity for Program Placement of Phase 2 Districts Note: Authors calculation from normalized score (N-Score) based on ranking from Planning Commission (2003). For each N-Score the probability of being assigned the MGNREGA in Phase 2 is calculated. 19

20 Figure 2: Seasonal variation in MGNREGA Employment !" %"#" $" " "&" 10 '"( " 0 Source: Ravi and Engler (2009) 20

21 Figure 3: Seasonal MPCE by NSS-Sub-Round Note: Authors calculation from NSS-CES rounds 61, 62, 63, and 64, respectively. Subround 1: July-September; sub-round 2: Ocotober - December; sub-round 3: January - March; sub-round 4: April - June. Phase-wise averages and total average of real MPCE in prices. A.2 Tables A.2.1 Summary Tables 21

22 Table 1: Phase-wise Means for Rural India between and Phase 1 Log. of MPCE (Rs.) Poverty Headcount Ratio (P.C.) (%) Poverty Gap Ratio (P.C.) (%) Poverty Headcount Ratio (T.C.) (%) Poverty Gap Ratio (T.C.) (%) Observations Phase 2 Log. of MPCE (Rs.) Poverty Headcount Ratio (P.C.) (%) Poverty Gap Ratio (P.C.) (%) Poverty Headcount Ratio (T.C.) (%) Poverty Gap Ratio (T.C.) (%) Observations Phase 3 Log. of MPCE (Rs.) Poverty Headcount Ratio (P.C.) (%) Poverty Gap Ratio (P.C.) (%) Poverty Headcount Ratio (T.C.) (%) Poverty Gap Ratio (T.C.) (%) Observations Note: All measures are de ated to real prices of using the CPIAL, Poverty Headcount Ratio and Poverty Gap (P.C.) are based on poverty lines as set by the Planning Commission. Poverty and Poverty Gap (T.C.) are based on poverty lines as per the Tendulkar Committee. The sample includes all major states, including Jammu & Kashmir and Assam excluding Union Territories and the remaining Northeastern States Sources: Rounds 61, 62, 63, and 64 of NSSO Consumption Expenditure Surveys. Table 2: MGNREGA Information on Sample Districts Phase 1 Phase Total Expenditure (Rs. in crores) 8,685 11,744 3,471 Exp. per HH employed under the MGNREGA (Rs.) 1,626 2,199 1,112 Exp. per HH employed under the MGNREGA (Rs.) 4,187 5,241 3,425 Expenditure on Wages (Rs. in crores) 5,758 7,877 2,460 Wage Exp. per Rural HH (Rs.) 1,078 1, Wage Exp. per HH employed under the MGNREGA (Rs.) 2,776 3,515 2,427 Total Person-days (in crores) Person-days per Rural HH Person-days per HH employed under the MGNREGA HHs employed under the MGNREGA (in millions) Number of Sample Districts Rounded gures. Calculated using Census 2001 data and MGNREGA data posted online at Districts from the Northeastern states, except for Assam s, are not included. 22

23 Table 3: Sample Summary Statistics for Analysis, rural households Full Sample Mean SD Max Min Nonmissing Log. MPCE * Poverty Headcount Ratio (P.C.) (%) Poverty Headcount Ratio (T.C.) (%) Poverty Gap (P.C.) (%) Poverty Gap (T.C.) (%) SC/ST population (%) HH-size N-score SC/ST Sample Log. MPCE * Poverty Headcount Ratio (P.C.) (%) Poverty Headcount Ratio (T.C.) (%) Poverty Gap (P.C.) (%) Poverty Gap (T.C.) (%) SC/ST population (%) HH-size N-score Calculated using NSS Consumption Expenditure Surveys, Rounds 63, 64. MPCE is Monthly per Capita Consumption. Poverty Headcount Ratio (P.C.) and Poverty Gap Ratio (P.C.) are calculated using Poverty Lines as per the Planning Commission. Poverty Headcount Ratio (T. C.) and Poverty Gap Ratio (T.C.) are calculated using Poverty Lines as per the Tendulkar Committee. All measures are de ated to prices using the CPIAL-index. For N-Score, refer to the text. 23