Microfinance for Startups: Experimental Evidence from Pakistan *

Transcription

1 Microfinance for Startups: Experimental Evidence from Pakistan * Farah Said, Mahreen Mahmud and Azam Chaudhry October 8, 2017 Abstract We use data over two years from a field experiment with 630 aspiring female entrepreneurs in Punjab, Pakistan to evaluate the impact of a start up loan for enterprises run by women. We find that the treatment increases the likelihood of setting up enterprise by 12% but only in the short run. Treated women whose husband s have their own business are significantly less likely to set up a business. We find small but significant effects on household assets and measures of financial access but no transformative effects on general household welfare or female agency. We also find weak negative effects on household assets and food consumption for treated households in the left tail of baseline distributions. *This project was funded by the International Growth Centre, Pakistan and received IRB approval from the Lahore School of Economics (RERC ). The project would not have been possible without the support of Zara Murtaza and Kamran Azim at Kashf Foundation, Pakistan, and Naved Hamid at the Centre for Research in Economics and Business at the Lahore School of Economics. Lahore School of Economics (farahs@lahoreschool.edu.pk) Centre for the Study of African Economies, Blavatnik School of Government, University of Oxford (mahreen.mahmud@economics.ox.ac.uk) Lahore School of Economics (azam@lahoreschool.edu.pk) 1

2 1 Introduction Over the last few decades, microfinance has emerged as one of the important tools used by policy makers to tackle poverty. Financial inclusion is thought to improve the lives of recipients by providing them with credit that allows them to invest in income generating activities, smooth consumption and mitigate financial risks. The general perception in the sector is that providing a portfolio of products to the previously unbanked population, especially women, would increase their economic empowerment and reduce inequality within and outside the household (Kabeer, 2001). For countries like Pakistan where female labor force participation rate is disproportionately low, and conservative norms and notions of purdah restricts female mobility, microfinance provides self-employment opportunities closer to home and those that do not interfere with her responsibilities at home (Ginè and Mansuri, 2011). This may explain why more women in Pakistan engage in home-based production rather than wage employment in the public space to earn an income. 1 Several large randomised impact evaluations on microfinance have indeed found a significant increase in investment in small business (Duflo et al., 2013; Angelucci et al., 2015; Attanasio et al., 2015; Crepon et al., 2015; Tarozzi et al., 2014; Augsburg et al., 2015; Banerjee et al., 2015). In general, they find that while access to finance led to increased profits of existing businesses, it did not translate into females starting their own enterprises. Further, there was limited evidence of microcredit significantly changing household or female welfare. However, while microfinance has often been promoted as a tool to promote female enterprise and economic equity, we have little evidence on the effectiveness of interventions directly aimed at enabling women to set up enterprise. Existing literature has looked at the impact of broad microcredit loans on female entrepreneurship; but not loans that target only women and are specifically for business creation. Studies have largely not differentiated between loans for new or existing enterprises making it difficult to disentangle the impact of the loan from the effects of experience and survivorship bias or having a business that has withstood the test of time. To investigate the role that access to finance can play in providing women with an income generating enterprise, this study looks at a field experiment conducted in Pakistan where female borrowers were provided loans to set up a new enterprise. Each applicant had to submit a business plan with their loan application. The business plan was evaluated for viability by loan officers at the local branch. Shortlisted applicants were then randomly assigned to receive the loan. Between May and August 2014, we worked with Kashf Foundation, one of the largest microfinance provider for women in Pakistan, to randomise which approved applicant receives a loan in three districts in Punjab, Pakistan where the product 1 At 22%, the female participation rate is one-third that of the males. The disparity between male and female participation is even greater in paid employment (13% for women vs. 43% for men) and formal microenterprises (19% for women vs. 41% for men). In the informal sector, the gender ratio is more equitable (albeit low) at 38% for women and 42% for men. All figures are from the Labor Force Survey Annual Report, prepared by the Pakistan Bureau of Statistics. 2 Said, Mahmud & Chaudhry

3 was going to be introduced - Bahawalpur, Gujrat and Sialkot. We use a balanced panel of 630 study participants, with data collected in three rounds of surveys over , to measure the impact of the loan on business creation and a host of household and individual level outcomes. Treatment assignment led to a significant increase in the likelihood of setting up a new business - loan recipients are 12% more likely to set up a business within one year. This effect is much larger than the modest effects found in other studies with loans that did not target start ups (Banerjee et al., 2015). In fact, the result is even more stark when one considers that most business outcomes in other studies were driven by enterprises owned by men (Ginè and Mansuri, 2011; de Mel et al., 2012; Fafchamps et al., 2009; Fiala, 2015). We find that the effect on business creation is not sustained and disappears over two years. Loan recipients are almost just as likely to shut down an existing business during the year as they are to open one suggesting that perhaps the loan in itself is not sufficient for the business or, as suggested by Banerjee et al. (2014), not all entrepreneurs are able to bear the opportunity cost of their labor. We also see a large negative effect of existing businesses in the household on the likelihood of female setting up a business. This effect is driven by women whose husbands have an existing business. This is a striking result - while the steady stream of income and experience from an existing business may be viewed as a positive for a lender providing enterprise finance, treated women are 17% less likely to set up a business under such circumstances. While we cannot say for certain if the applicant intended to use the loan for setting up her own business or if the intention was always to borrow for the husband, an appropriation of funds in such a manner can explain why the loan was found to have no impact on female agency, even among those who did manage to set up a business. On revisiting evidence from studies that found no impact of capital on profits earned by female entrepreneurs, Bernhardt et al. (2017) find, like we do, that this is attributable to the presence of another enterprise in the household run by the husband where the money was invested instead. Our results show that if the intention is to enable women to set up their own enterprise, then lending to women whose husbands have an existing business will likely be insufficient, especially if there are no explicit penalties on loan misuse. We find no effects on household expenditure but a mildly significant and positive effect on household assets. Access to finance improved, with treated clients more likely to take out other loans at the same time. On the whole, the loan product was not transformative on average - individual welfare indicators remained unchanged and household indicators were largely unaffected. We use quantile regressions to explore heterogeneous treatment effects and find that the treatment may have improved female agency for some women in the short run but reduced the asset base and food expenditures for others over longer time periods. There are two important caveats to these results. Our sample size is small relative to many other impact evaluations and we may be underpowered to detect otherwise significant effects. The economic magnitude of most effects for non-business outcomes are small despite the level of statistical significance. Second, we cannot comment on the external validity of 3 Said, Mahmud & Chaudhry

4 the results. The lender uses a similar appraisal strategy to other microfinance providers in the country and our sample is similar to a typical microfinance sample in Pakistan. 2 However, existing evidence on enterprise loans for new businesses by women is rare and a different context, market characteristics or socio-cultural norms towards female enterprise may yield different results. 2 Background on the Lender and Study Setting 2.1 Kashf Foundation The lending organisation, Kashf Foundation, is a specialized non-profit microfinance organization in Pakistan. Established in 1996, the organization offers microfinance services to women from low-income households, in an attempt to enhance the economic and social status of women in their households and the community at large. Kashf broadly attempts to accomplish these aims through lending directly to women, providing financial training and mentoring. At the time of the baseline, Kashf had nearly 250,000 active borrowers (12.5% of the total active female borrowers in the country), providing an average PKR 10,000 (~$ 100) entry loan at standard of 22% service charge. The flagship enterprise loan, called the Kashf Karobar Karza (KKK), involved loans of PKR 30, ,000 (~$ 300-1,000), backed by a promissory note and a cash flow analysis. Kashf Foundation does not require collateral for the loan. In 2014, Kashf Foundation claimed 97% of its enterprise loans had been used for productive business investments but it was likely that only one in five female recipients used the loan for an enterprise that was owned or operated by women Product Terms and Screening This study specifically focuses on evaluating the impact of a new microenterprise loan offered by Kashf Foundation between The Kashf Ibtada-e-Karobar Karza (KIKK) was a start up loan provided to women desirous of setting up a new business. The loan was smaller in size than the pre-existing KKK, ranging from PKR 10,000 - PKR 2 See for instance, Banerjee et al. (2015), Afzal et al. (2017) and Ginè and Mansuri (2011) for impact evaluations of microcredit and microsavings in Pakistan. Weber and Ahmad (2014), Salman (2008), Setboonsarng and Parpiev (2008) and Ghalib et al. (2011) evaluate microfinance programs using quasi-experimental techniques. 3 According to Kashf Foundation Annual Report , Annual_Report_ pdf and Pakistan Microfinance Review 2015, publications/category/pmr, accessed 30 January Said, Mahmud & Chaudhry

5 40,000 (~$ ). 4 The loan was to be repaid over a year, with repayments starting from the month after disbursement. The loan required applicants to submit a business plan, along with details of household income and expenditure. Applicants who were deemed to have a viable business plan, in addition to sufficient household income to repay the loan, were then randomly selected by the research team to receive the loan. Female beneficiaries of the loan also attended a three hour session on the importance of marketing, networking and capacity building for a new business. The workshop included discussions on setting goals and deadlines for their business; and keeping business and household expenditures separate. Each session was conducted at disbursement for a small group of 4 to 5 successful applicants by loan officers at the local Kashf office Study Setting The KIKK rolled out in January 2014 in peri-urban areas of four districts of Punjab, Pakistan. We conducted a randomized roll out and evaluation of this intervention in all 13 branches areas served by Kashf in three of the four districts - Bahawalpur, Gujrat and Sialkot. 6 The study districts provide us with sample representation in south and central Punjab. Bahawalpur, located in the south, lags behind the others in terms of educational performance. It is ranked 31 st out of 36 districts in Punjab (Memon et al., 2014) in terms of educational attainment. Gujrat and Sialkot fare better, ranked at 19 th and 13 th, respectively (Memon et al., 2014). The average monthly household income in Gujrat, Bahawalpur and Sialkot are PKR 51,854 ($520), PKR 30,294 ($300) and PKR 29,110 ($290), respectively. 7 At the time of the baseline, these districts were amongst the highest served districts in the sector, in terms of both MFI penetration and number of active borrowers. 8 4 The average and median loan size was PKR 30,000. All analysis in section 4 is robust to the inclusion of loan amount. 5 Since everyone in the treated sample attended the session, we cannot differentiate between the effect of the loan and effect of attending the session but we do not expect the short session to have had a significant independent effect. In fact one year after having received the loan and training, we asked respondents if they remember having attended such sessions at disbursement. Only 56% of the recipients reported in the affirmative. We do consider that remembering the session may mean something about individual retention, attention or understanding but find an insignificant correlation with business outcomes. 6 There were a total of 5 branches in Bahawalpur and 4 each in Gujrat and Sialkot. The fourth districts, Multan, was not selected because all branches piloting the product were located in the urban centre, as opposed to the peri-urban areas in the other districts. 7 Inflation adjusted estimates from Pakistan Social and Living Standards Measurement survey According to MicroWatch Issue 31, 2014 and MircoWatch Issue 37, 2015, publications/category/microwatch, Accessed 30 January Said, Mahmud & Chaudhry

6 Figure 1: Study timeline and respondents Random Assignment Baseline survey Midline survey Endline survey t = 0 t = 1 t = 2 May - Aug 2014 Aug - Sept 2015 Aug - Sept 2016 Number of respondents: (49%) (51%) (52%) Note: The figure displays months, duration and activities related conducted at t = 0, 1, 2. Proportion of sample in treatment group are reported in parenthesis 3 Research Design and Implementation 3.1 Study Design Our study uses an individual randomization design - every applicant approved by the microfinance partner was randomly selected by the research team to receive the KIKK loan. Every applicant in 13 branches of selected districts between May and July 2014 was vetted by the local branch staff. Applicants that were deemed to be eligible under the KIKK and Kashf criteria were then passed on to the research team to be randomly allocated to a treatment group that received the loan product (KIKK) and a control group that did not. Randomization was carried out separately for each branch. As a result, each branch had an approximately equal proportion of treatment and control participants at baseline. Figure 1 displays the study timeline. Study participants were surveyed thrice between May 2014 and September A baseline survey was filled in at the time of application, while Kashf was carrying out loan appraisals. A midline survey was conducted between July - September 2015 and an endline survey between July - September A total of 899 respondents were surveyed at baseline, out of which 440 were assigned to the treatment group. 630 original respondents could be located and successfully surveyed at endline of which 328 belonged to the treatment sample. Attrition is discussed in detail in section Experiment Implementation Kashf officers faced non-compliance from the treated individuals in 38 cases. 18 failed to complete paperwork required by Kashf operational policy and 10 refused the loan before disbursement. In these cases, the research team provided a random replacement from the control group. In 10 other instances, the applicant had refused the loan but the research 6 Said, Mahmud & Chaudhry

7 team was not informed until much later and a replacement was not provided. 11 individuals from the control group were provided the loan in violation of the research protocol. 9 We use a balanced panel of 630 individuals in our analysis. Table 1 presents the sample characteristics. Observable characteristics are strongly balanced across the control and treatment groups. The F-test of joint significance of treatment and baseline variables produces a p value of The average respondent is 37 years of age, married and can read or write. Most live in homes owned by one of the household member, with an average household expenditure of PKR 14,000 per month, which is well below the monthly averages for the study districts. Two in every five respondents had a current business at baseline or said they had a business in the past that has now shut down. On average, women feel they will not be allowed by family members to seek paid employment (outside home). In addition, indices for autonomy and female agency indicate the average respondent had low decision making power in the household. 10 Respondents reported low access to formal and informal finance at the start of the study. Table 1: Balance of randomization N Mean Median S. Dev Balance Test (1) (2) (3) (4) (5) Family 1: Demographics Age (years) Dummy: Respondent is currently married Dummy: Respondent can read and write Number of children (years < 17) in the household Household dependency ratio Family 2: Occupation and experience Dummy: Respondent has a business Dummy: Respondent has worked as a paid employee in the past Dummy: Respondent has had a business in the past Dummy: Household has existing business Average Intention to Treat results presented in section 4 are robust to the exclusion of these individuals. They are not significantly different from the randomly allocated sample on observables such as education, marital status, occupation, household expenditure or dependency ratio). 10 Variable construction and survey questions are described in the Online Appendix. 7 Said, Mahmud & Chaudhry

8 Family 3: Household assets and expenditure Household expenditure in an average month (PKR) Dummy: household home is owned by a household member Index: Assets owned by the household Family 4: Intra-household agency and autonomy Dummy: Respondent is confident she can support household for 4 weeks) Index: Respondent makes decisions in the household herself Inverse variance covariance index (Anderson, 2008) out of empowerment index and confidence variable Dummy: Respondent is not allowed by the household to seek employment Family 5: Access to formal or informal finance Dummy: Household has outstanding loans Dummy: Household member(s) have participated in ROSCAs Dummy: Household member(s) have a bank account Share of sample in treatment group 0.52 p value of F test of joint significance of explanatory 0.98 variables Note: Robust standard errors are show in column (4). Column (5) shows the result of the balance test. The cells show the p-values for statistical significance of the coefficient on the variable in the row when it is regressed on treatment assignment. The F test of joint significance is from a test of significance of all independent variables when all variables in rows are included in one regression with treatment assignment as the dependent variable. p < 0.01, p < 0.05, p < Said, Mahmud & Chaudhry

9 3.3 Attrition We use a balanced sample of 630 individuals for our analysis. We were unable to survey 210 of the initial 899 baseline respondents at midline, leading to an attrition rate of 30% from the original baseline sample. Almost two-third of the attrited sample belonged to the control sample. Local branch officers met with the treated clients once a month to collect loan instalments, but there was no such contact with control sample and tracking control clients one year later was difficult. Enumerators reported a high level of local migration. In fact, 80% of the attrited sample was reported to have migrated and could not be tracked despite assistance from staff at the local branch. In Appendix table A1, we use all 899 respondents at baseline to see if final attrition is related to observable characteristics. We find that attrition is not random - the probability of being surveyed at endline is positively correlated with household dependency ratio and negatively correlated with the respondent being married. As expected, the probability of being surveyed is also positively related to being in the treatment sample (column 1 and 2). However, probability of attrition is unrelated to treatment status once we control for observable characteristics and their interaction with treatment variables (column 3). In our analysis, we deal with attrition in two ways. First, our analysis in section 4 includes controls for all baseline characteristics that are systematically related to attrition. Second, we test the robustness of our results to differential treatment by constructing the upper and lower bounds of treatment effects using the Lee (2009) procedure. 3.4 Estimating Average Effects We use data collected over three rounds of surveys to study the effect of the loan product. Our primary variable of interest is business outcome, that is, if the treatment increases the likelihood of a business being set up. Our secondary variables of interest are household variables (expenditure, assets); female autonomy and decision making in the household; and access to finance. We measure impact over both one and two years of having received the treatment product or the short and long term, respectively. We have a limited sample due to budgetary constraints and because the loan product was still in a pilot phase with the implementing organization. Therefore, for our main regressions, we discuss the minimum detectable effect (MDE) size for each of our outcome variables. This is the ex post effect size given our sample size that is detectable at 5 percent significance level with 80 percent power (Duflo et al., 2008; Haushofer and Shapiro, 2016). Some variables may reflect the same channel of impact or proxy the same outcome. 11, We deal with the possible multiple inference problems in two ways: First, for each estimation, we report (i) the p-value for the estimated treatment effect, and (ii) a sharpened q-value, 11 In Table C1 in the Online Appendix, we describe how each variable has constructed and the family they belong to. 9 Said, Mahmud & Chaudhry

10 calculated within each listed outcome family (see (Benjamini et al., 2006)). Second, for each separate outcome family, we construct an index (following the method of (Anderson, 2008)) using the inverse of the covariance matrix. We then use this index as a separate summary outcome, and repeat our estimations using the summary index as the dependent variable. We estimate the average Intent to Treat (ITT) parameters of equation (1): y i1,2 = β 0 + β 1 Treatment i + β 2 y i0 + β 3 z i0 + φ s + ε i (1) Where y i1,2 is the midline (t = 1) and endline (t = 2) value for individual i for some outcome variable and y i0 is the baseline value. Then, for each outcome variable, we estimate an ANCOVA specification with Z i0 controls for baseline characteristics that predict attrition, φ s denoting the common parameter for branch stratum s and standard errors are clustered at the individual level. β 1 provides the average ITT effect on outcome y. We were able to revisit RCT participants for a total of three times over three years, allowing us to measure the impact of treatment over both one and two year periods. We look at impact on outcomes over both one and two year horizons, referring to them as short term and long term effects, respectively. 4 Results 4.1 Loan utilization One year later, 55% of the loan recipients claimed they had used loan funds to purchase business inventory. Another 5% claimed to have used it for purchasing a fixed asset for the business or for carrying out repairs of the building in which the business was located. 10% claimed to have used the money for buying a household appliance or for carrying out repairs of the house. Figure 2 provides a summary of the treatment loan utilization, as reported by respondents in the midline survey. For 40% of the borrowers, the loan was not used for what the lender intended (for the enterprise); however, it is not entirely surprising. The lender, much like other microfinance institutions in the country, does not impose an explicit penalty on misused funds. A borrower is only warned of misuse counting negatively in subsequent loan appraisals. In fact, the reported use of the treatment loan was not very different from other loans that the household may have. At baseline, expenditure related to businesses owned by household members dominated the use for other outstanding loans of the household; at midline, almost 80% of the respondents said these other loans were used for investment and expenditure related to the business; and at endline, 67% of the households with new (non-treatment) loans were 10 Said, Mahmud & Chaudhry

11 still borrowing predominantly to finance business expenditure. 12 The lender reported no defaults (non-repayment) in the study sample and all but two respondents reported they were able to repay the loan in time % of the recipients used the income from their business, 30% used wage income and further 25% used their own (15%) or their family members savings (10%) to repay the loan. 14 Figure 2: Treatment product - Reported expenditure items Note: x-axis shows the proportion (%) of treatment loan recipients who reported the item on the y-axis as the largest item the loan was used for. This question was asked only at midline, that is, one year after the disbursement of the loan. 4.2 Enterprise Creation At midline, we found 77 women had set up a business in the last one year but only 48 women had received the treatment product, implying finance alone may not the binding constraint 15 In addition, 52 women reported setting up a business in the last year that shut down before the first follow-up survey could be conducted. Finally, 149 women said they had thought about setting up a business but were ultimately unable to do so. A vast majority of this latter group of women reported insufficient finance (82%) and commitments at home 12 see Appendix Figure B2 for a summary of the most common purpose of loan as reported by respondents at each survey round. 13 Figure B1 summarizes the largest sources of income that were used to repay the treatment product % reported using the loan for other purposes. Unfortunately, we do not have data on what these purposes are. 15 Half of the women in the control group who set up an enterprise did so without taking out any loans. 11 Said, Mahmud & Chaudhry

12 (17%) kept them from setting up an enterprise. By endline, the number of businesses that had opened up since the start of the study had decreased to 40. However, these businesses had a larger asset base and reported slightly higher profits (approx. PKR or $ 75). Appendix Table A2 provides descriptive statistics of these new businesses. Most women chose to open a beauty parlour or a stitching and embroidery service at home. Appendix Figure B2 summarizes the type of new enterprise. The average business cost approximately PKR 20,000 (~$ 200) to set up, and earned an average monthly profit of PKR 6000 (~$ 60). Table 2 below tests whether the microenterprise loan led to a business being set up one and two years after the loan had been disbursed. Column (1) shows that treated women are 12% more likely to set up a new enterprise than the people in the control sample. This effect is relatively larger than that found in existing evidence on business creation. Other studies have looked at the impacts of finance on business outcomes but typically only for preexisting businesses. 16 Our findings suggest that the impact of finance on business creation can be significant when aspiring female entrepreneurs are provided with start-up loans. However, the effect of treatment was transitory at best, not lasting longer than a year (columns 3 and 4). The MDE are at least twice as large as our estimated effects. We also see that loan recipients are also 11% more likely to shut down a business over the last year (column 2) - an effect that is almost as large as the effect on business creation in column (1). While the impact on business creation is higher than the average affect found in literature, the short duration of micro-enterprises has been documented before. Indeed, given these results, we may have seen a larger impact of the loan had we surveyed more frequently than once a year. According to Banerjee et al. (2014), one possible reason for these results, is that the financial gains of a new enterprise are often offset by increased opportunity costs of the entrepreneur s labor. For instance, women may find increased demands on their time as they balance the time spent on her household chores with that available for her business. Others posit that finance alone is insufficient to sustain enterprise (Fafchamps et al., 2014) and that it must be complemented with skills, training (Blattman et al., 2015) and cooperation from household members (de Mel et al., 2009, 2012). There are other individual characteristics that we measured at the midline and endline surveys that provide us with further insights into who may have set up a business. We find that the likelihood of setting a business is positively correlated with basic numeracy, possessing a positive outlook about the future of the business and to a self reported eagerness to work in the short run. 17 However, though outlook and numeracy may help provide an 16 See for instance, Duflo et al. (2013), Angelucci et al. (2015), Attanasio et al. (2015), Crepon et al. (2015), Tarozzi et al. (2014) and Augsburg et al. (2015), amongst others. 17 Note, given that we do not have baseline values of these variables, we cannot measure the impact of these characteristics on the decision to set up an enterprise. However, without making this assumption about the balance of these covariates at baseline, we can still discuss how they correlate with the decision to set up an enterprise,controlling for baseline and other characteristics. These results are presented in Tables C2-C4 of the Online Appendix. 12 Said, Mahmud & Chaudhry

13 Table 2: Impact of treatment on business status Short term Long term Set up Shut down Set up Set up Shut down business last year business last year last year (1) (2) (3) (4) (5) Treatment A A (0.062)* (0.061)* (0.031) (0.020) (0.007) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Set up business is a binary variable equal to 1 if the respondent set up a business since baseline; Set up last year is a binary variable equal to 1 if the respondent set up a business in the last year; and Shut down last year is a binary variable equal to 1 if the respondent had a business that she shut down in the last year. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. immediate impetus to enterprise, they do not seem to matter in the long run. In the long run, a female is more likely to set up a business if she has seen someone else in her family run and operate a business in the past The role of pre-existing business in the household in business creation The effect on business creation is not as large as the loan utilization would suggest. Our surveys asked respondents if they had used the loan for business, but we did not ask them to specify which business(es) they had spent the funds on. For instance, it is also likely that the loan has indeed been used for business expenditure, as reported, but not spent on the expenditure related to a new business that the respondent sets up. Conversely, it is also possible that the experience of having a business in the household helps set up a new enterprise. Table 3 looks at the impact of the treatment product when the household has another business. On average, a respondent in the control sample is more likely to set up a new business if others in the household have a business. However, treated respondent who have been provided with finance to set up a new business are less likely to do so when the household has pre-existing businesses. Loan funds may have been captured by other members of the 13 Said, Mahmud & Chaudhry

14 household to finance pre-existing businesses than the new business that respondent wanted to set up. Indeed, 24% of the treatment sample who did not set up a business but reported wanting to said they were not able to do so because of insufficient funds. It is worthwhile to note that the vast majority of pre-existing businesses are owned by the respondent s husband % of the respondents at midline (and 71% at endline) report that that at least one other business in the household is owned by the husband. This can explain why so few of the treated women are able to set up an enterprise - their loan funds were used, perhaps as intended, for investment into a pre-existing business owned by another member of the family, most often the husband. This appropriation of funds can also explain why we see no impact of the loan product on female agency and empowerment measures (discussed in section 4.3.3). Indeed, we find the negative impact of existing businesses on the likelihood of the female respondent setting up an enterprise is driven by this subsample of women whose husbands own a business (see Appendix table A3). These women are borrowing for their husband s business and unlikely to set up a business of their own. On average, these business are old - at least 93% were at least one year old at the midline, pre-existing at the time of the baseline 19, with a median age of 7 years. The predominant businesses type (~40%) is Personal Service providers (tailoring, stitching, hairdresser, etc.), a type which is also the most common among new businesses set up by female respondents. Shared household experience may explain why pre-existing business encourages business creation in the control group. 4.3 Short term effects on other outcomes We next look at the average Intention to Treat impact of the treatment loan on outcomes one year after the disbursement of the loan. We discuss each outcome family separately Household Expenditure and Assets Table 4 reports the ITT effects of the treatment on household assets, home ownership and average monthly expenditure. It is difficult to say what the impact on expenditure can be ex-ante. Repayment of loans could increase the monthly expenditure or generation of 18 We do not have a clean measure of relationship to respondent at baseline but we do at the subsequent rounds of the survey. Though it is reasonable to assume that the ownership of old established businesses will not have changed dramatically between the baseline and midline, we report here only the measurement from the follow up surveys. 19 Enumerators misunderstood the question What year was this business started? at baseline, reporting the number of months of years since inception in many cases. Since this variable is noisy, we report here only midline and endline values. Note that with 93% of the businesses being at least a year old, most of these businesses existed at baseline and midline values provide a close approximation of baseline values. 14 Said, Mahmud & Chaudhry

15 Table 3: Impact of treatment and other businesses on business status Short term Long term Set up Shut down Set up Set up Shut down business last year business last year last year (1) (2) (3) (4) (5) Treatment (0.061)** (0.061)* (0.031) (0.021) (0.008) Household has existing business (0.055)*** (0.034) (0.042) (0.036) (0.033) Treatment*Household has existing business (0.069)** (0.053) (0.056) (0.045) (0.041) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Set up business is a binary variable equal to 1 if the respondent set up a business since baseline; Set up last year is a binary variable equal to 1 if the respondent set up a business in the last year; and Shut down last year is a binary variable equal to 1 if the respondent had a business that she shut down in the last year. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. income through business may lead to higher household expenditure. On the other hand, it is possible that an effect on the expenditures be transitory and not necessarily captured in annual surveys. In contrast, changes in assets and home ownership are likely to be durable and easier to capture. 20 Indeed, we see no effects of the treatment product on the average level of household expenditure (column 1) and home ownership (column 2) after a year. Table A4 in the appendix shows the one year impact of treatment on individual expenditure items. The only effect we do see after a year is in average monthly expenditure on recreation which increases by about PKR 68 per month - an increase of only 0.5% in total average monthly expenditure. This result is in contrast to the findings of other studies. Banerjee et al. (2015) find a de- 20 Our survey asked for the total income level of the household but many respondents chose not to answer, particularly at baseline. Using baseline values of income to look at the impact of treatment on income is likely to give biased results. Therefore, we concentrate here on household expenditure where we had only 31 instances of non-response at baseline. 15 Said, Mahmud & Chaudhry

16 crease in such expenditure and attribute it to greater self discipline imposed due to debt servicing. Another explanation given is that access to finance leads to an increase in female autonomy and women tend to impose greater discipline on the household s spending on temptation items (Angelucci et al., 2015). Conversely, if a female does not have sufficient say in the use loan funds are put to, then it is possible for these funds to be appropriated by other household members and used for recreation. The treatment results in a small but significant increase in household assets. With an MDE of 0.36, it is possible that we are unable to detect an effect due to the small size of the sample. Household asset index increases by 0.2σ for households in the treated sample. Our surveys only ask for the number of assets owned by the household and not the value of each assets owned and so we cannot say more about the nature of this change. For instance, we may see an increase in the asset index if treated households purchased many assets of small value. On the other hand, instances where the household replaced many assets of low value with a very valuable asset will be recorded as a decrease in the household index Credit and financial access Next, we look at the impact on access to formal or informal finance. The lender requires clients to provide postdated checks for the first three loan repayments. Although the check does not require the respondent to have a bank account, it does require someone from her household to have a bank account against which a check could be drawn. With only 2% of the households holding a bank account at baseline, the ex-ante expectation is an increase in the number of households in the treatment sample that have a bank account. We do not find this to be the case - once the baseline value of this variable is controlled for, the treatment sample is not more likely to have access to a bank account (Table 5, column 1). This may be because both control and treated sample opened up a bank account in anticipation of receiving a loan; however, with baseline surveys being conducted before or at disbursement we do not believe this to be the case. Documentation requirements across MFIs in the region are similar (Haq and Safavian, 2013) so it is possible that many in the sample opened up a bank accounts to fulfil requirements for other loans. For many new borrowers in a microfinance market, entry loans provide access to a line of credit that can be continuously drawn on. Most microfinance lenders allow repeat borrowers to borrow successfully larger amounts. Even when borrowing from a competing lender, a history of loans with no loans reduces the riskiness of the borrower for the lender. We expect a positive effect on borrowing in the long run. In the short run though, theory would suggest that the treatment loan will crowd out other new loans if the demand for finance has been met. Crowding out would suggest a decrease in the likelihood of taking out a loan during the year the treatment loan is outstanding. On the other hand, if the treatment loan does not fully relax credit constraints, it can increase demand for other funds to satisfy the demand for credit. We find that household members of treated clients are 13.5% more 16 Said, Mahmud & Chaudhry

17 Table 4: Short term impact: Households assets and expenditure Monthly household Home owner Asset index expenditure (PKR) (1) (2) (3) Treatment ( ) (0.075) (0.189)* Monthly household AAA expenditure t=0 (0.084)*** Home owner t= AAA (0.047)*** Asset index t= AAA (0.042)*** MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Monthly household expenditure is calculated by summing up the average monthly expenditure on different items, reported in PKR. Home owner is a binary variable equal to 1 if someone in the household owns the household home. Asset index is an index created from the number of assets owned by the household using Principal Component Analysis. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. likely to take out another loan the same year (Table 5, column 2). Note that very few respondents reported having to take out other loans to service existing debt (see Figure B1), leading credence to the explanation that in not fully relaxing household credit constraints, the treatment product crowds in other loans Female Agency and Autonomy Table 6 looks at the average ITT effect on measures of female autonomy and decision making power in the household. The loan can be expected to act on female autonomy in two ways. One, literature indicates that the very act of receiving funds in her name would empower a woman and give her a greater say in how that money is to be used (Kabeer, 2001). In addition, the woman can contribute the household income if uses the loan productively 17 Said, Mahmud & Chaudhry

18 Table 5: Short term impact: Access to finance Bank account Took loan(s) last year (1) (2) Treatment AAA (0.039) (0.034)*** Bank account t= (0.139) Took loan(s) last year t=0 (0.101) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Bank account is a binary variable that is 1 if someone in the household currently has a bank account. Took loan(s) last year is a binary variable equal to 1 if someone in the household took out a loan (other than the treatment loan) in the last year. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. and that can improve a female s role in household decision making (Grasmuck and Espinal, 2000). Two, an increase in the female controlled share of household resources can threaten male members of the family, leading to domestic violence and decrease female empowerment levels (Angelucci, 2008; Maldonado, 2005). It is also possible for the loan product to have no impact on female autonomy and decision making power. A year may not be a sufficiently long period of time for household norms and the level of autonomy afforded to a female in the household to change. Weber and Ahmad (2014) find that empowerment levels change slowly, increasing for women who are in higher loan cycles but not for first time borrowers. Finally, other studies argue that external or internal pressures on the female, coupled with greater fungibility of cash, may mean these loans are easily captured by household members (Fafchamps et al., 2009; Jakiela and Ozier, 2012). As a result, empowerment levels may not change due to the 18 Said, Mahmud & Chaudhry

19 loan. 21 The impact of loan under these circumstances is likely to be low. 22 Consistent with this last strand of literature, we find no impact of the treatment loan on various measures of female agency and autonomy within the household. 23 Other than the possible reasons provided in literature, consider also the fact that treated respondents in our sample who belong to households with pre-existing businesses are less likely to set up a business (Table 3). In fact, in regression not shown in table 6, we find that the presence of another business in the household significantly decreases the index value by 0.3σ. The lack of a treatment effect on empowerment is not surprising when we consider that their loan funds may have been appropriated for use in other household businesses. 4.4 Long term effects on other outcomes We also test for the long term (2 year) impact of the treatment loan on the outcomes discussed in section 4.3. Tables A5, A6 and A7 in the appendix show long term regression results. As with the short term effects before, we also report the MDE at a significance level of 5% and power of 80%. In general, we find no long term impacts on any outcome. For instance, households with a higher asset base are likely to increase their asset base further, those with higher expenditure at baseline are more likely to have increased expenditure over two years, but none of this increase is due to receiving the loan product in the first year. Similarly, those with prior loans are likely to take further loans, but the crowding in effect of the treatment product itself dissipates over the longer run. These results are not altogether surprising given the short term and small size of the treatment loan. In addition, households have recourse to other loans and microfinance providers, smoothing out their consumption over the long run. We also find that women who are more empowered at baseline are likely to become more empowered over longer periods of time. These results indicate that household and individual preferences are likely to be slow to change. However, MDE size for the empowerment index indicates we may be under-powered to detect a significant effect. Access to short duration loans, such as the treatment product, are unlikely to immediately bring about a 21 Empowerment levels may even decrease if the capture of funds was not expected by the women. However, it would be simplistic to assume many women would not be aware of household norms and dynamics for the appropriation of loans to be wholly unexpected. 22 The impact of access to finance on female empowerment levels has long been a topic of investigation in literature. See, for instance, Kabeer (2001); Nghiem et al. (2011); Zu Selhausen (2013); and Brauw et al. (2013) who discuss gender bias in welfare. Other interventions discussed in section 1, discuss a wide variety of outcomes, including empowerment of women, usually in the form of their role in household decision making and find little effect. Studies have shown that self help groups (Campbell, 2012), information on family planning and vocational training (Bandiera et al., 2014) may have greater success in improving empowerment than microfinance. 23 We allow for the fact that our estimation fails to detect an effect due to small sample size. Note, for instance, the MDE for empowerment index is not very different from the estimated effect. 19 Said, Mahmud & Chaudhry

20 Table 6: Short term impact: Agency and autonomy in decision making Confident Empowerment Agency Allowed to index index work (1) (2) (3) (4) Treatment (0.103) (0.462) (0.211) (0.073) Confidence t= (0.054) Empowerment index t=0 (0.037) Agency index t=0 (0.043) Allowed to work t=0 (0.105) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Confident is a binary variable equal to 1 if the respondent believes she can support her family on her own for 4 weeks. Empowerment index is an index created using Principal Component Analysis from variables that measure if the respondent can make household decisions (clothing, footwear, medical, recreation, social visits, joining credit groups, purchases for self, purchases for others, marriage, investment) on her own. Agency index is an inverse variance-covariance index (Anderson, 2008) created out of the Confident and Empowerment index variables. Allowed to work is a binary variable that is equal to 1 when the respondent feels her household members allow her to work or will allow her to seek work. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. change in agency and empowerment levels. 20 Said, Mahmud & Chaudhry

21 4.5 Quantile regression effects Next, we move from mean effects to testing how the treatment effects distribution of outcomes. Quantile regression results can inform us if the effects are spread consistently across the distribution or if most of the change occurs in the tails or the middle of the distribution. Figure 3 shows the short term, quantile treatment effects for continuous outcome variables: asset, empowerment and agency indices and monthly average household expenditure. Expenditure and assets seem to increase in the middle of the distribution though the increase is small and not statistically significant at any quantile. The agency index, which summarizes both the empowerment index and confidence variable, increases in the highest percentile as a result of the treatment product. This result implies that the treatment improved agency for at least some individuals in the short run, though we cannot infer how many individuals. Figure 3: Quantile Treatment Effects: Short term effects for outcome variables (a) Avg. monthly household expenditure (b) Asset index QTE Average Monthly Household Expenditure (PKR) QTE Asset Index Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (c) Empowerment index (d) Agency index Empowerment Index Agency Index QTE QTE Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Note: x-axis shows quantile in each graph, the Quantile Treatment Effects at each quantile. Vertical lines show the 95% confidence intervals. Although the treatment effect on overall expenditure is statistically insignificant, we also look at expenditure by items and reported household savings to see if the treatment had 21 Said, Mahmud & Chaudhry

22 a short term effect in certain percentiles. 24 Figure 4 shows that though expenditure are increasing, most individual QTE s are statistically insignificant. Medical expenditure is higher for treated sample at the twentieth percentile and lower at the 80th percentile. However, these results may have been due to chance given the generally insignificant treatment effect for the rest of the distribution. Figures 5 and 6 plot the long run quantile effects. The positive of treatment on the agency index is no longer statistically significant at any percentile in the long run. The effect on asset index, on the other hand, is now negative in the lower tail. This implies that the treatment made some households worse off in the long run. Food expenditure is also lower in the long run for those in the lower percentiles. Medical expenditure is higher at the 80th and the 90th, reversing the decrease in the short run, possibly offset by the decrease in expenditure on mobile phones for an overall insignificant effect on total expenditure. It is possible that households economised on medical expenditure while the loan was due, leading to a short term decrease in medical expenditure. Effects on other expenditure items and savings are generally insignificant. On the whole, the quantile results show us: (i) A positive but transient effect on agency and only for those with the best baseline outcomes. (ii) A positive, long run effect on medical expenditure for those in the right tail. This effect is much bigger than the initial decrease in medical expenditure for the same percentiles. (iii) On the other hand, there is a significant negative long term treatment effect on asset index in left tail. There is also a significant negative effect on food expenditure in the fifth and tenth percentile. This implies that the treatment product, possibly due to the burden of debt servicing, adversely effected households with the worst baseline outcomes. 4.6 Impact by outcome family As mentioned in section 3.4, we deal with multiple inference problems by (i) estimating sharpened q values using the False Discovery Rate (Benjamini et al., 2006) calculated within each outcome family and reporting statistical significance in the results presented in each regression table. In general, adjusting p values with this procedure did not change the significance of results discussed. (ii) For each outcome family we construct and index using the inverse of variance covariance matrix (Anderson, 2008). We then repeat our estimation using this family summary index as the outcome variable. For ease, we organised regression results for outcomes by family in section 4.3. We presented results for family 3: Household assets and expenditure in table 4, for family 4: Agency and autonomy in decision making in table 6, for family 5: Access to formal and 24 Two expenditure categories - expenditure on recreation and on gifts or loans to friends were modular and did not have a continuous distribution. Therefore, we could not estimate quantile regressions for these two categories. 22 Said, Mahmud & Chaudhry

23 Figure 4: Quantile Treatment Effects: Short term effects for average monthly expenditure by category (a) Food (b) Medical QTE Average monthly food expenditure QTE Average monthly medical expenditure Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (c) Schooling (d) Non food, non durables QTE Average monthly school expenditure QTE Average monthly expenditure on non food, non durable items Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (e) Mobile phone (f) Savings QTE Average monthly expenditure on mobile phone QTE Average monthly household savings Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Note: x-axis shows quantile in each graph, the Quantile Treatment Effects at each quantile. Vertical lines show the 95% confidence intervals. 23 Said, Mahmud & Chaudhry

24 Figure 5: Quantile Treatment Effects: Long term effects for outcome variables (a) Avg. monthly household expenditure (b) Asset index QTE Average Monthly Household Expenditure (PKR) Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 QTE Asset Index Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (c) Empowerment index (d) Agency index Empowerment Index Agency Index QTE QTE Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Note: x-axis shows quantile in each graph, the Quantile Treatment Effects at each quantile. Vertical lines show the 95% confidence intervals. 24 Said, Mahmud & Chaudhry

25 Figure 6: Quantile Treatment Effects: Long term effects for average monthly expenditure by category (a) Food (b) Medical QTE Average monthly food expenditure Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 QTE Average monthly medical expenditure Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (c) Schooling (d) Non food, non durables QTE Average monthly school expenditure QTE Average monthly expenditure on non food, non durable items Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 (e) Mobile phone (f) Savings Average monthly expenditure on mobile phone Average monthly savings QTE QTE Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Q5 Q10 Q20 Q30 Q40 Q50 Q60 Q70 Q80 Q90 Q95 Note: x-axis shows quantile in each graph, the Quantile Treatment Effects at each quantile. Vertical lines show the 95% confidence intervals. 25 Said, Mahmud & Chaudhry

26 informal finance in table 5 and finally, results for family 6: Business status in table 5. As expected for index outcome results, we find significant effect of treatment only for indices for the families with the strongest results - family 5, measuring the access to finance and family 6, which measures business status (Regression results for the short term are available in Table C5 in the Online Appendix) 4.7 Robustness to attrition We check for the robustness of our results to attrition in two ways. (i) We included as control, all variables that are significantly related to attrition in our estimations above. (ii) We use the bounding procedure by Lee (2009) to find out the lower and upper limits of treatment effects. The key identifying assumption for these bounds is to assume that the selection affects the sample in a single direction. In our case, we assume that individuals are more likely to refuse to answer or are less likely to be tracked for a follow-up survey if they are a part of the control group. This is a reasonable assumption given the context and reasons reported by the survey firm for non-response. At any rate, as explained by McKenzie (2008), it would be reasonable to assume an individual who drops out of the treatment sample would also drop out if she was a part of the control sample instead. Lee (2009) constructs bounds by trimming the distribution of outcomes by the differential level of attrition between treatment and control samples. In doing so, the techniques provides a range of average effects for a specific subsample, that is, the sample that never attrites, regardless of treatment status. The lower limit of treatment effect is constructed by trimming the upper tail of the distribution and the upper limit is estimated by trimming the lower tail of distribution. Our point estimates are closer to the upper limit, than to the lower limit. Indeed, the lower limit would be relevant if it is the respondents who are more likely to set up a business, who are home owners, to take a loan or to be more empowered be more likely to register as a survey non-response. This is not likely and so the upper limit is more relevant in our context. Thus it appears that the short term effects are robust to attrition. Results are available in tables C6 and C7 of the Online Appendix. 5 Conclusions Our results provide us with four key insights: first, borrowers in this sample used their loans mostly for investment in new or existing enterprise. This was true for the treatment product, but it was also largely true for other loans that their households had taken over the two years. This finding contradicts anecdotal evidence in this sector that says loan funds are 26 Said, Mahmud & Chaudhry

27 likely to be largely used for consumption purposes. Second, we find a relatively large effect on business creation. Third, almost all of the treatment impact comes from women whose husbands do not run their own business. We are explicitly able to measure the differential impact of women from households with pre-existing businesses and find that treated women are less likely to set up a business if their husband own a business; implying that either the woman borrowed on behalf of her husband or that the loan funds were captured by the husband for use in his own business. Fourth, the loan product did not have transformative effects. We found short term increase in assets and recreational spending. We also find the loan product crowds in other lending in the household, implying that the loan in itself was not sufficient to relax credit constraints. We find no significant average effects of the treatment product after a year. However, the quantile regression analysis shows some negative effects of debt servicing on assets and food expenditure in the left tail over two years. That is, some individuals with the worst baseline outcomes were made worse off by the treatment product over time. The results for individual and household level outcomes are not very different from existing evidence on microfinance. The effect on new businesses is larger than what is typically recorded in literature. In fact, most studies find comparable effects only for existing businesses and none for new businesses set up by women. We believe this was the product of marketing the loan as a start up loan for female run enterprise, as well as soliciting submission of business plans that reduced the likelihood of misuse of funds. The lender s implicit penalty on not lending in the future to borrowers who misuse funds may have also played a role in encouraging using funds as intended. The effect was largely transitory, significant only in the first year. As argued by Banerjee et al. (2014), we believe the short term nature of such businesses are because of opportunity costs of labor that we are unable to measure. These costs may also be largely unexpected by women who are setting up an enterprise for the first time and who later find out their household responsibilities impose a high cost on the time spent at her business. Our results come with several caveats: first, our sample had a 30% attrition rate that was significantly higher in the control sample. We show our results are robust to the inclusion of baseline characteristics that can predict attrition. We also show the robustness of our results by constructing bounds of treatment effects using Lee (2009). Second, the resultant balanced sample is small compared to other recent impact evaluations and we may be underpowered to detect certain effects. Third, though our sample is representative of a typical sample of female borrowers in Pakistan, we cannot say if the different contexts and social norms surrounding female enterprise will yield the same results. From a policy perspective, the existence of husband s business in a setting where women lack agency makes it highly unlikely for a woman to set up her own enterprise. Far from recommending that women not be lent funds if their husbands are entrepreneurs, our results imply that access to finance alone will be insufficient to promote the set up of female enterprise. Our findings complement those of Bernhardt et al. (2017) who find that that the reason for low average returns to capital earned by female entrepreneurs in India, Ghana 27 Said, Mahmud & Chaudhry

28 and Sri Lanka, are not due to a gap in aptitude but rather the existence of husband s enterprise where the capital is invested instead of her enterprise. These results, taken with results from recent studies documenting the effectiveness of peer support (Field et al., 2016), personal initiative training (Campos et al., 2017) and possibility of improving aspirations of female entrepreneurs (Lybbert and Wydick, 2016), imply that there is scope to encourage both set up and growth of female enterprise if additional measures are taken with provision of finance. 28 Said, Mahmud & Chaudhry

29 References Afzal, U., d Adda, G., Fafchamps, M., Quinn, S., and Said, F. (2017). Two sides of the same rupee? comparing demand for microcredit and microsaving in a framed field experiment in rural pakistan. Economic Journal. Anderson, M. L. (2008). Multiple Inference and Gender Differences in the Effects of Early ntervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects. Journal of the American statistical Association, 103(484). Angelucci, M. (2008). Love on the Rocks: Domestic violence and alchohol abuse in rural Mexico. The B.E. Journal of Economic Analysis and Policy, 8(1). Angelucci, M., Karlan, D., and Zinman, J. (2015). Microcredit impacts: Evidence from a randomized microcredit program placement experiment by Compartamos Banco. American Economic Journal: Applied Economics, 7(1): Attanasio, O., Augsburg, B., Haas, R. D., Fitzsimons, E., and Harmgart, H. (2015). The impacts of microfinance: Evidence from joint-liability lending in Mongolia. American Economic Journal: Applied Economics, 7(1): Augsburg, B., Haas, D., R., H., H., and Meghir, C. (2015). The impacts of microcredit: Evidence from Bosnia and Herzegovina. American Economic Journal: Applied Economics, 7(1): Bandiera, O., Buehren, N., Burgess, R., Goldstein, M., Gulesci, S., Rasul, I., and Suleiman, M. (2014). Women s empowerment in action: Evidence from a randomized control trial in Africa. CSAE Working Paper WPS / Banerjee, A., Duflo, E., and Hornbeck, R. (2014). (measured) profit is not welfare: Evidence from an experiment on bundling microcredit and insurance. Working Paper 20477, National Bureau of Economic Research. Banerjee, A., Karlan, D., and Zinman, J. (2015). Six randomized evaluations of microcredit: Introduction and further steps. American Economic Journal: Applied Economics, 7(1):1 21. Benjamini, Y., Krieger, A. M., and Yekutieli, D. (2006). Adaptive Linear Step-up Procedures that Control the False Discovery Rate. Biometrika, 93(3): Bernhardt, A., Field, E., Pande, R., and Rigol, N. (2017). Household matters: Revisiting the returns to capital among female micro-entrepreneurs. Technical report, National Bureau of Economic Research. Blattman, C. C. U., Green, E. P., Jamison, J. C., Lehmann, M. C., and Annan, J. (2015). The Returns to Microenterprise Support Among the Ultra-Poor: A field experiment in post-war Uganda. NBER Working Paper Series No Said, Mahmud & Chaudhry

30 Brauw, D., A., G., O., D., Hoddinott, J., and Roy, S. (2013). The impact of Bolsa Familia on women s decision-making power. World Development, 59: Campbell, H. (2012). Gender empowerment in microfinance: How SHGs in India exemplify the institutional potential. Jackson School Journal of International Studies Policy Brief, 3(1):6 18. Campos, F., Frese, M., Goldstein, M., Iacovone, L., Johnson, H. C., McKenzie, D., and Mensmann, M. (2017). Teaching personal initiative beats traditional training in boosting small business in west africa. Science, 357(6357): Crepon, B., Devoto, F., Duflo, E., and Parienté, W. (2015). Estimating the impact of microcredit on those who take it up: Evidence from a randomized experiment in morocco. American Economic Journal: Applied Economics, 7(1): de Mel, S., M., D., and Woodruff, C. (2009). Are women more credit constrained? Experimental evidence on gender and microenterprise returns. American Economic Journal: Applied Economics, 1(3):1 32. de Mel, S., M., D., and Woodruff, C. (2012). One-time transfers of cash or capital have long-lasting effects on microenterprises in Sri Lanka. Science, 335: Duflo, E., Banerjee, A., Glennerster, R., and Kinnan, C. G. (2013). The miracle of microfinance? Evidence from a randomized evaluation. NBER Working Paper Series No Duflo, E., Glennerster, R., and Kremer, M. (2008). Using Randomization in Development Economics Research: A Toolkit, volume 4. North Holland, Amsterdam and New York. This file is the version posted by the Centre for Economic Policy Research, CEPR Discussion Papers: 6059, Fafchamps, M., Kebede, B., and Quisumbing, A. R. (2009). Intrahousehold welfare in rural Ethiopia. Oxford Bulletin of Economics and Statistics, 71(4): Fafchamps, M., McKenzie, D., Quinn, S., and Woodruff, C. (2014). Microenterprise growth and the flypaper effect: Evidence from a randomized experiment in Ghana. Journal of Development Economics, 106: Fiala, N. (2015). Business is tough but family is worse: The role of family constraints on microenterprise development in Uganda. Unpiblished Manuscript. Field, E., Jayachandran, S., Pande, R., and Rigol, N. (2016). Friendship at work: Can peer effects catalyze female entrepreneurship? American Economic Journal: Economic Policy, 8(2): Said, Mahmud & Chaudhry

31 Ghalib, K., A., M., I., Imai, S., and K. (2011). The impact of micro finance and its role in easing poverty of rural households: Estimations from Pakistan. Kobe University Discussion Paper Series No Ginè, X. and Mansuri, G. (2011). Money or Ideas? A Field Experiment on Constraints to Entrepreneurship in Rural Pakistan. The World Bank Policy Research Working Paper Series, (September). Grasmuck, S. and Espinal, R. (2000). Market success or female autonomy? Income, ideology, and empowerment among microentrepreneurs in the Dominican Republic. Gender and Society, 14(2): Haq, A. and Safavian, M. (2013). Are Pakistan s Women Entrepreneurs Being Served by the Microfinance Sector? World Bank Publications. Haushofer, J. and Shapiro, J. (2016). The short-term impact of unconditional cash transfers to the poor: Experimental evidence from Kenya. The Quarterly Journal of Economics, 131(4): Jakiela, P. and Ozier, O. (2012). Does Africa need a rotten kin theore Experimental evidence from village economies. World Bank Policy Research Working Paper No Kabeer, N. (2001). Conflicts over credit: Re-evaluating the empowerment potential of loans to women in rural Bangladesh. World Development, 29(1): Lee, D. S. (2009). Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Effects. The Review of Economic Studies, 76(3): Lybbert, T. J. and Wydick, B. (2016). Hope as aspirations, agency, and pathways: Poverty dynamics and microfinance in oaxaca, mexico. Working Paper 22661, National Bureau of Economic Research. Maldonado, J. H. (2005). The Influence Of Microfinance On The Education Decisions Of Rural Households: Evidence From Bolivia. Documentes Cede , Universidad de Los Andes-Cede. McKenzie, D. (2008). Experimental evidence on returns to capital and access to finance in mexico. World Bank Economic Review, 22(3): Memon, A. S., Naz, S., Abass, H., Zahid, J., Tabbasum, R., and Zeshan, M. (2014). Alif Ailaan Pakistan District Education Rankings. Technical report, Alif Ailaan in collaboration with Social Development Policy Institute. Nghiem, S., Coelli, T., and Rao, P. (2011). Assessing the welfare effects of microfinance in Vietnam: Empirical results from a quasi-experimental survey. Journal of Development Studies, 48(5): Said, Mahmud & Chaudhry

32 Salman, A. (2008). Evaluating the impact of microcredit on women s empowerment in Pakistan. CREB Working Paper Series No 03/09. Setboonsarng, S. and Parpiev, Z. (2008). Micro finance and Milleninum development goals in Pakistan: Impact assessment using propensity score matching. ADB Institute Discussion Paper No Tarozzi, A., Desai, J., and Johnson, K. (2014). The impacts of microcredit: Evidence from Ethiopia. American Economic Journal: Applied Economics, 7(1): Weber, O. and Ahmad, A. (2014). Empowerment through microfinance: The relation between loan cycle and level of empowerment. World Development, 62: Zu Selhausen, F. (2013). Husbands and wives. The powers and perils of participation in a microfinance cooperative for female entrepreneurs. Utrecht School of Economics Working Paper No Said, Mahmud & Chaudhry

33 Appendix A Tables Table A1: Predicting attrition Outcome: Not Attrited Not Attrited Not Attrited Sharpened q-values (1) (2) (3) (4) Treatment Assignment 0.088*** 0.052*** Family 1: Demographics Age (years) Dummy: Respondent is currently married ** 0.082* 0.157** Dummy: Respondent can read and write Number of children (years <17) in the 0.018* household Household dependency ratio 0.037** 0.044** 0.053** Family 2: Occupation and experience Dummy: Respondent has a business Dummy: Respondent has worked as a paid employee in the past Dummy: Respondent has had a business in the past Family 3: Household assets and income Household expenditure in an average month (PKR) Dummy: household home is owned by a household member Index: Assets owned by the household 0.022* 0.030* 0.18 Family 4: Intrahousehold agency and autonomy Dummy: Respondent is confident she can support hh for 4 weeks Index: Respondent makes decisions in the household herself 0.070* 0.069* *** 0.023** Said, Mahmud & Chaudhry

34 Inverse variance covariance index (Anderson, 2008) out of empowerment index and confidence variable Dummy: Respondent is not allowed by the household to seek employment Family 5: Access to formal or informal finance Dummy: Household has outstanding loans Dummy: Household member(s) have participated in ROSCAs Dummy: Household member(s) have a bank account * N p-value of F test of joint significance of explanatory variables Above variables interacted with Treatment No No Yes Yes Note: p < 0.01, p < 0.05, p < 0.1. Column (1) reports the coefficient on the variable in the row when they are all included in a regression where the output is being successfully located and surveyed. Column (2) reports the coefficient on treatment status when the outcomes is being successfully located and surveyed. Column (3) reports the coefficient on row variable when included in a regression with treatment status and the interaction of each row variable with treatment status. The inverse covariance index variable in Family 4 drops out from the regression due to collinearity with variables in Family 4. Finally, column (4) reports critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 34 Said, Mahmud & Chaudhry

35 The following table describes performance variables for business status since treatment was offered. Table A2: Summary statistics: Business performance Midline (1 year later) Endline (2 years later) N Mean S. Dev. N Mean S. Dev. Busines start up costs (PKR) Business assets (PKR) Business average monthly expenditure (PKR) Business average monthly revenues (PKR) Business average monthly profits (1) (PKR) Business average monthly profits (2) (PKR) Note: Average monthly profit (1) is coded as the difference between reported business revenue and expenditure. Average monthly profit (2) is the self-reported level of profits. $1 PKR Said, Mahmud & Chaudhry

36 Table A3: Impact of treatment and husband s businesses on business status Short term Long term Set up Shut down Set up Set up Shut down business last year business last year last year (1) (2) (3) (4) (5) Treatment (0.024)*** (0.022)*** (0.019) (0.017) (0.011) Husband has existing business (0.066)*** (0.038) (0.040) (0.039) (0.045) Treatment*Husband has existing business (0.087)** (0.057) (0.061) (0.055) (0.058) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Set up business is a binary variable equal to 1 if the respondent set up a business since baseline; Set up last year is a binary variable equal to 1 if the respondent set up a business in the last year; and Shut down last year is a binary variable equal to 1 if the respondent had a business that she shut down in the last year. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 36 Said, Mahmud & Chaudhry

37 This tables shows the short term impact on individual expenditure items. 37 Said, Mahmud & Chaudhry

38 Table A4: Short term impact on individual expenditure items Food Non durable Medical School Recreation (Mobile Gifts/loans Set aside excl. food (bills) ( to others) (as savings) (1) (2) (3) (4) (5) (6) (7) (8) Treatment ( ) ( ) ( ) ( ) (35.083) ( ) ( ) ( ) Foodt= (0.051) Non durable foodt=0 excl (0.125) Medicalt= (0.089) Schoolt= (0.112) Recreationt= (0.027) Mobile billst=0 (0.050) Gifts/loans to othert=0 (0.044) Set aside savingst=0 as (0.100) N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Monthly household expenditure is calculated by summing up the average monthly expenditure on different items, reported in PKR. Home owner is a binary variable equal to 1 if someone in the household owns the household home. Asset index is an index created from the number of assets owned by the household using Principal Component Analysis. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 38 Said, Mahmud & Chaudhry

39 Table A5: Long term impact: Households assets and expenditure Monthly household Home owner Asset index expenditure (PKR) (1) (2) (3) Treatment ( ) (0.090) (0.588) Monthly household AA expenditure t=0 (0.104)* Home owner t= AAA (0.048)*** Asset index t= AAA (0.040)*** MDE N R Note: All regressions include controls for (married, marriedtreat and dep ratio for attrition imbalance). Mention here the clstering and se correction etc. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 39 Said, Mahmud & Chaudhry

40 Table A6: Long term impact: Access to finance Bank account Took loan(s) Insurance ATM card last year (1) (2) (3) (4) Treatment (0.101) (0.063) (0.074) (0.088) Bank account t= (0.111) Took loan(s) AA last year t=0 (0.087) Insurance t= AAA (0.061)*** ATM card t= AAA (0.062)*** MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Bank account is a binary variable that is 1 if someone in the household currently has a bank account. Took loan(s) last year is a binary variable equal to 1 if someone in the household took out a loan (other than the treatment loan) in the last year. Insurance and ATM are only available for the endline survey and are binary variables equal to 1 if someone in the household currently has insurance or an ATM card, respectively. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 40 Said, Mahmud & Chaudhry

41 Table A7: Long term impact: Agency and autonomy in decision making Confident Empowerment Agency index index (1) (2) (3) Treatment (0.097) (0.492) (0.218) Confidence t= (0.052) Empowerment AA index t=0 (0.043)*** Agency index t=0 (0.051) MDE N R Note: All regressions include controls for baseline characteristics that can predict attrition and branch dummies with errors clustered at the individual level. Confident is a binary variable equal to 1 if the respondent believes she can support her family on her own for 4 weeks. Empowerment index is an index created using Principal Component Analysis from variables that measure if the respondent can make household decisions (clothing, footwear, medical, recreation, social visits, joining credit groups, purchases for self, purchases for others, marriage, investment) on her own. Agency index is an inverse variance-covariance index (Anderson, 2008) created out of the Confident and Empowerment index variables. MDE is the ex post minimum detectable effect size at a significance level of 0.05 and power of 80 percent. p < 0.01, p < 0.05, p < 0.1. Adjusting critical values following the approach by Benjamini and Hochberg, 1995: AAA Significance at 1% level, AA Significance at 5% level, A Significance at 10% level. 41 Said, Mahmud & Chaudhry

42 Appendix B Figures Figure B1: Treatment product - Repayment sources Note: x-axis shows the proportion (%) of treatment loan recipients who reported the item on the y-axis as the largest source uses to repay the treatment loan. This question was asked only at midline, that is, one year after the disbursement of the loan and shortly after the loan had been repaid. 42 Said, Mahmud & Chaudhry

43 Figure B2: Other loans - Stated purpose (a) Purpose of other loans at t = 0 (b) Purpose of other loans at t = 1 43 Said, Mahmud & Chaudhry

44 (c) Purpose of other loans at t = 2 Note: x-axis shows the proportion (%) of respondents who report the item on the y-axis as the main purpose of current loans. The graphs are based on questions at each survey round about whether the household has outstanding loans (baseline) or additional loans last year (midline and endline) and what was the purpose of each loan. Figure B2: Type of business Note: x-axis shows the proportion (%) of respondents who report that their business was of the type specified on the y-axis. 44 Said, Mahmud & Chaudhry