Soap Operas for Female Micro Entrepreneur Training

Nakasone, Eduardo; Torero Cullen, Máximo Working Paper Soap Operas for Female Micro Entrepreneur Training IDB Working Paper Series, No. IDB-WP-564 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Nakasone, Eduardo; Torero Cullen, Máximo (2014) : Soap Operas for Female Micro Entrepreneur Training, IDB Working Paper Series, No. IDB-WP-564, Inter-American Development Bank (IDB), Washington, DC, https://hdl.handle.net/11319/6791 This Version is available at: https://hdl.handle.net/10419/115510 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by-nc-nd/3.0/igo/legalcode Soap Operas for Female M icro Entrepreneur Training Eduardo Nakasone Maximo Torero Multilateral Investment Fund (MIF) IDB-WP-564 IDB WORKING PAPER SERIES No. Inter-American Development Bank December 2014 Soap Operas for Female Micro Entrepreneur Training Eduardo Nakasone Maximo Torero 2014 Inter-American Development Bank h ttp://www.iadb.org T he opinions expressed in this publication are those of the authors and do not necessarily reflect th e v iews of the Inter-American Development Bank, its Board of Directors, or the countries they r epresent. T he unauthorized commercial use of Bank documents is prohibited and may be punishable under th e B ank's policies and/or applicable laws. C opyright © Inter-American Development Bank. This working paper may be reproduced for a ny non-commercial purpose. It may also be reproduced in any academic journal indexed by the A merican Economic Association's EconLit, with previous consent by the Inter-American Developmen t B ank (IDB), provided that the IDB is credited and that the author(s) receive no income from the p ublication. Eduardo Nakasone I nternational Food Policy Research Institute (IFPRI) e [email protected] M aximo Torero I nternational Food Policy Research Institute (IFPRI) [email protected] Cataloging-in-Publication data provided by the I nter-American Development Bank F elipe Herrera Library N akasone, Eduardo. Soap operas for female micro entrepreneur / Eduardo Nakasone, Maximo Torero. p. cm. — (IDB Working Paper Series ; 564) Includes bibliographic references. 1. Entrepreneurship—Training of—Peru. 2. Women executives—Training of—Peru. I. Torero, Maximo. II. I nter-American Development Bank. Multilateral Investment Fund. III. Title. IV. Series. I DB-WP-564 2014 Soap Operas for Female Micro Entrepreneur Training Eduardo Nakasone and Maximo Torero∗ December 9, 2014 Abstract This paper analyzes the impact of the Strengthening Women Entrepreneurship in Peru (SWEP) program. SWEP trained female micro entrepreneurs in business management practices (such as accounting and marketing). The training, which was provided in 4- to 5-hour sessions, used soap operas and practical exercises specifically designed for the program. A field experiment was conducted among a group of micro entrepreneurs based in two Peruvian cities (Lima and Piura) to investigate whether SWEP had a positive impact on its beneficiaries. The results show that the program positively affected the adoption of business practices taught by the program. In particular, those who received the training were 4 to 6 percentage points more likely to assign themselves a fixed salary (rather than taking cash from their businesses based on personal needs) and 6 to 11 percentage points more likely to keep better records of potential business contacts. Some positive impacts were found on the adoption of bookkeeping practices (4 to 6 percentage points), although this result is not significant across all of the specifications. Although these changes in adoption rates were large compared with their baseline levels, they were rather small in absolute terms. Therefore, the study did not find any impact on average business performance, household expenditures, or women’s empowerment in the household. Qualitative information suggests that micro entrepreneurs were satisfied with the training, but considered that many of the practices taught by the program were difficult to follow because of time constraints. ∗Markets, Trade and Institutions Division, International Food Policy Research Institute (IFPRI). 2033 K St NW, Washington DC 20006. Nakasone: [email protected]; Torero: [email protected]. We are grateful to Miriam Vasquez and Giancarlo Cafferata from APRENDA; Ximena Querol, Director of the 10000 Women Program - Peru; and Sandra Darville, Claudia Gutierrez, and Carmen Mosquera from the IADB’s Multilateral Investment Fund for their support in the evaluation of the SWEP program. USKAY EIRL facilitated the data collection. Yuri Soares, Miguel Almanzar, and anonymous reviewers from IADB provided insightful comments on a previous version of this paper. Mike Murphy provided research assistance with the data and the survey instrument. Maribel Elias and Luciana Delgado supervised all the logistical details of the field work. All errors remain our own. 1 Introduction There is increasing interest in understanding the role of micro enterprises in developing countries. First, micro enterprises constitute the vast majority of firms in developing countries. For example, Li and Rama (2012) find that this group comprises 61percent of firms in Chile; 83 percent in Turkey; 95 percent in South Africa; 85 percent in India; and 84 percent in Pakistan. Although some micro enterprises will disappear, others may transition into larger companies and provide the foundation for a modern private sector. Second, micro enterprises are the most important source of employment: 72 percent of jobs are created in micro enterprises in developing countries.1Third, they have relatively low productivity levels, and this is especially concerning given their large share of employment. In this line, Angelelli, Moudry, and Llisterri (2006) calculate that, even though 77 percent of total employment is generated by micro and small enterprises in Latin America, they only contribute to 30-60 percent of the gross domestic product. The economic literature provides a long list of potential culprits for the low levels of productivity of micro enterprises. A considerable proportion of previous studies has focused on the challenges that financial constraints can impose on small businesses: if firms are unable to borrow, they would be unable to finance optimal levels of capital (e.g., Evans and Jovanovic 1989; de Mel, McKenzie, and Woodruff 2008; Fafchamps et al. 2014; Banerjee et al. 2013). Restuccia and Rogerson (2008) argue that policies that create other distortions in input or output markets can lead to misallocation of resources and reductions in productivity. Low levels of human capital (at least, measured through formal schooling) have been another candidate to explain low levels of productivity in small firms in developing countries. De Soto (1989) argues that the 1There are some other calculations of the share of jobs in micro enterprises in developing countries (for example, see Ayyagari, Demirguc-Kunt, and Maksimovic 2011). However, most of these estimates are based on business surveys (which usually exclude firms in the informal sector) or are calculated for a single sector (e.g., manufacturing). We construct the average of the share of employment in micro enterprises for a sample of developing countries in the World Development Report 2013. (See the Report’s Online Appendix). The World Bank’s (2012) estimates are based on household surveys and, therefore, do include the informal sector. We estimate the average share of employment in micro enterprises (weighted by the size of their labor force) for 27 developing countries that have collected household surveys circa 2005 or 2010. 2 regulatory framework in developing countries creates an unnecessary burden for businesses in developing countries and promotes informality. More recently, another potentially constraining factor in micro enterprise development has gained attention: managerial capital (Bruhn, Karlan, and Schoar 2010). For example, Bloom and Reenen (2010) argue that persistent differences in firm productivity can be explained by management practices. The authors surveyed around 6,000 firms in 17 countries and measure their practices in three broad areas: monitoring (e.g., production process tracking), targets (e.g., goal setting), and incentives (e.g., promotion workers with high performance). Their findings suggest that there is considerable variation in firms’ practices within and between countries, and that better management is associated with stronger performance (in terms of size, productivity, and survival). The idea that managerial capital can spur firms’ growth has powerful policy implications: even within their capital limitations or adverse regulatory environment, micro enterprises can improve if they put their inputs to a better use. Not surprisingly, this notion has promoted many business training programs in developing countries (Cho and Honorati 2013). In this paper, we assess the impact of a large-scale program that provided short-term training to female micro entrepreneurs in Peru. The project, “Strengthening Women Entrepreneurship in Peru (SWEP)”, was implemented by the Inter-American Development Bank’s Multilateral Investment Fund, who partnered with APRENDA (a local institution that provides business training for micro entrepreneurs) and the Thunderbird School of Global Management (a United States-based business school). The general objective of the project was to improve the contribution of women-headed micro and small enterprises to family incomes and the economy of Peru, by providing assistance aimed at broadening access to business training. Since its deployment in 2010, SWEP has provided business training to 100,000 women in Peru. Training sessions were free for participants, short (one afternoon), and provided only once. They relied on a combination of media, games, practical exercises, and take-home guides. The contents used a soap opera format and were designed by APRENDA and Thunderbird for the needs of female Peruvian micro entrepreneurs . The soap opera depicts the struggles of the owner of a cash-strapped grocery shop who decides to open a catering business when her 3 husband gets fired from his job. The idea is that trainees could relate to the main character’s problems.2Among other topics, the soap opera illustrates the advantages of timely cash flows, setting a fixed salary for the owners (rather than using the business’s cash for personal needs), keeping record of important business clients and suppliers to develop a business network, and better work-life balance. The soap opera was complemented with instructors’ in-class instructions for group activities and workbook exercises. We designed a field experiment where we reproduced SWEP’s recruiting process and conducted a baseline survey with around 2,500 female micro enterprise owners. Participants were randomly assigned to one of two groups: 60 percent were assigned to the treatment group and were immediately invited to participate in SWEP’s training, while the remaining 40 percent were assigned to the control group and received the training a year later, at the end of the experiment. We conducted follow-up surveys 6 and 12 months after the baseline to investigate whether women who attended the SWEP sessions implemented the business management practices taught by the program and if these improved practices had any impact on their business outcomes (sales, expenditures, profits, size, and productivity). Given that the program should affect resources under women’s control, we examine if changes in business outcomes translate to their household. Therefore, to some extent, we also analyze whether enhanced business profitability induces improvements in broad household welfare indicators and female bargaining power. Recent papers have analyzed the impact of business training on micro entrepreneurial outcomes and have found mixed results.3For example, Calderón, Cunha, and Giorgi (2013) find that a 48 hour business training program in rural Mexico increased beneficiaries’ profits, revenues, and clients. Karlan and Valdivia (2011) conduct a field experiment program among 2Previous evidence has shown that access to television and media can alter women’s behavior. For example, Jensen and Oster (2009) show that access to cable TV increases female autonomy and lowers tolerance of domestic violence. La Ferrara, Chong, and Duryea (2012) show that areas where soap operas (novelas) are broadcasted experience reductions in fertility rates. The authors argue that the more “modern” attitude that female characters have in the novelas generates changes in women’s role models and aspirations. 3For a more detailed review of the available evidence on the impact of training programs for small firms in developing countries, see McKenzie and Woodruff (2014) and Cho and Honorati (2013). 4 clients of a microfinance institution (MFI) in Peru to analyze the impact of training. The training was provided in 30-60 minute sessions after the clients’ regular banking meetings over a period of one to two years. While they find that participants in the program improved their knowledge on business practices, they do not find changes in their business performance (i.e., revenue, profits, or employment). Field, Jayachandran, and Pande (2010) implement a similar field experiment with poor female bank clients in India, where a group of them participated in a two-day training on financial literacy, business skills, and aspirations. They find differential effects of the program for different castes: while Hindu women in upper castes increased their borrowing and business incomes, there was no effect among Muslims or Hindu women from historically disadvantaged castes. Bruhn, Karlan, and Schoar (2010) investigate the impact of a training program in Bosnia and Herzegovina and find increased financial basic knowledge among its beneficiaries. However, they also find that this increased knowledge did not translate into increases in the probability of business survival, business start-up, performance, or sales.4 This somewhat disappointing evidence led many to believe that training by itself would not suffice for businesses to thrive. One possibility is that, even with more entrepreneurial skills, financial constraints might still hinder firms’ growth. de Mel, McKenzie, and Woodruff (2014) test this hypothesis by analyzing the impact of a program that provided a group of women with either training alone or a combination of training and a cash grant. Their results suggest that training by itself led to the adoption of enhanced business practices, but did not improve firms’ profitability.5In contrast, the combination of human and financial capital led to increases in short-run improvements in business performance, but this effect dissipates a couple of years later. However, these effects were not found by Berge, Bjorvatn, and Tungodden (2011) and Giné and Mansuri (2011) in Tanzania and Pakistan, respectively. In their studies, male and 4The authors do find some improvements on sales and performance among entrepreneurs with ex-ante higher financial literacy. But, overall, they do not find any average effects of the program. 5Karlan, Knight, and Udry (2012) conduct a similar field experiment with tailors in urban Ghana, where they test the impact of a training program (business consultancy) and a cash grant among tailors. They find that the training led to a temporary change in business practices (which disappeared a year later) and that the cash grant led to business investments. However, neither intervention increased firms’ profitability. 5 from certain predetermined locations and provided them with lunch, a coffee break, and a symbolic present for their participation. Attendants would also enter a lottery where they could win appliances and other prizes. To maximize attendance, the training venues were carefully chosen considering the availability of public transportation, their central location, and closeness to the areas where the roster was collected.12 Despite all our efforts , only 703 (47 percent) out of the 1,500 women invited to the trainings attended. We collected a much more detailed baseline with the characteristics of the micro entrepreneurs in our field experiment. Our questionnaire included information on the women’s micro enterprises (i.e., business practices, sales, costs, number of employees, payroll, access to credit, etc.) and households (i.e., composition, household expenditures, women’s role in household decision making, etc.). The survey was administered with Android-based tablets to facilitate data collection. We took two approaches for the baseline collection. In one approach, women in the treatment group were interviewed in the training venues. Upon their arrival but prior to the beginning of the training the enumerators administered the survey. In the other approach, enumerators visited women in the control group at their houses or businesses to gather their information. Because enumerators were able to visit micro entrepreneurs multiple times and (re)schedule appointments as needed, 1,035 of the 1,100 women in our evaluation sample completed the baseline survey. We were not able to collect data (baseline or follow-up) for invited micro entrepreneurs who did not attend the training or from the women in the control group who refused to be interviewed.13 Therefore, our impact evaluation will be based on an effective sample of 703 micro entrepreneurs who did attend the training and 1,035 women in the control group. Because our sample and allocation deviated from our original random design, this creates some challenges to interpret our results. While the refusal rate (6 percent) in the control group was fairly small, 12There were ten training sessions in Lima in eight districts: Chorrillos (2), El Augustino (1), Independencia (2), Breña (2), Magdalena (1), Independencia (1), and Villa El Salvador (1). There were two training sessions in Piura. 13This prevents us from calculating standard Intention-to-Treat estimates as other papers in the literature (Karlan and Valdivia 2011; Valdivia 2011; Karlan, Knight, and Udry 2012; de Mel, McKenzie, and Woodruff 2014; Drexler, Fischer, and Schoar 2014). 12 the significantly larger no-show rates in the treatment group poses a problem: although women were randomly invited to the training, the actual decision to attend was not random. The last two columns of Table 1 shows that women in our final treatment group were 1.4 years older and more likely to work in retail. Table 2 presents more detailed characteristics of our effective sample at baseline. Women in both groups are similar in several dimensions (e.g., business value, education, access to formal credit, etc.). Importantly, they also have similar business outcomes in terms of sales, payroll, and other business expenditures. However, our effective treatment group has larger firms in terms of number of workers, has more access to informal credit, and is more likely to own their own stores. This seems to suggest that, if anything, women who attended the training were better off (and likely better able) than their counterparts in the control group. Therefore, the results of our paper constitute an upper bound of the actual effects of the training. The APRENDA training sessions were typically 4-5 hours long. During these sessions, attendants watched a soap opera about Vicky, a struggling bodega owner.14 The soap opera depicted wrongful management practices that are common among micro entrepreneurs.. Vicky’s bodega was not able to keep up with the increasing household expenditures of her son and younger sister, and usually she covered these personal expenses with the bodega’s day-to-day sales. Her situation turns critical when her husband is fired from his job and her sister unexpectedly gets pregnant. With the help of a friend, Vicky decides to start a catering business and get better organized. She sets a salary for herself (separating her personal expenses from the business’s profits), sets a cash flow to decide any new investments, and collects a list of potential clients . These changes in practices eventually pay off and she is able to open a successful restaurant after a few months. The soap opera was divided in six parts. After each part, the instructor reinforced the concepts the video illustrated with other examples and group games. Participants also received a workbook with exercises. After the soap opera and the instructor’s explanations, participants 14The soap opera videos can be found in this website: http://www.programasalta.org/Programa/Material. 13 Table 2: Baseline Characteristics by Treatment Status Characteristic Control Treatment Characteristic Control Treatment A. Business Characteristics (cont0d) Age of business 98.8 101.7 Size of most important 1,331 1,377 (months) (91.8) (91.4) informal loan 1(1,724) (1,792) Own store 0.48 0.52 ∗D. Household Variables (0.50) (0.50) Monthly per capita HH 358.5 363.7 Business value 26,674 25,780 expenditure (soles) (216.2) (220.2) (self-reported, soles) (34,008) (33,796) Value of household 4,132 4,927 B. Business Performance (soles) assets (5,998) (6,220) ∗∗∗ Yearly sales 43,887 42,003 D. Micro Entrepreneur’s Characteristics (41,364) (39,925) Married 0.62 0.63 Yearly payroll 1,736 1,863 (0.49) (0.48) (4,024) (4,068) Head of household 0.42 0.44 Yearly expenditures 27,615 24,869 (0.49) (0.50) (excl. payroll) (36,266) (32,354) Years of education 10.96 11.08 C. Employment (3.36) (3.28) Total number of workers 0.81 1.10 ∗∗∗ E. Does woman decide..? 3 (1.17) (1.36) How to spend money 0.70 0.73 Number of permanent 0.44 0.52 (0.46) (0.44) salaried workers (0.99) (1.15) Food purchases 0.83 0.85 Number of permanent 0.28 0.42 ∗∗∗ (0.38) (0.36) non salaried workers (0.63) (0.72) Furniture purchases 0.68 0.71 Number of salaried 0.06 0.10 ∗∗∗ (0.47) (0.46) temporary workers 1(0.24) (0.34) Family outings 0.67 0.68 Number of temporary 0.03 0.06 ∗∗∗ (0.47) (0.47) non salaried workers 1(0.15) (0.20) Children’s education 0.70 0.73 D. Credit (0.46) (0.44) Received formal loan 0.28 0.31 Family discipline 0.66 0.71 ∗∗ in last 12 months (0.45) (0.46) (0.47) (0.45) Size of most important 6,193 5,691 What to do if any HH 0.82 0.82 formal loan 2(7,476) (9,403) member is sick (0.38) (0.39) Received informal loan 0.16 0.23 ∗∗∗ Observations 1,035 703 in last 12 months (0.37) (0.42) 1Number of temporary workers was adjusted by the number of months worked, i.e., number of temporary workers x (number of worked months) / 12. 2For the sample of those who received a loan in the last 12 months. 3Whether each of these decisions is primarily made by the micro entrepreneur. Standard errors in parentheses. Significance levels denoted by: *** p<0.01, ** p<0.05, * p<0.1. 14 would start working on exercises. For example, the workbook would provide a hypothetical list of business sales, expenditures, and personal expenses. Participants would need to determine the cash flow, potential investments, and the salary they could assign themselves. The workbook also included take-home exercises where micro entrepreneurs were encouraged to work on their own cash flows and to create a list of potential clients to expand their business. To measure the impact of the program, we collected two follow-up surveys. The first one was collected six months after the training (January and February 2012) and the second one was collected another six months later (August and September 2012). This survey was collected among those in the effective evaluation sample (703 in the treatment group and 1,035 in the control group). Enumerators visited micro entrepreneurs either at their home or business, depending on their availability. The questionnaires were similar to those in the baseline for comparability reasons. 3 Empirical Approach and Results Our results focus on four main aspects on which the program might have affected micro entrepreneurs. First, we analyze whether training led to the adoption of the business practices that were taught to participants. The contents of the training prioritized three rules: (a) rather than taking cash based on the needs of their households, micro entrepreneurs should assign themselves a fixed salary every month; (b) they should register all sales and expenses in a cash flow to aid them in their investment and credit decisions; and (c) they should build a list of potential contacts (i.e. buyers, suppliers, etc.) to network. Second, we analyze whether there were any changes in business outcomes. In particular, we investigate if the training affected firms’ sales, payroll, and other expenses. Third, we analyze if the program had any impact on micro entrepreneurs’ household welfare as measured by participants’ increased expenditures and the degree of the micro entrepreneurs’ empowerment in the family. We analyze the data with several specifications. Our basic specifications exploit the difference in outcomes Yit for the post-intervention rounds of the data through a simple regression analy- 15 sis: Yi,t=T=βTreati+εit (1) where Treatiis an indicator variable for the treatment group and T=1 for the first followup (six months after the intervention) and T=2 for the second follow-up (a year after the intervention). We also estimate the average effect over the impact evaluation period by pulling together both post-intervention rounds and estimating: Yi,t=1,2 =βTreati+θDt=2+εit (2) where Dt=2is an indicator variable for the 12-month follow-up round of the data. Because our effective sample deviated from the original random design, we also estimate two additional versions of Equation 1 where we control for a set of characteristics of women in the control and treatment groups. In this spirit, we estimate ANCOVA (analysis of covariance) estimators, by including the baseline value of the dependent variable (Yi,t=0) as a regressor.15 We also estimate Equation 1 conditioning on a set Xiof micro entrepreneur’s characteristics. This set Xiincludes sector (i.e., services, manufacturing, or retail), city, age, head of household at baseline, whether the micro entrepreneur had received previous business training, and marital status. Yit=T=βTreati+γYi,t=0+εit (3) Yit=T=βTreati+δXi+εit (4) Finally, we fully account for any (time-invariant) unobservable differences between the treatment and control groups, by estimating a fixed effects regression. Because this estimator only relies on within micro entrepreneur variability, it accounts the most for pre-intervention differ- 15McKenzie (2012) shows that the ANCOVA estimator accounts for pre-treatment differences in outcomes. In particular, ˆ βANCOVA = (YT Post −YC Post)−ˆ γ(YT Pre −YC Pre). He also shows that this procedure can increase the efficiency of the estimator. 16 ences of participants, but is also less efficient. Yit =αi+β1TreatiDt=1+β2TreatiDt=2+θ1Dt=1+θ2Dt=2+εit (5) Table 3 shows the results of our estimation framework for business practices. We find a strong significant impact of the program on the probability that the micro entrepreneur assigns herself a fixed salary. The coefficient is highly significant across all specifications and is large in magnitude: there is an increase of 5 percentage points, with a baseline value of 4 percent (more than doubling the share of women who implement this practice). In most specifications, we also find a positive impact of the program on the other two practices: bookkeeping (4-6 percentage points, from a baseline value of 45 percent) and networking list (3-10 percentage points, from a baseline value of 28 percent). However, the effect on networking is not significant in the fixed effects models, and the impact on bookkeeping is not significant in the ANCOVA and fixed effects models (arguably the two models with lower power to detect any effect). Althoughwe find a general positive impact on the adoption of business practices, we do not find that these changes translate into any improvements in business performance. Table 4 shows the results for firms’ sales, payroll, and business expenditures (excluding payroll). Our results suggest that firms did not experience any improvements and are, in general, consistent with previous results in the literature. As suggested by McKenzie and Woodruff (2014), most studies find some changes in business practices taught to training participants, but no effect on outcomes. Finally, we estimate the impact of the program on household outcomes. First, we estimate whether there were any increases in household per capita expenditures (which is commonly used as a general measure of well-being). The results are shown in the top panel of Table 5, and suggest that the program did not have any significant impact on household expenditures. The coefficients are not statistically significant, and are rather small in magnitude compared with the means in the control group. Next, we also test the impact of the program on women’s empowerment. To capture the female micro entrepreneur’s degree of empowerment in the household, we captured a set of questions 17 Table 3: Impact of Training on Business Practices Specification (1) (2) (3) (4) (5) (6) (7) (8) Estimation Post1Post1Avg Post2ANCOVA3Post4ANCOVA3Post4F.E. 5 Rounds61 2 1,2 1 1 2 2 0,1,2 Controls No No No No Yes No Yes No N 1,480 1,371 2,851 1,480 1,480 1,371 1,371 4,589 A. Assigns Herself a Fixed Salary Treat 0.040*** 0.059*** 0.049*** 0.040*** 0.039*** 0.058*** 0.057*** (0.013) (0.014) (0.011) (0.013) (0.013) (0.014) (0.014) Treat x Dt=10.036** (0.017) Treat x Dt=20.056*** (0.017) ¯ Yctrl,t=00.040 ¯ Yctrl,t=10.051 0.051 0.051 0.051 ¯ Yctrl,t=20.044 0.044 0.044 0.044 B. Bookkeeping Treat 0.055** 0.046* 0.051** 0.038 0.044* 0.036 0.037 (0.026) (0.027) (0.021) (0.025) (0.026) (0.026) (0.026) Treat x Dt=10.000 (0.030) Treat x Dt=20.005 (0.032) ¯ Yctrl,t=00.446 ¯ Yctrl,t=10.413 0.413 0.413 0.413 ¯ Yctrl,t=20.446 0.446 0.446 0.446 C. Networking Treat 0.090*** 0.111*** 0.100*** 0.063*** 0.075*** 0.090*** 0.099*** (0.024) (0.026) (0.020) (0.023) (0.023) (0.025) (0.025) Treat x Dt=10.007 (0.029) Treat x Dt=20.028 (0.031) ¯ Yctrl,t=00.279 ¯ Yctrl,t=10.254 0.254 0.254 0.254 ¯ Yctrl,t=20.306 0.306 0.306 0.306 1Post estimators: Yi,t=βTreati+εit, for each t=1 and t=2. 2Average post estimator: Yi,t=1,2 =βTreati+θDt=2+εit. 3ANCOVA (analysis of covariance) estimators: Yi,t=βTreati+γYi,t=0+εit, for each t=1 and t=2. 4Post estimators including baseline controls: Yi,t=βTreati+δXi+εit, for each t=1 and t=2. Xiincludes sector, city, age, head of household at baseline, whether the micro entrepreneur had received previous business training, and marital status. 5Fixed effects estimator: Yit =β1TreatiDt=1+β2TreatiDt=2+θ1Dt=1+θ2Dt=2+αi+εit. 6Rounds: baseline (0), Jan-Feb 2012 follow-up (1), and Aug-Sep 2012 follow-up (2). Note: Standard errors in parentheses. Standard errors are clustered at the micro entrepreneur level in Columns 3 and 8. Significance levels denoted by: *** p<0.01, ** p<0.05, * p<0.1. Table 4: Impact of Training on Business Outcomes Specification (1) (2) (3) (4) (5) (6) (7) (8) Estimation Post1Post1Avg Post2ANCOVA3Post4ANCOVA3Post4F.E. 5 Rounds61 2 1,2 1 1 2 2 0,1,2 Controls No No No No Yes No Yes No N 1,468 1,366 2,834 1,453 1,468 1,361 1,366 4,536 A. Yearly Sales Treat -1,673.3 -1,786.4 -1,728.0 -2,064.4 236.5 -525.9 -543.5 (3,660.0) (3,174.7) (2,751.4) (2,234.5) (3,629.1) (3,028.6) (3,159.3) Treat x Dt=1-1,923.3 (2,539.3) Treat x Dt=2570.3 (3,042.8) ¯ Yctrl,t=043,886.7 ¯ Yctrl,t=152,243.7 52,243.7 52,243.7 52,243.7 ¯ Yctrl,t=245,950.3 45,950.3 45,950.3 45,950.3 B. Payroll Treat 394.2 -153.6 129.4 436.8 372.5 -175.0 -221.8 (397.0) (235.4) (269.0) (360.3) (395.5) (220.7) (233.6) Treat x Dt=1439.3 (407.9) Treat x Dt=2-41.9 (304.7) ¯ Yctrl,t=01,736.2 ¯ Yctrl,t=11,977.8 1,977.8 1,977.8 1,977.8 ¯ Yctrl,t=21,937.8 1,937.8 1,937.8 1,937.8 C. Business Expenditures (excluding payroll) Treat -1,532.5 -2,164.4 -1,837.0 -1,747.2 -144.2 -691.9 -1,272.4 (2,869.3) (2,554.2) (2,085.6) (1,265.7) (2,854.2) (1,220.7) (2,551.2) Treat x Dt=1-1,787.4 (1,487.4) Treat x Dt=2-797.4 (1,431.7) ¯ Yctrl,t=027,615.1 ¯ Yctrl,t=127,469.6 27,469.6 27,469.6 27,469.6 ¯ Yctrl,t=224,665.4 24,665.4 24,665.4 24,665.4 1Post estimators: Yi,t=βTreati+εit, for each t=1 and t=2. 2Average post estimator: Yi,t=1,2 =βTreati+θDt=2+εit. 3ANCOVA (analysis of covariance) estimators: Yi,t=βTreati+γYi,t=0+εit, for each t=1 and t=2. 4Post estimators including baseline controls: Yi,t=βTreati+δXi+εit, for each t=1 and t=2. Xiincludes sector, city, age, head of household at baseline, whether the micro entrepreneur had received previous business training, and marital status. 5Fixed effects estimator: Yit =β1TreatiDt=1+β2TreatiDt=2+θ1Dt=1+θ2Dt=2+αi+εit. 6Rounds: baseline (0), Jan-Feb 2012 follow-up (1), and Aug-Sep 2012 follow-up (2). Note: Standard errors in parentheses. Standard errors are clustered at the micro entrepreneur level in Columns 3 and 8. Significance levels denoted by: *** p<0.01, ** p<0.05, * p<0.1. to determine whether she, her husband, or other family member decides:16 (a) how to spend the household’s income; (b) food purchases; (c) furniture purchases; (d) family outings; (e) children’s education; (f) family discipline; and (g) what to do if a household member gets ill. Our analysis includes the estimation of seven empowerment measures. Independently testing such a large number of outcomes of the same family can lead to over-rejection: if we test a sufficiently large number of empowerment variables, we would considerably increase the probability that the treatment effect will be significant in at least one dimension (Duflo, Glennerster, and Kremer 2007; Schochet 2008). We are really more interested in the effect of the training on women’s empowerment, in general, rather than the strengthening of their bargaining power on specific household decisions. Therefore, we aggregate all seven decisions on a single index for this “family” of outcomes to avoid estimating too many outcomes. We adopt the framework proposed by Kling, Liebman, and Katz (2007). Consider Ghousehold decisions, each denoted by the subscript g. let σg be the standard deviation of the control group for each decision Dg. We estimate regressions Dg=θgTreati+εig for each g=1, ..., G. To account for the covariance of estimates θgacross equations, Kling, Liebman, and Katz (2007) propose to estimate them as a system of Seemingly Unrelated Regressions (SUR), rather than estimating the parameters individually.17 We then aggregate the effect on the family of decisions based on the treatment effects of each outcome 16Admittedly, measurement of women’s empowerment is a complex issue and presents several challenges. Although we did not design a set of original questions for this purpose, we used several items from the Mexican Intrahousehold Violence Survey (Encuesta sobre Violencia Intrafamiliar): http://tinyurl.com/MexicoEnvif (see page 3). 17Evaluation of business training usually entails a large number of outcomes and is prone to family-wise error (FWE) problems. Karlan and Valdivia (2011) and Valdivia (2011) also analyze the impact of training programs on aggregate indexes of outcomes. Their approach is somewhat different: they normalize each outcome with the mean and standard deviation and the control group and aggregate the normalized variables: 1 G∑g (Yig−µg) σg. They use this index as the dependent variable in a regression. This method is suitable when there are no missing variables in the analysis. In particular, (Kling and Liebman, 2004, p. 9) argue that “when an individual is missing data on an outcome however, the other non-missing outcomes implicitly are given more weight when the index is based on a simple average of non-missing standardized outcomes. The formulation described above based on the mean of estimated effects is a more direct summary of the estimates for each outcome.” Some of our decisions are prone to missing data problems. For example, the question of whether women decide about their children’s education is only answered by women with school-age sons or daughters. Therefore, we prefer to normalize the coefficients for different decisions rather than to create a single regressand based on normalized individual decisions. 20 and the control group standard deviations:18 τ=1 G"G ∑ g=1 θg σg#(6) The sample variance of τ(used to test its significance level) is based on the full variancecovariance matrix of θestimated through the SUR sytstem. Kling, Liebman, and Katz’s (2007) estimation on a family of outcomes allows us to tell whether women experience any positive effect on the set of decisions as a whole. The bottom panel of Table 5 shows the results for the aggregate impact of the program an women’s household decision-making. We find no statistically significant differences in the normalized aggregate empowerment index. 4 Discussion of Results We discuss four threats to the validity of the above findings: (a) missing data in the treatment group caused by selective attendance to the training; (b) attrition in follow-up surveys; (c) statistical power limitations; and (d) low quality of the training and levels of practice adoption. 4.1 Selective Attendance to the Trainings First, our identification strategy can be compromised by the high no-show rates among the invitees to the trainings. Because of the random assignment, had all the invitees attended the trainings, we would have expected very similar treatment and control groups. However despite all our efforts to increase attendance and our previous screening of each micro entrepreneur’s interest only 703 (47 percent of invited women) attended the trainings. We could not interview the remaining 797 women in the treatment group who did not attend, and they 18Because we use different samples in our estimations, σgvaries between specifications. In the cross-sectional post regressions with and without controls (i.e. Equations 1 and 4) we use the standard deviation of the control group in the period for which we estimate the regression (either t=1 or t=2). In the ANCOVA regressions, we standardize τusing the standard deviation in t=1. Standardization in the ANCOVA is based on the baseline values of σg. It is not possible to estimate a difference-in-differences specification with this framework. 21 ...). To preserve the dichotomous nature of the variable, we impute either 0s or 1s to the attrited control group to simulate each hypothetical rate pu c. For each value of δ, we assign 1s to nA c(po c+δ)(rounded up) observations, and assign 0s to the remaining nA c(1−po c−δ) observations. We are able to determine the value of ˜ δfor which the program no longer has any statistically significant effect, and calculate how much larger pu cwould need to be with respect to po cfor this to happen (i.e., po c+˜ δ po c). This procedure can also be accommodated for the treatment group. In this case, instead of increasing the proportion of adopting micro entrepreneurs in the control group, we could simulate decreasing rates of adoption among the treatment group. Suppose that nA Tmicro entrepreneurs attrited from the treatment group and that their (unobserved) adoption rate is pu T, where pu T= (po T−δ)<pc T. We can calculate the impact of the program for alternative values of δin: β=nA T N−Nc(po T−δ) + N−Nc−nA T N−Ncpo T−po c. We impute 1s in nA T(po T−δ)observations in the attrited treatment group, and impute 0s for the remaining nA T(1−po T+δ)observations. Again, this procedure will tell us the critical value of ˜ δand how much smaller the adoption rate would need to be among the attrited group to eliminate the effect of the program. We followed the same procedures for both post-intervention cross sectional data sets (t=1 and t=2). The estimates for assigning herself a fixed salary (A), keeping records of network contacts (B), and bookkeeping (C) are reported in Table 8. We find that it is very unlikely that attrition could eliminate the positive effect of the program on assignment of a fixed salary. For this to happen, the proportion of micro entrepreneurs that assigned themselves a fixed salary among attriters in the control group would need to be 3.1 to 5.3 times the proportion of those that we observe in the control group. Alternatively, even if no micro entrepreneurs in the attrited portion of the treatment group would have adopted this practice, the effect of the program would still be 3– 4 percent. The case is similar for keeping a record of potential network contacts: the proportion of micro entrepreneurs that adopt this practice among attriters in the control group would need to be twice what we observe in the data. We also find that even if no one among attriters in the treatment group adopts this practice, we would still get a 5 percent impact. All in all, we believe that it is very unlikely that those who attrited would be so substantially different that we would not find a positive impact of the program anymore. 28 However, our results for bookkeeping are not as robust. If those that attrited in the control group were 7 – 23 percent more likely to have a bookkeeping system (compared with those in the control group who remained in the sample), we would not be able to find a significant a positive effect any longer. Alternatively, this would also happen if the micro entrepreneurs in the attrited sample of the control group were 11% – 29 percent less likely to adopt this practice. Although this does not necessarily imply that the program did not have a positive impact on the adoption of bookkeeping practices, it could be the case that the positive effect that we previously found does not hold under some relatively plausible conditions. 4.3 Statistical Power An alternative explanation is that the program was successful and had a positive impact on entrepreneurs, but our intervention migt be underpowered to detect any statistically significant change. To explore this possibility, we re-estimated our power calculations to determine the minimum detectable effect from our experimental design. We focus on micro entrepreneurs’ adoption of practices and one outcome (yearly sales). Using the baseline information (i.e., 1,738 micro entrepreneurs: 703 in the treatment group and 1,035 in the control group), we simulate alternative hypothetical effects and determine the power to detect them at a 95 and 90 percent levels of significance. The results of these simulations are presented in Figure 2. The power of our intervention reaches 0.8 (at a 90 percent level of confidence) for a 3 percentagepoint increase in women’s self-assignment of a fixed salary and a 5 percentage-point increase the share of micro entrepreneurs who keep records for business contacts. Because the impacts reported in panels A and B of Table 3 are above these minimum detectable effects, we find significant results. Our power calculations also suggest that considering a power of 0.8 and a 90% level of significance we would only be able to find a significant impact on bookkeeping if adoption would increase by at least 5.5 percentage points. The impact on this practice seems to be around this level between 3.7 and 5.5 percentage points (in panel B of Table 3) and the coefficients are statistically significant in some specifications. In other words, it might be that the impact on bookkeeping is positive, yet we would not be able to significantly detect it. 29 Table 8: Sensitivity of Business Practice Adoption to Attrition A. Micro Entrepreneur Assigns Herself a Fixed Salary t=1 t=2 Adjustment (δ)1 Mean Effect 3Adjustment (δ)1 Mean Effect 3 Imputed Imputed Group2Group2 Impute missing observations in control 0% (po c)40.056 0.040∗∗∗ 0% (po c)40.047 0.059∗∗∗ (0.012) (0.012) 5% 0.102 0.033 ∗∗ 5% 0.097 0.047∗∗ (0.013) (0.013) 10% 0.153 0.024 ∗10% 0.144 0.035∗ (0.013) (0.013) 15% 0.203 0.015 15% 0.187 0.023 (0.014) (0.014) 20% 0.254 0.006 20% 0.195 0.010 (0.014) (0.015) No Effect ( ˜ δ=12%)50.175 0.020 No Effect ( ˜ δ=14%)50.245 0.025 (0.013) (0.014) (po c+˜ δ)/ po c63.10 (po c+˜ δ)/ po c65.25 Impute missing observations in treatment 0.0% (po T)40.099 0.040 ∗∗∗ 0.0% (po T)40.109 0.059∗∗∗ (0.012) (0.012) -2.5% 0.074 0.037 ∗∗∗ -2.5% 0.082 0.055∗∗∗ (0.012) (0.012) -5.0% 0.049 0.035∗∗∗ -5.0% 0.055 0.051∗∗∗ (0.012) (0.012) -7.5% 0.025 0.032 ∗∗∗ -7.5% 0.036 0.048∗∗∗ (0.012) (0.012) -10.0% 0.000 0.029 ∗∗ -10.0% 0.009 0.044∗∗∗ (0.012) (0.012) No Effect 5NA NA No Effect 5NA NA (po T−˜ δ)/ po T6NA (po T−˜ δ)/ po T6NA 1Simulated rates of adoption of po c+δ(or po T−δ) among attrited households in the control (or treament) group. 2We estimate a linear regression of the imputed variable on the treatment status and report the effect ˆ βand standard errors. 3While we impute a rate of adoption of po c+δ(for the control group) or po T−δ(for the treatment group), the rate in this column is not exactly the same because of rounding up when assigning 1s and 0s to the attrited observations. The rates reported here are: nA c(po c+δ)/nA cand nT c(po T−δ)/nA T, respectively. 5When δ=0, the coefficient of the adjusted estimate is the same as the one calculated on the non-attrited sample. 4˜ δis the minimum value of δfor which the effect of the program is not significant. 6This ratio indicates how much larger (smaller) the adoption of practices should be among the attrited micro entrepreneurs in the control (treatment) group with respect to the observed mean in the control (treatment) group to yield a statistically insignificant effect. 30 B. Keeps Records of Network t=1 t=2 Adjustment (δ)1 Mean Effect3Adjustment (δ)1 Mean Effect3 Imputed Imputed Group2Group2 Impute missing observations in control 0% (po c)40.258 0.089∗∗∗ 0% (po c)40.310 0.111∗∗∗ (0.022) (0.023) 10% 0.355 0.072∗∗∗ 10% 0.407 0.087∗∗∗ (0.022) (0.023) 20% 0.457 0.053∗∗ 20% 0.508 0.062∗∗∗ (0.023) (0.024) 30% 0.559 0.035 30% 0.609 0.037 (0.023) (0.024) 40% 0.656 0.018 40% 0.709 0.011 (0.023) (0.024) No Effect ( ˜ δ=29%)50.548 0.037 No Effect ( ˜ δ=30%)50.609 0.037 (0.023) 0.024 (po c+˜ δ)/ po c62.12 (po c+˜ δ)/ po c61.96 Impute missing observations in treatment 0% (po T)40.345 0.089∗∗∗ 0% (po T)40.421 0.111∗∗∗ (0.022) (0.023) -10% 0.250 0.078∗∗∗ -10% 0.325 0.095∗∗∗ (0.022) (0.023) -20% 0.155 0.066∗∗∗ -20% 0.219 0.078∗∗∗ (0.022) (0.023) -30% 0.048 0.054∗∗ -30% 0.123 0.063∗∗∗ (0.022) (0.023) -35% 0.000 0.048∗∗ -40% 0.026 0.047∗∗ (0.022) (0.023) No Effect5NA NA No Effect5NA NA (po T−˜ δ)/ po T6NA (po T−˜ δ)/ po T6NA 1Simulated rates of adoption of po c+δ(or po T−δ) among attrited households in the control (or treament) group. 2We estimate a linear regression of the imputed variable on the treatment status and report the effect ˆ βand standard errors. 3While we impute a rate of adoption of po c+δ(for the control group) or po T−δ(for the treatment group), the rate in this column is not exactly the same because of rounding up when assigning 1s and 0s to the attrited observations. The rates reported here are: nA c(po c+δ)/nA cand nT c(po T−δ)/nA T, respectively. 4When δ=0, the coefficient of the adjusted estimate is the same as the one calculated on the non-attrited sample. 5˜ δis the minimum value of δfor which the effect of the program is not significant. 6This ratio indicates how much larger (smaller) the adoption of practices should be among the attrited micro entrepreneurs in the control (treatment) group with respect to the observed mean in the control (treatment) group to yield a statistically insignificant effect. 31 C. Bookkeeping t=1 t=2 Adjustment (δ)1 Mean Effect 3Adjustment (δ)1 Mean Effect 3 Imputed Imputed Group2Group2 Impute missing observations in control 0% (po c)40.418 0.054∗∗ 0% (po c)40.447 0.047∗ (0.024) (0.024) 5% 0.463 0.047∗5% 0.498 0.035 (0.024) (0.024) 10% 0.514 0.038 10% 0.549 0.022 (0.024) (0.024) 15% 0.565 0.029 15% 0.599 0.010 (0.024) (0.024) 20% 0.616 0.021 20% 0.650 -0.003 (0.024) (0.024) No Effect ( ˜ δ=10%)50.514 0.038 No Effect ( ˜ δ=3%)50.479 0.039 (0.024) (0.024) (po c+˜ δ)/ po c61.23 (po c+˜ δ)/ po c61.07 Impute missing observations in treatment 0% (po T)40.469 0.054∗∗ 0% (po T)40.500 0.047∗ (0.024) (0.024) -5% 0.420 0.049∗∗ -5% 0.445 0.039 (0.024) (0.024) -10% 0.370 0.043∗-10% 0.400 0.032 (0.024) (0.024) -15% 0.321 0.037 -15% 0.345 0.023 (0.024) (0.024) -20% 0.272 0.032 -20% 0.300 0.016 (0.024) (0.024) No Effect ( ˜ δ=-13%)50.333 0.039 No Effect ( ˜ δ=-5%)50.445 0.039 (0.024) 0.024 (po T−˜ δ)/ po T60.71 (po T−˜ δ)/ po T60.89 1Simulated rates of adoption of po c+δ(or po T−δ) among attrited households in the control (or treament) group. 2We estimate a linear regression of the imputed variable on the treatment status and report the effect ˆ βand standard errors. 3While we impute a rate of adoption of po c+δ(for the control group) or po T−δ(for the treatment group), the rate in this column is not exactly the same because of rounding up when assigning 1s and 0s to the attrited observations. The rates reported here are: nA c(po c+δ)/nA cand nT c(po T−δ)/nA T, respectively. 4When δ=0, the coefficient of the adjusted estimate is the same as the one calculated on the non-attrited sample. 5˜ δis the minimum value of δfor which the effect of the program is not significant. 6This ratio indicates how much larger (smaller) the adoption of practices should be among the attrited micro entrepreneurs in the control (treatment) group with respect to the observed mean in the control (treatment) group to yield a statistically insignificant effect. 32 We reach a power of 0.8 (with a 90 percent confidence level) with an increase of about 10 percent in sales. Although this minimum detectable effect is somewhat larger, our results in Table 4 show that the impact was rather small and, in most cases, even negative. Our coefficients suggest that, if any, the effect of the program on sales was between -4 and 1 percent. Because of the magnitudes of our estimates, we do not believe that these results are driven by statistical power concerns. 4.4 Quality of the Training and Practice Adoption A possible explanation for our results is that the training was not useful for the participants. This could happen if the training did not explain how to implement their proposed business practices clearly enough or if it was unable to convey their usefulness. It could also be the case that even when participants did understand what they were taught the contents of the training were not helpful for their businesses or were difficult to implement. It does not seem the case that micro entrepreneurs did not understand the training. In general, program participants seemed highly satisfied with the training they received. At endline, we asked the women in the treatment group (that had participated in the program about a year before) to rate their satisfaction with the training from 0 to 10 (with 0 being highly dissatisfied and 10 being highly satisfied). On average they assigned a score of 7.8 and 68 percent of them rated the training 8 or more. Albeit subjective, this score does not support the idea that participants did not value the training or that they did not find it useful. If participants understood the training and considered there were any benefits from the contents they were taught, it might be that the contents of the programs were not practical for their businesses. This could explain why the the rate of practice adoption was relatively small: adoption rates of practices were relatively higher among those who received the training, but the increases were far from substantial in absolute terms. Table 3 shows that those who received the training were 4-6 percentage points more likely to assign themselves a fixed salary. This is an increase of 90–100 percent with respect to the baseline rate. Similarly, the share of micro entrepreneurs who kept a record of their business contacts was 6–11 percentage points larger 33 Figure 2: Power Calculations, Selected Variables1 (a) Assigning Self Fixed Salary2 .2 .4 .6 .8 1 Power 02468 Effect size (percentage points) P−value=5% P−value=10% (b) Keeping Record of Contacts2 0 .2 .4 .6 .8 1 Power 02468 Effect size (percentage points) P−value=5% P−value=10% (c) Bookkeeping2 0 .2 .4 .6 .8 1 Power 02468 Effect size (percentage points) P−value=5% P−value=10% (d) Sales3 0 .2 .4 .6 .8 1 Power 0 10 20 30 Effect size (percentage change) P−value=5% P−value=10% 1The power calculations determine the proportion of times that we would reject the null hypothesis H0: Effect=0 (with p-values of 5 percent and 10 percent) for each effect size. Using the baseline means and standard deviations, we perform 1,000 simulations for each effect size. The simulations are based on our effective sample size at baseline: 1,738 micro entrepreneurs (703 in the treatment group and 1,035 in the control group). 2Effect size is expressed in percentage points for discrete variables. 3Effect size is expressed in percentage change for continuous variables. 34 (which represents a difference of 21–39 percent when compared with their initial situation) and the proportion of those who implemented bookkeeping was 4–6 percentage points higher (10– 14 percent increase with respect to baseline) in the treatment group. Although all these changes were significant relative to the baseline levels, the changes are rather small in absolute terms. Hence, it is not completely surprising that these relatively mild changes would not necessarily translate into large improvements in the treatment group’s average business performance, household outcomes, or female empowerment. In other words, if we think of these practices as the drivers of performance outcomes, then only women who marginally adopt these techniques would experience improvements, and these increases would be swamped on average. We simulate how large the adoption rates of business practices would need to be to be able to detect a statistical impact on sales. Figure 3 estimates the statistical power of our intervention with hypothetical adoption rates and effects on sales. Although we do not estimate the power for any particular practice, we assume a general practice that would be adopted by 5, 10, 20, 30, 40, and 50 percent of the micro entrepreneurs in the treatment group. We suppose that the adoption of such practice would boost business sales of those who implement it by 1–90 percent of business sales (i.e., sales increases are generated by the adoption of the practice, so those who did not implement it would not experience such benefit). Assuming a significance level of 90 percent (i.e., a p-value of 10 percent), we find that any changes in sales values would be very difficult to detect with low levels of adoption. With a relatively high adoption rate of 50 percent, our intervention reaches a power of 0.8 when the effect of practices is about 20 percent of sales. With intermediate rates of adoption of 20–40 percent, practices would need to have an impact of 25–60 percent of income to reach that level of statistical power. When adoption is as low as 5–10 percent as in our intervention a power of 0.8 is not reached even if business sales double among adopters. Among micro entrepreneurs in the treatment group who did not adopt each business practice in the endline, we did ask why they did not do so. This allows us to present some qualitative evidence and explore why the training did not yield larger adoption rates. The results of this analysis are presented in Figure 4. In all the cases, there were few women who did not perceive any benefits from the business practices taught by the program or that did not understand the 35 Figure 3: Adoption Rates and Power to Detect Increases in Sales Values1 0 .2 .4 .6 .8 1 Power 0 20 40 60 80 100 Effect 5% Adoption Rate 10% Adoption Rate 20% Adoption Rate 30% Adoption Rate 40% Adoption Rate 50% Adoption Rate Note: Each simulation assumes the following: (a) X percent of micro entrepreneurs in the treatment group adopt a business practice; (b) the adoption of this practice leads to a 1%–50 percent increase in sales; (c) only micro entrepreneurs who adopt a practice experience sales increases. Potential values of X=5%, 10%, 20%,.. , 50%. We assume a p-value of 10% to determine the statistical power in these simulations. training (between 3.7 and 5.3 percent). There are some specific reasons for not assigning themselves a fixed salary: entrepreneurs might prefer to take a percentage of profits (rather than a fixed sum) or they might just want to take all business profits. Also, there are some particular reasons to avoid keeping a client list (arguably they rely on other informal mechanisms, although they were not specified).21 However, in all cases, lack of time was an important reason not to adopt the business practices recommended by the program (between 29 and 69 percent, depending on the practice). This is an interesting finding, as it suggests that micro entrepreneurs value their time very highly and may be reluctant to pay the opportunity costs associated with taking on a particular practice. It may simply be the case that the training did not adequately convey the benefits associated with 21More than 20 percent of women mentioned other reasons not to adopt each practice. The “Other Reasons” category includes many distinct answers. Some of the reasons not to assign themselves a fixed salary were: particular and unexpected economic shocks, businesses that were not profitable enough, husbands taking away all business profits, neglect, etc. Examples of reasons not to keep a record of business contacts were: blurry vision, clients / suppliers coming anyway, single buyer / seller, etc. Other examples of reasons not to keep a record of sales and expenditures were: low levels of sales and expenditures that did not require bookkeeping, reluctance to know when business declines because it would demoralize them, entrepreneurs who forget to do so by the end of the day, etc. 36 Figure 4: Reasons for Not Adopting Business Practices (a) Reasons for Not Monitoring Cash Flow Other (23.5%) Lack of time (68.5%) Lack of time (68.5%) Did not understand training (4.0%) Did not perceive any benefit (0.9%) Problems with Arithmetic (3.1%) (b) Reasons for Not Maintaining a Client List Did not understand the training (3.3%) Other (19.6%) Lack of time (61.6%) Did not perceive any benefit (2.2%) Has other method for maintaning contacts (13.4%) (c) Reasons for Not Assigning Self a Fixed Salary Lack of Time (28.7%) Other (25.8%) Did not understand the training (2.2%) Did not perceive any benefit (1.5%) Problems with Arithmetic (1.1%) Prefers to take a percentage of profits (16.4%) Prefers to take all profits (24.3%) 37 be uncorrelated with εt. However, in general, it is difficult to identify appropriate variables for this estimation, and there are no variables in our data that would credibly meet any necessary exclusion restrictions. Because we have a rich set of variables, we will instead suppose that our attrition exhibits selection on observables. This allows for the set of variables ztto be correlated with both At and yt. In particular, Fitzgerald, Gottschalk, and Moffitt (1998) state that, under selection on observables, this model is identified if εtand ηtare uncorrelated. This will hold if Pr(At= 0|yt,xt,zt) = Pr(At=0|xt,zt): the probability of attrition is independent of the dependent variable yt, once we account for xtand zt. In practical terms, our approach is to estimate weights wu=Pr(At=0|xt,zt)and normalize them by wr=Pr(At=0|xt), where xtis the treatment indicator and ztis a set of variables related to attrition. Our variables include: sector of the micro enterprise, sales at baseline, age of business, whether the micro entrepreneur had received previous business training, demographic composition of the household, whether the micro entrepreneur is the household head, her years of education, whether she is a migrant, her marital status, the per capita expenditure of her household, and a set of location (district) indicators. We calculate wuand wras the predicted probabilities from probit models. The IPW is given by the inverse of the normalized weight: W=wr wu. Then we can estimate 7 through weighted least squares with Was the regression weights. Intuitively, this estimation places higher weights on observations that share similar initial characteristics as those who attrit later on, and lower weights on those who are likely to remain in the sample. We report estimates (including marginal effects) of the probit models of attrition in t=1 and t=2 in Table 10. The probability of remaining in the sample increases by 5–9 percent if a micro entrepreneur is in the treatment group. Participants are also somewhat more likely to remain in the sample when they are in the retail and manufacturing sectors (relative to the service sector), when there are more working-age members in the household, when they are the head of their household, and when they are less educated. We use the fitted values of this regression to estimate weights ˆ W1and ˆ W2for t=1 and t=2, respectively. These are used as weights in linear 44 Table 10: Probability of Remaining in the Sample (Probit Model) Variable t=1 t=2 (1) (2) (3) (4) Probit Marginal Probit Marginal Coeff Effects1Coeff Effects1 Treatment 0.3613*** 0.0531*** 0.4476*** 0.0888*** (0.094) (0.013) (0.085) (0.016) Retail 20.0815 0.0126 0.1004 0.0211 (0.104) (0.016) (0.094) (0.020) Manufacturing 20.0702 0.0104 0.2306* 0.0434* (0.145) (0.021) (0.138) (0.023) Yearly sales (S/. 10,000s) -0.0124 -0.0019 0.0028 0.0006 at baseline (0.011) (0.002) (0.010) (0.002) Age of business (months, X100) 0.0163 0.0025 -0.0183 -0.0038 (0.050) (0.008) (0.044) (0.009) Micro entrepreneur had received 0.3671** 0.0454*** 0.2452 0.0452* previous training (0.180) (0.017) (0.151) (0.024) Number of members 0-14 y.o. 0.0660 0.0101 0.0078 0.0016 at baseline (0.046) (0.007) (0.039) (0.008) Number of members 15-60 y.o. 0.0843** 0.0129** 0.0403 0.0084 at baseline (0.034) (0.005) (0.030) (0.006) Number of members 61+ y.o. 0.0396 0.0061 0.1130 0.0235 at baseline (0.074) (0.011) (0.072) (0.015) Household head 0.3423*** 0.0510*** 0.1282 0.0264 (0.108) (0.016) (0.095) (0.019) Number of years of education -0.0108 -0.0017 -0.0300** -0.0062** (0.015) (0.002) (0.013) (0.003) Micro entrepreneur is a migrant 0.0742 0.0113 -0.0871 -0.0182 (not born in province) (0.092) (0.014) (0.083) (0.017) Married 0.1396 0.0219 0.1453 0.0308 (0.104) (0.017) (0.093) (0.020) HH per capita monthly 0.0058 0.0009 -0.0250 -0.0052 expenditure (S/. 100s) at baseline (0.020) (0.003) (0.016) (0.003) Constant 0.9476*** 1.3434*** (0.311) (0.284) District dummies (29) YES YES Observations 1,738 1738 Pseudo R-Squared 0.1045 0.1016 Note: The dependent variable of the regression takes a value of 0 if the micro entrepreneur attrited from the sample, and 1 otherwise. 1Marginal effects of discrete variables are calculated as the discrete difference in Φ(D=1)-Φ(D=0). 2Base category is services. Standard errors in parentheses. Significance levels denoted by: *** p<0.01, ** p<0.05, * p<0.1. cross sectional regressions: Yit =βTreati+εit for t=1 and t=2 and for each of our outcomes of interest (i.e., adoption of business practices, business performance, and household outcomes). Estimates ˆ W1and ˆ W2are predicted values from a probit model, and have their own error structure. They could be thought of as a parallel to the instrumental variables or sample selection modelsas “first stage estimates.” Therefore, we need to adjust the standard errors of the linear cross sectional regressions (the “second stage”) to incorporate the error structure of the first stage. To account for this, we bootstrap the estimation of both stages 1,000 times. Our estimates adjusted for the IPWs are reported in Table 11. With two exceptions (assignment of own salary and payroll), all the impacts are smaller than the ones reported in Tables 3, 4, and 5. Table 11: Cross Sectional Estimates (adjusted for IPW) Dependent Variable t=1 t=2 Own Salary 0.041*** 0.061*** (0.014) (0.014) Network 0.083*** 0.104*** (0.025) (0.027) Bookkeeping 0.050* 0.041 (0.027) (0.028) Yearly Sales -2,189.6 -2,390.6 (3,907.9) (3,149.4) Payroll 441.6 -198.2 (521.2) (234.5) Other Business Expenses -2,002.2 -2,330.7 (3,145.7) (2,410.0) Monthly HH per capita expenditure 5.23 3.76 (13.0) (10.5) Women Household Empowerment1-0.028 -0.053 (0.038) (0.041) N 1,738 Note: Bootstrapped standard errors (1,000 replications) to account for the error structure of the weights ( ˆ W) and the linear regression (second stage). Standard errors in parentheses. Significance levels denoted by: *** p<0.01, ** p<0.05, * p<0.1. 1We estimate a seemingly unrelated regressions system of decisions: Dig = θgTreati+εiwith decisions g=1,...,G, where each regression is weighted by W. We estimate the overall impact on this family of outcomes through τ= 1 GΣG g=1 θg σg.σgis also estimated using Wweights. 46 References Alderman, H., Behrman, J.R., Kohler, H.P., Maluccio, J.A., and Watkins, S.C. (2001). “Attrition in Longitudinal Household Survey Data.” Demographic Research 5, 79–124. Angelelli, P., Moudry, R., and Llisterri, J.J. (2006). “Institutional Capacities for Small Business Policy Development in Latin America and the Caribbean.” Sustainable Development Department Technical Paper Series, Inter-American Development Bank, Washington, DC. Ayyagari, M., Demirguc-Kunt, A., and Maksimovic, V. (2011). “Small vs. Young Firms across the World: Contribution to Employment, Job Creation, and Growth.” Policy Research Working Paper 5631, World Bank, Washington, DC. Banerjee, A., Duflo, E., Glennerster, R., and Kinnan, C. (2013). “The Miracle of Microfinance? Evidence from a Randomized Evaluation.” Working Paper 18950, National Bureau of Economic Research, Cambridge, MA. Berge, L.I.O., Bjorvatn, K., and Tungodden, B. (2011). “Human and Financial Capital for Microenterprise Development: Evidence from a Lab and Field Experiment.” Discussion Paper 1/2011, NHH Dept. of Economics, Bergen, Norway. Bland, J.M. and Altman, D.G. (1995). “Multiple Significance Tests: the Bonferroni Method.” BMJ: British Medical Journal 310(6973), 170. Bloom, N. and Reenen, J.V. (2010). “Why do Management Practices Differ across Firms and Countries?” Journal of Economic Perspectives 24(1), 203–224. Bruhn, M., Karlan, D., and Schoar, A. (2010). “What Capital is missing in Developing Countries?” American Economic Review 100(2), 629–33. Calderón, G., Cunha, J.M., and Giorgi, G.D. (2013). “Business Literacy and Development: Evidence from a Randomized Controlled Trial in Rural Mexico.” Working Paper 19740, National Bureau of Economic Research, Cambridge, MA. Cho, Y. and Honorati, M. (2013). “Entrepreneurship Programs in Developing Countries: A Meta Regression Analysis.” Policy Research Working Paper 6402, World Bank, Washington, DC. de Mel, S., McKenzie, D., and Woodruff, C. (2008). “Returns to Capital in Microenterprises: Evidence from a Field Experiment.” The Quarterly Journal of Economics 123(4), 1329–1372. de Mel, S., McKenzie, D., and Woodruff, C. (2014). “Business training and female enterprise start-up, growth, and dynamics: Experimental evidence from Sri Lanka.” Journal of Development Economics 106, 199–210. De Soto, H. (1989). The Other Path: The Economic Answer to Terrorism. Harper and Row Publishers, Inc, New York, NY. Diario Correo (2012). “Incendio consume cientos de puestos en mercado de Villa El Salvador.” Online newspaper version, published February 17, 2012: http://tinyurl.com/pnj9a4g. 47 Drexler, A., Fischer, G., and Schoar, A. (2014). “Keeping it Simple: Financial Literacy and Rules of Thumb.” American Economic Journal: Applied Economics 6(2), 1–31. Duflo, E., Glennerster, R., and Kremer, M. (2007). “Using Randomization in Development Economics Research: A Toolkit.” In T.P. Schultz and J.A. Strauss, eds., “Handbook of Development Economics,” Elsevier, volume 4, chapter 61, pp. 3895–3962. Evans, D.S. and Jovanovic, B. (1989). “An Estimated Model of Entrepreneurial Choice under Liquidity Constraints.” Journal of Political Economy 47(4), 808–827. 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, 211–226. Field, E., Jayachandran, S., and Pande, R. (2010). “Do Traditional Institutions Constrain Female Entrepreneurship? A Field Experiment on Business Training in India.” American Economic Review 100(2), 125–29. Fitzgerald, J., Gottschalk, P., and Moffitt, R. (1998). “An Analysis of Sample Attrition in Panel Data.” The Journal of Human Resources 33(2), 251–299. Giné, X. and Mansuri, G. (2011). “Money or Ideas? A Field Experiment on Constraints to Entrepreneurship in Rural Pakistan.” Unpublished Manuscript, World Bank, Washington, DC. Heckman, J. (1979). “Sample Selection Basis as a Specification Error.” Econometrica 47(1), 153– 161. Heckman, J.J., Lalonde, R.J., and Smith, J.A. (1999). “The Economics and Econometrics of Active Labor Market Programs.” In O. Ashenfelter and D. Card, eds., “Handbook of Labor Economics,” Elsevier, volume 3 of Handbook of Labor Economics, chapter 31, pp. 1865–2097. Jensen, R. and Oster, E. (2009). “The Power of TV: Cable Television and Women’s Status in India.” The Quarterly Journal of Economics 124(3), 1057–1094. Karlan, D., Knight, R., and Udry, C. (2012). “Hoping to Win, Expected to Lose: Theory and Lessons on Microenterprise Development.” Working Paper 312, Center for Global Development, Washington, DC. Karlan, D. and Valdivia, M. (2011). “Teaching Entrepreneurship: Impact of Business Training on Microfinance Clients and Institutions.” The Review of Economics and Statistics 93(2), 510– 527. Kling, J.R. and Liebman, J.B. (2004). “Experimental Analysis of Neighborhood Effects on Youth.” Working Paper 483, Princeton University, Industrial Relations Section (IRS), Princeton, NJ. Kling, J.R., Liebman, J.B., and Katz, L.F. (2007). “Experimental analysis of neighborhood effects.” Econometrica 75(1), 83–119. La Ferrara, E., Chong, A., and Duryea, S. (2012). “Soap Operas and Fertility: Evidence from Brazil.” American Economic Journal: Applied Economics 4(4), 1–31. 48 Li, Y. and Rama, M. (2012). “Firm Dynamics, Productivity Growth and Job Creation in Developing Countries: The Role of Micro- and Small Enterprises.” Background Paper for the World Development Report 2013, World Bank, Washington, DC. McKenzie, D. (2012). “Beyond baseline and follow-up: The case for more T in experiments.” Journal of Development Economics 99, 210–221. McKenzie, D. and Woodruff, C. (2014). “What are We Learning from Business Training and Entrepreneurship Evaluations around the Developing World?” World Bank Research Observer 29(1), 48 – 82. Restuccia, D. and Rogerson, R. (2008). “Policy Distortions and Aggregate Productivity with Heterogeneous Establishments.” Review of Economic Dynamics 11(4), 707–720. Sankoh, A., Huque, M., and Dubey, S. (1997). “Some Comments on Frequently used Multiple Endpoint Adjustment Methods in Clinical Trials.” Statistics in M 16, 2529–2542. Schochet, P.Z. (2008). “Guidelines for Multiple Testing in Impact Evaluations of Educational Interventions.” Unpublished manuscript, Mathematica Policy Research Inc., Princeton, NJ. Valdivia, M. (2011). “Training or Technical Assistance? A Field Experiment to Learn What Works to Increase Managerial Capital for Female Microentrepreneurs.” Unpublished manuscript, Lima, Peru. Wooldridge, J.M. (2002). Econometric Analysis of Cross-sectional and Panel Data. MIT Press, Cambrdige, MA. Wooldridge, J.M. (2007). “Inverse Probability Weighted Estimation for General Missing Data Problems.” Journal of Econometrics 141(2), 1281 – 1301. World Bank (2012). World Development Report 2013: Jobs. World Bank, Washington, DC. 49