SERIES

IZA DP No. 10024

PAPER

How to Help Poor Informal Workers to Save a Bit:

Evidence from a Field Experiment in Kenya

Merve Akbas
DISCUSSION

Dan Ariely
David A. Robalino
Michael Weber

June 2016

Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
 How to Help Poor Informal Workers
to Save a Bit: Evidence from a
Field Experiment in Kenya
Merve Akbas
Aliya Analytics

Dan Ariely
Duke University

David A. Robalino
The World Bank and IZA

Michael Weber
The World Bank

Discussion Paper No. 10024

June 2016

IZA

P.O. Box 7240

53072 Bonn
Germany

Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: iza@iza.org

Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The IZA research network is committed to the IZA Guiding Principles of Research Integrity.

The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.

IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 10024
June 2016

ABSTRACT

How to Help Poor Informal Workers to Save a Bit:

Evidence from a Field Experiment in Kenya1

Worldwide, the majority of workers hold jobs in the informal sector that do not provide access
to social insurance programs. We partnered with a savings product provider in Kenya to test
the extent to which behavioral interventions and financial incentives can increase the saving
rate through a voluntary pension program for informal workers with low and irregular income.
Our experiment lasted for six months and included a total of twelve conditions. The control
condition received weekly reminders and balance reporting via text messages. The treatment
conditions received in addition one of the following interventions: (1) reminder text messages
framed as if they came from the participant’s kid (2) a golden colored coin with numbers for
each week of the trial, on which participants were asked to keep track of their weekly
deposits (3) a match of weekly savings: The match was either 10% or 20% up to a certain
amount per week. The match was either deposited at the end of each week or the highest
possible match was deposited at the start of each week and was adjusted at the end. Among
these interventions, by far the most effective was the coin: Those in the coin condition saved
on average the highest amount and more than twice as those in the control condition. We
hypothesize that being a tangible track-keeping object; the coin made subjects remember to
save more often. Our results support the line of literature suggesting that saving decisions
involve psychological aspects and that policy makers and product designers should take
these influences into account.

JEL Classification: G21, B49, D03

Keywords: savings, field experiment, behavioral economics

Corresponding author:

Merve Akbas
Aliya Analytics
20 Marshall St., Suite 102
South Norwalk CT 06854
USA
E-mail: merve@aliya.com

1We thank the Trust Fund for Environmentally & Socially Sustainable Development (TFESSD) for
providing the funds. We would also like to thank our institutional partners RBA Kenya, Eagle Africa,
Kenya Commercial Bank, Jua Kali Association of Kenya and Safaricom for their generosity and
support and their interest in the research. We thank James Vancel, Chaning Jang and the Busara
Center for Behavioral Economics for the excellent fieldwork. We thank Anat Binur, Rebecca R.Kelley
and Chang Yuan Lee for their support during the project. We also thank Curtis Taylor, Erkut Ozbay,
Johannes Haushofer, Jeremy Shapiro and seminar participants at Behavior and Policy Lab at
Princeton University for their valuable comments. All errors are ours.
 1. Introduction

How do people decide how much to consume today versus how much to save for future
consumption? The answer to this question is central for many important economic
analyses as well as government policies. Savings behavior observed across many
situations exhibit numerous inconsistencies with standard models of inter-temporal
choice. For example, factors such as patience and risk tolerance, which should in
principle explain differences in observed retirement savings, fail to do so based on the
U.S. data (Bernheim, Skinner and Weinberg, 2001). Changing the offer for (401)K plan
participation from opt-out by default to opt-in by default influences the number of
employees who sign up and their total savings (Madrian and Shea, 2001). On the other
hand it is recognized since Strotz (1955) that psychological factors such as self-control
problems play a major role in savings decisions, and that people exhibit dynamic
inconsistencies even in simple inter-temporal decisions related to saving (Frederick et al.,
2001). The psychological view that emerges from this line of research has very different
implications for the factors that influence the savings behavior and ultimately for how
institutions and incentive mechanisms should be designed.

If individuals suffer from self-control problems, are impulsive and have a shortsighted
vision – the question of “how to get people make better saving decisions” becomes an
important economic problem. One approach to study this question is through field
experiments, by varying the features of an existing savings product. By testing whether
certain interventions increases saving rates, we gain insight about the underlying decision
making process, and propose alternative approaches and mechanisms. For example,
Ashraf, Karlan and Yin (2006) demonstrated that some individuals voluntarily take up
financial commitment devices, which limit their access to their own savings for a certain
time period even when these plans offer no direct benefit -- suggesting that self-control is
an issue for some people, and that some of these individuals are sophisticated enough to
take binding measures against it (O’Donoghue and Rabin, 1999). In other research it has
been shown that even simply announcing a savings goal and then regularly reporting to
peers the levels of achieved savings increase saving rates, suggesting that even non-

1
binding commitments can be effective when dealing with self-control problems (Kast,
Meier and Pomeranz, 2012).

Having an accurate model for what motivates people to save more becomes particularly
important when policy makers want to encourage their citizens to save more. For
example developing countries of Sub-Saharan Africa are looking for ways to increase
domestic savings in hopes of achieving a higher rate of economic growth and job
creation2. As another example, recently, some of the federal institutions in the United
States started to adopt policies to encourage low-income families to save more of their
regular income, as well as their tax refunds. 3 With the importance of such saving
programs, the question is what kinds of interventions would be most cost effective?
Would these be based on the standard economic model? Should they be based on some
behavioral findings and if so, on which ones?

Currently one of the most common interventions to encourage low-income individuals to

save more is financial literacy education. The underlying assumption for such financial
literacy programs is that the major barrier that is preventing low-income individuals from
saving is their lack of knowledge about the benefits of savings and the way saving
products work. Even though it is true that low-income individuals score low on financial
literacy tests, recent research suggests that the effectiveness of financial literacy programs
is negligible (Fernandes, Lynch and Netemeyer, 2014). Another common intervention to
encourage higher saving rates is the golden standard of economics: Providing incentives,
in the form of matching contributions.4 Undeniably, a match creates a motivation for
saving. For example, Duflo et al. (2006) reports that during tax-time in the United States,
a 20 percent match for IRA contributions increased the take up rate from 3 to 8 percent in
a sample of middle-and low-income families in St. Louis. However, since matching can

2 See the discussion paper “Incentivizing non-retirement savings” released by the National Treasury of
South Africa, retrieved at www.treasury.gov.za.
3 See “Assets for Independence Act” legislation at http://www.acf.hhs.gov/programs/ocs/resource/afi-
legislation-0
4 Matches are given sometimes given with certain pre-conditions, such as not withdrawing deposits for a
certain timeframe, or meeting the pre-set savings goals. See instructions on Individual Development
Accounts at http://cfed.org/programs/idas/.

2
be prohibitively costly, it is possible that psychological interventions might be cheaper
and hence more practical.

One effective psychological intervention is sending messages regularly to remind saving.

The idea behind reminder messages is that people have busy lives and cannot pay
attention to all of their future expenditures in the present, leading them to under-save
(Karlan et al., 2012). Thus, by making the relevant information regarding future needs
salient, a reminder message can help to make a better decision regarding saving. Though
this intervention is promising given its simplicity and low cost, it is also clear that in
many real world problems people do not always respond to mere information. We
illustrate this idea by two examples. First, in the charity context, it is widely documented
that people donate more to a charity when they see the picture of a particular child who
needs help, compared to when they receive the information demonstrating the severity of
the problem (Slovic, 2007). Armed with this knowledge, charities often use specific cues
to appeal potential donors’ emotions while conveying information about why they should
donate. Another example is about exercising. While there is abundant information about
the benefits of exercising, people usually fail to follow through even though they think
that they should exercise. However recent popularity of wearable activity trackers suggest
that the design of how relevant information is brought to mind can matter a lot for
directing behavior. While there are many features that make activity trackers popular, we
speculate that two of them are particularly important: First the wearable parts of the
trackers are appealing in design. Second, they set small daily, weekly goals and give
nonmonetary rewards such as badges when users meet them. With these two features,
users are constantly reminded about their tracking device, and rewarded for their on-
going achievements.

In this study, we designed two psychological interventions using similar approaches to

the ones explained above and we tested their effect on saving behavior in a field
experiment in Kenya. The first of these interventions is sending regular reminder text
messages to the savers framed as if sent by their kids; and asking them to save for their
future. Even though the savings product we used was an individual savings product, we
hypothesize that reminding individuals of their own kids through the text message

3
framing could appeal to their emotions, and thus motivate them to save more. The second
intervention aims to create a tangible and appealing representation of savings to serve as
a reminder and to provide small rewards for deposits and thus create a sense of
accomplishment among savers, boosting their motivation. To this end, we designed a
metal coin, which looks like gold to represent savings (see Figure 1). The coin displays a
picture representing continuous savings on the front and resulting increase in wealth on
the back. The coin also has numbers for each week of the six months period of our
experiment, so that users can mark each number in a specific way to keep track of the
weeks in which they saved and in which they did not. The idea behind the coin is that
savings, particularly small deposits towards long-term savings are abstract. It is very hard
to make sense of what each deposit means for the final goal, as well as the final result of
missing a deposit.

On top of testing the effectiveness of these two psychological interventions to motivate

low-income individuals save a bit more, the second aim of this study is to compare the
effectiveness of the psychological interventions to the effectiveness of providing financial
incentives. A policy issue of crucial importance is having a good understanding about the
effect size of financial incentives as well as psychological interventions and their
combinations. Since the setting of financial products exhibit very particular details, it is
very hard to compare the effects of psychological interventions measured with one study
in a particular country with a particular product to the effect size of financial incentives
measured in a different country with a different product. The difficulty of comparing the
effectiveness across studies makes it very difficult to understand the relative effects of
psychological interventions and financial incentives in general. By testing many types of
interventions within one study, we also aim to contribute to this line of literature.5

To test the effectiveness of these two approaches, appealing to emotions by reminding

one’s kids, representing savings with a tangible track-keeping object and to compare
these interventions to financial incentives, we partnered with the administrators of a
newly developed savings product in Kenya, Mbao Savings Plan, which resembles a

5 Ina loan repayment context, Cadena and Shoar (2013) find that sending reminder text messages three
days earlier than the loan repayment deadlines reduces late payments as much as a promised interest
reduction of 25% for the next loan if all deadlines are met for the loan repayment.

4
regular bank account with a commitment device, that is, a three year restriction on
withdrawals after registering.

Our experiment lasted for six months and included a total of twelve conditions, one
baseline, one control and ten treatment conditions. The baseline condition continued to
receive no regular communication from the savings product as it was the standard
procedure used by Mbao. The control condition received text messages twice every week,
one on the first day of the week, asking participants to save during the week, and one on
the last day, reporting their weekly savings and final savings balance. The treatment
conditions differed in one of the following ways from the control condition: (a) The kids
treatment received the text message on the first day of the week framed as if it came from
the participant’s kid. (b) The coin treatment received the “gold coin” at the start of the
intervention and each week, on top of the regular report message, participants received an
additional text message, asking them to scratch around that week’s number with a knife if
they saved that week and below the number if they did not. (c) The ex-post match 10%
treatment received a 10 percent match for weekly savings up to 100 Shillings per week,
deposited in accounts on the last day of the week and reflected in the balance report. The
first message of the week in this treatment was also modified to remind about the match
participants would receive if they saved that week. (d) The pre-match 10% treatment
received the highest possible match with the same 10% rate and 100 Shillings weekly
cap, that is 10 Shillings deposited in accounts on the first day of each week and the
reminder text message of the week informed participants about this and reminded them
that they would lose it if they did not save at all. This treatment was thus exactly the same
as the ex-post match 10% treatment in terms of incentives and the two only differed in the
framing of first weekly reminder text message. (e) The ex-post match 20% and pre-match
20% treatments were the same as the ex-post match 10% and pre-match 10%, but offered
a 20% match rate up to 100 Shillings a week, thus higher financial incentives to save. (f)
Four additional treatments involved a combination of ex-post match 10% and pre-match
10% treatments with either kids treatment or with the coin treatment. These four
treatments allow us to compare the effect of combining a financial incentive with a
psychological intervention to increasing the financial incentive 100 percent.

5
We tested the interventions explained above on a pre-selected sample of 2,400 Mbao
customers. We reached out to these customers via mobile phone at the beginning of the
experiment and they accepted a six months “promotional service” from Mbao. Each
individual from this sample was offered one randomly preselected intervention as “a
promotional service”. In addition, we offered the product to 1567 individuals in the slums
in Nairobi, offering each of them one randomly selected intervention as a “promotional
service”. We collected savings data for six months from the 2400 individuals in the
existing customers sample and the 1464 individuals who signed up for Mbao from the
new customers sample.

Our first general finding is that the coin was by far the most effective intervention in
increasing savings. In the existing customers sample, those in the control condition saved
on average 674 Shillings at the end of six months whereas those in the coin condition
saved more than twice on average, 1,461 Shillings. Average savings in the treatments
that involved a combination of the coin and ex-post match 10% and pre-match 10%, were
1,118 Shillings and 1,156 Shillings respectively. The difference between all three
conditions that involved the coin and the control condition is statistically and
economically significant. Second, we find that only sending weekly reminder text
messages increased the fraction of participants in the existing customers sample who
saved a positive amount in the six months of the experiment from 0.48 in the No
Intervention condition to 0.73 in the Reminder condition. Furthermore, average savings
in the Reminder condition were 674 Shillings whereas they were less than half in the No
Intervention condition, only 266 Shillings. Together, these two results confirm the
findings of previous research on the effect of reminders as well as support our hypothesis
that further psychological interventions can increase its effect significantly.

Our third finding is that in the existing customers sample the text message from kids
treatment and all treatments that involved financial incentives in the form of matching
slightly increased savings compared to the control 6 , however the increases are not
statistically significant in any of these treatments. Going back to the policy question we
posed earlier, our results provide evidence that a psychological intervention, in our study

6 With the exception of ex-post match 10% treatment, in which average savings were slightly lower than
the control.

6
the coin, can be more effective than financial incentives, and thus imply that policy
makers and product designers should consider such interventions as viable alternatives to
financial incentives whenever appropriate. Obviously the effectiveness of financial
incentives can always be increased with higher matching amounts, however it is
important to note that in our study, if we matched the actual savings of the coin condition
participants from the existing customers sample at 20% with a 100 Shillings weekly cap,
it would cost us 139 Shillings per person, which is worth about $1.5. On the other hand,
the coins cost us only about $0.6 per person.7 Thus, the coin was not only more effective
than matching savings at 20% level, but also cost less.

Finally, we find that in the new customers sample, the majority of the participants, about
85 percent, who registered for Mbao at the beginning of the experiment did not save at all
in the next six months. Average savings per saver in the six main treatments we
employed in this sample were not significantly different from average savings in the
control. We speculate that either these interventions were not strong enough to
substantially change the behavior of a sample with little interest in saving to begin with,
or that a low trust level in such a newly developed financial product suppressed any
effect we could observe in this sample.

The rest of this paper is organized as follows: Section 2 gives background information
about the mobile money system in Kenya and our partner institution. Section 3 explains
the experimental setup. Section 4 explains the implementation. Section 5 describes our
data. Section 6 reports the results and section 7 discusses the implications for policy and
future research.

2. Background on mobile money and partner institution

Kenya’s financial system experienced a major transformation when the first mobile
phone-based financial system in the country was launched in 2007, called M-Pesa.8 It is a
service offered by the leading mobile operator in the country, Safaricom, a subsidiary of

7 The coins were produced by a supplier from Alibaba.com and the cost per person mentioned here
includes the shipping cost, which was about the same as the total cost of the coins.
8 M stands for mobile and Pesa means money in Swahili. For detailed information on the MPesa system
see Jack and Suri (2014).

7
Vodaphone. M-Pesa is in principle a checking account linked to the mobile phone
number. In a country where less than 20 percent of the adult population has access to
bank accounts whereas access to a mobile phone even among the residents of rural areas
is 90 percent, M-Pesa has filled a major gap in the economy: Safe money transfer at a low
cost. Anyone who has a mobile phone account and a national ID number can sign up for
the M-Pesa account at a cost as much as 100 shillings (approx. $1.1). Once the account
balance is positive, one can transfer e-money to another phone by sending a text message,
receive e-money from another phone and cash-out the e-money in a certified M-Pesa
agent shop. Depositing and receiving money is free, however there is a transaction cost
associated with each transfer on the sender, following a step function. There is also a
fixed cost for withdrawing money. Each M-Pesa account has a private PIN number
required to be entered with each transaction, providing the security of accounts. Only
after four years of its launch, in 2011, nearly 70 percent of the adult population (14
million) adopted M-Pesa; making it the most widely adopted mobile money system in the
world (Jack and Suri, 2014). Partly, this success was possible by the quick spread of the
agents network, small booths converting cash to e-money and vice versa. The total
number of M-Pesa agents in the country reached 28,000 in 2011, making it easy for
anyone to exchange e-money for cash at any time. M-Pesa accounts do not pay interest
and do not lend money, thus it does not completely substitute a bank account.
Government regulations restrict each individual account balance to be at most 50,000
shillings (approximately $555).

Given the high adoption of M-Pesa, many financial institutions started to integrate with
it. Currently, many banks allow their customers to withdraw M-Pesa funds from ATMs
and insurance companies accept payments through M-Pesa. For this project, we partnered
with the administrators of a savings product, which is fully integrated with the M-Pesa
system. This savings product, called Mbao, resembles a regular savings account with a
few twists we will explain below.

Mbao was launched in late 2010 by Eagle Africa, an insurance company based in Nairobi
and a large private bank, Kenya Commercial Bank (KCB hereafter) with the support of
the government body responsible for regulating retirement plans (Retirement Benefits

8
Authority) and the national association of the informal sector workers (Jua Kali). The aim
of the product was to provide informal sector workers formal access to savings9 and to be
a substitute for formal pension plans such workers do not have access to. The product
was designed for this specific aim in mind: It does not offer a fixed interest rate like a
regular savings account, instead funds are invested by KCB which serves as the funds
manager, and returns for each account are calculated at the end of each year. It offers a
binding contract to the members, encouraging long-term savings: Withdrawing funds
before three years bears a cost and foregoing the returns of the year in which the
withdrawal is made. Because its target group is low-middle income workers with
irregular income stream, it does not impose a required minimum deposit either monthly
or yearly. Registering is easy, one just provides his national ID number, fills out a one-
page form and pays 100 shillings fee to sign up10. Since the product operates on M-Pesa
system deposits are made by sending a text message, thereby moving the specified fund
from the M-Pesa account to the Mbao account. Regular M-Pesa transfer fee applies and is
deducted form the M-Pesa balance, and one immediately gets an update via text message
about the new Mbao balance as well as the remaining M-Pesa balance. Mbao balance is
checked via text message as well, with a regular text message cost. There are no regular
balance updates other than upon a new deposit, thus checking the balance is costly. By
the time we partnered with the administrators, the product customers were not getting any
regular information or encouragement messages. With this operation model, it was very
easy that saving slipped away from minds of the majority of customers unless they were
very motivated.

According to the data the administrators shared with us, there were 35,204 unique
accounts in the Mbao database in June 2013, of which only 11,041 were active in the six
months prior to analysis. This low rate of active accounts is not surprising as the main
channel of growing the customer base was arranging social (e.g. concerts), or community
service events (e.g. free medical exams) and inviting the audience to sign up for free.
Since signing up is so easy, even those with little or no interest could have done it
because of reciprocity or social norm concerns. Prior to our experiment, we only had

9 Informal sector made up 90 percent of all the businesses in the country in 2005 according to www.ilo.org.
10 It is also offered for free at promotional events.

9
access to the unique identifiers for each account and their savings data. Not surprisingly
given the zero communication policy of the product, more than half of the active
accounts were only active in at most two weeks in the preceding six months of our initial
analysis. We implemented our experiment on a selected sample from active accounts, to
measure the effect of the interventions on a population with a revealed interest to save but
low actual savings. To measure the effects on the interventions on new customers, we
also offered the product to the residents of the slums in Nairobi. The following sections
explain the experimental design, implementation and the timeline.

3. Design

Our experiment involved 10 treatment conditions, one control condition and one baseline
condition, adding up to 12 conditions in total as shown in Table 1. Six treatments were
main interventions and four treatments were a combination of two main interventions.
We employed the full set of conditions with the existing customers sample and only the
control and main treatment conditions with the new customers sample. The text of the
weekly messages sent to each condition is summarized in Table 2. Baseline, control and
main intervention treatments are explained below:

No intervention (baseline): As mentioned before, the customers did not get any regular
message from the Mbao administrators before our experiment. Those in the no reminder
condition received during the experiment just the same service as what they received
before: No communication about their savings account.
Reminder (control): Those in this condition received two text messages every week, one
reminder message on the first day of the week and one report message on the last day of
each week. The reminder message read as: “Please don’t forget to save for your future
this week with MBAO PENSION PLAN.” and the report message was: “Between day1
and day7 you saved Ksh X with MBAO for account no. Your account balance is Ksh Y.”
The reason for choosing this particular setup for the control condition was that some of
the treatments required at least one message each week either on the first day or on the
last day of the week. To avoid any bias that could result from sending messages on
different times, we decided to include two messages in all conditions and only vary the
content across conditions.

10
Coin: The coin treatment provided individuals a gold-colored coin, with 4 cm diameter
and 0.6 pounds weight (approximately equal to the weight of 5 US Quarters). The coin
was designed to look like real gold coins to represent savings. To intensify this feeling its
surface displays a picture representing savings on the front and an increase in wealth on
the back. The coin has engraved numbers from 1 to 24, representing each week of the six
months intervention period. Those in this condition were told to think of this coin as
something to represent their savings and were asked to keep it in a safe place and were
told that they would use it to keep track of their weekly savings. During the experimental
period, they received the same reminder message on the first day of the week as those in
the control and the same report message at the end of each week. However, immediately
after the report message, they received an additional message asking them to scratch
around the corresponding week number if they saved that week and below if they did not.
Kids: This treatment differed from the control only in the reminder message. Those in
this condition received the reminder message as “Hi daddy/mommy. Please deposit as
much as you can this week to MBAO PENSION PLAN for our future. Thank you for
saving. Name”. The name of the kid was obtained from participants when they signed up
for the experiment and they knew that the message was coming from the Mbao
administrators. Those who did not have kids were asked to give the name of someone
they would like to receive the message from and instead of mommy (daddy), their text
message mentioned their name.
Ex-post match (10% and 20%): Those in the 10% ex-post match condition received 10
percent extra for their weekly savings up to 100 shillings. We implemented the cap to
discourage gaming the system by those who can afford to save more, as well as because
this is a common procedure for the match implementations in real world. We chose the
cap as 100 shillings ($1.1) because it is low enough that the majority could afford saving
that much each week, yet it is still much higher than the average savings per week of the
previous year that it would be high enough even if it functioned as an anchor. The
reminder message on the first day of the week read: “Save at least Ksh100 with MBAO
PENSION PLAN this week to receive Ksh10 extra. If you don’t save, you won’t receive
any extra money”. The report message included only an extra sentence “You received x
shillings extra from Mbao” on top of the control report message. The 20% ex-post match

11
treatment was exactly the same as the 10% ex-post match treatment except that the match
rate was 20 percent up to 100 shillings.
Pre-match (10% and 20%): The pre-match treatments provided the same financial
incentives as the ex-post match treatments. The difference was that the highest possible
match was deposited in the accounts on the first day of the week and the reminder
message provided this information with the following text: “MBAO PENSION PLAN
deposited Ksh10 in your account. Save at least Ksh100 this week to keep it. If you don’t
save, MBAO will take it back.” The report message was the same as the report message
of the ex-post match conditions.

4. Timeline and implementation

The experiment lasted from October 2013 to October 2014. Between October 2013 and
January 2013, we collected baseline survey data from all participants. Within a month
after the baseline survey data was collected, each participant started to receive the
intervention; and their saving data was collected for the next six months. At the end of six
months we contacted all participants via cell phone and collected endline survey data.
Endline survey data collection lasted from April to October in 2014. Even though
baseline survey data from both existing and new customers samples were collected
simultaneously, randomization was done separately within each sample and the
procedure was slightly different for each of the two samples, as we explain below:

Existing customers: From the 11,041 customers in the database of the institution who
were active in the six preceding months of June 2013, we selected the 4,596, who saved
on average less than 250 Shillings in the weeks they were active. The reason for selecting
such a specific group was to avoid including a distinct group of customers with higher
income than the rest and potentially employed in formal sector.11 This was also consistent
with our goal of testing our interventions on individuals with low to middle income.

We randomly assigned each customer from the pre-selected sample to one of twelve
conditions as explained earlier. Starting in October 2013, we called the customers in the

11 We had been informed that the employees of the bank and insurance company were invited to join the
plan in the early stages of development.

12
preselected group on their cell phones. After the introduction we informed them that they
were randomly selected to receive a promotional service from the institution for 24 weeks
and that we were conducting a research among customers to improve the product. After
we explained the intervention they were randomly assigned to, we obtained their consent
for research and collected baseline survey data. For those in the Coin condition, we
obtained the location of their residence during the introductory call, arranged a particular
day to deliver the coin to them and called again to confirm the delivery. During the
second call we explained once again in detail how they were going to use the coin. Each
customer was compensated for their time for the baseline survey with 150 Shillings,
which was sent to them electronically via M-Pesa. To exclude those with formal jobs we
first asked the customers in the preselected sample whether they were contributing to the
national social security, which is compulsory for the formal sector employees. If they
answered yes, we explained them that the research did not target formal sector
employees. With this setup, we only collected baseline survey data from those who were
willing to receive the promotional service from the institution.

Given that all interventions provided customers weekly messages including a weekly
balance report, which is a considerable service by itself, we do not suspect that different
interventions posed a selection bias problem. 12 In fact, the number of customers who
accepted to participate in each condition was very similar around 210.

New Customers: We also offered the savings product to the residents of the major slums
in Nairobi: Kibera, Makina and Mathare. In each slum, we opened a booth with Mbao
fliers on the street and invited people to the booth and explained the product features.
Then we informed potential customers that they are eligible to receive “a promotional
service” as well if they register. In order to ensure fairness, we explained that there are
seven different promotions (based on six treatments and one control condition) and that
they would randomly select their promotion by choosing one of the seven cards with
different shapes handed out to them. Based on the card they picked, we explained to them

12 With the exception of No Intervention condition. But given that those who accepted to participate in the
No Intervention condition could only be more motivated to save, this would only cause the potential
underestimation of the effect of sending reminder messages compared to no intervention. And it does not
pose any problem for estimating the effect of the treatment conditions compared to the control condition.

13
the relevant intervention and asked if they wanted to sign up. Signing up was free and it
only required them to fill out the sign up form with their national identification number
and their mobile phone number. If they signed up we also collected baseline survey data
from them and compensated them for their time with 150 Shillings.

5. Data and Descriptive Statistics

We report the results using three datasets: The baseline and endline survey data we
collected13 and the transaction level deposits data.

Of the 4,596 members we pre-selected for the existing members sample, those we could
reach 14 , who did not contribute to the national social security and who accepted the
promotional service were 2,643. 8 individuals out of this group changed their mind
during the survey and stated that they were not interested in the promotion anymore. In
the remaining sample, from 45 individuals we could not collect any baseline survey data.
We drop these observations. Finally, 190 customers who accepted to participate in the
study faced an issue with the institution within the first 2 weeks after they started to
receive the intervention. In order to eliminate the potential effects of this problem on
savings behavior, we drop these observations as well, noting that including them does not
change the results. We report the survey and savings data of the 2400 remaining
members who participated in the study.

In the slums, we offered the savings product to a total of 1567 individuals. Of these, 89
did not register, and 14 registered but did not want the promotion they were offered. We
surveyed all individuals we offered the product to, however we report the results from the
1464 individuals who registered and accepted the promotion they were offered. We note
that the high rate of the individuals who registered for the product might be due to the
150 Shillings we paid to those who took the baseline survey. Even though we only told
them about the payment after we asked whether they wanted to register or not, it is
possible that word had spread in the community and they already knew about the

13 We didn’t receive the endline survey data yet, but it will be included in the paper very soon.
14 Itis common that people change their phone numbers often, thus some phone numbers we called were
not active anymore or belonged to another person.

14
payment. As we will describe later, results form this sample reveal that only a small
fraction of those who registered actually saved with the product, suggesting that many of
those enrolled had no intention to save using Mbao.

Table 1 shows summary statistics for six demographic variables and two variables related
to income, separate for the two samples. Key differences between the two samples were
that (1) a larger fraction of the new customers sample consisted of female (0.44 versus
0.34), new customers sample was about five years younger than the existing customers
sample and (3) a lower fraction of the new customers were married compared to the
existing customers sample (0.50 versus 0.71). Table 2 and Table 3 show the summary
statistics of the same variables for each condition separately and reports the p-value of an
F-test after regressing the outcome variable on dummies for each condition, omitting the
control condition (Reminder). None of the variables were significantly different across
conditions except for the household size, confirming that our randomization was done
properly.

6. Results

Before we present the results of the experiment, we present the results of a “prediction
study” in which we collected online survey data from 119 individuals recruited on
Amazon M-Turk in exchange for 25 cents. The participants in this study were US
residents of average age 34.8 and 85% had some college degree. In this study we
described the interventions and asked participants to predict the effectiveness of each
treatment. On average participants predicted that the pre-match 20% and ex-post match
20% would yield the highest savings, followed by similar average ranking of pre-match
105 and ex-post match 10%, followed by kids treatment. The treatment predicted to be
least effective was the coin treatment. Below we present the results of the real experiment
and contrast them with these intuitions.

Given the differences between the two samples, we analyze the effects of treatments
separately for each sample. Overall, the response to the interventions was quite different
in each sample: Among the existing customers sample seventy-three percent of the

15
sample saved during the experimental period, however only 15.3 % of the new customers
sample saved any amount.

6.1 Treatment effects in the existing customers sample

We could not deliver the coin to a total of 47 participants in the three conditions
involving it. 11 participants in the “Coin”, 24 participants in the “Pre-match 10% &
Coin” and 12 participants in the “Ex-post match 10% & Coin” treatment did not receive
the coin. To avoid selection bias, we follow the intent-to-treat approach and report
general results by keeping these individuals in the sample. For the kids condition,
summary statistics from Table 1 shows that 81 percent of this sample actually had kids.
We report the results including all observations form this treatment in the analysis as
well.

We first present the results by looking at the total savings per person during the
experimental period. There was one outlier in the data with a particularly higher total
savings than the rest of the sample.15 This observation belonged to the No Intervention
condition. We drop this observation reporting the results in this section because it creates
a very high noise for and masks the underlying pattern of the No Intervention condition.

Figure 2 summarizes the main results from this section and shows the graph of average
savings during the experimental period in each condition. There is a very clear pattern in
this graph: First, average savings in the No Intervention condition is significantly lower
than the Control (Reminder) condition. Second, in all treatments excluding the three that
involve the coin, average savings are very close to the average savings in the control
(Reminder) and the difference between each of these treatments and the control is not
statistically significant. Third, those who only received the coin saved on average the
most among all conditions, 1461 Shillings, where average savings in the control
condition was only 674 Shillings. This implies that on average, the coin more than
doubled average savings. Finally, average savings in all three treatments, which involved
the coin, and only average savings in these treatments, were significantly higher than the

15 This observation was 20 standard deviations higher than the mean. The next highest total savings per
person was 13 standard deviations higher than the mean.

16
average savings in the control. It is interesting that in the two treatments combining ex-
post and pre-match interventions at 10% level with the coin, the average savings were
1,118 and 1,156 Shillings respectively, both of which is lower than the average savings in
the coin treatment. Thus, over the coin there was no additional benefit of adding financial
incentives to the intervention. While it is interesting that combining the coin with
financial incentives lowered the average savings, we hypothesize that this is caused by
the 100 Shillings-weekly cap imposed on the match. Such an unintended anchoring effect
is indeed a potential problem in every matching program (See Madrian, 2012 for a
discussion) and it is particularly interesting that we observe such a negative effect of the
match program in our sample, even though we paid particular attention not to be a victim
of it when we designed the experiment. We further investigate this later by looking at
individual savings patterns in the three coin treatments separately.

In summary, the most important result for this sample is that giving a tangible coin to the
participants and asking them to keep track of their weekly savings influenced their
savings behavior strongly, resulting in statistically and economically significant increase
in savings.

To further analyze this data, in Table 6 we report the results of the OLS regression with
the following model:

is the outcome variable for individual i. where ∈ 1,11 is a

dummy variable for each condition, omitting the control condition (Reminder). The
outcome variable in specification 4 is a dummy taking the value one if a participant saved
a positive amount during the experimental period and zero if he did not save at all. In all
the other specifications, the outcome variable is total savings per person during the
experimental period. Specification (1) conveys the same information as in the Figure 2,
while specifications (2) and (3) take the effect of the outliers into account with the
winsorized samples at top 1% and top 5% respectively. The main result of this part is that
the sharp increase in average savings caused by the Coin treatment is not driven by

17
outliers. Specification (4) reveals that about 73% of those in control condition saved a
positive amount during the experimental period. The share of those who saved positive
was about 7% higher in the two Pre-match conditions, which were marginally significant,
but almost the same in all other conditions except for the No Intervention. In No
Intervention condition, the fraction of savers was 25 percent less than the control,
highlighting the importance of reminders on savings behavior. The effect in this sample is
particularly higher than the effect from Karlan et al. (2012) which might be due to the
fact that the reminder condition in our experiment did not only remind to save but also
provided free balance reports. Finally, the last three specifications look at the effects of
the treatments on savings conditional on saving a positive amount. Since the fraction of
savers was not substantially different among conditions, as expected, the results form
these specifications follow the pattern observed in the first three specifications.

6.2 Treatment effects in the new customers sample

Because there is no such a clear pattern of results in the new customers sample, we only
report the OLS regressions following the same estimation strategy we used for the
existing customers sample in Table 7. In the first specification we see that the average
savings in the control condition is 80.7 Shillings, quite lower than the average savings we
observed in the existing customers sample. The main reason for this result is that the
majority of this sample, 85 percent of those who registered for Mbao, did not save at all
in the six months period. Even though the average savings in the coin treatment is the
highest compared to the control, the difference is not statistically significant and the
second and third specifications imply that it is driven by outliers. Average savings in the
ex-post match 10% treatment is 75 Shillings higher than the average savings in the
control treatment, which is largely driven by the higher fraction of participants saving in
this treatment. An interesting result from these estimations is that those in the kids
treatment saved on average less than those in the control. But given the very low share of
savers it is pre-mature to interpret these results.

We hypothesize that the low share of savers in this sample can be due to our payment of
150 Shillings to those who took the baseline survey and the possibility that those who

18
registered were not interested in saving with Mbao. If our hypothesis holds, the results
from this sample imply that none of the interventions had the needed power of
significantly increasing the interest in this product. However, it is important to note, that
it is also possible that the lack of interest in the Mbao saving plan could extend beyond a
general lack of interest (and an interest only in the 150 Shillings) to a deeper lack of trust
in slums in such financial product.

7. Discussion

In this study, we implemented two psychological interventions along with small financial
incentives in an attempt to increase savings rates on a sample of existing and new
customers of a newly developed savings product in Kenya, called Mbao. In general we
find that among those who had an initial interest in saving with the product, the coin
treatment led to an increase in average savings by more than 100% compared to the
control, thus being the most effective intervention among the many interventions we
tested. Furthermore, in line with previous research concerning the effectiveness of
reminder messages, we also find evidence that reminders are effective in increasing
savings. We found that sending regular reminder and balance report messages increased
average savings about 100% compared to not sending any messages. In contrast, the
small financial incentives we provided did not seem to have a big effect on savings rate in
the existing customer sample. We speculate that the ineffectiveness of these financial
incentives is because for some customers the matching financial incentives offered were
not high enough to change behavior while for other customers they were ineffective
because of the weekly cap we imposed on matching.

Our results have theoretical as well as practical implications and call for future research
to better understand the role of tangible reminders on saving rates, among the poor and
more generally. On the theoretical side, the fact that providing the coin was more
effective than simply reminding to save implies that the design of the way information is
presented can have a significant effect on behavior. Previous research has focused on the
effects of framing the language on behavior, however our finding suggests that there is
more to modify the information presentation to influence behavior. In particular, it

19
suggests that in the context of savings, a tangible representation of overall behavior (coin
representing overall savings) and small activities (scratches representing weekly deposits)
can motivate a behavior with immediate costs and delayed gratification.

Practically, our study suggests that policy makers and product designers could benefit
from this type of low-cost and practical interventions to target positive behavior change.
The results further suggest that the common-sense assumptions about the effectiveness of
policies might fail strongly and emphasizes the benefits of using controlled experiments
to guide policy decisions. The authors in fact speculate that a lot of policies impacting the
lives of millions are based on gut feelings of the few people in charge, and that by
experimentation many of such policies could be improved.

Finally, it would be interesting to explore how to improve the coin we used in this study.
For example, in a world where near field payments such as Apple Pay becomes more
prevalent, could we present a virtual coin displayed on a smartphone to influence savings
behavior of low-middle income individuals in the developed world? As the world moves
to electronic payment systems, the questions of how to represent money and savings
become more open and more important, and understanding how to do it right can be one
of the keys to a better financial future.

20
References

Ashraf, Nava, Dean Karlan, and Wesley Yin, 2006. “Deposit Collectors,” Advances in
Economic Analysis & Policy, 6 (2), Article 5.

, _ , _and _ , 2006 “Tying Odysseus to the Mast: Evidence from a Commitment Savings
Product in the Philippines,” Quarterly Journal of Economics, 121 (1), 635–672.

Bernheim, B. Douglas, Jonathan Skinner, and Steven Weinberg. 2001. "What Accounts
for the Variation in Retirement Wealth Among U.S. Households?" American Economic
Review, September 2001, 91(4), pp. 832-55.

Cadena, Ximena and Antoinette Schoar, 2011. “Remembering to Pay? Reminders vs.
Financial Incentives for Loan Payments,” NBER Working Paper No. 17020, 2011.

Daniel Fernandes, John G. Lynch Jr., Richard G. Netemeyer. 2014. Financial Literacy,
Financial Education, and Downstream Financial Behaviors. Management Science
60(8):1861-1883. http://dx.doi.org/10.1287/mnsc.2013.1849

Esther Duflo, William Gale, Jeffrey Liebman, Peter Orszag, and Emmanuel Saez. 2006.
“Savings Incentives for Low- and Middle-Income Families: Evidence from a Field
Experiment with H & R Block.” Quarterly Journal of Economics 121(4):1311–1346.

Frederick, Shane, George Loewenstein, and Edward D. O’Donoghue, 2001. “Time

Discounting: A Critical Review,” Journal of Economic Literature, XL (2001), 351– 401.

Jack, William, and Tavneet Suri. 2014. “Risk Sharing and Transactions Costs: Evidence
from Kenya’s Mobile Money Revolution” American Economic Review 2014, 104(1):
183–223

Karlan, Dean, McConnell, M., Mullainathan, S., and Zinman, J. 2012. “Getting on the
Top of Mind: How Reminders Increase Savings,” NBER Working Papers 16205,
National Bureau of Economic Research, Inc.

21
Kast, Felipe, Stephan Meier, and Dina Pomeranz. 2012. “Under-Savers Anonymous:
Evidence on Self- Help Groups and Peer Pressure as Savings Commitment Device”.
Working Paper (2012)

Madrian, Brigitte C. 2012. “Matching Contributions and Savings Outcomes: A

Behavioral Economics Perspective.” NBER Working Paper 18220, National Bureau of
Economic Research.

Madrian, Brigitte C. and Shea, Dennis F. "The Power of Suggestion: Inertia in 401(k)
Participation and Savings Behavior." Quarterly Journal of Economics, November 2001,
116(4), pp. 1149-87.

O’Donoghue, Edward D., and Matthew Rabin, 1999. “Doing it Now or Doing it Later,”
American Economic Review, LXXXIX 103–121.

Slovic, Paul. 2007. “If I look at the mass I will never act”: Psychic numbing and
genocide.” Judgment and Decision Making, 2, 79-95.

Stroz, Robert. H. 1955. “Myopia and Inconcsistency in Dynamic Utility Maximization.”

The Review of Economic Studies, (1955-56) 3: 165-180

22
 TABLES

Table 1: Design of the experiment and the sample size

Existing Customers N New Customers N

1. No Intervention 124 1. Reminder (Control) 203
2. Reminder (Control) 204 2. Coin 197
3. Coin 206 3. Kids 201
4. Kids 207 4. Ex-post match 10% 214
5. Ex-post match 10% 208 5. Ex-post match 20% 214
6. Ex-post matching 20% 211 6. Pre-match 10% 223
7. Pre-match 10% 201 7. Pre-match 20% 212
8. Pre-match 20% 210
9. Ex-post match 10% & Kids 209
10. Pre- match 10% & Kids 207
11. Ex-post match 10% & Coin 200
12. Pre- match 10% & Coin 213
Total Sample Size: 2400 Total Sample Size: 1464

23
Table 2: Text Messages

Conditions Reminder Message Report Message

Between <begin> and <end> you
Please don’t forget to save for
saved Ksh<save> with MBAO for
Reminder (Control) your future this week with MBAO
account <id>. Your account
PENSION PLAN.
balance is Ksh<balance>.
message
Coin same as control same as control
1:
You saved with MBAO PENSION
message
PLAN this week. Please scratch
2: if
AROUND <week> on your gold
saved
coin.
You didn’t save with MBAO
message
PENSION PLAN this week.
2: if not
Please scratch BELOW <week>
saved
on your gold coin.
Hi <daddy>, Please deposit as
much as you can this week to
Kids MBAO PENSION PLAN for our same as control
future! Thank you for saving.
<Jane>
Between <begin> and <end> you
Save at least Ksh100 with MBAO
saved Ksh<save> with MBAO for
PENSION PLAN this week to
account <id>. You received
Ex-post match 10% receive Ksh10 extra. If you don’t
<save/10> extra from MBAO.
save, you won’t receive any extra
Your account balance is
money.
Ksh<balance>.
Between <begin> and <end> you
MBAO PENSION PLAN
saved Ksh<save> with MBAO for
deposited Ksh10 in your account.
account <id>. You received
Pre-match 10% Save at least Ksh100 this week to
<save/10> extra from MBAO.
keep it. If you don’t save, MBAO
Your account balance is
will take it back.
Ksh<balance>.

24
Table 3: Summary statistics for key variables in each sample

Existing Customers New Customers

Female 0.342 0.443

(0.010) (0.013)
[2400] [1464]

Age 36.510 31.372

(0.194) (0.259)
[2351] [1464]

Years of education 10.703 10.072

(0.061) (0.079)
[2370] [1463]

Married 0.716 0.502

(0.009) (0.013)
[2376] [1464]

Have children 0.819 0.696

(0.008) (0.012)
[2375] [1464]

Householdsize 3.874 4.280

(0.041) (0.094)
[2371] [1018]

Renter 0.830 0.870

(0.008) (0.009)
[2373] [1464]

Irregular income 0.887 0.833

(0.007) (0.011)
[2219] [1121]
Notes:For each variable, standard errors are presented in
the parantheses and the number of observations are
presented in brackets.

25

All sample No Reminder Coin Kids Pre-match Post-match Pre-match 10% Post-match 10%
Intervention 10% 20% 10% 20% & Kids & Coin & Kids & Coin p-value

Female 0.342 0.363 0.392 0.364 0.343 0.308 0.338 0.356 0.322 0.329 0.305 0.364 0.325 0.829
(0.010) (0.043) (0.034) (0.034) (0.033) (0.033) (0.033) (0.033) (0.032) (0.033) (0.032) (0.033) (0.033)
[2400]
Age 36.510 37.439 36.915 36.722 36.920 37.327 35.800 35.351 36.137 36.468 35.948 36.780 36.755 0.625
(0.194) (0.897) (0.651) (0.656) (0.663) (0.693) (0.664) (0.663) (0.634) (0.651) (0.635) (0.635) (0.713)
[2351]
Years of education 10.703 10.672 10.545 11.020 10.581 10.625 10.816 10.788 10.641 10.704 10.769 10.820 10.426 0.890
(0.061) (0.238) (0.228) (0.209) (0.223) (0.226) (0.186) (0.187) (0.223) (0.199) (0.181) (0.215) (0.227)
[2370]
Married 0.716 0.734 0.729 0.699 0.672 0.725 0.734 0.663 0.758 0.689 0.744 0.716 0.735 0.588
(0.009) (0.040) (0.031) (0.032) (0.033) (0.032) (0.031) (0.033) (0.030) (0.032) (0.030) (0.032) (0.032)
[2376]
Have children 0.819 0.863 0.833 0.825 0.863 0.815 0.778 0.793 0.815 0.806 0.782 0.859 0.816 0.303
(0.008) (0.031) (0.026) (0.027) (0.024) (0.028) (0.029) (0.028) (0.027) (0.028) (0.028) (0.024) (0.028)
[2375]
Householdsize 3.874 3.734 4.192 3.888 4.186 4.065 3.922 3.812 3.941 3.568 3.814 3.778 3.526 0.010
(0.041) (0.178) (0.144) (0.141) (0.141) (0.144) (0.144) (0.142) (0.139) (0.130) (0.133) (0.133) (0.143)
[2371]
Renter 0.830 0.831 0.768 0.811 0.813 0.865 0.807 0.812 0.855 0.840 0.852 0.824 0.888 0.120

26
(0.008) (0.034) (0.030) (0.027) (0.027) (0.024) (0.028) (0.027) (0.025) (0.026) (0.025) (0.027) (0.023)
[2373]
Irregular income 0.887 0.904 0.899 0.882 0.881 0.867 0.897 0.858 0.886 0.918 0.874 0.878 0.910 0.818
(0.007) (0.028) (0.022) (0.024) (0.023) (0.025) (0.022) (0.025) (0.023) (0.020) (0.024) (0.024) (0.021)
[2219]
Notes: For each variabl, standard errors are presented in parantheses. In the first column, total number of observations for each variable is presented in brackets. The last column reports the p-value of the F-statistic
from regressing the variable of interest on condition dummies.

Table 4: Summary statistics for key variables in existing customers sample

All sample Reminder Coin Kids Pre-match Post-match

10% 20% 10% 20% p-value

Female 0.443 0.458 0.416 0.483 0.426 0.458 0.449 0.411 0.760
(0.013) (0.035) (0.035) (0.035) (0.033) (0.034) (0.034) (0.034)
[1464]
Age 31.372 32.296 30.279 30.592 31.704 31.014 32.879 30.734 0.089
(0.259) (0.731) (0.663) (0.662) (0.672) (0.657) (0.762) (0.633)
[1464]
Years of education 10.072 9.813 10.548 10.294 9.991 9.873 10.023 10.005 0.189
(0.079) (0.229) (0.215) (0.197) (0.193) (0.204) (0.198) (0.218)
[1463]
Married 0.502 0.517 0.437 0.547 0.538 0.524 0.505 0.444 0.137
(0.013) (0.035) (0.035) (0.035) (0.033) (0.034) (0.034) (0.034)
[1464]
Have children 0.696 0.724 0.629 0.701 0.704 0.689 0.720 0.701 0.509

27
(0.012) (0.031) (0.034) (0.032) (0.031) (0.032) (0.031) (0.031)
[1464]
Householdsize 4.280 4.449 4.137 4.057 4.433 4.786 4.344 3.727 0.021
(0.094) (0.211) (0.221) (0.174) (0.183) (0.467) (0.187) (0.155)
[1018]
Renter 0.870 0.862 0.878 0.856 0.892 0.854 0.911 0.832 0.195
(0.009) (0.024) (0.023) (0.025) (0.021) (0.024) (0.019) (0.026)
[1464]
Irregular income 0.833 0.832 0.811 0.825 0.856 0.843 0.826 0.835 0.961
(0.011) (0.030) (0.032) (0.031) (0.027) (0.029) (0.029) (0.029)
[1121]
Notes: For each variable, standard errors are presented in parantheses. In the first column, total number of observations for each variable is
Table 5: Summary statistics for key variables in new customers sample

presented in brackets. The last column reports the p-value of the F-statistic from regressing the variable of interest on condition dummies.
Table 6: Regressions with existing customers sample

Fraction of
Savings during trial savers Savings during trial
Conditions (1) (2) (3) (4) (5) (6) (7)

Coin 787.2*** 527.4*** 389.0*** 0.046 970.8*** 636.3*** 460.4***

(211.3) (129.7) (100.2) (0.043) (261.8) (154.8) (116.6)
Kids 162.9 109.0 47.92 0.018 185.9 115.3 36.73
(129.6) (112.1) (89.07) (0.043) (161.8) (138.3) (107.6)
Pre-match 10% 195.4 82.59 55.04 0.081* 161.5 22.41 -7.260
(171.5) (104.9) (85.82) (0.042) (208.7) (125.9) (101.4)
Ex-post match 10% -52.26 -63.51 -80.49 -0.053 -12.50 -28.75 -56.70
(104.2) (100.5) (83.64) (0.045) (134.8) (129.1) (104.4)
Pre-match 20% 122.2 119.0 136.3 0.079* 61.57 57.65 83.50
(97.69) (96.75) (86.95) (0.041) (116.2) (115.1) (101.9)
Ex-post match 20% 82.39 69.95 50.71 0.014 99.55 82.94 57.81
(107.1) (103.1) (87.66) (0.043) (132.7) (127.3) (106.1)
Pre-match 10%& Kids 213.0* 203.8* 180.5** 0.062 202.2 190.7 164.9
(109.4) (106.7) (92.02) (0.042) (131.1) (127.6) (108.4)
Ex-post match 10% & Kids 147.2 61.64 39.10 0.030 162.1 50.36 22.56
(138.1) (103.6) (87.03) (0.043) (173.5) (127.0) (104.8)
Pre-match 10% & Coin 482.9*** 397.8*** 311.9*** 0.044 582.9*** 473.1*** 364.3***
(143.2) (117.9) (93.56) (0.042) (174.8) (141.3) (109.0)
Ex-post match 10% & Coin 444.3*** 376.9*** 325.1*** 0.035 541.7*** 454.2*** 388.9***
(147.2) (117.1) (98.66) (0.043) (180.8) (140.2) (115.6)
No Intervention -407.7*** -407.7*** -371.2*** -0.299*** -364.7*** -364.7*** -315.4**
(89.71) (89.71) (83.28) (0.055) (133.4) (133.4) (126.2)
Constant 674.0*** 674.0*** 637.4*** 0.730*** 910.5*** 910.5*** 861.1***
(71.21) (71.21) (62.92) (0.031) (88.52) (88.52) (77.16)

Observations 2,399 2,399 2,399 2,399 1,810 1,810 1,810

R-squared 0.026 0.033 0.037 0.035 0.028 0.035 0.038
Winsorized 1% 5% 1% 5%
Among savers Yes Yes Yes
Notes: All monetary values are in Kenyan Shillings. 1 USD = 90.3 Kenyan Shllings approximately. Robust
standard errors are in parentheses. Significance levels are *** p<0.01, ** p<0.05, * p<0.1

28
Table 7: Regressions with new customers sample

Fraction of
Savings during trial savers Savings during trial
Conditions (1) (2) (3) (4) (5) (6) (7)

Coin 104.9 17.81 9.323 0.025 516.0 30.08 10.38

(69.31) (29.68) (13.55) (0.037) (356.2) (137.1) (45.28)
Kids -35.82 -32.88 -19.54* -0.028 -167.6 -150.5 -88.88*
(26.45) (23.88) (11.28) (0.034) (154.2) (136.3) (49.64)
Pre-match 10% 6.075 0.665 -3.835 -0.000 40.68 5.157 -24.68
(32.20) (27.24) (12.43) (0.035) (171.7) (138.2) (48.30)
Ex-post match 10% 75.44* 59.36* 25.28* 0.095** 101.9 48.15 -12.83
(41.04) (32.03) (13.94) (0.039) (163.4) (123.4) (41.93)
Pre-match 20% -0.153 -6.923 -4.655 -0.006 22.39 -25.19 -18.55
(32.81) (25.63) (12.46) (0.035) (182.6) (129.2) (48.37)
Ex-post match 20% 8.603 -9.276 -4.245 -0.003 68.76 -51.42 -22.02
(36.73) (25.77) (12.34) (0.035) (208.2) (131.3) (46.74)
Constant 80.72*** 76.29*** 45.76*** 0.153*** 528.6*** 499.5*** 299.7***
(21.18) (18.68) (9.067) (0.025) (108.9) (91.17) (32.95)

Observations 1,464 1,464 1,464 1,464 241 241 241

R-squared 0.008 0.008 0.010 0.010 0.031 0.011 0.019
Winsorized 1% 5% 1% 5%
Among savers Yes Yes Yes
Notes: All monetary values are in Kenyan Shillings. 1 USD = 90.3 Kenyan Shllings approximately.
Robust standard errors are in parentheses. Significance levels are *** p<0.01, ** p<0.05, * p<0.1

29
FIGURES

Figure 1:
Design of the coin

30

Figure 2:
Average savings (calculated as the sum of deposits during 24-week experiment period)
per person in each treatment

31