Between trust and trade: on informal credit networks in India

2024-02-01

2024-04-01

Our first set of primary outcomes has to do with the impact on the business. We will record indicators covering changes in customers’ overall purchasing patterns and credit repayment before vs. after the intervention. Additionally, we intend to capture a sense of the market share occupied by the store by looking at the diversity and total number of customers, total revenue, and volume of goods passing through the store.

We are also interested in seeing whether the interventions result in a deviation from customers’ existing relationships with stores. In particular, we will test whether:
- Treated customers deviate their standard expenditure shares and values from their usual stores and towards the stores where they were offered credit vs. discounts

In addition, we will also focus on the effects on customers, including the effects on consumption smoothing, and on the consumption baskets of these customers. We will focus on expenditure on food, dietary diversity, and overall nutritional intake.

We will test heterogeneous treatment effects on the following dimensions:
- Previously regular or active customers, i.e. whether a customer has been shopping at this (and other) stores in the past, and/or has received credit from the store in the past.
- Baseline (i.e. time-invariant) relational distance between customers and store owners (based on demographic characteristics like caste, gender, religion, migrant status, place of origin, and economic characteristics like informal work and wage status, steady state earnings, and frequency of earnings, and others.)
- Similarly, we will also test heterogeneity by spatial distance.
- We will also test HTEs through a data-driven approach using a machine learning algorithm.

Lastly, for the zero-stage experiment (detailed below), we will simply test whether customers act on the offer of credit and discounts to vary the stores they shop at. The clear outcome we will look at is whether treated customers shop at the stores they are informed about.

Controls:
- Baseline value of the outcome, where available
- We will use double LASSO to select other controls
- In an alternate specification, we will add fixed effects at the store level

We will also test the effects of the intervention on store-level activity. In particular, we will test the effects on:
- Predictability of the demand
- Frequency of stock-outs and ordering inventory
- Costs of repayment of the loans (time spent following up on loans, liquidity losses)

We will also attempt to characterize the relationships between storeowners and customers through the following measures we will collect in surveys with both groups after the intervention:
- Perceived social pressures
- Strength of relationship
- Perceived loyalty in the relationship
- Pro-sociality of the storeowners

First stage experiment:
T1 - Text message with store credit offer (33% of the sample): a third of customers will get a text message informing them to go to a store that is not their most commonly visited store, where they may have the opportunity to receive goods on credit. More specifically:
- We will determine which store each customer goes to most frequently; we will assume that this is the customer’s location
- We will calculate the distance between the most visited store and all other stores
- We will divide stores into different quantiles of distance to the most visited store
- We will randomly select 1 store from each quantile to include in the text message and encourage people to go to. This selection allows us to understand how distance impacts demand
- We will exclude the store most visited by the customer from the list of stores where they are eligible to get store credit

T2 - Text message with discount offer (33% of the sample): a third of customers will get a text message informing them to go to a store that is not their most commonly visited store, where they may have the opportunity to receive goods on discount.
- The selection of stores to include in the message will follow the same procedure as above
- Control group (33% of the sample): the last third of customers will not receive any text messages

Second stage experiment:

Store-level randomization
We will randomly allocate stores to one of the below conditions:
- Store Credit (8 stores): we will offer store credit to a random subset of these stores’ customers
- Discount (8 stores): we will offer price discounts to a random subset of these stores’ customers
- Store Credit and Discount (4 stores): we will offer either store credit or price discounts to random subsets of these stores’ customers
- Pure control (4 stores): we will not intervene in these stores

Individual-level randomization
Additionally, we will randomize customers into one of the conditions below:
- Within Credit stores (N=8): 50% of customers will be randomized to receive store credit
- Randomization will take place at the time of billing through an online app (Knack), after customers choose their basket
- The store credit offer will cover the entire customer basket up to a maximum of INR 100
- The store credit offer will follow usual terms of store credit repayment in these contexts (e.g., no-interest loans; repayment period is decided by the business owner )
- Store owners will be informed prior to the intervention that the research team is underwriting these loans and that they will be responsible for the recollection. The store owners will keep the full amount of the repayments and can choose to lend it again

- Within discount stores (N=8): 50% customers will be randomized to receive discount
Randomization will take place at the time of billing through an online app (Knack), after customers choose their basket
The discount amount will depend on the basket value. We will round the basket value to its floor, divisible by 5
- For example, if basket value is 28, the customer will pay 25 INR; if basket value if 32 the customer will pay 30

Within credit + discount stores (N=4): customers will be randomly assigned to receive credit, discount, or nothing with equal probability (~33%).
- Within control stores (N=4): customers will not be offered any discount / store credit
These interventions will last for 8 weeks, subject to us meeting the expected sample size listed below.

- Randomization will be conducted using a computer algorithm
- The randomization protocol is detailed as follows:
- Zero stage randomization: we will stratify treatment assignment by degree of loyalty to most commonly visited store and average monthly spend. In addition, the names/locations of stores we will list in text messages will be spatially stratified: we will include one store randomly from each quintile of stores in the Euclidean distance from the focal store.
- Store-level randomization: we will stratify treatment status by pre-existing revenue and credit activity.
- Customer-level randomization: this will not be stratified based on any indicators. We will use an online database platform called “Knack” on the standard ePOS device in each store to randomly order each transaction from a “new” customer with 50% (or 33%) probability into a treatment condition, as needed.

The unit of randomization is the customer for the first stage, and a new customer’s transaction for the intervention at the store.

No

26 for store-level outcomes
Approximately 10,000 clusters for transaction-level outcomes, where we will cluster at the level of the individuals

Approximately 10,000 observations over 2 months, totaling approximately 20,000 - 40,000 transaction-level observations.

10,000 per arm for the first stage, and approximately 5,000 per arm for the second stage