financing smallholder agriculture: an experiment with ... · motivation presentation outline 1...

Financing Smallholder Agriculture:An Experiment with Agent-Intermediated Microloans in

India

Pushkar Maitra, Sandip Mitra, Dilip Mookherjee, Alberto Motta and SujataVisaria

Presentation

June 2014

MMMMV (June 2014) Financing Smallholder Agriculture June 2014 1 / 60

Motivation

Financial Inclusion of Poor Farmers: The Challenge

Institutional finance does not reach smallholder agriculture owing to:

lack of collateralpoor information and enforcement capacityhigh transaction costs

Poor farmers are forced to rely on credit from informal lenders at high cost(owing either to high capital costs for informal lenders, or usury, or both)

Prevents poor farmers from escaping poverty by diversifying into high-valuecash crops

Restricts agricultural growth

Major development challenge: can financial institutions lend profitably toproductive small farmers?


Motivation

Microcredit: Does it Solve the Problem?

Microfinance has made impressive inroads in providing credit to the poor,while attaining high repayment rates

But it has not succeeded in financing productive needs in agriculture (inIndia)

Recent RCT-based evaluations of micro finance (Banerjee et al 2011, Fischer2013, Karlan and Zinman 2011) indicate it has had very limited success inpromoting entrepreneurship, risk-taking or borrower incomes

Mainly successful in allowing the poor to smooth consumption, acquireconsumer durables (Karlan and Zinman 2011, Islam and Maitra 2012,Banerjee et al 2013)


Motivation

Why Not?

Traditional microcredit is based on

high frequency (weekly-biweekly-monthly) repayment schedulesintensive peer-cum-MFI monitoring to limit risk-taking

Motivated by need to ensure high repayment rates

But these rule out scope for financing agriculture (crop cycles ≥ 3 months,high-risk)

Microfinance under the pump (AP crisis in India 2008-09)

Reserve Bank of India is urging development of direct lending by banks to poorborrowers in rural areas


Motivation

The Fundamental Problem

Key ingredients of traditional MF (intensive monitoring, peer pressure, limits onrisky choices, high frequency repayments) are viewed as necessary for selectionand repayment incentives for the poor.

Is there an inevitable trade-off between adequate selection/repayment incentivesand financing productive and risky activities?

How to ensure appropriate borrower selection and repayment incentives?

Can this be fiscally sustainable for lending institutions?


Motivation

Alternative Approaches

We design and implement a new approach to financing smallholderagriculture in two potato-growing districts of West Bengal, India

Distinctive features:

4 month repayment periods (timed according to crop cycles)built-in crop insuranceminimal monitoring by lender‘Low’ (18% p.a.) interest ratedynamic repayment incentives


Motivation

Alternative Approaches, contd.

We compare two different methods of selection/repayment incentives:

TRAIL: appointing a local agent from within the local community,recommends borrowers to receive individual liability loans

GBL: inviting five-member borrower groups to self-form, meetmonthly with MFI officials, attain savings targets, and applyfor joint liability loan

Within each method, we randomly select from those recommend/eligible andoffer them the corresponding loans

Survey households to evaluate loan impacts and selection patterns in TRAILand GBL


Motivation

Presentation Outline

1 Field Experiment Details

2 Theoretical Analysis and Testable Predictions

3 Impacts on Cultivation, Incomes and Rates of Return

4 Financial Performance: Take-up, Repayment rates, Administrative Costs

5 Conclusion


Experiment

The Field Experiment

We worked with a Kolkata based MFI to implement two treatments:

TRAIL: 24 villagesGBL: 24 villages

Design loans to facilitate financing of potato cultivation: high-value high-riskcash crop, in West Bengal

Work in two potato growing districts (Hugli, West Medinipur), select 48villages randomly in these districts, divided randomly between TRAIL andGBL


Experiment

Location


Experiment

Loan Features: Similarity and Differences

Similarity

interest rate (18%)

durations (4 month)

dynamic incentives (credit line in next cycle = 133% of current loan repaid)

index insurance

Differences:

TRAIL: loans are individual liability

GBL: joint liability and loan renewals are contingent on group repayment


Experiment

Random Allocation of Loans:

TRAIL: 1 agent per TRAIL village, 30 borrowers recommended byagent; loans given to 10 randomly chosen from this list

GBL: Groups organized as per MFI protocol: 5 member groupsself-form, meet fortnightly with MFI officials, each borrowerrequired to save Rs. 50 per month;2 groups out of those that survived on cut-off date selectedrandomly


Experiment

Agents’ Incentives to Select and/or Monitor Borrowers

Commission: portion of interest paid goes to agent (= 75% in experiment)

Bonus: payable at end of two years conditional on adequate repayment recordof recommended clients

Enhanced reputation within village: useful way to expand trade/market shares


Experiment

Possible Abuse of Power by Agents

Agents could:

charge high interest rates (if permitted or through kickbacks)

select cronies

select unprofitable clients

select non-poor clients

select those willing to pay bribes

extract borrower benefits by manipulating other contractual relationships

recommend non-repayment of loans to borrowers and divide up loan funds

coerce borrowers to repay


Experiment

Features to Limit Possible Abuse of Power

MFI lends directly to client rather than through the agent

Agent only recommends clients, MFI officials lend and collect payments

Can recommend only landless and marginal landowners (owning ≤ 1.5 acres)

Interest rate is pegged

Agent forfeits deposit he posted upfront if any client does not repay

Agent is terminated if less than 50% loans are repaid in any cycle

Not every household recommended by agent receives the loan(randomization)


Experiment

Other Important Features

Enhanced convenience for borrowers

No savings requirements or mandated meetings with agents or MFI officials

Door-step banking, no need to open a formal bank account

MFI monitors less than in traditional microcredit, so administrative costs are lower

Allows lower interest rates

Allows payment of commission to agents


Experiment

The Data

Household sample: 50 per village = 10 loan recipients (Treatment) + 10eligible non-recipients (Control 1) + 30 non-eligible (Control 2)

Two control groups allow separate identification of loan treatment andselection effects

Household data collected through 4-monthly surveys over 8 cycles (October2010- December 2012)

Matched to loan record data from MFI


Theory

Theory

Use simple model based on network effects

Abstract from adverse selection, moral hazard or credit rationing

Models based on adverse selection or moral hazard generate qualitativelysimilar results


Theory

Theoretical Framework

Segmented credit market within each village

Borrowers are classified into two categories:

Connected (C)Floating (F)

All lenders and connected borrowers are partitioned into different networks(caste/kinship/neighborhood groups with high social capital)

Each network consists of lenders and borrowers who behave in a cooperativefashion (i.e., maximize aggregate payoffs of network members)Share useful production and marketing information that raises farmproductivity: their projects have a high probability of success pc

Floating borrowers belong to groups with low social capital: behavenon-cooperatively, receive less help: projects have low probability pf ofsuccess, where pf (2− pf ) < pc


Theory

Loans

Borrowers borrow l ≥ 0 to finance cultivation at scale l

Revenue fi (l) conditional on success with f ′i > 0, fi ” < 0 (revenue = 0 iffailure) and f ′C (l) ≥ f ′F (l) for all l .

Lenders have cost of capital ρI , no capacity constraints; can lend to anyonein the village

Limited liability and no moral hazard: loans are repaid whenever projectssucceed

Lenders from different networks compete a la Bertrand in lending toout-of-network borrowers


Theory

Equilibrium of Informal Credit Market

No borrowing or lending across networks (since all networks are identical)

Network borrowers borrow from lenders within own-network

Select lcI cooperatively to maximize {pc fC (l)− ρI l} = pc (fC (l)− ρI

pcl)

Floating borrowers borrow at break-even interest rate ρpf

Select l fI non-cooperatively to maximize {pf fF (l)− ρI l} = pf (fF (l)− ρI

pfl)

l fI < lcI since ECC is lower for network borrowers ( ρI

pc< ρI

pf)


Theory

Agent-Intermediated Lending: TRAIL

MFI enters, select an agent randomly from set of network lenders, who isasked to recommend n borrowers

MFI offers TRAIL loans at interest rate rT (< ρI ) to m(< n) borrowersrandomly chosen from those recommended

Assume dynamic repayment incentives suffice to ensure all borrowers repayMFI whenever projects succeed (i.e., they are patient enough)

Agent commission: proportion K ∈ (0, 1) of repaid interest


Theory

Group-based Lending: GBL

Groups of size two self-form to apply for joint liability loan

Assume both loans are repaid iff at least one borrower’s project succeeds

Assortative matching may or may not arise, but mixed groups will be similarto homogenous groups

A type (i , i) group will select a loan l iG for each to maximizepi fi (l)− pi (2− pi )rT l

ECC in GBL is pi (2− pi )rT , higher than in TRAIL (rT ) owing to jointliability tax

Borrowers have to attend group meetings, make regular savings to qualify fora group loan: imposes additional cost γi for a borrower of type i ∈ {c , f }.


Theory

Agent-Intermediated Lending: TRAIL

Agent prefers own-network borrower to other-network borrower

own-network borrower chooses loansize to maximize entire network profit

Agent prefers other-network borrower to floating borrower

Repayment rate is higher for any network borrower


Theory

TRAIL Selection: Proposition

Under TRAIL, an agent who is a network lender prefers to recommend anown-network borrower, if either:

external borrowers cannot bribe the agent, or

they can bribe the agent, and the commission rate K is large enough


Theory

TRAIL Selection Incentives: Explanation

In absence of any bribes, agent prefers to recommend own-network lenderrather than floating borrower for two reasons:

(a) Higher probability of loan repayment/commission(b) Internalization of benefits accruing to own-networkborrower

If floating borrower pays bribe, reason (b) disappears (best case scenario) but(a) remains

Agent also prefers own-network borrower to other-network borrower owing to(b)

The best he can do is to extract all possible surplus from other-networkborrower: gives him same profit as in the case of own-network borrowerIf agent’s bargaining power is any lower he cannot extract all profit and preferown-network borrower


Theory

GBL Selection

Both (F ,F ) and (C ,C ) groups likely to gain from joining GBL, ignoringcosts of group meetings and savings targets

For (F ,F ): ECC drops from ρI

pfto (2− pf )rT

For (C ,C ): ECC drops from ρI

pcto (2− pc )rT

Cannot determine a priori which group gains more

Hence GBL likely to contain both (C ,C ) and (F ,F ) groups

No mechanism in GBL to exclude either type (unless group meetings andsavings targets are more costly for (F ,F ))


Theory

Table: Summary of Theoretical Predictions

Treatment Composition Observed Repayment EffectiveC=connected Interest rate Rate Cost of Credit

F=floaters

Treatment C rT pc (1− K)rT

TRAIL Control 1 CρIpc

pcρIpc

Control 2 C, FρIpc

,ρIpf

pc , pfρIpc

,ρIpf

Treatment CC, FF rT pc (2− pc ), pf (2− pf ) (2− pc )rT , (2− pf )rT

GBL Control 1 CC, FFρIpc,ρIpf

pc , pfρIpc,ρIpf

Control 2 C, FρIpc,ρIpf

pc , pfρIpc,ρIpf


Theory

Empirical Predictions

Borrower Selection: TRAIL selection patterns will be biased in favor of borrowerswho:

pay lower interest rates on the informal markethave higher productivity

Impact on Cultivation and Incomes: TRAIL borrowers will borrow andcultivate high-value crops more, earn higher increases in farmincomeGreater dispersion of loan sizes and treatment effects in GBLcompared to TRAIL


Theory

Predictions: Repayment and Take-up rates

Repayment rate comparisons: Ambiguous (opposing effects of differentialselection and group insurance in GBL)

Loan Take-up rate: Expect higher loan take up in TRAIL owing to avoidance ofjoint liability tax, burden of group meetings and savingsrequirements


Results

Descriptive Statistics and Randomization: Household LevelLand-Informal Loan Details

TRAIL GBL Difference

Landholding (acres) 1.00 (0.05) 1.05 (0.06) -0.05Landless .07 (0.01) 0.08 (0.01) -0.01Purchased inputs on credit 0.38 (0.02) 0.43 (0.02) -0.05*Total Borrowing 6579.42 (524.32) 6417.26 (489.24) 162.52Duration (Days) 124.88 (1.57) 122.47 (1.27) 2.40Interest Rate (Per annum) 20.48 (0.87) 20.89 (0.77) -0.40Number of Loans 2.17 (0.06) 2.24 (0.06) -0.06Collateralized 0.02 (0.00) 0.01 (0.00) 0.01*

Joint Significance: 27.07


Results

Loan Details

Source Proportion Interest Rate Duration Proportion(APR) (days) Collateralized

Informal Lenders 0.65 27 124 0.01Family and Friends 0.05 21 169 0.02Cooperative/SJSY 0.23 15 324 0.38Government Banks 0.05 12 300 0.63


Results

Selected Crop Characteristics

Potato Sesame Paddy Vegetables

Acreage (acres) 0.48 0.43 0.69 0.20Total cultivation cost (Rupees) 10335 695 4061 3285Family labor (Hours) 56 24 37 82Revenue (Rupees) 17782 2433 6696 8976Value-added (Rupees) 7245 1736 2843 5586Value-added per acre (Rupees) 15094.27 4037.21 4120.48 27933.95


Treatment and Selection Effects

Estimating Treatment and Selection Effects

Report results of following regression specification:

yi = β0 + β1TRAIL + β2TRAIL× Control 1 + β3TRAIL× Treatment

+ β4GBL× Control 1 + β5GBL× Treatment + γ Xi + εi (1)

Selection effects (TRAIL, GBL): β2, β4

Treatment effects, ITT estimates (TRAIL, GBL): β3 − β2, β5 − β4

Controls include land owned, year dummy, price information dummy



Steps

Start with treatment and selection effects on loan size and cost of borrowing

Then show corresponding effects on cultivation and incomes from variouscrops

Interpret these, comparing selection effects between TRAIL-GBL onproductivity and informal interest rate

Compare repayment rates, take-up and continuation rates

Effects on non-farm incomes; sensitivity; tests for possible extraction offarmer returns by TRAIL agent



Impacts on Borrowing

Unit Treatment Selection Sample MeanTRAIL GBL TRAIL GBL Size Control 1

All Loans:Loan Size Rs 7126*** 6464*** -417 -919 2758 7279Cost of Borrowing Percent -0.03** -0.07*** -0.01 0.04** 2428 0.24

Non Program Loans:Loan Size Rs -495 254 -372 -930 2601 7279Cost of Borrowing Percent 0.01 -0.01 -0.01 0.04*** 2159 0.24



Impacts on Potato Production


Cultivate 0.0545 0.0492 0.0949*** 0.0614 4163 0.677Acreage Acres 0.0896*** 0.0402 0.0010 -0.0421 2718 0.432Leased-in acres Acres 0.0467** 0.0222 -0.00265 0.00447 2718 0.111Output Kg 888.0*** 278 145.4 -417.9 2718 4760Cost of production Rs 1774** 1308 372.8 -1111 2718 9538Family labour hours Hours 6.03 4.91 -0.2 4.951 2718 57.86Revenue Rs 3429*** 1637 942 -2534 2718 19137Value added Rs 1687** 271.8 555.6 -1371 2718 9498



Impacts on Value Added of Other Crops, and on TotalFarm Income


Sesame Rs 180 -158.3 -115.7 73.41 2037 2126Paddy Rs 271.6 573.6 -469.9 -759.6* 3047 2506Vegetables Rs 1255 -1955 1329 -957.5 402 8325

Total Farm Income Rs 2621*** 53.24 11466*** 10066*** 4163 10328


Explanations

Explanation 1: Comparing Productivity of SelectedFarmers

Seek to explain preceding results by selection differences between TRAIL andGBL, as predicted by the theory

First step: test whether productivity of TRAIL-selected farmers exceeds thatof GBL-selected farmers

Estimating productivity requires strong identification assumptions

Higher productivity implies higher rate of return achieved with respect to anexogenous expansion in cultivation (costs)

Use treatment status as instrument for cultivation costs: obtain IV estimateof RoR in TRAIL and GBL, assuming Cobb-Douglas production functionspecification

Alternative approach: calculate ∆VA∆CC where ∆ denotes corresponding

treatment effect estimate, and bootstrap standard errors (with 600replications)


Explanations

Rate of Return Estimates

Potato Total Farmincome

Bootstrapped Estimates

TRAIL 1.05*** 1.15***(0.06) (0.02)

GBL 0.09 -0.10(0.37) (0.29)

IV Estimate of CD Production Function

TRAIL 0.72** 1.03***(0.33) (0.35)

GBL 0.37 0.38(0.97) (1.23)


Explanations

Explanation 2: Testing for Network Effects in TRAILselection

Test prediction that TRAIL agents recommend borrowers from their ownnetwork

Three alternate definitions of past transactions: purchased inputs, borrowed,worked from/for agent

Expect ‘borrowed’ to be the most relevant

Also check for caste affinity between agent and borrowers recommended(network based on caste)


Explanations

Selection: TRAIL

(Dependent Variable: Household was recommended)

(1) (2)

Bought from agent 0.023 0.016(0.044) (0.047)

Borrow from agent 0.139*** 0.142***(0.037) (0.035)

Work for agent 0.003 -0.005(0.049) (0.055)

Non Hindu 0.030(0.143)

Non Hindu × Agent Hindu -0.098(0.132)

SC 0.544***(0.031)

SC × Agent High Caste -0.610***(0.036)

ST -0.198*(0.108)

ST × Agent High Caste 0.218(0.166)

Sample Size 1031 1031


Explanations

Explanation 3: Comparing Informal Interest Rates Paid bySelected Farmers

Test the hypothesis that TRAIL agent recommended farmers withbelow-average default risk, while GBL invited applications from farmers withabove-average default risk

Present OLS and Heckman-selected regressions for informal interest rates(controlling for selection into borrowing, using household head occupationdummy as instrument)

Subsequently compare repayment rates between TRAIL and GBL


Explanations

Explanation 3 (continued):

We expect:

Among the agent’s own clientele, those he recommended pay a lower interestrate

Average riskiness of GBL groups is higher (than the general population)

Average riskiness of GBL groups is higher than TRAIL recommendedhouseholds


Explanations

Informal Interest Rate Comparisons between TRAIL andGBL Selection

TRAIL GBL TRAIL v GBLOLS Heckman OLS Heckman OLS Heckman(1) (2) (3) (4) (5) (6)

Recommend 0.022 0.022 0.053* 0.052*(0.016) (0.017) (0.027) (0.029)

Own-clientele 0.050 0.049*(0.033) (0.027)

Own-clientele × Recommend -0.071** -0.071**(0.026) (0.035)

TRAIL -0.064 -0.064**(0.046) (0.027)

High caste -0.058*** -0.059*** 0.134* 0.134*** 0.053 0.053*(0.016) (0.016) (0.071) (0.031) (0.044) (0.028)

Landholding 0.091 0.090 -0.103 -0.071 -0.047 -0.023(0.070) (0.078) (0.170) (0.142) (0.182) (0.129)

Landholding Squared -0.063 -0.062 0.065 0.050 0.053 0.042(0.044) (0.052) (0.129) (0.093) (0.136) (0.086)

Constant 0.238*** 0.240*** 0.196*** 0.151 0.271*** 0.235**

Sample Size 438 1,032 417 1,038 412 911MMMMV (June 2014) Financing Smallholder Agriculture June 2014 45 / 60

Explanations

Loan Repayment Rates


Explanations

Loan Take-Up Rates


Explanations

Loan Continuation Rates, conditional on eligibility


Additional Issues

Additional Issues

Adverse treatment effects on non-farm incomes?

Sensitivity to price fluctuations

Extraction of borrower benefits by TRAIL agent?

Year Specific Effects

Effect of Prior Experience


Additional Issues

Treatment Effect on Non-Farm Income.

Treatment Selection Sample MeanTRAIL GBL TRAIL GBL Size Control 1

Rental Income (Rupees) 153.6 784.4 -182.1 -427.9 4162 1508Income from Animal Products (Rupees) 166.8 49.18 62.66 -279.1 4162 771Labour income (year; Rupees) 393 -5642 -12729** -4941 4162 37465Wage employment (last 2 weeks; Hours) 0.615 -4.496 -6.855* 1.749 4162 40.24Self-employment (last 2 weeks; Hours) 6.884 4.294 0.215 5.914* 4162 121.8Reported profits (Rupees) 2343 2918 100.9 -1917 4162 5802Current value business (Rupees) 4917 6692 952.1 353.8 4162 10465Total Non-Farm Income (Rupees) 3056 -1890 -12748 -7565 4162 45546


Additional Issues

Sensitivity of Treatment Effects for Potato Value Added toPrice Changes

Dependent Variable: Value added (Actual/Imputed)

Treatment Selection Sample MeanTRAIL GBL TRAIL GBL Size Control 1

1 Actual 1687** 271.8 555.6 -1371 2718 94982 2011 prices 1654*** 55.11 318 -872.7 2718 82583 2012 prices 3187*** 500 254.8 -1907 2718 143114 2007 prices -194.7 -328.5 -45.25 -2744 2718 44235 2008 prices -1913** 1653 1079 -2886** 2718 -44346 2011 market wage 1672** 217.3 463.5 -1483 2718 82197 2012 market wage 1665** 182.6 460.4 -1416 2718 8134


Additional Issues

Treatment Effects for Input Transactions with TRAILAgent

Sample Size Mean Control 1 Treatment Effect

Ever Buy any Input from agent 12,448 0.0875 -0.00338Share of agricultural input purchased from agent 10,196 0.0760 -0.00359

Input Price (Rs/unit)

Inorganic fertilizer 1,672 13.78 -0.322Organic fertilizer 370 16.12 29.39Outside seeds 1,654 22.36 2.174Pesticide 2,691 533.5 -31.08Powertiller 1,403 195.2 -32.33***Water/irrigation 1,230 72.30 148.3


Additional Issues

Treatment Effects for Output Transactions with TRAILAgent


Ever sold output to agent 2,990 0.209 0.00559Share of output sold to agent 2,765 0.151 0.0152

Output Price (Rs/kg)

Potato 1,386 4.507 -0.0516Paddy 498 9.282 -0.0215Sesame 881 28.42 -1.003


Additional Issues

Treatment Effects for Credit Transactions with TRAILAgent


Ever borrowed from agent 1690 0.17 -0.072*Share of loan from agent 1690 0.05 -0.034**

Interest Rate (APR) 5278 0.14 -0.003


Additional Issues

Year Specific Effects on Potato


Additional Issues

Year Specific Effects on Overall Incomes


Additional Issues

Effect of Prior Experience on Value-Added in PotatoCultivation

Low experience High experience(1) (2)

TRAIL Treatment 730.7 2350*GBL Treatment 1068 -121.9TRAIL Selection 620.2 431.2GBL Selection -1142 -2331Mean Control 1 6192 12864

Sample Size 1247 1451


Discussion and Next Steps

Comparing Administrative Costs

TRAIL incurred substantially lower administrative costs: per-monthper-village costs were Rs 68, compared with Rs 1463 for GBL

Bulk of the cost savings of TRAIL came from reducing loan officers’ salariesand transport expenses, since there were no group meetings in TRAIL

These costs amounted to Rs 1125 per month (at 2012 prices) per GBL village

In contrast, loan officers visited TRAIL villages only once in four months,resulting in personnel and travel cost of only Rupees 31 per month per village



Summary

Designed and experimentally evaluated microcredit designed to financesmallholder agriculture of high-value cash crops

Key differences between TRAIL and GBL: selection mechanism, andindividual vs. joint liability loans

Evidence shows TRAIL was successful in selecting productive low-risk farmersto expand cultivation of potato, and their own incomes (RoR in excess of70% at 2011-12 prices)

GBL induced eligible farmers to expand cultivation of potato, butcorresponding effects on output, incomes were insignificant

TRAIL achieved superior repayment and take-up rates, while reducingadministrative costs significantly

No evidence of extraction of TRAIL borrower benefits by agent, but thesebenefits were sensitive to price fluctuations



Next Steps

Despite lowering administrative costs, TRAIL as implemented would notbreak even, if MFI had to borrow at commercial rates (12%)

However, rates for priority sector lending to agriculture for Indian banks hasbeen recently lowered to 4%

Banks would incur 7% costs of monitoring and capital provisions, as perexisting regulations

At this borrowing cost, TRAIL would break even only if commission rates arelowered significantly

TRAIL agents say they would be willing to participate if the program isscaled up (e.g., lend to 25 recommended borrowers per agent)

Unlikely to dilute quality of selection, but if TRAIL agents provide help andenforce loan repayments there may be adverse impacts

Need to evaluate these in trying to scale up the intervention


financing smallholder agriculture: an experiment with ... · motivation presentation outline 1...

Documents