Growth Research Podcast

In Uganda, every sale between two VAT-registered firms reaches the tax authority twice. The seller reports it, and so does the buyer. But the two numbers rarely match. In Turkiye, a supplier that installs its first industrial robot doesn’t make an announcement, but the administrative records can track when the firms automate, and what happens next.Justine Knebelmann and Nuriye Melisa Bilgin tell Tim Phillips how they used those records to find out more about the choices firms make, and the effects on the firms around them. Justine and her co-authors worked with the Uganda Revenue Authority to send letters to firms whose reports did not match their trading partner’s, to find out under what conditions the firms amended their tax returns. Nuriye mapped robot adoption across Turkiye's firm-to-firm network, to find out the competitive effects. They also discuss how researchers can access administrative data, and how simply compiling it can change policy.The research behind this episode:Almunia, Miguel, David J. Henning, Justine Knebelmann, Dorothy Nakyambadde, and Lin Tian. 2023. "Firm Networks and Tax Compliance: Experimental Evidence from Uganda." CEPR Discussion Paper 18151, revised October 2025. Bilgin, Nuriye Melisa, Ester Faia, and Gianmarco Ottaviano. 2024. "Technology Spillovers, Diffusion and Rivalry in Firm Networks." CEPR Discussion Paper 19804. To cite this episode:Phillips, Tim, Justine Knebelmann, and Nuriye Melisa Bilgin. 2026. "Follow the paper trail." PEDL (podcast). Centre for Economic Policy Research. About the guestsJustine Knebelmann is an Assistant Professor in the Department of Economics at Sciences Po, Paris, and an Associate at the Institute for Fiscal Studies. Her research spans state capacity in developing countries, tax administration, digitalisation and land. She has worked with tax administrations in Sub-Saharan Africa since 2015.Nuriye Melisa Bilgin is a Lecturer in Economics at Koç University and a Senior Fellow at the Microsoft AI Economy Institute. Her research spans technology adoption, from industrial robots to generative AI, and how new technologies and shocks travel through production networks and global value chains, using large-scale firm-to-firm transaction data from Turkiye. She held postdoctoral positions at Bocconi University and the University of Turin.About PEDLPrivate Enterprise Development in Low Income Countries (PEDL) is a joint research initiative of the Centre for Economic Policy Research (CEPR) and the Foreign, Commonwealth & Development Office (FCDO), and part of CEPR's Growth Research Platform. Established in 2011, it funds research on private-sector development in low-income countries. Find out more at grp.cepr.org/pedl.Research cited in this episodeValue-added tax (VAT). Firms charge VAT on their sales and reclaim the VAT they paid on inputs bought from other registered firms. Because every business-to-business sale appears on two returns, one from each side, the tax authority can in principle check one report against the other. Economists call this the VAT's self-enforcing property. In Uganda the VAT raises about 30% of total tax revenue, according to IMF figures cited by Almunia and co-authors; the IMF puts Uganda's VAT compliance gap at around 60% of potential VAT revenue.Seller shortfall. The Uganda team's term for a transaction in which the seller reports a smaller amount than the buyer. It lowers the tax bill and often signals evasion. In the 10 months before the experiment, from March to December 2017, seller shortfall appeared in 41.4% of monthly seller-buyer observations, and in more than 92% of those cases the seller had not reported the transaction at all.Almunia, Hjort, Knebelmann and Tian (2024). "Strategic or Confused Firms? Evidence from 'Missing' Transactions in Uganda," Review of Economics and Statistics 106 (1): 256-265. This earlier study measured the scale of the problem. Sellers and buyers reported different amounts for the same transactions 79% of the time, and the authors estimate that unilateral misreporting cost Uganda about USD 383 million in VAT revenue between 2013 and 2016.Randomising pairs, not firms. In a dense trading network, a letter sent to one firm can reach its partners by several routes and blur the results. The Uganda team randomised at the level of the seller-buyer pair instead, and selected 1,235 pairs so that no two shared a firm. Of these, 741 received letters, sent to the seller only, the buyer only, or both; the remaining 494 formed the control group. Because treated pairs were kept apart in the network, a correction by a firm that did not receive a letter points to communication between the two partners.The final sales loophole. Sales to consumers and to unregistered firms appear on a Ugandan VAT return as a single total, with no second report to check against. Sellers used this margin. When treated sellers amended their returns to add missing business-to-business sales, they cut their reported final sales by about 60% of that amount. The net gain in VAT was modest, but the letters still raised more than six times what they cost to send.Electronic invoicing in Uganda. After the 2018 experiment, the Uganda Revenue Authority expanded its data cross-checks. In 2021 it launched EFRIS (electronic fiscal receipting and invoicing), which requires input claims to be matched to VAT invoices, and in 2022 it began rolling out electronic billing machines that record transactions at the point of sale.The reflection problem. Charles Manski named it in "Identification of Endogenous Social Effects: The Reflection Problem" (Review of Economic Studies, 1993). When connected firms perform alike, it is hard to tell whether one influenced the other or whether similar firms simply chose to trade together. Nuriye's point about productive firms sorting into relationships with other productive firms is this problem in a supply chain. Her team addresses it by using the precise timing of each first-time robot purchase and checking for trends that predate it.Knowledge spillovers and product market rivalry. A firm can gain when a partner adopts a new technology and lose when a competitor does. Nicholas Bloom, Mark Schankerman and John Van Reenen set out the difficulty of separating the two effects in "Identifying Technology Spillovers and Product Market Rivalry" (Econometrica, 2013). Bilgin, Faia and Ottaviano separate them by the direction of the link. Suppliers pass gains down to their customers; rivals that sell to the same customers compete them away.Turkiye's matched firm data. The robot study links VAT returns covering firm-to-firm transactions, collected by the Ministry of Finance, with employer-employee records from the Social Security Institution, customs records from the Ministry of Trade, and registry and balance sheet data from the Ministry of Industry and Technology. The data run from 2007 to 2019. Robot purchases are identified by matching firms' fiscal codes with members of ENOSAD, the Turkish industrial automation manufacturers' association, and imported robots by their customs code, which also records the country of origin and whether an intermediary was involved.Robot quality and technical support. Robots from Switzerland and Germany, which sell at higher prices than those from China, generate larger spillovers to the adopter's customers. Robots bought through intermediaries that also provide technical support produce the largest effects. Four years after a supplier's first purchase, its customers' productivity is 14% higher when the robot came through an intermediary, against 6.3% when it came from a domestic producer.Linking surveys to administrative records. Justine mentions separate work by her co-author David Henning, who surveyed Ugandan firms and linked their answers to their tax returns. The combination reveals informal activity that administrative data alone cannot see, including among firms that had left the formal tax system.Listen nextNo taxation without administration: What makes tax authorities work, a VoxDev Talks episode in which Anders Jensen and Jonathan Weigel review the evidence on how tax authorities are organised, staffed and run, and why third-party data has limits where informality is high.Related readingTechnology spillovers, diffusion, and rivalry in firm networks, a VoxEU column in which Bilgin, Faia and Ottaviano summarise the robot study.Strategic or confused? Firm behaviour and missing millions in Uganda's VAT, a VoxDev article on the earlier study that showed how often Ugandan sellers and buyers disagree about the same transaction.VAT in developing countries: flawed, but irreplaceable, a VoxDev article by Anne Brockmeyer, Giulia Mascagni, Mazhar Waseem, Miguel Almunia and Vedanth Nair, which draws on VAT administrative data from 11 countries to show where the tax falls short of its textbook design.

Name a rich country that did not move its farm workers into factories, because Michael Peters says it would be hard to find one. Many years ago, Europe and the US took that road, then Japan and Korea, then China and Vietnam. But when workers in India, Kenya and Ethiopia leave the land, far fewer of them reach a factory. They are more likely to work in shops and restaurants, or sell from a street stall.Many of these countries, Peters argues, never industrialised in the first place, so there are fewer factories to go to. If services themselves become more productive, perhaps a country can grow without them. But how do you measure the productivity of services?Peters presented the research at STEG's sixth Annual Conference in Nairobi in January 2026. He tells Tim Phillips why a better supermarket in Delhi does little for a rural family, whether a street vendor has a good job or a disguised form of unemployment, and what AI means for countries that hope to export services.The research behind this episodePeters, Michael, Youdan Zhang, and Fabrizio Zilibotti. 2026. "Skipping the Factory: Service-Led Growth and Structural Transformation in the Developing World." NBER Working Paper 34692, prepared for the Handbook of Development Economics. The authors summarise the research in a VoxEU column, "Skipping the factory: Service-led growth in sub-Saharan Africa."To cite this episodePhillips, Tim, and Michael Peters. 2026. "Skipping the factory." Conversations on Transformation (podcast).About the guestMichael Peters is Associate Professor of Economics and Associate Chair of the Department of Economics at Yale University. He is a Research Associate of the National Bureau of Economic Research and a Research Fellow of the Centre for Economic Policy Research. He holds a PhD from MIT, and taught at the London School of Economics before he joined Yale. His research spans economic growth, firm dynamics, the long-run effects of migration, and structural change, with a focus on inequality between people and places.Research cited in this episodeThe canonical model of structural change describes development in stages. Agriculture sheds workers, manufacturing grows, and services then overtake manufacturing, so that manufacturing's share of jobs rises and falls in a hump. Peters says this fits the US and Europe in the 19th and early 20th centuries, Japan and Korea after the Second World War, and more recently China and Vietnam. The countries in his paper look different. In 16 fast-growing East and West African economies that do not depend on oil, agriculture's share of employment fell by an average of 18 percentage points between 2000 and 2019. Services took 15 of those points and industry took 3.The agricultural productivity gap is one reason why moving workers off the land can raise output. W. Arthur Lewis described an economy split between a subsistence sector and a modern sector in "Economic Development with Unlimited Supplies of Labour" (The Manchester School, 1954), and shared the 1979 Nobel Prize in economics. Douglas Gollin, David Lagakos and Michael Waugh measured the gap in "The Agricultural Productivity Gap" (Quarterly Journal of Economics, 2014). In national accounts, value added per worker outside agriculture is about three times the level inside it. A large gap remains after they adjust for hours worked and human capital.Innovation in manufacturing is the second link that Peters draws between factories and growth. The 2025 Nobel Prize in economics went to Joel Mokyr, Philippe Aghion and Peter Howitt for explaining innovation-driven growth. If innovation matters most in manufacturing, Peters argues, a large manufacturing sector gives firms the right incentives to invent and adopt new technology.Premature deindustrialisation describes countries that lose manufacturing jobs and output at much lower incomes than the early industrialisers did. Dani Rodrik documented the trend in "Premature Deindustrialization" (Journal of Economic Growth, 2016). Rodrik notes that the term seems to have first appeared in a 2006 paper by Sukti Dasgupta and Ajit Singh.Non-tradable consumer services must be delivered where the customer is. Unlike software or a call centre, a shop in Nairobi cannot sell to a village far away. In their estimates, the authors count retail, leisure and health. In the six African countries with two comparable surveys, the share of private employment in these services rose from 4% to 13% in rural areas between about 2000 and 2010, and from 38% to 43% in cities.Growing like India is the earlier study that Peters wrote with Tianyu Fan and Zilibotti, "Growing Like India: The Unequal Effects of Service-Led Growth" (Econometrica, 2023). It covers 360 Indian districts between 1987 and 2011. By 2011, services that mainly sell to consumers employed almost 55% of India's service workers, while ICT accounted for less than 1% of all employment. Rising productivity in consumer services explains almost one-third of the rise in welfare. The average resident of the most urbanised fifth of districts would have given up 37% of their 2011 income rather than go back to the 1987 productivity level in these services. In the three least urbanised fifths, the figure is 13%.Service-biased and service-led growth are the two explanations that the authors try to separate. Growth is service-biased when productivity rises elsewhere, people get richer, and they spend more on restaurant meals, entertainment and travel. Services expand, but only as a result of growth. Growth is service-led when services themselves become more productive. Peters's example is refrigeration, which transformed retail and raised what each worker in it could produce.Baumol's cost disease is the pessimistic case for services. It comes from William Baumol's "Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis" (American Economic Review, 1967). If resources shift towards industries where productivity grows slowly, productivity growth in the economy as a whole falls.Measuring productivity from willingness to pay avoids price data. Prices of local services differ from place to place, and they mix efficiency with quality. The authors instead use a model of demand, the local income distribution, and the number of people who work in services. If many people work in services given local incomes, consumers must value what they buy. Peters is clear about the limits, because the method rests on a structural model and on assumptions about how labour markets work. The African exercise pools rural and urban data from Benin, Ghana, Malawi, Mozambique, Rwanda and Tanzania, and borrows key parameters from the Indian study. The authors present it as an illustration of the method rather than a full estimate. It puts productivity growth in consumer services at 3.3% to 3.5% a year, against 1.7% to 2.7% in tradable sectors. The same method finds little productivity growth in Vietnam's consumer services.Equivalent variation measures what a change is worth to someone, as a share of their income. The authors compare people at the 5th and 95th percentiles of local income. Poor rural households buy almost no market services, so productivity growth in those services does not reach them. Rich urban households gain more from it than from productivity growth in tradable sectors. Poorer households gain mainly from productivity growth in agriculture and manufacturing.Durable goods as a proxy for income fill a gap in the data, because the census samples do not record household income. Across 13 African countries, households headed by consumer service workers are 4.1 percentage points more likely than manufacturing households to have electricity, 3.3 points more likely to have piped water, and 2.7 points more likely to have brick or metal walls. These gaps shrink to 2.3, 1.4 and 1.2 points after the authors control for urban location and education. Flush toilets and good roofs show no significant difference, and service households are less likely to own a radio. Manufacturing in this comparison includes construction, utilities and mining.Disguised unemployment is the case against this optimism. Suppose that frictions in labour or financial markets ration good jobs. The people who miss out become self-employed street vendors, not because demand is strong but because they have no better option. A large service sector would then be a symptom of what stands in the way of manufacturing-led growth. Peters says that data such as Brazil's matched employer-employee records could show how much of the rise in service jobs comes from demand and how much from frictions.Formalisation moves workers from self-employment into organised firms. Peters suggests it as one way to make service-led growth reach poorer people, because the typical consumer service worker has skills much like those of the typical manufacturing worker. A street vendor could become a shop employee, and a self-employed rickshaw driver could join a more organised transport firm. The evidence so far comes mainly from Brazil, in work by Matthew Schwartzman. In many African economies, the authors note, informal production still dominates and broad-based formalisation has yet to occur.South Africa's unemployment rate is Peters's example of the size of the task for services. The official rate reached 33.6% in the second quarter of 2026, up from 32.7% in the first quarter, according to Statistics South Africa.More Conversations on Transformation episodesStructural change in sub-Saharan Africa. Gaaitzen de Vries and Kei-Mu Yi use macro and trade data to ask whether the region follows the familiar pattern of transformation, the same puzzle from a different direction.Moving to profitability. Carolyn Pelnik asks why Kampala's street vendors and motorcycle taxi drivers stay in less profitable parts of the city, a close look at the kind of informal service work that Peters studies at a national scale.Further readingFrom Street Markets to Shopping Malls: T

Some studies find that AI destroys jobs, some find no effect, and some find that it creates them. Gustavo de Souza of the Chicago Fed argues that AI has different labour market impacts in different occupations. To test this, he uses records of Brazilian firms, who can register their software with the government for 50 years of copyright protection. The register holds a near-complete record of the AI programs written in the country. Almost half of them are for office work, but 44% are for production, on farms, in factories and on building sites. He tells Tim Phillips about the scale and direction of those different effects of AI on employment.De Souza presented the research at STEG's sixth Annual Conference in Nairobi in January 2026. The research behind this episodede Souza, Gustavo. 2025. "AI in the Office and the Factory: Evidence from Administrative Software Registry Data." Working Paper 2025-11, Federal Reserve Bank of Chicago. Revised 2026.To cite this episodePhillips, Tim, and Gustavo de Souza. 2026. "AI in the office and the factory." Conversations on Transformation (podcast).About the guestGustavo de Souza is an Economist in the economic research department of the Federal Reserve Bank of Chicago. He holds a PhD in Economics from the University of Chicago. His research spans macroeconomics, labour economics, development and public finance. The views in his paper and in this episode are his own, and do not necessarily reflect those of the Federal Reserve Bank of Chicago or the Federal Reserve System.Research cited in this episodeAI as a prediction technology is the idea at the centre of de Souza's argument. Ajay Agrawal, Joshua Gans and Avi Goldfarb set it out in "Artificial Intelligence: The Ambiguous Labor Market Impact of Automating Prediction" (Journal of Economic Perspectives, 2019). AI uses large amounts of data to predict something: which post a social media user will click on, whether a bank transfer is fraud, or which word comes next in a sentence. If a worker's main task is to make predictions, AI can replace them. If a worker makes decisions with the help of predictions, AI can make them more productive.Studies that find AI reduces jobs include Michael Webb's "The Impact of Artificial Intelligence on the Labor Market" (2020), one of the first papers to measure which occupations are exposed to AI, and Xiang Hui, Oren Reshef and Luofeng Zhou's "The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market" (Organization Science, 2024), which found that freelancers lost work after generative AI tools became available.Studies that find little or no effect include Anders Humlum and Emilie Vestergaard's "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI" (NBER Working Paper 33777, 2025), first circulated as "Large Language Models, Small Labor Market Effects". It uses Danish data on workers who use AI chatbots, and finds no meaningful effect on earnings or hours. Daron Acemoglu, David Autor, Jonathon Hazell and Pascual Restrepo, in "Artificial Intelligence and Jobs: Evidence from Online Vacancies" (Journal of Labor Economics, 2022), found that firms that adopt AI hire fewer people for other roles, but found no effect on employment in occupations or industries as a whole.Studies that find AI increases jobs look at firms that adopt it. Menaka Hampole, Dimitris Papanikolaou, Lawrence Schmidt and Bryan Seegmiller, in "Artificial Intelligence and the Labor Market" (NBER Working Paper 33509, 2025), use resume data and find that at firms that adopt AI, productivity gains outweigh the jobs it displaces. Philippe Aghion and co-authors, in "How Different Uses of AI Shape Labor Demand: Evidence from France" (AEA Papers and Proceedings, 2025), find that the effect depends on how firms use AI. Tania Babina, Anastassia Fedyk, Alex He and James Hodson, in "Artificial Intelligence, Firm Growth, and Product Innovation" (Journal of Financial Economics, 2024), find that firms that invest in AI grow faster.Experiments with AI assistants often find that less experienced and less skilled workers gain the most. In "Generative AI at Work" (Quarterly Journal of Economics, 2025), Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied customer support agents who were given an AI assistant. Productivity rose most for new and low-skilled agents. De Souza finds the same pattern in Brazilian factories.Brazil's software register is kept by the National Institute of Industrial Property (Instituto Nacional da Propriedade Industrial, INPI). A law of 1987 gave 50 years of copyright protection to software registered there. To register, a firm gives the name of the program, the names of its programmers, where it will be used, the methods and programming languages it uses, and a short description. The source code stays confidential. Registration is cheap and gives legal proof of ownership, so it has become standard practice. De Souza collected every registered program that uses AI and linked it to RAIS, Brazil's administrative records of all formal jobs, from 2003 to 2022.Brazil's informatics policy explains why the register exists. In 1984, the military government passed the National Informatics Policy to build a domestic computer industry, with restrictions on imports and support for local firms. Civilian rule returned in 1985, and the software law followed two years later. Today little software in Brazil is imported, so the register captures most of the software that Brazilian firms use.Manufacturing execution systems are the software that controls and tracks a production line: stock, output forecasts, workstations and quality. They account for nearly 40% of the AI programs for production in the register. One firm in the data, Techplus, has registered an industrial copilot: a chatbot with access to the manuals of a factory's machines, connected to live data from the production line, so that workers can ask it how to operate equipment and diagnose faults. The second most common use is quality control, in which computer vision checks products for defects as they leave the line.Measuring exposure to AI follows a method that Webb developed. De Souza compares the text of each occupation's task descriptions with the descriptions of AI programs in the register. The more similar the two, the more exposed the occupation is to AI.The instrument deals with the risk that software firms aim AI at occupations that are already growing or shrinking for other reasons. When a programming language becomes more popular in the United States, its community grows, along with its libraries and online answers, and software in that language becomes cheaper to write in Brazil too. De Souza measures popularity with the Tiobe index, and links languages to types of software with posts on Stack Overflow, the question and answer site for programmers. COBOL, created in the 1950s, still runs much of the world's banking; a Reuters estimate cited in the paper puts its share of in-person transactions at more than 80%. But it is no longer taught and its community is small, so AI for banking is expensive to build. R, created in the 1990s for statistics, has a large and growing community, so AI for data analysis is cheap. A bank teller is therefore less exposed to AI than an archivist, for reasons that have nothing to do with the Brazilian labour market.The missing intercept problem is why de Souza is cautious about the effect on the whole economy. His method compares occupations that are more exposed to AI with occupations that are less exposed. It can show that one group gains relative to the other, but not whether employment rose for everyone, fell for everyone, or rose for some and fell for others. His comparisons between regions, which are not in the paper, suggest that AI increases total employment and moves workers between occupations.More Conversations on Transformation episodesBroadband on the factory floor. Gaurav Chiplunkar finds that when Colombian manufacturing plants got broadband, they hired more production workers, not managers or technical staff.Digital technologies in developing countries. Robert Townsend asks whether digital tools can do the work of the social structures that support trade and credit in poor economies.Further readingFrom Skills to Occupations: Comparative Advantage and Cross-Country Income Differences, a STEG working paper by Charles Gottlieb, Jan Grobovsek and Alexander Monge-Naranjo, uses data from 50 countries to show how workers sort into occupations, and how technology that favours skilled workers widens the income gap between rich and poor countries.Mobile Internet Adoption and Inclusive Structural Change: Evidence from Nigerian Non-Farming Enterprises, a STEG working paper by Bernardo Caldarola, finds that mobile internet made small Nigerian firms more productive, mostly in services, but did not create new firms.

A tractor can plough an acre of Kenyan maize land in about an hour and a half. By hand, it takes a farmer nearly a week. Few smallholders can afford to buy a tractor, so they rent one. The tractor’s owner often lives in town, and employs an operator to drive the tractor to the fields and do business. To minimise risk, the owner keeps their tractor close to home, even when the best work is far away.Sophie Nottmeyer studied what changed when owners could observe exactly where the tractor was. She used GPS records from about 1,200 tractors fitted with a tracking app from the start-up Hello Tractor, covering close to a million fields, and satellite images of the land they worked. As owners learned to use the app, their tractors travelled further, finding better business opportunities. A year after adoption, they worked about 55 km further from home than in their first month.Nottmeyer presented the research at STEG's sixth Annual Conference in Nairobi in January 2026. She tells Tim Phillips what satellites reveal about the fields the tractors reached, and why a fuel subsidy that did the same job would cost hundreds of times more.The research behind this episodeNottmeyer, Sophie. 2025. "Trac(k)tors of Change: Monitoring, Tractor Mobility and Agricultural Mechanization in Kenya." Working paper, November 2025. A summary is available as a STEG policy brief: "Following the tracks: how digital monitoring expands access to agricultural mechanisation across space" (2026).To cite this episodePhillips, Tim, and Sophie Nottmeyer. 2026. "Tracking Kenya's tractors." Conversations on Transformation (podcast).About the guestSophie Nottmeyer is a Research Economist in the Directorate General of Economic Research at Banco de Mexico. She was previously a Postdoctoral Researcher at Universitat Pompeu Fabra in Barcelona, and holds a PhD in Economics from CEMFI in Madrid. Her research spans market access, organisational frictions and informal institutions, and how they interact with geography and climate. She has also worked with remote-sensing scientists to estimate maize yields in Kenya from satellite data.Research cited in this episodeHello Tractor is an agricultural technology company based in Nairobi. In 2018 it introduced a GPS tracking device for tractors, fitted by a mechanic, with a mobile and web app that shows owners where their tractor is in real time. The app reports how many jobs the tractor did, how many acres it worked, how long it ran and how far it drove, and it can replay the tractor's exact path across a field. In 2023 owners paid US$125 to install the device and US$60 a year to use the app. Nottmeyer's data cover every job done by tractors that adopted it between 2018 and 2025, about 917,000 in total.Moral hazard arises when one party cannot observe what another does on its behalf. Kenyan tractor owners hire operators, who are often paid from the revenue they collect. Because the owner cannot see how many acres were ploughed, an operator can under-report jobs, take unauthorised work or use fuel for other purposes. Owners respond with surprise visits, fuel checks and rules that the tractor must come home every night. All of these work only if the tractor stays nearby. The problem is sharper because many owners are not farmers at all; according to Kenya's 2019 census, 33% of tractor owners live in cities.Misallocation describes an economy in which capital or labour does not go where it would produce the most. Chang-Tai Hsieh and Peter Klenow showed in "Misallocation and Manufacturing TFP in China and India" (Quarterly Journal of Economics, 2009) how large the cost of such gaps can be. Nottmeyer measures misallocation as the spread in the returns that tractors earn in different places. If a tractor would earn more in a distant district than near home, but never goes there, the economy loses output.Counting Kenya's tractors is harder than it sounds. No official source records tractor ownership or use, Kenya's last nationwide agricultural census was in 1979, and the Food and Agriculture Organization stopped publishing its global series on agricultural machinery in 2009. The roughly 1,225 tractors in Nottmeyer's data represent between 1.73% of the tractors counted by the 2019 census and 8.75% of the last FAO figure, from 2008.Comparing each tractor with itself is how Nottmeyer deals with the lack of a control group. The data begin only when a tractor adopts the app, so she uses each tractor's first month as its baseline, on the assumption that owners take time to trust the information and change how they manage operators. The average distance from home was 70 km in that first month. It rose by about 30 km after six months, 55 km after one year and 112 km after two years. The pattern is similar for early and late adopters, for older tractors whose owners had run them before fitting the device, and for tractors that return home every night. Nottmeyer does not claim these estimates are strictly causal, but the alternative explanations she tests do not account for the pattern.Satellite measures of crop growth stand in for yields, which nobody records at field level in Kenya. Nottmeyer uses Sentinel-2 images from the European Union's Copernicus programme to calculate the Normalized Difference Vegetation Index (NDVI), which measures how green and dense vegetation is. Ploughing leaves a field bare at the start of the season, so the gap between the lowest and highest NDVI in a year captures both land preparation and crop growth. She matches 33,256 fields worked by monitored tractors with similar fields nearby. In the year of the visit, the gap rises by about 0.4 standard deviations more in the visited fields, and part of the gain persists in later years. Because some of the comparison fields may also have been ploughed by other tractors, this is a lower bound on what mechanisation does.A spatial model of tractor location turns these results into national figures. Owners choose where to send a tractor by weighing the returns in each place against the cost of distance, and monitoring lowers that cost. When all Hello Tractor tractors behave as if they had five years of experience with the app, the spread of returns across places falls by 15.9% and total output rises by 2.18%, compared with tractors that have just adopted it. If every tractor in Kenya had the technology, output would rise by 1.3%.A fuel subsidy is the obvious alternative policy, because it makes long trips cheaper. Nottmeyer calculates that to achieve the same gains, a subsidy would have to cover almost all the fuel a tractor uses per kilometre, at a total cost of about US$9.33 billion. Fitting all 70,930 tractors in the 2019 census with the device and a five-year subscription would cost about US$30.1 million.Monitoring in other industries follows the same logic wherever an owner hands a vehicle to someone else to operate. Thomas Hubbard's "The Demand for Monitoring Technologies: The Case of Trucking" (Quarterly Journal of Economics, 2000) studied the onboard computers that US trucking firms installed to track drivers. Erin Kelley, Gregory Lane and David Schonholzer ran an experiment with Nairobi's minibus owners, reported in "Monitoring in Small Firms: Experimental Evidence from Kenyan Public Transit" (American Economic Review, 2024), and found that tracking devices helped owners manage drivers and raised profits.More Conversations on Transformation episodesMoving to profitability. Carolyn Pelnik asks why street vendors and motorcycle taxi drivers in Kampala stay in less profitable parts of the city, another case in which mobile workers and assets do not go where they would earn most.Slack and economic development. Felix Samy Soliman explains why firms in Kenya leave labour and capital idle, and how much growth that unused capacity could deliver.Further readingMisallocation and Product Choice, a STEG working paper by Stepan Gordeev and Sudhir Singh, uses farm and crop data from India to estimate how much output is lost when inputs are spread unevenly between farms, and shows that the answer depends on which crops farmers can choose to grow.Tillers of Prosperity: Land Ownership, Reallocation, and Structural Transformation, a STEG working paper by Shuhei Kitamura, finds that when the land reform enforced by the Allies after the Second World War gave tenants ownership of their farms, they adopted low-cost farm technology, relied less on family labour, and more young people left for the cities.

In many cities in low-income countries, the grid fails often enough that businesses keep a diesel generator out the back. So an unreliable electricity grid looks like it is holding poor economies back. In many countries, lenders are putting money behind that idea.But Martin Shu and his co-authors, Jonathan Colmer and David Lagakos, set out to test it. They assembled data on power plants, fuel, workers and imported generators in 89 countries, and measured how productive each electricity sector was. In the poorest countries, electricity was about three-quarters as productive as in the US — in other words, much closer to the frontier than their economies. And in the economic model they built using their data, bringing every country up to the world frontier adds only about 1% to long-run GDP.Shu presented the research at STEG's sixth Annual Conference in Nairobi in January 2026. He tells Tim Phillips where the productivity gap in electricity lies, why a shoe factory that is better at left shoes than right shoes explains the small gains, and which part of the economy is the real bottleneck.The research behind this episodeColmer, Jonathan, David Lagakos, and Martin Shu. 2026. "Is the Electricity Sector a Weak Link in Development?" Working paper, January 2026 version presented at the STEG Annual Conference, Nairobi. An earlier version appeared as NBER Working Paper 32041 (2024).To cite this episodePhillips, Tim, and Martin Shu. 2026. "Is electricity a weak link in development?" Conversations on Transformation (podcast).About the guestMartin Shu is Assistant Professor of Economics at the School of Management and Economics and the Shenzhen Finance Institute of the Chinese University of Hong Kong, Shenzhen. His research spans macroeconomics, development and labour economics, including what household and firm data reveal about structural transformation, and why children in poor countries are more likely to follow their parents into the same occupation. He holds a PhD in Economics from Boston University and bachelor's and master's degrees from the Hong Kong University of Science and Technology.Research cited in this episodeWeak links are sectors whose problems hold back the whole economy. When inputs are strong complements, output depends heavily on the least productive link in the chain, so a small sector can matter far more than its size suggests. The idea draws on Albert Hirschman's The Strategy of Economic Development (1958), and Charles Jones formalised it in "Intermediate Goods and Weak Links in the Theory of Economic Development" (American Economic Journal: Macroeconomics, 2011), which uses electricity as a motivating example. Shu and his co-authors ask whether electricity fits that description in poor countries today.Total factor productivity (TFP) measures how much output an economy or a sector gets from its inputs. If two countries have the same capital and the same number of workers, but one produces more, it has higher TFP. Shu measures each country against the United States. The figures for whole economies come from the Penn World Table, hosted by the Groningen Growth and Development Centre and described by Robert Feenstra, Robert Inklaar and Marcel Timmer in "The Next Generation of the Penn World Table" (American Economic Review, 2015). The paper uses version 9.1; the current release, PWT 11.0, is supported by STEG.Agriculture's productivity gap is the comparison Shu uses to show why measuring TFP is not enough. Farming in poor countries is far less productive than the rest of the economy, as Francesco Caselli ("Accounting for Cross-Country Income Differences," Handbook of Economic Growth, 2005) and Diego Restuccia, Dennis Tao Yang and Xiaodong Zhu ("Agriculture and Aggregate Productivity: A Quantitative Cross-Country Analysis," Journal of Monetary Economics, 2008) document. Farmers are also the majority of workers in many developing countries. Yet few economists call agriculture a weak link. Whether a sector is a bottleneck depends on how it connects to the rest of the economy, and answering that needs a model as well as data.The electricity data combine several sources. Generating capacity comes from S&P Global Platts' World Electric Power Plants database, which lists more than 200,000 plants; the authors use generating capacity alone because it tracks transmission and distribution capacity closely. Power output and fuel use come from the International Energy Agency, and employment from national statistics offices. To count power from private generators, the authors use imports of generators recorded in the UN Comtrade database, an approach inspired by Francesco Caselli and Daniel Wilson ("Importing Technology," Journal of Monetary Economics, 2004). It works because few poor countries make their own generators. On the authors' conservative measure, generators supply about 7.6% of capacity on average.Labour's share of revenue in electricity is about 7%, in rich and poor countries alike. The authors collected annual reports, financial statements and SEC filings for 69 privately owned electricity companies in 31 countries, choosing private firms because their decisions follow profit rather than government mandate. Labour's share shows no relationship with GDP per capita. A share that does not change with income points to a Cobb-Douglas production function, in which each input's share of costs stays constant, and that simplifies the TFP calculation.Transmission and distribution losses are power that leaves a plant but never reaches a customer, lost as heat in wires and transformers, or stolen. According to IEA data, they average 14% of output across the sample, 22% in the poorest quarter of countries and 7% in the richest. They are the strongest predictor of low electricity TFP; once they are taken into account, the link between electricity TFP and national income largely disappears. Shu points to upgraded transformers close to end users, and ultra-high-voltage transmission lines, as the investments that would close the gap.Complementarity describes inputs that work together and cannot easily replace each other, like left and right shoes. The model sets the elasticity of substitution between electricity and everything else at 0.2, following Ernst Berndt and David Wood ("Technology, Prices, and the Derived Demand for Energy," Review of Economics and Statistics, 1975). Strong complementarity sounds as if it should make electricity more important. Shu's shoe factory shows why it does not: when a sector is already efficient, making it more efficient frees few resources, and output still waits on the less productive partner.Hulten's theorem, from Charles Hulten's "Growth Accounting with Intermediate Inputs" (Review of Economic Studies, 1978), predicts the economy-wide gain from a productivity improvement in one sector: multiply the improvement by the sector's Domar weight, its sales as a share of GDP. The authors treat any gain above that prediction as the sign of a weak link. Raising electricity TFP to the world frontier, defined as the 90th percentile, means an average increase of 43%, and by Hulten's rule it would add 1.7% to GDP. The model delivers about 1%, and only six countries beat the benchmark. The biggest exceptions are Ethiopia and Kenya, where frequent outages make power expensive or halt production. The authors also test their results against David Baqaee and Hannes Malmberg's "Long-Run Comparative Statics" (NBER Working Paper 33504, 2025), which extends Hulten's approach to allow for capital accumulation.Raising productivity outside electricity is the counterpart experiment. When the authors raise TFP in the rest of the economy by the same proportion instead, every country gains more than Hulten's theorem predicts, and the capital stock grows by 75% on average. Shu's conclusion is about balance. With complementary sectors, pushing one to an extreme achieves little, and today it is the non-electricity sector that lags.More Conversations on Transformation episodesA Microeconomic Perspective on Misallocation. Eric Verhoogen explains how economists diagnose the wedges that keep resources from their best use, the same kind of distortion Shu's model uses to capture outages and costly self-generated power.Structural change in sub-Saharan Africa. Gaaitzen de Vries and Kei-Mu Yi ask whether African economies follow the familiar path of structural transformation, a wider view of the sectors that Shu argues are holding growth back.Further readingCan Mass Electrification (Still) Cause Structural Change?, a STEG synthesis paper by Eric Hsu and Mitchell VanVuren, reviews the mixed evidence on whether connecting people to power still raises productivity and incomes in low- and middle-income countries, as it appears to have done in the past.Under the Grid: Price and Informational Barriers to Grid Densification in Benin, a STEG working paper by Arnaud Dakpogan, Megan Lang and Alpha Ly, finds that a connection subsidy did not raise formal connection rates on average, but information on costs and procedures nearly doubled them among households with no connection at all.