Enhancing Enforcement through Religious Institutions: Experimental Evidence from Pakistan’s Power Sector

(A) Religious Messaging recruits religious leaders/mosques to deliver messages on electricity theft during Friday sermons, a weekly congregation attended by most of the male population in the country. We encourage using references from the Quran and Sunnah of the Prophet (Peace Be Upon Him).

(B) Secular Messaging recruits village leaders to gather community members and deliver messages on electricity theft, analogous to a
Panchayat. We explicitly ask them not to make any religious references in their messages and instead to rely purely on logical reasons (e.g. less theft means more supply). This pilot arm is important to test if secular messages can be a viable alternative in less religious contexts.

For both (A) and (B), messages would be delivered once every two weeks for three months. Leaders have discretion over the precise message content but we do ask them to specifically mention: “stealing electricity is theft and not paying bills is equivalent to not honoring debt”. We will audit a random set of treatment sites each week and will record all messages delivered. From our initial field work we did not encounter any significant issues on recruiting mosques into the treatment or on compliance.

2024-01-01

2024-10-01

Changes in outright electricity theft, as measured by the difference between incoming feeder supply and billed power
Changes in bill recovery, as measured by changes in the proportion of monthly household electricity bills that get paid

Our experiment will take place across 140 distribution feeders in rural areas of Punjab, Pakistan over a 12-month period. The experiment consists of the following treatments:

A. Religious Messaging [Mosques] (65 Feeders)
B. Secular Messaging [Social gathering] (10 Feeders)
D. Control (65 Feeders)

Our level of randomization is large. A typical distribution feeder covers 5-6 villages with an average of over 3,500 customers per feeder. The largest mosques in each village will be recruited for the treatment. Akhuwat will help recruit communities into the various treatments, bringing down costs and setting us up for a future scale up. We randomize at the feeder level because this is the lowest level we can observe direct measurements of outright electricity theft via illegal hooks onto power lines, a common form of theft in our sample area.

Our focus is on rural areas. We stratify our sample based on the pre-treatment feeder category assigned by the government. The government classifies a feeder into different categories (I to VII) based on its historical aggregate technical and commercial losses. This category, in turn, determines the amount of scheduled power outages a particular feeder will face. Stratifying based on these bins will therefore ensure a cleaner comparison between treatment and control, especially since outages may influence payment or theft decisions. We control for further covariates, notably number of customers in a feeder and baseline incomes, that may influence our outcomes.

Monthly electricity billing and supply data at both the individual and feeder level will be provided by the government. This will let us measure our two core outcomes of interest: changes in outright electricity theft, as measured by the difference between incoming feeder supply and billed power, and changes in bill recovery, as measured by changes in the proportion of monthly household electricity bills that get paid. These outcomes provide a revealed preference measure for whether social norms change.

Household surveys (30-40 per feeder) will supplement this data to elicit beliefs on paying for electricity as well as measure religiosity. The household survey captures three major aspects:

- Basic Household Information: demographics, asset ownership, income and expenditure, numeracy and literacy tests, risk preferences and loss aversion, shocks
- Electricity: electricity sources, consumption, payments, reasons for partial/non-payment and theft, power outages, beliefs on enforcement and whether people believe that electricity is a right
- Religiosity: we elicit levels of religiosity (intrinsic, extrinsic, general) in a similar vein to Bryan, Choi & Karlan (2021). We included detailed questions on mosque and Friday sermon attendance.
- Networks: we are currently testing a survey instrument in a sample of villages to generate aggregate relational data that can identify likely network ties in local communities (Breza, Chandrasekhar, McCormick & Pan, 2020). This would enable us to trace out whether changes in outcomes following the receival of the treatment (e.g. listened to mosque sermon) extend via social networks to ‘untreated’ households (e.g. those who did not attend) within a community.

Not available

Randomization done in office by a computer.

Feeder.

Yes

140 distribution feeders in rural areas of Punjab, Pakistan.

140*3,500=490,000 customers/households, on average

Religious Messaging [Mosques]: 65 Feeders
Secular Messaging [Social gathering]: 10 Feeders
Control: 65 Feeders

Our primary outcomes of interest are outright theft (measured at the feeder level) and bill payments (measured either on individual bills or feeder level collection). We use administrative data for all feeders in our sample area over the past 6 months coupled with individual billing data to estimate the necessary sample sizes to obtain power. We stratify based on feeder category and control for the number of connections. With the standard assumption on significance level (0.05), we require a total sample size of 126-130 feeders to detect a meaningful effect (i.e. 10% change). To improve our measure of losses due to unpaid bills we can utilise our individual-level billing data by adapting a clustered design with distribution feeders as the cluster. We compute the intra-cluster correlation using consumer data from earlier fieldwork. Our power calculations for collection loss (individual level) tells us that in order to detect a 2.5 percentage point difference (6% change) with a cluster size of 50, we need 36 feeders in treatment and 36 in control. In total, 72 feeders and 3,600 individuals. Because we lack full pilot results on treatment effect sizes we opt for a conservative core total sample size of 130 feeders. The treatment effect sizes necessary (i.e. 6-10% changes) are reasonable and match those with other papers using religious interventions in other settings, such as paying back credit card debt (Bursztyn et al, 2019). To avoid possible concerns on spillovers we will not place treatment and control feeders next to each other.