Peer-to-peer lending and birth outcomes during national ...thew.web.unc.edu/files/2013/08/acquah.pdfPeer-to-peer lending and birth outcomes during national economic crises: Lessons

Page 1

Peer-to-peer lending and birth outcomes during national economic

crises: Lessons from Indonesia.

Joseph Kofi Acquah ∗

November 1, 2015

∗I am deeply indebted to Anna Aizer, Andrew Foster, Glenn Loury, Frank Sloan, Emily Oster, JesseShapiro and seminar participants at the Brown University economics department for their useful commentsand suggestions. Address: Department of Economics, Brown University, 64 Waterman St., Providence, RI02912. Contact: joseph [email protected]

1

Page 2

2

Abstract

National economic crises increase the incidence of low birth weight rates due to credit

constraints that prevent households from effectively smoothing their food consumption. In non-

crises years, evidence shows that peer to peer banking (P2P) schemes extend credit to individuals

that face idiosyncratic income shocks. However, during crises years, there is limited evidence

on the credit extension capabilities of P2P schemes. In this study, I explore the credit extension

capabilities of rotating savings and credit institutions (ROSCAs) during the 1998 Indonesian

financial crises. This study argues that heterogeneity in income shocks and differences in the

marginal utility of consumption, facilitate transfers within P2P schemes during crises. Results

show that in the midst of the 1998 Indonesian financial crisis, ROSCAs modified their structures

to establish and facilitate borrowing within the scheme to members facing economic hardships.

Consequently, mothers that participated in ROSCAs during the crises were able to maintain food

consumption and observed 8.1 percentage points fewer low birth weight babies. Findings suggest

that encouraging mothers to actively participate in peer-to-peer banking schemes in non-crisis

years may provide additional lines of credit that protect food consumption and consequently

child health during crises years.

Page 3

3

1 Introduction

As national economic crises become more frequent, understanding the capacity of institu-

tions to protect the most vulnerable members of society is a priority for policy makers. For

low-income households, national economic crises which result in large negative deviations

in GDP per capita reduce real household income and consequently food consumption (Al-

derman, 2011). Pregnant women in low-income households are especially vulnerable during

crises, since reductions in food consumption increase the probability of adverse birth out-

comes such as low birth weight. Estimates of increases in low birth weight during economic

crises, range from 1% (Argentina’s 2000 to 2002 economic crisis, (Cruces et al., 2010)) to

33% (Tanzania’s 2008 economic crisis, (Burlando, 2010)) . These statistics demonstrate that

national economic crises have a particularly strong impact on pregnant women.

Increases in low birth weight have long term implications on human capital accumulation

and long term economic growth (Bharadwaj et al., 2010). Using data from Norway’s birth

registry, Black and coauthors demonstrate that increasing birth weight by 10% increases the

probability of high school completion by 0.9% and later-life wages by 1% (Black et al., 2005).

Berhman confirms these findings using data from the Minnesota’s twins registry and shows

that reducing the birth weight gap among developed and developing countries can increase

per worker GDP in developing countries by as much as 9% (Behrman and Rosenzweig, 2004).

Thus, there are clear economic incentives for nations to devote resources to curtailing adverse

birth incomes during national economic crises.

Access to credit markets enable low-income households and pregnant women to smooth

their food consumption during crises (Skoufias, 2003). Unfortunately, among pregnant women

and the poor, collateral requirements at formal banks often exclude the most vulnerable mem-

bers within these groups from borrowing money during crises (Hoogeveen, 2002). Peer to peer

informal banks (P2P), which are increasingly common in the developing world, offer credit

by replacing collateral requirements with social sanctions as a means of enforcing repayment

(Bouman, 1995, 1977). These schemes offer an alternative for low-income households and

Page 4

4

especially pregnant women to secure credit to smooth food consumption during economic

crises.

Economic models of ROSCA institutions suggests that P2P banking schemes work well

when income shocks are idiosyncratic but less so when income shocks are partially correlated

(Townsend, 1995; Dercon, 2002; Hoogeveen, 2002; Fang and Ke, 2006; Klonner, 2000). In

the case of a homogeneous population with identical preferences, the situation logic is fairly

clear: with uncorrelated income shocks, the supply for credit should outpace demand, and

thus borrowers are able to find peers willing to lend their savings at a reasonable interest rate.

On the other hand, if income shocks are perfectly correlated and aggregate in an economy

with homogeneous agents, lenders will be scarce since all agents in the economy will want to

borrow at the same interest rate. Townsend suggests that these assumptions are too strong

even when one considers a national economic crises (Townsend, 1995). In particular, since

individuals obtain their income from dissimilar sources, income shocks are still idiosyncratic

during national economic crises. In addition, even if income shocks were aggregate differences

in preferences suggest that at a reasonable price (ie interest rates) transfers may occur within

P2P schemes from agents with a low marginal utility of consumption to agents with a high

marginal utility of consumption.

This study provides empirical evidence on the capabilities of P2P banking schemes as

alternative sources of credit during times of national economic crisis, especially for vulnerable

populations, including pregnant women. The relevance of this study is that it is the first

paper to provide empirical evidence on (1), the credit extension capabilities of P2P banking

schemes during national economic crises and (2), the role that P2P banking schemes may

play in curtailing the link between economic crises and adverse birth outcomes.

The 1998 Indonesian financial crisis has features that make it desirable to explore the link

between ROSCA participation and the mitigation of adverse birth outcomes during national

economic crises. First, the sudden and unexpected nature of the 1998 Indonesian financial

crisis and the fact that all provinces in Indonesia were simultaneously impacted means that

Page 5

5

it was unlikely that women made the decision to have a child or participate in a ROSCA or

immigrate in anticipation of the coming crisis (Iriana and Sjoholm, 2002; Levinsohn et al.,

2003; Sharma, 2001). Second, as in most national economic crises, income shocks differed

by economic sector and at different quantiles of the income distribution (Smith et al., 2002;

Wie, 2000)1 . Thus within a ROSCA, income shocks could be idiosyncratic and thus credit

extensions may have been viable. Third, the Indonesian family life survey (IFLS) was imple-

mented a few months before the onset of the crisis with two follow up surveys post crisis. The

panel feature and timing of the surveys allows this study to address endogenous sorting into

ROSCA by using a novel econometric technique (symmetric weighting difference in difference

method2) to control for differences in observables and time invariant unobservables between

ROSCA and non-ROSCA participants (Sylvain, 2014, 2015). The advantage of using this

method is that it allows for a Roy model where rational actors make optimizing decisions

on ROSCA participation (unobservables that may influence ROSCA participation that are

fixed over time are controlled for). Results from the 1998 Indonesian crisis demonstrates the

link between ROSCA participation and reductions in adverse birth outcomes for vulnerable

pregnant mothers. Study findings show that during the crisis, low birth weight increased by

4.2 percentage points with ROSCA participants experiencing 8.1 percentage points fewer low

birth weight births as compared to non-ROSCA participants. This study demonstrates that

the decrease in incidence of low birth weights for ROSCA participants is a result of ROSCAs

adapting to the crisis and facilitating borrowing within the schemes. Finally, this study

shows that loans obtained from ROSCAs were primarily used to stabilize food consumption

for low-income pregnant mothers, improving birth weights.

The organization of this paper is as follows: Section 2 develops a conceptual model which

predicts that individuals facing hardships and individuals with a high marginal utility of

consumption during crisis should be more likely to receive credit from their peers within

1The decline in wages for example for women from 1997 to 1998 at the 30th, 50th, and 90th quantile whencompared to the 10th quantile were 12.78%, 29.44% and 6.39% respectively (Smith et al., 2002)

2Analysis uses IPTW weights instead of 1 to 1 matching to minimize difference in observables and topreserve sample size.

Page 6

6

a ROSCA. Section 3 reviews data and sample selection criteria. Section 4 discusses the

identification strategy in this study while section 5 examines background information and

provides sample summary statistics. Section 6 and 7 presents results and robustness checks

respectively, and Section 8 concludes the study.

2 Conceptual Framework

2.1 ROSCA loans in Indonesia

In Indonesia, bidding ROSCAs called the “Arisan call” were first documented in the

1980s (Williams and Johnston, 1983; Prabowo, 1989). The model in this paper focuses on

bidding within a ROSCA institution as a mechanism to secure funds within the organiza-

tion. However, in Indonesia, there are two additional ways in which one can borrow from

a ROSCA. In an Arisan, one member is typically assigned to be the head of the ROSCA.

The head of the group is in charge of ensuring that members pay on time and settles any

disputes within the group. To compensate the head of the group for his Arisan duties, the

head typically receives the common fund first and may initiate short term loans with this

fund throughout the life of the ROSCA (typically to prevent members from defaulting on

their Arisan payments (Bouman, 1995)). In addition, different Arisan variants have been

documented in Indonesia, including Arisans where a portion of the members’ contributions

are used to create an emergency loan fund from which members can borrow (Hospes, 1992).

Regardless of how members borrow from a ROSCA, differences in the elasticity of marginal

utility of consumption (MU), as indicated in the model below, should increase the interest

rate that borrowers are willing to accept on a loan. In negotiating over interest rates on

loans, individuals with higher elasticities of MU should bid higher and all else equal have a

higher probability of securing a loan. The model provided in this paper focuses on bidding

within a ROSCA simply to illustrate this characteristic.

Page 7

7

2.2 The model

This section provides a simple conceptual framework to model the operation of a bidding

ROSCA. In doing so, it provides a structure that explains the insurance that ROSCA schemes

provide to pregnant women during economic crises. The model modifies the Fang and Ke

model of ROSCA institutions to allow for differences in the marginal utilities of consumption

across agents. (Fang and Ke, 2006). It assumes that social sanctions such as exclusion from

one’a social network, exclusion from future loans or benefits, threats of evil spells, and guilt

are used to ensure repayment as discussed in Hoogevean (Hoogeveen, 2002).

Here, there are two risk averse agents who live for two periods. In each period, an

individual receives risk free income y and income shock ε, where ε is independent across

agents and periods (idiosyncratic) and uniformly distributed with mean µ < 0. The mean

of ε is assumed to be negative since income shocks are generally negative during economic

crises. Each agent participates in a ROSCA institution with a fixed contribution of m in each

period. This implies that in a two-person ROSCA, the total common fund to be disbursed per

meeting is 2m. At each meeting, the ROSCA allocates the common fund through a bidding

process. In period 1, each agent observes her risk-free income and income shocks which are

private information. After observing risk-free total income in period 1, agents submit their

bids. The bid bi , is the amount that agent i is willing to pay to the other agent in order

to obtain the common fund in period 1. The person with the highest bid is allocated the

common fund but is no longer eligible to bid in the subsequent period. The agents make

consumption decisions in period 1 after the bidding process. We solve the model below in the

presence of credit markets that allow individuals to save and borrow from the formal market

system at interest rate r. Income shocks are assumed to occur with probability 1- p (p ≥ 0)

in the economy . The utility function of the two risk-averse agents are

Page 8

8

u1(c) =1− exp(−λ1c)

λ1

(1)

u2(c) =1− exp(−λ2c)

λ2

(2)

2.2.1 ROSCA participation: The agents problem

As in Fang and Ke (2006), consider the indirect utility function when an agent wins a bid

or loses a bid in period 1. The agents problem when she wins a bid is given by

V1w ≡ maxc11,a11

u(c11) + δE(u(c12) (3)

a.t. c11 + s11 = (y − ε1)−m+ 2m− b1 (4)

c12 = y12 + (1 + r)s11 −m (5)

where cij is consumption of agent i in period j, si1 is the saving or debt of agent i in

period 1 , yij is the income of agent i in period j, m is the fixed ROSCA contribution, bi is

the bid from agent i, and r is the interest on savings deposit or loan received from formal

credit institutions. Y12 is defined as the sum of risk-free income (y) and idiosyncratic shocks

(ε) for agent 1 in period 2. Solving the agents problem produces the indirect utility function

V1w =1 + γ

λ1

− 2 + r

λ1(1 + r)exp(−λ1c

∗w11 ) (6)

where

c∗w11 =(1 + r)(y − ε1) + (mr − b)

2 + r− ln[γ(1 + r)E(exp[−λ1y12])]

λ1

(7)

Likewise, the agents problem when she loses the bid is given by

Page 9

9

V1l ≡ maxc11,a11

u(c11) + δE(u(c12) (8)

a.t. c11 + s11 = (y − ε1 − γi)−m (9)

c12 = y12 + (1 + r)s11 −m+ 2m+ b2 (10)

Solving for the indirect utility function results in

V1l =1 + γ

λ1

− 2 + r

λ1(1 + r)exp(−λ1c

∗l11) (11)

where

c∗l11 =(1 + r)(y − ε1)− (mr − b)

2 + r− ln[γ(1 + r)E(exp[−λ1y12])]

λ1

(12)

Analogous solutions exists for agent 2 with λ1 replaced with λ2.

2.2.2 Symmetric bayesian nash equilibrium under heterogeneous preferences

Let bi(ε1) > bj(ε2) if ε1 > ε2 for i 6= j. Consider the bidding strategy for agent 1, b1(ε) for

some realized income shock ε > 0. This section uses truth telling requirements to pin down

the equation for the Symmetric Bayesian Nash equilibrium. The expected utility of agent 1

for reporting ε̄1 6= ε1 is

U1(ε1, ε̄1) = pV1w + (1− p){Prob(0 < ε2 <= ε̄1)× V1w + Prob(ε2 > ε̄1)× E[V1l|ε2 > ε̄1]}

(13)

In words, when agent 1 submits ε̄1 he wins the ROSCA fund in the first period if agent

2 does not receive an income shock, which happens with probability p. However if agent 2

Page 10

10

receives an income shock, which happens with probability 1-p , then agent 1 wins the ROSCA

fund if ε̄1 ≥ ε2 > 0 and loses the ROSCA fund in the first period if ε2 > ε̄1. Since V1l depends

on b2(ε2), agent 1 takes the expected value of his indirect utility function from losing on the

support [ε̄1, z] where z is the maximum income shock one can observe. Truth-telling requires

that

∂U(ε, ε̄)

∂ε̄

∣∣∣∣ε̄=ε

= 0 (14)

Solving the first order conditions, one can show as in Fang and Ke (2006 version: p. 26-27

: seminar publication at wisc.edu) that the Symmetric Bayesian Nash equilibrium bid for

agent 1 and 2 are

b1(ε, λ) = mr +2 + r

2λ1

ln{1− [1− exp(−2λ1mr

2 + r)][

p

p+ (1− p)F (ε)]2} (15)

b2(ε, λ) = mr +2 + r

2λ2

ln{1− [1− exp(−2λ2mr

2 + r)][

p

p+ (1− p)F (ε)]2} (16)

2.2.3 Comparative statics

It is interesting to note what happens to the equilibrium bid in a ROSCA when income

shocks rise (as they often do during an economic crisis). Taking the derivative of the equi-

librium bid with respect to ε for λ > 0, p > 0, r > 0,m > 0 one obtains

∂b(ε)

∂ε=

2 + r

2λ

({1− 1

e2λmr2+r

}[ p

p+ (1− p)F (ε)]2f(ε)

F [ε]2

)/A > 0 (17)

Page 11

11

where A ≡ 1− [1− exp(−2λ2mr2+r

)][ pp+(1−p)F (ε)

]2 > 0 .

The positive partial derivative of the equilibrium bid with respect to ε suggest that during

an economic crisis, individuals who realize larger incomes shocks should bid higher within

a ROSCA and -all else being equal-, secure funds in the current period to smooth their

consumption.

Given income shocks that are aggregate (similar income shock realizations), it may also

be interesting to compare how the equilibrium bid changes with respect to the elasticity of

marginal utility of consumption (u′′(c)cu′(c)

).

Theorem 2.1. Let λ1λ2> lnA(λ1)

lnA(λ2)then b1(ε, λ1) > b2(ε, λ2) where A[λi] ≡ 1−[1−exp(−2λimr

2+r)][ p

p+(1−p)F (ε)]2

Proof. Since lnA(λ) < 0 for all λ > 0 we have from the hypothesis that lnA(λ1)λ1

> lnA(λ2)λ2

.

Multiplying both sides by 2+r2> 0 and adding mr > 0 to the left and right hand side of the

equation we have that mr + 2+r2λ1

lnA(λ1) > mr + 2+r2λ2

lnA(λ2) which implies that b1(ε, λ1) >

b2(ε, λ2)

With no loss of generality, consider 2 agents in a ROSCA with identical consumption

levels3 (c1 = c2 = c). The hypothesis in theorem 2.1 can then be rewritten as λ1clnA(λ1)

> λ2clnA(λ2)

.

The above theorem argues that in making a bid, an agent considers how the bid increases his

marginal utility of consumption (elasticity of marginal utility of consumption is λc) and the

price she has to incur for the bid (lnA(λ) 4). If the gain in marginal utility of consumption

for a bid relative to its price for agent 1 exceeds that of agent 2 then agent 1 will have the

highest bid. In an oral auction, all else equal, an agent with a high elasticity of marginal

utility of consumption (λc) will win the ROSCA fund in period 1 5.

3One can argue that individuals within a ROSCA are homogeneous and have similar consumption levels.4Note that as lnA(λ) decreases, the interest rates within a ROSCA approach those of the formal sector

which is given by mr5If consumption levels are not identical then one obtains ( c2c1 ) λ1c1

lnA(λ1)> λ2c2

lnA(λ2)which simply states that

if c1 < c2 it is more likely for agent 1 to have the highest bid since her gain in the marginal utility ofconsumption for a bid relative to its price is upweighted by the fraction c2

c1(The consumption level also

matters in determining who has the highest bid).

Page 12

12

2.3 Implications of the model

The results above have implications for this study in understanding how ROSCAs allocate

the common fund during national economic crises. The model suggests that all else equal

agents with higher elasticities of marginal utilities of consumption (MU) are more likely to

secure the ROSCA fund even if income shocks are aggregate among members. The model

predicts that pregnant women who have a higher MU elasticity -either through a preference

change during pregnancy or an income shock from withdrawal from the labor market- should

bid higher than their peers to secure the pooled ROSCA contribution, which can then be

used to smooth their consumption. ROSCAs that facilitate borrowing within the scheme

therefore provide consumption protection for pregnant women since higher MU elasticities

during pregnancy can be revealed through a higher bidding price or interest rate. The

predictions of this model call for empirical analysis to determine if it holds true in an actual

case study.

3 Data and Sample Selection

3.1 Data

The data for this study comes from the Indonesia Family Life Survey (IFLS) which is a

representative longitudinal study of the Indonesian population 6. The survey has a wealth of

information on socioeconomic, demographic, and community characteristics. The waves of

the survey that are used in this study are IFLS 2, 3 and 4 which, were implemented in 1997,

2000, and 2007, respectively.

There are two main advantages of using this survey : first, the 1997 IFLS was initiated in

Indonesia from June to November 1997, before the sudden onset of the economic crisis. As a

result, this wave of the survey provides baseline information about individual and community

characteristics, as well as ROSCA participation leading up to the January 1998 crisis. Second,

6The survey with weights represents 83% of the Indonesian population

Page 13

13

the IFLS is a longitudinal study with an impressively low attrition rate (an average attrition

of about 6% from 1993 to 2007). Consequently, this survey allows for one to assess the

dynamics of ROSCA institutions during the 1998 Indonesian financial crisis and plays an

essential role in explaining the empirical results found in this study.

3.2 Sample selection and outcomes

The sample for this study consists of women aged 15 years and over who completed the

1997 Indonesian Family Life Survey (IFLS2). Women without anthropometric information

were excluded from this study 7. In addition, to ensure data quality, women who gave birth

at locations where birth weight measurements were of low frequency (< 100) were excluded

from the analysis. The final sample consists of 3,800 births from 1995 to 2006. 8

The three main outcomes analyzed in this study are birth weight, food consumption

expenditure and ROSCA loans initiated during the 1998 Indonesian financial crisis. These

outcomes provide consistent evidence for the role of ROSCA institutions in preventing adverse

birth outcomes during the 1998 Indonesian financial crises.

4 Identification issues and strategy

There are two potential challenges to the identification of the causal impact of ROSCA

participation in mitigating adverse birth outcomes during the 1998 Indonesian financial crisis.

First, it could be argued that, in anticipation of the crisis, women may have delayed preg-

nancy. If so, then differences in birth outcomes observed during the crisis may be attributable

to changes in parent characteristics and not the causal impact of the 1998 Indonesian financial

crises. Since the onset and magnitude of the 1998 Indonesian financial crisis were unexpected

and unprecedented, this is unlikely to be the case. To test for endogenous fertility or sorting

7Birthweight, birth month and birth year.8Robustness checks show that including births in locations where measuring birth weight was not the

norm does not change the results of this study

Page 14

14

into birth, one can search for significant deviations in mothers’ characteristics during the cri-

sis. Insignificant deviations in mothers’ characteristics provide evidence against endogenous

fertility. The econometric model below is estimated.

Yi = β0 + β1Y OB1998i + F (Y OBi, θ) + εi (18)

In equation 18, i denotes an individual, Y is the outcome of interest (mothers’ socioe-

conomic and household characteristics), Y OB1998 is a dummy for the cohorts born in 1998,

F (Y OB, θ) is a polynomial in year of birth that controls for cohort trends and, ε is the error

term. In equation 18, β1 is the coefficient of interest, since it estimates the deviation in the

outcome variable from trend in 1998 (crisis year). Estimates for B1 are relevant for assessing

deviations in parental characteristics and birth outcomes during the 1998 Indonesian financial

crisis.

The second challenge to the identification strategy is that people who participate in

ROSCAs may be fundamentally different than those who do not participate in a ROSCA. If

ROSCA participation is endogenous, this will bias results. To address endogeneity in ROSCA

participation, this study utilizes a symmetric weighting difference in difference method to

reduce differences in observables and to control for time invariant unobservables between

ROSCA and non-ROSCA participants (Sylvain, 2014, 2015) 9). This approach entails 4

steps as detailed below.

First, propensity scores for ROSCA participation are estimated for each individual in

the sample using the covariate balancing propensity score method (Imai et al., 2014). The

covariate balancing propensity score method models treatment assignment while optimizing

covariate balance using a general method of moments framework. The covariate balancing

propensity score method, as shown by Imai and Ratkovic, is robust to model misspecification,

9A symmetric weighting difference in difference (SWDID) method creates a synthetic sample where co-variates in the sample are independent of ROSCA participation. A symmetric matching difference in differ-ence(SMDID) method on the other hand , matches each treated individual to a set of controls. The SWDIDmethod is preferred in this study since all participants in the sample are used in the analysis while the SMDIDdiscards individuals for which a match from the control set cannot be found.

Page 15

15

prevents adhoc searches and improves the poor empirical performance of propensity score

matching and weighting methods reported in the literature. The choice of variables used in

the propensity score are selected after a literature review of covariates that differ between

ROSCA and non-ROSCA participants (Varadharajan, 2004). The covariates identified are

age, household size, occupational choices (farmers or business owners), wealth, education,

location of households and baseline access to formal and informal financial schemes. These

covariates are used to predict the probability of ROSCA participation for all individuals in

the sample using the covariate balancing propensity score method.

Second, the propensity scores are used to weight the sample to ensure that the distribution

of covariates in the sample are independent of ROSCA participation (the treatment group).

To obtain average treatment effects on the treated, non-ROSCA participants received a weight

of p1−p where p is the estimated propensity score from the covariate balancing propensity

score method. ROSCA participants received a weight of 1. The weights for non-ROSCA

participants were normalized to sum to 1 to reduce the influence of propensity scores that

are outliers (Austin, 2011).

Third, to control for unobservables that do not change over time, a difference in difference

model is estimated using the synthetic sample created with propensity score weights. The

equation estimated is

Yi = β0+β1ROSCA∗Crisisyear+β2Crisisyear+β3ROSCA∗PostCrisis+β4PostCrisis

+ β5ROSCA+Xiβ′6 + εi (19)

In equation (19), Y is the outcome of interest and denotes birth weight and food con-

sumption expenditure, ROSCA is a dummy equal to 1 if the person participated in a ROSCA

before the onset of the crisis, X is a vector of controls for season of birth as well as mother’s

household and socioeconomic characteristics, Crisisyear is a dummy equal to 1 in the year

Page 16

16

of the crisis (1998) , PostCrisis is a dummy equal to 1 in the year after the crisis (1999)

and ε is the error term. Observations are restricted to 24 months before and after the crisis.

Note that β1 is an intent to treat estimate since some ROSCA participants in 1997 may no

longer have been active in 1998. The intent to treat framework is suitable in this context,

since it minimizes endogenous sorting in and out of ROSCA during the crisis.

Finally, to account for unobservables that may impact the treatment and outcome vari-

able simultaneously, Rosenbaum bounds for B1 are calculated to determine how strongly an

unmeasured variable must influence the selection process to undermine the analysis (Becker

and Caliendo, 2007). Results from this exercise show that observables in the IFLS survey do

not have a strong enough effect on treatment to alter the implications/results of this study.

5 Sample summary and background

During the 1998 Indonesian financial crisis, the rupiah exchange rate fell dramatically,

causing inflation to rise to unprecedented levels. Figure 1 plots year over year data from the

St. Louis Federal Reserve on the Indonesian economy over time, based on data from the St.

Louis Federal Reserve. The trend line shows a sharp decline in the real broad exchange rate

from 104.68 to 51, and shows a 83.9% and 52.16% increase in the inflation rates of food and

non-food commodities, respectively. The relative increase in the prices of food suggests that,

as with other crises, expectant mothers’ food consumption may have been compromised.

Summary statistics for the study sample are provided in Table 1. Women in the sample

are on average 25 years of age, with 8% in poor health and 91% confirming that they can

read and write. On average, each household has about 6 members with each member owning

assets worth 9,050 rupiah. Households in the sample have diverse occupations with 38%

owning a side business and 29% owning a farm. Among the sample group, 53% of households

are situated in an urban area. Finally, 65% of households have access to a formal credit

scheme, while 15% had access to an informal credit scheme. Social networks also proved

to be an important source of loans: 35% of women in the survey reported that they could

Page 17

17

borrow funds from family or friends. Columns 2 and 3 of Table 1, tests for deviations in

mother and household characteristics for the 1998 birth cohort as detailed in Equation (18).

Results show that 13 out of 15 variables are statistically insignificant and with the existing

literature, provides additional evidence that the crisis was unexpected (Iriana and Sjoholm,

2002; Levinsohn et al., 2003; Sharma, 2001). These results provides direct evidence against

the hypothesis that women selectively sorted into child birth during the crisis. To assess

significant deviations in low birth weight, variables listed in table 1 are used as controls in

this study.

6 Results

6.1 The impact of the 1998 Indonesian crisis on birth outcomes

The economic effects of the 1998 Indonesian crisis had a strong impact on birth outcomes.

The demonstrable link between birth incomes and nutritional intake during the gestational pe-

riod makes expectant mothers a particularly vulnerable population during a crisis. Columns

2 and 3 of Table 2 allow for a comparison of birth outcomes in crisis and non-crisis years.

During the crisis, the incidence of low birth weight increased by 5 percentage points from

11% to 16% with mean birth weight declining by approximately 44 grams.

Figure 2 shows trends in low birth weight by birth cohort and the deviations from trend

for individuals born during the crisis (the 1998 birth cohort) 10. To formally assess the

magnitude of the deviations from trend for individuals born during the crisis year, Table 3

provides estimates for Equation (18), where the main outcome of interest is low birth weight.

Estimates indicate that during the crisis there was a 4.78 percentage point increase in the

incidence of low birth weights from trend. Controlling for gestation and other covariates as

specified in table 1, estimates of the increase in the incidence of low birth weight conservatively

falls to 4.1 percentage points. The robustness of the estimates to controls for gestation

10As shown in the graph the deviations are mostly observed by individuals that were born in the secondhalf of 1998.

Page 18

18

suggest that regardless of gestational length fetal growth was restricted. Health studies

demonstrate that restrictions in fetal growth are related to the mother’s nutritional intake

during pregnancy (Kelly, 2011). Thus, the robustness of the estimates to gestational length

provides direct evidence that reductions in mothers’ food consumption is a major factor

explaining the increase in low birth weight during the crisis year.

Panel B of Table 3 provides a breakdown of the increase in the incidence of low birth

weight by time of year for individuals who were born in 1998. Columns 1 through 3 illustrate

that the greatest increase in low birth weight during the crisis were for individuals born in the

second half of 1998. This observation is attributable to the fact that individuals born during

the second half of 1998 were in their 2nd or 3rd trimester at the height of the Indonesia crisis.

During this period, a fetus is the most sensitive to nutritional intake. Reductions in mean

birth weight, unlike low birth weight, were statistically not significant, suggesting that the

greatest impact of the crisis was to increase the left tail of the birth weight distribution. In

particular, by shifting women whose children were on the margin of low birth weight (low

income mothers) into low birth weight 11.

To assess the role of prenatal investments in explaining the rise in adverse birth outcomes

during the crisis, deviations in prenatal care from trend for the 1998 birth cohort are estimated

in Table 4. These estimates show no significant reductions in prenatal care during the crisis.

This is likely due to the fact that prices at community clinics (puskesmas and pustu) remained

low during the crises (Strauss et al., 2004).

To summarize, for the 1998 birth cohort, the Indonesian financial crisis increased the

fraction of low birth weight babies by 4.1 percentage points. The primary cause of this

increase is shown to be mother’s food consumption, rather than prenatal care.

11Insignificant results on birth weight could simply reflect the fact that the sample is underpowered. Notethat birth weight consistently decreases during the crises.

Page 19

19

6.2 Descriptive: Evolution of credit schemes within ROSCAs to protect food

consumption during the crisis

Pre-crisis, as shown in Table 5, 45% of women in the sample were participating in a

ROSCA and expected an average payout of $144.61 every 16 months. Members met on aver-

age every 3 weeks and paid $7.17 to the common fund at each meeting. Figure 3 shows that

ROSCAs modified their structures to establish and facilitate borrowing within the scheme to

members during the crisis. This result emphasizes that ROSCAs are dynamic and adaptable

institutions. Table 6 confirms the expansion of credit lines by ROSCAs during the financial

crisis based on the IFLS survey. Pre-crisis (1993 to 1997), only 1% of households reported

that they could borrow from a ROSCA. During the crisis (1997 to 2000), this number in-

creased to 8.4%. In the 12 months prior to the 2000 IFLS survey, the average amount loaned

to ROSCA participants was $49 for 5.8 months with a 3.2% monthly interest rate. Of the

ROSCA loans, 84% were of fixed length and 99% of the loans were made collateral-free.

The large expansion of credit occurred within ROSCA institutions where members had

uncorrelated income sources (not an office/work ROSCA) as shown in figure 4. The panel

shows that approximately 90% of all ROSCA loans were from ROSCA institutions where

members had uncorrelated income sources. Conditional on participating in a ROSCA at

work, 5% of members reported that they had the ability to borrow from a ROSCA. In

contrast, 9% of members participating in ROSCAs unaffiliated with the workplace reported

that they had an ability to borrow from a ROSCA. These results indicate that participating

in a ROSCA where members have uncorrelated income sources and where income shocks

were likely to be idiosyncratic increased the odds of a ROSCA facilitating borrowing during

the crisis by a factor of 2.

Finally, ROSCA loans during the crisis played an important role in smoothing food con-

sumption. Existing literature indicate that ROSCA funds are used to purchase durables.

Hoever table 7 shows that 12 months preceding the 2000 IFLS survey approximately 45%

of ROSCA loans were used to purchase necessary groceries during the crisis. The next four

Page 20

20

frequent categories were education at 16.6%, capital for other business at 9.96%, purchase of

household items at 6.64% and farm inputs at 4.56%. These top five categories represented

80% of survey responses and indicate a shift in the use of ROSCA funds during the crisis.

Information on the use of ROSCA loans were obtained from individuals who participated

in ROSCA and necessarily borrowed from it. This ignores in-kind food transfers that may

have also occurred within ROSCAs. To understand the overall impact of participating in a

ROSCA on food consumption via loans and in-kind transfers, table 8 provides difference in

differences estimates comparing real food expenditure per capita by ROSCA participation pre

and during crisis 12. Estimates show that pre-crisis (1993 to 1997) no differences in changes

in real food expenditures are observed for households that participated in ROSCAs in 1997 as

compared to households that did not. However, through the crisis (1997 to 2000), higher real

food expenditure per capita of 1102.6 Rupiah is observed for individuals that participated in

ROSCAs in 1997.

To summarize, ROSCAs adapted and facilitated borrowing during the crisis which was

used primarily to smooth food consumption. ROSCAs where members had uncorrelated

income sources were 2 times more likely to extend credit as compared to ROSCAs where

members had uncorrelated income sources. The next section investigates the advantages of

this protection in reducing adverse birth outcomes as proxied by low birth weight.

6.3 ROSCA participation and birth outcomes during crisis

ROSCA participation is likely to be endogenous. Figure 5 shows that the extremely poor

and wealthy have lower probabilities of participating in a ROSCA. Thus, it is not possible

in this study to simply draw causal inference from a simple comparison of outcomes between

ROSCA and non-ROSCA participants. To address non-random sorting into ROSCA, propen-

sity score weights are constructed (see Section 4 for details). Once the sample is weighted

with inverse probability weights, as shown in Figure 6, the differences in covariates between

12Real food expenditure was adjusted used CPI information at the province level. Adjusting for real foodexpenditure using average levels of prices in a province circumvents the problems of endogenous prices.

Page 21

21

ROSCA and non-ROSCA participants are no longer significant (standardized differences are

below the 10% threshold). Table 9 provides further information on the weighted sample

by providing the means and standardized differences before and after inverse probability

weighting.

To address endogenous sorting into ROSCA the study uses a symmetric weighted differ-

ence in difference method. The left column of Figure 7 provides some intuition behind the

symmetric weighting difference in difference method used in this study to address endogenous

sorting into ROSCA. Leading up to the crisis, ROSCA participants observe slightly higher

rates of low birth weight as compared to non-ROSCA participants. The differences in birth

outcomes between the two groups leading up to the crisis is largely stable in the pre-crisis

period. However, during the crisis, as shown in Figure 7, non-ROSCA participants had a

larger increase in the fraction of low birth weight relative to ROSCA participants. The sym-

metric weighting difference in difference estimate is graphically represented by the distance

between the horizontal dashed line 13 and differences in birth outcomes between ROSCA and

non-ROSCA participants during the crisis year. Post-crisis, the differences in birth outcomes

slowly return back to pre-crisis trend.

The results provided in the left column of Figure 7 and discussed above, depend on the

scale of the axis. The right column of figure 5 addresses this issue by providing results that

are not dependent on scale. The graph plots t-tests for significant deviations in differences

in low birth weight from mean trend between ROSCA and non-ROSCA participants. As

shown in the right column of figure 5 , significant deviations in differences in low birth weight

between ROSCA and non-ROSCA participants are observed from mean trend only in the

crisis year. These results further support the hypothesis that during the crisis, increases in

low birth weight for ROSCA participants were lower than those observed for non-ROSCA

participants.

Table 10 provides estimates for SWDID as specified in equation (19). The SWDID meth-

13which represents mean difference in birth outcomes between ROSCA and non-ROSCA participants pre–crisis

Page 22

22

ods is applied 12 and 24 months before and after the crisis. The most conservative estimates

show that ROSCA participants had 8.1 percentage points fewer low birth weight babies as

compared to non-ROSCA participants. The estimates are robust to controls for gestation

and suggest that reductions in food consumption during the crisis year are likely to be a

major factor explaining the differences in birth outcomes between ROSCA and non-ROSCA

participants. It is important to note that the estimates above do not imply that ROSCA

participants were totally insulated from the crisis. In general, estimates suggest that the

crisis increased the rates of low birth weight for ROSCA participants by 1 percentage point

during the crisis year. Rather, this study emphasizes that this small increase in incidences

of low birth weights is smaller for ROSCA participants than for non-ROSCA participants,

illuminating the comparative benefits of alternative lending for expectant mothers without

access to credit from financial institutions.

7 Robustness checks

So far, the results of this study have focused entirely on outcomes in the year 1998.

Replicating the analysis of this study using non-crisis years did not yield significant deviations

in low birth weight from trend. The fact that low birth weight did not increase in 1999 may

be surprising. As shown in Strauss (20004), this can be explained by household adjustments

which largely mitigated the effects of the crisis in the medium term (Strauss et al., 2004).

Propensity scores in this study were estimated using the covariate balancing propensity

score model, since this method is robust to model misspecification (Imai , 2014). The final

robustness check replicates the main results in this study by using a simple logit model to

estimate the propensity of ROSCA participation while ensuring that mean propensity scores

and covariates among treated groups are balanced within sample blocks 14(Garrido et al.,

2014). Results are similar across methods.

To ensure data quality, women who gave birth in locations where birth weight measure-

14Using the 10% acceptable threshold for standardized differences as measure of balance

Page 23

23

ments were of low frequency (< 100) were excluded from the analysis. Including these women,

the analysis was repeated, which resulted in similar estimates with larger standard errors.

Finally, Rosenbaum bounds for B1 were calculated to determine how strongly an un-

measured variable must influence the selection process to undermine the implications of the

analysis. The calculated bounds show that the study is insensitive to an omitted variable

that would increase the the odds of ROSCA participation by 30%. Using a hypothetical

individual with mean set of covariates across treatment and controls none of the observables

in the dataset satisfied this criteria (for an individual with mean levels of covariates, age

increases the odds of ROSCA participation by the largest amount which is 13%) , confirming

the robustness of the study estimates.

8 Discussion and conclusion

This study determines that ROSCA participants had 8.1 percentage points fewer low

birth weights as compared to non-ROSCA participants during the 1998 Indonesian financial

crisis. The lower rates of low birth weight amongst ROSCA participants are related to the

expansion of ROSCA loans during the crisis. ROSCA loans helped members to maintain

food consumption, as evidenced by the fact that 45% of ROSCA loans were used to purchase

groceries. This paper argues that if pregnancy increases the elasticity of the marginal utility

of consumption, then pregnant women should be more likely to secure ROSCA funds, since

they are willing to incur higher interest rates to borrow funds from ROSCA institutions in

order to maintain consumption. The lower incidence of low birth weight among ROSCA

participants as compared to non-ROSCA participants provides empirical evidence for this

hypothesis.

Economists have verified that adverse birth outcomes rise during national economic crises

(Alderman, 2011). Understanding the capacity of institutions to protect birth outcomes

during national economic crises has been and continues to be a priority for policy makers

(Skoufias, 2003). In the US, Medicaid, Food stamps and Supplemental Feeding Programs

Page 24

24

for Women, Infants, and Children (WIC) have been shown to protect birth outcomes for the

most vulnerable members during national economic crises (Almond and Currie, 2011). On

the other hand, in developing countries, there is limited empirical evidence on the types of

institutions that can protect birth outcomes during periods of crisis. This paper is the first to

confirm that credit extensions from ROSCA institutions, which are ubiquitous in developing

countries, can help pregnant women maintain their food consumption and thus protect birth

outcomes during national economic crises.

Peer to peer banking schemes during national economic crises face an enforcement prob-

lem, since borrowers are more likely to default on their credit payments. Demanding collateral

alleviates the enforcement problem, however this requirement often excludes individuals in

greatest need of loans. As a type of P2P banking scheme, ROSCAs are of particular interest

since they rely on social sanctions, rather than collateral to prevent payment defaults. Dur-

ing the 1998 Indonesian financial crisis, the fact that 99% of ROSCA loans were issued free

of collateral suggests that such schemes can be an alternative source of credit to individuals

who are credit constrained during economic crises as a result of collateral requirements. In

addition, the fact that ROSCA participation in Indonesia was immutable during the crisis

emphasizes the ability of social capital to not only sustain these schemes during national

economic crisis but also to facilitate credit extensions 15.

Existing studies demonstrate the sustainability of some informal insurance arrangements

through credit extensions, when income shocks are correlated as in economic crisis. For

instance, Platteau and Abraham, document regular credit transactions among fisherman in

a South Indian village to insure themselves from low proceed realizations, which are often

partially correlated (Platteau and Abraham, 1987). In South Africa, Carter and Maluccio

show that communities with larger social capital are better able to sustain informal insurance

arrangements and offer credit extensions when income shocks are partially correlated at the

community level (Carter and Maluccio, 2003). The findings of the present study confirm

15Repayment of ROSCA loans are similar to those from formal banks and emphasizes the viability andpotential of using social sanctions to expand credit lines within peer to peer banking schemes

Page 25

25

these local and regional findings at a national level, showing that ROSCAs have the capacity

to provide informal insurance during national economic crises.

Existing economics literature provides four main reasons for the formation of a ROSCA.

The reasons are to finance the purchase of indivisible durable goods (Besley at al. , 1993) , to

commit to save (Dagnelie and Lemay-Boucher, 2012, Ambec and Triech, 2007, Aliber,2001,

Gugerty, 2007), to protect savings against claims from kin (Dupas and Robinson, 2009, An-

derson and Baland, 2002), and as an insurance mechanism for unexpected events (Calomris

and Rajaraman, 1998, Fang and Ke, 2006, Klonner, 200). The expansion of credit lines

during the 1998 Indonesian financial crisis observed during the study confirms informal risk

arrangements within ROSCAs which are not easily observable in non-crisis years and empha-

sizes the importance of recognizing these institutions as flexible entities that serve multiple

roles.

To generalize the study findings, it will be important for future research to elucidate the

level of social capital required for P2P banking schemes to be sustainable during national

economic crises. The ability to replicate the success of ROSCA institutions during the 1998

Indonesian crisis in other countries, where social capital may be depleted, will require a

detailed model on how the design of such institutions can minimize its reliance on social

capital. The presence of ROSCAs globally, indicates that policies to bolster these institutions

in collaboration with other government programs 16 may be a powerful and cost effective way

of protecting pregnant women in developing countries during national economic crises.

16Other government programs are still needed since the absolute poor are still excluded from participatingin ROSCA institutions due to regular fixed payments required for membership

Page 26

26

References

Alderman, Harold, “No Small Matter: Human Development Perspectives,” World Bank

Publications, 2011.

Almond, Douglas and Janet Currie, “Killing me softly: The fetal origins hypothesis,”

The journal of economic perspectives: a journal of the American Economic Association,

2011, 25 (3), 153.

Austin, Peter C., “An introduction to propensity score methods for reducing the effects

of confounding in observational studies,” Multivariate behavioral research, 2011, 46 (3),

399–424.

Becker, Sascha O and Marco Caliendo, “Mhbounds-sensitivity analysis for average

treatment effects,” 2007.

Behrman, Jere R and Mark R Rosenzweig, “Returns to birthweight,” Review of Eco-

nomics and statistics, 2004, 86 (2), 586–601.

Bharadwaj, Prashant, Juan Eberhard, and Christopher Neilson, “Do initial endow-

ments matter only initially? the persistent effect of birth weight on school achievement,”

Department of Economics, UCSD, 2010.

Black, Sandra E, Paul J Devereux, and Kjell Salvanes, “From the cradle to the labor

market? The effect of birth weight on adult outcomes,” Technical Report, National Bureau

of Economic Research 2005.

Bouman, Frits JA, “Rotating and accumulating savings and credit associations: A devel-

opment perspective,” World development, 1995, 23 (3), 371–384.

Bouman, Fritz JA, “INDIGENOUS SAVINGS AND CREDIT SOCIETIES IN THE

THIRD WORLD. A MESSAGE/UN ENSEIGNEMENT: LES SOCIETES INDIGENES

D’EPARGNE ET DE CREDIT,” Savings and development, 1977, pp. 181–219.

Page 27

27

Burlando, Alfredo, “When the Lights Go Out: Permanent Health Effects of Transitory

Shocks,” Unpublished paper, Boston University, Boston, 2010.

Carter, Michael R and John A Maluccio, “Social capital and coping with economic

shocks: an analysis of stunting of South African children,” World Development, 2003, 31

(7), 1147–1163.

Cruces, Guillermo, Pablo Gluzmann, and Luis Felipe Lopez-Calva, “Permanent

effects of economic crises on household welfare: Evidence and projections from Argentina’s

downturns,” 2010.

Dercon, Stefan, “Income risk, coping strategies, and safety nets,” The World Bank Research

Observer, 2002, 17 (2), 141–166.

Fang, Hamming and Rongzhu Ke, “The Insurance Role of ROSCA in the Presence

of Credits Markets: Theory and Evidence,” unpublished paper. Department of Economic,

Yale University, 2006.

Garrido, Melissa M., Amy S. Kelley, Julia Paris, Katherine Roza, Diane E. Meier,

R. Sean Morrison, and Melissa D. Aldridge, “Methods for constructing and assessing

propensity scores,” Health services research, 2014, 49 (5), 1701–1720.

Hoogeveen, J. G. M., “Income risk, consumption security and the poor,” Oxford Devel-

opment Studies, 2002, 30 (1), 105–121.

Hospes, Otto, “PEOPLE THAT COUNT: THE FORGOTTEN FACES OF RO-

TATING SAVINGS AND CREDIT ASSOCIATIONS IN INDONESIA/LES GENS

QUI COMPTENT: LES FACES OUBLIEES DES ASSOCIATIONS ROTATOIRES

D’EPARGNE ET DE CREDIT EN INDONESIE,” Savings and Development, 1992,

pp. 371–401.

Iriana, Reiny and Fredrik Sjoholm, “Indonesia’s economic crisis: contagion and funda-

mentals,” The Developing Economies, 2002, 40 (2), 135–151.

Page 28

28

Kelly, Elaine, “The scourge of asian flu in utero exposure to pandemic influenza and the

development of a cohort of british children,” Journal of Human Resources, 2011, 46 (4),

669–694.

Klonner, Stefan, “Rotating Savings and Credit Associations as Insurance,” Sudasien In-

stitut der University Heldeberg, Working paper, 2000.

Levinsohn, James A., Steven T. Berry, and Jed Friedman, “Impacts of the Indonesian

economic crisis. Price changes and the poor,” in “Managing currency crises in emerging

markets,” University of Chicago Press, 2003, pp. 393–428.

Platteau, Jean-Philippe and Anita Abraham, “An inquiry into quasi-credit contracts:

The role of reciprocal credit and interlinked deals in small-scale fishing communities,” The

Journal of Development Studies, 1987, 23 (4), 461–490.

Prabowo, Dibyo, “The Role of Informal Financial Intermediation in the Mobilization of

Household Savings and Allocations in Indonesia,” in “in” Seminar on Informal Financial

Markets in Development 1989.

Sharma, Shalendra D., “The Indonesian financial crisis: From banking crisis to financial

sector reforms, 1997-2000,” Indonesia, 2001, pp. 79–110.

Skoufias, Emmanuel, “Economic crises and natural disasters: Coping strategies and policy

implications,” World Development, 2003, 31 (7), 1087–1102.

Smith, James P., Duncan Thomas, Elizabeth Frankenberg, Kathleen Beegle, and

Graciela Teruel, “Wages, employment and economic shocks: Evidence from Indonesia,”

Journal of Population Economics, 2002, 15 (1), 161–193.

Strauss, John, Kathleen Beegle, Agus Dwiyanto, Yulia Herawati, Daan Patti-

nasarany, Elan Satriawan, Bondan Sikoki, and Firman Witoelar, Indonesian liv-

ing standards: before and after the financial crisis., Institute of Southeast Asian Studies,

2004.

Page 29

29

Sylvain, Chabe-Farret, “Analysis of the bias of Matching and Difference-in-Difference

under alternative earnings and selection processes,” Journal of Econometrics, 2015, 185

(1), 110–123.

Sylvain, Chabe-Ferret, “why does difference in difference matching work,” Toulouse School

of Economics and Inra, Lerna, 2014.

Townsend, Robert M., “Consumption insurance: An evaluation of risk-bearing systems

in low-income economies,” The Journal of Economic Perspectives, 1995, pp. 83–102.

Varadharajan, Sowmya, “Explaining Participation in Rotating Savings and Credit Asso-

ciations (RoSCAs): Evidence from Indonesia,” Mimeograph, 2004.

Wie, THEE, “The impact of the economic crisis on Indonesia’s manufacturing sector,” The

Developing Economies, 2000, 38 (4), 420–453.

Williams, Glen and Mary Johnston, “The Arisan: A Tool for Economic and Social

Development?,” Prisma, 1983, 29, 66–73.

Page 30

30

Figure 1: Features of the Economic Crisis in Indonesia40

6080

100

120

Rea

l Effe

ctiv

e B

road

Exc

hang

e R

ate

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

Year

Real effective exchange rates are calculated as weighted averages of bilateral exchange rates−adjusted by relative consumer prices.

Real Broad Exchange Rate by Year

12

34

56

Inde

x

1990m1 1992m1 1994m1 1996m1 1998m1 2000m1 2002m1 2004m1 2006m1 2008m1 2010m1Month

Indonesia exchange rate plummeted in July 2007

Exchange rate by month

020

4060

8010

0In

flatio

n (%

)

1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

Year

Inflation : (all commodities) Inflation:(food)

Inflation (Yearly basis)

050

100

150

Infla

tion

1990m1 1992m1 1994m1 1996m1 1998m1 2000m1 2002m1 2004m1 2006m1 2008m1 2010m1Month of birth

Inflation Inflation : food

Increases in the cost of living spiked after january 1998

Inflation by month (yearly basis)

Figure shows unusual deviations in exchange rates and inflation during the 1998 Indonesian financial crisis.

Vertical line denotes the year in which the Indonesian crisis occured

Page 31

31

Figure 2: Trends in birth outcomes

.05

.1.1

5.2

.25

% L

ow b

irth

wei

ght (

< 2

500g

)

1995

h1

1996

h1

1997

h1

1998

h1

1999

h1

2000

h1

2001

h1

2002

h1

2003

h1

2004

h1

2005

h1

2006

h1

Year of birth by half year

Notes:Vertical lines 2nd half of 1998

Mean Trends by Year of Birth by Half Year% Low birth weight (< 2500g)

Figure shows an increase in low birth weight for cohorts born during the 1998 Indonesian financial crisis.

Page 32

32

Figure 3

0.0

2.0

4.0

6.0

8pr

opor

tion

of h

ouse

hold

s

Pre−crisis(’94−’97) During−crisis(’98−’00) Post−crisis(’01−’07)

Pre and post crisis% Households : ROSCA facilitates borrowing

Panel

shows the expansion of ROSCAs that facilitate borrowing during the crisis and its subsequent contraction

post crisis. The ability to obtain credit within a ROSCA institution was based on ROSCA participation pre

crisis.

Page 33

33

Figure 4

90.6%

9.4%

ROSCA not at work ROSCA at work

ROSCA at work (15.5%), ROSCA not at work (84.5%)

ROSCA credit line by type of ROSCA

0.0

2.0

4.0

6.0

8.1

Has

RO

SC

A c

redi

t lin

e (\

%)

Arisan not at work Arisan at workDifferences in credit extension by ROSCA type are significant with pvalue < 0.001

ROSCA credit line by type of ROSCA

0.0

2.0

4.0

6.0

8.1

Has

RO

SC

A c

redi

t lin

e (\

%)

ROSCA not at work ROSCA at work: office ROSCA at work: marketDifferences in credit extension by ROSCA type are significant with pvalue < 0.001

ROSCA credit line by type of ROSCA

Panel shows that approximately 90% of all ROSCA loans were from ROSCA institutions where members had

uncorrelated income sources. Conditional on participating in a ROSCA at work, 5% of members reported

that they had the ability to borrow from a ROSCA. In contrast, 9% of members participating in ROSCAs

unaffiliated with the workplace reported that they had an ability to borrow from a ROSCA.

Page 34

34

Figure 5: Probability of ROSCA participation by wealth

.3.3

5.4

.45

.5P

roba

bilit

y of

RO

SC

A p

artic

ipat

ion

0

5000

0

1000

00

1500

00

2000

00

2500

00

Household assets per capita in Rupiah

About 8% of the population have wealth less than 15000

household assets per capitaProbability of ROSCA participation by

Figure shows that participation in ROSCAs are not random. The poor and the wealthiest have the lowest

probabilities of participating in a ROSCA.

Page 35

35

Figure 6: Absolute Standardized differences before and after weighting

0 10 20 30 40Standardized % bias across covariates

Household sizeHas access to informal credit sc

In poor healthHousehold owns businessCompleted Middle school

Completed High schoolHas access to formal credit sche

Household owns farmCan read and write

Completed primary schoolCompleted Post Secondary

Urban household in 1997Household assets per capita

Age

Notes:Vertical line bounds accepted values for standardized differences

Before and after weighting Absolute Standardized differences

After Weighting Before weighting

0 .2 .4 .6 .8Propensity Score

Untreated Treated

Estimated probability of ROSCA Participation(ROSCA = Treated group)

Figure shows standardized differences in covariates before and after weighting between ROSCA and non-

ROSCA participants. After inverse probability weighting no significant differences exist across covariates

Page 36

36

Figure 7: Difference in Birth outcomes : ROSCA versus Non-ROSCA participants−.

10

.1.2

.3.4

.5

% L

ow b

irth

wei

ght (

< 2

500g

)

1997

h1

1997

h2

1998

h1

1998

h2

1999

h1

1999

h2

2000

h1

Year of birth

Year before and after the crisis

(Non−Rosca − ROSCA participants)Dashed vertical line denotes the crisis yearGraphs adjusts for seasonalitySample weighted with propensity scores

between ROSCA and Non−ROSCA participantsDifferences in % Low birth weight (< 2500g)

01

23

45

% L

ow b

irth

wei

ght (

< 2

500g

)

1997

h1

1997

h2

1998

h1

1998

h2

1999

h1

1999

h2

2000

h1

Year of birth

A Year before and after the crisis

(Non−Rosca − ROSCA participants)Dashed vertical line denotes the crisis yearGraphs adjusts for seasonalitySample weighted with propensity scores

T−test : between ROSCA and Non−ROSCA participantsTest for stable differences for % Low birth weight (< 2500g)

Figure shows that leading up to the crisis the mean difference in the fraction of low birth weights between

ROSCA and non ROSCA participants were similar. During the crisis however, Non-ROSCA participants

had a larger increase in the fraction of low birth weight relative to ROSCA participants

Page 37

37

Table 1: Sample Summary: Pre-crisis

Deviation from Trend for

1998 Birth cohort

Mean (Sd) Quadratic Trend Cubic Trend

Age 25.13 -0.48418 -0.48457

(6.57) (0.39871) (0.39857)

Poor Health 0.08 0.03778* 0.03780*

(0.27) (0.01957) (0.01957)

Completed primary school 0.13 0.01802 0.01801

(0.34) (0.02241) (0.02240)

Completed Middle school 0.31 0.03285 0.03280

(0.46) (0.02880) (0.02879)

Completed high school 0.23 -0.01784 -0.01778

(0.42) (0.02435) (0.02434)

Completed college 0.25 -0.02187 -0.02189

(0.42) (0.02611) (0.02610)

Can read and write 0.91 -0.01088 -0.01086

(0.28) (0.01888) (0.01887)

Total household assets per member(1000 Rupiah) 9.05 -0.09784 -0.09777

(1.81) (0.11601) (0.11598)

Own Business 0.38 -0.02633 -0.02630

(0.49) (0.02959) (0.02958)

Own Farm 0.29 0.00562 0.00561

(0.46) (0.02796) (0.02796)

Has access to formal credit schemes 0.65 -0.02671 -0.02670

(0.48) (0.02983) (0.02980)

Has access to informal credit schemes (Not Rosca) 0.15 0.00742 0.00739

(0.36) (0.02223) (0.02222)

Can borrow from friends & family 0.35 -0.06007** -0.06002**

(0.48) (0.02869) (0.02869)

Household size pre-crisis 6.15 -0.22483 -0.22483

(2.61) (0.15767) (0.15763)

Urban or rural 0.53 0.01821 0.01813

(0.50) (0.03065) (0.03064)

N 3800 3800 3800

∗∗∗ p<0.01, ** p<0.05, * p<0.10

Deviations in characteristics from trend are shown to be statistically insignificant and providence

evidence against sorting into birth in anticipation of the crisis.

Page 38

38

Table 2: Birth outcomes: Crisis and non-crisis years

Mean (Standard Deviation)

Birth Cohorts (N= 3800)

Full Sample 1998 Surrounding

Birth outcomes: Mean(sd)

Mean birth weight(grams) 3207.48 3167.3 3211.1

(542.80) (533.99) (543.51)

% Low birth weight(< 2500 grams) 11.4% 16%** 11%

(31.7%) (36.7%) (31.2%)

∗∗∗ p<0.01, ** p<0.05, * p<0.10,

Surrounding years are from 1995 to 2006 excluding the year 1998.

During the crisis, the fraction of low birth weight increased by 5 percentage

points from 11% to 16% with mean birth weight declining by approximately

44 grams. The fact that low birth weight is significant and mean birth weight

is not, suggests that much of the leftward shift in birth weight are for women

with births on the margin of low birth weight.

Page 39

39

Tab

le3:

Dev

iati

ons

ofB

irth

wei

ght

from

Coh

ort

Tre

nd

for

indiv

idual

sB

orn

Duri

ng

the

1998

Indon

esia

nF

inan

cial

Cri

sis

Low

Bir

thW

eight

Bir

thW

eight

PanelA.Full

sample

Bor

n19

980.0

478**

0.0

407*

0.0

413**

-28.4

9-1

9.9

0-2

5.6

2

(0.0

220)

(0.0

221)

(0.0

208)

(32.6

9)

(32.6

5)

(30.7

6)

PanelB.ByBirth

timin

g

Fir

sth

alf

0.0

0616

-0.0

0310

0.0

0532

-4.9

08

1.0

24

-13.6

5

(0.0

279)

(0.0

283)

(0.0

272)

(41.9

8)

(42.4

4)

(41.1

2)

Sec

ond

hal

f0.0

834***

0.0

778**

0.0

718**

-48.5

8-3

7.6

6-3

5.8

0

(0.0

314)

(0.0

314)

(0.0

291)

(45.9

3)

(45.5

8)

(41.9

9)

Con

trol

sfo

rG

esta

tion

No

No

Yes

No

Yes

Yes

Oth

erC

ontr

ols

No

Yes

Yes

No

No

Yes

N3800

3800

3800

3800

3800

3800

Nu

mb

ers

abov

esh

owa

4.1

per

centa

ge

poin

tin

crea

sein

low

bir

thw

eight

for

coh

orts

born

du

rin

gth

efi

nan

cial

cris

is.

Ch

ild

ren

wh

ow

ere

born

inth

ese

con

dh

alf

of

1998

wer

eth

e

wor

staff

ecte

dfr

omth

ecr

isis

sin

ceth

eyfa

ced

the

gre

ate

stn

etin

crea

sein

pri

ces

wh

ile

in

ute

ro(s

eeap

pen

dix

).

∗∗∗

p<

0.01

,**

p<

0.0

5,

*p<

0.1

0,

Page 40

40

Tab

le4:

Dev

iati

ons

ofP

renat

alC

are

from

Coh

ort

Tre

nd

for

indiv

idual

sB

orn

Duri

ng

the

1998

Indon

esia

nF

inan

cial

Cri

sis

Du

rin

gP

regn

an

cy1st

trim

este

r2n

dtr

imes

ter

3rd

trim

este

r

PanelA.Full

sample

Bor

n19

98-0

.00139

-0.0

0105

-0.1

39

-0.1

39

-0.0

502

-0.0

519

0.0

343

0.0

333

(0.0

0966)

(0.0

0940)

(0.0

903)

(0.0

902)

(0.0

985)

(0.0

983)

(0.1

66)

(0.1

60)

PanelB.ByBirth

timin

g

Bor

n19

98F

irst

hal

f-0

.00361

-0.0

0283

-0.1

48

-0.1

40

-0.1

53

-0.1

60

0.1

37

0.1

31

(0.0

142)

(0.0

138)

(0.1

16)

(0.1

16)

(0.1

22)

(0.1

20)

(0.2

39)

(0.2

33)

Bor

n19

98S

econ

dh

alf

0.0

00499

0.0

00455

-0.1

32

-0.1

38

0.0

374

0.0

396

-0.0

530

-0.0

489

(0.0

123)

(0.0

121)

(0.1

25)

(0.1

24)

(0.1

41)

(0.1

41)

(0.2

13)

(0.2

03)

N3800

3800

3708

3708

3701

3701

3693

3693

Con

trol

sN

oY

esN

oY

esN

oY

esN

oY

es

For

pre

nat

alca

re,

nu

mb

ers

ab

ove

show

no

dev

iati

on

from

coh

ort

tren

d.

∗∗∗

p<

0.01

,**

p<

0.05

,*

p<

0.1

0,

Page 41

41

Table 5: Characteristics of ROSCA pre-crisis: sample values

Median Mean Standard deviation

Panel A: ROSCA participation pre and post crisis:

ROSCA participation (Pre-crisis: 1997) - 45% 50%

ROSCA participation (Post-crisis: 2000) - 49% 50%

Panel B : ROSCA attributes pre-crisis(1997):

ROSCA Meeting intervals by weeks 2.71 3.09 2.89

ROSCA payments per month (1997 $) 3.33 7.17 13.10

ROSCA: Months to receive lump sum 11.77 16.40 13.54

ROSCA lump sum payments, previous year ($) 51.56 144.61 1099.91

Observations 2,920 2,920 2,920Payments to ROSCA are about 8-9% of monthly income

Table above provides basic characteristics of ROSCA institutions in Indonesia. In 1997, 45% ofwomen were participating in a ROSCA , met every 3 weeks with a median payment to the schemeof 3.33$. On average women received a median payout of around 51$ from the ROSCA institutionevery 12 months. Payments into ROSCA at each meeting represents 8-10% of monthly income.

Page 42

42

Tab

le6:

Char

acte

rist

ics

ofR

OSC

Alo

ans

pre

and

pos

tcr

isis

Du

rin

gC

risi

s(1

998-2

000)

Pre

Cri

sis(

1993-1

997)

ROSCA

asa

lineofcredit:

Hou

seh

old

s:R

OS

CA

faci

lita

tes

bor

row

ing

8.3

8%

***

0.8

8%

(27.7

1%

)(9

.33%

)

AttributesofROSCA

loans:

Previous12

months(IFLS

2000)

Hou

seh

old

s:b

orro

wed

from

aR

OS

CA

1.8

4%

***

0.4

4%

(13.4

5%

)(6

.61%

)

Fra

ctio

nco

llat

eral

free

99%

-(1

0%

)

Am

ount

loan

ed($

US

D)

49.1

269.0

4(1

01.6

9)

(114.6

2)

Fra

ctio

nfi

xed

len

gth

loan

0.8

40.9

1(0

.36)

(0.2

9)

Loa

nd

ura

tion

inm

onth

s5.7

67.4

5(6

.21)

(7.0

5)

Inte

rest

inm

onth

s3.2

32.6

1(6

.46)

(2.2

5)

Ob

serv

atio

ns

:F

ull

sam

ple

7510

7111

∗∗∗

p<

0.01

,**

p<

0.05

,*

p<

0.10

Tab

leab

ove

show

sra

tes

ofb

orro

win

gfr

omR

OS

CA

sch

emes

pre

cris

isan

dd

uri

ng

the

cris

is.

Nu

mb

ers

show

anex

pan

sion

oflo

ans

thro

ugh

the

cris

isan

dsh

ows

key

attr

ibu

tes

ofth

elo

ans

offer

ed.

Mos

tlo

ans

offer

edw

ere

mad

eco

llat

eral

free

du

rin

gth

ecr

isis

.

Page 43

43

Table 7: Top 10 uses of funds after borrowing from a ROSCAPrevious 12 months (IFLS 2000)

ROSCA funds use Percent frequency Cumulative Frequency

Necessary groceries 44.81 44.81

Education 16.60 61.41

Capital for other business 9.96 71.37

Purchase of household items 6.64 78.01

Farm inputs 4.56 82.57

Social ceremony 4.15 86.72

Other, specify 3.32 90.04

Home improvement 2.90 92.95

Death 1.66 94.61

Marriage 1.66 96.27

Observations 241 241

Greatest use of ROSCA funds for borrowers during the crisis was for

consumption loans. 44.81% of loans were used for necessary groceries.

Page 44

44

Table 8: Difference in Difference Estimates (DID) : Participation in ROSCA and foodconsumption per capita

Food consumption per capita

DID estimates 1997,2000 1993,1996

Has ROSCA * Post 1112.5* 1102.6* -3971.7 -3593.1

(658.0) (665.6) (3068.4) (3057.8)

Post -2358.2*** -2339.6*** -4329.8** -4738.8**

(611.3) (619.7) (2119.2) (2109.2)

Has ROSCA 477.1 465.4 3709.9 3237.3

(663.7) (579.6) (3166.2) (3246.6)

Constant 13681.6*** 14019.3*** 15414.2*** 16342.5***

(574.1) (601.8) (2057.4) (2300.9)

Observations 13713 13713 13286 13286

Controls No Yes No Yes

∗∗∗ p<0.01, ** p<0.05, * p<0.10,

Numbers show above that ROSCA households had larger increases in

food consumption capita relative to non-ROSCA households through

the crisis. Before the crisis however, changes in food consumption be-

tween ROSCA and non ROSCA participants were similar.

Page 45

45

Tab

le9:

Cov

aria

teB

alan

ceac

ross

Tre

atm

ent

and

Com

par

ison

Gro

ups

bef

ore

and

afte

rW

eigh

ting

onth

eP

rop

ensi

tySco

re

Ori

gin

al

Sam

ple

Wei

ghte

dS

am

ple

RO

SC

AN

on

-RO

SC

AS

tan

dard

ized

RO

SC

AN

on

-RO

SC

AS

tan

dard

ized

(n=

1886

)(n

=2273

)D

iffer

ence

(%)

(n=

1886

)(n

=2273

)D

iffer

ence

(%)

Age

24.0

526.4

40.3

7**

26.4

43

26.4

87

0.0

07

Inp

oor

hea

lth

0.0

80.0

80.0

30.0

85

0.0

87

0.0

07

Com

ple

ted

pri

mar

ysc

hool

0.1

60.1

10.1

5**

0.1

06

0.1

15

0.0

29

Com

ple

ted

Mid

dle

sch

ool

0.3

20.2

90.0

60.2

93

0.2

98

0.0

11

Com

ple

ted

Hig

hsc

hool

0.2

50.2

20.0

60.2

19

0.2

22

0.0

07

Com

ple

ted

Pos

tS

econ

dar

y0.2

10.2

90.1

7**

0.2

86

0.2

80

0.0

13

Can

read

and

wri

te0.8

90.9

30.1

4**

0.9

34

0.9

30

0.0

16

Urb

anh

ouse

hol

din

1997

0.4

80.5

80.2

1**

0.5

80

0.5

82

0.0

04

Hou

seh

old

size

6.1

66.1

30.0

16.1

34

6.0

25

0.0

42

Hou

seh

old

asse

tsp

erca

pit

a8.8

69.2

90.2

4**

9.2

85

9.2

85

0.0

00

Hou

seh

old

owns

bu

sin

ess

0.3

70.3

90.0

50.3

95

0.3

96

0.0

02

Hou

seh

old

owns

farm

0.3

20.2

60.1

4**

0.2

59

0.2

68

0.0

20

Has

acce

ssto

form

alcr

edit

sch

emes

0.6

20.6

80.1

3**

0.6

82

0.6

80

0.0

04

Has

acce

ssto

info

rmal

cred

itsc

hem

es0.1

50.1

60.0

20.1

58

0.1

58

0.0

00

Has

acce

ssto

cred

itfr

omfa

mil

yan

dfr

ien

ds

0.3

40.3

70.0

70.3

73

0.3

56

0.0

35

∗∗st

and

ard

ize

diff≥

0.20

,∗

stan

dar

diz

edd

iff≥

0.1

0&<

0.2

0

Tab

lesh

ows

bal

ance

inch

arac

teri

stic

saft

erin

ver

sep

rob

ab

ilit

yw

eighti

ng.

Page 46

46

Table 10: Low birth weight during the 1998 crisis by ROSCA participation

Weighted Symmetric DID

± 1 crisis year ± 2 crisis year

Crisis year * Rosca member -0.104** -0.107** -0.0846* -0.0811*

(0.0498) (0.0468) (0.0478) (0.0448)

Crisis year 0.116** 0.118** 0.0937** 0.0905**

(0.0398) (0.0373) (0.0421) (0.0394)

ROSCA member 0.0106 0.0303 -0.0108 0.00467

(0.0220) (0.0210) (0.0160) (0.0151)

Constant 0.120*** 0.0918** 0.134*** 0.110***

(0.0274) (0.0280) (0.0248) (0.0233)

Overall impact of Crisis year 0.0618** 0.0626** 0.0491* 0.0480*

Observations 1088 1088 1810 1810

Gestation No Yes No Yes

∗∗∗ p<0.01, ** p<0.05, * p<0.10,

DID ≡ Diifference in differences.

Estimates control for seasonality with dummies for quarter of birth