The University of Manchester, UK (

Transcription

1 The Effect of Social Safety Net Programs on the Calorie Consumption of Poor Households in Bangladesh: an Application of Regression Discontinuity Design By Mohammad Mahbubur Rahman December, 2011 Abstract Using the Bangladesh Household Income and Ependiture Survey 2005, this study eamines the effect of social safety net programs on per capita daily calorie consumption for each household. Through these programs, the Bangladesh government and some national and international agencies have been providing food, or cash, or both to poor households in Bangladesh since the famine in I seek to estimate how much these programs affect the well-being of poor households in Bangladesh. Most of the previous studies have produced negative impacts of these programs on calorie consumption, simply computing the raw differential. However, both observable and unobservable characteristics bias this treatment effect. Using fuzzy Regression Discontinuity (RD), I control for these selection effects and find a significant positive impact of the programs. Local Polynomial regression under RD setup with optimal bandwidth of Imbens and Kalyanaraman (2009) produces average treatment effect 843 kilo calorie (Kcal) as the best estimate, which is on average 37 percent of per capita daily calorie consumption for each household. JEL Classification: C21, C31 Key words: Social Safety Net; Regression Discontinuity; Treatment Effect The University of Manchester, UK ( Mohammad.Rahman-2@manchester.ac.uk).

2 1 Introduction Bangladesh is one of the disaster prone countries in the world. Many households in it, specially the poor households with reliance on agriculture, often appear in food insecurity 1 because of entitlement failure resulting from natural disasters such as flood, drought, cyclone and river erosion. Moreover, market failure, seasonal unemployment etc. force them in food insecurity. Therefore, protecting them from any food insecurity is a significant policy issue. Since the famine in 1974 the Bangladesh government and some national and international agencies have been supplying food, or cash, or both free of charge to a number of food insecure households under various programs called Social Safety Net (SSN). Now-a-days, there are many SSN programs running with the primary aim of rescuing poor households from food insecurity in the short-run. These programs target that they have continuous access to nutritionally adequate diet. However, some programs among them have also other objective that households themselves can cope up with any food insecurity in the long-run. Thus, these programs try to improve livelihood of treated households by increasing income earning capacity of them in addition to providing food or cash. In this study, I estimate average treatment effect of a number of SSN programs supported by the data used here on calorie consumption of treated households. I 1 Food insecurity refers to any interruption to continuous access to basic foodstuff. 2

3 focus only on calorie consumption because literature reveals a cyclical relationship among food insecurity manifested in terms of lower calorie intake, reduced human productivity, reduced income generation especially in cases where human capital is the major source of income, and thereby less inflow of calorie, resulting chronic food insecurity [Aromolaran, 2004, Stiglitz, 1976, Strauss, 1986]. Impact analysis of programs aiming to reduce food insecurity in poor countries therefore should focus mainly on calorie consumption. I use Household Income and Ependiture Survey (HIES) 2005, the largest survey in Bangladesh, where about 12% households are treated by at least any one SSN program. However, empirical analyses of treatment effects have not progressed in Bangladesh with the development of econometric evaluation designs. Most of the research papers that carry endogeneity problem in treatment dummy have estimated any specific program s effect using raw differential of outcomes between treatment and control groups [Ahmed and del Ninno, 2002, Matin and Hulme, 2003, Roy et al., 2008]. A few econometrics based evaluation [del Ninno et al., 2001, Ahmed et al., 2007], where outcome variable is regressed on the amount of cash or food received from programs along with income, household size, age of household head and some other household characteristics, can be criticized for carrying two problems [Chen and van der Klauuw, 2 First, the separate effects of the amount of benefits from programs and income on outcome are confounded. Second, the amount of benefits received depends on past income of households and other things, which produces an additional endogeneity 3

4 in the amount of benefits. I use here the so called fuzzy RD design that can remove endogeneity in treatment variable. After controlling endogeneity, this method produces 843 kcal (37%) from a IV form of local polynomial regression, as the best estimate. In the process of producing it, I have checked for different cutoffs rather than considering the official one only to achieve the highest discontinuity in the probability of treatment given the assignment variable. This is justified as the assignment variable is subject to measurement error and there remains the possibility of other covariates than the assignment variable itself that influence the treatment dummy. I have used local polynomial regressions, in addition to local linear regressions, under the optimal bandwidth of [Imbens and Kalyanaraman, 2009] (Hence from now this bandwidth will be called IK optimal bandwidth). Moreover, other covariates are included in local regressions to control for any discontinuity in the epected value of the outcome variable given the assignment variable. This ensures that the treatment effect estimated here is partially identified. Optimal polynomial terms for reduced form regressions are chosen using Bin test. Different models are compared to check for consistency of the result. My final result shows that SSN programs have substantial and significant positive impact on calorie consumption of poor households. This result can be justified that treated households do not change their non-food consumption after being treated. Although they have freedom to spend their benefits from programs on non-food 4

5 items, but they do not do, rather their food ependiture that includes imputed ependiture of benefits increases by 42% which is close to the average treatment effect (37%). On the other hand, epected per capita daily calorie consumption for each household from entitlements of benefits of programs is 859 kcal that is also close to 843. Therefore, the best estimate of the average treatment effect can be considered as a reliable estimate. The rest of the study is organized as follows. SSN programs whose effects are analyzed together are noted in section 2. The background and data are discussed in Section 3. The methodology, RD design, is described in Section 4. Estimated results are analyzed in Section 5. Section 6 discusses the final estimate of average treatment effect, and Section 7 concludes the study. 2 Social Safety Net Programs In Bangladesh, SSN programs are one of the key policy tools for poverty alleviation. These programs are also polictical agenda for governments. Therefore, day by day the number of SSN programs is increasing; up to 2008 that number became 66 (Bangladesh Economic Review (2008)). Money spent in SSN programs is increasing as well. For eample, the Bangladesh government spent around USD 366 million, 7% of its annual budget and 0.9% of real GDP, in (World Bank (2006)) which increased to USD 2450 million, 13.32% of its annual budget and 2.14% of real GDP, 5

6 in (Bangladesh Economic Review (2008)). World Food Program (WFP) and other agencies spend around twice of the government s allocation (Ministry of food and disaster management (2008)). On the other hand, coverage of programs is increasing too, though there is no eact figure about it. In fact, the government and all other agencies have the aim to cover all poor households. Some specific SSN programs whose effects are eamined here are briefly eplained one by one as follows. Vulnerable Group Feeding (VGF). This is first SSN program which was started by WFP during the famine in Now, the Bangladesh government and WFP are jointly operating the program. It supplies only rice to the poorest households and covers all regions of Bangladesh with the highest number of beneficiary households compared to other programs. Community leaders select eligible households for this program and provide them a card called VGF card to get benefits continuously after selection. Every selected household receives one VGF card which is given to household head. Up to 2009, there were around 1,04,67,000 card holding households. 10 kilo grams (kg) rice per month are provided for each card without any condition. This is mainly a continuous program where card holders receive rice every month until their eligibility is epired, but also acts as a transitory program at the time of disaster. When any disaster hits in any area, the program distributes VGF card immediately to the affected households for minimum three months. Old Age Pension (OAP). Under this program, the government provides 250 6

7 Taka per month to poor old aged individuals who are not entitled to formal pension (given to civil servants only) and whose age is at least 65 years. Local government selects eligible individuals. In this case, more than one member in a household can be selected. At present, in Bangladesh the number of individuals with at least 65 years of age is 10 million. Among them, only 2 million (of whom 50% are women) are treated under this program. Vulnerable Group Development (VGD). The program which was started by WFP in 1975 is now being jointly operated by the government, WFP and BRAC, a Non Government Organisation (NGO). It is targeted to woman headed households where no male members are present to earn sufficient money. At the beginning, those women were eligible who lost male member(s) during the liberation war in This program provides loan and training in addition to food to the woman, head of the beneficiary household, without any condition. Each beneficiary household receives 30 kg wheat per month. Moreover, it creates months employment for the eligible women and provides 150 hours of training to them to increase income earning capacity. Up to 2010, the number of beneficiary households was 5,00,000. Beneficiary households are selected by the local government officials. In some cases, VGF has been converted to this program. Test Relief (TR). It is an irregular program, which creates temporary employment in rural areas for poor people through repairing roads, bridges etc. It starts with a formal advertisement. A beneficiary of this program receives 3.5 kg rice daily 7

8 for a maimum of 30 days. Again, local government selects the eligibles. Freedom Fighters Pension (FFP). This program is designed by the government for households whose members participated at the liberation war in In , 1,25,000 households each received 900 Taka (about US $13) per month under this program. Food for Education (FFE). This is a conditional program operated by the government. Since 1993, it has been providing food to the poorest households subject to regular attendance of their children at school. Monthly allocation of per treated household is either a maimum of 20 kg of wheat or 16 kg of rice. Gratuitous Relief (GR). This program which is called transitory program is designed to provide emergency relief to households affected by natural disasters such as flood, cyclone. It was started by the government during the flood in 1998 when 51,200 metric tons of rice were distributed to flood victims. The average amount of rice received by per flood affected household was around 15 kg. Integrated Food Security (IFS). Since 2000, this program has been providing food to vulnerable people in addition to capacity building of them with the financial help from the government and other agencies. Up to 2005, 2,50,000 individuals were treated, where they were entitled to receive 20 kg wheat and 75 taka per month. Food for Work (FFW). Operated by the government, this program generates 100 days employment in the lean season for hard core poor through construction and reconstruction of rural infrastructure. Here, both food (rice and wheat) and 8

9 cash are provided to the beneficiary person. This program is targeted to poor individual, not household. Therefore, more than one member from a household can get benefits. The local government selects eligible persons. The program does not entitle beneficiaries any fied amount of benefit, as it depends on the level of work a beneficiary does. In , about 16,00,000 persons earn, on average, about 3,000 Taka (US $50) monthly under this program. Rural Maintenance Program (RMP). This program that was started in 1983 by CARE, an NGO, provides rural disadvantaged women with employment through regular maintenance of rural roads. Local government selects eligible women and makes a contract with them for four years. As per contract, women have to work 6 days a week for 6 hours a day at lower than the local minimum wage. However, banks disburse 4/5 of their wage and retain the rest of it as savings. Up to 2005, 1,81,000 women were treated under this program. 3 Regression Discontinuity Design 3.1 Identifying the Assignment Variable and the Cutoff We see that different programs have different target groups, e. g. women, old aged, freedom fighters etc. However, all groups belong to the poorest in the country, an ultimate target group of SSN programs. In 2005, the year of this study, the 9

10 Bangladesh government and WFP targeted to cover the 10% poorest households, the eligible group for treatment, by SSN programs (World bank, 2005). Like many other countries, Bangladesh identifies the poor households if per capita monthly income for each household lies below the poverty line income. Households with per capita monthly income below the first decile of the poverty line income are the 10% poorest households. Therefore, per capita monthly income for each household in 2005 is the assignment variable, i, and the first decile of the poverty line is the cutoff point, 0 on it. This i ecludes any amount of benefits received from programs. Households had no idea about the cutoff point, or even they did not know whether they belong to the eligible group or not. Local government with the help of local elected persons (who know about the local people well) selects households for treatment. Therefore, there is no chance of changing treatment though changing their position around the cutoff. This point makes a strong room to apply RD design. If that had happened, unobserved characteristics of households would have been discontinuous around the cutoff, which in turn would have violated the identification condition discussed later here, and would have made treatment effect invalid under RD design. On the other hand, beside i treatment is controlled by other factors such as corruption, measurement error in i, landholding etc. Therefore, the RD design is fuzzy here [Battistin et al., 2009]. 10

11 3.2 Identification Strategies of Treatment Effect Let, Y i is the potential outcome, per capita daily calorie consumption, of household i, and D i is a binary variable of treatment with 1 if household i is treated by at least one SSN program and 0 if household i is not treated by any SSN programs. Now, the potential outcome can be written in the following structural equation form [Angrist and Pischke, 2009, Hahn et al., 2001], Y i = f( i ) + ρ i D i + η i (1) where ρ i is the treatment effect on Y i, and is assumed to be heterogenous, as different (beneficiary) households receive different amount of benefits, and Y 1i = f( i ) + ρ i + η i if D i = 1 Y i = Y 0i = f( i ) + η i if D i = 0 where Y 0i is the potential outcome of household i without treatment eplained by i in f( i ) and other (observed and unobserved) factors in the error term η i, and Y 1i is the potential outcome of household i with treatment where ρ i is added to Y 0i. Since programs administrators select households for treatment eploiting the cutoff, 0, on i, it can be epected that E(D i ) will be discontinuous at it. However, as i is not the only determining factor of treatment, E(D i i ) will not be 11

12 sharply discontinuous at the cutoff. The size of discontinuity will be less than 1 that produces fuzzy RD design. Therefore, E(D i i ) can be written in the following form, g 1 ( i ) if i 0 E(D i i ) = P r(d i = 1 i ) = g 0 ( i ) if i > 0 where g 1 ( 0 ) > g 0 ( 0 ) indicates discontinuity in E(D i ) at the cutoff point, and households just below the cutoff have more chance to be treated than their counterparts just above the cutoff. Now, E(D i i ) can be written in the following functional form, E(D i i ) = g 0 ( i ) + [g 1 ( i ) g 0 ( i )]T i = g 0 ( i ) + πt i where g 1 ( i ) g 0 ( i ) = π, and T i = 1{ i 0 }, an instrument for D i, determines eligibility. Thus, D i is D i = g 0 ( i ) + πt i + ξ 1i (2) where ξ 1i is an error term that captures observed and unobserved factors plus measurement error in i influencing D i. Equation (2) is a reduced form equation, while equation (1) is a structural one. From equation (1), we can not identify the average 12

13 treatment effect, E(ρ i 0 ), as E(D i, η i ) 0 which indicates that D i is an endogenous variable. To identify the average treatment effect at the cutoff, E(ρ i 0 ), (2) needs to substitute into (1), and then the following conditions [Hahn et al., 2001]) are required to satisfy. Condition 1: f( i ) is continuous in i at 0. Condition 2: E(ρ i i ) is continuous in i at 0. Condition 3: D i is independent of ρ i conditional on i near 0 : D i ρ i i. After satisfying these conditions, the following average treatment effect can be identified. E(ρ i 0 ) = lim 0+{E[Y i i = 0 ] E[Y i i = 0 + ]} lim 0+ {E[D i i = 0 ] E[D i i = 0 + ]} = E(ρ i 0 )lim 0+ {g 0 ( i = 0 ) g 0 ( i = 0 + ) + π} lim 0+ {g 0 ( i = 0 ) g 0 ( i = 0 + ) + π} It is not possible to test conditions 2 and 3. Condition 1 can be tested by checking either whether other factors affecting Y i are continuous at the cutoff [Imbens and Lemieu, 2008], or whether density of i is continuous at the cutoff [McCrary, 2008]. In both cases, discontinuity indicates a violation of the condition. However, in the case of constant treatment effect, conditions 2 and 3 always hold, but condition 1 does not. Now, for the time being I assume that treatment effect is constant in ρ for each treated household. In this case, the limiting behaviour of treatment effect will be 13

14 same as we see in heterogenous case, but E(ρ i i ) will be replaced by ρ. However, we can identify ρ running the following IV regression, Y i = f( i ) + ρe(d i i ) + η i (3) On the other hand, ρ can be identified applying Indirect Least Squares (ILS). Under ILS, we need to substitute equation (2) into equation (1). After doing this, we have the following reduced form equation of Y i. Y i = f( i ) + ρ{g 0 ( i ) + πt i + ξ 1i } + η i = f( i ) + ρg 0 ( i ) + ρπt i + ρξ 1i + η i = k( i ) + ρπt i + ξ 2i (4) where f( i ) + ρg 0 ( i ) = k( i ), ρξ 1i + η i = ξ 2i. Now, we can estimate treatment effect, ρ, dividing ρπ, identified from (4), by π, identified from (2). Treatment effect identified in this way is called the Wald estimator, which will not necessarily be the same as in IV from (3). To be eactly equal, g 0 ( i ) is required to be continuous at the cutoff. It can be easily understood from the limiting presentation of average treatment effect. 14

15 3.3 Estimation Methods of Treatment Effect For estimating treatment effect, I run equation (3) (IV method), where the coefficient of E(D i i ) is the average treatment effect, because of two advantages. Any statistical software automatically produces standard error for it that is not in ILS case, and there is no need to assume that g 0 ( i ) is continuous at the cutoff. In addition, I run equations (2) and (4) to know the sizes of discontinuities in E(D i i ) and E(Y i i ) (π and ρπ) respectively. In spite of semi-parametric forms of them, running OLS regression is a practice taking different functional forms such as linear, quadratic, cubic in i. However, along with OLS regressions I run local linear [Hahn et al., 2001] and local polynomial [Porter, 2003] regressions at the IK optimal bandwidth. 4 Background and the Data Bangladesh Bureau of Statistics has been conducting the HIES, a cross section survey, on Bangladeshi households in every five years since 1991 with the financial help from the World Bank (WB). It collects data about households and their members characteristics, households income, ependiture, food and non-food consumption, agricultural production, and many other household related information from all regions of Bangladesh. In this study, the HIES 2005, the latest available one, where the number of households is 10,070, is used as it contains sufficient information 15

16 about SSN programs compared to the previous HIESs. There is one section about SSN programs in the questionnaire of HIES 2005 where questions are designed to collect information such as whether households are getting treatment under any of SSN programs, if they get treatment then for how long, reasons of getting treatment, how much cash or food they receive from the program, have they paid any bribes to be selected for treatment, and from whom they collect cash and food. Therefore, I estimate treatment dummy, D i, from the first question where 1,226 households (12% of total households) are treated by at least one SSN program. In this data, most of the treated households are under any one of the SSN programs, but only 48 households receive treatment from more than one. Table 1 shows the percentage of treated households participating in different SSN programs, where VGF program has the highest participation rate. Table 1: Partcipation Rate (%) of Treated Households in SSN Programs Program Partcipation Rate (%) VGF OAP VGD TR FFP FFE 8.16 GR 0.57 IFS 0.33 FFW 0.24 RMP 0.08 Others 0.82 Households food and non-food consumptions are collected using recall method 16

17 where households have to recall a specific period s item wise consumptions with ependitures on them. In this way, households provide information of last two weeks food consumptions and ependitures. For non-food items, some are provided in monthly and some are in yearly basis. In the case of food, unit of measurement is different for different items, such as kilo gram for rice, vegetables etc., litre for oil, milk etc., number for egg, banana etc. To estimate calorie consumption of households, all food items need to be converted into a single unit of measurement, gram, as calorie content of each food item available to Food and Nutrition department, Dhaka University, Bangladesh is for per gram consumption. Therefore, the following procedure is followed to estimate Y i, Y i = k (C ik c k ) H i where C ik is the amount of daily consumption in gram on food item k (from all sources such as own production, bought, gifts, SSN programs) by household i, c k is the per gram calorie content in kilo calorie (kcal) of food item k, and H i is the household size of household i. The assignment variable, i, of household i that ecludes any benefits received is estimated adding all types of monthly and other income in different frequencies converted to monthly and then divided by household size. This variable contains measurement error for not being collected from program administrators, like any 17

18 other survey data. On the other hand, in 2005 the poverty line income in Bangladesh was 860 taka (Bangladeshi currency), and therefore, 290 taka is the first decile of it on i. Figure 1 presents them under dotted vertical lines through truncated kernel density of i where values above 1000 are ignored. To note, I take a sample with i 1, 000 as a regression sample, as it is necessary to concentrate near the cutoff for estimating the treatment effect. In addition, I eamine other points near 290 to identify the biggest discontinuity in E(D i i ) with statistical significance to make T i as a relatively stronger instrument. Some nearest points of 290 are also shown in the Figure 1 by dotted vertical lines. Density of st Decile of Poverty Line Figure 1: Truncated Kernel Density of 1000, HIES

19 With the fuzzy design, the size of discontinuity of E(D i i ) can be bigger at any other points than at the known cutoff, due to measurement error in i. Therefore, if treatment effect is estimated at an unknown point with the biggest size of discontinuity in E(D i i ) with statistical significance, the number of compliers will increase, and thus, treatment effect estimation will be robust. Porter (2003) is the first to indicate that the treatment effect can be estimated in this way. Some other covariates called household characteristics that are determined prior to D i and affect D i or Y i or both are used for doing a test of identification condition 1. Descriptive statistics of all variables used in this study are presented in Table 2. Comparing means between treated and untreated households, it is clear that all characteristics of households differ between two groups both in the whole sample and regression sample. Although the gaps decrease in the case of latter sample, however, it is dangerous in both sample cases to use any other technique such as matching estimators where it is assumed that selection for treatment is determined by observed characteristics. Moreover, it will be totally misleading if anyone use raw differential to estimate treatment effect. The table shows that raw differential in Y i is negetive in the whole sample case, and just positive in regression sample case. Both are unreliable as a treatment effect, as the eperiment is not randomized. Previous studies can be criticised for not considering this issue. Whatever, these household characteristics can be equal between treated and untreated groups near the cutoff point, which can be known by tests of identification condition 1. With 19

20 holding this condition, randomized eperiment will eist near the cutoff, thereby RD design will be valid. Variable Table 2: Descriptives (Mean) Whole Sample Regression sample Treated Untreated Treated Untreated Y D T = 1( 0 = 290) Rice price per kg (Taka) Head s education (years of schooling) Se of head (1 if male, 0 otherwise) Household size Number of adult (18-60 aged) Location (1 if rural, 0 otherwise) Landholding (decimal) Total observations 1,226 8, ,172 To have an idea about discontinuity at 290 from raw data, in Figure 2 I plot mean values of D i at bins on i where bin sizes are 10 and 60. This type of graph is useful to understand discontinuity when data are noisy around the cutoff [Lee and Lemieu, 2010]. The graph shows that the size of discontinuity (ˆπ) for bin width 60 is with standard error which is estimated comparing two closest bins means of 290. The discontinuity size is bigger for bin width 10 than that for bin width 60. Thus, there is an indication of having discontinuity in P r(d i = 1 i ) at 290. Similarly, Figure 3 shows discontinuity in E(Y i i ) where for bin width 60 the size of discontinuity ( ρπ) ˆ is 89 with standard error 48. However, 20

21 for bin width 10, it is quite lower. Now, if 89 is divided by 0.094, treatment effect will be 947 (s.e. 680) which is called local Wald estimator [Hahn et al., 2001]. This effect is 47% which is substantial compared to previous studies. However, I can not rely on this estimator because of boundary bias problem [Hahn et al., 2001]. Pr(D=1 ) π = 0.094(0.044) Bin Width 10 Bin Width 60 Figure 2: Estimating P r(d = 1 ) by bin width ( 0 = 290) E(Y ) ρπ = 89(48) Bin Width 10 Bin Width 60 Figure 3: Estimating E(Y ) by bin width ( 0 = 290) 21

22 5 Estimated Results Using regression sample, OLS regression of equation 2 and 4 and IV regression of equation 3 with quartic form of i in g 0 ( i ), k( i ) and f( i ) are run to consider 250, 260, 270, 280, 285, 290 and 300 as cutoff points on i. Predicted values are plotted in Figure 9 and 10 of appendi A. Table 3 presents discontinuity parameters (ˆπ and ρπ) ˆ and treatment effect parameter (ˆρ) at these cutoff points. The value of ˆπ is seen here as the biggest at 285 which indicates that T i will be a stronger instrument at this cutoff point than any other points. On the other hand, t-ratio (ratio of estimate and standard error) is also the highest at this point not only for ˆπ but also for treatment effect, ˆρ. Therefore, to estimate treatment effect 285 can be considered as the cutoff point rather than the official cutoff, 290. However, final decision can be made after local regressions. Cutoff Table 3: Estimates from Quartic Regressions OLS ˆπ OLS ρπ ˆ (0.042) 63.54(53.36) 1,965.80( ) (0.041) 95.15(51.50) 1,425.70( ) (0.039) (49.58) 1,011.30(605.50) (0.038) (48.11) 1,226.10(695.00) (0.038) (47.44) (361.80) (0.037) 92.81(46.71) (478.10) (0.036) 63.12(45.84) (653.10) Note: Robust standard errors are in parentheses IV ˆρ Table 4 shows same estimates, but now local linear regression is run in all cases 22

23 Table 4: Discontinuities from Local Linear Regressions with Different Bandwidths Bandwidth ˆπ Cutoff=270 Cutoff=280 πρ ˆ ˆρ ˆπ πρ ˆ Cutoff 0.06(0.04) 116(46) 2053(1685) 0.05(0.04) 134(44) 2494(1980) (0.07) 140(71) 2277(2713) 0.05(0.06) 141(67) 2655(3389) (0.09) 156(101) 1232(1207) 0.15(0.09) 129(97) 888(866) (.013) -72(152) -841(2172) -0.11(0.13) 77(130) -674(1287) Bandwidth ˆπ Cutoff=285 Cutoff=290 πρ ˆ ˆρ ˆπ πρ ˆ Cutoff 0.11(0.04) 114(43) 1082(565) 0.09(0.04) 87(42) 942(590) (0.06) 124(66) 766(500) 0.08(0.06) 76(65) 959(1091) (0.09) 39(97) 107(266) 0.19(0.09) -23(95) -119(499) (0.12) 40(133) 177(607) 0.08(0.12) -30(134) -394(1810) Note: Robust standard errors are in parentheses. ˆρ is estimated from IV regression. Cutoff under bandwidth means that the size of bandwidth is equal to the size of a cutoff. ˆρ ˆρ with bandwidth 30, 60, 120 and the size of cutoff considering 270, 280, 285, 290 as cutoffs. Other cutoffs are ignored due to poor estimates as shown in Table 3. Now, at 285 estimates are showing consistent behaviour even in narrower bandwidths. As bandwidth decreases, ˆπ increases up to 60 bandwidth, ρπ ˆ decreases but remains positive, and ˆρ decreases as well but remains positive. ˆπ is significant in all bandwidth cases. On the other hand, at other cutoff points ˆρ becomes negetive in narrower bandwidth because of either ˆπ becomes negetive or ρπ ˆ becomes negetive. Thus, 285 is chosen as the cutoff for further analyses. IK optimal bandwidth for local linear regressions at the cutoff, 285, is 289, where I run some polynomial regressions along with local linear reegression. See Table 5 where same parameters are reported. Now, I run regressions incorporating 23

24 Table 5: Discontinuities from Local Regressions at the Cutoff 285 Model Type ˆπ πρ ˆ ˆρ Local Linear [0.04] [43] [509] Local Linear with Covariates [0.04] [41] [522] Local Quadratic [0.04] [48] [300] Local Quadratic with Covariates [0.04] [47] [336] Local Cubic [0.05] [57] [689] Local Cubic with Covariates [0.05] [55] [845] Local Polynomial a [0.07] [62] [422] Local Polynomial with Covariates [0.07] [60] [475] Note: Robust standard errors are in parentheses. All models are run within IK optimal bandwidth, 289. Household characteristics mentioned in Table 2 are used in models with covariates. ˆρ is estimated from IV regression. IV regressions with covariates uses covariates as instruments as well as eogenous variables. a Using Bin test [Lee and Lemieu, 2010], optimal polynomial terms are chosen as 8 in D i regression and 4 in Y i regression. other covariates listed in Table 2 too. This will control other covariates effects on outcome variable near to the cutoff, and therefore will produce valid treatment effect in case the first identification condition (Condition 1) is not satisfied [Lee and Lemieu, 2010]. I also choose optimal polynomial terms for D i and Y i regressions using Bin test [Lee and Lemieu, 2010]. Optimal polynomial term in D i regression is 8 and that in Y i regression is 4, and thus I have named them local 24

25 polynomial (regression). In this case, ˆπ contains the highest value that indicates the strength of T i instrument. Treatment effect, ˆρ, is here 623 kcal which increases to 843 kcal if other covariates are just incorporated here. I choose 843 kcal as my final estimate which is significant at 10% level. Figure 4 and 5 show predicted lines from these polynomial regressions along with bin means under 30 bin size. E(D ) Figure 4: Estimating E(D ) from polynomial regression mentioned in Table 5 with mean values of D under bins of 30 bin size ( 0 = 285) To test the first identification condition (Condition 1), other covariates are checked whether they are discontinuous at the cutoff (see Table 6). From local linear regressions under IK optimal bandwidth, it is seen that none of these covariates 25

26 E(Y ) Figure 5: Estimating E(Y ) from polynomial regression mentioned in Table 5 with mean values of Y under bins of 30 bin size ( 0 = 285) is significantly discontinuous at 285, which indicates that Condition 1 is satisfied. On the other hand, density of the assignment variable is seen as continuous in local quartic regression (see Figure 6). Histogram plot also suggests that density of i is not discontinuous at 285. Only local linear regression suggests discontinuity, but it has very bad fit. Thus, it can be concluded that Condition 1 is satisfied at 285. However, I chosen treatment effect 843 kcal as a final estimate from the local polynomial regression with covariates, as it controls any amount of discontinuities of covariates at

27 Table 6: Discontinuities of Covariates from Local Linear Regressions at the Cutoff 285 Variable Discontinuity Robust standard error Rice price per kg (Taka) Head s Education (years of schooling) Se of head (1 if male, 0 otherwise) Household size Number of adult (18-60 aged) Location (1 if rural, 0 otherwise) Landholding (decimal) Note: Local linear regression of each variable is run on and T at 285 cutoff, using IK optimal bandwidth, 289. For each variable, coefficient of T is reported here. Observation X Local Linear Local Quartic Histogram Figure 6: Estimating density of i from local (linear and quartic) regressions suggested by McCrary (2008) ( 0 = 285) 27

28 6 Discussion Let us consider per capita daily epected calorie consumption from SSN programs had the beneficiary households consumed all their benefit entitlements in food. In column (2) of Table 7, program wise participation rates are presented which are slightly different from that in Table 1. Here the number of individuals who are recipients of benefits from a program on behalf of their households is divided by the total number of beneficiary individuals rather than the total number of beneficiary households, in all programs listed here. These participation rates are used as weights to estimate per capita daily epected calorie consumption of benefits received from programs. Column (3) of the Table shows per capita monthly rice or wheat entitlement of a beneficiary household under a program. To estimate it, household size is assumed 5. Cash entitlements are also converted into rice using market price of rice in Column (4) shows calorie equivalence of per capita daily benefits entitlement for each household under a program. To estimate this column, per gram rice or wheat is set as equivalent to 3.69 kcal as the converter in Bangladesh. Thus, on average a beneficiary household can be epected to take 859 kcal per capita per day from benefits if it receives them as per entitlements and does not spend them on non-food items. Beneficiary households have freedom to spend programs benefits on non-food items. However, they are vulnerable group who face severe food insecurity. It is 28

29 Table 7: Epected Calorie Consumption from SSN Programs (2) (3) (4)=(2)/100*((3)*1000*3.69/30) Program Partcipation Per capita monthly Per capita daily epected Rate (%) rice/wheat entitlement (kg) calorie consumption (kcal) VGF OAP VGD TR FFP FFE GR IFS FFW RMP Others 0.75 Total likely that their first priority will be to fulfill their calorie deficiency, and therefore they will employ all or most of benefits received from programs on food consumption. In Figure 7, per capita monthly food ependiture for each household significantly increases by 14 percentage points at 285 cutoff. Predicted line here comes from quartic regression of log of per capita monthly food ependiture for each household on T i and quartic form of i under IK optimal bandwidth. Food ependiture here includes all types of food consumption. Those foods that are not bought from markets, such as own production, transfer receipts, are also converted into ependiture using market price of food in However, IV regression says that average treatment effect on a log of per capita monthly food ependiture for each household is 42 percent, which is quite close to the average treatment effect on the outcome variable 29

30 (37 percent). On the other hand, Figure 8 shows no significant discontinuity in the log of per capita monthly non-food ependiture for each household at the cutoff 285. The predicted line is generated from quartic regression as in food ependiture case. Therefore, as households do not change their non-food ependitures after getting treatment, average treatment effect on the outcome variable 843 kcal which is closer to the epected calorie consumption of benefits 859 kcal, can be accepted as a credible estimate. Log(Per Capita Food Ependiture for Each Household) Figure 7: Estimating predicted value of log of per capita monthly food ependiture for each household from quartic regression ( 0 =285) 30

31 Log(Per Capita Non-food Ependiture for Each Household) Figure 8: Estimating predicted value of log of per capita monthly non-food ependiture for each household from quartic regression ( 0 =285) 7 Conclusion In this study, using HIES 2005 I estimate the average treatment effect of a number of SSN programs in Bangladesh on calorie consumption of food insecure households with the application of fuzzy RD design. On average, their per capita daily calorie consumption increases by 843 kcal (37 percent) after being treated under these programs. This finding is substantially different from the treatment effects estimated by the previous studies. The statistically insignificant and/or even negative treatment effects of Bangladesh SSN programs estimated in eisting literatures are mainly because of the endogeneity problem with the treatment variable. Even the HIES 2005 yields a negative raw differential of the outcome variable. Fuzzy RD design removes 31

32 the endogeneity problem, and therefore this data produces such a substantial and significant result. The size of the effect estimated here can be considered as reliable, comparing the per capita daily epected calorie consumption from SSN programs had the beneficiary households consumed all their benefit entitlements (859 kcal), as it is seen that beneficiary households do not change their non-food ependiture after treatment. The technicalities of the study establish the estimated treatment effect as robust. I checked for different cutoffs rather than considering the official one only to achieve the highest discontinuity in the probability of treatment given the assignment variable. This is justified as the assignment variable is subject to measurement error and there remains the possibility of other covariates than the assignment variable itself that influence the treatment dummy. I used local polynomial regression including covariates to control for any discontinuity in the epected value of the outcome variable given the assignment variable. This ensures that the treatment effect estimated here is identified subject to Condition 1. Optimal polynomial terms for reduced form regressions are chosen using Bin test. Different models are compared to check for consistency of the result. Finally, the significant treatment effect of the SSN programs in Bangladesh estimated here has its own policy implication. The programs, though are limited in terms of the size of benefits, the money spent under them consists a significant part in government budget. The positive effect of the programs thereby provides 32

33 strong ground to increase, or at least maintain the current allocation and justifies the programs to be some effective policy tools for poverty alleviation. 33

34 References [Ahmed and del Ninno, 2002] Ahmed, A. U. and del Ninno, C. (2002). The food for education program in bangladesh: An evaluation of its impact on educational attainment and food security. Food Consumption and Nutrition Division Discussion Paper 138, International Food Policy Research Institute. [Ahmed et al., 2007] Ahmed, A. U., Quisumbing, A. R., and Hoddinott, J. F. (2007). Relative efficacy of food and cash transfers in improving food security and livelihoods of the ultra-poor in bangladesh. Technical report, International Food Policy Research Institute. [Angrist and Pischke, 2009] Angrist, J. D. and Pischke, J. S. (2009). Mostly Harmless Econometrics: An Empiricists Companion. Princeton University Press. [Aromolaran, 2004] Aromolaran, A. B. (2004). Intra-household redistribution of income and calorie consumption in south-western Nigeria. Discussion Paper 890, Yale University Economic Growth Center. [Battistin et al., 2009] Battistin, E., Brugiavini, A., Rettore, E., and Weber, G. (2009). The retirement consumption puzzle: Evidence from a regression discontinuity approach. American Economic Review, 99(5): [Chen and van der Klauuw, 2008] Chen, S. and van der Klauuw, W. (2008). The work disincentive effects of the disability insurance program in the 1990s. Journal of Econometrics, 142(2): [del Ninno et al., 2001] del Ninno, C., Dorosh, P. A., Smith, L. C., and Roy, D. K. (2001). The 1998 floods in bangladesh: Disaster impacts, household coping strategies, and response. Research Report 122, International Food Policy Research Institute. [Hahn et al., 2001] Hahn, J., Todd, P., and van der Klaauw, W. (2001). Identification and estimation of treatment effects with a regression-discontinuity design. Econometrica, 69(1): [Imbens and Kalyanaraman, 2009] Imbens, G. and Kalyanaraman, K. (2009). Optimal bandwidth choice for the regression discontinuity estimator. Unpublished manuscript, Department of Economics, Harvard University. [Imbens and Lemieu, 2008] Imbens, G. and Lemieu, T. (2008). Regression discontinuity designs: A guide to practice. Journal of Econometrics, 142: [Lee and Lemieu, 2010] Lee, D. S. and Lemieu, T. (2010). Regression discontinuity designs in economics. Journal of Economic Literature, 48(2): [Matin and Hulme, 2003] Matin, I. and Hulme, D. (2003). Programs for the poorest: Learning from the IGVGD program in Bangladesh. World Development, 31(3):

35 [McCrary, 2008] McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity design: A density test. Journal of Econometrics, 142(2): [Porter, 2003] Porter, J. (2003). Estimation in the regression discontinuity model. Unpublished Manuscript, Department of Economics, Harvard University. [Roy et al., 2008] Roy, S. K., Bilkes, F., Islam, K., Ara, G., Tanner, P., Wosk, I., Rahman, A. S., Chakraborty, B., Jolly, S. P., and Khatun, W. (2008). Impact of pilot project of rural maintenance programme (RMP) on destitute women: CARE, Bangladesh. Food and Nutrition Bulletin, 29(1): [Stiglitz, 1976] Stiglitz, J. E. (1976). The efficiency wage hypothesis, surplus labor, and distribution of income in ldcs. Oford Economic Papers, 28(2): [Strauss, 1986] Strauss, J. (1986). Does better nutrition raise productivity? Journal of Political Economy, 94(2):

36 Appendi A Pr(D=1 ) 0.5 Cutoff=250 Pr(D=1 ) 0.5 Cutoff= Pr(D=1 ) 0.5 Cutoff=270 Pr(D=1 ) 0.5 Cutoff= Pr(D=1 ) 0.5 Cutoff=285 Pr(D=1 ) 0.5 Cutoff= Pr(D=1 ) 0.5 Cutoff= Figure 9: Estimating P r(d = 1 ) using quartic regression and bin means (Bin width=30) 36

37 E(Y ) Cutoff= E(Y ) Cutoff= E(Y ) Cutoff= E(Y ) Cutoff= E(Y ) Cutoff= E(Y ) Cutoff= E(Y ) Cutoff= Figure 10: width=30) Estimating E(Y ) using quartic regressions and bin means (Bin 37