Expanding Access to Electricity in Kenya

Abstract

This case study tells the story behind a research project on the economics of rural electrification in Western Kenya. The chapter covers (1) aspects of the policy and technology environment that initially guided the course of the work; (2) how the project pivoted away from solar microgrids and focused instead on the expansion of the national electricity grid; (3) unexpected challenges encountered while implementing a randomized evaluation of electricity infrastructure; (4) how we interpreted the study findings in light of consequential, concurrent changes to Kenya’s electrification policies; and (5) possible directions for further research, motivated by our project experience.

For the full results of this study, see Lee et al. (2016a), and Lee, Miguel, and Wolfram (2016b, 2020a, b). I am grateful to Shipra Karan for excellent research assistance in preparing this chapter. All errors are my own.

1 The Development Challenge

In 2012, we began a study on solar microgrids in rural Kenya. Over time, it evolved into an experiment that randomized the expansion of the national electricity grid instead. In this chapter, I tell the story behind this project, focusing on the pivots and iterations that shaped the path of our research on the economics of electrification over nearly a decade.

When we started our project, access to electricity was widely seen as a major driver of economic development, just as it remains today. Then-United Nations Secretary General Ban Ki Moon famously referred to it as the “golden thread” connecting economic growth, social equity, and an environment where people could thrive. Supporting this outlook was the well-known, near-perfect correlation between electricity consumption and GDP per capita, which is shown in Fig. 5.1.

Fig. 5.1

The positive correlation between electricity consumption and GDP per capita

Notes: Both variables are presented on a logarithmic scale. 2014 data obtained from the World Bank DataBank Reprinted from Lee et al. (2020b)

The academic literature proposed many plausible channels through which electricity access could improve lives. Electric lighting, for instance, could extend the workday, increasing labor supply and income. Since lighting would also give children more time to do their homework, educational attainment might even improve. Some pointed out the potential issues with rural electrification programs as well. In a review of the history of electrification in sub-Saharan Africa, Bernard (2012) emphasized the limited productive uses of electricity that had often been observed across rural Africa. Although people seemed to use electricity for lighting and communications, they were less likely to use it for agriculture, handicrafts, and other activities that could be profitable. Khandker et al. (2014) proposed an additional issue that the gains from rural electrification could be much greater for wealthier households, which could exacerbate economic inequalities.

At the time, over a billion people still lacked access to electricity. The question of how governments could best expand access to power remained front and center. Moreover, developing countries were expected to drive a considerable amount of growth in global energy consumption (Wolfram et al., 2012). As a result, expanding access in these countries using conventional fossil fuel technologies would certainly accelerate global warming. The development challenge was clear: In countries with high rates of energy poverty, how could electricity access be expanded while mitigating the consequences on the global environment?

In the spring of 2012, UC Berkeley’s Development Impact Lab brought together a team of economists and engineers to work on this problem.¹ The basic goal of the collaboration was to improve the design of the solar microgrid technologies that were being developed for poor countries. By bringing together engineers and economists, the iterative process of engineering design could be merged with microeconomic survey data and evidence obtained using the randomized control trial (RCT) approach that had become widespread in development economics. We believed the results of such a collaboration could inform the design of technologies and public policies in unique ways.

A couple aspects of the partnership generated a great deal of excitement. First, the engineers had been working with start-up companies to design solar-powered microgrids that featured novel, prepaid, smart metering technologies. These devices offered a trove of high-frequency, electricity usage data. We thought about using these data to predict the kinds of appliances people were using in their homes. Combined with data collected through household surveys, we could perhaps unlock the precise mechanisms through which electrification improved well-being. And in places where electricity theft was an issue, we thought about measuring the impacts of the smart metering technologies on monitoring and enforcement. There were many possibilities for data science.

One of the start-up companies working on this technology was based nearby, in Oakland. They had just built a pilot microgrid in Kenya and were looking to scale up. After a series of meetings, we agreed to explore measuring the impacts of their solar microgrids in an RCT. Kenya was a good fit, since it was a place where the economists had extensive prior field experience. From the company’s perspective, an independent team documenting the beneficial impacts of their product could be useful in their public communications and marketing efforts, as well as in their search for venture capital funds to propel their rapid expansion across Africa. Of course, if they partnered with us, our research grant would pay for (or at least subsidize) the cost of some of their microgrid installations.

Second, from a research standpoint, we were excited about measuring the causal impacts of electrification in an experimental setting. Until then, the applied microeconomics literature covering this topic was limited. Nearly all of the existing work was nonexperimental and tended to rely on administrative or observational data, which made it a challenge to distinguish causal effects from correlations. In any situation, one could imagine a host of unobserved factors that could be correlated with someone’s access to electricity (the cause) and the changes they might experience over time (the possible effects). An exception at the time was Dinkelman (2011), which showed that rural electrification substantially improved female labor supply in South Africa. The study had creatively, and quite convincingly, applied an econometric technique to isolate a causal effect of electrification, using only administrative data. Thus, in the early days of our study, there was growing scholarly interest in building upon the evidence base on the impacts of electrification using ever more rigorous approaches to estimation.

More broadly, it seemed just a matter of time before billions of dollars would be directed towards electrification programs across the world. In the face of climate change, solar microgrids had great potential. We thought about how an infrastructure experiment could yield new benchmarks on the causal effects of electrification. Perhaps these could serve as useful inputs in the large-scale, infrastructure investment decisions that would surely be made in the future.

This case study chronicles our research timeline and how our views on the development challenge evolved along the way. Initially, we set out to build solar microgrids in off-grid villages in Kenya. But instead, we partnered with the Kenyan government and connected hundreds of randomly selected rural households to the national grid for the first time. We encountered a number of implementation challenges, many of which played a role in shaping our eventual conclusions. As it turned out, it was relatively expensive to build electricity network lines to rural homes. And around 3 years later, we had found no evidence indicating that household access to the grid had meaningfully changed a set of pre-defined economic and noneconomic outcomes. The project spurred a number of follow-up projects, which are still ongoing.

The bulk of our project was carried out between 2012 and 2017. Over this period, major policies would be introduced to accelerate the rural electrification of Kenya. In certain parts below, I bring up interactions we had with donors and policymakers at various points in time. I share these stories for a few reasons. First, I think these meetings imparted on us a number of timely perspectives that helped guide and refine our research focus. Second, I think our experience in Kenya offers an example of how the slow and deliberate process of academic research can sometimes be outpaced by rapidly shifting policy priorities in developing countries. Finally, and on a more personal note, I think the linkages between our research project and the wider policy environment in Kenya encapsulate what makes this line of work so exciting, and I hope some of this comes across in this chapter.

The remainder of this case study is organized as follows. The next section discusses the technology choices available to policymakers at the start of our project. Then, in the following sections, I describe the important decisions we needed to make to set up an experiment; the things we learned that influenced our research questions and intervention design; and how we made sense of our findings given the evolving policy context. The final section offers a view on some of the important research questions for the future.

2 Innovations in the Technology Landscape

There are several ways to address the development challenge of expanding access to electricity. Traditionally, governments have addressed this challenge by investing in expansions of their national grids. All developed countries have reached universal rural electrification in this way. The issue moving forward, of course, is the extent to which the grid can supply electricity from nonfossil fuel sources of energy.

The 2000s introduced various improvements to an array of decentralized, renewable energy alternatives, including solar lanterns, solar home systems, and renewable energy microgrids. There was hope that these novel technologies could allow people living in the Global South to gain access to electricity, while minimizing the negative consequences on the environment. Across sub-Saharan Africa, the rapid adoption of mobile phones had made landline telecommunications infrastructure obsolete. By 2012, many entrepreneurs, donors, and observers were talking about how this improved set of decentralized, renewable energy solutions would allow off-grid households to similarly leapfrog the grid.

A couple trends seemed to be driving this growing level of enthusiasm. First, it was becoming much cheaper to manufacture these products. With each passing year, there were increases in appliance efficiencies, reductions in the cost of photovoltaics, and improvements in battery capabilities. By around 2012, off-grid solar began to be seen as a potential alternative to the grid. Second, in countries like Kenya, rapid growth in mobile phone usage had been accompanied with widespread adoption of mobile money platforms like M-PESA. Around 2010, new start-up companies like M-KOPA began integrating pay-as-you-go technologies directly into their solar lanterns and solar home systems. Poor, rural consumers could now buy these products on credit, unlocking each day of usage with a small payment sent over their mobile phones. This was seen as a gamechanger in rural settings, where credit constraints had often limited the take-up of promising, new technologies.²

2.1 Prepaid, Smart Metered Solar Microgrids

Microgrids, which connect small networks of users to a centralized and stand-alone source of power generation and storage, were also generating substantial interest. Microgrids could provide longer hours of service and higher capacities than solar lanterns and solar home systems, making it feasible to use power more productively. Furthermore, they could be powered with clean energy sources, like solar, wind, and hydro. Despite their potential, microgrids had not yet been deployed at scale in developing countries. In fact, a number of early microgrid pilot deployments had completely failed.

For example, in the early 2000s, dozens of microgrids had been set up in rural villages across Rajasthan, India. The microgrids were built to connect rural hamlets to a 10-kilowatt capacity solar panel system. After several years, many of the microgrids had fallen into disrepair. There were a number of reasons why this happened. For instance, the Rajasthani microgrids, which offered users just enough electricity to power lighting and small appliances, did not meter households individually. Instead, each grid would be switched on and off at certain hours of the day. In return, each user needed to pay a monthly fixed fee. But when certain users refused to pay, there was no way to terminate their service, leading to a downward spiral in revenue collection. Making matters worse, the battery banks powering the microgrids could not withstand the hot and humid environmental conditions of Rajasthan. When the batteries failed, the lack of payment enforcement meant there was little to no cash reserve available to cover the cost of a replacement. The primary culprit for the failure of the Rajasthani microgrids was not some electrical engineering issue, but rather a misalignment of economic incentives. Simply put, the early microgrids were in need of a better business model.³

In Kenya, our microgrid partner had taken advantage of the technological trends to develop a next-generation, village-scale solar microgrid that allowed consumers to pay-as-they-go using their mobile phones. They marketed their technology as one that could empower consumers to make real-time decisions about their energy consumption, while alleviating credit constraints. Importantly, each user had their own smart meter that would send information about power consumption and credit balances over text messages. Depending on how each system was sized, they promised power that would be more reliable than the national grid.

In 2012, official estimates of the national household electrification rate in Kenya ranged from 18 to 26%.⁴ We were intrigued by the potential market for this microgrid. And as our discussions with our partner progressed, it became easy for us to imagine the thousands of off-grid villages across Kenya where this technology would thrive. We took it for granted that the people living in these off-grid villages would be receptive to this novel technology. Soon, we would discover that we were wrong.

3 Iterative Learning: A Major Pivot

In the summer of 2012, we traveled to Western Kenya to scope out a potential research project. The experiment we envisioned was straightforward. First, we would identify a hundred or so off-grid villages, randomly assigning half of them into a treatment group. In these villages, our microgrid partner, with contributions from the engineers on our team, would set up their prepaid, smart metered microgrids. As an additional experimental feature, we thought about varying the price of each connection in order to estimate a demand curve for electricity access.⁵ Later on, a team of enumerators would administer detailed, household-level microeconomic surveys. In theory, comparing survey data between households in the treatment communities and their counterparts in the control communities would yield unbiased, causal estimates of the impacts of electrification.

Our microgrid partner suggested that we find communities with a couple important features. One, we needed villages with a high density of potential users. The microgrids would be sized to supply power to roughly 50 customers. If customers were clustered close together, the line losses on the microgrid’s low-voltage network would be minimized. Two, we needed villages with many unelectrified businesses since these were likely to use more electricity, thus increasing revenue to our partner. From our standpoint, we also wanted villages that were far away from existing national grid infrastructure. The last thing we wanted was to invest our resources and time in villages that would soon receive grid electricity from Kenya Power, the national electricity distribution utility.

This was no easy task. After visiting a local Kenya Power office in the Western county of Busia, we learned that Kenya Power had yet to geotag the locations of its infrastructure, meaning there was limited administrative data that could help us locate a sample of off-grid villages. In lieu of actual data, we were given permission to photograph the aging infrastructure maps that were displayed on the walls. In addition, we were provided with an assortment of tips on where we could find the distant yet densely populated communities that would meet our criteria.

3.1 On-grid, Off-Grid, or Under-Grid?

As we drove across Western Kenya searching for rural, off-grid villages, we noticed something peculiar. Although the vast majority of rural homesteads lacked access to electricity, nearly every off-grid village we visited seemed to have a power line running nearby. Rather than being “off-grid,” much of what we observed appeared to be underneath the grid.

Why were so many rural households left unconnected to these electricity lines? We learned that a major barrier was the high cost of connection. In fact, during the decade leading up to the start of our study, any household in Kenya within 600 m of a low-voltage distribution transformer could apply for an electricity connection at a fixed price of 35,000 Kenya shillings (KES), which was worth roughly $398 USD at the time. This seemed far too expensive in Kenya, where annual per capita income was below $1,000 for most rural households. At the same time, the cost to the utility of supplying a single connection in an area with grid coverage was estimated to be several multiples higher.

Our trip to Kenya that summer was not much of a success. Instead of finding a hundred villages, we found just a handful. However, the experience triggered a shift in the way we viewed the development challenge. Until then, we had been thinking about electricity access as a binary variable. Households were either on-grid or off-grid. What naturally followed was an assumption that off-grid households were too far away to connect to a national electricity network and therefore required novel solutions. We wondered whether this assumption had played a role in the growing enthusiasm among engineers, entrepreneurs, and donors for the new generation of off-grid, distributed energy solutions, most of which were being designed and manufactured outside of the Global South. The solar microgrids that were being developed in Oakland, for instance, had essentially been designed with remote users and communities in Africa in mind.

Suddenly, it seemed plausible that a substantial share of the 600 million people lacking access to electricity were not off-grid but were instead “under-grid,” which we defined as being close enough to connect to a low-voltage line at a relatively low cost. This distinction seemed important because the policy implications for off-grid and under-grid communities were quite different. In under-grid communities, it might be preferable to design policies that could leverage existing infrastructure, as opposed to promoting an independent solution like a microgrid.

The argument against grid power seemed to hinge on the extent to which the grid delivered dirty, fossil fuel power. But across sub-Saharan Africa, installed generating capacities were still relatively low, and substantial capacity additions were slated for the future. Importantly, a large share of these additions was expected to feature nonfossil fuel technologies. In Kenya, where fossil fuels represented about a third of installed capacity at the time, several major geothermal and wind projects were already under development. Given the trends, why not focus on expanding electricity access through a grid that might soon be channeling a higher share of clean energy?⁶

3.2 Private Versus Public Infrastructure

Upon our return to Berkeley, we ran into problems agreeing on an acceptable research design with our microgrid partner. From a research standpoint, we needed to randomly select the villages where our partner would build their microgrids. In addition, we required access to all of the data generated by their smart meters; we needed to publish our findings, regardless of how favorable the conclusion; and of course, we planned to make all of the data and analysis involved in our work freely available to public.

All of this is highly undesirable for a start-up company. Our microgrid partner needed to prove that it had a viable and scalable business model. There was upside to having the benefits of their technology rigorously and independently measured and published. But there were obvious downsides to giving up control over consequential business decisions, like where they could build their microgrids.

We had other differences in opinion as well. For example, some of the communities that met our microgrid partner’s various criteria (e.g., density of structures, number of small businesses, etc.) looked to be—at least from our perspective—under grid. And while there may have been good business reasons for building a private microgrid directly underneath the public grid, this did not make sense to us. Moreover, to achieve a certain degree of statistical power in our experiment, we needed our microgrid partner to rapidly scale up its operations. New systems needed to be built in scores of villages, as soon as possible. But the difficulties we had faced in locating suitable sites seemed to portend inevitable delays and slow progress.⁷

We began to consider whether it made more sense to study the economics of expanding national grid access. After our summer travels, we had no doubt that Kenya’s future would revolve around its grid. Although the cost of a Kenya Power connection was exorbitantly high, it was roughly in line with the per household cost of our microgrid partner’s technology. If we shifted our focus to the grid, we could design a research study with a similar sample size, without requiring an increase to our budget.

In early 2013, our collaboration with our microgrid partner began to taper off. Fortunately, we had begun a promising round of discussions with Kenya’s Rural Electrification Authority (REA). Created in 2007, REA was a government agency that had been established to accelerate the pace of rural electrification. Using funds from the central government, international development agencies, and a small tax on every Kenya Power electricity bill, REA had been responsible for rapidly electrifying the majority of the country’s rural secondary schools, markets, and health clinics. It was the single reason why so many of the rural households we had observed that previous summer looked to be under-grid.

We proposed an experiment to REA to randomly connect households to the grid and then experimentally measure the impacts of electricity access on a variety of social and economic outcomes. They were interested in the idea. As a public infrastructure agency, REA could adopt a longer-term investment horizon, meaning they did not face the same pressures as a start-up company. Moreover, our project fit with their basic mandate to achieve universal rural electrification. In other words, there was a clear path towards justifying the incremental costs that would be incurred in order to participate in our research project. By the summer of 2013, the CEO of REA had agreed in principle to explore the viability of a randomized experiment.

4 Randomizing Access to Grid Electricity

Infrastructure investments tend to involve high fixed costs, relatively low marginal costs, and long investment horizons. As a result, they tend to be owned and regulated by the government. As mentioned earlier, it is difficult to measure the causal effects of electrification. There are a number of factors that tend to be correlated with the presence of electricity infrastructure (like the placement of roads), and many of these factors could also contribute to changes in key economic outcomes, like employment. In addition, it is only natural that governments would target expensive infrastructure investments towards regions predicted to enjoy the greatest rates of economic growth, or areas that were preferable to the ruling government party and were thus in line to reap a number of other rewards. In these situations, there is a problem of omitted variable bias (sometimes called confounding or selection bias). Whatever we may believe to be the effects of electrification may actually be caused by unobserved variables. If these are not addressed econometrically, the effects of electrification can be overemphasized.

At the outset of our project, rigorous microeconomic evidence in this area was limited. Two papers that caught our attention included the Dinkelman (2011) paper mentioned above, as well as Lipscomb et al. (2013), which estimates the long-term impacts of electrification in Brazil. Both studies relied on administrative data and addressed the issue of omitted variable bias using a similar econometric approach.⁸ From a research standpoint, we thought about how a field experiment could remove concerns about omitted variable bias entirely, yielding a new set of causal estimates that developing country policymakers could use as they weighed the relative costs and benefits of investments in health, education, energy, and other areas.

How could we randomly assign access to electricity, without forcing some people to connect to the grid and others to remain in the dark? In a methodological note that was greatly influential for our project, Bernard and Torero (2011) offered a solution: By providing treatment households with a subsidized electricity connection, we could randomly encourage them to connect to the grid. If many people responded to these offers, there would be enough variation in electricity access to measure impacts. On top of that, by offering different subsidy amounts to different households, we could randomly assign the effective price of a connection, allowing us to trace out a demand curve.⁹

4.1 Conducting a Census Across “Transformer Communities”

Now that we had an outline of our experimental intervention in place, several additional decisions needed to be made. The first set of choices involved defining a sample and a unit of randomization. We needed to consider the potential economic spillovers of our treatment. For example, a family living in an electrified household could easily share the benefits of their connection with some of their neighbors. In this case, a direct comparison of living standards between the treated households and neighboring control households would underestimate the size of impacts. To address this issue, we decided to assign our treatment at the community level. This choice would have the additional benefit of avoiding fairness concerns among neighbors, given the high monetary value of the subsidized connection offers.

We also needed to avoid situations in which households applied for power but were rejected for some technical reason, like being too far away from a transformer. So instead of defining communities along the lines of village boundaries—which is what most field experiments randomizing at a community level tended to do—we defined them as “transformer communities,” encompassing the universe of structures that were within 600 meters of a central, low-voltage distribution transformer. This, of course, had the added bonus of being consistent with Kenya Power’s existing fixed price connection policy.

Working with our partners at REA, we randomly drew a sample of 150 transformers, located at markets, secondary schools, and health clinics, spread across in the counties of Busia and Siaya in Western Kenya. Some of the transformers had been installed 4 to 5 years earlier, so they were not necessarily new.

The next step was to establish a sampling frame, or census, of all of the unconnected and "under-grid" households that could potentially be enrolled into our study. So, over the fall of 2013, a team of enumerators scouted the territory surrounding each transformer on foot, documenting the GPS location, some basic observable features (like the quality of roofs and walls), and the electrification status of every structure they were able to find. A number of iterations were required to establish a survey protocol that could produce maximum coverage of the area inside each transformer community. Examples of various transformer communities are provided in Fig. 5.2.

Fig. 5.2

Examples of transformer communities in Western Kenya

Notes: The white circle labeled “T” in the center of each transformer community identifies the location of the REA transformer. The larger white outline demarcates the 600-m radius boundary. Green circles represent unconnected households; purple squares represent unconnected businesses; and blue triangles represent unconnected public facilities. Yellow circles, squares, and triangles indicate households, businesses, and public facilities with visible electricity connections, respectively. Household markers are scaled by household size, with the largest indicating households with more than 10 members and the smallest indicating single-member households. In each community, roughly 15 households were randomly sampled and enrolled into the study. The average density of a transformer community is 84.7 households per community, and the average minimum distance between buildings (i.e., households, businesses, or public facilities) is 52.8 m. Reprinted from Lee et al. (2016a, b)

By December 2013, our field staff had geotagged over 20,000 structures, including households, enterprises, public facilities, transportation hubs, and other types of buildings. The data showed that electrification rates were extremely low, averaging just 5% for households and 22% for businesses. In addition, half of the unconnected households we observed were estimated to be within 200 m of a low-voltage power line.¹⁰

At that point, REA estimated it had connected 90% of the country’s public facilities to the grid. In the highly populated regions of Central and Western Kenya, it was believed that the vast majority of households were within walking distance of multiple public facilities. If true, then the under-grid pattern we documented in Western Kenya could very well extend across vast swathes of the country. For policymakers, connecting people to a network that was already in place seemed like low-hanging fruit on the path towards universal electrification.

We reached out to a number of organizations that were involved in rural electrification efforts across Africa. It was an exciting moment in time. Earlier that year, President Obama had launched Power Africa, a bold new initiative that targeted, as one of its initial goals, 20 million new electricity connections in six African countries including Kenya. We presented our findings to the coordinator of Power Africa, who was based in Nairobi at the time. Perhaps our under-grid narrative could be useful in shaping their overall strategy. They did not appear that interested in our work. Later, we learned that Power Africa’s strategy to meeting its short-term connection targets would be to support the distribution of decentralized solar technologies (mainly solar lanterns and solar home systems) through its Beyond the Grid initiative, as shown in Fig. 5.3.

Fig. 5.3

How USAID Power Africa will meet its target of 60 million new connections

Notes: Based on USAID Power Africa Annual Reports

We also presented our findings to the energy leads at the World Bank. From these interactions, we learned that the World Bank was in the process of planning a major commitment to modernize Kenya’s electricity system. Over the following months, we continued meeting with this team to share updates. As we later learned, this commitment from the World Bank would enable the Kenyan government to launch its transformative Last Mile Connectivity Project (LMCP), just a couple years later. The initiative, which is briefly mentioned later in the chapter, has already gone on to connect millions of people to the grid.

4.2 Binding Our Hands with a Pre-analysis Plan

In all of our discussions with potential donors, policymakers, and partners, there was always a great deal of interest in discussing the impacts of rural electrification. Everyone seemed to hold their own unique perspectives and predictions on what would happen when rural households accessed electricity for the first time. Existing academic studies tended to focus on the same major outcomes, like employment and educational attainment. This was in part due to their reliance on administrative records, as well as general household surveys, neither of which placed a large emphasis on energy outcomes, like the types of electrical appliances that people owned or wished to own. In our experiment, we would be free to draft our own survey instruments, meaning we could measure anything we wanted. In sum, this was a situation in which the conclusions of our experiment could seem arbitrary since they depended on the outcomes we chose to emphasize. With so many possibilities, how could we best assess the overall impacts of our electrification program?

There was also another issue. With REA, we were collaborating with individuals who had dedicated their entire working lives to expanding rural access to power. Like our former microgrid partner, there was upside to having the benefits of their work rigorously and independently evaluated. Yet in our interactions with REA officials, we sensed a high degree of confidence that our study would point to massive, positive effects. Furthermore, as a government agency, we could imagine how there may have been underlying political pressures or incentives to report a particular result. It was, after all, hard for any of us to picture rural life remaining the same after the introduction of grid electricity. But what if we were wrong?

Around that time, the use of pre-analysis plans in development economics was becoming increasingly common. The basic steps we followed were straightforward. Prior to accessing data, we wrote down our hypotheses; the econometric regression equations we planned to estimate; and the set of key outcomes we would consider. In our initial plan, which was filed before we analyzed the first round of follow-up survey data, we identified 77 outcomes of interest overall, across 10 broad families including energy consumption, productivity, education, and others. We then narrowed this list down to the 10 primary outcomes, shown in Table 5.1, that would guide our conclusions. We discussed these outcomes with our government partners, which allowed us to manage their expectations. Finally, the plan was registered online in order to be made accessible to future readers.¹¹ The pre-analysis plan disciplined our interpretations of impacts, limiting the scope for data mining or a biased presentation of results. As we would later find, our pre-analysis plan increased the space we had to draw our conclusions and made it easier to defend our findings, given the lack of impacts we eventually found.

Table 5.1 Defining ten primary outcomes of interest in a pre-analysis plan

5 Iterative Learning: Unexpected Field Challenges

In December 2013, we signed a Memorandum of Understanding (MOU) with REA’s Chief Executive Officer. REA agreed to honor the subsidized connection prices in our experiment, giving us the green light to proceed with the main component of our project. The MOU, which took 6 months to sign, was the result of countless in-person visits to the REA headquarters in Nairobi. At each visit, we would deliver a presentation to the CEO and other key managers at REA, updating them on what we were learning from our activities in Western Kenya and slowly building up our relationships with key individuals. At times, we would need to re-pitch the basic objectives of our experiment. At others, REA would provide helpful feedback on the information that would be most useful to them in our research.

In many of these meetings, a sticking point was the cost of the connection. We agreed to use a substantial amount of our research funds to cover the cost of subsidies, which guaranteed that REA would receive no less than 35,000 KES per connection. In turn, REA would need to cover the difference between the cost of construction and the 35,000 KES raised from each accepted offer. The uncertainty came from the wide variance in the cost of supplying each connection, since this depended on factors like the location, density, and terrain conditions of each community. Although there were economies of scale in connecting multiple households at the same time, it was difficult to predict how many households in each community would accept the offers. Given the many fixed commitments in REA’s organizational budget, there were understandable concerns about the total cost of what many at REA viewed as just an academic exercise (which was not inaccurate).

Over time, we established a stronger relationship with our government partners. Crucially, agreed on a compromise. Instead of offering a subsidized price to any unconnected household within the boundaries of a transformer community, we agreed to limit our sample to the households that were no more than 400 m away from a low-voltage line, which worked out to 84.9% of all of the households recorded in our census. According to REA’s estimates, this would substantially reduce the construction costs, improving the financial viability of the project.

Our experiment would require contributions from a number of individuals working out of REA’s offices in Nairobi and Kisumu at various levels of the organizational hierarchy, each facing a different set of incentives. Given the innate challenges of working in Western Kenya (e.g., travel distances, road conditions, etc.), there were many potential bottlenecks that could delay our progress. Although our MOU was not legally binding, it proved to be surprisingly effective as a document that we could point to in the face of an unanticipated challenge. To our luck, many of our interactions with REA would be facilitated by an ambitious, data-driven, and public-minded bureaucrat with whom we would work closely. This person not only took a keen interest in seeing the academic results of our study through but also proved instrumental in clearing several administrative roadblocks we faced. His involvement ensured that much of our work would be executed smoothly. Still, we encountered a number of unexpected problems, some of which required us to revise our intervention protocols and others that simply stalled our progress. Throughout the process, we carefully documented each of these issues, which helped us make sense of our eventual results.

5.1 Connecting Households in Areas with No Electricians

In the spring of 2014, our team of enumerators made rapid progress enrolling households into our study and collecting baseline social and economic data. Meanwhile, we worked out the details of our randomized pricing intervention, which was scheduled to commence that summer. The process was as follows. After completing the baseline survey, treatment households would receive a letter from REA describing a time-limited opportunity to connect to the grid at a subsidized price, which was randomly assigned at the transformer community level. At the end of the offer period, our staff would verify payments and provide REA with the list of households that needed to be connected. REA would then dispatch designers to each community to sketch out the low-voltage network that needed to be built. Based on these drawings, such as the one shown in Fig. 5.4, REA would earmark the appropriate amount of materials, including wooden poles, low-voltage feeder lines, and service drop lines. Next, a construction team would arrive on site to connect households. Each household would then register an account with Kenya Power in order to receive a prepaid meter. Once complete, households could use power.

Fig. 5.4

Example of a REA design drawing and the electrification of a treatment household

Notes: After receiving payment, REA designers visited each treatment community to design the local low-voltage network. The designs were then used to estimate the required materials and determine a budgeted estimates of the total construction cost. Materials (e.g., poles, electricity line, service cables) represented 65.9% of total installation costs. Reprinted from Lee et al. (2020a)

There was one unanticipated problem. Although REA could attach a service line to a building, households would still need outlets for plugging in their appliances, sockets for light bulbs, fuses, and other internal wiring. But in rural areas with low electrification rates, there were few electricians.

To solve this problem, we located a manufacturer in Nairobi that could produce a customized “ready-board,” an all-in-one household wiring solution. Each ready-board, which is shown in Fig. 5.5, featured a single light bulb socket, two power outlets, and two miniature circuit breakers. The ready-board was designed to be installed on the indoor side of an exterior wall, so that an outdoor service line could pass through a hole in the wall and connect directly into the back of the ready-board. The ready-boards were designed to be modular as well, thus providing households with the option of installing additional boards as their consumption needs grew. As it turned out, the wiring issue was a relatively easy challenge to address. However, it illustrated how the success of a particular technology in a developing country could be hindered by the lack of a key supporting service.

Fig. 5.5

Umeme Rahisi ready-board solution designed by Power Technics

Notes: Treatment households received an opportunity to install a certified household wiring solution in their homes at no additional cost. Each ready-board, valued at roughly $34 per unit, featured a single light bulb socket, two power outlets, and two miniature circuit breakers. The unit is first mounted onto a wall and the electricity service line is directly connected to the back. The hardware was designed and produced by Power Technics, an electronic supplies manufacturer in Nairobi. Reprinted from Lee et al. (2020a)

5.2 Major Supply-Side Issues: Blackouts and Construction Delays

Blackouts were another issue. Most of the time, the blackouts in Western Kenya would last just minutes or a few hours. But sometimes, a blackout would cut power from an entire community for months. In 2014, as REA carried out its connection work, we began noting the frequency, duration, and primary reason for all of the long-term blackouts experienced in our sample. In total, 19% of the transformers in our sample experienced at least one long-term blackout, which lasted 4 months, on average. The transformers seemed prone to burnouts, and other technical failures caused by severe weather conditions. In some cases, Kenya Power would temporarily relocate a transformer to another community, unannounced. There were also reports of vandalism, as well as theft due to the perceived value of the copper and oil components inside the transformers.

Overall, it took an extraordinarily long time to connect households in our experiment. The first household was metered in September 2014, just a couple months after it had completed payment. The last household was metered over a year later in October 2015. The average connection time was seven months.

Major delays arose at each stage of the process, as shown in Fig. 5.6. The longest average delays occurred during the design phase (57 days) and the metering phase (68 days). The design delays were caused in part by a sudden government announcement in 2015 to provide free laptops to all Grade 1 students, nationwide. A presidential election was on the horizon, and the incumbent would once again be running for office. Since only half of Kenya’s primary schools were electrified at the time, REA suddenly found itself under pressure to connect more primary schools. As a result, there were less REA designers available to work on less-prioritized projects, like ours.

Fig. 5.6

Timeline of the rural electrification process

Notes: Panel A summarizes the rural electrification process from the standpoint of the household, divided into three key phases. Panel B summarizes the process from the standpoint of the supplier, divided into four key phases. The numbers to the right of each bar report the average number of days required to complete each phase (standard deviations in parantheses). Households were first given 56 days (8 weeks) to complete their payments. Afterwards, it took on average 212 days (7 months) for households to be metered and electricity to flow to the household. Reprinted from Lee et al. (2020a)

The metering delays occurred because of unexpected issues at Kenya Power. There were lost meter applications, shortages in prepaid meters, competing priorities for Kenya Power staff, and in some cases, expectations that bribes would be paid. Other delays were caused by a general shortage in construction materials at REA storehouses. Heavy rains made roads impassable in some communities. Difficulties obtaining wayleaves, which permit electricity lines to pass through private property, required drawings to be reworked, additional trips to the storehouse, and further negotiations with contractors. In some cases, households that had initially declined the ready-board changed their minds; in an unfortunate case lightning struck, damaging a household’s electrical equipment; and so on.

Of course, these kinds of problems are not unusual in developing country settings. As researchers, there was little we could do to address what were primarily supply-side issues. So, we instructed our project staff to send weekly and persistent reminders to REA and Kenya Power, and we took notes. It’s possible that without these reminders, the delays could have been much worse.

5.3 Discovering Cost Data and Investigating Potential Leakage Issues

As we strove to better understand the root cause of each delay, we became deeply familiar with REA’s internal administrative procedures. For instance, we learned that once the design drawings were complete, the total cost of required materials was tallied up in an official budgeted cost slip. Similarly, following the construction work, a final invoiced cost slip was generated and stored in REA’s database. With REA’s permission, we were able to access and integrate these data into our analysis, which added an unexpected cost angle to our study.

Essentially, our intervention consisted of bundling electricity connection applications together. As a result, the pricing variation we had introduced at the community level—in order to estimate a demand curve—had the bonus effect of creating local construction projects at various scales. With REA’s cost data, we could now trace out the average and marginal costs at different community coverage levels, allowing us to study economies of scale. By combining the demand and cost data, we could assess the social surplus effects of electrification, using the textbook framework of electricity distribution as a natural monopoly.¹²

The cost data allowed us to think about another issue. Midway through our intervention, we began hearing stories that some of the poles had gone missing. There were rumors that a contractor had appropriated some of the assigned materials and sold them back to REA’s suppliers. In the cost data, however, the budgeted and invoiced cost slips reported nearly identical figures. So, to investigate, we asked our field staff to return to each community to count the number of electricity poles in the ground and compare these numbers with the budgeted and invoiced cost slips. As it turned out, more than 20% of the poles were missing in the field.

In economic theory, it can be debated whether this type of leakage is economically harmful. REA’s loss, after all, might be offset by the contractor’s private gain. However, we learned that using less poles could cause lines to sag, and this not only lowered service quality but also increased the risk that poles fell over. Thus, it seemed that leakage could be an important issue affecting the long-run reliability of the grid and the overall economic returns to grid infrastructure.

5.4 The Gap Between Demand and Costs

By the end of 2014, we had collected enough data to trace out the demand curve and plot a few points of the cost curve. In our pre-analysis plan, we had recorded our own prior estimates of demand at various prices. In addition, REA had shared an internal memo with us that included the government’s own predictions of demand. By comparing our experimental estimates with these two sets of priors, we could conclude that demand was much lower than expected. Our cost curve, which was incomplete at the time, appeared to show substantial economies of scale from connecting numerous households at the same time. The average cost of a connection in our experiment was around $1,200. However, we did not yet have the data to know that the economies of scale would quickly level off, and that at full community coverage, the average connection cost would still be two to three times higher than the status quo connection price of 35,000 KES. The final demand and costs curves are shown in Fig. 5.7.¹³

Fig. 5.7

Experimental evidence on the demand for and costs of rural electrification

Notes: The experimental demand curve is combined with the population-weighted average total cost per connection (ATC) curve corresponding to the predicted cost of connecting various population shares, based on the nonlinear estimation of ATC = b ₀/M + b ₁ + b ₂ M. Each point represents the community-level, budgeted estimate of ATC at a specific level of coverage. Reprinted from Lee et al. (2020a)

From our perspective, the gap between demand and cost, which we referred to as social surplus, was important. In economics, the area under our demand curve could be interpreted as the present value of all future benefits accruing from grid access over the population. If the demand curve was much lower than the supply curve, then this would suggest that the benefits of rural household grid access were not high enough to offset the immediate cost of supply.

We needed time to process and debate these early results. But in March 2015, we were encouraged to present our findings to a number of audiences in Nairobi, including the recently assembled National Electrification Strategy committee. The committee featured many familiar faces from REA and Kenya Power, as well as representatives from various government departments. The purpose of the committee was to determine the specific details of the government’s rural electrification plan, which included the work that would be supported by the World Bank. The response to our presentation was positive. Soon afterwards, a committee member shared with us a draft version of the national electrification strategy. The document had been written around that time, with key contributions from our collaborators at REA. It estimated the economies of scale from a mass connection program would yield cost savings of 30%, in line with the early results we had shared with our partners. It also recommended a reduction in the connection price from 35,000 to 15,000 KES, one of the experimental price points in our study. A couple months later, President Kenyatta announced the Last Mile Connectivity Project (LMCP), a $364 million program funded by the World Bank and the African Development Bank. The connection price would soon be reduced for everyone.

In July 2015, we were invited to present our findings at a workshop organized to launch the World Bank’s $458 million commitment to modernizing Kenya’s electricity system, which would provide support for the LMCP, among other investments. We pointed out the large and potentially problematic gap between demand and cost, as well as some of the field challenges described above. The response was again positive. Then, later in the year, additional details of the LMCP were released to the public. It was made clear that unlike in the past, clusters of potential customers would now be connected to the grid at the same time. In addition, all customers requiring internal wiring would be provided with a free ready-board.

In retrospect, 2015 introduced momentous changes to Kenya’s rural electrification outlook. And given the role of REA in drafting the national electrification strategy—as well as our frequent presentations to Kenya Power, the World Bank, and others—it is possible that our research influenced some of the assumptions and decisions that were being made at the time. We will never know for sure, but the timing of our discussions and the later policy decisions seems consistent with our research exerting at least some level of influence during the key moments.

That said, in 2015, we could not answer any questions about the economic impacts of our intervention. Due to the connection delays, which had been ongoing, follow-up survey data remained months away. Moreover, we had not yet agreed on a satisfying way to interpret the gap between demand and supply—an explanation that factored in the potential budget and credit constraints at the consumer level, as well as the organizational performance issues we were observing in the field. And with each passing month, it began to feel as though our research progress was falling more and more behind. We had been there in Kenya at the start of all of these consequential policy choices. But by the time we had our survey evidence on impacts a couple years later, the national electrification plan was set and the LMCP was well under way.

6 No Meaningful and Statistically Significant Impacts

By November 2016, we had completed our first round of follow-up household surveys, which took place 16 months after connection on average. With a pre-analysis plan in place, it was fairly easy to calculate the impacts of our intervention on our primary set of outcomes. As expected, energy consumption had increased for treated households, but only by a miniscule amount. People had not really acquired any new appliances either. In fact, there were no detectable effects on assets, consumption, health, student test scores, and a host of other outcomes. After nearly a year and a half, we had no evidence of any meaningful and statistically significant impacts of electricity access, a result that differed from much of the earlier literature on the topic. Although this was perhaps fascinating from a research standpoint, it was a result that also felt depressing and demoralizing on a human level.

In the spring of 2017, we met with the Principal Secretary of the Ministry of Energy and Petroleum in Nairobi, as well as the heads of Kenya Power, REA, and the other parastatals comprising Kenya’s electricity sector. We discussed tariff reform, and in particular, the need to eliminate a monthly fixed charge that had made the prepaid meters difficult for consumers to understand and use.

We wanted to know how the government planned to do to encourage productive use, as it approached its goal of universal access. We presented our preliminary impact results, which the secretary found convincing. He noted, quite memorably, that the question of universal electricity access was not an economic one, but a political one. It was an election year, and the LMCP was a daily topic in the news. Our conversation highlighted the political economy considerations that often determine which and when certain groups benefit from government programs.

By the end of the year, we completed a second round of household surveys, which took place roughly 32 months post connection. The data revealed a similar pattern of no meaningful and statistically significant impacts. The evidence was consistent with the large gap between demand and cost that we had estimated in our experiment, several years earlier.

We were challenged to consider what might happen to our demand and cost analysis if the surrounding institutional and economic context was more favorable. That is, how much higher would the demand curve be if we could eliminate the credit constraints, blackouts and delays that may have suppressed demand? And what would happen to our estimate of social surplus if we eliminated leakage and incorporated a moderate level of income growth?

To answer this question, we referred heavily to the notes we had kept during the construction process, which provided us with the average duration of connection delays, the average amount of time the transformers had blacked out, and other statistics. Using this information, we projected the incremental effect on social surplus of fixing each issue. For example, we estimated that reducing the waiting period from 188 to 0 days would increase the consumer surplus by about 30%. In our final calculation, we predicted that if a number of improvements were simultaneously introduced, the area under the demand curve would finally exceed total costs, reversing our conclusion. This was, of course, an ideal scenario. In the real world, the LMCP was being rolled out across rural Kenya at one of our subsidized price levels. Surely, the plethora of issues that had affected demand and costs in Western Kenya existed elsewhere across rural Kenya.

The overall interpretation of our experimental results was that providing poor households with electricity access alone was not enough to improve economic and noneconomic outcomes. This stood in stark contrast to the previous, non-experimental studies that had documented large and beneficial gains from electrification. Perhaps in those settings, there were other factors, either correlated with or visibly part of the electrification efforts, that had influenced the direction of the results.

7 Looking Ahead

Since 2000, there has been tremendous progress across the world in reducing the number of people living without access to electricity. But nearly all of the global gains have been achieved in India. By 2030, roughly 500 million people in sub-Saharan Africa will still be without power (IEA, 2019). It is likely that Kenya’s model of mass electrification will serve as a blueprint for a number of African countries in the years to come.

With our experiment largely complete, we began thinking about new areas that seemed ripe for further research and that would have relevance to other settings in sub-Saharan Africa. For example, there is the question of whether the impacts of electrification are concentrated in certain types of individuals, like people who have the means to make the most out of a new connection. There was some evidence of this in our data. The impacts appeared larger for households that had a high willingness to pay for a connection, which was positively correlated with income and education at baseline. But due to limitations in our sample size, our results were suggestive at best.

There are also questions about the interaction between a technology like grid electrification and surrounding contextual factors. Would the impacts of our experiment have been greater if electricity access were paired with some kind of complementary input? What if the beneficiaries of Kenya’s LMCP also received subsidies to purchase different kinds of electrical appliances? Would this encourage people to experiment with new activities, thus paving the way towards greater levels of consumption and impacts?

Moreover, there is the question of how to make it easier for new users to consume and pay for electricity consumption. This is perhaps the most pressing challenge facing Kenyan energy policymakers today. Consider for a moment that between 2012 and 2019, half of Kenya’s population gained access to grid electricity for the first time. (More than three quarters of Kenyans are now connected to power.) This is a historic achievement. Yet the meteoric rise in electricity connections (shown in Panel A, Fig. 5.8) has been matched by a plunge in average electricity consumption per connection (Panel B) as a greater share of poor, rural households are connected to the grid (as is perhaps predicted in our research).

The problem is that each new connection requires a small expansion in the national low-voltage network. But if the marginal customer is generating little to no revenue, there is less money, on average, to maintain high-quality connections across the expanded network. Panel C illustrates this point by plotting the recent, sharp decline in gross profit per customer and gross profit per kilometer of low-voltage network lines.¹⁴ Based on the reliability and leakage issues in our study, it is easy to imagine this problem spiraling into the future, until the grid is completely overrun with service quality issues. And as some have argued, poor-quality electricity service can lower incentives to pay for consumption, leading to a self-reinforcing, vicious circle of blackouts, and bankruptcies.¹⁵ How will Kenya encourage greater rates of paid consumption in order to minimize the consequences of electrifying too fast?

Fig. 5.8

Electricity access and consumption trends in Kenya

Notes: Based on Kenya Power annual reports and my own calculations. Panel A illustrates the recent expansion of the Kenyan electricity grid, particularly across residential consumers. Panel B shows how average consumption per customer is rapidly falling (since growth in the customer base is coming from poor, rural households). Panel C plots gross profit—defined as revenue minus power purchase costs—per customer, per kilometer of low-voltage network lines (data only available from 2016 and beyond), and per kilometer of 11 kVA network lines. Panel C suggests that the ability of Kenya Power to pay for line maintenance using operating revenue is declining

Finally, in the face of climate change, the development challenge looms ever as large. How can developing countries expand access to energy while minimizing the costs to the environment? As this case study describes, we initially set out to build solar microgrids but ended up focusing on the grid instead. Thinking back, I still believe this was the right move. But in a way, our conclusions on the economics of rural electrification are less than satisfying. Not only was the grid a costly intervention, but our experiment produced no evidence of meaningful and statistically significant impacts, at least in the medium run. Meanwhile, new innovations will continue pushing forward the technological frontier of decentralized, renewable energy alternatives. Did we turn away from solar microgrids too soon? Or is the future calling for a more coordinated deployment of private and public infrastructure? These are just a few of the questions that deserve further consideration.

8 Discussion Questions

• 1. The researchers find no evidence of meaningful and statistically significant impacts of household grid electrification in the short and medium run in rural Kenya. What do you think are the implications of this result on the broader United Nations Sustainable Development Goal (SDG7) of universal access to affordable, reliable, sustainable, and modern energy for all by 2030?

• 2. The author hypothesizes that the impacts of grid electrification may have been greater if electricity access had been paired with complementary inputs. What complementary inputs could have the highest potential for impact? How would you design a research study to measure the impacts of supplying these inputs? Assuming that the provision of these inputs is effective, how would you revise the design of a mass electrification policy, such as the LMCP, in Kenya?

• 3. The author suggests that making it easier, for new users to consume and pay for electricity consumption may be the most pressing challenge for rapidly electrifying countries such as Kenya. Which interventions, policies, and technologies do you think hold the greatest potential to address this challenge?

• 4. A recurring issue in this project was striking a balance between meeting the goals of the infrastructure provider, while meeting the standards of academic research in applied economics. What were the key decisions in designing and implementing this study? And what changes do you think could have improved the general study outcome, both in terms of the study results as well as the policy impact?

• 5. Did the researchers pivot away from solar microgrids too soon?