WPS8270 Policy Research Working Paper 8270 Mission and the Bottom Line Performance Incentives in a Multi-Goal Organization Xavier Giné Ghazala Mansuri Slesh A. Shrestha Development Research Group Finance and Private Sector Development Team & Poverty and Equity Global Practice Group December 2017 Policy Research Working Paper 8270 Abstract The impact of performance pay in institutions with mul- it undermined the social outcome. In contrast, the social tiple goals depends on complementarities in the disutility bonus advanced the social mission as well as the micro- cost of effort and how different tasks interact to achieve credit program, but only for employees working alone, each goal. Workers of a mission-oriented nonprofit were undermining the performance of employees working in randomly assigned to one of two bonus schemes, each teams. These results cannot be explained by a standard incentivizing one of its two main operational goals: the multitask principal-agent model featuring only comple- performance of its microcredit program and the strengthen- mentarities in the disutility cost of effort. Instead, they ing of community institutions of the poor. This study finds suggest that production complementarities are also relevant. that the credit bonus improved credit-related outcomes but This paper is a product of the Finance and Private Sector Development Team, Development Research Group and the Poverty and Equity Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at xgine@worldbank.org and gmansuri@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Mission and the Bottom Line: Performance Incentives in a Multi-Goal Organization e, Ghazala Mansuri, and Slesh A. Shrestha ∗ Xavier Gin´ JEL Classification: C93, D86, J33, M52, M55 Keywords: complementarities, incentives, intrinsic motivation, teamwork, field experiment ∗ Gin´ e: Development Research Group, The World Bank, xgine@worldbank.org. Mansuri: Global Practice on Poverty and Equity and Development Research Group, The World Bank, gmansuri@worldbank.org. Shrestha: De- partment of Economics, National University of Singapore, slesh@nus.edu.sg. This project was jointly funded by the Development Research Group at the World Bank, and the Pakistan Poverty Alleviation Fund. The views expressed herein are those of the authors and should not be attributed to the World Bank, its executive directors, or the countries they represent. 1 Introduction Economic theory has long suggested that monetary incentives can motivate agents to fulfill the goals of their principals. 1 In a seminal article, Holmstrom and Milgrom (1991), however, argue that incentives can produce poor results in settings where workers typically perform multiple tasks, as they devote more effort to those that are directly incentivized to the detriment of others. For-profit firms use profit-based incentives to overcome this multi-task problem (Murphy, 1999). But nonprofit organizations and public bureaucracies are driven by broader missions, often cast in diffuse and aspirational language (Dixit 2002). As Wilson (1989) has shown, the most successful bureaucracies are those that translate such aspirational missions into narrowly focused operational goals, which serve to increase accountability and make performance easier to evaluate and reward. However, missions often involve multiple goals, not all of which may translate into equally well measurable indicators (Dewatripont et al., 2000; Besley and Ghatak, 2005). The challenge of using performance incentives in organizations with multiple goals is perhaps the reason why evaluations of teacher incentives find mixed results on student’s performance in tests and critical thinking (Neal, 2011). Similarly, studies of health care providers find weak incentive effects on rewarded as well as unrewarded measures of health (Mullen et al., 2010). More generally, the impact of rewarding a subset of goals will depend on cost complementarities, that is, how effort provision affects the disutility cost of effort; and on production complementari- ties, namely, how effort in the different tasks interacts to achieve each goal. Despite a substantial theoretical literature on worker incentives in nonprofits and bureaucracies which has emphasized the role of complementarities (for example, Dixit 2002), there is little empirical evidence on how such complementarities affect an agent’s performance, and the implications of incentivizing specific tasks for the organization’s broader mission. In this paper, we examine the role of production and cost complementarities in designing incentive schemes for agents who work in a mission oriented nonprofit and are charged with more than one task. We first develop a stylized model of worker effort and then provide experimental evidence from a randomized trial that tests the effectiveness of incentivizing worker performance. In the model, workers choose effort over two tasks, each related to an organizational goal. Tasks are allowed to be strategic complements, substitutes, or neither in cost and production. The nonprofit we study is a prominent development organization in Pakistan called the Pakistan Poverty Alleviation Fund (PPAF). Its mission is to reduce poverty by supporting the development needs of poor communities and increasing their livelihood opportunities. PPAF works with partner 1 See Gibbons (1998) and Prendergast (1999) for a general overview of worker incentives used in organizations. 2 non-profit institutions who share its mission. The evaluation focuses on PPAF’s largest partner, the National Rural Support Program (NRSP). NRSP has two operational goals: building and strength- ening organizations of the poor and supporting poor households directly through small (micro) loans intended for investment in income generating activities. All NRSP field staff (Field Assistants or FAs) support both operational goals in their day to day interactions with poor communities. In order to assess the nature of complementarities across these two organizational goals, all FAs were assigned to one of two bonus schemes or to a control group. The “social bonus” rewarded effort on tasks related to social empowerment, such as working with communities to create and strengthen organizations of the poor, while the “credit bonus” incentivized the health of the microcredit port- folio, which required bringing in new community members for micro loans and ensuring the timely repayment of loans. Prior to the experiment, NRSP had been rapidly expanding the microcredit program, raising concerns about possible negative effects on the quality of the community level institutions. At the same time, most FAs were being moved from main branches to smaller satellite offices located in/near villages (labeled ‘village branches’) in an effort to improve their proximity to clients. While working directly from the main branches, FAs had been offered a fixed salary. NRSP was also concerned that the transfer of FAs to village branches, typically in teams of two FAs, could weaken the monitoring ability of supervisory staff and was therefore interested in testing the introduction of pay for performance incentives for FAs. Given that FA’s job description involves two related but separate tasks, the optimal incentive scheme would likely reward both goals simultaneously. Such a scheme, however, would not reveal the nature of the underlying cost and production complementarities even though the agent would take these into account when allocating effort optimally. At the same time, knowledge about the nature of cost and production complementarities can inform the institution about the need for pay for performance incentives and their potential impacts on its operational outcomes and mission. In this sense, the field experiment and the predictions of the model, taken together, allow us to answer three questions that are central to designing contracts for workers in multi-goal organizations: (1) In a context where workers have multiple tasks and direct worker supervision is not possible, can monetary incentives be used to improve worker performance? (2) What is the nature of production and cost complementarities between tasks? (3) Does working in teams affect the impact of monetary incentives for multi-task agents and, if so, how? Our results show that staff effort on social-related tasks did not harm (and in some instances, improved) the performance of the institution’s credit program but incentivized focus on credit- related tasks undermined the institution’s empowerment mission. In particular, the credit bonus 3 improved NRSP’s microcredit program but only for outcomes directly incentivized by the bonus. At the same time, it worsened the quality of community organizations (COs). In contrast, the social bonus increased CO formation, and did so without worsening microcredit outcomes or CO quality. In fact, among FAs working individually, the social bonus was as effective as the credit bonus at improving credit outcomes. These results indicate that both production and cost complementarities appear to be empirically relevant. While cost complementarities alone could explain why the credit bonus improved credit- related outcomes and worsened social-related outcomes, they cannot explain why at the same time the bonus that incentivized social outcomes also improved the health of the credit portfolio. However, for FAs working in teams, the social bonus had a negative impact on both credit and empowerment outcomes. The literature on intrinsic motivation provides some insights on why this might be so. Osterloh and Frey (2000) among others, argue that rewards for mission related tasks can undermine intrinsic motivation and exacerbate free-riding. 2 In our case, FAs clearly view the building and strengthening of community organizations as a bigger part of NRSP’s core mission. This is also evident in a significant decline in intrinsic motivation among all FAs offered the social bonus. For FAs working in teams, this also exacerbated the propensity to free-ride. In contrast, the credit bonus had no impact on intrinsic motivation or the odds of free-riding in teams. Our results contribute to several strands of the literature. Evaluations of microcredit programs have typically found small or negligible impacts on empowerment or other social-related outcomes (Banerjee et al. 2015, among others). 3 Our results offer a plausible explanation for this finding. We find that incentive structures that only reward the performance of the credit portfolio, similar to the incentives provided by many of the institutions reviewed in Banerjee et al. (2015), undermine social-related goals. They also make employees focus on the tasks directly related to the payment of incentives to the detriment of other tasks, such as mobilizing new community groups or increasing outreach to the poor. To be sure, incentive structures that focus on repayment or that rely on CO membership fees may be a response to the pressure that microfinance institutions face to become financially self-sustainable (Greaney et al., 2016), but under these schemes social empowerment may languish. Our results also contribute to the literature on the use of performance-based incentives to im- prove the delivery of public goods and services in both developing and developed countries (Lavy, o et al., 2013; Goodman 2002, 2009; Muralidharan and Sundararaman, 2011; Duflo et al., 2012; B´ and Turner, 2013; Olken et al., 2014; Imberman and Lovenheim, 2015). Anti-poverty programs in- 2 See Frey and Jegen (2001) and Gneezy et al. (2011) for an overview of intrinsic motivation literature. 3 Relatedly,several studies have found that incentives focused on credit-related tasks can change the composition of the borrower pool, favoring richer and more credit worthy individuals (McKim and Hughart, 2005; Aubert et al., 2009). 4 creasingly provide a holistic set of private and public services based on the idea that combinations of interventions are required to address the multidimensional problems of development (Mansuri and Rao, 2013; Banerjee et al., 2015). These programs often use a community participation approach, which relies on active and empowered communities. 4 When incentives can only be provided for a subset of goals, perhaps because effort towards unrewarded goals is too noisy, our results suggest that a careful assessment of task complementarities in implementation design will be critical for understanding the success or failure of such programs. Finally, our results relate to the broader literature on the role of performance pay in organizations (Lazear, 2000; Paarsch and Shearer, 2000; Shearer, 2004). Studies that contemporaneously vary worker incentive structure within a single firm are rare (with the exception of Bandiera et al. 2007, 2013; Friebel et al. 2017). We extend this literature by studying incentive design in mission-oriented organizations with staff that not only have multiple tasks, but may also be intrinsically motivated (Besley and Ghatak, 2005; Osterloh and Frey, 2000; Bowles and Polanina-Reyes, 2012) working sometimes in teams (Bandiera et al., 2010, 2013). In such settings, incentives have been shown to affect worker performance by changing, in some instances, the number and the quality of job o et al., 2013; Ashraf et al., 2014; Deserranno, 2016) and by inducing greater effort applicants (B´ among mission-motivated workers, albeit to a much smaller degree compared with their peers whose preferences are not aligned with the mission (Carpenter and Gong, 2016). Our results suggest that financial incentives that crowd-out intrinsic motivation can also affect performance by undermining the willingness of motivated employees to work in teams. The rest of the paper proceeds as follows. Section 2 describes NRSP’s organizational goals and its overall mission, and outlines the experimental design. Section 3 presents the model of worker effort choice in an organization with multiple objectives. Section 4 discusses the data, Section 5 describes the empirical strategy, and Section 6 reports the results of the experiment and discusses how the empirical findings match with the predictions of the model. Section 7 concludes. 2 Context and Experiment NRSP’s mission is to reduce poverty by empowering poor households and investing in their liveli- hoods.5 Central to this mission is a process of social mobilization—that involves the formation and strengthening of COs of the poor at the village or community level. Each CO is typically comprised 4 In the last 10 years, the World Bank alone has invested USD 85 billion in such participatory community driven development projects, with mixed results in improving a range of socio-economic outcomes and strengthening political institutions (Casey et al., 2012; Mansuri and Rao, 2013; Khanna et al., 2015). 5 NRSP is a development organization operating in Pakistan since 1991. Its activities have covered more than 2.5 million households, with 550,000 current clients in all four provinces, making it the largest rural support program in the country in terms of outreach, staff and development activities. 5 of 15 members, who live close to each other in the same village. 6 Once the CO is formed, its members meet regularly to save, attend skill-training programs, and work collectively on local infrastructure projects. NRSP also enhances the livelihood of CO members by offering microfinance services. Two-thirds of CO members are active borrowers. Microfinance is provided in the form of individual loans, usually with a single or monthly installments and a maturity of six to 12 months. 7 CO membership is a prerequisite to access these loans. NRSP views microcredit and social mobilization as the main instruments to improve their liveli- hoods and empower communities. At the same time, the health of its microcredit program is also critical to NRSP’s financial sustainability. NRSP’s dual focus on social mobilization and microcredit is reflected in its branch management structure. In each branch or Field Unit (FU), a Social Organizer (SO) oversees the social mobilization aspects of the program while a Credit Officer (CrO) is in charge of the microcredit program. Field Assistants (FAs), who are at the bottom of this hierarchy, report to the SO for social mobilization and to the CrO for all microcredit related issues. Both SO and CrO in all FUs are paid a flat wage that does not depend on the performance of FAs. FAs are the institution’s front line staff that engage directly with local communities and CO members on a daily basis. They carry out all on-the-ground activities related to NRSP’s two goals. Given borrowers are also CO members, FAs carry out both types of credit and social related activities. Among the social-related activities, FAs facilitate the formation and management of COs, which includes attending CO meetings, ensuring that COs maintain adequate records of meetings, attendance and savings, and gathering requests by CO members for skill training. Among credit- related activities, FAs screen loan applications and assess the creditworthiness of potential borrowers typically at the applicant’s home. FAs are also charged with ensuring timely loan repayment and visits to the home of delinquent borrowers. According to the baseline survey, FAs reported that social activities not only improve the quality of COs, but they also affect credit-related outcomes. More than 90 percent of FAs said that regular and on-time CO meetings improved the creditworthiness of CO members (see Table 1). Similarly, credit activities, such as enforcing strict repayment and following up with borrowers at their homes, may also affect CO quality if such activities discouraged borrowers, especially those in arrears, from 6 Depending on the local norms, CO members may be of the same or mixed gender. 7 All loans are individual loans with no joint liability at the CO level. CO members are eligible to receive new loans even if some members from the CO have overdue amounts. Each borrower is however required to find two guarantors, who can be members of the same CO. NRSP uses these guarantors and other CO members to exert pressure on the defaulting borrower to repay. The loans can be used for the purchase of agricultural inputs, livestock, and investments in household enterprises. A new borrower starts with a maximum loan size of PKR 10,000, which can increase in intervals of up to PKR 5,000 with each successful loan cycle. 6 attending and participating in CO meetings. According to the baseline survey, FAs regularly worked overtime hours, and thus they likely faced trade-offs between devoting their time towards credit- or social-related activities. 8 FAs work individually or in teams of two or three individuals. FAs working in teams co-manage a group of COs by dividing the monthly workload among them. Each team member is responsible for attending the meetings of the COs assigned to him or her that month and for collecting the repayments from CO members due that month. A given CO, therefore, will be managed by different FAs over time. Teamwork among FAs is encouraged by NRSP management to ensure continuity in case the FA falls sick, leaves NRSP, or is promoted. Teamwork also provides NRSP a useful way to train inexperienced FAs while they are on the job, by partnering them with relatively more experienced FAs. A detailed description of team formation is provided in Section 4 and in Appendix C. Until recently, all FAs had been working directly from the branch (FU) with the direct supervision of both the CrO and SO, while earning a fixed salary. At the time of the study, NRSP had transferred 85 percent FAs from FUs to village branches in an effort to have FAs be closer to the field. FAs continued to work alone or in teams, but this decentralization meant that direct supervision by the supervisors was no longer possible. Since the management was concerned that a fixed salary might no longer be optimal in absence of adequate monitoring, NRSP was willing to explore other ways of remunerating the staff. During this time, NRSP had also been scaling up its microcredit program thanks to its partnership with the Pakistan Poverty Alleviation Fund (PPAF), which provided financial support through both grants and loans. The health of the microcredit program was crucial for NRSP’s growth and survival, as the program funds were loans from the PPAF that needed to be repaid. But this microcredit expansion raised concerns among NRSP management about its implications for the quality of COs that were being formed and managed by FAs. 2.1 Bonus intervention We worked with NRSP to design and implement two pay-for-performance schemes for FAs to evaluate the cost and production complementarities on FA effort across credit and social activities. The study was conducted in all 35 FUs located in 15 districts across Sindh, Punjab, and Khyber Pakhtunkhwa provinces, where NRSP was active in March 2005. This provided us with a sample of 162 FAs who were working with NRSP at the time. FAs were randomly divided into three groups (two treatment groups, and one control group). To ensure that all FAs under a given CrO-SO management team were provided with the same bonus scheme, the randomization was done at the FU level. FAs 8 In the baseline survey, the self-reported average daily overtime among study-sample FAs is 2 hours (see Table 1). 7 that were already working from a village branch were also assigned to the treatment of the relevant FU. FAs in the treatment group received one of two bonus schemes. The credit bonus incentivized performance on disbursement and loan recovery. The social bonus incentivized performance on observable correlates of CO quality: new CO formation, regular CO meeting, and savings by CO members. The bonus scheme was designed to be easily understood, fair, and transparent. Each bonus had two triggers. The first trigger determined whether an FA was eligible to receive a bonus while the second trigger determined the bonus amount once the first trigger was achieved. During the intervention, slightly more than 25 percent of treatment FAs qualified for a bonus each month. In any given month, two-fifths of FAs offered the credit bonus and one-fifth of FAs offered the social bonus qualified for it. Appendix Table A.1 describes the triggers and provides more details about the bonus scheme. Appendix Figure A.1 presents the monthly frequency and the amount of bonus payments made during the study period. Since social-related tasks rely more heavily on the discretionary actions of CO members, the social outcomes might be less reflective of FA effort and may be more difficult to achieve. According to the baseline survey, 56.8 percent of FAs find CO quality difficult to improve while only 18.9 percent of FAs report repayment and disbursement difficult to improve. 9 For these reasons, we assigned more branches to the social bonus and control group compared to the number of FUs in the credit bonus. Appendix Table A.2 reports the list of FUs in the study and their bonus assignments, and the timeline of the study is presented in Appendix Figure A.2. NRSP management had been setting monthly targets on several credit and social empowerment outcomes even before the intervention, but prior to our intervention they were not linked to any performance incentive scheme. These targets were based on past performance and were meant to be achievable but were set at a higher level than current performance. Put differently, an FA working with the same intensity as before should not receive a bonus. The intervention only provided monetary incentives to FAs for achieving these targets and it did not affect their career progression or any other aspect of the program. While FAs could be promoted to Senior FA and earn a slightly higher salary, only one of the 162 study FAs had a Master’s degree, the required schooling level to be promoted to SO or CrO. Treatment FAs that had met their monthly target received the incentive as a bonus pay added to their base monthly salary. For FAs working in teams, the bonus was paid based on whether the joint performance of the team exceeded the target. The monthly base salary of an FA was about 9 In addition, social mobilization outcomes like savings and attendance are less costly for CO members to renege on compared to defaulting on their loans. 8 PKR 3,000 (USD 50.54) at the time of the study. 10 The largest bonus an FA could earn in any month was PKR 600 (20 percent of the base salary). FAs in control FUs continued to earn the (flat) base salary. The bonus scheme was announced in March 2005. Treatment FAs became eligible to receive the bonus starting on April 2005. The bonus intervention lasted for 15 months, and ended in June 2006. To discourage any intertemporal substitution of worker effort, FAs in treatment FUs were not informed in advance about whether and when the bonus would end. FAs in control FUs were never told about the intervention, and none of the control FAs interviewed in June 2006 reported having any knowledge of the bonus. 3 Theory Consider an employee of an organization that has two main tasks which produce outcomes y1 and y2 . The employee needs to decide how much effort ( e1 , e2 ) to allocate to each task. Both tasks carry a disutility cost of effort. For simplicity, we assume that the employee is risk neutral. The employee’s utility function is then given by W − C (e1 , e 2 ) where W is the employee’s salary and C (∙) is a convex function in both arguments denoting the disutility cost of effort. Employee effort generates outcome yi , associated with task i = 1, 2, according to the following production technology y1 = θ 1 e 1 + γ 1 e 2 + 1 and y2 = γ2 e 1 + θ 2 e 2 + 2, where ( 1 , 2) is a pair of observational noises. θ1 and θ2 are positive scalars, converting effort on task i into outcome i = 1, 2, and γ1 and γ2 are scalars which capture the effect of effort on task j on outcome i. We assume that θi > |γi |, i = 1, 2. In general, scalars γ1 and γ2 can be positive, negative or zero. The sign of γ1 (γ2 ) determines whether effort on task 2 (1) increases, has no effect on, or decreases outcome y1 (y2 ). If γ1 > 0, the two tasks are complements in producing y1 , in that effort on task 2 increases y1 . Conversely, if γ1 < 0, then the two tasks are substitutes in producing y1 , in that effort on task 2 decreases y1 . Finally, if γ1 = 0, then task 2 does not contribute to the production of y1 . 10 The exchange rate in March 2005 was USD 1 = PKR 59.36. 9 We now introduce a task specific bonus scheme ( b1 , b2 ). An employee offered a bonus for task 1 (2) earns an amount b1 y1 (b2 y2 ) in addition to the base salary w.11 An employee offered a bonus on task 1 chooses e1 and e2 to maximize max b 1 ( θ 1 e 1 + γ1 e2 ) + w − C ( e 1 , e 2 ) e1 ,e2 s.t. e1 ≥ 0, e2 ≥ 0. The first order conditions (incentive constraints) yield b 1 θ1 = C 1 (e 1 , e 2 ) and b1 γ1 ≤ C2 (e 1 , e 2 ) (= if e2 > 0), (1) ∂C (e1 ,e2 ) where Ci (e1 , e2 ) = ∂ei ≥ 0, i = 1, 2. When offered a bonus on task 2, the first order conditions are b2 γ 2 ≤ C 1 ( e 1 , e 2 ) (= if e1 > 0) and b2 θ2 = C 2 (e 1 , e 2 ). (2) Assuming an interior solution ( ei > 0, i = 1, 2) and differentiating (1) and (2) above with respect to bi , i = 1, 2 we obtain ∂b1 + C12 ∂b1 θ1 = C11 ∂e 1 ∂e2 and ∂b1 + C22 ∂b1 γ1 = C21 ∂e 1 ∂e2 ∂b2 + C12 ∂b2 γ2 = C11 ∂e 1 ∂e2 and ∂b2 + C22 ∂b2 θ2 = C21 ∂e 1 ∂e2 ∂2C where the term Cij = ∂ci ∂cj , i, j = 1, 2 is an element of the Hessian of the disutility cost function C (e1 , e2 ). The Hessian matrix is symmetric by definition, that is, C12 = C21 . If an increase in e1 makes e2 costlier in disutility terms, then the reverse must also be true. From the expressions above and simplifying, we obtain C22 C12 ∂e1 ∂b1 = D θ1 − D γ 1 C12 C11 ∂e2 ∂b1 = − D θ1 + D γ1 (3) C22 C12 ∂e1 ∂b2 = D γ2 − D θ2 C12 C11 ∂e2 ∂b2 = − D γ2 + D θ2 where D = C11 C22 − C12 2 > 0 is the determinant of the Hessian of the disutility cost function C (e1 , e2 ). Effort in each task, thus responds to the bonus depending on the sign of the technology scalars 11 The assumption of risk neutrality guarantees that linear incentive schemes are optimal. 10 γi , i = 1, 2, and on the cross-partial of the disutility cost function C12 . In addition, differentiating the production technologies with respect to bi , i = 1, 2 we obtain ∂y1 ∂bi ∂bi + γ1 ∂bi = θ1 ∂e 1 ∂e2 (4) ∂y2 ∂bi ∂bi + θ2 ∂bi = γ2 ∂e 1 ∂e2 The expressions in (4) above allows us to examine how the bonus bi , i = 1, 2 will impact outcomes y1 and y2 . We first note that given the assumptions made, a bonus on task i will always improve ∂yi outcome i. That is, ∂bi > 0, i = 1, 2. This can be verified by substituting in for ∂ei ∂bi > 0 and ∂ei ∂bj (i, j = 1, 2), and simplifying terms. Second, the impact of a bonus on task i on outcome j will, in general, depend on the nature of cost and production complementarities, that is, on the signs of C12 and the technology scalars γ1 and γ2 . We now explore each case in turn. Case 1: γ1 > 0 and γ2 > 0 The incentive constraints in (1) and (2) suggest that e1 > 0 and e2 > 0 irrespective of the bonus offered. The impact of either bonus on the other outcome is found by substituting the expressions for ∂ei ∂bj , i, j = 1, 2 from (3) in (4) and simplifying. We obtain ∂y2 ∂y1 C C C = = θ1 γ2 22 − (θ1 θ2 + γ1 γ2 ) 12 + θ2 γ1 11 (5) ∂b1 ∂b2 D D D The impact of a bonus on task 1 on outcome 2 is the same as the impact of a bonus on task 2 on outcome 1. When tasks are substitutes in the disutility cost of effort, that is, C12 > 0, such that an increase in e1 makes e2 costlier in disutility terms (and vice versa), the impact cannot be signed because the first and last terms are positive while the second term is negative. Intuitively, there are two opposing forces. On the one hand, effort in a particular task increases the disutility cost of effort in the other task because C12 > 0, and so effort on the unrewarded task may decline as the individual increases effort on the rewarded task, thus reducing production of the unrewarded outcome. On the other hand, the production complementarities ( γi > 0, i = 1, 2), suggest that effort in the rewarded task increases production of the unrewarded outcome. The net effect is therefore ambiguous. Alternatively, when tasks are complements in the disutility cost of effort, that is, C12 ≤ 0, then the impact is positive as the second term of the expression in (5) is now non-negative. In this case, effort in a particular task decreases the disutility cost of effort on the other task and so effort in the unrewarded task will increase with an increase in effort in the rewarded task. 11 Case 2: γ1 > 0 and γ2 ≤ 0 The incentive constraints in (1) suggest that e1 > 0 and e2 > 0 when a bonus on task 1 is offered. The impact of the bonus on task 1 on outcome 2 is given therefore by expression (5) above and is ambiguous.12 In contrast, the incentive constraints in (2) suggest that e1 = 0 and e2 > 0 when a bonus on task 2 is offered. In this case, the impact of the bonus on task 2 on outcome 1 is given by ∂y1 C C = γ1 θ2 11 − γ2 12 > 0, (6) ∂b2 D D and is positive because the expression in parenthesis is positive and thus the impact of the bonus has the same sign as γ1 . Intuitively, the bonus on task 2 has no direct impact on the effort on task 1 (that is, e1 = 0), yet outcome 1 will increase because of the production complementarity (scalar γ1 > 0). Case 3: γ1 ≤ 0 and γ2 > 0 This case is the opposite of Case 2. The incentive constraints in (1) suggest that e1 > 0 and e2 = 0 and so the impact of the bonus on task 1 on outcome 2 is then given by ∂y2 C C = γ2 θ1 11 − γ1 12 > 0, (7) ∂b1 D D which is positive because the expression in parenthesis is positive and thus the impact of the bonus has the same sign as γ2 . Intuitively, the bonus on task 1 has no direct impact on the effort on task 2 (that is, e2 = 0), and as a result, outcome 2 will increase because tasks are strategic complements in the production of outcome 2 (scalar γ2 > 0). The incentive constraints in (2) suggest that e1 > 0 and e2 > 0 when a bonus on task 2 is offered. The impact of the bonus on task 2 on outcome 1 is given therefore by expression (5) above and is ambiguous because the last term is non-positive and the second term will depend on the sign of C12 . Case 4: γ1 ≤ 0 and γ2 ≤ 0 The incentive constraints in (1) suggest that e1 > 0 and e2 = 0 when a bonus on task 1 is offered while those in (2) suggest that e1 = 0 and e2 > 0 when a bonus on task 2 is offered. In this case, the impact of the bonus on tasks 1 and 2 is non-positive and given in expressions (7) and (6), respectively. If the two tasks are substitutes in the production of each outcome, then the bonus will increase effort on the rewarded task thus decreasing the unrewarded outcome. 12 Given the assumptions, it is always the case that θ θ + γ γ > 0, but the second term of the expression in (5) 1 2 1 2 can be positive or negative depending on the sign of C12 . 12 3.1 Discussion Having explored all four cases of the technology scalars γi , i = 1, 2, Table 2 provides the predictions for how the bonuses affect the outcomes, separately for the case when C12 ≤ 0 (Panel A) and C12 > 0 (Panel B). A positive (negative) sign indicates that the impact of the bonus on the outcome is positive (negative). A question mark indicates that the impact is ambiguous. Several important patterns emerge in Table 2. First, the impact of a bonus on task i on outcome i is always positive. Second, each bonus affects the two goals differently depending on the assumptions about task complementarity. In particular, only when the technology scalars have different signs (in Cases 2 and 3) can the impact of the bonus on task 1 on outcome 2 have a different sign from that of the bonus on task 2 on outcome 1. In contrast, in Cases 1 and 4, the impact of the bonus on task 1 on outcome 2 will have the same sign as the impact of the bonus on task 2 on outcome 1. Finally, except for Case 1 where both technology scalars are positive, the impact of the bonus does not depend on the sign of the cross-partial of the disutility cost of effort C12 . In Section 6 we compare the actual impacts of the bonuses introduced in the field experiment with these predictions. Consider now a bonus that rewards both tasks simultaneously. The employee would choose e1 and e2 to maximize max b 1 ( θ 1 e 1 + γ 1 e 2 ) + b2 ( θ2 e2 + γ 2 e 1 ) + w − C ( e1 , e 2 ) e1 ,e2 s.t. e1 ≥ 0, e2 ≥ 0. The first order conditions (incentive constraints) yield b1 θ1 + b2 γ2 = C1 (e1 , e2 ) and b2 θ2 + b1 γ1 = C2 ( e 1 , e 2 ) , (8) The organization seeks to maximize the total surplus given by the sum of the organization’s objective function and the employee’s welfare. Letting the scalars αi , i = 1, 2 be the weight that the organization puts on outcome i, the problem of the organization is max α 1 (θ 1 e 1 + γ1 e2 ) + α 2 ( θ2 e2 + γ 2 e 1 ) − C (e1 , e 2 ) e1 ,e2 subject to the incentive constraints in (8) This problem is similar to the one solved by the employee above, and it is clear that the optimal bonus satisfies bi = αi , i = 1, 2, that is, the institution should reward a task inasmuch as it values 13 the outcome. The optimal bonus therefore does not depend on the nature of production and cost complementarities since the employee will internalize them when choosing the optimal level of effort. As a result, the only way to assess complementarities in production and disutility cost is by offering a bonus scheme that only rewards one task. 4 Data Data used in the empirical analysis come from multiple sources. Survey data were collected between January and February 2005, prior to the announcement of the bonus, and in June 2006, the last month of the intervention. These surveys asked each FA about his or her demographic and household characteristics, current employment conditions and work history, along with his or her level of motivation for working with NRSP. These two rounds of survey data are supplemented with administrative data from NRSP, in- cluding the monthly employee records with the employment status of each FA, salary and bonus information, and the name of FU or village branch where the FA worked. NRSP also provided us with a monthly record of COs managed (or co-managed) by each FA from June 2004 to June 2006. This FA-CO panel helps us construct the monthly portfolio of COs for each FA during the 10 months before and the 15 months of the bonus intervention. FA’s performance was tracked using two administrative datasets. Data on loan disbursements and recovery are obtained from NRSP’s Management Information Systems (MIS) database. The MIS digitally records all loans taken and repaid by all borrowing CO members by installment. A total of 5,364 unique COs appear in the MIS database during the 25 months that overlap with the FA-CO panel. Of them, 4,404 COs (82.1 percent) were managed by the 162 FAs who were working for NRSP at the time of the study. 13 Out of the 4,404 COs managed by the study-sample FAs, 4,008 COs show loan activity at least once during these 25 months, and have 5.81 active borrowers each month with repayment or disbursement in 14 out of the 25 months, on average. Information on social mobilization efforts is obtained from the Monthly Progress Reports (MPRs) submitted each month by each FA for the COs managed or co-managed by him or her. This re- port includes information on meeting attendance, member savings, and loans approved and denied during the meeting. The MPRs data are available for the 15 months when the bonus was imple- mented. These data were verified by a supervisor through random visits to a subset of scheduled 13 The rest of the COs were managed and formed by FAs hired after the bonus intervention began in March 2004. In our analysis, we focus on the FAs who were already working with NRSP prior to the bonus intervention because any differential selection of new hires on their characteristics/quality across the different experimental arms could confound the results. While financial incentives have also been shown to affect performance through differential selection into working in an organization (B´ o et al., 2013; Deserranno, 2016), the main focus of this paper is to understand the incentive effects on performance due to task complementarities. 14 CO meetings.14 We aggregate the CO level credit and social outcomes at the FA-month level using information from the FA-CO panel. We calculate the performance of an FA before and after the bonus, by taking the average across the 9 months prior to the bonus announcement and the 15 months during the bonus period, respectively. Data from March 2005 (the month when the bonus was announced but not yet implemented) is dropped from the analysis. In addition, CO members from a subset of COs managed by our study-sample FAs were in- terviewed in November 2006 (5 months after the end of our study) as part of a baseline survey for e and Mansuri (2017)). It covered 11 FUs, and interviewed 1691 CO members another study (see Gin´ from 214 COs managed by 57 (out of 162) study-sample FAs. The survey asked CO members about any changes since around the time the bonuses were introduced in their COs’ activities—such as discussions of non-credit related social issues during CO meetings, member’s ability to speak freely and actively participate in CO decisions, willingness to seek advice from CO leaders outside of CO meetings, and collective action taken by CO members to jointly purchase agricultural inputs or sell the harvest. We use these data to construct subjective quality measures of a CO as reported by its members. Lastly, we interviewed supervisors in June 2006 (last month of the bonus). The survey asked supervisors about the performance of all FAs working under them on various credit and social outcomes in the previous month. They also reported their subjective evaluation of each FA since the study began. We use the data from the supervisor survey to construct measures of supervisory effort, and of their assessment of FA performance. 4.1 Baseline characteristics and balance tests Columns 1, 2, and 3 of Table 1 report means of FA characteristics measured at baseline in the control, credit bonus, and social bonus groups, respectively. Columns 4, 5, and 6 report the p-value from the t-test of the difference in means between control FAs and credit bonus FAs (Cr-FAs); control FAs and social bonus FAs (Soc-FAs); and Cr-FAs and Soc-FAs, respectively. Across all the reported variables, we cannot reject that means are equal for any pairwise comparison at conventional levels of statistical significance. As indicated by the p-value of F-test at the bottom of Columns 4, 5, and 6, we again cannot reject that all covariates are not jointly different from zero in a regression where the dependent variable takes the value one if the FA is in the control group using the sample of FAs in the control and credit bonus groups in Column 4 and FAs in the control and social bonus groups 14 The random visits were carried out by the Credit Officer (CrO) to whom FAs report all issues related to micro- credit. All FAs working from the same FU report to the same CrO. 15 in Column 5. In Column 6, the dependent variable takes value 1 if the FA is offered a credit bonus using the sample of FAs offered a bonus. FAs in our study are on average 28 years old, roughly one-fourth are female, and slightly more than half of them have at least a high school degree (equivalent to 12 years of education). The average duration of employment with NRSP is 26 months, and NRSP was the first job for roughly two-fifths of FAs. FAs manage on average 14 COs every month. Their average monthly portfolio consists of 91.4 active loans (new and ongoing), with roughly PKR 100,000 (USD 1,685) disbursed each month. The mean recovery rate on installments due at the end of each month is around 98 percent, while only 70 percent of such installments is recovered fully by the 20th of that month. Slightly more than half of the FAs prefer a hypothetical bonus to be paid on credit outcomes as opposed to social outcomes. During the baseline interview, each FA was asked to rank what they liked most about working in NRSP. Roughly half of the FAs reported that the ability to help people is what they liked most. One-fifth of them had also done volunteer work before joining NRSP. Out of 162 FAs in the study, 132 FAs were successfully interviewed in the follow up survey. This attrition rate is almost identical and not statistically different across the three groups (see bottom of Table 1), suggesting that attrition bias is not a concern when examining impacts of bonus on outcomes measured in the follow up survey. In addition, CO meetings held by 31 FAs were never visited by a supervisor. These 31 FAs do not appear in the verified MPRS data, restricting our sample to 131 FAs when examining social outcomes. The selection into this restricted sample is again not statistically different across control FAs, Cr-FAs, and Soc-FAs (see bottom of Table 1). Similarly, 73 FAs for whom we have subjective assessments from supervisors are not differentially selected across the two bonus and control groups. Appendix Tables D.1, D.2, and D.3 present means of baseline variables in the control and the two bonus groups, and their differences for the verified MPRS, follow up, and supervisor evaluation restricted samples, respectively. None of the differences in means (out of 63 differences) in Appendix Table D.1 are statistically significant at the 10 percent conventional level. In Appendix Tables D.2 and D.3, four and three differences are statistically significant at the 10 percent level respectively. In all samples, the F-tests at the bottom of Columns 4-6 cannot reject the hypothesis that all variables are jointly insignificant in explaining assignment to an experimental arm. 4.2 Partnership Roughly two-fifths of FAs co-manage their entire CO portfolio with other FAs, while slightly less than one-fourth manage all their COs on their own. Appendix Figure C.1 plots the distribution 16 of FAs based on the share of their CO portfolio that are co-managed with other FAs during the 9 months prior to the bonus announcement. The median level of co-management is 73 percent, and we classify FAs who co-manage more than this median value as “partnered” FAs in the analysis. 15 Columns 1 and 2 of Appendix Table C.1 report the means of FAs’ baseline characteristics for non-partnered and partnered FAs respectively; Column 3 reports the p-values from the F-tests of the difference in means between the two groups. We find no statistically significant difference in any reported characteristics including education and work experience between partnered and non- partnered FAs. The p-value of the F-test (reported at the bottom of Column 3) cannot reject the equality of means across the two groups. Column 4 presents the correlation between the FA’s and his/her partner’s characteristics. Part- nerships are mainly formed between FAs of the same gender and the same level of education (cor- relation coefficients are 0.834 and 0.309 respectively). In the analysis, we take partnerships formed prior to the bonus as given. In Table 1, we do not find statistically significant difference in the propensity to work in a team prior to bonus nor in the share of co-managed COs between FAs in the different experimental arms. Appendix C contains a more detailed discussion on partnership. 5 Empirical Strategy Because the bonus assignment is random, we can estimate the causal impact of introducing the bonus scheme by estimating OLS with the following specification: Yi,1 = ηr + ψ Yi,0 + βC T Ci + βS T Si + (9) where Yi,1 is the post-treatment outcome of interest for FA i, ηr is a region dummy (one for each of the four NRSP’s administrative regions), and Yi,0 is the pre-treatment outcome for FA i. T Ci (T Si ) is an indicator variable that takes the value of one if FA i was offered the credit (social) bonus, and zero otherwise, and is a mean-zero error term. Because the offer of bonuses was done at the FU level and there are 35 FUs in the study, we conduct statistical inference using the t-asymptotic wild cluster bootstrap at the FU level described in Cameron et al. (2008). All tables discussed in Section 6 report the coefficients and the bootstrapped p-values. The coefficients of interest in the regression are βC and βS , which estimate the average treatment effects of the credit and social bonus on FA outcomes Yi,1 , respectively. 15 In the baseline, almost 90 percent of partnered FAs (71 out of 81 partnered FAs) co-manage their COs with one other FA, while the rest co-manage with two other FAs. We also find that FA teams are stable throughout the study period, unless one of the team members quits NRSP. 17 We also examine the impact of bonus separately for partnered and non-partnered FAs using the following specification: Yi,1 = ηr + ψ Yi,0 + π Pi + βC T Ci + βS T Si + δC Pi ∗ T Ci + δS Pi ∗ T Si + , (10) where Pi is an indicator variable that takes the value of one if FA i was partnered with another FA at the time the bonuses were introduced. The coefficients δC and δS on the interaction terms Pi ∗ T Ci and Pi ∗ T Si respectively, capture the differential impact of the bonus on FAs that work in a team, relative to those that work alone. The coefficients βC and βS estimate the impact of credit and social bonus, respectively, on FAs that work alone, while the sum of the coefficients βC +δC and βS +δS estimate their impact on partnered FAs. 6 Results Table 3 examines the effects of the credit and social bonus on the FA performance on microcredit outcomes. Columns 1 and 2 present the impact on the two outcomes that were directly incentivized: number of active loans and repayment by the 20th of the month. The number of active loans increased by 24.1, and the repayment improved by 8.4 percentage points for FAs offered the credit bonus (Cr-FAs) compared to control FAs. Both estimates are statistically significant at the 10 percent level. The size of these impacts is large, amounting to a 20 percent increase in active loans and a 12 percent improvement in repayment from the mean performance of control FAs. The impacts of the social bonus on the two trigger variables of the credit bonus are small (9.406 and -0.020), and not different from zero at conventional levels of statistical significance. Columns 3-5 estimate the impacts of the bonuses on other credit outcomes not directly incen- tivized: number of new loans, disbursement amount, and repayment by end of the month. In contrast to the impacts on the two trigger variables, Cr-FAs showed no improvements on any of these non-incentivized credit outcomes. In fact, the improved repayment rate at the 20th of the month made little difference to the repayment rates by the end of the month, partly because the end of the month repayment rates were already above 96 percent among control FAs. The impacts on new loans and disbursement amount are also small in magnitude (6.1 and 0.21 percent respectively compared to the means in the control group), and are not statistically significantly different from zero. The performance of Cr-FAs is also not statistically significantly different from that of Soc-FAs on any of these non-incentivized outcomes. Since the outcomes in Columns 1-5 of Table 3 may be correlated, we follow Kling et al. (2007) to 18 account for the problem of multiple hypothesis testing and construct a summary index that aggre- gates information over multiple outcomes. The credit index in Table 3, Column 6 is calculated by taking an equally weighted average across the standard distributions of all five microcredit outcomes. The impact of credit bonus on this index is positive and large, suggesting that Cr-FAs performed 0.238 standard deviations higher (on the credit index) than control FAs. However, it is not statis- tically significant at conventional levels. Moreover, this impact is largely due to improvements in the two specific outcomes that were directly incentivized by the credit bonus, rather than a general increase in performance in all credit-related tasks. The impact of the social bonus is virtually zero and thus not statistically significant. Table 4 estimates the impact of the bonuses on the objective measures used by NRSP to evaluate the quality of COs and social empowerment. Columns 1-3 report the impacts on the three measures directly incentivized and used as triggers for the social bonus, while Columns 4-6 report the impacts on other social measures not directly incentivized. We find that Soc-FAs formed 0.225 more new COs per month than control FAs (Column 1). The estimate is statistically significant at the 1 percent level, and amounts to a 58.6 percent increase in CO formation compared to the mean of 0.384 in the control group. It is however almost identical to and also not significantly different from Cr-FAs, who also increased CO formation by 0.284 compared to control FAs. This is not entirely unexpected since CO membership is a prerequisite for applying for microcredit loans. While Cr-FAs and Soc-FAs increased new CO formation relative to control FAs, the impact of the credit bonus on the rest of CO quality outcomes in Columns 2-6 is negative and large in magnitude among Cr-FAs. Cr-FAs decrease the share of savers in CO meetings by 12.7 percentage points, worsen attendance by 10.5 percentage points, and reduce the share of COs with multiple meetings in a month by 19.4 percentage points relative to control FAs. The estimates are statistically different from zero at the 10 percent level, and amount to 18.2, 13.5, and 46.0 percent decline in savings, attendance, and meetings relative to the mean of controls, respectively. This suggests a large negative effect of the credit bonus on social-related activities. In contrast, relative to control FAs we find no change in these measures of CO quality among Soc- FAs. The impacts of social bonus on the share of savers among CO members and their attendance in CO meetings, which make up the remaining triggers for the social bonus, are small (-0.03 and -0.027 respectively) and not statistically different from zero at conventional levels. But they are statistically different from those of Cr-FAs (p-values are 0.110 and 0.024, respectively). 16 For non-incentivized 16 We note that attendance levels among control FAs are so high (close to 80 percent) that Soc-FAs would have qualified for a positive bonus amount conditional on meeting their first trigger targets without any change in atten- dance. 19 social outcomes in Columns 4-6, the performance of Soc-FAs is not significantly different from those of control FAs. We construct a CO quality index similar to the credit index, by taking an equally weighted average of all six measures of CO quality standardized with mean zero and standard deviation one.17 The effect of social bonus on the CO quality index in Column 7 is slightly negative (-0.051 standard deviations) but not statistically different from zero. For Cr-FAs, however, this index is large and negative, and statistically significantly different from zero at the 1 percent level. Cr-FAs also performed worse on the CO quality index as compared to Soc-FAs (p-value is 0.014). Appendix Figure E.1 plots the bonus effects on FA outcomes (credit and social indices) month-by-month for the 15-month bonus period. In contrast to Jayaraman et al. (2016) that find a large productivity response immediately following a change in the incentive contract followed by a decline four months after the change, the plots in Appendix Figure E.1 suggest that monthly effects of the bonus were fairly consistent across the 15 months and similar in size. Table 5 examines the effects of the bonuses on subjective measures of CO quality constructed from a survey of CO members. Among the sample of CO members who were interviewed, we do not find any statistically significant difference on client characteristics between the two treatment arms and the control (see Appendix Table D.4). These client-level data provide us with complementary measures of CO quality and more detailed information on the CO and its members’ activities that are directly related to community mobilization and social empowerment. Client data were only collected for a subset of COs in 10 of the 35 FUs as part of the baseline for another study. The group of clients in each arm has similar client characteristics (see Appendix Table D.4). In Columns 3 and 4 of Table 5, we find that clients of Soc-FAs are more likely to engage in buying and selling agricultural inputs and outputs collectively with others in their village and that they are more likely to turn to their CO leaders for help or advice. The estimated impacts on these two outcomes are large (50.0 and 68.6 percent respectively compared to the control means), and they are statistically significant at the 1 percent level. Additionally, members of COs managed by Soc-FAs increased active participation in CO decisions and are also more likely to discuss non-credit related social issues during CO meetings. While these effects are not statistically significant at conventional levels, the point estimates however are not trivial in magnitude (77.4 and 86.9 percent respectively compared to their means in the control group). In contrast, the estimates of the credit bonus in Columns 1-4 are close to zero and not statistically significant, but the standard errors are large and thus the difference between the impact of the credit compared to the social bonus is not statistically significant. 17 While calculating the CO quality index, the sign of one of the outcomes, i.e. “Dead COs,” is reversed so that for all the outcomes, positive values represent an increase in social empowerment. 20 The effect of the social bonus on the empowerment index, constructed by taking an equally weighted mean of the four outcomes in Columns 1-4 of Table 5, is positive by 13.1 percentage points and suggests a 72.0 percent improvement compared to the control mean. It is also statistically significant at the 10 percent level. In comparison, the effect of credit bonus on this index is about half in magnitude (6.7 percentage points) and not statistically significantly different from zero. The difference in impact between the social and credit bonus is significant at conventional levels (p-value is 0.128). Finally, we examine whether the incentive scheme might have had an effect on supervisory effort and management quality by using the data from the supervisor survey. Our measure of supervisor effort is the absolute difference between the actual performance of FAs during the month of the survey and that reported by the supervisor on two credit outcomes (number of active loans and repayment) and one social outcome (savings). 18 Appendix Table E.3 reports the results. On both credit and social outcomes, the estimated effects of credit and social bonuses are close to zero and not statistically significantly different from zero at conventional levels. In other words, the introduction of the bonus does not influence the supervisors’ ability to correctly report their FA’s true performance on incentivized and non-incentivized outcomes. Overall, these results suggest that while the credit bonus improved the NRSP’s microcredit program, albeit only for outcomes directly incentivized, it also worsened the quality of COs thus undermining NRSP’s goal of empowering communities through social mobilization. In comparison, the social bonus increased CO formation without worsening the objective measures of CO quality and without adversely affecting microcredit outcomes. It also improved subjective measures of CO quality and client empowerment: COs managed by Soc-FAs are more cohesive and more likely to work collectively on economic and social activities. The negative and significant impact of the credit bonus on social outcomes and the non-negative impact of the social bonus on credit outcomes is consistent with Case 2 of the model described in Section 3 where task 1 refers to credit-related activities and task 2 to social-related activities. In this case, γ1 > 0 suggests that outcome 1, or a healthy credit portfolio, is more easily achieved when FAs work hard on organizing new COs or ensuring that existing COs are cohesive while γ2 ≤ 0 suggests that enforcing repayment discipline may discourage borrowers, perhaps those in arrears, ∂y2 from attending CO meetings thus undermining the social goal. With γ2 ≤ 0, the expression for ∂b1 will tend to be negative, for example, if the technology term γ1 > 0 is small and C12 > 0.19 18 Out of 162 study-sample FAs, we have supervisor cross-reports for 98 FAs (and for 55 FAs out of 132 FAs who also show up in the verified MPRs sample). We do not find evidence for a differential selection of FAs into the two restricted samples by treatment assignment. 19 This refers to the case presented in Panel B ( C 12 > 0), row 2 (γ1 > 0 and γ2 ≤ 0) of Table 2. 21 6.1 Impact of bonus by partnership Given that NRSP relies on teamwork among FAs and that the impacts of the bonuses may differ depending on whether FAs work alone or in teams, we examine the differential effects of the bonuses by FA partnership status in the baseline. Table 6, Columns 1 and 2 present the differential impacts on the credit and CO quality indices, respectively. The differential effects on the full sets of credit and social outcomes are presented in Appendix Tables E.1 and E.2. In Column 1 of Table 6, non-partnered Cr-FAs performed 0.277 standard deviations higher than non-partnered control FAs on the credit index (statistically significant at the 10 percent level). The improvement of credit-related outcomes when credit-related activities are incentivized is perhaps unsurprising and predicted by the model in Section 3. In Column 2, Cr-FAs performed 0.382 stan- dard deviations worse on the CO quality index than control FAs, and this difference is statistically significant at the 1 percent level. Non-partnered Soc-FAs also performed 0.297 standard deviations higher on the credit index compared to control FAs (Column 1 of Table 6). The estimate is large and statistically significant at the 10 percent level. In Column 2, non-partnered Soc-FAs perform 0.122 standard deviations higher than control FAs on the CO quality index. While this estimate is not statistically significant at conventional levels, the results are consistent with a positive impact of the social bonus on both the credit and social goals. More importantly, while the difference in the credit index between non-partnered Soc-FAs and Cr-FAs is negligible (0.02 standard deviation) and not statistically significant at any conventional levels, non-partnered Soc-FAs outperformed Cr-FAs on the CO quality index, and this difference is statistically significant at the 1 percent level. These estimates for non-partnered FAs are similar to those in Column 6 of Table 3 and Column 7 of Table 4 but are more precisely estimated. According to Columns 1 and 2 of Table 6, partnered Cr-FAs performed better in credit-related tasks and worse in social-related tasks than control FAs, but the differences are not statistically significant at conventional levels (p-values are 0.306 and 0.168 for the credit and social index, re- spectively). In contrast, partnered Soc-FAs performed 0.269 standard deviations lower in the credit index compared to partnered control FAs. The effect is statistically significant at the 12 percent level. They also performed statistically significantly worse than partnered Cr-FAs (p-value is 0.034). In Column 2, partnered Soc-FAs also underperformed compared to control FAs by 0.182 standard deviations, but the estimate is not statistically significant (p-value is 0.186). Overall, partnered Cr-FAs performed as well as control FA, while the performance of partnered Soc-FAs significantly worsened. But why would the performance of Soc-FAs working in teams suffer 22 under the social bonus? Following Osterloh and Frey (2000) among others, one may assume that rewarding effort in the social-related task may have undermined the intrinsic motivation of FAs and exacerbate free-riding. 20 Because the effort of each FA working in a team is unobservable to other team members, and social outcomes are less reflective of FA effort, Soc-FAs who were offered the social bonus could have experienced, on average, a significant decline in intrinsic motivation and this decline could have exacerbated free-riding in teams. In contrast, the credit bonus may not have affected negatively the intrinsic motivation nor teamwork among Cr-FAs. In the following subsection we provide evidence of the claim that the social bonus may undermine the intrinsic motivation of Soc-FAs. 6.2 Intrinsic motivation and teamwork Columns 1-3 of Table 7 present three different measures of intrinsic motivation. In the baseline and follow up surveys, FAs were asked to list the things that they liked most about working with NRSP. Column 1 thus reports a dummy variable that equals one if the FA mentioned the ability to help people as what they like most about NRSP. According to this measure, Soc-FAs are 23.4 percentage points less likely to be intrinsically motivated in the follow up survey, compared to control FAs. This decline in intrinsic motivation due to the introduction of the social bonus is statistically significant (p-value is 0.022). In contrast, the decline in intrinsic motivation for Cr-FAs is small (almost half in magnitude), and not statistically significant at conventional levels. In the follow up survey, FAs were also asked whether they identified with NRSP’s mission and whether they found their work with NRSP satisfying and important. The impacts of bonus on these two alternative measures are presented in Columns 2 and 3, respectively. Soc-FAs are 15.2 and 13.6 percentage points less likely to report being intrinsically motivated compared to control FAs on these two additional measures. The estimates are however not statistically significant at conventional levels, although they are large in magnitude and both indicate a 35 percent decline in intrinsic motivation compared to the control means. In contrast, the impacts on Cr-FAs are negligible and not statistically significant. The standard errors are again large and thus we cannot reject that the impacts of the bonuses in Columns 1-3 are different from each other. In Column 4, we construct a motivation index by taking an equally weighted mean of the three dummy variables in Columns 1-3. Based on this index, the social bonus decreased intrinsic motiva- tion by 16.7 percentage points (p-value is 0.074). In contrast, the impact of the credit bonus is close 20 More formally, one may assume that a motivated individual can exert effort e = e > 0 at no cost, that is, 2 C (0, e) = 0. Individuals are motivated so long as they are not offered a social bonus. Under a social bonus, the ˜ (e1 , e2 ) such that C individual is no longer motivated and facing a disutility cost of effort C ˜ (0, e) > 0. We assume that C12 and C ˜cs have the same sign. 23 to zero at -0.060 and not statistically significant. More importantly, the impacts of the two bonuses on the motivation index are statistically different from each other (p-value is 0.084). These results provide evidence that incentivizing effort on the social-related tasks likely under- mined FAs’ intrinsic motivation. Column 3 of Table 6 shows that partnered and non-partnered Soc-FAs alike experienced a statistically significant decline in the motivation index (p-values are 0.094 and 0.134, respectively) and that this decline does not vary by partnership (p-value = 0.522). We next examine the FA’s propensity to work in a team after bonuses were introduced. 21 An FA could change partnership status by the end of the intervention either because he or she requested to co-manage a smaller share of COs, work alone or because the supervisors decided to break-up a team. Column 5 of Table 7 reports a decline in the share of partnered Soc-FAs of 12.5 percentage points compared to control FAs. This negative impact of the social bonus on the likelihood of partnership is statistically significant at the 1 percent level, and implies roughly a 20 percent decline in the share of Soc-FAs working in teams relative to the mean in the control group. This effect is also statistically different from Cr-FAs at the 10 percent level. The credit bonus had no impact on the propensity of Cr-FAs to work in teams. In Column 6 of Table 7, we construct a measure of free riding in teams based on an incentivized trust game played by FAs in the follow up survey. Upon receiving an amount from another randomly chosen and unknown FA, each FA was asked to send back to the unkown FA some of the money. The average amount received was PKR 58.52. The dependent variable “Shares with a partner” is a dummy that takes value one if the FA sent some amount back to his or her randomly chosen partner, and zero otherwise. The effects of credit and social bonus on free riding are small in magnitude, and the estimates are not statistically different from zero and from each other at conventional levels. This is perhaps not surprising, given that the effects on free riding are likely to be concentrated among partnered FAs. In Column 5 of Table 6, we estimate the impact of credit and social bonus on free riding separately for partnered and non-partnered FAs. As expected, among partnered FAs, the social bonus more than doubled the share of FAs who did not send any money back to their partners, suggesting a considerable increase in free riding among partnered Soc-FAs (statistically significant at the 10 percent level). For non-partnered FAs however the social bonus did not have any impact on free-riding behavior. The difference in free riding between partnered and non-partnered Soc-FAs is statistically significant at the 5 percent level. Additionally, we do not observe changes in free riding behavior among neither partnered nor non-partnered Cr-FAs. In sum, the social bonus may have undermined the intrinsic motivation of employees. Such 21 We consider an FA as partnered in the months after the intervention if the share of COs that are co-managed with other FAs during the treatment months exceeds the pre-treatment median value of co-sharing (73 percent of COs). 24 decline in motivation, in turn, may have affected negatively the ability of Soc-FAs to perform in teams by worsening the free riding problem. These results provide suggestive evidence of this increased propensity to free ride, a decline in teamwork and highlight the role of intrinsic motivation for workers in mission-oriented organizations that work in teams. 6.3 Performance assessment by supervisors Table 8 estimates the impacts of the credit and social bonus on the assessment of FA performance provided by their supervisors. At the end of the study, supervisors of all study FAs were asked to evaluate each FA currently working under them on three specific dimensions: (1) the likelihood of being promoted to Senior FA; (2) perceived improvements in loan disbursement rates since the introduction of the bonus scheme; and (3) perceived improvements in savings by CO members. We use these measures to construct a supervisor assessment score by taking an equally weighted average of the three outcomes. Odd-numbered columns in Table 8 present the average treatment effects of the bonus on the three outcomes and on the assessment score; even numbered columns present the results by partnership status. In Column 7 supervisors of Soc-FAs increased their assessment by 13.7 percentage points com- pared to supervisors of control FAs. While the estimate is not statistically different from zero, the magnitude of the effect implies a 46.3 percent increase over the control mean. In comparison, super- visors of Cr-FAs increased their assessment score, on average, by only 5 percentage points compared to the assessment of control FAs, and this effect is also not statistically significant at conventional levels. Column 8 reports the supervisor’s assessment of Column 7 by partnership status. Supervisors increased their assessment of non-partnered Soc-FAs by 35.2 percentage points compared to the assessment of control non-partnered FAs. The estimate is statistically significant (p-value is 0.072). Supervisors also increased their assessment of non-partnered Soc-FAs by 45.2 percentage points relative to partnered Soc-FAs in the same FU. This difference is statistically significant (p-value is 0.072). Partnered and non-partnered Cr-FAs, on the other hand, scored 2.8 percentage points lower and 15.0 percentage points higher than their counterparts in control FUs, respectively. None of these estimates for Cr-FAs are statistically significantly different from zero, and from each other at conventional levels. These results are consistent and strengthen our previous findings. Table 6 reported that the social bonus improved performance in both credit- and social-related activities of non-partnered FAs, while partnered Soc-FAs performed significantly worse. The assessment of Soc-FAs by their supervisors also reflects these results. On the other hand, the impacts of the credit bonus on the 25 supervisor assessment of Cr-FAs, who improved their performance on microcredit but worsened their social outcomes, are mixed, and not very different in magnitude nor statistical power from that of FAs in control FUs. 7 Conclusion In this paper, we provide evidence of the role of cost and production complementarities in an organization with multiple goals. From a methodological perspective, an assessment of whether cost and production complemen- tarities exist requires researchers to isolate the impact of rewarding one task on the other outcome, ∂yi that is, ∂bj , i = j and this is only feasible with an experimental design that incentivizes one task at a time. Comparing the optimal contract that rewards both tasks simultaneously to a fixed wage contract would perhaps yield improvements in both outcomes but would tell us little about existing complementarities. Although cost complementarities may exist, e.g. effort on a particular task may increase the cost of exerting effort on another task, we uncover important production complementarities as well. The credit bonus unsurprisingly improves the performance of the microcredit program but worsens CO quality and thus social empowerment. In contrast, at least among staff working alone, the social bonus improves social outcomes and is as effective as the credit bonus at improving credit outcomes. This result cannot obtain with disutility cost complementarities alone. Given that the social bonus undermined the intrinsic motivation of all employees regardless of whether they worked alone or in teams, and it increased the propensity to free-ride among those working in teams, the results suggest that rather than incentivizing both tasks simultaneously, a fixed wage may be optimal. In fact, shortly after the study concluded NRSP stopped collecting the data needed to pay the social bonuses. Our results highlight the challenge that organizations with multiple goals face. We take as an example a development organization where an emphasis on sustainability may undermine its overarching development mission. As anti-poverty programs become increasingly multi-faceted by including microcredit, asset transfers, skill training and other interventions and are delivered through a community participation driven approach, the results suggest that the effectiveness of such pro- grams rely not only on how these interventions are designed but also on how the delivery of different services is incentivized. 26 References Ashraf, Nava, Oriana Bandiera, and B. Kelsey Jack , “No Margin, No Mission? A Field Experiment on Incentives for Public Service Delivery,” Journal of Public Economics , 2014, 120, 1-17. Aubert, C´ ecile, Alain de Janvry, and Elisabeth Sadoulet , “Designing Credit Agent Incen- tives to Prevent Mission Drift in Pro-poor Microfinance Institutions,” Journal of Development Economics, 2009, 90, 153-162. Bandiera, Oriana, Iwan Barankay, and Imran Rasul , “Incentives for Managers and Inequality Among Workers: Evidence From a Firm-Level Experiment,” Quarterly Journal of Economics , 2007, 122(2), 729-773. Bandiera, Oriana, Iwan Barankay, and Imran Rasul , “Social Incentives in the Workplace,” Review of Economic Studies , 2010, 77(2), 417-458. Bandiera, Oriana, Iwan Barankay, and Imran Rasul , “Team Incentives: Evidence From A Firm Level Experiment,” Journal of the European Economic Association , 2013, 11(5), 1079-1114. Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman , “Six Randomized Evaluations of Microcredit: Introduction and Further Steps,” American Economic Journal: Applied Economics , January 2015, 7(1), 1-21. Banerjee, Abhijit, Esther Duflo, Nathanael Goldberg, Dean Karlan, Robert Osei, William Pariente, Jeremy Shapiro, Bram Thuysbaert, and Christopher Udry , “A Mul- tifacted Program Causes Lasting Progress for the Poor: Evidence from Six Countries,” Science, 2015, 348(6236). Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan , “The Miracle of Microfinance? Evidence from a Randomized Evaluation,” American Economic Journal: Applied Economics, 2015, 7(1), 22-53. Besley, Timothy and Maitreesh Ghatak , “Competition and Incentives with Motivated Agent,” American Economic Review, 2005, 95 (3), 616-636. o, Ernesto Dal, Frederico Finan, and Martin A. Rossi , “Strengthening State Capabilities: B´ The Role of Financial Incentives in the Call to Public Service,” Quarterly Journal of Economics , 2013, 128 (3), 1169-1218. Bowles, Samuel and Sandra Polanina-Reyes , “Economic Incentives and Social Preferences: Substitutes or Complements?,” Journal of Economic Literature , 2012, 50(2), 368-425. Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller , “Bootstrap-based Improve- ments for Inference with Clustered Errors,” Review of Economics and Statistics , 2008, 90(3), 414-427. Casey, Katherine, Rachel Glennerster, and Edward Miguel , “Reshaping Institutions: Ev- idence on Aid Impacts Using a Preanalysis Plan,” Quarterly Journal of Economics , 2012, 127, 1755-1812. Carpenter, Jeffrey and Erick Gong , “Motivating Agents: How Much Does the Mission Matter?” Journal of Labor Economics , 2016, 34(1), 211-326. Deserranno, Erika, “Financial Incentives as Signals: Experimental Evidence from the Recruitment of Village Promoters in Uganda,” Working Paper, 2016. 27 Dewatripont, Mathias, Ian Jewitt, and Jean Tirole , “Multitask Agency Problems: Focus and Taskk Clustering,” European Economic Review, 2000, 44(4-6), 869-877. Dixit, Avinash, “Incentives and Organizations in the Public Sector: An Interpretative Review,” Journal of Human Resources , 2002, 37(4), 696-727. Duflo, Esther, Rema Hanna and Stephen P. Ryan , “Incentives Work: Getting Teachers to Come to School,” American Economic Review, 2012, 102(4), 1241-1278. Frey, Bruno S. and Reto Jegen , “Motivation Crowding Theory,” Journal of Economic Surveys , 2001, 15(5), 589-611. Friebel, Guido, Matthias Heinz, Miriam Krueger, and Nikolay Zubanov , “Team Incentives and Performance: Evidence from a Retail Chain,” American Economic Review, 2017, 107(8), 2168-2203. Gibbons, Robert, “Incentives in Organizations,” Journal of Economic Perspectives , 1998, 12(4), 115-132. e, Xavier and Ghazala Mansuri , “Money or Management? A Field Experiment on Con- Gin´ straints to Entrepreneurship in Rural Pakistan,” Working Paper, 2017. Gneezy, Uri, Stephan Meier, and Pedro Rey-Biel , “When and Why Incentives (Don’t) Work to Modify Behavior,” Journal of Economic Perspectives , 2011, 25(4), 191-209. When and Why Incentives (Don’t) Work to Modify Behavior Uri Gneezy, Stephan Meier and Pedro Rey-Biel The Journal of Economic Perspectives Vol. 25, No. 4 (Fall 2011), pp. 191-209 Goodman, Sarena F. and Lesley J. Turner , “The Design of Teacher Incentive Pay and Educa- tional Outcomes: Evidence from the New York City Bonus Program,” Journal of Labor Economics , 2013, 31(2), 409-420. Greaney, Brian P., Joseph P. Kaboski, and Eva Van Leemput , “Can Self-Help Groups Realyy Be Self-Help?” Review of Economic Studies , 2016, 83, 1614-1644. Holmstrom, Bengt and Paul Milgrom , “Multitask Principal-Agent Analyses: Incentive Con- tracts, Asset Ownership, and Job Design,” Journal of Law, Economics, & Organization , 1991, 7, 24-52. Imberman, Scott A. and Michael F. Lovenheim , “Incentive Strength and Teacher Produc- tivity: Evidence from a Group-Based Teacher Incentive Pay System,” Review of Economics and Statistics, 2015, 97(2), 364-386. Jayaraman, Rajshri, Debraj Ray, and Francis de V´ ericourt , “Anatomy of a Contract Change,” American Economic Review, 2016, 106(2), 316-358. Khanna, Madhulika, Nishtha Kochhar, and Nethra Palaniswamy , “A Retrospective Impact Evaluation of the Tamil Nadu Empowerment and Poverty Alleviation (Pudhu Vaazhvu) Project,” Journal of Development Studies , 2015, 51 (9), 1210-1223. Kling, Jeffrey, Jeffrey Liebman, and Lawrence Katz , “Experimental Analysis of Neighbor- hood Effects,” Econometrica, 2007, 75 (1), 83-119. Lavy, Victor, “Evaluating the Effect of Teachers’ Group Performance Incentives on Pupil Achieve- ment,” Journal of Political Economy , 2002, 110, 1286-1317. Lavy, Victor, “Performance Pay and Teachers’ Effort, Productivity, and Grading Ethics,” Ameri- can Economic Review, 2009, 99, 1979-2011. 28 Lazear, Edward P., “Performance Pay and Productivity,” American Economic Review, 2000, 90(5), 1346-1361. Mansuri, Ghazala and Vijayendra Rao , Localizing Development: Does Participation Work? , Policy Research Report, The World Bank: Washington, D.C., 2013. McKim, Andrew and Matthew Hughart , “Staff Incentive Schemes in Practice: Findings from a Global Survey of Microfinance Institutions,” Microfinance Network and CGAP, 2005. Mullen, Kathleen J., Richard G. Frank, and Meredith B. Rosenthal , “Can You Get What You Pay For? Pay-for-Performance and the Quality of Healthcare Providers,” RAND Journal of Economics, 2010, 41(1), 64-91. Muralidharan, Karthik and Venkatesh Sundararaman , “Teacher Performance Pay: Experi- mental Evidence from India,” Journal of Political Economy , 2011, 119(1), 39-77. Murphy, Kevin J., “Executive Compensation,” Handbook of Labor Economics , 1999, 3, 2485-2563. Neal, Derek, “The Design of Performance Pay in Education,” in Handbook of Economics of Edu- cation, 2011, ed. Eric Hanushek, Steve Machin, and Ludger Woessmann. Amsterdam: Elsevier. Olken, Benjamin A., Junko Onishi, and Susan Wong , “Should Aid Reward Performance? Evidence from a Field Experiment on Health and Education in Indonesia,” American Economic Journal: Applied Economics , 2014, 6(4), 1-34. Osterloh, Margit and Bruno S. Frey , “Motivation, Knowledge Transfer, and Organizational Forms,” Organization Science, 2000, 11 (5), 538-550. Paarsch, Harry J. and Bruce Shearer , “Piece Rates, Fixed Wages, and Incentive Effects: Statistical Evidence from Payroll Records,” International Economic Review , 2000, 41(1), 59-92. Prendergast, Canice, “The Provision of Incentives in Firms,” Journal of Economic Literature , 1999, 37(1), 7-63. Shearer, Bruce, “Fixed Rates, Fixed Wages and Incentives: Evidence from a Field Experiment,” Review of Economic Studies , 2004, 71, 513-534. Wilson, James, Bureaucracy: What Government Agencies Do and Why They Do It , 1989, Basic Books, New York, NY. 29 Table 1: Summary statistics and balance tests No Credit Social bonus bonus bonus P-values (C) (TC) (TS) C=TC C=TS TC=TS (1) (2) (3) (4) (5) (6) Demographic characteristics Age 27.39 27.53 28.03 0.932 0.628 0.734 Female 0.266 0.250 0.129 0.880 0.118 0.140 Married 0.375 0.361 0.435 0.890 0.572 0.532 Household head 0.125 0.167 0.177 0.588 0.572 0.932 Completed high school 0.562 0.500 0.565 0.566 0.928 0.556 Household consumption (PKR) 6531 5875 6874 0.426 0.572 0.192 Housing quality index 0.167 -0.094 0.265 0.162 0.618 0.140 Employment characteristics Employed with NRSP (months) 26.92 25.97 26.40 0.672 0.882 0.932 NRSP first job 0.547 0.667 0.597 0.138 0.450 0.478 Works from a village branch 0.781 0.889 0.903 0.428 0.362 0.870 Number of COs managed 12.29 16.59 14.62 0.362 0.292 0.686 Share of COs co-managed 0.565 0.494 0.597 0.596 0.846 0.496 Partnered FAa 0.500 0.389 0.565 0.488 0.742 0.266 Preferences and motivation Wants to work for next two years 0.906 0.889 0.903 0.792 0.982 0.790 Overtime work (hrs/day) 1.940 1.748 2.223 0.632 0.396 0.180 Prefers credit bonus 0.594 0.528 0.565 0.618 0.766 0.792 Thinks social-related tasks help credit goal 0.906 0.944 0.952 0.402 0.326 0.884 Did volunteer work before NRSP 0.250 0.194 0.290 0.672 0.690 0.480 Best about NRSP is ability to help others 0.484 0.528 0.565 0.618 0.386 0.766 Monthly performance Number of active loans 75.19 99.87 103.3 0.342 0.248 0.906 New disbursement (PKR) 61894 100675 92354 0.184 0.142 0.746 Repayment on dues at 20th of month 0.749 0.724 0.703 0.682 0.460 0.790 Repayment on dues at end of month 0.984 0.968 0.994 0.742 0.322 0.452 Number of field units (FUs) 11 9 15 - - - Number of field assistants (FAs) 64 36 62 - - - Number of credit organizations (COs) 1411 1217 1776 - - - FA attrition in followup survey 0.156 0.194 0.210 0.704 0.600 0.934 FA selection into verified MPRs b 0.719 0.944 0.823 0.288 0.608 0.244 FA selection into client survey 0.484 0.389 0.194 0.840 0.490 0.414 FA selection into supervisor evaluation 0.422 0.583 0.403 0.520 0.898 0.316 P-value for joint test of significance - - - 0.979 0.998 0.991 Notes: The p-values (in F-tests in Columns 4-6) are calculated using the t-asymptotic wild cluster bootstrap at the field unit level. a Partnered FA is defined as a dummy variable which equals one if an FA is co-managing more than 73 percent of her/his CO portfolio (median value of co-sharing) with other FAs during the 9 months prior to the bonus announcement. b Verified MPRs sample includes FAs whose CO meetings were visited by the CrO at least once during the bonus period. 30 Table 2: Impact of bonus on outcomes Bonus on task 1 Bonus on task 2 ∂y1 ∂y2 ∂y1 ∂y2 ∂b1 ∂b1 ∂b2 ∂b2 Panel A: C12 ≤ 0 γ1 >0 and γ2 >0 + + + + γ1 >0 and γ2 ≤0 + ? + + γ1 ≤0 and γ2 >0 + + ? + γ1 ≤0 and γ2 ≤0 + - - + Panel B: C12 > 0 γ1 >0 and γ2 >0 + ? ? + γ1 >0 and γ2 ≤0 + ? + + γ1 ≤0 and γ2 >0 + + ? + γ1 ≤0 and γ2 ≤0 + - - + Notes: A positive (negative) sign indicates that the impact of the bonus on the outcome is positive (negative). A question mark (?) indicates that the impact cannot be signed. Table 3: Impact of bonus on microcredit outcomes Bonus triggers Number Repayment New New Repayment Credit of active on dues at loans disburse- on dues at index loans 20th of month ment end of month (1) (2) (3) (4) (5) (6) Credit bonus (TC) 24.1* 0.084* 0.673 -284.0 0.008 0.238 (0.068) (0.072) (0.646) (1.000) (0.530) (0.160) Social bonus (TS) 9.406 -0.020 0.616 -444.8 -0.003 0.002 (0.546) (0.672) (0.732) (1.000) (0.810) (0.926) P-value of F-test: TC = TS 0.342 0.020 1.000 1.000 0.332 0.114 Observations 162 162 162 162 162 162 Mean dep. var., control 118.8 0.716 11.00 138347 0.964 -0.163 Notes: All specifications control for region dummies and the pre-treatment value of the dependent variable. Number of active loans is the monthly average number of active loans (new and on-going) managed by the FA. Repayment on dues at 20th of the month is the monthly average share of installment dues paid in full by the 20th. New loans is the monthly average number of new loans issued by the FA. New disbursement is the monthly average amount of new loans issued by the FA in Rupees. Repayment on dues at end of month is the monthly average share of installment dues that were paid in full by end of the month. Credit index is calculated by taking an equally weighted mean across the standard distributions of the five microcredit outcomes in Columns 1-5. Higher value on the credit index implies better performance on microcredit. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p<0.1, **p<0.05, ***p<0.01. 31 Table 4: Impact of bonus on CO quality Bonus triggers New Savers Attend- Dead Multiple Loan CO quality COs per ance COs meetings rejection index member rate (1) (2) (3) (4) (5) (6) (7) Credit bonus (TC) 0.284 -0.127* -0.105* 1.020 -0.194* -0.079 -0.384*** (0.118) (0.054) (0.062) (0.150) (0.074) (0.212) (0.008) Social bonus (TS) 0.225*** -0.030 -0.027 0.056 -0.040 -0.047 -0.051 (0.008) (0.460) (0.636) (0.892) (0.738) (0.372) (0.740) P-value of F-test: TC = TS 0.722 0.110 0.024 0.138 0 .158 0.420 0.014 Observations 131 131 131 131 131 131 131 Mean dep. var., control 0.384 0.699 0.777 -2.230 0.422 0.138 0.118 Notes: All specifications control for region dummies. New COs is the monthly average number of new COs formed by the FA. Savers per meeting is the monthly average share of CO members who saved during CO meetings conducted by the FA. Attendance is the monthly average share of CO members present at the CO meetings conducted by the FA. Dead COs is the monthly average number of COs managed by the FA without any active borrowers for the entire bonus period. Multiple meetings is the monthly average share of COs managed by the FA that had more than one monthly meetings. Loan rejection rate is the monthly average share of social appraisals rejected by the FA. Co quality index is calculated by taking an equally weighted mean across the standard distributions of the six CO-quality outcomes in Columns 1-6. Higher value on the social index implies better performance on social mobilization. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p <0.1, **p<0.05, ***p<0.01. 32 Table 5: Impact of bonus on social empowerment (client-level) Active CO Discussions on Collective CO leader Empowerment participation social issues action advice index (1) (2) (3) (4) (5) Credit bonus (TC) 0.082 0.086 0.011 0.088 0.067 (0.406) (0.494) (0.830) (0.186) (0.260) Social bonus (TS) 0.178 0.146 0.070*** 0.131*** 0.131* (0.178) (0.510) (0.000) (0.000) (0.072) P-value of F-test: TC = TS 0.248 0.604 0.488 0.566 0.128 Observations 1691 1691 1691 1691 1691 Mean dep. var., control 0.230 0.168 0.140 0.191 0.182 Notes: The sample includes CO members from a subset of COs managed by the study-sample FAs, who were interviewed in November 2006 (5 months after the end of the study). All specifications control for region dummies. Active CO participation is a dummy variable which equals one if a CO member reported that members in his/her CO expressed their opinions in meetings and participated in CO decisions more since July 2005 (3 months into the bonus period). Discussions on social issues is a dummy variable which equals one if a CO member reported that his/her CO discussed noncredit-related social issues like public goods and service provisions more frequently since July 2005. Collective action is a dummy variable which equals one if a CO member reported that he/she collectively bought and sold agricultural input and output with others in the village more frequently since July 2005. CO leader advice is a dummy variable if a CO member reported that he/she sought advice from CO leaders more frequently since July 2005. Empowerment index is calculated by taking an equally weighted mean across the four subjective CO quality variables in Columns 1-4. Higher value on the empowerment index implies higher CO quality. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p<0.1, **p<0.05, ***p<0.01. 33 Table 6: Differential impact of bonus by partnership Credit CO quality Motivation Works Shares index index index with a with a partner partner (1) (2) (3) (4) (5) Credit bonus (TC) 0.277* -0.382*** -0.096 -0.009 0.030 (0.084) (0.002) (0.360) (1.000) (0.724) Social bonus (TS) 0.297* 0.122 -0.214* -0.164 0.078 (0.100) (0.330) (0.094) (0.302) (0.114) P x TC -0.081 0.015 0.059 0.042 -0.071 (0.582) (0.954) (0.746) (0.906) (0.660) P x TS -0.566** -0.304* 0.075 0.074 -0.170** (0.012) (0.090) (0.522) (0.784) (0.030) Partnered FA (P) -0.049 -0.113 -0.029 0.514* 0.081 (0.684) (0.494) (0.812) (0.076) (0.156) P-value of F-test: TC + P x TC 0.306 0.168 0.772 0.798 0.556 TS + P x TS 0.116 0.186 0.134 0.356 0.106 TC = TS 0.928 0.002 0.324 0.246 0.504 TC+PxTC=TS+PxTS 0.034 0.424 0.398 0.292 0.570 Observations 162 131 132 162 132 Mean dep. var., control -0.163 0.118 0.426 0.716 0.944 Notes: All specifications control for region dummies and the pre-treatment value of the dependent variable (except Column 5). Partnered FA is a dummy variable which equals one if an FA co-manages more than 73 percent of her/his pre-treatment CO portfolio (median value of co-sharing) with other FAs. Credit index is calculated by taking an equally weighted mean across the standard distributions of the five microcredit outcomes in Table 2, Columns 1-5. Higher value on the credit index implies better performance on microcredit. CO quality index is calculated by taking an equally weighted mean across the standard distributions of the six social outcomes in Table 3, Columns 1-6. Higher value on the social index implies better performance on social mobilization. Motivation index is calculated by taking an equally weighted mean across the three intrinsic motivation variables in Table 6, Columns 1-3. Higher value on the motivation index implies higher intrinsic motivation. Works with a partner is a dummy variable which equals one if an FA co-manages more than 73 percent of his/her post-treatment CO portfolio with other FAs. Shares with a partner is a dummy variable which equals one if an FA decides to split the money with another unknown FA, after receiving a fixed amount from the unknown FA in a trust game. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p<0.1, **p<0.05, ***p<0.01. 34 Table 7: Impact of bonus on motivation and teamwork Best about Identify Finds work Motivation Works Shares NRSP: ability with NRSP important index with a with a to help mission partner partner (1) (2) (3) (4) (5) (6) Credit bonus (TC) -0.141 -0.068 0.005 -0.060 0.012 -0.011 (0.222) (0.498) (0.978) (0.354) (0.852) (0.880) Social bonus (TS) -0.234** -0.152 -0.136 -0.167* -0.125*** -0.012 (0.022) (0.226) (0.156) (0.074) (0.004) (0.724) P-value of F-test: TC = TS 0.390 0.238 0.198 0.084 0.092 0.978 Observations 132 132 132 132 162 132 Mean dep. var., control 0.500 0.426 0.370 0.426 0.716 0.944 Notes: All specifications control for region dummies and the pre-treatment value of the dependent variable (except Columns 4 and 7). Partnered FA is a dummy variable which equals one if an FA co-manages more than 73 percent of her/his pre- treatment CO portfolio (median value of co-sharing) with other FAs. Best about NRSP: ability to help is a dummy variable which equals one if an FA ranks the ability to help people as the best thing about working in NRSP in the followup survey. Identify with NRSP mission is a dummy variable which equals one if an FA reported that he/she identifies with the mission of NRSP. Finds work important is a dummy variable which equals one if an FA reported that he/she finds NRSP work to be important and satisfying. Motivation index is calculated by taking an equally weighted mean across the three intrinsic motivation variables in Columns 1-3. Higher value on the motivation index implies higher intrinsic motivation. Works with a partner is a dummy variable which equals one if an FA co-manages more than 73 percent of his/her post-treatment CO portfolio with other FAs. Shares with a partner is a dummy variable which equals one if an FA decides to split the money with another unknown FA, after receiving a fixed amount from the unknown FA in a trust game. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p <0.1, **p<0.05, ***p<0.01. 35 Table 8: Impact of bonus on performance assessment Promotion Loan CO Assessment disbursement savings score (1) (2) (3) (4) (5) (6) (7) (8) Credit bonus (TC) -0.016 0.070 0.060 0.091 0.105 0.288 0.050 0.150 (0.958) (0.706) (0.752) (0.646) (0.670) (0.208) (0.692) (0.392) Social bonus (TS) 0.127 0.373 -0.040 0.180 0.322 0.503** 0.137 0.352* (0.286) (0.124) (0.900) (0.332) (0.124) (0.016) (0.314) (0.072) P x TC - -0.134 - -0.004 - -0.397 - -0.178 (0.658) (0.934) - (0.316) - (0.538) P x TS - -0.522** - -0.476 - -0.358 - -0.452* (0.044) (0.166) - (0.238) - (0.072) Partnered FA (P) - 0.070 - -0.047 - 0.317 - 0.113 (0.660) (0.836) - (0.330) - (0.582) P-value of F-test: TC + P x TC - 0.804 - 0.806 – 0.724 - 0.890 TS + P x TS - 0.394 - 0.400 – 0.628 - 0.638 TC = TS 0.236 0.166 0.642 0.712 0.220 0.296 0.378 0.248 TC+PxTC=TS+PxTS - 0.540 - 0.234 - 0.418 - 0.628 Observations 73 73 73 73 73 73 73 73 Mean dep. var., control 0.222 0.222 0.444 0.444 0.222 0.222 0.296 0.296 Notes: All specifications control for region dummies. Partnered FA is a dummy variable which equals one if an FA co- manages more than 73 percent of his/her pre-treatment CO portfolio (median value of co-sharing) with other FAs. Promotion is a dummy variable which equals one if a supervisor thinks that an FA is likely to get a promotion in the next 2 years. Loan disbursement is a dummy variable which equals one if a supervisor reports that an FA has improved performance in credit disbursement rate, compared to a year ago. CO savings is a dummy variable which equals one if a supervisor reports that an FA has improved performance in ensuring COs save regularly, compared to a year ago. Assessment score is calculated by taking an equally weighted mean across the three assessment outcomes by supervisor in Columns 1-3. Higher value on the evaluation index implies higher score in the supervisor evaluation. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p <0.1, **p<0.05, ***p<0.01. 36 A Appendix: Study Design and Bonus Incentive Scheme Table A.1: Description of the credit and social bonus incentives Panel A: Credit bonus The first trigger is based on disbursement, measured by the number of active loans managed by the FA in any month. The second trigger is based on whether the repayment on the installment was made in full by the 20 th of the month due. The disbursement trigger can be satisfied at two target levels: High (A) or Low (B). If the FA meets at least target B for disbursement, he qualifies for a bonus based on his recovery rate at the 20 th in that month. If FA qualifies on target A, the size of the bonus is: 20% of base monthly salary if repayment is 100% 16% of base monthly salary if repayment is 99% 12% of base monthly salary if repayment is 98% 8% of base monthly salary if repayment is 97% 4% of base monthly salary if repayment is 96% 0 bonus if repayment is 95% or below If FA qualifies on target B, the size of the bonus is: 15% of base monthly salary if repayment is 100% 12% of base monthly salary if repayment is 99% 9% of base monthly salary if repayment is 98% 6% of base monthly salary if repayment is 97% 3% of base monthly salary if repayment is 96% 0 bonus if repayment is 95% or below The bonus cannot ever be negative. Panel B: Social bonus The first trigger is based on two outcomes: the number of new COs formed and the number of savers at CO meetings. High (A) and Low (B) target levels are set for both outcomes, and an FA needs to reach at least target B for both outcomes to satisfy the first trigger. The second trigger is based on the attendance of CO members at CO meetings. If an FA meets at least target B, he qualifies for a bonus based on member attend- ance at CO meetings. If the FA qualifies on target A, the size of the bonus is: 20% of base salary if average attendance is 85% or more (more than 60% in harvest months) 16% of base salary if average attendance is 80% to 84%(between 56% and 60% in harvest months) 12% of base salary if average attendance is 75% to 79% (between 50% and 55% in harvest months) 8% of base salary if average attendance is 70% to 74% (between 46% and 50% in harvest months) 4% of base salary if average attendance is 65% to 69% (between 40% and 45% in harvest months) 0 bonus if attendance is below 65% (0 bonus if attendance is below 40%) If the FA qualifies on target B, the size of the bonus is determined as follows: 15% of base salary if average attendance is 85% or more (more than 60% in harvest months) 12% of base salary if average attendance is 80% to 84% (between 56% and 60% in harvest months) 9% of base salary if average attendance is 75% to 79% (between 50% and 55% in harvest months) 6% of base salary if average attendance is 70% to 74% (between 46% and 50% in harvest months) 3% of base salary if average attendance is 65% to 69% (between 40% and 45% in harvest months) 0 bonus if attendance is below 65% (0 bonus if attendance is below 40%) The bonus cannot ever be negative. Notes: The bonus incentives were announced to the FAs in the treatment FUs in Mach 2005. The monthly bonuses were paid for 15 months during the study period, and terminated in June 2006. The average base monthly salary for an FA was PKR 3,000 (USD 50.54). 37 Table A.2: List of FUs and bonus assignments Region District Field Unit Panel A: No bonus (control group) Hyderabad Badin Matli Hyderabad Badin Talhar Hyderabad Hyderabad Hala Hyderabad Mir Pur Khas Digri Hyderabad Mir Pur Khas Ghulam Muhammad Hyderabad Thatta Mirpur Sakro Malakand Malakand Dargai Mianwali Bhakkar Bhakkar Mianwali Mianwali Mianwali (Swans) Rawalpindi Attock Hasanabdal Rawalpindi Gujar Khan Gujar Khan Panel B: Credit bonus Hyderabad Badin Badin II (Golarchi) Hyderabad Hyderabad Matiari Hyderabad Mir Pur Khas Hyderabad Malakand Malakand Thana Malakand Mardan Katlang Mianwali Khusab Quaidabad Rawalpindi Attock Attock Rawalpindi Jand Jand Rawalpindi Jand Pindi Gheb Panel C: Social bonus Hyderabad Badin Tando Bago Hyderabad Hyderabad Tando Allah Yar Hyderabad Hyderabad Tando M. Khan Hyderabad Thatta Mirpur Bathoro Hyderabad Thatta Sajawal Malakand Malakand Hero Shah Malakand Malakand Kabal Malakand Mardan Hatian Malakand Mardan Takhat Bhai Mianwali Bhakkar Dulle Wala Mianwali Bhakkar Mankera Mianwali Khusab Jauharabad Rawalpindi Attock Fateh Jang Rawalpindi Gujar Khan Doltala Rawalpindi Pind Dadan Pind Dadan Khan Notes: The study was conducted in 35 Field Units (FUs) of NRSP located in 15 districts and four regions of Pakistan, where NRSP was active in March 2005. The two treatment and control assignments were randomly allocated across these 35 FUs. All FAs working in an FU received the same type of bonus (or control group). 38 1 Figure A.1: Monthly payments of credit and social bonus 1 .8 .8 .6 .6 Credit bonus % % .4 .4 Social bonus .2 .2 0 0 May05 Jul05 Sep05 Nov05 Jan06 Mar06 May06 May05 Jul05 Sep05 Nov05 Jan06 Mar06 May06 (a) Qualified for a bonus (b) Target A 600 1 500 .8 400 .6 PKR 300 % .4 200 .2 100 0 0 May05 Jul05 Sep05 Nov05 Jan06 Mar06 May06 May05 Jul05 Sep05 Nov05 Jan06 Mar06 May06 (c) Target B (d) Bonus amount Notes: During the 15-months bonus period, 26.7 percent of treated FAs qualified for a bonus each month; and 62.9 percent of those qualified on target A. The average bonus amount was PRK 570.30 for target A and PKR 416.70 for target B. Comparing across the two types of bonus, two-fifths of Cr-FAs qualified for a bonus, while only one-fifth of Soc-FAs qualified for a bonus in any given month. Soc-FAs received PKR 124.30 less on monthly bonus payment compared to Cr-FAs, who on average earned PKR 249.30 in bonus each month. Soc-FAs therefore earned 50 percent less in bonus relative to Cr-FAs. 39 Figure A.2: Timeline of the study June 04 June 06 Jan 05-Feb 05 Mar 05 April 05 Nov 06 Monthly Bonus Payments Bonus Announced Client Survey Baseline Followup & Survey Supervisor Surveys FA-CO Panel and MIS data (25 months) MPRs data (15 months) 40 B Appendix: Selection into Supervisor Visits Data on CO-quality outcomes are based on the Monthly Progress Report (MPRs) filed by each FA for all COs visited that month. MPRs data are available for 15 months when the bonus was implemented. These data were verified by a supervisor i.e. a Credit Officer (CrO) through random visits to a subset of scheduled CO meetings. We restrict the analysis to MPRs data verified by supervisors. According to the FA-CO panel, 4,380 unique COs were managed by FAs in our study sample during the bonus period. Out of them, 1,807 COs (41.26%) were visited by a CrO at least once in 15 months, and 6 months out of 15 on average. We take all the COs that show up in the FA-CO panel for each month (during the bonus period), and estimate the rate of CrO visits across the two treatment and the control groups using the following specification: CrOc,t = α + βr + ωt + γ T Cc + λ T Sc + (11) where CrOc,t is an indicator variable that takes the value of one if a CO c was visited by a CrO in month t. βr and ωt are region and month dummies, respectively. The coefficients γ and λ estimate the propensity of CrO visits to CO meetings that are managed by credit and social bonus FAs, respectively (compared to CO meetings managed by control FAs). Appendix Table B.1, Column 1 presents the estimated results, with p-values calculated using the wild cluster bootstrap at the FU level. Among control FAs, 10.5 percent of CO meetings was visited by a CrO. The estimated coefficients γ and λ are both close to zero, 0.039 and -0.000 respectively, and not statistically different from zero and from each other at conventional levels. While we do not find evidence of a differential rate of CrO visits, we also test for any potential selection of COs visited by the supervisor on CO characteristics. For this purpose, we calculate CO’s disbursement and repayment outcomes for each month (during the bonus period) using information from the MIS data, and then estimate the following specification: Yc,t = α + βr + ωt + ζ CrOc,t + γ T Cc + λ T Sc + φ CrOc,t ∗ T Cc + ψ CrOc,t ∗ T Sc + (12) where Yc,t is the characteristics of a CO c in month t. The coefficients φ and ψ on the interaction terms CrOc,t ∗ T Cc and CrOc,t ∗ T Sc represent the difference in CO characteristics for those that were visited by a CrO compared with those that were not, among COs managed by credit and social bonus FAs, respectively (relative to the same difference among COs managed by control FAs). Appendix Table B.1, Columns 2-6 report the results on five different CO characteristics: the 41 Table B.1: Selection on frequency and the quality of CO meetings visited by a CrO CrO Number New New Repayment Repayment CO charac- visit of active loans disburs- on dues at on dues at teristics loans ement 20th of mth end of mth index (1) (2) (3) (4) (5) (6) (7) Credit bonus (TC) 0.039 0.335 0.005 -400.5 0.041 0.013 0.098 (0.420) (0.746) (0.976) (0.784) (0.210) (0.276) (0.472) Social bonus (TS) -0.000 0.087 0.013 -448.5 -0.002 0.008 0.022 (1.000) (0.900) (0.974) (0.764) (0.950) (0.604) (0.914) CrO visit (CrO) - 3.625*** 0.417*** 4997*** 0.006 0.015 0.426*** (0.000) (0.000) (0.000) (0.892) (0.350) (0.002) CrO * TC - -0.288 -0.018 -448.3 -0.011 -0.015 -0.084 (0.778) (0.932) (0.848) (0.770) (0.346) (0.528) CrO * TS - -1.158* -0.179 -1656 -0.015 -0.014 -0.190 (0.092) (0.194) (0.378) (0.742) (0.400) (0.182) p-value of F-test: TC = TS 0.388 - - - - - - CrO*TC = CrO*TS - 0.260 0.138 0.404 0.814 0.952 0.288 Observations 53,127 53,127 53,127 53,127 53,127 53,127 53,127 No. of COs 4,380 4,380 4,380 4,380 4,380 4,380 4,380 R-squared 0.089 0.069 0.013 0.014 0.041 0.048 0.039 Mean dep. var., control 0.105 5.998 0.558 7022 0.832 0.966 -0.0723 Notes: The above regressions control for region and month dummies. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the FU level; *p <0.1, **p<0.05, ***p<0.01. number of active loans, number of new loans, disbursement, and recovery rates at the 20th and at the end of the month. The estimated coefficient φ is close to zero (if anything, negative) and not statistically significantly different from zero at conventional levels for all five outcomes. The estimated coefficient ψ is also negative in sign for these five outcomes, although not statistically significantly different from zero at conventional levels (except for number of active loans, which is statistically significant at the 10 percent level). The two coefficients are also not statistically significantly different from each other at the conventional levels for all five outcomes. Column 7 presents the results on the CO characteristics index, which is calculated by taking an equally weighted average across the standard distributions of the five measures. The results in Column 7 also suggest no differential selection of CrO visits to CO meetings that are managed by credit and social bonus FAs (compared to control FAs). Both coefficients are not statistically signifi- cantly different from zero and from each other at the conventional levels. The negative signs on both coefficients φ and ψ suggest a plausibly negative selection (if anything), which would underestimate our main results on social outcomes in Table 4. 42 C Appendix: Partnership FAs may co-manage part of their CO portfolio jointly with other FAs. Partnership among FAs is encouraged by NRSP to ensure that services are delivered without interruption when an FA falls sick, leaves NRSP, or gets promoted. Partnership also provides NRSP a useful way to train inexperienced FAs while they are on the job, by partnering them with relatively more experienced FAs. In addition, it allows NRSP to keep check on corruption and fraudulent activities by FAs through peer monitoring. Appendix Figure C.1 depicts the distribution of FAs based on the percentage of their CO-portfolio that was co-managed with other FAs during the nine months before the bonus was announced. The median pre-treatment level of co-management is 73 percent. We use the median to identify FAs as either partnered or non-partnered in the main analysis. Our results are however robust to using alternate cut-offs for defining partnership. Almost two-fifths of FAs co-manage their entire CO- portfolio with other FAs, while slightly less than one-fourth manage all their COs independently. Columns 1 and 2 of Appendix Table C.1 present the mean of FA characteristics, preferences, and motivation separately for partnered and non-partnered FAs respectively; column 3 reports the p-values from the F-test of the difference in means between partnered and non-partnered FAs. For all variables including age, gender, and education, partnered FAs do not look any different from non-partnered FAs. Partnered FAs are slightly more experienced than non-partnered FAs (based on number of months of working at NRSP), though the difference is not statistically significant at the conventional level. As indicated by the F-test statistics at the bottom of Column 3, we can not reject equality of means across the full set of variables between partnered and non-partnered FAs. Non-partnered FAs on average co-manage 17 percent of their CO-portfolio with other FAs, while the average co-management rate is 96 percent among partnered FAs. The average number FAs in a partnership team (excluding the FA) is 1.243. Appendix Table C.1, Column 4 reports the correlation between FA’s and the partner’s char- acteristics (mean characteristics in case of two partners or more) among 81 partnered FAs. FA’s gender is highly correlated with his partner’s gender (correlation of 0.834), suggesting that most partnership-teams are formed between FAs of the same gender. More interestingly, partners are also positively sorted on their education. The correlation between FA’s and his partner’s education is 0.309. It is statistically significant at the 10 percent level. Partnered FAs are also positively sorted on experience (based on number of months working at NRSP) and preference for a type of bonus. The correlations are 0.162 and 0.209, but they are not statistically significant at the conventional level. The correlation between FA’s and the partner’s motivation is close to zero (corr=0.097), and it is also not statistically significantly different from zero. 43 Overall, we find no difference on FA characteristics between partnered and non-partnered FAs. In the paper, we take the pre-treatment selection into partnership and partnership formation as given. Figure C.1: Distribution of the share of CO-portfolio co-managed with other FAs Notes: The figure depicts the distribution of the share of FA’s CO-portfolio in the 9 months before the bonus was announced (June 2004 - March 2005) that was co-managed with other FAs. The dotted line in the graph shows the median value of co-management (73 percent of FA’s CO-portfolio). FAs with their share greater than the median value is categorized as “partnered” FA in the analysis. 71 out of 81 partnered FAs co-manage their COs with one other FA, while the rest (10 partnered FAs) co-manage with two other FAs. FA partnership-teams are stable throughout the study period (unless one of the partners quit NRSP). 44 Table C.1: Characteristics of partnered and non-partnered FAs Non-Partnered Partnered p-value Corr w/ (NP) (P) NP = P partner’s characteristicsa (1) (2) (3) (4) Age 28.47 26.86 0.244 -0.122 Female 0.173 0.247 0.240 0.834*** Married 0.407 0.383 0.770 0.224 Household head 0.173 0.136 0.584 0.010 Completed high school 0.531 0.568 0.650 0.309* Household consumption (Rs.) 6431 6602 0.730 0.147 Housing quality index 0.083 0.211 0.462 -0.244 Employed with NRSP (months) 27.17 25.85 0.556 0.162 NRSP first job 0.605 0.580 0.756 -0.022 Wants to work for next two years 0.889 0.914 0.724 -0.028 Prefers credit bonus 0.630 0.506 0.084 0.209 Thinks social helps credit 0.938 0.926 0.790 0.253 Did volunteer work before NRSP 0.296 0.210 0.380 0.028 Best about NRSP: ability to help 0.580 0.469 0.226 0.097 Number of field units (FUs) 32 23 - - Number of field assistants (FAs) 81 81 - - Share of COs co-managed 0.166 0.957 0.000 - Number of partners - 1.123 - - F-test statistics - - 0.696 - p-value - - 0.776 - Notes: The p-values (in F-tests in Column 3) are calculated using the t-asymptotic wild cluster bootstrap at the field unit level. FAs are categorized as partnered or non-partnered based on whether an FA is co-managing more than 73 percent of his/her CO portfolio (median value of co-sharing) with other FAs during the 9 months prior to the bonus announcement. a For partnered FAs with more than one partners, we take the mean value of characteristics across multiple partners. Starred value indicates a statistically significant correlation between FA’s and his partner’s characteristics. 45 D Appendix: Balance Tests on Restricted Samples Table D.1: Summary statistics and balance tests (restricted sample, verified MPRs) No Credit Social bonus bonus bonus p-values (C) (TC) (TS) C=TC C=TS TC=TS (1) (2) (3) (4) (5) (6) Demographic characteristics Age 28.04 27.85 27.82 0.888 0.942 0.982 Female 0.239 0.206 0.098 0.668 0.052 0.152 Married 0.370 0.382 0.431 0.940 0.652 0.746 Household head 0.174 0.176 0.176 0.964 0.930 1.000 Completed high school 0.587 0.500 0.569 0.494 0.812 0.558 Household consumption (Rs.) 7038 5927 6651 0.280 0.654 0.402 Housing quality index 0.095 -0.077 0.349 0.402 0.260 0.100 Employment characteristics Employed with NRSP (months) 26.17 26.15 25.51 0.980 0.704 0.826 NRSP first job 0.521 0.676 0.569 0.096 0.496 0.282 Work from a village branch 0.848 0.882 0.961 0.820 0.394 0.384 Number of COs managed 12.98 17.47 17.08 0.372 0.182 0.892 Share of COs co-managed 0.564 0.510 0.632 0.704 0.670 0.506 Partnered FAa 0.522 0.412 0.608 0.560 0.622 0.226 Preferences and motivation Wants to work for next two years 0.891 0.882 0.902 0.906 0.900 0.810 Prefers credit bonus 0.565 0.500 0.549 0.600 0.882 0.702 Thinks social-related task helps credit goal 0.957 0.941 0.941 0.738 0.802 0.948 Did volunteer work before NRSP 0.326 0.206 0.333 0.308 0.946 0.346 Best about NRSP is ability to help others 0.413 0.529 0.608 0.432 0.180 0.480 Monthly performance Number of active loans 86.51 105.6 122.9 0.528 0.196 0.502 New disbursement (Rs.) 65767 105845 107785 0.142 0.026 0.964 Repayment on dues at 20th of month 0.746 0.718 0.668 0.678 0.316 0.576 Repayment on dues at end of month 0.986 0.966 0.993 0.606 0.512 0.518 Number of field units (FUs) 10 9 14 - - - Number of field assistants (FAs) 46 34 51 - - - Number of credit organizations (COs) 764 792 1150 - - - p-value for joint test of significance - - - 0.971 0.983 0.992 Notes: The sample includes 131 out of 162 FAs whose CO meetings were visited by the supervisor at least once during the 15-month bonus period. The p-values (in F-tests in Columns 4-6) are calculated using the wild cluster bootstrap at the field unit level. a Partnered FA is defined as a dummy variable which equals one if an FA is co-managing more than 73 percent of his/her CO portfolio (median value of co-sharing) with other FAs during the 9 months prior to the bonus announcement. 46 Table D.2: Summary statistics and balance tests (restricted sample, follow up) No Credit Social bonus bonus bonus p-values (C) (TC) (TS) C=TC C=TS TC=TS (1) (2) (3) (4) (5) (6) Demographic characteristics Age 27.09 27.24 28.63 0.922 0.310 0.494 Female 0.315 0.310 0.143 1.000 0.116 0.180 Married 0.352 0.345 0.449 0.992 0.352 0.440 Household head 0.130 0.138 0.224 0.896 0.358 0.380 Completed high school 0.593 0.517 0.531 0.550 0.556 0.888 Household consumption (Rs.) 6486 6250 6769 0.836 0.602 0.544 Housing quality index 0.121 -0.047 0.300 0.430 0.352 0.144 Employment characteristics Employed with NRSP (months) 27.33 25.83 28.12 0.586 0.862 0.618 NRSP first job 0.574 0.655 0.592 0.344 0.826 0.426 Works from a village branch 0.796 0.862 0.898 0.654 0.472 0.708 Number of COs managed 12.70 17.66 15.35 0.330 0.344 0.650 Share of COs co-managed 0.587 0.522 0.572 0.616 0.944 0.756 Partnered FAa 0.519 0.414 0.531 0.548 1.000 0.500 Preferences and motivation Wants to work for next two years 0.889 0.862 0.918 0.678 0.662 0.318 Prefers credit bonus 0.630 0.483 0.551 0.326 0.554 0.680 Thinks social-related task helps credit goal 0.926 0.931 0.959 0.918 0.448 0.588 Did volunteer work before NRSP 0.241 0.207 0.327 0.826 0.446 0.466 Best about NRSP is ability to help others 0.519 0.448 0.531 0.474 0.968 0.536 Monthly performance Number of active loans 75.19 99.87 103.3 0.400 0.310 0.926 New disbursement (Rs.) 61894 100675 92354 0.132 0.190 0.672 Repayment on dues at 20th of month 0.749 0.724 0.703 0.698 0.718 0.580 Repayment on dues at end of month 0.984 0.968 0.994 0.266 0.324 0.468 Number of field units (FUs) 11 7 15 - - - Number of field assistants (FAs) 54 29 49 - - - p-value for joint test of significance - - - 0.977 0.997 0.992 Notes: The sample includes 132 out of 162 FAs who were interviewed in the followup survey in June 2006. The p-values (in F-tests in Columns 4-6) are calculated using the wild cluster bootstrap at the field unit level. a Partnered FA is defined as a dummy variable which equals one if an FA is co-managing more than 73 percent of his/her CO portfolio (median value of co-sharing) with other FAs during the 9 months prior to the bonus announcement. 47 Table D.3: Summary statistics and balance tests (restricted sample, supervisor eval.) No Credit Social bonus bonus bonus p-values (C) (TC) (TS) C=TC C=TS TC=TS (1) (2) (3) (4) (5) (6) Demographic characteristics Age 27.11 28.43 28.12 0.716 0.636 0.938 Female 0.185 0.238 0.0800 0.632 0.364 0.252 Married 0.333 0.381 0.440 0.836 0.626 0.826 Household head 0.111 0.190 0.280 0.524 0.084 0.502 Completed high school 0.741 0.476 0.440 0.198 0.078 0.818 Household consumption (Rs.) 6157 6238 5888 0.894 0.768 0.680 Housing quality index 0.091 0.011 0.546 0.630 0.162 0.104 Employment characteristics Employed with NRSP (months) 27.07 25.81 25.80 0.744 0.744 0.996 NRSP first job 0.519 0.667 0.600 0.236 0.388 0.514 Works from a village branch 0.926 0.810 0.960 0.366 0.732 0.180 Number of COs managed 11.29 20.24 15.75 0.132 0.316 0.370 Share of COs co-managed 0.589 0.552 0.526 0.878 0.714 0.794 Partnered FAa 0.519 0.429 0.480 0.712 0.800 0.776 Preferences and motivation Wants to work for next two years 0.852 0.857 0.920 1.000 0.592 0.544 Prefers credit bonus 0.556 0.524 0.480 0.846 0.592 0.764 Thinks social-related task helps credit goal 0.963 0.952 0.920 0.848 0.466 0.616 Did volunteer work before NRSP 0.333 0.286 0.200 0.720 0.264 0.572 Best about NRSP is ability to help others 0.444 0.476 0.520 0.838 0.654 0.722 Monthly performance Number of active loans 83.74 118.7 113.7 0.290 0.432 0.722 New disbursement (Rs.) 61202 117086 90765 0.022 0.232 0.334 Repayment on dues at 20th of month 0.709 0.705 0.716 0.708 0.758 0.922 Repayment on dues at end of month 0.993 0.948 0.993 0.782 0.982 0.644 Number of field units (FUs) 6 6 13 - - - Number of field assistants (FAs) 27 21 25 - - - p-value for joint test of significance - - - 0.958 0.946 1.000 Notes: The sample includes 73 out of 162 FAs for whom we have their supervisors’ evaluation of their performance from the supervisor survey conducted in June 2006. The p-values (in F-tests in Columns 4-6) are calculated using the wild cluster bootstrap at the field unit level. a Partnered FA is defined as a dummy variable which equals one if an FA is co-managing more than 73 percent of his/her CO portfolio (median value of co-sharing) with other FAs during the 9 months prior to the bonus announcement. 48 Table D.4: Summary statistics and balance tests (client sample) No Credit Social bonus bonus bonus p-values (C) (TC) (TS) C=TC C=TS TC=TS (1) (2) (3) (4) (5) (6) Age 34.33 41.14 36.56 0.268 0.108 0.324 Female 0.120 0.184 0.226 0.140 0.398 0.744 Household head 0.474 0.575 0.506 0.208 0.304 0.276 Years of education 6.150 5.544 6.483 0.516 0.518 0.572 Household size 9.294 7.774 8.309 0.168 0.180 0.548 Number of children 3.084 2.405 2.758 0.456 0.526 0.614 Number of CO members (clients) 758 548 385 - - - Number of field units (FUs) 3 4 4 - - - Number of field assistants (FAs) 31 14 12 - - - Number of credit organizations (COs) 83 71 59 - - - p-value for joint test of significance - - - 0.369 0.553 0.796 Notes: The sample includes 1691 clients of FAs (i.e. CO members) who were interviewed in Nov 2006 (5 months after the end of the study). The p-values (in F-tests in Columns 4-6) are calculated using the wild cluster bootstrap at the field unit level. E Appendix: Additional Tables and Figures 49 Table E.1: Differential impact of bonus on microcredit outcomes, by partnership Bonus triggers Number Repayment New New Repayment Credit of active on dues at loans disburse- on dues at index loans 20th of month ment end of month (1) (2) (3) (4) (5) (6) Credit bonus (TC) 14.88 0.117** -0.429 -9801 0.023 0.277* (0.456) (0.014) (0.812) (0.776) (0.194) (0.084) Social bonus (TS) 27.26 0.037 3.389 29389 0.020 0.297* (0.212) (0.276) (0.154) (0.390) (0.310) (0.100) P x TC 22.76 -0.069 2.729 22211 -0.027 -0.081 (0.392) (0.154) (0.196) (0.384) (0.320) (0.582) P x TS -33.18 -0.110*** -5.150* -55657 -0.045* -0.566** (0.252) (0.010) (0.096) (0.178) (0.104) (0.012) Partnered FA (P) -18.92 0.031* -3.214 -40066 0.034 -0.049 (0 .234) (0.048) (0.124) (0.158) (0.156) (0.684) p-value of F-test: TC + P x TC 0.028 0.400 0.294 0.656 0.862 0.306 TS + P x TS 0.882 0.192 0.598 0.570 0.150 0.116 TC = TS 0.660 0.052 0.114 0.168 0.806 0.928 TC+PxTC=TS+PxTS 0.000 0.104 0.020 0.058 0.378 0.034 Observations 162 162 162 162 162 162 Mean dep. var., control 118.8 0.716 11.00 138347 0.964 -0.163 Notes: All specifications control for region dummies and the pre-treatment value of the dependent variable. Partnered FA is a dummy variable which equals one if an FA co-manages more than 73 percent of his/her pre-treatment CO portfolio (median value of co-sharing) with other FAs. New COs is the monthly average number of active loans (new and on-going) managed by the FA. Repayment on dues at 20th of the month is the monthly average share of installment dues paid in full by the 20th. New loans is the monthly average number of new loans issued by the FA. New disbursement is the monthly average amount of new loans issued by the FA in Rupees. Repayment on dues at end of month is the monthly average share of installment dues that were paid in full by end of the month. Credit index is calculated by taking an equally weighted mean across the standard distributions of the five microcredit outcomes in Columns 1-5. Higher value on the microcredit index implies better performance on microcredit. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p <0.1, **p<0.05, ***p<0.01. 50 Table E.2: Differential impact of bonus on social mobilization outcomes, by partnership Bonus triggers New Savers Attend- Dead Multiple Loan Social COs per ance COs meetings rejection index member rate (1) (2) (3) (4) (5) (6) (7) Credit bonus (TC) 0.181 -0.121** -0.084** 1.494 -0.168 -0.054 -0.382*** (0.336) (0.030) (0.020) (0.186) (0.272) (0.424) (0.002) Social bonus (TS) 0.430*** -0.059 0.012 -0.326 0.013 -0.024 0.122 (0.010) (0.314) (0.732) (0.588) (0.878) (0.794) (0.330) P x TC 0.264 -0.028 -0.047 -1.30 -0.050 -0.057 0.015 (0.210) (0.796) (0.552) (0.238) (0.770) (0.384) (0.954) P x TS -0.356** 0.051 -0.068 0.662 -0.096 -0.041 -0.304* (0.048) (0.648) (0.388) (0.356) (0.492) (0.498) (0.090) Partnered FA (P) -0.173 -0.080 -0.039 -0.824 0.023 0.004 -0.113 (0.206) (0.392) (0.578) (0.246) (0.724) (0.964) (0.494) p-value of F-test: TC + P x TC 0.080 0.180 0.116 0.626 0.082 0.126 0.168 TS + P x TS 0.296 0.900 0.480 0.204 0.478 0.208 0.186 TC = TS 0.184 0.316 0.004 0.040 0.280 0.612 0.002 TC+PxTC=TS+PxTS 0.220 0.182 0.108 0.558 0.308 0.162 0.424 Observations 131 131 131 131 131 131 131 Mean dep. var., control 0.384 0.699 0.777 -2.230 0.422 0.138 0.118 Notes: All specifications control for region dummies. Partnered FA is a dummy variable which equals one if an FA co- manages more than 73 percent of his/her pre-treatment CO portfolio (median value of co-sharing) with other FAs. New COs is the monthly average number of new COs formed by the FA. Savers per meeting is the monthly average share of CO members who saved during CO meetings conducted by the FA. Attendance is the monthly average share of CO members present at the CO meetings conducted by the FA. Dead COs is the monthly average number of COs managed by the FA without any active borrowers for the entire bonus period. Multiple meetings is the monthly average share of COs managed by the FA that had more than one monthly meetings. Loan rejection rate is the monthly average share of social appraisals rejected by the FA. Social index is calculated by taking an equally weighted mean across the standard distributions of the six social mobilization outcomes in Columns 1-6. Higher value on the CO-quality index implies better performance on social mobilization. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p <0.1, **p<0.05, ***p<0.01. 51 Table E.3: Impact of bonus on supervisory effort |Difference| between actual and supervisor-reported FA performance on: Recovery Number of Attendance rate active loans (1) (2) (3) Credit bonus (TC) -0.029 12.15 0.021 (0.396) (0.588) (0.556) Social bonus (TS) -0.018 10.82 0.037 (0.646) (0.788) (0.630) p-value of F-test: TC = TS 0.658 0.928 0.892 Observations 96 96 55 Mean dep. var., control 0.050 92.64 0.200 Notes: The sample includes 96 FAs and (55 FAs from verified MPRs sample) whose supervisors were interviewed in June 2006. During the interview, the supervisors were asked about each of their FA’s performance in the previous month (i.e. May 2006) on two credit outcomes (number of active loans and repayment rates) and one social outcome (attendance of CO members in CO meetings). The dependent variables are constructed by taking the absolute difference between the supervisor’s reported performance and the actual performance on an FA in May 2006. All specifications control for region dummies. The p-values are reported in parentheses and are calculated using the t-asymptotic wild cluster bootstrap at the field unit level; *p<0.1, **p<0.05, ***p<0.01. 52 Figure E.1: Impact of credit and social bonus by month (a) Credit index: All FAs (b) Social index: All FAs (c) Credit index: Non-partnered FAs (d) Social index: Non-partnered FAs (e) Credit index: Partnered FAs (f) Social index: Partnered FAs Notes: The graphs above plot the estimated impact of credit bonus (solid line) and social bonus (dash line) on credit and social indices by month, for the 15-months bonus period. ATEs of credit and social bonus are estimated by using FA-month level data and by running an OLS regression with the following specification: Yit = α + βr + θt + Σ25 25 j =11 γj T Cij + Σj =11 σj T Sij + . γt s and σt s for 11 ≤ t ≤ 25 are plotted in (a) and (b) for the credit index and social indices (dependent variables) respectively. The effects on the subgroups by partnership are estimated by running an OLS regression with the following specification: Yit = α + βr + θt + Pi + Σ25 25 25 25 j =11 γj T Cij + Σj =11 σj T Sij + Σj =11 δj Pi T Cij + Σj =11 ωj Pi T Sij + . γt s and σt s, which represent the effects on the non-partnered FAs, are presented in (c) and (d); the effects on the partnered FAs given by γt + δt and σt + ωt are plotted in (e) and (f), for the dependent variables credit and social indices respectively. 53