The Persistent Power of Behavioral Change: Long-Run Impacts of Temporary Savings Subsidies for the Poor

Transcription

1 The Persistent Power of Behavioral Change: Long-Run Impacts of Temporary Savings Subsidies for the Poor Simone Schaner April 6, 2015 Abstract I use a field experiment in rural Kenya to study how temporary incentives to save impact long-run economic outcomes. Men who were randomly selected to receive a six-month interest rate subsidy on an individual bank account had significantly more total income and assets 2.5 years after the subsidies expired. These changes are much larger than the short-run impacts on experimental bank account use and almost entirely driven by increased rates of entrepreneurship. In contrast, interest subsidies on women s individual bank accounts had no long-run effects, while interest subsidies on joint bank accounts increased investment in household public goods and led to greater spousal consensus over financial matters. My results suggest that participants activated different behavioral savings rules in response to the different subsidies, the effects of which are highly persistent. I thank Erica Field, Dean Karlan, Rohini Pande, Joshua Schwartzstein, Jonathan Zinman, and seminar participants at Boston University, Brown University, IFPRI, the Tufts University, and the University of Missouri for numerous useful comments. This project would not have been possible without the tireless assistance, hard work, and commitment of many employees of Family Bank. I am particularly indebted to Victor Keriri Mwangi, Steve Mararo, and Michael Aswani Were. I also thank James Vancel, Noreen Makana, and Thomas Ginn for superb field management and the IPA field officers for their excellent assistance with the data collection. I gratefully acknowledge the financial support of the Russell Sage Foundation, the George and Obie Shultz Fund, MIT s Jameel Poverty Action Lab, the National Science Foundation s Graduate Research Fellowship, and the Yale Savings and Payments Research Fund at Innovations for Poverty Action, sponsored by a grant from the Bill & Melinda Gates Foundation. All errors are my own. simone.schaner@dartmouth.edu 1

2 1 Introduction Poor people in the developing world often use diverse, remarkably sophisticated financial strategies to make ends meet (Collins et al. 2009). At the same time, there is growing evidence that some of these individuals fail to exploit lucrative investment opportunities. For example, recent capital drop experiments have found very large (on the order of 5-30 percent per month) marginal returns to capital among microenterprises in contexts as varied as Sri Lanka, Ghana, India, Mexico, and Uganda 1. Other researchers have documented individuals regularly revolving debt at interest rates as high as 10 percent per day (Aleem 1990; Ananth et al. 2007; Banerjee and Duflo 2007). These observations conjure the puzzling image of a pauper sitting atop an untapped goldmine why don t these individuals save a little bit more today to take advantage of opportunities to reduce their debt, grow their businesses, and lift themselves out of poverty in the future? One possibility is that people are held back by their own behavioral biases an entrepreneur may struggle to realize his goals if he gives in to daily temptations to spend instead of regularly setting aside money for business investment, for example. 2 Most microeconomic research on behavioral barriers to saving has started with the premise that biases are driven by fixed preference parameters, and must therefore be addressed with permanent interventions, such as an always-available commitment device (see Karlan and Morduch 2010 for a review). In this paper, I take a different approach and ask whether individuals can sustain better savings and investment practices after exposure to temporary incentives for behavior change. To this end, I analyze the results of a field experiment that I conducted in rural Kenya in 2009, in which married couples were given the opportunity to open new formal bank accounts. All participating couples could open up to three accounts: an account in the name of the husband, an account in the name of the wife, and a joint account. Before participants decided which accounts to open, each account was randomly assigned a temporary interest subsidy, which lasted for six months and ranged from zero to 20 percent in annual terms. The randomization was conducted so that all subsidies were completely independent of one another thus, the experiment created variation in overall incentives to save and variation in the way that couples were incentivized to save (e.g. jointly versus individually). I find that in the short-run, the experiment had its intended effect: study participants 1 For Sri Lanka, see de Mel et al. (2008) and de Mel et al. (2012). For Ghana see Udry and Anagol (2006) and Fafchamps et al. (2014). For India see Banerjee and Duflo (2012) and Field et al. (2013). For Mexico see McKenzie and Woodruff (2008). For Uganda see Blattman et al. (2014). 2 Alternatively, investment could be hampered by credit or savings constraints, especially when production technologies are non-convex (Banerjee and Newman 1993). Intrahousehold issues may also limit some highreturn individuals ability to invest (Udry 1996). 2

3 were much more likely to open and use accounts with higher interest rates during the sixmonth subsidy period. I then explore the long-run effects of the subsidies in two phases. In my first set of main results, I show that the temporary subsidies changed long-run savings behavior, but that the nature of the impact depends how the household is incentivized to save. Endline data collected 2.5 years after the subsidies expired shows that subsidizing the husband s account reallocated assets from women to men, and as a result both male and overall household income increased. These gains are driven by growth in male entrepreneurship men who received the highest interest subsidy on their individual account were significantly (12 percentage points) more likely to be entrepreneurs and had substantially more business profits and capital at endline. In contrast, interest subsidies to women s accounts had no discernible long-run impacts, while interest subsidies to joint accounts increased investment in joint assets such as home renovations and livestock. Couples with higher joint subsidies also reported greater levels of spousal agreement regarding consumption and savings decisions. This suggests that the joint subsidy compelled couples to work together towards a mutually-agreeable savings goal, while the husband s individual subsidy spurred men to invest in more independent (and, based on the income results, higher return) ventures. One striking feature of the results for men is that the long-run impacts are much larger than the short-run changes in experimental bank account use. For example, moving from the lowest to the highest husband s subsidy increased short-run deposits into the husband s account by $8.74, while the long-run treatment effect on monthly male-owned business profit is $15. Although the long-run effects are large, they are quite consistent with treatment effects on short-term account use and the returns to capital suggested by the data, which are well within the range of returns found by other researchers. 3 Even so, these results beg the question: why didn t men revert back to their old savings and investment behaviors after the subsidies expired? What mechanism allowed them to translate modest short-run changes into meaningful long-run economic gains? My second set of main results seeks to answer these questions. Here, I consider three leading hypotheses in turn. The first hypothesis rests on non-convexities: if production requires large, lumpy investments then short-run savings subsidies could have persistent effects on income and capital if they help individuals make a big push to save past the nonconvexity. To test this hypothesis, I exploit the fact that 20 percent of respondents were randomly selected to receive a cash payout as part of the baseline discount factor elicitation procedure. The cash payments were delivered within three months of the baseline 3 For example, the male long-run treatment effects on income and assets could be attained by making an additional $1 investment in productive capital and reinvesting the proceeds at a 16 percent monthly return for the next 2.5 years. 3

4 sessions and could be deposited in respondents newly-opened bank accounts. Although the cash payment increased short-run bank account balances much more than the interest rate subsidies (thereby delivering a bigger push ), it had no impact on long-run economic outcomes. This suggests that nonconvexities alone are not enough to explain the long-run effects of the husband s subsidy. The second hypothesis postulates a banking channel. If entrepreneurs are savings constrained, then bank accounts could be valuable if they help ease these constraints (Dupas and Robinson 2013). 4 One appealing feature of the experimental design is that all participants gained access to a new formal bank account I am therefore able to rule out the hypothesis that my results are driven by access to new accounts per se. However, since the subsidies increased short-run use of the new accounts, the subsidies could have helped savings-constrained men learn about the benefits of bank accounts, or helped men gain access to bank loans. Yet men with and without baseline exposure to bank accounts benefit from the individual interest subsidy, and I find no impact of the male subsidy on either overall or bank debt. Given this, a banking channel mechanism seems unlikely. The third possibility I consider is that behavioral mechanisms are needed to explain my results. A notable prediction of models of habit formation is that behavior change induced by temporary incentives can persist after those incentives are removed (Becker and Murphy 1988). 5 The subsidies could have also induced individuals to make plans or adopt new financial heuristics to support sustained behavior change. I collected a second wave of followup data that focused on individuals budgeting behaviors and savings goals and attitudes to explore these hypotheses more directly. I find that men treated with higher individual interest rates were significantly more likely to explicitly budget for business expenses 3.5 years after the interest subsidies expired, and that this increase is entirely driven by growth in downwardly-rigid business budgets: that is, budgets in which the respondent stated he would not reduce the allocated amount to meet an unexpected expense. The husband s subsidy also increased the overall share of budgets deemed downwardly rigid. Thus, the data fit well with a model in which men used non-fungible mental accounts (Thaler 1999) as a means of sustaining better savings and investment habits in the long run. 4 More specifically, Dupas and Robinson (2013) find that giving female Kenyan entrepreneurs access to basic bank accounts substantially increases business investment, at least in the near term (4-6 months). Evaluations of bank accounts targeted to more general populations, such as Prina (2014), find more limited impacts on income and savings. 5 In this case, temporary incentives could be especially effective if individuals procrastinate paying the up-front costs of developing a virtuous habit due to time-inconsistent preferences (O Donoghue and Rabin 1999). Consistent with this, I find larger treatment effects for men who displayed quasi-hyperbolic preference reversals at baseline. I cannot, however, reject that the responses of hyperbolic and non-hyperbolic men are the same. 4

5 The main contribution of this paper is to demonstrate that temporary financial incentives can generate persistent changes in savings and investment behavior, and that these changes impact broader measures of economic welfare. To date, most research on temporary incentives for behavior change has focused on education or health-related behaviors (such as gym attendance or smoking) in developed-country contexts. Here, studies that track outcomes after incentives expire tend to do so for a limited period (e.g. one year or less after the intervention ends) and usually find modest-to-no impacts in the longer run, with treatment effects decaying over time. 6 This paper provides leading evidence that temporary incentives can have amplified effects in the long run, at least in the financial domain. 7 My results also contribute more broadly to existing research on saving in the developing world. Researchers have evaluated a range of services designed to address behavioral barriers to saving (Ashraf et al. 2006; Brune et al. 2013; Karlan et al. 2013; Kast et al. 2013; Dupas and Robinson 2014; Callen et al. 2014). Although this literature has identified a number of services that help individuals increase savings balances, there is little evidence as to whether, or under what conditions, behavioral changes persist after services are discontinued my paper helps fill this gap. It is also important to note that my results stand in sharp contrast to most literature on business training and financial literacy. These studies evaluate programs that utilize educational modules, rather than financial incentives, to alter financial behavior and generally find limited impacts on economic outcomes, especially in the long run (McKenzie and Woodruff 2014; Karlan et al. 2014). In this sense my results align with Cole et al. (2011), who find that financial subsidies are more effective than financial literacy training at boosting takeup and long-run use of savings accounts in Indonesia. One implication of my results is that some low-income entrepreneurs with high marginal returns to capital may have these high returns because they are stuck in a behavioral poverty trap (Kraay and McKenzie 2014). On the other hand, my results also present cause for 6 In terms of financial behavior, De Mel et al. (2013) find evidence that deposit collection services helped Sri Lankan study participants form a savings habit, but the authors also find that this effect attenuates over time. Studies of incentives to exercise or lose weight generally find that initial short-run effects dwindle over time, especially when habits are interrupted by holiday breaks (Charness and Gneezy 2009; John et al. 2011; Acland and Levy 2015; Royer et al. 2012). Volpp et al. (2009) find that financial incentives to quit smoking have decaying impacts over the longer-run, while Giné et al. (2010) find that a commitment savings account tied to smoking behavior had small effects that lasted at least six months after savings were released. Results in education are more promising: Jackson (2010) finds that paying students for passing advanced placement tests improved standardized test scores and increased college matriculation, while Baird et al. (2015) find that a conditional cash transfer improved long-run socio-economic outcomes for out-of-school girls in Malawi. Girls who were enrolled in school at program baseline did not experience long-run benefits, however. 7 I speculate that one reason for this difference could be the fact that some individuals in my sample had very high (and potentially very salient) returns to behavior change, which could provide dynamic feedback to help reinforce new behaviors. To the extent that women have lower marginal returns to capital than men, this could also help rationalize the gender differences that I observe. 5

6 optimism: individuals have the ability to break out of this trap on their own, provided they are given appropriate stimulus to change. However, I caution that my results should not be taken to imply that interest subsidies are a panacea. Two recent papers (Kast et al. 2013; Karlan and Zinman 2014) find that much lower interest subsidies (both in absolute and real terms) have no impact on economic outcomes. I therefore prefer to interpret my results as demonstrating that the right incentives targeted to the right population can have meaningful, lasting effects. Further work would be needed to generate additional evidence on optimal design and targeting of these incentives to maximize policy impact. The remainder of the paper proceeds as follows: Section 2 begins by describing the experimental context and design, then Section 3 presents the main results. Section 4 discusses potential mechanisms with a focus on men and Section 5 concludes. 2 Experimental Design and Data 2.1 Experimental Context The experiment was conducted between July and September 2009 in partnership with Family Bank of Kenya. Study participants were recruited from rural and semi-rural areas near the bank s Busia branch in Western Province. Even though six formal banks were operating in Busia at the time, most of these banks did not offer low-cost accounts suitable for a lowincome clientele. Importantly, Family Bank had begun to market a new, low-fee account just before the onset of the experimental activities. In contrast to traditional Kenyan bank accounts, which required relatively large minimum balances (around Ksh 1,000, or US$12.50 at an exchange rate of Ksh 80 per dollar) and charged monthly account maintenance fees, the new Family Bank account had a minimum deposit of Ksh 100 ($1.25), no maintenance fees, and no deposit fees. The only fees charged were for withdrawals, which cost Ksh 62 over-the-counter and Ksh 30 at the ATM. Like most bank accounts on the market, the new Family Bank account did not bear any interest. Just one other bank apart from Family Bank offered a similar low-cost account when the experiment began. The bank accounts in this study were therefore a relatively new technology and many study participants were unfamiliar with them. However, when interpreting the results it is important to keep in mind that Kenya s financial services landscape has evolved dramatically since 2009 and by 2012 (the time of the endline) nearly all banks offered low-cost accounts. Banking services have also been integrated into mobile money products, allowing individuals to save and apply for loans via a cell phone. Thus, while the experimental accounts dominated most other accounts on the market in 2009, this was no 6

7 longer the case at the time of the endline survey. 2.2 Experimental Design The experiment targeted married couples who did not have any accounts with Family Bank but expressed interest in opening one. Trained enumerators recruited couples in 19 communities surrounding the bank branch. Couples who expressed initial interest were issued invitations to attend a group meeting at a local primary school. All baseline interviews, account opening paperwork, and randomization activities were conducted at these meetings. Upon arrival, couples were informed that they could open up to three accounts with Family Bank at the meeting a joint account, an individual account for the husband, and an individual account for the wife. All accounts opened at the meetings were funded with the Ksh 100 minimum deposit to eliminate barriers to opening. This amount could not be withdrawn, so participants had little incentive to open accounts they knew they would never use. Before deciding which accounts to open, participants were given the opportunity to draw a temporary 6-month interest rate for each account. This process was designed so that interest rates on the three accounts were independent of one another. All randomization was conducted in the field, with respondents drawing folded envelopes from plastic bins. 8 Individual accounts were assigned an annual interest rate of either 0, 4, 12, or 20 percent with equal probability, while joint accounts were assigned an annual interest rate of either 4, 12, or 20 percent with equal probability. The interest rates in the experiment were purposely chosen to exceed market rates by a large margin, with the hope that such substantial subsidies would stimulate a short-run saving response. At the time, most formal financial institutions offered no interest on small-scale savings balances at best participants could have earned percent annually elsewhere. 9 Since many study participants had little-to-no experience with banks, project staff carefully explained what an interest rate was, provided numerical examples for each interest rate, and explained that the promotion would only last for six months. While very few couples chose to open all three bank accounts, all couples opened at least one account as a result 99 percent of study participants had access to either a newly-opened joint account or a newlyopened individual account in their own name (Appendix Table A1 shows the distribution of account opening choices). Thus, the experimental design allows me to study the impact of 8 Respondents took separate draws for each potential account. The field staff were carefully trained not to allow a respondent more than one draw for each treatment. I find no evidence of protocol problems when comparing the empirical distribution of treatments to the theoretical distribution of treatments. 9 For comparison, average inflation was 9.3 percent in 2009 and 2.0 percent in 2010 therefore most bank accounts offered a negative real rate of return. 7

8 interest subsidies on different account types holding access to a new bank account constant. Participants were also given a pocket-sized card for each account that they opened, which featured a reminder to save and, when applicable, the interest rate. 10 Before leaving the meeting all individuals participated in a final drawing for a cash prize. 11 This prize was the incentive for baseline questions on rates of time preference, which consisted of multiple price list style choices between a smaller monetary amount at time t and a larger amount at time t + τ (see the Data Appendix for additional detail). All individuals had a 20 percent chance of being selected for a cash prize selected individuals then drew one of their time preference questions at random for payout. Payouts ranged from Ksh 10 to Ksh 300, with an average payout of Ksh All payouts could be either picked up in person at the project field office or deposited automatically into the individual s newlyopened bank account. In practice, 77 percent of individuals elected to have their payouts deposited into a bank account, even though accessing these funds would require payment of the Ksh 62 withdrawal fee (there was no fee to pick up funds at the field office, although respondents only had a one-month window to claim their cash). This suggests that most individuals saw some value to the bank accounts and intended to use them for saving. 2.3 Data and Randomization Verification My analysis uses data from four sources. The first is a baseline survey conducted during the experimental sessions. The baseline collected basic demographic information, information on rates of time preference, and data on income and use of several popular savings devices. Second, I use three years of administrative data from the bank to get an accurate measure of short- and long-run use of the experimental accounts. This administrative data includes the date and amount of all transactions posted to experimental accounts. Finally, I use data from two waves of endline surveys. The first wave was conducted between August and November of 2012, approximately three years after the initial experiment. The wave 1 endline collected detailed information about respondents financial lives: in addition to basic demographic information it asked about income, savings, and debt by source as well as financial transfers and household decision-making. The wave 2 endline was conducted between July and August of 2013 and was much shorter. This secondary wave was informed by results from the wave 1 endline and was explicitly designed to collect additional detail on study participants 10 All cards, regardless of the interest rate, featured the following message in Swahili: keep this card as a reminder to use your new Family Bank account to build your savings for the future. 11 Individuals were also selected for information sharing treatments and free ATM card treatments at this stage. I do not discuss these interventions further as they have no impact on the results in this paper. 12 These amounts were designed to be substantial enough to ensure that individuals made choices carefully for comparison the median weekly income at baseline was Ksh

9 budgets, goals, and attitudes towards savings. The Data Appendix gives additional detail on the data sources and how key variables used in the analysis were constructed. A total of 1,558 individuals (779 couples) opened 1,152 bank accounts during the initial experiment. The survey team was able to re-interview 1,417 (91 percent) of these individuals during the wave 1 endline and 1,331 (85 percent) of these individuals during the wave 2 endline. Table 1 presents baseline demographic characteristics and verifies that these characteristics are not systematically correlated with the interest subsidies. The first column of the table shows means and standard deviations of the variables of interest. The next four columns present the coefficients and standard errors from regressions of individual characteristics on the treatment of interest. For ease of interpretation I divide the interest rate variables by 20 before running regressions thus, a change from 0 to 1 can be interpreted as the effect of moving from no interest subsidy to a 20 percent subsidy (this convention is maintained for the rest of the paper). The final row of the table presents the p-value from a chi-squared test that the coefficient on the treatment is jointly equal to zero across all outcomes. The first two rows verify that follow-up is uncorrelated with treatment status this is generally the case, though individuals in couples who received a higher subsidy on the husband s account were 5.6 percentage points less likely to be interviewed in the wave 2 endline. 13 The bulk of my analysis relies on the wave 1 endline I therefore limit my attention to the 1,417 individuals with endline survey data for the remainder of this table. Overall, participants have relatively low levels of human capital, income, and financial access. While three quarters of individuals are literate, education levels are low, with an average of 6.79 years. The most common occupations are subsistence farming and small-scale entrepreneurship, each accounting for 42 percent of respondents. 14 Individuals reported an average income of Ksh 4,595 ($57) per month, but the median income is much lower, at Ksh 2,167 ($27). 15 Just 22 percent of respondents reported owning a bank account at baseline. However, nearly all individuals reported saving in some way, with the most popular methods being storing cash at home (87 percent of respondents) and saving with a rotating savings and credit association, or ROSCA (58 percent of respondents). Individuals report saving Ksh 3,818 in the bank, in savings and credit cooperatives (SACCOs), and at home. The next three columns of Table 1 show that the randomization functioned well, with 13 Importantly, there is no evidence of selective attrition see Appendix Table A8. 14 I define entrepreneurship to include individuals operating an independent business. Common examples include market vendors, bicycle taxi drivers, shop owners, and commercial farmers. 15 Income, savings, and debt measures in my sample are all highly skewed I therefore topcode all variables of these types at the 99th percentile. Appendix Table A7 confirms that my main results are robust to using raw values, as well as alternative levels of topcoding and trimming. 9

10 none of the joint tests rejecting the null of no relationship at conventional levels. Some individual characteristics are correlated with the treatments, however participants who received higher joint interest rates are slightly older, more likely to be polygamous, have more children, and are more likely to participate in a ROSCA. Individuals in couples with higher male subsidies are somewhat better educated, less likely to be subsistence farmers, more likely to have a bank account at baseline, and accordingly have more cash savings. Individuals in couples that received higher female subsidies are less likely to save in SACCOs, more likely to save at home, and have less cash savings. Imbalance on savings-related variables is potentially important, since savings outcomes are of primary interest in my analysis. To address concerns that my results are driven by baseline differences, I therefore control for baseline use of the four savings devices listed in Table 1 (bank accounts, SACCOs, ROSCAs, and home). Importantly, adding these controls eliminates imbalance for baseline cash savings and my results are essentially unchanged when either omitting these controls or including all controls in Table 1. In results not shown, I also verify that the treatments are not correlated with one another. As expected, the treatments are generally independent of one another, with the notable exception that cash prize recipients received slightly higher joint interest rates (the average joint rate is 13.2 percent for cash prize recipients and 12.0 percent for non-recipients). I therefore control for cash prize receipt throughout the analysis as well. 3 Main Results 3.1 Impacts on Bank Account Use Short-Run Impacts Table 2 and Figure 1 show how the interest rate subsidies impacted short-term (6-month) use of the experimental bank accounts. The table and figures show results from regressions of the following form: y c = β 0 + β 1 intih c + β 2 intiw c + β 3 intj c + x cδ + ε c (1) where y c is a measure of account use for couple c, intih c is the husband s individual interest rate (divided by 20 so that it runs from zero to one), intiw c is the wife s individual interest rate, intj c is the joint rate, and x c is a vector of additional controls. 16 Note that if a couple did not open a given account, their use measure for that account is coded to zero thus, unopened accounts are treated the same as opened but unused accounts. 16 These controls include dummy variables for husband and wife cash prize selection, as well as dummy variables for husband and wife baseline savings device use. The regressions present results for all couples for whom at least one spouse was interviewed at endline. 10

11 The first three columns of Table 2 show the impact of the three interest subsidies on use of individual accounts for men. Account opening and short-run use was very sensitive to the husband s subsidy level just 31 percent of couples opened a man s account at no interest, while nearly half of participants opened this account at 20 percent interest. Receiving the most attractive interest offer increased the probability of making at least one transaction in the husband s account within the six month subsidy period by 9 percentage points (from a very low base of 6 percent), increased total deposits by Ksh 699, and increased the average daily balance in men s accounts by Ksh 97. The wife s interest rate had similar impacts on use of women s individual accounts, and the joint interest rate significantly increased takeup and use of joint accounts. Overall, there is strong evidence that when a certain account was subsidized, couples were more likely to use it. Table 2 also presents evidence of subsidy-induced crowd out. For example, higher subsidies for men reduced total deposits and average balances in women s accounts and reduced rates of opening and use of joint accounts. Similarly, couples who received higher joint subsidies were less likely to open individual accounts. Figure 1 depicts crowd out and the net impact of each interest rate treatment graphically. Panel A.i shows the impact of the male interest subsidy on total deposits in men s accounts (positive and significant), women s accounts (negative and significant), and joint accounts (negative and insignificant). The final bar in panel A.i shows the impact of the male interest subsidy on total deposits held in all three experimental accounts (i.e. the sum of the previous three bars). Panels A.ii and A.iii repeat this exercise for the female and joint interest subsidies. Panel B presents analogous results for average daily balances. The figure makes it clear that crowd out is quite important none of the interest subsidies significantly increased overall deposits or balances in the experimental bank accounts. Rather, the first order effect of the interest subsidies was to change where couples saved and by extension, who within the couple did the saving. Long-Run Impacts Figure 2 plots how the experimental accounts were used over the entire three year study period. Here, an account is coded as active if it received any deposit or withdrawal in the specified three month interval following the initial experiment (all unopened accounts are included in the means and coded as inactive). Panel A shows that men s accounts with higher interest rates were more likely to be active over both the short run and the long run, while long-run differences by interest rate are less apparent for women s and joint accounts. Also note that rates of account use decline rapidly after the first three months, and plateau at low levels around months The first three columns of Table 3 formally test whether the interest subsidies increased long-run use of experimental accounts. Here I run couple-level regressions in line with equation 1, where the outcome 11

12 is a dummy variable indicating whether each type of experimental account was active in the third year following account opening. Here we see that the male subsidy significantly increased long-run use of men s experimental accounts, while there is marginally significant evidence that the joint subsidy did the same for joint accounts. There is also some evidence that higher joint subsidies reduced long-run use of individual accounts. The vast majority of study participants abandoned the experimental accounts in the long-run just four percent of men s accounts, three percent of women s accounts, and five percent of joint accounts were used in their third year. When interpreting these numbers it is important to keep in mind that the Kenyan financial services landscape changed substantially between 2009 and 2012 as a result, study participants may have adopted other formal banking products in the long run. The last six columns of Table 3 use endline data to study ownership and use of all types of formal bank accounts in Here I use men s reports for ownership of men s accounts, women s reports for ownership of women s accounts, and both spouses reports for ownership of joint accounts. To do this, I switch to an individual-level regression of the following form: y ic = β 0 + β 1 intih c + β 2 intiw c + β 3 intj c + x icδ + ε ic (2) where standard errors are clustered at the couple level whenever both men s and women s reports are used. Columns 3-6 examine long-run impacts of the three interest subsidies on ownership of men s, women s, and joint accounts. Again we see that the husband s subsidy significantly increased ownership of men s accounts, while the joint subsidy increased ownership of joint accounts. In contrast, the wife s subsidy had no impact on ownership of women s accounts. In fact, just 44 percent of women who opened an individual account at baseline reported ownership of an individual account at endline, which suggests that bank accounts did not work well for many women in the long term. By way of comparison, 62 percent of men who opened an individual account at baseline still had one at endline, and 58 percent of joint account openers still had a joint account at endline. This ownership gap for women is especially striking given that men and women used individual accounts at very similar rates in the short run (see Table 2). The last three columns of Table 3 study the impact of the interest rate subsidies on total bank savings. Men who received the highest individual interest subsidy report Ksh 1,600 ($20) more in individual bank savings, which represents a near doubling of the amount for men in the lowest subsidy group. In contrast, there is limited evidence that the wife s or joint subsidy impacted long-run bank balances. Also note that although reported ownership of joint accounts is relatively high, most people do not use these accounts to store much value 12

13 the average joint account has just Ksh 1,097 in balances, while the average individual account has Ksh 5,128 in balances. Overall, Table 3 suggests that both the male and joint subsidies had persistent impacts on how couples save, but that changes induced by the male subsidy were larger in magnitude. The next subsection asks whether these changes had implications for broader outcomes. 3.2 Long-Run Impacts on Economic and Household Outcomes Figure 3 and Table 4 use data from the wave 1 endline to study the long-run impact of the interest subsidies on total assets, net assets (total assets minus debt), and total income. Panel A of Figure 3 begins with impacts on assets (this includes cash savings, the value of livestock, and the value of business capital). The first graph (A.i) illustrates the impact of the husband s interest subsidy on both the intrahousehold allocation of resources and overall resources. The first bar shows that men with the highest interest subsidy reported nearly Ksh 12,000 more in assets at endline. 17 The second bar shows that part of this represents a reallocation of resources from women to men, as women in couples assigned the highest male interest subsidy report Ksh 6,000 less in total assets. The third bar sums these two effects together to obtain the overall impact on household assets, which is large and positive but not significantly different from zero. 18 Panel A.ii shows that the female interest subsidy had no impact on assets, which is not surprising given its lack of effect on banking outcomes. Finally, Panel A.iii shows that the joint subsidy significantly increased female-owned assets, but not male-owned assets. In results not shown here I find that most of the impact on female-owned assets is driven by women reporting more jointly held resources. Since there is no similar effect on jointly held resources reported by men, this suggests that use of the joint account may have increased women s perceived ownership of household assets, while leaving overall asset levels unchanged. Panel B shows that the impacts in Panel A are not driven by households taking on more debt this is confirmed by Appendix Table A4, which reports treatment effects on debt by type. Finally, Panel C graphs treatment effects on total monthly income, which includes farm, business, wage, and other earnings. Panel C.i shows that the husband s subsidy significantly increased male income while having no impact on female earnings 17 In cases of jointly-owned assets, such as joint bank balances or livestock, I assign half of the reported value to the husband and half of the reported value to the wife. I follow this convention for the rest of the paper. 18 In order to obtain standard errors on these sums I jointly estimate the male and female regressions (which follow equation 2), clustering standard errors at the couple level. I prefer this approach to adding male and female reports together and running a couple level regression, since the field team was only able to interview both spouses in 85 percent of couples and additional missing data is generated in cases where responses are missing for one of two spouses. 13

14 as a result, overall household income increases by Ksh 2,079 ($26), which represents an 18 percent increase relative to couples who received the lowest husband s subsidy. In contrast, neither the wife s nor the joint interest subsidy had a meaningful impact on household income. Table 4 reports the same estimates graphed in Figure 3 and also presents p-values on F-tests of whether the interest subsidies had equivalent impacts on outcomes. I am able to strongly reject that the subsidies had equivalent impacts on the intrahousehold allocation of assets, but I am not able to formally reject that they had equivalent impacts on couple-level assets or income. Note that the overall impact of the husband s interest subsidy on men s assets is much larger than the impact on men s bank account savings. What, then, is the source of the remaining growth in male-owned assets (and income)? To answer this question I disaggregate impacts by income and asset classes. Table 5 shows that the husband s subsidy had large and important effects on business outcomes, while Appendix Tables A2 and A3 show that estimated impacts on all other types of assets and income are much smaller in magnitude and not significantly different from zero. Table 5 shows that receiving the highest husband s subsidy increased men s probability of having any business capital or profits by 14 percentage points (a 64 percent increase) and increased the probability that a man reported entrepreneurship as his main occupation by 12 percentage points (a 36 percent increase). Moreover, the male subsidy increased husbands (household) business capital by Ksh 7,088 (Ksh 6,193) and husbands (household) business profits by Ksh 1,190 (Ksh 1,743) per month. There is also some evidence of spillover effects on women: women in couples who received the highest husband s subsidy are slightly more likely to report entrepreneurship as their main occupation and consequently report more business income. In contrast, estimated impacts for the wife s and the joint subsidy are generally close to zero and insignificant. 19 The data also reveal a great deal of business churning just 50 percent of individuals who reported entrepreneurship as their main occupation at baseline reported the same at endline, while 23 percent of non-entrepreneurs transitioned into entrepreneurship. Appendix Table A5 explores heterogeneity in main effects by baseline occupation. Here we see that the impact of the husband s subsidy is concentrated in the subsample of men who were entrepreneurs at baseline. Thus, the husband s interest subsidy helped existing entrepreneurs stay in business, but did not facilitate much business creation. The results also suggest that the husband s subsidy helped existing businesses grow in order for the extensive margin to entirely account for the treatment effects on capital and profits, the average business kept 19 Again, the impact of the joint subsidy on women s reports of business assets is driven by joint assets. This impact is not found in men s reports of joint assets, consistent with the perceived ownership hypothesis. 14

15 alive by the higher subsidy treatment would need to have been around the 85th percentile of the business profits and assets distributions. These results mirror recent findings from a number of capital drop experiments and microfinance evaluations, which suggest that only individuals with some pre-existing entrepreneurial skill are able to benefit from interventions to grow self-employment income (Crépon et al. 2011; Angelucci et al Banerjee et al. 2013; Field et al. 2013; Fafchamps et al. 2014). Although I find no significant evidence that the joint subsidy impacted overall income and assets, I do find that the joint subsidies impacted outcomes beyond joint bank account use: Table 6 uses wave 1 endline data to explore how the interest subsidies affected a number of measures meant to capture investment in household public goods and levels of spousal alignment. Since all the outcomes in this table capture joint/public investment or shared spousal outcomes, I pool men and women in all specifications. The first column of the table focuses on livestock holdings. Livestock are arguably the most important class of assets held by study households nearly all households (95 percent) own some sort of livestock, and their value accounts for half of total assets. Livestock are also inherently joint investments, as they are easily observed and accessed by both members of a couple. Although the joint interest subsidy did not significantly impact average livestock holdings (see Appendix Table A2), a specification using the inverse hyperbolic sine transformation in Table 6 suggests that the subsidy did substantially increase livestock holdings at the lower end of the distribution. 20 It is common practice in the study area for households to make periodic, incremental investments in their homes. Home improvement expenditures are also inherently joint, as all members of the household benefit from them. Column 2 of Table 6 shows that nearly half of individuals reported making some investment in home renovation in the past year, and moving from the lowest to the highest joint subsidy increased this share by 6 percentage points (recall that the joint interest rate variable runs from 0.2-1). There is also evidence of increased home investment based on actual home quality. Column 3 shows that individuals who received higher joint subsidies were more likely to live in a home with a permanent (i.e. iron sheet as opposed to thatch) roof at endline. 21 The next two columns of Table 6 test whether the joint subsidy increased spousal alignment over decision making: at endline, all individuals were asked to rate, on a scale of 0 to 10, how much they and their spouse agreed about different consumption and savings ( 20 The inverse hyperbolic sine of x is given by ln x + ( x ) ).5 ln (2x). Results are very similar if I use ln(x + 1) instead. The impact of the joint interest rate on livestock holdings is confirmed by quantile regressions on level values, which reveal significant, positive impacts on the 28th to the 70th quantiles. 21 I do not, however, find any significant changes in the quality of the home s walls or floor. This is not entirely surprising as upgrading the walls/floors from mud to brick/cement generally requires a bigger investment than replacing a thatch roof with iron sheets. 15

16 decisions. 22 Columns 5 and 6 show that individuals who received the highest joint subsidy report greater levels of agreement about both consumption and savings decisions. There is also some evidence that the husband s subsidy increased alignment over spending decisions this could reflect that fact that households treated with this subsidy have more income, which could reduce conflict over how to spend money. Note that although the joint subsidy increased spousal agreement, I do not observe any significant impacts on self-reported joint decision making regarding savings (column 6). Finally, column 7 of Table 6 studies impacts on self-reported net spousal financial transfers (transfers received from a spouse minus transfers given to a spouse). Given how the variable is constructed, net transfers should be zero when averaging over men and women in the same set of couples. Instead, couples who received the lowest joint subsidy report a negative net transfer (on average, individuals report giving more than they receive). However, moving from the lowest to the highest joint subsidy eliminates this bias one way of interpreting this result is that higher joint rates increase intra-couple alignment and recognition of spousal contributions to the household. Taken together, the results in Table 6 suggest that the joint interest subsidy moderately increased investment in public goods such as livestock and the home, and slightly increased spousal alignment. 3.3 Robustness Since the estimated impacts of the husband s subsidy on economic outcomes are so large, it is important to ask (a) whether the results could be driven by spurious factors, such as reporting bias, and (b) whether the effect sizes are reasonable, given short-run impacts on savings behavior. Appendices A and B address these questions in detail I provide a short summary of my findings here. Appendix A performs a number of specification checks to evaluate the robustness of my main results. I find no evidence that the results are driven by reporting bias (Appendix Table A6) and the main results are robust to various top-coding and trimming schemes. Appendix Table A7 shows that the results for male business outcomes are robust to top-coding up to the 90th percentile and trimming up to the 98th percentile. I also note that the magnitudes of my point estimates do decline with top-coding and trimming observations in the right tail clearly have an important influence on point estimates. Another concern is missing data I lack follow-up data due to both attrition (an individual could not be contacted for an interview at endline) and missing values (an individual gave a don t know or refuse response at endline). This second source is very important for 22 Individuals who were no longer married were asked refer back to when they were married. 16

17 variables like total assets and income, which aggregate multiple survey responses. Appendix Table A8 presents p-values from F-tests of whether the treatments of interest are correlated with missing observations and F-tests of whether there is evidence of differential selection on observables with respect to missing data and the treatments. I find no evidence of problems. Appendix Table A9 evaluates the robustness of my main results to a variety of methods of imputing missing data, including methodologies that impute systematically lower values to observations in higher individual subsidy groups. My results for business outcomes are very robust to alternative imputation schemes, while estimated impacts for overall income and assets are somewhat attenuated and become close in magnitude to the estimates for business income and assets. Given this, I prefer to interpret estimates of the male subsidy s impact on business income and assets as indicative of impacts on overall income and assets. Appendix B discusses whether the men s estimates are reasonable in magnitude. Appendix Table A10 shows that my results for the men s subsidy imply a return to capital between percent per month. These numbers, while large, are consistent with existing estimates in the literature, which range from 4 to 33 percent per month (Blattman et al. 2014; McKenzie and Woodruff 2008). Moreover, given these returns my point estimates on income and assets can be generated by very modest short-run investments. For example, suppose that at the end of the subsidy period, an individual invested a lump sum of Ksh 80 ($1) at a 16 percent monthly rate of return and re-invested the proceeds each month. Then after 32 months (i.e. the time of the wave 1 endline survey) monthly income would be Ksh 1,478 higher, while assets would be Ksh 9,241 higher. Both these numbers exceed the point estimates on men s business income and assets. Table 2 shows that at the 20 percent subsidy level just 50 percent of men chose to open an individual account, and only 15 percent of men (or 30 percent of account openers) saved in the individual account. If one assumes that only men who opened an individual account could benefit from the husband s interest subsidy, then the absolute maximum fraction of compliers in the population would be 50 percent, while a more reasonable maximum fraction of compliers would be 15 percent. 23 Even if only 15 percent of men complied, the short-run investment among compliers needed to generate my treatment effects would be Ksh 533 ($6.67), which is still quite modest. Another way to cross-check my results is to ask whether short-run compliers (those who saved more in the experimental accounts during the subsidy period) are also driving 23 It is possible that some participants who opened an individual account but did not save still benefitted if, for example, individuals accumulated a stock of money at home earmarked for the bank account but never got around to actually making a deposit. Given the experimental context, this is plausible: most individuals had to commute into town to use the bank, so making small frequent deposits was less practical than making a smaller number of large deposits. 17