2
MIT LIBRARIES
_
DUPL
r"
3 9080 02898 1444
lA^\i>
^.^1-i^
/I
Massachusetts
Institute of
Technology
Department of Econonnics
Working Paper Series
A
.__-.l
DEWEYl
lilllllilllllllll
a Large-Scale QuasiExperimental Microfinance Initiative
Structural Evaluation of
Joseph
P.
Kaboski
Robert M. Townsend
Working Paper 09-1
December 1, 2008
Room
E52-251
50 Memorial Drive
Cambridge,
MA 02142
downloaded without charge from the
Network Paper Collection at
http://ssrn.com/abstractJd=1 395349
This paper can be
Social Science Research
Digitized by the Internet Archive
in
2011 with funding from
Boston Library Consortium IVIember Libraries
http://www.archive.org/details/structuralevaluaOOkabo
A
Structural Evaluation of a Large- Scale
Quasi- Experimental Microfinance Initiative
Joseph
P.
Kaboski and Robert M. Townsend^
December
1,
2008
Abstract
This paper uses a structural model to understand, predict, and evaluate the impact
of an exogenous microcredit intervention program, the Thai Million
We
program.
model household decisions
and high-yield
uncertainty,
in the face of
Baht Village Fund
borrowing constraints, income
indivisible investment opportunities. After estimation of
parameters using pre-program data, we evaluate the model's
and
ability to predict
Simulated predictions from
interpret the impact of the village fund intervention.
the model mirror actual data in reproducing a greater increase in consumption than
credit,
which
is
interpreted as evidence of credit contraints.
A
cost-benefit analysis
using the model indicates that some households value the program
its
much more than
per household cost, but overall the program costs 20 percent more than the
sum
of these benefits.
*This research
is
funded by
NICHD
R03 HD04776801.
grant
collaboration and data acquisition, and Aleena
Lee,
Audrey
comments.
Light,
We
Masao Ogaki,
received helpful
comments
MIT, lUPUI, Michigan, NIH, Ohio
NEUDC
State,
2006, 2006 Econometric Society,
UC-UTCC.
Adam, Ben
.A.nan Pawasutipaisit,
in
We
thank Sombat Sakuntasathien
for his
Moll, Francisco Buera, Xavier Gine,
Donghoon
Mark Rosenzweig, and Bruce Weinberg
for helpful
presentations at FRB-Chicago, FRB-Minneapolis, Harvard-
UW
BREAD
Milwaukee,
NYU,
2008, and World
Yale and the following conferences:
Bank Microeconomics
of
Growth 2008,
Bin Yu, Taehyun Ahn, and Jungick Lee provided excellent research assistance on this project.
^Email: Kaboski <kaboski.l(3!osu.edu>, Townsend <rtownsen@mit.edu>
Introduction
1
^
r
=
This paper uses a structural model to understand, predict, and evaluate the impact of
an exogenous microcredit intervention program, the Thai Million Baht Village Fund proUnderstanding and evaluating microfinance interventions, especially such a large
gram.
scale
government program,
is
a matter of great importance.
Proponents of microfinance
argue that the unique policies of microfinance enable institutions to bring credit and savings services to underdeveloped areas
to formal financial systems.
is
both
and
to people with otherwise insufficient or
The hope and claim
effective in fighting poverty
and more
no access
that the provision of saving and credit
is
financially viable
than other means, while
detractors point to high default rates, reliance on (implicit and explicit) subsidies,
the lack of hard evidence of their impacts on households.
The few
and
efforts to evaluate the
impacts of microfinance institutions using reduced form methods and plausibly exogenous
data have produced mixed and even contradictory
first
structural attempt to
results.'
To our knowledge,
this is the
model and evaluate the impact of microfinance. Three key ad-
vantages of the structural approach are the potential for quantitative interpretation of the
data, counterfactual policy/out of sample prediction, and well-defined normative
program
evaluation.
The Thai
Million Baht Village fund program
microfinance initiatives of
its
is
kind.^ Started in 2001, the
are run by a committee of villagers
and Khandker (1998), Pitt
and lend to
et al (2003),
program involved the transfer of
in
Thailand to start village banks that
village
members. The transfers themselves
one miUion baht to each of the nearly 80,000 villages
'Pitt
one of the largest scale government
Morduch
(1998),
Moretti (2003), and Karlan and Zinman (2006) are examples.
Coleman
(1999), Gertler, Levine
and
Kaboski and Townsend (2003) estimates
positive impacts of microfinance in Thailand using non-experimental data.
^The Thai program
the rural sector
about S2
through
in
is
billion in
much
involves approximatel}' $1.8 billion in initial funds.
smaller than Brazilian experience in the 1970s, which saw a growth in credit from
1970 to $20.5
village institutions
billion in 1979.
However,
in
terms of a government program implemented
and using micro-lending techniques, the only comparable government program
terms of scale would be Indonesia's
cost of $20 million
This injection of credit into
KUPEDES
village
bank program, which was started
and supplemented by an additional $107 million
in 1987.
in
1984 at a
(World Bank, 1996)
sum
GDP
to about 1.5 percent of Thai
and substantially increased available
We
credit.
study a panel of 960 households from sixty-four rural Thai villages in the Townsend Thai
Survey (Townsend
et al, 1997). In these villages, funds
were founded between the 2001 and
2002 survey years, and village fund loans amounted to eighty percent of new short-term
loans and one third of total short-term credit in the 2002 data.
we count
If
as part of the formal sector, participation in the formal credit sector
village
funds
jumps from 60 to 80
percent.
We
view this injection of
an
credit,
initiative of (then newly-elected,
now) former Prime
Minister Thaksin Shinawatra, as a quasi-experiment that produced exogenous variation
over time and across villages.
More importantly,
the total
The program was unanticipated and
amount
million baht) regardless of the
of funding given to each village
number
Indeed, using
and are often subdivided or joined
GIS maps, we have
was the same (one
of households in the village. Although village size
shows considerable variation within the rural regions we study,
geopolitical units
rapidly introduced.
villages are administrative
for administrative or political purposes.
verified that village size patterns are not related to
underlying geographic features and vary from year to year in biannual data. Hence, there
are a priori grounds for believing that this variation
intervention
is
and the magnitude of the per capita
exogenous with respect to the relevant variables.
Our companion paper, Kaboski and Townsend
(2008), examines the exogeneity
impacts of the program using a reduced form regression approach.
village size
credit
is
not significantly related to pre-existing differences
market or relevant outcome
variables.'^
explicit theory of credit-constrained behavior.
A
perfect credit model, such as a
show indeed that
(in levels or trends) in
After the program, however,
variables are strongly related to village size, and
borrowing and their consumption roughly one
We
many
many outcome
of these are puzzling without
an
In particular, households increased their
for
one with each dollar put into the funds.
permanent income model, would have trouble explaining
the large increase in borrowing, since reported interest rates on borrowing did not
This companion paper also uses reduced form analyses to examine impacts
for general
and
equilibrium effects on wages and interest rates.
in greater detail
fall
and look
Similarly, even
as a result of the program.
income rather than a
loan, they
if
households treated loans as a shock to
would only consume the
interest of the shock (roughly
seven percent) perpetually. Moreover, households were not
after the
program was introduced, despite the increase
investment
is
initially
more
in borrowing.
We
an important aspect of household behavior.
difficult to discern.
credit constraints are
The
structural
deemed
al.,
is
observe increase in the
2008),
we
level of
investment
a priori puzzling in a model with divisible investment,
to plaj' an important role.
model we develop
Gi^'en the prevalence of
appori et
This
household
Finally,
frequency of investment, however, oddly impacts of the program on the
were
likely in default
in this
,;
paper here sheds
income shocks that are not
-
>
light
on many of these
We
findings.
fully insured in these villages (see
start with a standard precautionary savings
1994, Carroll, 1997, Deaton, 1991).
if
model
(e.g.,
Chi-
Aiyagari,
then add important features designed to capture
other key aspects of the economic environment and household behavior in the pre-program
data. In particular, short-term borrowing exists but
borrowing but only up to
tiation of
limits.
payment terms, and
allow households to
size follows
limited,
and so we naturally allow
Similarly, default exists in equilibrium, as does renego-
so our
relatively infrequent in the data but
we
is
is
model incorporates
sizable
make investments
when
it
default.
occurs.
Finally, investment
To capture
this lumpiness,
in indivisible, illiquid, high yield projects
an unobserved stochastic process.^
Finalty,
income growth
is
growth requires writing a model that
component of income,
is
homogeneous
whose
high but variable,
averaging 7 percent but varying greatly over households, even after controlling for
trends. Allowing for
is
in the
life
cycle
permanent
so that a suitably normalized version attains a steady state solution,
gi\dng us value functions and time-invariant (normalized) policy functions. These features
are not only central features of the data, but central to the evaluation of microfinance.
Our approach
is
indeed to estimate the model in an attempt to closely mimic relevant
features of the pre-program data by matching income growth volatility, default rates, sav''An important literature in development has examined the interaction between financial constraints
and
indivisible investments. See, for example, Banerjee
and Townsend
(2001), Lloyd-Ellis
and
Newman
(1993), Galor
and Zeira (1993), Gine
and Bernhardt (2001), and Owen and Weil (1997).
ing/borrowing
levels of
rates,
observed income and hquid assets.
Simulated Moments
many important
els,
and consumption and investment behavior of people with
(MSM)
across 16
moment
estimate 11 parameters using a
conditions.
Method
of
The model broadly reproduces
features of the data, closely matching consumption and investment lev-
and investment and default
less successful,
We
different
probabilities.
and the overidentifying
Nonetheless, two features of the model are
restrictions of the
model are
rejected.
process of the model has trouble replicating the variance in the data, which
the Thai financial
crisis in
The income
is
affected by
the middle of our pre-intervention data, and the borrowing and
lending rates differ in the data but are assumed equal in the model. Using the model to
match year-to-year
fluctuations
is
also difficult.
For our purposes, however, a more relevant test of the usefulness of the pre-program-
estimated model
is its ability
to predict responses to an increase in available credit,
the village fund intervention. Methodologically,
we model the microfinance
namely
intervention as
an introduction of a borrowing/lending technology that relaxes household borrowing
These
limits are relaxed differentially across villages in order to induce
limits.
an additional one
million baht of short-term credit in each village; hence, small villages get larger reductions
of their borrowing constraint.
Given the relaxed borrowing
process to create 500
artificial
limits,
we simulate the model with the
stochastic income
datasets of the same size as the actual Thai panel.
We
then
examine whether the model can reproduce the above impact estimates. The model does
remarkably
well. In particular,
it
predicts an average response in consumption that
to the dollar-to-dollar response in the data. Similarly, the
effects
on average investment
the data.
levels
model reproduces the
and investment probabilities are
.::.:,
,
;,
difficult to
",,.-
much
econometricians. Increases in consumption
of which
we
in their
fact that
measure
in
mask
consid-
treat as unobservable to us as
come from roughly two groups.
hand-to-mouth consumers, who are constrained
close
V-
In the simulated data, however, these aggregate effects can be seen to
erable heterogeneity across households,
is
First, there are
consumption either because they
have low current liquidity (income plus savings) or are using current (pre-program)
liquid-
ity to finance
lumpj^ investments. These constrained households use additional availability
of credit to finance current consumption. Second, households
increase their consumption even without borrowing.
savings, which
for
is less
needed
of these households
credit induces
may
They simply reduce
them
their bufferstock
Third,
to invest in their high yield projects.
we present
for
in order to finance sizable indivisible projects.
such behavior in the pre-intervention data
is
tant motivation for modeling investment indivisibility.) Finally, for households
have defaulted without the program, available credit
loans and so have
may
actually reduce their consumption, however, as they sup-
plement credit with reduced consumption
(Again, the evidence
are not constrained
in the future given the increased availability of credit.
some households, increased
Some
who
may
an impor-
who would
simply be used to repay existing
on consumption or investment. Perhaps most surprising
little effect
that these different types of households
may
all
is
appear ex ante identical in terms of their
observables.
The model not only
highlights this underlying heterogeneity, but also shows the quanti-
tative importance of these behaviors. Namely, the large increeise in
the relative importance of the
first
consumption indicates
two types of households, both of
whom
increase their
consumption. Also, the estimated structural parameters capture the relatively low invest-
ment
rates
and
large
skew
in
investment
sizes.
Hence, overall investment relationships are
driven by a relatively few, large investments, and so very large samples are needed to accurately measure effects on average investment.
The model generates
these effects but for
data that are larger than the actual Thai sample. Second and related, given the lumpiness
of projects, small
on the
We
amounts of
credit are relatively unlikely to
change investment decisions
large projects that drive aggregate investment.
use the model and data for several other alternative analyses. First,
the model and borrov-nng constraint parameters using both
post-intervention data, and
we
(i.e.,
quite similar to a "best
ulate long run predictions
fit"
pooling) the pre-
and
verify that these estimates are quite similar to our baseline
estimates and calibrated borrowing constraints. Hence, the model
is
we re-estimate
model
in the overall data.
we use
to do predictions
Second, we use the model to sim-
and show that measured impacts from reduced form regressions
6
fall
substantially with the
number
we model a
insignificant. Third,
implemented policy but
costs of a direct transfer
The
is.
when they borrow. The
Such a policy has larger
not as large as the effect on consumption.
program that
is
equivalent in the sense of providing the
role,
program vary substantially across households and
are two major differences between the microfinance
program.
is
on investment than the
effects
heterogeneity of households plays an important
benefits of the
use of fund per se
normative evaluation compares the costs of the Million Baht program to the
Finally, our
benefit.
still
statistically
counterfactual policy in which the microfinance funds only
lend to households that invest in the period
not observable, but investment
and become
of years of post-treatment data
First, the
microfinance program
is
same
utility
and indeed the welfare
villages. Essentially, there
program and a well-directed transfer
potentially less beneficial because households
face the interest costs of credit. In order to access liquidity, households
borrow more, and
while they can always carry forward more debt into the future, they are
left
with larger
interest payments. Interest costs are particularly high for otherwise defaulting households,
whose debts
augmented to the more
is
interest charges.
cial
On
borrowing
liberal
the other hand, the microfinance program
than a direct transfer program because
it
and so they bear higher
limit,
is
potentially
more benefi-
can also provide more liquidity to those
who
potentially have the highest marginal valuation of liquidity by lowering the borrowing constraint.
Hence, the program
with urgent liquidity needs
is
relatively
for
more
cost-effective for non-defaulting households
consumption and investment.
Quantitatively, given the
high level of default in the data and the high interest rate, the benefits
transfer) of the
program are twenty percent
this interesting variation
The paper
among
losers
and
less
than the program
(i.e.,
costs,
the equivalent
but this masks
gainers.
contributes to several literatures.
First,
we add a
structural modeling ap-
proach to a small literature that uses theory to test the importance of credit constraints in
developing countries
literature
(e.g.,
Banerjee and Duflo, 2002). Second,
on consumption and
ular. Studies
liquidity constraints,
we
contribute to an active
and the bufferstock model,
in partic-
with U.S. data have also found a high sensitivity of consumption to current
liquidity (e.g., Zeldes, 1989, Aaronson, Agarwal,
and French, 2008), but we believe ours
is
the
first
to study this response with quasi experimental data in a developing country. Third,
methodologically,
and experiments
al.,
2005b,
we
build on an existing literature that has used out-of-sample prediction,
models
in particular, to evaluate structural
Todd and Wolpin,
and interpreting treatment
2006). Finally,
effects (e.g.,
we
(e.g.,
Lise et
al.,
2005a, Lise et
contribute to the literature on measuring
Heckman, Urzua, and
Vytlacil, 2004),
emphasized unobserved heterogeneity, non-linearity and time-varying impacts.
an
explicit behavioral
The remainder
model where
of the paper
is
all
which has
We
three play a role.
develop
'
-
The next
organized as follows.
section discusses the
underlying economic environment, the Million Baht village fund intervention, and reviews
The model, and
the facts from reduced form impact regressions that motivate the model.
resulting value
and policy functions, are presented
and presents the
in Section 3. Section 4 discusses the
GMM estimation procedure and resulting estimates.
data
Section 5 simulates
the Million Baht intervention, performs policy counterfactuals, and presents the welfare
analysis. Section 6 concludes.
2
-
>
.
Thai Million Baht Credit Intervention
The exogenous
institutions
intervention that
we
consider
is
the founding of village-level microcredit
by the Thai government, the Million Baht Fund program. Former Thai Prime
Minister Thaksin Shinawatra implemented the program in Thailand in 2001, shortly after
winning
election.
villages in
One
million baht (about $24,000)
to each of the 77,000
Every
Thailand to found self-sustaining village microfinance banks.
whether poor or wealthy, urban^ or rural was
transfers alone, about $1.8 billion,
The
was distributed
amounts
eligible to receive the funds.
to about 1.5 percent of
GDP
design and organization of the funds were intended to allow
all
The
village,
size of the
in 2001.
existing villagers
equal access to loans through competitive application and loan evaluation handled at the
^The
village
(moo ban)
is
an
official political
sub-district (tambon), district (amphoe),
unit in Thailand, the smallest such unit,
and province (changwat)
levels.
of as just small communities of households that exist in both urban
and
is
under the
Thus, "villages" can be thought
and rural
areas.
village level. For these rural villages, funds
of Agriculture
withdrawal
were disbursed to and held at the Thai
Bank
and Agricultural Cooperatives, and funds could only be withdrawn with a
slip
from the
large (consisting of 9-15
village
Village fund committees were relatively
fund committee.
members) and representative
half
(e.g.,
women, no more than one
per household) with short, two year terms. Residence in the village was the only
member
requirement for membership, and so although migrating villagers or new-
official eligibility
comers would
likely
not receive loans, there was no
official
targeting of any sub-population
within villages. Loans were uncollateralized, though most funds required guarantors. Re-
payment
first
rates were quite high; less than three percent of funds lent to households in the
year of the program were 90 days behind by the end of the second year. Indeed, based
on the household
in the first,
level data, ten
percent more credit was given out in the second year than
presumably partially
There were no
reflecting repaid interest plus principal.
firm rules regarding the use of funds, but reasons for borrowing, ability to repay,
need
for
funds were the three most
common
loan criteria used. Indeed
and the
many households
were openly granted loans for consumption. The funds make short-term loans - the vast
majority of lending
This was about a
average
2,1
is
five
money market
annual - with an average nominal interest rate of seven percent.
percent real interest rate in 2001, and about
percent above the
rate in Bangkok.^
Quasi- Experimental Design of the
As described
in the introduction, the
ways.
it
First,
five
Program
program design was
beneficial for research in
two
arose from a quick election, after the Thai parliament was dissolved in
November, 2000, and was rapidly implemented
in 2001.
None
of the funds
had been
founded by our 2001 (May) survey date, but by our 2002 survey, each of our 64 villages
had received and
^More
details of the funds
^We know
baht on average.'' Households would not have
lent funds, lending 950,000
the precise
and program are presented
month that the funds were
uncorrelated with the amount of credit disbursed, but
the impacts of credit.
.
,
,.
.
in
Kaboski and Townsend (2008).
received, which varies across villages. This
,<
may be an
month was
additional source of error in predicting
.:.
.,
,,-
,,
;
.-.^
anticipated the program in earlier years.
We
therefore
model the program as a surprise.
Second, the same amount was given to each village, regardless of the
size,
so villages with
fewer households received more funding per household. Regressions below report a highly
significant relationship
between household's credit from a
There are strong a priori reasons
for
expecting variation in inverse village size to be fairly
exogenous with respect to important variables of
and
villages are divided or
because inverse village size
is
analysis
is
interest.
merged based on
First, villages are geopolitical
Second,
fairly arbitrary redistricting.
the variable of interest, the most important variation comes
from a comparison among small
form
and inverse village
2002 after the program.
size in
units,
village fund
villages (e.g.,
between 50 and 250 households). That
is,
not based on comparing urban areas with rural areas, and indeed, the reduced
results
below are robust to the exclusion of the few
villages outside of this
50-250
range.* Third, no obvious geographic correlates with village size are discernible from
Reduced Form Impacts
2.2
In the analysis that follows
size as driven
or trends.
we view
by the program
Indeed, this
is
the relationships between outcome measures and village
itself,
first
post-program year, 2002)
**Thus, an}' populist bias
analysis.
result of pre-existing differences in levels
Townsend
,
we run
tambon
which
and "One Tambon-One Product"
level, since
is
a first-stage regression to predict village fund
toward rural areas and against Bangkok
Neither were operated at the village
(2008),
Using seven years of data (1997-2003, so that t=6
is
not likely to contaminate our
Other policies initiated by Thaksin included the "30 Baht Health Plan" (which
at 30 baht per medical visit),
at the
and not the
established in depth by Kaboski and
presents reduced form regressions.
is
GIS
and Townsend, 2008).
plots of village size in our area of study (Kaboski
the
our
the former
(sub-district) level.
10
is
(a
marketing policy
an individual
level
set a price control
for local
products).
program while the
latter
credit of household
._„„
n
in year
VFCRn,t'
t,
1,000,000
Y-
^
1,000,000
^,
# HHs
# ^^^ ^'^ Village^
j^7
+lVFCR^nt + (^VFCR,t + dvFCR,n + ^VFCR.nt
t
Here the crucial instrument
latter
is
is
a vector of demographic household
and Ovfcrx ^nd 9vFCR,n are time and household
Second stage outcome equations
tively.
in levels
and
+'yz^rit
+ 9z,t +
9z,n
Here Znt represents an outcome variable of interest
and household
specific-fixed effectsm respec-
differences are of the form:
# HHs m
dz,n are time
^
inverse village size in the post-intervention years (the
captured by the indicator function It=j), ^nt
controls,
in village^
village^
(
+ £Z,nt
for
household n in year
specific- fixed effects respectively.
t,
and
Although there
geneity across households and non-linearity in the impact of credit, dj
^
is
6z,t
and
hetero-
captures (a linear
approximation of) the relationship between the average impact of a dollar of credit on the
outcome of Znt
The
estimates of Q2,z capture any other relationship between village size and credit
outcome variables that
or
and time
program
years, after controlling for
Znt, including credit disaggregated
rates, income, investment, savings, lending,
and
by source and stated
assets. °
below the rate of Type
1
control in equations
(1)
^
^
'
"An
size.
(2),
^
''
„J'
#„ HHs
These
additional specification that
37 regressions, which
is
Another robustness check includes an additional
errors expected.
and
use, interest
Using a conservative ten percent
level of significance, the Q2,z estimates are only significant in 3 of the
vary by inverse village
household
Using these and other specifications, we analyze a wide range of
fixed effects.
outcome variables
exists before the
*
'
.
.,,
in village^
i,'
which allows
for linear-time
trends to
(
coefficients are significant in only 1 of 37 regressions.^"
we consider included a geographic control variable that captures the
average size of neighboring villages.
'°Over the
fertilizer
full
sample, smaller villages were associated with higher levels of short-term credit where
was the stated reason
They were
for
borrowing, and higher fractions of income from rice and other crops.
also associated with higher
growth
in
the fraction of income coming from wages.
11
Thus, our regressions do not seem to
sufiFer
from pre-existing differences in
levels or
trends
associated with inverse village size.
The
in our
regressions produce several interesting "impact" estimates
companion paper, Kaboski and Townsend
program expanded
fund credit roughly one
village
close to one. Second, total credit overall appears to have
ai^z i^ear one
and there
no evidence of crowding out
is
expansion did not occur through a reduction in interest
though small
as reported in detail
With regard
(2008).^^
for one,
aiz
the
to credit,
with the coefficient
aiypcR
had a similar expansion, with an
in the credit
rates.
market. Finally, the
Indeed the ai^z
is
positive,
for interest rates.
Household consumption was obviously and
The higher
a di_2 point estimate near one.
consumption and
services, rather
significantly affected
level of
by the program, with
consumption was driven by non-durable
than durable goods. While the frequency of agricultural
investments did increase mildly, total investment showed no significant response to the
program.
The frequency
of households in default increased mildly in the second year,
but default rates remained
less
than 15 percent of loans. Asset
levels (including savings)
^^
declined in response to the program, while income growth increased weakly.
Together, these results are puzzling. In a perfect credit, permanent income model, with
no changes
in prices, unsubsidized credit should
simply have an income
effect. If credit
bounded above by the amount
have no
effect,
while subsidized credit would
did not need to be repaid, this income effect would be
of credit injected. Yet repayment rates were actually quite
high, with only 3 percent of village fund credit in default in the last year of the survey.
again, even
of the
if
market
'^The sample
credit were not repaid,
an income
interest rate (less than 0.07),
in
i.e.,
Kaboski and Townsend (2008) varies
necessarily exclude 118 households
who
effect
would produce
at
most a
But
coefficient
the household would keep the principle of
slightly
from the sample
did not have complete set of data for
in this
all
paper. Here
seven years.
we
To avoid
we do not report the actual Kaboski and Townsend (2008) estimates here.
^^Wage income also increased in response to the shock, which is a focus of Kaboski and Townsend (2008).
confusion,
The
increase
is
quite small relative to the increase in consumption, however,
in explaining the puzzles.
We
and so
this has little
promise
abstract from general equilibrium effects on the wage and interest rate in
the model we present.
12
the one-time wealth shock and consume the interest.
The
fact that households
have simply increased their consumption by the value of the funds lent
Given the positive
is
appear to
therefore puzzling.
observed investment, the lack of a response to investment might
level of
point to well-functioning credit markets, but the large response of credit and consumption
indicate the opposite. Thus, the coefficients overall require a theoretical and quantitative
explanation.
2.3
Underlying Environment
Growth, savings/credit, default, and investment are key features
in the
Thai
villages
during
the pre-intervention period (as well as afterward). Households income growth averages 7
percent over the panel, but both income levels and growth rates are stochastic. Savings and
credit are important buffers against
but credit
is
income shocks (Samphantharak and Townsend, 2008),
Income shocks are neither
limited (Puentes, 2008).
smoothed (Chiappori
et al, 2008),
fit
the data better than alternative
mechanism design models. High income households appear
is,
among borrowing
income yield a
nor fully
and Karaivanov and Townsend (2008) conclude that
savings and borrowing models and savings only models
That
fully insured
to have access to greater credit.
households, regressions of log short-term credit on log current
coefficient of 0.32 (std. err. =0.02).
Related, default occurs in equilibrium, and appears to be one
way
shocks. In any given year, 19 percent of households are over three
payments on short-term
(less
than on year) debts. Default
income, but household consumption
is
is
of smoothing against
months behind
in their
negatively related to current
substantial during periods of default, averaging 164
percent of current income, and positively related to income. Using only years of default,
regressions of log consumption on log income yield a coefficient of regression of 0.41 (std.
error =0.03).
Finally, investment plays
hold's income.
It is
an important
lumpy, however.
given year. Investment
is
On
large in years
role in the data, averaging 10 percent of house-
average only 12 percent of households invest in any
when investment occurs and
13
highly skewed with a
mean
and
of 79 percent of total income and a median of 15 percent.
tiie
size of
investment
frequency of investment are strongly positively associated with income, but high
income households
still
poorer households but
trated
Both the
among
invest infrequently.
still
The top income
more often than
tercile invest
only 22 percent of the time. Related, investment
the same households each year.
is
not concen-
the average probability of investing (0.12)
If
were independent across years and households, one would predict that
(1
—
0.88^
=)47 per-
cent of households would invest at least once over the five years of pre-intervention data.
This
is
quite close to the 42 percent that
The next
We
observed.
section develops a model broadly consistent with this underlying environment.
Model
3
is
'
'
:
address these key features of the data by developing a model of a household facing
permanent and transitory income shocks and making decisions about consumption, low
3'ield liquid
savings, high yield illiquid investment
tractability requires that
problem
is
written so that
we make strong
all
and
functional form assumptions. In particular, the
constraints are linear in the permanent
so that the value function and policy functions can
We
all
component of income,
be normalized by permanent income.
do this to attain a stationary, recursive problem.
Sequential Problem
3.1
At
default. In order to allow for growth,
t
-I-
1,
liquid wealth Lt+i includes the principle
pre\'ious period (1
r) St (negative for
-I-
'
-
and
interest
on
liquid savings
from the
borrowing) and current realized income ^f+i:
L,+i=y,+i +
5,(l
Following the literature on precautionary savings
+
(e.g.,
r)
(3)
Zeldes, 1989, Carroll, 1997, Gourin-
chas and Parker, 2001), current income Yt+i consists of a permanent component of income
Pt+i and a transitory one-period shock, Ut+\, additive in logs:
'
Y,+,
=
Pt+^Ut,+^
14
_
(4)
We follow the same literature in modeling an exogenous component of permanent income
that follows a
in this
paper
random walk
is
to allow for
Investment
investment.
(again in logs) based on shock Nt with drift G, but our addition
permanent income to
indivisible
is
also
- the household makes a choice
whether to undertake a lumpy investment project of
Pt+i
Investment
is
also illiquid
and
=
be increased endogenously through
P,G7Vt+i
irreversible,
size
it
{0,
1} of
In sum,
all.
r,
and
permanent income,
sufficiently high to
at a
induce
Having investment increase the
liquidity.
simplifies the
(5)
increases
rate R, higher than the interest rate on liquid savings,
permanent component of future income
or to not invest at
7^*
G
+ RDi^j;
but again
investment for households with high enough
£)/_f
model by allowing us
to track only
While we have endogenized an important
Pt rather than multiple potential capital stocks.^''
element of the income process, we abstract from potentially endogenous decisions such as
labor supply, and the linearity in
R
abstracts from any diminishing returns that
would
follow from a non-linear production function.
Project size
is
stochastic, governed
by an exogenous shock
permanent component of income:
We
is
assume that investment opportunities
I^
y
increase with
consistent with the empirical literature, where investment
large firms invest higher amounts.
and proportional to the
,_
'
Il-^lPt
-:
.
i*
It also
is
•"
^
^
(6)
permanent income
typically scaled
by
This
Pt.
size, i.e.,
ensures that high permanent income alone will
not automatically eliminate the role of credit constraints and lumpiness in investment by
This approach ignores many issues of investment "portfolio" decisions and
risk diversification.
Still,
the lumpy investment does capture the important portfolio decision between a riskless, low yield, liquid
asset
We
and a
risky, illiquid asset,
which
can show this by defining At
At
=
is
already beyond what
Pt/R and using
(3), (4), (5),
(1
+r).
studied
and
-',
in
a standard bufferstock model.
(9) to write:
+ St--{RUt + GNt)At-i+St-i{l+r)-Ct]
Physical assets At pay a stochastic gross return of {RUi
of
is
15
+ GNt),
-
..
/;
.
while liquid savings pay a fixed return
allowing for investment every period.
We
do not observe
this in the data, as discussed in
the previous section.
s of
•
bounded by a
Liquid savings can be negative, but borrov/ing
is
the permanent component of income. That
when
and the more negative
it is,
is,
s is
\
i
-StysPt
-
a multiple
is
allowed,
we
__:
..
,
:
(7)
.
For the purposes of this partial equilibrium analysis, this borrowing constraint
It is
is
the key parameter that
is
.
.
which
negative, borrowing
the more can be borrowed. This
calibrate to the intervention:
limit
is
exogenous.
not a natural borroAving constraint as in Aiyagari (1994) and therefore somewhat ad
hoc, but such a constraint can arise endogenously in models with limited
where lenders have rights to garnish a fraction of future wages
(see Wright, 2002) or
Lochner and Monge-Naranjo, 2008). Most importantly,
which
is
commitment
it
(e.g.,
allows for default (see below),
observed in the data and of central interest to microfinance interventions.
In period
0,
the household begins with a potential investment project of size /q, a
permanent component of income
Pq,
and
liquid wealth Lq
maximize expected discounted
all
as initial conditions.
The
household's problem
is
consumption Ct >
savings St, and decisions Dj^t £ {0, 1} of whether or not to invest:
0,
to
utility
by choosing a sequence of
1-p
V{Lo,Io,Pq;s)
=
max Eq
{Ct>0}
s.t.
0^^
V^
^-^t=0
1-/3
eq. (3), (4), (5), (6), (7),
and
(8)
"
Ct
The expectation
income shocks
[/(,
to one another:
•
Nt
is
+
is
St-VDi^ti;
<
U
.
(9)
taken over sequences of permanent income shocks Nt, transitory
and investment
size
shocks
i*.
These shocks are each
i.i.d.
'
random walk shock
to
permanent income.
16
In A^t
^
A''(0, o"^).
and orthogonal
•
[/f is
•
i* is
If s
<
a temporary (one period) income shock. Ut
an agent with debt,
need to default.
That
investment, equation
to
permanent income).
project size (relative to
0,
is,
(9)
permanent income
<
St-\
i.e.,
with Lt
=
+
Yt
could imply St
(i.e.,
=
<
and a
0,
5'/-i(l
+
In
[/j
Ini*
~
~
A^(0, a^).
N{fj,^,af)
sufficiently
low income shock
may
even with zero consumption and
r),
bad enough shock
sPt. Essentially, given (7), a
a low Nt) can produce a "margin call" on credit that exceeds
current liquidity.
we assume
In this case,
default allows for a
minimum consumption
level that is pro-
portional to permanent income {cPt). Defining the default indicator, Ddef,t S {0,1}, this
condition for default
is
expressed:
;
+
c)Pf
'
L(
(10)
for the period
'
<
otherwise
0,
and the defaulting household's policy
(s
'-
if
1,
D<ieu=
Ct
=^
cPt
St
=
sPt
'
becomes:
- J.
•
.':'.
•;
^-V:
•
^
•'
This completes the model. The above modeling assumptions are strong and not without
costs.
Still,
ical benefits
as
we have
seen, they are motivated by the data,
beyond allowing us
to deal easily with growth.
and they do have analytFirst, the
model
is
simple
and has limited heterogeneity, but consequently has a low dimension, tractable state space
{L, /*,
P} and parameter space
and parameter can be more
{r,
a^,
cr^,
G,
c, /?, p, /ij, ai,
s}.
Hence, the role of each state
easily understood. Furthermore, the linearity of the constraints
in Pt reduces the dimensionality of the state space to two,
sentation of policy functions (in Section 5.2).
which allows
The next subsection
for graphical repre-
derives the normalized,
recursive representation.^'*
Conditions
for
the equivalence of the recursive and sequential problems and existence of the steady
17
Normalized and Recursive Problem
3.2
Above, we have explicitly emphasized the value function's dependence on
s,
since this will
be the parameter of most interest in considering the microfinance intervention in Section
We
drop
this
emphasis in the simplifying notation that
Using lower case variables
follows.
by permanent income, the recursive problem becomes:
to indicate variables normalized^"
= max^^+PElip'y-'vil^i*')]
—
v{l,i*)
5.
1
c,s',di
p
''
(11)
.
J
L
'
"
S.t.
:;
A
\..
c
:
(p
-:
,
-;
,
+ s + dii*<l
s>s
:
-.
from
from
(9)
(7)
..
•
(12)
(13)
,
"
= GN' +
p'
,
,
.:
i>
y
We
(16),
^
y'
=
U'
+
further simplify by substituting
and substituting out
s
/'
Rd,i*
from
(5)
(14)
£ii±!:l
P
from
from
(3)
(15)
and
y'
(4)
into the continuation value using (15)
<
1,
G
and
RE
[i'\
'^Here the decision whether to invest
earlier
states
problem.
I
and
i*.
We
(16)
and
using the liquidity budget constraint (12), which will hold with
state are straightforward extensions of conditions given in Alvarez
In particular, for p
•
-
have denoted
it
:
in
must be
d, is
sufficiently
and Stokey (1998) and Carroll (2004).
bounded.
not a normalized variable and
lower case to emphasize that
>.
18
it
will
is
in fact identical to I?, in
the
depend only on the normalized
equality, to yield:
.i-p
max
v{l, i*
c,
—
1
di
p
+
{I
rj,
''v[U ^
+
,.\-p
,
+15E {p)
I
-
r)il
c
-
dji*)
(17)
,^
s.t.
{I
—
—
c
GN' +
V
The normalized form
dii*)
>
s
(18)
Rd!i*
(19)
of the problem has two advantages.
sionality of the state variable to two. Second,
solution.
consumption
c«
and investment decisions
-p
ic)
_
P{l
+ r)E
(
lowers the dimen-
it
allows the problem to have a steady state
it
Using * to signify optimal decision
First,
rules, the necessary conditions for
optimal
dj^ are:^^
(l
'yp^^m'
+
^)(^-^* -dj,i*)
.,,^
I
+
(20)
p'
.i-p
i-p
v{U' +
ip'
.i-p
+ PE
{1
'\i-p,
+ PE
'\i-p,
V
iP
U'
:i
+
+
+
r)(l
-
c,
- djX)
..,
>
'^
',
r)[/-c., -(l -d;.)r]
.,,
(21)
P'
Equation (20)
non-zero
when
is
the usual credit constrained Euler equation.
the credit constraint (18) binds,
i.e.,
c,
—
I
ensures that the value given the optimal investment decision
value given the alternative,
1
alternative investment decision
—
^^
c^,^,
(i.e., c*,
In practice, the value function
and indeed the
d/.,
—
The
s
c?/,,
—
constraint
dj^i*
(f)
is
only
Equation (21)
exceeds the
maximum
indicates the optimal consumption under this
satisfies
the analog to (20) for
1
—
d/,).
and optimal policy functions must be solved numerically,
indivisible investment decision complicates the computation.^^
Although the value function
is
kinked,
it is
differentiable almost everywhere,
and the smooth expecta-
tion removes any kink in the continuation value.
^^Details of the computational approach and codes are available from the authors
19
upon
request.
Figure
1
presents a three-dimensional graph of a computed value function.
portion at very low levels of liquidity
option.
The dark
line highlights
/
The
fiat
comes from the minimum consumption and default
a groove going through the middle of the value function
surfaces along the critical values at which households
first
decide to invest in the
lumpy
project. Naturally, these threshold levels of liquidity are increasing in the size of the project.
The
slope of the value function with respect to
utility of
consumption increases
as liquidity increases
Figure
2,
panel
A
beyond
/
increases at this point because the marginal
Consumption actually
at the point of investment. ^^
this threshold.
more
illustrates this
clearly
consumption policy as a function of normalized
by showing a cross-section of the optimal
liquidity for a given value of
much
as possible given the borrowing constraint,
borrowing constraint holds with equality. At higher liquidity
longer binding as savings levels s exceed the lower
U, the household chooses to invest in the
lumpy
bound
project, at
and the marginal propensity to consume out of additional
pictured, for
some parameter values
(e.g.,
very high
i?),
is
made,
and hence the
levels, this constraint is
At some
s.
At the
i*.
lowest values, households are in default. At low values of liquidity, no investment
households consume as
falls
no
crucial level of liquidity
which point consumption
liquidity increases.
falls
Although not
the borrowing constraint can again
hold with equality, and marginal increases in liquidity are used for purely for consumption.^^
Panel
'^
B
of Figure 2 shows the effect of a surprise
Given the convex kink
in the value function,
permanent decrease
in s
on the
households at or near the kink would benefit from lotteries,
which we rule out consistent with the idea that borrowing and lending subject to
limits
is
the only form
of intermediation.
"'Using a bufTerstock model, Zeldes (1989) derived reduced form equations for consumption growth,
and found that consumption growth was
significantly related to current income, but only for low wealth
households, interpreted as evidence of credit constraints.
that also contain investment
eis
We
run similar consumption growth equations
an explanatory variable:
In Cn,( + i/Cn,(
=
-Y„,(/?i
+
/^a^'n.t
For the low wealth sample, we find significant estimates
/32
<
+
/?3-^n,(
and
+
/3^
£n,t
>
0,
which
is
consistent with the
prediction of investment lowering current consumption (thereby raising future consumption growth).
20
optimal consumption policy for the same given value of
liquidity levels in every region, except for the region that
i*.
is
Consumption increases
for
induced into investing by more
access to borrowing.
An
<
{st
additional interesting prediction of the model
a household that invests
0),
{dj^t
=
1)
that for a given level of borrowing
has a lower probability of default next period.
Conditional on investing, the default probabihty
further decreasing in the size of invest-
is
Thus, other things equal borrowing to invest leads to
ment.
to
is
consume because investment
maximum amount
than borrowing
less default
increases future income and therefore ability to repay.
of debt that can be carried over into next period
(i.e.,
—sPt)
tionate to permanent income. Because investment increases permanent income,
is
it
The
propor-
increases
the borrowing limit next period, and therefore reduces the probabihty of a "margin call"
on outstanding debt.
One can
(4),
see this formally by substituting the definitions of liquidity (3)
and the law of motion
for
permanent income
(5) into
and income
the condition for default (10) to
yield:
E{Ddeu+i\St,Pt,Dj^t,i:)
=
Pr
Ut+i
<
(s
St
+ c){PtNt+,G
Since St
both
is
is
Dj^t
negative and
and
/*.
R
is
+
(22)
RDj,tin
positive, the right-hand side of the inequality
is
decreasing in
Since both Nt+i and Ut+i are independent of investment, the probability
therefore decreasing in Djj and I^.
Estimation
4
This section addresses the data used and then the estimation approach. The model
is
quite
parsimonious with a total of 11 parameters. Due to poor identification, we calibrate the
return on investment parameter, R, using a separate data source. After adding classical
measurement
9
—
error
on income with
log variance
{r,af^,au,crE,G,c,/3,p,iJ.^,ai,s} via
GMM
21
ge we estimate
,
the remaining parameters,
using the optimal weighting matrix. This
estimation
is
performed using
years (1997-2001) of pre-intervention data, so that
five
corresponds to the year 1997.
Tlie data
ified,
y-'"/:
Data
4.1
:;,:-:.
:--'
'
.
come from the Townsend
clustered,
''",•;
-/^
random sample
Tliai
.,
,
t
=
1
,'..:
'
';•
'..::--..^-
data project, an ongoing panel dataset of a strat-
of institutions (256 in 2002), households (960 each year,
715 with complete data in the pre-experiment balanced panel used for estimation, and 700
2002 and 2003, respectively, which are used to evaluate the model's prediction), and key
in
informants
for the village (64,
villages in four provinces:
Chachoengsao
one
in
each village).
Buriram and Srisaket
in the Central region.
The data
in the
are collected from sixty- four
Northeast region, and Lopburi and
The components used
in this
study include detailed
data from households and household businesses on their consumption, income, investment,
credit, liquid assets
and the
interest
income from these
assets, as well as village
population
data from the village key informants. All data has been deflated using the Thai consumer
price index to the middle of the pre-experiment data, 1999.
The measure
t)
is
of household consumption
we use (denoted
Cn,t for
household n at time
calculated using detailed data on monthly expenditure data for thirteen key items,
and scaled up using weights derived from the Thai Socioeconomic Survey.^° In addition, we
include household durables in consumption, though durables play no role in the observed
increases in consumption.
The measure
of investment
(/n,t)
we use
is
total
farm and business
investments, including livestock and shrimp/fish farm purchases.
We
impute default each year
for
households
who
report one or
more loans due
previous 15 months that are outstanding at least three months. Note that
all
loans,
(1) this
in the
includes
and not just short-term, since any (non-voluntary) default indicates a lack of
available liquidity,
and
(2)
due dates are based on the original terms of the loan, since
changes in duration are generally a result of default.-^ This only approximates default in
'^"The tildes represent raw data which will be normalized in Section 4.3.1.
"^According to this definition, default probability
produce
different results.
The
is
about 19 percent, but alternative definitions can
probability for short-term loan alone
22
is
just 12 percent, for example.
On
the model, and
it
may
underestimate default because of underreporting, but overestimate
model or
default as defined in the
The income measure we
to the extent that late loans are eventually repaid.
use (denoted Yn,t) includes
agricultural, wage, business
all
income (net of agricultural and business expenses) but excludes
financial
on liquid assets such as savings deposits as
in the
model.
Our
interest
and
income
savings measure (Sn^t)
includes not only savings deposits in formal and semi-formal financial institutions, but also
the value of rice holdings in the household. Cash holdings are unfortunately not available.
The measure
less.
of liquid credit (CRn^t)
The measurement
income
credit
in year
t
on savings
the data
is
short-term credit with loan durations of one year or
of interest income
{OW E D _I NT„^t)
"WTiile
is
is
in formal
on
liquid savings
and semi- formal
{E ARN E D _I NTn^t)
institutions.
The
is
interest
interest
owed on
the reported interest owed on short-term credit.
high quality and detailed, measurement error
is
an important concern.
Net income measures are complicated when expenditures and corresponding income do not
coincide in the
same
income fluctuations
be
year, for example. If
will
income
will
error will just
may
also suffer
is
may appear
to
from measurement
error,
but
add additional variation to these endogenous variables
A
not effect the moments, only the weighting matrix.
error for interest
amount of true
income shocks, hence credit constraints may not appear to
be important. Consumption and investment
measurement
error, the
be overstated in the data, and household decisions
less closely tied to transitory
classical
measured with
is
that savings and borrowing
may
the annual flow of both earned and paid interest
the end-of-year stocks contained in the data.
major source of measurement
fluctuate within the year, so that
may
not accurately reflect interest on
This measurement error
will assist in
the
estimation.
Table
1
presents key
the other hand, relabeling
summary
all
statistics for the data.
loans from non-family sources that have no duration data whatsoever as in
default yields a default probability of 23 percent.
Our
higher rates of default.
23
results for
consumption and default hold
for
the
Adjusting the Data for Demographic and CycUcal Variation
4.1.1
The model
is
of infinitely lived dynasties that are heterogeneous only in their liquidity,
permanent income, and potential investment. That
of variation
among households come from
is,
in the
model, the exogenous sources
given differences in
initial liquidity or
perma-
nent income, and histories of shocks to permanent income, transitory income, and project
size.
Clearly,
the data, however, contain important variation due to heterogeneity in
household composition, business cycle and regional variation, and unmodeled aspects of
unobserved household heterogeneity.
Ignoring these sources of variation would be prob-
lematic. For household composition, to the extent that changes in household composition
are predictable, the variance in income changes
predictable changes in household composition.
may
not be capturing uncertainty but also
Likewise, consumption variation
may
not
be capturing household responses to income shocks but rather predictable responses to
changes in household composition. Failure to account
for this
would
likely
exaggerate both
the size of income shocks and the response of household consumption to these shocks. In
the data, the business cycle (notably the financial crisis in 1997 and subsequent recovery)
also plays an important role in household behavior, investment
particular.
Although our post-program analysis
will focus
and savings behavior
on the across- village
in
differential
impacts of the village fund program, and not merely the time-changes, we do not want to
confound the impacts with business cycle movements.
for
example, across households
more about unobserved
We
may
tell
consumption,
us less about past and current income shocks, and
differences in preferences or
therefore follow Gourinchas
Finally, differences in
consumption needs.
and Parker (2002)
in
removing the business cycle and
household composition variation from the data. In particular, we run linear regressions of
log income, log consumption,
since
it
and
liquidity over income.
(We do not take
logs of liquiditj',
takes both positive and negative values, but instead normalize by income so that
24
The estimated equations
high values do not carry disproportionate weight.)'-
where X„,(
number
is
In Yn,t
=
7vX„,t
+ 9Y,j,t + eY,n,t
Ln,t/yn,t
=
Jl'^n.t
+
d L,j,t
+
&L,n,t
In Cn,t
=
Tc^n.t
+
^CJ,(
+
ec,n,f
In Dn,t
=
lD^n,t
+
dD„j,t
+
eD,n,t
a vector of household composition variables
of adult females,
number
are:
(i.e.,
number
of children, male head of household
of adult males,
dummy,
and
linear
squared terms of age of head of household, years education of head of household, and a
household-specific fixed effect) for household n at time
t
and 9
j^t is
a time f-specific effect
that varies by region j and captures the business cycle.
These regressions are run using
only the pre-program data, 1997-2001, which ensures that
we do
program
for the
filter
out the effects of the
Unfortunately, the pre-program, time-specific effects cannot be extrapolated
itself.
post-program data, so we rely on across
the model's predictions.
The
R"^
village, within-year variation to
evaluate
and
values for the four regressions are 0.63, 0.34, 0.76,
0.31, respectively, so the regressions are indeed accounting for a great deal of heterogeneity
and
variation.
For the
mean
full
sample, 1997-2003,
we
construct the adjusted data for a household with
(X and
values of the explanatory variables
6,j) using the
residuals:
,
'
•
.
•
;',
where Qy and
,;
estimated coefficients and
.
lny„t
= %-X + ey^j+gy{t-
1999)
+ eY,n,t
InCnt
= 7cX +
^cj
+
5c(i
-
1999)
+
Dnt
= 1d^ +
^D,j
+
eD,n,t
Qc are the average
''
'
:
,
'
growth rates of the trending
sumption, respectively, in the pre-program data.
ec,„,t
'
variables,
,.
.
income and con-
Next, we use a multiplicative scaling
^^As noted before, 79 of the original 960 households realized negative net income at some point
pre-intervention sample.
The model
yields only positive income,
25
in
the
and so these households were dropped.
consumption, and default are equal in
term to ensure that average income,
liquidity ratios,
the raw and adjusted data. Finally,
we construct investment data
measured investment/income
In principle,
by multiplying the
by the newly constructed income data
ratios (/nt/V'nt)
Returns on Investment
4.2
/„_t
Y-n^t-
i
income growi^h and investment data should
tell
us something about the return
on investment, R. In practice, however, the parameter cannot be well estimated because
investment data
itself is
We
relativelj' infrequently.
we
calibrate
R
to
match
To separate the
instead use data on physical assets rather than investment,
cross-sectional relationship between assets
effect of assets
capital investments are
the
endogenous to current income, and also because investment occurs
made
permanent income, equation
equation
and income.
and labor quality on income, we assume that
prior to investments in physical assets.
year of investing in physical assets. That
first
(5),
J times
and
Let
t
—
all
human
J, indicate
substituting the law of motion for
is,
recursively into the definition of actual income,
(4), jdelds;
J
Ut
+ R E^-^^^-^nL^^
-^'-^
k=i
Ut
Lj=i
income of mvestinent prior
to
t~J
income from investment
The
first
after
t
—J
term captures income from the early human capital investments, which we
measure by imputing wage income from
teristics (sex, age, education, region).
linear regressions of
The second term
wages on household charac-
involves the return
R
multiplied
by the some of the past J years of investments (weighted by the deterministic and random
components of growth.)
We
measure
this
term using current physical
assets.
That
is,
R
is
calibrated using the following operational formula:
£r
We
=
^t
" imputed
labor income^
—R
(physical assetsj
have the additional issue of how to deal with the value of housing and unused land.
Neither source of assets contributes to
V;,
so
we would
26
ideally exclude
them from the stock
of assets. ^'^ Using data on the (a) value of the home, (b) value of the plot of land including
the home, and
the value of unused or community use land,
(c)
we construct
three variants
of physical assets.
We
return.
Townsend Thai Monthly Survey,
use a separate data set, the
The data
is
to calibrate this
obtained from different villages, but the same overall survey area, and
the monthly has the advantage of including wage data used to impute the labor income
portion of total income.
We
us a procedure which
is
The
to zero in the sample of households.
R=
from assets) yields
(c)
R = 0.04, respectively.
R values are identical
households, however.
If
R
is
likely
Not
due
to set the average
sr
baseline value (which excludes categories (a)-
to solve e/j
to our estimates.
R
choose
0.11, while including (c), or (b)
we choose
This
GMM. We
analogous to
=
for
and
(c),
yield
R =
0.08 and
each household, then the median
R
surprisingly,
substantially varies across
because permanent shock histories and
in part
current transitory shocks differ across households, but also in part because households face
different ex ante returns to investment.
Method
4.3
In estimating,
of Simulated
we introduce
Moments
multiplicative
log normally distributed with zero log
is
Lt
is
measurement
error in
mean and standard
income which we assume
deviation as- Since liquidity
calculated using current income, measurement error will also produce
error in liquidity.
We
measurement
-
therefore have eleven remaining parameters 9
=
{r,
G,
ct/v, cr^, erg, c, /?, p, /i,,
a^js}
,
which are estimated using a Method of Simulated Moments. The model parameters are
identified jointly
of the specific
23
Our measure
by the
moments
full set
of
moments.
We
include, however, an intuitive discussion
that are particularly important for identifying each parameter.
of Yt does not include
imputed owner occupied
27
rent.
The
first
two types of moments help identify the return to Hquid savings,
-
'
In £s, St-i
is
= OWED_INTt-rCRt-i
Scr{X,r)
liquid savings in the previous year, while
received on this savings.
previous year, and
Likewise, in
OWED _I NT
:
EARNED _INTt-rSt-i
=
es{X,r)
r
Ect-,
CR
is
^
-
"
-'
,
EARNED _INTt
is
interest
income
outstanding short-term credit in the
the subsequent interest owed on this short-term credit
is
in the following year."'^
The remaining moments
investment decisions, Dj{Lt,
ddef {U'i9)
,
strategy
decisions
ditional
Pt,
where we have now
the parameter set
Our
require solving for consumption, C{Lt, Pt, I^; 6)
is
9.
We
—
9)
/j";
dj{lt, i^; 9),
explicitly
and default
—
Ptc{lt, i^; 0),
=
decisions, D^efiLt, Pt',6)
denoted the dependence of policy functions on
observe data on decisions, Ct,
It,
and states Lt and
Ddej,t,
Yt-
to use these policy functions to define deviations of actual variables (policy
and income growth) from the corresponding expectations of these variables con-
on Lt and
Ij."^
By
the
expectation and therefore valid
Law
of Iterated Expectations, these deviations are zero in
moment
conditions.
With simulated moments, we
these conditional expectations by drawing series of shocks for
error for a large sample, simulating,
[/(,
A^^, i',
calculate
and measurement
and taking sample averages. Details are available upon
•
request.
.
.
The income growth moments
help to identify the income process parameters and are
derived from the definition of income and the law of motion for permanent income, equations (4)
and
(5)."''
Average income growth helps identify the
drift
component
of
growth
income growth, G:
E, {Lt, Yt, Yt+^; 9)
^"'In
the data there are
many
rates on short term borrowing
=
In
{Yt+./Y)
- E [In
is
small, just 2 percent.
Samwick (1997) provide techniques
and au without solving the policy function.
however, since income
is
7,]
low interest loans, and the average difference between households interest
and saving
^^Since L( requires the previous years savings St~i, these
"^Carroll and
[Yt+^/Y) |L„
for
moments
are not available in the
These techniques cannot be directly applied
depends on endogenous investment decisions.
28
first
year.
estimating the income process parameters G,
in
cri\j,
our case,
The
variance of income growth over different horizons {k =1... 3-year growth rates, respec-
tively) helps identify
standard deviation of transitory and permanent income shocks,
cju
and a^, since transitory income shocks add the same amount of variance to income growth
regardless of horizon k, whereas the variance contributed by permanent income shocks increases with k.
role in
The standard
deviation of measurement error
measured income growth. The deviations are defined
^v,k{Lt,Yt,Yt^k]0)
will also play a
strong
as:
]n(Y,.JYA
In(rtW^t)
r
=
ge
1
-E[lniYt+,/Yt)\Lr,Yt]
-,
2
ln{Yt,^,/Yt)
-E
Lt,Yt
-E[ln{Yt+,/Yt)\Lt,Yt]
for
fc
=
1,2,3
We identify minimum consumption,
and
ing
af,
the preference parameters
moments on consumption
/3
c;
and
the investment project size distribution parameters,
p,
and the variance of measurement error a^
decisions, investment decisions,
Focusing on both investment probability and investment
mean
identifying the
(/Zj)
and standard deviation
((7^)
and the
size
us-
size of investments.
should help in separately
of the project size distribution. Fo-
cusing on deviations in log consumption, investment decisions, and log investments (when
'
,
investments are made):
'
^
= Ct-E[Ct\LuYt]
ec{Ct,LuYt-0)
--
we
•
.
-E
eDiDr,t,Lt,Y,;e)
=
Dj,t
ei{Di^tJuLuYue)
=
Duh
are left with essentially three
moment
E[sc]^0
-
]''''''
[Di,t\Lt,Yt]
.,
,
E[Dj,j;\L,,Yt],
conditions for five parameters:
E[6d]
=
E[6j]
=
However, we gain additional moment conditions by realizing that since these deviations are
conditional on income and liquidity, their interaction with functions of income and liquidity
29
fj.^
should also be zero in expectation. Omitting the functional dependence of these deviations,
we express below the remaining
-
E\ec {L,IY^)\
=
E[eD
and
conditions:
{Lt/Y,)]
log investment across all
^
E[ei\nYt]
=
=
E[e, {Lt/Y,)]
model should match average
Intuitively, in expectation, the
ity of investing,
moment
E\eD\nYt\=Q
E[sc\nYt]=Q
:
,
six valid
income and
Q
.
:••
=
log consumption, probabil-
liquidity levels, e.g., not over-
predicting at low income or liquidity levels, while underpredicting at high levels.
moments play
ular. If the
particular roles in identifying
data shows
less
,
measurement error shocks ue and
c,
These
in partic-
response of these pohcy variables to income then predicted, that
could be due to a high level of measurement error in income. Similarly, high consumption
at
low levels of income and liquidity in the data would indicate a high
consumption
c,
default decision
moments
which can be clearly seen from equation
Sdef {Lt,Yt, Ddefj)
In total,
4.4
minimum
c.
Finally, given
s,
level of
we have
16
moments
=
are used to identify the borrowing constraint
(10):
Ddef,t
—
E[Ddef,t\Lt,Yt]
to estimate 11 parameters.
Estimation Results
Table 2 presents the estimation results for the structural model as well as some measures of
model
fit.
The
and savings
interest rate r (0.054)
(0.035),
and
is
is
midway between
the average rates on credit (0.073)
quite similar to the six percent interest rate typically charged
by village funds. The estimated discount factor 3 (0.915) and
elasticity of substitution
(1.16) are within the range of usual values for bufferstock models.
deviations of permanent
cta'
(0.31)
and transitory ay
(0.42)
p
The estimated standard
income shocks are about twice
those for wage earners in the United States (see Gourinchas and Parker, 2002), but reflect
the higher level of income uncertainty of predominantly self-employed households in a rural,
developing economy. In contrast, the standard deviation of measurement error as (0.15)
30
is
much
smaller than that of actual transitory income shocks, and
parameter that
greatly exceeds the average size of actual investments
and there
vs. -1.96),
in the
data (2.50
much
a greater standard deviation
is
vs.
1.22).
(i.e.,
less likely to
log project size
jl^
log h/Yt) in the data (1.47
in project size
ct,
than in investments
In the model, these difference between the average sizes of
and potential projects stem from the
realized investment
are
The average
not significantly different from zero.
is
the only estimated
is
fact that larger potential projects
be undertaken. ^^ The estimated borrowing constraint parameter |
indicates that agents could borrow
up to about
8 percent of their annual
summary
(In the
as short-term credit in the baseline period.
permanent income
statistics of
Table
1,
credit
averages about 20 percent of annual income, but liquid savings net of credit, the relevant
measure,
actually positive and averages 9 percent of income.)
is
consumption
in default
is
The
value of c indicates
roughly half of the permanent component of income.
Standard errors on the model are relatively small.
We
attempt to shed
light
on the im-
portance of each of the 16 moments to identification of each the 11 parameters, but this
not trivial to show. Let e be the (16-by-l) vector of
moments and
metric weighting matrix, then the criterion function
matrix
is
then
is [e'
We] "
2e'W|| =
moment
to
involves 2£'
.
The minimization condition
0.
is
e'We and
W,
the (16-by-16) sym-
the variance-covariance
for the derivative of the criterion function
Table 3 presents ||, a 16-by-ll matrix showing the sensitivity of each
any given parameter. The infiuence of the parameter on the criterion function
W,
which has both positive and negative elements, however
.
Hence, the mag-
nitudes of the elements in Table 3 very substantially across parameters and moments.
is
also not a simple diagonal matrix so that the
moments
are strongly affected by
^'^In
is
(e.g.,
down
the parameters
we
many moments
(e.g., r,
associate with them.
the only parameter in the interest
identified.
G, and
31
Some
/?).
moments play a
In particular, the interest
moments (rows £s and
Ecr)-
While
the model, the average standard deviation of log investment (when investment occurs)
to the 1.22 in the data.
W
income growth and variances),
the partial derivatives confirm the intuition above, in that the
role in pinning
rate r
parameters are jointly
many parameters
while some parameters have strong effects on
Still,
is
is
ajv is
1.37, close
more important
relatively
and
important
CTy is
cv',3),
important
for the variance of
effects
for the variance of
\nY) and
more strongly
L/Y
(rows sc
affected by
*
L/Y
Also, while
/Zj
effect
p,
Sj *
however.)
s
and
£d*L/Y, ei*\nY, and
£-£)*lny",
L/Y)
sj*
has
but also the
[ec * InF, ed * InF,
and
L/Y). (These moments are even
The
parameters
utility function
moments (rows eo-e/
L/Y)
*
sv,2
and
L/Y)
also
but have opposite-signed
c affect default similarly,
liquidity ratios with investment (rows
and, especially, consumption
fit,
(rowsec^lny and
the model does well in reproducing average default probability, con-
sumption, investment probability and investment
levels (presented in
deviations are uncorrelated with log income or liquidity ratios.
the overidentifying restrictions in the model, which
The
tells
Still,
Table
is
2),
we can
us that the model
large J-statistic in the bottom-right of Table 2
First, the estimation rejects that the savings
ments."*
cte
decisions.
In terms of
world.
£1/3),
income growth variance (rows eyj,
on the interaction of measured income and
effects
Y
and
Sv,i)-
e\/2
on consumption and investment moments (rows
a^ also affect
help identify them. Finally, both
*
L/Y, and
*
the investment probability and investment level
ev'.s),
£c
sd
G, P, and
r, aj^,
and
,
£y,2
ev^.i,
and investment decisions with
P and p have the most important
£c-c/.L/y-).
one-year growth rates (row
on the variance of income growi^h (rows
interaction of consumption
£/ *
two and three-year growth rates (rows
is
and indeed
easily reject
not the real
driven by two sets of mo-
and borrowing rates are equal. ^^
Second, the model does poorly in replicating the volatility of the income growth process,
yielding too
We
little volatility.
suspect this
'^^The J-statistic
this as e'e.
is
is
the
the result of the income process and our statistical procedures failing
number
The major elements
all less than
would be straightforward
while the others are
^®It
a kink
in
of households (720) times
of the
summation
e'We.
Since,
W
is
symmetric, we can rewrite
i'i are 0.02 [Ss), 0.02 (scr), 0.03
(ei,,i),
0.04
(£t,,2),
0.01.
to allow for different
the budget constraint, however.
The
effect
borrowing and saving
rates.
This would lead to
would that one would never observe simultaneous
borrowing and saving and there would be a region where households neither save nor borrow. In the data,
simultaneous short-term borrowing and saving
neither savings nor credit
is
observed
in
is
observed
only 12 percent.
32
in
45 percent of observations, while having
1
to adequately capture cyclical effects of income growth, in particular the Thai financial crisis
and recovery
and 1998 (survey years 1998 and 1999,
of 1997
varying volatility
is
respectively).
Only mean time-
extracted from the data using our regression techniques, but the crisis
presumably affected the variance as
well.'^"
Excluding the
from the pre-sample
crisis
is
not
just
one year of income growth to identify both transitory
and permanent income shocks. An
alternative estimation that uses only data from 2000
possible, since
would leave us
it
and 2001, except
1999 data used to create two-year income growth variance moments,
for
produced estimates with wide standard errors that were not
estimates above.
The only economically
constraint (I
=
Thai
even pre-intervention.
villages
—0.25), which
is
from the
statistically different
significant difference
was a much lower borrowing
consistent with an expansion of credit observed in the
Recall that this trend
is
not related to village
size,
however.
Another way of evaluating the within-sample
fit
of the
model
is
to notice that
it is
com-
parable to what could be obtained using a series of simple linear regressions estimating
coefficients (rather
tion, the nine
levels could
than 11 parameters estimated by the structural model).
moments
defined on consumption, investment probability, and investment
would use nine
coefficients.
The two remaining
ply be linear regressions of growth and default on constant terms
On
linear regressions
coefficients could sim(i.e.,
simple averages).
would exactly match the eleven moments that we only nearly
fit.
the other hand, these linear predictors would predict no income growth volatility, and
would have nothing
So the
to
construc-
be set equal to zero by simply regressing each on a constant, log income, and
liquidity ratios. This
These
By
1
make
on the
fit
of the
'
credit.
model are mixed. However, we view the model's
We
ability
its
consider this in the next section.
know from
alternative estimation techniques that the
fluctuations in variables. In the estimation
etc.,
on savings and
policy predictions on the impact of credit as a stronger basis for evaluating
usefulness.
''"We
result
to say about the interest
model does poorly
we pursue, we construct moments
for
in
matching year-to-year
consumption, investment,
that are based only on averages across the four years. For income growth volatitility, the
necessarily have a year-specific component.
33
moments
Million Baht
5
Fund Analysis
This section introduces the MiUion Baht fund intervention into the model, examines the
model's predictions relative to the data, presents a normative evaluation of the program,
and then presents alternative analyses using the structural model.
Relaxation of Borrowing Constraints
5.1
We
is,
incorporate the injection of credit into the model as a surprise decrease in
for
each of sLxty four villages, indexed by
under the million baht fund intervention
dict
v,
we
That
calibrate the new, reduced constraint
as the level for which our
s^''
s.^^.
one million baht of additional credit relative to the baseline at
model would pre-
We
s.
explain this
mathematically below.
Define
first
the expected borrowing of a household n with the Million Baht Fund inter-
vention:
^
mbl
[^n,t,v\^n,t, in,t,Sy
Ell.<o
\Ln,t,
1 n,t
-D,[LuPuIt;s^')i;
and
in the baseline
without the intervention:
E [Bn,t,v\Ln,t, yn,t\s] = E
Lt-C{Lt,Pt.j;;s)
<
J<o
\^n,ti
J n,\
-D,[LuPui;;s)I*
where X<o
is
negative
baht
is
(i.e.,
An
is
for the indicator function that the bracketed expression
borrowing and not savings).
in the first year, so
'^Microfinance
in s.
shorthand notation
we choose
s^''
so that
On
average, village funds lent out 950,000
we would have hypothetically predicted an
often viewed as a lending technology innovation which
alternative would be to
model the expansion of
credit
borrowing, but recall that we did not measure a decline
program.
34
in
is
consistent with the reduction
through a decrease
in
short-term interest rates
the interest rate on
in
response to the
additional 950,000 baht of borrowing in each village in the pre-intervention data
±f-i
AT^I
E [B^lJK^Yn^uS^"]
1
-E[B^,t,v\Lr.,uYr,,t;s]
950,000
^
# HHs
J
^~:
in village.
Here
M represents the number of surveyed households
The
resulting s^'' values average -0.28 across the villages, with a standard deviation
of 0.14, a
program
minimum
ability to
of -0.91
borrow
and a maximum of
is
after the introduction of the
Using the calibrated values of borrowing
(i.e., t
=
Q
and
7)
on
five
limits,
and income data
series of Un,t, Nn,t,
(Ki.s) for year five
inf
and simulate the model
a'^'^
villages, the post-
=
—0.08), averaging
program.
'^'^
for
we evaluate the model's
predictions for 2002
and income growth. Using the observed
liquidity
2001), the last pre-intervention year,
we draw
(i.e.,
measurement error shocks from the estimated distributions,
2002 and 2003.
with the actual pre-intervention data,
We
most
dimensions: log consumption, probability of investing,
log investment levels, default probability,
(Ln^s)
for
Power
Predictive
and 2003
Hence,
substantial relative to the baseline (s
about one-fifth of permanent income
5.2
-0.09.
in the pre-intervention data.
We
do
this
in order to create
500 times, and combine the data
500
artificial datasets.
then ask whether reduced-form regressions would produce similar impact estimates
using simulated data as they would using the actual post-intervention data, even though
statistically the
model
village fund, so rather
is
rejected.
We
do not have a theory of actual borrowing from the
than instrumenting
for village
fund credit, we put
„
HHif^n^vma
e
'
^^^
average injection per household, directly into the outcome equations in place of predicted
'^Since 1999
similar
if
is
the base year used, the 950,000 baht
we use the one
is
deflated to 1999 values.
Predicted results are
million baht which might have been predicted ex ante.
^''These large changes are in line with the size of the intervention, however. In the smallest village, the
ratio of
for
program funds
the 0.83 drop
to village
income
in
2001
is
0.42. If half the households borrow, this
in s.
35
would account
—
village
fund credit. The following reduced form regressions are then:
,,
950,000
Y^
''#
HHs
in village/
=6,7
'
'
'
'
r^
Dn,t
^
«Dj
/
-^
„ TTTT
V--
V
# HHs
in Village^
'# HHs m
''
^.^
^d,j, o:jj, ctoEFj,
grov.'th, respectively.
regression with
regressions, equation (2), in
:.
-
^
,
,_
,
•
''
:
°
'^
^^^
.^'^
j
impact of the
probability, average investment, default probability,
Beyond replacing
village
'
village fund credit
{V FCRn^t) and
^^® above equations differ from the motivating
two other ways.
First,
but recall Section 4.1.1, where we
with household level demographic data.
pre-program data, so the year fixed
as well
els
any
cyclical
We
impact coefficients q^j are now vary
filtered the
data of variation correlated
also filtered year-to-year variation out of the
effects will
post-program years, however, the year
We
,
Second, the regressions above omit the household level controls and household
fixed-effects,
data.
'
and aAinVj would be estimates of the year
program on consumption, investment
program
eD,n,t
950,000
""
.
+
in Village^
j=6,7
by year j
'
Village,,
# HHs
V^
its first-stage
"D,t
950,000
Y^
and log income
+
^
950,000
:
^ti^r
Here Qcj'
^t=j^
T^
'
: ,
'-"-,..
950,000
sr^
=
be zero
for the
pre-program years. For the
fixed-effects will capture the aggregate effect of the
component not
filtered
out of the the actual post-program
run these regressions on both the simulated and actual data and compare the
'
estimates and standard errors.
Table 4 compares the regression results of the model to the data, and shows that the
model does generally quite
well in replicating the results, particularly for
consumption,
investment probability, and investment.
The top panel
presents the estimates from the actual data.
These regressions yield
the surprisingly high, and highly significant, estimates for consumption of 1.39 and 0.90
36
in the first year
is
significant
village
and second
and
positive,
The estimate on investment
year, respectively.
but only in the
For a
year.
first
with the average
village,
fund credit per household of 9600, the point estimate of 6.3e-6 would translate into
an increase in investment probabiHty of
surprising in a world without
sbc
percentage points. Nonetheless, and perhaps
lumpy investment, the
regressions find no significant impact
The impact
investment, and very large standard errors on the estimates.
are significant, but negative in the
first
significant,
efTects
on
on default
year and positive in the second year reflecting
program on
transitional dynamics. Finally, the impact of the
and
probability
log
income growth
is
positive
but only in the second year. Again, given the average village fund credit
per household, this coefficient would translate into a ten percentage point higher growth
rate in the second year.
The second panel
first
of Table 4 presents the regressions using the simulated data.
row shows the average (across 500 samples) estimated
in the
data are typical of the estimates the model produces
probability,
and investment. In
of the coefficient
particular, the
on consumption that
large estimate in the second year.
still
observed.
The model
also finds a
is
model
close to
The standard
comparably sized
probabilities, although its average coefficients are
years,
first
whereas the data show a steep drop
year.
that the estimates
consumption, investment
and
first year,
significant estimate
and a smaller though
errors are also quite similar to
off in the
is
on the investment
significant coefficient
more
what
similar in both the
first
and second
magnitude and significance
after the
predictions for default and income volatility growi;h are less aligned with
For default, both the model and data show a marked and significant decrease
in default in the first year,
though the model's
is
much
significant increase in default in the second year, the
also
is
:.
The model's
the data.
for
yields a large
one in the
and the second row
coefficient
shows the average standard error on these estimates. The main point
The
shows a significant increase
in
income growth
For the alternative definition of default, where
are considered in default, the data actually
all
Wliile the data
model produces no
in the
effect.
'''^
show a
The data
second year, whereas regressions
loans not from relatives with an unstated duration
show a small decrease
37
larger.
in
the second year.
from the model measure no impact on income growth. Perhaps, both of these shortcomings
are results of the model's inability to fully capture year to year fluctuations in the volatility
of the income growth process in the estimation.
The
final
shows the fraction of simulations
It
with both the actual and simulated data
where
for the
in
which a
further note
is
on a sample
test
rare.
Except
for
investment
found at a rate comparable to
outliers can drive results, structural breaks are
the level of significance.
Chow
post-program years finds a structural break
Such occurrences are quite
at a 5 percent level of significance.
One
'
panel shows formally that the estimates from the model are statistically similar
to those in the data.
levels,
'
'
-
'
-
that while the impact coefficients in the data are quite similar
to those in the simulated structural model, they differ substantially from
predicted using reduced form regressions.
For example,
if
we added
what would be
(CRnj) as a
credit
right-hand side variable in a regression on consumption, a reduced form approach might
use the coefficient (say 5i) on credit to predict the per baht impact of the village fund
credit injection.
However,
That
is,
we might
predict a change in consumption of 6i
in the following regression:
reject that 5i
^-^^^^
y°;^^ ^
,
,
=
So-
village^
Parallel regressions that replace credit with
consumption, investment probability, or default also reject this restriction, and these
strictions are also rejected
In sum,
we measure
if
credit
is
•
•
# HHs m
an F-test does indeed
„
re-
replaced with liquidity or income.
large average effects on
consumption and insignificant
effects
on
investment, but the structural model helps us in quantitatively interpreting these impacts.
First, these average coefficients
mask a
great deal of unobserved heterogeneity.
Figure 3 which shows the estimated policy function
manent income)
/.
Again, the
c as
cliff-like
for
consumption (normalized by per-
a function of (normalized) project size
drop
in
i*
and (normalized)
liquidity
consumption running diagonally through the middle of the
graph represents the threshold
level of liquidity that
induces investment.
tions, households in a village are distributed along this graph,
-
Consider
38
In the simula-
and the distribution depends
on the observables {Y and
We
L),
and stochastic draws of the shocks
have plotted examples of
five potential
ante identical in terms of their observables,
households,
Y
and L.
decrease in s can yield qualitatively different responses to the
I's
income
is
U, since
P—
^).
could appear ex
but
their state) constant,
resembles a leftward shift in the graphed decision (recall Figure
Household
and
whom
of
all
(i.e.,
(i*
2,
five
panel B).
A
small
households labeled.
lower than expected, and so would respond to small decrease in s
by borrowing to the limit and increasing consumption. Household
II is
a household that
had higher than expected income. Without the intervention, the household invests and
not constrained in
increases
its
its
consumption. Given the lower
s, it
Household
in the future,
though not investing,
III,
increase consumption without borrowing by reducing
in s.
does not borrow, but nevertheless
consumption. Given the lower borrowing constraint
requires as large a bufferstock today.
its
it
no longer
will similarly
bufferstock given a small decrease
Thus, in terms of consumption, Household I-HI would increase consumption, and
Households
and
II
III
would do so without borrowing.
these households were the only
If
households, the model would dehver the surprising result that consumption increases
than
is
credit,
in default.
but Households IV and
A
small decrease in
s
V
work against
would have no
Household IV
this.
on
affect
its
is
more
a household
consumption or investment,
but simply increase the indebtedness of the household and reduce the amount of credit
that would have been defaulted. Finally, Household
V
is
perhaps the target household of
microcredit rhetoric a small increase in credit would induce the household to invest. But
if
(as
drawn) the household would invest
in a sizable project,
it
would finance
this
by not
One can
also
see that the effects of changes in s are not only heterogeneous, but also nonlinear.
For
only increasing
example,
if
its
borrowing but also by reducing
its
current consumption.
the decrease in s were large enough relative to
i*,
Household
V would
not only
invest but also increase consumption.
Quantitatively, draws from the distributions of
distribution of
3.
The high
L/Y) determine
level of transitory
i*
and
U
(together with the empirical
the scattering of households in each village across Figure
income growth
hence a diffuse distribution in the
L/P
volatility lead to a high variance in
U,
dimension (given L/Y). There are a substantial
39
number
(like
of defaulters in the baseline data, but the changes in s lead to fewer defaulters
Household IV) and more hand-to-mouth consumers
(like
Household
I).
Similarly, the
low investment probability but sizable average investment levels in the data lead to high
estimated
in the
mean and
model have
variance of the
verj' large
i*
Given these estimates, most households
distribution.
projects (with a log
mean
of 6.26), but investment
infrequent (11.6 percent of observations in the model and data).
is
relatively
The median investment
14 percent (22 percent) of annual income in the data (model), so that most investments
is
are relatively small, but these constitute only 4 percent (8 percent) of
the data (model). '^^
In contrast, a few very large
warehouse) have large
effects
i*
on overall investment
investments
levels.
of investments accounts for 36 percent (24 percent) of
Hence, while some households
close
lie
all
(e.g.,
all
investment in
a large truck or a
For example, the top percentile
investment in the data (model).
enough to the threshold that changes
in s
induce
investment, the vast majority of these investments are small. At high levels of i*, the density
of households lying just left of the threshold
is
relatively small.
That
few households
is,
resemble Household V.
Since a lower
on investment
it.
s
can never reduce investment, the theoretical
levels
is
clear.
It
is
effect of increased liquidity
simply that the samples are too small to measure
Given enough households, a small amounts of credit available
whether a
ver}' large
the decrease in
s.
investment
Indeed,
when
is
made
or not,
and
this will occurs
will eventually
more
decide
often the larger
the 500 samples are pooled together, the pooled estimates
of 0.40 (standard error=0.04) for
-]'j
2002 is
highly significant.
Given the average credit injection per household,
this
The estimate
would be an increase
is
also sizable.
in
investment
of 3800 baht per household (relative to a pre-sample average of 4600 baht/household).
^^.^n alternative interpretation of the data
is
that most households do not have potential projects that
are of the relevant scale for microfinance. Households with unrealistically large projects
in
the real world, to households that simply have no potential project
40
in
which to
invest.
may
correspond,
Normative Analysis
5.3
We
evaluate the benefits of the Million Baht program by comparing
As our
liquidity transfer.
its
benefits to a simple
analysis of Figure 3 indicates reductions in s (leftward shifts in
the policy function from the Million Baht program) are similar to increases in liquidity
(rightward shifts in the households from the transfer). Both provide additional liquidity.
The advantage
baht
of the Million
in potential liquidity
(—
Baht program
(s^*"
—
That
£) P).
that
is
provides more than a million
it
(by construction) borrowers choose
is,
to increase their credit by roughly a million baht, but non-borrowers also benefit
from the
More
increased potential liquidity from the relaxed borrowing constraint in the future.
borrow have access to a disproportionate amount of liquidity relative
generally, those that
to
what
ould get
theyvv'
The disadvantage
and hence there are
receives a transfer
the
if
money were
of the Million
interest costs
distributed equally as transfers.
Baht program
uidity.
relative
in s, since they
may
some
for quite
by a marginal reduction
to pay
if it
more
=
r
in s, since the
in
s,
On
consumer
have no need
household
will
II
for
now
and
is less
Households IV, who
time.
interest next period.
from the reduction
is
III,
is
are not locally
than one), benefit
defaulting,
is
little
and
actually hurt
hold more debt, and be forced
since both are locally constrained, in
I
V
and
benefit greatly
consumption and investment
.
s)
to a transfer
is
done by comparing the cost of the program (the
program (an increase
in
/)
that
providing the same expected level of utility (given L^f and
program
who
for liq-
in the current period,
it
the other hand. Households
quantitative cost benefit analysis
reduction in
household that
indefinitely.
respectively.
A
A
0.054.
can be borrowed
Household
3.
their marginal propsensity to
from a marginal decrease
it
provides this liquidity as credit,
importance of these two differences depends on household's need
(i.e.,
not need
it
10,000 baht earns interest on that transfer relative to a household
of, say,
Consider again the household in Figure
constrained
that
which are substantial given
that has access to 10,000 baht in credit, even
The
is
introduced).
That
is,
we
is
equivalent in terms of
Yn,t in
2001, just before the
solve the equivalent transfer T„ for each household
41
using the following equation:
E
The
which
r sf
;
)
|y„,5,., L„,5,„]
=£;[K(L + T„,p,r;
about twenty percent
is
cost.'^^
less
Ten percent
value the program at more than
just 5900 baht.
of households value the
its
program
many households
Thus,
less.
cost (10,100 baht), but the
because of the increased availability of
liquidity,
able to offer the typical household
—
s)
P) =13,400 baht
(— (s^^
the average cost per household in that village
interest costs to households.
at 19,500 baht or
for a
is
more,
28 percent of households
median equivalent transfer
but most benefit
is
much
more
less.
program
Although
liquidity (e.g., in the
household with average income, while
9100 baht),
this benefit
is
swamped
b}^
the
^
Alternative Structural Analyses
5.4
The
just 8200 baht
benefit disproportionately from the
the Million Baht program
village,
„
Again, this average masks a great deal of heterogeneity across households,
in expectation.
median
is
'
than the 10,100 baht per household that the Million Baht
while another ten percent value the program at 500 baht or
is
5) |y„,5,.,L„,5,.]
average equivalent liquidity transfer per household in the sample
program
even
[K(L, p,
-
structural
model allows
for several alternative analyses including
comparison with
reduced form predictions, robustness checks with respect to the return on investment
i?,
estimation using post-intervention data, long run predictions and policy counterfactuals.
We
briefly
5.4.1
Our
summarize the
results here, but details are available
upon
request.
Return on Investment
baseline value of
R
was
0.11.
Recall that two alternative calibrations of the return
on assets were calculated based on the whether our measure of productive assets included
uncultivated or
the
home {R =
community use land {R =
0.04).
We
0.08) or the value the plot of land containing
redid both the estimation and simulation using these alternative
'^This includes only the seed fund, and omits any administrative or monitoring costs of the village banks.
42
For
values.
R =
were quite
0.08, the estimates
similar;, only a higher
/3
(0-94), a lower
r (0.032); and a lower risk aversion (1.12) were statistically different than the baseline.
The model had even more
overall
difficulty
matching income growth and
was substantially worse (J-statistic=200
fit
113 in the baseline).
vs.
regression estimates were nearly identical. For the low value of i?
=
substantially higher (0.97) than typical for bufferstock models.
P be
substantially worse (J-statistic=324).
Finally, the regression estimates
data were qualitatively similar but smaller
first
year.)
(e.g.,
The simulation
estimation
0.04, the
required that the return on liquidity be substantially lower than in the data
that
that the
volatility, so
(r
=
The
0.018)
,
was
fit
and
also
on the simulated
a consumption coefficient of 0.68 in the
Indeed, only the reduction of default in the
first
year was statistically significant
at a 0.05 percent level.
5.4.2
Estimation Using Ex Post Data
In this analysis, rather than use the post-intervention data to test the model using calibrated
borrowing constraints, we use
it
to estimate the
identify the other parameters in the model.
but parametric function for
s„^(,
in the
we add
Sj, $2,
and
s^ are
and better
proceed by specifying a reasonably flexible
UhHs in village,,
the parameters of
additional year-specific
constraints
post-program years:
-'"''•^"-^ "^-2
where
We
new borrowing
moments
for
interest.'^''
Third, for the post-program years,
income growth and income growth
volatility;
consumption, investment probability, investment, and their interactions with measured
income and
liquidity ratios;
and
default. In total, the estimation
now
includes 41
moments
and 14 parameters.
The estimated
results
from the
full
sample are strikingly similar to the baseline
esti-
mates from the pre-program sample and the cahbration from the post-program sample,
If all
all
households borrowed every period and had identical permanent income, then the extra borrowing
per household (950,000/#
HHs
in village^,)
pre-intervention borrowing constraint), s^
=
would translate
——
-k
,
and
43
S3
into borrowing constraints with
=
1.
s-^
=
s (the
with two standard deviation
and
=
I3
—1.17.
bands.'^^
The model
same dimensions and not
standard deviation,
The
resulting estimates are Sj
=
—0.18, §2
—
comparable to the baseline, performing well along the
fit is
well at
all
along the same dimensions.
minimmn and maximum
of
Finally, the average,
implied by the estimates are -0.39
s„^i,„
0.28 in baseline calibration), -0.19 (-0.14), -1.04 (-0.91), and -0.18 (-0.09) respectively.
correlation between the two approaches one by construction, since both increase
ically
with village
That
size.
is,
—46,
the estimated s^j
down by roughly
values, except that they are shifted
^
(-
The
monoton-
are quite similar to the calibrated
0.1.
This
is
not particularly surpris-
ing since time-specific fixed effects for the latter years were not taken out, and
we know
that the model had difficulty matching year-to-year fluctuations. For example,
consump-
tion in
all
villages
was high
in
2002 and 2003, and so the estimation wants to
borrowing constraint. Nonetheless, the
fact that the estimates
fit
a lower
and calibrated values are
otherwise quite close indicates that cross-sectionally the simulated predictions of the model
on average approximate a best
to the variation in the actual data.
Long Run Predictions
5.4.3
The
fit
between azj estimates
differences
in the first
program indicate that impacts are time-varying,
and second year
(i.e.,
We
—
since there are transitional
households approach desired bufferstocks. The structural model allows
longer run horizon estimates of impact.
j
1,2) of the
dynamics as
for simulation
and
therefore simulate datasets that include five
additional years of data and run the analogous regressions. Seven years out, none of the
azj estimates
are statistically significant on average.
Wliile the average point estimates
are quite small for investment probability (0.23), investment (0.10),
ity (0.01) relative to the first year, the
and
(0.58)
and default probabil-
average azj for consumption remains substantial
close to the estimate in the second year (0.73).
In the model, the impacts
'^For comparison, the point estimates of the fuh-sample (basehne) estimation are f
cr^'
=
(0.926)
0.24 (0.31),
,
p
=
ay =
=
0.42 (0.42), c^e
1.17 (1.20), A,
=
1-35 (1.47)
,
a,
0.38 (0.15),
=
G =
2.70 (2.50)
44
,
1.06 (1.047), c
and |
=
=
=
0.061 (0.054),
0.49 (0.52),
-0.07 (-0.08).
on
/3
=
0.928
consumption
somewhat
fall
effect. Still, alternative regression
post-program years
consumption
in
estimates that simply measure a single
az do not capture any
j) coefficient
when seven
remains a substantial persistent
after the first year, but there
(common
statistically significant
for all
impact on
years of long run data are used. This shows the importance of
considering the potential time-varying nature of impacts in evaluation.
Policy Counterfactual
5.4.4
From
the perspective of policymakers, the Million Baht Village
problematic along two fronts.
investment, and
may simply go
households.
it
appears
An
less cost-eff'ective
money
than a simple transfer mainly because funds
and the increased borrowing
limit actually hurts defaulting
would be to only
alternative policy that one might attempt to implement
We
allow borrowing for investment.
but since
most discernible impacts are on consumption rather than
Its
to prevent default
Fund Program may appear
is
fungible,
it
would assume that the
village
can observe investment,
would be unclear whether these investments would have been
undertaken even without the loans, in which case the loans are really consumption loans.
Since defaulting households cannot undertake investments,
default from borrowing. Nevertheless, such a policy
Household
The
In a
I
in Figure 3
ability to
model with
would
it
would prevent households
also eliminate households like
x
from borrowing.
model policy counterfactuals
is
;•
another strength of a structural model.
this particular policy, households face the constraint
period in which they decide to invest, while facing the baseline
invest.
The
default threshold
is
from investing and borrowing
also
in
moved
to
s^''^"-^ternauve
^
The
is
^^^
they decide not to
new borrowing
constraints are even lower
from -0.16 to
-4.78.
who
borrow.
The new
.,.",.
policy increases both the impact on consumption and increase the impact on in-
vestment. Pooling
that
^^
however, to prevent households
(averaging -0.67 vs. -0.28 in the actual pohcy) but only for those
is
s if
s^'',aiternative
one period, and then purposely not investing in the next
period in order to default. Under this policy, the
range of borrowing constraints
in
all
500 simulated samples yields a significant estimate
similar to the actual million baht intervention (1.40 vs.
45
.-''
for
1.38 in the
consumption
first
year).
It
also yields a
year)
.
much
and
larger
significant estimate for investment levels (0.62 in the first
Clearly, the counterfactual policy channels funds only to investors
relax borrowing constraints
much more
substantially for investors,
large investments but relaxes the borrowing constraint even
than the actual
policy. Finally, the negative
more
households
variation in the benefits across households
equivalent transfer
is
We
(e.g.,
some
able to
with
households
Although
exists.
rather not invest,
There
of this loss.
is
the standard deviation of the
14,000 baht in this counterfactual vs. 11,000 in the baseline policy),
but the average equivalent transfer
6
for investing
who would
the benefits are larger to defaulters and investors help outweigh
it is
in turn to help
impact on default no longer
this policy offers less flexibility for constrained
much more
and
and so
is
actually lower (7500 vs. 8200).
'
;
Conclusions
have developed a model of bufferstock saving and indivisible investment, and used
to evaluate the impacts of the Million
prediction of consumption increasing
"smoking gun"
it
Baht program as a quasi-experiment. The correct
more than one
for
for the existence of credit constraints,
one with the credit injection
and
is
is
a
strong support for the impor-
tance of bufferstock savings behavior. Nevertheless, the microfinance intervention appears
to be less cost effective
on average than a simpler transfer program because
households with interest payments.
it
saddles
This masks considerable heterogeneity, however,
cluding some households that gain substantially. Finally,
we have emphasized the
in-
relative
strengths of a natural experiment, a structural model, and reduced form regressions.
One
limitation of the
model
is
investments, modeled through R,
the data,
R
that although project size
is
is
stochastic, the quality of
assumedconstant across projects and households. In
varies substantially across households.
Heterogeneity in project quality
be an important dimension
for analysis, especially since microfinance
position of project quality.
The
potential projects
may be
may change
may
the com-
process for project sizes was also extremely stylized. Also,
arrive less often, be less transient (which allows for important
anticipatory savings behavior as in Buera, 2008), and multiple projects
46
may be
ordered by
their profitability.
Such extensions might help explain the gap between positive predicted
impacts on investment probability, but no impact in the data.
Related, the analysis has also been purely partial equilibrium analysis of household
behavior.
we have emphasized heterogeneous impacts
Wliile
across households,
it
may
be possible that even a given relaxation of credit constraints would have heterogeneous
impacts across
villages.
For example, with a large scale intervention, one might suspect
that general equilibrium effects on income, wage rates, rates of return to investment, and
interest rates
we
on
liquidity
may be important
(see
Kaboski and Townsend, 2008). Finally,
did not consider the potential interactions between villagers or between villages, nor
was the intermediation mechanism
explicitly modeled.
These are
all
avenues for future
research.
References
[1]
Aaronsen, D., Agarwal,
Wage
[2]
Increases."
FRB
and French, E. "The Consumption Response to Minimum
Chicago working paper 2007-23, May, 2008
Ahlin, C. and Townsend, R. "Using
Liability Lending."
[3]
S.
Repayment Data
Aiyagari, S. R., "Uninsured Idiosyncratic Risk
Alvarez, F.
Models of Joint
mimeo, 2000
and Aggregate Saving." Quarterly Jour-
nal of Economics, 109 {August 19M): 659-684
[4]
to Test Across
.,
.
and Stokey, N. "Dynamic Programming with Homogeneous Functions",
Journal of Economic Theory, 82 (1998): 167-189
[5]
Aportela, F. "Effect of Financial Access on Savings by Low-Income People." mimeo,
Massachusetts Institute of Technology, 1998
[6]
Banerjee, A. and
Newman,
.
.
'
A. "Occupational Choice and the Process of Development."
Journal of Political Economy,
(^
April 1993/' 274-98
47
[7]
Banerjee, A. and Duflo, E. "Do Firms
Want
to
Borrow? Testing Credit Constraints
Using a Direct Lending Program." mimeo, Massachusetts Institute of Technology,
2002
[8]
.
.
Blossfeld, H.
and Sha\dt,Y. Persistent
Westview
in Thirteen Countries. (^Boulder, Co:
[9]
Buera, F.
J.
Inequality:
.
Changing Educational Attainment
Press), 1993.
"Persistency of Poverty, Financial Frictions, and Entrepreneurship."
mimeo, Northwestern University, 2006
[10]
May
Carroll, C. D. "Buffer Stock Sa\ang
-
and the Lifecycle Permanent Income Hypothesis."
Quarterly Journal of Economics, 107 (1) 1997
[11]
Carroll, C. D. "Buffer Stock Saving:
"
University 2004
[12]
Chiappori,
P.,
Carroll, C. D.
Schulhofer-Wohl,
(1)
Coleman, B. "Microfinance
Asian Development Bank
[15]
S.,
,
Samphantharak, K. and Townsend, R. "On Ho-
Thai Villages." manuscript, January 2008
in
and Samwick, A. A. "The Nature of Precautionary Wealth." Journal of
Monetary Economics, 40
[14]
_
.
mogeneity and Risk Sharing
[13]
Some Theory." Manuscript, The Johns Hopkins
1997
in
;
Northeast Thailand: Wlio Benefits and
ERD
Working Paper
Coleman, B. "The Impact of Group Lending
in
How Much?"
Series No. 9, 2001
Northeast Thailand." Journal of De-
velopment Economics, (October 1999): 105-141
[16]
Galor, O. and Zeira,
nomic
[17]
J.
Studies, 60 (1 1993): 35-52
Gertler, P., Le\dne, D.
ilies
"Income Distribution and Macroeconomics." Review of Eco-
and Moretti, E. "Do Microfinance Programs Help Insure Fam-
Against Illness?" Center for International and Development Economics Research
Paper C03-129 (February 2003)
48
[18]
Gine, X. and R.
Townsend "Evaluation
of Financial Liberalization:
A Model
with
Constrained Occupation Choice." mimeo, University of Chicago, 2001
[19]
Gourinchas, P.-O., and Parker,
J.
"Consumption over the
Life Cycle." Econometrica,
71 (January 2002) .-47-91.
[20]
Heckman,
ables in
J.
J.,
Urzua,
S.,
Vytlacil, E., 2004.
"Understanding Instrumental Vari-
Models with Essential Heterogeneity". Stanford University, Stanford, CA.
http://www.stanford.edu/~vytlacil/
[21]
Kaboski,
J.
and Townsend, R. "The Impacts of Credit on Village Economies." mimeo,
2008, http://kaboski.econ.ohio-state.edu/impactofcredit.pdf
[22]
Kaboski,
J.
and Townsend, R. "Borrowing and Lending
in
Semi-Urban and Rural
Thailand." mimeo. University of Chicago, 1998
[23]
Kaboski,
J.
and Townsend, R.
"Policies
and Impact: An Analysis of Village Microfi-
nance Institutions." mimeo, University of Chicago, 2003
[24]
Karlan, D. and Zinman,
Decisions
[25]
J.
"Expanding Credit Access: Using Randomized Supply
To Estimate the Impacts." mimeo, August, 2006
Lloyd-Ellis, H.
•
and Bernhardt, D. "Enterprise, Inequality, and Economic Develop-
ment." Review of Economic Studies, (January 2000): 147-68
[26]
Lise, J., Seitz, S.,
and Smith,
of Social Programs."
[27]
Lise,
J.,
Seitz,
S.,
mimeo,
J.
"Equilibrium Policy Experiments and the Evaluation
April, 2005
and Smith,
J.
.
.
.
"Evaluating Search and Matching Models Using
Experimental Data." mimeo, April, 2005
[28]
Lochner, L. and Monge-Naranjo, A. "The Nature of Credit Constraints and
Capital."
NBER working
paper W13912, April 2008
49
,
Human
[29]
Owen, A. and Weil, D. "Intergenerational Earnings
NBER
[30]
Paulson, A. and Townsend, R. M. "Entrepreneurship and Liquidity Constraints in
Pitt,
M. and Khandker,
Households
Pitt,
S.
in Bangladesh:
.
.
"The Impact of Group-Based Credit Programs on Poor
Does the Gender of Participants Matter?" Journal of Po-
Economy, (October 1998): 958-96
litical
[32]
and Growth."
Working Paper 6070 (June 1997)
Rural Thailand." mimeo, 2000
[31]
Mobility, Inequality,
M., Khandker,
S.,
Chowdury, O. and Millimet, D. "Credit Programs
Poor and the Health Status of Children
in
for the
Rural Bangladesh." International Economic
Review, 44 (February 2003): 87-118
[33]
Puentes, E. Thai Project Note, 2008
[34]
Rosenzweig, M. and Wolpin. K. "Natural 'Natural Experiments' in Economics." Journal of
[35]
Economic Literature 38 (December
2000): 827-874
Samphantharak, K. Internal Capital Markets
in Business Groups. Ph. D. thesis, Uni-
versity of Chicago: Chicago, 2003
[36]
.
.
Samphantharak, K. and Townsend, R. ""Households as Corporate Firms: Constructing Financial Statements from Integrated Household Surveys" mimeo, 2008
[37]
Todd,
P.
and Wolpin, K. "Assessing the Impact of a School Subsidy Program
in
Mexico: Using a Social Experiment to Vahdate a Dynamic Behavioral Model of Child
Schooling and
[38]
Towmsend,
Fertilit}'."
American Economic Review, 96 f December 2006): 1384-1417
R., principal investigator with Paulson, A., Sakuntasathien, S., Lee, T. J.,
and Binford, M., Questionnaire Design and Data Collection
Insurance and the Family and
NSF
grants.
50
The
for
NICHD
Grant Risk,
University of Chicago, 1997
[39]
World Bank "KUPEDES: Indonesia's Model Small Credit Program."
Number
[40]
104,
OED
Precis,
February 1996, 4 pages
Wright, M. "Investment and Growth with Limited Commitment." mimeo, October
2002
[41]
Zeldes, S.
"Consumption and Liquidity Constraints:
Journal of Political Economy, 97 (March 1989): 305-46
51
An
Empirical Investigation."
w
s
ID
o
g
O
3
3
O)
(a) Ajiiiin 3LU!Ja;n
papadxg
amooui jusuBuuad/uoTjduinsuoo
91UO0UI jusuBuusd/uouduinsuoo
'o
A',
^*.
^^
•
*IS^
--
a
-C
».
•
r
a>
3
I
•aO
I
III
I
ll1llli.lllll:ll iiiC.
•3
o
.jssMiP**''
e
o
O
00
>o
^
c-a
-^
(d/D=') 3UI03UI }n3UEUU9<j/uopdiDnsuo3
i
•3
(X,
U
fc
((j/0=o)
amooui ;n3UBUU3(i/uoi4duinsuo3
x
o
o
>n
u->
IT)
o
o
en
o
"^ o
roo
(N
(N
VO
--
O
O
t^
--
W)
^-<
—
>yn
—
•*
CO
r-
o
O O O O O
(N
r^ vD
OS
'
^-i
o-
r<-)
*
o
o
o
o
o
o o
o ON
•* -o o
o
^^ oo ^^ oo rn
(N o o
o
c^
^•^
ooooooooo
—
o
_
^
—
o
o
o o
m
Q oo
ON
O ^
lO
•3
QJ
o-^ooooooo
o
m
ON
«
«
§
2
"o
.a
a
in
3
o
a:
B
o -^ o lO o o o o
o 00
Ov o rn o ON o o o o
o
o
o
(N rn — ON o
m rn rrM O m o O o
to
ON
ro
o
^^
fN
>
u
CnI
^~->
Q
On
lO
On
(N
iri
rO o
CnI
^^
ra
—
^
^
o
o
-a
•d
CO
^
O-^OCnIO-^OOOO^
OOO — N00nOOOO>0
_o
s
u
>
M
o —
o
oo
_z
^
^
•^
ON
in
m
r^
in
m
in
r^
in
iTi
rNi^rN)f-^t^-^ONmor~- —
rt
r--
CnI
j3
S3
B
i
L.
ioooinuni/^ooi/~iir-)00>/nio
mtNmrnmtNmmfNrnm
Cm
;«
in
r--
in
rn
ID
m
in in in
m m m m
ly-i
r~-
t-~
t--
o
t«
t«
o
(L>
>
5
o
CL
CU
BON
ON
Ov
-2
in
b
S
.o
s
(U
o
u
tn
3 o
O
>
Ov
ON
ON
•x
3
o O
,13
(U
l4-l
OJ
CO
es
c
o
o
c
H
CQ
TO *
12
O
o
o
O
>
O
s < ^
^
3
o
=
o
5o
:^
dj
*
^ O
2^
o
^
3
P
fc
1-
3
*
u
E
t1-
^
-s:
3:
C3
TO
> Q ^ 5 J
=
-S -a
§--§
<
O ^
Si*^ -I -o
o o «
it!
-2
-'-
^ c
O
*-
^
p
g
o
<<
-J
^
2P
.=
a
— > < > g
o >> H " ^
_4;
.2
2o
<2 *
*
$
TO
r.
a:
^ ?
o
Z
^
-
e:
O
-
Xi
^
E
~
E
~,
I*-
o
^
P
3
•I
^
o
O
Z Z Z Q
re
I/)
^
>
p
P
3
X)
_3
>
OJ
>-
OX)
<
>
0^
"O
o
IWD
03
^
S
0^
^ o On
§
oo -*
§ ^ o
o
[Co-* o o
^
CN
o >
U
9i
>
CD
^
•
s
o
M
WD
c«
-^^
C5
Q
-^
t^
o "^ o ^ 2:
^ o ^ d d
S
-(—
."2
o
s
a
s:
o
<
fi
U
Plh
O
^
2
«
.2
r
V3
w
'b^
fa
U Q
u
S,
.o
>
a
•
in
CO
>
H
^^
^^
(«
w
u
<v
a>
_H r^ n- ^ ^ "^ ^ ON
O O g
o
g
g
o o
o :^
o S
d
d
o
o
o
o
^
^
^
<^
.
2
^
.
.
en
00
o
O
oo
.
e«
W
2
(N
DDbO
wioao-i
o
I
(/3|
i/3|
6"
=L
o o
o
^ o
o o o
o o
o o
lO
o o
o o
00
_oo
50
1—t
Os
1
o
1
<^
<?
^
=?
1
"^
r^
'^ O O
o 9 o o
,—
I
(^
o o o "^ m
o o o o _; o 9
r*-)
rj
1
'-J
o o o ^H
o o o o
o o
1
^ (N o
pg ^
o O O O o O o
0^
-^
— O
I
o o
rsi'^o<^'-:'^oo'sf^^^f~^^o
o
o^^^l^-^'-Hf~~:o^'*<^oo-^'^
oo OfNdfNO—
;d9—d9dd'^
"do
u
<^.
a.
S
U
.2
ei
I
5^^
S^
at
2
s
2
t;
.^
==>
r-l
r\i
I/S
•
-
<r<-)Q.^<3,(N^^<;^(N\^iy-,(^
«uOO oooo'^'—
99'T<~^'9dd9dd9ddd
s
S
^o
..
O
^ -^ ^
oo O rn <N
m
d
d
T^9o'-;Ji/Sd^M2?jq2=^==^
d
d
d
&Ht»dd
'
.
>/^^'-;oo—if^o°°.r)^^<^.-H'^
300
d d 9^'Tdd9d9<~^f^Jd9'^
b
I
,>
«
«
.2
zoo >r)-ironoo^-^f^'^^9;Q»oo
H9dOo^J9?2l?5^2?5^d
bdd
—
...
^^
"J^
--^
I
(N
u
CnI
=-^2
o__d^^o-;r.^ocgc^
c«
"
c^
"
w
w
w
" *y ^„ "
W CO
SJU3UI0J\[
--^
^^
w
w"
"
r r
w
W
««
1
>
c
.o
V£)
*+-»
o
I
1)
a»
o
u
C3
o;
^
^
a>
ON
o
:?
in
4-«
o
E
o
u
s
Q
On
Hg
Qn
fS
4^
O
^
a
?^
"a
o
(U
1)
03
O
"3
(D
-C3
3
o
«
^
M3
>.
0)
^
s
E
3
u.'
^o
u
I
I
ON
^
CI
c
CU
u.
a
15
-o
:g
a>
^
^V
oV
a
>-
1)
(N
1
4>
ID
rn
O
ri
CO
a
D
vo
fN^
_C
•o
>^
(N
o
o
vo
VO
ou
fo
'o
eD
03
J3
^
O
o
^
0^
ON
'
a>
•T3
0)
x;
*-*
o
2
u
*-»
J3
4-'
w
>
_a
-H
(U»
en
s
H
as
in
o
o
00
J2
'>
00
a*
o
c^
o
(L>
a
a;
>
&
O
o
>-
-^
3
o
0)
en
!-.
o
OJ
3
O
J3
tM
I
1)
o
03
s
3
03
rt
01
01
s
o
_o
(S
3
a
o
u
O
ON
O
o
r
1
ON
rn
-H
00
On
fn
o
o
f
o
Q
55
Q
s
4-*
u
o
0>
*
£
S
r"'
'o
ac
C3
5?
2
<
o
Ml
yi
<
O.
.S
03
<
(U
«
dj
o
3
o
g
rt
tn
03
0)
J3
U
en
3
O
J=
+-
C/1
.S5
r-"
*
a.
*
's
.SP
en
j3
•±3
2o 1 213
x:
H
*
o
3
O
03
(D
>
^o
O
03
-o
D.
3
03
*tn
l+H
-*-»
aj
'i
C
O
O
o
m
-4-»
,^j
o
^
13
aj~
.2i
tu
o
3
5
3
crt
rt
4-»
4—1
en
O
o
M
*
>
a3
;-
(L>
o
^
r-*
^
c
3
"in
o
s
o
3
aj
(L)
O
-4—
JS
(U
*"*
3
3
o
<s
t/3
r^
00
0\
(^
_a>
w
O
o
o
o"
o
o^
in
so
03
-a
T3
"5?
_C
en
H
Q-
V5
X3
1
H
u
o
o
-4-<
(U
O
-S
_aj
Uri
01
S
o
03
O
in
e
>
n.
in
-o
"O
-a
5^
.S
'c^
en
-^
aj
en
Q.
(U
0>
u
;2
"B
"o