Improving Access to Savings through Mobile Money:
Experimental Evidence from Smallholder Farmers in Mozambique*
Catia Batista†, Pedro C. Vicente‡, and Dean Yang§
December 2016
Abstract:
Investment in improved agricultural inputs is infrequent for smallholder farmers
in Africa. One barrier may be limited access to formal savings. We designed and
conducted a field experiment in rural Mozambique that randomized access to a
savings account through mobile money to a sample of smallholder farmers. All
subjects were given access to mobile money and information about fertilizer
use. We also randomized whether closest farming friends were targeted by the
same intervention. We find that the savings account increased savings, the
probability of fertilizer use, by 31-36 pp, the use of other agricultural inputs. We
show that the savings account increased household expenditures, in particular
non-frequent ones. We also suggest that the network intervention decreased
social pressure to share resources and that the savings account protected farmers
against this network pressure.
JEL Codes: D14, D85.
Keywords: Mobile Money, Savings, Agriculture, Smallholder Farmers, Social
Networks, Field Experiment, Mozambique, Africa.
*
We wish to thank Simone Bertoli and Maya Shankar for helpful suggestions on the analysis of this paper.
We are also grateful to Luke Crowley, Matilde Grácio, Diogo Mendes, and Guilherme Rodrigues for
excellent field coordination. We thank seminar participants at CERDI, CSAE, IGC, and NOVAFRICA for
useful comments. We are particularly grateful to Carteira Móvel, Mcel, and the International Growth
Centre for fruitful collaboration in Mozambique. Nadean Szafman and his team, and Mindy Hernandez
offered important inputs to this project for which we are most thankful. Finally, we would like to extend an
appreciative word to the group of enumerators with whom we worked. IRB approval was secured from
Innovations for Poverty Action. We wish to acknowledge financial support from ATAI and USAID. All
errors are our responsibility.
†
Universidade Nova de Lisboa, NOVAFRICA, and IZA. Email: catia.batista@novasbe.pt.
‡
Universidade Nova de Lisboa, NOVAFRICA, and BREAD. Email: pedro.vicente@novasbe.pt.
§
University of Michigan. Email: deanyang@umich.edu.
1
1 Introduction
African farmers have a hard time saving. First, they are typically poverty-ridden smallholder
farmers. Second, they are usually unbanked. Without access to formal financial products, namely
those entailing some degree of commitment, they are easy prey to the pressures of their families
and neighbors, and to their own temptations. Importantly, saving seems crucial to break the cycle
of low investment and low agricultural productivity that is typical of many rural settings in
Africa. Improved agricultural technologies, with the potential to have clear impacts on
productivity, have yet to arrive in the African continent (e.g., fertilizer use is the lowest in the
world). Enabling access to formal savings seems to be part of the solution to this development
challenge.
At the same time, the so-called mobile money revolution is making its way in the African
continent. The first mobile money service, M-PESA, was launched in 2007 in Kenya and was
quickly adopted by a majority of the adult population of the country.1 Other countries have
followed. Based on a network of agents, standard mobile money platforms enable users to save
money in their accounts and to send money to other people. All they need is a mobile phone with
network coverage. Mobile money has an enormous potential to expand access to formal financial
products – a lot is going to happen in the next few years. However, the way to tailor mobile
money services to help farmers to save is not obvious. Indeed, it is possible that mobile money by
itself de-incentivizes savings by facilitating transfers to other people. Commitment savings are
yet to be introduced in many mobile money platforms, often because regulators have limited
evidence about their potential impact.
In this paper we report on a field experiment we designed and conducted in rural Mozambique in
2013-2015 with a sample of smallholder farmers cultivating maize in non-irrigated plots. Mobile
money has recently been launched in Mozambique. We ask whether a custom-made savings
account offered through mobile money, which remunerated average balances held, increases
savings by farmers. We also verify whether investment in improved inputs, with an emphasis on
fertilizer, and household expenditures increase as a result. Our design also enables venturing into
why access to the savings account may change farmer behavior: we aimed to induce exogenous
changes in social pressure to share resources within farming networks, by varying whether
1
See Jack and Suri (2011) and Mbiti and Weil (2011) for a detailed description of the introduction of MPESA in Kenya.
2
individual farmers and their closest farming friends are given similar opportunities. This
exogenous variation is interacted with the introduction of the savings account.
Specifically, our experimental design is a 2×2 design, based on two randomized treatments:
access to the savings account through mobile money, and the symmetric treatment of the closest
two farming friends of primary experimental subjects. All primary experimental subjects were
given: an introduction to mobile money services, which included a free mobile phone, mobile
money registration, and trials offering seed money; and an information module on urea fertilizer
use, which included the opportunity to sell maize and to purchase urea fertilizer. The savings
account was active before planting season until the time when urea fertilizer should be applied (at
the time the maize had grown knee-high).
Our measurement of outcomes is based on: (i) administrative data for mobile money transactions
made available by the mobile money operator that partnered with us in our experiment, and for
sales of maize and purchases of fertilizer; (ii) survey data related to savings; (iii) survey data for
fertilizer use; (iv) survey data for household expenditures; and (v) survey data for transfers sent
and received. Since we conducted both baseline and follow-up surveys, we are able to estimate
difference-in-difference regressions with individual fixed effects for some of the outcome
variables.
We find clear positive effects of the savings account treatment on savings in the mobile money
platform, especially during the first year, when the account was active. We also observe a positive
impact of the savings treatment on the total and average values of mobile money transactions.
The role of transfers received in the mobile money system seems to be particularly important in
mediating these effects. Overall household savings seem to increase as well. On agricultural
inputs, we report that the probability of using fertilizer increased by 31-36 percentage points at
the 1 percent level of statistical significance, when considering farmers approached individually
by experimenters. The use of other improved inputs also increased as a result. Importantly, we see
that household expenditures increased, in particular non-frequent ones, which is indicative of
welfare improvements. We have evidence that the likelihood that individuals lent/borrowed
money to/from their connections decreased with the savings treatment. The second dimension of
our experimental variation, on symmetric treatment of closest farming friends, enables us to
propose that network social pressure to share resources is a possible force working against
savings. We find some evidence, though not always statistically significant, that the symmetric
3
treatment increased the general use of mobile money, and reduced household expenditures and
lending to social networks. These patterns are consistent with lower social pressure induced by
the symmetric (network) treatment. Significant interaction effects between the two treatments hint
that the savings account enabled counteracting social pressure with regards savings.
This paper contributes to the literatures on risk sharing within social networks in rural settings,
commitment savings, input investment by smallholder farmers, and mobile money. Seminal work
by Townsend (1994) and Udry (1994) documented the importance of informal risk sharing in
rural settings for insuring against idiosyncratic risk. Indeed, social pressure to share resources
within networks is a powerful force at work in many developing countries.2 In order to counteract
these pressures but also to limit individual time-inconsistency problems, access to savings
accounts with various degrees of commitment has been documented to increase savings and
investment in different settings (e.g., Ashraf et al., 2006, Dupas and Robinson, 2013).
Consistently, Duflo et al. (2011) have shown that small discounts for fertilizer purchase just after
harvest time could significantly increase agricultural investment in fertilizer.3 Specifically for
Mozambique, in a similar setting to ours, recent contributions have tested the impact of input
subsidies and saving incentives (Carter et al., 2013, 2014, 2016). These authors find that a
matched savings program can be particularly effective at raising farmer consumption while
controlling risk. The recent mobile money literature has focused on M-PESA and its risk sharing
consequences. Jack et al. (2012) and Jack and Suri (2013) show that the consumption of
households with access to M-PESA does not suffer from idiosyncratic shocks, which implies that
decreased transaction costs for transfers drive more risk sharing. Our paper tests the impact of a
very different mechanism of change for the introduction of mobile money, in line with the
commitment savings literature. The recent paper by Jack and Suri (2016) also documents positive
effects of mobile money on savings, along with impacts on occupational choices of women.
The paper is organized as follows. In section 2 we present the context of our field experiment. In
section 3 we fully develop the experimental design, with treatments, hypotheses, sampling and
assignment to treatment, measurement, and estimation strategy. The following section provides
the econometric results, including balance tests, treatment effects on use of mobile money,
2
A study of credit cooperatives in Cameroon by Baland et al. (2011) shows members bearing significant
costs to protect their savings from friends and relatives.
3
Duflo et al. (2008) show that, in the same setting of their other study, fertilizer use in specific quantities
has a clear impact on farmer welfare through increased productivity.
4
savings, agricultural inputs use, expenditures, and transfers from/to closest farming friends. We
conclude in section 5.
2 Context
Mozambique, a country with 25.8 million inhabitants, is one of the poorest countries in the world
with GDP per capita of 1105 USD (current, PPP) in 2013 - it ranks 175 in 181 countries in terms
of GDP per capita. Despite substantial natural resource discoveries and exploration in recent
years, it is still a country with clear dependence on official aid assistance, which accounts for 57
percent of central government expenses. Agriculture is considered the key sector in Mozambique
for those interested in pro-poor economic policies. This is because it accounts for 81 percent of
employment in the country. Despite this impressive figure, the contribution of the agricultural
sector for the value added of the country is only 29 percent.4
Cereal agricultural productivity for 2011 in Mozambique was 10.4 thousand hectograms per
hectare, well below the world average, 36.6, and even below the African average, 14.4.5 Two
factors may help explaining this low agricultural productivity. First, smallholder farmers
constitute the vast majority of farmers in the country: data from the National Agricultural Survey
(TIA) in 2008 indicate that only 0.58 percent of Mozambican farmers cultivate more than 10
hectares of land. Second, investment in improved inputs is very limited: e.g., FAO Statistical
Yearbook reports that in 2011 the Mozambican average for nitrogen fertilizer use was 6.4
kilograms per hectare, which is well below both the world average (73.3) and the African average
(13.3). In an extremely poor setting like rural Mozambique, it is likely that smallholder farmers
have difficulties in saving resources, from harvest to planting, for investing in improved
agricultural inputs like fertilizers.6
Access to financial services is very limited in Mozambique, specifically in rural areas. In 2013,
only 24 bank accounts existed for each 100 Mozambican adults, and the number of bank branches
per 100,000 adults was 3.9. Both figures were below their corresponding African averages, which
4
World Development Indicators, 2015, latest available years.
FAO Statistical Yearbook, 2014.
6
Cunguara and Darnhofer (2011) show that improved agricultural technologies are associated with higher
household incomes for smallholder farmers in Mozambique when these farmers have secured access to
markets.
5
5
were 55 and 7.7, respectively.7 Saving methods for the rural population are often limited to
keeping money at home, keeping money informally with someone, and to participating in savings
groups.8 The introduction of mobile money in 2011 created expectations that the level of financial
inclusion could improve quickly in the country. Mozambique had around five million subscribers
of mobile phone services in a competitive market, and geographical coverage included 80 percent
of the population.9 mKesh became the first mobile money service operating in Mozambique: it is
offered by Carteira Móvel, a financial institution created by Mcel, the main mobile
telecommunication operator in the country. During the first years most of mKesh’s expansion
efforts, specifically in terms of agent coverage, were concentrated in urban locations.10 Even
though the Mozambican Central Bank does not allow mobile money operators to offer saving
products of their own (i.e., earning interest), mobile money can still be seen as an attractive
saving method, namely for farmers who live far from bank branches.
3 Experimental design
3.1 Treatments
Our experiment encompasses two treatments, interacted in a 2×2 design, submitted at the
individual level. The pool of primary experimental subjects includes 196 farmers at the baseline.
All primary experimental subjects were given two information modules after the baseline, one on
mobile money and one on the use of fertilizer. The first treatment is a savings treatment, which
allows individuals to receive a bonus (or interest) depending on average balances they have on
their mobile money account. The second treatment is a network treatment, by which farmers had
their two closest farming friends given the information modules on mobile money and fertilizer
use. The interaction of the savings and network treatments provides farmers’ connections with
access to the bonus as well.
7
IMF, Financial Access Survey, 2015.
Batista and Vicente (2012) report for a large sample of rural households surveyed in 2012 the following
statistics: 63 percent save money at home, 30 percent save money informally with someone, and 21 percent
participate in a savings group. Only 21 percent report any money saved in a bank account.
9
Computed from data made available by Mcel and Vodacom, the two existing Mozambican mobile
telecommunication operators in 2011. In 2012 a third operator entered the mobile phone market (Movitel).
10
M-Pesa, operated through a financial institution controlled by Vodacom, entered the mobile money
market in late 2013, after our experiment had started.
8
6
The information module was a general introduction to mKesh, the only mobile money service
being offered in Mozambique at the time of the experiment. The module started by offering a
simple mobile phone (worth 750 Meticais, close to USD 30 in an Mcel shop) and a leaflet
explaining how to use mKesh (giving an overview of all possible services on offer). See Figure 1
for a depiction of this leaflet. Verbally, enumerators stressed what saving means, what a bank is,
and some details of mKesh (ability to save there, safety based on a PIN, no need to go to the
bank). After this verbal introduction, enumerators registered each individual on mKesh using the
self-registration feature of the service, and gave 55 Meticais (close to USD 2) for cash-in
(deposit) in mKesh. In the process of the cash-in operation, enumerators introduced the local
mKesh agents to the experimental subjects. After the cash-in operation, enumerators helped each
individual to check his/her mKesh balance. Enumerators also explained how to cash-out
(withdraw) the money from mKesh (making sure individuals understood that they had to pay a
commission of 5 Meticais for that operation).
<Figure 1 near here>
The module on fertilizer use was based on a leaflet (see Figure 2), which was delivered by
enumerators and explained verbally. It focused on maize production and its main message was
‘Using fertilizer is good! This year take good care of your machamba [plot]. Increase your
production by increasing your soil fertility’. Details of fertilizer use were explained on one of the
sides of the leaflet: they included information on what farmers already do well (prepare the soil,
place seeds after first rains, use of organic fertilizer, remove unwanted plants from plot during
maize growth), and added information on how to apply urea as inorganic fertilizer two to three
weeks after germination. These details were verbally discussed at length with experimental
subjects. At the end of the module on fertilizer use, farmers were given information on the
possibility of selling their maize to a local buyer (Desenvolvimento e Comercialização Agrícola DECA). Note that the survey team was available to mediate these sales, i.e., for the maize that
farmers had to sell from the previous season (harvested in July). mKesh was one of the payment
methods made available to farmers. The survey team was also available to mediate the sale of
fertilizer, for the growing season starting in November 2013. These were resources available
during visits performed until planting season.
<Figure 2 near here>
7
The savings treatment was based on the offer of a bonus of 20% interest for the average mKesh
balance held by an individual over the period from the end of the survey team visits before
planting season to the follow-up survey in January/February 2014 (just after when urea fertilizer
should be applied). This bonus was paid in urea fertilizer. The leaflet that was distributed
announcing this treatment scheme is depicted in Figure 3. Note that, even though experimental
subjects could cash-out the money on their mobile money account, they faced a strong incentive
to maintain a high balance for as much time as possible until the follow-up survey. Interest rates
paid by banks in Mozambique approached but did not reach 10% on a full year in 2013 (as given
by the reference rate of the Bank of Mozambique). Hence the savings treatment can be seen as a
commitment-savings intervention.
<Figure 3 near here>
The network treatment gave the two closest farming friends of each treated primary experimental
subject the modules on mobile money and fertilizer use. In addition to these modules, when
interacted with the savings treatment, the network treatment also enabled access of closest friends
to the bonus for average mKesh balances. Closest farming friends were defined through questions
on identifying a primary experimental subject’s (i) closest friends in the same community that
farm non-irrigated plots, (ii) connections in the same community that farm non-irrigated plots
from whom that individual had a loan granted, and (iii) connections in the same community that
farm non-irrigated plots to whom that individual granted a loan. In addition, the selected closest
friends had to have a mobile phone number (as indicated by the primary experimental subject)
and could not be in the list of primary experimental subjects. When more than two farming
friends were mentioned across the referred questions, priority was given to connections that were
mentioned in more than one of the (i) to (iii) lists for a given individual.
3.2 Hypotheses
The experimental design that is offered by our structure of treatments is depicted in Figure 4. The
group that is not exposed to either the savings or network treatments is labeled CI (for control,
i.e., no savings treatment, and for individual, i.e., no network treatment). The group that is
exposed to the savings treatment but not to the network treatment is labeled SI. The group that is
not exposed to the savings treatment but is exposed to the network treatment is labeled CN.
8
Finally, the group facing the interaction between the two treatments is labeled SN. We denote
below (!) as our outcome variables of interest.
<Figure 4 near here>
Our first hypothesis is that the savings treatment adoption of mobile money services, savings in
particular, and investment on improved inputs, with a potential impact on household
expenditures. This is because experimental subjects face, with the savings treatment, a clear
incentive to save and, therefore, to invest in their farming businesses, with possible improvements
in household welfare. Note that all primary experimental subjects are subject to the modules on
mKesh and on fertilizer use, which guarantee that they are familiar with the specific savings
treatment proposed and with the benefits of fertilizer. Hence, taking the group of experimental
subjects that is treated individually, we expect SI to have higher levels of the referred outcomes
than CI, i.e., !(!") − !(!") > 0.
Hypothesis 1a: The savings treatment increases adoption of mobile money services, savings in
particular, investment on agricultural inputs, and household expenditures, when taking the group
of experimental subjects that is approached individually, i.e., !(!") − !(!") > 0.
Applying a similar rationale, we expect that the savings intervention increases the level of the
referred outcomes for the group of experimental subjects that is treated together with closest
connections. Consistently, we anticipate TN to have higher levels of the outcomes than CN, i.e.,
!(!") − !(!") > 0. However, if social pressure from closest farming friends is minimal when
the full network is treated, and the motive for adoption of the mKesh possibilities is protection
from one’s network, this difference could be zero, i.e., ! !" − ! !" = 0.
Hypothesis 1b: The savings treatment weakly increases the level of our outcomes of interest,
when taking the group that is approached together with closest connections, i.e., !(!") −
!(!") ≥ 0.
It is possible that faced with the network treatment alone our experimental subjects are more
likely to adopt mobile money services. We propose two possible motives. The first is that
subjects treated individually face more social pressure from their connections to lend them
money, as primary individuals were given a set of opportunities (the mKesh information and
9
fertilizer modules) that was not available to their connections. This is what we call the social
pressure explanation. The second is that primary individuals feel more confident about mKesh
services since they know other people around them know about mobile money and are likely
users as well. This is what we call the network imitation/information explanation.11 For the group
of individuals that was not given the savings treatment, this corresponds to !(!!) − !(!") > 0.
Hypothesis 2a: The network treatment increases adoption of mobile money services, when
taking the group of experimental subjects that is not given the savings treatment, i.e., !(!") −
!(!") > 0.
If one takes the first motive described above (social pressure), then it is unclear whether we
should observe the same difference as for Hypothesis 2a when taking the group of experimental
subjects that was given the savings treatment. This is because the savings treatment gives primary
subjects access to commitment savings, which could act as a shield against social pressure to
share resources. However, the second motive would simply yield the same difference as before,
i.e., !(!") − !(!") > 0.
Hypothesis 2b: The network treatment weakly increases adoption of mobile money services
when taking the group that is given the savings treatment, i.e., ! !" − ! !" ≥ 0.12
In our experimental design, we are also able to test what happens to our outcomes of interest
when individuals are exposed to the savings treatment and to the network treatment at the same
time. The social pressure explanation would yield ! !" − ! !"
− ! !" − ! !"
< 0.
Otherwise, the sign of this double difference would be unclear.13
11
Conley and Udry (2010) show the importance of social networks in learning about and adopting new
agricultural technologies. Note that a positive effect of the network treatment could also be due to an
increased perception of the value of the network in face of the introduction of mobile money. It is however
unlikely that there is an increased network externality at the level of closest (locally) farming friends.
12
Note that for both Hypothesis 2a and Hypothesis 2b we could also consider the possibility that negative
differences are observed. The motive could be that subjects treated together with their networks would freeride on each other in terms of savings, since they are aware the others were given the same set of new
opportunities.
13
Note that imitation and information would in principle yield positive interaction effects. However, if
information is not assumed additive, i.e., people either learn alone about mKesh or through their network
but not in both ways, it could yield a negative interaction effect.
10
Hypothesis 3: The savings and network treatment interaction decreases the adoption of mobile
money services when the savings treatment is taken as a shield against social pressure to share
resources, i.e., ! !" − ! !"
− ! !" − ! !"
< 0. Otherwise, the sign of the interaction
is unclear.
3.3 Sampling and assignment to treatment
This project was implemented in the districts of Manica, Mossurize, and Sussundenga, in the
Mozambican province of Manica. In each district, a set of localities, 15 in total, was identified
namely as having farmer associations. We asked for lists of farmers in each of the localities and
surveyed these farmers in a pre-project survey. 240 farmers operating non-irrigated plots who
also provided information about their connections were surveyed at that point in June-July 2013.
Within these, we were able to identify a set of 196 farmers in the same 15 localities with two
connections each (both willing to participate in the study). These 196 farmers were interviewed
during our baseline survey, which took place in July-August 2013, and form our list of primary
experimental subjects.
Each triplet at the baseline (defined as one primary experimental subject and his/her two
connections) was assigned to one of the four comparison groups (control, savings treatment,
network treatment, and interaction between savings and network treatments). The procedure was
the following. We first composed blocks of four triplets within the same locality and using
observable characteristics of primary farmers collected in the pre-project survey (type of
secondary occupation, whether he/she operated irrigated plots, whether he/she had used
fertilizer). We then randomly assigned each member of each block to a different comparison
group.
The post-intervention survey was implemented in January-February 2014, after the planting
season was over, and after the urea fertilizer could be applied in that season. Of the 196 primary
farmers, we were able to survey 186 individuals, which entails an attrition rate of 5%. We check
below balance in observable characteristics for both baseline and post-intervention samples.
3.4 Measurement
11
Our measurement can be divided in three different types of data: (i) administrative data from
mKesh; (ii) survey data from pre-project, baseline, and post-intervention surveys; and (iii) sale
records of maize and purchase records of urea fertilizer mediated by enumerators.
The administrative data from mKesh includes balance and transaction data for all experimental
subjects for the relevant period of study, starting with the end of the survey team visits before
planting season in 2013 to the end of June 2015, for approximately two years.
Sale records of maize regard sales of maize through enumerators under the arrangement with
DECA after the baseline survey, and purchase records of fertilizer regard purchases of urea
fertilizer through enumerators during the planting season in November-December 2013.
Enumerators recorded all details of these transactions.
Finally, the survey data we employ ahead, which focus on baseline and post-intervention survey
data, include information on respondent and household characteristics, mobile phone use and
mKesh literacy, agricultural practices, financial literacy and practices (including savings),
household expenses and assets, relationship with the two connected farmers, information on
transfers sent/received.
3.5 Estimation strategy
Our empirical approach is based on estimating treatment effects on the variety of outcome
variables that we have available. We now describe the main econometric specifications we
employ for the estimation of these parameters.
Our design allows us to estimate average treatment effects in different ways. Most simply, the
effect of interest (!) could be estimated through the specification:
!!,!,!"#$ = ! + !!!,! + !!,!,!"#$ , (1)
where ! is an outcome of interest, !, !, 1!are identifiers for locations, individuals, and time specifically, 1!represents the follow-up measurement -, ! = !!!
effects of interest, and !!,! = [!!!,!
!!,!
!!
!!" ! is the vector of
!!,! ×!!,! !]′ is a vector of dummy variables representing
the treatments (savings, !, and network, !) and their interaction.
12
Looking at the working hypotheses presented above, we can easily relate the hypothesized
differences between comparison groups, under the social pressure mechanism, with the
coefficients of interest in our specification, as follows:
H1a: ! !" − ! !" > 0 ⇔ !! > 0
H1b: ! !" − ! !" = 0 ⇔ !! +!!" = 0
H2a: ! !" − ! !" > 0 ⇔ !! > 0
H2b: ! !" − ! !" = 0 ⇔ !! +!!" = 0
H3: ! !" − ! !"
− ! !" − ! !"
< 0 ⇔ !!" < 0
In this setting, because of limited sample size, we add controls to compose our main
specification: although controls do not generally change the estimate for the average treatment
effect, they can help explaining the dependent variable, and therefore typically lower the standard
error of the coefficient of interest. We then have the following core specification:
!!,!,!"#$ = ! + !!!,! + !!!,! + !!,!,!"#$ , (2)
where !!,! is a vector of location and individual (demographic) controls.
We also employ a difference-in-differences specification, where baseline data are included,
assuming parallel trends in the different comparison groups. We use specifications with location
and individual controls, or with individual fixed effects, which control for time-invariant
unobservable individual characteristics. The latter is displayed here:
!!,!,! = !! + !! + !! + !"!,! ×! + !!,!,! , (3)
where !! is an individual fixed effect and ! is a time (before-after) dummy variable.
For ease of interpretation and transparency, we employ OLS estimations throughout the paper.
Given our randomization procedure at the individual level, we estimate robust standard errors in
all regressions.
13
4 Econometric results
4.1 Balance
We begin by showing balance tests for the primary farmers in the comparison groups of our
experiment. These are displayed in Tables 1. We present average values for a wide range of
observable individual characteristics of the control group (CI), and differences of this group to the
other three groups. We test the statistical significance of these differences. We also test the
overall significance of all differences by employing a joint F-test, for which we report p-values.
Note that in the first columns of the tables, we focus on the full baseline sample of primary
farmers. Because we have some attrition regarding this sample in the follow-up survey (186 out
of 196 individuals were surveyed at that point), we focus on the follow-up sample in the second
set of columns in the tables. This allows seeing whether unbalances between the comparison
groups were associated with the panel attrition.
<Tables 1 near here>
From the 37 individual characteristics that we tested, we can observe unbalances for age, number
of children, number of plots, and whether the farmer used improved seeds. These differences to
the control group concern the network or the interaction groups. They are significant at the 10
percent level, except for the number of children, which is significant at the 5 percent level. The Fstat on the null hypothesis of joint no differences is only rejected for the numbers of household
members and children. When we look at the follow-up sample, we face similar results: gender
becomes significant for both the savings and network differences to the control group, and
whether the farmer saves at home becomes significant for the network difference to the control
group, but several statistically significant differences in the baseline disappear, namely for age,
number of plots, and whether the farmer used improved seeds. Overall, we can conclude that we
do not seem to identify differences across comparison groups beyond what is statistically
acceptable (in the baseline only 4 out of 111 differences tested are found to be statistically
significant). In any event, we will employ basic demographic controls in our regressions, namely
variables for which we found statistically significant differences across comparison groups.
An additional note goes to the characterization of our sample of primary farmers. We can observe
in Tables 1 that the control group is mainly male (90 percent), average age is 43 years, most
14
farmers were born in Manica province (92 percent), the average number of household members is
6.9, and the average number of children is 4.3. As farmers, this group, on average, has been
cultivating a plot for 10 years, has 2 different plots, and their main plots have 4.3 hectares; in the
last year before the experiment, 22 percent used improved seeds for maize, 16 percent used
fertilizer for maize, and 76 percent of the maize produced was sold. On finance, 26 percent report
having a bank account, for 7 years on average; 82 percent save at home, and 14 percent contribute
to a saving group. Finally, 25 percent of these households have an improved latrine, 28 percent
have access to electricity, and 50 percent have access to piped water or a protected spring.
4.2 Administrative data: mKesh use, maize and fertilizer sales
We now turn to our analysis of treatment effects. We begin by showing results related to the use
of mKesh by exploring administrative data per transaction made available by the mKesh operator.
These are displayed in Tables 2. These data concern the period from the last visit of the survey
team before planting season in 2013 to the end of June 2015, for the total of approximately two
years. We focus our attention on the following outcome variables: log average daily savings in
mKesh (separating between the first and the second years of data), log total value of transactions
in mKesh, and the log average value of transactions in mKesh. We also look at different types of
mKesh transactions in value by using squared-roots of total value transacted (not to drop zeros),
specifically cash-ins, top-ups of airtime, transfers received, transfers sent, payments, and cashouts. Note that experimentation transactions made as part of the module on the introduction to
mKesh are not taken into account in our analysis except for savings (since experimental subjects
were faced with a decision on savings regarding trial money). For each outcome, we display two
one-difference specifications for the sample of primary farmers, since there is no baseline mKesh
data in our design, the first employing district dummies only, and the second adding individual
controls (like in specification 2 above).14 We test all five hypotheses proposed above: H1a, H2a,
and H3 correspond to individual significance tests of the three main regression coefficients; H1b
and H2b correspond to tests of linear hypotheses displayed at the bottom panel of the table. We
also add the test of joint significance of all three main regressors.
<Tables 2 near here>
14
Individual controls are basic demographic variables: gender, age, whether the individual was born in
Manica province, whether the individual has completed primary school, number of household members,
and number of children.
15
We observe that the savings treatment increased average daily savings in mKesh in the first year
of the experiment, when the experimental savings account was active. This is specifically for the
sample of primary experimental farmers that did not receive the network treatment. The
magnitude of this effect was 38-44 percent, statistically significant at the 5 or 10 percent levels.
Note that the point estimate in the second year is still positive and sizeable, even though statistical
significance is not achieved. We also report positive treatment effects of the savings treatment for
individually approached farmers when considering the total value of transactions and the average
value of transactions in mKesh. The magnitudes are 78-84 percent (total value) and 69 percent
(average value), statistically significant at the 10 percent value. These findings indicate that the
savings treatment was effective at increasing mKesh adoption for the group not given the network
treatment, in particular for savings. We can then conclude that H1a is confirmed but not a strictly
positively signed H1b, in line with our social pressure explanation. When looking at the different
types of transactions, we note that the savings treatment only produced significant positive effects
for the value of transfers received, for both the group not given the network treatment and the
group given that treatment. These effects are significant at the 1 or 5 percent levels. We also
observe that the network treatment, only when considering the sample of individuals that was not
subject to the savings treatment, seems to increase the total and average values of mKesh
transactions, even though these effects are not statistically significant. These results are
suggestive of H2a, which is consistent with the idea that when faced with the network
intervention alone, subjects feel less social pressure from their connections to share resources,
and use more mKesh. Finally, we observe that the interaction coefficient is robustly negative for
average daily savings, total value of transactions, and average value of transactions, even if
statistical significance is not achieved (borderline for total value of transactions). These effects
are suggestive that H3 is supported, namely that when the savings treatments is applied together
with the network treatment, there is less of a social pressure motive for mKesh use and savings,
and both decrease.15
We now report on the maize sales and fertilizer purchases by our primary experimental subjects,
which were mediated by our survey team. Our data are based on the actual records of these
15
The alternative explanation for a negative interaction effect, related to information that is not additive,
seems not be verified as we do not find statistically significant network or interaction effects when
employing variables related to information about mKesh (specifically behavioral variables built from
observing whether subjects were able to check mKesh balances and from asking whether they wanted to
withdraw trial mKesh money). These results are available upon request.
16
sales/purchases. From the baseline survey to the planting season, all experimental subjects had
the option to sell maize through our survey team to a local buyer. mKesh was available as a
payment method for these sales. The possibility of purchase of urea fertilizer was also made
available before planting by the survey team to all experimental subjects. We consider three
outcome variables in Table 3: whether maize was sold through the survey team, the percentage of
maize value that was sold through the survey team using mKesh, and whether fertilizer was
purchased through the survey team. The specifications are the same as the ones displayed in
Tables 2.
<Table 3 near here>
We find that the network treatment increases the likelihood that maize was sold through the
survey team, by 14-15 percentage points (significant at the 5 percent level). This is possibly due
to a network imitation/information explanation, by which networked individuals influence each
other towards selling maize through the survey team. The social pressure explanation is also
consistent with this result as individuals treated jointly with their networks of closest farming
friends may have felt lower incentive to hide their maize sales from potential pressures to share
resources. We should note, consistently with the social pressure explanation, that this effect is
only significant for subjects not facing the savings treatment. We also observe that among those
selling maize through the survey team, there is a clear impact of the savings treatment on the
likelihood of selling maize through the survey team using mKesh. This is a positive effect of 7792 percentage points for the sample of individuals not facing the network treatment, significant at
the 1 percent level. For the sample facing the network treatment, effects vary between 0.7 and
0.83, with the same level of statistical significance. These effects of the savings treatment, which
are in accordance with the effects reported on transfers received in the administrative data from
the mobile money operator, reinforce the idea that this treatment induced the use of mKesh for
saving purposes. Finally, we do not find statistical significant effects on the likelihood of buying
fertilizer through the survey team, even though we can report positive point estimates for the
savings treatment alone.
4.3 Survey-based measures of savings
We now turn to survey outcomes. In Tables 4 we show our treatment effects for information
outcomes related to the definition of savings, extensive margin of different types of savings,
17
specifically saving at home, saving in a bank account, saving with a shopkeeper, saving in gold
and saving in animals, and intensive margin of savings (in log value at the time of surveying),
namely at home, in a bank account, and aggregate over all survey-measured types of savings.
Hence, we are covering here the main alternatives to saving in mKesh as depicted in Table 2a.
We expect that the savings treatment effect is not as clearly positive as it is for mKesh savings: in
fact it could be negative if there is substitution between savings in mKesh and other types of
savings. We are also agnostic with respect to the effect of the network treatment: if one takes
social pressure as the main driving force behind this treatment, it could be positive due to lower
social pressure to share resources; however, it could also be negative if this treatment nudges
individuals to move (or to report moving) their savings out of standard (non-mKesh) savings.
Since we have available baseline data for saving methods beyond mKesh measured through both
their extensive and intensive margins, we display difference-in-difference specifications including
both district dummies and full individual controls, or including individual fixed effects (as in
specification 3 above). Note that difference-in-difference specifications may be most reliable for
saving at home, given the statistically significant difference between two comparison groups for
this variable at the baseline (shown in Table 1b).
<Tables 4 near here>
We begin by observing positive effects of the savings treatment (in the sample not targeted by the
network treatment) on the information variables related to the definition of savings, namely on
whether subjects mentioned plan and wise spending and referred friends’ help when asked about
the definition of savings. The magnitude of these effects is 10-11 (plan and wise spending) and 67 (friends’ help) percentage points, statistically significant at the 5 or 10 percent levels. We also
note positive and significant effects of the network treatment (for the sample not targeted by the
savings treatment) for the variable related to planning and wise spending, and negative and
significant interaction effects for the variable related to friends’ help. Both are consistent with the
social pressure explanation proposed in this paper.
Turning to non-mKesh types of saving, we find a range of different results. Looking at the
extensive margin results, while we see some positive and significant point estimates of the effect
of the savings treatment, namely for saving in animals (sample with network treatment), we find
some negative and significant point estimates as well, namely for saving at home (sample with
network treatment) and saving with shopkeeper (sample with no network treatment). These
18
findings indicate that the savings treatment may incentivize other types of savings beyond
mKesh, but may also lead to substituting away from some standard methods of saving. Looking
at the intensive margin results, we see positive effects of the savings treatment on the aggregate
value of savings as reported in the surveys by aggregating non-mKesh methods of saving. This is
statistically significant at the 1 or 5 percent levels when considering the sample of primary
experimental subjects that were given the network treatment: the magnitude of this effect is 131181 percent. We also identify mostly negative and significant effects of the network treatment,
consistently with the idea that experimental subjects could be moving or reporting to move their
savings out of some standard saving methods nudged by the symmetric network intervention
(possibly as free-riding on their treated networks in terms of savings). We can conclude that,
while we see positive effects on the non-mKesh aggregate value of savings, we see a mixed
pattern of treatment effects on specific methods of saving.
4.4 Agricultural inputs
We now report the treatment effects related to the use of agricultural inputs for the primary
experimental farmers. Table 5 shows effects on synthetic fertilizer use in the previous season,
both in terms of extensive and of intensive (kilograms) margins. Note that urea was the fertilizer
type that was central to the module on fertilizer use that was submitted to experimental subjects.
In this context, the survey team mediated purchases of urea. It distributed saving bonuses in urea
as well. We expect that the treatment centered on mKesh savings produces a clear positive effect
on the take-up of fertilizer. Table 5 also shows treatment effects on knowledge about using urea
fertilizer (assessed through an index of four equally weighted binary variables constructed from
four different survey questions: one asking about the appropriate distance to the plant for the
application of fertilizer, one asking about the appropriate depth for the application of fertilizer,
one asking about the appropriate quantity of fertilizer per plant, and one asking about the
appropriate timing for the application of fertilizer), on the number of workers in the subject’s
farm, on owning irrigation pumps, on employing improved seeds, and on using organic fertilizer
(all questions about the use of inputs referred to the previous season). We show one-difference or
difference-in-difference specifications depending on the outcome variable, following the models
presented in the previous tables. Baseline data are available and so difference-in-difference
specifications are possible for fertilizer use (intensive margin), and owning irrigation pumps. For
the remaining outcomes we employ one-difference specifications.
19
<Table 5 near here>
We find that the likelihood that fertilizer was used clearly increased with the savings treatment.
This effect ranges between 31 and 36 percentage points, when considering the sample not facing
the network intervention, and between 40 and 41 percentage points, when considering the sample
facing the network treatment. It is significant at the 1 percent level for all specifications
employed. This is clear evidence indicating that the savings treatment was particularly effective at
increasing fertilizer use. We do not find, however, statistically significant effects on the intensive
margin of fertilizer use, despite positive point estimates. Note that the average report of urea
fertilizer acquired was 47.3Kgs, while the average quantity of urea fertilizer given in bonuses for
the savings treatment was 0.9Kgs, and the average quantity of urea fertilizer bought from the
survey team was 2.8Kgs. We also note that in the group of farmers that received a bonus in
fertilizer, i.e., through the savings treatment, only 51 percent reported having used fertilizer. This
evidence shows that the impact of the savings treatment on fertilizer use is unlikely to have been
driven by the specific quantities of fertilizer distributed as bonuses, or even by fertilizer sold
through the survey team. Our outcome variable related to knowledge about using urea fertilizer
reinforces this picture: we find a positive effect of the savings treatment on the likelihood of
knowing about how to use urea fertilizer (for the sample not given the network treatment) ranging
between 15 and 18 percentage points, significant at the 5 or 10 percent levels. We do not find
clear patterns of network or interaction effects on use/knowledge of fertilizer.
Looking at other agricultural inputs, we find positive effects of the savings treatment for the
sample of subjects not given the network treatment for number of workers in farm, owning
irrigation pumps, using improved seeds, and using organic fertilizer. However, these are only
statistically significant for irrigation pumps: the probability of owning theses inputs increases by
7 percentage points, significant at the 10 percent level (relative to just 1 percent of CI farmers
owning these pumps). Note that the probability of using improved seeds also increases
significantly with the savings treatment when considering the sample of farmers given the
network treatment: this is a 17 percentage-point increase in the specification without individual
controls. We can conclude that we have some evidence that the savings treatment was able to
cause positive changes in agricultural investment beyond the use of fertilizer.
4.5 Household expenditure
20
We present in Table 6 treatment effects for different types of household expenditure, specifically
day-to-day and non-frequent expenditures, as well as total expenditures. These data were
collected in both surveys conducted during this experiment. We expect that the intervention
offering the savings account in mKesh increases the level of non-frequent expenditures more
clearly than the level of day-to-day expenditures. This is because it is more likely savings are
used for buying non-frequent goods and services. It is possible that the network intervention
decreases expenditures due to diminished social pressure. We present difference-in-difference
specifications in line with the previous tables.
<Table 6 near here>
Results in Table 6 show that the savings treatment increased non-frequent expenditures by 56
percent when considering the sample that was not given the network treatment, and by 83-88
percent when considering the sample that was given the network treatment. These are statistically
significant effects at the 5 or 10 percent levels. Day-to-day expenditures yield insignificant effects
for the savings treatment. Overall, this is evidence that the savings treatment was more effective
at increasing non-frequent purchases of goods and services. Interestingly, we find robust effects
of the savings treatment on total expenditures, which is suggestive of a positive welfare effect.
The magnitude of these effects ranges between 58 and 64 percent for the sample without the
network treatment (significant at the 5 or 10 percent levels), and between 61 and 62 percent for
the sample with the network treatment (significant at the 5 percent level). The network treatment
had an overall negative effect for all types of expenditures, when taking into account the sample
of individuals not given access to the savings account in mKesh. However, this effect is never
statistically significant. This finding is suggestive that lower social pressure to share resources
within social networks induced lower expenses. Overall, we find evidence that the savings
treatment induced higher aggregate household expenditures, driven by non-frequent ones.
4.6 Transfers sent and received
Finally, we analyze treatment effects for whether primary farmers lent to their closest farming
friends, and for whether they borrowed from their closest farming friends. We also report the
treatment effects on the log amounts of transfers sent and received, as well as on whether
lending/borrowing to/from closest friends were for current consumption. These results are
displayed in Table 7. This information was collected at the time of the follow-up survey. We
21
expect the savings treatment to decrease lending and the sending of transfers in particular, as
treated individuals gain an attractive option for their savings. We also expect the network
treatment to decrease lending, as social pressure to share resources decreases. However, the same
effects could be expected for borrowing and receiving transfers as well, if consumption/savings
become less/more important. We show regressions employing district dummies only, and
regressions employing district dummies and individual controls.
<Table 7 near here>
As expected, we observe that the savings treatment decreases all outcomes related to
lending/borrowing in Table 7, when considering the sample of primary experimental subjects that
was not subject to the network treatment. We see a statistically significant effect for borrowing
from closest farming friends: this is a 10 percentage-point effect, significant at the 10 percent
level. We also find a positive and significant effect on borrowing for current consumption
purposes, when considering only the individuals that borrowed from closest farming friends.16
When considering the sample that was given the network intervention, we actually observe some
significant positive effects for lending/borrowing, driven by the interaction coefficient. A possible
explanation is that individuals that are given access to the mKesh savings account jointly with
their networks are nudged to pool more resources. Also as expected, we report consistently
negative effects of the network treatment in the sample of experimental subjects that was not
given the savings intervention. These are statistically significant for log amount of transfers sent
and for borrowing from closest farming friends for consumption purposes. Note that when we
consider the sample of individuals that was faced with the savings intervention, again driven by
the interaction coefficient, we identify positive and significant effects for borrowing. The
interaction coefficient is clearly positive for both lending and borrowing. For lending it reaches
16 percentage points, significant at the 10 percent level, and for borrowing it ranges between 25
and 29 percentage points, significant at the 1 percent level. We also see positive and significant
interaction effects on borrowing for consumption purposes. In face of the social pressure
explanation of our results, this is the expected sign, as the mKesh savings account decreases
lending/borrowing but less so when networks are less demanding.
16
In face of the savings treatment, for farmers not exposed to the network treatment, it could be the case
that primary farmers would share access to the new savings account with their closest friends. Our findings
on borrowing from closest farming friends suggest that possibility did not happen.
22
5 Concluding remarks
In this paper, we analyzed the results of a field experiment where a savings account operated
through a mobile money platform was introduced to smallholder farmers in rural Mozambique.
We find that access to the savings account increased the amount of money saved in mobile
money, total transactions in that platform, and aggregate household savings. Investment in
agricultural inputs increased as result. We are also able to identify a positive effect on total
household expenditures, which is suggestive of a welfare impact. The experiment also varied
whether closest farming friends were given access to the same opportunities offered to primary
experimental subjects. Beyond the savings account, these opportunities included being introduced
to mobile money and given information about fertilizer use. We interpret asymmetry within
networks as increased social pressure to share resources. Our results are consistent with this
hypothesis. We have suggestive evidence that the symmetric network treatment increased the use
of mKesh and decreased lending/borrowing to/from closest friends. Overall, our evidence hints
that the savings account could be used to counteract social pressure.
This research shows that mobile money can be used effectively to increase financial inclusion in a
country like Mozambique. Farmers are particularly in need of financial products that enable them
to save resources from harvest to planting, namely to invest in improved inputs like fertilizer.
Access to a tailored savings account can lead to higher levels of investment by farmers. Our study
suggests that this savings account may constitute an instrument to protect savers from the
pressure their social networks can exert to share resources. Since many central banks in
developing countries see mobile money as a risky innovation, in need of tight regulation, many of
the possibilities of mobile money are still not allowed. Our evidence suggests that enabling
saving accounts to be offered through mobile money platforms is likely to be a pro-poor policy,
bringing improvements for important populations. A final note on ways forward for the research
in this paper: while we have found suggestive evidence in support social network pressure as a
central impediment for savings, our structure of treatments, which was limited by statistical
power, can be refined further, namely in terms of producing variation that can be interpreted
robustly as social network pressure.
23
References
Ashraf, Nava, Dean Karlan, and Wesley Yin (2006), Tying Odysseus to the Mast: Evidence from
a Commitment Savings Product in the Philippines, Quarterly Journal of Economics, 121(2),
pp. 635-672;
Baland, Jean-Marie, Catherine Guirkinger, and Charlotte Mali (2011), Pretending to Be Poor:
Borrowing to Escape Forced Solidarity in Cameroon, Economic Development and Cultural
Change, 60(1), pp. 1-16;
Batista, Cátia, and Pedro C. Vicente (2012), Introducing Mobile Money in Rural Mozambique:
Evidence from a Field Experiment, Universidade Nova de Lisboa, Working Paper;
Carter, Michael R., Rachid Laajaj, and Dean Yang (2013), The Impact of Voucher Coupons on
the Uptake of Fertilizer and Improved Seeds: Evidence from a Randomized Trial in
Mozambique, American Journal of Agricultural Economics, 95(5), pp. 1345-1351;
Carter, Michael R., Rachid Laajaj, and Dean Yang (2014), Subsidies and the Persistence of
Technology Adoption: Field Experimental Evidence from Mozambique, University of
Michigan, Working Paper;
Carter, Michael R., Rachid Laajaj, and Dean Yang (2016), Subsidies, Savings and Sustainable
Technology Adoption: Field Experimental Evidence from Mozambique, University of
Michigan, Working Paper;
Cunguara, Benedito, and Ika Darnhofer (2011), Assessing the Impact of Improved Agricultural
Technologies on Household Income in Rural Mozambique, Food Policy, 36, pp. 378-390;
Conley, Timothy G., and Christopher R. Udry (2010), Learning About a New Technology:
Pineapple in Ghana, American Economic Review, 100(1), pp. 35-69;
Duflo, Esther, Michael Kremer, and Jonathan Robinson (2008), How High Are Rates of Return to
Fertilizer? Evidence from Field Experiments in Kenya, American Economic Review, 98(2),
pp. 482-488;
Duflo, Esther, Michael Kremer, and Jonathan Robinson (2011), Nudging Farmers to Use
Fertilizer: Theory and Experimental Evidence from Kenya, American Economic Review,
101, pp. 2350-2390;
Dupas, Pascaline, and Jonathan Robinson (2013), Savings Constraints and Microenterprise
Development: Evidence from a Field Experiment in Kenya, American Economic Journal:
Applied Economics, 5(1), pp. 163-192;
Jack, William, Adam Ray, and Tavneet Suri (2012), Transaction Networks: Evidence from
Mobile Money in Kenya, American Economic Review, Papers and Proceedings, 103(3), pp.
356-361;
Jack, William, and Tavneet Suri (2011), Mobile Money: The Economics of M-PESA, NBER
Working Paper No. 16721;
Jack, William, and Tavneet Suri (2014), Risk Sharing and Transactions Costs: Evidence from
Kenya’s Mobile Money Revolution, American Economic Review, 104(1), pp. 183-223;
Jack, William, and Tavneet Suri (2016), The Long Run Poverty and Gender Impacts of Mobile
Money, Science, 354(6317), pp. 1288-1292;
Mbiti, Isaac, and David N. Weil (2011), Mobile Banking: The Impact of M-Pesa in Kenya, NBER
Working Paper No. 17129;
24
Townsend, Robert (1994), Risk and Insurance in Village India, Econometrica, 62(3), pp. 539591;
Udry, Christopher (1994), Risk and Insurance in a Rural Credit Market: An Empirical
Investigation in Northern Nigeria, Review of Economic Studies, 61(3), pp. 495-526.
25
Appendix
Figure 1: mKesh leaflet
Front.
Main operations conducted:
Self-registration.
Deposit.
26
Figure 2: Leaflet on fertilizer use
27
Figure 3: Leaflet on savings bonus
28
Figure 4: 2x2 experimental design
Individual Treatment - I
Network Treatment - N
Control – C
CI
CN
Savings Treatment – S
SI
SN
29
Table 1a: Primary farmers' individual characteristics - differences across treatment and control groups; for both baseline and follow-up samples
baseline sample
CI
basic
demographics
agriculture
female
0.100
age
43.388
born in Manica province
0.920
complete primary school
0.280
number of household members
6.820
number of children
4.340
time cultivating plot (months)
116.851
number of plots
2.220
size of main plot (hectares)
4.293
number of crops last year
2.520
land fertility (1-4)
2.900
used improved seeds for maize
last year
used organic fertilizer for
maize last year
used fertilizer for maize last
year
maize production last year
(Kgs)
maize production value last
year (MZN)
% maize for sale last year
0.220
0.200
0.160
2,555.789
21,050.357
0.760
savings * joint F-stat
savings
network
network
p-value
0.124
0.057
0.052
0.499
(0.085)
(0.072)
(0.064)
3.910
4.436*
-0.127
0.165
(2.410)
(2.646)
(2.673)
-0.048
-0.038
-0.007
0.845
(0.060)
(0.060)
(0.051)
0.047
-0.005
0.003
0.946
(0.100)
(0.092)
(0.091)
0.343
1.317
-0.711
0.054
(0.786)
(0.825)
(0.666)
0.742
1.738**
-0.557
0.013
(0.667)
(0.732)
(0.624)
34.924
16.443
29.460
0.462
(22.854)
(21.462)
(25.010)
-0.077
-0.220
-0.437*
0.340
(0.282)
(0.216)
(0.259)
-0.508
0.763
-0.091
0.527
(0.670)
(0.996)
(0.765)
0.092
-0.108
0.176
0.701
(0.267)
(0.267)
(0.283)
-0.063
-0.057
-0.117
0.905
(0.127)
(0.128)
(0.160)
0.127
0.172
0.193*
0.292
(0.095)
(0.112)
(0.111)
0.147
0.035
0.104
0.330
(0.094)
(0.090)
(0.098)
-0.058
0.016
0.014
0.475
(0.067)
(0.075)
(0.068)
287.589
178.655
27.446
0.965
(559.173)
(566.536)
(577.998)
2,466.071
2,365.310 10,570.512
0.818
(7,920.370) (6,615.428) (11,176.000)
0.036
0.044
0.044
0.918
(0.074)
(0.074)
(0.084)
follow-up sample
CI
0.045
44.568
0.909
0.273
6.864
4.568
122.595
2.114
4.329
2.386
2.909
0.250
0.205
0.182
2,662.222
21,780.400
0.750
savings * joint F-stat
network
p-value
0.167**
0.111*
0.091
0.132
(0.082)
(0.067)
(0.067)
2.810
3.255
-0.318
0.379
(2.506)
(2.757)
(2.814)
-0.020
-0.027
-0.000
0.966
(0.064)
(0.064)
(0.056)
0.046
0.002
0.000
0.949
(0.101)
(0.092)
(0.095)
0.434
1.274
-0.614
0.065
(0.820)
(0.881)
(0.658)
0.602
1.510*
-0.636
0.026
(0.736)
(0.776)
(0.649)
27.469
10.699
28.847
0.618
(24.143)
(22.559)
(26.912)
0.057
-0.114
-0.295
0.551
(0.275)
(0.290)
(0.222)
-0.504
0.728
-0.073
0.564
(0.727)
(1.031)
(0.826)
0.273
0.025
0.295
0.637
(0.329)
(0.261)
(0.280)
-0.079
-0.066
-0.136
0.878
(0.137)
(0.138)
(0.171)
0.112
0.142
0.182
0.443
(0.099)
(0.116)
(0.118)
0.136
0.031
0.114
0.362
(0.098)
(0.095)
(0.105)
-0.075
-0.005
-0.000
0.502
(0.071)
(0.078)
(0.072)
237.921
72.222
-73.434
0.978
(584.756)
(576.931)
(598.421)
2,295.896
1,635.267
9,840.470
0.859
(8,447.872) (7,083.919) (11,358.225)
0.059
0.054
0.045
0.856
(0.072)
(0.077)
(0.089)
savings
network
Note: Standard errors of the differences reported in parenthesis; standard errors are corrected by clustering at the location level. * significant at 10%; ** significant at 5%; *** significant at 1%.
30
Table 1b: Primary farmers' individual characteristics - differences across treatment and control groups; for both baseline and follow-up samples
baseline sample
CI
has bank account
0.260
time having a bank account
(months)
79.154
contributes to a saving group
0.140
number of saving groups
2.143
time contributing to saving
groups (months)
48.857
saving at home
0.820
total expenditure
(MZN/month)
1,407.204
owns barn
0.880
owns fridge
0.040
owns sewing machine
0.200
owns radio
0.820
owns tv
0.429
owns bike
0.700
owns motorcycle
0.100
owns generator
0.060
owns animals
0.900
owns pump
0.020
owns improved latrine
0.245
has access to electricity
0.280
has access to piped water or
protected spring
0.500
savings
expenditure
and assets
savings
network
0.128
(0.094)
-28.321
(29.651)
0.044
(0.064)
-0.921
(0.998)
-18.302
(19.522)
-0.106
(0.087)
261.097
(479.768)
0.059
(0.049)
0.062
(0.050)
-0.016
(0.086)
-0.004
(0.069)
0.020
(0.104)
-0.027
(0.099)
-0.018
(0.066)
0.062
(0.060)
0.018
(0.060)
-0.020
(0.020)
0.020
(0.098)
0.026
(0.089)
0.031
(0.105)
0.054
(0.071)
-10.621
(39.861)
0.036
(0.077)
-1.032
(1.005)
-17.857
(18.080)
-0.140
(0.085)
589.231
(411.227)
0.061
(0.055)
-0.020
(0.034)
-0.020
(0.096)
0.060
(0.073)
-0.109
(0.098)
-0.140
(0.093)
0.080
(0.077)
0.020
(0.051)
0.000
(0.062)
-0.020
(0.020)
-0.025
(0.098)
0.020
(0.090)
0.040
(0.100)
follow-up sample
savings *
network
0.066
(0.077)
-14.354
(34.955)
0.012
(0.054)
-1.000
(0.869)
-21.143
(18.092)
0.024
(0.066)
-109.973
(290.928)
0.077
(0.051)
0.003
(0.040)
-0.004
(0.086)
0.071
(0.081)
-0.016
(0.106)
0.061
(0.079)
0.030
(0.055)
0.070
(0.045)
-0.030
(0.065)
0.023
(0.036)
0.038
(0.102)
0.003
(0.080)
0.043
(0.101)
joint F-stat
p-value
CI
0.586
0.273
0.752
82.750
0.921
0.136
0.375
2.333
0.560
50.500
0.167
0.864
0.396
1,373.589
0.417
0.864
0.372
0.045
0.994
0.159
0.695
0.841
0.581
0.364
0.183
0.682
0.361
0.068
0.226
0.045
0.862
0.886
0.209
0.023
0.929
0.273
0.991
0.250
0.969
0.523
savings
network
0.132
(0.104)
-31.917
(31.547)
0.034
(0.065)
-1.083
(1.157)
-16.875
(22.558)
-0.119
(0.087)
329.716
(507.955)
0.073
(0.054)
0.040
(0.051)
0.011
(0.067)
-0.032
(0.075)
0.083
(0.099)
-0.001
(0.102)
0.017
(0.059)
0.082
(0.061)
0.029
(0.064)
-0.023
(0.023)
0.004
(0.106)
0.048
(0.090)
0.009
(0.108)
0.041
(0.077)
-14.217
(41.830)
0.040
(0.078)
-1.222
(1.157)
-19.500
(21.041)
-0.184**
(0.086)
622.845
(422.624)
0.078
(0.059)
-0.025
(0.038)
0.021
(0.086)
0.039
(0.075)
-0.044
(0.094)
-0.122
(0.096)
0.112
(0.069)
0.035
(0.049)
0.014
(0.065)
-0.023
(0.023)
-0.053
(0.102)
0.050
(0.096)
0.017
(0.097)
savings *
network
0.068
(0.085)
-17.950
(36.878)
0.023
(0.053)
-1.190
(1.023)
-22.786
(21.216)
-0.003
(0.068)
-47.557
(316.114)
0.091
(0.056)
0.000
(0.044)
0.045
(0.087)
0.045
(0.083)
0.023
(0.106)
0.068
(0.086)
0.045
(0.063)
0.068
(0.046)
0.000
(0.066)
0.023
(0.039)
0.023
(0.110)
0.045
(0.081)
0.045
(0.096)
joint F-stat
p-value
0.637
0.720
0.947
0.294
0.428
0.104
0.403
0.373
0.522
0.959
0.726
0.675
0.233
0.337
0.256
0.954
0.207
0.876
0.925
0.973
Note: Standard errors of the differences reported in parenthesis; standard errors are corrected by clustering at the location level. * significant at 10%; ** significant at 5%; *** significant at 1%.
31
Table 2a: mKesh use - administrative data (log)
dependent variable ------>
average daily savings
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
first year
(1)
(2)
0.444**
0.376*
(0.204)
(0.207)
-0.001
-0.059
(0.226)
(0.220)
-0.263
-0.115
(0.297)
(0.299)
4.333
4.343
0.400
0.231
0.162
0.397
0.339
0.309
0.035
0.052
146
142
no
yes
second year
(3)
(4)
0.348
0.301
(0.281)
(0.314)
-0.052
-0.095
(0.309)
(0.302)
-0.581
-0.353
(0.462)
(0.486)
4.018
4.018
0.524
0.882
0.067
0.231
0.393
0.646
0.000
0.024
144
140
no
yes
total value of
transactions
(5)
0.782*
(0.458)
0.417
(0.441)
-0.836
(0.608)
5.109
0.895
0.339
0.436
0.079
116
no
(6)
0.837*
(0.457)
0.411
(0.455)
-0.852
(0.635)
5.109
0.973
0.312
0.396
0.078
115
yes
average value of
transactions
(7)
0.633
(0.387)
0.142
(0.352)
-0.712
(0.524)
2.707
0.825
0.155
0.867
-0.003
116
no
(8)
0.687*
(0.360)
0.088
(0.363)
-0.620
(0.529)
2.707
0.858
0.162
0.683
0.041
115
yes
Note: All regressions are OLS. All dependent variables are based on transaction data made available by the mKesh operator for the period between the end of the survey team
visits before planting season in 2013 to the end of June 2015. Transactions include cash-ins, cash-outs, transfers sent and received, and payments; transactions that were part of
the mKesh dissemination module are not included when computing total amounts transacted. All regressions include district dummies. Controls are gender, age, whether the
individual was born in Manica province, whether the individual has completed primary school, number of household members, and number of children. Robust standard errors
reported in parenthesis. * significant at 10%; ** significant at 5%; *** significant at 1%.
Table 2b: mKesh use - administrative data (squared root)
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
total value of cash-ins
(1)
1.494
(2.243)
1.470
(2.680)
-3.393
(3.422)
6.338
0.456
0.351
0.834
0.117
196
no
(2)
1.528
(2.370)
1.285
(2.814)
-3.453
(3.605)
6.467
0.466
0.314
0.761
0.098
191
yes
total value of top-ups
(3)
-0.220
(0.983)
1.268
(1.255)
-0.595
(1.665)
2.952
0.544
0.536
0.694
0.214
196
no
(4)
0.275
(1.058)
1.715
(1.380)
-1.615
(1.869)
3.012
0.361
0.930
0.753
0.216
191
yes
total value of transfers
received
(5)
(6)
2.739***
3.125***
(0.849)
(0.916)
0.455
0.586
(0.532)
(0.617)
0.943
0.485
(1.580)
(1.756)
0.000
0.000
0.008
0.016
0.347
0.493
0.003
0.003
0.221
0.215
196
191
no
yes
total value of transfers
sent
(7)
(8)
0.067
-0.007
(0.138)
(0.194)
1.237
1.169
(0.843)
(0.849)
-0.208
-0.095
(1.076)
(1.113)
0.000
0.000
0.892
0.923
0.077
0.079
0.066
0.076
0.032
0.010
196
191
no
yes
total value of payments
(9)
0.313
(0.290)
0.061
(0.243)
-0.540*
(0.324)
0.172
0.180
0.045
0.350
0.001
196
no
(10)
0.330
(0.342)
0.105
(0.304)
-0.668
(0.450)
0.176
0.113
0.047
0.247
0.029
191
yes
total value of cash-outs
(11)
0.343
(2.144)
-0.294
(2.311)
-0.183
(3.240)
6.019
0.948
0.832
0.953
0.072
196
no
(12)
0.600
(2.290)
-0.155
(2.490)
-0.626
(3.494)
6.142
0.992
0.738
0.938
0.055
191
yes
Note: All regressions are OLS. All dependent variables are based on transaction data made available by the mKesh operator for the period between the end of the survey team visits before planting season in 2013 to the end of June
2015. Transactions that were part of the mKesh dissemination module are not included when computing total amounts transacted. All regressions include district dummies. Controls are gender, age, whether the individual was born in
Manica province, whether the individual has completed primary school, number of household members, and number of children. Robust standard errors reported in parenthesis. * significant at 10%; ** significant at 5%; *** significant
at 1%.
32
Table 3: Maize sold and fertilizer received through the survey team
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
whether maize was
sold through the
survey team
(1)
0.019
(0.057)
0.144**
(0.062)
-0.103
(0.083)
0.119
0.162
0.455
0.333
0.439
176
no
(2)
0.041
(0.060)
0.152**
(0.066)
-0.126
(0.090)
0.119
0.174
0.651
0.286
0.442
173
yes
% maize value sold
through the survey
team using mKesh
(3)
0.771***
(0.154)
0.003
(0.003)
0.059
(0.189)
0.000
0.000
0.747
0.000
0.815
27
no
(4)
0.917***
(0.179)
0.055
(0.098)
-0.105
(0.241)
0.000
0.000
0.849
0.000
0.776
27
yes
whether fertilizer was
purchased through the
survey team
(5)
0.044
(0.047)
0.007
(0.040)
-0.009
(0.063)
0.026
0.422
0.973
0.359
-0.001
174
no
(6)
0.054
(0.049)
0.010
(0.036)
-0.021
(0.063)
0.026
0.467
0.833
0.355
-0.030
170
yes
Note: All regressions are OLS. All dependent variables are based on transaction data registered by the survey team during all visits before
planting season. All regressions include district dummies. Controls are gender, age, whether the individual was born in Manica province, whether
the individual has completed primary school, number of household members, and number of children. Robust standard errors reported in
parenthesis. * significant at 10%; ** significant at 5%; *** significant at 1%.
33
Table 4a: Savings definition
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
when asked about
definition of savings
mentioned plan and
wise spending
when asked about
definition of savings
mentioned friends'
help
(1)
0.105**
(0.046)
0.059*
(0.033)
-0.091
(0.069)
0.000
0.786
0.614
0.076
0.000
181
no
(3)
0.066*
(0.036)
0.004
(0.006)
-0.072*
(0.040)
0.000
0.401
0.072
0.738
0.055
181
no
(2)
0.097**
(0.048)
0.063*
(0.034)
-0.083
(0.067)
0.000
0.799
0.740
0.076
-0.026
177
yes
(4)
0.064*
(0.035)
-0.008
(0.011)
-0.058*
(0.034)
0.000
0.574
0.063
0.799
0.091
177
yes
Note: All regressions are OLS. All dependent variables are based on transaction data registered by the survey team
during all visits before planting season. All regressions include district dummies. Controls are gender, age, whether
the individual was born in Manica province, whether the individual has completed primary school, number of
household members, and number of children. Robust standard errors reported in parenthesis. * significant at 10%;
** significant at 5%; *** significant at 1%.
Table 4b: Saving methods beyond mKesh - extensive margin
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
difference-in-differences
fixed effects
saving at home
(1)
0.161
(0.121)
0.135
(0.125)
-0.351**
(0.177)
0.785
0.141
0.085
0.658
0.034
371
yes
yes
no
(2)
0.136
(0.112)
0.156
(0.127)
-0.342**
(0.166)
0.777
0.095
0.082
0.660
0.028
380
no
yes
yes
saving in bank account
(3)
-0.030
(0.132)
0.064
(0.125)
0.008
(0.190)
0.247
0.872
0.614
0.751
0.053
373
yes
yes
no
(4)
-0.021
(0.059)
0.059
(0.069)
-0.015
(0.099)
0.245
0.650
0.536
0.765
0.000
382
no
yes
yes
saving with shopkeeper
(5)
-0.157**
(0.071)
-0.085
(0.060)
0.242**
(0.101)
0.043
0.229
0.052
0.986
0.012
373
yes
yes
no
(6)
-0.152**
(0.064)
-0.085
(0.060)
0.237**
(0.093)
0.043
0.215
0.036
1.000
0.032
382
no
yes
yes
saving in gold
(7)
-0.069
(0.058)
-0.084*
(0.051)
0.086
(0.080)
0.032
0.753
0.977
0.244
-0.002
373
yes
yes
no
(8)
-0.065
(0.058)
-0.082
(0.052)
0.079
(0.081)
0.032
0.812
0.963
0.258
0.016
382
no
yes
yes
saving in animals
(9)
-0.094
(0.091)
-0.160*
(0.088)
0.314**
(0.129)
0.086
0.017
0.102
0.508
0.019
373
yes
yes
no
(10)
-0.091
(0.093)
-0.150*
(0.091)
0.309**
(0.134)
0.085
0.026
0.109
0.478
0.031
382
no
yes
yes
Note: All regressions are OLS. All dependent variables are based on survey questions asked in both the follow-up and baseline surveys. All regressions without fixed effects include district dummies.
Controls are gender, age, whether the individual was born in Manica province, whether the individual has completed primary school, number of household members, and number of children. Robust
standard errors reported in parenthesis. * significant at 10%; ** significant at 5%; *** significant at 1%.
34
Table 4c: Saving methods beyond mKesh - intensive margin (log)
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
difference-in-differences
fixed effects
saving at home
(1)
-0.212
(0.334)
-0.606**
(0.308)
0.272
(0.476)
7.290
0.863
0.362
0.119
0.052
133
no
no
no
(2)
-0.329
(0.302)
-0.822***
(0.296)
0.805*
(0.444)
7.290
0.145
0.958
0.289
0.245
132
yes
no
no
saving in bank account
(3)
0.393
(0.865)
0.269
(0.911)
0.411
(0.958)
7.613
0.176
0.190
0.261
-0.038
61
no
no
no
(4)
0.675
(0.851)
0.225
(0.925)
0.333
(0.942)
7.613
0.164
0.274
0.229
-0.044
59
yes
no
no
overall savings
(aggregate)
(5)
(6)
0.199
0.119
(0.478)
(0.471)
-1.066*
-1.381**
(0.572)
(0.543)
1.107
1.691**
(0.795)
(0.793)
7.665
7.665
0.046
0.006
0.940
0.591
0.679
0.454
0.017
0.099
166
163
no
yes
no
no
no
no
Note: All regressions are OLS. All dependent variables are based on survey questions asked in both the follow-up and baseline surveys. All
regressions without fixed effects include district dummies. Controls are gender, age, whether the individual was born in Manica province,
whether the individual has completed primary school, number of household members, and number of children. Robust standard errors reported in
parenthesis. * significant at 10%; ** significant at 5%; *** significant at 1%.
35
Table 5: Agricultural inputs
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
difference-in-differences
fixed effects
fertilizer
(1)
0.311***
(0.117)
-0.143
(0.105)
0.095
(0.164)
0.194
0.000
0.701
0.033
0.167
373
yes
yes
no
(2)
0.359***
(0.091)
-0.124
(0.078)
0.038
(0.136)
0.191
0.000
0.439
0.008
0.238
382
no
yes
yes
kgs fertilizer
(3)
12.107
(29.687)
5.611
(28.525)
-12.652
(48.759)
46.000
0.988
0.844
0.865
0.021
49
no
no
no
(4)
25.322
(36.633)
20.043
(33.351)
-16.825
(53.767)
46.000
0.824
0.935
0.491
-0.030
47
yes
no
no
knowledge about using
urea fertilizer
(5)
(6)
0.154*
0.175**
(0.080)
(0.078)
-0.084
-0.059
(0.076)
(0.075)
-0.044
-0.075
(0.112)
(0.114)
0.545
0.545
0.160
0.210
0.124
0.106
0.728
0.605
0.026
0.086
184
180
no
yes
no
no
no
no
number of workers in
farm
(7)
(8)
0.093
0.323
(1.344)
(1.253)
-0.112
-0.459
(1.339)
(1.234)
0.123
0.636
(1.682)
(1.558)
7.386
7.386
0.842
0.370
0.992
0.867
0.944
0.733
-0.013
0.054
186
182
no
yes
no
no
no
no
irrigation pumps
(9)
0.069*
(0.040)
0.022
(0.021)
-0.090
(0.055)
0.011
0.583
0.180
0.989
-0.005
370
yes
yes
no
(10)
0.067*
(0.038)
0.023
(0.023)
-0.090
(0.055)
0.011
0.566
0.183
1.000
0.012
379
no
yes
yes
improved seeds
(11)
0.028
(0.105)
-0.097
(0.105)
0.146
(0.147)
0.523
0.092
0.642
0.475
0.019
185
no
no
no
(12)
0.072
(0.108)
-0.033
(0.108)
0.078
(0.151)
0.523
0.135
0.654
0.277
0.091
181
yes
no
no
organic fertilizer
(13)
0.115
(0.106)
0.044
(0.104)
-0.120
(0.147)
0.372
0.960
0.462
0.717
0.024
182
no
no
no
(14)
0.154
(0.107)
0.077
(0.104)
-0.194
(0.152)
0.372
0.706
0.282
0.734
0.025
178
yes
no
no
Note: All regressions are OLS. All dependent variables are based on survey questions asked in the follow-up survey or both the follow-up and baseline surveys. All regressions without fixed effects include district dummies. Controls are gender, age, whether the individual was born in
Manica province, whether the individual has completed primary school, number of household members, and number of children. Robust standard errors reported in parenthesis. * significant at 10%; ** significant at 5%; *** significant at 1%.
36
Table 6: Household expenditures (log)
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
difference-in-differences
fixed effects
day-to-day expenditures
(1)
0.243
(0.339)
-0.075
(0.330)
0.125
(0.461)
6.934
0.252
0.880
0.341
0.047
308
yes
yes
no
(2)
0.160
(0.277)
-0.120
(0.276)
0.268
(0.410)
6.919
0.159
0.624
0.273
0.029
315
no
yes
yes
non-frequent
expenditures
(3)
(4)
0.350
0.565*
(0.374)
(0.337)
-0.194
-0.236
(0.337)
(0.346)
0.479
0.319
(0.526)
(0.490)
5.592
5.580
0.027
0.014
0.484
0.810
0.083
0.062
0.220
0.356
328
334
yes
no
yes
yes
no
yes
total expenditures
(5)
0.581*
(0.330)
-0.095
(0.309)
0.026
(0.449)
7.045
0.049
0.832
0.100
0.166
353
yes
yes
no
(6)
0.638**
(0.299)
-0.103
(0.268)
-0.023
(0.397)
7.034
0.019
0.666
0.075
0.320
360
no
yes
yes
Note: All regressions are OLS. All dependent variables are based on survey questions asked in both the follow-up and baseline surveys. All regressions
without fixed effects include district dummies. Controls are gender, age, whether the individual was born in Manica province, whether the individual has
completed primary school, number of household members, and number of children. Robust standard errors reported in parenthesis. * significant at 10%;
** significant at 5%; *** significant at 1%.
37
Table 7: Lending/borrowing and transfer amounts
dependent variable ------>
coefficient
savings - β S (H1a)
standard error
coefficient
network - β N (H2a)
standard error
coefficient
savings*network - β SN (H3)
standard error
mean dep. variable (CI group)
β S + β SN = 0 (H1b)
F-stat p-value
β N + β SN = 0 (H2b)
F-stat p-value
β S + βN + β SN = 0
F-stat p-value
r-squared adjusted
number of observations
controls
lent to closest farming
friends
(1)
-0.032
(0.064)
-0.093
(0.063)
0.163*
(0.089)
0.193
0.035
0.261
0.572
0.013
186
no
(2)
-0.012
(0.068)
-0.083
(0.064)
0.129
(0.093)
0.193
0.067
0.474
0.618
0.010
182
yes
log amount transfers sent
(3)
-0.456
(1.315)
-0.550
(1.196)
0.267
(1.454)
7.637
0.770
0.720
0.526
-0.208
21
no
(4)
-0.760
(2.029)
-2.059*
(1.174)
1.255
(2.265)
7.637
0.571
0.678
0.186
-0.218
21
yes
lent to closest farming
friends for current
consumption
(5)
-0.013
(0.243)
-0.152
(0.263)
0.447
(0.340)
0.400
0.077
0.204
0.201
0.083
36
no
(6)
0.166
(0.236)
-0.057
(0.259)
0.167
(0.359)
0.400
0.175
0.650
0.196
0.238
36
yes
borrowed from closest
farming friends
log amount transfers
received
borrowed from closest
farming friends for
current consumption
(7)
-0.104*
(0.055)
-0.092
(0.056)
0.286***
(0.082)
0.170
0.003
0.001
0.190
0.061
186
no
(9)
-0.117
(1.019)
-0.735
(1.022)
0.002
(1.272)
7.453
0.879
0.325
0.433
-0.112
29
no
(11)
-0.846***
(0.177)
-0.516**
(0.250)
1.180***
(0.285)
0.750
0.161
0.000
0.361
0.221
34
no
(8)
-0.087
(0.055)
-0.063
(0.055)
0.246***
(0.079)
0.170
0.007
0.002
0.155
0.087
182
yes
(10)
-0.155
(1.079)
-1.577
(1.115)
0.387
(1.363)
7.453
0.763
0.121
0.216
-0.143
27
yes
(12)
-1.124***
(0.311)
-0.378
(0.343)
1.256**
(0.502)
0.750
0.714
0.009
0.269
0.189
33
yes
Note: All regressions are OLS. Dependent variables are based on survey questions asked in the follow-up survey. The dependent variables in regressions (1)-(2) and (9)-(10) take value 1 if both friends were lent or borrowed by/from the
individual; they take value 0.5 if just one friend was lent or borrowed; they take value 0 if no friend was lent or borrowed. The remaining dependent variables are binary except log amounts. All regressions include district dummies. Controls
are gender, age, whether the individual was born in Manica province, whether the individual has completed primary school, number of household members, and number of children. Robust standard errors reported in parenthesis. *
significant at 10%; ** significant at 5%; *** significant at 1%.
38