L52 WD-2019 Huis Impacts of te gender and entrepreneurship female microfinance in Vietnam.pdf

World Development 120 (2019) 46–61 Contents lists available at ScienceDirect World Development...

World Development 120 (2019) 46–61 Contents lists available at ScienceDirect World Development journal homepage: www.elsevier.com/locate/worlddev Impacts of the Gender and Entrepreneurship Together Ahead (GET Ahead) training on empowerment of female microfinance borrowers in Northern Vietnam Marloes Huis a, Robert Lensink b,⇑, Nhung Vu c, Nina Hansen d a Department of Social Psychology, University of Groningen, Grote Kruisstraat 2/1, 9712TS Groningen, The Netherlands b Department of Economics, Econometrics and Finance, University of Groningen, Development Economics Group, Wageningen University, Nettelbosje 2, 9747 AE Groningen, The Netherlands c Faculty of Economics and Management, Vietnamese – German University, Thu Dau Mot City, Binh Duong Province, Vietnam d Department of Social Psychology, University of Groningen, Grote Kruisstraat 2/1 9712TS, Groningen, The Netherlands a r t i c l e i n f o a b s t r a c t Article history: Across the world the Gender and Entrepreneurship Together (GET Ahead) training originally developed Accepted 1 April 2019 by the International Labour Organization has been implemented to improve business outcomes and Available online 11 April 2019 enhance women’s empowerment. This randomized controlled trial is the first rigorous attempt to exam- ine the impact of the GET Ahead training on women’s empowerment. We focus on the impact of offering Keywords: this training to female microfinance borrowers of TYM, the largest microfinance organization in North Women empowerment Vietnam. A major contribution of this study is that it focuses on different dimensions of women’s empow- GET Ahead training erment: (1) personal empowerment, measured by control beliefs, and (2) relational empowerment, mea- Randomized controlled trial Vietnam sured by relational friction and intra-household decision making power. This study also stands out in that Asia we explicitly study whether involving husbands affects the impact of the training. We find that the GET Ahead training improves women’s empowerment on all three aspects: increased control beliefs and intra-household decision making power (only on larger expenditures), and decreased relational friction. However, the results on relational frictions should be taken with care due to possible underreporting. Moreover, in general, we find no additional impacts of inviting husbands to the training. Finally, our results suggest that it takes some time before the training starts to improve women’s empowerment. We observe no short-term but only mid-term effects from before the training to 12 months after the training. Ó 2019 Elsevier Ltd. All rights reserved. 1. Introduction opment. Conversely, enhancing gender equity could encourage economic development. This assumption is supported by scholars Throughout human history, as societies grew and expanded, who stress the importance of womeńs empowerment for economic men maintained positions of supremacy (Brown, 1991). Today development (e.g., Duflo, 2012) and provide evidence that the sus- women worldwide still face disadvantages in several domains, tained growth in economic development observed in modern soci- including education, employment, political representation, and eties can be attributed in part to improvements in gender equity intra-household oppression due to social norms that encourage (Diebolt & Perrin, 2013). Even the former Secretary General of such status differences (e.g., Armendáriz & Morduch, 2010). Gen- the United Nations Kofi Annan has argued that gender equity is a der inequity is especially prevalent in nations with lower incomes prerequisite to achieving other development goals (UNDP, 2005). (Sanyal, 2009). This is also supported by statistics of the Gender Whether encouraging women’s empowerment directly results in Inequality Index which indicates the extent to which national durable economic development remains an open question, but achievements are eroded by gender inequity [United Nations the two concepts clearly are interrelated. It seems plausible that Development Program (UNDP), 2014]. The persistence of gender empowering women could change existing decision-making pat- inequity appears to hinder the advancement of sustainable devel- terns and thereby impact development (e.g., Duflo, 2012). Different programs have been developed and implemented around the ⇑ Corresponding author. world to improve the position of women in society such as E-mail addresses: [email protected] (M. Huis), [email protected] (R. Lensink), strengthening women’s political participation or decreasing vio- [email protected] (N. Vu), [email protected] (N. Hansen). lence against women (for examples see UNDP, 2017). Another https://doi.org/10.1016/j.worlddev.2019.04.001 0305-750X/Ó 2019 Elsevier Ltd. All rights reserved.  M. Huis et al. / World Development 120 (2019) 46–61 47 prominent approach is to offer microfinance services to women indicated by (2) less relational friction and (3) more intra- (for a recent debate see Garikipati, Guérin, Johnson, & Szafarz, household decision making power. In general, no differential 2017). impact of inviting the husbands of female microfinance borrowers The current research reports results of a randomized control to the training compared to inviting female microfinance borrow- trial examining the impacts of the Gender and Entrepreneurship ers alone was observed. All impacts increased over time from Together – GET Ahead for women in Enterprise – training offered before the training to 12 months after women participated in to a large group of female microfinance borrowers in Vietnam, training. where microfinance is widespread (Asian Development Bank The next section contains a short overview of the theoretical [ADB], 2014). The study’s main contributions are threefold. First, background and relevant literature. We then describe the study we examined the short- and medium-term impacts of the GET context in Section 3 and the methods in Section 4, including bal- Ahead training on different aspects of women’s empowerment. ancing tests, attrition analyses, and uptake. Section 5 contains The program’s aim was to help women to develop their entrepre- the estimations, and Section 6 presents the results, and possible neurial skills and receive social support through groups, networks, mechanisms linking access to training to women’s empowerment. and institutions to promote women’s empowerment. In Section 7 we review the relevant findings, and discuss the impli- The current training was a culturally adapted and shortened cations and limitations of the current research and suggestions for version of the GET-Ahead for women in enterprise training devel- future research. oped by the International Labour Organisation (ILO; Bauer et al., 2004), which has been implemented in more than 18 countries around the world. This training differs from conventional business 2. Theoretical background training by highlighting essential entrepreneurial skills from a gen- dered perspective (Bauer et al., 2004).1 Whereas many NGO’s offer 2.1. Gender relations and women’s empowerment combined gender and business programs, to the best of our knowl- edge previous research has only focussed on examining the effects Societies are organized as group-based social hierarchies in of business training on economic outcomes (e.g., McKenzie & which men on average hold more power than women (Sidanius Woodruff, 2014). To the best of our knowledge, this study is the first & Pratto, 1999). According to the social dominance theory to investigate the impact of the GET Ahead training on various (Sidanius & Pratto, 1999) the relatively stable inequity between aspects of women’s empowerment.2 men and women is a consequence of persistent unequal access Second, previous research in the field of microfinance services to resources. Expanding on this idea, the gendered power model has mainly focused on women’s intra-household decision making (Pratto & Walker, 2004; Pratto, Lee, Tan, & Pitpitan, 2008) suggests power as the main indicator of womeńs empowerment (for an that gender inequity has four bases: strength (e.g., physical and overview see Duvendack, Palmer-Jones, & Vaessen, 2014). In this emotional power), access to resources (e.g., control over financial paper we extend this research by studying women’s empower- resources), social obligations (e.g., distribution of household ment more broadly by assessing personal empowerment with con- responsibilities), and gender ideology (e.g., culturally shared trol beliefs and relational empowerment with relational friction as beliefs that legitimize power asymmetries). Men compared to well as intra-household decision making power (Huis, Hansen, women hold more power on all four bases. The model assumes Otten, & Lensink, 2017). dynamic relations between these four bases. Third, though microfinance services are available to both men Whether women gain more power in the different bases of and women, most borrowers are women (Kaur, 2014). Supporters power is related to the concept of empowerment. Women’s argue that offering access to microfinance services to women is a empowerment is defined as women’s ability to make strategic life promising means to empower women and increase gender equity choices in settings in which this ability was previously denied (e.g., Armendáriz & Morduch, 2010; Kulkarni, 2011). However, them (Kabeer, 1999). Empowerment is thus a process of change some research also provides evidence for negative effects for that enables women with limited choice, freedom, and power to female borrowers. More precisely, qualitative research suggested gain and leverage power, which enhances their ability to exercise that female microfinance borrowers experienced more intimate choice and freedom in ways that enhances their well-being (e.g., partner violence (Rahman, 1999). Other research showed that Ganle, Afriyie, & Segbefia, 2015; Kabeer, 1999). Accordingly, men felt excluded from these programs because they were mainly women’s empowerment results from both a change away from a offered to women (Rahman, Hoque, & Makinoda, 2011). To avoid condition of disempowerment, and towards a sense of personal this potential source of friction (i.e., exclusion of men), we ran- agency or choice (Malhotra, Schuler, & Boender, 2002). domly selected a group of female microfinance borrowers whose Comparing different literatures, women’s empowerment has husbands were also invited to participate in the training sessions. been studied as a multifaceted concept which includes a variety In sum, we compared female borrowers who were not invited with of aspects such as access to resources, psychological feelings of effi- those, who were invited to the GET Ahead training alone, and those cacy, decision making power (e.g., Ali & Hatta, 2012; Klein, 2014), invited together with their husbands. freedom of movement and gender beliefs (Arestoff & Djemai, Overall, the results indicated that offering the GET Ahead train- 2016), feelings of agency and social networks (Hansen, 2015), bar- ing to female microfinance borrowers increased women’s empow- gaining power (e.g., Mishra & Abdoul, 2016), and women’s labor erment on three aspects: (1) more personal empowerment as force participation (e.g., de Jong, Smits, & Longwe, 2017). These dif- indicated by control beliefs and more relational empowerment as ferent aspects can be categorized in three different dimensions of women’s empowerment, namely personal, relational, and societal 1 empowerment (for a review see Huis et al., 2017). Personal empow- The reported results are part of a larger research project which tested the impacts of the same GET Ahead training in Vietnam. Bulte et al. (2017) studied the impact of erment encompasses aspects of empowerment related to the indi- the training on business outcomes, Bulte et al. (2016) provide a preliminary analysis vidual self, such as self-efficacy, control beliefs, or feelings of of its impact on women’s intra-household decision making power in a behavioral personal agency. Relational empowerment is indicated by women’s game among a small subsample, and Bulte and Lensink (2019) focus on the impact on level of empowerment with respect to their relationships such as physical assault using the item count technique. 2 Important to note, in this paper we focus specifically on the impact of training in their marriage, household, or family, and includes aspects such offered to female microfinanceborrowers on women’s empowerment and do not as bargaining power as well as aspects of relationship quality examine the impact of offering microfinance credit. (e.g., freedom of movement, intimate partner violence). Societal 48 M. Huis et al. / World Development 120 (2019) 46–61 empowerment, finally, refers to women’s position in society on a results at least suggest that business training may not have the macro level and is indicated by aspects such as women’s labor intended effects for women’s empowerment. force participation or the number of women leading a business. Two studies reported positive impacts of training female entre- In the current paper we focussed on the personal and relational preneurs on different aspects of women’s empowerment. First, level as we studied an intervention at the micro level and involved participating in intensive technical and social awareness training, husbands as well. We assessed personal empowerment with con- but not only receiving credit, was positively related to intra- trol beliefs and relational empowerment in terms of relational fric- household decision making power among female microfinance tion as well as intra-household decision making power about borrowers in South India (Holvoet, 2005). Furthermore, longer expenditures in the daily domain and larger expenditures. retention in a lending group and more intensive training changed Our study focusses on the impact of a training offered to micro- norm-guided male decision making to more joint and female credit borrowers on women’s empowerment, and thus does not decision-making. Second, female members of microfinance institu- deal with the women empowerment effects of microcredit. Yet it tions who participated in diverse trainings in soft skills, business, is relevant to briefly discuss the literature on microcredit and and vocational skills reported higher levels of personal control women empowerment. Proponents of microfinance services beliefs and larger networks in Sri Lanka (Hansen, 2015). Berge assume that offering women access to these services (i.e., micro- and colleagues (2014) suggest that trainings for female entrepre- credit, savings) should increase women’s empowerment neurs should increase women’s motivation to manage business (Armendáriz & Morduch, 2010; Duvendack et al., 2014). While activities and consider external constraints (i.e., household respon- there is abundant research on the impact of offering microcredit sibilities) to increase the effectiveness of training among women. on business outcomes, the relation between offering access to In the current study we examine the impact of a training tailored microfinance services and empowerment remains unclear (e.g., to the needs of women by highlighting essential entrepreneurial Kabeer, 2001; Weber & Ahmad, 2014). For example, on the per- skills from a gendered perspective (Bauer et al., 2004). sonal level of empowerment, women participating in microcredit Based on the brief literature survey above we expect that invit- interventions tend to exhibit more financial security and self- ing female microfinance borrowers to this training should increase confidence compared to a control group (Kim et al., 2009). How- women’s internal control beliefs as well as their intra-household ever, several recent studies report no relation between access to decision making power compared to a control group of female microcredit and women’s decision-making power within their microfinance borrowers (no training). However, as mentioned households, which is an indicator of relational empowerment above, microfinance services may also have negative impacts (Banerjee, Duflo, Glennerster, & Kinnan, 2015; Crépon, Devoto, (e.g., if men feel excluded, Rahman, 1999). For example, research Duflo, & Pariente, 2015; Tarozzi, Desai, & Johnson, 2015). Other showed an increased risk of intimate partner violence if women research reports that microloans procured by women may even profit from education which increased more liberal ideas about result in women’s disempowerment, possibly because the loans gender roles (Jewkes, Levin, & Penn-Kekana, 2002). In general, are used for household assets over which women hold no control women’s transgression of conservative gender roles or challenges (e.g., Garikipati, 2008; Johnson, 2004). Interestingly, the type of to male privilege is related to increased relational friction loan acquired by women also seems to influence empowerment (Jewkes, 2002). This implies that inviting female microfinance bor- outcomes. Previous research suggests that small loans for day-to- rowers alone to the GET Ahead training should increase relational day expenses, which are obtained via informal networks and are friction. perceived as socially dishonourable, are more likely than planned loans to increase women’s bargaining power. It is reasoned that 2.3. The additional impact of inviting husbands to attend the training if women undergo the social humiliation associated with these loans, they are met with more appreciation and thus bargaining Inequality between members of an advantaged and a disadvan- power within the household (Garikipati, Agier, Guérin, & Szafarz, taged group, such as men and women, is shaped by their relation- 2017). The current research specifically examines the impact of ship (Prentice & Shelton, 2012). This implies that to achieve change training – not credit – offered to female microfinance borrowers, towards more equity, both parties need to be involved (for a who have access to microcredit and thus are unlikely to be credit review, see Dixon, Levine, Reicher, & Durrheim, 2012; Prentice & constrained. Shelton, 2012). More precisely, men would need to understand the perspective of women and be willing to change, whereas 2.2. The impact of training on women’s empowerment women would need to become aware of their disadvantaged posi- tion and actively strive for more equity. Research conducted in Over the years microfinance services have evolved to not only India has also stressed the need to engage both men and women offer microcredit but also different financial and non-financial ser- to achieve gender equity (Chowdhury & Patnaik, 2010). Inviting vices (e.g., Armendáriz & Murdoch, 2010). Many microfinance men to training could contribute to more egalitarian gender rela- institutions provide additional training on the management of tions as men’s attitudes and behaviours need to change to achieve income-generating activities to increase the recipients’ human gender equality (Howson & Flood, 2015). In fact, men may influ- capital (Augsburg, De Haas, Harmgart, & Meghir, 2015). While ence women’s agency both in direct interactions (i.e., the house- some research has provided evidence on the positive impacts of hold) and indirectly as decision-makers and community leaders training on business outcomes (for reviews see Frese, Gielnik, & (e.g., Fleming, Barker, McCleary-Sills, & Morton, 2013; Howson & Mensmann, 2016; McKenzie & Woodruff, 2014), other research Flood, 2015). Thus, involving men may be an important prerequi- reports only small or no impacts in low income countries (for a site to changing existing gender relations. review see McKenzie & Woodruff, 2014). In the context of microcredit, two studies investigated the idea It is interesting to note that, even though other studies do not of involving both women and their husbands. First, to avoid that explicitly focus on impacts of training on women empowerment, female borrowers may face intimate partner violence (as men- three studies provide evidence that a business training improves tioned above, see Rahman, 1999; Rahman et al., 2011) inviting hus- business outcomes for male entrepreneurs, but not for female bands is suggested to help to overcome this relational friction. entrepreneurs (Berge, Bjorvatn, & Tungodden, 2014; De Mel, Second, a study investigated the interest of female borrower to McKenzie, & Woodruff, 2009; Giné & Mansuri, 2014). If business invite their husbands to solidarity-groups (Allen, Armendáriz, outcomes turn out to be positively related to empowerment, these Karlan, & Mullainathan, 2010). Only a few women invited their  M. Huis et al. / World Development 120 (2019) 46–61 49 husbands. The authors speculated that women might only involve lion Vietnamese Dong ($43–$1000) with repayment terms from 25 their husbands if they hoped to solve arguments about the loan or to 100 weeks depending on the type of loan (TYM, 2017). to increase their decision-making power. One promising way to transform existing gender relations in 3.2. Intervention general could be inviting spouses to training (e.g., Mbweza, Norr, & McElmurry, 2008). Therefore, we expect that inviting female All female borrowers have access to microcredit loans. In addi- microfinance borrowers and their husbands to the GET Ahead tion, female borrowers in the two treatment conditions received training improves relational empowerment – either by means of access to additional business training tailored to the needs of a decrease in relational friction or by an increase in bargaining women. In the first treatment condition female borrowers received power – compared to female borrowers who were invited to par- access to training for which their husbands were also invited. In ticipate alone in training and even more so for women who were the second treatment condition female borrowers received access not invited (control group). to training individually. Female borrowers in the two treatment arms were invited to nine monthly training sessions that lasted 3. Context 45–60 min each. The training sessions were based on the GET Ahead program (ILO, 2011). This training program has been used Although Vietnam, our study setting, has made clear develop- in 18 countries since the late 1990s. Its ultimate goal is to increase mental progress in recent years—achieving economic growth, women’s economic and social empowerment (ILO, 2011). The shifting toward industrialization and modernization, and increas- training program consists of nine training modules. Each module ing social equity (World Bank, 2014)—it still ranks 121 out of focuses on a different theme (see Table 1 for an overview). This 187 nations on the Human Development Index and 58 out of 149 training program was culturally adapted to the Vietnamese con- on the Gender Equality Index (UNDP, 2014). Vietnam is a patriar- text. The sessions were shortened from approximately 240 min chal society, with strong traditional gender roles (Duvvury, to 45–60-min sessions selecting and adopting examples and exer- Carney, & Nguyen, 2012). We conducted our research in two pro- cises to the present cultural context.4 The gender component vinces in northern Vietnam: Vinh Phúc and Hà N i. These pro- included three modules. The first module focused on gender beliefs, vinces are neighbouring provinces in the Red River Delta which roles, prejudice, and gender equity. The second module centred on are very similar in their geographical characteristics and are com- female entrepreneurs’ business skills, confidence, and identifying parable with respect to their economic development. The Red River successful business goals. The third module covered the difficulties Delta covers an area of 21,060 km2 and is somewhat more affluent and challenges for women in doing business and setting up a coop- in comparison to other areas in Vietnam (for example see poverty eration. The pure business component included six different mod- rates in 2015: 7.0% in Vietnam, 3.2% in the Red River Delta; General ules. The fourth module centred on identifying and selecting Statistics Office of Vietnam, 2015). Furthermore, the two provinces business ideas and opportunities. The fifth module considered the rank among the highest out of the 63 provinces in Vietnam in importance of product, price, promotion, and place in marketing terms of good business environments for the private sector (Vinh and business sales. The sixth module focused on calculating interest Phúc ranks 4th; Hà N i ranks 24th; PCI, 2015). rates and the possibility of saving. The seventh module discussed opening and managing cash books. The eighth module focused on opening and managing account receivable and account payable 3.1. Microfinance in Vietnam books. The last module centred on calculating purchasing and pro- duction costs and costs of sold goods. In total, 16 TYM loan officers Only 21% of the adult population has access to formal financial received a training by the TYM headquarter staff to be able to teach services in Vietnam (International Finance Corporation [IFC], the modules. All of these 16 trained loan officers provided training in 2014). However, microfinance is widespread (e.g., Asian all centres and both treatment conditions. To support trainees’ Development Bank [ADB], 2012). Microfinance services are mainly understanding of content, loan officers used various teaching meth- offered by two state-owned banks; the Vietnam Bank for Agricul- ods and materials, such as role playing, cards, and pictures. In addi- ture and Rural Development and the Vietnam Bank for Social Poli- tion to monthly training sessions, the TYM staff offered clients access cies (Asian Development Bank [ADB], 2014). In addition, the to consultations repeating the discussed training material at weekly People’s Credit Fund and two licensed microfinance institutions, meetings when borrowers arrived to make payments on their loans. M7-MFI and TYM, are the leading formal providers of financial ser- The training sessions were free of charge and voluntary. We com- vices to the poor. However, the two licensed microfinance institu- pared female borrowers who received additional training on their tions and 48 semi-formal microfinance institutions account for less own with female borrowers whose husbands also received access than 2 percent of the market share of outstanding loans to the training; to encourage the husbands to participate, we offered (International Finance Corporation [IFC], 2014). In 2015, 28 finan- them financial compensation for participation in the training cial service providers financed 7.5 million borrowers with total sessions. loan amounts of $6.7 billion (MIX Market, 2017).3 The largest insti- tution in northern Vietnam, the Tao Yeu May fund, is owned by the Vietnam Women’s Union and started operating in 1992, with the 4. Methods and data inspection aim of improving the quality of life and status of poor women and their families. At the time of this study, TYM operated in 56 districts 4.1. Methods in Northern and Central Vietnam offering microloans to 119,653 bor- rowers adding to a total of 906,82 billion Vietnamese Dong in out- 4.1.1. Research design standing loans ($39 million). The TYM fund offers microcredit We randomly assigned 187 lending centres from three branches loans, savings, assistance funds, training, and social support activi- in Vinh Phúc and one in Hà N i to one of three conditions: (1) ties. Individual borrowers can receive loans from 1 million to 25 mil- female borrowers receiving a loan and access to training to which their husbands were invited, (2) female borrowers receiving a loan 3 These figures are based on available information provided to MixMarket by and access to training individually, or (3) a control group of female financial services providers. Important to note is that this information does not 4 include all institutions in Vietnam but it offers an estimate. The training module in Vietnamese can be obtained upon request. 50 M. Huis et al. / World Development 120 (2019) 46–61 Table 1 Table 2 Modules of the GET Ahead Training package. Personal control belief scale. Module 1: Gender and gender equity Internal locus of control External locus of control Module 2: The business woman and her self-confidence 1 What happens to me is my own Sometimes I feel that I don’t have Module 3: The business woman and her environment doing. enough control over the direction Module 4: The business project: Business ideas my life is taking. Module 5: The business project: Marketing and how to sell with success 2 When I make plans, I am almost When I make plans, it is not always Module 6: The business project-finance: Calculations and how to calculate certain that I can make them work. wise to plan too far ahead, because interest rate many things turn out to be a matter Module 7: The business project-finance: Managing cash of good or bad fortune anyhow. Module 8: The business project-finance: How to record accounts receivable 3 Getting what I want has little or Many times, I might just as well and accounts payable nothing to do with luck. decide what to do by flipping a Module 9: The business project: How to calculate costs of production and cost coin. of goods sold 4 It is impossible for me to believe Many times I feel that I have little that chance or luck plays an influence over the things that important role in my life. happen to me. borrowers receiving only a loan. The data were collected at three measurement points: baseline data in October–November 2011, or three months before the training started; midline data in be incorrectly translated, we removed it from the scale and used March–May 2013, or six months after the last training module; the remaining three items for the measure. We assigned one point and endline data in October–November 2013, one year after the for each statement reflecting an internal locus of control and last training module. On average, each interview lasted approxi- summed the total, so the scale ranged from 0 (no internal control mately one hour (ranging from 45 to 90 min). belief) to 3 (strong internal control belief). The average score on The randomization was stratified by lending branch; each personal control beliefs before the intervention varies between branch experienced the same proportion of two training and one 1.78 and 1.90 (see Table 4). control conditions. To account for the low expected participation Relational empowerment was assessed with two aspects; rela- by husbands, we oversampled the training condition group in tional friction and intra-household decision making power. which husbands were invited. We selected 70 credit centres in First, we assessed relational friction with seven items. The which female borrowers received a loan and access to training to female borrowers were asked to indicate how often, in the previ- which their husbands were invited, 31 credit centres in which ous six months, their spouse engaged in the seven acts: ‘‘verbal female borrowers received a loan and access to training individu- aggression”, ‘‘physical assault (pushed, slapped, beat or hit with a ally, and 86 credit centres in which female borrowers received only fist)”, ‘‘threatened and used an object like sticks, knife etcetera”, a loan. In addition, we excluded clients who received permission ‘‘kept you from seeing your family members or friends”, ‘‘insisted from TYM to miss monthly compulsory centre meetings, due to on knowing where you are at all times”, ‘‘wanted you to ask per- their work, because they could not attend any training. In each mission before doing anything”, and ‘‘insulted or humiliated you centre, we interviewed a random selection of a maximum of 23 in front of other people” (Straus, 1979; World Health female borrowers; if the centre had fewer than 23 members, all Organization [WHO], 2005). The female borrowers indicated how of them were interviewed. A total of 4041 female borrowers were often, in the previous six months, they had experienced these sampled at the baseline assessment, of whom 1509 received a loan seven acts on a five-point scale, from 0 (never) to 4 (very often). and access to training to which their husbands were invited (T1), We computed an overall scale for relational friction by averaging 673 received a loan and access to training individually (T2), and the score on these seven items. The average score on relational fric- 1859 received only a loan (C). tions before the intervention varies between 0.12 and 0.14 (see Table 4). This implies that, on average, women indicated that they 4.1.2. Procedure never experienced any form of relational frictions. The very low Interviews were conducted at all three data collection points score on relational friction is somewhat surprising and may be (baseline, midline, and endline). Twenty-three native enumerators due to some underreporting on one or more of the seven items interviewed the participants one-on-one to gather information of relational frictions. Women may especially be tempted to under- about each client’s individual characteristics, households, socio- report on the sensitive questions related to relational aggression.6 economic status, business practices, gender awareness, cognitive For this reason, we also present estimates on an index which and non-cognitive skills, household decision making, loans, and only included items which assessed relational oppression, the relational friction.5 All study materials were carefully pretested extent to which female borrowers were limited in their personal and adapted to the cultural context. freedom by their husbands (four items: kept you from seeing your family members or friends, insisted on knowing where you are at 4.1.3. Measures all times, wanted you to ask permission before doing anything, We assessed personal control beliefs to reflect personal empow- and insulted or humiliated you in front of other people). We call erment, using four items adapted from Rotter’s (1966) Locus of this additional index relational oppression. This index excluded Control scale. The items assess the extent to which participants three items assessing relational aggression against female borrow- believe that they are in control of their lives (internal control ers (verbal aggression, physical assault and threatened and used an beliefs) or that their lives are controlled by environmental factors object like sticks, knife etcetera). outside their control, such as fate or chance (external control beliefs). Participants selected a statement from pairs to match their 6 beliefs; one statement reflected internal and the other represented By using the item count technique Bulte and Lensink (2019) estimate that external control beliefs (see Table 2). Because one item appeared to between 10 and 20 percent of women in our sample are hit by their husband. Unfortunately, our dataset doesn’t enable to test whether underreporting may also be a problem for other items in our relational frictions index. However, it seems 5 We used self-reported data for all our measures. We believe it is unlikely that plausible to assume that underreporting is especially problematic for items related to participants will answer the questions differently due to their participation in the relational aggression, and does not or to a much lower extent affect items related to training or apply knowledge learned in the training that could impact the measures. relational oppression.  M. Huis et al. / World Development 120 (2019) 46–61 51 Second, to assess intra-household decision making7, we used ten Table 3 items that measured who made financial decisions (adapted from Intra-household decision making scales. Banerjee, Duflo, et al., 2015). For ten expenditures, the female bor- Large expenditure decision making rowers indicated who made most decisions: their husband, they Who makes most decisions about asking for a loan? and their husband together as a couple, or they alone.8 The items Who makes most decisions about consumer durable items? (TV, Fridge, Tape recorder etc.) assessed decision making about various fields, such as food, tuition, Who makes most decisions about what health expenditures to make? loans, investments, and financial assistance for family members. We Who makes most decisions about saving for business and for household? computed a scale by assigning 0 points for each decision made by Who makes most decisions about expenses for home purchase, improvement the husband, 0.5 points for each decision made by the couple or repair? Who makes decisions about where to invest surplus money? together, and 1 point for each decision made by the wife, then Who makes decisions about how to assist family members? summed these points for all ten items. The average score on intra- Daily domain decision making household decision making before the intervention varies between Who makes most decisions about what food items to purchase? 6.21 and 6.47 (see Table 4). Who makes most decisions about what educational expenditures to make However, decision making power among couples may differ (tuition, etc.)? with respect to more male dominated domains and the daily Who makes most decisions about what clothing items to purchase? domain. More precisely, previous research suggests that women’s financial decision making in domains that are traditionally male dominated and concern larger sums of money reflect a different type of decision making power compared to decisions in domains The results of the balancing test indicated no reason to doubt the that are traditionally female dominated (e.g., Dutt, Grabe, & randomization (see Table 4). We also conducted joint significance Castro, 2016; Johnson, 2016). Therefore, we asked female borrow- tests across all baseline characteristics by regressing treatment sta- ers to indicate the extent to which they participated in decision tus on the set of baseline characteristics. These estimates also sug- making in two domains: decision making on larger expenditures, such gest that all groups are similar (p >.67). as loans, savings, and investments, and decision making in the daily domain, such as food or clothing. To test whether this differentia- tion can also be empirically supported in our data, we conducted 4.2.2. Attrition a principle component analysis. As expected, seven items loaded Table 5 shows that 4041 female microfinance borrowers were on a factor assessing decision making power on larger expenditures, interviewed initially, at the midline 3511 female microfinance bor- and three items loaded on a factor assessing decision making power rowers were interviewed again. However, the attrition rate from in the daily domain.9 Therefore, we computed two scales. The sub- baseline to midline was higher in the condition where female bor- scale of decision making on larger expenditures included seven deci- rowers were invited to receive training individually. To compen- sions related to expenditures such as ‘‘Who makes most sate for this larger attrition, we randomly selected and decisions about asking for a loan?” (abaseline = 0.91, amidline = 0.89, interviewed an additional 315 female microfinance borrowers to aendline = 0.87). The average score on decision making on larger increase the sample size in this condition. Together this resulted expenditures before the intervention varies between 3.88 and 4.06. in 3826 female microfinance borrowers at the midline. A sample The subscale of decision making in the daily domain included three of female borrowers who dropped out and indicated why revealed decisions related to expenditures such as ‘‘Who makes most that 10 women were no longer clients of TYM, 95 refused to partic- decisions about what food items to purchase?” (abaseline = 0.84, ipate in the study, and 40 provided different reasons (e.g., time amidline = 0.88, aendline = 0.71). The average score on decision making constraints). At the endline, supported by TYM, we interviewed in the daily domain before the intervention varies between 2.32 almost all female microfinance borrowers who were interviewed and 2.41. We confirmed the two factors at the midline and endline, at the baseline and/or at the midline resulting in 4350 borrowers using confirmatory factor analysis (see Table 3). at the endline. 3412 female borrowers were interviewed at all three measurement occasions (T1 = 1311; T2 = 549; C = 1552). 4.2. Data inspection We created an attrition dummy; equal to 1 if a household was not interviewed at the midline or endline, and then used a logistic 4.2.1. Balancing test regression analysis to check for non-random attrition (see We tested for baseline differences in observable characteristics Table 6)10. We regressed attrition on the treatment groups T1 and across the three groups by regressing the demographic variables, T2, age, lending branches, marital status, household size, baseline control variables, and outcome variables at the baseline on dum- measures of all dependent variables, and interactions between treat- mies, to determine the assignments to the two training conditions. ment indicators and demographic characteristics. The analysis sug- Because the randomization took place at the microfinance centre gests that attrition was random. We only found that in one of the level, we clustered the standard errors at the centre level. We also lending branches attrition was higher. Yet, Wald tests showed no dif- test for baseline differences between the two treatment groups. ference in attrition for lending branches interacted with the two treatment conditions. 7 In the construction of our measures we have used a principle component analyses only for household decision-making power as we theorized that this construct should 4.2.3. Training acceptance be differentiated between two ‘categories’ of decision-making: decision-making in It appears that most of the invited female borrowers partici- the daily domain and decision-making on larger expenditures (p. 20). For our other pated in the offered training, suggesting minor non-compliance measures we had no a-priori reason to expect the necessity of a differentiation between two or more components, thus we used an equally-weighted index. 8 10 At the baseline, the participants were asked to decide among husband/other, Please note that the sample in the attrition-table differs from the sample reported couple, and themselves. However, at the midline and end line, ‘‘others” was presented in the tables in the main-analyses. We have added tables reflecting attrition analysis as a fourth category. Both the husband and others options were coded as 0 for the for the three samples (midline, endline, double difference) in the appendix (Table A3). scale construction. As can be seen, the results are almost the same. Please note that we conducted the 9 The principle component factor analysis indicated that one item (health analyses below based on the sample for the control beliefs-variable. Moreover, in the expenditures) was not related to one of the two factors. Because the results were attrition analysis, some variables with missing variables in the baseline are taken into unaffected by the addition or removal of this item, we included it in the scale account. Thus, our sample sizes presented in table A3 still differ somewhat from the construction. tables in the main text. 52 M. Huis et al. / World Development 120 (2019) 46–61 Table 4 Descriptive overview of control variables, descriptive statistics, and dependent variables and a balancing test between conditions. Female borrowers Female borrowers Female borrowers Test of equality [Regress receiving training with receiving training receiving only a loan (C) var_baseline T1 T2, vce husbands (T1) individually (T2) (cluster centre-level)] Mean (SD) Range Mean (SD) Range Mean (SD) Range T1 = C T2 = C T1 = T2 Age 43.47 (10.32) 19–71 43.88 (10.50) 20–71 44.02 (10.26) 20–72 0.39 0.82 0.58 Educational level 1.93 (0.68) 1–5 1.97 (0.66) 1–5 1.95 (0.65) 1–5 0.75 0.71 0.57 Region (Hà N i) 0.25 (0.43) 0–1 0.27 (0.45) 0–1 0.26 (0.44) 0–1 0.89 0.91 0.83 Married 0.81 (0.39) 0–1 0.83 (0.38) 0–1 0.82 (0.38) 0–1 0.52 0.65 0.32 Household size 4.74 (1.56) 1–12 4.71 (1.64) 1–15 4.77 (1.53) 1–13 0.68 0.57 0.82 Control beliefs 1.78 (0.79) 0–3 1.78 (0.76) 0–3 1.90 (0.81) 0–3 0.08* 0.21 0.97 Relational friction 0.13 (0.22) 0–2.3 0.14 (0.22) 0–1.3 0.12 (0.19) 0–1.6 0.61 0.23 0.41 Relational oppression 0.08 (0.24) 0–2.75 0.08 (0.22) 0–2.5 0.10 (0.25) 0–2.5 0.29 0.06** 0.27 Overall decision making 6.47 (2.19) 0–10 6.21(2.15) 0–10 6.45(2.08) 0–10 0.90 0.18 0.17 Decision making on larger expenditures 4.06 (1.61) 0–7 3.88 (1.54) 0–7 4.05 (1.50) 0–7 0.99 0.14 0.15 Decision making in daily domain 2.41 (0.78) 0–3 2.32 (0.81) 0–3 2.39 (0.76) 0–3 0.75 0.45 0.34 Notes. Descriptive statistics of control variables, descriptives, and dependent variables at baseline are displayed per condition. The test of equality displays p-values observed in ordinary least square regressions to test for possible differences at baseline between the conditions. Standard errors are clustered at the centre level (187 centres). We also conducted joint significance tests across all baseline characteristics by regressing treatment status on the set of baseline characteristics. These estimates strongly suggest that all groups are similar on average as the F statistics were highly insignificant (p >.67; results can be obtained on request). *p <.10. **p <.05. ***p <.01. Table 5 Attrition per condition per measurement-time. Female borrowers receiving Female borrowers receiving Female borrowers receiving Total training with husbands (T1) training individually (T2) loan but no training (C) Female borrowers at baseline 1509 673 1859 4041 Female borrowers at midline 1328 565 1618 3511 Additional borrowers at midline – 315 – 3826 Female borrowers at end line 1507 988 1855 4350 Attrition from baseline – midline 181 108 241 530 Attrition rate baseline – midline 12.0% 16.1% 13.0% 15.09% Attrition from baseline – end line 2 0 4 6 Attrition rate baseline – midline 0.1% 0.0% 0.2% 0.1% Note. The attrition rate is calculated based on the originally sampled dataset, excluding the 315 female borrowers assigned to treatment condition T2 who were interviewed at the midline - post facto- to compensate for the higher attrition-rate in this condition. problems. More specifically, on average 80.2% of the female bor- Table 6 Non-random attrition. rowers receiving a loan and access to training individually partici- pated in each of the offered training modules (range: 74.6–83.3%). B (SE) A similar pattern was observed for female borrowers receiving a T1 0.041 (0.09) loan and access to training to which their husbands also were T2 0.08 (0.17) invited: on average 83.4% of the invited female borrowers partici- Age 0.072 (0.092) Control beliefs at baseline 0.01 (0.01) pated in each of the offered training modules (range: 80.9–84.6%). Relational friction at baseline 0.01 (0.06) However, on average only 24.1% of all invited husbands partici- Larger expenditures at baseline 0.011 (0.01) pated in each of the offered training modules (range: 7.2–39.1%). Daily domain expenditures at baseline 0.01 (0.02) Note that a larger proportion of the invited husbands participated Age*T1 0.072 (0.011) Age*T2 0.072 (0.021) in the gender component modules which were offered first (see Control beliefs at baseline*T1 0.031 (0.02) Table 7 for an overview of the attendance per training module). Control beliefs at baseline *T2 0.04 (0.03) Relational friction*T1 0.01 (0.07) 5. Identification strategy Relational friction*T2 0.06 (0.10) Larger expenditures*T1 0.041 (0.01) Larger expenditures*T2 0.02 (0.02) We first conducted intention to treat (ITT) estimates at the mid- Daily domain expenditures*T1 0.01 (0.02) line and endline, with post-treatment regressions, using control Daily domain expenditures*T2 0.09 (0.05)* variables.11 The controls helped increase the precision of our esti- Constant 0.18** (0.06) Observations 3286 mates of the impact of the GET Ahead gender component training R-squared 0.016 on outcomes Yijt, measured for individual i in centre j at time t (sep- arately for midline and endline), according to the following Notes. Coefficients indicate differences in attrition. We have reported branch-fixed specification: effects. Standard errors adjusted for cluster effects at the centre level (187 centres) are in parentheses. We also conducted these analyses with a Logit regression, which Y ij ¼ b0 þ b1 T1ij þ b2 T2ij þ b3 Z ij þ eij ð1Þ yielded the same results. We conducted a joint significance test by regressing all characteristics on attrition, this estimate suggests that there are no differences based on the included characteristics [F(26, 185) = 1.35, p =.129]. We also con- ducted a joint significance test by regressing all interactions for T1 and T2 on 11 All these results report the findings of analyses conducted with control variables attrition, this estimate suggests that these interactions do not explain attrition (age, education, lending branches), which did not differ from the results obtained [F(12, 185) = 0.65, p =.797]. T1 = treatment condition in which female borrowers through analyses without control variables. These latter results are available on were invited to training with their husband, T2 = treatment condition in which request. female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01.  M. Huis et al. / World Development 120 (2019) 46–61 53 Table 7 Overview of attendance at the training sessions by female borrowers and their husbands.1 Female borrowers Female borrowers receiving training with their husband (T1) receiving training alone (T2) Participation wife Participation wife Participation husband n % n % n % Gender training module 1 693 79.0 1095 84.6 442 39.1 Gender training module 2 654 74.6 1058 81.8 403 35.6 Gender training module 3 715 81.5 1094 84.5 363 32.1 Business training module 4 727 82.9 1084 83.8 342 30.2 Business training module 5 688 78.5 1068 82.5 295 26.1 Business training module 6 727 82.9 1095 84.6 292 25.8 Business training module 7 722 83.3 1087 84.0 137 12.1 Business training module 8 710 81.0 1089 84.2 98 8.7 Business training module 9 686 78.2 1047 80.9 82 7.2 1 In addition to the reported ITT analyses we conducted IV-regressions to obtain insight in the impact of participation and not invitation to the training modules by female borrowers. However, since the uptake of the training is consistently high (>74.5%) these analyses are not meaningful and showed very similar results. The results can be obtained upon request. where T1ij is a dummy equal to 1 if a woman has access to training tional level, and lending branch in time t; and eijt is an error term. for which her husband is invited; T2ij is a dummy equal to 1 if a The coefficients of interest are b1, b2, b3, and b4. That is, b2 and b4 woman has access to training individually; Zij is a vector of variables estimate the impact of providing female microfinance borrowers age, educational level, and region measured at the baseline. As we access to the GET Ahead training individually at the midline and used a stratified randomization at branch level, Z also includes at the endline, respectively, whereas b1 and b3 estimate the impact dummies for the different branches, and eij is an error term. In these of providing female microfinance borrowers access to the GET and all following analyses, the standard errors are clustered at the Ahead training to which their husbands were invited at the midline centre level to control for possible dependency between female bor- and endline, respectively. rowers in the same centres. The coefficients of interest are b1 and b2, which measure the impact of providing female borrowers access to 6. Results training individually or training for which their husbands are invited, respectively. 6.1. Personal control beliefs Next, we conducted double difference estimations with the age, educational level, and lending branch control variables.12,13 We We expected that female borrowers who received access to the also examined the impact of being assigned to the two training con- GET Ahead for women in enterprise training would develop more ditions, compared with the control condition, and differentiated internal control beliefs; Table 8 partly confirms this expectation. midline and endline effects. We thus estimated the impact of the The post-treatment regressions indicated that training increased GET Ahead training on outcome Yijt for individual i in centre j at time internal control beliefs at the endline but not at the midline for t using the following specification: female borrowers with access to training (see Table 8, Columns 1 and 2). The double difference regressions provided slightly differ- Yijt ¼ b0 þ b1 Midijt  T1ijt þ b2 Midijt  T2ijt þ b3 Endijt  T1ijt ent results though, suggesting a small positive impact of individual þ b4 Endijt  T2ijt þ b5 T1ijt þ b6 T2ijt þ b7 Midijt þ b8 Endijt training at the midline and an impact of both the individual and of þ b9 Zijt þ eijt ð2Þ the joint training at the endline (see Table 8, Column 3)14. The results clearly suggest that it may take some time before training where Midijt and Endijt are dummies equal to 1 when the observa- changes women’s personal control beliefs; the endline effects are tion belongs to the midline or endline survey, respectively; T1ijt is a stronger than the midline effects. We do not find that inviting hus- dummy equal to 1 if a woman has access to training for which her bands significantly changes the impact of the training on control husband is invited; T2ijt is a dummy equal to 1 if a woman has beliefs of women who have been offered the training. access to training individually; t indicates whether an observation belongs to the baseline, midline, or endline survey; b5 and b6 are 6.2. Relational friction fixed effects for T1 and T2 respectively measuring whether T1 and T2 differ from the control group in the baseline; b7 and b8 are fixed Table 9 shows that training reduces relational friction at the effects for the midline and endline respectively measuring whether endline but not at the midline, across both estimation methods. the midline and endline differ from the baseline estimate for the However, as the relational friction index may be biased due to control group; Zijt is a vector of the control variables age, educa- underreporting on some items, specifically those related to rela- tional aggression, we also consider the impact on relational 12 We conducted double difference estimations on balanced data, including only participants who reported data on all relevant variables at all three assessment 14 points, to control for possible fixed effects at the group level. We use a consistent Please note that the point estimates for the post-treatment regressions "T2 x sample for all dependent variables, only including responses from women who Midline" differ substantially from the point estimates for the double-difference provided responses on all variables at all time points. No important differences arose specifications. In order to test whether these differences are due to differences in the for the double difference estimations with balanced versus unbalanced data; thus, sample used for the post-treatment vis a vis the double difference specification, we both estimations yielded similar results. also estimated the post-treatment models using the sample used for the double 13 We have also estimated double difference models including fixed effects at the difference specification, see Table A1 in the appendix. It appeared that these estimates individual level, see Table A2 in the appendix. However, as expected, these results are give very similar results as the post-treatment estimates presented in the text almost the same as the (sum of the) fixed effects at the individual level are perfectly (estimates can be obtained upon request). Thus, it is unlikely that the differences in collinear with the treatment dummies (fixed effects at treatment level), in case of a the point estimates are not due to varying samples. Rather, it should be explained by balanced panel. the ‘‘fixed effects” that are taken into account in the double difference specification. 54 M. Huis et al. / World Development 120 (2019) 46–61 Table 8 Results on women’s control beliefs. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.06 (0.09) 0.15 (0.10) T2  Midline 0.11 (0.12) 0.26** (0.12) ** T1  Endline 0.18 (0.09) 0.30** (0.125) T2  Endline 0.33*** (0.10) 0.52*** (0.14) Constant 1.80*** (0.12) 2.45*** (0.12) 1.91*** (0.10) N 3386 3325 7218 R2 0.035 0.146 0.070 T1 Mid = T2 Mid 0.19 0.87 T1 End = T2 End 1.76 0.14 T1 Mid = T1 End 3.31 T2 Mid = T2 End 5.30** Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. Table 9 Results on women’s experienced relational friction. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.031 (0.02) 0.02 (0.03) T2  Midline 0.03 (0.02) 0.10** (0.04) T1  Endline 0.12*** (0.03) 0.17*** (0.04) T2  Endline 0.11** (0.04) 0.16*** (0.06) Constant 0.28*** (0.03) 0.51 *** (0.06) 0.34*** (0.03) N 3369 3070 7218 R2 0.062 0.199 0.123 T1 Mid = T2 Mid 2.23 3.32* T1 End = T2 End 0.14 0.95 T1 Mid = T1 End 19.56*** T2 Mid = T2 End 2.15 Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. Table 10 Results on women’s experienced relational oppression. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.02 (0.02) 0.01 (0.04) T2  Midline 0.02 (0.02) 0.10** (0.05) *** T1  Endline 0.10 (0.03) 0.16*** (0.05) T2  Endline 0.09** (0.04) 0.17*** (0.06) Constant 0.17*** (0.03) 0.40*** (0.07) 0.22*** (0.03) N 3391 3095 7218 R2 0.062 0.158 0.098 T1 Mid = T2 Mid 2.56 2.80* T1 End = T2 End 0.13 0.03 T1 Mid = T1 End 15.96*** T2 Mid = T2 End 1.90 Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. oppression, which assesses the extent to which female borrowers time before the training starts to exert an impact. In contrast to were limited in their personal freedom by their husbands.15 our expectations, we do not find different impacts if husbands were Table 10 shows that these results are the same. Again, it may take invited to participate (see Table 10, Columns 1–3). 15 Bulte and Lensink (2019) analyse the impact of the training on physical assault, 6.3. Intra-household decision making one item of the relational frictions index, ignored in the relational oppression index. They find opposite results when the item is based on a List experiment (the item count technique) or based on survey data. The results based on the List experiment The post-treatment regressions in Table 11 show a significant suggest that the training accentuated physical assault. increase in women’s intra-household decision making at both  M. Huis et al. / World Development 120 (2019) 46–61 55 Table 11 Results on women’s intra-household decision making power. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.21* (0.11) 0.27 (0.17) T2  Midline 0.011 (0.17) 0.09 (0.24) * T1  Endline 0.33 (0.14) 0.39** (0.19) T2  Endline 0.05 (0.01) 0.26 (0.22) Constant 5.33*** 4.73*** 6.11*** N 3318 3395 7218 R2 0.069 0.066 0.047 T1 Mid = T2 Mid 1.62 0.60 T1 End = T2 End 3.10* 0.42 T1 Mid = T1 End 0.77 T2 Mid = T2 End 0.27 Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. Table 12 Results on women’s decision making about larger expenditures. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.13 (0.08) 0.20* (0.11) T2  Midline 0.05 (0.12) 0.05 (0.16) T1  Endline 0.34*** (0.11) 0.40*** (0.13) T2  Endline 0.10 (13) 0.33* (0.17) Constant 2.93*** (0.19) 2.65*** (0.21) 3.69*** (0.15) N 3312 3396 7218 R2 0.047 0.044 0.032 T1 Mid = T2 Mid 2.24 0.95 T1 End = T2 End 3.88* 0.25 T1 Mid = T1 End 3.16* T2 Mid = T2 End 4.89** Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. Table 13 Results on women’s decision making in the daily domain. Midline post-treatment regression Endline post-treatment regression Double difference regression (1) (2) (3) T1  Midline 0.08 (0.06) 0.06 (0.09) T2  Midline 0.06 (0.09) 0.04 (0.11) T1  Endline 0.01 (0.06) 0.01 (0.10) T2  Endline 0.05 (0.09) 0.07 (0.11) *** Constant 2.40 (0.12) 2.07*** (0.11) 2.42*** (0.11) N 3416 3406 7218 R2 0.088 0.142 0.076 T1 Mid = T2 Mid 0.09 0.06 T1 End = T2 End 0.28 0.27 T1 Mid = T1 End 1.09 T2 Mid = T2 End 1.68 Notes. Standard errors are in parentheses, adjusted for clustering at the centre level. The variation in the reported N is caused by differences in women’s non-response to the specific dependent variables and/or included control variables. We conducted the double difference analyses using a consistent sample for all dependent variables. However, the reported results were very similar to the results we observed when only balancing the data for each dependent variable separately. For all analyses Wald tests are reported to test for equality of the two treatment condition coefficients. The Wald statistic is reported in the table. T1 = treatment condition in which female borrowers were invited to training with their husband, T2 = treatment condition in which female borrowers were invited to training individually. *p <.10. **p <.05. ***p <.01. 56 M. Huis et al. / World Development 120 (2019) 46–61 midline and end line, if the training is offered to female borrowe