international conference on rural finance …...international conference on rural finance research:...
TRANSCRIPT
International Conference on Rural Finance Research:
Moving Results into Policies and PracticeMoving Results into Policies and PracticeMoving Results into Policies and PracticeMoving Results into Policies and Practice
FAO Headquarters
Rome, Italy 19-21 March 2007
Field Experiments in Rural Finance: An Example Field Experiments in Rural Finance: An Example Field Experiments in Rural Finance: An Example Field Experiments in Rural Finance: An Example
from Tamil Nadu, Indiafrom Tamil Nadu, Indiafrom Tamil Nadu, Indiafrom Tamil Nadu, India
by Michael Faye and Sendhil Mullainathanby Michael Faye and Sendhil Mullainathanby Michael Faye and Sendhil Mullainathanby Michael Faye and Sendhil Mullainathan
This paper was chosen through an open call for research in rural finance, whereby the
selected individuals were invited to Rome, Italy, to share their results during the
conference and to discuss key issues in shaping the rural finance research agenda as well
as ways of strengthening the ties between research, policy and practice.
Field Experiments in Rural Finance: An Example from Tamil Nadu, India
Michael Faye and Sendhil Mullainathan1
Introduction: Reasons to Conduct Randomized Controlled Trials (RCTs)
Suppose you’re in charge of the rural finance portfolio at a private bank in a developing
country. You’re reviewing the performance of your rural branches, and would like to
consider some policy changes. Some staff members propose new products, while others
suggest changing interest rates and the borrowing limit. Remembering your old
economics courses, you realize that while raising the interest rate increases margins,
higher interest rates can also decrease take-up and lead to more default. The effect on
profits is ambiguous. You are also aware that increasing the loan to value (LTV) ratio on
collateralized products could increase your portfolio size but might also have adverse
effects on default. Again, it is not clear whether an increase in the LTV ratio will have a
positive impact on profits. So what should you do?
One option would be simply to change the LTV ratio in several branches and measure the
relevant effects on default rate, take-up, and ultimately profits. Along the same lines, you
might try several of the newly proposed products at different branches. While this option
might initially seem appealing for its relative ease, the results will be of limited value.
1 Department of Economics, Harvard University.
For example, imagine there was a drought in the same region where you experimented
with increasing interest rates. If default rates among farmers skyrocket, should we
conclude that the higher interest rates were the cause? Probably not. While droughts
might be accounted for, there are generally lots of other confounding factors that make
inference impossible in such cases. Moreover, we must also concern ourselves with the
external validity of our trial. If default rates increase in a region of paddy farmers, should
we conclude that this will be the case among sugar cane farmers?
Another option would be to use historical data from either the branches themselves or
another setting. This solution could avoid many of the pitfalls discussed above, but it
also creates some of its own. At a basic level, the past is not always a good predictor of
the future. On a more operational level, the data requirements for such analysis are quite
stringent. For example, we might be tempted to look at the effects of the Rural Bank of
India’s historical increases of interest rates. Yet, these reforms took place amidst a sea of
change. It is not clear if the observed effects are a product of the interest rate increase or
another confounding factor. Moreover, the RBI rate increase presumably affected all
banks, but we are only interested in the effect of changing interest rates at one. And in
the case of evaluating new products, historical data on the specific product will not exist.
Given the challenges of the proposed solutions above, it is perhaps not surprising that
most managers choose to rely on instinct and gut feelings. These gut decisions are not
made from complete ignorance. Rather they are informed by thousands of historical data
points and anecdotes from the individual’s past. But this begs the question of whether
people are good mental statisticians. The statistical evidence suggests not. Kahneman
and Tversky, amongst others, have shown that people are poor intuitive statisticians.
They under-estimate the importance of sample-size and draw too strong a conclusion
from small samples (the representativeness heuristic, Kahneman and Tversky). They
anchor too strongly on their prior beliefs and over-read new data as consistent with what
they already believe (confirmatory bias, Lepper, Ross 1986). These insights are in fact
leading economists to re-think the rational foundations of economics (Mullainathan and
Thaler 2000).
Despite all these challenges, there does remain one infrequently used option in the
analytical toolbox: randomized controlled trials (RCT). Long the gold standard in
medical trials, RCTs have recently begun to permeate various dimensions of the social
sciences and business. They have been used to answer questions ranging from the impact
of the negative income tax to the impact of an extra teacher program. And unlike the
strategies described above, RCTs allow us to answer these questions with precision. By
spending some resources, and avoiding the traditional guess and implement strategy, an
organization can dramatically improve the quality of choices it makes.
Precision is not the only advantage of RCTs. By offering an unbiased solution to many
contentious questions, RCTs allow teams to resolve impasses more easily. In our initial
example, each team member might have a different suggestion for increasing uptake of
the loan product. Some team members might recommend increasing interest rates, while
others argue fervently against it. The marketing representative on the team might suggest
offering free gifts to people when they open an account. Rather than engage in the
unnecessary (and inevitable) arguing, the team could simply run a RCT and find the right
answer. An RCT also allows us to test these various hypotheses simultaneously. In large
teams, lots of ideas might be put on the table and testing each of these separately could
take years. RTCs provide a way around this and can teach us things that we don’t expect.
And perhaps most importantly, RCTs are not subject to the psychological biases
discussed earlier.
The situation outlined above was precisely that faced by one of South Africa’s largest
lenders. The lender did not know whether it was setting rates optimally and several other
proposals existed on the table. It turns out that simply changing the picture on the
mailing – an action that there was much resistance to initially because of minor printing
costs– increased uptake by the equivalent of lowering interest rates by 3-5%. This is just
one of the surprising results of this study and illustrates how simple observation is often
not enough to uncover even very powerful effects.
When to experiment?
Not all questions can be answered through RCTs. For decisions on whether to fire the
rural branch manager or what market to enter, the age-old tradition of going with the best
informed guess must be relied upon. Yet, for countless other decisions RCTs are ideally
suited. Organizations often fail to realize the massive scope for RCTs; anything from
compensation policies to optimal fundraising techniques can be addressed with RCTs.
To help assess whether a particular problem is suitable, we propose a simple three-tiered
framework.
Diagnostic 1: Can the uncertain elements of the choice be clearly delineated?
While this restriction might initially seem overly stringent, it is not quite as restrictive as
it appears. Take the question of what the optimal LTV ratio should be. The uncertainty
in this example all involves the level of the LTV, a clearly quantifiable and manipulable
variable. Should the bank keep the LTV ratio at 70%, increase it by 10% or increase it by
50%?
Diagnostic 2: Are the uncertain elements of the choice frequent enough to allow for
several attempts?
Some actions like the hiring of a new branch manager or the expansion into a new region
occur infrequently. It would be infeasible to run an experiment addressing policy changes
around these actions. Yet, there are many actions that occur across a large population or
repeated interactions. In the case of our experiment, the jewelry loan transaction occurs
dozens of times per day across many rural branches in Tamil Nadu, India. This allows us
to randomize across loan transactions: in other words, each individual that comes to the
bank is offered a different loan schedule. In another example, you might consider
looking at the effects of additional classroom supplies on educational outcomes in Kenya
(Glewwe, Kremer and Moulin 1998). If there were enough schools in the state, you
could randomly assign different schools different levels of supplies and measure the
outcomes.
When thinking about the level of randomization (e.g., individual, school district) and
whether there are enough units of observation, one must also worry about two possible
confounds: peer effects and the Hawthorne effect. Peer effects is the notion that the
treatment that one individual receives might affect the outcome of another. In a seminal
paper by Edward Miguel and Michael Kremer (2004), peer effects play a central role.
The authors consider the effects of deworming drugs on health and education. The level
of randomization is the school: some schools received the deworming program, while
others did not. Yet, one cannot simply look at the differences in the outcomes since there
are potentially large cross-school externalities. In other words, by lowering the worm
burden in the treatment areas, the intervention might also have an effect on incidence on
non-treatment schools. Such an effect is referred to as a peer effect and requires extreme
caution if the design and interpretation of experiments.
The other possible concern is the Hawthorne effect or the idea that the simple knowledge
that an individual is involved in an experiment might alter his or her behavior. In the
case of our project, this meant that it was crucial to prevent individuals from informing
others of the loan schedule that they were offered. After sufficient surveying, and several
example of individuals themselves forgetting what LTV they borrowed at last time, we
were convinced that there was nearly no recognition that an experiment was being
conducted.
Diagnostic 3: Is it operationally feasible?
The last and perhaps most difficult hurdle to be crossed is that of operational feasibility:
A well designed experiment should involve as few actors and required actions as
possible. In the next section we will describe exactly how to make RCTs feasible even in
harder to imagine contexts.
Managing the RCT Process: The Six-Step Pathway
The RCT process can be extremely effective at identifying successful new interventions
or unlocking additional profits. It can help us get home early from interminable meetings
and save us from unpleasant arguments with our colleagues. Yet, if not done properly, it
can also have (pernicious) results. A poorly done RCT could provide false confidence in
one project or unfairly lead to the squashing of another. We outline below the six steps
of a successful RCT
Step 1: Identify key questions
The first step in the process is to identify a business question that would be appropriate
for a RCT, as discussed above. This process requires focused and creative brainstorming,
usually involving senior level management along with external consultants. The more
ideas generated at this stage, the more potential there is for the experiment to reveal
something truly surprising. Recall the South Africa experiment above where the idea of
changing the gender of the person in the advertisement was initially dismissed as silly but
ultimately led to a dramatic increase in loan demand (the equivalent of a 5% decrease in
interest rates)
Step 2: Design Experiment
Once the focus questions have been agreed upon, the process must be designed around
the operational realities. Essential to this process, the entire management team must first
agree on the experiment and the action steps that will follow; you don’t want to find
yourself in a position after the results have been analyzed, where one person on the
management team refuses to believe the results and stalls the implementation. In that
case, you might want to experiment with removing that person from the meeting. One
must also check what forms of consent are required for the experiment and if disclosure
agreements are required. In the case of the India bank experiment, we had to go through
several rounds with the risk department and the university Institutional Review Board
(IRB) before the experiment was approved. We also needed to decide on the level of
randomization: should we offer different loan products to different individuals or should
we offer different products at different bank branches? Should the bank tellers know
about the experiment or would this compromise the design? This process requires some
judgment calls. It requires creativity and an obsession for detail. The key guiding
principles are the desire to maintain the validity of the experimental design.
Step 3: Implement Design
It is, of course, impossible to design a perfect experiment. Every experiment will involve
unforeseen implementation problems. It is, therefore, critical to run a small scale pilot of
the experiment to identify and fix these kinks. For example, the experiment might prove
to be too burdensome on the staff; low levels of compliance might compromise the
results; or customers might complain about a change in procedures. In our recent India
experiment, we encountered a particularly obstinate branch manager who, despite
assuring us of his enthusiasm for the experiment, decided to inform the entire bank staff
of the project. The bank staff, in turn seems to have informed the customers. Needless to
say, we were forced to move the experiment to other branches. And this leads us to the
next point.
Step 4: Monitor Compliance Constantly
The centerpiece of a well-executed RCT is strong monitoring. Audits, site-visits and
constant communication with the teams running the RCT are imperative. It is also
possible to use statistical tools to track the experiments’ progress and identify any
flagrant violations of the design. To give an example of what could go wrong, consider
an experiment that a colleague of ours ran with a health club to help them decide on the
type membership pricing. The health club owner was given precise instructions on
administering the various treatments. After a few weeks, however, he noticed that one
was performing slightly better than another and began to offer it to all his customers.
Identifying such randomization failures immediately is imperative as they can invalidate
the entire experiment.
Step 5: Analyze Data
The analysis is by far the most technical step in the process and requires an understanding
of statistics and econometrics. It can be used not only to measure causal responses, but
can used to test the quality of the randomization. Moreover, it is often valuable to
separate out different treatment affects by subgroup (e.g., differential effects of cash
grants to males and females), and identify peer effects. However, each of these effects
should be of specified interest at the design stage. It is a dangerous practice to mine the
data in search of an effect on a given subpopulations. To build intuition on why this can
lead to bad inference, consider a population which can be split into 21 non-overlapping
age subpopulations (e.g. <1, 1-5, 6-10, …95-100). If we were to test the null hypothesis
that there was no effect at standard levels of statistical significance (i.e., there is a 5% of
rejecting the null hypothesis when it is true) for each subpopulation, we would quite
likely find a significant effect for at least one of these groups (i.e., there is a a 5% chance
of falsely rejecting the null for each group but the chance of rejecting it for one is much
higher). Thus, we should be careful to only consider theoretically or intuitively motivated
hypothesis testing on subpopulations.
Step 6: Adopt Policy Changes Accordingly
At this stage of the analysis, most of the work should be done. All that remains is the
adoption of the interventions with a net positive effect. Or if the experiment was testing
mutually exclusive policies, the adoption of the most effective. Yet, this can be one of
the most contentious stages of the process. Team members whose initial suggestions
were not adopted may claim that the experiment was flawed or unfair. Other members
might simply be skeptical of using experimental results to dictate important
organizational decisions. This is precisely why it is so imperative to get agreement
among the team on not only the experimental design (to ensure fairness) but on the
adoption of the results. To further assuage concerns, it is often helpful to have the
experiment designed and possibly run by an outside consultant or academic to ensure
objectivity. An insider might be tempted to bias the results in his or her own favor. In
addition to the immediate answers that the experiment provides, the results might also
suggest hypotheses for further experimentation.
An Example: Relaxing Borrowing Constraints In Rural India
To give a better sense of what an experiment looks like in the field, we present below the
design and results of an experiment in several rural branches of one of India’s largest
private banks. The experiment was initially setup in order to help the bank answer a
fundamental question around one of its collateralized loan products, jewelry loans: Is the
bank being too cautious in setting the loan-to-value (LTV) ratio? Alternately put, can the
bank increase the LTV ratio without increasing default rates? If so, the bank could
dramatically increase its portfolio of collateralized loans without bearing additional risk.
In addition to this question, we were also interested in testing for the existence and
prevalence of credit constraints. Much anecdotal suggests extremely high levels of credit
constraints in rural areas, yet there has been relatively little academic work in the area.
We might also expect such credit constraints to affect anything from entrepreneurship to
health outcomes. To test for these links, we follow up the experiment with a household
survey that allows us to measure many of the outcomes that we are interested in.
Context
The bank with which we are working has been an innovator in the rural finance sector in
India, introducing several new products in the last five years. One of these products,
jewelry loans, offers individuals the opportunity to pledge their gold jewelry as collateral
in exchange for a loan. Much of this jewelry has been bought in preparation for a
daughter’s wedding or as an investment, received as a gift or was simply passed between
generations. The minimum allowed loan size is 5000 Rs. (~$110) and the average loan in
the branches we have been working is about 19000 Rs (~$400). The median is slightly
lower at 14000 Rs suggesting that a few outliers with large loans are inflating the mean.
Customers are offered a schedule of different loan options, each of which specifies an
interest rate, LTV ratio and a term for the loan. A sample schedule is shown below:
Interest Rate Loan-to-Value Ratio Maturity
10.0% 460 18 months
10.5% 500 12 months
11.5% 510 6 months
13.0% 525 3 months
15.0% 550 45 days
Loan Schedule
The loan to value is expressed in a Rs. to grams ratio. Thus, an individual who brings 20
g of gold will be able to borrow 9200 Rs. at the 10% interest rate. The price of gold was
approximately 650 Rs. per gram at the time so the highest slab represented a LTV of
about 85%. The average amount of jewelry loaned against is 34 g and the median
amount, 28 g. Nearly half of the customers borrow at the lowest interest rate and the most
frequently stated uses of the money are working capital (e.g., extra fabric before Divali),
general consumption (especially to smooth lumpy income streams), and weddings and
other festivals. And perhaps the most surprising fact, and the motivating factor for the
initial experiment: default rates on these loans was close to zero.
The formal sector banks are not the only players in the rural finance market: pawnbrokers
and local moneylenders often control a large share of the market. The interest rates
charged by this sector are significantly higher than those in the formal sector; reported
rates in the area were approximately two to three times higher than those mentioned
above. Despite this, anecdotal evidence suggests that the use of jewelry loans to
refinance outside loans is relatively limited. We discuss this puzzle below.
Experimental Design
The focus of the initial experiment was to measure the effect of increased credit limits on
default rates and various social indicators. We also hoped to explore the extent to which
individuals were credit constrained (i.e., would liked to have borrowed more at the
prevailing interest rate but were unable to do so because of credit limits). One of the key
features of developing countries is that the poor borrow at very high interest rates.
Banerjee (2001) for example surveys a variety of empirical studies and finds that “an
annual interest rate above 35% is standard and those above 75% are by no means rare.”
Such high interest rates have profound implications for our understanding of the
consequences and pathways out of poverty. If the poor face such high rates of capital, it
suggests that they may not be able to smooth many adverse shocks to consumption. On
the flip side, it suggests that the poor may have easier pathways out of poverty if only
given easier access to credit. Since the marginal rate of return on investment should equal
the borrowing rates, high interest rates imply high rates of returns on investment. This in
turn suggests that the poor may be sitting on a large set of quite profitable investments
that they are unable to undertake due to lack of credit. This view has obvious policy
implications and also raises questions about why international and domestic capital
markets cannot arbitrage interest rate differentials.
To test the prevalence of credit constraints, we offer jewelry loan customers randomly
assigned loan schedules by varying the LTV ratio. We choose to randomize at the
individual level both to maximize the power of the statistical test, since there were not
many rural branches, and to allow us to interpret the results as the effect of loosening
individual level credit constraints. The control group is offered the schedule presented
above The first treatment group is offered an additional 50 Rs. per gram on their jewelry.
The schedule that they are presented with is thus
Interest Rate Loan-to-Value Ratio Maturity
10.0% 510 18 months
10.5% 550 12 months
11.5% 560 6 months
13.0% 575 3 months
15.0% 600 45 days
Loan Schedule
The second treatment group is offered an additional 100 Rs. per gram.
We chose to hire and train independent tellers to run the experiment, rather than letting
the individual branch employees do so. We did this because we were concerned about a
possible conflict of interest: the branch employees, especially the jewel appraiser, might
be tempted to simply offer the highest credit limit to everyone (or offer it to the
individuals with the most jewelry), since they get compensated as a fraction of the loans
they distribute. By separating the appraising of the jewelry and the process of
randomization into two stages, we have more confidence that the assignment of credit
limits (LTVs) is truly random.
Each customer first approaches the appraiser and has his or her jewelry weighed. Once
the jewelry has been weighed, the customer continues to a different area of the bank
where one of the centrally trained staff shows the customer the appropriate schedule. The
schedule has been randomly assigned and is explained to the customer by the independent
teller. The customer then selects an amount that he would like to borrow and receives his
money from a separate teller.
Results of Baseline Experiment
While the experiment is still running and the results not yet conclusive, we can already
point to several interesting findings. We present the data on the initial 390 observations
Most individuals borrow at the lowest interest rate
In the baseline loan schedule (no extra amount offered), nearly 55% of the customers
borrowed at the lowest interest rate. And of these customers, approximately 34% left
more than 100 Rs. on the table. If individuals borrow up to the point where the return to
capital no longer exceeds the cost of borrowing, such evidence would suggest lower
returns than is normally assumed. We present a different explanation below.
Nearly 40% of individuals who are offered higher LTV do not take any of it
Perhaps the most surprising finding is the high fraction (40%) of individuals who are
offered an additional amount but do not take more they were initially eligible for.
Moreover, a t-test comparing the total loan amount taken by the customers receiving a
higher LTV and those receiving the baseline amounts, delivers a statistically insignificant
difference (p-value .41). Alternately put, there is no difference, on average, in the loan
amount taken by individuals who are offered the extra amount and those who are not.
Furthermore, only 30% of individuals who are offered a higher LTV take the full (defined
as within 100 Rs of product of net weight and chosen LTV ration) amount for which they
were eligible. The other 70% of the sample left money at the table; this 70% of the
sample left on average 23% of the amount for which they were eligible. By the
traditional definition of credit constrained, 30% thus serves as an upper bound on the
fraction of credit constrained customers. This is truly surprising given the anecdotal
evidence which suggests that a large fraction of the population is credit constrained. The
number of individuals leaving money on the table should also be surprising since
anecdotal evidence suggests that many of these same individuals have informal loans at
much higher interest rates. In a related work, however, Kochar (1997) uses a structural
model to estimate that the probability of not having access to formal sector credit
conditional on the household's demand for credit and choosing the formal over the
informal sector is about 26%.
Individuals in treatment cells borrow at lower interest rates
Consistent with the results above, we find that individuals who were offered extra borrow
at a lower interest rate than those who weren’t. This is consistent with the idea that
individuals have a fixed amount that they would like to borrow and choose the lowest
interest that allows them to borrow this amount. By offering higher LTV ratios, we allow
individuals to borrow the same amount at a lower interest rate. The point estimate for
the difference in interest rates is .27 and the p-value .04.
Mental Accounting
The low rates of credit constraints and minimal refinancing is surprising. Has the
literature been placing too much emphasis on credit constraints? We suggest that mental
accounting, a behavioral economics theory, explains such unexpected borrowing patterns.
What is mental accounting? While the literature has struggled to place a precise definition
on the term, mental accounting is generally used to capture the idea that individuals
process ideas and activities in separate accounts, not unlike the way accountants process
financial information. For example, an individual might have separate mental accounts
for housing, food, business, and festival expenses. In other words, mental accounting
suggests that individuals do not have, as standard economic theory suggests, a single
inter-temporal budget constraint which they use to make trade-offs across investments,
consumption, debt and time. Instead, the theory suggests that individuals may have
narrow accounts for specific purposes. Each account may in turn have its own “budget”
and implied discount rate.
This alternative view of consumer finance, if true, suggests that the standard exercise of
using the interest rate that the individual borrowed at to estimate the marginal return to
capital may be flawed. Finding high interest rates in some mental accounts does not in
anyway suggest that individuals face those rates across all accounts. Most importantly, in
this world view, these high rates would likely not be the ones they would face if they
were interested in another loan (or for that matter what they faced in their last loan). This
is because accounting only locally—i.e. the non-equalization of rates across accounts—
means that individuals will not have a pecking order of borrowing moving from the
lowest source to the highest. A mental accounting theory, as it relates to credit
constraints, would imply this non-pecking order fact and also that loans taken in a
moment of crisis, though replaceable later, will tend not be revisited and stay at a high
interest rate. However, circumstances that lead the individual to revisit these accounts—
such as taking a loan of a related activity or highlighting the account at the time of
another loan can lead to revisiting and reduction of the rate at a later time.
In our specific case, we suggest that individuals might have placed their previous loans
with moneylenders or pawnbrokers in a separate mental account than those they are
taking from the bank. If this were the case, the individuals might not be thinking across
their mental accounts when they are taking an individual loan. The individual borrows in
reference to a specific account; he does not think in terms of a global account (e.g.,
money is not treated as fungible). Such behavior would explain both the limited amount
of refinancing that appears to be happening and also the relatively few people borrowing
the full loan amount.
We therefore propose an additional experiment to test whether this theory can explain the
surprising results above. Assuming the theory is correct, we expect that an intervention
that leads an individual to think across accounts and beyond the account he is initially
borrowing in, should increase the loan amount taken. The specific intervention that we
use to push individuals across accounts is a simple reminder by the teller that the loan can
be used for various purposes. The reminder is scripted and mentions three alternate uses
of the loan, beyond the purpose for which the individual has stated he is borrowing. To
control for the possibility that the simple interaction between teller and customer might
drive the customer to borrow more, we compare the above treatment to a treatment where
the teller mentions three additional uses of the money related to the stated purpose of the
loan. For example, if an individual is borrowing for a wedding, the teller would mention
using the loan for other wedding expenses in the first treatment; and in the second
treatment, she would mention using it for refinancing other loans, household consumption
or investment in a business. The results from this experiment are still inconclusive at this
stage due to the relatively low number of observations.
Conclusion
As we’ve tried to argue above, RCTs can be applied to many areas of rural finance and
agriculture. They could be used for anything from understanding the impact of subsidized
fertilizer on local market outcomes to the impact of relaxing the traditionally rigid
microfinance contracts. They provide the most scientifically founded answers to many of
the most difficult policy questions and should be taken very seriously in policy
discussions. In the experiment described above, we will not only able to inform the bank
whether they can safely increase credit limits without much concern for increased default,
but we can also say something about the prevalence of credit constraints and the
importance of mental accounting in understanding borrowing patterns.
References
BANERJEE, A. (2004): "Contracting Constraints, Credit Markets and Economic Development," in Advances in Economics and Econometrics: Theory and
Applications, Eight World Congress of the Econometric Society, Volume Iii, ed. by L. H. a. S. T. M. Dewatripoint: Cambridge University Press, 1-46.
BERTRAND, M., D. KARLAN, S. MULLAINATHAN, E. SHAFIR, and J. ZINMAN (2005):
"What's Psychology Worth? A Field Experiment in the Consumer Credit Market," NBER Working Papers.
GLEWWE, P., M. KREMER, and S. MOULIN (1998): "Textbooks and Test Scores: Evidence
from a Prospective Evaluation in Kenya." KAHNEMAN, D., and A. TVERSKY (1973): "On the Psychology of Prediction,"
Psychological Review, 80. KOCHAR, A. (1997): "An Empirical Invesitgation of Rationing Constraints in Rural Credit
Markets in India," Journal of Development Economics, 53, 339-71. KREMER, M. (2003): "Randomized Evaluations of Educational Programs in Developing
Countries: Some Lessons," American Economic Review, 93, 102-106. LEPPER, M., L. ROSS, and R. LAU (1986): "Persistence of Inaccurate Beliefs About the
Self: Perseverance Effects in the Classroom," Journal of Personality and Social
Psychology, 50, 482-491. MIGUEL, E. A. M. K. (2004): "Worms: Identifying Impacts on Education and Health in the
Presence of Treatment Externalities," Econometrica, 72, 159-217. MORDUCH, J. (1998): "Does Microfinance Really Help the Poor? New Evidence from
Flagship Programs in Bangladesh," mimeo. MULLAINATHAN, S., and R. THALER (2000): "Behavioral Economics," NBER Working
Papers. THALER, D. (1999): "Mental Accounting Matters," Journal of Behavioral Decision
Making, 12, 183-206.