ESSAYS ON POVERTY, MICROFINANCE AND LABOR ECONOMICS by SANDARADURA INDUNIL UDAYANGA DE SILVA, B.Sc., M.A.

A DISSERTATION IN ECONOMICS Submitted to the Graduate Faculty of Texas Tech University in Partial Fulfillment of the Requirements for the Degree of DOCTOR OF PHILOSOPHY Approved Masha Rahnama
Chairperson of the Committee

Thomas Steinmeier Robert McComb

Accepted John Borrelli

Dean of the Graduate School

August, 2006

Copyright 2006, Sandaradura Indunil Udayanga De Silva

ACKNOWLEDGEMENTS I would like to extend my gratitude and thanks to my dissertation committee chair, Dr. Masha Rahnama, for his guidance throughout my work. I also wish to extend my sincere gratitude to other members of my committee, Dr. Thomas Steinmeier and Dr. Robert McComb for their helpful comments, discussion and guidance. Deep appreciation goes to my parents. Without their encouragement, devotion and sacrifices, my education would not have reached this level. Finally, I owe a debt of gratitude to Dr. Joseph King, Chairman of the Department of Economics at Texas Tech University, for providing continuous encouragement during my Ph.D. studies.

ii

TABLE OF CONTENTS

ACKNOWLEDGEMENTS ii LIST OF TABLES... LIST OF FIGURES.. CHAPTER 1. INTRODUCTION... 2. A ROBUST POVERTY PROFILE FOR SRI LANKA IN A MULTIVARIATE FRAMEWORK 2.1 Introduction.. 4 4 1 v vii

2.2 Data and Methodology.. 5 2.3 Unconditional Poverty Profile: Cross Tabulations 6 2.4 Conditional Poverty Profile: Marginal Effects.. 14 2.5 Conclusion. 24 3. EVALUATING THE IMPACT OF MICROFINANCE ON SAVINGS AND INCOME: QUASI-EXPERIMENTAL APPROACH USING PROPENSITY SCORE MATCHING 34 3.1 Introduction. 34 3.2 Microfinance Institutions and Impact Studies. 37 3.3 Econometric Methodology and Data... 53 3.4 Results.. 66

iii

3.5 Conclusion 71 4. RETURNS TO EDUCATION IN SRI LANKA: QUANTILE REGRESSION ANALYSIS 103 4.1 Introduction 103 4.2 Human Capital Framework, Signaling and the Returns To Schooling.. 4.3 Econometric Methodology. 4.4 Empirical Results 4.5 Conclusion.. 105 111 114 118

REFERENCES. 134

iv

LIST OF TABLES

2.1 2.2 2.3 2.4 2.5 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8

Unconditional Poverty Profile (cross tabulations) Inequality Indices. Variable definitions and means Logit regression estimates Quantile and OLS regression estimates.... Social Transfers and social expenditure,1999.. Income Transfer Component

26 27 28 29 30 73 74

Bottom 20th percentile probit model for the propensity score. 75 20th-40th percentile probit model for the propensity score 40th-60th percentile probit model for the propensity score 60th-80th percentile probit model for the propensity score 80th-100th percentile probit model for the propensity score.. Matching quality indicators (covariate balancing) for the 20th percentile. 80 76 77 78 79

3.9

Matching quality indicators (covariate balancing) for the 20th-40th percentile.. 83

3.10

Matching quality indicators (covariate balancing) for the 40th-60th percentile. 86

LIST OF TABLES 3.11 Matching quality indicators (covariate balancing) for the 60th-80th percentile.. 89 3.12 Matching quality indicators (covariate balancing) for the 80th-100th percentile 3.13 3.14 3.15 4.1 4.2 4.3 4.4 4.5 4.6 4.7 Individuals lost due to common support requirement.. 92 95

Impact of microfinance on household savings.. 96 Impact of microfinance on household income 97 Cross Country evidence on the Returns to Schooling (year-1995) 120 Quantile and OLS regression estimates (model 1-full sample).. 121 Quantile and OLS regression estimates (model 1-Sinhala) 122 Quantile and OLS regression estimates (model 1-Tamil).. 123 Quantile and OLS regression estimates (model 2-full sample).. 124 Quantile and OLS regression estimates (model 2-Sinhala).125 Quantile and OLS regression estimates (model 2-Tamil)126

vi

LIST OF FIGURES 2.1 2.2 2.3 3.1 Cumulative poverty gap curves 31 Lorenz curves 32

Ordinary Least Squares and Quantile Regression Estimates 33 Histogram of the estimated propensity score (bottom 20th percentile) 98

3.2

Histogram of the estimated propensity score (20th-40th percentile). 99

3.3

Histogram of the estimated propensity score (40th-60th percentile). 100

3.4

Histogram of the estimated propensity score (60th-80th percentile). 101

3.5

Histogram of the estimated propensity score (80th-100th percentile) 102

4.1 4.2 4.3 4.4 4.5 4.6

Returns to Schooling in Europe, Men and Women 127 OLS and quantile regression estimates (model 1-full sample)....... 128 OLS and quantile regression estimates (model 1-full Sinhala).. 129 OLS and quantile regression estimates (model 1-full Tamil) 130 OLS and quantile regression estimates (model 2-full sample)... 131 OLS and quantile regression estimates (model 2Sinhala) 132

vii

4.7

OLS and quantile regression estimates (model 2Tamil)... 133

viii

Copyright 2006, Sandaradura Indunil Udayanga De Silva.

CHAPTER 1 INTRODUCTION

This dissertation presents three essays examining some of the issues concerning poverty, microfinance and returns to education. The first essay examines the micro-level determinants and correlates of poverty, and presents a poverty profile for Sri Lanka. This is the first study that examines the probable determinants and correlates of Sri Lankan poverty in a multivariate framework employing both logit and quantile regressions. The most appealing feature of quantile regression is that it does not impose constant parameters over the entire distribution. The empirical findings are broadly encouraging. The estimation results show poverty is more acute in rural areas than in urban areas. Furthermore, since 76% of the population lives in rural areas, the rural shares in the total composition of poverty are more higher. However, the degree of inequality is much more greater in urban areas, compared to the rural sector. Therefore priority need to be given for policy initiatives aimed at reducing poverty in rural areas, while recognizing the need to tackle urban inequality. Results also indicate that the pay off to smaller families is higher, and larger families are more likely to be poor. . The study also found evidence to support the hypothesis of the feminization of poverty. Female-headed households are significantly worse off compared to male-headed households, especially in poorer households. Household head's education level also had an instrumental effect in determining the vulnerability to poverty. Poverty incidence declined monotonically with years of education. The finding of a strong correlation between poverty and children,

suggests that the presence of children need to be considered as a strong indicator candidate for targeting. The beauty of enacting poverty alleviation programs through targeting key poverty indicators is that both administrative costs and leakage can be lowered. These findings indicate the importance of a set of policies which are super propoor, namely increasing school enrollment and achievement, effective family planning programs to reduce the birth rate and dependency load within households, and granting priorities for specific cohorts (children, elderly, rural and female headed households) in targeted interventions. The second essay applied recent advances in propensity score matching to assess the impact of microfinance on household per-capita income and savings. Microfinance for the poor has become a focus of attention in the Sri Lankan development community over the last several years. The microfinance revolution has built on innovations in financial intermediation that lowers the cost of risks of lending to poor households. Using Data from a nationally representative household survey, this paper analyzed the impact of participating in microfinance on household per-capita income and savings employing a quasi-experimental approach. Since a baseline survey or randomization are not feasible options in this case, the study is well suited to matching methods. There are several attractive features associated with propensity score matching, including the potential to allow for heterogeneous impacts, while optimally weighting observed characteristics when constructing a comparison group. The technique is well suited due to its flexible (non-parametric) nature, not imposing exclusion restrictions or ad hoc assumptions about the functional form of impacts. The method eliminates selection bias due to observable differences between program participants and non-participants. Although a rich data set

has been used, permitting to match on a wide range of household characteristics, the likelihood always remains of latent unobserved factors being correlated with program participation and outcome variables. Results suggests that program participation increases with household size, being a Sinhalese, living in a rural area, and employed as a casual worker or self-employed. Overall program participants benefit incidence is indeed propoor. With respect to household per-capita savings, program participation definitely has a positive impact for all low-income households. Household per-capita savings are

significantly higher on average for participants of microfinance than for observationally identical non-participants. However, the overall results are rather discouraging for household per-capita income, since the impact estimates are negative for all estimated income quintiles. Finally, the principle message that emerges form the study is; there are quantitatively non-negligible, average gains from microfinance on household savings, especially for the poor. The third essay investigates the returns to education in Sri Lankan labor market using the latest Consumer Finance and Socio-economic Survey. This essay employs the quantile regression technique for each conditional quantile wage group rather than mean regression analysis used in most labor market analysis. Quantile regression results suggest that returns to education are positive and significant across all quantiles. However, a comparison of wage returns to education between ethnic groups reveals that returns are higher for Sinhalese workers than for Tamil Workers.

CHAPTER 2 A ROBUST POVERTY PROFILE FOR SRI LANKA IN A MULTIVARIATE FRAMEWORK

2.1. Introduction Notwithstanding its achievements in human development, poverty in Sri Lanka is still a pervasive phenomenon. According to the World Bank (2002), "Sri Lanka's success in reducing income poverty is less noteworthy, especially when contrasted with that of East Asian countries that were at comparable levels of development only a few decades ago". During the past decade, there has been a renewed sense of urgency for poverty reduction strategies in Sri Lanka by the government, non-governmental organizations and international donors. The design of effective poverty reduction strategies requires the knowledge of who are the poor, where they live and what their socio-economic profile is. Ideally, policy makers and program designers would like to know- 1) the income generating activities of the poor (e.g., whether they are self employed, earning wages, traders, microentrepreneurs, etc.), 2) to what degree do the poor have access to services and infrastructure (e.g., piped water electricity sanitation facilities, etc.), 3) housing conditions (e.g., owns a house, lives in a shanty or line room, etc.). The current paper zeroes on this aspect, with the objective of identifying the poor using a microeconometric approach. The specific questions addressed in this chapter are: firstly, is poverty more prevalent among female-headed households than among male headed households? Typically in developing countries feminization of poverty occurs mainly due to women

being relatively less educated and also as a result of discrimination in the labor market. Grootaert and Braithwaite (1998), finds that female headed households have a higher probability of being in poverty than their male headed counterparts. On the other hand, Szekely (1998) found no evidence claiming that female-headed households are more likely to be in poverty. The second question examined is that whether and what levels of education contributes positively to higher living standards? Findings of Schultz (1988) and Psacharapoulous (1985) indicate that there is a positive relationship between education and higher earnings. The third major question addressed is whether households in rural and estate (plantation) sectors face a higher probability being in poverty. There is a vast amount of literature demonstrating that poverty is a predominantly rural phenomena in developing countries, World Bank (1990). The final key question examined is whether the occupation of the household head shows a significant association with the likelihood of being in poverty. More specifically, an examination of how living standards vary across households in salaried employment, casual wage and business. The rest of the chapter is organized as follows. The next section briefly describes the data and methodology. Section 2.3 presents the results of the unconditional poverty profile (cross-tabulations) and some sectorial inequality indices. Sections 2.4 presents the conditional poverty profile based on logit and quantile regression estimates. Section 5 concludes. 2.2 Data and Methodology The study is based on the latest Sri Lanka Integrated Survey (SLIS), commissioned by the World Bank in 1999/2000. The survey is nationally representative

and consists of 7500 households and a 34,330 individual population. The SLIS is unique in the sense that it is the first integrated survey that covered the entire island. The survey collects information on a broad range of topics including demographic characteristics, household income and expenditure, literacy and education, household amenities and employment. The single monetary indicator of household welfare (or living standard) used is real per-capita consumption1. Although the survey collects information on both household income and consumption, consumption rather than income is used as the welfare indicator due to many reasons. Firstly there is a relatively high rate of underreporting of income which biases reported household aggregate income. Secondly consumption captures welfare achievement more precisely then income, since the latter is a more appropriate measure of welfare opportunity. In other words, consumption is a better outcome indicator then income. Furthermore, income tends to fluctuate more than consumption, especially in agrarian economies according to the harvest cycle. Throughout the chapter, I use four poverty lines; Rs 1206 (national), Rs 1391 (Urban), Rs 1189 (rural), and Rs 1067 (estate), estimated by Siddhisena and Jayathilaka (2004)2.

2.3 Unconditional Poverty Profile: Cross Tabulations Poverty is frequently considered as the defining characteristic of

underdevelopment and its reduction is the ultimate goal of development policy. To reduce poverty, policy makers first need to know the incidence, depth and severity of poverty. Three different poverty measures nested in the Foster-Greer-Thorbecke (FGT)
1 2

Total household consumption divided by the number of household members. The paper provides a detailed explanation for derivation of poverty lines.

class were utilized to capture the different dimensions of poverty. The FGT indices combines income and the poverty line in to poverty gaps, and aggregate these gaps to evaluate the extent of poverty. The FGT poverty index can be expressed as;
P ( z; ) = [ z Q* ( p : z )] dp
0 1

Incomes censored at the poverty line z, is given by Q* ( p : z ) . Thus, the poverty gap at percentile p is g ( p; z ) = z Q* ( p : z ) . When = 0 , the FGT index reduces to the simple headcount poverty measure. Poverty headcount is the share of population with incomes falling below the poverty line. Using the poverty lines and per capita consumption levels, the poverty headcount figures show that 25.2 percent of the Sri Lankan population are in poverty (Table 2.1). Furthermore the highest incidence of poverty is in the estate sector followed by the rural and urban sectors. However, looking only at the percentage of people falling below the poverty may gloss over some vital variation in the depth and severity of poverty in different sectors. The depth in poverty across the three sectors was captured using the poverty gap index. The average poverty gap, P ( z; = 1) , is the average extra consumption that would be required to bring each poor household up to the poverty line. The second column in Table 2.1, presents the normalized average poverty gap estimates. The national unnormalized average poverty gap (derived from the normalized average poverty in Table 2.1)3 was Rs.69.95. After extrapolating the poverty gap of the survey population to that of the nation, the total annual poverty gap in Sri Lanka was estimated at Rs.315 million. However, in reality this figure will be much higher after accounting for targeting inefficiencies and administration costs in poverty reduction programs. Furthermore, in practice closing the total gap solely through income transfers
3

Unnormalized average poverty is equal to the normalized poverty multiplied by the absolute poverty line.

is not feasible. A more prudent way is through poverty reduction programs that raise the income of the poor via income generating activities. Both the headcount index and the poverty gap violates the transfer principle since they are insensitive to transfers among the poor. To overcome this shortcoming, the squared poverty gap was used to depict the severity of poverty. The squared poverty gap, P( z; = 2) , applies more weight on the

poverty gaps of those households whose consumption fall further below the poverty line and takes in to account the inequality of the poor. According to Table 2.1, the estate sector has the worst situation, with a poverty incidence of 28 percent and an average poverty gap of almost 7 percent. The remaining two measures (poverty gap and squared poverty gap) also indicate that the highest level of poverty is in the estate sector followed by the rural and urban sectors. Poverty profile cross tabulations with respect to the characteristics of the household, housing and access to services are also summarized in Table 2.1. Firstly, it is important to explore the age and gender dimensions of poverty. Since household characteristics such as age and gender are easily observable, they serve as important targeting variables. Results show that the incidence, depth and severity of poverty varies significantly with respect to the gender of the household head. Female headed households are six percent more likely to be in poverty compared to male headed households. Since this analysis is based solely on the headship of, it might be also be reasonable to believe that the "average welfare" of women is much more lower than men after accounting for gender wage gaps in the labor marker and/or intra-household distribution of resources4. The age of the household does not seem to be significant correlate of poverty. In Table
4

An elegant paper by Deaton (1989) on intra-household resource allocation.

2.1, increase in the poverty incidence with age is negligible. This fact is proven to a further extent in the conditional poverty profile (based on logit regression) discussed in the next section, which reveals that the marginal effect of age on poverty is statistically insignificant after controlling for other factors such as education and household size. One of the most significant and extremely pronounced negative correlates of poverty is the level of education of the household. According to Table 2.1, poverty declines monotonically with years of education. Households with no schooling has a forty three percent probability of being in poverty, while a household with tertiary education has only a five percent chance of being in poverty. Another important correlate of poverty is the household size. As indicated in Table 2.1, households consisting of four or more persons being in poverty on average is more than twice, compared to a household with one or two members. Even after the household size was disaggregated in to different sub-groups (number of children, women, men and elders), the size of each group shows a positive relationship with poverty incidence. The correlation is strongest with the number children, where households with three or more children have a poverty incidence of more than double the national rate. Furthermore, for households with children in the age group of 0-6 years have a higher probability of being in poverty than households with the same number of children aged 7-16 years. Similarly if one considers two households with the same number of men and women, the one consisting of men has a lower poverty incidence than the household with women. This shows again that gender is critical factor with respect to poverty. As regards employment status, households with salaried employment have the lowest headcount, compared to the ones in business (including trade and manufacture) and in

casual labor (wage). Households engaged in casual labor have thirty eight percent probability of being below the poverty line, while the salary employed households have a ten percent probability. Furthermore, results indicate that household heads with retirement benefits are less likely to be in poverty than ones without. The poverty profile with respect to housing characteristics and access to services need to analyzed cautiously, since it reveals only the association between variables and not casual relationships. According to Table 2.1, houses with electricity as the main source of lighting has a seventeen percent poverty incidence, while houses using kerosene as the main source of lighting have a poverty incidence of thirty seven percent. Poverty incidence is also high for houses which uses firewood or sawdust as fuel for cooking than houses using gas. Households living in shanties and line rooms have a poverty incidence twice as much as households in single houses. An insightful way to depict the incidence, intensity and inequality of poverty is through cumulative poverty gap (CPG) curves or TIP-curves5 shown in Figure 2.1. The cumulative poverty gap curve aggregates the average poverty gaps of the bottom p percentiles of the population and is expressed as:
G ( p : z ) = g (q : z )dq.
0 p

The poverty gap g ( p : z ) , at a given value of p is given by the slope of G ( p : z ) . The average poverty gap equals the cumulative poverty gap at p = 1 . Figure 2.1 indicates that when p = 1 , that the unnormalized average poverty gap are Rs.75.11, Rs.70.15 and Rs.71.5 for the urban, rural and estate sectors, respectively. The percentile at which the cumulative poverty gap curve becomes horizontal indicates the poverty headcount. For
5

The "Three I's of Poverty" (Jenkins and Lambert 1997).

10

the urban sector, the percentile in which the CPG curve becomes horizontal is 0.207, implying that 20.7 percent of urban household are in poverty. Similarly, in the rural sector 25.7 percent and in the estate sector 28.1 percent of the households are in poverty. Furthermore, the inequality dimension among the poor is captured by the degree of concavity in the non-horizontal section of the CPG curve. Thus, if income among the poor were equal or if the poverty gaps were the same the non-horizontal section in Figure 2.1 would be a straight line. According to Figure 2.1, among the bottom p = 0.25 proportion of the population the highest level of inequality is found in the urban sector followed by the rural and estate sectors. Household level consumption inequality across sectors was measured using Lorenz curves and Gini indices. Lorenz curve plots the cumulative percentages of total consumption against the cumulative percentage of households, starting with the poorest household. Lorenz curve can be defined as:

Q(q)dq = 1 L( p) = Q(q)dq
0 1 0

Q(q )dq

The numerator is the sum of income of the bottom percentile, p . The denominator sums the income of all households. When the size of the population is normalized to one, the denominator can be viewed as the average income . Figure 2.2 shows the Lorenz curves for the three sectors (urban, rural and estate). From the graph it is evident that the highest level of consumption inequality is in the urban sector followed by rural and estate. In other words the consumption in the estate sector is more egalitarian than in the rural and urban sector since the poorest people in the estate sector receive a share superior to that of their equivalents in the rural and urban sectors. Furthermore the Lorenz

11

curve for the estate sector can viewed as having been obtained from rural or urban Lorenz curves through a series of equalizing Pigou-Dalton transfers6. When p = 0.5 the sectorial estimates for urban, rural and estate are L( p ) = 0.256 , L( p) = 0.308 and L( p) = 0.336 respectively. These values can be interpreted, for instance in the urban sector as the poorest 50% households holding 25.6% of the total consumption in the total population. The ratio between the area enclosed by the line of equality and the Lorenz curve can be summarized by the Gini coefficient. The Gini index can be expressed as:
Gini Index of Inequality = 2 ( p L( p ))dp
0 1

One of this implicit assumptions in the Lorenz curve is that the distance, p L( p ) , from the line of perfect equality in consumption is weighted equally across percentiles, p . A more general version is the class of S-Gini (single parameter Gini) inequality indices which applies percentile dependent weights to the distance p = L( p) . The S-Gini inequality indices can be expressed as:
I ( ) = ( p L( p )) ( p :)dp
0 1

Where, ( p : ) are percentile dependent weights expressed as:

( p : ) = ( 1)(1 P)( 2)

Larger the value of , larger will be the weight placed on the inequality of the bottom percentile (or the poorest people). Therefore larger the value of , greater will be the
Dalton transfers principle states that a transfer of income from someone lower in the income distribution to someone higher in the income distribution, holding everyone else's income constant, should increase the numerical value of an inequality index. If vector y which is a transformation of the vector y obtained by a transfer
6

from y j iff I ( y) I ( y ) .

to

yi , and yi + > y j , then the transfer principle is satisfied

12

ethical concern felt for the poor by the social decision maker. Note, when = 2 , we have the standard Gini index which gives equal weight to all percentiles, p . An alternative inequality measure that explicitly incorporate normative judgments about the social welfare is the Atkinson index. The Atkinson inequality index is based on an additive social welfare function7 and is expressed as;
1 1 (1 ) 1 ( 0 Q( p ) dp ) 1 , when 1 I ( ) = 1 exp( ln(Q( p )dp ) 0 1 , when = 1

Q( p) is the standard of living of the individual whose rank or percentile in the

distribution is p . The parameter reflects the strength of society's preference for equality and is bounded by zero and infinity. When = 0 an increase in a poor individual's income has the same effect on social welfare as increasing the income of a rich individual by the same amount. However, when > 0 increasing the income of a poor individual is socially more enviable than increasing the income of a rich individual. Table 2.2 reports the Gini coefficients (when , is equal to 1.5, 2.0, and 2.5) and Atkinson measures (when , equal to 0.5, 1.0, and 2.0) for the urban, rural and estate sectors. According to both the Atkinson and the standard Gini index (when =2), the highest degree of inequality is in the urban sector followed by the rural and estate sector.

xi1 1 i =1 7 , 1where x is the standard of living. Social welfare function (W) = N 1 1 N ln W = ln xi , = 1 N i =1

13

2.4 Conditional Poverty Profile: Marginal effects Although, there are numerous studies on the measurement of poverty in Sri Lanka, literature on the determinants and correlates of poverty in a multivariate framework is best at scanty. Siddhisena and Jayathilaka (2004) looked at the composition of the poor according to several demographic and socio-economic characteristics on a one-to-one basis (bivariate analysis). The primary drawback of unconditional bivariate analysis is that it often erroneously oversimplifies complex relationships. For an example, if poverty is higher in rural area, it is not clear if the observed relationship should be attributed to rural areas per se, or to some factor that is correlated with rural areas such as low educational attainment. Bivariate unconditional poverty profiles is useful to a certain extent in the case of geographical targeting, but multivariate conditional poverty profiles are highly desirable for evaluating proposed policy interventions. The primary objective of this section is to assess the relative importance of various correlates of poverty, and where possible attribute causality to them. Conditional poverty profile is constructed on the basis of a multivariate analysis of poverty correlates. Partial correlates of poverty are computed using two comparable methodologies. Firstly, a logistic regression was estimated, with the probability of a household being in poverty as the dependent variable and a set of economic and demographic variables as correlates. The response variable is a dummy defines as:
1, if the household is belowthe poverty line POV = 0, if otherwise and Pr POV = 1 X = F ( X , ) Pr POV = 0 X = 1 F ( X , )

14

Where X is the vector of economic and demographic variables. is the set of parameters reflecting the impact of changes in X on the probability. The vector of

economic and demographic variables ( X ) that are hypothesized to determine consumption and hence poverty can be categorized under, demographic, education, employment, region and dwelling characteristics. The demographic data include household size disaggregated by age and sex: number of children 0-6 years, number of children 7-16 years, number of women 17-60 years, number of men 17-60 years and number of elderly 60+ years. In regressions, a quadratic term in household size is included to capture the nonlinearities in the relationship between household size and living standards. Based on the findings of other developing country studies, (Lanjouw and Raviallion, 1995; Deaton and Paxton, 1998), the expectation is a positive relationship between household size and the probability being in poverty (or a negative relationship between household size and total consumption per capita). The level of educational attainment was measured on three different levels, based on the assumption that human capital (as measured by education) contributes negatively to the probability of being in poverty. The three different levels that was used to measure the maximum level of education attained by the household head are: primary education (studying in year 1passed year 6), secondary education (passed year 7- GCE O/L), and tertiary education (beyond GCE O/L). In the employment category four variables were used: household head is engaged in casual labor, household head is in salaried employment, if any member of the household receives or entitled to receive pension income(EPF or ETF), and whether the household head is engaged in business(including trade and manufacturing). Four variables were selected to reflect the housing characteristics and

15

access to services: Ownership of the dwelling tenure, type of housing, main source of lighting utilized for dwelling and main type of fuel used for cooking. And finally, regional heterogeneity was controlled by allowing for the sector and province in which the household resides. Variable definitions and means are provided in Table 2.3. From the stand point of econometric purity, the set of independent variables used in this study are fairly generous and the argument for exogenity is stronger especially in a short time horizon model. As the time horizon gets longer, most of the economic variables at the household level become endogenous. Except for a few variables (such as gender and age), all other variables end up being a function of the household welfare level to some extent. Even though the ideal solution is instrumental variable technique, reasonable instruments were not available in the survey data. Therefore, special care must be taken when interpreting coefficients, since the regressions will only return results for the degree of association or correlation and not for casual relationships. The probability model is the regression;
E POV X = 0[1 F ( X )] + 1[ F ( X )] = F ( X ]

Based on the logistic distribution, e X Pr POV = 1 X = 1 + e X Table 2.4 presents the parameter estimates for the logistic regression. The column,

dy , dx

is the marginal effect of a change in a specific element of X on the probability of being poor. Since almost all the variables (except for age, ageSQR, urbundum, south and nwest) in the model have estimated parameters significantly different from zero, the model does point at a sharply defined set of potentially useful targeting variables in the

16

context of policy intervention to alleviate poverty. Firstly, there is a significant positive and concave relationship between household size and being in poverty. All five variables measuring the household size, disaggregated by age and sex are positive and highly significant. In developing countries, due to low savings and underdeveloped social security systems, fertility rates among the poor are high, since parents receive some economic support from children once they reach old age. Being consistent with the bivariate unconditional poverty profile, the number of children in the age group of 0-6 years has the strongest positive correlation with poverty. After controlling for other factors, age of the household head does not have a significant effect on the probability of being in poverty. This finding is not surprising, since the unconditional poverty profile also indicated that age to be weak correlate of poverty. However even after controlling for other factors, the gender of the household head is a significant correlate of poverty. According to Table 2.4, the probability of being in poverty declines monotonically with years of education. All three educational level variables are statistically significant and have the expected negative association with poverty. Tertiary education has the largest impact on poverty (followed by secondary and primary), reflecting the fact that education increases the stock human capital, which in turn increases labor productivity, earnings and consumption. Turning then to the marginal effects of employment related variables, the findings are policy-wise imperative. Firstly, household heads working as casual laborers increases the probability of being in poverty, while working in a salaried occupation decreases the probability of being in poverty. This fact is not surprising, since occupations which requires low amount of human capital (casual wage jobs) will be associated with low earnings and thereby increases the likelihood of being below the

17

poverty line. Furthermore, the results indicate that the probability of being poor is significantly lower for household heads engaged in business (including trade and manufacturing). Finally with respect to employment benefits, households with members receiving or entitled to receive EPF or ETF are less likely to be in poverty than those who are not receiving or entitled to receive pension income. The unconditional poverty profile revealed earlier that the highest incidence of poverty is in the estate sector followed by the rural and urban sectors. According to the multivariate poverty regression, still living in the rural sector significantly increases the probability of being in poverty. But, after controlling for other factors, urban and estate sector dummies turns out as insignificant correlates of poverty. For any permutation of sector dummies included in the poverty regression, urban and estate sector dummies remained statistically insignificant all the time, while the rural sector dummy was positive and significant. The finding suggests two salient features about poverty in Sri Lanka. Firstly, the high incidence of poverty in the estate sector should not be attributed to the estate sector per se, but for some other factor[s] (such as low educational attainment), that might be correlated with the estate sector. Secondly, poverty in Sri Lanka is certainly a rural phenomenon, that needs to be explained by many other factors, which deserves continuing attention and scrutiny. The estimated coefficients the dwelling tenure and type of housing are both statistically significant and positive. There is a positive correlation between poverty and households not owning their house, living in a shanty room or line room. The marginal effects are strongest for the ones living in shanty and line rooms - both in terms of the magnitude of the coefficient and statistical significance. Ownership of a house is important, especially

18

in rural areas, since it provides the location for a household enterprise and also acts as a collateral for a loan. The two variables used to capture a households access to services are also statistically significant and have plausible signs. Firstly, there is a negative correlation between households using electricity (as the main source of lighting) and the probability of being in poverty. Secondly, households using firewood or sawdust as the main fuel for cooking has a positive association with poverty. It is important to note here that the caveat about interpreting the estimated coefficients as partial correlation coefficients is particularly important, since the direction of causation is most certainly from poverty to variables related to housing and access to services. As a robustness check, it is important to note here that the results of the multivariate poverty regression had corroborated with the findings of the bivariate unconditional poverty profile in the preceding section. Substantively, the pattern of the partial correlates of poverty in the poverty regression is entirely consistent with the pattern of correlates that was revealed by the bivariate poverty profile. All factors which are correlated with an increase (decrease) in the poverty headcount are correlated with an increase (decrease) in probability to be poor. Next, the quantile regression8 approach was utilized to examine the correlates of per capita consumption at different points on the distribution. The most appealing feature of quantile regression is that it does not impose constant parameters over the entire distribution. It assumes the effect of economic and demographic characteristics of the i th household to differ across the welfare spectrum.

Koenker and Basset (1978).

19

The quantile regression model can be expressed as:

yi = xi + i ,

Where yi is the log of per capita consumption of the i th household, and xi represents the economic and demographic characteristics of the i th household. By Imposing the assumption that the th quantile of the error term conditional on the regressors is zero,

(Q (ui , | xi ) = 0) , the th conditional quantile of expressed as:

Q ( yi | xi ) = xi For any (0,1), the parameter can be estimated by
= arg min

R k

yi with respect to xi can be

i{i| yi xi }

| yi xi | +

(1 ) | | y x i i i{i| yi xi }

Note, that when = 0.5 , we have the special case known as the median regression or the least absolute deviation (LAD) estimator. Five quantile regressions were estimated at the 10th, 25th, 50th, 75th and 90th quantiles. The standard errors were computed by bootstrapping with 100 replications. OLS regression was also estimated for the purpose of comparison. Table 2.5 reports the results of the OLS and quantile regressions. Firstly, at all quantiles estimated on the conditional expenditure distribution, household size is inversely related with the standard of living as measured by consumption. All five variables measuring the household size, disaggregated by age and gender are highly significant at all estimated quantiles. Furthermore, additional children have a much larger effect on per capita expenditure than adults. These results reconfirm the earlier findings of both the unconditional poverty profile and the logistic regression.

20

A more comprehensive manner of presenting the results is in the form of a graph. Figure 2.3 (panel 1-5) shows the development of the coefficients representing household size over the entire conditional consumption distribution. The estimated coefficient for each percentile is plotted as a continuous line and its 95%-confidence interval is the shaded area. The OLS estimate is the dark horizontal line and parallel to it is the 95%-confidence bands. With the exception of elders, all the other variables reflecting household size tends to have an increasingly larger impact on consumption as one moves up on the expenditure distribution. At the two extreme end of the distribution, estimates for these variables fall outside the confidence interval of the OLS estimate and is quite different from the OLS estimate. According to the Table 2.5 age tends to have an extremely weak negative relationship with the standard living of at all estimated quantiles on the conditional distribution. This again confirms the findings of the unconditional poverty profile and the logit regression. However, the gender of the household head has a significant association with the standard of living . At all quantiles there is a negative relationship between per capita expenditure and being a female-headed household. But the gender effect on welfare tends to weaken as one moves up the conditional distribution. Below the median (50-th quantile), being a female headed household reduces per capita expenditure by at least 4%, but the fall in per capita expenditure is nearly three times less on the top of the distribution. The significance of the gender effect begins to fade beyond the median and eventually becomes insignificant at the 90-th quantile. Quantile regression results also indicate that households residing in the rural sector are worse off. Figure 2.3 shows that households in the bottom quantiles and upper quantiles

21

are less sensitive to the rural sector compared to the households in the median quantiles. This implies that the poorest and the least poorest people experience less of the negative impact of living in a rural area than the median poor. Also, as Table 2.5 indicates, for the households on the top of the conditional distribution (90-th quantile), living in the rural sector has no significant effect on the standard of living. With respect to education, all three variables indicating the levels of education shows a positive relationship with per capita consumption. Figure 2.3 reveals that the impact of primary education does not vary a lot between quantiles and the quantile coefficients do not differ much from the least-square results. In other words, returns to primary education are not different for the poor and non-poor. Table 2.5 also shows that the primary education variable is not significant for the households in the bottom 10-th quantile, implying that the pay off from primary education to the poorest is not significant for the poorest is not significant. But secondary and tertiary education significantly increases the standard of living across all quantiles. Figure 2.3 also shows that the premium on tertiary education is relatively high for less poor households. It is important to note that at the two extreme ends of the distribution, quantile regression estimates fall outside of the OLS estimates. Thus, the conventional least squares confidence interval does a poor job representing this range for the tertiary education variable. Turning to the employment related variables, with the exception of the bottom 10-th quantile, the casual wage coefficient is significant and negative across all other quantiles. Being consistent with the earlier findings from the unconditional poverty profile and logit regression, household heads engaged in casual labor are associated with lower per capita expenditure. However the impact of being in casual labor on per capita is relatively

22

negligible at upper quantiles than at the bottom quantiles. Figure 2.3 also indicates that the OLS method underestimates the effects of being in casual labor on per capita consumption on the upper quantiles of the conditional distribution. According to Table 2.5 and Figure 2.3, for the poorer households (quantiles 10 and 25), the marginal effect of being in a salaried occupation is relatively high, compared to the less poor households. Results also suggest that households in businesses (including and manufacturing) experience higher levels of per capita consumption. The coefficient for a household being engaged in business is highly significant (Table 2.5) and stable across all quantiles (Figure 2.3). With respect to retirement benefits, not receiving or not being entitled to receive pension income has a strong negative effect on per capita expenditure for the households in lower quantiles. The impact of retirement benefits on consumption and its significance level is relatively less at higher quantiles. With regard to the four variables (ShantyLineroom, HHnotowninghouse, Firewood, Electricity) reflecting housing characteristics and access to services, the quantile regression estimates only return the results of the degree of association with per capita expenditure and no influence of causation can be made. All four coefficients are significant across the conditional distribution and have signs consistent with the unconditional poverty profile. Finally, the interpretations of the casual effects of regional dummies are somewhat difficult and can only be described as dramatic. The regional dummies were included primarily for controlling regional heterogeneity.

23

2.5 Conclusion

This chapter investigated the probable determinants and correlates of poverty in Sri Lanka. It is worth summarizing some of the main results of this chapter. First, poverty remains more acute in rural areas than in urban areas. Furthermore, since 76% of the population live in rural areas, the rural shares in the total composition of poverty is more higher. However, the degree of inequality is much more greater in urban areas, compared to the rural sector. Therefore priority need to be given for policy initiatives aimed at reducing poverty in rural areas, while recognizing the need to tackle urban inequality. Results also indicate that the pay off to smaller families is higher, and larger families are more likely to be poor. Furthermore, the costs of dependents are significant for all expenditure groups. An extra child or elderly creates a greater economic burden than an extra man or woman in household. With regard to the age of the household, the unconditional bivariate poverty profile indicated age to be a weak correlate of poverty. Confirming this result, multivariate analysis also found the age of the household head to be insignificant correlate of poverty, even after controlling for other factors. The study also found evidence to support the hypothesis of the feminization of poverty. Femaleheaded households are significantly worse off compared to male-headed households, especially in poorer households. Household head's education level also had an instrumental effect in determining the vulnerability to poverty. Poverty incidence declined monotonically with years of education. Furthermore, quantile regression results indicate that the pay off from primary education to the poorest is not significant. With respect to the labor market, the incidence and probability of being in poverty is higher for households in casual labor, compared to

24

the ones in salaried employment. Finally, the poor are more likely to live in shanty and line rooms and to use kerosene and firewood for lighting and cooking. All of the above suggests the need for increasing school enrolment; supplemental educational programs and upgrading of schools are sensible components of a poverty reduction strategy. Clearly, programs of information, micro-credit, marketing, small business incubators, etc. deserves special attention in the design of national poverty reduction strategies. Findings regarding the link between welfare and household size, employment status and the access to services, are invaluable in the realm of indicator targeting. For an example, finer targeting can be done on the basis of household size and composition (eg., number of children, number of female members). The finding of a strong correlation between poverty and children, suggests that the presence of children need to be considered as a strong indicator candidate for targeting. The beauty of enacting poverty alleviation programs through targeting key poverty indicators is that both administrative costs and leakage can be lowered. In conclusion, these findings indicate the importance of a set of policies which are super pro-poor, namely increasing school enrollment and achievement, effective family planning programs to reduce the birth rate and dependency load within households, and granting priorities for specific cohorts (children, elderly, rural and female headed households) in targeted interventions.

25

Table 2.1: Unconditional Poverty Profile (cross tabulations) = 0 =1 = 2

Sri Lanka Urban Rural Estate Characteristics of the Head Male Female Age 29 years Age: 30-59 years Age 60 Education No schooling (primary.edu) (secondary.edu) (tertiary.edu) Household Size 0-2 3-4 4+ Number of Children (0-6 years) 1 2 3 4 Number of Children (7-16 years) 1 2 3 4 0.226 0.304 0.407 0.444 0.050 0.068 0.097 0.122 0.017 0.024 0.035 0.048 0.296 0.389 0.508 0.800 0.066 0.097 0.136 0.234 0.023 0.036 0.048 0.074 0.143 0.178 0.342 0.038 0.036 0.082 0.018 0.012 0.030 0.426 0.336 0.196 0.052 0.116 0.079 0.041 0.008 0.048 0.029 0.013 0.003 0.242 0.302 0.224 0.248 0.270 0.054 0.078 0.043 0.056 0.066 0.019 0.031 0.013 0.019 0.026 0.252 0.207 0.257 0.281 0.058 0.054 0.059 0.067 0.021 0.020 0.021 0.025 Number of men (17-60 years) 1 2 3 4 Number of Women (17-60 years) 1 2 3 4 Number of Elders 1 2 3 Employment Salary Casual Wage

= 0 =1 = 2
0.235 0.252 0.308 0.347 0.051 0.058 0.077 0.086 0.017 0.020 0.030 0.030

0.237 0.259 0.319 0.351 0.263 0.278 0.269 0.101 0.376

0.052 0.059 0.077 0.091 0.062 0.069 0.117 0.018 0.093 0.027 0.052 0.060 0.052 0.077 0.053 0.117 0.161 0.036 0.088 0.004 0.065

0.018 0.020 0.030 0.037 0.024 0.025 0.054 0.007 0.035 0.009 0.020 0.022 0.018 0.029 0.019 0.058 0.067 0.012 0.032 0.002 0.023

Business 0.139 Receiving (or entitled) for pension income 0.216 Not receiving (or not entitled) for pension income 0.261 Housing Tenure and Type Owned by household head 0.234 Not owned by household head 0.312 Single house Annexe Shanty or line room Main Source of Lighting Electricity Kerosene Main Fuel Used for Cooking Gas Firewood or sawdust 0.022 0.284 0.173 0.366 0.241 0.250 0.550

26

Table 2.2: Inequality Indices

S-Gini Index
=1.5 =2 =2.5

Atkinson Index
=0.5 =1.0 =2.0

Urban

0.260

0.375

0.438

0.116

0.213

0.817

Rural

0.197

0.288

0.345

0.072

0.138

0.716

Estate

0.162

0.243

0.296

0.052

0.101

0.222

27

Table 2.3: Variable definitions and means

Variable LogPCE Children (1) Children (2) Men Women Elders HsizeSQR Age AgeSQR Dummy Variables POV Female Head Rural Urban Primary.edu Secondary.edu Tertiary.edu Salary Casual Wage Business NoRetBenifit ShantyLineR Household is below the poverty line Household head is female Household resides in the rural sector Household resides in the urban sector Year 1-Year 6 Year 7-GCE (O/L) Year 12 and above Household head in salaried employment Household head works for a casual wage Household head in business (including trade and manufacture) Household head is not receiving or not entitled to receive retirement benefits. Household lives in shanty or line room 0.2818 0.9719 0.3968 0.1644 0.0597 0.1094 0.0766 0.1258 0.1930 0.0936 0.0528 0.2093 0.8248 0.6385 0.1103 0.1269 0.1503 0.0743 0.0750 0.1007 0.0941 0.0731 0.2276 0.8619 0.5775 0.1240 0.1100 0.1400 0.0720 0.0878 0.1240 0.0940 0.0680 by household Firewood Electricity South North East Uva Sabara Central Nwest Ncentral Household uses firewood or sawdust as the main of fuel for cooking Main source of lighting is electricity Household in Sourthen province Household in Northern province Household in Eastern province Household in Uva province Household in Sabaragamuwa province Household in Central Province Household in North Western province Household in North Central Province HHnotowningHouse Dwelling unit is not owned by 0.8290 0.7896 0.7995 1 0.2009 0.8138 0.1132 0.4706 0.3527 0.0211 0.0645 0.3702 0.0851 0 0.1564 0.7471 0.2159 0.3138 0.4862 0.1289 0.1965 0.2077 0.1776 0.2521 0.1676 0.7639 0.1900 0.3533 0.4525 0.1017 0.1632 0.2486 0.1542 Definition Logarithm of real per capita consumption Number of children (0-6 years) Number of children (7-16 years) Number of men (17-60 years) Number of women (17-60) Number of persons (60+ years) Household size squared age of household head Age squared Poor 2.9466 0.5896 1.2062 1.4569 1.5711 0.3882 30.9974 49.6345 2660.517 Non-poor 3.3233 0.3631 0.8256 1.353 1.4336 0.3545 21.444 49.2157 2602.498 Total 3.2282 0.4201 0.9216 1.3793 1.4683 0.3629 23.853 49.3213 2617.128

28

Table 2.4: Logit regression estimates

Independent Variable Children (1) Children (2) Men Women Elders HsizeSQR Age AgeSQR Female Head Rural Urban Primary.edu Secondary.edu Tertiary.edu Salary Casual Wage Business NoRetBenifit ShantyLineR HHnotowningHouse Electricity Firewood South North East Uva Sabara Central Nwest Ncentral Dy/dx (marginal effects) 0.126 0.089 0.073 0.082 0.089 -0.002 -0.003 0.000 0.068 0.105 0.031 -0.035 -0.085 -0.217 -0.068 0.055 -0.123 0.059 0.174 0.024 -0.113 0.198 0.031 -0.262 -0.095 -0.067 0.043 0.111 -0.022 -0.132 z-value 11.44 9.55 7.18 8.02 6.72 -3.70 -1.38 1.52 5.47 4.24 1.18 -2.40 -5.37 -7.36 -3.79 4.92 -8.04 4.44 -11.12 2.18 -11.12 8.73 1.83 -12.11 -5.21 -3.17 2.31 6.38 -1.19 -5.94

N=7481 Pr>=0.000

Pseudo R=0.225 Loglikelihood =-3274.5261

Notes

1, if the household is belowthe poverty line 1) Dependent variable: POV = 0, if otherwise 2) Poverty line = Rs.1206 3) Variable definitions and means are given in Table 2.3.

29

Table 2.5: Quantile and OLS regression estimates (standard errors in parentheses)
Independent Variable Constant Children(1) Children(2) Men Women Elders HsizeSQR Age AgeSQR FemaleHead Rural Urban Primary.edu Secondary.edu Tertiary.edu Salary Casual Wage Business NoRetBenifit ShantyLineR HHnotowningHouse Electricity Firewood South North East Uva Sabara Central Nwest Ncentral 0.1 3.19 (0.065) -0.087 (0.009) -0.065 (0.008) -0.06 (0.007) -0.062 (0.009) -0.064 (0.010) 0.003 (0.000) 0.007 (0.002) 0.000 (0.000) -0.042 (0.009) -0.068 (0.023) -0.025 (0.023) 0.020 (0.015) 0.068 (0.016) 0.142 (0.019) 0.035 (0.110) -0.017 (0.010) 0.080 (0.012) -0.037 (0.010) -0.144 (0.027) -0.028 (0.009) 0.078 (0.008) -0.125 (0.013) -0.050 (0.014) 0.116 (0.013) 0.069 (0.012) -0.043 (0.016) -0.046 (0.015) -0.109 (0.012) -0.004 (0.014) 0.096 (0.023) Quantile 0.25 3.366 (0.043) -0.095 (0.006) -0.074 (0.006) -0.072 (0.006) -0.073 (0.006) -0.070 (0.008) 0.004 (0.000) 0.006 (0.002) 0.000 (0.000) -0.04 (0.007) -0.008 (0.016) -0.029 (0.016) 0.021 (0.009) 0.056 (0.011) 0.139 (0.013) 0.024 (0.007) -0.026 (0.006) 0.008 (0.009) -0.039 (0.007) -0.119 (0.019) -0.022 (0.007) 0.066 (0.006) -0.138 (0.008) -0.047 (0.009) 0.112 (0.011) 0.037 (0.009) 0.030 (0.012) -0.055 (0.011) 0.104 (0.011) 0.001 (0.009) 0.058 (0.010) 0.5 3.473 (0.039) -0.106 (0.006) -0.078 (0.005) -0.075 (0.006) -0.078 (0.005) -0.066 (0.007) 0.004 (0.000) 0.005 (0.001) 0.000 (0.000) -0.043 (0.008) -0.057 (0.015) 0.007 (0.015) 0.025 (0.008) 0.058 (0.009) 0.164 (0.014) 0.012 (0.006) -0.033 (0.006) 0.078 (0.007) -0.029 (0.007) -0.087 (0.014) -0.015 (0.006) 0.060 (0.005) -0.159 (0.008) -0.022 (0.009) 0.138 (0.011) 0.024 (0.008) 0.019 (0.008) -0.019 (0.008) -0.064 (0.010) 0.018 (0.012) 0.046 (0.010) 0.75 3.591 (0.040) -0.114 (0.007) -0.085 (0.007) -0.079 (0.008) -0.088 (0.007) -0.078 (0.010) 0.004 (0.001) 0.004 (0.001) 0.000 (0.000) -0.033 (0.009) -0.028 (0.013) 0.042 (0.015) 0.027 (0.009) 0.062 (0.008) 0.181 (0.016) -0.002 (0.009) -0.042 (0.007) 0.078 (0.009) -0.016 (0.007) -0.062 (0.014) -0.019 (0.006) 0.065 (0.007) -0.173 (0.011) -0.016 (0.009) 0.144 (0.011) 0.001 (0.009) 0.001 (0.012) -0.038 (0.013) -0.055 (0.011) 0.014 (0.013) 0.036 (0.012) 0.9 3.698 (0.056) -0.113 (0.009) -0.089 (0.008) -0.079 (0.010) -0.086 (0.009) -0.066 (0.012) 0.004 (0.001) 0.005 (0.002) 0.000 (0.000) -0.015 (0.0012) -0.021 (0.029) 0.081 (0.030) 0.028 (0.001) 0.075 (0.015) 0.221 (0.021) -0.025 (0.012) -0.069 (0.010) 0.089 (0.015) -0.009 (0.011) -0.059 (0.027) -0.037 (0.010) 0.059 (0.009) -0.180 (0.016) -0.032 (0.014) 0.122 (0.014) -0.010 (0.017) -0.024 (0.017) -0.043 (0.021) -0.050 (0.019) 0.002 (0.019) 0.048 (0.024) OLS 3.431 (0.036) -0.097 (0.005) -0.074 (0.004) -0.069 (0.005) -0.069 (0.005) -0.063 (0.006) 0.003 (0.000) 0.006 (0.001) 0.000 (0.000) -0.037 (0.006) -0.056 (0.013) 0.013 (0.014) 0.031 (0.008) 0.072 (0.009) 0.179 (0.011) 0.007 (0.007) -0.044 (0.006) 0.081 (0.006) -0.031 (0.006) -0.110 (0.013) -0.029 (0.005) 0.069 (0.005) -0.156 (0.007) -0.025 (0.009) 0.137 (0.009) 0.029 (0.009) 0.024 (0.010) -0.026 (0.009) -0.070 (0.009) 0.006 (0.009) 0.070 (0.010)

30

Figure 2.1.Cumulative poverty gap curves

31

Figure 2.2 Lorenz curves

32

0.05

uva 0.00 0.05 0.10

No_Ret_Benifit 0.08 0.06 0.04 0.02 0.000.02

Salary 0.10 0.05 0.00 0.05 0.10

Casual_Wage 0.10 0.08 0.06 0.04 0.02 0.00

men 0.10 0.08 0.06 0.04 0.02

0 .2 .4 .6 Quantile .8 1
sabara 0.100.05 0.00 0.05 0.10 0.00 0.00 business 0.04 0.06 0.08 0.10 0.12 ageSQR 0.00 Rural 0.15 0.10 0.05 0.00 0.05

0 .2 .4 .6 Quantile .8 1 1 1 1 .8 .8 .8 .4 .6 Quantile .4 .6 Quantile .4 .6 Quantile .2 .2 .2

0
women 0.10 0.08 0.06 0.04

0 .2 .4 .6 Quantile .8 1
0.15 central 0.10 0.05 0.00 0.00 0.01 HsizeSQR 0.00 0.00 HHnotowningHouse 0.08 0.06 0.04 0.02 0.00

0 .2 .4 .6 Quantile .8 1 1 1 .8 .8 .4 .6 Quantile .4 .6 Quantile

FemaleHead 0.08 0.06 0.04 0.02 0.00 0.02

0 .2 .2

0 .2 .4 .6 Quantile .8 1
children1 0.140.120.100.080.06

0 .2 .4 .6 Quantile .8 1
nwest 0.04 0.02 0.00 0.02 0.04 0.06 0.10 0.05 south 0.05 0.00

0 .2 .4 .6 Quantile .8 1 1
ShantyLineR 0.200.150.100.05 0.00

0 .2 .4 .6 Quantile .8 .4 .6 Quantile .8 1 .2

0
PrimaryEdu 0.02 0.00 0.02 0.04 0.06

0 .2 .4 .6 Quantile .8 1
children2 0.100.080.060.040.02

Figure 2.3: Ordinary Least Squares and Quantile Regression Estimates

33
0 .2 .4 .6 Quantile .8 1
0.00 ncentral 0.05 0.10 0.15 north 0.08 0.10 0.12 0.14 0.16 0.18

0 .2 .4 .6 Quantile .8 1 .8 1
0.04

0 .2 .4 .6 Quantile
electricity 0.06 0.08 0.10

0 .2 .4 .6 Quantile .8 1
SecondaryEdu 0.04 0.06 0.08 0.10 0.12

0 .2 .4 .6 Quantile .8 1
elders 0.10 0.08 0.06 0.04 0.02

0 .2 .4 .6 Quantile .8 1
Urban 0.05 0.00 0.05 0.10 0.15

0 .2 .4 .6 Quantile .8 1
east 0.05 0.00 0.05 0.10 0.15

0 .2 .4 .6 Quantile .8 1
Firewood 0.25 0.20 0.15 0.10

0 .2 .4 .6 Quantile .8 1
TertiaryEdu 0.10 0.15 0.20 0.25 0.30

0 .2 .4 .6 Quantile .8 1
age 0.000.000.000.010.010.01

0 .2 .4 .6 Quantile .8 1

0 .2 .4 .6 Quantile .8 1

0 .2 .4 .6 Quantile .8 1

0 .2 .4 .6 Quantile .8 1

0 .2 .4 .6 Quantile .8 1

CHAPTER 3 EVALUATING THE IMPACT OF MICROFINANCE ON SAVINGS AND INCOME: QUASI-EXPERIMENTAL APPROACH USING PROPENSITY SCORE MATCHING

3.1. Introduction The microfinance revolution has changed attitudes towards helping the poor and has provided a large amount of credit, often to very low-income households, usually who would have been excluded by conventional financial institutions. There is no precise estimate of the number of microfinance service providers worldwide. According to the A worldwide Inventory of Microfinance Institutions. Sustainable banking with the poor (World Bank, 1996), 101 microfinance programs were surveyed in 101 developing countries. The report indicates that there are more than 1000 microfinance institutions, consisting of commercial banks, savings banks, credit unions, and non-governmental organization (NGOs). According to the report microfinance is more prevalent in Asia compared to other regions in the world. Out of the total loans disbursed, 76 percent are in Asia, 21 percent in Latin America, and 3 percent in Africa. The report further suggests that total funding mostly comes from external donors. Foreign donors provide 55 percent in Latin America, 47 percent in Asia, and 39 percent in Africa. However, still the academic development community is rather more skeptical about the impact of microfinance programs given the enthusiasm shown for these programs in donor and policy-making circles. To quote from Zeller and Meyer (2002):

34

"MFI field operations have far surpassed the research capacity to analyze them, so excitement about the use of microfinance for poverty alleviation is not backed up with sound facts derived from rigorous research. Given the current state of knowledge, it is difficult to allocate confidently public resources to microfinance development." This is a very strong statement of doubt on what proportion of income, savings and other effects on the beneficiaries of microfinance can be actually attributed to programs themselves. In spite of the large amount of subsidized resources that is pumped into microfinance programs, still there is a lack of statistically and scientifically robust impact assessment of such programs. Statistically robust evaluations have been limited due to several problems, such as selection bias, non-random program placement, difficulties in finding instrument variables, and paucity of reliable data. Moreover, methodological issues such as attribution and fungibility also poses problems. One of the major obstacles in evaluating microfinance programs is related to the attribution of specific impacts (effects) to specific causes (microfinance treatments). Fungibility of loans is referred to as the use of a loan by an individual else than the borrower or use of a loan for a purpose other than the one for which the loan was issued in the first place. Recently, an increasing effort is made to measure the impact of microfinance-not just to demonstrate the effectiveness of microfinance but to develop as well. As the concept of microfinance expands, it is sensible to ask the big question: does microfinance work at all as a poverty reduction tool. It is important to note that microfinance impact evaluations vary immensely in both quality and rigor. Copestake et al. (2001) describes how the concept of impact evaluation is scrutinized in three different schools of thought:

35

"The first accepts the case for doing a limited number of rigorous studies but argues that it is a specialized and expensive task. The second trusts more in the ability of practitioners to interpret and be guided by a mixture of routine monitoring and qualitative studies, more akin to market research than to academic research. A third view seeks an intermediate or `middle range' level of assessment: cheap enough to be carried out quite widely, but sufficiently rigorous to be credible." In developing countries, access to formal saving instruments for low-income people is non-universal. In these countries, opening a bank account is associated with high transaction costs and fees. Usually it is believed that low-income people does not save. However according to Mansell (1995) and Robinson (1992) if suitable financial instruments are available to low-income people, they become eager and regular savers. Unfortunately, no reliable estimates of the exact effect of these instruments on individual saving behavior exists. The belief that low income people do not save can also be refuted on the fact that their savings are not necessarily financial assets. It can be anything that can be used to preserve value and even increase it from the present to the future. For an example, it might be gold, coins, jewelry, cash and even animals. Rosenzweig and Wolpin (1993) illustrated the use of bullocks as assets in India. They showed that bullocks served as investment assets to generate income and also to smooth consumption. The lack of access to financial services could also lead households to adopt very inefficient forms of savings, ranging from cash under the mattress to extreme forms such as children (with the expectation of receiving old age support from children). Thus, all these adverse savings methods have significant effects on and in the perpetuation of poverty. Deaton (1990) states at least three reasons for studying savings in developing

36

countries. Firstly, in a microeconomic context, households are large and poor with income prospects being more uncertain compared to developed countries. Secondly, in a macroeconomic context, the fiscal system of low-income countries is not well developed enough to control the level disposable income to stabilize output and employment. Thirdly postwar literature suggests that savings is at very low levels, and this acts as an impediment to development.

3.2 Microfinance Institutions and Impact studies In Sri Lanka the concept of formal microfinance originated in 1906, with the creation of the first Sanasa society, the beginnings of a movement which was formalized subsequently in 1911, with the passing of the Cooperative Societies Ordinance. However if microfinance was to include informal credit providers as well, then very much the major proportion of micro-loans has been obtained from such informal sources, such as family, moneylenders, merchants, and produce buyers. Even as recently as 1986/87 a central bank survey revealed that some three quarters of credit transactions in Sri Lanka originated in this informal sector (Senanayake 1999). During the last twenty years, a wide range of institutional sources of microfinance developed, and the presidential commission on finance and banking encouraged the development of a "pluralistic" approach to rural finance. According to the Presidential Commission on Financing and Banking: provides scope for the involvement of a wide array of financial institutions, cooperatives, NGOs (nongovernmental organizations) and a range of governmental and informal agencies, each employing different techniques and strategies based on their different credit cultures, together with market-based interest rates and adequate

37

support services (Presidential Commission on Finance and Banking 1992, 122). According to Attanayake (1997), providers of microfinance services in the rural sector can be classified under three categories. (I) regulated financial institutions --- consist of commercial banks, both state and private, savings and development banks, and the regional development banks (RDBs). (ii) cooperatives; and (iii) "other formal nongovernmental institutions." The next subsection examines the providers of microfinance in Sri Lanka, using the above classification.

2.2.1. Regulated Financial Institutions in Microfinance Commercial banks: The Sri Lankan government had been using both commercial banks and regional development banks for channeling credit in poverty alleviation programs. State commercial banks and RDBs disbursed large amounts for microcredit in Janasaviya program from 1989 to 1994. The Janasaviya Trust Fund (from 1991) and its successor, the National Development Trust Fund (to end-1997 depended mostly on rural branch banking networks for disbursing credit. The main government program, Samurdhi (beginning 1995), relied on two state banks , while the Small Farmers and Landless Credit (SFLC) project (beginning 1990) employed RDBs for channeling credit. Samurdhi and the SFLC project will be discussed in sections 2.3 and 4.2 respectively. Apart from "policy lending", the only commercial bank that was engaged in broad range of microfinance activities was Hatton National Bank (HNB), the leading private commercial

38

bank in Sri Lanka. In addition to channeling credit in government poverty reduction programs, HNB has formed an novel microfinance program called Gami Pubuduwa Upadeshaka (GPU). Originating in 1989, GPU served small and micro-savers and borrowers in rural and semi-urban areas. To the end of 1998, GPU had advanced SLR 876 million for 27,500 small and micro activities, while 10,900 small projects had an outstanding amount of SLRs277 million. Furthermore, HNB's collection performance is remarkable compared to other commercial banks lending in rural areas. HNB has been very successful in mobilizing rural savings, recognizing the potential for growth within this sector that accounts for less than 5 percent of the total savings in the formal banking sector. Even though estimating rural savings involves some difficulties, available data suggest that in the late 1990s, national and rural savings experienced a growth of 57 percent in constant price terms. During the same period, HNB's rural branches recorded a nominal increase in savings of greater than 200 percent. Also, deposits within the GPU "micro" program increased over the same period by 183 percent, at an annual compound rate of 23.2 percent (Wijesundera 1999). HNB extended their services to rural areas by adopting innovative approaches to rural banking. Most banking transactions are processed inside the village rather than at bank branches. Most loans are issued on an individual basis, while many smaller loans are processed through nongovernmental organizations (NGOs). The average loan size of HNB's micro-program is around $700800 (country's where per capita GNP is around $800). In conclusion, HNB has developed the capacity to provide convenient financial services to the rural poor and genuine microcredit clients. From the viewpoint of HNB, this is a profitable strategy of resource mobilization; at the end of 2000, HNB recorded a deposit/loan ratio of around 2.5, while

39

having an average deposit balance of approximately SLRs10, 500 or $150. Unfortunately, HNB's is constrained both by the relatively small size of its branch network and consequent lack of rural infrastructure to expand its microfinance activities in rural areas. Savings and Development Banks: One of the recent developments was the establishment of two small specialized banks, Sanasa Development Bank and Pramuka Savings and Development Bank. Sanasa Development Bank was established in 1997 and was based on the Sanasa movement of TCCSs, and is has a micro-lending portfolio in its business operations. It is also be engaged in overseas-funded poverty lending activities operated by the CBSL. However, the operations of Pramuka Bank is rather different, since it is urban based and as a result the capability to reach the rural network is very much limited. The Pramuka Bank is mainly engaged in savings mobilization from urban and corporate sources. Regional Development Banks: The RDBs were created in 1997, by amalgamating the previously operating RRDBs (eight RDBs were to be formed from the seventeen RRDBs). Several reasons were behind the amalgamation; firstly to be consistent with provincial boundaries and political decentralization. Secondly, to specialize RDB management and realize economies of scale. Thirdly, and to increase the CBSL independence over RDB's operations, since the central bank owned all shares under the RRDBs system, and their chairmen usually politically appointed after the Monetary Board nominating board members for approval by the minister of finance. The RDBs plays an active role in government poverty alleviation lending. Furthermore the RDBs have the obligation to develop the governments Samurdhi program; " promoting and developing projects of Samurdhi beneficiaries " within the province (Parliament of

40

Sri Lanka 1997). RDBs are also involved in lending to beneficiaries of the SFLC project, both directly to groups of the poor and indirectly to NGOs and cooperatives which onlend to the poor.

3.2.2 Cooperatives in Microfinance Under cooperatives, there are two types of institutions involved in providing financial services to the poor: cooperative rural banks (CRBs) and thrift and credit cooperative societies (TCCs). Both institutions are regulated by the Department of Cooperatives. Cooperative rural banks: CRBs and TCCCs are very effective in mobilizing savings, and account for more than 5.3 million individual savings accounts in the country. They also accounted for more deposits in total than RRDBs. However they are less noteworthy as lending institutions, usually having a loan/deposit ratio of less than 40 percent. According to Attanayake(1997), the performance of CRB loans are not satisfactory, since 75 percent of the loans have been given non-income-generating purposes, such as pawning, housing, and consumption. The study notes: [They] have not played an important role in providing credit facilities for their rural membership to undertake income-generating activities [and] as in the case of commercial banks have not been innovative in rural lending. They do not have adequately trained staff to handle development credit. They also lack an effective credit supervision and follow-up system to ensure the proper use of credit by the borrowers and to undertake recovery at field level.

41

Furthermore, CRB are exposed to a high degree of politicization as a result of being attached to multi-purpose cooperatives, thus hindering the capacity to adopt a sustainable lending approach. It is also reasonable to believe that CRBs could become even more vulnerable, if the institution is expand its lending activities. Even with all the disadvantages, the potential for effective microcredit is enormous, given the institution's outreach and infrastructure. Therefore, removal of CRB's from a politicized environment and rebuilding the institutional capacity are critical in attaining a significant achievement in rural financial intermediation. Thrift and credit societies: The thrift and credit cooperative societies in Sri Lanka are structured according to a federation referred to as the Sanasa movement. According to Kiriwandeniya (1998), at the end of 1997, the total membership was 786,000 in 8400 TCCs. This indicates the ability of the Sanasa movement to reach a significant proportion of households in country with a population less than 20 million people. The primary societies are mainly engaged in mobilizing savings and providing loans for small business activities and consumption. However, still a larger proportion of its membership consists of the village elite and the salary employed, despite the movements effort to include the village poor. The Sanasa movement also serves as a channel for several microcredit schemes funded by both government and internationally. Sanasa expanded its activities available to its members by establishing an insurance company in 1992, and later the Sanasa Development Bank in 1997. The Sanasa Development Bank started with an initial subscribed capital of $1.43 million, and has established a large number of "agency banks", while preserving their society status.

42

2.2.3. NGOs in Microfinance There is a large number of NGOs in Sri Lanka, comprising of small, local, and with religious or social welfare objectives. Furthermore, there are also community-based organizations created by government agencies for the purpose of serving as agents in official programs. The most well-established movement working in the sphere of microfinance and rural development is the Sarvodaya Economic Enterprises Development Services (SEEDS). SEEDS was established as a company limited by guarantee, with the purpose of building a specialized development bank under CBSL regulation by 2003. This is to be expanded on the base of an existing network of "village banks," of which some 250 were operating in mid-1999, with a further 2,200 "societies" (potential village banks) altogether serving some 294,000 people. SEEDS accumulated savings amounted to SLRs323 million ($4.6 million) in late 2000, an increase of SLRs31 million for the quarter ended 30 June 1999, while loan disbursements were SLRs97 million ($1.4 million) for the quarter of 2000. The repayment rate (amounts paid on current loans as a proportion of loans currently due) was 93 percent for the quarter, while portfolio at risk (defined as the proportion of current portfolio more than 12 months overdue) was 14.5 percent (SEEDS 1999). According to Gunatilaka(1997), SEEDS acting primarily as a provider of savings should necessarily be considered noteworthy: "Some of the better-established small savings programs may not be accessible to individual small savers. This is because they prefer to exclude themselves from the credit and microenterprise components of an on-going program because of the risks involved In such cases, the catchment of savers could be widened by making it possible for such poor to save regularly through an easily accessible savings

43

program" (Gunatilaka 1997, 32). The societies and "village banks" are listed as membership institutions under the Societies Ordinance, 1891. SEEDS adopted a "credit plus" approach of microfinance, by establishing separate divisions for banking and training operations. This is evident in the loan portfolio for the June quarter of 1999; borrowers for the quarter amounted to 5,839 (of which 60 percent were women) from a society membership of 294,000 individuals. It appears that most of the members were participating for access to savings services and other benefits, rather than credit. The mean loan size was estimated around SLRs16,300 ($230) which, when compared to GNP per capita of $800, imply that the bulk of loans can be categorized under the microcredit range for Sri Lanka.

2.2.4 Samurdhi Poverty Alleviation Program In 1995 the Sri Lankan government launched Samurdhi poverty alleviation program. The Samurdhi program is Sri Lanka's largest microfinance and welfare expenditure program implemented in 22 districts of the island. Currently 51 percent of the total population are recipients of the program. The program is financed by general revenue and has consumed about half of the total welfare budget of the economy excluding health and education. The total cost of the program is estimated to be 4 percent of the government revenue, or almost 1 percent of GDP (Table 3.1). The Samurdhi program consists of 3 main components, 1) welfare componentfood stamp, insurance 2) group savings and credit component 3) integrated rural development. The monthly welfare grant( food stamp) and the insurance scheme falls under the protect ional component of the program. The protectional component of the

44

program acts as a temporary method of cushioning poverty by the use of an income transfer. The savings/credit schemes and the integrated rural development element fall under the promotional component of the program. The promotional element of the program contains medium and long-term goals for eradicating poverty mainly through institutional support. An important note to point is that the program is a poverty alleviation program rather than a social security scheme. The main goals of the Samurdhi Program are: 1) Broadening opportunities for income enhancement and employment for poor. 2) Integrate youth, woman and other disadvantaged groups in to economic and social development activities. 3) Developing latent talents and skills of the youth and increasing their marketability. 4) Increase the productiveness of rural assets to generate additional income and employment opportunities. The administration of the program is under the ministry of Samurdhi, Youth and Sports and the department for poor relief. The administration structure of the program starts from the top national bodies and extends down to village level animators or development officers. Since the program was launched in 1995, it has recruited approximately 25000 youths for various administrative tasks. However one of the key drawbacks of the program is that the administrative costs are very high. Salaries and wages alone accounted for around 10 percent of the total expenditure in the program. With respect to the administrative structure, several key points need to be noted. Firstly from the launch of the program (in 1995) until 2001 administrative jobs/recruits increased by approximately 25 percent. Most of the increase is as a result of the expansion of the

45

program to the North and East regions of the country. However it is also evident that most of the recruitment was done to gain political mileage, since there was a 20 percent increase in the recruitment of development officers in the run-up to the general elections. Secondly salaries paid to development officers and managers have doubled since the launch of the program in 1995. Furthermore all positions have been made permanent as a result of strikes and requests made. Thirdly as a result of politicization of the program, most of the development officers employed are political appointees. Due to this fact, several problems arose in the execution of the program in areas where the popularity for the government is weak. There were occasions when programs and projects were even sabotaged in such areas. 1) Group-Savings and Credit Component: The main objective behind group savings is to increase the savings habit of the poor. Under this scheme a group of 5 people contribute a small amount weekly or fortnightly towards the group fund. The group fund is then either deposited at a Samurdhi bank for an interest rate or invested in a common project. By June 2000, the total group savings of the program was Rs. 2264 million. The main weakness of the scheme are ; 1) most of the members are too poor to save 2) Since the group is involuntary formed by development officers the members are heterogeneous. As a result the breakdown rate is high. 2) Rural Infrastructure Development Component: The objective of this scheme is to increase the stock of rural infrastructure. For small scale projects(gravel a road, re roof a school) 80 percent of the funds are contributed by the government. Labor contribution by the members account for the remaining 20 percent. Most of the projects are identified, designed and implemented by the members with some technical assistance provide by the

46

government. Large-scale projects(irrigation canal, dams, roadways and bridges) are fully financed by the government. Most often program beneficiaries are hired for these projects, which as a result provides off-season employment in the villages. Since the inception of the program, approximately Rs. 2 billion has been spent on nearly 41000 projects. The main weaknesses encountered with project implementation are, 1) ad-hoc planning 2) poor maintenance 3) haphazard use of land and resources. 3) Welfare Component (food stamps, insurance): The main element of the welfare component is the consumption grant transfer(food stamp). Monthly coupons are issued to beneficiaries which can be later used to purchase goods from the local co-operative store. The face value of a coupon given to a household is either Rs 1000, Rs 500, Rs 250, Rs 200, Rs 100. Firstly households earning less than Rs 1500 a month are identified. The identified families are then divided in to 4 categories (see table 3.2). Finally the beneficiaries can claim their entitlements in either cash or kind depending on the category they belong to. Deductions for savings and insurance are made at the point of collection. Furthermore, a requirement for beneficiaries is contributing "voluntary" labor. The amount of labor that should be contributed is based on the size of the grant. For an example, a beneficiary need to contribute 4-5 man-days towards community development projects if the recipient's grant is Rs 500. When benefiting household's income exceeds Rs 2000 and remains as such for 6 continuous months or when a family member finds employment, the household must exit the program. However most of the time rules and limits of the exit criteria are difficult to enforce.

47

One of the key problems associated with the food transfer program is related to sufficiency. The perception of most recipients was that the grant was not adequate enough to have a significant impact. When a beneficiary receives Rs 1000 or Rs 500 the real net amount received is only Rs 755 or Rs 375, since deductions for savings and insurance are made at the point of collection. Moreover, the real value is further reduced as a result of efficiency issues via the delivery system. One positive aspect of the program is that it is unlikely to a disincentive to labor supply. The welfare program (Janasaviya) launched by the former political regime issued out a larger grant but also created a disincentive to supply labor. The co-operative system is used to deliver grants in the program. However there are several inefficiencies associated with grant deliveries. Most of the food stamp recipients are not satisfied with the quality if food that are being issued to them. Recipients also complain that corrupt practices in terms of weights and measures are common in the co-operative delivery system. Furthermore a prosperous secondary market is in existence for the publicly provided goods. Most often the recipients sell their goods purchased from the co-operative to private traders (at prices below the market value). The cash they get are in turn is used to purchase goods not provided by the cooperatives. With respect to targeting the household eligibility threshold was set at approximately one-third of the national poverty line. However the central bank reported 50 percent of the households in the country are grant recipients. While the poverty rate was 20 percent. Therefore it can be concluded that many non-poor households receives grants. Almost all the practical problems in the targeting occurs in the political economic framework. Most of the targeting is politically biased since the development officers are political appointees and are subject to political pressure. This explains the fact that why

48

half of the country's population are program beneficiaries when less that a quarter is below the poverty line. In 1999 a survey was done based on a multistage stratified sample to judge the distributional effectiveness of the program. In the sample 40 percent of households reported receiving food stamps. It was revealed that the incidence of being a transfer recipient is progressive than any of the country's other functioning public transfer programs. Unfortunately the program did not capture the 36 percent of the households in the lowest expenditure quintile. It was revealed that the lowest two expenditure quintiles accounted for 60 percent of the total food stamp budget. While quintiles three, four and five accounted for the remaining 40 percent. The best way to judge the effectiveness of the program is to compare it with a similar program in another country. According to Grosh (1994) across Latin American countries in a targeted transfer program the two expenditure quintiles accounted for 70 percent of the total budget. In an untargeted program (primary health care and public education) lowest two quintiles accounted for 60 percent of the total budget. Therefore this suggests that the outcomes of the Samurdhi program is more similar to an untargeted program even though it is a targeted program However, according to Grosh (1994) a targeted transfer program has a administrative cost of around 9 percent of the total program cost, which is therefore comparable to the Samurdhi program administration cost of nine percent. Since microfinance is viewed as a "win-win strategy" for poverty alleviation, it is important to explore the existing evidence from the literature of the impact of microfinance programs. Evidence at micro-level: In spite of most microfinance institutions view income and employment generation as the explicit objective, poverty reduction remains the principal goal. The impact of program participation on income,

49

consumption and net worth of households is generally used as the criteria for judging the success of poverty reduction. The first microfinance institution to attract international attention was the Grameen Bank in Bangladesh. As a result being a pioneer in microfinance, most of the earliest impact evaluations were focused on the Grameen Bank. The first impact assessment of the Grameen Bank was done by Mahabub Hossain (1988). Hossain compared the welfare Grameen clients to eligible non-clients in Grameen villages as well as non clients in non-Grameen villages. His findings indicate that Grameen members average household income to be 43 percent higher than non-members in non-Grameen villages and 28 percent higher than eligible non-members in Grameen villages. Furthermore, he found out that per capita spending on food for Grameen members to be 8 percent higher than non-members in Grameen villages, and 35 percent more on food and 32 percent more on clothing than non-members in non-Grameen villages. The first most sophisticated econometric study on the impact of micro-credit on poor households was done by Pitt and Khandker (1998). The study was influential in a econometric context, since it was the first serious attempt to evaluate the impacts of micro-credit after controlling for selection bias and non-random program placement. The study was based on data collected by the World Bank and the Bangladesh Institute of Development Studies(BIDS) in 1991-1992. The survey includes 1978 households who were either Grameen, Bangladesh Rural Advancement Committee's (BRAC) or Bangladesh Rural Development Board's Rural Development Project 12 (RD-12), as well as non-participant households. The results of the study uplifted the confidence of microfinance after demonstrating highly positive effects on bank members and their families.

50

The main conclusions of the study are: Every additional taka lent to a woman adds an additional 0.18 additional taka to annual household expenditure- an 18 percent return to income from borrowing. The probability of girls school enrolment increases by 1.86 percent for every one percent increase in credit to Grameen women. A 10 percent increase in credit to member women increased the arm circumference of girls by 6 percent. One percent increase in credit to women increased the height-for-age of girls by 1.16 percent and 1.42 for boys. Later, Khandker (2005) extended the study by employing panel data (1998/1999 resurvey of World Bank-BIDS data). Khandker (2005) estimates revealed that for each additional 100 taka of credit to women increased total annual household expenditure by more than 20 taka. Furthermore, the use of panel data permitted to compare the poverty rates in 1991/1992 and 1998/1999. Results showed that moderate poverty in all villages declined by 17 percentage points, and 18 points decline in villages with microfinance, and a decline of 13 percentage points in non-program villages. The most unique microfinance evaluation was done by Asian Development Bank (ADB) economist Brett Coleman(1999). Coleman study involved two microfinance institutions in Northeast Thailand, the Rural Friends Association and the Foundation for Integrated Agricultural Management. Cole tackled the selection bias issue in a very clever manner. In order to determine who in the comparison villages would have chosen to participate in microfinance programs had they been available, he actually had

51

interested individuals sign up a year in advance- now it is possible to compare borrowers to people with the same `entrepreneurial spirit' who had not been offered credit. After controlling for selection bias, findings showed zero impact from microfinance programs but positive impact using naive estimates. Coleman's technique is highly credible and provides strong for the need to address selection bias. Later, the follow-up article of the study-Coleman (2002), disaggregated participation and impact by type of client and found: self-selected program participants are significantly wealthier than nonparticipants even prior to program intervention, and the wealthiest villagers are almost twice as likely to participate in the program than the poorer villagers. Moreover, some of the wealthiest villagers obtain a disproportionate share of program loan volume by virtue of holding influential positions as village bank committee members. Positive impact is seen largely in this wealthier group. Impact on rank and file members is significantly smaller than impact on the wealthy, and is largely insignificant. However, Coleman himself recognized why Thailand is too atypical for the results of such a study to provide evidence for the impact of microfinance in other countries. He noted that in the villages surveyed, 63 percent of households held the membership the Bank for Agriculture and Agricultural Cooperatives (BAAC), a state bank that provides subsidized credit to rural households with significantly larger average loans than village banks. Coleman claimed that since only 30 percent of BAAC members in the survey are women, only 19 percent of households (30 percent of 63 percent) in the survey included women who are BAAC members. Furthermore the fungibility issue of credit within the household is also not accounted. In essence, 63 percent of surveyed

52

households---male and female members of BAAC--- already had access to a significant amount of credit than the village banks provide. Coleman observed that the average household low-interest debt, not including village bank debt, was 31,330 baht16 (9,342 of which was held by women), while village bank loans only amounted to 7,500 baht. Coleman wrote: "In such an environment, it should not be surprising that loans of 1,500 to 7,500 baht would have a negligible impact."

3.3 Econometric Methodology and Data 2.3.1. The evaluation problem The basic problem in treatment evaluation involves the inference of a casual relationship between the treatment and outcome. In a canonical single treatment setting, one can observe (Yi , i , Di ), i,..., N , and the impact on Y from a hypothetical change in D, while holding constant. Such inference is the key feature of a potential outcome model, where the outcome variable of the treated state is compared to the outcome variable of the untreated state. However, it is impossible to simultaneously observe both states for any given individual. Thus, the problem is akin to one of missing data, which can be solved by techniques of casual inference carried out in terms of counterfactuals. The counterfactual question is: what would have happened to people who participated in a program (or received treatment) if they had not done so (or else had participated in another program). First, assume the setup of a randomized treatment assignment, where no one is included in the treatment group because the benefits of the treatment to that individual would be large, and no one is excluded because the expected benefit is small. Let the vector of observables be (Yi , i , Di ; i,..., N ) . Where Y is the scalar-value

53

outcome variable, is a vector of observables, and D a binary indicator of treatment (D takes the value of 1 if the treatment is applied, 0 otherwise). In the potential outcome framework, one can define as the difference between the outcome in the treated and untreated states = Y1 Y0 Where, Y1 is the outcome of the treated individual, and Y0 is the outcome of the untreated individual. It is important to note that is not directly observable since an individual cannot be observed in both states. The two key evaluation parameters that will be used in this study are average treatment effect (ATE) and average treatment effect on the treated (ATT). ATE is the expected effect of treatment on a randomly drawn person from the population. ATT is the mean effect of those who actually participate in the program. The ATE is important when the treatment has universal applicability and it is useful to consider the hypothetical gain from treatment to a randomly selected individual of the population. The ATT is important when considering the average gain from treatment for the treated. In sample analogues, the ATE and ATT can be defined as,
ATE =

1 N

[ ]
i =1 i

ATT =
N

1 NT

[
i =1

NT

| Di = 1]

where NT = i =1 Di . The computation of both ATE and ATT is straight-forward if i can be estimated. The empirical strategy to estimate the average outcomes for the participants had they in fact not participated (the counterfactual), is through substituting

54

the outcome of untreated individuals for whom the observable characteristics, , match those of the treated up to some selected degree of closeness. The treatment evaluation problem can be easily understood by writing the ATT as ATT=E(|D=1)=E(Y1|D=1)-E(Y0 |D=1) From the above equation, the problem of selection bias is straightforward, since the second term on the right side- E(Y0 |D=1) the counterfactual mean of the treated, is not observable. If the condition E(Y0 |D=1)=E(Y0 |D=0) holds, one can use the nonparticipants as the comparison group. But with non-experimental data this condition will not hold, since the components which determines the participation decision also determines the outcome variable of interest. Thus, the outcomes of the participants would differ even in the absence of program participation, leading to a selection bias. It may be the case that selection bias can be fully accounted for by observables characteristics (such as age, skill differences, etc.). In this case, selection bias can be eliminated simply by including the relevant variables in the outcome equation. But in practice, unobservable characteristics effecting participation can also influence outcomes. For an example, it may be that highly motivated individuals are more likely to participate in microfinance programs and are also more likely to have higher income and savings. On the other hand, program entry may be a function of administrator selection. It is reasonable to believe that administrators are discriminating between the less and better able, as a basis for program selection. If administrators are `creamskimming' by selecting the best for the program, then program effects will be overstated. Equally, if program administrators are targeting program resources on the least able then program effects may be understated.

55

The ATT can also be expressed as

E[Y1 | D = 1] E[Y0 | D = 0) = ATT + E (Y0 | D = 1] E[Y0 | D = 0]

The difference between the left hand side of the equation and the ATT is the self selection bias. The true parameter ATT is only identified if
E[Y0 | D = 1] E[Y0 | D = 0] = 0

The crux of the evaluation problem is that E[Y0 | D = 1] is unobservable. Practically three solutions exist: experimental, quasi-experimental, and non-experimental strategies identification strategies. The experimental approach solves the evaluation problem by gathering a set of individuals equally eligible and willing to participate and then randomly dividing in to two groups: the set of individuals who receive the treatment (treatment group) and a set of individuals who are denied treatment (control group). In this setting, random assignment serves as a perfect counterfactual, free from the bothersome self-selection problem, thus ensuring that no bias arises in comparing observed outcomes for treated and control units. Even though the experimental approach is the optimal method to estimate project impacts, in practice several problems exist. For an example, randomization may be unethical due to the denial of benefits (treatment) to otherwise eligible members, or it can be politically infeasible to provide an intervention to one group and not another. The second identification strategy relies on non-experimental evaluation methods. In non-experimental methods, program participants are compared to non-participants by controlling statistically for differences between participants and non-participants. The non-experimental approach is appropriate when it is not possible to randomly select a control group or identify a suitable comparison group through using techniques such

56

matching (discussed later). The non-experimental method applies statistical and econometric techniques such as instrumental variable(IV) and Heckman selection estimators. The primary drawback with the IV technique is due to the extreme difficulty in finding suitable instruments. To identify the treatment effect, one has to find at least one regressor which determines program participation but is not itself determined by factors which affect outomes. Practically, in a survey data set finding such instruments is a ardent task. Similarly, again for the Heckman selection estimator the most significant practical obstacle to successful implementation is the identification of suitable instruments. Additionally, Heckman selection estimates are highly dependent on the underlying distributional assumptions relating to the unobserved variables. Goldberger (1983) and Puhani (2000) shows that estimates can surprisingly sensitive if these assumptions are not met. The quasi-experimental approach identifies the true parameter ATT by constructing a comparison group using reflexive or matching techniques. In a reflexive comparison, the counterfactual is constructed by taking into account the situation of program participants before the program. The basic idea of reflexive comparison is to compare the outcome set of a group of individuals after participation in a program with outcomes of before participation and to view the difference as the difference as the estimate of ATT. The most commonly used econometric technique in reflexive comparisons is the difference in differences (DD) estimator. The DD technique estimates the ATT as the difference between the before-after estimate of participants and beforeafter estimates of non-participants. Alternatively, matching involves identifying nonprogram participants comparable in essential characteristics to participants, and then

57

compare differences in mean outcomes between these two groups to identify the impact of the program. The primary econometric technique that will be utilized in this study to solve the evaluation problem is matching or more precisely what's referred to as propensity score matching (PSM).

2.3.2. Propensity Score Matching The essential idea of propensity score matching (PSM) is to match participants and non-participants on their observable characteristics. Propensity score matching assumes that selection can be explained purely in terms of observable characteristics and that any selection on unobservables is trivial and they do not affect outcomes in the absence of treatment. The mean effect of treatment (participation) can be estimated as the average difference in outcomes between the treated and non-treated. Both nonexperimental and quasi-experimental approaches share one thing in common: when the counterfactual mean for the treated- E[Y0 | D = 1] , is not observed, one has to invoke `identifying assumptions' to estimate the casual effect of a program on the outcome. The first identification assumption in propensity score matching is referred to as the conditional independent assumption (CIA), and is expressed as
Y0 , Y1 D |

It states that outcomes are independent of program participation, after controlling for the variation in outcomes induced by differences in 9. Under the conditional independence assumption, the matching process is analogous to constructing an

Using Dawid's (1979) notation, represents independence.

58

experimental dataset in that, conditional on observed characteristics, the selection process becomes random. The second identification assumption is referred to as the overlap or matching assumption, written as 0 < Pr[ D = 1| ] < 1 This assumption implies that for each value of there are both treated and untreated individuals. In other words, for each participant there is another non-participant with a similar . A practical constraint that exists in matching is that when the number of covariates i increases, the chances of finding a match reduces. However, Rosenbaum and Rubin (1983) showed that matching on the propensity score p ( ) - the probability of participating in a program, could achieve consistent estimates of the treatment effect the same way as matching on all covariates.
Proposition 1 (Rosenbaum and Rubin 1983);

Let p( i ) be the probability of unit i having been as s i gn ed to treatment , defined as p( i ) Pr( Di = 1| i ) = E ( Di | i ). Assume that 0 < P( i ) < 1, for all i , and

Pr( D1, D2 ,...DN | 1, 2 ,... N ) = i =1,... N

the sample. Then

p ( i ) Di (1 p( i ))(1 Di ) for the N units in

{(Yi1 , Yi 0 ) Di }| i {(Yi1 , Yi 0 ) Di }| p ( i )

Corollary; If {(Yi1 , Yi 0 ) Di }| i and the asumptions of proposition 1 hold , then

|D =1 = E{E (Yi | Di = 1, P( i )) E (Yi | Di = 0, P( i )) | Di = 1},

59

assuming the expectations are defined. The outer expectation is over the distribution of
P( i ) | Di = 1 . The proposition implies that observations with the same propensity score

have the same distribution of the full vector of covariates, i . When estimating the propensity score, two decisions have to be made. Firstly, the evaluator needs to choose the appropriate model and the functional form to estimate the propensity score. In general any discrete model (such as a logit or a probit) can be used to estimate the propensity score. In a single treatment framework, logit and probit models gives similar estimates for the probability of participation. After choosing the model, the evaluator needs to decide the covariates to be included in the model. This is a knife-edge decision since including too many as well as too few variables might lead to undesirable consequences. Heckman, Ichimura and Todd (1997) show that that omitting important variables can seriously increase the bias in the estimate. However there also strong reasons to avoid over-parameterized models. According to Bryson, Dorset and Purdon (2002), there are two reasons why over-parameterized models should be avoided. Firstly, including too many extraneous variables might exacerbate the support problem. Secondly, although the inclusion of non-relevant variables will not bias the estimates or make them inconsistent, it can increase the variance. On the other hand, Rubin and Thomas (1996), strongly criticizes `trimming models' in the name of parsimony. If there are doubts about whether a variable is related to the outcome or not a proper covariate, they explicitly advise to include the variable in the propensity score estimation. It is important to note that the standard regression based method and propensity score matching differs significantly with regard to the choice of control variables. In a standard regression, preference is usually given to variables that one can argue are exogenous o

60

outcomes, but in propensity score matching the primary interest is in covariates (not good predictors) and thus including variables even when they are poor predictors. Analytic results and simulations of Rubin and Thomas (1996) suggests that variables with weak predictive ability for outcomes can still help minimize bias in estimating casual effects with propensity score matching. In essence, the main purpose of the propensity score estimation is not to predict selection in to treatment but to balance covariates and get closer to the observationally identical non-participant. Finally it should be noted that with `too good' data also problems can arise. If P ( ) = 0 or P ( ) = 1 , this implies that individuals with such characteristics either always or never participate. Therefore some random is needed to ensure that individuals can be observed in both states. After estimating the propensity score, the next decision to be made concerns the common support region(s). Enforcing the common support region ensures that any combination of characteristics observed in the participation group can also be observed among non-participants. The approach referred to as the `minima and maxima' condition will be used in all estimations in this chapter. The basic idea of this condition is to delete all participants, whose propensity score is smaller than the minimum and higher than the maximum in the non-participants. Therefore participants who fall outside the common support region will be discarded and for these individuals the treatment effect will not be estimated. When the proportion of lost individuals is small, this poses few problems. However, if there is a significant reduction in the sample size, then there are doubts about whether the estimated effect on the remaining individuals can be viewed as a representative of the full sample.

61

Having enforced the common support region, the next step is to choose the matching algorithm. The general formula for the matching estimator is
BM =

1 NT

i{ D =1}

[Yi1 w(i, j )Y j 0 ],
j

where BM denotes the matching estimator for the bias, 0 < w(i. j ) 1 , is the set of treated individuals, and j is an element of the set of matched comparison units. w(i, j ) represents a weighting function that depends on the specific matching estimator. Results will be presented for four matching algorithms: nearest-neighbor matching, caliper matching, radius matching and kernel matching. The nearest-neighbor matching method assigns a weight equal to one [ w(i, j ) = 1] , and takes each participant in turn and identifies the nonparticipant with the closest propensity score. The nearest neighbor method will be implemented with replacement, so that a non-participant can be used more than once as a match. A variant of the nearest-neighbor matching is caliper matching. The caliper matching method chooses the nearest-neighbor within a caliper of width , so that { j :| P( X i ) P( X j ) |< } where P( X ) is the propensity score. Therefore caliper matching imposes a form of quality control on the match by setting a tolerance level on the maximum propensity score distance. Dehejia and Wahba (2002) introduced a variant of caliper matching which is referred to as radius matching. In radius matching the idea is to use not only the nearest-neighbor within each caliper but all of the comparison members (non-participants) within the caliper. The final matching algorithm that will be used in the study is referred to as kernel matching. Kernel matching uses all the nonparticipants for each participant in the matching process. The kernel is a function that weights the contribution of each non-participant, so that more importance is attached to

62

those non-participants providing a better match. The Gaussian and the Epanechnikov will be used as weighting functions with kernel matching. The final step in propensity score matching is to assess the matching quality. Since the conditioning is not done on all covariates but on the propensity score, one has to check the ability of the matching procedure to balance the relevant covariates. Three measures will be used to judge the performance of the match: standardized bias, PseudoR and the t-Test. The standardized bias for each covariate as suggested by Rosenbaum and Rubin (1985) is defined as the percentage of the square root of the average sample variances in both groups. The standardized bias before matching is given by

SBbefore = 100

1 0 0.5 (V1 ( ) + V0 ( )) 1M 0 M , 0.5 (V1M ( ) + V0 M ( ))

SBafter = 100

where 1 (V1 ) is the mean (variance) in the treatment group before matching and 0 (V0 ) the analogue for the comparison group. 1M (V1m ) and 0 M (V0 M ) are the corresponding values for the matched samples. The basic idea of the Pseudo-R is to re-estimate the propensity score on the matched sample, that is only on participants and matched non-participants and compare the Pseudo-R's before and after matching. The Pseudo-R indicates how well the regressors X explain the participation probability. After matching there should be no systematic differences in the distribution of covariates. Furthermore one can also perform an t-Test to check if there are significant differences in covariate means of treated and

63

comparison units. Before matching differences are expected, after matching the covariates should be balanced in both groups and hence no significant differences should exist. There are two main advantages of propensity score matching over standard linear regression models. Firstly, propensity score matching does not rely on any functional form assumptions for the outcome equation. Regression models depend on the form of relationship (linear, log-linear, etc.) which may be inaccurate and which propensity score matching avoids. This is especially important when functional form restrictions are not justified by economic theory. Secondly, propensity score matching does not make the assumption of constant additive treatment effects across individuals as it is done in simple regression. In propensity score matching heterogeneous treatment effects are permitted.

2.3.3. Data The study is based on the latest Consumer Finances and Socio-Economic Survey (CFS) 2003/2004 conducted by the Central Bank of Sri Lanka. The CFS was initiated in 1953 and is the eighth multipurpose household survey that helps to capture long-term changes in the living standards of the country. The foremost achievement of the survey was its coverage of the Northern and Eastern provinces in the country after a time lag of 20 years. The CFS used the Census 2001 in constructing the distribution of housing units across sectors and districts within provinces. Unfortunately, the CFS population frame excluded three districts, Killinochchi, Mannar and Mulaitivu due to the ongoing civil conflict. The under-coverage due to excluding the three districts is estimated around 2

64

percent of the total estimated housing units in the country. The CFS used a two-stage stratified sampling procedure for the sample design. The survey is nationally representative and consists of 11722 households and a 50545 individual population. The survey collects information on a broad range of topics including demographic characteristics, household income and expenditure, literacy and education, household amenities and employment. The variables used in this study can be broadly categorized as following: 1) Household per-capita income 2) Household per-capita savings 3) Household per-capita expenditure 4) Household size (household size 3-4, household size 5-6, household size 7) 5) Gender (male, female) 6) Age 7) Ethnicity (Sinhala, Tamil) 8) Sector (urban, rural) 9) Education (primary, secondary, tertiary) 10) Employment employed) 11) Housing particulars ( household not owning house, living in a single house, living in a shanty/lineroom, floor type-cement, floor type-clay/mud/earth, main source of lighting-electricity, main source of lighting-kerosene, energy for cooking-LP gas, energy for cooking-firewood, source of water-common well) sector (regular worker-formal, casual worker-informal, self-

65

12) Microfinance participant/client (if any member of a household has participated in a microfinance program or had any financial transaction with a microfinance institution) 13) Province (central, southern, northern, eastern, north western, north central, Uva, Sabaragamuwa)

2.4. Results Results are presented for the individual samples stratified by income quintiles: bottom 20th percentile, 20th-40th percentile, 40th-60th percentile, 60th-80th percentile, top 20th percentile. Table 3.3 reports the 20th percentile estimates of the probit regression where the binary outcome takes a value one if the household is a client of a microfinance institution or a microfinance program. The results are generally unsurprising and reveals a number of significant covariates of program participation. The probability of a household participating in microfinance tends to increase with household size, and decrease with the age. Microfinance participants are also relatively more likely to be Sinhalese households and live in rural areas. The coefficient for Sinhalese is greater in both the magnitude and statistical significance relative to the Tamil. With regard to employment, program participation increases when the household is engaged in casual labor or self-employment. Self-employed households have a greater propensity in participation compared to the casual workers, which might be a indication of their better entrepreneurial skills that is needed to reap the full benefits of microfinance. Table 3.4 presents the results for the 20th-40th percentile estimates of the probit regression. Again the most significant covariate of program participation is household

66

size. Clearly, there is a positive and monotonic effect of household size on the probability of being a microfinance client. Household being a Sinhalese also has a positive impact on program participation. However, in contrast to the bottom income percentile, the significance of the estimate for age vanishes. But, still the coefficient for a household being a casual laborer remains statistically significant. There are also several regional and housing characteristic dummy variables functioning as important covariates for the probability of program participation. Table 3.5 reports the 40th-60th percentile estimates of the probit regression. Being consistent with the previous two income categories, household size still remains a positive covariate of program participation and the likelihood of being microfinance client increases when the household is Sinhalese. According to table 3.5, both the estimates for a household being a casual laborer and being self-employed increases the probability of program participation. Furthermore, results suggests that households with higher per capita expenditure are less likely to be clients of microfinance. Table 3.6 presents the results for the 60th-80th percentile estimates of the probit regression. Results indicate that the probability of being a microfinance client decreases with the household age. Being consistent with the previous three lower percentile estimates, yet again the likelihood of program participation increases when household belongs to the Sinhala ethnicity. Results also suggests that casual laborers and selfemployed house are more likely to be participants of microfinance. According to table 3.6, households not owning a house tends to exhibit a lower probability of being a microfinance participant.

67

Table 3.7 reports the 80th-100th percentile estimates of the probit regression. Results show that the probability of program participation to decrease with household age and to increase for households with seven members or more. According to table 3.7, the probability of Sinhalese households participating in microfinance programs is higher compared to households belonging to the Tamil ethnic minority. Households in selfemployment, living in houses with cement or clay floors are also more likely to be microfinance clients. However the likelihood of program participation decreases for households living in urban areas. Again it is important to emphasize that all the variables with weak predictive ability included in the probit regressions can be still helpful to minimize bias in estimating casual effects in propensity score matching, since the ultimate goal is to not to predict selection in to treatment but to balance covariates and get closer to the observationally identical non-participant. Next, the common support region can be examined by plotting a histogram of the propensity score. The common support is the region where the propensity score has a positive density for both treated and non-treated units. Figures 3.1-3.5 gives the frequency distribution of the propensity scores based on tables 3.3-3.7 for the participants (treated) and non-participants (untreated) of microfinance. Except for the 80th-100th percentile histogram (figure 3.5), all other histograms reveals that there is a substantial region of overlap, and a severe common support problem does not exist. Since the main purpose is not on the probability estimations (probit estimations) but to match households, it is encouraging to see that a large fraction of households from both groups (treated and untreated) gets an estimated probability in the range of 0.2 to 0.6.

68

Unfortunately, for the 80th-100th percentile untreated group is large relative to the treated group and there is a limited overlap between the two groups. As a result of this limited overlap, the impact estimates for the 80th-100th percentile will be statistically less robust relatively. Tables 3.8-3.12 present the results on covariate balancing. Each cell reports the average standardized bias of the different covariates before and after matching. It evident that the differences between the households in the treated and untreated groups are quite small before matching, and matching removes most of the existing bias for almost all covariates. A t-test of equality of means in the two samples of participants and nonparticipants was also conducted for each covariate. Results indicates that there is no systematic patterns of significant differences between the covariates in the treatment and non-treated groups after conditioning on the propensity score. Additionally, after matching the pseudo- R 2 is fairly low implying that the matching procedure is able to balance the characteristics of the treated and non-treated groups. Table 3.13 reports the percentage of all deleted treated observations, whose propensity score is smaller than the minimum and higher than the maximum in the untreated group. It can be seen that the number of lost individuals is very low, except when the tolerance level is set very high (when = 0.001 ). Therefore, in general the estimated effect on the remaining individuals can be viewed as a representative, since the proportion of lost individuals is small. The empirical analysis led in this chapter will focus on two parameters of interest when estimating treatment effects. First, the impact of microfinance on household percapita income and savings who were actually treated- i.e., the average treatment effect on

69

the treated (ATT). Second, what effect microfinance would have on a household drawn randomly from the population- i.e. the average treatment effect (ATE). If we assume homogeneous responses to treatment among households, these two effects will be identical. The ATT and ATE will differ, should the responses be allowed to vary across households. The first of policy makers concern is to of course to determine whether microfinance had any impact on household per-capita income and savings. Another important concern is whether the expansion of microfinance programs is worth considering. While the ATT provides answers to the question of the impact, ATE is required to go further and assess the opportunity of expanding microfinance. For instance, if only individuals with the largest expected gains participate in microfinance, then ATE will be smaller than ATT. Thus, a generalization of the program may generate a lower effect than the one indicated by ATT. Table 3.14 reports the estimated mean impacts on household per capita savings. The estimates of the average treatment effect on the treated (ATT) and the average treatment effect (ATE) obtained via propensity score matching, using four matching algorithms and imposing the minima and maxima common support. The results for the mean impact indicate that program participation significantly increases household per capita savings for the bottom four quintiles, though the magnitude varies by matching method. However, these gains are not visible for the households in the 80th-100th percentile, implying that even without participating in microfinance programs there would be no difference in savings for the richest quintile. But, it is important to note that the estimates of the richest quintile (80th-100th percentile) are highly unreliable, due to the common support problem revealed by the propensity score histogram (figure 3.5). Table

70

3.15 reports the estimated mean impacts on household per capita income. Even though there are sizable gains in household per capita savings for program participants, this is not evident for household per capita income. Results suggest that there is no impact on household per capita income across all quintiles.

3.5 Conclusion Microfinance for the poor has become a focus of attention in the Sri Lankan development community over the last several years. The microfinance revolution has built on innovations in financial intermediation that lowers the cost of risks of lending to poor households. Replications of the movement flagship, the Grameen Bank of Bangladesh, have now expanded universally. To date, there has been no comprehensive investigation of their impact on Sri Lankan household income and savings. Using Data from a nationally representative household survey, this chapter analyzed the impact of participating in microfinance on household per-capita income and savings employing a quasi-experimental approach. The study applied recent advances in propensity score matching methods to assess the impact microfinance on household income and savings. Since a baseline survey or randomization are not feasible options in this case, the study is well suited to matching methods. There are several attractive features associated with propensity score matching, including the potential to allow for heterogeneous impacts, while optimally weighting observed characteristics when constructing a comparison group. The technique is well suited due to its flexible (non-parametric) nature, not imposing exclusion restrictions or ad hoc assumptions about the functional form of impacts. The method eliminates selection bias due to observable differences between

71

program participants and non-participants. Although a rich data set has been used, permitting to match on a wide range of household characteristics, the likelihood always remains of latent unobserved factors being correlated with program participation and outcome variables. Results suggests that program participation increases with household size, being a Sinhalese, living in a rural area, and employed as a casual worker or selfemployed. Overall program participants benefit incidence is indeed pro-poor. With respect to household per-capita savings, program participation definitely has a positive impact for all low-income households. Household per-capita savings are significantly higher on average for participants of microfinance than for observationally identical nonparticipants. However, the overall results are rather discouraging for household per-capita income, since the impact estimates are negative for all estimated income quintiles. Finally, the principle message that emerges form the study is; there are quantitatively non-negligible, average gains from microfinance on household savings, especially for the poor.

72

Table 3.1: Social Transfers and social expenditure,1999 Social transfers Sri Lanka Percent of Rupees (million) Samurdhi Fertilizer subsidy Refugees Other Sub-total Other social expenditures Pensions Education Health Total 20,723 22,231 10,651 67,601 8145 1451 2661 1739 13,996 the sub-total 58.20 10.37 19.01 12.42 100.00

Percent of the total 12.05 2.15 3.94 2.57 20.70

30.65 32.89 15.76 100

73

Table 3.2: Income Transfer Component Category of family Amount received Families earning less Rs. 1000 than Rs. 500 and having more than 5 members

Distribution of coupons 1) Rs. 400-only to purchase food items 2) Rs. 375-to buy goods or to encash 3) Rs. 25- insurance premium 4) Rs.200- compulsory saving Only for the purchase of food Only for the purchase of food 1) Rs. 200- only to purchase food items 2) Rs. 175- to buy goods or to encash 3) Rs. 25- insurance premium 4) Rs. 100- compulsory savings Can either buy goods or encash

Families with only 1 member Families with 2 members Others

Rs. 100 Rs. 200 Rs. 500

Janasaviya recipients

Rs. 250

74

Table 3.3: Bottom 20th percentile probit model for the propensity score
Variable Age Female Married Household size 3-4 Household size 5-6 Household size 7 Sinhala Tamil Primary education Secondary education Tertiary education Regular employee Casual employee Self-employed Urban sector Rural sector Household expenditure Household not owning house Lives in a single house Lives in a slum/shanty/lineroom Floor type- clay/mud/earth Floor type-cement Main source of lighting- electricity Main source of lighting- kerosene Energy for cooking- LP gas Energy for cooking- firewood Source of water- common well Central province Southern province Northern province Eastern province North western province North central province Uva province Sabaragamuwa province Constant Coefficient -0.0072* 0.0130 0.0390 0.3200* 0.5533* 0.5074* 0.4563* 0.2720 0.0233 -0.0058 0.1010 -0.0044 0.2050** 0.2764* 0.4456** 0.5409** -4.4100 -0.1116 0.3917 0.3551 0.0204 -0.0500 -0.4440 -0.3452 -0.2634 0.0949 0.0114 0.2025 0.5200* -0.2272 -0.0768 0.2237 0.0246 0.2390 0.3268* -1.5600* Standard Error 0.0022 0.0866 0.0873 0.0760 0.0923 0.1646 0.1250 0.1477 0.0780 0.0836 0.1973 0.1622 0.0885 0.0860 0.2119 0.1710 2.8000 0.1144 0.2202 0.2563 0.2787 0.2760 0.2327 0.2314 0.3950 0.2720 0.0595 0.1222 0.1210 0.2342 0.1435 0.1282 0.1396 0.1290 0.1205 0.5822

Number of observations = 2346 Pseudo R 2 = 0.0921 1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.

LR chi 2 (35) = 288.46 prob > chi 2 = 0.0000

75

Table 3.4: 20th-40th percentile probit model for the propensity score
Variable Age Female Married Household size 3-4 Household size 5-6 Household size 7 Sinhala Tamil Primary education Secondary education Tertiary education Regular employee Casual employee Self-employed Urban sector Rural sector Household expenditure Household not owning house Lives in a single house Lives in a slum/shanty/lineroom Floor type- clay/mud/earth Floor type-cement Main source of lighting- electricity Main source of lighting- kerosene Energy for cooking- LP gas Energy for cooking- firewood Source of water- common well Central province Southern province Northern province Eastern province North western province North central province Uva province Sabaragamuwa province Constant Coefficient -0.0017 0.1148 0.0420 0.3066* 0.4746* 0.5233* 0.4101* 0.1237 -0.0538 -0.1347 -0.1972 0.1395 0.1945 0.1438 0.1396 0.2428 -2.6400 0.0749 0.1823 -0.2090 0.6070** 0.4668 0.0448 0.0091 -0.2382 0.0931 0.0568 -0.1179 0.2375** 0.3606 -0.0803 0.2783* 0.0375 -0.2050 0.0589 -1.8470* Standard Error 0.0022 0.0952 0.0998 0.0966 0.1040 0.1433 0.1176 0.1493 0.0969 0.0980 0.1653 0.1314 0.1014 0.1010 0.2007 0.1697 2.4900 0.1113 0.1842 0.2182 0.2681 0.2629 0.1744 0.1763 0.3150 0.2373 0.0605 0.1046 0.1010 0.2075 0.1361 0.1021 0.1220 0.1214 0.1040 0.5093

Number of observations = 2343 Pseudo R 2 = 0.0629 1)Bold indicates 10% significance level. 2)** indicates 5% significance level. 3) * indicates 1% significance level.

LR chi 2 (35) = 201.33 prob > chi 2 = 0.0000

76

Table 3.5: 40th-60th percentile probit model for the propensity score
Variable Age Female Married Household size 3-4 Household size 5-6 Household size 7 Sinhala Tamil Primary education Secondary education Tertiary education Regular employee Casual employee Self-employed Urban sector Rural sector Household expenditure Household not owning house Lives in a single house Lives in a slum/shanty/lineroom Floor type- clay/mud/earth Floor type-cement Main source of lighting- electricity Main source of lighting- kerosene Energy for cooking- LP gas Energy for cooking- firewood Source of water- common well Central province Southern province Northern province Eastern province North western province North central province Uva province Sabaragamuwa province Constant Coefficient -0.0030 0.0200 0.0435 0.1384 0.2303 0.2761 0.3754* 0.1630 0.0251 -0.0519 -0.2448 0.0031 0.3144* 0.2691* 0.1025 0.2538 4.0400** 0.0337 0.1167 -0.1310 0.1900 0.0678 -0.0298 0.1415 -0.1678 0.1014 0.0701 -0.0510 0.5240* -0.2410 0.1897 0.5160* 0.0918 0.0647 0.2510** -1.5070* Standard Error 0.0024 0.0988 0.1064 0.1130 0.1180 0.1483 0.1256 0.1681 0.1238 0.1228 0.1577 0.1174 0.1067 0.1020 0.2361 0.2273 2.0100 0.1200 0.1777 0.2479 0.2140 0.2018 0.1897 0.1953 0.2021 0.1760 0.0658 0.1007 0.0890 0.2218 0.1496 0.0941 0.1295 0.1341 0.1040 0.4960

Number of observations = 2344 Pseudo R 2 = 0.0807 1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.

LR chi 2 (35) = 251.47 prob > chi 2 = 0.0000

77

Table 3.6: 60th-80th percentile probit model for the propensity score
Variable Age Female Married Household size 3-4 Household size 5-6 Household size 7 Sinhala Tamil Primary education Secondary education Tertiary education Regular employee Casual employee Self-employed Urban sector Rural sector Household expenditure Household not owning house Lives in a single house Lives in a slum/shanty/lineroom Floor type- clay/mud/earth Floor type-cement Main source of lighting- electricity Main source of lighting- kerosene Energy for cooking- LP gas Energy for cooking- firewood Source of water- common well Central province Southern province Northern province Eastern province North western province North central province Uva province Sabaragamuwa province Constant Coefficient -0.004 0.0489 0.0856 0.0605 0.1135 0.1271 0.6440* 0.2379 0.1781 0.1424 0.0230 0.1587 0.2330** 0.2437** -0.3357 -0.1485 -5.8400 -0.3016** -0.0361 0.0997 0.3068 0.1645 0.0253 0.2571 -0.1377 0.1230 0.0900 -0.0426 0.4667** -0.2266 0.2042 0.4763* -0.1475 0.0112 0.1735 -1.5480* Standard Error 0.0024 0.1004 0.1122 0.1350 0.1391 0.1585 0.1333 0.1933 0.1663 0.1628 0.1780 0.1087 0.1102 0.1010 0.2512 0.2455 9.1200 0.1295 0.1541 0.2583 0.1886 0.1594 0.2416 0.2534 0.1816 0.1721 0.0793 0.1026 0.0900 0.2638 0.1681 0.0915 0.1283 0.1450 0.1090 0.5315

Number of observations = 2344 Pseudo R 2 = 0.0765

LR chi 2 (35) = 224.81 prob > chi 2 = 0.0000

1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.

78

Table 3.7: 80th-100th percentile probit model for the propensity score
Variable Age Female Married Household size 3-4 Household size 5-6 Household size 7 Sinhala Tamil Primary education Secondary education Tertiary education Regular employee Casual employee Self-employed Urban sector Rural sector Household expenditure Household not owning house Lives in a single house Lives in a slum/shanty/lineroom Floor type- clay/mud/earth Floor type-cement Main source of lighting- electricity Main source of lighting- kerosene Energy for cooking- LP gas Energy for cooking- firewood Source of water- common well Central province Southern province Northern province Eastern province North western province North central province Uva province Sabaragamuwa province Constant Coefficient -0.0073* -0.0846 0.0303 0.0479 0.1567 0.2891 0.8650* 0.3912 0.2080 0.1169 -0.0490 0.0721 0.2968** -0.0101 -0.6509 -0.4700 3.3900 -0.2006 0.0115 -0.5230 0.4039** 0.2512** 0.0949 0.2888 -0.1400 0.1309 0.0322 -0.1470 0.5799* -0.0225 0.2050 0.4065* -0.1909 0.1496 0.1950 -1.3960** Standard Error 0.0026 0.1125 0.1320 0.1485 0.1505 0.1654 0.1705 0.2287 0.2359 0.2303 0.2389 0.0987 0.1160 0.0983 0.3592 0.3570 3.3000 0.1312 0.1367 0.4249 0.1937 0.1050 0.2550 0.2860 0.2272 0.2277 0.1089 0.1225 0.1084 0.2870 0.1808 0.0957 0.1428 0.1759 0.1316 0.6222

Number of observations = 2344 Pseudo R 2 = 0.1089 1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.

LR chi 2 (35) = 278.73

79

Table 3.8: Matching quality indicators (covariate balancing) for the 20th percentile
Mean Variable age Sample Unmatched Matched female Unmatched Matched married Unmatched Matched Household size 3-4 Unmatched Matched Household size 5-6 Unmatched Matched Household size Treated 46.086 46.117 0.19362 0.19426 0.81298 0.81236 0.52145 0.52318 0.25633 0.25386 0.0374 0.03753 0.85149 0.85099 0.10231 0.10265 0.42024 0.42053 0.38724 0.38631 0.0242 0.02428 0.033 0.03311 0.42024 0.42053 0.41804 0.41722 0.033 0.03311 Untreated 49.888 46.638 0.26583 0.17439 0.74182 0.82119 0.46625 0.5287 0.17467 0.24945 0.03271 0.03753 0.67919 0.87748 0.21851 0.08389 0.42171 0.45585 0.36395 0.3819 0.02296 0.01987 0.0675 0.03753 0.36534 0.39735 0.3222 0.43377 0.06541 0.02649 standardized Bias% -26.8 -3.7 -17.2 4.7 17.2 -2.1 11.1 -1.1 19.9 1.1 2.6 0 41.5 -6.4 -32.1 5.2 -0.3 -7.2 4.8 0.9 0.8 2.9 -15.8 -2 11.3 4.8 19.9 -3.4 -15 3.1 79.6 82.7 57.8 87.2 -256.7 81 -2302.9 83.9 84.6 100 94.6 90 87.6 72.5 86.3 %reduction In |bias| -6.23 -0.73 -4.01 0.97 4 -0.43 2.61 -0.21 4.78 0.19 0.61 0 9.51 -1.47 -7.32 1.23 -0.07 -1.35 1.14 0.17 0.19 0.57 -3.61 -0.45 2.66 0.9 4.73 -0.64 -3.43 0.74 t t-test p>|t| 0 0.468 0 0.33 0 0.664 0.009 0.834 0 0.847 0.544 1 0 0.141 0 0.22 0.944 0.176 0.256 0.863 0.847 0.568 0 0.649 0.008 0.37 0 0.524 0.001 0.459

Unmatched Matched

Sinhala

Unmatched Matched

Tamil

Unmatched Matched

Primary education

Unmatched Matched

Secondary education

Unmatched Matched

Tertiary education

Unmatched Matched

Regular employee

Unmatched Matched

Casual employee

Unmatched Matched

Self-employed

Unmatched Matched

Urban sector

Unmatched Matched

80

Table 3.8 continued

Mean Variable rural sector Sample Unmatched Matched Household expenditure Treated 0.92409 0.92384 control 0.8142 0.92936 standardized Bias % 33 -1.7 95 %reduction In |bias| 7.49 -0.4 t t-test p>|t| 0 0.687

Unmatched Matched

7662.2 7664

8273.3 7382

-5.8 2.7 53.9

-1.31 0.73

0.19 0.465

Household not owning house

Unmatched Matched

0.08141 0.08168

0.18163 0.07395

-30 2.3 92.3

-6.82 0.55

0 0.583

Lives in house

single Unmatched Matched 0.94169 0.9415 0.85943 0.93488 27.7 2.2 91.9 6.29 0.52 0 0.601

Lives in a slum/shanty/lineroom

Unmatched Matched

0.0484 0.04857

0.11065 0.05298

-23.1 -1.6 92.9

-5.25 -0.38

0 0.702

Floor clay/mud/earth

typeUnmatched Matched 0.43234 0.43157 0.55776 0.5585 0.34586 0.40177 0.6444 0.57616 17.8 6.1 -17.8 -3.6 79.6 65.5 4.22 1.15 -4.21 -0.68 0 0.25 0 0.498

Floor type-cement

Unmatched Matched

Main source lighting- electricity

of Unmatched Matched 0.41034 0.4117 0.49408 0.43488 -16.9 -4.7 72.3 -3.98 -0.89 0 0.372

Main source lighting- kerosene

of Unmatched Matched 0.56876 0.56954 0.49548 0.55188 14.7 3.5 75.9 3.47 0.68 0.001 0.499

Energy for cookingLP gas

Unmatched Matched

0.0055 0.00552

0.01461 0.00331

-9.1 2.2 75.8

-2.06 0.63

0.04 0.527

Energy for cookingfirewood

Unmatched Matched

0.9879 0.98786

0.96381 0.99338

15.7 -3.6 77.1

3.52 -1.09

0 0.276

Source of common well

waterUnmatched Matched 0.38614 0.38631 0.14741 0.1479 0.17492 0.17439 0.34029 0.38852 0.16562 0.16998 0.10369 0.16446 9.5 -0.5 -5 -6.1 20.7 2.9 86.1 -21.2 95.2 2.26 -0.09 -1.18 -1.15 5 0.5 0.024 0.931 0.24 0.251 0 0.615

Central province

Unmatched Matched

Southern province

Unmatched Matched

81

Table 3.8 continued

Variable Northern province Sample Unmatched Matched Eastern province Unmatched Matched North province western Unmatched Matched North central province Unmatched Matched Uva province Unmatched Matched Sabaragamuwa province 0.11441 0.11479 0.07811 0.07837 0.13641 0.13687 0.10647 0.12804 0.07933 0.09272 0.10786 0.11921 2.5 -4.2 -0.5 -5.3 8.7 5.4 38.1 -1072.2 -66.8 0.6 -0.77 -0.11 -0.98 2.08 1.01 0.549 0.441 0.915 0.329 0.037 0.315 Treated 0.0154 0.01545 0.09021 0.09051 control 0.04036 0.00773 0.15031 0.07837 Bias % -15.2 4.7 -18.5 3.7 79.8 69 In |bias| -3.42 1.37 -4.27 0.83 t p>|t| 0.001 0.17 0 0.406

Unmatched Matched Pseudo- R

2

0.17382 0.17219

0.13848 0.17108

9.7 0.3 96.9

2.32 0.06

0.02 0.956

Unmatched Matched

0.092 0.012

82

Table 3.9: Matching quality indicators (covariate balancing) for the 20th-40 percentile
Variable Sample age Unmatched Matched female Unmatched Matched married Unmatched Matched Household size 3-4 Unmatched Matched Household size 5-6 Unmatched Matched Household size 45.708 45.708 0.16087 0.16087 0.87488 0.87488 0.50348 0.50348 0.34657 0.34657 0.07249 0.07249 0.84806 0.84806 0.0993 0.0993 0.37239 0.37239 0.48163 0.48163 0.03674 0.03674 0.07646 0.07646 0.46773 0.46773 0.36544 0.36544 0.04469 0.04469 0.91063 Mean Treated Control 46.124 45.535 0.17066 0.18371 0.85629 0.85799 0.51347 0.51043 0.29042 0.33664 0.06362 0.07547 0.69237 0.84608 0.21781 0.1003 0.36228 0.37537 0.49551 0.47071 0.04491 0.03178 0.12201 0.10526 0.40344 0.47071 0.34431 0.33664 0.07111 0.04767 0.81063 standardized Bias% -3.3 1.4 -2.6 -6.1 5.4 4.9 -2 -1.4 12.1 2.1 3.5 -1.2 37.6 0.5 -32.9 -0.3 2.1 -0.6 -2.8 2.2 -4.1 2.5 -15.3 -9.7 13 -0.6 4.4 6 -11.3 -1.3 29.2 88.7 -36.3 95.4 36.8 39.2 21.3 70.6 99.2 98.7 66.4 82.3 30.5 9.2 -133.4 58.5 %reduction In |bias| -0.77 0.28 -0.63 -1.21 1.3 0.99 -0.48 -0.28 2.9 0.42 0.85 -0.23 8.87 0.11 -7.71 -0.07 0.5 -0.12 -0.67 0.44 -0.98 0.55 -3.61 -2.01 3.12 -0.12 1.06 1.21 -2.67 -0.28 6.84 t t-test p>|t| 0.439 0.781 0.529 0.227 0.194 0.321 0.632 0.781 0.004 0.676 0.397 0.82 0 0.912 0 0.947 0.615 0.902 0.506 0.662 0.326 0.585 0 0.045 0.002 0.905 0.29 0.228 0.008 0.777 0

Unmatched Matched

Sinhala

Unmatched Matched

Tamil

Unmatched Matched

Primary education

Unmatched Matched

Secondary education

Unmatched Matched

Tertiary education

Unmatched Matched

Regular employee

Unmatched Matched

Casual employee

Unmatched Matched

Self-employed

Unmatched Matched

Urban sector

Unmatched Matched

rural sector

Unmatched

83

Table 3.9 continued

Variable Sample Household expenditure Unmatched Matched Household owning house not Unmatched Matched Lives in house a single Unmatched Matched Lives in a slum/shanty/lineroom 0.94935 0.94935 0.86228 0.94141 30.1 2.7 90.9 7.02 0.7 0 0.485 0.09037 0.09037 0.16392 0.1003 -22.2 -3 86.5 -5.23 -0.68 0 0.499 Mean Treated 10831 10831 control 11236 10505 standardized Bias% -3.7 3 19.4 %reduction In |bias| -0.89 0.62 t t-test p>|t| 0.371 0.538

Unmatched Matched

0.03277 0.03277

0.10329 0.0427

-28.3 -4 85.9

-6.55 -1.04

0 0.298

Floor typeclay/mud/earth

Unmatched Matched

0.28004 0.28004 0.71301 0.71301

0.21482 0.26415 0.76796 0.73188

15.2 3.7 -12.6 -4.3 65.7 75.6

3.65 0.71 -3.02 -0.84

0 0.476 0.003 0.4

Floor type-cement

Unmatched Matched

Main source of lighting- electricity

Unmatched Matched

0.57895 0.57895

0.63174 0.59285

-10.8 -2.8 73.7

-2.59 -0.56

0.01 0.573

Main source of lighting- kerosene

Unmatched Matched

0.39523 0.39523

0.34356 0.38332

10.7 2.5 76.9

2.57 0.49

0.01 0.625

Energy for cookingLP gas

Unmatched Matched

0.00993 0.00993

0.02695 0.00596

-12.7 3 76.7

-2.94 0.89

0.003 0.371

Energy for cookingfirewood

Unmatched Matched

0.98014 0.98014

0.95135 0.9861

15.9 -3.3 79.3

3.7 -0.92

0 0.355

Source of watercommon well

Unmatched Matched

0.36445 0.36445 0.12115 0.12115 0.18173 0.18173 0.01986 0.01986

0.29566 0.37041 0.18413 0.11619 0.11452 0.16882 0.04491 0.02085

14.7 -1.3 -17.6 1.4 19 3.6 -14.2 -0.6 96 80.8 92.1 91.3

3.53 -0.25 -4.16 0.31 4.61 0.68 -3.31 -0.14

0 0.805 0 0.759 0 0.497 0.001 0.888

Central province

Unmatched Matched

Southern province

Unmatched Matched

Northern province

Unmatched Matched

84

Table 3.9 continued

Variable Sample North province western Unmatched Matched North province central Unmatched Matched Uva province Unmatched Matched Sabaragamuwa province 0.0864 0.0864 0.07051 0.07051 0.06886 0.08937 0.10105 0.07051 6.6 -1.1 -10.9 0 100 83 1.58 -0.21 -2.59 0 0.114 0.833 0.01 1 0.1857 0.1857 0.11751 0.19662 19.1 -3.1 84 4.64 -0.56 0 0.579 Mean Treated control standardized Bias% %reduction In |bias| t t-test p>|t|

Unmatched Matched Pseudo- R

2

0.13406 0.13406

0.11826 0.13505

4.8 -0.3 93.7

1.14 -0.06

0.253 0.954

Unmatched Matched

0.109 0.018

85

Table 3.10: Matching quality indicators (covariate balancing) for the 40th-60th percentile
Mean Variable age Sample Unmatched Matched female Unmatched Matched married Unmatched Matched Household size 3-4 Unmatched Matched Household size 5-6 Unmatched Matched Household size Treated 46.056 46.17 0.13982 0.14172 0.89374 0.89229 0.47204 0.47619 0.38479 0.38435 0.08725 0.08277 0.88591 0.88435 0.06488 0.06576 0.32662 0.32313 0.54586 0.55215 0.06152 0.06236 0.11633 0.11791 0.39374 0.38662 0.40828 0.4127 0.06711 0.06803 0.91387 Control 46.537 46.219 0.17655 0.13265 0.87241 0.89796 0.49655 0.45692 0.33517 0.36168 0.07379 0.10317 0.77034 0.87868 0.13655 0.06236 0.26759 0.36054 0.56897 0.52721 0.10414 0.05329 0.19586 0.10884 0.31241 0.35147 0.35379 0.44785 0.13724 0.0771 0.81241 standardized Bias% -3.9 -0.4 -10.1 2.5 6.6 -1.8 -4.9 3.9 10.3 4.7 4.9 -7.5 31 1.5 -24 1.1 12.9 -8.2 -4.7 5 -15.5 3.3 -22 2.5 17.1 7.4 11.2 -7.2 -23.3 -3 29.8 87.1 35.5 56.8 88.6 78.7 -8 36.6 95.3 95.1 -51.7 54.3 21.4 73.4 75.3 89.9 %reduction In |bias| -0.92 -0.08 -2.34 0.5 1.55 -0.35 -1.15 0.73 2.44 0.89 1.17 -1.33 7.06 0.33 -5.44 0.26 3.06 -1.49 -1.09 0.95 -3.55 0.74 -5.06 0.54 4.04 1.38 2.65 -1.34 -5.29 -0.66 6.76 t t-test p>|t| 0.359 0.939 0.019 0.618 0.122 0.726 0.249 0.465 0.015 0.375 0.241 0.184 0 0.74 0 0.793 0.002 0.135 0.274 0.344 0 0.462 0 0.588 0 0.168 0.008 0.179 0 0.508 0

Unmatched Matched

Sinhala

Unmatched Matched

Tamil

Unmatched Matched

Primary education

Unmatched Matched

Secondary education

Unmatched Matched

Tertiary education

Unmatched Matched

Regular employee

Unmatched Matched

Casual employee

Unmatched Matched

Self-employed

Unmatched Matched

Urban sector

Unmatched Matched

rural sector

Unmatched

86

Table 3.10 continued

Mean Variable Household expenditure Sample Unmatched Matched Household owning house not Unmatched Matched Lives in house a single Unmatched Matched Lives in a slum/shanty/lineroom 0.96421 0.96372 0.91172 0.95692 21.9 2.8 87 4.92 0.66 0 0.51 0.06264 0.06349 0.11379 0.06349 -18.1 0 100 -4.13 0 0 1 Treated 14940 14985 Control 16653 15724 standardized Bias % -11.3 -4.9 56.9 %reduction In |bias| -2.54 -1.13 t t-test p>|t| 0.011 0.261

Unmatched Matched

0.01678 0.01701

0.04828 0.01814

-17.8 -0.6 96.4

-3.97 -0.16

0 0.87

Floor typeclay/mud/earth

Unmatched Matched

0.18456 0.17347 0.79866 0.80952

0.11586 0.178 0.86207 0.80272

19.3 -1.3 -16.9 1.8 89.3 93.4

4.65 -0.23 -4.06 0.33

0 0.822 0 0.745

Floor type-cement

Unmatched Matched

Main source of lighting- electricity

Unmatched Matched

0.71477 0.72449

0.81034 0.69955

-22.6 5.9 73.9

-5.4 1.04

0 0.297

Main source of lighting- kerosene

Unmatched Matched

0.26063 0.25057

0.17034 0.28231

22.1 -7.8 64.8

5.29 -1.36

0 0.174

Energy for cookingLP gas

Unmatched Matched

0.03579 0.03628

0.09379 0.04082

-23.7 -1.9 92.2

-5.32 -0.45

0 0.656

Energy for cookingfirewood

Unmatched Matched

0.94519 0.94444

0.85931 0.94898

29.2 -1.5 94.7

6.57 -0.38

0 0.702

Source of watercommon well

Unmatched Matched

0.29418 0.29138 0.09955 0.10091 0.22371 0.21882 0.0179 0.01814

0.22207 0.29932 0.15448 0.11338 0.12138 0.21315 0.04828 0.02154

16.5 -1.8 -16.5 -3.8 27.3 1.5 -17 -1.9 88.8 94.5 77.3 89

3.93 -0.33 -3.81 -0.76 6.62 0.26 -3.81 -0.46

0 0.742 0 0.446 0 0.794 0 0.644

Central province

Unmatched Matched

Southern province

Unmatched Matched

Northern province

Unmatched Matched

87

Table 3.10. continued.

Mean Variable North province Sample western Unmatched Matched North province central Unmatched Matched Uva province Unmatched Matched Sabaragamuwa province 0.05481 0.05556 0.05145 0.05215 0.05586 0.03968 0.05655 0.04649 -0.5 6.9 -2.3 2.5 -11.2 -1408.5 -0.11 1.41 -0.53 0.5 0.914 0.158 0.598 0.62 0.19239 0.18934 0.10207 0.19728 25.7 -2.3 91.2 6.24 -0.38 0 0.704 Treated Control standardized Bias % %reduction In |bias| t t-test p>|t|

Unmatched Matched Pseudo- R

2

0.11409 0.11565

0.0931 0.14512

6.9 -9.7 -40.4

1.64 -1.66

0.102 0.098

Unmatched Matched

0.081 0.015

88

Table 3.11: Matching quality indicators (covariate balancing) for 60th-80th percentile
Mean Variable age Sample Unmatched Matched female Unmatched Matched married Unmatched Matched Household size 3-4 Unmatched Matched Household size 5-6 Unmatched Matched Household size Treated 47.207 47.251 0.13733 0.13673 0.90533 0.90483 0.44133 0.44236 0.396 0.3941 0.12 0.12064 0.916 0.91555 0.04667 0.04692 0.23867 0.23727 0.58933 0.58981 0.14133 0.14209 0.22533 0.22654 0.24933 0.24665 0.43333 0.43432 0.08667 0.08713 Control 48.216 47.531 0.15496 0.12869 0.88833 0.90617 0.46989 0.38874 0.35006 0.40885 0.11606 0.15952 0.79548 0.92091 0.10916 0.04021 0.20075 0.24263 0.57779 0.60054 0.18444 0.13941 0.27415 0.22788 0.2064 0.21046 0.38331 0.48257 0.1744 0.09651 standardized Bias% -8.3 -2.3 -5 2.3 5.6 -0.4 -5.7 10.8 9.5 -3.1 1.2 -12 34.8 -1.5 -23.5 2.5 9.2 -1.3 2.3 -2.2 -11.7 0.7 -11.3 -0.3 10.2 8.6 10.2 -9.8 -26.3 -2.8 89.3 3.5 15.7 97.3 93.8 7.1 85.9 89.3 95.6 -886.7 67.9 -87.8 92.1 54.4 72.3 %reduction In |bias| -1.85 -0.42 -1.12 0.42 1.25 -0.08 -1.29 1.92 2.16 -0.53 0.28 -1.98 7.4 -0.34 -4.98 0.58 2.09 -0.22 0.53 -0.39 -2.59 0.14 -2.52 -0.06 2.34 1.52 2.31 -1.71 -5.64 -0.57 t t-test p>|t| 0.064 0.676 0.264 0.676 0.213 0.936 0.196 0.055 0.031 0.596 0.782 0.048 0 0.73 0 0.563 0.036 0.825 0.597 0.7 0.01 0.892 0.012 0.955 0.019 0.129 0.021 0.088 0 0.567

Unmatched Matched

Sinhala

Unmatched Matched

Tamil

Unmatched Matched

Primary education

Unmatched Matched

Secondary education

Unmatched Matched

Tertiary education

Unmatched Matched

Regular employee

Unmatched Matched

Casual employee

Unmatched Matched

Self-employed

Unmatched Matched

Urban sector

Unmatched Matched

89

Table 3.11. continued.

Mean Variable rural sector Sample Unmatched Matched Household expenditure Treated 0.88933 0.88874 Bias % 0.79235 0.88606 standardized Bias % 26.7 0.7 97.2 %reduction In |bias| 5.78 0.15 t 0 0.881 t-test p>|t|

Unmatched Matched

23305 23344

24988 23619

-4.4 -0.7 83.7

-0.89 -0.2

0.375 0.839

Household owning house

not Unmatched Matched 0.05067 0.05094 0.101 0.03753 -19.1 5.1 73.4 -4.09 1.15 0 0.25

Lives in house

single Unmatched Matched 0.95067 0.9504 0.91656 0.96113 13.7 -4.3 68.6 2.97 -0.92 0.003 0.358

Lives in a slum/shanty/lineroom

Unmatched Matched

0.02133 0.02145

0.02949 0.01475

-5.2 4.3 17.8

-1.14 0.89

0.255 0.375

Floor typeclay/mud/earth

Unmatched Matched

0.104 0.0992 0.86933 0.87399

0.06148 0.09786 0.89523 0.88204

15.5 0.5 -8 -2.5 68.9 96.8

3.65 0.08 -1.85 -0.43

0 0.937 0.065 0.665

Floor type-cement

Unmatched Matched

Main source of lighting- electricity

Unmatched Matched

0.83733 0.84182

0.90715 0.85255

-21 -3.2 84.6

-4.96 -0.53

0 0.599

Main source of lighting- kerosene

Unmatched Matched

0.148 0.14343

0.0803 0.13807

21.4 1.7 92.1

5.08 0.27

0 0.786

Energy for cookingLP gas

Unmatched Matched

0.11867 0.1193

0.23149 0.13673

-30 -4.6 84.6

-6.48 -0.92

0 0.358

Energy for cookingfirewood

Unmatched Matched

0.85867 0.85791

0.72334 0.84853

33.7 2.3 93.1

7.3 0.47

0 0.64

Source of watercommon well

Unmatched Matched

0.19067 0.18767 0.08933 0.08981 0.192 0.19035

0.14053 0.19437 0.12986 0.07373 0.09975 0.19571

13.5 -1.8 -13 5.2 26.3 -1.5 94.2 60.3 86.6

3.12 -0.3 -2.85 1.04 6.26 -0.24

0.002 0.764 0.004 0.301 0 0.811

Central province

Unmatched Matched

Southern province

Unmatched Matched

90

Table 3.11. Continued.

Mean Variable Eastern province Sample Unmatched Matched North province western Unmatched Matched North province central Unmatched Matched Uva province Unmatched Matched Sabaragamuwa province 0.05467 0.05496 0.04133 0.04155 0.06399 0.06971 0.04391 0.0429 -3.9 -6.2 -1.3 -0.7 48.1 -58.2 -0.88 -1.08 -0.29 -0.12 0.379 0.282 0.774 0.906 0.19733 0.19571 0.10289 0.17828 26.7 4.9 81.5 6.33 0.79 0 0.431 Treated 0.03333 0.03351 Bias % 0.05332 0.03351 standardized Bias % -9.8 0 100 %reduction In |bias| -2.14 0 t 1 t-test p>|t| 0.033

Unmatched Matched Pseudo- R

2

0.096 0.09651

0.07403 0.09786

7.9 -0.5 93.9

1.82 -0.08

0.069 0.936

Unmatched Matched

0.077 0.013

91

Table 3.12: Matching quality indicators (covariate balancing) for the 80th-100 percentile
Mean Variable age Sample Unmatched Matched female Unmatched Matched married Unmatched Matched Household size 3-4 Unmatched Matched Household size 5-6 Unmatched Matched Household size Treated 49.014 49.025 0.11775 0.11818 0.91304 0.91273 0.37862 0.38 0.4058 0.40727 0.17572 0.17273 0.94022 0.94 0.03623 0.03636 0.17754 0.17455 0.55254 0.55455 0.25181 0.25273 0.31341 0.31455 0.18478 0.18182 0.37681 0.37818 Bias % 50.049 49.033 0.15123 0.08727 0.9029 0.92545 0.44029 0.40727 0.36998 0.38727 0.12333 0.16909 0.81417 0.94545 0.0904 0.03636 0.11272 0.20545 0.51674 0.51273 0.35156 0.25091 0.36105 0.27091 0.10882 0.21273 0.35379 0.40727 standardized Bias% -8.5 -0.1 -9.8 9.1 3.5 -4.4 -12.6 -5.6 7.4 4.1 14.7 1 39.1 -1.7 -22.4 0 18.5 -8.8 7.2 8.4 -21.8 0.4 -10.1 9.2 21.6 -8.8 4.8 -6 -26.4 59.3 8.4 98.2 -16.8 52.3 100 95.7 93.1 44.2 55.8 -25.5 7.7 99.3 %reduction In |bias| -1.75 -0.01 -1.96 1.57 0.71 -0.72 -2.56 -0.86 1.52 0.63 3.15 0.15 7.21 -0.36 -4.17 0 4 -1.21 1.47 1.29 -4.38 0.06 -2.05 1.48 4.71 -1.2 0.99 -0.92 t t-test p>|t| 0.081 0.993 0.05 0.117 0.477 0.473 0.01 0.39 0.129 0.529 0.002 0.882 0 0.718 0 1 0 0.225 0.141 0.197 0 0.949 0.04 0.14 0 0.232 0.325 0.359

Unmatched Matched

Sinhala

Unmatched Matched

Tamil

Unmatched Matched

Primary education

Unmatched Matched

Secondary education

Unmatched Matched

Tertiary education

Unmatched Matched

Regular employee

Unmatched Matched

Casual employee

Unmatched Matched

Self-employed

Unmatched Matched

92

Table 3.12 continued.

Mean Variable Urban sector Sample Unmatched Matched rural sector Unmatched Matched Household expenditure Treated 0.13587 0.13636 0.85326 0.85273 Variable 0.29855 0.15273 0.6942 0.83818 standardized Bias % -40.2 -4 38.7 3.5 90.9 89.9 %reduction In |bias| -7.71 -0.72 7.46 0.62 t 0 0.474 0 0.536 t-test p>|t|

Unmatched Matched

47187 47281

51170 42732

-4.5 5.2 -14.2

-0.88 0.99

0.381 0.32

Household owning house

not Unmatched Matched 0.0471 0.04727 0.09821 0.04727 -19.8 0 100 -3.75 0 0 1

Lives in house

single Unmatched Matched 0.95471 0.95455 0.89732 0.94727 22 2.8 87.3 4.15 0.52 0 0.605

Lives in a slum/shanty/lineroom

Unmatched Matched

0.00362 0.00364

0.01004 0.00364

-7.8 0 100

-1.43 0

0.152 1

Floor typeclay/mud/earth

Unmatched Matched

0.06341 0.06 0.86413 0.86727

0.02734 0.07273 0.80301 0.83818

17.4 -6.1 16.5 7.8 52.4 64.7

4 -0.79 3.26 1.26

0 0.431 0.001 0.206

Floor type-cement

Unmatched Matched

Main source of lighting- electricity

Unmatched Matched

0.88768 0.89091

0.94922 0.87091

-22.6 7.3 67.5

-5.14 0.95

0 0.342

Main source of lighting- kerosene

Unmatched Matched

0.09601 0.09273

0.04018 0.10364

22.3 -4.4 80.5

5.13 -0.56

0 0.573

Energy for cookingLP gas

Unmatched Matched

0.25362 0.25455

0.48828 0.24182

-50 2.7 94.6

-9.93 0.45

0 0.65

Energy for cookingfirewood

Unmatched Matched

0.73188 0.73091

0.48605 0.75091

52 -4.2 91.9

10.37 -0.7

0 0.483

Source of watercommon well

Unmatched Matched

0.11957 0.11818 0.06159 0.06182

0.07533 0.11636 0.09542 0.05636

14.9 0.6 -12.6 2 83.9 95.9

3.25 0.09 -2.46 0.36

0.001 0.931 0.014 0.722

Central province

Unmatched Matched

93

Table 3.12 continued.

Mean Variable Northern province Sample Unmatched Matched Eastern province Unmatched Matched North province western Unmatched Matched North province central Unmatched Matched Uva province Unmatched Matched Sabaragamuwa province 0.05435 0.05455 0.03623 0.03636 0.05357 0.06 0.02734 0.05091 0.3 -2.4 5.1 -8.3 -63.7 -602.5 0.07 -0.36 1.08 -1.1 0.944 0.718 0.28 0.273 0.19384 0.19273 0.09208 0.18909 29.4 1 96.4 6.58 0.14 0 0.887 Treated 0.01268 0.01273 0.03442 0.03455 Variable 0.0279 0.02182 0.04967 0.03636 standardized Bias% -10.8 -6.5 -7.6 -0.9 88.1 40.3 %reduction In |bias| -2.03 -1.07 -1.49 -0.15 t t-test p>|t| 0.042 0.283 0.135 0.88

Unmatched Matched Pseudo- R

2

0.07246 0.07273

0.05134 0.07091

8.8 0.8 91.4

1.88 0.11

0.06 0.914

Unmatched Matched

0.109 0.018

94

Table 3.13: Individuals lost due to common support requirement (%)

Matching Algorithm NN NN (caliper) =0.01 =0.001 Radius =0.01 =0.001 Kernel Epanechnikov (bw=0.1) Epanechnikov (bw=0.2) Gaussian (bw=0.1) Gaussian (bw=0.2) 0-20th percentile 0.255 0.68 15.68 0.68 15.68 0.255 0.255 0.255 0.255 20th-40th percentile 0.59 0.72 8.83 0.72 8.83 0.59 0.59 0.59 0.59 40th-60th percentile 0.59 0.63 9.85 0.63 9.85 0.59 0.59 0.59 0.59 60th-80th percentile

80th-100th percentile 1.74 1.91 20.26 1.91 20.26 1.74 1.74 1.74 1.74

1.27 1.57 11.51 1.57 11.51 1.27 1.27 1.27 1.27

95

Table 3.14: Impact of microfinance on household savings

Matching Algorithm NN NN (caliper) =0.01 =0.001 Radius =0.01 =0.001 Kernel Epanechnikov (bw=0.1) Epanechnikov (bw=0.2) Gaussian (bw=0.1) Gaussian (bw=0.2) 0-20th percentile ATE ATT 293.09 57.60 20th-40th percentile ATT ATE 393.67 423.17 393.73 423.24 429.06 344.64 297.52 431.54 427.05 426.34 298.05 391.71 244.77 303.28 236.60 296.97 186.37 231.80 40th-60th percentile ATT ATE 60.98 69.62 60.98 62.95 72.05 -40.27 252.24 507.07 214.44 45.48 216.10 347.12 168.78 140.34 153.3 116.52 2.90 -67.91

60th-80th percentile ATT ATE 678.71 54.74 678.71 54.49 861.58 403.52 286.79 113.73 857.09 704.32 242.88 142.03 195.75 194.47 195.14 202.00 137.67 194.00

80th-100th percentile ATT ATE -1616.28 -1387.18 -1617.07 -1389.55 -291.68 -1538.20 -1329.95 -1628.91 -1063.78 -1893.89 -1325.03 -1755.08 -1449.91 -1812.34 -1469.92 -18176.38 -1723.51 -1920.53

293.49 56.40 287.05 70.82 251.03 313.88 371.42 333.58 215.26 261.43 158.06 210.17 148.13 192.93 37.73 58.06

96

Table 3.15: Impact of microfinance on household income

Matching Algorithm NN NN (caliper) =0.01 =0.001 Radius =0.01 =0.001 Kernel Epanechnikov (bw=0.1) Epanechnikov (bw=0.2) Gaussian (bw=0.1) Gaussian (bw=0.2) 0-20th percentile ATE ATT -20.87 -11.70 20th-40th percentile ATT ATE -9.84 14.07 -8.33 14.75 -4.92 9.81 -18.94 -28.06 -16.86 -12.05 -25.64 -33.65 -63.93 -61.54 -68.86 -67.93 -123.33 -123.47 40th-60th percentile ATT ATE -154.95 -65.52 -154.95 -67.12 -175.1 -81.35 -39.35 -34.78 -90.09 -68.90 -62.34 -48.72 -90.88 -104.84 -96.79 -112.34 -161.48 -180.82

60th-80th percentile ATT ATE -100.70 71.95 -100.70 68.21 -101.56 80.54 -58.47 -82.05 -27.72 12.44 -82.59 -97.25 -132.17 -148.64 -142.39 -159.24 -226.94 -233.12

80th-100th percentile ATT ATE -951.69 -1047.76 -936.33 -1043.14 -985.76 -877.35 -911.85 -1276.17 -1152.63 -1016.00 -1060.08 -1541.74 -1397.12 -1853.59 -1471.09 -1893.93 -2181.90 -2346.53

-21.71 -12.50 -16.62 -19.95 4.85 -8.91 -12.12 -14.81 -13.48 -19.04 -44.37 -45.96 -50.74 -51.48 -113.91 -111.53

97

Treated
80 80

Untreated

60

Frequency 40

20

.2

.4 Estimated p(x)

.6

.8

0
0

20

Frequency 40

60

.2

.4 Estimated p(x)

.6

.8

Figure 3.1: Histogram of the estimated propensity score (bottom 20th percentile)

98

Treated
100 100

Untreated

80

Frequency 60

40

20

.2

.4 Estimated p(x)

.6

.8

0
0

20

40

Frequency 60

80

.2

.4 Estimated p(x)

.6

.8

Figure 3.2: Histogram of the estimated propensity score (20th-40th percentile)

99

Treated
100 100

Untreated

Frequency 50

.2

.4 Estimated p(x)

.6

.8

0
0

Frequency 50

.2

.4 Estimated p(x)

.6

.8

Figure 3.3: Histogram of the estimated propensity score (40th-60th percentile)

100

Treated
150 150

Untreated

100

Frequency

50

.2

.4 Estimated p(x)

.6

.8

0
0

50

Frequency

100

.2

.4 Estimated p(x)

.6

.8

Figure 3.4: Histogram of the estimated propensity score (60th-80th percentile)

101

Treated
150 150

Untreated

100

Frequency

50

.2

.4 .6 Estimated p(x)

.8

0
0

50

Frequency

100

.2

.4 Estimated p(x)

.6

.8

Figure 3.5: Histogram of the estimated propensity score (80th-100th percentile)

102

CHAPTER 4 RETURNS TO EDUCATION IN SRI LANKA: QUANTILE REGRESSION ANALYSIS

4.1. Introduction Two centuries ago Adam Smith stated that the capital stock of a country is partly made out of ..the acquired and useful abilities of all the inhabitants or members of the society. The acquisition of such talents, by the maintenance of the acquirer during his education, study, or apprenticeship, always costing a real expense, which is a capital fixed and realized, as it were, in his person. Those talents, as they make a part of his fortune, so do they likewise of that of the society to which he belongs. The improved dexterity of a workman may be considered in the same light as a machine or instrument of trade which facilitates and abridges labor, and which, though it costs a certain expense, repays that expense with a profit. ( The Wealth of Nations, 19776, book 2, chapter 1). Unfortunately, Smiths words were quickly forgotten, and the concept of capital was defined narrowly in terms of nonhuman elements in production. But, about 50 years ago the concept of human capital reemerged like a new finding, giving rise to range of research topics. Several inquiries were made in relation to human capital, such as; why invest in education and training? What are the returns to the individual and to the economy? How much should be spent on education, and who should bear the expenses?

103

Analyzing the changes in skill differentials across cohorts (genders, ethnicities, sectors of work, etc.) are imperative for skill formation, school enrolment and training efforts. Most of the wage distribution studies are concerned with estimating the unexplained wage gap for groups of people. Unexplained wage gaps arise when two individuals possessing equal abilities and skills are employed in comparable jobs, but exhibit differences in earnings. These wage gaps arising from wage and employment discrimination leads to a reduction in earnings and total product. Therefore policies that help to eliminate discrimination in labor markets are able to raise both the earnings of the discriminated group (generally more vulnerable to poverty than the dominant group) and the total product. Previous studies rarely dealt with the returns to schooling by race and gender. Duraisamy (2000) estimated the returns to schooling by gender and location for India. The main findings were that the private rate of return to schooling increases as the level of education increases up to the secondary level. According to the study, returns to women education exceed that to men at the secondary level. Rural-urban comparison of returns reveal that there are higher returns to education in rural than in urban areas for primary and secondary levels, but returns for college education are higher for individuals in urban areas. Chase (1998) investigates how returns to education in the Czech Republic and Slovakia changed after the collapse communism in these countries. According to the study, there was a significant increase in returns to schooling with the transition to noncommunist countries. Results indicated that returns to schooling rose 2.4% to 5.2% for Czech men between 1984 and 1993. Furthermore, results suggested the returns to schooling for men increases more with regime change. Taber (2001) estimates a dynamic

104

programming selection model to examine the return to schooling in the U.S. during the 1980s. The study finds a large increase in the gap between returns to high school and college during the decade. According to Taber (2001), the increase in college earnings premium is due to the rise in the relative demand for high ability workers and not due to an increase in demand for skills accrued in college (which is revealed by a significant increase in the variance of wage residuals). Barrow and Rouse (2005) used the U.S. Decennial Census and the National Longitudinal Surveys to examine whether returns to schooling differ by race and ethnicity. The study found little evidence of differences in returns to schooling across racial and ethnic groups, even after controlling for ability and measurement biases. According to Barrow and Rouse (2005), there is no evidence that returns to schooling are lesser for African-American or Hispanics compared to nonminorities. The principle objective of this chapter is to analyze the returns education in the Sri Lankan labor market by ethnicity employing the quantile regression technique for each conditional quantile wage group rather than mean regression analysis used in most labor market analysis. Quantile regression results suggest that returns to education are positive and significant across all quantiles. However, a comparison of wage returns to education between ethnic groups reveals that returns are higher for Sinhalese workers than for Tamil Workers.

4.2. Human Capital Framework, Signaling and the Returns to Schooling The analysis of returns for education is based on the concept of human capital which was first introduced by Gary Becker, Jacob Mincer and Theodore Schultz. The

105

human capital approach views education as an investment of current resources with the expectation of future returns. According to Becker (1964), the amount of education (s) is undertaken to maximize the expected present value of future income until the retirement date (T), after accounting for the costs of education (c). Thus, the condition is given by:

T s

Ws Ws 1 = Ws 1 + Cs t =1 (1 + rs )

Where rs is referred to as the internal rate of return. Internal rate of return can be interpreted as the discount rate that equates the present value of benefits to the present value of costs. Optimal rule is to invest in the s th year of education if rs > i , where i is the market interest rate. Investment in schooling takes place until the marginal cost of education is equal to the marginal return to education. An individual will prefer more education if the internal rate of return is greater than the market interest rate. It can also be seen that individuals with higher discount rates prefer more value current income than future income, and hence will prefer less schooling. The net benefits of schooling are also lower when the direct costs of schooling increases. The level education will also rise if the wage gap between the educated and the uneducated increases or if more income received while being in school. In the theory of human capital, the empirical estimation of the earnings function takes the form:
log wi = i + edui + exp + exp 2 + i Where wi is the hourly wage, edu years of schooling, exp is the years of experience and i represents set of other variables such gender, sector of work, etc. Experience is

106

included with a quadratic term to account for the concave earnings pattern. The

above

earnings function is commonly referred to as the Mincerian earnings specification. Mincer (1974) analyzed the returns to education using 1960 census data and estimated that returns to education to be 10% and returns to experience as 8%. The Mincerian specification is also used in many other studies including, gender and racial discrimination, effectiveness of labor training programs and in the estimation of returns to beauty. In the Mincerian specification, schooling is identified as exogenous even when education is an endogenous variable. Furthermore, the error term captures the unobservable factors (such as ability), and some of these factors affect schooling leading to a correlation between the schooling and the error term. One solution to the problem is to include a proxy for ability in the earnings function. Some studies have also used twins to analyze differences in earnings and education on the assumption that the unobserved factors are additive and common among twins. There are two effects when differences in ability are introduced to the earnings function. Firstly, the internal rate of return will increase for the more able individuals since they are capable of converting schooling into human more efficiently. Secondly, the possibility also exists for a reduction in the internal rate of return when the ability to succeed in school is positively correlated with the ability to increase earnings in the labor market (high opportunity costs education). The optimal amount of schooling will also differ between individuals for reasons other than differences in ability. For an example, marginal rates of substitution may differ between individuals resulting in different discounts across individuals. Discount rates might vary for several other reasons such as differences in access to funds or different tastes and preferences for schooling. However if ability levels are the same for all

107

individuals then the outcome is straightforward, where individuals with low discount rates choosing more schooling. In an empirical analysis, the estimated returns to education may be influenced by sample selection. Since low productive individuals are most of the time non-participants, there is very good chance of OLS estimates being bias. If the selected sample of workers truncates the lower part of the wage distribution then the outcome would be to inflate the mean of the distribution. As a result of OLS going through the mean of the sample, the estimates will be affected by the truncation of data. For an example, if the participation of women is less compared to men then the returns to education for women would be biased downwards. But, since the median is not affected by truncation, estimates based on median regressions will not encounter this problem. Furthermore, returns to education for individuals in the lower segment of the wage distribution may not be the same as the returns to education for individuals in the upper segment of the wage distribution. If this is the case, then an alternative technique to OLS is the use of quantile regressions. Quantile regressions permit to estimate the returns to education at various points the wage distribution. Table 4.1 presents the rate of return to education for several countries using the International Social Survey Program (ISSP) data. The estimates are based on OLS regressions, after controlling only for age and union status. Thus it is reasonable to believe that these returns to fall if additional variables (such as experience) are included in the model. Results indicate that Great Britain, Northern Ireland and the republic of Ireland are experiencing high returns to education relative to other countries. Figure 4.1 gives some cross-country rates of return to education broken down by gender. Estimates

108

indicates that for countries such as UK, Ireland, Germany, Greece and Italy experience a significant variation in the returns to education between men and women. Since the supply of educated workers have increased in the last two decades (especially in OECD countries), there are doubts that the supply of educated workers have exceeded the demand for these workers, thus leading to an over-education in the workforce. Over-education implies that skills workers possess are more than what is required for the job. Under this situation, workers will experience lower returns to education. Therefore in order to obtain a correct picture one has to break up an individuals total years of schooling in to required years and surplus years of education. Over-education can be measured in several ways such as directly posing a question on workers of whether they are over-educated. Alternatively, one can take the difference between the actual years of education and the education required by the employer on chapter. Three hypotheses concerning the over-education in the United Kingdom graduate labor market was tested by Dolton and Vignoles (2000). Firstly, the study concluded that the return to surplus years of education quite different from the return to required years of education. Secondly, the study revealed that return to surplus years of education is the same for all different degree classes. For an example, an over-educated worker with first-class degree has the same return as an over-educated worker with a second-class. Final conclusion of the study was that returns to surplus education are the same across private and public sectors. Spence (1973) developed an alternative view of the labor market based on job market signaling. According to the signaling model, employers have limited and uncertain information about potential workers and they use education as a sorting devise

109

to find the potential productivity of a worker. According to Spence (1973), it is difficult for employees to determine the productivity of a new worker until the worker has been inside the firm for some time. Therefore, when a firm evaluates job candidates it has to foretell how productive workers will be after training and the likelihood they are going to stay with the firm after training. In this instance, firms use the level and type of schooling to predict the productiveness of workers. In the model schooling acts as a signal only if two conditions are met. Firstly, the costs incurred to an individual must be inversely related to the value a firm places on a worker. The underlying idea is that the cost of schooling will be high for slow learners, and slow learners are most likely to be slow learners on the job too. This condition ensures that slow and unproductive workers have a relatively less incentive of devoting their time and investing in school. Secondly, there should not be any less expensive or more accurate signals available to the firm when evaluating job candidates. For an example test scores might be a cheap alternative years of education for measuring cognitive skills. But test scores are less important when trying to ascertain attitudinal characteristics compared to other methods such as interviews and references. In the signaling model, the estimates returns to education will not only produce the productivity enhancing effect of schooling but also the underlying ability that education signals. Even after controlling for education, if schooling still shows a strong effect on earning then the difference can be interpreted as signaling effect of education. Alternatively, one can estimate the returns to education for the self-employed where education does not act as a signal since individuals know their own skills and have no incentive to signal themselves by obtaining more schooling. The main drawback with self-employed individuals is that they possess specific (and usually unobservable)

110

characteristics that makes them self-employed so self-employment is not random. The main problem with ability measures is that they should not be contaminated with education. Furthermore, ability measures need to capture the aptitude to make money and increase earnings rather than measuring ability in an intelligence quotient (IQ) aspect.

4.3. Econometric Methodology and Data The standard model used to returns to education is based on the human capital earnings function developed by Mincer (1974). The first human capital wage equation to be estimated has the form: Model 1: ln wi = 0 + 0 edu + 0 expi + 1 exp sqri + i Where ln wi is the natural log of the hourly wage, edu are years of schooling, exp are years of experience and exp sqr is the square of the level of experience. It is important to note that model-1 is based on some restrictive assumptions. Firstly, the equation assumes that workers have equal abilities and face equal opportunities. Secondly, the coefficient for edu is an estimate of the impact of schooling on wages, and it is not an internal rate of return on investment. If it is to be an internal rate if return, still it would be a private one, since model disregards any subsidizing of schooling and also overlook any positive or negative externalities associated with schooling. Next, a model that disaggregates the years of schooling in to education levels was estimated in order to capture differences in returns to education between education levels. Generally, it is observed that returns to an additional year of education are not identical across levels of schooling. The model is based on three levels of the Sri Lankan school

111

system; primary, secondary and tertiary education. Thus, the empirical estimation takes form; Model-2: ln wi = 0 + 0 primaryi + 1 sec ondaryi + 2tertiaryi + 0 expi + 1 exp sqri + i The estimates on the splines are interpreted in an identical manner as an estimate of a continuous education variable (as in model-1). An appropriate empirical strategy to estimate the above earnings models across different points on the conditional wage distribution is through the quantile regression method (Koenker and Basset, 1978). Quantile regression analysis is an attractive estimation method to overcome various shortcomings of OLS analysis. OLS estimates the effect of education on the mean of the conditional wage distribution, and assumes that possible differences in terms of the impact of education along the conditional wage distribution are unimportant. In contrast to OLS, quantile regression technique assumes the effect of education of the i th individual to differ across the earnings spectrum. The quantile regression model can be expressed as:
ln wi = xi + i ,

Where ln wi is the natural log of the hourly wage of the i th individual, and xi represents the demographic and employment related characteristics of the i th individual. By Imposing the assumption that the th quantile of the error term conditional on the regressors is zero, (Q (ui , | xi ) = 0) , the th conditional quantile of ln wi with respect to xi can be expressed as: Q (ln wi | xi ) = xi For any (0,1), the parameter can be estimated by

112

 = arg min

R k

i{i| yi xi }

| yi xi | +

y x (1 ) | | i i i{i| yi xi }

Note, that when = 0.5 , we have the special case known as the median regression or the least absolute deviation (LAD) estimator. Five quantile regressions were estimated at the 10th, 25th, 50th, 75th and 90th quantiles. The standard errors were computed by bootstrapping with 100 replications. OLS regression was also estimated for the purpose of comparison. The study is based on the latest Consumer Finances and Socio-Economic Survey (CFS) 2003/2004 conducted by the Central Bank of Sri Lanka. The CFS was initiated in 1953 and is the eighth multipurpose household survey that helps to capture long-term changes in the living standards of the country. The foremost achievement of the survey was its coverage of the Northern and Eastern provinces in the country after a time lag of 20 years. The CFS used the Census 2001 in constructing the distribution of housing units across sectors and districts within provinces. Unfortunately, the CFS population frame excluded three districts, Killinochchi, Mannar and Mulaitivu due to the ongoing civil conflict. The under-coverage due to excluding the three districts is estimated around 2 percent of the total estimated housing units in the country. The CFS used a two-stage stratified sampling procedure for the sample design. The survey is nationally representative and consists of 11722 households and a 50545 individual population. The survey collects information on a broad range of topics including demographic characteristics, household income and expenditure, literacy and education, household amenities and employment. The survey collects information on earnings and hours of work, thus making it possible to obtain hourly wage rates. The working sample in this

113

study contains workers who worked at least 30 hours per week and were not selfemployed. Self-employed workers were omitted due to the difficulty in separation of income in to returns to labor and returns to capital.

4.4. Empirical Results Table 4.2 presents the results of the full sample (both ethnic groups) estimates for model-1. OLS estimates of all variables are statistically significant, and returns to education are positive. Quantile regression results also suggest that returns to education are positive and significant across all quantiles. However, the returns to education vary significantly across all estimated quantiles. Figure 4.2 shows the development of the coefficients over the entire wage distribution. The estimated coefficient for each quantile is plotted as a continuous line and its 95% confidence interval is the shaded area. The OLS estimate is the dark horizontal line and parallel to it is the 95% confidence band. Results suggests that whereas the average return to education is of 1.9%, the return at the bottom 10th quantile is 0.63% and the return at the top 90th quantile reaches 4.6%. Returns to experience are also positive and significant across all quantiles, and shows a slight upward trend when moving from the lower quantiles to upper quantiles. The average return to experience is 1.6%, the return at the bottom 10th quantile is 1.1% and the return at the top 90th quantile is 1.9%. Thus, the returns to experience exhibit less variation across quantiles compared to returns to education (as seen clearly in figure 4.2). Table 4.3 reports the estimated returns to education and experience in model-1 for the Sinhala ethnic group. According to Table 4.3 all the estimated coefficients are statistically significant, and returns to education and experience are positive for all

114

quantiles for the Sinhala ethnic group. OLS estimates indicate that the average return to education for Sinhalese workers to be 2.4%. Across all quantiles, returns to education for the Sinhalese ethnic group tend to increase as one move up the conditional wage distribution. At the bottom 10th quantile, returns to education is 0.8% while workers in the upper 90th quantile exhibit a return of 4.8%. However, returns to experience tends to be relatively more uniform across quantiles for the Sinhala ethnic group. The average return to experience as indicated by OLS is 1.8%, while the bottom and top quantiles shows a return of 1.4% and 2.1% respectively. Table 4.4 reports the estimated returns to education and experience in model-1 for the Tamil ethnic minority. Results indicate that returns to education in the 10th quantile are not statistically significant for the Tamil ethnic minority. Table 4.4 also suggest that both the statistical significance and the magnitude of returns to education to increase systematically moving up the wage distribution. The average return to education for the Tamil ethnic minority in model-1 is 0.8%, while the returns at the bottom 10th and 90th quantiles are 0.3% and 3.5% respectively. According to Table 4.4 returns to education are positive for the along the entire wage distribution, even though there is a slight variation in the statistical significance across quantiles. Model-1 results indicate that returns to education are substantially different for Sinhalese and Tamil workers across quantiles. Firstly, for both ethnic groups returns to education tend to increase when moving from the bottom to the top part of the conditional wage distribution. Furthermore, Sinhalese workers have higher returns education than for Tamil workers especially in the lower quantiles of the distribution, but the difference in returns tends to be lesser in higher quantiles. The returns to education

115

also seem more homogeneous across the lower quantiles for Tamil workers than for Sinhalese workers. Similarly, the returns for experience are also higher and much more statistically significant for the Sinhalese workers than for Tamil workers. Table 4.5 reports the OLS returns as well as conditional returns at five representative quantiles: 0.10, 0.25, 0.5, 0.75, and 0.9, for model-2. The estimated results are for the full sample consisting both Sinhalese and Tamil ethnic groups. OLS estimates reveal that all the estimated coefficients are statistically significant at the 1% level. However according to the quantile regression estimates, the primary education coefficient is significant only at the 50th quantile. At all quantiles estimated on the conditional wage distribution, primary education is positively related with the hourly wage rate. Figure 4.5 shows the development of the coefficient representing primary education over the entire conditional wage distribution. In general, it is clear that the returns for primary education do not vary a lot along the wage distribution. The returns to secondary education are constant until the 80th quantile, while an upward trend is shown for the upper 90th quantile. Unsurprisingly, results indicate that tertiary education to have very strong effect on the hourly wage, conformed by both OLS and quantile regression estimates. Returns to tertiary education are especially higher at the upper 90th quantile. With respect to work experience, returns are statistically significant at all quantiles but the magnitude is less compared to schooling. Table 4.6 reports the estimated returns to education and experience in model-2 for the Sinhala ethnic group. For all estimated quantiles the statistical significance of primary education is low compared to secondary and tertiary education. As to be expected, tertiary education has the greatest effect on earnings followed by secondary and primary

116

education across all quantiles. With respect to experience, the returns are statistically significant and tend to be uniform across the entire wage distribution. Table 4.7 reports the estimated returns to education and experience in model-2 for the Tamil ethnic minority. Both OLS and quantile estimates suggest that primary education is not statistically significant across all quantiles for the Tamil workers. Furthermore, coefficient for the secondary education becomes statistically significant beyond the 25th quantile. The estimate for tertiary education is statistically significant and systematically increases as we move from the lowest to the highest quantile. Model-2 results indicate that returns to education at different levels are significantly different for Sinhalese and Tamil workers across quantiles. The main distinction between the two ethnic groups is that the returns to secondary education are slightly higher for Sinhalese workers for all estimated quantiles. Furthermore, there is upward trend in secondary education returns around the center of the wage distribution for Tamil workers, where as the returns to secondary education are generally uniform for the Sinhalese. However, the returns to tertiary education are greater for Tamil workers from the 50th quantile upwards. Figures 4.6 - 4.7 shows the development of all estimated coefficients over the entire wage conditional wage distribution for the two ethnic subsamples. For the sub-sample of Tamil workers it evident that the returns to tertiary education differs significantly between the lower and upper quantiles. Furthermore, at the extreme upper end of the distribution, the estimate of tertiary education falls outside the confidence interval of the OLS estimate and is quite different from the OLS estimate.

117

4.5. Conclusion Any extensive program of macroeconomic stabilization and rapid liberalization of trade in a developing economy are likely to generate crucial and widespread consequences for the labor market. In the latter half of the 1980s Sri Lanka experienced an increased openness of the economy leading to significant exportation of labor intensive products. Sri Lanka having a comparative advantage in labor-intensive sectors requiring unskilled labor would suggest an increase in demand for low-unskilled rather than for skilled labor. In view of such economic changes, it certainly would be interesting to observe the returns to education in the Sri Lankan labor market. This chapter undertakes an empirical examination of returns to education in the Sri Lankan labor market using the latest Consumer Finance and Socio-Economic Survey. The study employs the quantile regression method for each conditional quantile wage group, rather than the mean regression analysis used in most labor market analysis. The quantile regression technique fits hyperplanes through out the conditional wage and is ideal for characterizing the entire wage distribution. The standard Mincerian wage equation was estimated for the full sample of male workers and separately for the two ethnic groups in Sri Lanka. Quantile regression results suggest that average returns to education for both ethnic groups differs significantly from the returns at the two extreme ends of the wage distribution. In general the returns to education are positive for both groups, but the returns are higher for Sinhalese workers than for Tamil. An increasing trend in returns to education is evident for both ethnic groups when moving up wage distribution. Sinhalese workers experience higher returns to education than for Tamil especially at the bottom of the wage distribution, but the difference becomes less at the upper part of the distribution.

118

Estimated results with spline in years of education suggest that returns to secondary education are higher for Sinhalese workers, but the returns to tertiary education are greater for Tamil workers at the upper part of the wage distribution. Findings indicate that returns to experience are also higher for Sinhalese workers than for Tamil workers.

119

Table 4.1: Cross Country evidence on the Returns to Schooling (year-1995)

Source: Harmon, Oosterbeek and Walker (2000). Note: Standard Errors in italics.

120

Table 4.2: Quantile and OLS regression estimates (model 1-full sample)
Bootstrap standard error 0.001139 0.002136 5.660005 0.023498 0.001593 0.001683 4.980005 0.019620 0.001799 0.001388 4.520005 0.020842 0.001537 0.002179 7.210005 0.019220 0.001819 0.002074 6.180005 0.022178 0.001111 0.001321 3.680005 0.014925

coefficient q10 edu exp expsqr consant q25 edu exp expsqr constant q50 edu exp expsqr constant q75 edu exp expsqr constant q90 edu exp expsqr constant OLS edu exp expsqr constant 0.006319 0.011061 -0.000340 1.064279 0.011317 0.017371 -0.000490 1.154758 0.025822 0.019514 -0.000470 1.172323 0.035738 0.019776 -0.00040 1.232041 0.046208 0.019536 -0.000300 1.298972 0.019337 0.016851 -0.000380 1.246002

t-value 5.550 5.180 -5.960 45.29 7.100 10.32 -9.780 58.86 14.35 14.06 -10.39 56.25 23.25 9.080 -5.560 64.10 25.41 9.420 -4.890 58.57 17.40 12.76 -10.29 83.48

P>|t| 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

121

Table 4.3: Quantile and OLS regression estimates (model 1-Sinhala)

Bootstrap standard error 0.002173 0.003162 9.390005 0.029267 0.002322 0.001942 6.180005 0.027466 0.002308 0.001585 0.000054 0.026159 0.002080 0.002173 6.590005 0.023665 0.002029 0.002091 6.250005 0.024245 0.001355 0.001497 4.230005 0.017709

coefficient q10 edu exp expsqr _cons q25 edu exp expsqr _cons q50 edu exp expsqr _cons q75 edu exp expsqr _cons q90 edu exp expsqr _cons OLS edu exp expsqr _cons 0.008043 0.014276 -0.000460 1.068700 0.015461 0.019509 -0.000540 1.127452 0.030930 0.021360 -0.000480 1.128302 0.038618 0.020451 -0.000380 1.203375 0.048383 0.021267 -0.000330 1.270688 0.023714 0.018623 -0.000410 1.209412

t-value 3.700 4.520 -4.880 36.52 6.660 10.04 -8.690 41.05 12.99 13.48 -8.940 43.13 18.56 9.410 -5.780 50.85 23.84 10.17 -5.260 52.41 17.50 12.44 -9.690 68.29

P>|t| 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

122

Table 4.4: Quantile and OLS regression estimates (model 1-Tamil)

Bootstrap standard error 0.003189 0.004570 0.000108 0.054497 0.001845 0.003504 8.290005 0.037250 0.003252 0.004017 9.560005 0.051122 0.005574 0.004414 0.000131 0.058955 0.006039 0.006323 0.000198 0.063310 0.002080 0.003030 0.000078 0.031635

coefficient q10 edu exp expsqr constant q25 edu exp expsqr constant q50 edu exp expsqr constant q75 edu exp expsqr constant q90 edu exp expsqr constant OLS edu exp expsqr constant 0.002938 0.005744 -0.000110 1.006615 0.003880 0.007455 -0.000180 1.140844 0.008106 0.007632 -0.000200 1.275534 0.021085 0.013077 -0.000300 1.302319 0.034554 0.011229 -0.000100 1.347976 0.008564 0.008709 -0.000170 1.272718

t-value 0.920 1.260 -1.010 18.47 2.100 2.130 -2.200 30.63 2.490 1.900 -2.060 24.95 3.780 2.960 -2.270 22.09 5.720 1.780 -0.520 21.29 4.120 2.870 -2.240 40.23

P>|t| 0.357 0.209 0.311 0.000 0.036 0.034 0.028 0.000 0.013 0.058 0.039 0.000 0.000 0.003 0.024 0.000 0.000 0.076 0.605 0.000 0.000 0.004 0.025 0.000

123

Table 4.5: Quantile and OLS regression estimates (model 2-full sample)
Coefficient q10 primary secondary tertiary exp expsqr constant q25 primary secondary tertiary exp expsqr constant q50 primary secondary tertiary exp expsqr constant q75 primary secondary tertiary exp expsqr constant q90 primary secondary tertiary exp expsqr constant OLS primary secondary tertiary exp expsqr constant 0.048946 0.205679 0.443819 0.018826 -0.000420 0.859709 0.052829 0.208309 0.430995 0.018944 -0.000430 1.038040 0.052460 0.227014 0.461729 0.020518 -0.000450 1.167785 0.035789 0.213596 0.479334 0.018359 -0.000350 1.342088 0.038085 0.246998 0.577881 0.021994 -0.000370 1.454727 0.056244 0.239790 0.497390 0.019650 -0.000390 1.154330 Bootstrap Standard Error. 0.050912 0.049757 0.049191 0.002060 6.230005 0.049047 0.033157 0.031062 0.030209 0.001414 5.040005 0.030342 0.018224 0.016367 0.017562 0.001243 3.490005 0.016027 0.022852 0.021513 0.023509 0.002441 0.000078 0.023643 0.031410 0.028517 0.031498 0.002351 7.030005 0.028247 0.020290 0.019035 0.020172 0.001200 0.000030 0.020340 t-value 0.960 4.130 9.020 9.140 -6.720 17.53 1.590 6.710 14.27 13.40 -8.630 34.21 2.880 13.87 26.29 16.51 -12.90 72.86 1.570 9.930 20.39 7.520 -4.450 56.76 1.210 8.660 18.35 9.360 -5.290 51.50 2.770 12.60 24.66 16.28 -11.68 56.73 P>|t| 0.336 0.000 0.000 0.000 0.000 0.000 0.111 0.000 0.000 0.000 0.000 0.000 0.004 0.000 0.000 0.000 0.000 0.000 0.117 0.000 0.000 0.000 0.000 0.000 0.225 0.000 0.000 0.000 0.000 0.000 0.006 0.000 0.000 0.000 0.000 0.000

124

Table 4.6: Quantile and OLS regression estimates (model 2-Sinhala)

Coefficient q10 primary secondary tertiary exp expsqr constant q25 primary secondary tertiary exp expsqr constant q50 primary secondary tertiary exp expsqr constant q75 primary secondary tertiary exp expsqr constant q90 primary secondary tertiary exp expsqr constant OLS primary secondary tertiary exp expsqr constant 0.094427 0.269427 0.494878 0.021972 -0.000510 0.806024 0.081863 0.243767 0.456219 0.020349 -0.000470 1.009352 0.040893 0.212035 0.434629 0.022127 -0.000490 1.184314 0.042947 0.212639 0.464373 0.020530 -0.000410 1.339696 0.015132 0.226079 0.542847 0.023075 -0.000380 1.472891 0.061440 0.249500 0.499660 0.020880 -0.000410 1.147780 Bootstrap standard error 0.0542914 0.0546298 0.0555607 0.0027689 0.0000887 0.0564095 0.0317356 0.0307653 0.0319011 0.0015455 0.0000512 0.0309662 0.0342083 0.0342209 0.0364486 0.0013746 0.0000419 0.0339347 0.0294690 0.0304455 0.0330742 0.0023610 0.0000796 0.0329567 0.0488720 0.0497276 0.0529322 0.0021202 0.0000587 0.0482806 0.0264700 0.0247800 0.0257500 0.0013850 0.0000300 0.0262000 t-value 1.740 4.930 8.910 7.940 -5.780 14.29 2.580 7.920 14.30 13.17 -9.120 32.60 1.200 6.200 11.92 16.10 -11.58 34.90 1.460 6.980 14.04 8.700 -5.130 40.65 0.310 4.550 10.26 10.88 -6.550 30.51 2.320 10.07 19.40 15.07 -10.70 43.81 P>|t| 0.082 0.000 0.000 0.000 0.000 0.000 0.010 0.000 0.000 0.000 0.000 0.000 0.232 0.000 0.000 0.000 0.000 0.000 0.145 0.000 0.000 0.000 0.000 0.000 0.757 0.000 0.000 0.000 0.000 0.000 0.020 0.000 0.000 0.000 0.000 0.000

125

Table 4.7: Quantile and OLS regression estimates (model 2-Tamil)

Coefficient q10 primary secondary tertiary exp expsqr constant q25 primary secondary tertiary exp expsqr constant q50 primary secondary tertiary exp expsqr constant q75 primary secondary tertiary exp expsqr constant q90 primary secondary tertiary exp expsqr constant OLS primary secondary tertiary exp expsqr constant 0.049911 0.144278 0.435165 0.012022 -0.000220 0.880291 -0.024360 0.080193 0.365616 0.007349 -0.000160 1.135356 -0.003780 0.134977 0.452884 0.012599 -0.000250 1.206258 -0.005400 0.177777 0.538524 0.014200 -0.000270 1.328006 0.006410 0.186644 0.755078 0.008360 -0.000110 1.507489 0.026551 0.178914 0.508884 0.013254 -0.000230 1.162422 Bootstrap Standard error 0.064335 0.069195 0.111695 0.004306 0.000104 0.081788 0.045348 0.049733 0.065839 0.003440 9.730005 0.046479 0.025997 0.029174 0.056329 0.002366 6.550005 0.024628 0.048577 0.048616 0.073098 0.004617 0.000127 0.055418 0.052708 0.055268 0.124952 0.005918 0.000174 0.056284 0.033083 0.032569 0.041317 0.002708 6.930005 0.035884 t-value 0.780 2.090 3.900 2.790 -2.070 10.76 -0.540 1.610 5.550 2.140 -1.650 24.43 -0.150 4.630 8.040 5.320 -3.840 48.98 -0.110 3.660 7.370 3.080 -2.170 23.96 0.120 3.380 6.040 1.410 -0.610 26.78 0.800 5.490 12.32 4.900 -3.340 32.39 P>|t| 0.438 0.037 0.000 0.005 0.038 0.000 0.591 0.107 0.000 0.033 0.100 0.000 0.884 0.000 0.000 0.000 0.000 0.000 0.911 0.000 0.000 0.002 0.031 0.000 0.903 0.001 0.000 0.158 0.544 0.000 0.422 0.000 0.000 0.000 0.001 0.000

126

Source: Harmon, Oosterbeek and Walker (2000)

Figure 4.1 Returns to Schooling in Europe, Men and Women

127

0.06

.2

.4 .6 Quantile

.8

exp 0.01 0.01 0.01 0.02 0.03 0.03

0

0.00

0.02

EDU 0.04

.2

.4 .6 Quantile

.8

expsqr 0.00 0.00 0.00 0.00

0.00
0

.2

.4 .6 Quantile

.8

Figure 4.2 OLS and quantile regression estimates (model 1-full sample)

128

0.06

.2

.4 .6 Quantile

.8

exp 0.01 0.01 0.01 0.02 0.03 0.03

0

0.00

0.02

EDU 0.04

.2

.4 .6 Quantile

.8

expsqr 0.000.000.000.00 0.00 0.00

.2

.4 .6 Quantile

.8

Figure 4.3 OLS and quantile regression estimates (model 1-full Sinhala)

129

0.06

EDU 0.02 0.04

.2

.4 .6 Quantile

.8

0.01 0.00

0.02 0.00

exp 0.01
0

0.02

0.03

.2

.4 .6 Quantile

.8

expsqr 0.000.000.00 0.00 0.00

0

.2

.4 .6 Quantile

.8

Figure 4.4 OLS and quantile regression estimates (model 1-full Tamil)

130

secondary 0.10 0.15 0.20 0.25 0.30 0.35

0.15

primary 0.05 0.10

0.05

0.00

.2

.4 .6 Quantile

.8

.2

.4 .6 Quantile

.8

0.30
0

0.40

tertiary 0.50

0.60

0.70

.2

.4 .6 Quantile

.8

0.03

0.03

0.01

.2

.4 .6 Quantile

.8

0.00

0.01

0.00
0

expsqr 0.00

exp 0.02

0.00

.2

.4 .6 Quantile

.8

Figure 4.5 OLS and quantile regression estimates (model 2-full sample)

131

0.20

secondary 0.15 0.20 0.25 0.30 0.35 0.40

primary 0.00 0.10

0.10

.2

.4 .6 Quantile

.8

.2

.4 .6 Quantile

.8

0.30
0

0.40

tertiary 0.50

0.60

0.70

.2

.4 .6 Quantile

.8

0.04

.2

.4 .6 Quantile

.8

expsqr 0.00 0.00 0.00 0.00 0.00 0.00

0

0.01

0.02

exp

0.03

.2

.4 .6 Quantile

.8

Figure 4.6 OLS and quantile regression estimates (model 2Sinhala)

132

0.00
0 .2 .4 .6 Quantile .8 1 .8 1 .4 .6 Quantile .2 0

0.01

exp 0.02 0.03 0.04

primary 0.20 0.10 0.00 0.10 0.20 0.30

expsqr 0.00 0.00 0.00 0.00 0.00 0.00

0 .2 .4 .6 Quantile

secondary 0.10 0.00 0.10 0.20 0.30 0.40

0 .2 .4 .6 Quantile

Figure 4.7 OLS and quantile regression estimates (model 2Tamil)

133
.8 1

.8 1

tertiary 0.00 0.20 0.40 0.60 0.80 1.00

0 .2 .4 .6 Quantile .8 1

REFERENCES

Ackland, R., J. Braithwaite, M. Foley, T. Garner, C. Grootaert, B. Milanovic, T. Oh, S. Sipos, S. Tsirounian, Y. Ying (1997). Household Expenditure and Income Data for Transition Economies (HEIDE) mimeo, Policy Research Group, Washington, DC: World Bank. Appleton, S (2001). "The rich are just like us, only richer: Poverty functions or consumption functions?" Journal of African Economies 10 (4): 433--469. Ashenfelter, O. and Rouse, C. (1998), Income, Schooling, and Ability: Evidence from a New Sample of Identical Twins, Quarterly Journal of Economics, 113. Angrist, J. and Krueger, A. B. (1991), Does Compulsory School Attendance Affect Schooling and Earnings, Quarterly Journal of Economics, 106, pp. 979-1014. Attanayake, M. M. (1997). An Evaluation of the Role of Finance Institutions in Poverty Alleviation: Sri Lanka Experience. Economic Research Department. Colombo: Central Bank of Sri Lanka. Barro, R. and Lee, J. W. (1993), International Comparisons of Educational Attainment, Journal of Monetary Economics, 32(3), pp.363-394. Barrow and Rouse (2005) Do Returns to Schooling Differ by Race and Ethnicity? American Economic Review, Vol. 95, No. 2, May 2005. Becker, G. (1964), Human Capital: A Theoretical and Empirical Analysis, with Special Reference to Education, Columbia University Press: New York. Benhabib, J. and Speigel, M. (1994), The Role of Human Capital in Economic Development: Evidence from Aggregate Cross Country Data, Journal of MonetaryEconomics, 34(2), pp.143-173. Besley, T., S. Coate and G. Loury (1992b), On the Allocative Performance of Rotating Savings and Credit Associations, Discussion Paper # 163, Research Program in Development Studies, Woodrow Wilson School, Princeton, NJ, July Blackburn, M. and Neumark, D. (1995), Are OLS Estimates of the Return to Schooling Biased Downward?, Review of Economics and Statistics, 77(2), May. Blanchflower, D. and Elias, P. (1999), Ability, Schooling and Earning Are Twins Different, Mimeo, Dartmouth College. Blundell, R. W., Dearden, L., Goodman, A. and Reed, H. (1998), Higher Education,

134

Employment and Earnings in Britain, Institute for Fiscal Studies Paper. Buchinsky, M. (1994), Changes in the U.S. Wage Structure 1963-1987: Application of Quantile Regression, Econometrica, 62(2), March, pp.405-458. Bryson A., Dorsett R., and Purdon S. (2002). The Use of Propensity Score Matching in Evaluation Of Labor Market Policies. Working Paper No.4, department for Work and Pensions. Cameron A. and Trivedi P. (2005). Microeconometrics. Cambridge University Press. Cawley, J., Heckman, J. and Vytlacil, E. (1998), Cognitive Ability and the Rising Return to Education, NBER Working Paper No. 6388. Chase. R (1998) Markets for communist human capital: Returns to education and experience in the Czech Republic and Slovakia, Industrial and Labor Relations Review, Vol 51 No.3. Chen, Martha, (1992). Impact of Grameen Bank's Credit Operations on its Members: Past and Future Research. Report to the Grameen Bank. Coleman B. (1999). The impact of group lending in Northeast Thailand. Journal of Development Economics, Vol 60 (1), 105-141. Conroy J. (2002). The role of central banks in microfinance in Asia and the Pacific. Copestake J., Sonia B., and Susan J. (2001). Assessing the impact of microcredit: A Zambian case study. Journal of Development Studies 37 No.4 81-100. Dearden, L. (1998), Ability, Families, Education and Earnings in Britain, Institute for Fiscal Studies Working Paper No. W98/14. Deaton A. (1990). Savings in developing countries: Theory and review. The World Bank Economic Review, 61-96. Deaton, A (1989). "Looking for Boy-Girl Discrimination in Household Expenditure Data", World Bank Economic Review, Vol. 3, pp. 183-210. Deaton, A (1997). The analysis of household surveys: A Microeconometric approach to development policy. Baltimore, U.S.A.: Johns Hopkins University Press for the World Bank. Dehejia R. and Wahba S. (2002). propensity Score Matching Methods for Nonexperimental Casual Studies. The review of Economics and Statistics, 84(1), 151-161.

135

De Jong, M. and N. Kleiterp (1991), Credit for Small Businesses and Microenterprises in Developing Countries, Small Enterprise Development, Vol. 2, No. 4, December, p. 21- 30 Denny, K. and Harmon, C. (2000b), Educational Policy Reform and the Return to Schooling from Instrumental Variables, IFS Working Paper No. W00/06. Dolton, P. and Vignoles, A. (2000), The Incidence and Effects of Overeducation in the UK Graduate Labour Market, Economics of Education Review, 19, pp.179-198. Duraisamy, P. (2000). Changes in returns to education in India: By gender, age-cohort and location, Center discussion paper No.815. Economic Growth Center. Yale University. Ferreira, Francisco, Lanjouw, Peter and Neri, Marcelo (2001) A New Poverty Profile for Brazil Using PPV, PNAD and Census Data. Proceedings of the 2001 ANPEC Meeting, Salvador, Brasil. Fidler, P. (1998), Bolivia: Assessing the Performance of Banco Solidario, Case Studies in Microfinance, Sustainable Banking with the Poor, World Bank, August, Washington, DC Fong, M. (1996). Gender Barriers in the Transition to a Market Economy. Poverty and Social Policy Department Discussion Paper No. 87. Washington, DC: World Bank. Foster, J., J. Greer, and E. Thorbecke (1984). "A class of decomposable poverty measures." Econometrica 52 (3): 761--765. Glewwe, P (1991). "Investigating the determinants of household welfare in the Cte d'Ivoire."Journal of Development Economics35(2):307-337. Glewwe P. and G. Hall. (1995). Who is Most Vulnerable to Macroeconomic Shocks? Hypotheses Tests using Panel Data from Peru. Living Standards Measurement Study Working Paper No. 117. Washington DC: World Bank. Goetz, Anne Marie, and Rina Sen Gupta, (1996). Who Takes the Credit? Gender, Power, and Control over Loan Use in Rural Credit Programs in Bangladesh. World Development, Vol. 24, No. 1, pp. 45-63. Goldberge N. (2005). Measuring the impact of microfinance: Taking the stock of what we know Grameen foundation, USA Publication Series. Goldberger A. (1983). Abnormal Selection Bias. Studies in Econometrics. Time Series and Multivariate Statistics. New York: Academic Press.

136

Greene, W. H (1997). Econometric analysis. 3rd ed. Upper Saddle River, N.J., U.S.A.: Prentice-Hall. Grootaert, Christiaan, and Jeanine Braithwaite (1998): Poverty Correlates and Indicatorbased Targeting in Eastern Europe and the Former Soviet Union, ms. Poverty Reduction and Economic Management Network, World Bank, May. Grosh M. (1994). Administering targeted social programs in Latin America. From platitude to practice. World Bank. Washington D.C. Gunatilaka, Ramani. 1997. Credit-Based, Participatory Poverty Alleviation Strategies in Sri Lanka: What Have We Learned? Poverty and Income Distribution Series No. 2. Colombo: Institute of Policy Studies. Hartgog C., Pereira P., and Viera (2001) Changing returns to education in Portugal during the 1980s and early 1990s: OLS and quantile regression estimators. Applied Economics, 2001, 33, 1021-1037. Hamermesh, D. and Biddle, J. (1998), Beauty, Productivity and Discrimination: Lawyers Looks and Lucre, Journal of Labor Economics, 16(1), pp.172-201. Harmon, C., Oosterbeek H. and Walker N. (2000), The returns to education: A review of evidence, issues and deficiencies in literature, Center for the Economics of Education, LSE. Harmon, C. and Walker, I. (1995), Estimates of the Economic Return to Schooling for the United Kingdom, American Economic Review, 85, pp.1278-1286. Heckman, J.J., Ichimura, H. and Todd, P.E. (1998). Matching as an Econometric Evaluation Estimator. Review of Economic Studies, 65, 261-294. Heckman, J.J., Ichimura, H. and Todd, P.E. (1997). Matching As An Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme. Review of Economic Studies, 64, 605-654. Hossain M. (1988). Credit for the alleviation of rural poverty: The Grameen Bank in Bangladesh. Washington, D.C. IFPRI, Research report No. 65. Hulme, D. (2000). Impact Assessment Methodologies for Microfinance: Theory, Experience and Better Practice. World Development Vol. 28, No. 1, pp. 79-98. Inter-American Development Bank (1998b), Evaluation of Global Microenterprise Credit (GMC): Who gets access?, How much does it cost?; Summary of Three Case Studies: Ecuador, Paraguay, El Salvador, Revised CF-4/98, Evaluation Office, Washington, DC

137

Jenkins, Stephen P, and Peter J Lambert (1997). "Three 'I's of Poverty Curves, with an Analysis of UK Poverty Trends." Oxford Economic Papers 49.3: 317-27. Kakwani, N. C (1980). "On a Class of Poverty Measures." Econometrica 48: 437-46. Khandker S. (2005). microfinance and poverty: Evidence using panel data from Bangladesh. World Bank Economic Review 2005. Kiriwandeniya, P. A. 1998. Sanasa as an Alternative Development Approach. The Federation of Thrift and Credit Cooperative Societies in Sri Lanka, Colombo. Klasen, Stephan (2000) Measuring Poverty and Deprivation in South Africa, Review of Income and Wealth, 46 pp.33-58 Klugman and Braithwaite (1998). Poverty in Russia During the Transition: An Overview. World Bank Research Observer. Forthcoming, February. Koenker, R. and G. Bassett (1978): "Regression Quantiles." Econometrica, 46, 33-50. Kolev, A. (1996). Poverty Analysis in Russia: What Can We Learn from the RLMS Round 6? Florence, Italy: European University Institute. Lang, K. (1993), Ability Bias, Discount Rate Bias and the Return to Education, Mimeo, Boston University. Lanjouw, P. and Ravallion, M (1995). "Poverty and household size." Economic Journal, 105:433. Lanjouw, P. and N. Stern. (1991). Poverty in Palanpur. World Bank Economic Review. 5(1). Layard, R. and Psacharopoulos, G. (1979), Human Capital and Earnings: British Evidence and a Critique, The Review of Economic Studies, Vol. 46, No. 3. Martins S. and Pereira P. (2004). Does education reduce wage inequality? Quantile regression evidence from 16 countries. Labor Economics (11). 355-371. Meghir, C. and Palme, M. (1999), Assessing the Effect of Schooling on Earnings Using a Social Experiment, IFS Working Paper No. W99/10. Milanovic, B. (1990). Poverty in Poland, Hungary, and Yugoslavia in the Years of Crisis 1978-87. Policy Research Working Paper No. 507. Washington, DC: World Bank.

138

Mincer, J. (1974), Schooling, Experience and Earnings, Columbia University Press: New York. Miracle, M.P., D.S. Miracle, and L. Cohen (1980), Informal Savings Mobilization in Africa, Economic Development and Cultural Change, p. 701-724 Montenegro C. (2001). Wage distribution in Chile: Does gender matter? A quantile regression approach. Working paper series No.20. World Bank. Mosley, P. (1997). The use of control groups in impact assessment for microfinance mimeo. Washington, DC: Consultative Group to Assist the Poorest (CGAP) and also (2001) Working Paper 19, ILO, Geneva. Neal, Derek A., and William A. Johnson. (1996) The Role of Pre-Market Factors in Black- White Wage Differences. Journal of Political Economy 104, pp. 869-895. OBoyle, E.J. (1999) Toward an Improved Definition of Poverty. Review of Social Economy, Vol. LVII, n. 3, September, pp. 281-301 Park, J. H. (1996), Measuring Education Over Time, Economics Letters, 60, pp.425428. Paxton, J. and C. Fruman (1997), Outreach and Sustainability: A Comparative Analysis of Savings-First vs. Credit-First Financial Institutions, Sustainable Banking with the Poor, World Bank, January, Washington, DC. Peace, Gill and David Hulme, (1994). Microenterprise and Children - What are the Intra-Household Impacts of Income Generating Programs? Small Enterprise Development, Vol. 5. Pitt M. and Khandker S. (1998). The impact of group-based credit programs on poor households in Bangladesh: Does the gender of participants matter? Journal of Political Economy 106 No.5 958-996. Psacharopoulos, G. (1994), Returns to Investment in Education: A Global Update, World Development, 22, pp.1325-1343. Psacharopoulous, George (1985) "Returns to Education: A Further International Update and Implication." Journal of Human Resources, 20:4, pages 583-604. Pudney, S (1999). "On some statistical methods for modeling the incidence of poverty." Oxford Bulletin of Economics and Statistics 61 (3): 385--408. Puhani P. (2000). The Heckman Correction for Sample selection and Its Critique- A Short Survey. Journal of Economic Surveys 14, 53-68.

139

Quisumbing, A., Haddad, L., & Pena, C. (1995). Gender and poverty: New evidence from ten developing countries. IFPRI - FCND Discussion Paper 9, Washington, DC. Ravallion, Martin (1996) How Important to Indias Poor is the Sectoral Composition of Economic Growth?, World Bank Economic Review, 10(1): 1-25. Ravallion, M., and B. Bidani (1994). "How robust is a poverty profile?" World Bank Economic Review 8 (1): 75--102. Robinson M. (1992). Rural financial intermediation: Lessons from Indonesia. The Bank of Rakyat Indonesia: rural Banking 1970-91. Development discussion paper No.434. Harvard Institute of International Development. Rosenbaum, P.R. and Rubin, D.B. (1985). Constructing a Control Group Using Multivariate Matched Sampling Methods that Incorporate the Propensity Score. Biometrica, 70, 41- 50. Rosenbaum, P.R. and Rubin, D.B. (1983). The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika, 70, 1,41-55. Rubin D. and Thomas N. (1996). Matching Using Estimated Propensity Score: Relating Theory to Practice. Biometrics, 52, 249-264. Ruggeri Laderchi, C. (2000), The Monetary Approach to Poverty: A Survey of Concepts and Methods, Queen Elizabeth House, Working Paper 58 Sen, A. K (1976). "Poverty: An Ordinal Approach to Measurement." Econometrica 44.2: 219- 31. Senanayake, S. M. P. (1999). Microfinance Sector in Sri Lanka. Mimeo. Department of Economics, University of Colombo, Colombo. Schultz, T. P (1988). "Education investments and returns." In Handbook of development economics, Vol. 1, ed. H. Chenery and T. N. Srinivasan. Amsterdam: NorthHolland Press. Shorrocks, Anthony F (1995). "Revisiting the Sen Poverty Index." Econometrica 63.5: 1225-30. Sianesi, B. and Van Reenen, J. (2000), The Returns to Education A Review of the Macroeconomic Literature, Centre for the Economics of Education, London School of Economics, Discussion Paper No. 6.

140

Siddhisena, K.A.P, and Jayathilaka, M.D.R.K (2004). "Identification of the Poor in Sri Lanka: Development of Composite Indicator and Regional Poverty Lines." MIMAP Sri Lanka Series. Spence, M. (1973), Job Market Signaling Quarterly Journal Of Economics, 87(3), pp.355-373. Swain R. (2004). Is microfinance a good poverty alleviation Strategy? Evidence from impact assessment. The Department for Infrastructure and Economic Cooperation. SIDA. Szkely, Miguel (1998). The Economics of Poverty, Inequality and Wealth Accumulation in Mexico. St.Antony's Series. NewYork: St.Martin's Press; London: Macmillan Press; in association with St.Antony's College,1998. Taber. C (2001) The rising college premium in the eighties: Return to college or returns to unobserved ability? Review of Economic Studies, 68, 665-691. Uusilato, R. (1999), Returns to Education in Finland, Labor Economics, 6(4), pp.569580. Weiss J. and Montgomery H. (2004) Great Expectations: Microfinance and poverty reduction in Asia and Lation America. Discussion paper No.15. ADB Institute. World Bank (1999), Poverty and Social Developments in Peru, 1994-1997, World Bank Country Study 0253-2123 World Bank (1990). The World Bank Annual Report 1990, Washington DC: The World Bank. World Bank (1998). Public expenditures for poverty alleviation in northeast Brazil: Promoting growth and improving services. World Bank LAC Report. World Bank (1997). Introducing Savings in Microcredit Institutions: When and How? Consultative Group to Assist the Poorest, CGAP Note No. 8, April, Washington, DC. World Bank (1996), A Worldwide Inventory of Microfinance Institutions, Sustainable Banking with the Poor, July, Washington, DC. Zeller M. and Meyer R. (2002). The triangle of microfinance: Financial sustainability, outreach and impact. John Hopkins for the International Food Policy Research Institute. Baltimore.

141