Decentralized Targeting of Agricultural Credit

Programs: Private versus Political Intermediaries ∗

Pushkar Maitra† Sandip Mitra‡ Dilip Mookherjee§ Sujata Visaria ¶

February 14, 2019

Abstract

We conducted a field experiment in rural West Bengal, India to compare alternative ap-proaches to delegating beneficiary selection for an agricultural credit program. In bothapproaches, a local agent recommends borrowers for individual liability loans, and is in-centivized by commissions that depend on repayments. The agent is a randomly chosentrader in TRAIL and is appointed by the local government in GRAIL. TRAIL loans ex-hibited greater take-up, similar default rates and larger impacts on farm incomes. Whilethe TRAIL agent recommended more productive farmers, selection differences accountedfor a small fraction of the observed difference in average treatment effects. Instead, higherconditional treatment effects accounted for the bulk of the difference. To explain this, wedevelop and test a model where differences in agents’ incentives caused them to help andmonitor borrowers differently.

Key Words: Targeting, Intermediation, Decentralization, Community Driven Develop-ment, Agricultural credit, Networks.

JEL Codes: H42, I38, O13, O16, O17

∗Funding was provided by the Australian Agency for International Development, the International Growth Centre,United States Agency for International Development, the Hong Kong Research Grants Council (GRF Grant Number16503014) and the HKUST Institute for Emerging Market Studies (Grant Number IEMS15BM05). We thank the directorand staff of Shree Sanchari for collaborating on the project. Jingyan Gao, Arpita Khanna, Clarence Lee, Daijing Lv, FoezMojumder, Moumita Poddar and Nina Yeung provided exceptional research assistance, and Elizabeth Kwok providedexcellent administrative support. We thank audiences at a large number of seminars and conferences for their commentsand suggestions, which have significantly improved the paper. We received Internal Review Board clearance from MonashUniversity, Boston University and the Hong Kong University of Science and Technology. The authors are responsible forall errors.†Pushkar Maitra, Department of Economics, Monash University, Clayton Campus, VIC 3800, Australia.

Pushkar.Maitra@monash.edu.‡Sandip Mitra, Sampling and Official Statistics Unit, Indian Statistical Institute, 203 B.T. Road, Kolkata 700108,

India. Sandip@isical.ac.in.§Dilip Mookherjee, Department of Economics, Boston University, 270 Bay State Road, Boston, MA 02215, USA.

dilipm@bu.edu.¶Sujata Visaria, Department of Economics, Lee Shau Kee Business Building, Hong Kong University of Science and

Technology, Clear Water Bay, Hong Kong. svisaria@ust.hk.

1 Introduction

In developing countries, public programs are increasingly being implemented at the local level.

For example, communities are monitoring teachers and health professionals, and budgets for

local public goods are being devolved to local governments. Such programs aim to leverage the

local community’s superior information about potential beneficiaries, and its greater incentives

to target accurately. In practice, however, both communities and local government are subject

to elite capture (see, for example World Development Report 2004; Mansuri and Rao 2013).

There is also growing evidence that local governments looking to get re-elected target their

vote bank, rather than those who stand to benefit the most (Stokes 2005; Robinson and Verdier

2013; Bardhan et al. 2015; Bardhan and Mookherjee 2016; Devarajan and Khemani 2016; Dey

and Sen 2016).

Different approaches to community-mediated development are likely to vary in their effective-

ness. For instance, responsibility for selecting beneficiaries could be delegated to important

members of the local communities. These communities are typically characterized by differ-

ent (though possibly overlapping) networks of economic and social or political interactions;

those occupying central roles in economic networks may be different from those more central

in social or political networks. This paper aims to evaluate different methods of community

involvement in the targeting of beneficiaries. Specifically, we examine the choice faced by an

external lender between appointing nodes of two different networks as its commission agents

in the rollout of a subsidized agricultural credit program intended for poor farmers. In one

treatment called TRAIL, a local private trader is appointed the agent, randomly from a set

of important private traders with past experience in trading and lending within the village.

In the other treatment called GRAIL, the elected local government was asked to appoint an

agent, ostensibly a political operative of the incumbent political party. The experiment was

carried out in rural West Bengal, at a time when local government elections were character-

ized by intense competition between two rival parties, the Left Front (LF) and the Trinamul

Congress (TMC), where both parties had extensive local cadres that they used to mobilize

voters. Hence it is likely that the GRAIL agent would be someone in (or close to) the party

hierarchy, responsible in mobilizing local farmers for political purposes and delivering them

benefits on behalf of the local government.

Apart from choice of the agent, the interventions were similar in all other respects. The formal

role of the agent was to recommend 30 farmers within the village owning less than 1.5 acres

of cultvable land for the subsidized crop loans. 10 out of the 30 recommended were chosen

randomly to receive offers of the loan. Agents were incentivized through commissions set at

75% of interest repaid by chosen borrowers, besides forfeiting deposits posted on their behalf

1

in the event of default. The loans were based on individual liability, involved the same interest

rate (with an APR 7 percent below the average informal credit rate), timing and duration

(4 months, offered at planting time of important crops). The loans started with small credit

limits (Rs 2000, equivalent to $30), but rising fast (133% of the amount repaid in the previous

4 month cycle), in order to create dynamic repayment incentives. The experiment involved 48

villages in two potato growing districts of West Bengal, started in October 2010 and continued

for three years, ending in August 2013. Loans were designed to be borrower-friendly, with

transactions in cash at the doorstep of each household, and built in insurance against covariate

price and yield risk in potato, the key cash crop in these districts. We collaborated with a

MFI based in Kolkata, the state capital, in the rollout of the credit program, while our survey

teams conducted detailed farm surveys for a sample of 50 farming households in each village

three times a year.

Owing to differences in their own occupations and nature of local interactions, the TRAIL and

GRAIL agent are likely to have different kinds of information about productivity and reliability

of individual borrowers. The TRAIL agent is a private trader with extensive experience lending

within the community, buying crops and selling farm inputs, so is likely to know track records

of loan repayments, crop production and sales of individual farmers. The GRAIL agent is

less likely to have this kind of detailed business information. Hence their respective ability to

identify relevant characteristics of potential recipients could be quite different.

Actual selections they make would additionally depend on their respective incentives. Private

traders typically act as middlemen in crop supply chains, buying crops from farmers in order

to resell them to wholesale buyers for a profit. Middlemen margins in potato tend to be

quite large in these areas: in 2008 Mitra et al. (2018) estimate farmgate prices were 45% of

wholesale prices, with middlemen earning at least 50-70% of this gap. So the TRAIL agent

would be motivated to direct credit towards farmers producing and selling more potatoes to

them. These trade-related motivations would be unlikely to be relevant in GRAIL, where the

agent would likely select beneficiaries in accordance with priorities of the incumbent political

party. Given the redistributive ideology and intense political competition between the two

competing parties, it is plausible that either party would seek to target poorer borrowers,

rather than the most productive ones. They could be subject to clientelistic vote mobilization

objectives, and thus direct loans towards swing voters. They would also be expected to ensure

low default rates, given the repayment based commissions as well as the need to avoid greater

distress among poor farmers.

Besides their formal role in selection, the agent could conceivably play an important role by

interacting more intensively with treated borrowers, monitoring them or providing assistance.

Monitoring inputs, production methods, labor applied and loan repayments could affect loan

2

default risks. Farmers could face a trade-off between risk and mean returns in input choices,

cropping or marketing patterns: e.g., applying more pesticides lower risks of crop failure while

raising cultivation costs, high yielding crop varieties are often riskier, storing and selling the

crop later frequently generates gains from arbitrage but is vulnerable to crop price volatility. If

the primary objective of the agent is to lower default risk, which seems plausible for the GRAIL

agent, they would be motivated to monitor and restrict risky choices made by borrowers. The

TRAIL agent on the other hand shares both upside and downside risk with the farmer, as

higher outputs translate into higher potato sales to the agent. Hence the TRAIL agent may

be reluctant to restrict risky choices if they lower mean outputs at the same time. The TRAIL

agent may instead be motivated to provide assistance in the form of advice on crop varieties,

input quality, or timing of purchases and crop sales, which help farmers reduce unit costs and

raise realized crop prices. In GRAIL this would not apply owing to lack of business expertise

and involvement of the GRAIL agent. Hence the nature of subsequent ‘engagement’ of the

agent with treated borrowers could vary between the two interventions: taking the form of

greater monitoring in GRAIL and greater help in TRAIL, which have different effects on costs

and profits of treated farmers.

Besides assessing loan take-up and repayment rates, we estimate the average treatment effects

(ATEs) of the loans on borrowers’ farm production and farm incomes.1 We find that TRAIL

loans had higher take-up rates, and were repaid at the same high rate (93%) as GRAIL loans.

Both interventions raised the scale of farmer borrowing and potato cultivation, with a similar

increase in potato output of approximately 25%. However, TRAIL treated farmers achieved

lower production costs and sold their output at higher prices, and achieved a 40% increase in

profit from potatoes, as compared with an insignificant 4% increase in GRAIL. Farm income

aggregating across all principal cash crops rose 21% in TRAIL, compared to 1.4% in GRAIL,

the difference being significant at the 10% level.

The rest of the paper seeks to explore the underlying reasons for these differences. We start by

estimating respective roles of selection and incentive effects in generating the observed ATE

differences. Only 10 out of the 30 recommended borrowers are chosen via lottery to receive

the program loan offers, thereby creating two control groups: Control 1, those recommended

but not treated, and Control 2, those not recommended. This allows us to estimate loan

treatment effects conditional on selection (comparing treated and Control 1 subjects), and

selection differences (comparing Control 1 and Control 2, neither of whom were treated). Based

on a Cobb-Douglas specification of the production function, we develop a simple procedure for

1Our previous research (Maitra et al. 2017) compares TRAIL with the group-based lending (GBL) approachwhere borrowers self-selected into joint liability groups. The GBL intervention was implemented in 24 other(randomly chosen) villages in the same two districts. In the current paper we focus on comparisons of theTRAIL and GRAIL schemes.

3

estimating productivity of control subjects from farmer fixed effects in a panel regression of area

cultivated or output of potato, after controlling for village and year dummies. This enables us

to assess and compare selection patterns between TRAIL and GRAIL. We find TRAIL agents

selected more productive farmers, with the distribution of productivity among recommended

farmers in TRAIL dominating (in the first-order stochastic sense) that in GRAIL.

However, a decomposition of the ATE differences shows that a small fraction (less than 10%)

was accounted for by selection differences. Approximately 75% owed to differences in condi-

tional (on productivity) treatment effects (CTEs). In other words, borrowers of the same type

and offered a loan with exactly the same features, achieved very different outcomes across the

two interventions. While both TRAIL and GRAIL treated borrowers raised potato area cul-

tivated and output by similar extents, TRAIL borrowers realized higher crop prices and lower

unit costs, resulting in substantially larger profit impacts. Since the only difference between

a TRAIL and GRAIL borrower is the nature of the agent involved, this indicates that the

differential engagement of the two types of agents with the treated farmers drove the superior

performance of the TRAIL scheme.

In the final stage of our analysis, we develop a theoretical model of differences in the nature

of engagement between the agent and borrowers between the two schemes. The model is

based on the hypothesis that the respective agents were motivated differently to alter the

nature and scale of their engagement with treated farmers, with consequent effects on risk,

productivity and unit costs. The model generates detailed predictions concerning variation

across productivity types of default rates, CTEs on production, incomes, reported interactions

with the agent, and on political support expressed by farmers for rival political parties in an

opinion poll we conducted. In particular we expect higher CTEs in GRAIL on loan repayment

rates and on political support for the incumbent, particularly among less productive and poorer

farmers. On the other hand, CTEs on potato production, unit cost reductions and profits

would be higher in TRAIL. We verify that these predictions match the patterns in the data.

In GRAIL, treatment effects on repayment rates and interactions with the agent were larger

among less productive borrowers; in TRAIL the CTEs on interactions with borrowers and on

production and incomes were larger for the more productive recipients.

Our paper contributes to the literature on the targeting of development interventions, bene-

ficiary selection for public goods, and the targeting of creditworthy entrepreneurs in the ab-

sence of well-functioning financial markets and information asymmetries. Bandiera and Rasul

(2006); Alatas et al. (2012, 2016); Fisman et al. (2017); Hussam et al. (2018) all argue that

using community information effectively can reduce the information constraints that hamper

the selection of beneficiaries and hence development outcomes. Berg et al. (2018) and Debnath

and Jain (2018) find that the social network plays an important role in spread of information on

4

take-up of public health services.2 Relatedly, Banerjee et al. (2013) and Chandrasekhar et al.

(2018) show that identifying and effectively using central individuals in a network can lead to

wider diffusion of information and knowledge, and hence improve the take-up of development

interventions. Our paper complements this work by comparing the effectiveness of appointing

central nodes in two different local networks: economic, and political.

However, it is not clear what constitutes the relevant “network”, when there are many different

networks that coexist within any given community. Not much is known either about the

precise role of program intermediaries, whether this takes the form of selection or subsequent

engagement with recipients, and differences in their incentives to conduct either role. To our

knowledge, our paper is the first to focus on and model engagement between the beneficiary and

program intermediary, and quantify its role vis-a-vis selection in explaining observed outcomes.

This paper is organized as follows. In the next section we describe the two loan intervention

schemes that we will analyse in this paper. In Section 3 we describe our data. In Section

4 we present the estimates of the average treatment effects of the two schemes on borrower

outcomes; in Section 5 we present statistics on the financial performance of the schemes: take-

up and repayment rates. Next, we attempt to explain our empirical findings. We start in

Section 6 by estimating the farm productivity of borrowers, examining how the distribution

of borrower productivity differed across the two schemes, and then decomposing the difference

in average treatment effects into the part that can be explained by these selection differences,

and the part that is due to differences in treatment effects, conditional on productivity. Since

we find that this latter part is substantial, in Section 7 we develop the aforementioned model

of agent incentives and engagement. The predictions of this model are validated in Section 8.

Section 9 concludes.

2 Details of the Interventions

The TRAIL and GRAIL interventions were conducted across 48 randomly-chosen villages in the

Hugli and West Medinipur districts of West Bengal in India, over the three-year period 2010–

2013.3 The scheme, which provides individual-liability agricultural loans, began in October-

November 2010. The loan disbursals were timed to coincide with the planting of potatoes,

the major cash crop in this area.4 Similar to many microcredit programs, the scheme had a

2Beaman and Magruder (2012) and Beaman et al. (2018) also present evidence of the importance of thesocial network in alleviating informational constraints in different contexts.

3This was the same random sample of villages where Mitra et al. (2018) conducted their potato price infor-mation intervention experiment in 2007–08.

4These districts belong to a major potato-growing region, producing approximately 30% of all potatoescultivated in India. More on this below.

5

“progressive lending” feature designed to generate dynamic repayment incentives. The starting

loan was Rs. 2000, repayable as a single lumpsum with 6 percent interest at the end of four

months. If the amount due was fully repaid at the end of any cycle, the borrower was offered

a new four-month loan 33% larger than the previous one. Borrowers who repaid less than

50% of the repayment obligation in any cycle were not allowed to borrow again. Those who

paid less than 100% but more than 50% of the repayment due were eligible for a loan 33%

larger than the amount they had repaid. In order to ensure that borrowers would not need

to sell the potato crop prematurely to repay the loan, we allowed repayments in the form of

potato bonds rather than cash. In this case the repaid amount was calculated at the prevailing

price of potato bonds. Although borrowers were told the loans were for agricultural purposes,

households were not required to provide any evidence of how they actually used the loan.

Each of the 48 villages in the sample was randomly assigned to one intervention: either TRAIL

or GRAIL. Table 1 shows that there are no significant differences in village size, number of

potato cultivators in the village or the number of potato cultivators in the different landholding

categories across the two treatment groups. Our partner MFI had not operated in any of these

villages and in general, there was very little MFI penetration in this area at the time our

intervention began. The role of the MFI was limited to carrying out loan transactions on our

behalf with treated farmers; they did not provide the loan capital nor were they involved in

selecting borrowers, or monitoring them subsequently.5

In TRAIL villages, a trader based in the local community was selected to be the agent.6 In

GRAIL villages, a member of the Gram Panchayat (elected village council) selected the agent.7

In each intervention arm, the agent was required to recommend 30 potential borrowers, all of

whom had to be resident in the village and own less than 1.5 acres of land. We imposed

this landholding restriction in order to ensure that the loans were given to the poorer farmers

in the village, and to help avoid cronyism and bribery by the well-to-do farmers. Ten of

these 30 recommended were randomly chosen at a lottery we conducted in the office of the

local government, while keeping the names of recommended borrowers confidential (in order

to avoid any spillover effects on informal credit access or other relationships within the village

of control 1 subjects). Those selected in the lottery were then approached and and offered the

individual-liability loans.

Agents had both monetary and non-monetary incentives to participate in the scheme. At the

5A research grant provided the loan funds for the interventions.6We created a list of local traders who had at least 50 clients in the village, or had been operating in the

village for at least 3 years, and then offered the contract to a randomly chosen trader from this list. If he refusedto participate, we would have offered the position to a second randomly chosen trader. In practice the firsttrader whom we approached always accepted the contract.

7We required that the recommended individual had lived in the village for at least 3 years, was personallyfamiliar with farmers in the village; and had a good local reputation.

6

end of each loan cycle, they stood to earn a commission equal to 75% of all interest earned from

borrowers whom they had recommended. This high commission rate was meant to incentivize

the agent to select productive borrowers who would repay the loan and benefit from it, and to

discourage collusion between the agent and potential applicants. Further details are provided

in our previous paper (Maitra et al. 2017). There was also a system of deposits and bonuses.

In particular, the agent was required to provide a deposit of Rs 50 per borrower. At the end

of any cycle, if more than one-half of the borrowers recommended by any agent had defaulted

on the loan, the agent was terminated from the scheme and could earn no further commission.

At the end of two years, the deposit was returned to all agents who had not been terminated,

and the agent was offered a paid holiday to a seaside resort. In conversations during our field

visits, some TRAIL agents also remarked that they expected to become more prominent in

the village, or boost their own business. GRAIL agents may also have viewed the scheme as

an extension of government anti-poverty programs, or as a means to increase the dominance

of their political party.

3 Data and Selected Descriptive Statistics

In each village we collected data from a sample of 50 households. This sample included the

households of all 10 borrowers that were randomly chosen to receive the loan. We call these

the Treatment households. Of those that had been recommended but did not receive the loan,

we included in our sample a random subset of 10 households; these are called the Control

1 households. We also included 30 households randomly chosen from the pool of the non-

recommended. We refer to them as Control 2 households. Note that only households that

owned less than 1.5 acres of land could be recommended. Thus all our Treatment and Control

1 households own at most 1.5 acres of land.8 In Table 4 we present summary statistics of

selected household-level characteristics. The F-statistic presented in the bottom row indicates

that these household characteristics do not jointly explain assignment to treatment.

As stated above, all Treatment households received their first loan in October 2010. The first

round of surveys was conducted in December 2010; surveys were repeated every four months.

The high frequency of data collection helped minimize measurement error. In each sample

household, the same person answered the survey in each round and there was no attrition in

the sample over the eight survey cycles.

The districts where we conducted our intervention are two of the largest producers of potatoes

8However some Control 2 households whom we surveyed owned more than 1.5 acres of land. In our empiricalanalysis we only include households that owned at most 1.5 acres of land.

7

in West Bengal and India. Potatoes accounted for a significant proportion of acreage in our

sample villages. Potatoes had the largest working capital needs and generated nearly three

times as much value-added per acre as the other major crops (see Maitra et al. 2017, Table 2).

In Table 2 we present summary statistics on the characteristics of the TRAIL and GRAIL

agents. Over 95% of the TRAIL agents owned a shop or business, compared to about 20%

of GRAIL agents. Instead GRAIL agents were more likely to be cultivators, or to have a

government job. They were also more likely to be involved in civil society or politics: 30

percent said they were members of a village organization, 17 percent were political party

workers, and 13 percent had been members of the local government. None of the TRAIL

agents were directly involved in politics in this way. TRAIL agents also owned more land and

had higher incomes, but were less likely to have been educated above primary school.

Table 3 presents descriptive statistics on the extent of pre-intervention social and economic

interactions with the agent, separately for sample households that were recommended by the

agent (Treatment and Control 1 group) and those that were not recommended by the agent

(Control 2 group). First, there are considerable differences in economic interactions between

households and the type of the agent. Consistent with our hypothesis that the TRAIL agent

had more economic links with the villages, we see that the farmers in TRAIL villages are

more likely to report that the agent is one of the two main money lenders, input buyers and

output buyers than farmers in GRAIL villages. While farmers in both TRAIL and GRAIL

villages report knowing the agent and meeting him at least once every week, farmers in TRAIL

villages are much more likely to report having bought from the agent, borrowed from the agent

or worked for the agent. However, at a social level they seem to have been connected in similar

ways to either kind of agent: more than 90% reported either knowing the agent or meeting

them at least once a week. In both sets of villages, farmers reported similar likelihoods of

receiving benefits from the local government, with over 60% receiving at least some benefit

over the past three years.

Through our surveys we collected information on all the outstanding loans that our sample

households held at the time of the survey, or had taken since the previous survey cycle. In

survey cycle 1 we specifically asked about outstanding loans taken before September 2010.

Table 5 summarizes salient characteristics of these outstanding loan data to describe the credit

market in these villages before our intervention began. 67% of the households reported that

they had an outstanding loan. As the second column suggests, the majority of these loans

had been taken for agricultural purposes. Nearly two-thirds of these loans had been taken

from traders or moneylenders; the mean duration was 4 months. Most likely this is because

agricultural cycles in this region are four months long. Credit cooperatives accounted for nearly

one quarter of the loans. Formal sources like government banks accounted for a very small

8

percentage. MFI loans were also rare. The average annual interest rate on loans from traders

and moneylenders was 25 percent. Cooperatives and government banks charged substantially

lower interest rates, but as the last panel shows, 75-80 percent of these loans were collateralized.

This may also explain why our sample households, who owned very little land, were unlikely to

hold such loans, and why traders and moneylenders were the dominant source of agricultural

credit.

Every village in India is governed by a village council (or Gram Panchayat, hence forth GP).

Members are elected through direct elections once every five years. In the West Bengal village

council elections, candidates tend to be affiliated with state-level political parties. The Left

Front led by the Communist Party of India (Marxist) held power in both the West Bengal

state legislature and village councils continuously from 1977 until 2011. In the state legislative

assembly elections in 2011, the Left Front lost its majority and the Trinamool Congress formed

the state government. Similarly, after the 2013 village council elections, most village councils

now have a Trinamool Congress majority. Village councils are not authorised to raise revenues,

but receive devolved funds from upper-level governments through various national or state-level

schemes, that they are charged with implementing. They are often involved in identifying the

beneficiaries for these programs.

4 Average Treatment Effects of the Schemes

We start our empirical analysis by examining the impacts of the two loan schemes on borrower

outcomes. We are able to identify these average treatment effects because only a random

subset of the recommended household were offered the loans. The difference in the outcomes

of households that were recommended as well as offered the loan (Treatment households) and

those that were recommended but were not offered the loan (Control 1) is an estimate of the

treatment effect of the loan, conditional on the household being selected into the scheme. Our

regression specification therefore takes the following form:

yhvt = β0 + β1TRAILv + β2(TRAILv × Control 1hv) + β3(TRAILv × Treatmenthv)

+ β4GRAILv + β5(GRAILv × Control 1hv) + β6(GRAILv × Treatmenthv) (1)

+ γXhvt + I(Yeart) + εhvt

Here yhvt denotes the outcome variable of interest for household h in village v at time t. The

omitted category is Control 2 households in GBL villages. The average treatment effects of

9

TRAIL are GRAIL are estimated by β̂3 − β̂2 and β̂6 − β̂5 respectively.9 β̂2 is the difference in

outcomes of Control 1 and Control 2 borrowers in TRAIL villages, so represents the TRAIL

selection effect; analogously, β̂5 measures the GRAIL selection effect. Xhvt is a set of additional

controls, including land owned by the household, religion and caste of the household, and age,

education and occupation of the oldest male in the household. I(Yeart) denotes two year

dummies to control for secular changes over time. Standard errors are clustered at the hamlet

level.10

4.1 Treatment Effects on Agricultural Borrowing

In Table 6, we examine the treatment effects on agricultural borrowing, estimated using equa-

tion (1). In column 1 we see that the total agricultural borrowing of TRAIL Treatment

households was Rs. 7517 (or 134%) larger than that of TRAIL Control 1 households over the

three-year study period, significant at the 1% level. In GRAIL villages we see a similar increase

in agricultural borrowing of Rs 7341. Column 2 makes it clear that the program loans did not

crowd out agricultural loans from other sources: although the point estimates on non-program

agricultural borrowing are negative, they are small and statistically insignificant. Overall, it

appears that in both schemes borrowers used the program loans to increase their net agricul-

tural borrowing, and did not merely substitute away from expensive non-program to cheaper

program credit.11

4.2 Treatment Effects on Potato Cultivation and Farm Incomes

Table 7 shows ATEs on farm outcomes. We focus on potatoes, since as described above, the

loan terms were designed to facilitate potato production.12 We aggregate the data collected

from the four-monthly surveys into a farmer-year level dataset measuring the amount of land

planted with potatoes, potato output, sales, revenues, production costs, value-added and im-

puted profits.13 The TRAIL loans led to large and statistically significant increases in potato

9All treatment effects are intent-to-treat estimates because they compare the outcomes of households assignedto the Treatment and Control 1 groups, regardless of whether the Treatment households actually took the loan.

10 The administrative definition of a village in our study corresponds to a collection of hamlets or paras.Households within the same hamlet tend to be more homogenous, are more likely to interact with each other,and arguably experience correlated shocks to cultivation and market prices. The results are robust to clusteringat the village-level instead.

11We conjecture this is explained by reluctance of farmers to disrupt their traditional informal credit rela-tionships within the village, as they may have been uncertain about the permanence or duration of the programinterventions.

12Estimated effects on the other three major crops are available on request.13This year-level aggregation is necessary to correctly line up the costs that a farmer incurred on the crop,

with the revenues earned from the crop, since harvested potatoes can be stored and sold at different points in

10

cultivation. The effect is concentrated on the intensive but not the extensive margin: the

likelihood of cultivating potatoes did not change significantly (column 1), but TRAIL Treat-

ment households devoted 0.09 acres (or 27%) more land to potatoes (column 2), and produced

950 kg (or 26%) more potatoes than Control 1 households. There are related effects on the

implied potato sales revenue: TRAIL Treatment households received Rs. 3900 (or 27%) more

than TRAIL Control 1 households in potato sales revenue (column 5).14 After we subtract the

corresponding increase in the cost of production (Rs. 1846 or 22%, column 4), they earned Rs.

2060 (or 36%) more in value-added (column 6). When we further subtract the imputed cost of

family labor employed in potato farming, this works out to a statistically significant Rs. 1907

or 40% increase in profit (column 7).15

Next, consider the effects of the GRAIL scheme. GRAIL loans appear to have changed the

extensive margin for potato farming by drawing in households that would not have grown

potatoes otherwise. Only 64 percent of GRAIL Control 1 farmers planted potatoes, but 77%

of GRAIL Treatment households did. This increase of 13% points (or 20%) is statistically

significant at the 1% level (column 1). Average acreage (column 2) and output (column 3) also

increased significantly. Although the point estimates for the GRAIL scheme are smaller than

for the TRAIL scheme, the difference between the TRAIL and GRAIL treatment effects is

not statistically significant. Revenue increased by 19% (column 5), and the cost of production

increased by a larger 28% (column 4). The result is that unlike TRAIL, GRAIL Treatment

households did not earn significantly larger value-added or imputed profit than their Control 1

counterparts, with a 4% increase (compared to 40% in TRAIL).16 The difference in the TRAIL

and GRAIL treatment effects on imputed profits (column 7) is statistically significant at 10%.

In sum therefore, both schemes led borrowers to increase potato cultivation, but only TRAIL

borrowers did this profitably. This appears to be partly because GRAIL borrowers’ potato

output increase was accompanied by substantially larger increases in production costs, and a

greater decline in potato price realized.

the year.14Note: Sales revenues are calculated from data on each individual potato sales transaction. For amounts

that farmers held for self-consumption, we impute the sales revenue by pricing the potatoes at the median saleprice in the village.

15To calculate the shadow cost of family labour, we use the reported family labor time (male, female andchild labor separately) spent on the crop and price it at the median wage paid for that type of hired labor inthat village in that year for that crop.

16Since we analyse a large number of dependent variables, we correct for the increased chance of findingstatistically significant results. Following Hochberg (1988), we report a conservative p-value for an index ofvariables in a family of outcomes taken together (see Kling et al. 2007). The variables are normalized bysubtracting the mean in the control group and dividing by the standard deviation in the control group; theindex is the simple average of the normalized variables. To adjust the p-value of the treatment effect for anindex, the p-values for all indices are ranked in increasing order, and then each original p-value is multiplied bym− k + 1, where m is the number of indices and k is the rank of the original p-value. If the resulting value isgreater than 1, we assign an adjusted p-value of > 0.999.

11

Table 8 shows ATEs on aggregate farm and non-farm incomes of the household. In column

1, farm value-added is computed as the sum of value-added for all four major crops grown

by our sample farmers: potatoes, paddy, sesame and vegetables. The TRAIL loans resulted

in a 21% increase in average farm value-added, significant at the 1% level, compared to 1.3%

increase in GRAIL which was not statistically different from zero. In column 2, we examine

effects on non-agricultural income: a sum of rental, sales, labour and business income. We

are unable to obtain precise estimates of ATEs on non-agricultural income, possibly due to

higher measurement error for these income sources. Nevertheless, column 2 indicates that

neither loan scheme increased non-agricultural incomes significantly: the point estimate for

the GRAIL scheme is negative. Column 3 shows aggregate income (aggregating farm and

non-farm sources) rose by 11.3% in TRAIL, as against a 10% drop in GRAIL; this difference

is significant at the 10% level.

5 Loan Performance

The financial sustainability of a lending program depends on loan acceptance and repayment

rates. Column 1 shows effects on loan take-up rates, which are computed as follows. Condi-

tional on being eligible for a loan, borrowers are more likely to accept the loan if they expect

to benefit from it. After cycle 1, in each subsequent loan cycle eligibility of a Treatment house-

hold depended on past repayment.17 Therefore in any given loan cycle a Treatment household

may not have a loan either because it had defaulted on a previous loan, or because it chose

not to take the program loan this cycle despite being eligible for it. In column 1 of Table 9

we disregard those who had defaulted on a previous loan and instead focus only on those who

were eligible to borrow in the loan cycle. The sample means in Panel A show that TRAIL

Treatment households accepted the program loan 94% of the time, whereas GRAIL Treatment

households took it 87% of the time (column 1). The difference of 6.5% points is statistically

significant at the 1% level. In Panel B we present results from a regression that follows the

equation below:

yhvt = α0 + α1GRAILv + γXt + εhvt (2)

As column 1 shows, the difference in loan acceptance rates is statistically significant even after

we control for cycle fixed effects and other household characteristics (landholding, religion,

caste, age and educational attainment of the oldest male member of the household) through

17In particular, a borrower needed to repay at least 50% of the amount due before he/she became eligible toreceive a new program loan.

12

Xt to account for seasonal variation.

We define a loan to be in default if less than 50% of the amount due was repaid by the due

date. In both schemes, the average default rate over the three-year intervention period is

an identically low 7%. The regression result in Panel B confirms that there is no difference

in default rates in the two schemes. Despite the fact that the scheme did not increase their

incomes significantly, GRAIL Treatment households repaid their loans at similarly high rates

as TRAIL Treatment households.

6 Explanations

We now examine explanations for our empirical result so far: the TRAIL scheme generated

significant increases in Treatment households’ farm value-added, whereas the GRAIL scheme

did not. We consider two possibilities: (i) differences in selection (henceforth selection effect);

and (ii) differences in agents’ actions, conditional on agent type (henceforth conditional treat-

ment effect). In order to do this, we need a way to represent and estimate relevant dimensions

of underlying heterogeneity of farmers. We next explain how this is done.

6.1 Estimation of Control Households’ Ability

Consider farmer i in village v in year t.18 The production function takes the Cobb-Douglas

form and can be written as follows:

Yivt = PvtAi1

1− αl1−αivt (3)

where Pvt denotes a village-level yield shock, Ai is the farmer’s total factor productivity and an

increasing function of farmer ability θi, livt is the farmer’s chosen scale of cultivation, and the

parameter α ∈ (0, 1). Farmer ability θi follows a common, given, distribution in both GRAIL

and TRAIL villages. We also assume that the cost of production per unit area ci is decreasing

in farmer ability. In this section, we take Ai and ci as given for each farmer; in Section 7 we

allow them to depend also on help received and monitoring which are endogenously determined.

The control group farmer borrows from informal lenders, who charge the farmer interest at the

18 Here we provide a simplifed version of the more general model that we will present in Section 7. The currentmodel abstracts from the agent engagement effects occurring through endogenous help and monitoring whichaffect farmer productivity and costs. Instead, it assumes that farmer productivity and costs are correlated witha measure of exogenous ability. We shall see that this simpler model is a special case of the more general model.It is also essentially the same model as in Maitra et al. (2017).

13

rate ρvt.19 To cultivate potatoes, the farmer must incur a fixed cost F > 0. Then, the control

group farmer chooses l = lcivt to maximize

PvtAil1−α

1− α− ρvtcil − F

This yields

log lcivt =1

αlog

Aici

+1

α[Pvt − ρvt] (4)

A similar expression obtains for the logarithm of output produced.

Observe that log Aici

is increasing in ability. This serves as a suitable measure of productivity

which is exogenous for each farmer under the maintained assumptions that productivity and

cost depend on ability alone. It follows from equation (4) that this productivity measure can

be estimated by the household fixed effect in a household-year level panel regression, where

the scale of potato cultivation (acreage or output) is regressed on farmer, village and year

dummies.20

Equation (4) applies only to farmers with ability above some threshold θvt who end up earning

profits from cultivation that cover the fixed cost F . Others less able would choose not to

cultivate potatoes. Roughly 30 percent of control (Control 1 and Control 2) farmers in our

sample either did not plant potatoes even once in the three years of our study period, or

planted them only once in three years. Hence we are able to obtain a continuous productivity

estimate only for farmers who cultivated potatoes at least two out of three years. For the rest

(who we henceforth refer to as non-cultivators (NC)), we can only estimate an upper bound to

their productivity by the lower endpoint of the estimated productivity distribution among the

cultivators, those who cultivate at least two years. We pool the NCs into a single group, and

treat the productivity distribution as having a mass point at this lower endpoint. This is an

upper bound to the average productivity of the NC group, but none of the comparisons below

will change if it were replaced by any lower estimate.

The right panel of Table 10 presents descriptive statistics on the farm productivity of cul-

tivators. Again, the sample is restricted to Control 1 households with at most 1.5 acres of

land in TRAIL and GRAIL villages. Observe the high dispersion in productivity: at the 75th

19We assume that traders and lenders know the ability of each farmer. One specific version of this modelinvolves Bertrand competition among lenders, all of whom have cost of capital ρ′vt in village v in year t.Alternatively, they charge a markup on the cost of capital at a fixed rate. If the probability that a farmerdefaults on his loan depends on his ability, then the default risk is incorporated into the markup. In the generalmodel to be presented later, we will allow for interlinked credit-cum-marketing contracts between farmers andtraders, and will find similar results.

20Note that since households’ landholding tended to remain fixed over our three-year sample period, the effectof landholding is subsumed in the household fixed effect.

14

percentile a farmer cultivates between three to four times the area cultivated by a farmer at

the 25th percentile. With an elasticity of acreage in the production function is at least 0.5,

this implies that Aici

was at least 150% -200% higher at the 75th percentile.

It is interesting to check the extent to which household observable characteristics correlate

with the productivity estimate. This tells us whether productivity can be estimated from

demographic characteristics of the household that could be collected through a household

survey by formal lenders, or the government. In the left panel of Table 10 (see column 1) we

regress the household productivity estimate on household characteristics, and find significant

correlations with socio-economic characteristics. Households with higher landholding, Hindu

households, households that belonged to the “general” castes and male headed households

had higher farm productivity. Note the low R-squared of this regression, indicating that the

observable demographic characteristics of the household can only explain 18% of the variation

in household productivity. This underscores the fact that these community-based agents likely

had other local information about farmer productivity that would be difficult for an outsider

to observe.21

6.2 Selection Differences between TRAIL and GRAIL Schemes

We can now examine whether the TRAIL and GRAIL agents systematically recommended

households of different productivity levels. For this exercise we first examine whether, in each

scheme, recommended households were positively selected. Next, we check how the produc-

tivity distributions of recommended households compare in the two schemes. We do this by

focusing only on Control 1 households, since they were recommended by the agent but did not

receive any program loans. This helps avoid the concern that the receipt of the loan may itself

have changed the estimated productivity of the household.

The right panel of Table 10 shows the proportion of cultivators was higher in TRAIL (72.65%)

than GRAIL (65.53%). The mean farm productivity and the maximum farm productivity are

both higher in TRAIL households as is the standard deviation. The value of farm productivity

at the three quartiles of the distribution is consistent with the notion that TRAIL Control 1

households were more productive on average than GRAIL Control 1 households.

In Figure 1, we consider more detailed assessments of selection performance of either scheme,

based on comparing respective productivity distributions of Control 1 and Control 2 subjects.

The left panel presents the cumulative distribution function (CDF) of estimated productivity,

21A LASSO estimator performs only marginally better than the ordinary least squares estimator. Under theExtended Bayesian Information Criterion the selected LASSO model has an R-squared of 0.20.

15

separately for TRAIL Control 1 and TRAIL Control 2 households, with the convention that

we pool NCs at the bottom end at their estimated upper bound; pooling them at any lower

productivity level would not change any of the comparisons to follow. This corresponds to the

flat segment in each of the plotted CDFs at the bottom end. As we can see, the CDF of Control

1 households first order stochastically dominates that of Control 2 households: a smaller

fraction of Control 1 households had productivity below any arbitrarily chosen threshold.

Using a two sample Kolmogorov-Smirnov test, we reject the null hypothesis that the two

distributions are identical (p-value = 0.005).22 This indicates that TRAIL Control 1 farmers

were more productive than TRAIL Control 2 farmers, subject to the qualification that the

distributions cannot be compared below the bottom end-point as a point estimate of the

productivity of NCs cannot be obtained.

A very similar pattern emerges in the right panel, suggesting that GRAIL Control 1 farmers

were more productive than GRAIL Control 2 farmers. The two-sample Kolmogorov-Smirnov

test rejects the hypothesis of equality of distributions (p-value = 0.011.23 Thus both types of

agents recommended the more productive households as borrowers for the loan scheme.

To check how the productivity of recommended households compares across the two schemes,

in Figure 2 we plot the productivity distributions of the Control 1 households in TRAIL and

GRAIL villages. We can now see that the CDF for GRAIL Control 1 households first order

stochastically dominates that for TRAIL Control 1 households. A two-sample Kolmogorov-

Smirnov test rejects the null hypothesis that the two distributions are identical (p-value =

0.06).24 Since villages were randomly assigned to TRAIL or GRAIL treatment arm, and

village and household characteristics are balanced across treatment arm, any differences in the

productivity levels of recommended households in the two schemes cannot be due to inherent

differences across villages. Instead, the results indicate that TRAIL agents selected borrowers

of higher productivity more often than GRAIL agents did.

6.3 Conditional Treatment Effect Differences between TRAIL and GRAIL

Schemes

As discussed above, the TRAIL and GRAIL agents may also have played other informal roles

in the scheme, beyond selecting borrowers. They may have engaged further with the Treatment

22Since our productivity estimates are generated variables, we also simulate 2000 bootstrap samples and runthe Kolmogorov-Smirnov test for each Control 1 vs Control 2 CDF comparison. In 87 percent of the simulations,we can reject the null hypothesis that the two distributions are identical.

23Again, the Kolmogorov-Smirnov test rejects the null hypothesis that the two distributions are identical in83 percent of our bootstrap simulations.

24The Kolmogorov-Smirnov test rejects the null hypothesis that the two distributions are identical in 74percent of our 2000 bootstrap simulations.

16

households and influenced their farming practices and costs. If the two types of agents engaged

differently or varied in their effectiveness, this could have led treatment households of equal

productivity to have different treatment effects. Next we examine to what extent we see

differential conditional treatment effects.

In order to estimate the conditional treatment effects we first need to estimate the productivity

of treatment households. Here, we cannot use the same procedure as we used for control

households, because the treatment households’ scale of cultivation is affected by the treatment

itself, both directly through the loan they received and indirectly, through engagement with

the agent which could also respond to the treatment.

To infer the productivity of treated households, we make an Order-Preserving Assumption

(OPA), viz. that the rank order of treatment farmers by cultivation scale or output remains

the same as the corresponding rank order of control households. In other words, we assume

that the treatment does not change the relative position of a household in the distribution of

cultivation scale or output. See Athey and Imbens (2006) for a similar assumption which is

key to identifying treatment effects in non-linear difference-of-difference settings. Under this

assumption, in each treatment arm (TRAIL and GRAIL) respectively, we rank Treatment

farmers by cultivation scale (or output), and assign to them the counterfactual productivity

estimate Ai of the farmer at the same rank in the Control 1 distribution. Hence this is to

be interpreted not as an actual productivity estimate, but instead a latent (counter-factual)

productivity had the farmer in question not been treated. This is what we need in order

to estimate heterogenous treatment effects, which involves comparing outcomes for Control 1

subjects at some productivity level with treated subjects with the same latent productivity.

Such a procedure can be used to estimate CTEs at every productivity level in the support of

the continuous part of the estimated distribution. However the small scale of the experiment

implies that we do not have sufficient scale to estimate these precisely enough. We therefore

group subjects into three “productivity bins”. All households who cultivate potatoes less than

twice in the three survey years are placed in Bin 1. Among the rest, those with produc-

tivity estimate below and above the median (of the cultivators) are placed in Bins 2 and 3

respectively.25

This allows us to estimate bin-specific TRAIL treatment effects according to the following

25 Figure A1 presents the fraction of Treatment (T) and Control 1 (C1) households in each productivity bin,separately in the TRAIL and GRAIL villages. These results are consistent with those presented in Figure 2 andare indicative of superior selection by TRAIL agents, relative to GRAIL agents.

17

specification:

yivt =

3∑k=1

ξ1k B̂inik +

3∑i=1

ξ2k (Control 1iv × B̂inik) +

3∑k=1

ξ3k (Treatmentiv × B̂inik)

+3∑

k=1

ξ4k B̂inik ×GRAILv +3∑

k=1

ξ5k (Control 1iv × B̂inik ×GRAILv) (5)

+3∑

k=1

ξ6k (Treatmentiv × B̂inik ×GRAILv) + γX′ivt + εivt

From equation (5), the TRAIL CTEs for the three bins are ξ31−ξ21, ξ32−ξ22 and ξ33−ξ23, and

ξ61 − ξ51, ξ62 − ξ52 and ξ63 − ξ53 in GRAIL. Note that we are abstracting from heterogeneity

of treatment effects within productivity bins. However, the decomposition results we report

below turn out to be robust with respect to using finer partitions of the productivity range.

Note also that for subjects that did not cultivate potatoes the corresponding outcomes are

replaced by zero when the outcome in question is potato area cultivated, output produced

or profits earned. When we examine unit costs, treatment effects are defined by difference

between unit costs for treated subjects, and control 1 subjects who cultivated at least one out

of the three years.

The results are presented in Table 11. Standard errors are bootstrapped using 2000 iterations,

since the regression is based on the classification of subjects into different bins on the basis of

their estimated (latent) productivity. In column 3 we see the heterogenous treatment effects

of the program loans on potato output, and in columns 7 and 8 we see the effects on potato

value-added and imputed profit, respectively. In Panel A, we see positive, significant TRAIL

CTEs on potato output, value-added and imputed profit in Bins 2 and 3, with point estimates

higher in Bin 3. In Panel B, we see significant and positive CTEs only for potato output and

revenues for these these two bins, but not for value-added or profit. The reason is that the

corresponding CTEs on costs were also large, resulting in a negative point estimate for profits

in Bin 1 and small, insignificant effects in the other two bins. With regard to the difference in

CTEs between TRAIL and GRAIL, we see a significantly higher TRAIL effect on potato price

realized in Bin 1, and on unit cost in Bin 3. The value added CTEs are substantially larger in

TRAIL for all three bins, but is not statistically significant.

6.4 Decomposition

We have seen that TRAIL achieved superior selection, as well as higher conditional treatment

effects on potato profits and value-added. Now we seek to measure the respective roles of these

18

two differences in accounting for the observed ATE difference.

If σR(ζ) denotes the proportion of borrowers of productivity ζ that were recommended in loan

scheme R ∈ {T,G}, the ATE of Treatment R can be written as:

ATER =

∫ζσR(ζ)TR(ζ)dζ (6)

Therefore the difference in the ATEs of the two schemes is

ATET −ATEG =

∫ζ[σT (ζ)− σG(ζ)]T T (ζ)dζ +

∫ζσG(ζ)[T T (ζ)− TG(ζ)]dζ

where the first term on the right-hand-side is the selection effect, or how much lower the TRAIL

ATE would have been if selection had been as in the GRAIL scheme, while the CTEs were as in

TRAIL. The second term represents the “agent engagement effect”, or what the GRAIL ATE

would have been if borrowers were selected as in GRAIL but CTEs had been as in TRAIL.

In Table 12 we present the results of this decomposition exercise where we partition borrowers

into the three productivity bins described earlier, and ignore variations within each bin. In

columns 1 and 2 we compute the proportion of the TRAIL and GRAIL pools of Control 1

households that belong to each productivity bin (σR(ζ)). In column 3, we compute the TRAIL

v. GRAIL difference in these proportions, or measure the difference in the likelihood of finding

a household of this productivity bin in the TRAIL as opposed to the GRAIL recommended

household pool. As suggested by our discussion of Figures 1, compared to the GRAIL pool, a

smaller proportion of the TRAIL pool lies in Bin 1, but a larger proportion lies in Bin 2 and

Bin 3. We use the differences from column 3 to compute a weighted average of the TRAIL

conditional treatment effects in column 7. This average is Rs. 131.20 and represents the

difference in the TRAIL v. GRAIL average treatment effects we might have seen, if superior

borrower selection alone were driving the differential effects. This accounts for only 8.4% of

the actual estimated difference in average treatment effects of Rs. 1566.84.

Next, for each productivity bin, we compute the difference in the estimated conditional treat-

ment effects between the TRAIL and GRAIL schemes (see column 6). We weight these by

the proportion of GRAIL Control 1 households that belong to this productivity bin (σG(ζ)).

The weighted average, seen in column 8, indicates that conditional on productivity, differences

in the TRAIL and GRAIL treatment effects cause the average treatment effects of the two

schemes to differ by Rs. 1182.30. In other words, 75.4% of the estimated ATE difference

can be attributed to differences in treatment effects conditional on borrower productivity.26

26Since we ignore intra-bin heterogeneity in productivity, our decomposition exercise is unable to explain partof the difference in average treatment effect. When we instead use the continuous measure of productivity we

19

We have repeated this exercise with finer partitions of the continuous productivity range (for

cultivators) to examine whether the low estimated contribution of the selection effect owes

to intra-bin differences in selection that are ignored in the above procedure; even with a fully

continuous distribution the contribution of the selection effect remains small, below 20%, while

the role of CTE differences remain substantial (above 60%).

Clearly then, selection differences are not the main driver behind the superior performance of

the TRAIL scheme. Instead, our results indicate that farmers of similar productivity levels

increased their farm incomes by more in TRAIL. Since TRAIL and GRAIL differ only in the

nature of the agent, this suggests TRAIL agents engaged differently with client farmers in ways

that enabled them to realize higher profits.

7 A Model Of Agent-Farmer Engagement

To investigate further the engagement between farmers, traders and agent, we formulate a

model of interactions between them. These interactions could be conversations about the

weather, market prices, cultivation practices and techniques and harvesting times. They could

be used to provide technical or marketing assistance, or monitor and regulate farmers’ actions,

which would affect productivity and cultivation costs.

As explained in the Introduction and in Section 2, besides the commissions, TRAIL agents

are likely to be motivated to earn higher profits from middleman activities based on crop

purchases from farmers, net of costs of credit they provide and time they spend engaging

with the farmers. They share both downside and upside risk with the farmers. The former

applies when the farmer’s crop is unsuccessful or low, the farmer is likely to default on loans

(both program loans and informal loans from the TRAIL agent) resulting in loss of the TRAIL

commission, default on the informal loan, and low potato purchases. Conversely when the

farmer plants more potatoes, is successful in growing a larger output and covering his costs,

the farmer is likely to repay program and informal loans and sell more potatoes to the TRAIL

agent. These business related motives do not apply in the case of GRAIL agent, whose main

motivation (besides earning commissions) is securing political or mission-oriented objectives

of the incumbent political party. These objectives are likely to target poorer farmers of lower

productivity, and then ensuring that they do not default. The GRAIL agent’s incentives

would then resemble that of a classic banker, sharing mainly the downside risk but not the

corresponding upside risk, owing to the lack of a potato trading motive. If there is a trade-off

estimate that 17.5% of the estimated ATE difference can be explained by selection differences, and 82.5% is dueto differences in conditional treatment effects.

20

between lowering default risk and raising mean returns, the TRAIL agent would be inclined

to favor of the higher mean returns while the GRAIL agent would want to minimize default

risk. The model we develop incorporates this trade-off.

7.1 Assumptions

Farmers vary in intrinsic ability, denoted by θ. Less able farmers are less likely to grow a

successful crop. Farmers can be monitored by local intermediaries such as traders or agents.

When monitored, farmers choose less risky production techniques, and do more to prevent pest

attacks. Therefore the crop success rate increases in monitoring. Formally, the probability of

crop success p = p(θ,m), pθ > 0, pm > 0, pmm < 0. Moreover pθm < 0, i.e, monitoring is more

effective for less able farmers who are higher default risks.

Conditional on crop success, the farmer’s output depends on his total factor productivity

(TFP) a(θ,m), and a concave production function f that depends on the scale of cultivation

l. Monitoring lowers productivity by discouraging adoption of high return but risky crop

varieties or production methods. Hence expected TFP A(θ,m) ≡ p(θ,m)a(θ,m) decreases in

monitoring: at any given scale of cultivation, monitoring has a negative effect on TFP that

outweighs the positive effect on the crop success rate. Monitoring also raises unit production

costs, because it causes farmers to use expensive inputs such as pesticides and water to increase

the chance that their crop will succeed.

On the other hand, traders can also help farmers by providing advice about the types or sources

of inputs, and the timing of transactions. This lowers farmers’ unit costs. The unit cost of

production c(h,m) is thus decreasing in help h and increasing in m.

Monitoring and help are assumed to be time-consuming activities, so are measured in units of

time. Traders have an opportunity cost of γT per unit time, while GRAIL agents’ opportunity

cost is γG. These may differ owing to differences in their occupation and wealth; we impose

no restriction on this.

A specific parametric example involves constant elasticity (CE) success, production, produc-

tivity and cost functions: p(θ,m) = 1− (1−π(θ))(1+m)−µ0 , where π is increasing and µ0 > 0;

f = l1−α

1−α with 1 > α > 0; a(θ,m) = z(θ)(1 + m)−µ1 and c(h,m) = (1 + h)−ν(1 + m)µ2 where

z is increasing, µ1, µ2 > 0, 1 > ν > 0. We also impose the conditions (i) ν < α1−α , to ensure

diminishing returns to help, and (ii) 1 + µ0µ1

< 11−π(θ) where θ is a lower bound to θ, which

implies expected TFP is falling in m.

All inputs are purchased upfront, so the farmer incurs cultivation cost c(h,m)l when he plants

21

the crop, and earns revenue a(θ,m)f(l) at harvest time conditional on avoiding a crop failure.27

The farmer has no wealth to self-finance production, and thus needs to borrow c(h,m)l upfront.

Borrowers have limited liability, so they repay loans only if the crop is successful. We also

assume that there is no strategic default, owing to the dynamic repayment incentives created

for borrowers.

Traders in the village face a constant cost of capital ρ. They enter into “interlinked” contracts

with farmers; these contracts specify the scale of cultivation, extent of help and monitoring. In

turn these determine the farmers’ borrowing and output levels. Farmers and traders are risk-

neutral. The farmer sells the output to a trader who earns a middleman margin of τ per unit

transacted. Each farmer enters into a contract with a representative village trader. Traders

have perfect information about farmers’ types, and there are no frictions, so the interlinked

contract maximizes joint expected payoff of trader and farmer. The division of payoffs between

the two depends on the degree of competition in the market for contracts, represented by a

suitable lumpsum side-payment, which will have no productive consequences.28

7.2 Control Farmers

A contract between farmer F of ability θ and trader T is represented by a scale of cultivation

l, help h, monitoring m, an interest rate r and a side-payment s. The first three determine the

size of the loan c(h,m)l. The farmer repays the loan if his crop succeeds. Hence the farmer’s

expected payoff (excluding fixed cost F ) is

p(θ,m)[a(θ,m)f(l)− rc(h,m)l] + s (7)

while the trader’s payoff is

τp(θ,m)a(θ,m)f(l) + [rp(θ,m)− ρ]c(h,m)l − γT (m+ h)− s (8)

where τ represents the product of the share of output sold to the trader, and middleman

margin earned by the trader per unit output. We take τ to be exogenous. An efficient contract

maximizes the joint payoff given by

(1 + τ)A(θ,m)f(l)− ρc(h,m)l − γT [m+ h] (9)

27The product price is treated as exogenous and normalized to unity. We abstract from the possible role ofhelp and monitoring in determining the product price, since that would give us qualitatively similar results asfor their effect on unit cost.

28An implicit assumption here is that the limited liability constraint does not bind. This will be true if farmersare on the short side of the market and therefore have all the bargaining power.

22

To start with, note that it is optimal for the trader to not monitor the farmer at all (mc(θ) = 0),

since monitoring lowers expected productivity A, imposes a monitoring cost, and increases the

production cost.

Next, given a certain level of help h, the optimal scale of cultivation lc(θ, h) maximizes (1 +

τ)A(θ, 0)f(l)− ρc(h, 0)l, so satisfies the first order condition

f ′(lc(θ, h)) =ρc(h, 0)

(1 + τ)A(θ, 0)(10)

The trader chooses the optimal level of help hc(θ) to maximize (1 + τ)A(θ, 0)f(lc(θ, h)) −ρc(h, 0)lc(θ, h)− γTh, which (using the Envelope Theorem) satisfies the first order condition

[−ch(hc(θ), 0)]ρlc(θ, hc(θ)) = γT (11)

Observe also that the choice of scale of cultivation can be delegated to the farmer, if the interest

rate is set at

rc(θ) =ρ

(1 + τ)p(θ, 0)(12)

This interest rate equals the cost of capital ρ adjusted upwards for default risk, and then

subsidized by the trader so as to induce the farmer to internalize the effect of cultivation scale

on T’s profits. It therefore follows that:

Proposition 1 Among control farmers, interest rates are decreasing in ability.

In the constant elasticity (CE) case, we can derive closed form expressions for the efficient

contract:

lc(θ, h) = [(1 + τ)A(θ, 0)(1 + hc(θ))ν

ρ]1α (13)

1 + hc(θ) = [(1 + τ)A(θ, 0)

ρ1−α { νγT}α]

αα−ν(1−α) (14)

while mc(θ) = 0 and rc(θ) is given by (12). Note that higher ability farmers receive more

help and thus end up with higher TFP, a larger cultivation scale, higher expected output and

higher expected profit. Hence productivity is monotone increasing in ability, enabling us to

proxy farmer ability by observed scales of cultivation or output.

23

7.3 TRAIL Treatment Effects

In TRAIL, a trader is appointed the agent, and recommends borrowers for TRAIL loans.

These loans are offered at interest rate rT , which is lower than the informal cost of capital for

traders ρ. Agents earn a commission of ψ ∈ (0, 1) per rupee interest paid by the borrowers

they recommended. We assume that any farmer whom the agent selects is already committed

to cultivating lc, financed by informal loans taken before the TRAIL loan was offered to

him/her.29 As a result the TRAIL loan finances an increase in the cultivation scale.30 This

applies to farmers in productivity Bins 2 and 3; for those in Bin 1 there are no pre-existing

plans for cultivating potatoes. In what follows, we present calculations for farmers in Bins 2

and 3; for those in Bin 1 we set the pre-existing cultivation scale Lc(θ) to zero.

The efficient contract between T and F will now involve a supplementary cultivation scale of

lt, resulting in total scale of lT ≡ lc + lt. The levels of monitoring and help will be adjusted to

mT , hT . Then the joint payoff of T and F is

(1 + τ)A(θ,m)f(Lc(θ) + lt)− [ρLc(θ) + p(θ,m)rT (1− ψ)lt]c(h,m)− γT [h+m] (15)

where Lc(θ) ≡ lc(θ, hc(θ)).

The TRAIL agent continues to find it optimal not to monitor the farmer: mT (θ) = 0.

The treatment effect on cultivation scale lt satisfies

f ′(Lc(θ) + lt(θ)) =c(hT (θ), 0)p(θ, 0)(1− ψ)rT

(1 + τ)A(θ, 0)(16)

The level of help is given by

[−ch(hT (θ), 0)][ρLc(θ) + p(θ, 0)(1− ψ)rT lt(θ)] = γT (17)

Comparing (17) and (12), it is evident that TRAIL treated farmers will receive more help than

TRAIL control farmers of the same ability level: hT (θ) > hc(θ). Hence the TRAIL treatment

will increase agent engagement, thereby lowering the unit costs for Treatment farmers in Bins

2 and 3. Since rT < ρ, it follows that TRAIL treated farmers have a lower expected marginal

cultivation cost than corresponding control farmers in bins 2 and 3: c(hT (θ), 0)p(θ, 0)rT <

c(hc(θ, 0))ρ. Hence the TRAIL treatment effect on scale of cultivation and expected output

29This is in order to explain the lack of treatment effects on informal borrowing.30Recall that in Table 6 we did not see any evidence that the TRAIL loans crowded out informal loans.

24

will be positive. This is the result of both the lower marginal borrowing costs and the induced

increase in the TRAIL agent’s help, which lower unit production costs. This is true for Bins 2

and 3; for Bin 1 the result is obvious since control farmers in Bin 1 do not cultivate potatoes.31

32

Proposition 2 TRAIL conditional treatment effects on agent interactions, area cultivated and

output are positive, and on unit cost are negative.

7.4 GRAIL Treatment Effects

In the GRAIL scheme, the political incumbent appoints an agent who is not a trader. This

agent does not lend, or trade in inputs or crop output, and so does not have the same business-

related incentives as a TRAIL agent. Instead, his objectives are political or ideological, rep-

resented by welfare weight v(θ). In the context of West Bengal, it is natural to suppose that

v is decreasing in θ, representing a redistributive ideology or a motivation to garner support

from low-ability farmers for the incumbent political party. In addition, this agent would earn

commissions that depended on the interest payments by the borrowers he recommended. Both

the political and non-political payoffs of the GRAIL agents depend on whether Treatment

farmers successfully repay their loans. Hence the objective of the GRAIL agent, conditional

on selecting a farmer of ability θ, is to maximize

[χ+ v(θ)]p(θ,m)− γGm (18)

where χ denotes the agent’s return from the repayment based commission. We abstract here

from loan size and its effect on the agent’s commission, which complicates the analysis con-

siderably is likely to be of second order importance relative to political or ideological motives.

The results derived below would hold in an extended model where loan size is incorporated, if

the commissions are small relative to the political weight v(θ).

In contrast to the TRAIL agent, the GRAIL agent is motivated to minimize default risk, so

has no incentive to help Treatment farmers, but instead seeks to monitor them. The optimal

level of monitoring (positive if γG is small enough) satisfies

[χ+ v(θ)]pm(θ,mG(θ)) = γG (19)

31Unfortunately, closed form expressions for the TRAIL treatment effects in the CE case can no longer beobtained for farmers in Bins 2 and 3, so we are unable to provide theoretical results showing how TRAILtreatment effects vary with θ.

32Treatment effects for Bin 1 are defined as the difference between unit costs of treated farmers with thesubsample of Control 1 farmers that produce in one out of the three years.

25

Since monitoring is more effective when farmers are less able, and the welfare weights are

decreasing in ability, mG(θ) is decreasing in ability. Hence GRAIL treated farmers are less

likely to default on their loans than GRAIL control farmers of the same ability. The GRAIL

control farmers’ default rates equal the default rates of both control and treated farmers in

the TRAIL scheme, since as we saw above, the TRAIL agent does not monitor in equilibrium.

Finally the ratio of interest rates r1r2

paid on informal loans by borrowers of ability θ1, θ2 where

θ1 < θ2, equalsRI2RI1

, where RIi denotes the repayment rate by θi on informal loans. This relative

repayment rate on informal loans will be higher than on GRAIL loans, since repayment rates

for lower ability borrowers increase by more owing to the more intensive monitoring by the

GRAIL agent. Hence r1r2

will be higher thanRG2RG1

, where RGi denotes the repayment rate on

GRAIL loans.

Proposition 3 (i) GRAIL Treatment effects on agent engagement are decreasing in produc-

tivity.

(ii) At any given productivity level, default rates on program loans are lower in GRAIL com-

pared with TRAIL

(iii) r1r2>

RG2RG1

, where ri, RGi denote the informal interest rate and GRAIL repayment rates of

borrowers of ability θi, if θ1 < θ2.

Now we turn to treatment effects on help, scale of cultivation, outputs and farm profits.

Monitoring by the GRAIL agent affects the payoffs of treated farmers and the trader they

contract with. Their joint payoff is given by

(1+τ)A(θ,mG(θ)+m))f(Lc(θ)+lg)−[ρLc(θ)+p(θ,mG(θ)+m)rT (1−ψ)lg]c(h,mG(θ)+m)−γT [h+m]

(20)

where lg denotes the additional area that the GRAIL treated farmer cultivates, and (h,m)

continues to denote help and monitoring activities of the trader. The trader continues to have

no incentive to monitor. Hence the contract involves a treatment effect lg(θ) on area cultivated

and a revised level of help hG(θ) which maximize

(1 + τ)A(θ,mG(θ)))f(Lc(θ) + lg)− [ρLc(θ) + p(θ,mG(θ))rT (1− ψ)lg]c(h,mg(θ))− γTh (21)

It is evident that for any choice of contract, monitoring by the GRAIL agent lowers the treated

farmer’s productivity and raises unit costs above those of the TRAIL treated farmer of the

same ability. This causes GRAIL treatment effects on farmer outcomes to be smaller that

TRAIL treatment effects.

26

To obtain the specific expressions for the treatment effects on area cultivated, output and unit

costs, observe that the GRAIL treatment effect on area cultivated GRAIL lg(θ) satisfies

f ′(Lc(θ) + lg(θ)) =c(hG(θ),mG(θ))(1− ψ)rT

(1 + τ)A(θ,mG(θ))(22)

When we compare (22) with (16), we see that the GRAIL treatment effect on area cultivated,

and hence output, is smaller than the corresponding TRAIL treatment effect if the unit cost for

a GRAIL treated farmer is larger than that for a TRAIL treated farmer of the same ability:

cG ≡ c(hG(θ),mG(θ)) ≥ cT ≡ c(hT (θ), 0). In turn the unit costs depend on the treatment

effects on the trader’s help. In the GRAIL scheme this is given by

[−ch(hG(θ),mG(θ))][ρLc(θ) + p(θ,mG(θ))(1− ψ)rT lg(θ)] = γT (23)

Greater monitoring makes help more effective at lowering unit costs, so it is difficult to compare

the help provided by the GRAIL and TRAIL agents. The comparison also depends on the

respective treatment effects on area cultivated and the probability of crop success.

In the CE case, we can obtain the following result.33

Proposition 4 In the constant elasticity case, TRAIL treated farmers in Bin 1 have smaller

unit costs, cultivate larger area, produce more output and earn higher value added/profit than

GRAIL farmers of the same ability in Bin 1. TRAIL treated farmers in Bins 2 and 3, have

lower unit costs, produce more output and earn higher value added/profit if the TRAIL treat-

ment effects on area cultivated (weighted by repayment rate) are at least as large as the GRAIL

treatment effects:

p(θ,mG(θ))lg(θ) ≤ p(θ, 0)lt(θ) (24)

In this case, TRAIL treatment effects on value added and farm profit are higher than corre-

sponding GRAIL treatment effects in Bins 2 and 3.

The comparison of the help provided between TRAIL and GRAIL turns out to be ambiguous

even in Bin 1, since it depends on whether α(1 + ν) is larger or smaller than 1, and either is

possible. Intuitively, this is because of two conflicting effects. On the one hand, since TRAIL

treated farmers cultivate a larger area they benefit more on the margin from lower unit costs

that result from more help. On the other hand, less monitoring directly lowers the marginal

effectiveness of help in lowering unit costs.

33The proof is presented in Section A1.

27

Nevertheless, unit costs end up being smaller for TRAIL farmers. This is because the TRAIL

agent does not monitor, and monitoring raises unit costs directly, which outweighs the induced

indirect effects on help, even if they happen to be larger in the GRAIL scheme. For Bins 2

and 3, the same result obtains if the additional condition (24) holds. This condition involves

endogenous variables, but can be checked empirically. It ensures that the benefits of unit cost

reductions are spread over a larger base in TRAIL, which motivates the trader to push unit

costs lower in TRAIL.

Finally, we compare effects of the two schemes on political support for the incumbent party.

In TRAIL the agent has a business profit motivation rather than a political one, so we expect

no treatment effects on political support for the incumbent. In GRAIL we expect recipients

to either be grateful for being selected for a low interest loan by the local incumbent party,

or reciprocate the benefit in a clientelistic arrangement.34 The incumbent is likely to target

low ability borrowers because they are poorer on average, and both welfarist and clientelistic

missions would be expected to target the poor more since the welfare impacts would be larger.

So we expect:

Proposition 5 Treatment effects on the likelihood of voting for the political incumbent in

GRAIL are positive on average, and decreasing in borrower productivity. Corresponding treat-

ment effects in TRAIL are zero.

8 Testing Predictions of the Model

We now test the predictions made in the preceding Lemmas.

Test of Proposition 1

We consider informal loans taken before our intervention began, in order to avoid the concern

that the intervention itself might have changed the interest rate they were charged.35 The

left panel of Figure 3 presents the average interest rate paid on informal loans by productivity

bin. This average is 21% per annum by households in Bin 1, which is significantly greater

than that for households in Bins 2 and 3 (15% (p-value = 0.03) and 16% (p-value = 0.04)

respectively). The right panel takes into account the variation in productivity. Recall that we

34This is a standard feature of the literature on political clientelism: poorer voters are likely to sell their votefor a lower price, so parties tend to target swing voters among the poor in their efforts to mobilize vote share.

35However, since only 10 households in each village received program loans, we do not believe there were anyspillover or general equilibrium effects.

28

do not have a continuous measure of productivity for those who did not cultivate potatoes;

for this group we compute the mean informal interest rate and plot it as a single point in the

right panel of Figure 3. For Bins 2 and 3, we run a locally-weighted polynomial regression

of the mean interest rate on farmer productivity. As the predicted values plotted in Figure 3

show, higher-productivity control farmers paid lower interest rates, confirming that Lemma 1

matches our data.

Test of Proposition 2

In each survey round, we asked sample households three questions to ascertain whether they

had spoken with the local trader in the previous three days, about cultivation, harvesting and

sales. We aggregate these data over the three survey cycles in a year, to arrive at a measure

of how much the trader engaged with farmers.

The left panel of Figure 4 presents the heterogeneous treatment effects of TRAIL on agent

engagement by ability bin. The treatment effects are positive, indicating that agents were

more likely to engage with Treatment households than with Control 1 households; the CTEs

were larger for more productive farmers.

Panel A of Table 11 presents the heterogenous treatment effects estimates on potato cultivation

in TRAIL, by productivity bin. In all three bins, TRAIL loans caused a significant increase in

the acreage that farmers devoted to potatoes (column 2) and potato output (column 3). The

treatment effects are larger for the more able farmers. In contrast, and consistent with Lemma

2, the point estimates of the TRAIL treatment effects on potato input costs er acre (column

10) are negative. They become increasingly negative as we move to higher productivity bins,

and ultimately the point estimate is statistically significant in Bin 3.

Test of Proposition 3

The right hand panel of Figure 4 presents estimates of heterogeneous treatment effects on agent

engagement by productivity bin. Consistent with part (i) of Proposition 3, GRAIL treatment

effects on agent engagement are decreasing in productivity. To examine (ii), in Figure 5 we

show the rate at which borrowers in each productivity bin defaulted on the program loans.36

TRAIL borrowers in productivity Bin 1 defaulted on 9.3 percent of their loans, whereas GRAIL

Bin 1 borrowers defaulted on a significantly lower 5 percent (p-value = 0.03). In Bins 2 and

36If a borrower did not repay the entire amount due on the pre-arranged repayment date, we record this as adefault.

29

3 the differences in default rates across TRAIL and GRAIL go the other way, although they

are not statistically significant. Finally, (iii) holds when we compare Bin 1 with either Bin 2

or 3, since GRAIL loan repayment rates are higher in Bin 1, while the corresponding informal

interest rates are higher in Bin 1. Default and interest rates do not vary much between Bins

2 and 3, so it is not meaningful to compare across them.

Test of Proposition 4

Table 11 presents estimates of the TRAIL and GRAIL heterogeneous treatment effects on

potato cultivation by ability and also the TRAIL–GRAIL difference in treatment effects for

each ability bin. While the differences are not statistically significant, the results in the last

panel of Table 11 show that TRAIL treated farmers in Bin 1 cultivated a larger area (column

2), produced more output (column 3), had smaller unit costs (column 10) and had higher

value-added (column 7) and imputed profit (column 8) than GRAIL treated farmers in Bin 1.

For Bins 2 and 3, Proposition 4 predicts that TRAIL treated farmers would have smaller

unit costs and would produce more output if treatment effects on area cultivated (weighted

by repayment rate) in the TRAIL scheme are at least as larger as in the GRAIL scheme.

This condition is satisfied in the data.37 Consistent with the predictions of Proposiiton 4, the

TRAIL treated farmers have smaller unit costs; the difference is statistically significant for

Bin 3. TRAIL treatment effects on output, value added and profits exceed GRAIL treatment

effects, although the differences are not statistically significant.

Test of Proposition 5

In our final survey round in 2013, all sample households were asked to participate in a straw

poll. On a sheet of paper resembling a ballot, the respondent was asked to mark the symbol

of their preferred political party and place the paper in a box. We use the response to this

question as our indicator of the political party they support.38 Our working assumption is that

since the GRAIL agents were selected by the local government, they were affiliated with the

37Recall from Figure 5, the repayment rates in Bins 2 and 3 are respectively 93.4% and 94% in TRAIL villagesand 91.8% and 92.7% in the GRAIL villages. So for households in Bins 2 and 3, the treatment effects on areacultivated, weighted by the repayment rate, are always larger in TRAIL scheme than the GRAIL scheme: therepayment weighted acreage treatment effects are 0.084 and 0.1034 for Bins 2 and 3 respectively versus 0.0734and 0.1019.

38Although our surveyors did not look at the respondents’ choice at the time the ballot was cast, respondentsmay have realised that they would look at the choice later and record it. However, surveyors made it clearthat the response would only be used for research purposes and would be kept confidential. Less than 1% ofhouseholds refused to participate, so we have responses from 1194 TRAIL and 1188 GRAIL households.

30

political party in power at the time. A straw poll vote cast in favor of this incumbent party

then indicates that the voter supports the incumbent. We use data from the state election

commission to identify this incumbent political party in each local government constituency.

This allows us to run the regression

Ihv = ξ0 + ξ1Treatment hv + ξ2Control 1hv + γXhv + εhv (25)

The dependent variable Ihv takes the value 1 if the respondent h in village v voted for the

incumbent party in the straw poll. Xhv includes the same set of additional controls that were

included in previous regressions and discussed in Section 6. We run the regression separately

for TRAIL and GRAIL villages. The treatment effect ξ1 − ξ2 indicates whether, conditional

on being recommended by the agent, Treatment households supported the incumbent more

or less than Control 1 households. If they were more likely to support the incumbent, this

suggests that households that benefited from the loan scheme were supporting this party in the

subsequent election. In turn, the agent may have behaved clientelistically and recommended

borrowers who he knew would reward the party in this fashion.

As column 2 of Table 13 shows, the treatment effect is positive and statistically significant in

GRAIL villages. This is consistent with the hypothesis that GRAIL agents were clientelistic

and recommended borrowers who would repay the favor by voting for them. However it is

also possible that the primary motive of the GRAIL agent was to alleviate poverty, and the

households that received the loan then reciprocated by supporting his political party. Both

explanations are also consistent with the finding in Column 4: a Bin 1 borrower who received a

GRAIL loan was 13 percentage points more likely to vote for the incumbent party than a Bin

1 borrower who was also recommended, but did not receive the loan. In contrast the loan had

small and non-significant treatment effects on Bin 2 and Bin 3 borrowers. Again, this could be

because better-off voters are less responsive to such favors, and political parties tend to focus

clientelistic efforts on poorer or less productive voters. Alternatively, the poorer households

tend to reciprocate more when they receive the loan, because the marginal benefit of the loan

is much higher for them.39

Columns 1 and 3 show that the loans did not change voting patterns in TRAIL villages. This

is consistent with the idea that TRAIL agents were primarily non-political individuals and did

not select households on the basis of their voting patterns or propensities.

39Although our model does not specifically predict this, note also that we estimate a positive significantselection effect in GRAIL villages (column 2) that is concentrated among the most productive borrowers (column4). In the clientelism interpretation we might infer that the GRAIL agent recommended some loyalists: personswho might have voted for his party regardless of whether they received the loan. These loyalists appear to havebeen highly productive: either because the GRAIL agent only had networks with highly productive farmers, orbecause he chose the more productive from among his loyalists.

31

9 Conclusion

We briefly summarize our results. First, the average treatment effects in TRAIL exceed those

in GRAIL: this is true for value-added in potato and also on aggregate farm value added.

Our decomposition analysis shows that selection differences provide only a limited explanation

for the superior performance of the TRAIL scheme. Instead, we find that farmers of similar

productivity levels increase their farm incomes by more if they participate in the TRAIL scheme

than in the GRAIL scheme. To understand these differences in the conditional treatment effects

of the two schemes, we focus on the non-scheme private incentives of the two agent types.

TRAIL agents are traders who earn profits by buying and reselling farmers’ crop output. This

gives them an incentive to help farmers produce more. In contrast, the GRAIL agent’s interests

are aligned with that of the village government, and may want the loan scheme to benefit poor

farmers: either because this helps fulfill the poverty alleviation mission of his government, or

for clientelistic reasons: households that gain from the scheme may be more likely to vote for

his party in the subsequent election. The GRAIL agent may then monitor borrowers to ensure

that they repay the loans and maintain good standing within the program.

These differences in the nature of agent-farmer engagement can then explain our findings.

Due to the TRAIL agent’s help, the TRAIL treatment households might use superior farming

techniques and lower production costs, and increase mean output and value-added. The GRAIL

agent’s monitoring might induce the GRAIL treatment farmer to prevent crop failure. This

would ensure high loan repayment rates but would also increase costs and hence prevent him

from increasing value-added. Our theoretical model formalizes these key arguments. These

are then corroborated in our data.

Our paper shows that local community members can play a role that goes beyond benefi-

ciary selection. Driven by their private occupations and incentives that are independent of

the program, the two different types of agents engaged with program beneficiaries in different

ways. This engagement function appears to have been important enough that despite iden-

tical incentives provided within the program, the two schemes led to very different outcomes

for beneficiaries. This underscores the importance of considering the private motivations of

community members when designing decentralized community-driven development programs.

32

References

Alatas, V., A. Banerjee, R. Hanna, B. A. Olken, R. Purnamasari, and M. Wai-Poi (2016). Self-targeting:Evidence from a field experiment in indonesia. Journal of Political Economy 124 (2), 371 – 426.

Alatas, V., A. Banerjee, R. Hanna, B. A. Olken, and J. Tobias (2012). Targeting the poor: Evidence from afield experiment in indonesia. American Economic Review 102 (4), 1206 – 1240.

Athey, S. and G. W. Imbens (2006). Identification and inference in nonlinear difference-in-differences models.Econometrica 74 (2), 431 – 497.

Bandiera, O. and I. Rasul (2006). Social networks and technology adoption in northern mozambique. EconomicJournal 116 (514), 869 – 902.

Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. Jackson (2013). The diffusion of microfinance. Science 341,363 – 372.

Bardhan, P., S. Mitra, D. Mookherjee, and A. Nath (2015). Resource transfers to local governments: Politicalmanipulation and voting patterns in west bengal. Technical report, IED Working Paper, Boston University.

Bardhan, P. and D. Mookherjee (2016). Clientelistic politics and economic development: An overview. Technicalreport, IED Working Paper, Boston University.

Beaman, L., A. BenYishay, J. Magruder, and M. Mobarak (2018). Can network theory-based targeting increasetechnology adoption? Technical report, NBER Working Paper 24912.

Beaman, L. and J. Magruder (2012). Who gets the job referral? evidence from a social networks experiment.American Economic Review 102 (7), 3574 – 3593.

Berg, E., M. Ghatak, R. Manjula, D. Rajasekhar, and S. Roy (2018). Motivating knowledge agents: Canincentive pay overcome social distance? Economic Journal 129 (617), 110 – 142.

Chandrasekhar, A. G., A. Banerjee, E. Duflo, and M. Jackson (2018). Using gossips to spread information:theory and evidence from two randomized controlled trials. Review of Economic Studies Forthcoming.

Debnath, S. and T. Jain (2018). Social connections and public healthcare utilization. Technical report, Mimeo,Indian School of Business.

Devarajan, S. and S. Khemani (2016). If politics is the problem, how can development assistance be part of thesolution? Technical report, World Bank.

Dey, S. and K. Sen (2016). Is partisan alignment electorally rewarding? evidence from village council electionsin india. Technical report, University of Manchester Working Paper.

Fisman, R., D. Paravisini, and V. Vig (2017). Social proximity and loan outcomes: Evidence from an indianbank. American Economic Review 107 (2), 457 – 492.

Hochberg, Y. (1988). A sharper bonferroni procedure for multiple tests of significance. Biometrika 75 (4), 800– 802.

Hussam, R., N. Rigol, and B. Roth (2018). Targeting high ability entrepreneurs using community information:Mechanism design in the field. Mimeo, Harvard Business School.

Kling, J. R., J. B. Liebman, and L. F. Katz (2007). Experimental analysis of neighborhood effects. Economet-rica 75, 83 – 119.

Maitra, P., S. Mitra, D. Mookherjee, A. Motta, and S. Visaria (2017). Financing smallholder agriculture: Anexperiment with agent-intermediated microloans in india. Journal of Development Economics 127, 306 – 337.

Mansuri, G. and V. Rao (2013). Localizing development: Does participation work? Technical report, Washing-ton, DC: World Bank.

33

Mitra, S., D. Mookherjee, M. Torero, and S. Visaria (2018). Asymmetric information and middleman margins:An experiment with west bengal potato farmers. Review of Economics and Statistics C (1), 1 – 13.

Robinson, J. and T. Verdier (2013). The political economy of clientelism. Scandivanian Journal of Eco-nomics 115 (2), 260 – 291.

Stokes, S. (2005). Perverse accountability: A formal model of machine politics with evidence from argentina.Americal Political Science Review 99 (3), 315 – 325.

World Development Report (2004). Making Services Work for Poor People. World Bank, Washington D. C.

34

A1 Appendix: Proof of Proposition 4

In the CE case, we can obtain closed form expressions for area cultivated, help and unit costs

of Bin 1 farmers, as a function of agent monitoring (m(θ) = mG(θ) in GRAIL, zero in TRAIL):

l(θ) =[{(1 + τ)A(θ,m(θ))}1+ν{ νcT}ν{(1− ψ)p(θ,m(θ))rT }−1

(1 +m(θ))−µ2(1−ν)]1

α−ν(1−α) (A1)

[1 + h(θ)]1+ν =[{(1 + τ)A(θ,m(θ))}1+ν{ νcT}α(1+ν){(1− ψ)p(θ,m(θ))rT }−(1−α)(1+ν)

(1 +m(θ))µ2[α(1+ν)−1]]1

α−ν(1−α) (A2)

c(θ) =(1 +m(θ))µ2

α−ν(ν−α)(ν+1)[α−ν(1−α)]K

1(ν+1)[α−ν(1−α)] (A3)

where

K ≡ [ν

cT]αν1+ν [(1− ψ)p(θ,m(θ)rT ](1−α)(1+ν)ν [(1 + τ)A(θ,m(θ))]−(1+ν)ν

The result concerning area cultivated and unit cost comparisons follow from observing that

(A1) is falling and (A3) is rising in m(θ) (the latter holds since α−ν(ν−α) > α−ν(1−α) > 0).

For Bins 2 and 3 the comparison of unit cost is more complicated, owing to the absence of

closed-form expressions. Comparing (17, 23) and using (24), we obtain (1 + hG)−ν−1(1 +

mG)µ2 ≥ (1 + hT )−ν−1, where we suppress the notation for dependence on θ. Hence unit

cost of a GRAIL treated farmer cG = (1 + hG)−ν(1 + mG)µ2 > (1 + hG)−ν(1 + mG)µ2ν1+ν ≥

(1 + hT )−ν = cT . The rest follows from (22).

Finally, the comparison of value added follows from the result that unit cost is lower and

productivity is higher in TRAIL (owing to less monitoring) for any given type, since value

added for any given type is the result of choosing scale l to maximize

Aif(LC + l)− [ρLC + pirT (1− ψ)l]ci (A4)

where Ai, pi, ci denote expected productivity, repayment rate and unit cost in intervention

i = T,G and LC denotes pre-program scale for the given type, and we have AT > AG, pT <

pG, cT < cG.

35

Figure 1: Productivity estimates for Selected v. Non-selected households inTRAIL and GRAIL villages

Notes: The sample is restricted to Control 1 and Control 2 households in TRAIL and GRAIL villages. Productivityis computed using the logarithm of acreage under potato cultivation. The null hypothesis that the distributions ofestimated productivity for Control 1 and Control 2 households are equal is rejected by a two sample Kolmogorov-Smirnov (KS) test with p-value = 0.005 for TRAIL and 0.011 for GRAIL.

36

Figure 2: Productivity estimates for Selected households in TRAIL and GRAILvillages

Notes: The sample is restricted to Control 1 households in TRAIL and GRAIL villages. Productivity is computedusing the logarithm of acreage under potato cultivation. The null hypothesis that the distributions of estimatedproductivity for TRAIL Control 1 and GRAIL Control 1 households are equal is rejected by a two sample Kolmogorov-Smirnov (KS) test with p-value = 0.06.

37

Fig

ure

3:

Inte

rest

Rate

on

Info

rmal

Loan

san

dP

rod

ucti

vit

y

Avera

ge

Info

rmal

Inte

rest

Rate

sby

Pro

du

cti

vit

yB

inV

ari

ati

on

inIn

form

al

Inte

rest

Rate

sby

Pro

du

cti

vit

y

Note

s:T

he

ver

tica

laxis

mea

sure

sth

eaver

age

inte

rest

rate

paid

on

info

rmal

loan

sby

hou

seh

old

s.T

he

hori

zonta

laxis

show

sth

ep

rod

uct

ivit

yes

tim

ate

.In

the

left

pan

el,

we

com

pu

teth

eaver

age

inte

rest

rate

for

hou

seh

old

sin

each

pro

du

ctiv

ity

bin

.T

he

aver

age

inte

rest

rate

paid

on

info

rmal

loan

sby

hou

seh

old

sin

pro

du

ctiv

ity

Bin

1is

sign

ifica

ntl

yh

igh

erth

an

that

paid

by

hou

seh

old

sin

pro

du

ctiv

ity

Bin

2(p

-valu

e=

0.0

3)

an

dp

rod

uct

ivit

yB

in3

(p-v

alu

e=

0.0

4).

Inth

eri

ght

pan

elw

ep

rese

nt

the

loca

lly

wei

ghte

dre

gre

ssio

ns

inof

inte

rest

paid

on

info

rmal

loan

son

pro

du

ctiv

ity.

Th

eaver

age

inte

rest

rate

paid

by

hou

seh

old

sin

pro

du

ctiv

ity

Bin

1is

show

nas

asi

ngle

poin

t.T

he

sam

ple

isre

stri

cted

toC

ontr

ol

1an

dC

ontr

ol

2h

ou

seh

old

sin

TR

AIL

an

dG

RA

ILvilla

ges

wit

hat

most

1.5

acr

esof

lan

d.

Pro

du

ctiv

ity

isco

mp

ute

du

sin

gth

elo

gari

thm

of

the

acr

eage

un

der

pota

tocu

ltiv

ati

on

.

38

Figure 4: Heterogeneous Treatment Effects on Agent Engagement (TRAIL vsGRAIL)

Notes: Agent Engagement measured by the number of times in the past year the responder talked to the agentabout agricultural cultivation related matters. The bars denote the heterogeneous treatment effects by ability bin.The lines denote the bootstrapped 95% confidence interval (2000 iterations). Estimating sample includes Treatmentand Control1 households

39

Figure 5: Default Rates on TRAIL and GRAIL Loans, by Productivity Bin

Notes: The height of each bar measures the fraction of program loans that were not repaid fully by the due date.The sample is restricted to Treatment households in TRAIL and GRAIL villages with at most 1.5 acres of land.Productivity is computed using the logarithm of the acreage under potato cultivation.

40

Table 1: Randomization at the Village Level

TRAIL GRAIL DifferenceMean SD Mean SD TRAIL - GRAIL

(1) (2) (3) (4) (5)

Number of Households 276.04 201.59 252.21 238.36 23.83Number of Potato Cultivators 164.63 130.30 160.75 168.39 3.88Total Landless 15.96 18.98 27.96 75.63 -12.00Total 0− 1.25 113.88 103.22 99.67 78.00 14.21Total 1.25− 2.50 25.58 16.27 24.63 25.20 0.96Total 2.50− 5.00 10.88 7.39 12.83 17.11 -1.96Total 5.00− 12.50 1.38 1.79 1.17 1.95 0.21Total Above 12.50 0.00 0.00 0.04 0.20 -0.04

Notes: Total refers to total number of potato cultivators

41

Table 2: Agent Characteristics

GRAIL TRAIL Difference(1) (2) (3)

Male 1.00 0.958 0.042(0.00) (0.042) (0.042)

SC/ST 0.208 0.083 0.125(0.085) (0.058) (0.102)

Non-Hindu 0.125 0.083 0.042(0.069) (0.058) (0.090)

General caste 0.667 0.833 -0.167(0.098) (0.078) (0.125)

Occupation: Cultivator 0.375 0.042 0.33***(0.101) (0.042) (0.109)

Occupation: Shop/business 0.208 0.958 -0.667***(0.095) (0.042) (0.104)

Occupation: Other 0.417 0.000 0.125*(0.690) (0.000) (0.690)

Owned agricultural land 2.63 3.29 -0.667**(0.198) (0.244) (0.314)

Total owned land 4.08 5.04 -0.958**(0.248) (0.292) (0.383)

Has pucca house 0.375 0.458 -0.083(0.101) (0.104) (0.145)

Educated above primary school 0.958 0.792 0.167*(0.042) (0.085) (0.094)

Weekly income (Rupees) 1102.895 1668.75 -565.855(138.99) (278.16) (336.78)

Village society member 0.292 0.083 0.208*(0.095) (0.058) (0.111)

Party hierarchy member 0.167 0.000 0.167**(0.078) (0.00) (0.079)

Panchayat member 0.125 0.000 0.125*(0.069) (0.00) (0.069)

Self/family ran for village head 0.083 0.000 0.083(0.058) (0.00) (0.058)

Notes: Sample consists of 24 agents in TRAIL villages and 24 agentsin GRAIL villages. Standard errors in parenthesis. ∗∗∗ : p < 0.01;∗∗ :p < 0.05;∗ : p < 0.1.

42

Table 3: Pre-Intervention Interactions with Agent and GP Benefits Received

TRAIL GRAILControl 2 Recommended Control 2 Recommended

(1) (2) (3) (4)

Social and Economic Interactions†

Relationship with Agent: Moneylender 0.17 0.21 0.08 0.10(0.016) (0.019) (0.012) (0.014)

Relationship with Agent: Input supplier 0.18 0.24 0.08 0.09(0.016) (0.020) (0.011) (0.014)

Relationship with Agent: Output buyer 0.18 0.25 0.02 0.04(0.016) (0.020) (0.006) (0.009)

Relationship with Agent: Landlord/tenant 0.01 0.03 0.02 0.02(0.005) (0.008) (0.006) (0.006)

Relationship with Agent: Labour employer 0.12 0.10 0.08 0.07(0.013) (0.014) (0.011) (0.012)

Bought from Agent 0.33 0.41 0.05 0.07(0.020) (0.023) (0.009) (0.012)

Borrowed from Agent 0.14 0.25 0.04 0.09(0.015) (0.020) (0.008) (0.013)

Worked for Agent 0.10 0.13 0.09 0.11(0.012) (0.016) (0.012) (0.014)

Agent and Farmer same social category 0.45 0.53 0.61 0.68(0.021) (0.023) (0.020) (0.022)

Know Agent 0.90 0.96 0.91 0.97(0.013) (0.009) (0.012) (0.008)

Meet Agent at least once a week 0.98 1.00 0.98 1.00(0.007) (0.002) (0.006) (0.003)

Invited by Agent on special occasions 0.31 0.45 0.27 0.39(0.020) (0.024) (0.019) (0.023)

Sold Potato to Agent (% of Transactions) 0.109 0.171 0.020 0.027(0.312) (0.377) (0.142) (0.163)

Sold Potato to Agent for more than one year♣ 0.870 0.877 0.556 0.615(0.339) (0.331) (0.527) (0.506)

GP Benefits Received‡

Housing 0.01 0.03 0.05 0.02(0.004) (0.007) (0.008) (0.007)

BPL card 0.14 0.16 0.11 0.13(0.013) (0.017) (0.012) (0.015)

NREGA employment 0.30 0.41 0.37 0.48(0.017) (0.023) (0.018) (0.023)

IRDP loan 0.00 0.00 0.00 0.00(0.000) (0.002) (0.002) (0.003)

Minikit 0.04 0.06 0.03 0.04(0.007) (0.011) (0.006) (0.009)

Toilet 0.08 0.10 0.13 0.16(0.010) (0.014) (0.013) (0.017)

Drinking water 0.18 0.21 0.26 0.26(0.014) (0.018) (0.016) (0.020)

Medical help 0.14 0.14 0.17 0.19(0.013) (0.016) (0.014) (0.018)

Notes: Recommended households include Treatment and Control 1 households. Sample restricted to allhouseholds with 1.5 acres of land in TRAIL and GRAIL villages. †: Economic and Social Relations as of2010 (pre-intervention). Agent and Farmer same social category denotes the farmer belongs to same thereligion category (Hindu/Non-Hindu) or caste category (Scheduled caste/Scheduled tribe, or General)as the agent. ‡: GP benefits received in the past 3 years as of 2010. ♣: Sold to agent for more than 1year conditional on selling to the agent.

Table 4: Household Characteristics and RandomisationCheck

TRAIL GRAIL TRAIL–GRAIL(1) (2) (3)

Head: More than Primary School 0.407 0.420 -0.013(0.015) (0.015)

Head: Cultivator 0.441 0.415 0.026(0.015) (0.015)

Head: Labourer 0.340 0.343 -0.003(0.015) (0.015)

Area of house and homestead (Acres) 0.052 0.052 0.000(0.001) (0.002)

Separate toilet in house 0.564 0.608 -0.044(0.015) (0.015)

Landholding (Acres) 0.456 0.443 0.013(0.013) (0.013)

Own a motorized vehicle 0.124 0.126 -0.002(0.010) (0.010)

Own a Savings Bank Account 0.447 0.475 -0.028(0.015) (0.015)

F-test of joint significance (p-value) 0.996

Notes: Sample uses household survey data and restricts the sample to house-holds with atmost 1.5 acres of land in TRAIL and GRAIL villages. Standarderrors in parenthesis. ∗∗∗ : p < 0.01;∗∗ : p < 0.05;∗ : p < 0.1.

44

Table 5: Baseline Credit Market characteristics

All Loans Agricultural Loans(1) (2)

Household had borrowed 0.67 0.59Total Borrowing† 6352 (10421) 5054 (8776)

Proportion of Loans by Source‡

Traders/Money Lenders 0.63 0.66Family and Friends 0.05 0.02Cooperatives 0.24 0.25Government Banks 0.05 0.05MFI and Other Sources 0.03 0.02

Annualized Interest Rate by Source (percent)Traders/Money Lenders 24.93 (20.36) 25.19 (21.47)Family and Friends 21.28 (14.12) 22.66 (16.50)Cooperatives 15.51 (3.83) 15.70 (2.97)Government Banks 11.33 (4.63) 11.87 (4.57)MFI and Other Sources 37.26 (21.64) 34.38 (25.79)

Duration by Source (days)Traders/Money Lenders 125.08 (34.05) 122.80 (22.43)Family and Friends 164.08 (97.40) 183.70 (104.25)Cooperatives 323.34 (90.97) 327.25 (87.74)Government Banks 271.86 (121.04) 324.67 (91.49)MFI and Other Sources 238.03 (144.12) 272.80 (128.48)

Proportion of Loans Collateralized by SourceTraders/Money Lenders 0.02 0.01Family and Friends 0.04 0.07Cooperatives 0.79 0.78Government Banks 0.81 0.83MFI and Other Sources 0.01 0.01

Notes:Statistics are reported for all sample households in TRAIL andGRAIL villages with at most 1.5 acres of land. All characteristicsare for loans taken by the households in Cycle 1. Program loans arenot included. For the interest rate summary statistics loans wherethe principal amount is reported equal to the repayment amount arenot included. To arrive at representative estimates for the studyarea, Treatment and Control 1 households are assigned a weight of30N

and Control 2 households are assigned a weight of N−30N

, were N

is the total number of households in their village. †: Total borrowing= 0 for households that do not borrow. ‡: Proportion of loans interms of value of loans at the household level. All proportions arecomputed only over households that borrowed. Standard deviationsare in parentheses.

45

Table 6: Treatment Effects on Borrowing

All Agricultural Loans Non Program Agricultural Loans†

(Rs) (Rs)(1) (2)

TRAIL Treatment 7517*** -385.9(817.6) (653.8)

Mean TRAIL C1 5590 5590% Effect TRAIL 134.472 -6.903

GRAIL Treatment 7341*** -250.7(816.3) (570.4)

Mean GRAIL C1 5016 5016% Effect GRAIL 146.352 -4.998

TRAIL v. GRAIL Treatment 176.5 -135.2(1162) (869.5)

Sample Size 6,156 6,156

Notes: Treatment effects are computed from regressions that follow equation (1) in the text and arerun on household-year level data for all sample households in TRAIL and GRAIL villages with at most1.5 acres of land. Regressions also control for the religion and caste of the household, age, educationalattainment and occupation of the eldest male member of the household, household’s landholding, aset of year dummies and an information village dummy. % Effect: Treatment effect as a percentageof the mean of the relevant Control 1 group. †: Non-Program loans refer to loans from sources otherthan the TRAIL or GRAIL schemes. Standard errors in parentheses are clustered at the hamlet level.∗∗∗ : p < 0.01,∗∗ : p < 0.05,∗ : p < 0.1.

46

Tab

le7:

Avera

ge

Tre

atm

ent

Eff

ects

inP

ota

toC

ult

ivati

on

Cu

ltiv

ate

Acr

eage

Pro

du

cton

Cost

of

Pro

du

ctio

nR

even

ue

Valu

eA

dd

edIm

pu

ted

Pro

fit

Ind

exV

ari

ab

le(a

cres

)(k

g)

(Rs)

(Rs)

(Rs)

(Rs)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

TR

AIL

Tre

atm

ent

0.0

40

0.0

93***

946.3

68***

1846.0

25**

3898.4

01***

2059.2

35***

1906.6

96***

0.1

80***

(0.0

28)

(0.0

28)

(302.2

39)

(722.5

61)

(1204.2

71)

(631.2

41)

(600.5

95)

(0.0

54)

Mea

nT

RA

ILC

10.7

15

0.3

36

3646.1

24

8481.7

51

14285.4

67

5732.3

57

4733.7

70

%E

ffec

tT

RA

IL5.6

327.5

825.9

621.7

627.2

935.9

240.2

8

GR

AIL

Tre

atm

ent

0.1

30***

0.0

69**

770.9

16**

2007.1

17**

2500.9

96*

492.3

95

190.2

92

0.1

42**

(0.0

29)

(0.0

29)

(340.0

84)

(779.8

52)

(1377.8

91)

(744.3

49)

(720.0

68)

(0.0

62)

Mea

nG

RA

ILC

10.6

41

0.2

96

3236.7

30

7070.6

53

12964.7

76

5827.6

63

4941.5

11

%E

ffec

tG

RA

IL20.3

323.2

523.8

228.3

919.2

98.4

53.8

5

TR

AIL

v.

GR

AIL

Tre

atm

ent

-0.0

90**

0.0

24

175.4

52

-161.0

92

1397.4

05

1566.8

39

1716.4

04*

0.0

38

(0.0

40)

(0.0

41)

(457.8

36)

(1067.9

44)

(1838.4

10)

(979.8

76)

(941.2

79)

(0.0

83)

Sam

ple

Siz

e6,1

50

6,1

50

6,1

50

6,1

50

6,1

50

6,1

50

6,1

50

Note

s:T

reatm

ent

effec

tsare

com

pu

ted

from

regre

ssio

ns

that

follow

equ

ati

on

(1)

inth

ete

xt

an

dare

run

on

hou

seh

old

-yea

rle

vel

data

for

all

sam

ple

hou

seh

old

sin

TR

AIL

an

dG

RA

ILvilla

ges

wit

hat

most

1.5

acr

esof

lan

d.

Reg

ress

ion

sals

oco

ntr

ol

for

the

religio

nan

dca

ste

of

the

hou

seh

old

,age,

edu

cati

on

al

att

ain

men

tan

docc

up

ati

on

of

the

eld

est

male

mem

ber

of

the

hou

seh

old

,h

ou

seh

old

’sla

nd

hold

ing,

ase

tof

yea

rd

um

mie

san

dan

info

rmati

on

villa

ge

du

mm

y.%

Eff

ect:

Tre

atm

ent

effec

tas

ap

erce

nta

ge

of

the

mea

nof

the

rele

vant

Contr

ol

1gro

up

.Im

pu

ted

pro

fit

=V

alu

eA

dd

ed–

shad

ow

cost

of

lab

ou

r.In

colu

mn

8,

the

dep

end

ent

vari

ab

leis

an

ind

exof

z-sc

ore

sof

the

ou

tcom

evari

ab

les

inth

ep

an

el;

the

p-v

alu

esfo

rtr

eatm

ent

effec

tsin

this

colu

mn

are

com

pu

ted

acc

ord

ing

toH

och

ber

g(1

988)’

sst

ep-u

pm

eth

od

toco

ntr

ol

for

the

fam

ily-w

eighte

der

ror

rate

acr

oss

all

ind

exou

tcom

es.

Sta

nd

ard

erro

rsin

pare

nth

eses

are

clu

ster

edat

the

ham

let

level

.∗∗∗

:p<

0.0

1,∗∗

:p<

0.0

5,∗

:p<

0.1

.

47

Table 8: Average Treatment Effects on Farm Value Added, Non Agri-cultural Income and Total Household Income

Farm Value Added Non Agricultural Income Household Income(Rs) (Rs) (Rs)(1) (2) (3)

TRAIL Treatment 2147.874*** 1506.785 5612.479(723.935) (3655.543) (4125.591)

Mean TRAIL C1 10336.511 33617.764 49765.238% Effect TRAIL 20.779 4.482 11.278

GRAIL Treatment 138.905 -4326.522 -5434.789(929.675) (3225.654) (4496.421)

Mean GRAIL C1 10501.064 37170.751 56082.489% Effect GRAIL 1.323 -11.640 -9.691

TRAIL v. GRAIL Treatment 2008.969* 5833.307 11047.268*(1183.349) (4879.593) (6110.838)

Sample Size 6,156 6,159 6,156

Notes: Treatment effects are computed from regressions that follow equation (1) in the text and arerun on household-year level data for all sample households in TRAIL and GRAIL villages with at most1.5 acres of land. Regressions also control for the religion and caste of the household, age, educationalattainment and occupation of the eldest male member of the household, household’s landholding, aset of year dummies and an information village dummy. % Effect: Treatment effect as a percentage ofthe mean of the relevant Control 1 group. Standard errors in parentheses are clustered at the hamletlevel. ∗∗∗ : p < 0.01,∗∗ : p < 0.05,∗ : p < 0.1.

48

Table 9: Loan Performance

Continuation Default(1) (2)

Panel A: Sample Means

TRAIL 0.937 0.070(0.006) (0.007)

GRAIL 0.872 0.070(0.009) (0.008)

Difference (TRAIL–GRAIL) 0.065*** 0.000(0.011) (0.010)

Panel B: Regression Results

GRAIL -0.066 *** 0.005(0.011) (0.010)

R2 0.08 0.06Sample Size 2667 2422

Notes: The sample consists of household-cycle level ob-servations of Treatment households in TRAIL and GRAILvillages with at most 1.5 acres of landholding. The depen-dent variable in column 1 takes value 1 if the householdtook the program loan. The dependent variable in column2 takes value 1 if a borrowing household repays at least50% the amount due on the loan taken in that cycle onthe due date. In Panel B, the treatment effects are com-puted from regressions that follow equation 2 in the text.Regressions also control for landholding, religion and casteof the household and age and educational attainment ofthe oldest male in the household. Robust standard errors.Significance ∗∗∗ : p < 0.01,∗∗ : p < 0.05,∗ : p < 0.1.

49

Table 10: Variation of Farm Productivity with Observable HouseholdCharacteristics

Dependent Variable: farm productivity

Regression Results (1) Descriptive Statistics (2)

Landholding 1.559*** TRAIL(0.491)

Non Hindu -0.999** % Cultivators 72.65(0.429) Mean of Productivity (Cultivators) 1.670

Low Caste -1.055*** SD of Productivity (Cultivators) 1.109(0.278) Minimum Productivity (Cultivators) -1.294

Household Size -0.093 Productivity Quartile 25% 0.786(0.074) Productivity Quartile 50% 1.937

Female Headed Household -1.443** Productivity Quartile 75% 2.482(0.568) Maximum Productivity (Cultivators) 3.702

Age of oldest Male -0.004(0.011) GRAIL

Oldest Male: Completed Primary School 0.146(0.287) % Cultivators 65.53

Constant 0.469 Mean of Productivity (Cultivators) 1.602(0.672) SD of Productivity (Cultivators) 1.219

Minimum Productivity (Cultivators) -1.111Sample Size 464 Productivity Quartile 25% 0.625R-squared 0.184 Productivity Quartile 50% 1.885

Productivity Quartile 75% 2.628Maximum Productivity (Cultivators) 3.651

Notes: Standard errors in parenthesis clustered at the hamlet level. Estimating sample includes Control1 households in TRAIL and GRAIL villages with at most 1.5 acres of land. Farm Productivity computedusing log(acreage under potato cultivation). Significance ∗∗∗ : p < 0.01,∗∗ : p < 0.05,∗ : p < 0.1. Harmers inour sample who either did not plant potatoes even once in the three years of our study period, or plantedthem only once in three years are categorized as non-cultivators. See discussion in Section 6.1.

50

Tab

le11:

Hete

rogen

eou

sT

reatm

ent

Eff

ects

.P

ota

toC

ult

ivati

on

on

ly

Cu

ltiv

ate

Acr

eage

Ou

tpu

tC

ost

(Tota

l)P

rice

Rev

enu

eV

alu

eA

dd

edIm

pu

ted

Pro

fit

Inp

ut

Cost

Inp

ut

Cost

(Per

acr

e)(0

/1)

(Acr

es)

(Kg)

(Rs)

(Rs/

Kg)

(Rs)

(Rs)

(Rs)

(Rs)

(Rs/

Acr

e)(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)(1

0)

Pan

elA

:T

RA

ILB

in1

0.0

8†

0.0

4†

383.8

2696.4

60.2

71740.2

41040.4

3930.0

21404.3

1-1

701.3

4(0

.03)

(0.0

2)

(179.6

6)

(444.7

1)

(0.1

9)

(799.8

2)

(461.0

1)

(439.5

9)

(867.5

8)

(5217.8

5)

Bin

20.0

10.0

9†

909.2

8†

2132.5

4†

-0.1

33701.2

2†

1561.1

9†

1458.5

8†

3629.6

6†

-2320.5

3(0

.02)

(0.0

3)

(274.2

5)

(638.2

0)

(0.1

1)

(1116.5

0)

(616.8

8)

(590.2

0)

(1380.7

8)

(1624.2

9)

Bin

30.0

00.1

1†

1126.3

8†

1814.6

80.0

34617.7

6†

2834.0

5†

2668.1

6†

2614.4

0-3

737.3

4†

(0.0

0)

(0.0

4)

(488.6

2)

(1252.2

5)

(0.1

6)

(2127.1

2)

(1296.7

0)

(1268.4

9)

(2162.7

3)

(1334.2

6)

Pan

elB

:G

RA

ILB

in1

0.2

6†

0.0

3†

287.0

4892.4

0†

-0.2

1923.1

330.3

7-2

65.0

81882.2

0†

6768.5

5†

(0.0

4)

(0.0

2)

(191.7

1)

(392.0

2)

(0.1

6)

(754.4

2)

(436.8

2)

(419.1

7)

(782.3

4)

(2949.5

9)

Bin

20.1

2†

0.0

8†

844.4

0†

2461.0

4†

-0.2

73022.0

3†

551.2

2301.5

54002.0

5†

-1881.2

6(0

.02)

(0.0

3)

(277.3

9)

(622.1

8)

(0.2

2)

(1214.5

3)

(776.0

4)

(752.0

9)

(1350.7

0)

(1708.2

3)

Bin

30.0

00.1

1†

1418.7

4†

3210.2

0†

-0.1

24499.3

8†

1291.4

1900.7

36777.3

2†

1552.8

7(0

.00)

(0.0

5)

(570.7

3)

(1254.9

1)

(0.1

0)

(2438.5

5)

(1705.7

8)

(1713.9

7)

(2271.2

6)

(1561.1

8)

Pan

elC

:T

RA

IL-

GR

AIL

Diff

eren

ceB

in1

-0.1

8†

0.0

196.7

9-1

95.9

40.4

7†

817.1

11010.0

61195.1

0-4

77.8

9-8

469.9

0(0

.05)

(0.0

2)

(269.9

0)

(608.3

0)

(0.2

5)

(1135.2

5)

(652.1

5)

(622.9

4)

(1197.9

4)

(5981.1

3)

Bin

2-0

.11†

0.0

064.8

8-3

28.4

90.1

4679.1

91009.9

71157.0

3-3

72.3

8-4

39.2

7(0

.03)

(0.0

4)

(388.5

8)

(887.6

3)

(0.2

4)

(1650.4

3)

(997.5

1)

(962.3

1)

(1921.0

3)

(2374.9

4)

Bin

30.0

00.0

0-2

92.3

6-1

395.5

20.1

6118.3

81542.6

31767.4

4-4

162.9

2-5

290.2

1†

(0.0

0)

(0.0

6)

(751.6

2)

(1773.0

3)

(0.1

9)

(3235.0

1)

(2147.6

1)

(2137.3

8)

(3134.6

1)

(2061.1

9)

Sam

ple

Siz

e6,1

50

6,1

50

6,1

50

6,1

50

3,8

20

6,1

50

6,1

50

6,1

50

6,1

50

4,0

38

Note

s:T

he

tab

lep

rese

nts

het

erogen

eou

str

eatm

ent

effec

tsof

acr

eage,

reven

ue

an

dvalu

ead

ded

inp

ota

tocu

ltiv

ati

on

by

pro

du

ctiv

ity

bin

.T

he

esti

mati

ng

equ

ati

on

follow

seq

uati

on

(5)

inth

ete

xt.

Th

ees

tim

ati

ng

sam

ple

incl

ud

esall

sam

ple

hou

seh

old

sin

TR

AIL

an

dG

RA

ILvilla

ges

wit

hat

most

1.5

acr

esof

lan

d.

Th

ed

epen

den

tvari

ab

les

inco

lum

ns

(1)–

(9)

take

the

act

ual

valu

ere

port

edby

the

hou

seh

old

ifit

did

,or

take

the

valu

eze

roif

itd

idn

ot

cult

ivate

pota

toes

inth

at

yea

r.In

colu

mn

(10),

hou

seh

old

sth

at

did

not

cult

ivate

pota

toes

ina

yea

rare

dro

pp

edfr

om

the

esti

mati

ng

sam

ple

.T

he

regre

ssio

ns

als

oco

ntr

ol

for

the

religio

nan

dca

ste

of

the

hou

seh

old

,age,

edu

cati

on

al

att

ain

men

tan

docc

up

ati

on

of

the

eld

est

male

mem

ber

of

the

hou

seh

old

,h

ou

seh

old

’sla

nd

hold

ing,

ase

tof

yea

rd

um

mie

san

dan

info

rmati

on

villa

ge

du

mm

y.P

rod

uct

ivit

yis

com

pu

ted

usi

ng

the

logari

thm

of

the

acr

eage

un

der

pota

tocu

ltiv

ati

on

.S

tan

dard

erro

rsin

pare

nth

eses

are

clu

ster

edat

the

ham

let

level

.† :

Boots

trap

ped

90%

con

fid

ence

inte

rval

(2000

iter

ati

on

s)d

oes

not

incl

ud

e0.

51

Tab

le12:

Decom

posi

tion

of

AT

ED

iffere

nces

inV

alu

e-a

dd

ed

from

Pota

toC

ult

ivati

on

;T

RA

ILv.

GR

AIL

σT k

σG k

σT k−σG k

TR

AIL

HT

Es

GR

AIL

HT

Es

TR

AIL

–G

RA

ILH

TE

s(σT k−σG k

)×TT

σG k×

(TT−TG

)(TT k

)(TG k

)(TT k−TG k

)(1

)(2

)(3

)(4

)(5

)(6

)(7

=(3

)×

(4))

(8=

(2)×

(6))

Bin

10.2

70.3

4-0

.07

1040.4

30.4

1010.1

-74.1

348.2

Bin

20.3

30.3

30.0

01561.2

551.2

1010.0

-4.5

335.2

Bin

30.4

00.3

20.0

72834.1

1291.4

1542.6

209.8

498.9

AT

E2059.2

492.4

1566.8

%of

AT

Ed

ue

toS

elec

tion

8.3

8%

of

AT

Ed

ue

toC

TE

75.4

6

Note

s:σT k

den

ote

sth

ep

rop

ort

ion

of

TR

AIL

hou

seh

old

sin

bink;σG k

den

ote

sth

ep

rop

ort

ion

of

GR

AIL

hou

seh

old

sin

bink.

52

Tab

le13:

Eff

ect

of

Tre

atm

ent

on

Voti

ng

Patt

ern

sin

Str

aw

Poll

Average

Treatm

ent

Eff

ect

Hete

rogen

eou

sT

reatm

ent

Eff

ect

TR

AIL

GR

AIL

TR

AIL

GR

AIL

(1)

(2)

(3)

(4)

Tre

atm

ent

Eff

ect

0.0

241

0.0

782**

(0.0

496)

(0.0

340)

Tre

atm

ent

Eff

ect:

Ab

ilit

yB

in1

0.0

915

0.1

30†

(0.0

868)

(0.0

697)

Tre

atm

ent

Eff

ect:

Ab

ilit

yB

in2

-0.0

741

0.0

309

(0.0

805)

(0.0

702)

Tre

atm

ent

Eff

ect:

Ab

ilit

yB

in3

0.0

568

0.0

135

(0.0

564)

(0.0

743)

Sel

ecti

on

Eff

ect

-0.0

649

0.0

825**

(0.0

447)

(0.0

369)

Sel

ecti

on

Eff

ect:

Bin

1-0

.133

0.0

217

(0.0

610)

(0.0

580)

Sel

ecti

on

Eff

ect:

Bin

2-0

.0291

0.1

17†

(0.0

738)

(0.0

664)

Sel

ecti

on

Eff

ect:

Bin

3-0

.0343

0.1

05†

(0.0

594)

(0.0

718)

Sam

ple

Siz

e1,0

11

1,0

26

1,0

21

1,0

44

Note

s:E

stim

ati

ng

sam

ple

incl

ud

esall

sam

ple

hou

seh

old

sin

TR

AIL

an

dG

RA

ILvil

lages

wit

hatm

ost

1.5

acr

esof

lan

d.

OL

Sre

gre

ssio

nre

sult

sp

rese

nte

d.

Sta

nd

ard

erro

rsin

pare

nth

esis

clu

ster

edat

the

ham

let

level

.E

stim

ati

ng

sam

ple

incl

ud

esall

sam

ple

hou

seh

old

sin

TR

AIL

an

dG

RA

ILvilla

ges

wit

hat

most

1.5

acr

esof

lan

d.

Ab

ilit

yco

mp

ute

du

sin

glo

g(a

crea

ge

un

der

pota

tocu

ltiv

ati

on

).S

ign

ifica

nce

∗∗∗

:p<

0.0

1,∗∗

:p<

0.0

5,∗

:p<

0.1

.†

:B

oots

trap

ped

con

fid

ence

inte

rval

wit

h2000

iter

ati

on

sd

oes

not

incl

ud

e0.

53

Figure A1: Percentage of households in each Productivity Bin

Notes: The height of the bars denote the fraction of households in each productivity Bin. Productivityis computed using the logarithm of acreage under potato cultivation.

54