CAF - The Origins of Structural Transformation

C A F - W O R KI N G PA P E R # 2 0 2 5 / 0 3

F i r s t v e r s i o n : J u l y 3 1 , 2 0 2 5

The Origins of Structural Transformation Julieta Caunedo1 | Mayara Felix2 | Kristina Manysheva3 1Associate Professor, Cornell University and CEPR. julieta.caunedo@cornell.edu 2Assistant Professor, Yale University and NBER. mayara.felix@yale.edu 3Assistant Professor: Columbia University. km3924@columbia.edu We study how labor market shocks originating in non-agriculture affect the organization of agricultural production. Using rich data from Brazil between 1986 and 2017, we show that the entry of large non-agricultural firms leads to persistent increases in local wages, declines in agricultural employment, and a shift toward more capital-intensive farming. Farms consolidate, the number of small operations declines, and mechanization increases. To study the magnitude of different channels contributing to this reorganization of production, we develop a general equilibrium model that features heterogeneous firms and farms through decreasing returns, an endogenous mechanization margin in agriculture and income effects. Farms optimally substitute capital for labor in response to rising wages following labor demand increase. At the same time, household demand relatively less agricultural products as their income raises with higher productivity in non-agriculture. Calibrated to match the Brazilian economy, our model predicts that a reduction in entry costs in non-agriculture–sufficient to generate the wage increase observed in the data–leads to labor reallocation out of agriculture, farm exit, and capital deepening. When we hold mechanization fixed, these adjustments are substantially attenuated, highlighting the role of endogenous technology adoption as an important

amplification mechanism. Finally, we use the model to explore how heterogeneity in farm sizes and mechanization can rationalize disparate responses to similar non-agriculture labor demand shocks. K E Y W O R D S Structural Transformation, Agriculture, Firm Entry, Labor Markets Small sections of text that are less than two paragraphs may be quoted without explicit permission as long as this document is acknowledged. Findings, interpretations and conclusions expressed in this publication are the sole responsibility of its author(s) and cannot be, in any way, attributed to CAF, its Executive Directors or the countries they represent. CAF does not guarantee the accuracy of the data included in this publication and is not, in any way, responsible for any consequences resulting from its use. ©2025 Corporación Andina de FomentoC A F - D O C U M E NT O D E T R AB A J O # 2 0 2 5 / 0 3

P r i m e r a v e r s i ó n : 3 1 d e j u l i o d e 2 0 2 5

Los Orígenes del Cambio Estructural Julieta Caunedo1 | Mayara Felix2 | Kristina Manysheva3 1Associate Professor, Cornell University and CEPR. julieta.caunedo@cornell.edu 2Assistant Professor, Yale University and NBER. mayara.felix@yale.edu 3Assistant Professor: Columbia University. km3924@columbia.edu Este trabajo estudia cómo los shocks del mercado laboral originados en el sector no agrícola afectan la organización de la producción agrícola. Utilizando datos exhaustivos de Brasil entre 1986 y 2017, se muestra que la entrada de grandes empresas no agrícolas conduce a aumentos persistentes en los salarios

locales, disminuciones en el empleo agrícola y un cambio hacia una agricultura más intensiva en capital. La explotación agrícola se consolida, disminuye el número de pequeñas operaciones y aumenta la mecanización. Para estudiar la magnitud de los distintos canales que contribuyen a esta reorganización de la producción, se desarrolla un modelo de equilibrio general que incluye empresas y fincas heterogéneas a través de rendimientos decrecientes, un margen de mecanización endógeno en la agricultura y efecto renta. Las fincas sustituyen óptimamente el capital por mano de obra en respuesta al aumento de los salarios que sige al incremento de la demanda de mano de obra. Al mismo tiempo, los hogares demandan relativamente menos productos agrícolas a medida que sus ingresos aumentan con una mayor productividad en el sector no agrícola. Calibrado para ajustarse a la economía brasileña, este modelo predice que una reducción de los costos de entrada en el sector no agrícola -suficiente para generar el aumento salarial observado en los datosconduce a una reasignación de la mano de obra fuera de la agricultura, al abandono de las fincas y a una intensificación del capital. Cuando mantenemos fija la mecanización, estos ajustes se atenúan sustancialmente, destacando el papel de la adopción endógena de tecnología como importante mecanismo de amplificación. Por último, el modelo es utilizado para explorar cómo la heterogeneidad en el tamaño de fincas y la mecanización puede racionalizar respuestas dispares a perturbaciones similares de la

demanda de mano de obra no agrícola. K E Y W O R D S Cambio Estructural, Agricultura, Entrada de Empresas, Mercado Laboral Pequeñas secciones del texto, menores a dos párrafos, pueden ser citadas sin autorización explícita siempre que se cite el presente documento. Los resultados, interpretaciones y conclusiones expresados en esta publicación son de exclusiva responsabilidad de su(s) autor(es), y de ninguna manera pueden ser atribuidos a CAF, a los miembros de su Directorio Ejecutivo o a los países que ellos representan. CAF no garantiza la exactitud de los datos incluidos en esta publicación y no se hace responsable en ningún aspecto de las consecuencias que resulten de su utilización. ©2025 Corporación Andina de FomentoCAUNEDO ET AL . 2 1 | INTRODUCTION The process of structural transformation—the reallocation of labor and resources from agriculture to manufacturing and subsequently to services—is a defining feature of economic development. Classical theories predict that as economies develop, differences in sectorial productivity trends and income effects lead the reallocation of labor into non-agricultural sectors (Herrendorf et al., 2014). Evidence from industrialized economies suggests that changes in industrial technology attracting labor out of agriculture were key to labor reallocation until the 1920s, while improvements in agricultural technology lead labor reallocation post 1960s (Alvarez-Cuadrado and Poschke, 2011). While either technological change lead to productivity improvements in agriculture, the underlying forces driving those improvements differ, and so the channels driving labor reallocation and its implications for wages. The study of recent developers has put particular emphasis on technological change and shocks to agriculture (Bustos et al., 2016; Imbert and Ulyssea, 2023). In contrast, we know

relatively little about the impact of positive shocks in non-agriculture for the organization of production in agriculture, despite their growing relevance in the context of industrial development, urbanization, and firm expansion. In this paper, we study how entries of large-scale non-agricultural firms that shifts labor demand affect the organization of agricultural production, including changes in farm size, capital intensity, and labor engagement. Canonical models of structural change suggest that firm entry, and the associated improvements in non-agriculture productivity, leads to labor reallocation into agriculture (rather than away from it).1 Combining rich panel data from Brazil’s municipalities —including agricultural censuses, population censuses, and employer-employee matched data for the overall economywe show that non-agriculture firm entry is associated with labor reallocation out of agriculture and increases in wages. At the same time, non-agriculture entry is associated with more mechanization in agriculture and small farm exit.These effects are identified from staggered different and different design comparing municipalities that received large entry shocks, and those in the same microregion that did not. We show that most of these entries are accounted for by the construction and the real-state sectors, and that those are in turn correlated with a government program for low-income housing.2 We develop a general equilibrium model in which the organization of agricultural production changes endogenously to relative prices, and therefore, developments in the non-agricultural sector. Decreasing returns in production technologies lead to a non-trivial size distribution of firms in non-agriculture and farms in agriculture (Hopenhayn, 1992) but nests, in its aggregate form, the properties of standard models of structural change. One key departure form a canonical model is that capital intensity in the agricultural sector is endogenous, Zeira (1998). As in Caunedo and Kala (2021), the model is tractable despite

heterogeneity in farm productivity and potentially, incentives to mechanize. The measure of operating firms and farms as well as the average productivity in each sector are equilibrium outcomes that depend on the respective operating and entry costs. When entry costs in non-agriculture fall, average productivity in this sector increases as does the measure of firms in the market. This entry shifts the price of output as well as the relative price of capital to labor, inducing capital intensification in agriculture. We calibrate our benchmark model to match key features of the Brazilian economy in the 1990s, at the beginning of our sample. We then run comparative static exercises where we lower the cost of entry in non-agriculture inducing new establishment entry, and study 1Herrendorf et al. (2014) provide a review of research on structural transformation. 2Our results are robust to controlling for soy production, which is a salient commodity during the study period both because of the surge in international prices, as well as the expansion of production in Brazil.CAUNEDO ET AL . 3 its implications for the organization of production in agriculture. Our model predicts that following a a reduction in the non-agricultural entry cost—calibrated to generate a 2.2% increase in wages—agricultural employment falls by 2.4%, the number of farms declines by 2.1%, and capital per farm rises by over 8%. The large response in capital intensification is accompanied by land consolidation, with an increase in land-per farm of 4%. These effects are economically meaningful. The increase in wages is consistent with the effects of large firms entries observed over the last three decades in Brazil. Our empirical strategy exploits rare episodes of large firm entry—defined as the opening of new establishments with 100 or more workers—and tracks their effects at the municipality level. Using a staggered difference-in-differences framework, we compare treated municipalities (those whose large entry shocks make up a large share of baseline formal sector employment) with never-treated controls, weighted by entropy balancing on a rich set of baseline characteristics. We find evidence that large non-agricultural firm entry causes formal non-agricultural employment to increase and raises wages in both agriculture and non-agriculture. Agricultural labor declines persistently, while the number of farms falls and land per farm increases. Consistent with capital-labor substitution, we also find increases in capital intensity, measured by tractor use, and a decline in the share of small farms. To understand the contribution of the mechanization channel to these adjustments, we contrast the baseline results with a counterfactual in which the level of mechanization is held fixed. In this environment, the decline in agricultural employment is less than half as large, and capital per farm is unchanged. Farm exit still occurs, but labor intensity remains high, and production adjusts only through contraction. These results highlight that endogenous technology adoption is a critical mechanism through which “pull” shocks in non-agriculture generate reorganization and reallocation in agriculture. Importantly, capital deepening in agriculture emerges as both a response to and a mediator of labor market shocks, reinforcing the structural transformation process. Finally, using our quantitative model we plan to extend the analysis by exploring how initial differences in agricultural organization—such as landholding patterns and capital intensity—shape the response to non-agricultural entry. Brazil’s regions vary widely in their reliance on laborversus capital-intensive farming, as well as prevalence of smallholder versus large-scale farms, suggesting that identical shocks may generate heterogeneous adjustments. To capture this, we are working on recalibrating the model to reflect regional characteristics and simulate the same entry shock across these settings. This will allow us to test the model’s predictions under different baseline conditions and shed light on how local production structures mediate the effects of labor market shocks on agricultural

transformation. Literature Review The debate about the role of technological progress in the agricultural and industrial sectors for the process of structural change dates back to the 1950s at least. Classical references like Harris and Todaro (1970) sustain that improvement in urban wages are fundamental to the shift in employment away from agriculture. At the same time, the proponents of the “food problem" as a central deterrent to sectorial reallocation, Schultz (1953), argue that agricultural productivity growth is key to releasing workers out of agriculture. More recent work is also divided, Gollin et al. (2007) emphasizes agricultural productivity growth, while Boppart et al. (2023) emphasizes productivity growth elsewhere in the economy as driver of value-added and employment reallocation. Our study leverages microlevel evidence to argue that these two processes are linked together through endogenous technology adoption. On the technical side, we propose a theory amenable to the discipline of modern micro-CAUNEDO ET AL . 4 data. This discipline allows us to expense away from pure inference through relative prices, (Alvarez-Cuadrado and Poschke, 2011), while being consistent with aggregate patterns. Our analysis underscores the importance of relative input price changes influencing technology adoption in agriculture, building on insights from Manuelli and Seshadri (2014), Caunedo and Keller (2020) and Caunedo and Kala (2021). An important implication of our analysis, is that sluggish productivity growth in agriculture may not have its genesis in the agriculture sector. Hence, perhaps the puzzle to the sluggish adoption of technology in agriculture Suri and Udry (2022) resides to some extent, elsewhere in the economy. 2 | THE IMPACT OF NON-AGRICULTURE ENTRY ON THE AGRICULTURAL SECTOR 2.1 | Data

This section describes the data sources we use the estimate the effects of new large firm entry on the local labor markets and agricultural sector, and to estimate our model. Our main data sources are the Brazilian agricultural census and linked employer-employee confidential dataset Relação Anual do informações Sociais (RAIS).3 The unit of analysis are municipalities and the unit of time is a decade, since the Agricultural census runs every ten years.4 Agricultural Sector We use the Brazilian agricultural census data for years 1995-96, 2006, and 2017 to obtain information on the organization of the agricultural sector. We study farm area, agricultural labor, value of produced output, and number of tractors employed. We also obtain information on input expenditure shares as well as on the number of agricultural establishments (farms). We further classify these establishments by land size: a small farm if less than 100 hectares (ha henceforth), a medium farm if between 100 and 500 ha, and a large farm if holding more than 500 ha. In addition, we obtain land usage rights to determine how much of used land is rented and how much is owned. Labor Markets We use three complementary data sources to derive labor market outcomes: RAIS, which covers nearly the universe of formal employees on an annual basis since 1986; the Agricultural Census for agricultural labor; and the Brazilian Population Census conducted in 1991, 2000, and 2010 to obtain information on informal employment as well as internal migration across municipalities. From RAIS we aggregate total formal employment by sector at the municipality level. We also obtain information on the average years of schooling for workers in a given sector. Finally, using information on earnings, we compute average wage per hour and average monthly earnings in a given sector in units of the minimum wage. Wage per hour is

3We use the RAIS data cleaned by Engbom and Moser (2022). 4While we refer to the unit of analysis as municipalities, the actual unit of analysis is Minimum Comparable Area (MCA) during study period, which are the smallest consistent sets of municipalities. There are 5,352 MCAs, covering 5,507 municipalities, between 1997 and 2017. We construct MCAs using information from the Brazilian Census on which municipalities split and/or emerged as combinations of which municipalities between 1991 and 2010. See https://www.ibge.gov.br/geociencias/organizacao-do-territorio/estruturaterritorial/15771-evolucao-da-divisao-territorial-do-brasil.html . 1,016 of of the 5,507 observed in the Census and RAIS datasets in 1997-2017 are municipalities that split off from or joined other municipalities during this period.CAUNEDO ET AL . 5 computed from monthly earnings of a worker multiplied by 12, to obtain annual earnings, and divided by the total hours worked in a given year for each employee. The Agricultural Census (Ag Census henceforth) provides information on the total agricultural labor at the municipality level. RAIS also provides information for agricultural labor, but only includes formal employment. Our empirical analysis includes annual estimates using data from RAIS, as well as decadal estimates for the outcomes from Agricultural Census. To match the timing of Agricultural Census, we use annual data or data for the entire decade preceding years 1997, 2006 and 2016.5 From the population Census we obtain information on the number of workers in agricultural and non-agricultural sectors, including informal ones. We further disaggregate those workers based on the level of their education into three groups – those with no or

only (some) primary education, those with some secondary education, and those with some tertiary education. For this disaggregation we restrict the sample to individuals who are older than 10 years old and were working at some point of the year when the survey was conducted. Since both censuses are conducted every ten years, and our main outcomes stems from the Ag census, we match information from the 1991 population census with Ag census data in 1995-’97, the 2000 population census with Ag Census for 2006, and the 2010 population census with agricultural data for 2017. The measure of informality follows the standard in the literature, Ulyssea (2018), i.e. the share of employed workers aged 18-64 with non-zero earnings that do not have a signed Carteira de Trabalho(employee registration). We use the 1991, 2000, and 2010 census to measure informality at the baseline entry decade as follows. Shocks between 2011 and 2016 have the 2010 as baseline, shocks between 2001 and 2010 have the 2000 census as baseline, and shocks before 2000 have 1991 as the baseline. Finally, we build measures of internal migration across municipalities from the population census, which we further disaggregated by individual socio-economic characteristics, like age or education level. International Trade We use data on imports and exports at the municipality-year-product level from Comex Stats, available for 1997-2022.6 We aggregate these data at the municipality and year level, separately for agricultural and non-agricultural products. Additional Data Sources We complement our analysis with microregion-level exposure to import competition during Brazil’s 1990-1994 trade liberalization, from Dix-Carneiro and Kovak (2017), which is used as an additional variable to balance control and treated units.

2.2 | Motivating Facts Before turning to our empirical strategy, we document a set of stylized facts that motivate our analysis and contextualize our identification approach. These facts, drawn from national and municipal-level data, illustrate the ongoing process of structural transformation in Brazil and provide suggestive evidence that localized labor demand shocks originating in non-agriculture can affect agricultural outcomes. Figure 1 shows standard patterns of structural transformation, with total employment in agriculture declining and (formal) employment in non-agriculture increasing, Herrendorf 5We use 1997 as the first year instead of 1996 as, prior to this year, not all municipalities were covered in RAIS. 6Downloaded from https://basedosdados.org/dataset/74827951-3f2c-4f9f-b3d0-56e3aa7aeb39?table= f4b08023-5530-4dc9-bced-3321e8928fd7 .CAUNEDO ET AL . 6 et al. (2014). At the same time, the number of tractors per agricultural worker rises steadily, reflecting increased mechanization in the agricultural sector. These patterns are consistent with a broader structural transformation, whereby labor reallocates away from agriculture while farms adopt more capital-intensive modes of production, Chen (2020). .16.18.2.22.2416.516.751717.2517.5199720062017Agricultural Census YearLog employment in agriculture (Agricultural Census, left axis)Log formal employment in non-agriculture (RAIS, left axis)Tractors per farm (Agricultural Census, right axis) F I G U R E 1 National trends in tractors per farm and employment. Notes: Data on tractors and agricultural employment are from the Agricultural Censuses for 1997, 2006, and 2017. Data for

non-agricultural (formal) employment are from RAIS for the respective years. At the municipal level, we can further investigate the proximate sources of such variation by bringing in data on wages. Higher agricultural wages are associated with greater mechanization, lower agricultural employment and higher formal non-agricultural employment. Figure 2 presents binned scatters plots using municipality-level data for each year in which the Ag Census was conducted. Panel A shows a positive relationship between agricultural wages and tractors per farm, and a negative relationship between agricultural wages and agricultural employment. Hence, municipalities where labor is relatively more expensive display more capital intensive farms and lower agricultural employment. At the same time, Panel B of Figure 2 shows that municipalities with higher levels of formal non-agricultural employment also exhibit higher agricultural wages. Plausible explanation of observed patterns is that tighter labor markets in non-agriculture, or higher productivity in non-agriculture spills over into agriculture, putting upward pressure on rural wages. Hence, labor demand shocks originated in non-agriculture can affect agricultural production decisions through general equilibrium wage effects.CAUNEDO ET AL . 7 F I G U R E 2 Organization of Agricultural Production and Local Labor Markets Panel A: Municipal-level mechanization, wages, and employment in agriculture 0.05.1.15.2234567Log (Agricultural wage) year=1997 year=2006 year=2017Tractors per agricultural worker and Agricultural wages, by year (a) Tractors per worker and ag wage 6.577.588.5234567Log (Agricultural wage) year=1997 year=2006 year=2017Agricultural employment and Agricultural wages, by year (b) Agricultural employment and ag wage

Panel B: Municipal-level wages and employment in non-agriculture 3.544.555.54681012Log(Non-Agricultural employment) year=1997 year=2006 year=2017Agricultural wages and Non-Agricultural employment, by year

(c) Ag wage and non-ag employment 77.27.47.64681012Log(Non-Agricultural employment) year=1997 year=2006 year=2017Non-Agricultural wages and Non-Agricultural employment, by year (d) Non-ag wage and non-ag employment Notes: This figure plots binned scatters showing cross-municipality variation, separately by year and conditional on region fixed effects, for various municipal-level outcomes. Data on tractors, agricultural wages, and agricultural employment are from the Agricultural Censuses for 1997, 2006, and 2017. Data for non-agricultural wages and employment are for the formal sector only, from RAIS for the respective years. Agricultural wages are calculated as the sum of nominal labor expenses divided by the total number of workers employed in agriculture in each municipality. Non-agricultural wages are measured as average worker monthly earnings and excluding workers in government. All wages are converted to 2017 reais using the iPCA inflation index. Taken together, these patterns motivate our empirical approach. First, structural transformation is ongoing in Brazil over our study period, with agricultural employment declining and production becoming increasingly capital-intensive. Second, this process is heterogeneous across regions, and local labor market conditions —as summarized by wages in the agriculture and non-agricultural sector— correlate with the relative employment shares across sectors and importantly, with capital intensity. Since these are equilibrium outcomes in the labor and output markets, it is not possible to interpret these wage differentials, and associated outcomes, as causal. To learn more about the plausible causal role of wage increases on the organization of production in

the agriculture sector, we study large firm entries into local labor markets and build a difference-in-differences research design.CAUNEDO ET AL . 8 2.3 | Large Firm Entry Shocks Our goal is to identify non-agricultural firm entry events that are large relative to the local labor market and therefore generate a shift in labor demand, bidding up local wages. We define a large entry shock as a non-agricultural new establishment entry comprising more than 0.01% of its microregion’s population in the year preceding the shock and with at least 200 employees at entry. 7 The former restriction identifies new entries that are large enough relative to a potential labor market of a region, whereas the second removes very small regions, for which even a 4-employee entry (5th percentile of new entering establishments) would otherwise be considered a large entry. One concern with interpreting the effects of these large entry shocks as exogenous to the agriculture sector is that the firms entering these local labor markets may also rely heavily on agricultural goods as inputs, and, hence, choose the location based on inputs availability and proximity. To address this, in our baseline estimates we exclude large entries in sectors directly demanding agricultural products, such us Food and Beverages, Textiles, and Rubber and Tobacco. Figure 3 displays the size distribution of all firm entries that account for at least 0.01% of the population in the microregion, by the number of employees at entry. It also presents the distribution conditional on at least having 200 workers at entry. The median employment share accounted for by entry of new firms across municipalities is 238 employees, but the distribution is highly skewed with the 75th percentile of the distribution being 473 employees at entry, and many entry events with 1000+ employees. Our cutoff of 200 workers to define an entry shock is slightly below the median employment at entry.

5th percentile: 4 employees50th percentile: 238 employees75th percentile: 473 employees05010015020001000200030004000Number of employees at entryAll entriesEntries with at least 200 employeesDistribution of non-ag new establishment entries with more than 0.01% of microregion population F I G U R E 3 Distribution of large entry shocks by number of employees ant entry. These large entries are unevenly distributed along industries and space. Figure 4 displays the distribution of large entry shocks through the years in our sample. Between the late ’80s and the early 2000, most of these entries were concentrated in the real-state (in terms 7This restriction is necessary to exclude new entries at very small microregions. Not doing so results in the inclusion of municipalities like the island of Fernando de Noronha, which is a natural reserve and would otherwise be the municipality with the largest number of ”large entries”, 28, per the microregion population cutoff in the data. We explore robustness to alternative thresholds, such as 100 or 300 employees.CAUNEDO ET AL . 9

Panel A: By Industry 05,00010,0001986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016AutomobileChemical, pharmaceuticalConstructionEducationElectronics, communicationFootwearHospitalityMetalMineralMiningOther manufacturingReal estateRetailTransport, communicationUtilitiesWholesaleWood, furniture Panel B: By Region 05,00010,0001986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016Central-WestNorthNortheastSouthSoutheast F I G U R E 4 Total Employment from Large Entries at Entry Year.Notes: A large entry is a new

non-agricultural establishment with more than 0.1% of total microregion population at entry and at least 200 workers. Excludes entries in manufacturing activities that directly demand agricultural inputs (i.e., food, beverages, textiles, and paper). of employment contributions), followed by the construction sector and the transport and communication. Since the early 2000s, most of the employment contribution of these entry shocks stems from the construction sector. It is not surprising then, that the nature of our shocks is mostly concentrated in non-tradables, rather than in tradable industries, see Figure A.2. In terms of spatial distribution,