Determinants of demand and repayment in microcredit lending groups in rural Bangladesh

We conduct lab-in-the-field experiments with clients of a microfinance institution that participate in a randomized evaluation of credit contracts. We invite borrowers of 76 borrowing groups to participate in three different experimental games. In the first two games we elicit behavioral drivers of group performance and leadership quality of the group leader. In the third game we conduct choice experiments for different loan contract characteristics.

We conduct lab-in-the-field experiments with clients of a microfinance institution that participate in a randomized evaluation of credit contracts. We invite borrowers of 76 borrowing groups to participate in three different experimental games.

In the first two games we elicit behavioral drivers of group performance and leadership quality of the group leader.

The first game is a Trust Game with third party punishment with the following specifications:
• initial endowment of 40 BDT
• sender (A) can send 0, 10, 20, 30 or 40 BDT
• amount is tripled and send to receiver
• receiver (B) decides how much to return to sender (strategy method)
• action space of third party (C) is varied in 3 treatments
We have a 3x2 design:
Treatment Dimension 1 (within subjects): Action space of player C
o Base game: no action - elicit C's beliefs
o Monitoring: C observes A's & B's decisions at no costs
o Punishment: C can punish B at a cost: pay 10 to punish B by 20 BDT
Treatment Dimension 2 (between subjects): Identity of player C
o actual group leader or
o randomly selected (and announced) other member
We elicit the following decisions in partially anonymous groups (composition of groups and identity of A and B not known; identity of C known to A and B):
Player A:
o amount sent to B
o belief about back transfer from B (strategy method)
Player B
o amount sent back to A (strategy method)
o belief about transfer from A
Player C:
o belief about transfer from A
o belief about back transfers from B (strategy method)
o punishment choices (strategy method for amounts sent by A and choice of cut-off value)

The second game is a Public Good Game with sequential decisions (Leader - Follower) with the following characteristics:
• initial endowment of 22 BDT
• first mover (A) can contribute 20 BDT to public good
• second mover (B) observes A's action before own contribution decision
• amount in public good is ....
increased by factor 3
increased by factor 1.5
depleted completely
and split equally across A and B
• A's knowledge of the Marginal Per Capita Return (MPCR) varies in 2 treatments. (MPCR is 150%, 75% or 0%)
We have a 2x2 Design:
Treatment Dimension 1 (within subjects): Knowledge of MPCR
o Base game: A has full information on MPCR before contributing
o Info acquisition: A can acquire information on MPCR at cost 2 BDT
Treatment Dimension 2 (between subjects): Identity of player A
o actual group leader or
o randomly selected (and announced) other member
We elicit the following decisions in partially anonymous groups (composition of groups and identity of B not known; identity of A known to B):
Player A:
o Contributions (yes/no) for different MPCR
o Decision to acquire information on MPCR
Player B
o Contributions (yes/no) after observing A’s action (strategy method)


In the third game we conduct choice experiments for different loan contract characteristics. In discrete-choice experiments (DCEs) based on the randomized evaluation “Reaching the unreached: Credit contract design for the ultra poor” by Kazushi Takahashi, Abu Shonchoy, Seiro Ito, and Takashi Kurosaki the following aspects of a loan contract are studied:
• small repeated loans vs. bigger one time loans (e.g. 3 consecutive one-year loans of 5,000 BDT vs. 1 3-years loan of 15,000 BDT)
• a grace period to delay loan repayment by one year for the 3-years loan of 15,000 BDT
• providing the credit in kind, that is a dairy cattle, rather than in cash for the 3-years loan of 15,000 BDT with a grace period
The general setup of the loans studied consist of four loan contract characteristics (attributes) for a total loan amount of 15,000 BDT over a 3-years period, i.e. 150 weeks.
1. Number of loan disbursements (1 or 3)
2. Grace period (in weeks) by which loan repayment is delayed after loan disbursement (0, 25 or 50 weeks)
3. Interest rate calculated as a flat interest rate (interests rate x loan amount which is divided in equal installments across all repayment weeks) (8, 10, 12, 14, 16, 18, 20, 22, or 24 %)
4. Form of credit disbursement (in cash or in kind (cow))
From the specifications of the attributes, hypothetical loan contracts are formed by combining different specifications of each attribute. Using the number of identified characteristics (also called attributes) and different levels (the different specifications or each attribute), the full factorial design gives 2 x 3 x 9 x 2 =108 possible combinations of the attributes and hence as many hypothetical loan contracts.
The set of all possible factorial combinations yields 108 x (108-1)/2=5778 possible binary choice sets. From these possible binary choices a subset is selected based on the following criteria:
• Orthogonality: Minimal correlation between the attribute levels that appear in the DCE (measured by correlation coefficients and D-efficiency for statistical efficient designs)
• Level balance: each attribute level should appear roughly an equal number of times in the DCE
• Minimal overlap: two loan contracts that appear together in a choice set should rarely have the same attribute levels
For this we followed the following procedure:
1) Use SPSS Orthoplan command: A fractional factorial design resulted in an orthogonal design matrix with 27 alternative binary choices
2) Generate a set of D-efficient alternatives (using the “Algorithmic Experimental Design” package of the statistical environment R. This package offers an implementation of Fedorov's exchange algorithm). Resulting orthogonal matrix has 29 binary choices
3) Check for the orthogonality and attribute level balance
4) Allocate alternatives into four blocks and construct all possible binary choice sets per block
5) Eliminate sets due to complexity (e.g. if contracts vary on all attribute levels) and domination
6) Randomly select 8 choices for each block (ensure that at least one choice set with the common anchor contract is in each block
The four blocks were randomly assigned across sessions. Per session, all 8 binary choices of the assigned block were played, that is two loan contracts are presented to participants at a time. The participant has to state which of the loan contracts she prefers. This binary choice is repeated for a series of loan contract pairs. The binary decisions are easy for participants even though they have to consider several loan contract attributes at this same time. Precisely this feature allows us to elicit the trade-off between different attributes.

Outcomes:
For the Trust Game:
- Transfers
- Back transfers
- Punishment
- Beliefs
For the Public Good Game:
- Contribution
- Information Acquisition
For the repayment data (from the randomized evaluation)
- Default 1: binary variable if loan is not repaid at the end of the loan cycle
- Default 2: amount in default at end of the loan cycle, and various intervals after the end of the loan cycle, e.g. 1 month after end of loan cycle, 2 months after, etc.
- Repayment discipline 1: number of weekly repayments missed
- Repayment discipline 2: share of repayments made on time (relative to total number of repayments)
- Savings 1: amount of savings accumulated
- Savings 2: amount of savings withdrawn
- Savings 3: binary indicator if savings have been used for loan repayment

2016-04-05

2016-06-05

Trust, Trustworthiness, Contribution to public good, Beliefs, Default, Repayment discipline, Savings

Repayment discipline 1: number of weekly repayments missed
Repayment discipline 2: share of repayments made on time (relative to total number of repayments)

We conduct lab-in-the-field experiments with various treatments. In our 3x2 and 2x2 design, we have a within subject design along the first treatment dimension and a between subject design along the second treatment dimension.

We conduct lab-in-the-field experiments with various treatments. In our 3x2 and 2x2 design, we have a within subject design along the first treatment dimension and a between subject design along the second treatment dimension.
In the Trust Game we have a 3x2 design:

Player Type:
- Random assignment of participants to player A, B, and C
- Exception: real life group leader is always player C.

Treatment Dimension 1 (within subjects): Action space of player C:
Treatment 1: Base game: no action - elicit C's beliefs
Treatment 2: Monitoring: C observes A's & B's decisions at no costs
Treatment 3: Punishment: C can punish B at a cost: pay 10 to punish B by 20 BDT
--> all subjects play all treatments in the same sequence in order to increase subjects' understanding
Treatment Dimension 2 (between subjects): Identity of player C
Treatment 1: actual group leader or
Treatment 2: randomly selected (and announced) other member

In the Public Good Game we have a 2x2 Design:
Player Type:
- Random assignment of participants to player A and B
- Exception: real life group leader is always player A
Treatment Dimension 1 (within subjects): Knowledge of MPCR
Treatment 1: Base game: A has full information on MPCR before contributing
Treatment 2: Info acquisition: A can acquire information on MPCR at cost 2 BDT
--> sequence of treatments is randomly assigned across sessions
Treatment Dimension 2 (between subjects): Identity of player A
Treatment 1: actual group leader or
Treatment 2: randomly selected (and announced) other member

In the Discrete Choice Experiments, the four decision blocks with eight binary choices each are randomly assigned across sessions.

Randomization method:
- Assignment of players: draw of colored chip from opaque bag at the beginning of Trust Game and Public Good Game: for the Trust Game: 7 players as sender, 7 players as receiver, 6 players as third party (the group leader is automatically assigned this position), for the Public Good Game: 10 first movers (the group leader is automatically assigned this position) and 10 second movers. If less than 20 participants in session, in TG, remove one sender or receiver first, in PGG remove first mover first.
- Assignment of Leader vs. Member Monitoring and Punishment (Player A and B) in Trust Game and Leader vs. Member as first mover (Player B) in Public Good Game by random allocation of decision sheets
- Randomization of treatment sequence in Public Good Game and decision block in discrete choice games done in office by a computer
- Treatment sequence and block assignment clustered at session level

- Assignment of players in game by draw of colored chip from opaque bag before game or by random allocation of decision sheets
- Randomization of treatment sequence and decision block done in office by a computer clustered at session level

Borrowing group

Yes

76 borrowing groups

1520 borrowers

1520 borrowers: treatments in lab-in-the-field experiments are mainly varied within subject. If not, procedures are specified in the experimental design description.

We test several hypothesis for which we calculate the minimum detectable effect size based on our sample size, a power level of 80% and a significance level of 5%. The minimum detectable effect size of whether monitoring and punishment increase transfers is an increase of 0.2818 and 0.2930, respectively. The minimum detectable effect size of whether monitoring and punishment increase relative back transfers is 0.0295 percentage points and 0.0312 percentage points respectively. For testing whether leader monitoring and leader punishment are more effective in increasing transfers/ back transfers, the minimum detectable effect size is 0.4315 and 0.4372 for transfers and 0.0455 and 0.0491 percentage points for relative back transfers. In order to test whether the leader is followed more, the minimum detectable effect size is 0.0797 if the leader contributes, and 0.1290 if not. For the repayment data, we have a intra group correlation. When using a median split technique to identify groups with good and bad leaders, we have 50 % of groups with good and bad leaders. The minimum detectable effect size is 0.0882 when accounting for intra group correlation.