95. Tinh Doan Khoa KTCT 2014

95. Tinh Doan Khoa KTCT 2014 tài liệu, giáo án, bài giảng , luận văn, luận án, đồ án, bài tập lớn về tất cả các lĩnh vực...

Journal of Southeast Asian Economies Vol 31, No 1 (2014), pp 103–19 ISSN 2339-5095 print / ISSN 2339-5206 electronic

DOI: 10.1355/ae31-1g

OTHER ARTICLES

Impact of Household Credit on Education and Healthcare Spending

by the Poor in Peri-Urban Areas,

Vietnam Tinh Doan, John Gibson and Mark Holmes

There is an ongoing debate about whether microfinance has a positive impact on education and health for borrowing households in developing countries To understand this debate,

we use a survey designed to meet the conditions for propensity score matching (PSM) and examine the impact of household credit on education and healthcare spending by the poor

in peri-urban areas of Ho Chi Minh City, Vietnam In addition to matching statistically identical non-borrowers to borrowers, our estimates also control for household pre-treatment income and assets, which may be associated with unobservable factors affecting both credit participation and the outcomes of interest The PSM estimates show a significant and positive impact of borrowing on education and healthcare spending However, further investigation of the effects of the treatment reveals that only formal credit has a significant and positive impact

on education and healthcare spending, while informal credit has an insignificant impact on spending This paper contributes to the limited literature on peri-urban areas using evidence from one of the largest and most dynamic cities in Southeast Asia.

Keywords: Matching, education and healthcare spending, household credit, the poor, peri-urban.

I Introduction

Microfinance has increasingly attracted attention

from the global development community because

it is considered a powerful tool for alleviating

poverty in developing countries An argument commonly made in support of microfinance is that

it may help keep household production stable and mitigate adverse shocks, thus preventing school

(Armendariz and Morduch 2010; Dehejia and

Gatti 2002; Edmonds, 2006; Jacoby and Skoufias

1997; Maldonado and Gonzalez-Vega 2008)

Education and health are critical to sustainable

poverty reduction since they affect the formation

of quality human capital and the productivity of

future generations

There is ongoing debate about the impact of

microfinance (Cull, Kunt and Morduch 2009) on

borrowing households’ access to education and

healthcare On the one hand, microcredit has

a positive impact on education; girls received

more schooling if households borrowed from the

Grameen Bank (Pitt and Khandker 1998) On

the other hand, some studies find no effects or

adverse effects on children’s education (Hazarika

and Sarangi 2008; Islam and Choe 2009; Morduch

1998) Likewise, in terms of health, Pitt et al

(2003) find higher weight-for-age and

height-for-age levels amongst children of Grameen Bank

borrowers, but Coleman (2006) observes a negative

impact of microcredit on healthcare spending by

households in northeast Thailand These differing

outcomes are discussed in greater detail in the

literature review (section II)

One difficulty in evaluating the impact of

microcredit is that borrowing and non-borrowing

households typically differ in both observable

and unobservable characteristics Borrowers may

self-select into borrowing activities due to their

“better” characteristics, thus making it challenging

to form a counterfactual of what would have

happened to borrowers in the absence of credit

If studies fail to correct for this problem of self-selection, the estimates will be nạve and overstated

(Coleman 2006) The propensity score matching

(PSM) method may help avoid this problem In

this approach, the effects of the treatment (i.e.,

borrowing/credit participation) are estimated

by simulating a randomized experiment with a

treatment and control group Households in the

treatment group are matched, based on observable

characteristics/factors, to other similar households

that will then form the control group It is assumed

that the matched households in the control group

would have no systematic differences in response to

the treatment, thus offering a valid counterfactual

When used appropriately, PSM can replicate the advantages of a randomized experiment (Dehejia and Wahba 2002)

In this paper, a survey, designed by the authors

to meet the conditions under which PSM works well, is used to examine the impact of household credit on education and healthcare spending by the poor in the peri-urban areas of Ho Chi Minh City (HCMC), Vietnam In addition to matching statistically identical non-borrowers to borrowers, our estimates also control for pre-treatment household income and assets These pre-treatment variables may be associated with unobservable factors affecting both credit participation and the outcomes of interest; therefore the inclusion of these variables helps reduce bias

Apart from the use of PSM, four other important features of the current analysis warrant comment

First, the analysis takes both formal and informal credit into consideration Most studies have examined the impact of formal or programme credit but have not considered the effects of credit from other sources (Coleman 2006; Khandker 2005; Morduch 1998; Pitt and Khandker 1998)

Our survey contributes to the exiting literature by capturing all sources of credit; the results reported below compare the effects of formal and informal credit (provided by relatives, friends, neighbours and informal moneylenders) Policy-makers often influence access to formal credit but have less leverage over informal credit; hence distinguishing their separate impact is of interest

Secondly, we provide evidence from a newly industrializing peri-urban area (District 9) in one of the largest and most dynamic Southeast Asian cities Our findings may therefore have external relevance for peri-urban areas in the rest

of Southeast Asia Peri-urban areas are defined

as places where industrialization has led to rapid urbanization and in-migration Rapid urbanization displaces poverty from rural to urban areas, resulting in the rapid urbanization of poverty In Vietnam’s urban areas, poverty levels have risen from 5.99 per cent in 1998 to 6.63 per cent in

2008 due to rapid population growth in these areas over the same period (VDR 2010)

Thirdly, most non-experimental research designs

found in the existing literature on microcredit and

human capital lack valid counterfactuals of the

outcome for borrowers if they had not borrowed

Further, they typically make comparisons between

borrowers and non-borrowers without adopting

a plausible mechanism for dealing with

self-selection

Finally, although a comparison of the outcomes

for borrowers and non-borrowers using PSM is at

risk of being unreliable if the groups that are being

compared are too different, our empirical approach

minimizes the differences between the two groups

in order to enhance the validity of the PSM

approach (Dehejia and Wahba 2002) Therefore,

this paper also provides recommendations for

overcoming the potential weakness of the PSM

method

The rest of this paper is organized as follows:

Section II reviews the literature on household

credit and its impact on education and healthcare;

section III discusses the methodology employed;

empirical results are reported in section IV;

followed by concluding remarks in section V

II Literature Review

As was mentioned above, credit may affect

household demand for education and healthcare

in two ways (Armendariz and Morduch 2010)

Microcredit may help households earn higher

incomes, in turn, raising consumption and

increasing demand for healthcare and children’s

education In a contrasting example, if access to

credit raises female economic activity, children

may be taken out of school to replace maternal

input in the care of younger siblings or to work in

expanded household businesses

The evidence on these opposing effects of credit

participation is mixed Dehejia and Gatti (2002),

Edmonds (2006) and Jacoby and Skoufias (1997)

find that inadequate schooling in poor countries

can often be attributed to the lack of access to

credit — households facing adverse shocks may

pull children out of schools to reduce expenditure

and increase household income by increasing

children’s working hours Yet, in the same vein,

Hazarika and Sarangi (2008) find that borrowing households may take children out of school to work in family businesses Because small loans are often associated with higher interest rates and short-term repayment conditions, poor borrowers may reduce costs by using their own labour, which may include child labour, in order to accrue high returns for the repayment of loans and interest For example, a study on Vietnam finds that households that borrowed from lenders with higher interest rates used more child labour (Beegle, Dehejia and Gatti 2004)

There is also some interplay between healthcare and education If borrowing enables parents to promptly provide medication when children fall sick; children may recover faster and can stay in school Healthier children not only perform better

in school but are more likely to stay in school for longer such that they grow up to be more productive adults In contrast, low school achievement and attendance rates are associated with malnutrition (Glewwe, Jacoby and King 2000) It has been observed that microfinance clients consume healthcare services such as pasteurization, health insurance, family planning and pregnant-mother care more often than non-clients (CGAP 2003)

III Analytical Framework

III.1 Data

Four hundred and eleven borrowing and non-borrowing households were interviewed in early 2008 in the peri-urban District 9 (HCMC, Vietnam).1 Since our focus is on the impact of microcredit on poor households, our sample was selected from a list of households with an initial per capita income below the HCMC general poverty line of VND6 million (approximately US$1 per day).2 The sampling was carried out in two steps:

we first selected wards within District 9 and then selected households from those wards The target sample size was set at 500 households, including

100 reserves, to achieve a realized sample of

400 In fact, 411 households were successfully interviewed, accounting for 26 per cent of the total number of poor households in each of the selected

wards in the district The sample comprises 304

borrowing households and 107 non-borrowing

households, with a total of 2,062 members: 955

(46.3 per cent) males and 1,102 (53.7 per cent)

females More information on the sources of credit

and the households’ per capita income prior to the

treatment (borrowing) can be found in Appendix 1

The survey was designed to collect data on

household-level and individual-level variables:

commune characteristics, household durable

and fixed assets, borrowing and expenditure on

healthcare, food and non-food items, and children’s

schooling and education We also utilized GPS

receivers to collect data on the location of the

households and to measure their distances from

facilities, such as schools

III.2 Challenges to Evaluating Impact

The main challenge to evaluating the impact of

credit is the difficulty of separating the causal

effect of credit from selection and reverse causation

biases, which are common to nearly all statistical

evaluations (Armendariz and Morduch 2010)

In order to separate the effects of the treatment

from other factors, we have to ask how borrowers

would have managed without credit (Armendariz

and Morduch 2010) This question is not easy to

answer because researchers are unable to observe

the virtual outcomes needed to construct such a

counterfactual As such, estimating the impact

of credit participation requires measuring the

difference in the outcome between the treatment

and control groups:

E(Y|D=1) – E(Y|D=0) Where, Y is the outcome and D is the treatment

taking a value of 1 if the treatment has been

received, and 0 if it has not The difference in

the outcome, however, may result from either

the treatment (credit participation), differences

in observable characteristics or differences in

unobservable characteristics Estimates will be

biased if one does not control for the differences

in observable and unobservable characteristics

Differences in observable characteristics result

in an “overt bias”, which can be removed by controlling for observables (Xi) in estimation models (Lee 2005) Thus, the estimation is now formulated as follows:

E(Y| D=1, Xi) – E(Y|D=0, Xi) The estimated impact may also include a “hidden bias” resulting from unobservable characteristics

Models using a randomized selection of treatment and control groups are helpful in this regard, as the randomization allows us to cancel out the differences in both observable and unobservable characteristics However, it is very hard to conduct a randomized test evaluating the impact

of credit due to motivation and contamination problems (Mosley 1997) Hence, there are usually some problems with using non-experimental data due to the non-random implementation of credit programmes and self-selection into credit participation by borrowers Estimates of causality may include a selection bias if credit participation

is correlated with unobserved characteristics that also affect the outcome For instance, households that are better motivated to invest in children’s schooling may have a greater demand for credit

Without an adequate measure of motivation, this hidden factor may make an observable correlation between credit and schooling seem like a causal effect has taken place

In the case of our sample, non-randomized credit participation is not a crucial concern because all the surveyed households have a per capita income of less than VND6 million This means that they are all eligible for preferred credit (i.e., subsidized interest and easy conditions) from government funds Selection by informal lenders and self-selection into credit borrowing due to unobservable factors, however, may still occur

If data on pre-treatment variables of interest are available, researchers may examine differences in these variables in order to determine whether there

is a positive or negative selection on unobserved characteristics (conditional on the observed characteristics) For instance, if YT

0 and YC

0 are the respective outcomes for the treatment and control groups at time 0 (before the treatment),

and if E(YT | D=1, Xi) ≠ E(YC

0 | D=0, Xi) is the result after controlling for observable factors, one

should suspect that unobservable confounders

are affecting the treatment and outcomes That

is to say that a “hidden bias” has resulted from

unobservable confounders Lee (2005, p 125)

suggests that controlling for Y0 (together with Xi

on the right-hand side) may to some extent reduce

the hidden bias For the purposes of our study,

pre-treatment data on the variables of interest is not

available but we can use (baseline) pre-treatment

income per capita as a control variable (Mosley

1997; Heckman and Smith 1999):

Yij,t-1 = a + b.Dij,t + I.Xij,t + eij,t-1 (1) Where, Yij is the outcome of interest of household

i in ward j; D is a dummy variable representing

(1) if a household borrows and (0) if it does

not; X is a set of unchanged (or little changed)

control variables over time, such as household

characteristics The coefficient b shows whether

borrowers had a higher or lower income per

capita than non-borrowers prior to participating in

borrowing activities (conditional on their observed

characteristics) If b is positive, that means a

positive selection on unobserved attributes exists:

borrowers tend to be richer than non-borrowers,

which will lead the non-experimental estimators

to overstate the impact of credit participation

We ran a regression of equation (1) and found a

significantly positive b coefficient

III.3 Emprical Method

The propensity-matching method is the most

suitable candidate for cross sectional

non-experimental datasets without good instrument

variables This method forms a control group of

non-participants with observed characteristics that

are similar to participants (the treatment group)

(Dehejia and Wahba 1999, 2002) The main

advantage of the matching method is that one can

draw on existing data sources and so it is quicker

and cheaper to implement Nevertheless, matching

does not control for unobservable characteristics

that may cause a selection bias, and as a result, the

reliability of estimates is reduced (Smith and Todd 2005) The most widely used matching method is propensity score matching (PSM) Other methods

of matching each X (covariate matching) create a problem of high dimensionality, which requires large datasets

The PSM method first estimates the propensity score for each participant and non-participant on the basis of observed characteristics It then compares the mean outcome for participants with the outcome for the matched (similar in terms of scores) non-participants In other words, the purpose of the PSM is to first select comparable non-borrowing households among all non-borrowing households

to generate a control group, and then compare the outcomes for the treatment and matched control groups The crucial assumption is that the outcomes for non-borrowers in the matched control group represent what the borrowers would have experienced without credit participation This

is referred to as unconfoundedness or a conditional independence assumption (CIA) (Rosenbaum and Rubin 1983) In summary, this is the underlying point of propensity score matching: control and treatment units with the same propensity score have the same probability of assignment to the treatment as in randomized experiments (Dehejia and Wahba 1999)

The PSM method may produce estimates with low bias if datasets satisfy three conditions (Dehejia and Wahba 2002): (i) the data for treatment and control groups are collected using the same questionnaire; (ii) the treatment and control groups are drawn from the same locality;

and (iii) the dataset contains a rich set of variables relevant to modelling credit participation and its outcomes Since all households surveyed in this study were poor prior to credit participation, the PSM method should produce less biased estimates than it would for a general sample of households with highly divergent per capita income rates

While the PSM method also allows controlling for potential biases such as non-placement and self-selection (Dehejia 2005; Dehejia and Wahba 2002), it fails to control for unobservable characteristics which may create hidden biases because the scores are calculated on the basis of

observed characteristics only (Dias, Ichimura and

Berg 2007) As was mentioned above, observable

characteristics may not fully capture individual

motivation, ability and skills — all of which may

affect the treatment participation Ultimately,

the success of the PSM method depends on how

close the control and treatment group are in terms

of space and time Further, the two groups should

have as little baseline differences as possible

(Lee 2005)

IV Empirical Results

The PSM estimates of the impact on education

and healthcare expenditure are presented in

subsection IV.1 while subsection IV.2 looks at the

impact of formal and informal credit on household

expenditure

IV.1 PSM Estimation

Kernel (with the default bandwidth of 0.06)

and radius matching (with the default radius

of 0.1) results of the impact of microcredit on

education and healthcare spending are discussed

in this subsection.3 Imbens (2004), Lee (2005) and

Rosenbaum and Rubin (1983) note that sets of

controlling covariates should meet the conditions

of matching controlling variables In this paper,

the use of covariates in the score estimation stage

follows discussions in Lee (2005) and Rosenbaum

and Rubin (1983), and Caliendo and Kopienig

(2008) and Bryson, Dorsett and Purdon (2002)

In some cases, interaction terms were also used to

balance the estimated propensity scores

Impact on Education Expenditure Our base

specifications (S1 and S3 in Table 1) use a set of

covariates of household characteristics such as:

the head of the household’s gender, age, education

and marital status; school-aged child ratio, the

number of children; and ward dummies to estimate

the scores Although we do not have panel data

to apply the difference-in-difference matching

estimator (believed to be considerably better than

cross-sectional matching estimators), our attempts

at reducing bias associated with unobservable

characters by including pre-treatment household income and assets may help offset any disadvantage (Imbens and Wooldridge 2009; Mosley 1997) The effects that occur when pre-treatment income and assets are included in the matching are reported in the second (S2) and fourth rows (S4) of Table 1

The changes in the model specifications between

S1 and S3 and between S2 and S4, are to test for the sensitivity of the effect

Figure 1 displays the kernel densities of the propensity scores when pre-treatment income and assets are included alongside the other controlling variables (S4 in Table 1) The propensity scores range from 0.418 to 0.943 and from 0.174 to 0.940 for borrowers and non-borrowers, respectively,4 but the mean scores are not much different (0.761 and 0.675 for borrower and non-borrower groups)

The figure also illustrates a substantial overlap

in the distributions The following estimation of the average effect of the treatment is restricted

to the area of common support, where the two distributions overlap Thus, some non-borrowers who are dissimilar to borrowers are not used in the comparison

The estimates of the average effect of the treatment, on the treated (ATT), are reported in Table 1 There is little difference in the results

of the two matching approaches used When matching just household characteristics and location dummies (S1 and S3), the effect of credit participation on education spending is observed to

be statistically significant at the 1 per cent level

After including the pre-treatment income and assets (S2 and S4) the estimated impact of credit participation declines, but is still significant at the

5 per cent level

According to these PSM estimates, the borrowers, on average, spent about VND81,000

to VND99,000 more on education per month than comparable non-borrowers

Impact on Healthcare Expenditure Figure 2

displays the kernel densities of the propensity scores estimated for evaluating the impact of credit participation on healthcare expenditure.5 The scores are from when the pre-treatment income and assets are included alongside the

FIGURE 1 Propensity of Scores for Borrowers and Non-borrowers to Estimate the Effects of

the Treatment on the Treated for Education Expenditure

N ote : The propensity scores of control units outside the area of common support are cut off.

TABLE 1 The Average Effect of the Treatment on Monthly Average Education Expenditure in VND1,000

Using Matching Estimators for the Entire Sample

Control variables in the propensity score estimation Kernel matching Radius matching

Head’s gender, head’s age, head’s education, marital status, 92.696 (31.967)** 98.696 (32.393)**

school-aged child ratio, and ward dummies (S1)

S2=S1plus initial income in log, initial assets in logarithm 85.020 (34.027)* 93.022 (31.506)**

Head’s gender, head’s age, head’s education, marital status, 87.447 (33.875)** 93.179 (34.182)**

number of children from 6 to 18, and ward dummies (S3)

S4=S3 plus initial income in log, initial assets in logarithm 81.232 (34.621)* 86.861 (34.448)*

N otes : Bootstrapped standard errors in parentheses with 1,000 repetitions, statistically significant at 10 per cent (+);

5 per cent (*); and 1 per cent (**) Only a few households (10 households) have 4 or more children aged 6 to 18 years

old To simplify the process, we grouped them with households having 4 kids S i are model specifications.

1 2 3 4

Predicted Probability

matches (S4 in Table 2) The propensity scores

range from 0.348 to 0.989 for borrowers and from

0.195 to 0.962 for non-borrowers Once again,

the estimation of the average treatment effect is

restricted to the area of common support where

the distributions overlap

The estimates of the impact of credit participation

on healthcare expenditure are reported in Table 2

The results show that credit participation positively

affects healthcare expenditure They are also

statistically significant regardless of the sets of

covariates and the types of matching approaches

used Borrowers spent about at least VND93,000

more on healthcare than their non-borrowing

counterparts

IV.2 Impact of Formal and Informal Household Credit

In this subsection, the effects of multiple treatments are estimated to contrast the impact

of informal and formal credit on education and healthcare expenditure (Lee 2005) When applying these multiple treatments, we treat credit from formal sources (F) as a full “dose” and credit from informal sources (I) as a partial “dose”.6 Here, we directly compare the formal and informal credit groups, taking the informal credit group as the control group

Estimates of the effects of the multiple treatments

on education expenditure are reported in Table 3

The estimation procedure used here is similar to

FIGURE 2 Propensity of Scores for Borrowers and Non-borrowers to Estimate ATT for Healthcare Expenditure

N ote : The propensity scores of control units outside the area of common support are cut off.

–4

2 1

3 4

Predicted Probability

TABLE 2 Average Effect of the Treatment on Monthly Average Healthcare Expenditure in

VND1,000 Using Matching Estimators

Control variables in the propensity score estimation Kernel matching Radius matching

S2=S1 plus initial income in log, initial assets in logarithm 93.082 94.016

S4=S3plus initial income in logarithm, initial assets in log 108.313 112.895

N otes : Bootstrapped standard errors in parentheses with 1000 repetitions, statistically significant at 10 per cent (+);

5 per cent (*); and 1 per cent (**).

S 1 : Head’s gender, head’s age, head’s education, marital status, household size in log, head’s age*gender, ward

dummies.

S 3 : Head’s gender, head’s education, marital status, dummy of child below 6, number of children from 6 to 18 years

old, persons from 18 to 60 years old, dummy of person older than 60 years old, head’s age*education, and ward

dummies.

the one discussed in subsection IV.1: household

characteristics are first used to construct the scores

(S1 and S3) and thereafter, pre-treatment income

and assets are controlled for (S2 and S4) The

estimated effects of informal credit are reported in

columns 2 and 3, and the effects of formal credit

effect are reported in columns 4 and 5

The estimates show that informal credit has

no significant effect on household education

expenditure In contrast, formal credit strongly

affects education expenditure Both kernel

and radius matching estimators display similar

estimates that are statistically significant at the

1 per cent level

To test the results further, we directly compared

the impact of formal credit to informal credit

Estimates of the difference between the impact of

formal and informal credit are shown in the last

column of Table 3 The estimates are consistent

across the specifications of the matching variables

The higher credit level (or treatment level) leads to

a greater positive impact

Likewise, we looked at the impact of formal and informal credit on healthcare spending The estimates for informal credit and formal credit

on health care expenditure are reported in Table

4 The results of the difference in the impact of formal and informal credit are presented in the last column of Table 4 Informal credit has a positive but only marginally significant impact at the 10 per cent level of statistical significance In contrast, the impact of formal credit is more than double that of informal credit (statistically significant at the 10 per cent level)

The use of multiple ordered treatments shows that the higher treatment level (formal credit) has a greater positive impact on healthcare and education expenditure, thus confirming the positive effects of credit on healthcare and education

TABLE 3 The Average Effect of Treatment on Monthly Average Education Expenditure in

VND1,000 using Matching Estimators for the Entire Sample

N otes : Bootstrapped standard errors in parentheses with 1,000 replications, statistically significant at 10 per cent (+);

5 per cent (*); 1 per cent (**).

S 1 : Head’s gender, head’s age, head’s education, marital status, ward dummies, school-aged child ratio, and head’s

age*head’s gender.

S 2 : Head’s gender, head’s age, head’s education, marital status, ward dummies, school-aged child ratio, head’s

age*head’s education, initial income in logarithm, initial assets in logarithm.

S 3 : Head’s gender, head’s age, head’s education, marital status, ward dummies, number of children aged 6 to 18 years

old, and head’s age*head’s gender.

S 4 : Head’s gender, head’s age, head’s education, marital status, ward dummies, number of children aged 6 to 18 years

old, head’s age*education, initial income in logarithm, initial assets in logarithm In some cases to meet the balancing

property in estimation of propensity scores we added interaction terms in specifications of the propensity score.

ATTK: Average Treatment Effect on the Treated, Kernel matching.

ATTR: Average Treatment Effect on the Treated, Radius matching.

V Finding Summary and Concluding

Remarks

This paper estimates the impact of credit

participation on education and healthcare

expenditure by the poor in the peri-urban District

9 of Ho Chi Minh City, Vietnam, using data from

a survey designed to meet the conditions for

propensity score matching.7 The results illustrate

that borrowers spent more on education and

healthcare than their non-borrowing counterparts, thus credit participation has a highly positive and significant effect on healthcare and education spending

We focussed on the poor in order to ensure that disparities between the treatment and control groups would be minimal We also controlled for pre-treatment income levels, which, in turn, control for unobservable attributes such as motivation,