The World Bank Economic Review, 37(4), 2023, 620–639 https://doi.org10.1093/wber/lhad012 Article Nudging Payment Behavior: Evidence from a Field Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Experiment on Pay-as-You-Go Off-Grid Electricity Jacopo Bonan, Giovanna d’Adda, Mahreen Mahmud, and Farah Said Abstract This paper reports results from a randomized control trial with a pay-as-you-go (PAYG) solar system provider in Pakistan. In the default treatment, customers are told the amount to pay every month to keep the system active. In a first treatment, customers are assisted in planning this monthly payment. A second treatment dis- closes that payments can be made flexibly within the month. This disclosure may reduce contract cancellation by helping minimize transaction costs but may increase contract complexity and reduce discipline. A third treatment combines flexibility with assistance in planning payments. Disclosing flexibility increases contract cancellation relative to the default, but combining flexibility with planning offsets this effect. Treatment effects appear stronger among users facing high mental constraints and transaction costs. These findings support the idea that behavioral factors, such as inattention and commitment problems, lay behind the negative impact of flexibility on cancellation. The results suggest that providers of PAYG systems may face a trade-off between disclosing complex contractual features and customer retention. Planning helps customers handle the added complexity. JEL classification: C93, D12, P18, O13 Keywords: consumer behavior, behavioral nudges, solar energy, field experiment, Pakistan 1. Introduction The high costs and slow pace of grid expansion mean that developing countries must rely on off-grid solu- tions for basic electricity access in short to medium run. Solar home systems dominate this sector, reaching Jacopo Bonan is Assistant Professor at Politecnico di Milano and Affiliated Scientist at RFF-CMCC European Institute on Economics and the Environment (EIEE), Centro Euro-Mediterraneo sui Cambiamenti Climatici; his email address is ja- copo.bonan@polimi.it. Giovanna d’Adda (corresponding author) is Associate Professor at the University of Milan and Af- filiated Scientist at RFF-CMCC European Institute on Economics and the Environment (EIEE), Centro Euro-Mediterraneo sui Cambiamenti Climatici; her email address is giovanna.dadda@unimi.it. Mahreen Mahmud is Assistant Professor at the University of Exeter; her email address is m.mahmud@exeter.ac.uk. Farah Said is Assistant Professor at the Lahore Uni- versity of Management Sciences; her email address is farah_said@lums.edu.pk. We acknowledge financial support from the International Growth Centre (project code 37403) and the European Research Council under the European Union’s Seventh Framework Programme (ERC grant agreement no. 336155). We thank Jeremy Higgs, Faez Shakil, and the EcoEnergy sales team for the fruitful collaboration. We also thank Ali Habib, Ali Dehlavi, Altaaf Hussein Sheikh, Abdul Rehman, and the WWF Pakistan enumerator team. We thank Muhammad Meki, Imran Rasul, Kate Vyborny, and the participants at ADE 2018 and NCID Workshop 2022 for their very useful comments. Faryal Manzoor provided valuable research assistance. This RCT was registered in the American Economic Association Registry for randomized control trials under trial num- ber AEARCTR-0002543 (https://www.socialscienceregistry.org/trials/2543). This study has ethics approval from the Lahore School of Economics (RERC-082016-01) and the University of Oxford (SSD/CUREC1A/BSG_C1A-17-010). A supplemen- tary online appendix is available with this article at The World Bank Economic Review website. C The Author(s) 2023. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com The World Bank Economic Review 621 about 20 percent of the off-grid population in Africa and South Asia (Lightening Global Program 2020) and representing a market that has been growing steadily over the past decade.1 Pay-as-you-go (PAYG) has emerged as a dominant business model for home solar systems since the flexibility allows low-income households to set their payment schedules. Flexibility can help low-income customers tailor payments to irregular cash flows and minimize waste and transaction costs, but the literature suggests varying de- grees of effectiveness under different contexts (McIntosh 2008; Field and Pande 2008; Labie, Laureti, and Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Szafarz 2017; Barboni 2017). This may be because flexibility is complex for customers to understand and can harm payment discipline (Brune, Ginè, and Karlan 2022). Therefore, from the provider’s perspective, encouraging regular payments simplifies the presentation of contractual features and could reduce default. This study uses a randomized control trial (RCT) to test the impact of two nudges on customers’ prod- uct demand and payment behavior through administrative and survey data on 726 customers of a provider of PAYG home solar systems in rural Pakistan.2 The first treatment arm (“Flex”) discloses the possibility for customers to set a flexible payment schedule of their choosing rather than sticking to the rigid monthly one suggested in the default condition by the provider. The second treatment (Implementation Intention Plan, “IIP”) nudges customers to formulate a plan for when and how they will make payments. Imple- mentation plans are effective in helping users identify the schedule best suited to them and sticking to commitments (Nickerson and Rogers 2010; Milkman et al. 2013; Abel et al. 2019). When combined with flexibility in payment, implementation plans can compensate for a flexible schedule’s higher complexity and reduced commitment features: the third treatment arm (“Flex × IIP”) tests this by combining the Flex and IIP treatments. All treatments are administered only once, at the start of the contract. The extensive margin of customers’ product demand, measured as the likelihood of customers can- celling their contracts, is high within the sample: more than half of the customers cancel their contracts due to default. The impact of the treatments on the extensive margin provides support for the relevance of behavioral factors in observed customer behavior. First, disclosing flexibility leads to marginally higher contract cancellation by 9.8 percentage points—approximately a 19 percent increase over the control group. Second, the negative effect of flexibility disclosure on cancellation is more than offset when flex- ibility is combined with the planning nudge to address its negative consequences on payment discipline. Planning in the absence of flexibility has no impact on cancellation. Third, cancellation appears to be higher among individuals in the Flex treatment who report difficulties maintaining financial discipline at baseline. Customers in the Flex treatment who face high transaction costs (proxied by the distance to the nearest payment location) are also more likely to cancel. These results suggest that flexibility on its own may increase the complexity of payment decisions and reduce payment discipline, mainly when the barriers to payment are high, and that the addition of planning may help customers better manage these challenges. Administrative data on daily inactivity and payments are leveraged to estimate the effect of the treat- ment patterns on customer payment behavior. Treatment effects on the likelihood of being inactive over the contract duration and the number of monthly inactive days appear consistent with those on cancella- tion. No impact is instead observed on other measures of the intensive margin of product demand, such as the average duration of inactivity spells and the number of monthly payments (top-ups). The primary specification used in the analysis alleviates concerns that these results are due to selective attrition. Still, 1 Focusing on the pre-pandemic period, GOGLA et al. (2018) reports that sales of home solar systems in 2018 have increased by 77 percent compared to 2017 and by 133 percent compared to 2016. Eight million households worldwide are estimated to have received access to electricity via solar PAYG products between 2015 and 2020 (IRENA 2020). However, the quality of off-grid systems varies greatly. 2 Pakistan represents a potentially massive market for PAYG solar solutions on account of its population density, the relatively low proportion of the population living in absolute poverty, and the high costs of existing alternative energy sources (e.g., kerosene) (Lightening Global Program 2019). 622 Bonan et al. the possibility that the null results are due to low statistical power from high cancellation rates cannot be ruled out. This paper relates to the literature on disclosing contractual features’ impact on consumer behavior. Customers have limited attention and tend to overlook complex information (Lang 2022), and providers exploit this by giving limited or opaque information on product features (Chetty, Looney, and Kroft 2009; Grubb and Osborne 2015; Karlan et al. 2016). The evidence on the impact of information disclosure is Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 mixed. In finance, providing customers with better contract information improves their financial behavior (Bertrand and Morse 2011; Stango and Zinman 2014). In the field of resource usage, Wichman (2017) finds that increasing the frequency of price and consumption information reduces resource conservation. The study looks at households’ response to disclosing products’ contractual features in a developing country with low literacy levels. Information disclosure widens customers’ opportunities but increases the complexity of effective usage of the product. In the context of the present study, providers face a trade-off between disclosure of complex contractual features and customer retention. Contract disclosure in our setting concerns the flexibility allowed by the PAYG scheme. Evidence on the relative impact of rigid and flexible contracts on repayment is mostly restricted to the microfinance lit- erature (McIntosh 2008; Field and Pande 2008; Field et al. 2013; Czura 2015; Labie, Laureti, and Szafarz 2017; Barboni 2017; Barboni and Agarwal 2019). This paper differs from this literature in three main respects. First, the flexibility treatment discloses an existing feature rather than introducing and testing a new flexible feature. Second, flexibility yields different benefits in our setting compared to microfinance. Third, the design of this study combines flexibility disclosure and planning nudges. Implementation inten- tions have provided cost-effective means of increasing the likelihood of follow-through when targeted to simple and short-term behaviors and when follow-through requires overcoming an obstacle (Prestwich, Lawton, and Conner 2003; Nickerson and Rogers 2010; Milkman et al. 2011, 2013; Abel et al. 2019; Mazar, Mochon, and Ariely 2018).3 Our results indicate that planning’s effectiveness is detectable only when flexibility disclosure makes the formulation of plans important, and suggest that the combination of nudges can have different effects relative to the effect of each one in isolation (Muralidharan, Romero, and Wüthrich 2020; Banerjee et al. 2021). Given the technological advances that lower the administrative cost of small, frequent payments and poor customers’ preference for schedules that match their irregular cash flows (Jack and Smith 2015; Suri 2017; Afzal et al. 2018), the form of flexibility examined here and our results are relevant in the context of microfinance, service payment, and possibly other realms. This study focuses on the policy-relevant domain of energy provision in developing countries (Bonan, Pareglio, and Tavoni 2017) and particularly on solar off-grid (Girardeau, Oberholzer, and Pattanayak 2021). Pre-paid electricity is rapidly expanding in developing countries due to technological innovations in metering and payment technologies. Existing evidence shows the advantages of pre-paid grid electric- ity over traditional post-paid monthly billing for liquidity-constrained users in urban contexts (Jack and Smith 2015, 2020). The perishable nature of pre-paid electricity in PAYG solar systems introduces impor- tant distinctions compared to pre-paid metering. The viability of these businesses is still under-researched (Guajardo 2016, 2019; Bensch et al. 2018; Grimm et al. 2020; Groenewoudt and Romijn 2022) and the debate is open on whether PAYG offers a viable way to cover the last mile of electricity access, particularly for the poorest (Barry and Creti 2020; Grimm et al. 2020). In the closest study to ours, Lang (2020) fo- cuses on the trade-off PAYG users face between liquidity constraints and transaction costs when deciding payment frequency. Our study offers a broader overview of obstacles to customer retention and regular payment and assesses their empirical relevance and interaction with behavioral nudges. The analysis has limitations. The design does not address users’ demand for flexibility or planning nudges. This prevents our analysis from speaking to the literature on the demand for soft commitment 3 However, other works find null effects of planning on public transport usage (Gravert and Olsson Collentine 2019). See Hagger and Luszczynska (2014) and Rogers et al. (2015) for reviews. The World Bank Economic Review 623 devices (Bryan, Karlan, and Nelson 2010; Laibson 2018). With survey data collected only at baseline, the impact of the nudges (Allcott and Kessler 2019; Bicchieri and Dimant 2019) and of access to off-grid technology on household welfare (Grimm et al. 2017; Aklin et al. 2017; Wagner et al. 2021; Stojanovski et al. 2021) cannot be evaluated. The paper first presents the study context. Then, it discusses the experimental design, presents the data, outlines the estimation strategy, and reports empirical results and robustness checks. Finally, it concludes. Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 2. Context 2.1. Setting The study is the product of a collaboration with EcoEnergy (EE), a for-profit company supplying solar energy solutions in rural Pakistan. EE targets areas that are off-grid (no electricity) or have a “bad” grid (more than 12 hours of load-shedding daily). The sample includes 726 EE customers whose systems were activated between March 2017 and December 2018. Our study followed EE’s expansion in new areas in the province of Sindh, specifically the southern districts of Thatta, Badin, Sujawal, Mirpur Khas, and Tando Muhammad Khan. Except for Mirpur Khas, these are some of the poorest districts of Pakistan, with approximately half of their population living below the official poverty line.4 Average household income in sample districts is approximately PKR 9,000 (USD 267 PPP) per month.5 Economic activity in these districts is predominantly agricultural, employing between 50 and 70 percent of the labor force, and a small percentage of the labor force is self-employed. Travel between villages and larger city centers is often on unclassified roads, adding to travel time and costs. The average distance to the nearest market/commercial center in the study sample is approximately 6 km, which respondents can travel to in under 20 minutes using a combination of private and public transport. 2.2. The Product The solar units provided by EE are capable of charging, depending on the model, up to a 17 Ah battery. They can power multiple bulbs or fans, two mobile phones via a USB charger, a radio, or a 15” TV. The systems’ electricity-generating capacity depends on their size and the weather. Depending on the cloud cover, efficiency can drop between 10 and 25 percent of the energy output on a sunny day. A system’s capacity determines the daily rate charged to users, which may range between USD 0.26 and 1.7 PPP. EE first conducts product demonstrations at the village or bazaar (market) level in each area where it enters. At the end of the demonstrations, EE field staff meet interested individuals and businesses one to one. Applicants are screened twice: first through a quick questionnaire conducted by the salesperson to determine whether they can afford the system that meets their needs based on their current energy expenditures, and second by the local sales manager to determine whether the application should be approved. If approved, an “order” is created, and a visit by an operational staff member to install the system and sign the contract is scheduled. When signing the contract, customers choose between perpetual rental and rent-to-buy. The rent-to-buy contract transfers ownership of the unit after the customer has made payments roughly equivalent to its sales value. When signing the contract, customers must make a down 4 As reported in the Government of Pakistan’s Data for Pakistan Portal (http://www.data4pakistan.com/). The proportions of the population below the poverty line for Thatta, Badin, Sujawal, Mirpur Khas, and Tando Muhammad Khan are 51 percent, 47 percent, 52 percent, 41 percent, and 49 percent, respectively, as of 2014. The official poverty line in Pakistan in 2014 was based on recommended nutritional requirements of 2,350 calories per person per day. 5 All PKR values reported in USD PPP use the 2018 World Bank PPP conversion factor for the private consumption rate of USD 1 = PKR 33.54. 624 Bonan et al. payment proportional to the system’s value. Rental customers need to make a security deposit equivalent to one month’s rent and pay upfront the rent for the first month, which includes the cost of electricity. After the installation, a client profile is created in EE’s billing system, providing details of customer balance and payments. This implies that the study sample does not include individuals who did not com- plete the installation process. In the study area, EE has, on average, five customers per village, likely to have adopted systems of different capacities. Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 The contract involves a PAYG payment system. Customers purchase access to the electricity generated by the system through top-ups. Customers top up at mobile money (Easypaisa) agents, typically located in larger villages and market centers. Smartphone and mobile money penetration allowing direct customer top-up was very low at the time of the study. The size of the payment, divided by the system’s daily rate, determines the number of days of electricity paid for through the top-up. The daily rate gets deducted from the outstanding credit, regardless of whether or how much electricity the system uses or produces daily. Credit, therefore, runs down continuously, and customers cannot store their credit when they do not use electricity or when the system’s generation capacity is low due to bad weather. SMS reminders notify customers when their credit is about to expire and invite them to top up at the monthly rate. SMSs are also sent to acknowledge when a payment is made and report the number of days before the next payment is due. Once credit expires, the solar unit is remotely disconnected and produces no electricity. After about 30 consecutive inactive days, customers have their status turned into “default.” At this point, EE is free to repossess the product and cancel the contract. In practice, EE’s sales team contacts customers after the 30th consecutive inactive day to notify them that the system will be repossessed unless a payment is made. The possibility of negotiations between EE and the customer implies that inactivity spells longer than 30 consecutive days may not always be followed by a cancellation. Besides eventual contract cancellation, there was no financial penalty for not immediately topping up once the credit expired. 3. The Experiment 3.1. Experimental Design The experimental design is articulated in two dimensions: the disclosure of the possibility of setting the repayment schedule flexibly and the administration of a planning intervention. This results in a 2 × 2 factorial design. Both treatments are administered once at the start of the contract. Default contract: Under EE’s default, customers are informed of the daily rate associated with their system. They are then told that they are expected to pay every month an amount corresponding to the monthly rate, computed from the daily rate under the assumption of no inactive days, and are provided with examples.6 This forms the control group, referred to as the “Fixed-no IIP” group, in the following discussion. Flexibility disclosure (Flex): In the first treatment dimension, customers are explicitly informed that the monthly payment, corresponding to their system’s daily rate, can be paid in smaller and more frequent installments within each month. These customers are still told that they are expected to pay a certain amount each month, but in addition to the information provided in the Fixed-no IIP group, they learn 6 Specifically, the script reads: “Your plan costs a daily rate of Rs [calculated rate]. You are expected to pay the total amount of Rs XX every month. This means that, for example, if you are connected on February 3rd, you are expected to pay Rs XX by March 3rd,” where XX corresponds to the monthly rate. Supplementary online appendix S1 reports the script used to present the treatments to the customers. The World Bank Economic Review 625 that they are essentially free to plan their payment schedule. For illustration purposes, these customers are given examples of payments at different frequencies (e.g., weekly, bi-weekly, monthly).7 Customers in both groups (Fixed-no IIP and Flex) can pay at any frequency. The flexibility disclosure treatment makes the possibility of setting the payment schedule autonomously clearer to the customer at the start of the contract relative to the default. Implementation Intention Plan (IIP): A second treatment dimension offers a planning nudge to cus- Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 tomers, drawing from the psychology literature on the use of implementation plans (Gollwitzer and Sheeran 2006). Customers in the IIP treatment are asked to state their commitment to making timely payments; identify the main obstacles they face in meeting this commitment; formulate strategies for overcoming each obstacle; and consolidate the resulting saving plan and payment schedule by circling the corresponding dates on a calendar, delivered by the enumerator, to be kept by the customers in their workplace or house. Combining Flex and IIP (Flex × IIP): A third treatment arm combines the Flex and the IIP treatments, i.e., customers are explicitly informed about flexibility in making payments and given a planning nudge. Flexibility should help customers match their payments with their cash flow. It can also minimize trans- action costs for making payments. By paying early or late, they could better exploit synergies with trips to the market center rather than going there solely to stick to a rigid payment schedule. In the same way, flexibility may help customers reduce waste by keeping the system inactive when its generation capacity is low. If these were the only mechanisms through which flexibility worked, then it should affect payment and short-term inactivity without leading to higher rates of prolonged inactivity and cancellation. How- ever, flexibility may cause users to unwillingly become inactive and default if it increases the complexity of managing payments and reduces the commitment to timely top-ups. Flexibility may be conducive to cancellation also if users face difficulties resuming payment after periods of inactivity. Planning should foster timely payments and reduce cancellations by encouraging customers to foresee and plan against the main barriers to making top-ups. Also, the planning process could encourage some ex ante learning of the contractual terms. Planning should boost the benefits of flexibility and potentially reduce its negative consequences in several ways: first by helping customers identify and stick to the best payment schedule; second by making it easier for individuals to commit to a payment schedule of their choosing and anticipate potential barriers to adhering to it; third by making payments easier after an inactivity spell, thus mitigating any negative effect of flexibility. However, by drawing users’ attention to the barriers to timely payment, the planning exercise may make them realize how large the costs associated with overcoming such barriers are. Treatment effects should be stronger among users who face greater barriers to timely payment. For instance, if flexibility reduces payment discipline, its negative effects may be stronger for individuals with commitment problems or for whom making payments takes more effort due to high transaction costs. Similarly, the benefits of planning and flexibility may be larger for these individuals. 3.2. Implementation EE’s salespersons administered the flexibility disclosure treatment after all other contractual aspects— electric items, price, and rent versus ownership—were explained, agreed upon, and accepted. The treat- ment assignment, therefore, occurred after the customer had accepted the general contractual conditions. This prevents our design from generating selection into contractual features. Treatment was assigned through a random number generator incorporated into the software used by the salespersons to register new customers. Hence, the flexibility disclosure treatment is stratified by salesperson. 7 Specifically, the flexibility disclosure treatment script adds to the Fixed-no IIP script the following sentences: “You can pay in different installments. This means, for example, that you can pay Rs XX/4 every week or Rs XX/2 every two weeks. In sum, you can top up your credit as many times as you like.” 626 Bonan et al. Following signing a new contract and administering the flexibility disclosure treatment, EE transferred the customer’s information to the research team. An enumerator then visited the customer to administer a survey and the planning treatment. The research team randomized the planning nudge via the survey software and conducted it within a few weeks of the contract start date. Out of the 726 customers in the sample, 155 are in the Flex treatment, 216 are in the IIP treatment, and 195 received both the Flex and IIP treatments. The control group, who received the default, Fixed-no Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 IIP contract, is formed from 160 customers. The slight imbalance in the number of customers assigned to each cell is due to the randomization of each treatment occurring at the individual level through two separate procedures. 4. Data 4.1. Administrative Data EE’s administrative records detail customers’ subscriptions, the type of system installed, the amount due, deadlines, and flows of payments made daily. These data allow the timely monitoring of payments and system status (active, inactive, cancelled) from each customer’s installation date, i.e., from March 2017 for the first customers, until the end of the monitoring period, in September 2019. This means that each customer is observed for at least one year after the contract was signed and the treatments administered unless they cancelled their contract in less than a year. The analysis focuses on five main outcomes. Customers’ extensive and intensive product demand is measured through cancellation and inactivity, respectively. Cancellation is a binary variable indicating whether the customer has cancelled their contract in the sample period. Inactivity is captured by the likelihood that a customer’s balance falls to PKR 0 during the contract period, which leads to the system becoming inactive, and by the average number of days in a month when the system is inactive. Inactive days are set equal to 1 for all days following cancellation until the end of the monitoring period. This approach follows the one used in McKenzie and Puerto (2021) for firms whose decision to exit the market is influenced by a treatment: such firms are assigned profits and sales equal to 0 in the post-exit period and are retained in the sample. To capture mechanisms of consumer behavior, the analysis explores inactive spell duration and monthly top-ups. Inactive spell duration is defined as the average number of successive days a customer has zero balance before they make a top-up over the sample period. Since, by definition, a spell has a start and an end date, this variable is defined only over the contract period. Monthly top-ups is the average number of payments made by the customer in a month over the observation period, where top-ups are set equal to 0 in the post-cancellation period. In robustness analysis, inactivity and top-ups are set to “missing” after cancellation. 4.2. Survey Data The second data source is the survey, administered within a few weeks from contract signing and con- ducted by an independent survey firm with the customers. The survey data provide information on cus- tomers’ demographic and socioeconomic characteristics; their energy usage and other household expendi- tures; their performance in repaying loans in the past; and on a set of behavioral measures, including time preferences, measures of grit, locus of control, self-control, and willpower to resist temptations. Measures constructed from the survey variables are used for heterogeneity analysis and as controls in the regression models. The analysis focuses on two measures expected to be associated with cancellation and payment and heterogeneous treatment impacts. First is an index for mental constraints, constructed from measures of (a) cognitive skills, (b) self-reported ability to pay bills on time, (c) an index for the (in)ability to resist The World Bank Economic Review 627 temptations, self-control, locus of control, grit and discipline with previous loans.8 Second is a measure of transaction costs, proxied through distance from the nearest Easypaisa agent where payments can be made. 5. Descriptive Analysis Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Descriptives of customers’ characteristics and behavior regarding contract cancellation, inactivity, and top-ups, and the correlation between payment behavior and cancellation are first provided. This is impor- tant for two reasons. First, given the novelty of these products, little is known about the characteristics and behavior of their users. Second, detecting user behavior patterns can inform the interpretation of the empirical results.9 5.1. Sample Characteristics Sample characteristics are generally balanced across treatment arms, as shown through F-stats for joint significance of treatment groups dummies (column 5 of table 1).10 Almost all (95 percent) of the individuals in the sample are residential customers. On average, cus- tomers are 36 years old, 82 percent are literate, about 25 percent have savings, and 17 percent have access to credit. Nearly a third of respondents earn income primarily from agricultural activities, another 27 percent are government employees (with regular monthly salaries), while 19 percent are laborers, earn- ing irregular, often weekly wages. A small percentage (13 percent) are self-employed. Overall, almost half of the members of respondents’ households are recipients of regular income flows. Grid connectivity is low, and the electricity quality is poor. About 20 percent of individuals in the sample have no access to any electric power source, while 66 percent of them are connected to the national grid or mini-grids, and 14 percent have home solar systems. The vast majority (99 percent) of those connected to the grid experience load-shedding daily: 33 percent between 1 and 6 times, 24 percent between 7 and 10 times, and 41 percent more than 10 times a day. Off-grid households live on average 7 km from the closest on-grid location. Regarding the home solar system, users report high levels of understanding of contract terms and mostly do not anticipate problems in making timely top-ups. The average distance between a customer and the closest Easypaisa agent is 6 km: it takes customers 17 minutes and PKR 47 (USD 1.40 PPP), on average, to cover this distance. Remembering to buy credit, putting aside money for payment, resisting the temptation to divert the money to other uses, and actually making the payment are the main obstacles to timely payment anticipated by customers. Almost all (98.3 percent) systems installed can power a light, while 82.8 percent can power a fan. Only 6.3 percent of customers have systems able to support a TV, while 5.6 percent are endowed with a mobile charger. On average, keeping systems active costs customers USD 37.6 PPP per month, i.e., about 6 percent of monthly household income. In addition, 67.9 percent of the sample choose a perpetual rental over rent-to-own contracts. 8 Each variable, source, and construction is described in detail in supplementary online appendix S2. The variables are first reverse-coded if needed, then aggregated following Anderson (2008) so that the index increases in the degree of mental constraints. The Fixed-no IIP group is used as the “reference/control” group for standardizing. 9 The following analysis is conducted on the whole study sample. However, results are qualitatively similar if restricted to the control sample, i.e., no Flex- no-IIP. Results are not shown but are available upon request. 10 Balance tests are run on 59 respondent and contract characteristics. The share of unbalanced variables with p < 0.1 is 0.086, and the share of unbalanced variables with p < 0.05 is 0.052. Table S3.1 in the supplementary online appendix shows the comprehensive list of contract and respondent characteristics. 628 Bonan et al. Table 1. Contract and Respondent Characteristics at Baseline (1) (2) (3) (4) (5) All Flex IIP Flex × IIP F-stat p Panel A: Respondent characteristics Type of customer: business 0.052 0.052 0.056 0.062 0.738 Respondent age 35.566 36.110 35.194 35.503 0.702 Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Can read and write 0.822 0.832 0.829 0.815 0.954 Any savings 0.253 0.297 0.185 0.262 0.028 Access to credit 0.171 0.174 0.181 0.159 0.949 Share of HH members with regular income 0.457 0.495 0.462 0.442 0.541 Experiences > 10 hours loadshed/day 0.858 0.897 0.880 0.846 0.105 Understands product type 0.815 0.826 0.824 0.805 0.931 Understands payment procedure 0.912 0.897 0.903 0.913 0.512 Does not know daily rate 0.090 0.084 0.079 0.077 0.458 Distance from Easypaisa (payment) agent (km) 6.161 6.103 6.361 6.323 0.662 Anticipate problems to repay on time 0.180 0.174 0.185 0.154 0.555 Main constraint to pay: set aside money 0.394 0.413 0.398 0.390 0.918 Main constraint to pay: keep safe from others 0.326 0.310 0.333 0.303 0.645 Main constraint to pay: resist temptations 0.394 0.458 0.417 0.338 0.108 Main constraint to pay: remember payments 0.431 0.426 0.463 0.374 0.243 Main constraint to pay: go and pay 0.366 0.374 0.370 0.328 0.555 Panel B: Contract characteristics Perpetual (vs rent-to-own) 0.675 0.742 0.685 0.636 0.131 Daily rate (USD PPP) 1.258 1.199 1.221 1.293 0.077 Observations 726 155 216 195 Source: Authors’ analysis based on survey data collected by the authors and EcoEnergy’s administrative data. Note: The table shows mean values of respondent and contract characteristics for the whole sample (column 1) and for each treatment group (columns 2–4). Column 5 reports the p-value of a test of joint significance (F-stat) of two treatment dummies and their interaction on the characteristic in each line. 5.2. User Behavior This subsection discusses the main patterns in contract duration, inactivity, and payment. Their distri- bution is shown in fig. 1, overall and for users who do and do not cancel over the sample period, while descriptive statistics for the main outcome variables of interest for the study sample are provided in table S3.2. Overall, 56 percent of the customers have their contract cancelled due to default. Of the customers who cancel during the sample period, 25 percent and 50 percent cancel in the first 3 and 5 months of the contract, respectively.11 The dynamics of cancellation over time confirm the higher cancellation rates during the early contract months (fig. S3.1). Cancellation is almost exclusively the result of inactivity: only 4 percent of those who cancel do so while holding a positive balance. The analysis restricts attention to the first 18 months of the contract, given that the number of customers in our sample drops below 100 afterwards. Payment quality is generally low: the cross-sectional data show that 96 percent of customers experience at least one inactive day throughout the contract, and their systems are inactive for more than a quarter 11 We cannot comment on whether these cancellation rates are higher or lower than those other providers face. Globally, the definition of default varies, and information on client retention is not publicly available. However, there is a recent push to adopt consistent performance indicators and publish data for greater transparency. EE is the only PAYG off-grid electricity provider in Pakistan. The observed cancellation rate is higher than the one estimated by EE at the time of designing the study. At that time, EE was in the early stages of product expansion, and hence their estimate of customer retention was imprecise. The World Bank Economic Review 629 Figure 1. Distribution of Contract Duration and Payment Behavior Contract duration (days) Monthly inactive days .15 0 .001.002.003.004 .1 Density Density Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 .05 0 0 100 200 300 400 500 600 700 800 900 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 x x Inactivity cycle duration (days) Monthly top−ups 0 .02 .04 .06 .08 .1 2 1 1.5 Density Density .5 0 0 5 10 15 20 25 30 35 40 45 50 0 1 2 3 4 x x All No cancel Cancel Source: Authors’ analysis based on survey data collected by the authors and EcoEnergy’s administrative data. Note: The figures show kernel density plots for the whole sample (“All”) and by cancellation status. The top-left panel depicts the contract duration, in days. The top-right panel shows the distribution of the average number of inactive days in a month. The bottom-left panel depicts the average inactivity cycle duration (excluding the last cycle before cancellation). The bottom-right panel shows the average number of top-ups in a month. All figures are generated from the cross-section (N = 726). of the time (26.2 percent of contract days). This corresponds to an average of 6.1 monthly inactive days. Customers who eventually cancel report, on average, almost twice as many inactive days per month than those who do not cancel during the contract period. This is also confirmed when looking at the dynamics of inactivity over time (fig. S3.1). The average monthly inactive days distribution also differs between users who cancel and those who do not (fig. 1). The median duration of any inactivity cycle before a top-up is made is 4.6 days. Users top up on average more than once a month (mean = 0.94, median = 0.7), and the likelihood of topping up is lower for those who eventually cancel. The average number of days paid for at top-up is 23, and the average amount is PKR 1022 (USD 31 PPP, median USD 27 PPP). The median top-up occurs on the day when the credit expires. 6. Estimation Strategy The following equation is used to investigate treatment effects in the cross-section data set: yi = α + β1 FLEXi + β2 IIPi + β3 FLEXi × IIPi + Xi + γt + εi , (1) where yi are the cancellation, inactivity, and payment outcomes described previously for individual i. These variables are regressed on the FLEX and IIP treatment dummies, and their interaction FLEX × IIP. The 630 Bonan et al. term Xi includes location, salesperson, and enumerator fixed effects and individual controls, including all unbalanced covariates across treatment groups, selected using the post-double LASSO regularization approach of Belloni, Chernozhukov, and Hansen (2011).12 The coefficients of the Flex and IIP dummies in equation (1) capture the main effect of each treatment when administered in isolation. At the same time, the interaction terms report the effect of the combined treatment relative to the sum of the two main effects. A fully interacted model allows us to detect whether Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 the combination of the Flex and IIP treatments has a different impact relative to what one would expect from the effect of the two treatments when administered in isolation. Equation (1) is pre-registered. The pre-analysis plan (PAP) reported additional regression specifications, omitting the interaction term. Results from this estimation strategy are not reported in light of recent work recommending estimating a saturated model with interactions between treatment cells in the presence of fully factorial experimental designs like ours. This is important both for correct inference (Muralidharan, Romero, and Wüthrich 2020) and for policy implications (Banerjee et al. 2021). Regression tables report whether the marginal effect of the combined treatment is significantly different from the control (Fixed-no IIP) by displaying the p-value of a Wald test of equality of coefficients. While the regression specification adheres to the pre-specified one, the analysis presented here differs in some respects from the one featured in the PAP. This subsection summarizes and motivates these de- partures. Readers are referred to supplementary online appendix S4 for a detailed description. A first set of deviations from the PAP is motivated by the desire to conform to best practices estab- lished since writing the PAP. The first, already mentioned, concerns the correct estimation of treatment effects from factorial designs. A second, discussed above, concerns our attrition treatment due to con- tract cancellation. Setting inactivity and top-ups equal to zero in the post-cancellation period aligns with the literature (McKenzie and Puerto 2021).13 This approach is preferable to the one envisioned in the PAP, i.e., Lee bounds on treatment effects (Lee 2009), as Lee bounds are not the correct tool to deal with selection on the extensive margin, particularly in the presence of multiple treatments with potentially non-monotonic effects. A third deviation relates to the approach used for multiple hypothesis testing. The approach adopted in the paper is more conservative than the one specified in the PAP: it reports sharp- ened q-values (Benjamini, Krieger, and Yekutieli 2006) across all the outcomes variables of interest, not just within the family of extensive and intensive margin outcomes, and across the dimensions of hetero- geneity. Finally, while the PAP does not specify the control variables to be included in the regressions, the analysis avoids cherry-picking them and instead selects them through a data-driven approach—the post-double-selection LASSO procedure (Belloni, Chernozhukov, and Hansen 2014). A second source of deviation from the pre-registered analysis is due to the focus on two pre-specified dimensions of heterogeneity in the main text, with results for the rest of the dimensions shown in the appendix.14 The PAP pre-specified many heterogeneity dimensions speculated ex ante as potentially rel- evant. The leading view nowadays on pre-registration is that it should be used with moderation and be focused only on core outcomes and heterogeneity dimensions (Banerjee et al. 2020). The analysis in the main text presents only a subset of these results, also because our ability to pursue any heterogeneity analysis is hampered by low power. 12 Individual controls are selected among age, literacy, knowledge of contract rules and rate, availability of savings, men- tal constraints (index), ability to smooth consumption (index), time-inconsistent preferences, the system’s daily rate, rental versus rent-to-own contract, and distance from the Easypaisa agent. Supplementary online appendix S2 provides descriptions of variables. 13 We thank an anonymous reviewer for encouraging us to address the issue of selective attrition through the correct definition of our variables and sample. 14 As discussed in supplementary online appendix S4, results on two dimensions are not reported because they display no variability in answers. The World Bank Economic Review 631 Table 2. Treatment Effects on Cancellation and Payment Behavior (1) (2) (3) (4) (5) Extensive margin Intensive margin Mechanisms At least one inactive Avg. monthly Avg. duration of Avg. number of Cancel day inactive days inactive spells (days) monthly top-ups Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Flex 0.098∗ 0.040∗∗ 2.042∗ − 0.932 − 0.153 (0.052) (0.020) (1.094) (1.291) (0.120) [0.206] [0.206] [0.206] [0.434] [0.231] IIP 0.064 0.030 1.310 − 0.627 0.010 (0.048) (0.020) (0.987) (1.283) (0.117) [0.231] [0.221] [0.231] [0.442] [0.491] Flex×IIP −0.208∗∗∗ − 0.061∗∗ − 3.324∗∗ 0.027 0.308∗ (0.070) (0.029) (1.456) (1.746) (0.170) [0.053] [0.206] [0.206] [0.491] [0.206] Observations 726 726 726 726 726 no Flex-no IIP group 0.525 0.950 13.68 9.013 0.971 mean p-val of 0.368 0.700 0.979 0.225 0.278 Flex+IIP+Flex×IIP Source: Authors’ analysis based on survey data collected by the authors and EcoEnergy’s administrative data. Note: The unit of analysis is a customer. The extensive margin variable, “Cancel,” takes a value of 1 if contract cancellation occurs and 0 otherwise; the intensive margin variables “At least one inactive day” and “Avg. monthly inactive days” are dummy variables for whether customers experienced at least one inactive day and the number of inactive days in the month, respectively. “Avg. duration of inactive spells (days)” is the average duration of inactivity spells over the contract period; “Avg. number of monthly top-ups” is the average number of top-ups per month over the contract period. “Average monthly inactive days” and “Avg. number of monthly top-ups” are set equal to the total days in a month and to zero after cancellation, respectively. Flex takes the value of 1 if the individual is offered the flexibility disclosure treatment and 0 otherwise; IIP takes the value of 1 if the individual is offered the intention implementation plan nudge and 0 otherwise. Estimates are obtained via OLS. All specifications include individual controls selected through LASSO between daily rate at the contract start, rental contract at start, respondent’s age, respondent can read and write, any savings, knows the contract rate, knowledge of system rules, distance from Easypaisa agent, index for mental constraints, index for ability to smooth consumption, time-inconsistent preferences; month, location, enumerator, and salesperson fixed effects; robust standard errors are in parentheses. No Flex-noIP group mean refers to the control group mean. Sharpened q-values in square brackets control the false discovery rate for tests across pre-specified outcomes. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1. A third, large share of deviations from the PAP are due to errors or insufficient knowledge of the administrative data during pre-registration. Indeed, the PAP was registered before having access to the administrative data. These deviations encompass the inclusion in the PAP of (a) outcome variables for which no administrative data are available, e.g., contract cancellation before the installation of the system; (b) multiple variables to capture the same phenomenon, e.g., delayed payments and system deactivation due to missed payment are both inactivity; (c) heterogeneity dimensions for which there are insufficient data or power, such as the distinction between business and residential customers. 7. Results 7.1. Treatment Effects This subsection examines treatment effects on pre-specified dimensions of consumer behavior. Specifically, it reports intent to treat results from the estimation of equation (1) on the extensive and intensive margins of consumer demand: cancellation, the probability of experiencing inactivity, and the number of inactive days per month (table 2). Cancellation: Disclosing flexibility marginally reduces the extensive margin of demand, increasing monthly cancellation rates by about 10 percentage points, almost a 19 percent increase over the con- trol means. The planning treatment alone has no significant effect on cancellation. The combination of 632 Bonan et al. flexibility and planning more than offsets the effect of the former, and the cancellation rates in the com- bined treatment do not significantly differ from those in the control treatment. This result is robust to correcting the p-values for testing for multiple hypotheses (sharpened q-values are displayed in square brackets in table 2). The behavioral literature on planning prompts outlines two conditions behind their effectiveness (Rogers et al. 2015). First, prompting people to focus on the obstacles to follow through on their in- Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 tentions and devise a plan to overcome them works only if individuals consolidate their strategy into a plan. Second, planning helps when there are obstacles to overcome. In our setting, flexibility increases the complexity of payment decisions. Consistent with the greater need to make a plan in the presence of a self-imposed payment schedule, the share of customers assigned to the Flex × IIP treatment consolidating their plan on the provided calendar is significantly higher than those in the IIP treatment (p = 0.034).15 Our treatment effects also support this claim: only individuals assigned to the flexible payment schedule benefit from planning, i.e., planning only impacts cancellation when subjects have to devise their payment schedule. Since both treatments in our experiments were administered once at the beginning of the contract and never repeated, one may expect their effects to vary over time. Decreasing impacts over time may be expected if behavioral interventions, such as planning, have short-lived effects (Gneezy and List 2006; Allcott 2016); if customers learn about the flexibility built into the standard contract over time; or if selective attrition makes the individuals who are most responsive to the treatments drop out. On the contrary, one would expect persistent treatment effects if behavior at the contract’s start informs different habits (Schaner 2018). Though not pre-specified, estimating equation (1) using a monthly panel over non- overlapping three-month periods with month fixed effects allows us to study treatment dynamics. The effect of the combined treatment does not appear to fade over time, both when considering cumulative effects by retaining in the sample the customers who cancel and setting cancellation to 1 for the months post-cancellation (table S3.3), and when focusing on the selected sample of customers who are still active in each period (table S3.4). Inactivity: First, the intensive margin of product demand is studied in terms of the probability of expe- riencing at least one inactive day over the contract period. Results in column 2 of table 2 show that the treatment impacts on the likelihood of ever experiencing inactivity mirror those on cancellation: flexibility increases such likelihood, but combining it with planning offsets this effect. This makes sense, as inactiv- ity is the main cause of cancellation. A possible interpretation of these results is that flexibility reduces payment discipline and increases inattention, thus causing users to miss payment deadlines. However, planning prevents this negative consequence of flexibility by focusing users’ attention on their payment deadlines. If the inactivity induced by flexibility were exclusively beneficial, i.e., a way to minimize waste and transaction costs, then one would not expect it to disappear, thanks to planning. Similar treatment effects are observed on the average number of inactive days in a month (column 3). Flexibility increases them by two days on average while combining it with planning restores inactivity levels to those in the Fixed-no IIP group. The results on the two inactivity measures are not robust to the multiple hypothe- sis correction. This suggests taking the evidence on the treatment impacts on the intensive margin with caution. The results on the number of monthly inactive days are also no longer statistically significant when the days after cancellation are set to “missing” (table S3.5). This suggests that users whose cancellation 15 The data for this analysis come from the survey with the customers where the enumerators noted whether they used the calendar provided by the enumerators themselves to consolidate their planned payment schedule. Note that no data are available on the exact dates circled by customers on the calendars. The World Bank Economic Review 633 decision is not influenced by the treatments do not change their inactivity levels on average.16 Results are similarly different between the balanced and selected sample when looking at the dynamics of treatment effects over non-overlapping periods: in the balanced panel, the combined treatment has persistent neg- ative cumulative effects on inactivity (table S3.7), while in the panel of active customers, it only affects monthly inactive days in the first three months of the contract (table S3.8). Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 7.2. Exploring Mechanisms The analysis of potential mechanisms takes two approaches. The first consists in estimating treatment effects on two outcomes that could explain consumer behavior: the average duration of inactive spells and the number of monthly top-ups. Inactive spell duration can reveal whether the treatments allow users to use inactivity to reduce transaction costs—e.g., through brief inactivity spells until the next trip to the market center—or whether they cause customers to fall into long inactivity spells and eventual default. The treatments should directly influence the number of monthly top-ups. The analysis of these two additional outcomes produces null results, except for a marginally significant increase in top-ups in the combined treatment, relative to the sum of the effects of flexibility and planning in isolation (columns 4–5 of table 2).17 The second approach involves exploring heterogeneous treatment effects on the outcomes of interest. Heterogeneity is examined by augmenting equation (1) with the interaction of treatment indicators and two dimensions of heterogeneity: indicators for above-median mental constraints and transaction costs. Mental constraints is an index of various measures (detailed in the Data section) that could pose difficulties in maintaining discipline in payments. “High” mental constraints is a dummy equal to 1 when the value of this index is greater than the sample median. “High” transaction costs are captured by a dummy equal to 1 if a customer lives further than 5 km—the median distance—from the closest Easypaisa agent. The results do not survive corrections for testing multiple hypotheses, but they provide suggestive insights into the characteristics of customers who benefit the least from the product and the treatments. The flexibility disclosure treatment increases cancellation rates more among customers with above- median than below-median mental constraints (column 1 of table 3). This supports the hypothesis that people with high mental constraints may struggle to keep track of payments when the possibility of making them flexibly is salient. Users facing higher transaction costs are also more likely to cancel than those facing low transaction costs when exposed to the flexibility disclosure treatment, and also significantly less likely to cancel when flexibility and planning are combined (column 2 of table 3).18 Similarly, among users with high mental constraints and transaction costs, flexibility in isolation increases the number of monthly inactive days (columns 5–6). However, flexibility increases the likelihood of experiencing inactivity among low transaction cost users but not among high transaction cost ones (columns 3–4). There is no significant heterogeneity on other outcome variables—the average duration of inactive spells and the number of monthly top-ups—likely due to lack of statistical power.19 The heterogeneity analysis presented here focuses on a subset of the potential sources of heterogeneous treatment effects mentioned in the PAP. Specifically, the following other sources of heterogeneity were pre- specified: the ability to smooth consumption, financial management, and time inconsistency. There is no 16 Treatment effects are null in the cross-section for another pre-specified measure of inactivity, the average number of monthly inactivity spells, reported in column 1 of table S3.6. This variable also considers only the pre-cancellation period. 17 An additional outcome variable considered, though not pre-specified, is the average size of the payments. One hypothesis could be that those in the flexibility disclosure treatment make smaller, more frequent payments. This is not the case (column 2 of table S3.6): if anything, flexibility leads to a marginally significant increase in the average payment size. 18 There is no evidence of heterogeneous treatment effects by these dimensions on average payments. 19 The heterogeneous effects on the average number of inactivity cycles in a month are also insignificant (table S3.9). 634 Table 3. Heterogeneity in Treatment Effects on Cancellation and Payment Behavior (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Extensive margin Intensive margin Mechanisms Cancel At least one inactive day Avg. monthly inactive days Avg. duration of inactive spells (days) Avg. number of monthly top-ups H: High MC High TC High MC High TC High MC High TC High MC High TC High MC High TC Flex − 0.007 0.013 0.055∗ 0.072∗∗∗ − 0.734 0.239 − 0.883 − 0.765 0.029 − 0.166 (0.077) (0.069) (0.030) (0.025) (1.662) (1.485) (2.067) (1.672) (0.186) (0.152) IIP 0.137∗ 0.057 0.054∗ 0.053∗ 1.880 1.073 − 2.521 1.253 0.073 − 0.047 (0.070) (0.063) (0.031) (0.028) (1.487) (1.321) (1.760) (1.994) (0.175) (0.154) Flex×IIP − 0.210∗∗ − 0.085 − 0.087∗∗ − 0.102∗∗∗ − 2.101 − 0.682 1.298 − 1.151 0.195 0.344 (0.102) (0.092) (0.042) (0.038) (2.142) (1.943) (2.571) (2.640) (0.250) (0.243) Flex×H 0.200∗ 0.173∗ − 0.028 − 0.076∗∗ 5.323∗∗ 3.624∗ − 0.149 − 0.460 − 0.346 0.028 (0.104) (0.105) (0.041) (0.036) (2.206) (2.172) (2.691) (2.639) (0.268) (0.226) [0.485] [0.485] [0.835] [0.485] [0.485] [0.485] [1.000] [1.000] [0.568] [1.000] IIP×H − 0.148 0.011 − 0.046 − 0.056 − 1.243 0.362 3.720 − 3.853 − 0.112 0.113 (0.097) (0.099) (0.040) (0.038) (1.991) (2.007) (2.734) (2.636) (0.252) (0.225) [0.485] [1.000] [0.65] [0.485] [0.835] [1.000] [0.533] [0.485] [0.945] [0.907] Flex×IIP×H 0.013 − 0.258∗ 0.051 0.097∗ − 2.171 − 5.387∗ − 2.521 2.159 0.207 − 0.057 (0.140) (0.140) (0.060) (0.051) (2.926) (2.883) (3.372) (3.482) (0.369) (0.334) [1.000] [0.485] [0.825] [0.485] [0.835] [0.485] [0.835] [0.835] [0.871] [1.000] Observations 726 726 726 726 726 726 726 726 726 726 Source: Authors’ analysis based on survey data collected by the authors and EcoEnergy’s administrative data. Note: The unit of analysis is a customer. Dependent variables are as defined as in table 2. Flex takes the value of 1 if the individual is offered the flexibility disclosure treatment and 0 otherwise; IIP takes the value of 1 if the individual is offered the implementation intention plan nudge and 0 otherwise. Dimensions of heterogeneity are binary variables. The dimension is “high mental constraint” (High MC) for odd-numbered columns, which is a binary variable = 1 if the value of the index is greater than the sample median. The dimension is “high transaction costs” (High TC) for even-numbered columns, which is a binary variable = 1 if the distance to an Easypaisa agent is greater than 5 km. All specifications include individual controls selected through LASSO between daily rate at the contract start, rental contract at start, respondent’s age, respondent can read and write, any savings, knows the contract rate, knowledge of system rules, distance from Easypaisa agent, index for mental constraints, index for ability to smooth consumption, time inconsistent preferences; month, location, enumerator, and salesperson fixed effects; robust standard errors are in parentheses. Sharpened q-values in square brackets control the false discovery rate for tests across heterogeneous effects (the test only considers the two dimensions in this table). ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1. Bonan et al. Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 The World Bank Economic Review 635 compelling evidence of heterogeneity along these dimensions. Results are available in table S3.10 and S3.11. 8. Robustness This section discusses factors potentially influencing the magnitude and statistical significance of the em- Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 pirical results: statistical power, selective attrition, and spillover effects. Assessing statistical power is critical in experimental research, as low power increases the risk of false negatives, but also of false positives, where non-existent effects are detected. Low statistical power may be due to the sample size or the treatment effects being relatively small (Ioannidis 2005). The present study may suffer from both sources of low power. First, the initial sample size, determined through power calculations, was based on lower attrition rates due to contract cancellation than the ones observed. Second, the treatments were light touch and administered only once at the contract’s start. Therefore, ex post minimum detectable effects (MDEs) are presented and discussed in what follows.20 The estimated MDEs for the average duration of inactive spells and the number of top-ups range be- tween 34 and 54 percent. These are quite high and may indicate that the results could not reach statistical significance due to low power rather than due to precisely estimated null effects. Looking at the effect of broadly similar interventions on different outcomes, these MDEs appear to fall short of power.21 Selective attrition may attenuate treatment effects if it drives the individuals most responsive to the treatments out of the sample. Selective attrition is a serious concern in our setting. Our results show that the treatments affect contract cancellation and that the treatment effects on cancellation are stronger for customers with specific characteristics. The analysis addresses this issue in three ways. First, as already mentioned, the primary definition of the outcome variables related to inactivity and top-ups retains in the sample all customers, even after they cancel their contracts. Therefore, the main sample of analysis is not the selected sample of active customers. Second, the potential role of selective attrition is assessed through robustness analysis, where all outcome variables in the post-cancellation period are set to “missing”. The comparison between the (statistically significant) results from our main analysis and the (insignificant) results from the robustness check suggests that selective attrition due to the treatments results in a sample where customers, whose inactivity and payment decisions are not significantly affected by the treatments, to be observed for more extended periods. Third, balance tests reveal whether and how selective attrition affects observable characteristics within our sample. The same balance tests, displayed in table 1 for the baseline sample, are run for customers still active after 12 months of contract (table S3.12). Customers in the flex treatment who cancelled within the first year are less literate, live further from Easypaisa (although non-significantly), and are more subject to cognitive bias (e.g., perceiving memory issues as a constraint to pay). This is consistent with the results from the analysis of heterogeneity. Finally, spillovers may attenuate the treatments’ effect by reducing the observed differences in behavior across customers assigned to different treatments. Customers assigned to the Flex treatment may have discussed their preferences over payment frequency and inactivity with customers in the control group, increasing the similarity in behavior across the two groups. The concern of spillovers across customers is reduced by two features of the setting, which make learning across customers unlikely. First, the systems’ daily rates vary within each village, as each customer’s needs primarily drive them. Second, the small 20 The MDE is the effect that we would have been able to detect with 80 percent power at the 5 percent rate of significance level ex post with the study sample. Ex post MDEs are computed as the standard error (β )∗2.8 and reported as a percentage of the control group mean. We follow Haushofer and Shapiro (2016) and report MDE only for non-significant parameters. 21 Della Vigna and Linos (2020) review behavioral intervention trials from academic literature and nudge units and find average MDEs of 33.4 and 8 percent, respectively. Specific trials on reminders and planning prompts have an average relative MDE of 9.3 percent within nudge units’ trials and 20 percent within academic papers. 636 Bonan et al. number of customers per village (approximately 5, amounting to less than 10 percent of the average number of households in a village) makes customer interactions unlikely. 9. Conclusion Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 This study uses administrative data to study payment behavior for a PAYG solar system provided by a for-profit organization to individuals living in off-grid areas of rural Sindh, Pakistan. Credit runs down continuously, and the system is remotely disconnected when customer credit expires. A randomized con- trol trial with this sample is used to study the impact of disclosing flexibility on customer behavior. While flexibility should help poor households reduce waste and transaction costs and manage payments in the presence of volatile income flows, it may hurt repayment quality if rigidity provides discipline. The com- plexity of contractual terms increases the cognitive costs of adhering to them. The combination of flexi- bility disclosure with a planning intervention aims to address flexibility’s potentially negative behavioral consequences. Overall, the empirical results support the relevance of the negative behavioral repercussions of flexibil- ity on the extensive margin of customer demand, measured by contract cancellation. Disclosing flexibility marginally increases the likelihood of cancellation compared to the default contract, where customers are told to make monthly payments. The planning intervention has no significant effect on any dimension of payment behavior when added to the default contract. Combining flexibility disclosure with the planning nudge significantly reduces cancellation rates compared to flexibility alone. The analysis produces sugges- tive evidence of similar treatment effects on the likelihood that a customer ever experiences inactivity and on the number of inactive days. The impact of flexibility disclosure on cancellation is more pronounced for customers with difficulties maintaining financial discipline and living far from the payment agent. There are no treatment effects on the average duration of inactive spells and the number of monthly payments. While our main specification should attenuate concerns of selective attrition, high contract cancellation rates may make the analysis underpowered to detect impacts on these outcomes. Future research can build on these results to formulate and test a theory of change linking contractual features, behavioral outcomes in terms of inactivity and payment, and contract cancellation. In a world where behavioral nudges are increasingly used in the public and private sectors, individu- als are likely to be exposed to multiple nudges at any one time: understanding the impacts of different policy combinations is therefore important. Our results show that potential unintended consequences of nudges could be prevented by combining them with other behavioral interventions. From a method- ological standpoint, these results confirm the importance of considering interactions between orthogonal treatment dimensions. The study results on the relevance of behavioral payment determinants also imply that PAYG system providers face a trade-off between providing information on complex contractual features and foster- ing timely payments and customer retention. Contract information transparency may harm customers if access to electricity is welfare enhancing. Our results suggest that a potential solution to these adverse effects may be complementing contractual disclosure with soft interventions that encourage making pay- ment plans. Unfortunately, this study cannot speak to the welfare effects of our treatments or of electricity access more broadly. Similarly, our design does not allow us to evaluate users’ demand for flexibility or planning aids in our setting. These represent interesting and important areas for future research. Declarations of Interest None. The World Bank Economic Review 637 Data availability Anonymised, unit-level data along with the do files that are required for replication of the findings pre- sented in this paper can be provided on request. Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 References Abel, M., R. Burger, E. Carranza, and P. Piraino. 2019. “Bridging the Intention-Behavior Gap? The Effect of Plan- Making Prompts on Job Search and Employment.” American Economic Journal: Applied Economics 11(2): 284– 301. Afzal, U., G. d’Adda, M. Fafchamps, S. Quinn, and F. Said. 2018. “ Two Sides of the Same Rupee? Comparing Demand for Microcredit and Microsaving in a Framed Field Experiment in Rural Pakistan.” Economic Journal 128(614): 2161–90. Aklin, M., P. Bayer, S. P. Harish, and J. Urpelainen. 2017. “ Does Basic Energy Access Generate Socioeconomic Benefits? A Field Experiment with Off-Grid Solar Power in India.” Science Advances 3(5): e1602153. Allcott, H. 2016. “ Paternalism and Energy Efficiency: An Overview.” Annual Review of Economics 8(1): 145–76. Allcott, H., and J. B. Kessler. 2019. “ The Welfare Effects of Nudges: A Case Study of Energy Use Social Comparisons.” American Economic Journal: Applied Economics 11(1): 236–76. Anderson, M.L. 2008. “ Multiple Inference and Gender Differences in the Effects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects.” Journal of the American Statistical Association 103(484): 1481–95. Banerjee, A., A. G. Chandrasekhar, S. Dalpath, E. Duflo, J. Floretta, M. O. Jackson, H. Kannan et al., 2021. “Selecting the Most Effective Nudge: Evidence from a Large-Scale Experiment on Immunization.” NBER Working Paper No. 28726. Banerjee, A., E. Duflo, A. Finkelstein, L. F. Katz, B. A. Olken, and A. Sautmann. 2020. “In Praise of Moderation: Suggestions for the Scope and Use of Pre-analysis Plans for RCTs in Economics.” Working Paper 26993, National Bureau of Economic Research. Barboni, G. 2017. “Repayment Flexibility in Microfinance Contracts: Theory and Experimental Evidence on Take Up and Selection.” Journal of Economic Behavior & Organization 142: 425–50. https://www.sciencedirect.com/scienc e/article/pii/S0167268117301981. Barboni, G., and P. Agarwal. 2019. “How do flexible microfinance contracts improve repayment rates and business outcomes? Experimental Evidence from India, with Parul Agarwal” IFMR-LEAD Working Paper. Barry, M. S., and A. Creti. 2020. “Pay-as-You-Go Contracts for Electricity Access: Bridging the ‘Last Mile’ Gap? A Case Study in Benin.” Energy Economics 90: 104843. https://www.sciencedirect.com/science/article/pii/S0140988 320301833. Belloni, A., V. Chernozhukov, and C. Hansen. 2013. “Inference for High-Dimensional Sparse Econometric Mod- els.”In D. Acemoglu, M. Arellano, & E. Dekel (Eds.) Advances in Economics and Econometrics: Tenth World Congress (Econometric Society Monographs) pp. 245–295. Cambridge: Cambridge University Press. doi 10.1017/CBO9781139060035.008. ———. 2014. “ High-Dimensional Methods and Inference on Structural and Treatment Effects.” Journal of Economic Perspectives 28(2): 29–50. Benjamini, Y., A. M. Krieger, and D. Yekutieli. 2006. “ Adaptive Linear Step-Up Procedures that Control the False Discovery Rate.” Biometrika 93(3): 491–507. Bensch, G., M. Grimm, M. Huppertz, J. Langbein, and J. Peters. 2018. “Are Promotion Programs Needed to Establish Off-Grid Solar Energy Markets? Evidence from Rural Burkina Faso.” Renewable and Sustainable Energy Reviews 90: 1060–68. Bertrand, M., and A. Morse. 2011. “ Information Disclosure, Cognitive Biases, and Payday Borrowing.” Journal of Finance 66(6): 1865–93. Bicchieri, C., and E. Dimant. 2022. “Nudging with Care: The Risks and Benefits of Social Information.” Public Choice: 191: 443–464. Bonan, J., S. Pareglio, and M. Tavoni. 2017. “ Access to Modern Energy: A Review of Barriers, Drivers and Impacts.” Environment and Development Economics 22(5): 491–516. 638 Bonan et al. Brune, L., Ginè X., and Karlan D.. 2022, November. “Give Me a Pass: Flexible Credit for Entrepreneurs in Colombia.” Working Paper 30634, National Bureau of Economic Research. (unpublished) Bryan, G., D. Karlan, and S. Nelson. 2010. “ Commitment Devices.” Annual Review of Economics 2(1): 671–98. Chetty, R., A. Looney, and K. Kroft. 2009. “ Salience and Taxation: Theory and Evidence.” American Economic Review 99(4): 1145–77. Czura, K. 2015. “Do Flexible Repayment Schedules Improve the Impact of Microcredit? Evidence from a Randomized Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Evaluation in Rural India.” Technical report, University of Munich Discussion Papers in Economics no. 26608. (unpublished) Della Vigna, S., and E. Linos. 2022. “RCTs to Scale: Comprehensive Evidence from Two Nudge Units.” Econometrica 90: 81–116 Field, E., and R. Pande. 2008. “Repayment Frequency and Default in Microfinance: Evidence from India.” Journal of the European Economic Association 6(2–3): 501–9. Field, E., R. Pande, J. Papp, and N. Rigol. 2013. “ Does the Classic Microfinance Model Discourage Entrepreneurship among the Poor? Experimental Evidence from India.” American Economic Review 103(6): 2196–226. Girardeau, H., A. Oberholzer, and S. K. Pattanayak. 2021. “The Enabling Environment for Household Solar Adoption: A Systematic Review.” World Development Perspectives 21: 100290. https://www.sciencedirect.com/science/articl e/pii/S2452292921000047. Gneezy, U., and J. A. List. 2006. “ Putting Behavioral Economics to Work: Testing for Gift Exchange in Labor Markets Using Field Experiments.” Econometrica 74(5): 1365–84. GOGLA, Lightening-Global, EAC, and Berenschot. 2018. “Global Off-Grid Solar Market Report Semi-annual Sales and Impact Data.” Technical report, GOGLA and Lightening-Global and EAC and Berenschot. Gollwitzer, P. M., and P. Sheeran. 2006. “Implementation Intentions and Goal Achievement: A Meta-Analysis of Effects and Processes.” Advances in Experimental Social Psychology 38: 69–119. https://www.sciencedirect.com/ science/article/abs/pii/S0065260106380021. Gravert, C. A., and L. Olsson Collentine. 2021. “When Nudges Aren’t Enough: Incentives and Habit Formation in Public Transport Usage.” Journal of Economic Behavior & Organization 190: 1–14 Grimm, M., L. Lenz, J. Peters, and M. Sievert. 2020. “ Demand for Off-Grid Solar Electricity: Experimental Evidence from Rwanda.” Journal of the Association of Environmental and Resource Economists 7(3): 417–54. Grimm, M., A. Munyehirwe, J. Peters, and M. Sievert. 2017. “ A First Step Up the Energy Ladder? Low Cost Solar Kits and Household’s Welfare in Rural Rwanda.” World Bank Economic Review 31(3): 631–49. Groenewoudt, A. C., and Romijn H. A.. 2022. “Limits of the Corporate-Led Market Approach to Off-Grid Energy Access: A Review.” Environmental Innovation and Societal Transitions 42: 27–43. https://www.sciencedirect.com/ science/article/pii/S2210422421001052. Grubb, M. D., and M. Osborne. 2015. “ Cellular Service Demand: Biased Beliefs, Learning, and Bill Shock.” American Economic Review 105(1): 234–71. Guajardo, J. A. 2021. “Repayment performance for pay-as-you-go solar lamps”. Energy for Sustainable Development 63: 78–85 ———. 2019. “ How Do Usage and Payment Behavior Interact in Rent-to-Own Business Models? Evidence from Developing Economies.” Production and Operations Management 28(11): 2808–22. Hagger, M. S., and A. Luszczynska. 2014. “ Implementation Intention and Action Planning Interventions in Health Contexts: State of the Research and Proposals for the Way Forward.” Applied Psychology: Health and Well-Being 6(1): 1–47. Haushofer, J., and J. Shapiro. 2016. “ The Short-Term Impact of Unconditional Cash Transfers to the Poor: Experi- mental Evidence from Kenya.” Quarterly Journal of Economics 131(4): 1973–2042. Ioannidis, J.P. 2005. “ Why Most Published Research Findings Are False.” PLoS medicine 2(8): e124. IRENA. 2020. “Innovation Landscape Brief: Pay-as-You-Go Models.” International Renewable Energy Agency, Abu Dhabi. Jack, K., and G. Smith. 2015. “ Pay as You Go: Prepaid Metering and Electricity Expenditures in South Africa.” American Economic Review 105(5): 237–41. ———. 2020. “ Charging Ahead: Prepaid Metering, Electricity Use, and Utility Revenue.” American Economic Jour- nal: Applied Economics 12(2): 134–68. The World Bank Economic Review 639 Karlan, D., M. McConnell, S. Mullainathan, and J. Zinman. 2016. “ Getting to the Top of Mind: How Reminders Increase Saving.” Management Science 62(12): 3393–411. Labie, M., C. Laureti, and A. Szafarz. 2017. “ Discipline and Flexibility: A Behavioural Perspective on Microfinance Product Design.” Oxford Development Studies 45(3): 321–37. Laibson, D. 2018. “Private Paternalism, the Commitment Puzzle, and Model-Free Equilibrium.” American Economic Review Papers and Proceedings 108: 1–21. Downloaded from https://academic.oup.com/wber/article/37/4/620/7206770 by University of Oxford user on 12 December 2023 Lang, M. 2020. “Consuming Perishable Goods in the Presence of Transaction Costs and Liquidity Constraints.” Working paper ———. 2022. “Using Incentives to Understand Intensive-Margin Demand for Electricity.” Working paper Lee, D.S. 2009. “ Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Effects.” Review of Economic Studies 76(3): 1071–102. Lightening Global Program. 2019. “Pay-as-You-Go Market Attractiveness Index 2019.” International Finance Cor- poration. ———. 2020. “The 2020 Global Off-Grid Solar Market Trends Report.” International Finance Corporation. Mazar, N., D. Mochon, and D. Ariely. 2018. “ If You Are Going to Pay within the Next 24 Hours, Press 1: Automatic Planning Prompt Reduces Credit Card Delinquency.” Journal of Consumer Psychology 28(3): 466–76. McIntosh, C. 2008. “ Estimating Treatment Effects from Spatial Policy Experiments: An Application to Ugandan Microfinance.” Review of Economics and Statistics 90(1): 15–28. McKenzie, D., and S. Puerto. 2021. “ Growing Markets through Business Training for Female Entrepreneurs: A Market-Level Randomized Experiment in Kenya.” American Economic Journal: Applied Economics 13(2): 297– 332. Milkman, K. L., J. Beshears, J. J. Choi, D. Laibson, and B. C. Madrian. 2011. “ Using Implementation Intentions Prompts to Enhance Influenza Vaccination Rates.” Proceedings of the National Academy of Sciences 108(26): 10415–20. ———. 2013. “Planning Prompts as a Means of Increasing Preventive Screening Rates.” Preventive Medicine 56(1): 92–3 Muralidharan, K., M. Romero, and K. Wüthrich. 2023. “Factorial Designs, Model Selection, and (Incorrect) Inference in Randomized Experiments.” The Review of Economics and Statistics https://doi.org/10.1162/rest_a_01317 Nickerson, and T. Rogers. 2010. “ Do You Have a Voting Plan?: Implementation Intentions, Voter Turnout, and Organic Plan Making.” Psychological Science 21(2): 194–99. Prestwich, A., R. Lawton, and M. Conner. 2003. “ The Use of Implementation Intentions and the Decision Balance Sheet in Promoting Exercise Behaviour.” Psychology & Health 18(6): 707–21. Rogers, T., K. L. Milkman, L. K. John, and M. I. Norton. 2015. “ Beyond Good Intentions: Prompting People to Make Plans Improves Follow-Through on Important Tasks.” Behavioral Science & Policy 1(2): 33–41. Schaner, S. 2018. “ The Persistent Power of Behavioral Change: Long-Run Impacts of Temporary Savings Subsidies for the Poor.” American Economic Journal: Applied Economics 10(3): 67–100. Stango, V., and J. Zinman. 2014. “ Limited and Varying Consumer Attention: Evidence from Shocks to the Salience of Bank Overdraft Fees.” Review of Financial Studies 27(4): 990–1030. Stojanovski, O., M. C. Thurber, F. A. Wolak, G. Muwowo, and K. Harrison. 2021. “ Assessing Opportunities for Solar Lanterns to Improve Educational Outcomes in Off-Grid Rural Areas: Results from a Randomized Controlled Trial.” World Bank Economic Review 35(4): 999–1018. Suri, T. 2017. “ Mobile Money.” Annual Review of Economics 9(1): 497–520. Wagner, N., M. Rieger, A. S. Bedi, J. Vermeulen, and B. A. Demena. 2021. “The Impact of Off-Grid Solar Home Systems in Kenya on Energy Consumption and Expenditures.” Energy Economics 99: 105314. Wichman, C. J. 2017. “Information Provision and Consumer Behavior: A Natural Experiment in Billing Frequency.” Journal of Public Economics 152: 13–33.