WPS8633

Policy Research Working Paper 8633

The Limits of Commitment
Who Benefits from Illiquid Savings Products?

Niklas Buehren
Markus Goldstein
Leora Klapper
Tricia Koroknay-Palicz
Simone Schaner

Africa Region
Gender Innovation Lab
October 2018
Policy Research Working Paper 8633

Abstract
Working with a private bank in Ghana, this study examines that clients with above-median baseline overdraft histories
the impacts of a commitment savings product designed do not accrue new savings during the commitment period.
to help clients taking repeated overdrafts break their debt Rather, they draw down other savings to offset the commit-
cycles. Overall, the product significantly increased savings ted amount and take on new debt. In contrast, individuals
with the bank without increasing overdrafts. However, after with below-median overdraft histories significantly increase
accounting for other sources of savings, the study finds savings during and after the commitment period.

This paper is a product of the Gender Innovation Lab, Africa Region. It is part of a larger effort by the World Bank to
provide open access to its research and make a contribution to development policy discussions around the world. Policy
Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted
at mgoldstein@worldbank.org.

The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.

Produced by the Research Support Team
The Limits of Commitment: Who Beneﬁts from Illiquid
Savings Products?∗
Niklas Buehren Markus Goldstein Leora Klapper
Tricia Koroknay-Palicz Simone Schaner†

∗
We thank Emmanuel Apiors, Virginia Ceretti, and Ervin Dervisevic for excellent research assistance,
Innovations for Poverty Action Ghana for expert ﬁeld implementation, and Robert Osei for helpful comments
and advice. This project would not have been possible without the support of the North Volta Rural Bank and
its staﬀ. We are particularly indebted to Patrick Ata and S. S. Mohenu, who provided invaluable advice and
leadership throughout the project. This paper is an output of the Africa Gender Innovation Lab. Financial
support for this study was provided by the World Bank’s Umbrella Facility for Gender Equality (UFGE), the
World Bank’s Development Research Group Finance and Private Sector Development Team’s Knowledge
for Change Program (KCP II), the World Bank’s Africa Region’s Vice President’s Oﬃce, the World Bank’s
Ghana Country Management Unit, and the Bill and Melinda Gates Foundation as part of their work on
ﬁnancial inclusion. The study protocol was approved by IRBs at the University of Ghana, Legon (CPN
009/13-14) and Dartmouth College (CPHS # 24177) and registered in the AEA RCT registry (AEARCTR-
0001630) and the International Initiative for Impact Evaluation’s Registry for International Development
Evaluations (study ID 5409d4c83af40). The ﬁndings, interpretations, and conclusions expressed in this
paper are entirely those of the authors. They do not necessarily represent the views of the International
Bank for Reconstruction and Development/World Bank and its aﬃliated organizations, or those of the
Executive Directors of the World Bank or the governments they represent.
†
Buehren, Goldstein, Klapper, and Koroknay-Palicz: The World Bank. Schaner: The Uni-
versity of Southern California, NBER, and BREAD. Buehren: nbuehren@worldbank.org, Goldstein:
mgoldstein@worldbank.org, Klapper: lklapper@worldbank.org, Koroknay-Palicz: tgonwa@worldbank.org,
Schaner: schaner@usc.edu.
1 Introduction
Temporary short-term loans, such as checking account overdrafts or payday credit, are
double-edged swords: while easy access to such credit may help individuals make ends meet
when faced with shocks (Morse, 2011; Islam and Maitra, 2012), there is also a risk that
high-cost debt triggers further ﬁnancial distress (Melzer, 2011) and draws people into debt
cycles. The risk of negative consequences is particularly acute when individuals exhibit be-
havioral biases like hyperbolic discounting, which could lead individuals to continuously put
oﬀ paying down high-cost debt in order to enjoy more consumption in the present (Angeletos
et al., 2001). When agents are time inconsistent, there is scope for ﬁnancial products that
oﬀer commitment to restrict current spending and improve long-run welfare.
Yet the success of commitment in such a context is by no means assured: individuals
who are (partially) naïve about their self-control problem may not sign up, or sign up for
too little commitment (Laibson, 2015). Furthermore, individuals who do choose to commit
may take costly measures to undo their commitments in the future. This could take the
form of drawing down other savings, paying fees to opt out of the commitment product, or
taking on additional debt to support current-period consumption. Ultimately, assessing the
extent to which commitment products increase welfare and build net savings is an empirical
question. Given the risk that commitment crowds out other savings or crowds in debt, it
is particularly important to collect comprehensive data on consumer balance sheets when
evaluating commitment products – yet rigorous evaluations that do this are surprisingly rare,
especially in developing country contexts.1
We conducted a randomized trial with salaried workers in Ghana to ﬁll this evidence
gap and shed light on whether commitment has the potential to lift individuals out of debt
cycles. All the workers had access to an overdraft facility on the bank account linked to
their salary payments, and many made regular use of the facility at a substantial cost. Half
the workers were randomly oﬀered a novel savings product that automatically deducted
a pre-speciﬁed savings contribution from their monthly paycheck. The contribution cycle
lasted for 18 months, after which all contributions, plus a “completion bonus” equal to one
month’s contribution, were released to the saver. Workers faced a penalty if they withdrew
any money before the end of 18 months. The product was designed with debt cycles in mind:
by building up a sizable lump sum with the partner bank, users could release themselves
from an overdraft debt cycle at the end of the commitment period.
1
Accounting for crowd out is important, as even the very poor typically have access to some type of
liquidity (Banerjee and Duﬂo, 2007; Collins et al., 2009). Note that access to liquidity in and of itself is not
a sign that commitment will fail, especially when individuals have precautionary savings motives – in this
case, people can be pushed to a corner before drawing down their liquidity (Laibson, 1997).

2
We combine administrative data from the bank with six waves of detailed survey data
to study how commitment impacted respondents’ ﬁnancial behavior – both at the partner
bank and more broadly. First, we ﬁnd robust demand for commitment. Seventy-two percent
of individuals in the treatment group signed up for the product, and just 13 percent of those
who signed up dropped out before the 18-month savings cycle ended. Second, the product
more than doubled balances at the partner bank both during and after the commitment
period. We also observe a marginally signiﬁcant 6 percentage point (18 percent) decline in
overdrafts in the six months following the release of the commitment amount. In the longer
run, however, treatment eﬀects on overdrafts disappear.
Our third main result underscores the importance of taking a broader view of consumers’
ﬁnances when evaluating savings products. Although the commitment product did not crowd
out other cash savings, we do observe a signiﬁcant increase in the share of respondents who
took on new debt during the commitment period. Point estimates on debt stocks are roughly
equal to point estimates on total savings, and as a result our treatment estimates on net
savings (savings less debt) are close to zero and not statistically signiﬁcant.
This interpretation changes dramatically when studying heterogeneity in treatment ef-
fects by baseline overdrafter status, however. “Heavy overdrafters” (those with above median
overdraft histories at baseline) did save more at the partner bank during the commitment
period, but this treatment eﬀect disappears after the commitment amount is released. More-
over, the product signiﬁcantly decreased heavy overdrafters’ other formal sector savings dur-
ing the commitment period while increasing the rate of debt taking. These results are con-
sistent with the hypothesis that heavy overdrafters are naïve hyperbolic discounters, who
end up undoing their commitment by borrowing. “Light overdrafters” (those with below-
median histories) exhibited a very diﬀerent response: the commitment product signiﬁcantly
increased other formal sector savings both during and after the commitment period, with no
signiﬁcant impact on debt. This suggests that light overdrafters beneﬁted from the savings
product through a channel other than time inconsistency.
The main contribution of our paper is to explore the extent to which commitment is
crowded out by other sources of liquidity, while directly exploring the pitfalls of commitment.
To date, much of the research on behavioral savings products in development economics
focuses on savings held at partner ﬁnancial institutions, or savings without accounting for
debt (for a review of recent work see Karlan et al. 2014). Our ﬁndings highlight the need for
future work to cast a broader net when assessing the promise of commitment for improving
the ﬁnancial lives of the poor.2 Our work also builds on John (2017), who ﬁnds that over
2
In this way our paper is related to a growing literature that seeks to understand the net eﬀects of
retirement savings programs in developed country contexts (Poterba et al., 1996; Benjamin, 2003; Engelhardt

3
half of individuals who sign up for a similar contribution-based commitment product end up
paying a fee to drop out early. We go beyond this to show that negative eﬀects need not be
limited to program dropout – in our case, agents turn to debt to oﬀset savings commitments,
which is costly and runs the risk of propagating debt cycles. We also demonstrate that
negative side eﬀects are concentrated among people who exhibit behavior most consistent
with self-control problems at baseline. This suggests that when individuals are naïve about
the extent of their self-control problem, commitment may do more harm than good.
Another contribution of our paper is to show that illiquid savings products can improve
individuals’ ﬁnancial lives for reasons unrelated to a direct commitment eﬀect. Here, our
work relates to Brune et al. (2016), who ﬁnd that a commitment savings product increased
savings, agricultural investment, agricultural output, and consumption of Malawian farmers
even though amounts actually committed were quite low. We build on this by showing that
beneﬁts accrue to individuals least likely to be liquidity constrained, and by showing that
the commitment product crowds in other savings both during and after the commitment
period.
The remainder of the paper proceeds as follows: We ﬁrst describe our sample and the
experiment in Section 2. We present our results in Section 3, which includes an analysis
of overall impacts, heterogeneity by baseline overdraft status, and discussion. Section 4
concludes.

2 Experimental Context and Design
2.1 Partner Bank and Study Sample
The study sample is comprised of 320 salaried workers receiving electronic salary payments
into bank accounts at the North Volta Rural Bank (NVRB).3 NVRB is a small bank that
mostly serves individuals living in rural areas, where the majority of economically active
individuals are self-employed in agriculture (Ghana Statistical Service, 2013).
By virtue of being salaried workers, our sample is better oﬀ than the average citizen in the
bank’s catchment area. Column 1 of Appendix Table A1 illustrates baseline characteristics
in the control group: three-quarters of study participants were men, two-thirds had a post-
secondary education, and 90 percent were government employees, with teacher the most
common occupation.4 In comparison, across Ghana and in the Volta Region, approximately
and Kumar, 2007; Gelber, 2011; Chetty et al., 2014; Beshears et al., 2017).
3
Appendix A provides additional detail on sampling and inclusion/exclusion criteria.
4
The remaining 10 percent of workers are NVRB employees. Our results are unchanged if we drop these
individuals from the analysis.

4
6 percent of the economically active population is employed by the government and just 8
percent of individuals aged 15 and older have a post-secondary education (Ghana Statistical
Service, 2013, 2012). Our respondents reported an average monthly income of GHS 770,
or $358 at a 2013 exchange rate of GHS 2.15 per $1, had GHS 240-250 in savings with
NVRB, and GHS 1,628 in cash savings across all sources.5 Although our respondents have
desirable jobs, they are still economically vulnerable: one-third of the respondents reported
that children in their care miss school due to late school fees “often” or “sometimes”, 71
percent reported that they often or sometimes worry about having enough money to pay
monthly bills, and 61 percent stated that they are “socially taxed”, in that they often or
sometimes have diﬃculty saying no to requests from others.
Salaried workers are valuable bank customers in Ghana. Regularly-recurring salary pay-
ments allow banks to reliably deduct loan fees, interest, and principal directly from the
customer’s account, and bank accounts oﬀering salary advances or overdraft facilities to cus-
tomers who receive electronic payments are common. At NVRB, salaried customers may
receive an overdraft by completing a requisition form at the bank branch. NVRB then per-
mits the customer to take a cash withdrawal equal to the overdraft amount, which turns
the customer’s bank balance negative. At the time of the experiment, overdrafts incurred
a ﬁxed GHS 5 service fee plus a variable fee equal to 18 percent of the overdraft amount.
Since most overdrafts are cleared in less than one month, these loans have exceptionally high
eﬀective interest rates.6 Workers in our sample made extensive use of overdrafts at baseline:
Fifty-eight percent of respondents overdrafted at least once in the year before baseline, with
the average overdrafter paying GHS 226 ($105) in fees and interest over the same period.
Appendix Figure A1 graphs the distribution of overdrafts.
Individuals who use costly, short-term credit are particularly likely to have time incon-
sistency problems (Meier and Sprenger, 2010; Gathergood, 2012). As such, we stratiﬁed
our randomization by baseline overdrafting behavior (as well as gender and bank branch)
because we were interested in studying heterogeneity in treatment eﬀects by overdrafter sta-
tus.7 Before randomizing we constructed an index of overdrafting behavior and deﬁned those
5
Total baseline cash savings include savings at NVRB, savings at other banks, microﬁnance institutions,
and credit unions, savings with group savings clubs, savings with agricultural co-ops, savings held by susu
collectors, money lenders, other individuals, and cash kept at home.
6
Seventy-six percent of the overdrafts taken out by our study sample were cleared within 31 days and 90
percent were cleared in 93 days. As a matter of policy, NVRB only formally approves overdrafts for one or
two months.
7
The baseline survey also asked respondents to make hypothetical choices between diﬀerent amounts of
money at diﬀerent times, but there were only two opportunities for respondents to make a quasi-hyperbolic
preference reversal and choices were unincentivized. In practice, just 3.8 percent of respondents made two
reversals, while 20 percent of respondents made one reversal. Given the weaknesses of the survey-based
measure, we do not use it in the main analysis and instead focus on heterogeneity in overdrafting behavior.

5
with an above-median score to be “heavy overdrafters”.8
Table 1 illustrates diﬀerences in baseline characteristics by overdrafter status. The ﬁrst
row shows that, on average, light overdrafters only took an overdraft in 6 percent of the
months in the year prior to the baseline, while heavy overdrafters were in overdraft more
than 50 percent of the time. Heavy overdrafters were more likely to be men, more likely
to be teachers, less likely to work in the private sector, and had more children and lived in
larger households. Although they had similar incomes to light overdrafters, they reported
roughly GHS 400 less in savings and GHS 1,200 more in total debt and were 12 percentage
points more likely to report diﬃculty paying bills.

2.2 Experimental Design
NVRB worked with the research team to design a new savings product called Salary Susu
Plus (SSP) for the purposes of the study. Similar services were not oﬀered by other banks
in the study area at the time. SSP participants speciﬁed a monthly contribution amount,
which would be deducted from future salary deposits for a period of 18 months. The monthly
contribution had to be equal to at least GHS 30 and could not be changed after it was set.
Customers could not access their contributions during the savings cycle unless they opted
out of the program at a cost of one month’s contribution. Customers who completed the
commitment cycle received all their contributions plus an additional bonus payment equal to
one month’s contribution (or 5.6 percent of the total contribution amount). This amounts
to an annual percentage yield (APY) of 7 percent, assuming monthly compounding. By
way of comparison, NVRB oﬀered no interest on current accounts and a 3 percent APY on
savings accounts at the time of the baseline. While SSP oﬀered an attractive rate of return
relative to NVRB’s other oﬀerings, the real return on savings was still negative, since annual
inﬂation ranged from 10-19 percent over the course of our study period (Bank of Ghana,
2014, 2017).
Half the sample in each overdrafter×branch×gender stratum was assigned to the treat-
ment group, which was oﬀered the opportunity to sign up for SSP. Appendix Table A1 con-
ﬁrms that baseline characteristics are balanced across treatment and control groups, both
overall and when the sample is split by overdrafter status.
Appendix Figure A2 illustrates the timeline of all study activities. SSP debits began in
8
Overdraft scores were constructed for each person using their current account transaction records avail-
able at the time of the randomization. The score was calculated as the fraction of balance entries that were
negative over the life of the account. The sample was sorted from lowest score (never had a negative balance)
to highest (negative balance 66 percent of the time).

6
December 2013, continued through May 2015, and payouts were made to SSP clients in June
2015.9

2.3 Data
Our analysis combines administrative data from NVRB with six waves of survey data. The
administrative data contain information on all transactions posted to respondents’ NVRB
accounts from the time of account opening to one year after the end of the intervention (32
months after baseline).
The baseline took place before SSP was introduced. The ﬁrst three follow-ups took place
while SSP was ongoing – during program months 6, 10, and 14. The objective of these
rounds was to collect data on savings and debt during the commitment period, and assess
how commitment impacted expenditure. The last two follow-ups took place 3 and 5 months
after the SSP payout. These surveys were designed to assess how respondents’ ﬁnancial lives
had changed after the commitment amount was released.10
In total, 91 percent of study participants participated in all survey rounds and 98 percent
of study participants participated in at least one follow-up survey round. Appendix Table
A2 veriﬁes that attrition is uncorrelated with treatment status.

3 Main Results
3.1 Take Up
Seventy-two percent of individuals in the treatment group signed up for SSP. The median
monthly contribution was GHS 30, while the average contribution was GHS 42. Completion
rates were also high, with just 13 percent of respondents who signed up for SSP dropping
out before the savings cycle was over. Appendix Table A3 studies correlates of take-up (in
the treatment group) and dropout (among those in the treatment group who signed up).
The ﬁrst two columns study pairwise correlations between take-up/dropout and baseline
characteristics, while the last two columns present results of regressions where all baseline
9
All study participants – in both the treatment and the control group – who did not have a savings
account with NVRB were given the opportunity to open one in May 2015. Treatment group individuals
were signiﬁcantly more likely to open a savings account, largely because the bank stated to customers that
SSP deposits were meant to be sent to a savings account. As a result the treatment group had 0.45 more
accounts with NVRB than the control group by endline. This should be kept in mind when interpreting
treatment eﬀects in the post-intervention period.
10
Additional details on each survey round, including timing, participation, and modules contained, can be
found in Appendix A.

7
characteristics are entered simultaneously. SSP had broad appeal, with few baseline char-
acteristics predicting take-up and dropout. Strikingly, while heavy overdrafters were just as
likely as light overdrafters to sign up for SSP, they were 12 percentage points more likely to
drop out (column 2). This is consistent with the hypothesis that some heavy overdrafters
are partially-naïve and underestimate their future desire to undo commitments.

3.2 Direct Eﬀects: Results from Administrative Data
We now analyze how SSP impacted study participants’ use of NVRB accounts. Here, we
focus on use of overdrafts, overdraft fees paid, net deposits (total deposits minus total debits,
a measure of savings ﬂow), and the average daily account balance (a measure of savings
stock).11 Our administrative data span the commitment period, the post-commitment period
covered by follow-up surveys, and a post-commitment, post-survey period. We therefore
estimate separate treatment eﬀects for these three periods to understand how eﬀects evolve
over time.12 We use the following regression speciﬁcation:

yit = β1 treati × duringt + β2 treati × af ter1t + β3 treati × af ter2t + β4 yi0 + γs + δt + εit (1)

Where yit is the outcome of interest for account holder i in month t, treati is a dummy variable
equal to one if account holder i is in the treatment group, duringt identiﬁes the commitment
period, af ter1t identiﬁes the post-commitment, follow-up survey period, af ter2t identiﬁes
the post-commitment, post-survey period, γs are strata ﬁxed eﬀects, δt are month ﬁxed
eﬀects, and yi0 is the average outcome of interest, measured in the year before the survey.
The ﬁrst two columns of Table 2 show that during the commitment period SSP had no
impact on overdrafts. After SSP savings were released, the share of accounts in overdraft
declined by 6 percentage points (18 percent), while overdraft charges declined by GHS 3.77
(21 percent) per month – both signiﬁcant at the 10 percent level. In the longer run, however,
participants slid back into debt cycles: we ﬁnd no treatment eﬀect on overdrafts 7-12 months
after the commitment amount was released.
During the commitment period, SSP increased net deposits by GHS 19.5 (column 3 of
Table 2). The 72 percent takeup rate implies a treatment-on-the-treated eﬀect of GHS 27,
roughly two-thirds of the average SSP contribution of GHS 42 – thus, there is evidence
that participants were able to partially undo their commitments by drawing down other
11
When constructing variables we sum across all accounts each respondent holds with NVRB. Note that
average daily balances can be negative, e.g. when an account holder takes an overdraft.
12
The commitment period lasted from December 2013 to May 2015, the post-commitment, follow-up
survey period covered June to November 2015, and the post-commitment, post-survey period covers six
months from December 2015 to May 2016.

8
savings at NVRB. Column 3 also shows that the treatment eﬀect on net deposits dropped
to zero during the follow-up, post-commitment period. This is not surprising, since many
participants withdrew their SSP savings during this period. Interestingly, in the post-survey,
post-commitment period the treatment eﬀect on net deposits (GHS 32.5) exceeds the com-
mitment period treatment eﬀect. This suggests that SSP had persistent eﬀects on savings
behavior at NVRB, subject to the caveat that the point estimate is not signiﬁcantly diﬀerent
from zero.
Finally, column 4 shows that SSP generated a signiﬁcant and highly persistent increase
in NVRB balances: the average daily balance during the commitment period increased by
GHS 244, and remained GHS 140-141 higher during the two post-commitment periods.
These treatment eﬀects are large relative to control groups means ranging from GHS 90-111
across the three focal time periods.

3.3 Overall Eﬀects: Results from Survey Data
We now turn to the survey data to assess whether impacts on NVRB balances reﬂect new
savings or crowd out. We use a regression speciﬁcation that mirrors equation 1, except t
now references survey round, the interaction between treatment and the post-commitment,
post-survey period is dropped, and yi0 references values measured at the baseline survey.
The ﬁrst two columns of Table 3 present treatment eﬀects on NVRB savings. The ﬁrst
column uses the same administrative data used in the earlier section, but focuses on NVRB
savings on the day of the relevant survey. In order to make the measure comparable with
survey reports, negative NVRB account balances are coded as zero NVRB savings. The
second column studies savings based on respondent self-reports. Both variables paint a
similar picture: SSP more than doubled NVRB savings during the commitment period.
The administrative data also show a marginally signiﬁcant impact on NVRB savings in the
post-commitment period, while the point estimate for the survey data is smaller and not
signiﬁcantly diﬀerent from zero. Columns 3 and 4 show that SSP had no signiﬁcant impact
on other formal or informal savings.
Individuals may also oﬀset commitments by taking on more debt. Columns 6 and 7 of
Table 3 study SSP’s impacts on debt.13 Here, we ﬁnd a signiﬁcant, 7.6 percentage point (40
percent) increase in the likelihood of taking on new debt during the commitment period.
This could reﬂect increased ﬁnancial strain driven by the regular monthly commitment.14
Recall that we do not see any parallel impacts on the overdraft rate or overdraft fees at
13
Appendix Table A4 shows impacts on debt by source.
14
We do not, however, ﬁnd that SSP increased self-reported markers of ﬁnancial strain, like food insecurity
and diﬃculty paying bills (Appendix Table A5).

9
NVRB during the commitment period (Table 2). This could be for one of two reasons:
either marginal debtors have access to cheaper forms of credit than overdrafts, or marginal
debtors are already maxed out on overdrafts.
While the eﬀect of SSP on the debt stock is not signiﬁcantly diﬀerent from zero, the point
estimate is economically meaningful and very close in magnitude to the treatment eﬀect on
savings. As a result, we ﬁnd no evidence that SSP increased savings net of debt (Table 3,
column 8). One caveat here is that our survey-based estimates – while close to zero – are
noisy. The 95 percent conﬁdence interval for the impact of SSP on savings net debt easily
includes the analogous point estimate on savings with NVRB.

3.4 Heterogeneity by Baseline Overdrafting
The main eﬀects mask strikingly diﬀerent treatment eﬀects by baseline overdrafter status.
This can be seen in Figure 1, which graphs monthly treatment eﬀects on the average daily
balance at NVRB by overdrafter status. Here, the ﬁrst vertical dashed line demarcates the
payout month and the second vertical line demarcates the end of survey coverage. Panel
B shows that the treatment eﬀect for heavy overdrafters peaked roughly six months before
the SSP payout. This could reﬂect a combination of program dropout (19 percent of heavy
overdrafters dropped out early, as compared to 7 percent of light overdrafters) and draw-
down of liquid, non-committed NVRB balances. In contrast, the treatment eﬀect for light
overdrafters peaked the month before the payout (Panel A). Moreover, treatment eﬀects for
light overdrafters re-emerge in the post-endline period after dropping immediately after com-
mitment amounts are released. These patterns are consistent with the hypothesis that light
overdrafters became accustomed to making smaller withdrawals from their NVRB accounts
and continued with this behavior after the SSP program ended.15
These stark diﬀerences in treatment eﬀects by baseline overdraft behavior are also ap-
parent in the survey data. We examine heterogeneous treatment eﬀects by augmenting
equation 1 to include interactions between the diﬀerent treatment eﬀects and either a “heavy
overdrafter” (above median overdraft score at baseline) or “light overdrafter” (below me-
dian overdraft score) dummy.16 Table 4 shows that during the commitment period, light
overdrafters saved more both with NVRB and with other banks and credit unions – com-
mitment crowded additional savings into the banking sector. While there is some crowd out
of informal savings during the commitment period, the point estimate is small compared to
treatment eﬀects on formal savings. Consequently, we reject the null of no impact on total
15
In contrast, treatment eﬀects on overdrafts are similar for the two groups – see Appendix Table A6 for
detail.
16
We also include interactions between overdrafter status and time period.

10
savings at the 5 percent level, both during and after the commitment period. The point
estimates for both periods are just short of GHS 600, very close to total program savings of
the median SSP participant (GHS 570).
In contrast, heavy overdrafters saved signiﬁcantly less in other formal sector accounts,
more than oﬀsetting the additional savings stored with NVRB – as a result, point estimates
on total savings are negative and not statistically signiﬁcant. We reject equality of treatment
eﬀects on total savings for heavy and light overdrafters at the 1 percent level both during
and after the commitment period.
Columns 6-8 of Table 4 show that this heterogeneity persists when accounting for debt.
Only heavy overdrafters took on signiﬁcantly more debt during the commitment period. As
a result, treatment eﬀects on savings net debt are large and negative for heavy overdrafters,
while treatment eﬀects on savings net debt are similar to treatment eﬀects on total savings
for light overdrafters. Here, we reject equality of eﬀects by overdrafter status at the 5 percent
level during and after the commitment period.
Given that SSP helped light overdrafters save more without taking on new debt, where
did the additional money come from? Appendix Table A8 presents impacts on income and
expenditure. Given the modest monthly changes needed to generate our point estimates, we
cannot deﬁnitively identify consumption or income as the source of the new savings, although
We do ﬁnd some signiﬁcant survey-based evidence that light overdrafters in the treatment
group earn more. We view this result with caution, as we only collected income data in
one follow-up survey. Table A8 also helps rule out the hypothesis that heavy overdrafters
leverage SSP payouts and new debt to make lumpy investments in the post-commitment
period: SSP had no signiﬁcant impact on investment for either group.

3.5 Mechanisms and Discussion
Overall, our results for heavy overdrafters are consistent with a model of partially-naïve
hyperbolic discounting, e.g. where workers sign up for SSP expecting that it will help with
their commitment problems, but fail to account for the fact that they will resort to costly
borrowing to undo their commitment in the future. The behavior of light overdrafters is
more diﬃcult to rationalize. What is clear is that their response to SSP is not consistent
with standard models of hyperbolic discounting, where savings outside SSP should weakly
decrease during the commitment period. How, then, did SSP catalyze savings among light
overdrafters?
One possibility is that SSP helped light overdrafters manage ﬁnancial demands from
extended family or community members. This type of “other control” problem could be

11
particularly important in our context, where salaried workers are often asked to support less
well-oﬀ relatives – recall that roughly 60 percent of our sample reported diﬃculty saying no
to requests from others at baseline. In theory, SSP could have had a direct eﬀect on transfer
requests, e.g. if SSP reduced liquidity available for transfers, and an indirect eﬀect, e.g. if
signing up for SSP gave respondents an accepted justiﬁcation to deny transfer requests. In
practice, the indirect channel is needed to explain our results for light overdrafters, given
that the treatment crowds in other formal sector savings and that treatment eﬀects persist
into the post-commitment period. We ﬁnd no strong evidence that other control is driving
heterogeneity by overdrafter status: Appendix Table A8 shows that the treatment had no
signiﬁcant eﬀect on money spent on others, and heterogeneity by other control does not
mirror heterogeneity by overdrafter status (Appendix Table A9).
Alternatively, psychological mechanisms could have catalyzed the SSP treatment eﬀects
for light overdrafters. For example, the product may have helped light overdrafters build
better savings habits – the fact that treated light overdrafters continued to save more with
NVRB after withdrawing the SSP lump sum is consistent with this channel. A related
hypothesis is that SSP helped individuals “learn” how to save, or helped them learn that
saving was less diﬃcult than they initially thought. The spillover eﬀects on other types of
saving are consistent with these theories, which could transition individuals to an entirely
new savings equilibrium (Becker and Murphy, 1988).
An alternative way to rationalize crowd in would be if individuals have lumpy savings
goals with high returns/marginal utility (e.g. a business asset or a high-value durable good
like a vehicle), but there is some uncertainty over whether the savings goal can be met. If
SSP increased the probability of reaching the savings goal, this could increase the return
to saving in other places as well. We do have some suggestive evidence that SSP helped
respondents reach savings goals. First, 69 percent of people who signed up for SSP reported
that they did so with a speciﬁc goal in mind. Second, experience with SSP led participants
to positively update their assessment of the product’s usefulness for meeting goals.
Appendix Table A10 reports on how treatment impacted respondents’ post-commitment
assessments of SSP, both overall and by baseline overdrafter status.17 In the control group,
21 percent of respondents strongly agreed that SSP would help to meet savings goals and
19 percent of respondents strongly agreed SSP would help build money for a large expense.
Assignment to the treatment group increased these assessments by 14 and 19 percentage
17
In our ﬁnal survey round, we asked all respondents whether they thought SSP was useful for a range of
purposes. SSP users answered based on their experience, while the control group had the product explained
to them and were then asked to speculate as to whether the product would be useful for the given purposes.
Here we focus on the share of respondents who “strongly agree” that SSP helps with a given aim, since the
default response for most respondents was “agree”.

12
points respectively, with similar eﬀects for both heavy and light overdrafters. SSP also
signiﬁcantly increased the share of individuals who strongly agreed that the product helps
manage transfer requests from people within the household (column 1) – this is broadly
consistent with results in Appendix Table A9.
However, we cannot reject the null that the eﬀects in Appendix Table A10 are the same
for heavy and light overdrafters. Why are eﬀects on downstream outcomes so diﬀerent by
overdrafter status when perceived beneﬁts of the product are so similar? One possibility is
that diﬀerential treatment eﬀects are not driven by underlying behavioral mechanisms so
much as they are driven by diﬀerences in ﬁnancial fundamentals. Consider, for example,
the “lumpy savings goal” hypothesis: in order for SSP to have an impact on the likelihood
of meeting a savings goal, a saver’s baseline probability of meeting a goal would have to be
neither too low nor too high. It may be that light overdrafters, by virtue of being in a better
savings position to begin with, were more likely to be marginal savers.
Ultimately, data constraints prevent us from testing the hypotheses discussed in this
section more directly – we therefore leave further investigation of these questions to future
research.

4 Conclusion
We use six rounds of detailed survey data as well as administrative bank data to study the
impact of a commitment savings product for salaried workers in Ghana. Overall, the product
increased savings at the partner bank, both during and after the commitment period, with
limited crowd out of cash savings held in other places. These savings gains are oﬀset by an
increase in debt – as a result, the point estimate on net savings is small and insigniﬁcant.
This suggests that people with access to liquidity may fail to beneﬁt from commitment
savings products because they can undo their commitment with other ﬁnancial instruments.
However, we ﬁnd divergent results when we examine heterogeneity in treatment eﬀects
by baseline propensity to take overdrafts with our partner bank. Heavy overdrafters – who
had less liquid savings and more debt at baseline – see (marginally signiﬁcant) declines in
net savings when oﬀered the product, while light overdrafters – who had considerably better
access to liquidity – save more, both with the partner bank and in other savings devices,
and do not take on more debt. Strikingly, these treatment eﬀects manifest both during and
after the commitment period.
Our results for heavy overdrafters are consistent with the hypothesis that time incon-
sistent agents make ineﬃcient savings commitments, especially when they have access to
liquidity that allows them to unwind their commitments. The results for light overdrafters

13
present a much bigger puzzle, especially because commitment crowded in liquid savings
during the commitment period. Although optimally chosen commitment can increase the
savings of time inconsistent agents, theory predicts that this will be at the expense of liquid-
ity held elsewhere. Hence, our light overdrafter results underscore that commitment savings
products can beneﬁt consumers for reasons unrelated to time inconsistency.
What remains less clear is why commitment products beneﬁt savers for other reasons.
While we cannot isolate a single mechanism for our eﬀects, we do identify several potential
channels, including habit formation and learning. Ultimately, additional research is needed to
explore the multiple pathways through which illiquid savings products transform individuals’
ﬁnancial lives.

References
Angeletos, G.-M., D. Laibson, A. Repetto, J. Tobacman, and S. Weinberg (2001). The
Hyperbolic Consumption Model: Calibration, Simulation, and Empirical Evaluation. The
Journal of Economic Perspectives 15 (3), 47–68.

Banerjee, A. V. and E. Duﬂo (Winter 2007). The Economic Lives of the Poor. The Journal
of Economic Perspectives 21 (27), 141–167.

Bank of Ghana (2014). Statistical Bulletin: March 2013. Technical report, Bank of Ghana
Statistics Department.

Bank of Ghana (2017). Monthly Statistical Bulletin: December 2016. Technical report, Bank
of Ghana Statistics Department.

Becker, G. S. and K. M. Murphy (1988). A Theory of Rational Addiction. The Journal of
Political Economy , 675–700.

Benjamin, D. J. (2003). Does 401 (k) Eligibility Increase Saving?: Evidence from Propensity
Score Subclassiﬁcation. Journal of Public Economics 87 (5), 1259–1290.

Beshears, J., J. J. Choi, D. Laibson, B. C. Madrian, and W. L. Skimmyhorn (2017). Bor-
rowing to Save? The Impact of Automatic Enrollment on Debt.

Brune, L., X. Gine, J. Goldberg, and D. Yang (2016). Facilitating Savings for Agricul-
ture: Field Experimental Evidence from Malawi. Economic Development and Cultural
Change 64 (2), 187–220.

14
Chetty, R., J. N. Friedman, S. Leth-Petersen, T. H. Nielsen, and T. Olsen (2014). Active
vs. Passive Decisions and Crowd-Out in Retirement Savings Accounts: Evidence from
Denmark. The Quarterly Journal of Economics 129 (3), 1141–1219.

Collins, D., J. Morduch, S. Rutherford, and O. Ruthven (2009). Portfolios of the Poor: How
the World’s Poor Live on Two Dollars a Day. Princeton University Press.

Engelhardt, G. V. and A. Kumar (2007). Employer Matching and 401 (k) Saving: Evidence
from the Health and Retirement Study. Journal of Public Economics 91 (10), 1920–1943.

Gathergood, J. (2012). Self-Control, Financial Literacy and Consumer Over-Indebtedness.
Journal of Economic Psychology 33 (3), 590 – 602.

Gelber, A. M. (2011). How Do 401(k)s Aﬀect Saving? Evidence from Changes in 401(k)
Eligibility. American Economic Journal .

Ghana Statistical Service (2012). 2010 Population and Housing Census: Summary of Final
Results.

Ghana Statistical Service (2013). 2010 Population and Housing Census: Regional Analytical
Report: Volta Region. Technical report.

Islam, A. and P. Maitra (2012). Health Shocks and Consumption Smoothing in Rural House-
holds: Does Microcredit Have a Role to Play? Journal of Development Economics 97 (2),
232 – 243.

John, A. (2017). When Commitment Fails - Evidence from a Field Experiment.

Karlan, D., A. L. Ratan, and J. Zinman (2014). Savings by and for the Poor: A Research
Review and Agenda. Review of Income and Wealth 60 (1), 36–78.

Laibson, D. (1997). Golden Eggs and Hyperbolic Discounting. Quarterly Journal of Eco-
nomics 112 (2), 443–477.

Laibson, D. (2015). Why Don’t Present-Biased Agents Make Commitments? The American
Economic Review 105 (5), 267–272.

Meier, S. and C. Sprenger (2010). Present-Biased Preferences and Credit Card Borrowing.
American Economic Journal: Applied Economics 2 (1), 193–210.

Melzer, B. T. (2011). The Real Costs of Credit Access: Evidence from the Payday Lending
Market. The Quarterly Journal of Economics 126 (1), 517–555.

15
Morse, A. (2011). Payday Lenders: Heroes or Villains? Journal of Financial Eco-
nomics 102 (1), 28–44.

Poterba, J. M., S. F. Venti, and D. A. Wise (1996). How Retirement Saving Programs
Increase Saving. The Journal of Economic Perspectives 10 (4), 91–112.

16
Table 1: Demographic Diﬀerences By Baseline Overdrafter Status

(1) (2) (3)
Light Heavy-
Overdrafter Light
Mean Diﬀerence N
Fraction Months in Overdraft: Past Year 0.0578 0.481∗∗∗ 318
[0.141] (0.0242)
Female 0.3 -.129∗∗∗ 318
[0.46] (0.0472)
Age 45.5 4.53∗∗∗ 318
[9.63] (0.945)
Teacher 0.469 0.202∗∗∗ 318
[0.501] (0.0545)
Other Govt. Employee 0.35 -.0272 318
[0.478] (0.0531)
Formal Private Sector Worker 0.188 -.181∗∗∗ 318
[0.392] (0.0316)
Education: Secondary Graduate 0.344 -.0526 318
[0.476] (0.0523)
Education: Postsecondary 0.606 0.0583 318
[0.49] (0.054)
Married 0.85 0.0487 318
[0.358] (0.0372)
Biological Children Ever Born 3.14 1.32∗∗∗ 318
[2.1] (0.252)
Householders 14 and Under 2.15 -.0677 318
[1.73] (0.19)
Householders 15 and Older 3.5 0.69∗∗∗ 318
[2.04] (0.237)
Income Last Month 729 15.4 302
[372] (45.3)
Savings - NVRB (Reported) 324 -186∗∗∗ 283
[612] (58.9)
Savings - NVRB (Adminstrative) 356 -243∗∗∗ 318
[699] (63.2)
Total Cash Savings 1672 -399 249
[3249] (358)
Total Debt 1930 1252∗∗∗ 308
[2428] (329)
Economic Shock 0.481 0.0377 318
[0.501] (0.0562)
Financial Strain: Missed School 0.313 -.00237 318
[0.465] (0.0521)
Financial Strain: Pay Bills 0.675 0.116∗∗ 318
[0.47] (0.0493)
Financial Strain: Pay Debt 0.563 0.0957∗ 318
[0.498] (0.0546)
Socially Taxed 0.619 -.0301 318
[0.487] (0.055)
Any Food Insecurity 0.444 0.0562 318
[0.498] (0.0561)
Notes: Standard deviations in brackets, heteroskedasticity robust standard errors
in parentheses. All variables denominated in Ghanaian Cedis top-coded at the 99th
percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.

17
Table 2: Impacts on NVRB Account Use, Administrative Data

(1) (2) (3) (4)
Acct. in Overdraft Net Average
Over- Charges Deposits Daily
draft Balance
Treatment × During 0.0095 0.55 19.5∗∗ 243.5∗∗∗
(0.020) (1.36) (8.11) (43.9)
Treatment × Endline -0.060∗ -3.77∗ -3.96 140.4∗∗
(0.032) (2.17) (26.7) (63.2)
Treatment × Post-Endline -0.0056 -0.55 32.5 141.2∗
(0.030) (2.36) (23.2) (84.1)

Control Mean (During) 0.30 16.3 -19.3 89.0
Control Mean (Endline) 0.33 18.3 1.99 110.1
Control Mean (Post-Endline) 0.26 17.5 18.8 110.6
N 9540 9540 9540 9540

Baseline Control? Yes Yes Yes Yes
Notes: All regressions control for strata and month ﬁxed eﬀects as well as the average
outcome in the year before the baseline survey. Robust standard errors clustered at
the individual level in parentheses. Data is at the person-month level. Overdrafter
is a dummy equal to one if the participant paid either overdraft interest or the fee
for an overdraft form in a given month. Monthly overdrafter charges topcoded at
the 99th percentile, average daily balance topcoded at the 1st and 99th percentile.
Net deposit is the diﬀerence between total deposits and withdrawals, both topcoded
at the 99th percentile. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.

18
Table 3: Impacts on Total Cash Savings, Debt, and Savings Net Debt

(1) (2) (3) (4) (5) (6) (7) (8)
Savings Debt
Savings Savings Other Informal Total Took Any Total Savings
at NVRB at NVRB Formal Savings Savings New Debt Debt Net Debt
(Admin) (Survey) Savings
Treatment × During 303.1∗∗∗ 288.2∗∗∗ -47.7 -20.7 215.1 0.076∗∗∗ 283.3 -82.9
(43.7) (42.0) (118.1) (22.3) (139.4) (0.029) (283.3) (346.5)
Treatment × After 91.3∗ 25.0 127.5 2.57 145.0 0.046 195.1 1.38
(52.8) (49.3) (142.6) (45.6) (167.7) (0.031) (269.1) (354.6)

Control Mean (During) 164.6 214.2 979.9 172.9 1377.4 0.19 3080.8 -1687.8
Control Mean (After) 199.5 256.7 921.1 236.2 1424.1 0.15 2636.7 -1244.9
N 1528 1499 1518 1527 1488 1528 1519 1481
19

Baseline Control? Yes Yes Yes Yes Yes No Yes Yes
Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed eﬀects, and the baseline outcome of the dependent variable whenever
possible. Robust standard errors clustered at the individual level in parentheses. Administrative NVRB savings is the balance on the
relevant survey day, with negative balances coded to zero. All variables in Ghanain Cedis top coded at the 99th percentile. In 2013
GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.
Table 4: Impacts on Total Cash Savings, Debt, and Savings Net Debt - Heterogeneity by Baseline Overdraft Behavior

(1) (2) (3) (4) (5) (6) (7) (8)
Savings Debt
Savings Savings Other Informal Total Took Any Total Savings
at NVRB at NVRB Formal Savings Savings New Debt Debt Net Debt
(Admin) (Survey) Savings
Treatment × During × Light Overdrafter 322.4∗∗∗ 326.6∗∗∗ 341.7∗∗ -81.2∗∗ 573.8∗∗∗ 0.043 72.5 621.8
(76.8) (71.4) (166.5) (36.4) (200.2) (0.028) (369.9) (471.3)
Treatment × After × Light Overdrafter 101.2 54.7 567.3∗∗∗ -8.97 596.4∗∗ 0.021 -76.5 794.4∗
(97.4) (85.7) (205.1) (63.9) (242.9) (0.036) (338.1) (459.1)
Treatment × During × Heavy Overdrafter 284.2∗∗∗ 251.2∗∗∗ -428.6∗∗∗ 38.7 -136.6 0.11∗∗ 492.6 -776.1
(43.0) (46.8) (158.6) (26.1) (185.8) (0.050) (426.0) (472.2)
Treatment × After × Heavy Overdrafter 81.0∗∗ -4.04 -312.0∗ 15.3 -297.4 0.069 463.3 -771.4
(39.4) (51.8) (186.9) (65.9) (221.5) (0.051) (418.0) (499.4)

P-value: Heavy=Light, During 0.665 0.382 0.001∗∗∗ 0.009∗∗∗ 0.009∗∗∗ 0.253 0.456 0.031∗∗
P-value: Heavy=Light, After 0.847 0.561 0.002∗∗∗ 0.793 0.006∗∗∗ 0.437 0.316 0.017∗∗
P-value: Treat=0, Light 0.000∗∗∗ 0.000∗∗∗ 0.021∗∗ 0.079∗ 0.011∗∗ 0.308 0.892 0.221
20

P-value: Treat=0, Heavy 0.000∗∗∗ 0.000∗∗∗ 0.026∗∗ 0.333 0.406 0.060∗ 0.440 0.195

Control Mean (During, Light) 292.1 293.2 780.2 234.1 1311.7 0.081 2552.9 -1257.1
Control Mean (After, Light) 336.7 349.4 731.2 246.9 1328.1 0.088 2322.0 -989.5
Control Mean (During, Heavy) 42.0 137.9 1174.4 113.9 1442.0 0.29 3590.2 -2114.5
Control Mean (After, Heavy) 65.9 165.3 1105.9 225.7 1518.8 0.20 2943.2 -1496.9
N 1528 1499 1518 1527 1488 1528 1519 1481

Baseline Control? Yes Yes Yes Yes Yes No Yes Yes
Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed eﬀects, interactions between overdrafter status and time period, and the baseline outcome
of the dependent variable whenever possible. Robust standard errors clustered at the individual level in parentheses. Administrative NVRB savings is the
balance on the relevant survey day, with negative balances coded to zero. All variables in Ghanain Cedis top coded at the 99th percentile. In 2013 GHS 2.15
≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.
Figure 1: Administrative Data – Impact on NVRB Balance by Month and Overdrafter Status

A. Non-Overdrafters

1000
500
0
-500

2014m1 2014m7 2015m1 2015m7 2016m1 2016m7
Month

B. Overdrafters
1000
500
0
-500

2014m1 2014m7 2015m1 2015m7 2016m1 2016m7
Month

Notes: Whiskers give 90 and 95 percent conﬁdence intervals on estimates.
Average monthly balance is top-coded at the 99 percent level by month.

21
A Appendix
A.1 Sampling
NVRB is a small bank, with eight branches in ﬁve districts serving roughly 10,000 account
holders at the time of the experiment.18 We attempted to enroll all workers who received
electronic salary payments into an NVRB account. Of the 420 workers we approached, 320
met our inclusion criteria and were willing to participate in the study. Workers were excluded
if (a) their electronic payments were not salaries (e.g. pension recipients), (b) they expected
to retire over the study duration, (c) they expected to move out of the study area, or (d)
they had all deposits to the NVRB account immediately transferred to another bank. Of
the 100 individuals from the initial list who were not included in the study sample, 63 met
one of the exclusion criteria, 10 were on chronic sick bed or mentally ill, 3 were deceased,
and 24 were not interested in participating in the study.
Half of individuals enrolled in the study were oﬀered the opportunity to sign up for SSP.
We stratiﬁed the randomization by heavy overdrafter×branch×gender, which generated 28
distinct strata. The heavy overdrafter dummy was constructed by ﬁrst creating an index of
overdrafting behavior. Overdraft scores were constructed for each person using their current
account transaction records available at the time of the randomization. The score was
calculated as the fraction of balance entries that were negative over the life of the account.
The sample was sorted from lowest score (never had a negative balance) to highest (negative
balance 66 percent of the time), with the top 50 percent of the sample classiﬁed as “heavy
overdrafters”.

A.2 Data Collection
Appendix Figure A2 shows the study timeline. In April 2013 we used administrative data
from NVRB to compile a list of 420 salaried workers receiving salary deposits into an NVRB
account. We contacted and enrolled these workers between April and September 2013. The
baseline followed in September-October 2013, before SSP was marketed to the treatment
group. During November 2013, we conducted the randomization and enrolled interested
treatment group clients in SSP. The ﬁrst three follow-up surveys took place while SSP was
ongoing – during June 2014, October 2014, and February 2014 (months 6, 10, and 14 of the
commitment cycle). SSP savings were released in June 2015. The last two follow-ups took
place between September-November 2014 (months 3 and 5 after the SSP payout). The key
follow-up survey modules included savings, debt, expenditures, and assets. We collected
18
The districts are Kratchi East, Biakoye, Jasikan, Kadjebi, and Nkwanta South.

22
data on income by source in the baseline and fourth follow-up. The baseline collected
additional detail on respondents’ demographic characteristics. Selected follow-up survey
rounds also asked about economic shocks, ﬁnancial well-being, intra-household decision-
making, and (unincentivized) time preferences. The ﬁnal follow-up also asked respondents
about perceived beneﬁts of SSP and interest in signing up for future SSP cycles.

A.3 Attrition
Study attrition was very low, with 91 percent of the sample interviewed in each survey round.
Appendix Table A2 formally veriﬁes that attrition is uncorrelated with treatment status –
here, we see that treatment-control diﬀerences in attrition are small in magnitude and almost
never statistically signiﬁcant, for both heavy and light overdrafters.

A.4 Summary of Key Variables
This subsection summarizes some of the key outcome variables in the study, and provides
additional detail on how we constructed them. All variables that were winsorized were done
so by survey round/month.

Administrative Data

• Account in Overdraft – A dummy variable equal to one in months where an individual
paid for an overdraft form or paid overdraft interest. Equal to one if an overdraft is
taken on any NVRB account.

• Overdraft Charges – Charges associated with issuance of an overdraft or overdraft
interest. Sums across all NVRB accounts.

• Net Deposits – Total monthly deposits less total monthly debits (including withdrawals
and fees), summed across all NVRB accounts. Total monthly deposits and withdrawals
are each winsorized at the 99th percentile before taking the diﬀerence.

• Average Daily Balance – The average daily balance across all NVRB accounts. Ac-
counts in overdraft are included with a negative balance. Winsorized at the ﬁrst and
99th percentile.

• NVRB Savings at Survey Date – Total balance across all accounts on date of survey.
Accounts with negative balances (e.g. due to an overdraft) are coded to zero before
summing. Winsorized at the 99th percentile.

23
• Salary Deposited into NVRB Accounts – Monthly salary deposits made into NVRB
accounts, based on account narration text. Winsorized at the 99th percentile.

Survey Data

• NVRB Savings – Respondent self-reported NVRB balances. In the survey, respondents
were asked to report non-SSP balances. SSP balances were added on top of this amount
based on administrative data. Winsorized at the 99th percentile.

• Other Formal Savings – Includes self-reported savings at other banks, microﬁnance
institutions, and credit unions. Winsorized at the 99th percentile.

• Informal Savings – Includes self-reported savings in cash and in savings clubs. Addi-
tional categories covered varied from survey round to survey round. At baseline, this
measure also includes savings in agricultural co-ops, savings held by other individuals
and susu collectors, and moneylenders. Follow-ups 1, 2, and 3 include mobile money
savings and “other informal cash savings”, which was collected in a free answer format.
This mostly includes money hidden at home and money stored with other individu-
als. Follow ups 4 and 5 includes the same categories as follow ups 1-3, but explicitly
prompted respondents to report savings held with other individuals including friends
and relatives. Winsorized at the 99th percentile.

• Total Savings – Sum of (winsorized) self-reported NVRB savings, other formal savings,
and informal savings.

• Took Any New Debt – Dummy variable equal to one if individual reported taking on
any new debt in the past 30 days.

• Debt with NVRB – Self-reported debt held at NVRB. Winsorized at the 99th percentile.

• Other Formal Debt – At baseline includes self-reported debt held at other banks, credit
unions, and microﬁnance organizations. Measures for all follow up surveys include these
categories as well as insurance companies and the teachers fund. Winsorized at the
99th percentile.

• Informal Debt – At baseline includes self-reported debt with individual and group
susus. Follow-up surveys include this, as well as other debt (collected in a free form
answer) from informal sources, such as shop keepers, traders, and individuals. Win-
sorized at the 99th percentile.

• Total Debt – Sum of (winsorized) debt with NVRB, other formal, and informal debt.

24
• Savings Net Debt – Diﬀerence between total savings and total debt, both as deﬁned
above.

• Child Missed Class for Unpaid School Fees – Dummy variable equal to one if a child
missed school because the respondent was not able to pay their school fees. The way
this variable is constructed varies from round to round. At the baseline, the variable
is equal to one if the respondent reports that children have to miss school “often” or
“sometimes” due to missed school payments. lookback period on this question varied.
In the ﬁrst, second, and third follow up, the dummy is equal to one if a child missed
school due to a late payment in the past 30 days. Follow up 4 asked respondents to
report on the period between SSP release (June 1, 2015) and the survey and follow up
5 asked about the past 60 days.

• Concerned About Paying Bills – Dummy variable equal to one if the respondent re-
ported not having enough money to pay normal monthly expenses. The construction
is similar to the “missed school” variable, in that the baseline dummy identiﬁes individ-
uals who report concern often or sometimes, while the follow ups identify respondents
who report actual diﬃculty. The follow up look back periods are the same as they are
for the missed school variable.

• Concerned About Repaying Debt – Dummy variable equal to one if the respondent
reported not having enough money to service scheduled debt payments. The construc-
tion is similar to the “missed school” variable, in that the baseline dummy identiﬁes
individuals who report concern often or sometimes, while the follow ups identify re-
spondents who report actual diﬃculty. The follow up look back periods are the same
as they are for the missed school variable.

• Experienced Food Insecurity – Dummy to identify any kind of food insecurity. Equal
to one if respondent reports days when there was not enough food to meet the needs
of the family or days when adults cut or skipped meals because there was not enough
food rarely, sometimes, or often. The lookback period varied by survey round: it was
12 months at baseline, 30 days for follow ups 1-3, since June 1 2015 for follow up 4,
and 60 days for follow up 5.

• Personal Consumption – Respondent’s spending on his/her self (excludes household
expenses such as rent, utilities and food). The lookback period was 30 days for follow
ups 1-3, since June 1 2015 for follow up 4, and 60 days for follow up 5. Values are
re-normalized to correspond to a 30 day window for each survey round. Winsorized at
the 99th percentile.

25
• Expenditure on Household and Others – Respondent’s spending on the household and
dependents (includes expenses such as rent, utilities, food, and education). The look-
back period was 30 days for follow ups 1-3, since June 1 2015 for follow up 4, and 60
days for follow up 5. Values are re-normalized to correspond to a 30 day window for
each survey round. Winsorized at the 99th percentile.

• Expenditure on Investments – Respondent’s spending on investments. The lookback
period was 30 days for follow ups 1-3, since June 1 2015 for follow up 4, and 60 days
for follow up 5. Values are re-normalized to correspond to a 30 day window for each
survey round. Winsorized at the 99th percentile.

• Monthly Salary and Wages – Income in the past month from salaried and wage jobs.
Winsorized at the 99th percentile. Note that income modules were only administered
at baseline and at follow up survey 4.

• Monthly Self-Employment Income – Income in the past month from self employment.
Winsorized at the 99th percentile. Note that income modules were only administered
at baseline and at follow up survey 4.

• Total Monthly Income – Sum of salary/wage and self employment income as deﬁned
above. Note that income modules were only administered at baseline and at follow up
survey 4.

• SSP Reduces Transfer Pressure in Household – Dummy variable equal to one if respon-
dent strongly agrees that SSP makes it easier to say no to requests for money within
the household.

• SSP Reduces Transfer Pressure Outside Household – Dummy variable equal to one if
respondent strongly agrees that SSP makes it easier to say no to requests for money
coming from outside the household.

• SSP Helps Meet Savings Goals – Dummy variable equal to one if the respondent
strongly agrees that SSP makes it easier to meet savings goals.

• SSP Helps Gather Money for Large Expense – Dummy variable equal to one if the re-
spondent strongly agrees that SSP makes it easier to gather money for a large expense.

• SSP Helps Build Funds for Emergency – Dummy variable equal to one if the respondent
strongly agrees that SSP makes the respondent more likely to have funds available in
case of an emergency.

26
Table A1: Demographic Characteristics and Randomization Veriﬁcation

(1) (2) (3) (4) (5) (6) (7) (8) (9)
Whole Sample Heavy Overdrafters Light Overdrafters
Control Treat-Control Control Treat-Control Control Treat-Control
Mean Diﬀerence N Mean Diﬀerence N Mean Diﬀerence N
Fraction Months in Overdraft: Past Year 0.286 0.0209 318 0.525 0.0274 158 0.0475 0.0204 160
[0.319] (0.0362) [0.274] (0.043) [0.121] (0.0222)
Female 0.234 0.00332 318 0.165 0.0127 158 0.304 -.0075 160
[0.425] (0.0478) [0.373] (0.0603) [0.463] (0.0729)
Age 47.3 0.965 318 50.1 -.0633 158 44.5 2.04 160
[8.98] (0.979) [7.04] (1.12) [9.84] (1.52)
Teacher 0.563 0.0117 318 0.684 -.0253 158 0.443 0.0508 160
[0.498] (0.0557) [0.468] (0.0752) [0.5] (0.0793)
Other Govt. Employee 0.335 0.00206 318 0.316 0.0127 158 0.354 -.00875 160
[0.474] (0.0532) [0.468] (0.0749) [0.481] (0.0759)
Formal Private Sector Worker 0.101 -.00752 318 0 0.0127 158 0.203 -.0297 160
[0.303] (0.0334) [0] (0.0127) [0.404] (0.0621)
Education: Secondary Graduate 0.304 0.0275 318 0.304 -.0253 158 0.304 0.0789 160
[0.461] (0.0523) [0.463] (0.0727) [0.463] (0.0753)
Education: Postsecondary 0.665 -.0583 318 0.671 -.0127 158 0.658 -.103 160
[0.474] (0.054) [0.473] (0.0756) [0.477] (0.0773)
Married 0.861 0.0267 318 0.886 0.0253 158 0.835 0.0288 160
[0.347] (0.0373) [0.32] (0.0483) [0.373] (0.0568)
Biological Children Ever Born 3.84 -.0793 318 4.51 -.0759 158 3.18 -.0661 160
[2.34] (0.262) [2.39] (0.379) [2.1] (0.334)
Householders 14 and Under 2.2 -.159 318 2.23 -.291 158 2.16 -.0288 160
[1.78] (0.19) [1.73] (0.262) [1.84] (0.276)
Householders 15 and Older 3.78 0.128 318 4.16 0.0506 158 3.39 0.213 160
[2.27] (0.24) [2.28] (0.348) [2.21] (0.324)
Income Last Month 770 -67.1 302 746 -3.33 148 794 -128∗∗ 154
[421] (45) [425] (68.1) [419] (59.4)
Savings - NVRB (Reported) 249 -33.2 283 125 26.3 139 368 -89.5 144
[528] (60.3) [330] (59.1) [645] (102)
Savings - NVRB (Adminstrative) 243 -14.9 318 106 14.4 158 380 -46.8 160
[614] (64.9) [411] (61.7) [743] (111)
Total Cash Savings 1628 -315 249 1314 -83.3 124 1937 -543 125
[3269] (357) [1825] (423) [4231] (572)
Total Debt 2506 82.3 308 3127 111 152 1886 86.7 156
[2877] (336) [3108] (535) [2495] (390)
Economic Shock 0.468 0.0629 318 0.494 0.0506 158 0.443 0.0755 160
[0.501] (0.0561) [0.503] (0.0799) [0.5] (0.0793)
Financial Strain: Missed School 0.342 -.0605 318 0.329 -.038 158 0.354 -.0828 160
[0.476] (0.052) [0.473] (0.074) [0.481] (0.0735)
Financial Strain: Pay Bills 0.709 0.0474 318 0.747 0.0886 158 0.671 0.00813 160
[0.456] (0.0497) [0.438] (0.0647) [0.473] (0.0745)
Financial Strain: Pay Debt 0.589 0.0426 318 0.633 0.0506 158 0.544 0.0359 160
[0.494] (0.0548) [0.485] (0.0758) [0.501] (0.0789)
Socially Taxed 0.614 -.0202 318 0.557 0.0633 158 0.671 -.103 160
[0.488] (0.055) [0.5] (0.0786) [0.473] (0.0768)
Any Food Insecurity 0.456 0.0318 318 0.481 0.038 158 0.43 0.0264 160
[0.5] (0.0561) [0.503] (0.08) [0.498] (0.079)
Notes: Standard deviations in brackets, heteroskedasticity robust standard errors in parentheses. All variables denominated in Ghanaian Cedis top-coded
at the 99th percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.

27
Table A2: Tests for Diﬀerential Attrition by Treatment Status

(1) (2) (3) (4) (5) (6) (7) (8) (9)
Whole Sample Heavy Overdrafters Non-Heavy Overdrafters
Control Treat-Control Control Treat-Control Control Treat-Control
Mean Diﬀerence N Mean Diﬀerence N Mean Diﬀerence N
In Baseline 0.994 0.0000781 320 0.988 0 160 1 0 160
[0.0793] (0.00884) [0.112] (0.0177) [0] (0)
In Endline 1 0.962 0.00668 320 0.975 0 160 0.949 0.0136 160
[0.191] (0.0204) [0.157] (0.0248) [0.221] (0.0326)
In Endline 2 0.962 -.00574 320 0.975 -.025 160 0.949 0.0136 160
[0.191] (0.0221) [0.157] (0.0302) [0.221] (0.0326)
In Endline 3 0.937 0.0194 320 0.95 0 160 0.924 0.0389 160
[0.244] (0.0252) [0.219] (0.0347) [0.267] (0.0367)
In Endline 4 0.943 0.0255 320 0.95 0 160 0.937 0.0509∗ 160
[0.232] (0.0229) [0.219] (0.0347) [0.245] (0.0302)
In Endline 5 0.943 0.00691 320 0.95 -.0125 160 0.937 0.0263 160
[0.232] (0.0252) [0.219] (0.0366) [0.245] (0.0347)
28

In All Survey Rounds 0.906 0.00738 320 0.913 -.0125 160 0.899 0.0272 160
[0.293] (0.0322) [0.284] (0.0464) [0.304] (0.045)
Notes: Standard deviations in brackets, heteroskedasticity robust standard errors in parentheses. * p≤ 0.10, ** p≤ 0.05, ***
p≤ 0.10.
Table A3: Predictors of SSP Take Up and Drop Out

(1) (2) (3) (4)
Pairwise Joint
Take Up Drop Out Take Up Drop Out
Heavy Overdrafter 0.005 0.124∗∗ -0.001 0.063
(0.072) (0.063) (0.082) (0.072)
Female 0.024 -0.078 0.090 0.003
(0.083) (0.062) (0.090) (0.071)
Age 0.013∗∗∗ 0.002 0.016∗∗∗ -0.003
(0.004) (0.004) (0.006) (0.007)
Teacher -0.054 0.093 -0.361∗∗∗ 0.033
(0.072) (0.061) (0.132) (0.141)
Other Govt. Employee 0.033 -0.085 -0.337∗∗ -0.043
(0.075) (0.060) (0.154) (0.153)
Education: Secondary Graduate -0.003 -0.038 -0.176 0.049
(0.076) (0.064) (0.158) (0.078)
Education: Postsecondary -0.019 0.072 -0.181 -0.014
(0.073) (0.061) (0.196) (0.135)
Married 0.059 0.053 0.118 0.014
(0.118) (0.087) (0.135) (0.107)
Biological Children Ever Born 0.038∗∗ 0.024 -0.002 0.022
(0.017) (0.016) (0.021) (0.021)
Householders 14 and Under 0.039∗ -0.009 0.042 -0.026
(0.022) (0.021) (0.027) (0.025)
Householders 15 and Older 0.014 0.024 0.001 0.009
(0.019) (0.017) (0.024) (0.024)
Income Last Month / 1000 0.065 0.223∗∗ 0.143 0.100
(0.103) (0.105) (0.149) (0.121)
Total Cash Savings / 1000 0.002 -0.012∗ -0.010 -0.013
(0.014) (0.007) (0.017) (0.009)
Total Debt / 1000 -0.007 0.035∗∗ -0.005 0.022
(0.013) (0.015) (0.013) (0.018)
Economic Shock 0.023 -0.003 0.043 0.004
(0.072) (0.064) (0.080) (0.082)
Financial Strain: Missed School 0.113 0.012 0.157∗ 0.009
(0.074) (0.069) (0.084) (0.079)
Financial Strain: Pay Bills 0.103 -0.191∗∗ 0.188∗ -0.187∗
(0.087) (0.096) (0.098) (0.107)
Financial Strain: Pay Debt -0.123∗ -0.031 -0.161∗ -0.006
(0.071) (0.066) (0.088) (0.075)
Socially Taxed 0.019 0.072 0.036 0.098
(0.073) (0.061) (0.075) (0.068)
Any Food Insecurity -0.077 0.003 -0.065 0.019
(0.071) (0.064) (0.078) (0.072)
DV Mean 0.719 0.130 0.719 0.130
N 160 115 160 115
Notes: Heteroskedasticity robust standard errors in parentheses. The ﬁrst two columns
show results of pairwise regressions (each cell represents a separate regression). The last
two columns show results from regressions where all demographic variables are jointly
included as covariates. All variables denominated in Ghanaian Cedis top-coded at the
99th percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.

29
Table A4: Impacts on Debt by Source

(1) (2) (3)
Debt Other Informal
With Formal Debt
NVRB Debt
Treatment × During 29.6 240.6 5.68
(174.6) (216.4) (5.97)
Treatment × After -153.5 338.5 0.96
(157.1) (225.6) (8.81)

Control Mean (During) 1563.7 1490.6 10.6
Control Mean (After) 1471.6 1150.1 15
N 1520 1527 1528

Baseline Control? Yes Yes Yes
30

Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed
eﬀects, and the baseline outcome of the dependent variable whenever
possible. Robust standard errors clustered at the individual level in
parentheses. All variables in Ghanain Cedis top coded at the 99th
percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, ***
p≤ 0.10.
Table A5: Impacts on Financial Strain

(1) (2) (3) (4)
Child Concerned Concerned Experienced
Missed About About Food
Class for Paying Bills Repaying Insecurity
Unpaid Debt
School Fees
Treatment × During -0.012 0.050 0.020 -0.019
(0.034) (0.037) (0.032) (0.038)
31

Treatment × After 0.021 0.0084 0.030 -0.0015
(0.037) (0.044) (0.038) (0.041)

Control Mean (During) 0.27 0.49 0.22 0.40
Control Mean (After) 0.21 0.35 0.22 0.28
N 1516 1510 1497 1528

Baseline Control? Yes Yes Yes Yes
Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed eﬀects, and the baseline
outcome of the dependent variable whenever possible. Robust standard errors clustered at
the individual level in parentheses. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.
Table A6: Impacts on NVRB Account Use, Administrative Data - Heterogeneity by Baseline
Overdraft Behavior

(1) (2) (3) (4)
Acct. in Overdraft Net Average
Over- Charges Deposits Daily
draft Balance
Treatment × During × Light Overdrafter 0.018 0.62 35.7∗∗∗ 339.0∗∗∗
(0.019) (1.06) (12.5) (66.9)
Treatment × Endline × Light Overdrafter -0.058∗ -3.85∗∗ -11.0 213.6∗∗
(0.029) (1.89) (44.8) (103.8)
Treatment × Post-Endline × Light Overdrafter -0.020 -1.19 89.5∗∗ 278.2∗∗
(0.029) (2.32) (35.1) (132.0)
Treatment × During × Heavy Overdrafter 0.0015 0.48 2.87 147.1∗∗∗
(0.036) (2.52) (10.2) (55.7)
Treatment × Endline × Heavy Overdrafter -0.062 -3.72 3.16 65.9
(0.057) (3.93) (28.0) (71.0)
Treatment × Post-Endline × Heavy Overdrafter 0.0078 0.089 -25.3 2.22
(0.052) (4.13) (29.3) (102.8)

P-value: Heavy=Light, During 0.691 0.959 0.044∗∗ 0.028∗∗
P-value: Heavy=Light, Endline 0.948 0.976 0.788 0.241
P-value: Heavy=Light, Post-Endline 0.643 0.788 0.012∗∗ 0.100
P-value: Treat=0, Light 0.030∗∗ 0.052∗ 0.003∗∗∗ 0.000∗∗∗
P-value: Treat=0, Heavy 0.312 0.527 0.837 0.072∗

Control Mean (During, Light) 0.070 3.22 -20.1 263.0
Control Mean (Endline, Light) 0.12 6.57 3.56 322.3
Control Mean (Post-Endline, Light) 0.099 5.96 -13.9 288.9
Control Mean (During, Heavy) 0.54 29.3 -18.4 -85.0
Control Mean (Endline, Heavy) 0.54 30.0 0.42 -102.1
Control Mean (Post-Endline, Heavy) 0.43 29.0 51.5 -67.7
N 9540 9540 9540 9540

Baseline Control? Yes Yes Yes Yes
Notes: All regressions control for strata and month ﬁxed eﬀects, interactions between overdrafter status
and period (endline, post-endline), and the average outcome in the year before the baseline survey.
Robust standard errors clustered at the individual level in parentheses. Data is at the person-month
level. Overdrafter is a dummy equal to one if the participant paid either overdraft interest or the fee
for an overdraft form in a given month. Monthly overdrafter charges topcoded at the 99th percentile,
average daily balance topcoded at the 1st and 99th percentile. Net deposit is the diﬀerence between total
deposits and withdrawals, both topcoded at the 99th percentile. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.

32
Table A7: Impacts on Debt by Source - Heterogeneity by Baseline Over-
draft Behavior

(1) (2) (3)
Debt Other Informal
With Formal Debt
NVRB Debt
Treatment × During × Light Overdrafter -38.4 135.8 -5.11
(269.8) (277.2) (7.82)
Treatment × After × Light Overdrafter -386.9 313.4 3.67
(234.6) (255.2) (14.2)
Treatment × During × Heavy Overdrafter 98.0 344.7 16.3∗
(219.6) (328.8) (9.03)
Treatment × After × Heavy Overdrafter 78.7 361.3 -1.81
(205.1) (372.4) (10.3)

P-value: Heavy=Light, During 0.694 0.626 0.075∗
P-value: Heavy=Light, After 0.135 0.915 0.755
P-value: Treat=0, Light 0.115 0.445 0.659
P-value: Treat=0, Heavy 0.883 0.500 0.176

Control Mean (During, Light) 1504.3 1021.6 17.6
Control Mean (After, Light) 1598.3 709.0 14.7
Control Mean (During, Heavy) 1621.1 1941.4 3.83
Control Mean (After, Heavy) 1348.2 1579.7 15.3
N 1520 1527 1528

Baseline Control? Yes Yes Yes
Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed eﬀects, in-
teractions between overdrafter status and time period, and the baseline outcome
of the dependent variable whenever possible. Robust standard errors clustered
at the individual level in parentheses. All variables in Ghanain Cedis top coded
at the 99th percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, ***
p≤ 0.10.

33
Table A8: Impacts on Expenditure and Income - Heterogeneity by Baseline Overdraft Behavior

(1) (2) (3) (4) (5) (6)
Expenditure Income
Personal Spending Expenditure Salary Self Em- Total
Con- on House- on Invest- and ployment
sumption hold and ments Wages
Others
Treatment × During × Light Overdrafter -11.2 0.65 -24.8
(18.4) (83.4) (42.7)
Treatment × After × Light Overdrafter -2.12 -102.5 0.93 114.3∗∗ 29.8 161.9∗∗
(13.6) (80.7) (43.1) (55.2) (27.3) (66.3)
Treatment × During × Heavy Overdrafter 31.9 144.2 7.65
(20.1) (100.8) (36.4)
Treatment × After × Heavy Overdrafter -0.48 93.8 -1.36 4.81 23.3 27.5
34

(12.7) (79.6) (38.5) (81.8) (28.8) (83.1)

P-value: Heavy=Light, During 0.115 0.273 0.563
P-value: Heavy=Light, After 0.930 0.084∗ 0.968 0.270 0.867 0.206
P-value: Treat=0, Light 0.831 0.311 0.818 0.039∗∗ 0.276 0.015∗∗
P-value: Treat=0, Heavy 0.262 0.294 0.973 0.953 0.420 0.740

Control Mean (During, Light) 139.2 789.9 155.2
Control Mean (After, Light) 81.8 732.0 106.3 712.8 54.1 766.9
Control Mean (During, Heavy) 128.9 842.2 128.5
Control Mean (After, Heavy) 83.4 658.4 87.0 843.7 32.3 876.0
N 1528 1528 1528 306 306 306

Baseline Control? No No No Yes Yes Yes
Notes: All regressions include strata ﬁxed eﬀects, time period ﬁxed eﬀects, interactions between overdrafter status and time
period, and the baseline outcome of the dependent variable whenever possible. Robust standard errors clustered at the individual
level in parentheses. All variables in Ghanain Cedis top coded at the 99th percentile. In 2013 GHS 2.15 ≈ USD 1. Survey data
on income only available at baseline and in follow up survey 4. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.
Table A9: Impacts on Overall Cash Savings by Baseline Other Control

Savings Debt
(1) (2) (3) (4) (5) (6) (7) (8)
Savings Savings Other
at NVRB at NVRB Formal Informal Total Took Any Savings
(Admin) (Survey) Savings Savings Savings New Debt Total Debt Net Debt
Treatment × During 320.669∗∗∗ 307.723∗∗∗ -70.619 -33.662 178.277 0.057 11.572 173.230
(65.958) (67.313) (160.979) (26.211) (187.036) (0.038) (356.207) (404.700)
Treatment × During × Other Control -24.511 -20.091 77.412 21.291 89.924 0.033 526.606 -465.372
(76.088) (77.384) (181.515) (30.331) (207.053) (0.045) (415.209) (459.911)
Treatment × After 215.657∗∗ 127.123 145.120 -74.361 168.866 0.054 -59.809 304.677
(87.049) (84.844) (183.638) (51.846) (215.301) (0.043) (347.078) (407.878)
35

Treatment × After × Other Control -210.609∗∗ -160.434∗ 4.668 130.586∗∗ -9.620 -0.009 472.037 -523.364
(90.014) (86.444) (220.950) (64.381) (250.668) (0.047) (399.027) (466.946)

P-Val, During=0, Other Control Problem 0.000∗∗∗ 0.000∗∗∗ 0.963 0.654 0.110 0.011∗∗ 0.121 0.477
P-Val, After=0, Other Control Problem 0.923 0.493 0.402 0.335 0.439 0.218 0.205 0.607

DV Mean (Control, No Problem) 185.542 242.651 1212.544 204.877 1662.356 0.165 2792.135 -1073.859
DV Mean (Control, Other Control Problem) 175.536 226.206 803.076 195.684 1238.958 0.172 2969.259 -1770.425
N 1519 1490 1509 1518 1479 1519 1510 1472
Notes: Robust standard errors clustered at the individual level in parentheses. All regressions control for the baseline value of the outcome, strata ﬁxed
eﬀects, time period ﬁxed eﬀects, an other control dummy, and interactions between the other control dummy and time period. The other control dummy is
equal to one if the respondent reports diﬃculty saving no to transfer requests from others at baseline. All variables in Ghanain Cedis top coded at the 99th
percentile. In 2013 GHS 2.15 ≈ USD 1. * p≤ 0.10, ** p≤ 0.05, *** p≤ 0.10.
Table A10: Does Experience Impact Respondents’ Assessments of SSP?

(1) (2) (3) (4) (5)
Reduce Reduce Helps Me Helps Me Helps Build
Transfer Transfer Meet Gather Funds for an
Pressure in Pressure Savings Money for a Emergency
Household Outside Goals Big Expense
Household
Panel A. Overall Impacts
Treatment 0.13∗∗∗ 0.039 0.14∗∗∗ 0.19∗∗∗ 0.076
(0.047) (0.046) (0.051) (0.052) (0.052)

Control Mean (After)XXXXXXXXXXXX 0.15 0.18 0.21 0.19 0.24

Panel B. Impacts by Baseline Overdraft Behavior
36

Treatment × Light Overdrafter 0.17∗∗∗ 0.047 0.15∗∗ 0.13∗ 0.052
(0.064) (0.070) (0.071) (0.074) (0.068)
Treatment × Heavy Overdrafter 0.085 0.030 0.14∗ 0.25∗∗∗ 0.099
(0.069) (0.059) (0.073) (0.072) (0.078)

P-value: Non-OD=OD 0.386 0.855 0.922 0.214 0.651
Control Mean (After)XXXXXXXXXXXX
Control Mean (After, Non-OD) 0.12 0.23 0.19 0.22 0.20
Control Mean (After, OD) 0.18 0.13 0.22 0.16 0.28
N 303 303 303 303 303

Baseline Control? No No No No No
Notes: All regressions control for strata ﬁxed eﬀects and the baseline outcome of the dependent variable whenever possible.
Heteroskedasticity robust standard errors in parentheses. All regressions control for strata ﬁxed eﬀects. * p≤ 0.10, ** p≤ 0.05, ***
p≤ 0.10.
Figure A1: Administrative Data – Overdrafts in Year Before Experiment

A. Share Months in Overdraft Year Before Intervention B. Overdraft Fees Year Before Intervention (Among Overdrafters)

.15
.4
.3

.1
Fraction

Fraction
.2

.05
.1
37

0
0 .2 .4 .6 .8 1 0 200 400 600 800
Share Months in OD Annual OD Fees (GHS)

Note: Dashed lines mark median values.
Figure A2: Study Timeline

Enrollment of eligible salaried
workers Intervention (18 months)

Sept. - Oct. June Sept.-
April 2013 2013 Nov. 2013 Dec. 2013 June 2014 Oct. 2014 Feb. 2015 May 2015 2015 Oct. 2015 Nov. 2015 May 2016

Listing of Baseline Random- Beginning of Endline 1 Endline 2 Endline 3 End of SSP Endline 4 Endline 5 Last
target survey ization SSP intervention Payout extraction of
population contributions account data
using admin
data Opening of
new savings
account for
payout
Note: Activities took place in sequential order from left to right.
Notes: Activities took place in sequential order from left to right.
38