Understanding Demand for Electricity Among Rural Consumers Using Pay as You Go Solar

I offer promotions to existing customers of a pay as you go (PAYG) solar company that randomly vary the effective price that the customer pays for a day of electricity.

I work with a pay as you go (PAYG) solar provider operating in Rwanda and Kenya. Under PAYG, customers adopt a solar home system, which is comprised of solar panels, a battery to store power, and appliances. At a minimum, the solar home system includes three lights, a phone charger, and one other small appliance which could be a radio, rechargeable torch, an extra light, or a charging station for charging many phones at once. Later, customers can upgrade to systems that power more and larger appliances including a television, smartphone, and electric shaver. After selecting a solar home system, customers make payments on a long-term loan for their solar home system which are tied to their use of the system. Specifically, customers purchase system "on" time, or time in which they can use their system as much as they like conditional on the capacity of the system. For example, if a customer purchases a week on their system then they can use their system at full capacity for the next seven continuous days, but if they do not make another payment by the end of those seven days then they are remotely locked out of their system. If customers fail to make a payment for an extended period of time (typically a few months in a row), BBOXX sends a technician to repossess and redeploy their system; however, repossession is costly for BBOXX so even customers with relatively poor payment behavior do not get repossessed.

In the RCT, I am offering two types of payment incentives to its clients in Rwanda. Both types of incentives provide ways for clients to earn bonus days, or days in which their systems will not be switched off even if they do not pay. One type of incentive is a bulk incentive, in which customers earn bonus days by purchasing a minimum number of weeks in a single transaction. The other type of incentive is a reliability incentive, in which customers must purchase a minimum number of weeks over the course of a month to earn bonus days. Through a randomized control trial (RCT), I will induce variation in the price of electricity by randomly varying the effective price that customers face under each type of incentive.

2018-07-01

2018-12-31

Days of electricity purchased in a single month, marginal utility of expenditure.

I plan to estimate the marginal utility of expenditure using the methods in Ligon (2016). This entails eliciting household composition (number of boys, girls, men, and women living in the household) and eliciting expenditures on a range of income-elastic goods. For each good, I regress expenditures on the number of boys, girls, men, and women living in the house as well as a time by location fixed effect to account for price variation. This leaves me with a matrix of residuals, on which I perform a singular value decomposition (SVD). The marginal utility of expenditure is an object of rank one, so the SVD yields an estimate that is proportional to the marginal utility of expenditure, providing a measure of welfare for the household.

Frequency of remote lockouts, duration of remote lockouts, customer satisfaction.

I plan to ask a series of questions eliciting customer satisfaction with the company, particularly focusing on issues of fairness and value for money. I will combine these into a customer satisfaction index.

I randomly offer payment incentives to existing customers of a solar company in Rwanda and Kenya. PAYG solar customers purchase time on their solar home systems rather than kilowatt hours. Therefore, incentives take one of two forms. The first incentive is of the form "purchase X days and receive Y days free." The second incentive is of the form "purchase X days over the course of a month and receive Y days free at the end of the month." For each type of incentive, I randomly offer some customers a high discount, or a high number of free days, and others a low discount. The discounts range between 3.5-15% on the low discount schedule and 7-24% on the high discount schedule, with discounts increasing in the number of days purchased for both types of incentives. Finally, I offer some customers discounts starting at a low minimum threshold and others discounts starting at a high minimum threshold. This cross-randomized experimental design induces random variation in the effective price of electricity. Finally, I offer a simple information treatment that merely informs customers about how well they have to pay to be considered "good customers" by the solar company.

I am randomly offering 800 customers the bulk incentives and 800 customers the reliability incentives. Within each type of incentive, half of the customers will be randomly selected to be on a low incentive schedule and half on a high incentive schedule. Effective discounts range from 3.5-15% on the low discount schedule and from 7-24% on the high discount schedule. Finally, half of all customers within each incentive-type/incentive-level combination will be able to qualify for an incentive by purchasing at least four weeks, while the other half will only qualify upon purchasing 5 weeks. Therefore, I have eight groups of 200 customers each facing different types of incentives at a high or low incentive and at a high or low qualifying threshold. I have randomly selected an additional 800 customers to serve in the control group. I stratify all randomization by past payment performance. I am stratifying on past payment behavior to ensure that my sample contains a sufficient number of high electricity demand customers and low electricity demand customers.

Customers are being notified of the incentives that they face by a representative from the solar company's call center, and they also receive a SMS message with their incentive schedule for their future reference. Customers who are offered bulk incentives receive a SMS each time they make a qualifying purchase that informs them about how many bonus days they earned with their purchase. Customers who are offered reliability incentives will receive a SMS at the end of each month that informs them of how many bonus days they earned that month. Finally, customers receive a SMS notification when they come within three days of beginning to use their bonus days and when they come within three days of running out of time on their system.

I am collecting data in three ways. For one, I have access to the universe of administrative data from the solar company which includes information on basic customer demographics, appliances owned by each customer, system maintenance and customer calls to the call center, and the complete payment history of each customer. In addition, I am sending customers in Rwanda short monthly SMS surveys on consumption expenditures on a select set of goods, which I identified using the methods outlined in Ligon (2016). In effect, I select foods that are the most income elastic and which constitute a non-neglibile share of household expenditures among rural households in the lowest income quartile in a nationally representative survey. These SMS surveys will enable me to measure households' marginal utility of expenditures at a relatively high frequency. The other data collection I plan to do is an endline survey, which will take place with an enumerator over the phone at the end of the experiment in December, 2018 or January, 2019. This survey will provide one final measure of the marginal utility of expenditures, elicit more information about energy usage and customer satisfaction, and collect other data deemed relevant by stakeholders who I plan to engage with as I design the final survey instrument.

The randomization is done in an office by a computer.

The randomization is done at the level of the individual customer, there is no clustering. I stratify by past payment performance so that there are sufficient numbers of previously well-paying and previously poor-paying customers to examine effects on each.

No

2400 individual customers.

2400 individual customers.

400 - pure control
400 - information treatment
200 - reliability incentive, low discount rate, low minimum threshold
200 - reliability incentive, low discount rate, high minimum threshold
200 - reliability incentive, high discount rate, low minimum threshold
200 - reliability incentive, high discount rate, high minimum threshold
200 - bulk incentive, low discount rate, low minimum threshold
200 - bulk incentive, low discount rate, high minimum threshold
200 - bulk incentive, high discount rate, low minimum threshold
200 - bulk incentive, high discount rate, high minimum threshold

Note that when performing power calculations, I am presenting power calculations for 800 treated individuals (in a given type of incentive) compared to 800 control individual (the pure control combined with the information treatment). All sub-analyses will have less power in accordance with the reduced sample.

For days purchased in a single month and the duration of remote lockouts, the relevant units are days. If I assume that the data are distributed normally with mean of 23.44 days and a standard deviation between one and a half and three days for the days purchased in a single month, this indicates that I will have a minimum detectable effect size of 0.19-0.38 standard deviations, which would be 7-14% of the mean. When examining the marginal utility of expenditures, the relevant unit will be utils/RWF, as I am measuring expenditures in Rwandan Francs. I have very little upon which to base my power calculations for this outcome, as it is a novel measure of welfare and I have insufficient data on the individuals in my sample to extrapolate from nationally representative surveys of Rwanda. However, using the data from Ligon (2016), I estimate that assuming my data are distributed normally with a mean MUE of 0.14 and a standard deviation of between 0.01 and 0.05, I will have a MDE of 0.001 to 0.006 standard deviations, or 8-40% of the mean.