Preliminary Draft: Please do not cite

Preliminary Draft: Please do not cite

1

NEUDC 2004 Northeast Universities Development Consortium Conference HEC Montréal October 1-3, 2004 NEUDC Home > Call for papers -------------------------------------------------------------------------------- A. Presenter Identification Last Name: McKim First Name: Andrew Home university: University of California, Berkeley Status at university: Ph.D. Candidate Email: [email protected] Fax: Personal Web Site: B. Paper description Extended abstract file to upload (2 pages max.): Attached Name of presenter: Andrew McKim Field: EmpiricalMicro 3 keywords:

1. Microfinance 2. Incentives 3. Personnel

JEL classification:

G1 - General Financial Markets M5 - Personnel Economics


2

Andrew McKim University of California, Berkeley

May 4, 2004

Staff Incentive Policies for Credit Officers: An analysis of their effects within heterogeneous

firms and contexts and across multiple tasks Extended Abstract: This paper explores how changes in monetary staff incentive contracts affect the productivity of credit officers within microfinance institutions (MFIs). Credit officers are responsible for the promotion, screening, monitoring, and enforcement of loans and account for about half of an institution’s input costs in the form of salaries. Over the past three to five years, a large percentage of microfinance institutions have switched from a fixed salary compensation scheme to a pay-for-performance scheme that links compensation to various performance indicators. Despite the importance of staff incentive policies for MFIs, we know relatively little about the actual design and effectiveness of such policies The purpose of this paper is to empirically examine how an increase in the power of staff incentives affects credit officer productivity. In particular, we are interested in estimating the overall magnitude of the effect and whether there is a differential effect across different performance indicators (number of clients, value of outstanding portfolio, delinquency rate). We are also interested in whether there are differential effects of the incentives across credit officers, contexts, and/or firms. We use a unique data set from five microfinance institutions that includes monthly productivity data at the credit officer level for several months before and after a change in the incentive schemes. In each of the institutions, the monetary incentives were increased from a relatively weak-powered scheme to a moderately strong-powered scheme. The estimated effects were quite significant on various measures of productivity but were notably heterogeneous across tasks, individuals, branches, and institutions. The results are consistent with an expanded principal-agent framework that also accounts for non-contractible differences in local market conditions.


1

Andrew McKim University of California, Berkeley

May 4, 2004

Staff Incentive Policies for Credit Officers: An analysis of their effects within heterogeneous firms and contexts and

across multiple tasks1 Abstract: This paper explores how changes in monetary staff incentive contracts affect the productivity of credit officers within microfinance institutions. The study uses a unique data set from five microfinance institutions that includes productivity data at the credit officer level for several months before and after a change in the incentive schemes. In each of the institutions, the monetary incentives were increased from a relatively weak-powered scheme to a moderately strong-powered scheme. The effects were quite significant on various measures of productivity but were notably heterogeneous across tasks, individuals, branches, and institutions. The results are consistent with an expanded principal-agent framework that also accounts for non-contractible differences in local market conditions. 1. Introduction 1.A. Staff incentives within microfinance institutions Significant resources have been employed over the past several decades by governments, international donor organizations, as well as private investors towards the strengthening and expansion of micro-finance institutions (MFIs). Microfinance institutions are institutions (both for-profit and non-profit) that specialize in loans and savings products for individuals who do not have access to the formal banking sector. Although there are no reliable estimates on how many MFIs exist, more than 1,500 MFIs serving over 30 million families responded to a recent survey by the Microcredit Summit Campaign.2 The key difference between a microfinance institution and a bank is that in a bank, loans are processed primarily in the branch office and are based upon formal financial records and legal guarantees. In microfinance institutions, the authorization process is primarily based upon: a field-based socio-economic analysis aimed at determining capacity and willingness to repay the loan, and/or guarantees from other individuals or groups of individuals in the community. Within microfinance institutions, credit officers are responsible for the promotion, screening, monitoring, and enforcement of loans. They account for about half of an institution’s input costs (in the form of salaries) and critically affect the overall performance of the institution. Direct 1 This research project received financial support from the Ford Foundation and the Aspen Institute’s Nonprofit Sector Research Fund. 2 Microcredit Summit Campaign Report, 2001.


2

monitoring is difficult as credit officers spend about 75% of their time in the field. Over the past three to five years, a large percentage of microfinance institutions have switched from a fixed salary compensation scheme to a pay-for-performance scheme that links compensation to various performance indicators. Despite the importance of staff incentive policies for MFIs, we know relatively little about the actual design and effectiveness of such policies: knowledge to date is restricted to qualitative case studies from a handful of high profile MFIs.3 In a Micro-Banking Bulletin publication devoted to the topic of monetary incentive policies, Martin Holtmann discusses the importance of staff incentives, general types of incentive schemes, and some preliminary findings of their effects on productivity.4 In the same issue, Eduardo Bazobery's article relates PRODEM's negative experiences with an incentive scheme design that is remarkably similar to a "common" scheme described by Holtmann.5 The two featured authors agree that incentive systems can be a valuable tool but warn that they need to be tailored to each institution as the context and objectives of MFIs vary significantly. Failure to do so can lead to discouraged employees and/or harmful distortions in their behavior. The author has conducted in-depth interviews with managers, branch managers and credit officers from fourteen microfinance institutions in four countries in Latin America. Most managers have reported significant increases in productivity while some institutions have reported unanticipated negative effects. The primary concerns expressed by managers are that a switch to monetary incentives can produce a decline in the quality of service and a resulting increase in the delinquency rate. Through the in-depth interviews with more than 200 credit officers, additional concerns were identified. One of the principal concerns and complaints by credit officers of the incentive schemes is that the schemes do not account (i.e. adjust target parameters) for differences in market opportunities across sometimes very heterogeneous branches and neighborhoods. In a presentation on staff incentives at a recent international microfinance conference, Nanci Lee identified three specific topics of interest to practitioners that merit further study: What are the effects of incentives on performance? How are incentive schemes tailored to context? And how can the effectiveness of incentive schemes be monitored?6 These questions are not unique to MFIs; economists have been wrestling with these issues for years. 1.B. Staff Incentives and agency theory While this paper focuses on the use of staff incentives in MFIs, it strives to make a contribution towards empirically testing theories in the broader moral hazard and personnel economics literature. A number of agency (or “moral hazard”) models have been developed to analyze the choice of contracts and their effects when monitoring of employee effort is imperfect and/or costly. There have been several important extensions to the standard principal agent model, particularly in the

3 See Robinson (1997) and El Shami (1997). 4 Holtmann, 2001. 5 Bazoberry (2001). 6 Lee (2001).


3

past five to ten years. The standard principal agent model that focused on the efficiency vs. insurance trade-off (Grossman and Hart, 1983 and Stiglitz, 1974) has been expanded to include: multiple agents, multiple principals, multiple periods, multiple tasks, intrinsic motivation, and several other special cases. Each extension brings us closer to modeling the complexities of reality. Unfortunately, the empirical studies to test such improved theories have been remarkably limited.7 Empirical work within the incentives literature can be divided into studies that have focused on either:

• Directly measuring the effects of incentive contracts on productivity; or • Analyzing whether the design and/or utilization of observed contracts coincides with

theoretical predictions. This study, which falls into the first category, will analyze how credit officers in five microfinance institutions responded to a significant change in their monetary incentives contract. 2. Research Questions and Motivations The overall research question is:

What is the magnitude and nature of the effects of staff incentive policies on performance indicators of credit officers in microfinance institutions?

More specifically, we are interested in investigating the following three research questions (each of which is discussed in more detail below):

• What is the average effect on various indicators of performance due to an increase in the power of an incentive scheme?

• Do the incentives affect all performance indicators (number of clients, value of outstanding portfolio, delinquency rate) equally, or is there a differential effect (possibly due to substitution between tasks) in terms of magnitude and/or timing of effects across indicators?

• Are there differential effects of incentives across credit officers (tenure in institution, gender, initial productivity indicators), contexts, and/or firms?

2.A. What is the average effect on various indicators of performance due to an increase in the power of an incentive scheme? Much of the interest in agency theory is focused on the question of whether incentives matter: do agents respond to incentives? High productivity is a goal shared by all microfinance institutions, regardless of whether they are for-profit firms or nonprofit socially motivated non-governmental organizations. The principal reason for adopting incentive schemes is to improve productivity and efficiency. We will explore the average overall treatment effect of an increase in incentives across institutions.

7 See Gibbons (1998), Prendergast (1997), and Chiappori and Salanie (2000) for fairly comprehensive reviews of recent contract literature.


4

Empirical studies of the effects of incentives on productivity have been very limited due to scarcity of data. Only a few studies in very specific workplaces have been conducted over the past decade. Among these, studies of employees of a windshield installation company (Lazear, 1996) and horse jockeys (Fernie and Metcalf, 1996) found very significant increases in productivity following the adoption of incentive or “piece rate” contracts. Using a more structural approach, Paarsch and Shearer (1996) find increases in productivity of between 6% and 30% among tree planters working under an incentive versus a fixed wage scheme. Shearer (1999) documents a 20% productivity differential among tree planters who were randomly assigned to piece rate versus fixed wages. Qualitative results from interviews with credit officers suggest that incentives can have powerful effects over productivity. When asked if they had changed their behavior following the implementation or a recent major change in the incentive scheme only about forty percent of the credit officers said they had. Among the 60% that did not report an increase in their effort, most did not meet the minimum cut-off levels to qualify for the incentive compensation. Many felt that the cut-off levels and benchmarks for performance indicators were unfair given the difficulties of their operating area. For many, the increased salary potential can be a significant motivating force that produces very immediate responses: “Yes. Of course I did- I realized how much more I could earn.” When pressed to explain exactly how they changed their behavior, some of the more common responses were:

• Work later hours. Credit officers can begin to work more on paperwork in the evenings in order to save time during the day. For institutions that lock the doors at a certain time, credit officers take paperwork home (even though this is often against institutional policy).

• Spend more time in the field. Many feel the best way to make a good credit evaluation, find new customers, and to control delinquency, is to spend as much time in the field as possible.

• Organize work more efficiently. This can be achieved through more efficient means of processing paper work or through concentrating efforts in a specific geographic area. Much of the time of a credit officer is spent traveling from one client to another. By concentrating in certain areas, the credit officer can save travel time while increasing monitoring (can visit each client more often through drop-by visits) and promotion (most credit officers prefer to attract new clients through referrals of current clients).

• Concentrate on wealthier clients and larger loans. Some credit officers said they could increase their outstanding loan portfolio much quicker through say ten new $2,000 loans than through forty $500 loans. The evaluation process (fixed costs of processing a new client) is often similar in terms of time and effort. However, there are significant checks on this strategy. First, the effort to find and attract wealthier clients might actually be more difficult if there is greater competition from other institutions for this same market segment. Second, credit officers feel that their portfolio is more susceptible to risk when it is concentrated in a smaller number of larger loans (this would be interesting to check quantitatively). Often, they feel it is


5

easier to control delinquency among the forty $500 loans than among the ten $2,000 loans.

• Allocate work among tasks more efficiently. The principal trade-off in a credit officer’s allocation of effort is between controlling delinquency and increasing loan portfolio. If they are not receiving incentives because of insufficiently large loan portfolio (while the delinquency rate is in control), then they will allocate more time towards generating new clients. If their delinquency rate is too high they would choose to actively pursue the easiest delinquent cases first and the more difficult ones later. If they feel they will not have much success in pursuing the delinquent cases, they could work on increasing their loan portfolio as a means of reducing the delinquency rate (i.e. increasing the denominator to reduce the overall percentage of delinquent portfolio).

2.B. Do the incentives affect all performance indicators (number of clients, value of outstanding portfolio, delinquency rate) equally, or is there a differential effect (possibly due to substitution between tasks) in terms of magnitude and/or timing of effects? Although there have been very few empirical studies of multi-tasking theory, the results have been very interesting. Using probit models for the type of contract between oil firms and service station, Slade (1996) finds evidence that the level of incentives varies depending upon whether or not the service station performs complementary tasks. Preyra and Pink (2001) find that CEOs of non-profit hospitals (who have multi-dimensional objectives) receive less incentives than those directing for-profit hospitals. Cockburn, et. al. (2000) investigate incentive schemes for researchers in pharmaceutical laboratories. They find a positive correlation between the use of incentives for different tasks (basic vs. applied research): an increase in incentives for one task is matched by increasing incentives for the other task, in order to avoid unintended distortions in effort. In our case, credit officers have two primary tasks: to improve the quantity and quality of their outstanding loan portfolio. These tasks are fully contractible and are directly related to the institutional objectives of generating revenue. A third task pursued by nonprofit firms is to expand outreach to poorer clients. Due to high monitoring costs (and potential for fraud in reporting less verifiable indicators), this is weakly proxied by some institutions by simply using the total number of active clients an agent serves. The reasoning is that although the cost of effort per dollar lent is more for poorer clients, the cost of effort per client is relatively lower for poorer clients. If agents serve poorer clients, they are able to serve more of them for the same amount of total effort. Potential differences in the cost functions (i.e. repayment rate might have a higher second derivative of cost function) and the timing of the tasks (efforts to affect repayment rate might not be rewarded as quickly as efforts to increase portfolio or number of clients) could result in undesirable substitution of effort between the tasks. 2.C. Is there a differentiated effect of an increase in incentives across branches, credit officers (tenure in institution, gender), and contexts (rural vs. urban)? A key objective of this research is to investigate in greater detail how differences in the geographic areas affect a credit officer’s productivity and compensation under incentive schemes. I have visited 14 microfinance institutions and interviewed more than 250 credit officers over the past several months. The primary concern expressed by the credit officers is


6

that they feel the incentive schemes are unfair. Generally, there exists a uniform incentive scheme (with established parameters and payoffs) for all credit officers within an institution despite the fact that the institution might operate in very heterogeneous contexts. Credit officers are assigned certain geographic areas in which they are to operate. Credit officers that operate in especially poor areas or areas that are very distant from the branch office (clients are less willing to make deposits if they live very far from the branch office) feel like they are being treated unfairly by the incentive scheme (they have to work much harder to attain the same performance parameters). Rather than being motivated by the incentive scheme, a considerable portion of credit officers become less motivated. This is a potential problem for virtually all incentive- or commission-based schemes, but has yet to be addressed by the theoretical or empirical literature. The shift of a waiter, shift of a taxi driver, geographical area of a salesman or manager of a chain restaurant, and the call-list (some are “hotter” than others) of a telemarketer, all significantly influence the opportunity to earn incentives/commissions/tips. Many industries handle this fairness problem by giving the choice of best areas/shifts to those with more tenure (as a form of promotion incentive). However, this is very difficult to do for credit officers and employees in several industries: there are often very high costs associated with transferring employees who have highly specialized knowledge of local markets. 3. Contributions of this research This study is important primarily because the richness of the data set collected permits a rare opportunity to empirically test some of the predictions of the multi-tasking literature and the potential effects of heterogeneous contexts on incentive schemes. Previous work has hinted at a possible effect of heterogeneous contexts8, but there has yet to be a thorough analysis to gauge the potential importance of this phenomenon in affecting the success of incentives. Existing empirical studies on the multi-tasking literature have also been remarkably limited in terms of scope. It is also important to highlight that this study will utilize actual data on contractible productivity measures as well as the incentives paid; virtually every other empirical study of the effects of incentives has proxied the effects using payroll data (Asche, 1990 and Lazear, 2000 are two exceptions). Besides the contributions to the incentives literature, the results will be particularly relevant to MFI managers and their estimated 200,000 credit officers, worldwide. Key microfinance networks (MicroFinance Network, SEEP Network) have identified staff incentive policies as a priority area for further research.

8 For example, in Asche (1990), there is a reference to differences in ability to meet quotas for navy recruiting across geographic areas of the U.S. However, the data set utilized in the study was restricted to the Chicago area.


7

4. Data 4.A. Description of the data set The panel data I have collected includes monthly productivity indicators at the credit officer level for five institutions. The data corresponds to several months before and after a significant increase in the incentive scheme. Data from an additional two institutions will not be used because of other confounding institutional changes during the period of study. None of the five remaining institutions have experienced any significant internal or external changes during the period of study. Summary information on the exact dates of the corresponding data and sizes of the institutions is presented in Table 1. Basically, I have placed all of the institutional data on a single timeframe beginning with June 2000 and ending in July 2003. The months for which I have data are labeled according to this timeframe. The placement on a single timeframe will enable us to better correct for seasonal or macro-level shocks that affect all institutions. Table 1: Composition of Data Set Governance

Structure # of Branches # of credit

officers9 Months of Data

Availability: 1 = 6/00

31= 12/02

Month of Major Change

in Scheme

MFI 1 Non-regulated 11 127 1 – 28 20 MFI 2 Regulated 12 84 15 – 31 23 MFI 3 Non-regulated 2 14 8 – 31 19 MFI 4 Regulated 6 54 8 – 31 23 MFI 5 Regulated 2 37 15 – 31 25 Totals for 5 MFIs

33 316

Data at the credit officer level was unavailable prior to the implementation of the institutions' first monetary incentive scheme. However, the increase in the power of the scheme is very significant in all cases (in some cases, the original scheme was so weak it could almost be considered non-existent). In three of the MFIs, the managers replaced a quarterly scheme with a monthly scheme (i.e. tripling the power of the incentives). The other two institutions had a monthly scheme in place and significantly increased the value of the bonuses across all indicators. It is important to point out that the inflation during the time of the increase in incentives was close to 0%- therefore the increase was not meant to adjust for inflation (note that all values in the dataset have been corrected for inflation on a monthly basis). Two of the five MFIs are non-regulated NGOs and the other three MFIs are regulated privately owned enterprises. The regulated MFIs are forced to abide by stricter regulations and tend to be more commercially oriented. Only MFI1 utilizes the community banking lending technology while the other four MFIs lend to individuals.

9 The number of credit officers refers to the total number of credit officers over the entire study period for each MFI. The number of credit officers for each MFI at a given time is considerably less.


8

4.B. Summary Statistics Summary statistics of credit officer performance across institutions, before and after treatment, are presented in Table 2 (on the following page). These statistics (and all analysis in this paper) reflect averages across credit officers and are not weighted by portfolio size. Since our focus is on the effects of incentives on credit officers, this data format is suitable for our purposes. However, it is important to point out that these statistics are not directly comparable to a weighted average of the institution.10 We see that there are very large differences across MFIs, particularly for outstanding portfolio, number of clients, and average loan size. Performance indicators seem to substantially improve after the increase in incentives (treatment). However, it is important to point out that much of the increase in performance indicators could also be attributed to an increase in average tenure. Staff turnover was quite high both before and after the increase in incentives, with about an average drop-out or attrition rate of 1.05% per month. The increase in compensation in the form of incentives was relatively important for all institutions except for MFI4, where the average increase was less than 50 Soles. To give an idea of the relative “power” of the incentives, the monthly base salary ranged between S/900 and S/1500 across institutions.

10 For example, if an MFI had 2 observations for loan delinquency rate: .05 and .15, the average delinquency rate across credit officers would .10. However, in actuality, if the first credit officer had a larger portfolio, then the average institutional delinquency rate would be less than .10.


9

Table 2: Summary Statistics Across Institutions, Before and After Treatment Effect (T) MFI 1 MFI 2 MFI 3 MFI 4 MFI 5 Total

Before T After T Before T After T Before T After T Before T After T Before T After T Outstanding portfolio (in Real Soles)

88,672 (62,062)

136,250 (66182)

783,256 (433,750)

894,436 (427,442)

262,230 (43,847)

369,326 (104,568)

403,203 (183,760)

495,742 (212,866)

312,278 (127,369)

350,155 (104,459)

377,706 (374048.6)

Number of clients

145 (83.2))

211 (86.30)

279 (150.5)

311.1 (151)

171 (20.8)

201 (33.4)

152 (65.6)

176 (65.7)

129 (51.8)

141 (39.9)

197 (115.3)

Delinquency rate**

5.2% (.099)

6.3% (.116)

13.8% (.091)

12.3% (.092)

9.4% (.052)

6.9% (.060)

7.2% (.041)

7.0% (.042)

9.5% (.05093)

8.8% (.04681)

**

Average Loan Size (Real Soles)

570 (118.5)

628 (123.97)

2,762 (602.5)

2,876 (625.4)

1,535 (188.9)

1,821 (404.5)

2,622 (393.97)

2,764 (410.44)

2,428 (195.12)

2,480 (206.46)

1,753 (1091.4)

Average tenure (in months)

8.4 (5.85)

15.1 (7.19)

16.7 (7.52)

22.7 (10.14)

14.9 (5.87)

24.3 (9.70)

13.9 (7.78)

21.3 (10.89)

15.5 (8.77)

20.9 (10.39)

15.4 (9.53)

Average rate of staff attrition (per month)

1.29% (.113)

1.57% (.125)

0.35% (.059)

1.46% (.1198)

1.10% (.1045)

0.55% (.0741)

0.76% (.0870)

1.03% (.1010)

1.00% (.0993)

0.39% (.0621)

1.05% (.1018)

Avg monthly change in Outstanding Portfolio (Real Soles)

7,520 (11,261)

7,152 (15,574)

10,449 (50,020)

25,006 (44,833)

4,032 (12,418))

12,428 (15,596)

10,743 (31,945)

16,061 (27,639)

4,843 (12,757)

7,394 (9,948)

11,042 (28822.7)

Avg monthly change in Number of Clients

11.2 (14.14)

10.3 (17.5)

3.8 (17.19)

7.3 (16.41)

2.0 (3.60)

3.6 (4.86)

3.4 (7.65)

4.5 (8.32)

1.5 (4.94)

2.6 (3.90)

7.1 (14.07)

Avg. monthly change in Delinquency Rate

.0082 (.041)

.00003 (.028)

.0015 (.0161)

-.0019 (.0097)

-.0005 (.0086)

-.0017 (.0043)

.0016 (.0060)

-.0008 (.0041)

.00029 (.0024)

-.00082 (.0021)

.0019 (.0247)

Avg. percentage of credit officers receiving incentives

16.8% (.3738)

67.1% (.4701)

49.6% (.5005)

59.0% (.4922)

2.2% (.1482)

43.8% (.4979)

26.7% (.4430)

31.5% (.4651)

45.8% (.5012)

55.2% (.4985)

40.4% (.4907)

Average monthly incentive compensation (Real Soles)

35.4 (137.9)

172.2 (154.4)

177.3 (196.4)

473.3 (435.9)

2.3 (28.78)

243.7 (263.1)

34.81 (71.4)

82.64 (137.4)

77.5 (279.4)

322.21 (306.8)

197.9 (275.1)

Observations: Number of observations 1,216 827 532 612 110 137 519 368 277 201 4,799 Number of credit officers 117 104 72 79 12 12 45 47 35 30 316 Avg. # of obs. per credit officer

10.4 7.95 7.39 7.75 9.17 11.42 11.53 7.83 7.91 6.7 15.2

Standard deviations are presented in parentheses * The delinquency rate is not directly comparable across institutions due to different definitions. Note, the values presented are Real Soles adjusted on a monthly basis using month 19 as the base value. The exchange rate was approximately 3.5 Soles per US$.

10

All of the indicators are comparable across institutions except for the delinquency rate. For most institutions, the statistic I am presenting is the value of portfolio that is more than 30 days past due divided by the entire outstanding portfolio. For one institution the number of days past due is different. The data for another institution corresponds to portfolio at risk, which is essentially a weighted measure of the delinquency rate (with substantially higher weights to amounts that are more days past due) The indicators of average monthly changes (for outstanding portfolio, number of clients, and delinquency rate) are generated for each month by subtracting last month’s data from the indicator for this month (i.e. change in number of clients for month 2= number of clients in month 2 minus number of clients in month 1). By construct, a new credit officer will not have this indicator until his/her second month. I did not include any new portfolio or clients that were transferred to a credit officer during a month (this affected 18 cases). 11 4.C. Staff Rotation There is a degree of staff rotation over the 31 month study period with an average rate of staff attrition of 1.05%. When a credit officer leaves an institution (either voluntarily or fired), in about 90% of the cases, the credit officer’s portfolio is transferred to a new credit officer. In about 10% of the cases, the old portfolio is divided whereby an experienced credit officer receives a portion and/or a new credit officer receives a portion. I have information that permits me to track where the portfolio of a departing credit officer goes. In order to better understand this phenomenon of staff rotation, we can take a closer look at the reasons for staff leaving an institution. For two of the institutions, I have information on why a total of 49 employees left the institution over the study period.12 The principal reasons were as follows: Table 3: Reasons for Staff Attrition Principal reason for leaving 1st MFI 2nd MFI Total

Fired for theft, fraud, or gross negligence 6 6 Fired for poor performance or behavior 13 5 18 Left voluntarily for personal reasons 4 3 7 Left voluntarily for a job in another institution 11 2 13 Transferred or promoted within institution 5 5

Total 39 10 49

A comparison of the pattern of these reasons before and after treatment does not appear to reveal any systematic differences. In fact, the reasons for staff rotation are somewhat more random than we might originally expect. Even among the credit officers that were fired for “poor performance”, the reasons were as varied as: showing up to work drunk, being rude to clients, sloppy handling of paperwork, and performance (measured by quantitative parameters) below 11 I would like to isolate this flow indicator from excessive noise and to try to ensure that it would better reflect a credit officer’s effort (i.e. it is much harder to generate 100 clients than to have them transferred to you from a leaving credit officer). 12 This represents 85% of the total number of employees that left in those institutions and 57% of the total number of employees that left all five institutions.

11

expectations. Employees leaving for personal reasons included credit officers who: had to move because of spouse’s employment, wished to spend more time with children, left for health reasons, and those who quit because they did not like the nature of the work or the work schedule. There is also considerable competition for credit officers among microfinance institutions; although incentives are part of the agent’s decision problem, it seems that differences in the base salary are even more important (based upon the qualitative interviews). The base salary in the 1st MFI listed above was significantly lower than that of the other MFIs and several employees did in fact leave to work for another MFI (both before and after the increase in incentives). In one of the MFIs I studied, not a single employee had left to work for another institution, because they were already receiving the highest base salary in the market. We are particularly interested in the nature of staff rotation because we might expect the increase in incentives to affect (either increase or decrease) staff rotation. In particular, we might expect staff turnover to decline following an increase in the incentives as the better workers are more willing to remain with the institution. On the other hand, we might expect either a constant or a higher rotation rate among the lower ability workers as a result of the increase in incentives. Overall, the monthly attrition rate increased slightly from an average 0.90% before the treatment effect to 1.31% after the increase in incentives (statistically significant at the 10% level). In Annex 1, there is a detailed tracking of the turnover rate by institution and level of employee productivity.13 Each month, by institution, the workers were classified by quartile according to the value of their outstanding portfolio.14 The average monthly turnover rate was calculated among credit officers of each quartile of each institution both before and after the increase in the incentive scheme. The aim is to see if there exist patterns whereby either the best or worst employees are leaving an institution. The results show that there are some slight tendencies (not quite statistically significant) in MFI1 and MFI2 towards an increase in staff attrition among workers in the lowest quartile following the treatment effect. In MFI5, there is a hint (not statistically significant) of a decrease in staff attrition among the highest quartile employees. Overall, this approach does not provide strong evidence of a consistent and powerful selection effect at least on the attrition of employees. This probably stems from the observation discussed above that there are several very diverse causes for staff attrition. The factors guiding the laying off of employees would most likely remain unchanged before and after the treatment effect. Similarly, policies for transfers and promotions and many of the personal reasons would not be affected by the incentive scheme.

13 Following the same approach used by Lazear (2000). 14 A parallel exercise was carried out for the delinquency rate parameter; however, the results were almost universally insignificant across all quartiles.

12

5. Theoretical framework Our theoretical framework is motivated by Holmstrom and Milgrom’s (1991) multi-tasking principal agent model. However, our focus is not on deriving the optimal contract but simply analyzing agents’ responses to a change in the contract offered by the principal. We will take the observed contracts offered by the principal to be constrained-optimal and analyze how agent performance and welfare are affected by specific changes in these contracts. Although both our theoretical framework and our empirical analysis are focused on agent responses to changes in contracts, any results or findings on agent response will be of direct interest towards a better understanding of the broader context of the principal-agent contract design problem. In Section 5.1, the features of the contracting environment and the principal’s maximization problem are presented first to frame our agent’s problem in the broader agency literature. A generic form of the actual contracts offered by the principals is presented in Section 5.2. We will then analyze the agent’s decision problem and the expected effects of a change in the contract in Sections 5.3 and 5.4. 5.1 Features of contracting environment In this model, the principal (MFI) employs an exogenously determined number (N) of identical risk-averse agents (credit officers) to implement a credit program. We assume that each agent is assigned to work in a specific geographic area; some areas are exogenously more or less profitable to work in than others because of local market conditions that can affect productivity for a given level of effort. The principal knows the overall distribution of market conditions but only the agents know their own local market condition. The principal would like the agents to exert effort towards three tasks. Two of the tasks generate financial profits for the principal: disbursement of loans and recovery of loans.15 The principal also has a social objective and would like agents to exert effort towards ensuring that the program has an impact on poverty by targeting poorer clients. All three tasks are direct substitutes. For example, if agents spend time disbursing loans, then they have less time available to monitor and enforce repayment of outstanding loans. In fact if too many loans are disbursed, then the agent will not be able to adequately enforce loan repayment. Similarly, if agents devote too much time towards monitoring and enforcing the repayment of loans, then they have less time available to disburse new loans. Maintaining an optimal balance of effort for each of these two tasks is very important for the principal. There also exist trade-offs between the financial and social objectives: serving poorer clients and/or providing them with complimentary services is less profitable.16 Poorer clients require more effort per dollar lent because they receive smaller loans while a considerable portion of the costs

15 In reality, the agent has more tasks but let us simplify tasks of promotion, screening, and disbursement into simply "disbursement" and the tasks of monitoring and enforcement of loans into "recovery." 16 While some publicly contest the existence of this trade-off, interviews with credit officers of several institutions confirm this trade-off in most all circumstances.

13

of administering and monitoring loans are fixed. Similarly, serving poorer clients often entails higher training costs: poor clients have lower literacy and less experience with credit. We assume that the benefits related to attainment of the financial and social objectives can both be expressed in monetary units. A function of the value of outstanding loans (at a given interest rate) and the repayment rate will yield revenue for the MFI. A monetary measure of the social benefits could be represented as: a) the amount of money the principal is willing to forgo in profits to attain such social output, or b) the expected value the principal will receive from a donor for expending effort on social tasks. The principal’s total profit (π) is the sum of benefits (where en is the vector of effort across the three tasks for agent n and m is the general market conditions faced by agent n) for each agent minus the salaries (S) and per-unit incentives (αTen) paid summed up over all N agents:

(1) ∑=

−−=N

nn

Tnn eSemB

1]),([ απ

The principal observes some information signals of each agent’s effort but cannot observe the actual effort or actions. These signals (Xi) are generated by the agent’s effort (ei), the general market conditions (m) faced by each agent, and by a random variable (εi), are: (2) X1= me1 + ε1 disbursing loans: value of current outstanding loans

X2= me2 + ε2 recovery of loans: current repayment rate of agent's portfolio X3= e3 + ε3 impact on poverty: number of clients and average loan size (measures impact with significant error)

All three signals are contractible, while the third one has considerably greater noise than the first two (ε3 > ε1, ε2). Let X1 represent value of outstanding loans and X2 represent the repayment rate. Due to high monitoring costs of measuring impact on poverty (and potential for fraud in reporting less verifiable indicators), this is weakly proxied by contracting on the total number of active clients an agent serves and the average loan size (a lower average loan size weakly proxies for poorer clientele). Each credit officer is randomly assigned a distinct geographic area in which to operate. Some areas are more profitable to operate in (i.e. more concentrated and dynamic commercial activity) than others. Let m∈(0,1], reflect these differences in market conditions; a higher m represents better market conditions.17 Each credit officer knows his parameter of local market conditions, while the principal only knows the overall distribution of m across all credit officers. Let us further assume that the geographic area does not have an effect on the productivity of the third task (impact on poverty): credit officers can always find poor clients with equal difficulty, regardless of market conditions.

17 In a similar model, the m parameter could capture some of the innate ability or knowledge of the worker that affects how effort is translated into output. For our purposes, we will maintain that all agents are identical but face different local market conditions. The role of local market conditions will be tested in the empirical strategy.

14

An additional and important assumption is that costless but imperfect monitoring by the principal can ensure that a minimum level of effort (e1min, e2min, e3min) is exerted by every agent for each task. Incentives are required to encourage effort beyond this minimum for each task. 5.2 Contract offered by principal As explained above, we do not need to derive a theoretically optimal contract since we observe the actual contracts offered by the principals (different MFIs), both before and after the changes in their incentive schemes. We are interested in the agents’ responses to actual changes in contracts. Although each scheme is unique, the five schemes that will be analyzed share a common structural form which will be summarized in generic form for two tasks as follows. Definitions:

X1 – level of output for task 1 X2 – level of output for task 2

X1benchmark – level of output for task 1 beyond which incentives are paid X2benchmark – level of output for task 2 beyond which incentives are paid X1cut-off – level of output for task 1 necessary to be eligible for incentives

X2cut-off – level of output for task 2 necessary to be eligible for incentives α1 – incentives paid for output in task 1

α2 – incentives paid for output in task 2 Generic contract form: (3) W(X) = S + α1X1 + α2X2 Where α1X1 is paid only if X1 > X1benchmark and X2 > X2cut-off Where α2X2 is paid only if X2 > X2benchmark and X1 > X1cut-off As we see, the contract is piecewise linear18: if certain conditions are met, the payment scheme is a linear function in each task. For each task, there are two breakpoints: a lower cut-off point needed to become eligible for incentives for either task, and a higher (benchmark) level above which incentives are earned. In Figure 1, these conditions are displayed graphically. Agents receive no incentives in the light gray area because they have not met the minimum benchmark levels to begin earning incentives for either task. The dark gray areas represent situations in which the agent performs above the benchmark for one task but does not meet the minimum cut-off for the other task. The white areas of the figure represent situations in which the agent actually receives incentives corresponding to one or both tasks. This contract form seems rational from the principal’s perspective. The principal only earns a profit if both outputs are met (i.e. multiplicative function of the two tasks). An agent could have a 100% repayment rate, but this is only important to the principal if the agent has a significant

18 In some institutions, the “linear” section of the incentive scheme is actually a series of payoffs corresponding to much smaller cut-off ranges. As long as these cut-off ranges are perceived by the agent to be relatively small, then these schemes can be considered comparable. The importance of such cut-off ranges (as opposed to a linear scheme) can be tested empirically (in a later version).

15

volume of loans. Similarly, the agent could have a very large loan portfolio, but this is only of value to the principal if the repayment rate is very high.

Output for task 1 (X1)X1cutoff X1benchmark

X2cutoff

X2benchmark

Output for

Figure 1: Eligibility to Receive Incentives

task 2 (X2) Incentives

Incentives

Incentives for

No incentives receivedin shaded areas

for task 2

for task 1

tasks 1 & 2

5.3 Agents' Problem Each of the n agents must choose a level of effort for each task that maximizes his/her utility given the contract offered by the principal. Each agent in the MFI faces the same problem since they are identical (in this version of the model).19 Let us assume that the agent’s utility function exhibits constant absolute risk aversion and is of the following form: (4) u = -exp[-r(W-C(e)]

where: u is the agent’s utility r is the agent’s coefficient of absolute risk aversion W is the sum of wages paid to the agent (both fixed and variable) e is the vector of effort exerted for each of the three tasks C(e) is the total cost of agent’s action, assumed to be strictly convex

If we assume error terms are normally distributed, the certainty equivalent of the agent’s utility can be expressed as a function of mean and variance of the error distribution20: The agent’s maximization problem is as follows:

19 Note that we are not considering any possible strategic group behavior on the part of the agents.

20

2( )

2The general formula when wealth is normally distributed is see Varian (1992).: ( ) ( )wrr wrwEu w e f w dw e

σ− −−= − =−∫

)',,(),,(2

)(max)5( 321321332211,, 321

ααααααααα mmmmreCeememSeee

Σ−−+++

16

The sum of the first four terms represents the agent’s expected wages. The fifth and six terms are the total cost of effort and the agent’s cost of risk exposure which is increasing in its coefficient of risk aversion and the variance of expected wages. The three first order conditions are: (6) *)(11 eCm ≥α *)(22 eCm ≥α *)(33 eC≥α

where Ci is the partial derivative of cost with respect to effort in each of the three tasks We see that the first order conditions, or agent’s incentive compatibility constraints, are independent of the fixed salary and risk.21 We obtain the standard result that agents will increase effort for each task until the marginal benefit of effort is equal to the marginal cost of effort. However, this is assuming that we are at an interior solution. The agent must decide whether he/she wishes to operate at the minimum level of effort and receive only a base salary or exert extra effort to obtain the incentives as well. This decision depends upon a comparison of these two equations:

Figure 2 illustrates the agent’s problem along one task. The agent can choose to exert a minimal level of effort such that output reaches (Xmin) and receive fixed salary S. For any output greater than Xbenchmark the agent receives the fixed salary plus an incentive payment per additional unit of output. The U1 and U2 curves represent the indifference curves of agents where U1 corresponds to an agent in an area of worse local market conditions. The U1 curve is much steeper because of a steeper cost function: each additional level of output requires a proportionally greater level of effort.

21 This is primarily due to our assumption of constant absolute risk aversion.

)()()8(

)',,(),,(2

*)(***).(int)7(

min

321321332211

eCSsolutioncornerU

mmmmreCeememSsolutionU

−=

Σ−−+++= ααααααααα

17

With a degree of heterogeneity in market conditions, we can easily imagine a scenario whereby a percentage of the agents receive no incentives and a large percentage does receive incentives of varying amounts. 5.4 Effects of an increase in incentives on output We now consider the effects on output when the incentive payments for all three outputs are increased by an equal proportion. In predicting the results, we need to consider the effects of several separate intermediating factors: the magnitude of the increase in incentives, the shape of the cost functions, the non-linearity of the payment scheme, heterogeneity of local market conditions, possible substitution between tasks, uncertainty of the environment, and possible selection/sorting effects. When considering the effects of an increase in incentives, we see from the first order conditions (6), that for an interior solution, the increase in output is dependent upon both the magnitude of the increase in incentives and shape of the cost function. If the cost of additional effort is very high at the margin, then an increase in incentives will have very little impact on output. We see in Figure 3 a case (see curves U2 and U2’) whereby an increase in incentives resulted in an increase in welfare of the agent, but no increase in output. Given the non-linearity of the payment scheme, it is possible that a portion of agents will not respond to an increase in incentives. For these agents, the increase in potential compensation is not enough to outweigh the marginal cost of increasing effort. This is particularly true for individuals who were producing at levels far below Xbenchmark to begin with. In Figure 3, we see that an increase in incentive payments from α to α’ would not result in any increase in output for an agent with a steep indifference curve such as U1. Returning to the multi-tasking framework of Figure 1, we see that even if an agent is highly productive in one output, they will not receive incentives until output is above the minimum cut-off for all tasks. The stereotypical case would

U1 U2

Xmin Xbenchmark X2*

Compensation

Figure 2: The role of cost functions and market conditions in determining optimal compensation and output levels

Output

Fixed salary S

Incentive payment α

18

be an agent with high numbers of clients, a large outstanding portfolio, but a delinquency rate on loans that is worse than the minimum cut-off. If the agent was relatively close to the cut-off, we could predict that the marginal benefit of increasing effort towards recovery of delinquent loans might outweigh the costs. However, if the agent was relatively distant from the minimum cut-off requirement, we might not expect a reduction in delinquency rate even after a significant increase in incentives.

Our assumption of heterogeneous local market conditions is a significant extension of the standard agency literature. It also implies that the effects of an increase in the incentive scheme could result in very different effects among agents. Even without detailed information on local market conditions, we could predict that agents who had similar levels of output before the change in incentives will react similarly. In other words, if two agents have reacted similarly to the first contract, we might expect that they would respond in a similar fashion to a uniform change in the contract. Agents who work in rural areas or in branches in particularly difficult areas would be less likely to earn incentives both before and after the change in the incentive scheme. Even when the increase in incentives is uniform across all tasks, the increase could result in substitution between tasks. This is possible due to the differences in the cost function, particularly at the margin. If we assume that the cost function for reducing the delinquency rate is steeper than the cost function for expanding loan portfolio, we would expect to see a slight shift in effort towards the task of generating portfolio. Although risk and uncertainty do not affect output under the interior solution, it does affect the choice of whether an agent chooses to provide minimal effort or work for the incentives (equation 7). The term “risk” is broadly interpreted as the loss of precision (or increase in noisiness) in the translation of effort into output. This can be separated into riskiness of environment (external shocks that affect output) and the precision of the performance indicator. Agents who work in relatively riskier environments (rural areas in particular) might be less likely

U1

U2

Xmin Xbenchmark X2’*

Compensation

Figure 3: An increase in incentives might not result in any increase in output

Output

Fixed salary S

Incentive payment α

Incentive payment α’

U2’

19

to respond to incentives. Also, if one task were less precisely measured than other, we might expect a lower treatment effect for that task. Finally, if we were to ease our restriction that all agents are identical and if we also assumed that both potential and current employees have outside alternatives, the increase in incentives could induce a selection effect whereby higher skilled workers are drawn to the institution and lower skilled workers would either remain or choose alternative jobs.

Table 4: Summary of Predicted Effects of an Increase in the Power of the Incentive Scheme

I. Average treatment effects 1. Magnitude of the effects on productivity will depend upon the magnitude of the

increase in the power of incentives 2. The degree of the effects will depend upon the percentage of employees who receive

incentives or who are very close to the benchmark for receiving incentives, before the change

II. Heterogeneous effects 1. Local market conditions: agents in rural areas and/or in difficult branches will not

increase output as much as branches in more vibrant areas. 2. Differences in the response to the incentives could vary across credit officer

characteristics such as gender, tenure in the institution, and initial equilibrium productivity levels.

3. There is a potential for a sorting or selection effect that results in an increase in productivity.

III. Differences across tasks 1. Differences in the effects across performance indicators are possible due to

differences in the cost functions and precision of performance indicator. In particular, changes in performance for the portfolio indicator might be quicker or larger than changes in the delinquency indicator.

2. Since institutions utilize number of clients rather than average loan size to proxy for social outreach, it is possible that the higher incentives for portfolio and delinquency rate might cause agents to shift from poorer clients (causing an increase in the average loan size).

IV. Timing of the effects 1. Since we are moving from one steady-state equilibrium to another one, we might

expect that any increases in productivity to be one-time increases in levels followed by a period of “steady-state” at this new level. The timing of such changes could vary across tasks (delinquency rate would probably be slower than outstanding portfolio or number of clients).

6. Empirical Strategy The empirical strategy entails credit officer fixed effects regressions on the main performance indicators and incentive compensation of credit officers. The new incentive scheme is implemented at a certain point in time throughout each institution. The aim is to measure whether the implementation of the incentive scheme resulted in an effect on performance levels immediately following this event after controlling for individual fixed effects, seasonality, and tenure in the institution. The discontinuity in time is the assignment mechanism and the

20

identifying assumption is that no other shocks (internal or external) occurred around the time of the implementation of the incentive scheme (these assumptions are addressed below). 6.1 Estimation of average treatment effects The basic fixed effects regression equation is: (1) ijttijtijijt MTy εγδα +++=

yijt is the output of individual i of firm j in time t: it will either be outstanding portfolio, number of clients, delinquency rate, change in outstanding portfolio, change in number of clients, or change in delinquency rate

αij is a fixed effect for each individual of firm j meant to capture all time invariant observed and unobserved individual characteristics that affect productivity

Tijt is the treatment dummy coded as 0 for all months prior to treatment and 1 for all months following treatment at the level of institution

Mt is the vector of 11 monthly dummy variables to capture seasonal effects that could affect productivity

ε ijt is the error term which captures unobservable effects on productivity at the individual level within each time period

This basic equation can be enriched to include time variant individual characteristics: (2) ijttijtijtijtijijt MXXTy εγββδα +++++= 2

21 In particular, the tenure of an individual in the institution is likely to affect productivity. To capture possible non-linear effects of tenure, we introduce Xijt which represents each individual’s months of tenure in the institution (which varies over time) as a quadratic function. 6.2 Exploring interactions with treatment effects We are interested in the coefficient δ, the average treatment effect across all individuals. However, we are also interested in possible heterogeneity in the impacts across institutions, branches, and individuals. Equation 3 includes additional treatment interaction terms (both time-invariant and time variant) that could capture possible heterogeneity in the treatment effects. (3) ijttijtijtijtijtijijijt MXXTXXy εγββδδδα +++++++= 2

21210 )( In particular, Xij will include: firm dummy variables, branch dummy variables, and whether the credit officer was “eligible” to receive incentives. Credit officers must meet minimum cut-off levels for quantity and quality of portfolio in order to qualify for incentives. I have derived this variable based upon whether the credit officer had a delinquency rate of less than 9%, three months prior to the treatment effect (therefore it is a time invariant variable). The Xijt variable could include the individual’s tenure in the institution. 6.3 Correction for autocorrelation of error terms Since the major output indicators (log of outstanding portfolio, delinquency rate, and number of outstanding clients) are “stock” variables, there is a real concern about possible autocorrelation of the error terms. For example, a shock that affects the delinquency rate (or either of the other

21

indicators) in time t would also likely affect the delinquency rate in time t+1. Even the flow variables (change in portfolio, change in delinquency rate, and change in outstanding clients) are susceptible since shocks that affect the ability to produce output in time t are likely to affect production in time t+1. Equation 4 incorporates an assumption of autocorrelation of the first degree:

(4) ijtijtijt

ijttijtijtijtijtijijijt

where

MXXTXXy

ερµµ

µγββδδδα

+=

+++++++=

−1

221210 )(

The hypothesis that ρ = 0 can be tested using a procedure that is similar to the Durbin Watson test, but specifically developed for panel data by Bhargava, Franzini, and Narendranathan.22 If ρ is significantly different from zero, we can correct for the autocorrelation using the Cochrane-Orcutt estimator. Two main assumptions in using this estimator are that the correlation coefficient ρ is constant over the study period and across groups of individuals (we could verify this assumption). 6.4 Incentive and selection effects In estimating the effects of the change in incentive scheme on productivity, we must consider both potential incentive and selection effects. In Section 4.B, we found that there is no strong evidence that the increase incentives (treatment) affected the rate or types of workers that leave the institution. However, there remains the possibility that the treatment has had an effect on the type of worker joining the institution (i.e. higher ability workers are attracted to the institution after the increase in incentive scheme). By including our full sample of credit officers before and after the treatment, we are essentially estimating the combined incentive and selection effects. Given our data set, we are unableto separately identify these two effects. Even if we were to find differences in the productivity of newly hired workers immediately before and after the treatment, we would not be able to conclude whether the differences were attributed to different quality workers or the same quality of workers who are now more motivated by the incentives. 6.5 Exploring heterogeneity of the fixed effects The estimated coefficients of αij from equation (4) represent the fixed effects of the credit officer: including ability, knowledge, and local market conditions. We will regress these estimated fixed effects on credit officer characteristics and branch dummy variables. We are interested in seeing how these characteristics affect underlying “productivity” as well as the response to the incentives. 6.6 Assumptions and limitation of empirical approach A potential limitation of the data set is the fact that the changes in the incentive schemes occur simultaneously throughout each institution. For the case when we run all institutions together, we can use the differences in implementation dates across institutions to identify and control for

22 Bhargava, A., L. Franzini, and W. Narendranathan. “Serial Correlation and the Fixed Effects Model.” Review of Economic Studies, Volume XLIX, 1982, pages 533-549.

22

macro-economic and/or seasonal shocks that might coincide with the implementation of an incentive scheme through the inclusion of time dummy variables. There are in effect three specific assumptions underlying the general empirical approach to estimate the average effect of the new incentive scheme: 1. We implicitly assume that there are no other simultaneous unexplained shocks at the level

of the institution (other than the seasonal and macro-economic shocks that affect all institutions) that that could affect performance data at the credit officer level. The most likely form of an institution-specific shock could be the introduction of other new policies by the MFI (i.e. changes in: credit products, management of delinquency cases, new management information system, etc.). This potential problem is being addressed by our decision to only include relatively stable MFIs that have not implemented any other major initiatives or policies over the study period. Two additional institutions were excluded from the study because of simultaneous changes of either an administrative (reorganization to include additional supervisors) or programmatic (changes in product rules) nature.

2. The second major assumption is that there is no endogeneity problem whereby the factors affecting the decision to implement/increase an incentive scheme also affect the efficiency/productivity of workers. For example, if the introduction/increase of staff incentive policies were implemented in response to a crisis or only in periods of high profitability, then it would be impossible to separate out the treatment effects. Evidence from the field (in-depth interviews with managers of 25 MFIs) suggests that there is little evidence of either problem (implementing due to crisis or in periods of high profitability). Institutions decide to implement/change incentives if they feel that it will be cost-effective. Managers base their decision in large part on what they hear from consultants, donors, other institutions, and especially their own workers and branch managers. It is also important to note that there is a significant time lag (minimum of two months) between when an MFI chooses to adopt or significantly change a monetary incentive policy and its actual implementation. We will test this assumption by looking for any significant positive or negative trends in the two-three months immediately prior to implementation of the new incentive scheme in each institution (after controlling for seasonality). The sample includes virtually all microfinance institutions (all of the ones I am aware of) in Peru that: • implemented a new scheme or significantly altered an existing incentive scheme in the

past 3 years; and • for which data was available for several months before and after the implementation or

change; and • in which no other significant administrative or programmatic changes occurred. There is an inherent self-selection bias for choosing treatment (adoption of the new incentive scheme) whereby the only institutions that choose to adopt incentives are the ones that expect to benefit the most from it. However, this should not affect our analysis, since we are interested in assessing the effects of incentive schemes among institutions that have chosen to implement or change them.

23

3. The third assumption is that a significant portion of the effect of the incentive policy can be accounted for within the study time frame. This assumption can be explored through analysis of the data. It is probably reasonable to assume that the majority of the effects of the incentives could be accounted for within about 6-8 months following implementation of the incentive scheme. The average loan period is 4-5 months and the maximum loan period is 6 months. Loans are generally repaid in equal monthly (sometimes bi-monthly) installments, so a new client could become delinquent with one month.

7. Results 7.1 Graphical Analysis of Effects Before proceeding with the regression results, it might be interesting to analyze some graphs of the trends over time for several indicators. Such an analysis provides an overview and some intuition about the results, but will not enable us to clearly isolate any treatment effects. In Annex II.1, we can see that the value of outstanding portfolio and the number of clients increases with tenure in the institution (in months). There is a strong association between the number of months a credit officer has been employed and his/her indicators for outstanding portfolio, number of clients, and delinquency rate. This makes sense since it takes time to build up a clientele and a portfolio. Therefore, it is difficult to separate the tenure effect from the treatment effect of any increases in outstanding portfolio or number of clients over time. All of the remaining graphs in Annex II are also locally weighted smoothed curves, but the curves are created separately for the periods before and after the treatment effect for each institution. In Annex II.2, we see how the value of outstanding portfolio has evolved in each of the five institutions. Outstanding portfolio was increasing for all institutions even before the treatment effect. For most institutions, there appears to be a slight increase in the slope of the increase after the treatment effect (and a jump in levels for MFI3). The effect of the treatment on the delinquency rate seems quite dramatic for most institutions (see Graphs 10-14). Delinquency rate had been stable or even increasing for most institutions before the treatment effect, and is decreasing for all institutions afterwards. The apparent impact of the treatment on the number of clients (Graphs 15-19) seems very similar in pattern to what we saw for value of portfolio. The final series of graphs (20-24) in Annex II shows the evolution of the incentive compensation over the study period. In all cases, we see an immediate jump in levels of incentives beginning with the treatment effect: for most institutions, the incentive compensation seems to immediately double or triple (with the exception of MFI4 which seems to have a much smaller gain). Of greater interest is the very pronounced increase in the months following the treatment effect: this implies that credit officers are responding to the incentives and are benefiting from them. We next explore whether we can more clearly identify any treatment effects using a fixed regression approach.

24

7.2 Regression Procedures The regression results are presented in Annex 3 in the following tables: Estimations using pooled sample of all institutions (with institution dummy variables)

Table 1: Outstanding portfolio and change in outstanding portfolio Table 2: Number of clients and change in number of clients Table 3: Delinquency rate and change in delinquency rate Table 4: Average loan size Table 5: Incentive compensation

Regressions of institutions estimated separately Table 6: Outstanding loan portfolio Table 7: Growth in outstanding loan portfolio Table 8: Number of clients Table 9: Growth in number of clients Table 10: Delinquency rate Table 11: Change in delinquency rate Table 12: Average loan size Table 13: Incentive compensation

Limited model (no treatment interactions) Table 14: Timing of incentive effects across indicators

Analysis of fixed effects Table 15: Heterogeneity across credit officer fixed effects

The first 5 regression tables utilize a pooled sample of five institutions to estimate each of the dependent variables. For each dependent variable, the results of 2 model specifications are given using the sample of all MFIs. The first specification includes tenure, tenure-squared, month dummy variables to capture seasonality effects, a treatment dummy to capture the average treatment effects across all institutions, and firm-treatment interaction dummies to capture firm specific effects. The second specification includes the same covariates plus the following treatment interaction dummy variables: tenure, tenure squared, “eligibility” (for all regressions except the delinquency regressions). “Eligibility” is a dummy variable defined as “1 if in the period three months prior to treatment effect, the agent had a delinquency rate < 9%.” If a credit officer reduces his/her delinquency rate below the cut-off, then he/she becomes eligible again for incentives. This variable helps to capture our predicted result that agents that are below the minimum cutoff will be less likely to respond to a change in the incentive scheme. All MFIs have an actual cut-off (established in the rules of the incentive policy) that is between 8-8.25% above which, an agent is unable to receive any incentives for that month. The regression results presented in Tables 6 – 13 are institution-specific regressions. The model specification for each of the dependent variable includes tenure, tenure-squared, month dummy variables, a treatment dummy to capture the average treatment effect, and the following treatment interaction dummy variables: tenure, tenure squared, eligible, gender (for MFI 1), and branch offices. In Table 14, we investigate the timing of the treatment effects. We estimate the most simplified treatment model with no interaction dummy variables and vary our definition of treatment

25

period. The treatment dummy variable is estimated to detect effects that occur within the first month, first two months, first three months, first four months, and the first five months. We also check for possible endogeneity by testing for significant effects one, two, and three months before the treatment period. All of the regressions are estimated as fixed effects AR(1) models utilizing a Cochrane-Orcutt panel-by-panel transformation to correct for autocorrelation. The Bhargava modified Durbin-Watson statistic (modified for use with panel data) ranged .44 and .91 for the following variables: outstanding portfolio, loan size, and delinquency rate. We can reject the hypothesis of no autocorrelation for these variables; the coefficient for the correlated error ranged between .41 and .92. There was little or no autocorrelation (can not reject it, and very low correlated errors) for the three flow indicators or for the number of clients. In Table 15, the estimated fixed effects from the regression in Table 6 are used to examine patterns of variation in productivity of credit officers. The estimated fixed effects are regressed on branch and individual characteristics. 7.2 Fixed effects regression results for pooled sample The results from the fixed effects pooled regressions of all institutions can be found in Annex 3, Tables 1-5. Let us first discuss some of the common trends across performance indicators and types of specifications. Seasonal effects- We see a fairly strong seasonal effect on the variables for portfolio, change in portfolio, average loan size, and number of clients. In particular, November and December have slight peaks versus slight dips for March-June. We see little or no systematic seasonal effect on change in number of clients, delinquency rate, or change in delinquency rate. These results are robust across model specifications. Effect of tenure- Tenure has a consistently significant impact across virtually all specifications. In general, it has a positive impact on the stock variables and a negative impact on the flow variables. This is consistent with a story of decreasing marginal returns to tenure. Each additional month of tenure adds an average of about $1,000 to $2,000 or between 12 and 15 clients. Treatment effects across firms- When we consider the firm-treatment interaction dummy variables for our simplest specification, we find some very interesting patterns. The specific effects for different regressions is discussed below, but in most cases, the treatment effect is limited to only one or two MFIs. Treatment effects between eligible and in-eligible credit officers- As expected, we do see a significant difference in the treatment effect between eligible and in-eligible officers for: change in portfolio, number of clients, change in number of clients, and incentive compensation. However, the magnitude of the effect is much less than we might have predicted, and it is surprisingly absent for the parameter for value of outstanding portfolio. 7.2.A. Outstanding portfolio

26

In Regression Table 1, we see that the only MFI that had a significant improvement in outstanding portfolio due to the increase in incentives was MFI1. The marginal impact of the treatment, after controlling tenure, seasonality, and credit officer fixed effects for MFI1 was effect is between S/17,000 and S/18,000 (depending upon specification). This represents about a 19% increase in outstanding portfolio for MFI1. 7.2.B. Change in outstanding portfolio In the third and fourth columns of Regression Table1, we see that only MFIs 2 and 3 had a significantly higher growth rate of outstanding portfolio due to the treatment. It is interesting that the treatment effect for outstanding portfolio would be higher in MFI1, while the monthly growth is higher in MFIs 2 and 3. There are at least two contributing factors to this apparent dichotomy. The first is that in MFI 1, both before and after treatment, the rate of growth was very high relative to the size of outstanding portfolio (MFI1 was in a phase of expansion). For MFI1, the effect on growth rate was very short-term, until a new steady-state was achieved (as predicted in the theory section). For MFIs 2 and 3, the initial stock of portfolio was relatively high and the monthly growth rate relatively low. Thus, although the treatment had an effect on growth of portfolio, this effect does not significantly change the level of portfolio. Overall, the treatment effect was stronger among eligible credit officers and those with less tenure. The magnitude of the effects are quite impressive: an increase in outstanding portfolio of almost 20% for MFI1 and more than a 100% increase in the monthly growth rate of outstanding portfolio for MFIs 2 and 3 after controlling for tenure. Credit officers who had been “sitting” on their portfolios (not generating new clients) became motivated to generate new clients after the increase in incentives. 7.2.C. Number of clients and change in number of clients We see (in Regression Table 2) that when controlling for all other variables, the treatment response of eligible credit officers on number of clients and growth in number of clients was significantly higher than among non-eligible credit officers. MFI1 was the only one that had a positive treatment effect on the number of clients about 7%, but not on growth in clients (as was the case for outstanding portfolio). Meanwhile, MFI 2 was the only one that experienced a positive impact on growth in number of clients. The average credit officer in MFI2 more than doubled the monthly rate of bringing in new clients after the treatment. 7.2.D. Delinquency rate and change in delinquency rate In Regression Table 3, we see that there is no significant treatment effect on delinquency rate, regardless of tenure or institution. The sign of the coefficients point to a reduction in delinquency rate and this is relatively close to being significant for MFI 1. In Regression Table 6, we see that there is a significant reduction in delinquency when all institutions are run together with no treatment interaction effects. In sorting out these differences, it is first important to note that our model for delinquency and change in delinquency has much less explanatory power than our models for other performance indicators (overall adjusted R-square of around .03 for delinquency rate versus .35 for outstanding portfolio). Therefore, our failure to detect an effect on delinquency rate in Table 3 while we did in Table 6, could be related to the increased number of observations we have to

27

predict a treatment effect in Table 6 (i.e. we are not dividing the observations on treatment across the five institutional dummy variables and other treatment interaction terms). 7.2.E. Average loan size Recall from the theory section, that we are interested in tracking the average loan size as a rough proxy for outreach to the poor. There seems to be little effect of the treatment on average loan size except among credit officers in MFI 3 who increased there average loan size by S/.108. Even though this is significant, the magnitude (about a 7% increase) is not overly worrisome, given the average loan size. 7.2.F. Incentive compensation received We see that, after controlling for tenure, the increase in incentives resulted in an actual increase in incentive payments only for institutions 1, 2, and 3. There is a very large difference (as would be expected) between “eligible” and “ineligible” credit officers. In MFI1, the ineligible credit officers did not increase their compensation, while on average, the eligible credit officers increased their incentives by S/ 125. Even ineligible credit officers increased their incentives in MFIs 2 and 3 (they must have become eligible for incentives at some point after the increase in incentives took place). It is important to remember that these estimates have controlled for tenure: before controlling for tenure, the actual increases in incentive payments is much higher. 7.3. Discussion of firm-specific regressions One key difference between the pooled and firm-specific regressions is that the pooled regression assumes that tenure, eligibility, and seasonality have the same effect across institutions and the same interactions with treatment. The firm-specific regressions are more efficient if these underlying equations are very different across institutions. However, we lose a large number of observations that could help detect treatment effects if they are similar in structure across institutions. We have included the branch treatment effects in the firm-specific regressions.23 MFI 5 was dropped due to collinearity in the regression. This is primarily due to the fact that there were much fewer months after the treatment effect for this MFI. In comparing the pooled and firm-specific approaches, we see that there are some differences- primarily with regard to MFI 3. The treatment effects (or lack of) for institutions 1, 2, and 4 remained relatively unchanged between the two approaches. However, there are several differences for MFI 3. We now detect a significant treatment effect (whereas we did not under the pooled approach) for the loan portfolio (at the .1 level of confidence), number of clients, and change in number of clients. On the other hand, MFI 3 no longer has an estimated treatment effect on change in loan portfolio and average loan size. One of the more interesting findings of the firm specific estimations is the significant differential effect of the treatment across branches, especially for MFIs 1 and 2. This confirms our suspicion (theoretical framework and qualitative interviews) that local market conditions are important determinants of response to the incentives. The magnitude of the branch effect can be very large; for example, the treatment effect in Branch 5 of MFI2 was more than S/ 60,000 greater than the treatment effect in Branch 10. 23 We could have included them in the pooled sample, but comparing branch effects across institutions would have held less meaning than comparing branch effects within an institution.

28

The following table summarizes, by indicator, which institutions had significant differences in the treatment effect across branches: Heterogeneous treatment effects across branches Indicator MFIs with significant

differences across branches Loan portfolio 1, 2 Change in loan portfolio 1, 2, 4 Number of clients 1 Change in number of clients 2, 4 Delinquency rate Change in delinquency rate Average loan size 1, 2 Incentive compensation received 1, 2, 4

7.4 Exploring the Timing of Treatment Effects Up until now, we have been considering the treatment effect as corresponding to all months after the change in the incentive scheme. We can gain further insight into the nature of the effect by a closer examination as to exactly when do the effects of the treatment become evident. From a practitioner’s perspective, it is important to see whether the effects on performance occur rapidly (1-2 months) or if credit officers respond slowly (4-5 months) to the change in incentives. Also, this analysis can serve as a robustness check for possible concerns of endogeneity regarding the decision to increase the incentive scheme. In the empirical strategy section, we discussed the possibility that other factors might affect the decision to increase an incentive scheme which could also affect the efficiency/productivity of workers. We can test this assumption by looking for any significant positive or negative trends in the two-three months immediately prior to implementation of the new incentive scheme (after controlling for seasonality). Regression Table 14 in Annex 2 illustrates the coefficients and t-statistics corresponding to different specifications of the treatment effect time period and across different performance indicators. For example, for treatment period corresponding to month mT (in the chart), months 1 though (T-1) are coded as Treatment= 0, month mT is coded as Treatment=1 and months mT+1 through the end of the study period are coded as Treatment= missing. The table of full regression results is not presented here, just the coefficients of interest. We see that when the treatment has an effect, it occurs very rapidly, beginning in the first month. The one exception is the effect on delinquency rate which takes five months to reach its full effect. The flow variables of change in portfolio and change in number of clients exhibit the greatest increase in the first month of treatment and then gradually decrease over time. The flow variable for change in delinquency rate shows a steady rate of decrease over the five months following treatment. These results are consistent with the hypothesis that the increases for portfolio and number of clients would be one-time improvements in levels: i.e. increasing until

29

marginal benefit equals marginal cost and then remaining at the new steady-state level of effort/output. However, the effect on delinquency rate and change in delinquency rate continues through the end of the study period. We also test the significance of other months just before the treatment period. It is interesting to note that the coefficients for the months immediately prior to the treatment period were significant and negative for the indicators of change in portfolio, number of clients, and change in number of clients (and not significantly different for portfolio). This suggests that there might in fact be some endogeneity in the choice to implement the incentive scheme. If so, a possible story is that the managers were experiencing lower performance and/or lower motivation of their credit officers and decided to increase the incentives to increase output. The direction of this potential bias (the significantly negative performance in the months leading up to the treatment effect) is not clear. If it represents a very short-term dip, then our analysis would be over-stating the effects of the treatment (our treatment effect would be capturing part of the natural recovery of the dip). On the other hand, if the negative period is caused by a longer term period of stagnation and poor productivity, then our true treatment effect might be stronger than our estimates. 8. Conclusions 8.1. Average treatment effects Our findings show that an increase in the power of an incentive scheme can significantly improve productivity, but not necessarily. There were significant improvements in productivity in three of the five institutions following the increase in incentive compensation. The magnitude of the effects in the three MFIs was quite large: reaching as much as a 20% average increase in stock variables or a 100% increase in flow variables (i.e. monthly growth in portfolio) after controlling for individual fixed effects, seasonality, and tenure in the institution. Our theoretical framework predicted that the response to incentives will be affected by the magnitude of the increase in incentives and the percentage of credit officers that receive or were close to receiving incentives. Although we don’t have enough institutional observations to formally test our theoretical predictions, we can at least see if the prediction holds for the five institutions we studied. The only MFIs that had a significant treatment effect on productivity were MFIs 1, 2, and 3. From Table 2, we see that MFIs 1, 2, 3, and 5 all had rather large average increases in incentive compensation (between 10% and 20% of fixed wages) while MFI 4 only increased average incentive compensation by about 5% of fixed wages. Furthermore, MFI 4 had the lowest increase (and lowest overall after treatment) in the percentage of credit officers receiving incentives. The actual effects in MFIs 1, 2, 3, and 4 were at least consistent with our theoretical predictions. However, it is unclear why there was no significant treatment effect in MFI 5. Of course, the power of the incentive scheme is only one of several factors affecting its success; there are several design and implementation issues that would greatly affect the success of the incentive scheme for all institutions.

30

8.2 Heterogeneity of effects The effect of the treatment varied significantly across individuals and branches. In particular, we found that individuals who met the minimum cut-off for delinquency rate (our “eligible” variable) just before the treatment were much more responsive to the treatment. This differential treatment effect across eligibility disappears when we include branch-treatment interaction effects. In other words, the eligibility criterion is important and this is correlated with differences across branches. While the differences in treatment effects exist across branches only within MFIs 1, 2, and 4 the differences are quite large in magnitude, at least for MFIs 1 and 2. MFIs 1 and 2 have branches in multiple cities and rural areas; MFI 4 has branches in multiple cities but not rural areas; and MFIs 3 and 5 only operate in the capital city. Thus, the more heterogeneous the local market conditions between branches, the greater the differences in treatment effect. Our comparison of which types of credit officers leave before and after the treatment (by quartile of productivity) showed some weak evidence of a sorting effect (though neither very large nor significant). However, given our dataset, we can not separate out the treatment effects into incentive and selection effects. Our estimates reflect the combined incentive and selecting effect. 8.3 Differences across tasks We are also interested in the multi-dimensional nature of the contract and possible substitution between tasks. The theoretical framework suggests that substitutions between tasks are possible even if the payments of incentive scheme are increased by the same amount for each task. The central level managers that were interviewed were very concerned about a possible substitution in effort towards disbursing more loans at the expense of trying to reduce the delinquency rate. There is another concern expressed by donor agencies and managers of certain very socially oriented MFIs that the incentive scheme will result in a shift from poor clients to non-poor clients. We found no evidence of an increase in the delinquency rate; in fact, the evidence suggests a small (but insignificant) decrease in delinquency rates. The structure of the incentive contracts, with the cut-off level for delinquency rate, is probably an important contributing factor for this check on an increase in the delinquency rate. We do not have data on the wealth level of clients and can not truly address the concern about a shift towards nonpoor clients. However, there is a belief (among practitioners) that average loan size is a rough proxy for wealth level (poorer clients are unable to borrow as much as nonpoor clients). Our analysis shows very little evidence (only MFI 3, when estimated within the pooled sample) of an increase in average loan size over the study period. 8.4 Timing of the effects across indicators Our analysis of the timing of the effects across indicators (Table 14) confirm our hypothesis that the incentives will produce one-time increases in productivity and then remain at the new steady-state level of output. Also, as expected the effect on flow (growth) variables was relatively

31

quicker and shorter than on stock (levels) variables. The effect on delinquency rate and change in delinquency rate took longer to fully materialize than the other variables. One disturbing finding, from an econometric perspective, is that there was a slight dip in certain productivity indicators (particularly for change in number of clients and change in delinquency rate) in the one to three month period prior to the treatment effect. As discussed above, the direction of this potential bias (in estimating our treatment effect) is not clear. If it represents a short-term dip, then we are over-estimating our treatment effect; if it represents a longer term period or trend towards stagnation or poor productivity, then we are under-estimating our treatment effects. One reassurance is that the magnitude of this difference is not very large and it is only present in certain indicators. 8.5 Concluding remarks One conclusion we can draw from this research is that an increase in the power of incentives will not produce equal effects across individuals, branches, or tasks. The increase in incentives did have significant and quite powerful effects in some institutions, some branches, and among some individuals. However, there are many credit officers who did not respond to the incentives. The effects, when they were present, occurred relatively rapidly (within a couple of months) and then productivity stabilized at a new steady-state level. Although we have been analyzing treatment effects within the context of a multi-tasking contract, we saw little or no substitution effects between tasks. This finding should be taken within the context of our study. Each of the five institutions already had a multi-dimensional incentive scheme, even if it was very weak-powered. Our treatment effect was essentially to increase the power of this contract equally across all tasks. There have been several anecdotal accounts of institutions that indeed suffered very costly substitutions between portfolio and delinquency following the implementation of a new incentive scheme (presumably because they did not correctly assign weights between tasks or establish appropriate minimum cut-off levels for performance in each task). Furthermore, we are unable to capture costly long-term substitution effects away from quality. Interviews with credit officers and managers have reported a decrease in the quality of attention to clients as credit officers become more focused upon output indicators. Individual actions can have significant externalities for the institution that would not be captured by our data: if one credit officer treats potential clients in a very brusque or demanding manner that decreases the time needed to process new clients, a percentage of potential clients will be turned off and will never apply for another loan from the institution (and might even convince other individuals to not apply). Similarly, a credit officer might decide to ignore delinquent clients if their outstanding amounts are very low (i.e. he is maximizing his benefit from incentives minus cost of effort), however, this could produce a negative externality where the reputation of the institution as an enforcer of all delinquent loans is undermined. Perhaps one of the most important contributions of this work is to shed some light on the potential role of heterogeneous local market conditions within a standard principal agent framework. There are many real world situations where local market conditions greatly affect the productivity of the workers but are not sufficiently distinguishable or exogenous to contract upon. The application of a uniform one-size-fits-all contract, even if it is constrained-optimal,

32

will not necessarily (perhaps even rarely) provide incentives for all employees to increase effort and output. Although our findings in this paper were not able to document this: interviews revealed that a considerable percentage of credit officers in certain branches became less motivated following the increase in incentives since they were not able to benefit from it even though they felt they were working just as hard as other credit officers. This phenomenon of heterogeneous local market conditions could even be an additional contributing factor to the long-standing theoretical puzzle of why we see very few real world contracts with high- or even low-powered incentives. This situation is conceptually different from a standard reason for low-powered incentives: the degree of riskiness of the environment but reasonably plausible.

33

Bibliography Bazoberry, Eduardo. "We Aren't Selling Vacuum Cleaners: PRODEM's Experiences with Staff

Incentives." Microbanking Bulletin, April 2001, 11-13. Bhargava, A., L. Franzini, and W. Narendranathan. “Serial Correlation and the Fixed Effects

Model.” Review of Economic Studies, Volume XLIX, 1982, pages 533-549. Chiappori, P.A. and B. Salanie. “Testing Contract Theory: A Survey of Some Recent Work.”

Presented at the World Congress of the Econometric Society in Seattle, August 2000. Cockburn, Iain, Rebecca Henderson, and Scott Stern. “Balancing Incentives: The Tension

Between Basic and Applied Research. Working Paper, August 2000. Deckup, John and Carol Cirka. “The Risk and Reward of a Double-Edged Sword: Effects of a

Merit Pay Program on Intrinsic Motivation.” Nonprofit and Voluntary Sector Quarterly, September 2000, 29(3), 400-418.

El Shami, Nabil. "Staff Incentive Schemes." Presented at the 5th Annual Conference of the

MicroFinance Network in Alexandria, Egypt, 1997. Fernie, S. and D.Metcalf. “It’s not what you pay it’s the way that you pay it and that’s what gets

results” Mimeograph: London School of Economics, 1996. Gibbons, Robert. “Incentives in Organizations.” Journal of Economic Perspectives, Fall 1998,

12(4), 115-132. Holmstrom and Milgrom. “Multitask Principal-Agent Analyses: Incentive Contracts, Asset

Ownership, and Job Design.” Journal of Law, Economics, and Organization, 7 (1991), 24-51.

Holtmann, Martin. “Designing Financial Incentives to Increase Loan Officer Productivity:

Handle With Care!” Microbanking Bulletin, April 2001, 5-10. Kreps, David. “Intrinsic Motivation and Extrinsic Motivation.” American Economic Review.

May 1997 (Papers and Proceedings), 87(2), 359-364. Lazear, Edward. “Salaries and Piece Rates.” Journal of Business, 1986, vol. 59 no. 3, pp. 405-

431. Lazear, Edward. “The Power of Incentives.” American Economic Review, May 2000 (Papers

and Proceedings), 90(2), 410-414. Lee, Nanci. “Staff Incentives to Retain Talent in Your MFI: The Case of CRESCE,

Mozambique." Presented at the Fourth Frankfurt Seminar on New Development Finance held in Frankfurt, Germany from September 3 to 8, 2001.

34

Marsden, David. “Teachers and performance management: one year on (Provisional results).”

Working paper, 2001 (http://www.lse.ac.uk/Depts/industrial/teachers-study/). Micocredit Summit. “2001 State of the Microcredit Summit Campaign Report.”

www.microcreditsummit.com Paarsch, H. and B. Shearer. “The Response of Worker Effort to Piece Rates: Evidence from

British Columbia Tree-Planting Industry,” in Journal of Human Resources, 34, 643-667. Prendergast, Canice. “What Happens Within Firms? A Survey of Empirical Evidence on

Compensation Policies.” in Labor Statistics and Measurement, Robert Topel and Marilyn Manser (eds.). University of Chicago Press for the National Bureau of Economic Research, 1997, 329-354.

Preyra, Colin and George Pink. “Balancing incentives in the compensation contracts of

nonprofit hospital CEOs.” Journal of Health Economics, 20 (2001), 509-525. Robinson, Marguerite."Staff: Preliminary Thoughts on How to Retain Good Staff in MFIs."

Presented at the 5th Annual Conference of the MicroFinance Network in Alexandria, Egypt, 1997.

Shearer, B. “Piece Rates, Fixed Wages and Incentives: Evidence from a Field Experience.”

Mimeo, Laval University, 1999. Slade, Margaret. “Multitask Agency and Contract Choice: An Empirical Exploration.”

International Economic Review, Volume 37, Issue 2 (May, 1996), 465-486.

35

ANNEX I: Analysis of Staff Attrition Analysis of Staff Attrition by Quartile of Productivity Dummy variable=1 for employees last full month of employment, 0 for all other months that they are employed. The mean represents the percentage of attrition (percentage of employees leaving voluntarily or involuntarily) per quartile of outstanding portfolio. Institution 1

T=0 T=1 T test difference Quartile Mean Std. Dev Obs. Mean Std. Dev Obs. Mean P value 1 (Lowest) 0.74% .0860651 269 2.86% .1669967 210 .0211365 0.0735 2 2.31% .1504374 260 3.40% .1816206 206 .0109037 0.4789 3 2.27% .1493158 264 0 0 208 -.0227273 0.0287 4 (Highest) 3.49% .1838419 258 5.42% .2269464 203 .0193035 0.3136 1.29% .1130298 1778 1.57% .12454 1524 .0028121 0.4966

Institution 2

T=0 T=1 T test difference Quartile Mean Std. Dev Obs. Mean Std. Dev Obs. Mean P value 1 (Lowest) 1.47% .1208178 136 4.52% .2083309 155 -.0304554 0.1353 2 1.49% .1217093 134 2.61% .1600868 153 -.0112184 0.5092 3 0.75% .0863868 134 0 0 153 .0074627 0.2861 4 (Highest) 0 0 128 0 0 151 0 0.35% .0590896 1428 1.46% .119823 756 .0110489 .0040

Institution 3

T=0 T=1 T test difference Quartile Mean Std. Dev Obs. Mean Std. Dev Obs. Mean P value 1 (Lowest) 0 0 33 0 0 39 0 2 4.35% .2085144 23 3.03% .1740777 33 -.0131752 0.7983 3 0 0 32 0 0 39 0 4 (Highest) 4.55% .2132007 22 0 0 26 -.0454545 0.2817 1.09% .1045385 182 0.55% .0741249 182 -.0054945 0.5633

36

Institution 4

T=0 T=1 T test difference Quartile Mean Std. Dev Obs. Mean Std. Dev Obs. Mean P value 1 (Lowest) 2..24% .1484971 134 3.13% .174906 96 .0088619 0.6792 2 0.77% .0877058 130 0 0 90 -.0076923 0.4066 3 1.49% .1217093 134 1.05% .1025978 95 -.0043991 0.7742 4 (Highest) 0.83% .0909091 121 0 0 87 -.0082645 0.3978 0.76% .0870367 918 1.03% .101011 486 .0051672 0.6064

Institution 5

T=0 T=1 T test difference Quartile Mean Std. Dev Obs. Mean Std. Dev Obs. Mean P value 1 (Lowest) 1.43% .1195229 70 1.85% .1360828 54 .0042328 0.8543 2 0 0 70 0 0 49 0 3 4.29% .2039973 70 0 0 49 -.0428571 0.1446 4 (Highest) 4.48% .2083729 67 0 0 49 -.0447761 0.1358 0.99% .099359 703 0.39% .062137 259 -.0066049 0.3562

37

ANNEX II: Graphical Analysis II.1 Analysis of the Effect of Tenure on Performance Indicators (all credit officers from all MFIs are included) Graph 1 Graph2 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

050

0000

1000

000

1500

000

2000

000

Portf

olio

(Rea

l Sol

es)/l

owes

s v

t

0 10 20 30 40 50Tenure (in months)��

��Portfolio (Real Soles) lowess v t

Value of Portfolio by Tenure

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

040

060

080

0N

umbe

r of C

lient

s/lo

wes

s n

t0 10 20 30 40 50

Tenure (in months)�� Number of Clients lowess n t

Number of Clients by Tenure

Graph 3 Graph4 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.2

.4.6

.8D

elin

quen

cy ra

te/lo

wes

s d

t

0 10 20 30 40 50Tenure (in months)��

�� Delinquency rate lowess d t

Delinquency Rate by Tenure

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

010

2030

4050

Tenu

re (i

n m

onth

s)/lo

wes

s t

mon

th

0 10 20 30month��

�� Tenure (in months) lowess t month

Tenure Pattern over Study Period

38

II.2 Analysis of the Effect of Treatment on Outstanding Portfolio (MFIs are presented separately) Graph 5 MFI1 Graph 6 MFI2 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

010

0000

2000

0030

0000

0 10 20 30 0 10 20 30

��

0

��

1

�� Portfolio (Real Soles) lowess v month

Valu

e of

Por

tfolio

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

1

month

Graphs by Treatment dummy

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

050

0000

1000

000

1500

000

2000

000

0 10 20 30 0 10 20 30

��

0

��

1

�� Portfolio (Real Soles) lowess v month

Valu

e of

Por

tfolio

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

2

month


Graph 7 MFI3 Graph 8 MFI4 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

0000

4000

0060

0000

0 10 20 30 0 10 20 30

��0

��1

��

Portfolio (Real Soles) lowess v month

Valu

e of

Por

tfolio

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

3

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

050

0000

1000

000

0 10 20 30 0 10 20 30

��0

��1

��


Valu

e of

Por

tfolio

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

4

month


Graph 9 MFI5 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

0000

4000

0060

0000

8000

00

0 10 20 30 0 10 20 30

��0

��1

��


Valu

e of

Por

tfolio

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

5

month


39

II.3 Analysis of the Effect of Treatment on Delinquency Rate (MFIs are presented separately) Graph 10 MFI1 Graph 11 MFI2 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.2

.4.6

.8

0 10 20 30 0 10 20 30

��0

��1

��Delinquency rate lowess d month

Del

inqu

ency

Rat

e Be

fore

and

Afte

r Tre

atm

ent:

MFI

1

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.5

0 10 20 30 0 10 20 30

��0

��1

��Delinquency rate lowess d month

Del

inqu

ency

Rat

e Be

fore

and

Afte

r Tre

atm

ent:

MFI

2

month



��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.1

.2.3

0 10 20 30 0 10 20 30

��0

��1

��

Delinquency rate lowess d month

Del

inqu

ency

Rat

e Be

fore

and

Afte

r Tre

atm

ent:

MFI

3

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.0

5.1

.15

0 10 20 30 0 10 20 30

��0

��1

��


Del

inqu

ency

Rat

e Be

fore

and

Afte

r Tre

atm

ent:

MFI

4

month


Graph 14 MFI5 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

0.1

.2.3

0 10 20 30 0 10 20 30

��0

��1

��


Del

inqu

ency

Rat

e Be

fore

and

Afte

r Tre

atm

ent:

MFI

5

month


40

II.4 Analysis of the Effect of Treatment on Number of Clients (MFIs are presented separately) Graph 15 MFI1 Graph 16 MFI2 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

010

020

030

040

0

0 10 20 30 0 10 20 30

��0

��1

��

Number of Clients lowess n month

Num

ber o

f Clie

nts

Befo

re a

nd A

fter T

reat

men

t: M

FI1

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

040

060

080

0

0 10 20 30 0 10 20 30

��0

��1

��


Num

ber o

f Clie

nts

Befo

re a

nd A

fter T

reat

men

t: M

FI2

month



��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

5010

015

020

025

0

0 10 20 30 0 10 20 30

��0

��1

��


Num

ber o

f Clie

nts

Befo

re a

nd A

fter T

reat

men

t: M

FI3

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

010

020

030

0

0 10 20 30 0 10 20 30

��0

��1

��


Num

ber o

f Clie

nts

Befo

re a

nd A

fter T

reat

men

t: M

FI4

month


Graph 19 MFI5 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

010

020

030

0

0 10 20 30 0 10 20 30

��0

��1

��Number of Clients lowess n month

Num

ber o

f Clie

nts

Befo

re a

nd A

fter T

reat

men

t: M

FI5

month


41

II.5 Evolution of Incentive Compensation over Time (MFIs are presented separately) *Note that for the pre-treatment phase of MFIs 1, 3, and 5, the incentives were only paid quarterly- the other months are considered missing. Therefore, average incentives received over this period is actually one-third as high as the graph shows (if average in 0 payments). Graph 20 MFI1* Graph 21 MFI2 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

050

0

0 10 20 30 0 10 20 30

��0

��1

��

inctot lowess inctot month

Ince

ntiv

es B

efor

e an

d Af

ter T

reat

men

t: M

FI1

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

050

010

0015

00

0 10 20 30 0 10 20 30

��0

��1

��

inctot lowess inctot monthIn

cent

ives

Bef

ore

and

Afte

r Tre

atm

ent:

MFI

2month


Graph 22 MFI3* Graph 23 MFI4 ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

040

060

080

0

0 10 20 30 0 10 20 30

��0

��1

��


Ince

ntiv

es B

efor

e an

d Af

ter T

reat

men

t: M

FI3

month


��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

020

040

060

080

0

0 10 20 30 0 10 20 30

��0

��1

��


Ince

ntiv

es B

efor

e an

d Af

ter T

reat

men

t: M

FI4

month


Graph 24 MFI5* ��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��050

010

00

0 10 20 30 0 10 20 30

��0

��1

��inctot lowess inctot month

Ince

ntiv

es B

efor

e an

d Af

ter T

reat

men

t: M

FI5

month


42

ANNEX III: Regression Results

43

Regression Table 1: Fixed effects Regressions of Outstanding Portfolio and Monthly Change in Outstanding Portfolio (in Real Soles) Outstanding Portfolio Change in Outstanding Portfolio 1 2 3 4

Tenure (in months) 6118 (2.37)

3,683.3 (1.26)

-1,386.5 (-4.87)

-2.057.6 (-4.69)

Tenure squared 32.16 (0.67)

76.8 (1.33)

7.8 (1.20)

31.1 (2.28)

Feb -3,635.5 (-2.48)

-3,819.7 (-2.60)

5,051.0 (2.22)

4,988.1 (2.20)

Mar -6,453.1 (-3.28)

-6,791.1 (-3.44)

5,660.7 (2.55)

5,573.1 (2.51)

Apr -5,872.1 (-2.42)

-6,391.5 (-2.63)

6,195.40 (2.71)

6,277.1 (2.73)

May -3,840.7 (-1.48)

-4,445.3 (-1.71)

7,458.1 (3.26)

7,549.7 (3.28

Jun -1,663.3 (-0.62)

-2,304.9 (-0.85)

7,869 (3.41)

8058.7 (3.45)

Jul 5,222.3 (1.93)

4,564.6 (1.68)

13,606.6 (5.89)

13,779.2 (5.89)

Aug 5,372.3 (2.00)

4,763.4 (1.77)

6,614.2 (2.91)

6788.1 (2.94)

Sep 3,861.2 (1.50)

3,336.8 (1.29)

5,236.5 (2.27)

5,472.7 (2.34)

Oct 3,584 (1.49)

3,209.2 (1.33)

7,538.2 (3.37)

7,911 (3.50)

Nov 10,540.1 (4.81)

10,311.5 (4.70)

16,582.8 (7.35)

16,968.9 (7.45)

Dec 13,900.8 (7.30)

13,833.3 (7.26)

12,976.4 (5.80)

13,340.0 (5.93)

Treatment dummy 1,203.2 (0.22)

-20,048.4 (-2.19)

6513 (2.11)

-207.2 -(0.03)

Treatment * tenure 1,951.1 (2.22)

687.1 (1.31)

Treatment * tenure squared

-43.6 (-1.83)

-24.4 (-1.84)

Treatment * MFI 1 17,064.7 (2.61)

18,321.4 (2.73)

5,558.6 (1.57)

4,979.1 (1.40)

Treatment * MFI 2 6863.8 (1.02)

6,904.8 (1.02)

12,377.9 (3.56)

13,753.9 (3.90)

Treatment * MFI 3 13,937.1 (1.35)

11,640.6 (1.12)

12,871.4 (2.57)

12,115.3 (2.40)

Treatment * MFI 4 -5225.5 (-0.72)

-6127.3 (-0.84)

3973.8 (1.08)

3,580.2 (0.97)

Treatment * Eligible 6,264.0 (1.67)

5,881.4 (2.97)

Constant 293,913.7 (76.81)

325,174.9 (77.3)

15,429.1 (6.54)

18,556.0 (5.69)

# of observations 4401 4401 4089 4089 Number of credit officers 310 310 306 306

Adj. R2 within 0.0718 0.0755 0.0432 0.0463 between 0.4289 0.4171 0.1021 0.1287

overall 0.3520 0.3399 0.0431 0.0515 t-statistics are presented in parentheses

44

Regression Table 2: Fixed effects Regressions of Number of Clients and Monthly Change in Number of Clients Number of Clients Change in Number of Clients 1 2 3 4

Tenure (in months) 6.3 (4.79)

4.89 (3.35)

-1.47 (-10.44)

-1.93 (-9.01)

Tenure squared -.070 (-2.91)

-.045 (-1.58)

.013 (4.10)

.030 (4.56)

Feb .333 (0.48)

.227 (0.33)

1.11 (1.05)

1.07 (1.02)

Mar -.136 (-0.15)

-.331 (-0.36)

.158 (0.15)

.130 (0.12)

Apr -.483 (-0.42)

-.762 (-0.67)

-1.56 (-1.44)

-1.40 (-1.29)

May .306 (0.25)

-.034 (-0.03)

.336 (0.31)

.524 (0.48)

Jun 1.20 (0.95)

.833 (0.66)

.344 (0.31)

.612 (0.55)

Jul 3.53 (2.77)

3.14 (2.46)

2.67 (2.44)

2.94 (2.66)

Aug 2.45 (1.94)

2.10 (1.66)

-.852 (-0.79)

-566 (-0.52)

Sep 1.03 (0.85)

.729 (0.60)

-.775 (-0.71)

-.444 (-0.40)

Oct .717 (0.63)

.507 (0.45)

.934 (0.88)

1.31 (1.22)

Nov 3.58 (3.48)

3.46 (3.37)

5.03 (4.70)

5.38 (5.0)

Dec 4.13 (4.62)

4.11 (4.60)

2.99 (2.86)

3.32 (3.15)

Treatment dummy .659 (0.26)

-14.59 (-3.40)

6.72 (4.45)

5.06 (1.70)

Treatment * tenure 1.29 (3.13)

.285 (1.12)


-.028 (-2.48)

-.014 (-2.13)

Treatment * MFI 1 10.05 (3.27)

11.20 (3.56)

1.15 (0.67)

.7043 (0.41)


1.22 (0.38)

2.88 (1.69)

3.62 (2.10)


2.84 (0.58)

3.10 (1.26)

2.72 (1.11)

Treatment * MFI 4 -1.06 (-0.31)

-1.80 (-0.53)

1.75 (0.97)

1.54 (0.86)

Treatment * Eligible 5.35 (3.04)

2.81 (2.92)

Constant 144.54 (76.33)

162.64 (79.29)

21.45 (16.40)

23.39 (15.27)




45

Regression Table 3: Fixed effects Regressions of Delinquency Rate and Monthly Change in Delinquency Rate Delinquency Rate Change in Delinquency Rate 1 2 3 4

Tenure (in months) .0038 (3.33)

.0044 (3.22)

-.00051 (-1.85)

-.00082 (-1.95)


-.00010 (-3.08)

.00001 (1.94)

.00002 (1.53)

Feb -.00137 (-1.03)

-.00133 (-1.00)

.00072 (0.34)

.00066 (0.31)

Mar -.00049 (-0.28)

-.00041 (-0.23)

.00320 (1.52)

.00302 (1.44)

Apr .00224 (1.04)

.00231 (1.07)

.00445 (2.05)

.00415 (1.91)

May .00102 (0.43)

.00211 (0.91)

.00301 (1.39)

.00262 (1.20)

Jun .00102 (0.43)

.00103 (0.43)

.00263 (1.20)

.00223 (1.01)

Jul .00060 (0.25)

.00058 (0.24)

.00324 (1.48)

.00277 (1.25)

Aug .00047 (0.20)

.00041 (0.17)

.00254 (1.18)

.00204 (0.93)

Sep -.00203 (-0.89)

-.00211 (-0.93)

-.00096 (-0.44)

-.00144 (-0.65)

Oct .00084 (0.39)

.00074 (0.35)

.00456 (2.15)

.00434 (2.02)

Nov .00142 (0.73)

.00134 (0.68)

.00228 (1.07)

.00213 (0.99)

Dec -.00050 (-0.30)

-.00057 (-0.34)

-.00106 (-0.51)

-.00121 (-0.57)

Treatment dummy -.00071 (-0.15)

-.00147 (-0.18)

-.00066 (-0.22)

-.01255 (-2.13)

Treatment * tenure -.00040 (-0.52)

.00098 (1.93)


.00001 (0.72)

-.000022 (-1.75)

Treatment * MFI 1 -.00903 (-1.59)

-.00828 (-1.42)

-.00616 (-1.81)

-.00632 (-1.85)

Treatment * MFI 2 -.00546 (-0.94)

-.00468 (-0.80)

-.00304 (-0.91)

-.00223 (-0.66)

Treatment * MFI 3 -.00324 (-0.36)

-.00280 (-0.31)

.00023 (0.05)

-.00070 (-0.14)

Treatment * MFI 4 -.00093 (-0.15)

-.00077 (-0.12)

-.00207 (-0.58)

-.00247 (-0.69)

Treatment * Eligible .00299 (0.92)

.00494 (2.59)

Constant .0577 (24.47)

.0524 (19.44)

.00545 (2.07)

.00806 (2.63)




46

Regression Table 4: Fixed effects Regressions of Average Loan Size (in Real Soles) 1 2


3.90 (0.94)

Tenure squared .071 (0.99)

.181 (1.71)

Feb -7.25 (-1.37)

-7.47 (-1.41)

Mar -16.26 (-2.34)

-16.77 (-2.41)

Apr -2.51 (-0.30)

-2.88 (-0.34)

May 4.94 (0.56)

4.69 (0.53)

Jun 2.74 (0.30)

2.66 (0.29)

Jul 18.77 (2.03)

18.80 (2.02)

Aug 15.06 (1.63)

15.24 (1.65)

Sep 8.61 (0.98)

9.03 (1.03)

Oct 11.04 (1.33)

11.61 (1.39)

Nov 15.09 (1.97)

15.67 (2.04)

Dec 23.03 (3.44)

23.58 (3.52)

Treatment dummy -14.01 (-0.78)

-28.36 (-0.92)


Treatment * tenure squared -.092 (-1.16)


27.51 (1.25)


25.48 (1.15)

Treatment * MFI 3 110.50 (3.31)

108.40 (3.23)


5.08 (0.22)


Constant 1,612.77 (198.02)

1,633.00 (171.16)

# of observations 4398 4398 Number of credit officers 310 310

Adj. R2 within 0.0448 0.0463 Between 0.2151 0.2231

Overall 0.1626 0.1693 t-statistics are presented in parentheses

47

Regression Table 5: Fixed effects Regressions of Incentive Compensation (in Real Soles) 1 2


4.97 (1.89)


-.011 (-0.14)

Feb 2.40 (0.40)

2.00 (0.33)

Mar 25.43 (3.35)

24.90 (3.32)

Apr -2.82 (-0.32)

-5.60 (-0.64)

May 3.03 (0.32)

-.778 (-0.08)

Jun 27.42 (2.89)

23.12 (2.49)

Jul 13.92 (1.48)

9.30 (1.01)

Aug 15.24 (1.61)

10.08 (1.09)

Sep 48.91 (5.39)

43.16 (4.83)

Oct 6.02 (0.68)

1.08 (0.12)

Nov 13.77 (1.64)

9.85 (1.19)

Dec 48.03 (6.47)

44.93 (6.10)


-133.28 (-3.55)


Treatment * tenure squared .076 (1.02)


101.33 (3.47)

Treatment * MFI 2 245.99 (8.34)

274.50 (9.67)


75.01 (2.02)

Treatment * MFI 4 -10.64 (-0.35)

-16.12 (-0.55)


Constant 20.83 (2.76)

16.60 (1.88)

# of observations 4166 4166 Number of credit officers 304 304

Adj. R2 within 0.2266 0.2674 Between 0.2975 0.3802

Overall 0.3334 0.4003 t-statistics are presented in parentheses

48

Regression Table 6: Fixed Effects Regression Results on Loan Portfolio by MFI (in Real Soles) MFI1 MFI2 MFI3 MFI4 MFI5

Tenure (in months) 11,476 (6.31)

26,071 (2.75)

13,365 (0.84)

25,473 (2.48)

Dropped due to collinearity

Tenure squared -216.6 (-4.59)

-289.0 (-1.83)

56.2 (0.21)

-219.8 (-1.29)

Month Dummies included

YES YES YES YES

Treatment dummy -20,412 (-2.15)

6,972 (2.05)

47,803 (1.92)

-17,909 (-0.71)

Treatment * tenure 1451 (1.89)

-164 (-1.84)

-2143 (-0.60)

202 (0.08)


-24.76 (-0.98)

11,971.85 (0.93)

31.74 (0.29)

2.90 (0.05)

Treatment * Eligible 7763 (2.42)

8,022 (0.31)

-589 (-0.05)

6,194 (0.52)

Treatment*Branch1 11,340 (1.82)

15,786 (0.42)

1,597 (0.14)

8,585 (0.47)


13,189 (0.49)

-7,295 (-0.32)


-6030 (-0.23)

-632 (-0.03)


-5697 (-0.22)

655 (0.05)


63,839 (2.45)

-25,117 (-1.65)


41,006 (1.60)

-3,616 (-0.18)


17,753 (0.69)

-32,816 (-1.39)


20,151 (0.73)


16,826 (0.63)


Constant 8,418 (3.38)

428,732 (19.19)

-78,903 (-3.77)

37,929 (2.62)

# of observations 1,851 1,041 233 807 Number of credit

officers 122 82 14 53



49

Regression Table 7: Fixed Effects Regression Results on Growth in Loan Portfolio by MFI (in Real Soles) MFI1 MFI2 MFI3 MFI4 MFI5

Tenure (in months) -976.4 (-2.95)

-13275 (-3.85)

-3065.5 (-1.77)

-30.1 (-0.02)


Tenure squared 7.03 (0.56)

188.9 (2.95)

50.9 (0.95)

6.5 (0.20)


YES YES YES YES

Treatment dummy 8627.2 (1.34)

6978.7 (3.35)

17912 (1.22)

-10799.5 (-0.55)

Treatment * tenure -615.8 (-1.17)

-195.8 (-3.58)

1070 (0.66)

172.2 (0.12)


12.8 (0.84)

21860.2 (2.84)

-37.5 (-0.74)

-8 (-0.23)


14426.0 (0.93)

3889.8 (0.79)

14076.8 (2.25)

Treatment * gender -1464.5 (-0.73)

Treatment*Branch1 4230.1 (1.41

30582.7 (1.27)

-7358.7 (-1.51)

-14572 (-1.54)

Treatment*Branch2 -3016.8 (-0.76))

28496.9 (1.81)

1962.2 (0.17)

Treatment*Branch3 6115.8 (2.09)

-11002.0 (-0.73)

5268.8 (0.45)


22951.4 (1.53)

4739.5 (0.66)


22839.9 (1.54)

-11864.9 (-1.48)

Treatment*Branch6 7735 (1.96)

22368.8 (1.49)

10423 (1.07)

Treatment*Branch7 -4239.7 (-1.19)

22523.6 (1.50)

-33958.5 (-3.07)


41472.2 (2.48)

Treatment*Branch9 4721.3 (1.24

29344.2 (1.90)


Constant 15,169.4 (7.74)

135,953.5 (3.16)

40,122.2 (3.00)

-5075.1 (-0.40)

# of observations 1727 959 219 754 Number of credit

officers 122 80 13 53



50

Regression Table 8: Fixed Effects Regression Results on Number of Clients by MFI MFI1 MFI2 MFI3 MFI4 MFI5


7.1 (2.05)

5.8 (2.14)

9.3 (2.68)



-.085 (-1.47)

-.06 (-1.06)

-.11 (-2.09)


YES YES YES YES


1.7 (1.41)

15.1 (2.17)

.50 (0.08)


-.04 (-1.12)

-1.38 (-1.38)

-.50 (-0.77)


-.05 (-1.83)

3.6 (0.78)

.04 (1.18)

.02 (0.98)


4.0 (0.43)

.75 (0.23)

-.67 (-0.22)

Treatment * gender 6.8 (1.56)


3.8 (0.28)

5.6 (1.72)

9.2 (1.93)


.13 (0.01)

1.03 (0.17)


-1.3 (-0.13)

2.2 (0.36)


2.9 (0.31)

-.37 (-0.10)


4.6 (0.49)

-1.4 (-0.36)


2.2 (0.24)

-.66 (-0.13)


7.3 (0.78)

4.5 (0.73)


6.2 (0.62)


5 (0.52)


Constant 100.4 (31.85)

190.8 (23.78)

84.2 (20.01)

23.9 (5.28)


officers 122 82 14 53



51

Regression Table 9: Fixed Effects Regression Results on Change in Number of Clients by MFI MFI1 MFI2 MFI3 MFI4 MFI5

Tenure (in months) -2.2 (-5.80)

-4.6 (-3.71)

-1.1 (-2.33)

.13 (0.45)



.06 (2.56)

.01 (0.88)

-.01 (-0.98)


YES YES YES YES


2.2 (2.94)

9.4 (2.32)

.25 (0.05)

Treatment * tenure .30 (0.48)

-.06 (-3.01)

.07 (0.16)

-.18 (-0.51)


-.01 (-0.79)

9.0 (3.27)

-.01 (-0.44)

.01 (0.79)


4.3 (0.78)

.20 (0.15)

3.1 (2.04)

Treatment * gender .89 (0.38)


10.3 (1.20)

-.004 (-0.0)

-5.2 (-2.25)


6.4 (1.14)

-.33 (-0.12)


-5.8 (-1.08)

1.2 (0.43)

Treatment*Branch4 .35 (0.08)

7.3 (1.36)

.63 (0.36)


.80 (0.15)

-5.3 (-2.74)


.50 (0.09)

1.9 (0.82)


9.6 (1.79)

-10.8 (-4.06)


16.7 (2.80)


10.9 (1.97)


()

Constant 27.6 (12.97)

48.35 (3.16)

13.48 (3.73)

3.25 (0.99)


officers 122 80 13 53



52

Regression Table 10: Fixed Effects Regression Results on Delinquency Rate by MFI MFI1 MFI2 MFI3 MFI4 MFI5


-.00 (-0.32)

.00 (0.31)

.002 (2.51)



-.00 (-0.75)

-.00 (-1.08)

-.00 (-2.92)


YES YES

YES YES


.00 (0.44)

.01 (0.98)

.002 (0.42)

Treatment * tenure -.00 (-0.36)

-.00 (-0.26)

-.003 (-1.60)

-.000 (-0.09)


.00 (0.32)

.003 (0.97)

.00 (1.48)

00 (0.21)


.001 (0.26)

.01 (0.98)

-.001 (-0.57)

Treatment * gender -.01 (-0.89)


.01 (0.75)

.002 (0.44)

.004 (1.15)


-.00 (-0.44)

.006 (1.50)


-.01 (-1.35)

-.00 (-0.96)


-.00 (-0.05)

-.00 (-0.24)


-.00 (-0.60)

.003 (1.05)


-.00 (-0.12)

.004 (1.11)


-.00 (-0.02)

.006 (1.48)


.00 (0.08)

Treatment*Branch9 -.00 (-0.01)

.00 (0.00)


Constant .02 (3.70)

.17 (25.21)

.10 (17.27)

.06 (31.56)


officers 122 82 14 53



53

Regression Table 11: Fixed Effects Regression Results on Change in Delinquency Rate by MFI MFI1 MFI2 MFI3 MFI4 MFI5

Tenure (in months) -.001 (-1.79)

-.003 (-2.63)

.002 (3.08)

-.000 (-0.25)



.00 (1.79)

-.00 (-2.53)

-.000 (-0.29)


YES YES

YES YES


.001 (2.05)

.002 (0.27)

.006 (1.70)

Treatment * tenure .003 (2.38)

-.00 (-1.76)

-.001 (-2.30)

-.000 (-1.90)


-.00 (-2.26)

.004 (1.51)

.000 (2.40)

.000 (1.57)


.002 (0.47)

.003 (1.52)

.000 (0.45)

Treatment * gender -.01 (-1.44)


.002 (0.32)

.002 (1.14)

.000 (0.02)


.001 (0.15)

.002 (0.99)


-.012 (-2.45)

.001 (0.39)


.001 (0.30)

-.002 (-1.29)


-.001 (-0.29)

-.001 (-0.50)


.002 (0.42)

-.001 (-0.34)


.001 (0.24)

-.002 (-1.25)


.001 (0.20)


.001 (0.17)


Constant .01 (2.21)

.031 (2.44)

-.02 (-3.31)

.006 (2.46)


officers 122 80 13 53



54

Regression Table 12: Fixed Effects Regression Results on Average Loan Size by MFI (in Real Soles) MFI1 MFI2 MFI3 MFI4 MFI5


30.0 (1.78)

4.3 (0.10)

21.2 (1.62)



-.30 (-1.05)

.66 (0.69)

.03 (0.08)


YES YES

YES YES


8.05 (1.15)

139.7 (0.96)

-221.7 (-1.85)

Treatment * tenure -6.1 (-1.83)

-.25 (-1.35)

-2.0 (-0.10)

12.4 (1.11)


.18 (1.71)

14.7 (0.56)

-.05 (-0.08)

-.30 (-1.07)


-33 (-0.62)

-25.6 (-0.38)

71.7 (1.38)



28.9 (0.37)

-24.3 (-0.36)

-34.3 (-0.43)


122.7 (2.24)

-87.8 (-0.89)


-11.2 (-0.21)

-67.6 (-0.67)


-29.6 (-0.56)

78.3 (1.24)


162.4 (3.05)

-151.5 (-2.26)


146.4 (2.79)

-45.1 (-0.53)


.05 (0.00)

-269.3 (-2.71)


15.3 (0.27)


23.5 (0.43)


Constant 447.6 (59.86)

2334.2 (49.23)

1204.6 (15.80)

2278.5 (55.0)


officers 122 82 14 53



55

Regression Table 13: Fixed Effects Regression Results on Incentive Compensation by MFI (in Real Soles) MFI1 MFI2 MFI3 MFI4 MFI5


-18.1 (-1.33)

-34.2 (-0.89)

-3.4 (-1.32)



.28 (1.22)

1.6 (1.54)

.10 (1.54)


YES YES

YES YES


28.7 (4.92)

-155.4 (-0.78)

36.4 (1.16)


-.45 (-2.94)

28.7 (1.02)

-4.4 (-1.62)


-.16 (-1.32)

200.1 (9.12)

-.99 (-1.14)

.07 (1.05)


-231.1 (-5.21)

78.1 (0.86)

24.7 (2.02)



-266.3 (-4.13)

-19.3 (-0.22)

53.1 (2.79)


-138.6 (-3.04)

6.1 (0.26)

Treatment*Branch3 37 (1.45)

-58.3 (-1.29)

17.3 (0.74)


-25 (-0.56)

6.2 (0.42)


87.8 (1.98)

41.1 (2.59)


-58.5 (-1.34)

68 (3.48)


13.8 (0.31)

58.1 (2.57)


-132.3 (-2.80)


113 (2.49)


Constant -31.1 (-2.64)

401.6 (10.03)

-20.2 (-0.21)

57.3 (4.90)


officers 122 82 13 53



56

Regression Table 14: Timing of Treatment Effects: Coefficients and standard deviations of treatment dummy by different definitions of treatment periods (Pooled sample and no treatment interaction dummy variables are included) Portfolio Change

in Portfolio

Number of clients

Change in # of clients

Avg. Loan Size

Del. Rate

Change in Del. Rate

Treatment Periods

Month mT 15,687 (7.63)

10,897 (6.92)

6.199 (6.63)

4.786 (5.53)

28.521 (3.75)

-.00292 (-1.35)

-.00666 (-2.78)

Months mT , mT+1 16,026 (8.01)

8,446 (7.32)

6.018 (6.53)

2.972 (4.50)

32.051 (4.49)

-.00329 (-1.57)

-.00692 (-4.05)

Months mT , mT+1, mT+2

15,52 (7.82)

7,348 (7.64)

5.623 (6.12)

2.129 (3.68)

35.803 (5.21)

-.00310 (-1.53)

-.00550 (-3.89)

Months mT , mT+1, mT+2 , mT+3

16,147 (7.33)

6,285 (7.20)

5.268 (5.39)

1.643 (3.07)

35.657 (5.33)

-.00395 (-2.02)

-.00509 (-4.02)

Months mT , mT+1, mT+2 , mT+3, mT+4

14,934 (6.92)

5,869 (7.21)

5.081 (5.27)

1.435 (2.83)

36.687 (5.61)

-.00433 (-2.28)

-.00539 (-4.64)

Non-treatment periods

Month mT-1 2,127 (1.01)

-4,187 (-2.66)

-1.674 (-1.78)

-3.898 (-4.49)

18.472 (2.39)

-.00030 (-0.13)

-.00529 (-2.12)

Months mT-1, mT-2 788.9 (0.37)

-4,897 (-3.99)

-2.100 (-2.22)

-4.375 (-6.38)

32.350 (4.34)

.00418 (1.86)

-.00277 (-1.45)

Months mT-1, mT-2, mT-3

2,227 (1.00)

-4,994 (-4.56)

-.659 (-0.66)

-4.333 (-7.14)

32.290 (4.26)

.0101 (4.32)

.00056 (0.33)

t-statistics are presented in parentheses

Preliminary Draft: Please do not cite

Documents

Transcript of Preliminary Draft: Please do not cite