Upload
damith-chandimal
View
41
Download
1
Embed Size (px)
DESCRIPTION
This dissertation presents three essays examining some of the issues concerningpoverty, microfinance and returns to education. The first essay examines the micro-leveldeterminants and correlates of poverty, and presents a poverty profile for Sri Lanka. Thisis the first study that examines the probable determinants and correlates of Sri Lankanpoverty in a multivariate framework employing both logit and quantile regressions. Themost appealing feature of quantile regression is that it does not impose constant parameters over the entire distribution. The empirical findings are broadly encouraging.
Citation preview
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
ii
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.
iii
TABLE OF CONTENTS
ACKNOWLEDGEMENTS………………………………………………… ii
LIST OF TABLES………………………………………………………... v
LIST OF FIGURES……………………………………………………….. vii
CHAPTER
1. INTRODUCTION………………………………………... …… 1
2. A ROBUST POVERTY PROFILE FOR SRI LANKA IN A MULTIVARIATE FRAMEWORK……………………… 4
2.1 Introduction…………………………………………….. 4
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
iv
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…………………………………………….. 105
4.3 Econometric Methodology………………………………. 111
4.4 Empirical Results………………………………………… 114
4.5 Conclusion……………………………………………….. 118
REFERENCES………………………………………………………. 134
v
LIST OF TABLES
2.1 Unconditional Poverty Profile (cross tabulations)…………………… 26
2.2 Inequality Indices……………………………………………………. 27 2.3 Variable definitions and means……………………………………… 28 2.4 Logit regression estimates…………………………………………… 29 2.5 Quantile and OLS regression estimates…………………………….... 30 3.1 Social Transfers and social expenditure,1999……………………….. 73 3.2 Income Transfer Component………………………………………… 74 3.3 Bottom 20th percentile probit model for the propensity score………. 75 3.4 20th-40th percentile probit model for the propensity score…………… 76 3.5 40th-60th percentile probit model for the propensity score…………… 77 3.6 60th-80th percentile probit model for the propensity score…………… 78 3.7 80th-100th percentile probit model for the propensity score………….. 79 3.8 Matching quality indicators (covariate balancing) for the 20th percentile……………………………………………………. 80 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
vi
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……………………………………………… 92 3.13 Individuals lost due to common support requirement……………….. 95 3.14 Impact of microfinance on household savings……………………….. 96 3.15 Impact of microfinance on household income………………………… 97 4.1 Cross Country evidence on the Returns to Schooling (year-1995)…… 120 4.2 Quantile and OLS regression estimates (model 1-full sample)……….. 121 4.3 Quantile and OLS regression estimates (model 1-Sinhala)…………… 122 4.4 Quantile and OLS regression estimates (model 1-Tamil)…………….. 123 4.5 Quantile and OLS regression estimates (model 2-full sample)……….. 124 4.6 Quantile and OLS regression estimates (model 2-Sinhala)…………….125 4.7 Quantile and OLS regression estimates (model 2-Tamil)………………126
vii
LIST OF FIGURES
2.1 Cumulative poverty gap curves……………………………………… 31
2.2 Lorenz curves………………………………………………………… 32
2.3 Ordinary Least Squares and Quantile Regression Estimates………… 33
3.1 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 Returns to Schooling in Europe, Men and Women…………………… 127
4.2 OLS and quantile regression estimates (model 1-full sample)……....... 128 4.3 OLS and quantile regression estimates (model 1-full Sinhala)……….. 129 4.4 OLS and quantile regression estimates (model 1-full Tamil)………… 130 4.5 OLS and quantile regression estimates (model 2-full sample)………... 131 4.6 OLS and quantile regression estimates (model 2–Sinhala)…………… 132
viii
4.7 OLS and quantile regression estimates (model 2–Tamil)……………... 133
Copyright 2006, Sandaradura Indunil Udayanga De Silva.
1
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,
2
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 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.
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
3
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 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 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.
4
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
5
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
6
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 under-
reporting 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 Total household consumption divided by the number of household members. 2 The paper provides a detailed explanation for derivation of poverty lines.
7
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;
1 *
0( ; ) [ ( : )]P z z Q p z dpαα = −∫
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 0α = , 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, ( ; 1)P z α = , 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.
8
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, ( ; 2)P z α = , 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.
9
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
10
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:
0
( : ) ( : ) .p
G p z g q z dq= ∫
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 1p = . Figure 2.1 indicates that
when 1p = , 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).
11
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 0.25p =
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:
01 0
0
( ) 1( ) ( )( )
ppQ q dq
L p Q q dqQ q dq µ
= =∫∫
∫
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 incomeµ . 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
12
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 0.5p = the sectorial
estimates for urban, rural and estate are ( ) 0.256L p = , ( ) 0.308L p = and ( ) 0.336L p =
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:
1
0Gini Index of Inequality = 2 ( ( ))p L p dp−∫
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:
1
0( ) ( ( )) ( :)I p L p p dpρ κ= −∫
Where, ( : )pκ ρ are percentile dependent weights expressed as:
( 2)( : ) ( 1)(1 )p P ρκ ρ ρ ρ −= − −
Larger the value of ρ , larger will be the weight placed on the inequality of the bottom
percentile (or the poorest people). Therefore larger the value of ρ , greater will be the 6 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 δ from jy to iy , and i jy yδ δ+ > − , then the transfer principle is satisfied
( ) ( )iff I y I y′ ≥ .
13
ethical concern felt for the poor by the social decision maker. Note, when 2ρ = , 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;
11 (1 ) 10
1
0
( ( ) )1 , 1
( )exp( ln( ( ) )
1 , 1
Q p dpwhen
IQ p dp
when
ε ε
εµε
εµ
− − − ≠ =
− =
∫
∫
( )Q p is the standard of living of the individual whose rank or percentile in the
distribution is p . The parameter ε reflects the strength of society's preference for equality
and is bounded by zero and infinity. When 0ε = 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 0ε > 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 ρ , is equal to 1.5, 2.0, and 2.5) and
Atkinson measures (whenε , 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 ρ =2), the
highest degree of inequality is in the urban sector followed by the rural and estate sector.
7 Social welfare function (W)
1
11 , 1where is the standard of living.1
N
ii
xx
N
ε
εε
−
== ≠−
∑
1
1ln ln , 1N
ii
W xN
ε=
= =∑
14
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:
1,0,
Pr 1 ( , )
Pr 0 1 ( , )
if the household is belowthe poverty linePOV
if otherwiseand
POV X F X
POV X F X
β
β
=
= =
= = −
15
Where X is the vector of economic and demographic variables. β 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 1-
passed 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
16
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;
0[1 ( )] 1[ ( )] ( ]E POV X F X F X F Xβ β β′ ′ ′= − + =
Based on the logistic distribution,
Pr 11
X
X
ePOV Xe
β
β
′
′= =+
Table 2.4 presents the parameter estimates for the logistic regression. The column, dydx
,
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
17
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
18
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
19
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 household’s 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 thi
household to differ across the welfare spectrum.
8 Koenker and Basset (1978).
20
The quantile regression model can be expressed as:
,i i iy x τ τβ µ′= +
Where iy is the log of per capita consumption of the thi household, and ix represents the
economic and demographic characteristics of the thi household. By Imposing the
assumption that the thτ − quantile of the error term conditional on the regressors is zero,
,( ( | ) 0)i iQ u xτ τ = , the thτ − conditional quantile of iy with respect to ix can be
expressed as:
( | )i i iQ y x xτ τβ′=
For any τ ∈(0,1), the parameter τβ can be estimated by
{ } { }| |
ˆ arg min | | (1 ) | |k
i i i i
i i i ii i y x i i y x
y x y xτ τ τ
τβ β β
β τ β τ β∈ ′ ′∈ ≥ ∈ ≥
′ ′= − + − − ∑ ∑
R
Note, that when 0.5τ = , 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.
21
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
22
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
23
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.
24
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. 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. 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
25
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.
26
Table 2.1: Unconditional Poverty Profile (cross tabulations) 0α = 1α = 2α = 0α = 1α = 2α =
Sri Lanka 0.252 0.058 0.021 Number of men (17-60 years)
Urban 0.207 0.054 0.020 1 0.235 0.051 0.017
Rural 0.257 0.059 0.021 2 0.252 0.058 0.020
Estate 0.281 0.067 0.025 3 0.308 0.077 0.030
Characteristics of the Head 4 0.347 0.086 0.030
Male 0.242 0.054 0.019 Number of Women (17-60 years)
Female 0.302 0.078 0.031 1 0.237 0.052 0.018
Age ≤29 years 0.224 0.043 0.013 2 0.259 0.059 0.020
Age: 30-59 years 0.248 0.056 0.019 3 0.319 0.077 0.030
Age ≥60 0.270 0.066 0.026 4 0.351 0.091 0.037
Education Number of Elders
No schooling 0.426 0.116 0.048 1 0.263 0.062 0.024
(primary.edu) 0.336 0.079 0.029 2 0.278 0.069 0.025
(secondary.edu) 0.196 0.041 0.013 3 0.269 0.117 0.054
(tertiary.edu) 0.052 0.008 0.003 Employment
Household Size Salary 0.101 0.018 0.007
0-2 0.143 0.038 0.018 Casual Wage 0.376 0.093 0.035
3-4 0.178 0.036 0.012 Business 0.139 0.027 0.009
4+ 0.342 0.082 0.030 Receiving (or entitled) for pension income 0.216 0.052 0.020
Number of Children (0-6 years) Not receiving (or not entitled) for pension income 0.261 0.060 0.022
1 0.296 0.066 0.023 Housing Tenure and Type
2 0.389 0.097 0.036 Owned by household head 0.234 0.052 0.018
3 0.508 0.136 0.048 Not owned by household head 0.312 0.077 0.029
4 0.800 0.234 0.074 Single house 0.241 0.053 0.019
Number of Children (7-16 years) Annexe 0.250 0.117 0.058
1 0.226 0.050 0.017 Shanty or line room 0.550 0.161 0.067
2 0.304 0.068 0.024 Main Source of Lighting
3 0.407 0.097 0.035 Electricity 0.173 0.036 0.012
4 0.444 0.122 0.048 Kerosene 0.366 0.088 0.032
Main Fuel Used for Cooking
Gas 0.022 0.004 0.002 Firewood or sawdust 0.284 0.065 0.023
27
Table 2.2: Inequality Indices
S-Gini Index Atkinson Index
ρ=1.5 ρ=2 ρ=2.5 ε=0.5 ε=1.0 ε=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
28
Table 2.3: Variable definitions and means Variable Definition Poor Non-poor Total
LogPCE Logarithm of real per capita consumption 2.9466 3.3233 3.2282
Children (1) Number of children (0-6 years) 0.5896 0.3631 0.4201
Children (2) Number of children (7-16 years) 1.2062 0.8256 0.9216
Men Number of men (17-60 years) 1.4569 1.353 1.3793
Women Number of women (17-60) 1.5711 1.4336 1.4683
Elders Number of persons (60+ years) 0.3882 0.3545 0.3629
HsizeSQR Household size squared 30.9974 21.444 23.853
Age age of household head 49.6345 49.2157 49.3213
AgeSQR Age squared 2660.517 2602.498 2617.128
Dummy Variables
POV Household is below the poverty line 1 0 0.2521
Female Head Household head is female 0.2009 0.1564 0.1676
Rural Household resides in the rural sector 0.8138 0.7471 0.7639
Urban Household resides in the urban sector 0.1132 0.2159 0.1900
Primary.edu Year 1-Year 6 0.4706 0.3138 0.3533
Secondary.edu Year 7-GCE (O/L) 0.3527 0.4862 0.4525
Tertiary.edu Year 12 and above 0.0211 0.1289 0.1017
Salary Household head in salaried employment 0.0645 0.1965 0.1632
Casual Wage Household head works for a casual wage 0.3702 0.2077 0.2486
Business Household head in business (including 0.0851 0.1776 0.1542
trade and manufacture)
NoRetBenifit Household head is not receiving or not 0.8290 0.7896 0.7995
entitled to receive retirement benefits.
ShantyLineR Household lives in shanty or line room
HHnotowningHouse Dwelling unit is not owned by 0.2818 0.2093 0.2276
by household
Firewood Household uses firewood or sawdust as 0.9719 0.8248 0.8619
the main of fuel for cooking
Electricity Main source of lighting is electricity 0.3968 0.6385 0.5775
South Household in Sourthen province 0.1644 0.1103 0.1240
North Household in Northern province 0.0597 0.1269 0.1100
East Household in Eastern province 0.1094 0.1503 0.1400
Uva Household in Uva province 0.0766 0.0743 0.0720
Sabara Household in Sabaragamuwa province 0.1258 0.0750 0.0878
Central Household in Central Province 0.1930 0.1007 0.1240
Nwest Household in North Western province 0.0936 0.0941 0.0940
Ncentral Household in North Central Province 0.0528 0.0731 0.0680
29
Table 2.4: Logit regression estimates Independent Variable Dy/dx (marginal effects) z-value Children (1) 0.126 11.44
Children (2) 0.089 9.55
Men 0.073 7.18
Women 0.082 8.02
Elders 0.089 6.72
HsizeSQR -0.002 -3.70
Age -0.003 -1.38
AgeSQR 0.000 1.52
Female Head 0.068 5.47
Rural 0.105 4.24
Urban 0.031 1.18
Primary.edu -0.035 -2.40
Secondary.edu -0.085 -5.37
Tertiary.edu -0.217 -7.36
Salary -0.068 -3.79
Casual Wage 0.055 4.92
Business -0.123 -8.04
NoRetBenifit 0.059 4.44
ShantyLineR 0.174 -11.12
HHnotowningHouse 0.024 2.18
Electricity -0.113 -11.12
Firewood 0.198 8.73
South 0.031 1.83
North -0.262 -12.11
East -0.095 -5.21
Uva -0.067 -3.17
Sabara 0.043 2.31
Central 0.111 6.38
Nwest -0.022 -1.19
Ncentral -0.132 -5.94
N=7481 Pseudo R²=0.225 Pr>χ²=0.000 Loglikelihood =-3274.5261 Notes
1,1) Dependent variable:
0,if the household is belowthe poverty line
POVif otherwise
=
2) Poverty line = Rs.1206 3) Variable definitions and means are given in Table 2.3.
30
Table 2.5: Quantile and OLS regression estimates (standard errors in parentheses) Quantile Independent Variable 0.1 0.25 0.5 0.75 0.9 OLS
Constant 3.19 (0.065)
3.366 (0.043)
3.473 (0.039)
3.591 (0.040)
3.698 (0.056)
3.431 (0.036)
Children(1) -0.087 (0.009)
-0.095 (0.006)
-0.106 (0.006)
-0.114 (0.007)
-0.113 (0.009)
-0.097 (0.005)
Children(2) -0.065 (0.008)
-0.074 (0.006)
-0.078 (0.005)
-0.085 (0.007)
-0.089 (0.008)
-0.074 (0.004)
Men -0.06 (0.007)
-0.072 (0.006)
-0.075 (0.006)
-0.079 (0.008)
-0.079 (0.010)
-0.069 (0.005)
Women -0.062 (0.009)
-0.073 (0.006)
-0.078 (0.005)
-0.088 (0.007)
-0.086 (0.009)
-0.069 (0.005)
Elders -0.064 (0.010)
-0.070 (0.008)
-0.066 (0.007)
-0.078 (0.010)
-0.066 (0.012)
-0.063 (0.006)
HsizeSQR 0.003 (0.000)
0.004 (0.000)
0.004 (0.000)
0.004 (0.001)
0.004 (0.001)
0.003 (0.000)
Age 0.007 (0.002)
0.006 (0.002)
0.005 (0.001)
0.004 (0.001)
0.005 (0.002)
0.006 (0.001)
AgeSQR 0.000 (0.000)
0.000 (0.000)
0.000 (0.000)
0.000 (0.000)
0.000 (0.000)
0.000 (0.000)
FemaleHead -0.042 (0.009)
-0.04 (0.007)
-0.043 (0.008)
-0.033 (0.009)
-0.015 (0.0012)
-0.037 (0.006)
Rural -0.068 (0.023)
-0.008 (0.016)
-0.057 (0.015)
-0.028 (0.013)
-0.021 (0.029)
-0.056 (0.013)
Urban -0.025 (0.023)
-0.029 (0.016)
0.007 (0.015)
0.042 (0.015)
0.081 (0.030)
0.013 (0.014)
Primary.edu 0.020 (0.015)
0.021 (0.009)
0.025 (0.008)
0.027 (0.009)
0.028 (0.001)
0.031 (0.008)
Secondary.edu 0.068 (0.016)
0.056 (0.011)
0.058 (0.009)
0.062 (0.008)
0.075 (0.015)
0.072 (0.009)
Tertiary.edu 0.142 (0.019)
0.139 (0.013)
0.164 (0.014)
0.181 (0.016)
0.221 (0.021)
0.179 (0.011)
Salary 0.035 (0.110)
0.024 (0.007)
0.012 (0.006)
-0.002 (0.009)
-0.025 (0.012)
0.007 (0.007)
Casual Wage -0.017 (0.010)
-0.026 (0.006)
-0.033 (0.006)
-0.042 (0.007)
-0.069 (0.010)
-0.044 (0.006)
Business 0.080 (0.012)
0.008 (0.009)
0.078 (0.007)
0.078 (0.009)
0.089 (0.015)
0.081 (0.006)
NoRetBenifit -0.037 (0.010)
-0.039 (0.007)
-0.029 (0.007)
-0.016 (0.007)
-0.009 (0.011)
-0.031 (0.006)
ShantyLineR -0.144 (0.027)
-0.119 (0.019)
-0.087 (0.014)
-0.062 (0.014)
-0.059 (0.027)
-0.110 (0.013)
HHnotowningHouse -0.028 (0.009)
-0.022 (0.007)
-0.015 (0.006)
-0.019 (0.006)
-0.037 (0.010)
-0.029 (0.005)
Electricity 0.078 (0.008)
0.066 (0.006)
0.060 (0.005)
0.065 (0.007)
0.059 (0.009)
0.069 (0.005)
Firewood -0.125 (0.013)
-0.138 (0.008)
-0.159 (0.008)
-0.173 (0.011)
-0.180 (0.016)
-0.156 (0.007)
South -0.050 (0.014)
-0.047 (0.009)
-0.022 (0.009)
-0.016 (0.009)
-0.032 (0.014)
-0.025 (0.009)
North 0.116 (0.013)
0.112 (0.011)
0.138 (0.011)
0.144 (0.011)
0.122 (0.014)
0.137 (0.009)
East 0.069 (0.012)
0.037 (0.009)
0.024 (0.008)
0.001 (0.009)
-0.010 (0.017)
0.029 (0.009)
Uva -0.043 (0.016)
0.030 (0.012)
0.019 (0.008)
0.001 (0.012)
-0.024 (0.017)
0.024 (0.010)
Sabara -0.046 (0.015)
-0.055 (0.011)
-0.019 (0.008)
-0.038 (0.013)
-0.043 (0.021)
-0.026 (0.009)
Central -0.109 (0.012)
0.104 (0.011)
-0.064 (0.010)
-0.055 (0.011)
-0.050 (0.019)
-0.070 (0.009)
Nwest -0.004 (0.014)
0.001 (0.009)
0.018 (0.012)
0.014 (0.013)
0.002 (0.019)
0.006 (0.009)
Ncentral 0.096 (0.023)
0.058 (0.010)
0.046 (0.010)
0.036 (0.012)
0.048 (0.024)
0.070 (0.010)
31
Figure 2.1.Cumulative poverty gap curves
32
Figure 2.2 Lorenz curves
33
−0
.10−
0.0
8−
0.0
6−
0.0
4−
0.0
2m
en
0 .2 .4 .6 .8 1Quantile
−0
.10−
0.0
8−
0.0
6−
0.0
4w
om
en
0 .2 .4 .6 .8 1Quantile
−0
.14−
0.1
2−
0.1
0−
0.0
8−
0.0
6ch
ild
ren
1
0 .2 .4 .6 .8 1Quantile
−0
.10−
0.0
8−
0.0
6−
0.0
4−
0.0
2ch
ild
ren
2
0 .2 .4 .6 .8 1Quantile
−0
.10−
0.0
8−
0.0
6−
0.0
4−
0.0
2e
lde
rs
0 .2 .4 .6 .8 1Quantile
0.0
00
.000
.000
.010
.010
.01
ag
e
0 .2 .4 .6 .8 1Quantile
−0
.10−0
.08−0
.06−0
.04−0
.020
.00
Ca
su
al_
Wa
ge
0 .2 .4 .6 .8 1Quantile
−0
.15−
0.1
0−
0.0
50
.00
0.0
5R
ura
l
0 .2 .4 .6 .8 1Quantile
−0
.08−
0.0
6−0
.04−
0.0
20.0
00
.02
Fe
ma
leH
ea
d
0 .2 .4 .6 .8 1Quantile
−0
.02
0.0
00
.02
0.0
40
.06
Prim
ary
Ed
u
0 .2 .4 .6 .8 1Quantile
0.0
40
.06
0.0
80
.10
0.1
2S
eco
nd
ary
Ed
u
0 .2 .4 .6 .8 1Quantile
0.1
00
.15
0.2
00
.25
0.3
0T
ert
iary
Ed
u
0 .2 .4 .6 .8 1Quantile
−0
.10−
0.0
50
.00
0.0
50
.10
Sa
lary
0 .2 .4 .6 .8 1Quantile
−0
.00
−0
.00
0.0
0a
ge
SQ
R
0 .2 .4 .6 .8 1Quantile
−0
.08−
0.0
6−
0.0
4−
0.0
20
.00
HH
no
tow
nin
gH
ou
se
0 .2 .4 .6 .8 1Quantile
−0
.20−
0.1
5−
0.1
0−
0.0
50
.00
Sh
an
tyL
ine
R
0 .2 .4 .6 .8 1Quantile
0.0
40
.06
0.0
80
.10
ele
ctr
icity
0 .2 .4 .6 .8 1Quantile
−0
.25−
0.2
0−
0.1
5−
0.1
0F
ire
wo
od
0 .2 .4 .6 .8 1Quantile
−0
.08−
0.0
6−0
.04−
0.0
20.0
00
.02
No
_R
et_
Be
nifit
0 .2 .4 .6 .8 1Quantile
0.0
40
.06
0.0
80
.10
0.1
2b
usin
ess
0 .2 .4 .6 .8 1Quantile
0.0
00
.00
0.0
00
.01
Hsiz
eS
QR
0 .2 .4 .6 .8 1Quantile
−0
.10
−0
.05
0.0
00
.05
so
uth
0 .2 .4 .6 .8 1Quantile
0.0
80
.10
0.1
20
.14
0.1
60
.18
no
rth
0 .2 .4 .6 .8 1Quantile
−0
.05
0.0
00
.05
0.1
00
.15
ea
st
0 .2 .4 .6 .8 1Quantile
−0
.05
0.0
00
.05
0.1
0u
va
0 .2 .4 .6 .8 1Quantile
−0
.10−
0.0
50
.00
0.0
50
.10
sa
ba
ra
0 .2 .4 .6 .8 1Quantile
−0
.15
−0
.10
−0
.05
0.0
0ce
ntr
al
0 .2 .4 .6 .8 1Quantile
−0
.04−
0.0
20.0
00
.02
0.0
40
.06
nw
est
0 .2 .4 .6 .8 1Quantile
0.0
00
.05
0.1
00
.15
nce
ntr
al
0 .2 .4 .6 .8 1Quantile
−0
.05
0.0
00
.05
0.1
00
.15
Urb
an
0 .2 .4 .6 .8 1Quantile
Figure 2.3: Ordinary Least Squares and Quantile Regression Estimates
34
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):
35
"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:
36
"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
37
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
38
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
39
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 $700-
800 (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
40
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
41
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.”
42
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.
43
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
44
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 component-
food 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
45
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
46
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
47
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.
48
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 co-
operatives. 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
49
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,
50
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.
51
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 girl’s 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
52
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
53
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 ( , , ), ,..., ,i i iY D i NΧ and the impact on Y from a hypothetical change in
D, while holding Χ 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 ( , , ; ,..., )i i iY D i NΧ . Where Y is the scalar-value
54
outcome variable, Χ 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
1 0Y Y∆ = −
Where, 1Y is the outcome of the treated individual, and 0Y 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,
1
1 [ ]N
ii
ATEN =
= ∆∑
1
1 [ | 1]TN
i iiT
ATT DN =
= ∆ =∑
where 1
NT ii
N D=
=∑ . 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
55
the outcome of untreated individuals for whom the observable characteristics, Χ , 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
1 0ATT=E( |D=1)=E(Y |D=1)-E(Y |D=1)∆
From the above equation, the problem of selection bias is straightforward, since the
second term on the right side- 0E(Y |D=1) ≡ the counterfactual mean of the treated, is not
observable. If the condition 0 0E(Y |D=1)=E(Y |D=0) holds, one can use the non-
participants 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.
56
The ATT can also be expressed as
1 0 0 0[ | 1] [ | 0) ( | 1] [ | 0]E Y D E Y D ATT E Y D E Y D= − = = + = − =
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
0 0[ | 1] [ | 0] 0E Y D E Y D= − = =
The crux of the evaluation problem is that 0[ | 1]E Y D = 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
57
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 before-
after estimates of non-participants. Alternatively, matching involves identifying non-
program participants comparable in essential characteristics to participants, and then
58
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 non-
experimental and quasi-experimental approaches share one thing in common: when the
counterfactual mean for the treated- 0[ | 1]E Y D = , 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
0 1, |Y Y D⊥ Χ
It states that outcomes are independent of program participation, after controlling
for the variation in outcomes induced by differences in Χ 9. Under the conditional
independence assumption, the matching process is analogous to constructing an
9 Using Dawid's (1979) notation, ⊥ represents independence.
59
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[ 1| ] 1D< = Χ <
This assumption implies that for each value of Χ there are both treated and untreated
individuals. In other words, for each participant there is another non-participant with a
similar Χ . A practical constraint that exists in matching is that when the number of
covariates iΧ increases, the chances of finding a match reduces. However, Rosenbaum
and Rubin (1983) showed that matching on the propensity score ( )p Χ - 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);
(1 )
1, 2 1, 2 1,...
( ) s gn ,( ) Pr( 1| ) ( | ). 0 ( ) 1, ,
( ) (1 ( ))Pr( ,... | ,... )
i i
i
i i i i i i iD D
i iN N i
Let p be the probability of unit i having been as i ed to treatment definedas p D E D Assume that P for all and
p p for the N units inD D D
−
=
ΧΧ ≡ = Χ = Χ < Χ < Χ
Χ − ΧΧ Χ Χ =
1 0 1 0
.{( , ) } | {( , ) } | ( )
N
i i i i i i i i
the sample ThenY Y D Y Y D p⊥ Χ ⇒ ⊥ Χ
∏
Corollary;
1 0
1
{( , ) } | 1 ,
| { ( | 1, ( )) ( | 0, ( )) | 1},
i i i i
D i i i i i i i
If Y Y D and the asumptions of proposition hold then
E E Y D P E Y D P D=
⊥ Χ
∆ = = Χ − = Χ =
60
assuming the expectations are defined. The outer expectation is over the distribution of
( ) | 1i iP DΧ = . The proposition implies that observations with the same propensity score
have the same distribution of the full vector of covariates, iΧ .
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
61
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 ( ) 0P Χ = or ( ) 1P Χ = , 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.
62
Having enforced the common support region, the next step is to choose the
matching algorithm. The general formula for the matching estimator is
1 0{ 1}
1 [ ( , ) ],M i ji D jT
B Y w i j YN ∈ =
= −∑ ∑
where MB denotes the matching estimator for the bias, 0 ( . ) 1w i j< ≤ , 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 [ ( , ) 1]w i j = , and takes each participant in turn and identifies the non-
participant 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 δ , so that
{ :| ( ) ( ) | }i jj P X P X δ− < 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 non-
participants 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
63
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, Pseudo-
R² 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
1 0
1 0
1000.5 ( ( ) ( ))beforeSB
V VΧ −Χ
= ⋅⋅ Χ + Χ
1 0
1 0
100 ,0.5 ( ( ) ( ))
M Mafter
M M
SBV VΧ −Χ
= ⋅⋅ Χ + Χ
where 1 1( )VΧ is the mean (variance) in the treatment group before matching and 0 0( )VΧ
the analogue for the comparison group. 1 1( )M mVΧ and 0 0( )M MVΧ are the corresponding
values for the matched samples.
The basic idea of the Pseudo-R² is to re-estimate the propensity score on the
matched sample, that is only on participants and matched non-participants and compare
the Pseudo-R²'s before and after matching. The Pseudo-R² 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
64
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
65
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 sector (regular worker-formal, casual worker-informal, self-
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)
66
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
67
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 self-
employed 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.
68
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 self-
employment, 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.
69
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 non-
participants 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- 2R 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 0.001δ = ). 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 per-
capita income and savings who were actually treated- i.e., the average treatment effect on
70
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
71
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
72
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 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 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.
73
Table 3.1: Social Transfers and social expenditure,1999 Social transfers Sri Lanka Percent of Percent of
Rupees (million) the sub-total the total
Samurdhi 8145 58.20 12.05
Fertilizer subsidy 1451 10.37 2.15
Refugees 2661 19.01 3.94
Other 1739 12.42 2.57
Sub-total 13,996 100.00 20.70
Other social expenditures
Pensions 20,723 30.65
Education 22,231 32.89
Health 10,651 15.76
Total 67,601 100
74
Table 3.2: Income Transfer Component Category of family Amount received Distribution of coupons Families earning less Rs. 1000 1) Rs. 400-only to purchase than Rs. 500 and food items having more than 5 2) Rs. 375-to buy goods or members to encash 3) Rs. 25- insurance premium 4) Rs.200- compulsory saving Families with only Rs. 100 Only for the purchase of food 1 member Families with 2 members Rs. 200 Only for the purchase of food Others Rs. 500 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 Janasaviya recipients Rs. 250 Can either buy goods or encash
75
Table 3.3: Bottom 20th percentile probit model for the propensity score Variable Coefficient Standard Error Age -0.0072* 0.0022 Female 0.0130 0.0866 Married 0.0390 0.0873 Household size 3-4 0.3200* 0.0760 Household size 5-6 0.5533* 0.0923 Household size ≥ 7 0.5074* 0.1646 Sinhala 0.4563* 0.1250 Tamil 0.2720 0.1477 Primary education 0.0233 0.0780 Secondary education -0.0058 0.0836 Tertiary education 0.1010 0.1973 Regular employee -0.0044 0.1622 Casual employee 0.2050** 0.0885 Self-employed 0.2764* 0.0860 Urban sector 0.4456** 0.2119 Rural sector 0.5409** 0.1710 Household expenditure -4.4100 2.8000 Household not owning house -0.1116 0.1144 Lives in a single house 0.3917 0.2202 Lives in a slum/shanty/lineroom 0.3551 0.2563 Floor type- clay/mud/earth 0.0204 0.2787 Floor type-cement -0.0500 0.2760 Main source of lighting- electricity -0.4440 0.2327 Main source of lighting- kerosene -0.3452 0.2314 Energy for cooking- LP gas -0.2634 0.3950 Energy for cooking- firewood 0.0949 0.2720 Source of water- common well 0.0114 0.0595 Central province 0.2025 0.1222 Southern province 0.5200* 0.1210 Northern province -0.2272 0.2342 Eastern province -0.0768 0.1435 North western province 0.2237 0.1282 North central province 0.0246 0.1396 Uva province 0.2390 0.1290 Sabaragamuwa province 0.3268* 0.1205 Constant -1.5600* 0.5822
Number of observations = 2346 2 (35) 288.46LR chi = 2 0.0921Pseudo R = 2 0.0000prob chi> =
1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.
76
Table 3.4: 20th-40th percentile probit model for the propensity score
Variable Coefficient Standard Error
Age -0.0017 0.0022 Female 0.1148 0.0952 Married 0.0420 0.0998 Household size 3-4 0.3066* 0.0966 Household size 5-6 0.4746* 0.1040 Household size ≥ 7 0.5233* 0.1433 Sinhala 0.4101* 0.1176 Tamil 0.1237 0.1493 Primary education -0.0538 0.0969 Secondary education -0.1347 0.0980 Tertiary education -0.1972 0.1653 Regular employee 0.1395 0.1314 Casual employee 0.1945 0.1014 Self-employed 0.1438 0.1010 Urban sector 0.1396 0.2007 Rural sector 0.2428 0.1697 Household expenditure -2.6400 2.4900 Household not owning house 0.0749 0.1113 Lives in a single house 0.1823 0.1842 Lives in a slum/shanty/lineroom -0.2090 0.2182 Floor type- clay/mud/earth 0.6070** 0.2681 Floor type-cement 0.4668 0.2629 Main source of lighting- electricity 0.0448 0.1744 Main source of lighting- kerosene 0.0091 0.1763 Energy for cooking- LP gas -0.2382 0.3150 Energy for cooking- firewood 0.0931 0.2373 Source of water- common well 0.0568 0.0605 Central province -0.1179 0.1046 Southern province 0.2375** 0.1010 Northern province 0.3606 0.2075 Eastern province -0.0803 0.1361 North western province 0.2783* 0.1021 North central province 0.0375 0.1220 Uva province -0.2050 0.1214 Sabaragamuwa province 0.0589 0.1040 Constant -1.8470* 0.5093
Number of observations = 2343 2 (35) 201.33LR chi = 2 0.0629Pseudo R = 2 0.0000prob chi> =
1)Bold indicates 10% significance level. 2)** indicates 5% significance level. 3) * indicates 1% significance level.
77
Table 3.5: 40th-60th percentile probit model for the propensity score Variable Coefficient Standard Error Age -0.0030 0.0024 Female 0.0200 0.0988 Married 0.0435 0.1064 Household size 3-4 0.1384 0.1130 Household size 5-6 0.2303 0.1180 Household size ≥ 7 0.2761 0.1483 Sinhala 0.3754* 0.1256 Tamil 0.1630 0.1681 Primary education 0.0251 0.1238 Secondary education -0.0519 0.1228 Tertiary education -0.2448 0.1577 Regular employee 0.0031 0.1174 Casual employee 0.3144* 0.1067 Self-employed 0.2691* 0.1020 Urban sector 0.1025 0.2361 Rural sector 0.2538 0.2273 Household expenditure 4.0400** 2.0100 Household not owning house 0.0337 0.1200 Lives in a single house 0.1167 0.1777 Lives in a slum/shanty/lineroom -0.1310 0.2479 Floor type- clay/mud/earth 0.1900 0.2140 Floor type-cement 0.0678 0.2018 Main source of lighting- electricity -0.0298 0.1897 Main source of lighting- kerosene 0.1415 0.1953 Energy for cooking- LP gas -0.1678 0.2021 Energy for cooking- firewood 0.1014 0.1760 Source of water- common well 0.0701 0.0658 Central province -0.0510 0.1007 Southern province 0.5240* 0.0890 Northern province -0.2410 0.2218 Eastern province 0.1897 0.1496 North western province 0.5160* 0.0941 North central province 0.0918 0.1295 Uva province 0.0647 0.1341 Sabaragamuwa province 0.2510** 0.1040 Constant -1.5070* 0.4960
Number of observations = 2344 2 (35) 251.47LR chi = 2 0.0807Pseudo R = 2 0.0000prob chi> =
1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.
78
Table 3.6: 60th-80th percentile probit model for the propensity score Variable Coefficient Standard Error Age -0.004 0.0024 Female 0.0489 0.1004 Married 0.0856 0.1122 Household size 3-4 0.0605 0.1350 Household size 5-6 0.1135 0.1391 Household size ≥ 7 0.1271 0.1585 Sinhala 0.6440* 0.1333 Tamil 0.2379 0.1933 Primary education 0.1781 0.1663 Secondary education 0.1424 0.1628 Tertiary education 0.0230 0.1780 Regular employee 0.1587 0.1087 Casual employee 0.2330** 0.1102 Self-employed 0.2437** 0.1010 Urban sector -0.3357 0.2512 Rural sector -0.1485 0.2455 Household expenditure -5.8400 9.1200 Household not owning house -0.3016** 0.1295 Lives in a single house -0.0361 0.1541 Lives in a slum/shanty/lineroom 0.0997 0.2583 Floor type- clay/mud/earth 0.3068 0.1886 Floor type-cement 0.1645 0.1594 Main source of lighting- electricity 0.0253 0.2416 Main source of lighting- kerosene 0.2571 0.2534 Energy for cooking- LP gas -0.1377 0.1816 Energy for cooking- firewood 0.1230 0.1721 Source of water- common well 0.0900 0.0793 Central province -0.0426 0.1026 Southern province 0.4667** 0.0900 Northern province -0.2266 0.2638 Eastern province 0.2042 0.1681 North western province 0.4763* 0.0915 North central province -0.1475 0.1283 Uva province 0.0112 0.1450 Sabaragamuwa province 0.1735 0.1090 Constant -1.5480* 0.5315
Number of observations = 2344 2 (35) 224.81LR chi = 2 0.0765Pseudo R = 2 0.0000prob chi> =
1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.
79
Table 3.7: 80th-100th percentile probit model for the propensity score Variable Coefficient Standard Error Age -0.0073* 0.0026 Female -0.0846 0.1125 Married 0.0303 0.1320 Household size 3-4 0.0479 0.1485 Household size 5-6 0.1567 0.1505 Household size ≥ 7 0.2891 0.1654 Sinhala 0.8650* 0.1705 Tamil 0.3912 0.2287 Primary education 0.2080 0.2359 Secondary education 0.1169 0.2303 Tertiary education -0.0490 0.2389 Regular employee 0.0721 0.0987 Casual employee 0.2968** 0.1160 Self-employed -0.0101 0.0983 Urban sector -0.6509 0.3592 Rural sector -0.4700 0.3570 Household expenditure 3.3900 3.3000 Household not owning house -0.2006 0.1312 Lives in a single house 0.0115 0.1367 Lives in a slum/shanty/lineroom -0.5230 0.4249 Floor type- clay/mud/earth 0.4039** 0.1937 Floor type-cement 0.2512** 0.1050 Main source of lighting- electricity 0.0949 0.2550 Main source of lighting- kerosene 0.2888 0.2860 Energy for cooking- LP gas -0.1400 0.2272 Energy for cooking- firewood 0.1309 0.2277 Source of water- common well 0.0322 0.1089 Central province -0.1470 0.1225 Southern province 0.5799* 0.1084 Northern province -0.0225 0.2870 Eastern province 0.2050 0.1808 North western province 0.4065* 0.0957 North central province -0.1909 0.1428 Uva province 0.1496 0.1759 Sabaragamuwa province 0.1950 0.1316 Constant -1.3960** 0.6222
Number of observations = 2344 2 (35) 278.73LR chi = 2 0.1089Pseudo R =
1) Bold indicates 10% significance level. 2) ** indicates 5% significance level. 3) * indicates 1% significance level.
80
Table 3.8: Matching quality indicators (covariate balancing) for the 20th percentile Mean standardized %reduction t-test Variable Sample Treated Untreated Bias% In |bias| t p>|t| age Unmatched 46.086 49.888 -26.8 -6.23 0 Matched 46.117 46.638 -3.7 86.3 -0.73 0.468 female Unmatched 0.19362 0.26583 -17.2 -4.01 0 Matched 0.19426 0.17439 4.7 72.5 0.97 0.33 married Unmatched 0.81298 0.74182 17.2 4 0 Matched 0.81236 0.82119 -2.1 87.6 -0.43 0.664 Household size 3-4 Unmatched 0.52145 0.46625 11.1 2.61 0.009 Matched 0.52318 0.5287 -1.1 90 -0.21 0.834 Household size 5-6 Unmatched 0.25633 0.17467 19.9 4.78 0 Matched 0.25386 0.24945 1.1 94.6 0.19 0.847 Household size ≥ 7 Unmatched 0.0374 0.03271 2.6 0.61 0.544 Matched 0.03753 0.03753 0 100 0 1 Sinhala Unmatched 0.85149 0.67919 41.5 9.51 0 Matched 0.85099 0.87748 -6.4 84.6 -1.47 0.141 Tamil Unmatched 0.10231 0.21851 -32.1 -7.32 0 Matched 0.10265 0.08389 5.2 83.9 1.23 0.22 Primary education Unmatched 0.42024 0.42171 -0.3 -0.07 0.944 Matched 0.42053 0.45585 -7.2 -2302.9 -1.35 0.176 Secondary education Unmatched 0.38724 0.36395 4.8 1.14 0.256 Matched 0.38631 0.3819 0.9 81 0.17 0.863 Tertiary education Unmatched 0.0242 0.02296 0.8 0.19 0.847 Matched 0.02428 0.01987 2.9 -256.7 0.57 0.568 Regular employee Unmatched 0.033 0.0675 -15.8 -3.61 0 Matched 0.03311 0.03753 -2 87.2 -0.45 0.649 Casual employee Unmatched 0.42024 0.36534 11.3 2.66 0.008 Matched 0.42053 0.39735 4.8 57.8 0.9 0.37 Self-employed Unmatched 0.41804 0.3222 19.9 4.73 0 Matched 0.41722 0.43377 -3.4 82.7 -0.64 0.524 Urban sector Unmatched 0.033 0.06541 -15 -3.43 0.001 Matched 0.03311 0.02649 3.1 79.6 0.74 0.459
81
Table 3.8 continued Mean standardized %reduction t-test Variable Sample Treated control Bias % In |bias| t p>|t|
rural sector Unmatched 0.92409 0.8142 33 7.49 0 Matched 0.92384 0.92936 -1.7 95 -0.4 0.687 Household expenditure Unmatched 7662.2 8273.3 -5.8 -1.31 0.19 Matched 7664 7382 2.7 53.9 0.73 0.465 Household not owning house Unmatched 0.08141 0.18163 -30 -6.82 0 Matched 0.08168 0.07395 2.3 92.3 0.55 0.583 Lives in a single house Unmatched 0.94169 0.85943 27.7 6.29 0 Matched 0.9415 0.93488 2.2 91.9 0.52 0.601 Lives in a slum/shanty/lineroom Unmatched 0.0484 0.11065 -23.1 -5.25 0 Matched 0.04857 0.05298 -1.6 92.9 -0.38 0.702 Floor type- clay/mud/earth Unmatched 0.43234 0.34586 17.8 4.22 0 Matched 0.43157 0.40177 6.1 65.5 1.15 0.25 Floor type-cement Unmatched 0.55776 0.6444 -17.8 -4.21 0 Matched 0.5585 0.57616 -3.6 79.6 -0.68 0.498 Main source of lighting- electricity Unmatched 0.41034 0.49408 -16.9 -3.98 0 Matched 0.4117 0.43488 -4.7 72.3 -0.89 0.372 Main source of lighting- kerosene Unmatched 0.56876 0.49548 14.7 3.47 0.001 Matched 0.56954 0.55188 3.5 75.9 0.68 0.499 Energy for cooking- LP gas Unmatched 0.0055 0.01461 -9.1 -2.06 0.04 Matched 0.00552 0.00331 2.2 75.8 0.63 0.527 Energy for cooking- firewood Unmatched 0.9879 0.96381 15.7 3.52 0 Matched 0.98786 0.99338 -3.6 77.1 -1.09 0.276 Source of water- common well Unmatched 0.38614 0.34029 9.5 2.26 0.024 Matched 0.38631 0.38852 -0.5 95.2 -0.09 0.931 Central province Unmatched 0.14741 0.16562 -5 -1.18 0.24 Matched 0.1479 0.16998 -6.1 -21.2 -1.15 0.251 Southern province Unmatched 0.17492 0.10369 20.7 5 0 Matched 0.17439 0.16446 2.9 86.1 0.5 0.615
82
Table 3.8 continued
Variable Sample Treated control Bias % In |bias| t p>|t| Northern province Unmatched 0.0154 0.04036 -15.2 -3.42 0.001 Matched 0.01545 0.00773 4.7 69 1.37 0.17 Eastern province Unmatched 0.09021 0.15031 -18.5 -4.27 0 Matched 0.09051 0.07837 3.7 79.8 0.83 0.406 North western province Unmatched 0.11441 0.10647 2.5 0.6 0.549 Matched 0.11479 0.12804 -4.2 -66.8 -0.77 0.441 North central province Unmatched 0.07811 0.07933 -0.5 -0.11 0.915 Matched 0.07837 0.09272 -5.3 -1072.2 -0.98 0.329 Uva province Unmatched 0.13641 0.10786 8.7 2.08 0.037 Matched 0.13687 0.11921 5.4 38.1 1.01 0.315 Sabaragamuwa province Unmatched 0.17382 0.13848 9.7 2.32 0.02 Matched 0.17219 0.17108 0.3 96.9 0.06 0.956
Pseudo-2R
Unmatched 0.092 Matched 0.012
83
Table 3.9: Matching quality indicators (covariate balancing) for the 20th-40 percentile Variable Mean standardized %reduction t-test Sample Treated Control Bias% In |bias| t p>|t| age Unmatched 45.708 46.124 -3.3 -0.77 0.439 Matched 45.708 45.535 1.4 58.5 0.28 0.781 female Unmatched 0.16087 0.17066 -2.6 -0.63 0.529 Matched 0.16087 0.18371 -6.1 -133.4 -1.21 0.227 married Unmatched 0.87488 0.85629 5.4 1.3 0.194 Matched 0.87488 0.85799 4.9 9.2 0.99 0.321 Household size 3-4 Unmatched 0.50348 0.51347 -2 -0.48 0.632 Matched 0.50348 0.51043 -1.4 30.5 -0.28 0.781 Household size 5-6 Unmatched 0.34657 0.29042 12.1 2.9 0.004 Matched 0.34657 0.33664 2.1 82.3 0.42 0.676 Household size ≥ 7 Unmatched 0.07249 0.06362 3.5 0.85 0.397 Matched 0.07249 0.07547 -1.2 66.4 -0.23 0.82 Sinhala Unmatched 0.84806 0.69237 37.6 8.87 0 Matched 0.84806 0.84608 0.5 98.7 0.11 0.912 Tamil Unmatched 0.0993 0.21781 -32.9 -7.71 0 Matched 0.0993 0.1003 -0.3 99.2 -0.07 0.947 Primary education Unmatched 0.37239 0.36228 2.1 0.5 0.615 Matched 0.37239 0.37537 -0.6 70.6 -0.12 0.902 Secondary education Unmatched 0.48163 0.49551 -2.8 -0.67 0.506 Matched 0.48163 0.47071 2.2 21.3 0.44 0.662 Tertiary education Unmatched 0.03674 0.04491 -4.1 -0.98 0.326 Matched 0.03674 0.03178 2.5 39.2 0.55 0.585 Regular employee Unmatched 0.07646 0.12201 -15.3 -3.61 0 Matched 0.07646 0.10526 -9.7 36.8 -2.01 0.045 Casual employee Unmatched 0.46773 0.40344 13 3.12 0.002 Matched 0.46773 0.47071 -0.6 95.4 -0.12 0.905 Self-employed Unmatched 0.36544 0.34431 4.4 1.06 0.29 Matched 0.36544 0.33664 6 -36.3 1.21 0.228 Urban sector Unmatched 0.04469 0.07111 -11.3 -2.67 0.008 Matched 0.04469 0.04767 -1.3 88.7 -0.28 0.777 rural sector Unmatched 0.91063 0.81063 29.2 6.84 0
84
Table 3.9 continued
Variable Mean standardized %reduction t-test Sample Treated control Bias% In |bias| t p>|t| Household expenditure Unmatched 10831 11236 -3.7 -0.89 0.371 Matched 10831 10505 3 19.4 0.62 0.538 Household not owning house Unmatched 0.09037 0.16392 -22.2 -5.23 0 Matched 0.09037 0.1003 -3 86.5 -0.68 0.499 Lives in a single house Unmatched 0.94935 0.86228 30.1 7.02 0 Matched 0.94935 0.94141 2.7 90.9 0.7 0.485 Lives in a slum/shanty/lineroom Unmatched 0.03277 0.10329 -28.3 -6.55 0 Matched 0.03277 0.0427 -4 85.9 -1.04 0.298 Floor type- clay/mud/earth Unmatched 0.28004 0.21482 15.2 3.65 0 Matched 0.28004 0.26415 3.7 75.6 0.71 0.476 Floor type-cement Unmatched 0.71301 0.76796 -12.6 -3.02 0.003 Matched 0.71301 0.73188 -4.3 65.7 -0.84 0.4 Main source of lighting- electricity Unmatched 0.57895 0.63174 -10.8 -2.59 0.01 Matched 0.57895 0.59285 -2.8 73.7 -0.56 0.573 Main source of lighting- kerosene Unmatched 0.39523 0.34356 10.7 2.57 0.01 Matched 0.39523 0.38332 2.5 76.9 0.49 0.625 Energy for cooking- LP gas Unmatched 0.00993 0.02695 -12.7 -2.94 0.003 Matched 0.00993 0.00596 3 76.7 0.89 0.371 Energy for cooking- firewood Unmatched 0.98014 0.95135 15.9 3.7 0 Matched 0.98014 0.9861 -3.3 79.3 -0.92 0.355 Source of water- common well Unmatched 0.36445 0.29566 14.7 3.53 0 Matched 0.36445 0.37041 -1.3 91.3 -0.25 0.805 Central province Unmatched 0.12115 0.18413 -17.6 -4.16 0 Matched 0.12115 0.11619 1.4 92.1 0.31 0.759 Southern province Unmatched 0.18173 0.11452 19 4.61 0 Matched 0.18173 0.16882 3.6 80.8 0.68 0.497 Northern province Unmatched 0.01986 0.04491 -14.2 -3.31 0.001 Matched 0.01986 0.02085 -0.6 96 -0.14 0.888
85
Table 3.9 continued Variable Mean standardized %reduction t-test Sample Treated control Bias% In |bias| t p>|t| North western province Unmatched 0.1857 0.11751 19.1 4.64 0 Matched 0.1857 0.19662 -3.1 84 -0.56 0.579 North central province Unmatched 0.0864 0.06886 6.6 1.58 0.114 Matched 0.0864 0.08937 -1.1 83 -0.21 0.833 Uva province Unmatched 0.07051 0.10105 -10.9 -2.59 0.01 Matched 0.07051 0.07051 0 100 0 1 Sabaragamuwa province Unmatched 0.13406 0.11826 4.8 1.14 0.253 Matched 0.13406 0.13505 -0.3 93.7 -0.06 0.954
Pseudo-2R
Unmatched 0.109 Matched 0.018
86
Table 3.10: Matching quality indicators (covariate balancing) for the 40th-60th percentile Mean standardized %reduction t-test Variable Sample Treated Control Bias% In |bias| t p>|t| age Unmatched 46.056 46.537 -3.9 -0.92 0.359 Matched 46.17 46.219 -0.4 89.9 -0.08 0.939 female Unmatched 0.13982 0.17655 -10.1 -2.34 0.019 Matched 0.14172 0.13265 2.5 75.3 0.5 0.618 married Unmatched 0.89374 0.87241 6.6 1.55 0.122 Matched 0.89229 0.89796 -1.8 73.4 -0.35 0.726 Household size 3-4 Unmatched 0.47204 0.49655 -4.9 -1.15 0.249 Matched 0.47619 0.45692 3.9 21.4 0.73 0.465 Household size 5-6 Unmatched 0.38479 0.33517 10.3 2.44 0.015 Matched 0.38435 0.36168 4.7 54.3 0.89 0.375 Household size ≥ 7 Unmatched 0.08725 0.07379 4.9 1.17 0.241 Matched 0.08277 0.10317 -7.5 -51.7 -1.33 0.184 Sinhala Unmatched 0.88591 0.77034 31 7.06 0 Matched 0.88435 0.87868 1.5 95.1 0.33 0.74 Tamil Unmatched 0.06488 0.13655 -24 -5.44 0 Matched 0.06576 0.06236 1.1 95.3 0.26 0.793 Primary education Unmatched 0.32662 0.26759 12.9 3.06 0.002 Matched 0.32313 0.36054 -8.2 36.6 -1.49 0.135 Secondary education Unmatched 0.54586 0.56897 -4.7 -1.09 0.274 Matched 0.55215 0.52721 5 -8 0.95 0.344 Tertiary education Unmatched 0.06152 0.10414 -15.5 -3.55 0 Matched 0.06236 0.05329 3.3 78.7 0.74 0.462 Regular employee Unmatched 0.11633 0.19586 -22 -5.06 0 Matched 0.11791 0.10884 2.5 88.6 0.54 0.588 Casual employee Unmatched 0.39374 0.31241 17.1 4.04 0 Matched 0.38662 0.35147 7.4 56.8 1.38 0.168 Self-employed Unmatched 0.40828 0.35379 11.2 2.65 0.008 Matched 0.4127 0.44785 -7.2 35.5 -1.34 0.179 Urban sector Unmatched 0.06711 0.13724 -23.3 -5.29 0 Matched 0.06803 0.0771 -3 87.1 -0.66 0.508 rural sector Unmatched 0.91387 0.81241 29.8 6.76 0
87
Table 3.10 continued Mean standardized %reduction t-test Variable Sample Treated Control Bias % In |bias| t p>|t| Household expenditure Unmatched 14940 16653 -11.3 -2.54 0.011 Matched 14985 15724 -4.9 56.9 -1.13 0.261 Household not owning house Unmatched 0.06264 0.11379 -18.1 -4.13 0 Matched 0.06349 0.06349 0 100 0 1 Lives in a single house Unmatched 0.96421 0.91172 21.9 4.92 0 Matched 0.96372 0.95692 2.8 87 0.66 0.51 Lives in a slum/shanty/lineroom Unmatched 0.01678 0.04828 -17.8 -3.97 0 Matched 0.01701 0.01814 -0.6 96.4 -0.16 0.87 Floor type- clay/mud/earth Unmatched 0.18456 0.11586 19.3 4.65 0 Matched 0.17347 0.178 -1.3 93.4 -0.23 0.822 Floor type-cement Unmatched 0.79866 0.86207 -16.9 -4.06 0 Matched 0.80952 0.80272 1.8 89.3 0.33 0.745 Main source of lighting- electricity Unmatched 0.71477 0.81034 -22.6 -5.4 0 Matched 0.72449 0.69955 5.9 73.9 1.04 0.297 Main source of lighting- kerosene Unmatched 0.26063 0.17034 22.1 5.29 0 Matched 0.25057 0.28231 -7.8 64.8 -1.36 0.174 Energy for cooking- LP gas Unmatched 0.03579 0.09379 -23.7 -5.32 0 Matched 0.03628 0.04082 -1.9 92.2 -0.45 0.656 Energy for cooking- firewood Unmatched 0.94519 0.85931 29.2 6.57 0 Matched 0.94444 0.94898 -1.5 94.7 -0.38 0.702 Source of water- common well Unmatched 0.29418 0.22207 16.5 3.93 0 Matched 0.29138 0.29932 -1.8 89 -0.33 0.742 Central province Unmatched 0.09955 0.15448 -16.5 -3.81 0 Matched 0.10091 0.11338 -3.8 77.3 -0.76 0.446 Southern province Unmatched 0.22371 0.12138 27.3 6.62 0 Matched 0.21882 0.21315 1.5 94.5 0.26 0.794 Northern province Unmatched 0.0179 0.04828 -17 -3.81 0 Matched 0.01814 0.02154 -1.9 88.8 -0.46 0.644
88
Table 3.10. continued. Mean standardized %reduction t-test Variable Sample Treated Control Bias % In |bias| t p>|t| North western province Unmatched 0.19239 0.10207 25.7 6.24 0 Matched 0.18934 0.19728 -2.3 91.2 -0.38 0.704 North central province Unmatched 0.05481 0.05586 -0.5 -0.11 0.914 Matched 0.05556 0.03968 6.9 -1408.5 1.41 0.158 Uva province Unmatched 0.05145 0.05655 -2.3 -0.53 0.598 Matched 0.05215 0.04649 2.5 -11.2 0.5 0.62 Sabaragamuwa province Unmatched 0.11409 0.0931 6.9 1.64 0.102 Matched 0.11565 0.14512 -9.7 -40.4 -1.66 0.098
Pseudo-2R
Unmatched 0.081 Matched 0.015
89
Table 3.11: Matching quality indicators (covariate balancing) for 60th-80th percentile
Mean standardized %reduction t-test Variable Sample Treated Control Bias% In |bias| t p>|t| age Unmatched 47.207 48.216 -8.3 -1.85 0.064 Matched 47.251 47.531 -2.3 72.3 -0.42 0.676 female Unmatched 0.13733 0.15496 -5 -1.12 0.264 Matched 0.13673 0.12869 2.3 54.4 0.42 0.676 married Unmatched 0.90533 0.88833 5.6 1.25 0.213 Matched 0.90483 0.90617 -0.4 92.1 -0.08 0.936 Household size 3-4 Unmatched 0.44133 0.46989 -5.7 -1.29 0.196 Matched 0.44236 0.38874 10.8 -87.8 1.92 0.055 Household size 5-6 Unmatched 0.396 0.35006 9.5 2.16 0.031 Matched 0.3941 0.40885 -3.1 67.9 -0.53 0.596 Household size ≥ 7 Unmatched 0.12 0.11606 1.2 0.28 0.782 Matched 0.12064 0.15952 -12 -886.7 -1.98 0.048 Sinhala Unmatched 0.916 0.79548 34.8 7.4 0 Matched 0.91555 0.92091 -1.5 95.6 -0.34 0.73 Tamil Unmatched 0.04667 0.10916 -23.5 -4.98 0 Matched 0.04692 0.04021 2.5 89.3 0.58 0.563 Primary education Unmatched 0.23867 0.20075 9.2 2.09 0.036 Matched 0.23727 0.24263 -1.3 85.9 -0.22 0.825 Secondary education Unmatched 0.58933 0.57779 2.3 0.53 0.597 Matched 0.58981 0.60054 -2.2 7.1 -0.39 0.7 Tertiary education Unmatched 0.14133 0.18444 -11.7 -2.59 0.01 Matched 0.14209 0.13941 0.7 93.8 0.14 0.892 Regular employee Unmatched 0.22533 0.27415 -11.3 -2.52 0.012 Matched 0.22654 0.22788 -0.3 97.3 -0.06 0.955 Casual employee Unmatched 0.24933 0.2064 10.2 2.34 0.019 Matched 0.24665 0.21046 8.6 15.7 1.52 0.129 Self-employed Unmatched 0.43333 0.38331 10.2 2.31 0.021 Matched 0.43432 0.48257 -9.8 3.5 -1.71 0.088 Urban sector Unmatched 0.08667 0.1744 -26.3 -5.64 0 Matched 0.08713 0.09651 -2.8 89.3 -0.57 0.567
90
Table 3.11. continued. Mean standardized %reduction t-test Variable Sample Treated Bias % Bias % In |bias| t p>|t| rural sector Unmatched 0.88933 0.79235 26.7 5.78 0 Matched 0.88874 0.88606 0.7 97.2 0.15 0.881 Household expenditure Unmatched 23305 24988 -4.4 -0.89 0.375 Matched 23344 23619 -0.7 83.7 -0.2 0.839 Household not owning house Unmatched 0.05067 0.101 -19.1 -4.09 0 Matched 0.05094 0.03753 5.1 73.4 1.15 0.25 Lives in a single house Unmatched 0.95067 0.91656 13.7 2.97 0.003 Matched 0.9504 0.96113 -4.3 68.6 -0.92 0.358 Lives in a slum/shanty/lineroom Unmatched 0.02133 0.02949 -5.2 -1.14 0.255 Matched 0.02145 0.01475 4.3 17.8 0.89 0.375 Floor type- clay/mud/earth Unmatched 0.104 0.06148 15.5 3.65 0 Matched 0.0992 0.09786 0.5 96.8 0.08 0.937 Floor type-cement Unmatched 0.86933 0.89523 -8 -1.85 0.065 Matched 0.87399 0.88204 -2.5 68.9 -0.43 0.665 Main source of lighting- electricity Unmatched 0.83733 0.90715 -21 -4.96 0 Matched 0.84182 0.85255 -3.2 84.6 -0.53 0.599 Main source of lighting- kerosene Unmatched 0.148 0.0803 21.4 5.08 0 Matched 0.14343 0.13807 1.7 92.1 0.27 0.786 Energy for cooking- LP gas Unmatched 0.11867 0.23149 -30 -6.48 0 Matched 0.1193 0.13673 -4.6 84.6 -0.92 0.358 Energy for cooking- firewood Unmatched 0.85867 0.72334 33.7 7.3 0 Matched 0.85791 0.84853 2.3 93.1 0.47 0.64 Source of water- common well Unmatched 0.19067 0.14053 13.5 3.12 0.002 Matched 0.18767 0.19437 -1.8 86.6 -0.3 0.764 Central province Unmatched 0.08933 0.12986 -13 -2.85 0.004 Matched 0.08981 0.07373 5.2 60.3 1.04 0.301 Southern province Unmatched 0.192 0.09975 26.3 6.26 0 Matched 0.19035 0.19571 -1.5 94.2 -0.24 0.811
91
Table 3.11. Continued. Mean standardized %reduction t-test Variable Sample Treated Bias % Bias % In |bias| t p>|t| Eastern province Unmatched 0.03333 0.05332 -9.8 -2.14 0.033 Matched 0.03351 0.03351 0 100 0 1 North western province Unmatched 0.19733 0.10289 26.7 6.33 0 Matched 0.19571 0.17828 4.9 81.5 0.79 0.431 North central province Unmatched 0.05467 0.06399 -3.9 -0.88 0.379 Matched 0.05496 0.06971 -6.2 -58.2 -1.08 0.282 Uva province Unmatched 0.04133 0.04391 -1.3 -0.29 0.774 Matched 0.04155 0.0429 -0.7 48.1 -0.12 0.906 Sabaragamuwa province Unmatched 0.096 0.07403 7.9 1.82 0.069 Matched 0.09651 0.09786 -0.5 93.9 -0.08 0.936
Pseudo-2R
Unmatched 0.077 Matched 0.013
92
Table 3.12: Matching quality indicators (covariate balancing) for the 80th-100 percentile Mean standardized %reduction t-test Variable Sample Treated Bias % Bias% In |bias| t p>|t| age Unmatched 49.014 50.049 -8.5 -1.75 0.081 Matched 49.025 49.033 -0.1 99.3 -0.01 0.993 female Unmatched 0.11775 0.15123 -9.8 -1.96 0.05 Matched 0.11818 0.08727 9.1 7.7 1.57 0.117 married Unmatched 0.91304 0.9029 3.5 0.71 0.477 Matched 0.91273 0.92545 -4.4 -25.5 -0.72 0.473 Household size 3-4 Unmatched 0.37862 0.44029 -12.6 -2.56 0.01 Matched 0.38 0.40727 -5.6 55.8 -0.86 0.39 Household size 5-6 Unmatched 0.4058 0.36998 7.4 1.52 0.129 Matched 0.40727 0.38727 4.1 44.2 0.63 0.529 Household size ≥ 7 Unmatched 0.17572 0.12333 14.7 3.15 0.002 Matched 0.17273 0.16909 1 93.1 0.15 0.882 Sinhala Unmatched 0.94022 0.81417 39.1 7.21 0 Matched 0.94 0.94545 -1.7 95.7 -0.36 0.718 Tamil Unmatched 0.03623 0.0904 -22.4 -4.17 0 Matched 0.03636 0.03636 0 100 0 1 Primary education Unmatched 0.17754 0.11272 18.5 4 0 Matched 0.17455 0.20545 -8.8 52.3 -1.21 0.225 Secondary education Unmatched 0.55254 0.51674 7.2 1.47 0.141 Matched 0.55455 0.51273 8.4 -16.8 1.29 0.197 Tertiary education Unmatched 0.25181 0.35156 -21.8 -4.38 0 Matched 0.25273 0.25091 0.4 98.2 0.06 0.949 Regular employee Unmatched 0.31341 0.36105 -10.1 -2.05 0.04 Matched 0.31455 0.27091 9.2 8.4 1.48 0.14 Casual employee Unmatched 0.18478 0.10882 21.6 4.71 0 Matched 0.18182 0.21273 -8.8 59.3 -1.2 0.232 Self-employed Unmatched 0.37681 0.35379 4.8 0.99 0.325 Matched 0.37818 0.40727 -6 -26.4 -0.92 0.359
93
Table 3.12 continued. Mean standardized %reduction t-test Variable Sample Treated Variable Bias % In |bias| t p>|t| Urban sector Unmatched 0.13587 0.29855 -40.2 -7.71 0 Matched 0.13636 0.15273 -4 89.9 -0.72 0.474 rural sector Unmatched 0.85326 0.6942 38.7 7.46 0 Matched 0.85273 0.83818 3.5 90.9 0.62 0.536 Household expenditure Unmatched 47187 51170 -4.5 -0.88 0.381 Matched 47281 42732 5.2 -14.2 0.99 0.32 Household not owning house Unmatched 0.0471 0.09821 -19.8 -3.75 0 Matched 0.04727 0.04727 0 100 0 1 Lives in a single house Unmatched 0.95471 0.89732 22 4.15 0 Matched 0.95455 0.94727 2.8 87.3 0.52 0.605 Lives in a slum/shanty/lineroom Unmatched 0.00362 0.01004 -7.8 -1.43 0.152 Matched 0.00364 0.00364 0 100 0 1 Floor type- clay/mud/earth Unmatched 0.06341 0.02734 17.4 4 0 Matched 0.06 0.07273 -6.1 64.7 -0.79 0.431 Floor type-cement Unmatched 0.86413 0.80301 16.5 3.26 0.001 Matched 0.86727 0.83818 7.8 52.4 1.26 0.206 Main source of lighting- electricity Unmatched 0.88768 0.94922 -22.6 -5.14 0 Matched 0.89091 0.87091 7.3 67.5 0.95 0.342 Main source of lighting- kerosene Unmatched 0.09601 0.04018 22.3 5.13 0 Matched 0.09273 0.10364 -4.4 80.5 -0.56 0.573 Energy for cooking- LP gas Unmatched 0.25362 0.48828 -50 -9.93 0 Matched 0.25455 0.24182 2.7 94.6 0.45 0.65 Energy for cooking- firewood Unmatched 0.73188 0.48605 52 10.37 0 Matched 0.73091 0.75091 -4.2 91.9 -0.7 0.483 Source of water- common well Unmatched 0.11957 0.07533 14.9 3.25 0.001 Matched 0.11818 0.11636 0.6 95.9 0.09 0.931 Central province Unmatched 0.06159 0.09542 -12.6 -2.46 0.014 Matched 0.06182 0.05636 2 83.9 0.36 0.722
94
Table 3.12 continued. Mean standardized %reduction t-test Variable Sample Treated Variable Bias% In |bias| t p>|t| Northern province Unmatched 0.01268 0.0279 -10.8 -2.03 0.042 Matched 0.01273 0.02182 -6.5 40.3 -1.07 0.283 Eastern province Unmatched 0.03442 0.04967 -7.6 -1.49 0.135 Matched 0.03455 0.03636 -0.9 88.1 -0.15 0.88 North western province Unmatched 0.19384 0.09208 29.4 6.58 0 Matched 0.19273 0.18909 1 96.4 0.14 0.887 North central province Unmatched 0.05435 0.05357 0.3 0.07 0.944 Matched 0.05455 0.06 -2.4 -602.5 -0.36 0.718 Uva province Unmatched 0.03623 0.02734 5.1 1.08 0.28 Matched 0.03636 0.05091 -8.3 -63.7 -1.1 0.273 Sabaragamuwa province Unmatched 0.07246 0.05134 8.8 1.88 0.06 Matched 0.07273 0.07091 0.8 91.4 0.11 0.914
Pseudo-2R
Unmatched 0.109 Matched 0.018
95
Table 3.13: Individuals lost due to common support requirement (%) Matching Algorithm
0-20th
percentile
20th-40th percentile
40th-60th percentile
60th-80th percentile
80th-100th percentile
NN 0.255 0.59 0.59 1.27 1.74 NN (caliper) δ=0.01 0.68 0.72 0.63 1.57 1.91 δ=0.001 15.68 8.83 9.85 11.51 20.26 Radius δ=0.01 0.68 0.72 0.63 1.57 1.91 δ=0.001 15.68 8.83 9.85 11.51 20.26 Kernel Epanechnikov (bw=0.1)
0.255 0.59 0.59 1.27 1.74
Epanechnikov (bw=0.2)
0.255 0.59 0.59 1.27 1.74
Gaussian (bw=0.1)
0.255 0.59 0.59 1.27 1.74
Gaussian (bw=0.2)
0.255 0.59 0.59 1.27 1.74
96
Table 3.14: Impact of microfinance on household savings Matching Algorithm
0-20th
percentile ATT ATE
20th-40th percentile ATT ATE
40th-60th percentile ATT ATE
60th-80th percentile ATT ATE
80th-100th percentile ATT ATE
NN 293.09 57.60 393.67 423.17 60.98 69.62 678.71 54.74 -1616.28 -1387.18 NN (caliper) δ=0.01 293.49 56.40 393.73 423.24 60.98 62.95 678.71 54.49 -1617.07 -1389.55 δ=0.001 287.05 70.82 429.06 344.64 72.05 -40.27 861.58 403.52 -291.68 -1538.20 Radius δ=0.01 251.03 313.88 297.52 431.54 252.24 507.07 286.79 113.73 -1329.95 -1628.91 δ=0.001 371.42 333.58 427.05 426.34 214.44 45.48 857.09 704.32 -1063.78 -1893.89 Kernel Epanechnikov (bw=0.1)
215.26 261.43 298.05 391.71 216.10 347.12 242.88 142.03 -1325.03 -1755.08
Epanechnikov (bw=0.2)
158.06 210.17 244.77 303.28 168.78 140.34 195.75 194.47 -1449.91 -1812.34
Gaussian (bw=0.1)
148.13 192.93 236.60 296.97 153.3 116.52 195.14 202.00 -1469.92 -18176.38
Gaussian (bw=0.2)
37.73 58.06 186.37 231.80 2.90 -67.91 137.67 194.00 -1723.51 -1920.53
97
Table 3.15: Impact of microfinance on household income Matching Algorithm
0-20th
percentile ATT ATE
20th-40th percentile ATT ATE
40th-60th percentile ATT ATE
60th-80th percentile ATT ATE
80th-100th percentile ATT ATE
NN -20.87 -11.70 -9.84 14.07 -154.95 -65.52 -100.70 71.95 -951.69 -1047.76 NN (caliper) δ=0.01 -21.71 -12.50 -8.33 14.75 -154.95 -67.12 -100.70 68.21 -936.33 -1043.14 δ=0.001 -16.62 -19.95 -4.92 9.81 -175.1 -81.35 -101.56 80.54 -985.76 -877.35 Radius δ=0.01 4.85 -8.91 -18.94 -28.06 -39.35 -34.78 -58.47 -82.05 -911.85 -1276.17 δ=0.001 -12.12 -14.81 -16.86 -12.05 -90.09 -68.90 -27.72 12.44 -1152.63 -1016.00 Kernel Epanechnikov (bw=0.1)
-13.48 -19.04 -25.64 -33.65 -62.34 -48.72 -82.59 -97.25 -1060.08 -1541.74
Epanechnikov (bw=0.2)
-44.37 -45.96 -63.93 -61.54 -90.88 -104.84 -132.17 -148.64 -1397.12 -1853.59
Gaussian (bw=0.1)
-50.74 -51.48 -68.86 -67.93 -96.79 -112.34 -142.39 -159.24 -1471.09 -1893.93
Gaussian (bw=0.2)
-113.91 -111.53 -123.33 -123.47 -161.48 -180.82 -226.94 -233.12 -2181.90 -2346.53
98
020
4060
80F
requ
ency
0 .2 .4 .6 .8Estimated p(x)
Treated
020
4060
80F
requ
ency
0 .2 .4 .6 .8Estimated p(x)
Untreated
Figure 3.1: Histogram of the estimated propensity score (bottom 20th percentile)
99
020
4060
8010
0F
requ
ency
0 .2 .4 .6 .8Estimated p(x)
Treated
020
4060
8010
0F
requ
ency
0 .2 .4 .6 .8Estimated p(x)
Untreated
Figure 3.2: Histogram of the estimated propensity score (20th-40th percentile)
100
050
100
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Treated
050
100
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Untreated
Figure 3.3: Histogram of the estimated propensity score (40th-60th percentile)
101
050
100
150
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Treated
050
100
150
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Untreated
Figure 3.4: Histogram of the estimated propensity score (60th-80th percentile)
102
050
100
150
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Treated
050
100
150
Fre
quen
cy
0 .2 .4 .6 .8Estimated p(x)
Untreated
Figure 3.5: Histogram of the estimated propensity score (80th-100th percentile)
103
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, Smith’s 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?
104
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 non-
communist 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
105
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 non-
minorities.
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
106
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:
11
1 (1 )
T ss s
s st s
W W W Cr
−−
−=
−= +
+∑
Where sr 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 ths year of education if sr 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:
2log i i i iw edu exp expβ ϕ δ γ µ= Χ + + + +
Where iw is the hourly wage, edu years of schooling, exp is the years of experience and
iΧ represents set of other variables such gender, sector of work, etc. Experience is
107
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
108
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
109
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
110
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)
111
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: 0 0 0 1ln exp expi i i iw edu sqrα β δ δ ε= + + + +
Where ln iw 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
112
system; primary, secondary and tertiary education. Thus, the empirical estimation takes
form;
Model-2: 0 0 1 2 0 1ln sec exp expi i i i i i iw primary ondary tertiary sqrα β β β δ δ ε= + + + + + +
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 thi individual to differ across the earnings spectrum.
The quantile regression model can be expressed as:
,ln i i iw x τ τβ µ′= +
Where ln iw is the natural log of the hourly wage of the thi individual, and ix represents
the demographic and employment related characteristics of the thi individual. By
Imposing the assumption that the thτ − quantile of the error term conditional on the
regressors is zero, ,( ( | ) 0)i iQ u xτ τ = , the thτ − conditional quantile of ln iw with respect
to ix can be expressed as:
(ln | )i i iQ w x xτ τβ′=
For any ε ∈(0,1), the parameter τβ can be estimated by
113
{ } { }| |
ˆ arg min | | (1 ) | |k
i i i i
i i i ii i y x i i y x
y x y xτ τ τ
τβ β β
β τ β τ β∈ ′ ′∈ ≥ ∈ ≥
′ ′= − + − − ∑ ∑
R
Note, that when 0.5τ = , 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
114
study contains workers who worked at least 30 hours per week and were not self-
employed. 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
115
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
116
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
117
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 sub-
samples. 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.
118
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.
119
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.
120
Table 4.1: Cross Country evidence on the Returns to Schooling (year-1995)
Source: Harmon, Oosterbeek and Walker (2000). Note: Standard Errors in italics.
121
Table 4.2: Quantile and OLS regression estimates (model 1-full sample)
coefficient
Bootstrap standard error t-value
P>|t|
q10 edu 0.006319 0.001139 5.550 0.000 exp 0.011061 0.002136 5.180 0.000 expsqr -0.000340 5.660005 -5.960 0.000 consant 1.064279 0.023498 45.29 0.000 q25 edu 0.011317 0.001593 7.100 0.000 exp 0.017371 0.001683 10.32 0.000 expsqr -0.000490 4.980005 -9.780 0.000 constant 1.154758 0.019620 58.86 0.000 q50 edu 0.025822 0.001799 14.35 0.000 exp 0.019514 0.001388 14.06 0.000 expsqr -0.000470 4.520005 -10.39 0.000 constant 1.172323 0.020842 56.25 0.000 q75 edu 0.035738 0.001537 23.25 0.000 exp 0.019776 0.002179 9.080 0.000 expsqr -0.00040 7.210005 -5.560 0.000 constant 1.232041 0.019220 64.10 0.000 q90 edu 0.046208 0.001819 25.41 0.000 exp 0.019536 0.002074 9.420 0.000 expsqr -0.000300 6.180005 -4.890 0.000 constant 1.298972 0.022178 58.57 0.000 OLS edu 0.019337 0.001111 17.40 0.000 exp 0.016851 0.001321 12.76 0.000 expsqr -0.000380 3.680005 -10.29 0.000 constant 1.246002 0.014925 83.48 0.000
122
Table 4.3: Quantile and OLS regression estimates (model 1-Sinhala)
coefficient
Bootstrap standard error t-value
P>|t|
q10 edu 0.008043 0.002173 3.700 0.000 exp 0.014276 0.003162 4.520 0.000 expsqr -0.000460 9.390005 -4.880 0.000 _cons 1.068700 0.029267 36.52 0.000 q25 edu 0.015461 0.002322 6.660 0.000 exp 0.019509 0.001942 10.04 0.000 expsqr -0.000540 6.180005 -8.690 0.000 _cons 1.127452 0.027466 41.05 0.000 q50 edu 0.030930 0.002308 12.99 0.000 exp 0.021360 0.001585 13.48 0.000 expsqr -0.000480 0.000054 -8.940 0.000 _cons 1.128302 0.026159 43.13 0.000 q75 edu 0.038618 0.002080 18.56 0.000 exp 0.020451 0.002173 9.410 0.000 expsqr -0.000380 6.590005 -5.780 0.000 _cons 1.203375 0.023665 50.85 0.000 q90 edu 0.048383 0.002029 23.84 0.000 exp 0.021267 0.002091 10.17 0.000 expsqr -0.000330 6.250005 -5.260 0.000 _cons 1.270688 0.024245 52.41 0.000 OLS edu 0.023714 0.001355 17.50 0.000 exp 0.018623 0.001497 12.44 0.000 expsqr -0.000410 4.230005 -9.690 0.000 _cons 1.209412 0.017709 68.29 0.000
123
Table 4.4: Quantile and OLS regression estimates (model 1-Tamil)
coefficient
Bootstrap standard error t-value
P>|t|
q10 edu 0.002938 0.003189 0.920 0.357 exp 0.005744 0.004570 1.260 0.209 expsqr -0.000110 0.000108 -1.010 0.311 constant 1.006615 0.054497 18.47 0.000 q25 edu 0.003880 0.001845 2.100 0.036 exp 0.007455 0.003504 2.130 0.034 expsqr -0.000180 8.290005 -2.200 0.028 constant 1.140844 0.037250 30.63 0.000 q50 edu 0.008106 0.003252 2.490 0.013 exp 0.007632 0.004017 1.900 0.058 expsqr -0.000200 9.560005 -2.060 0.039 constant 1.275534 0.051122 24.95 0.000 q75 edu 0.021085 0.005574 3.780 0.000 exp 0.013077 0.004414 2.960 0.003 expsqr -0.000300 0.000131 -2.270 0.024 constant 1.302319 0.058955 22.09 0.000 q90 edu 0.034554 0.006039 5.720 0.000 exp 0.011229 0.006323 1.780 0.076 expsqr -0.000100 0.000198 -0.520 0.605 constant 1.347976 0.063310 21.29 0.000 OLS edu 0.008564 0.002080 4.120 0.000 exp 0.008709 0.003030 2.870 0.004 expsqr -0.000170 0.000078 -2.240 0.025 constant 1.272718 0.031635 40.23 0.000
124
Table 4.5: Quantile and OLS regression estimates (model 2-full sample)
Coefficient
Bootstrap Standard Error. t-value
P>|t|
q10 primary 0.048946 0.050912 0.960 0.336 secondary 0.205679 0.049757 4.130 0.000 tertiary 0.443819 0.049191 9.020 0.000 exp 0.018826 0.002060 9.140 0.000 expsqr -0.000420 6.230005 -6.720 0.000 constant 0.859709 0.049047 17.53 0.000 q25 primary 0.052829 0.033157 1.590 0.111 secondary 0.208309 0.031062 6.710 0.000 tertiary 0.430995 0.030209 14.27 0.000 exp 0.018944 0.001414 13.40 0.000 expsqr -0.000430 5.040005 -8.630 0.000 constant 1.038040 0.030342 34.21 0.000 q50 primary 0.052460 0.018224 2.880 0.004 secondary 0.227014 0.016367 13.87 0.000 tertiary 0.461729 0.017562 26.29 0.000 exp 0.020518 0.001243 16.51 0.000 expsqr -0.000450 3.490005 -12.90 0.000 constant 1.167785 0.016027 72.86 0.000 q75 primary 0.035789 0.022852 1.570 0.117 secondary 0.213596 0.021513 9.930 0.000 tertiary 0.479334 0.023509 20.39 0.000 exp 0.018359 0.002441 7.520 0.000 expsqr -0.000350 0.000078 -4.450 0.000 constant 1.342088 0.023643 56.76 0.000 q90 primary 0.038085 0.031410 1.210 0.225 secondary 0.246998 0.028517 8.660 0.000 tertiary 0.577881 0.031498 18.35 0.000 exp 0.021994 0.002351 9.360 0.000 expsqr -0.000370 7.030005 -5.290 0.000 constant 1.454727 0.028247 51.50 0.000 OLS primary 0.056244 0.020290 2.770 0.006 secondary 0.239790 0.019035 12.60 0.000 tertiary 0.497390 0.020172 24.66 0.000 exp 0.019650 0.001200 16.28 0.000 expsqr -0.000390 0.000030 -11.68 0.000 constant 1.154330 0.020340 56.73 0.000
125
Table 4.6: Quantile and OLS regression estimates (model 2-Sinhala)
Coefficient
Bootstrap standard error t-value
P>|t|
q10 primary 0.094427 0.0542914 1.740 0.082 secondary 0.269427 0.0546298 4.930 0.000 tertiary 0.494878 0.0555607 8.910 0.000 exp 0.021972 0.0027689 7.940 0.000 expsqr -0.000510 0.0000887 -5.780 0.000 constant 0.806024 0.0564095 14.29 0.000 q25 primary 0.081863 0.0317356 2.580 0.010 secondary 0.243767 0.0307653 7.920 0.000 tertiary 0.456219 0.0319011 14.30 0.000 exp 0.020349 0.0015455 13.17 0.000 expsqr -0.000470 0.0000512 -9.120 0.000 constant 1.009352 0.0309662 32.60 0.000 q50 primary 0.040893 0.0342083 1.200 0.232 secondary 0.212035 0.0342209 6.200 0.000 tertiary 0.434629 0.0364486 11.92 0.000 exp 0.022127 0.0013746 16.10 0.000 expsqr -0.000490 0.0000419 -11.58 0.000 constant 1.184314 0.0339347 34.90 0.000 q75 primary 0.042947 0.0294690 1.460 0.145 secondary 0.212639 0.0304455 6.980 0.000 tertiary 0.464373 0.0330742 14.04 0.000 exp 0.020530 0.0023610 8.700 0.000 expsqr -0.000410 0.0000796 -5.130 0.000 constant 1.339696 0.0329567 40.65 0.000 q90 primary 0.015132 0.0488720 0.310 0.757 secondary 0.226079 0.0497276 4.550 0.000 tertiary 0.542847 0.0529322 10.26 0.000 exp 0.023075 0.0021202 10.88 0.000 expsqr -0.000380 0.0000587 -6.550 0.000 constant 1.472891 0.0482806 30.51 0.000 OLS primary 0.061440 0.0264700 2.320 0.020 secondary 0.249500 0.0247800 10.07 0.000 tertiary 0.499660 0.0257500 19.40 0.000 exp 0.020880 0.0013850 15.07 0.000 expsqr -0.000410 0.0000300 -10.70 0.000 constant 1.147780 0.0262000 43.81 0.000
126
Table 4.7: Quantile and OLS regression estimates (model 2-Tamil)
Coefficient
BootstrapStandard error t-value
P>|t|
q10 primary 0.049911 0.064335 0.780 0.438 secondary 0.144278 0.069195 2.090 0.037 tertiary 0.435165 0.111695 3.900 0.000 exp 0.012022 0.004306 2.790 0.005 expsqr -0.000220 0.000104 -2.070 0.038 constant 0.880291 0.081788 10.76 0.000 q25 primary -0.024360 0.045348 -0.540 0.591 secondary 0.080193 0.049733 1.610 0.107 tertiary 0.365616 0.065839 5.550 0.000 exp 0.007349 0.003440 2.140 0.033 expsqr -0.000160 9.730005 -1.650 0.100 constant 1.135356 0.046479 24.43 0.000 q50 primary -0.003780 0.025997 -0.150 0.884 secondary 0.134977 0.029174 4.630 0.000 tertiary 0.452884 0.056329 8.040 0.000 exp 0.012599 0.002366 5.320 0.000 expsqr -0.000250 6.550005 -3.840 0.000 constant 1.206258 0.024628 48.98 0.000 q75 primary -0.005400 0.048577 -0.110 0.911 secondary 0.177777 0.048616 3.660 0.000 tertiary 0.538524 0.073098 7.370 0.000 exp 0.014200 0.004617 3.080 0.002 expsqr -0.000270 0.000127 -2.170 0.031 constant 1.328006 0.055418 23.96 0.000 q90 primary 0.006410 0.052708 0.120 0.903 secondary 0.186644 0.055268 3.380 0.001 tertiary 0.755078 0.124952 6.040 0.000 exp 0.008360 0.005918 1.410 0.158 expsqr -0.000110 0.000174 -0.610 0.544 constant 1.507489 0.056284 26.78 0.000 OLS primary 0.026551 0.033083 0.800 0.422 secondary 0.178914 0.032569 5.490 0.000 tertiary 0.508884 0.041317 12.32 0.000 exp 0.013254 0.002708 4.900 0.000 expsqr -0.000230 6.930005 -3.340 0.001 constant 1.162422 0.035884 32.39 0.000
127
Source: Harmon, Oosterbeek and Walker (2000) Figure 4.1 Returns to Schooling in Europe, Men and Women
128
0.00
0.02
0.04
0.06
ED
U
0 .2 .4 .6 .8 1Quantile
0.01
0.01
0.01
0.02
0.03
0.03
exp
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00
−0.
000.
000.
00ex
psqr
0 .2 .4 .6 .8 1Quantile
Figure 4.2 OLS and quantile regression estimates (model 1-full sample)
129
0.00
0.02
0.04
0.06
ED
U
0 .2 .4 .6 .8 1Quantile
0.01
0.01
0.01
0.02
0.03
0.03
exp
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00−
0.00−
0.00
0.00
0.00
exps
qr
0 .2 .4 .6 .8 1Quantile
Figure 4.3 OLS and quantile regression estimates (model 1-full Sinhala)
130
−0.
020.
000.
020.
040.
06E
DU
0 .2 .4 .6 .8 1Quantile
−0.
010.
000.
010.
020.
03ex
p
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00−
0.00
0.00
0.00
exps
qr
0 .2 .4 .6 .8 1Quantile
Figure 4.4 OLS and quantile regression estimates (model 1-full Tamil)
131
−0.
050.
000.
050.
100.
15pr
imar
y
0 .2 .4 .6 .8 1Quantile
0.10
0.15
0.20
0.25
0.30
0.35
seco
ndar
y
0 .2 .4 .6 .8 1Quantile
0.30
0.40
0.50
0.60
0.70
tert
iary
0 .2 .4 .6 .8 1Quantile
0.01
0.01
0.02
0.03
0.03
exp
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00
−0.
00−
0.00
exps
qr
0 .2 .4 .6 .8 1Quantile
Figure 4.5 OLS and quantile regression estimates (model 2-full sample)
132
−0.
100.
000.
100.
20pr
imar
y
0 .2 .4 .6 .8 1Quantile
0.15
0.20
0.25
0.30
0.35
0.40
seco
ndar
y
0 .2 .4 .6 .8 1Quantile
0.30
0.40
0.50
0.60
0.70
tert
iary
0 .2 .4 .6 .8 1Quantile
0.01
0.02
0.03
0.04
exp
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00
−0.
00−
0.00
−0.
000.
00ex
psqr
0 .2 .4 .6 .8 1Quantile
Figure 4.6 OLS and quantile regression estimates (model 2–Sinhala)
133
−0.
20−
0.10
0.00
0.10
0.20
0.30
prim
ary
0 .2 .4 .6 .8 1Quantile
−0.
100.
000.
100.
200.
300.
40se
cond
ary
0 .2 .4 .6 .8 1Quantile
0.00
0.20
0.40
0.60
0.80
1.00
tert
iary
0 .2 .4 .6 .8 1Quantile
0.00
0.01
0.02
0.03
0.04
exp
0 .2 .4 .6 .8 1Quantile
−0.
00−
0.00
−0.
00−
0.00
0.00
0.00
exps
qr
0 .2 .4 .6 .8 1Quantile
Figure 4.7 OLS and quantile regression estimates (model 2–Tamil)
134
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,
135
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.
136
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 Côte 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.
137
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 Indicator- based 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
138
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.
139
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. O’Boyle, 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.425- 428. 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.
140
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 India’s 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: North- Holland 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.
141
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. Székely, 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.569- 580. 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.