WPS8087 Policy Research Working Paper 8087 Beyond Dualism Agricultural Productivity, Small Towns, and Structural Change in Bangladesh M. Shahe Emran Forhad Shilpi Development Research Group Environment and Energy Team June 2017 Policy Research Working Paper 8087 Abstract This paper uses a framework that goes beyond rural-urban higher employment in the town-goods sector. Using rainfall dualism and highlights the role of small town economy in as an instrument, the empirical analysis finds a significant understanding structural change in a developing country. positive effect of agricultural productivity on rice yield and It provides a theoretical and empirical analysis of the role agricultural wages. Productivity shock increases wages more of agricultural productivity in structural transformation in in the rural sample compared with the small town economy the labor market. The empirical work is based on a general sample, but structural change in employment is more pro- equilibrium model that formalizes the demand and labor nounced in the small-town economy. In the rural sample, market linkages: the small-town draws labor away from the it increases employment only in small-scale manufacturing rural areas to produce goods and services whose demand and services. In contrast, a positive productivity shock has may depend largely on rural income. The theory clarifies large and positive impacts on employment in construc- the role played by the income elasticity of demand and the tion and transport, education, health and other services, wage elasticity with respect to productivity increase in agri- and manufacturing employment in larger scale enterprises culture. For productivity growth to lead to a demand effect, located in small towns and cities. Agricultural productivity the wage elasticity has to be lower than a threshold. When growth induces structural transformation within the services the demand for goods and services produced in small towns sector in small towns, with employment in skilled services comes mainly from the adjacent rural areas, the demand growing at a faster pace than that of low skilled services. effect can outweigh the negative wage effect, and lead to This paper is a product of the Environment and Energy Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at fshilpi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Beyond Dualism: Agricultural Productivity, Small Towns, and Structural Change in Bangladesh M. Shahe Emran IPD, Columbia University Forhad Shilpi World Bank Key Words: Agricultural Productivity, Small Town Economy, Dualism, Employment in Large Firms, Employment Growth, Structural Transformation JEL Codes: O13, O14, O17, O18 1. Introduction A substantial literature in development economics focuses on structural change where economic de- velopment is characterized by labor reallocation from a contracting agricultural sector to the expanding manufacturing and services sectors. In much of the literature, agriculture is equated with the rural econ- omy, and manufacturing and services with urbanization. Although such dualistic perspectives have been at the center of a large body of literature starting from Lewis (1954), and Kuznets(1973) the more recent work emphasizes the need for a broader framework, beyond the binary conceptualization of dualistic models, to capture the richness of structural change in a developing economy. In the context of rural-urban dualism, a growing literature underscores the fact that the “rural” and “urban” are two polar cases in a broader spectrum, and many geographic and administrative units are better characterized as partially urbanized. A simple yet useful framework that goes beyond the canonical rural-urban dualism is where the focus is on areas that contain a small town surrounded by significant rural population and agricultural activities. Contrary to the popular impression of most urban people being crammed into mega-cities, recent evidence shows that a large share of urban population and rural- migrants live in the smaller cities and towns (Ferre, Ferreira and Lanjouw (2010);Christiaensen, De Weerdt and Kanbur (2015)). In the spatial spectrum from metropolitan cities to villages, the small towns occupy a space in the middle in terms of population and employment density. As suppliers of goods and services and destinations of rural workers, the small towns have ties with both metropolitan cities and rural areas- the ties being perhaps closer with the rural areas (Haggblade, Hazell and Reardon (2006)). This paper aims to analyze structural change in employment in a ‘small-town economy’ with a focus on the role played by higher agricultural productivity in the surrounding villages. The focus on agricultural productivity change as a driver of employment specialization and structural change is primarily motivated by a longstanding debate about the role of agricultural productivity in industrialization and structural change. The earliest view in development literature identified a trade- off between poverty reduction through agricultural development and long-run structural change, because 1 higher agricultural productivity is likely to increase the wages faced by the nonfarm sectors (Lewis (1954), Foster and Rosenzweig (2004)). The competing view emphasizes the Engel curve effect in consumer demand reflecting inelastic demand for food, and higher income elasticity of manufacturing goods and services. Since much of the rural population are engaged in agricultural activities, a rise in agricultural productivity can increase rural income and thus have a positive demand effect on non-farm activities, including those located in small towns (Mellor (1976), Ranis and Stewart (1973), Haggblade, Hazell and Reardon (2006)). Recent empirical studies suggest a bigger role of labor market linkages. Foster and Rosenzweig (2004) find a negative relationship between manufacturing employment and agricultural productivity growth providing support to the contention that there is a trade-off between agricultural growth and manufacturing growth. Bustos, Caprettini, Ponticelli (2016) on the other hand find that labor saving technical change in soybean production in Brazil led to industrial growth. In this paper, we provide causal evidence on the impacts of agricultural productivity shock on employment growth and specialization in a more general context where both labor market and demand linkages interact with each other, and may lead to contrasting outcomes across different areas within the same country. We develop a simple model of a small-town economy that formalizes the dual roles that a town plays in such an economy: it draws labor away from the rural areas to produce goods and services whose demand may depend largely on rural income. There are three types of goods produced in this economy: food and informal good (low quality manufacturing and services) are produced in the villages, and a formal good (high quality manufacturing or services) is produced in the small town.1 A positive productivity shock to agriculture, however, does not necessarily increase village income; given a labor endowment, the village income increases only when the response of the wage (i.e, elasticity of wage with respect to productivity shock) is lower than a threshold.2 When the elasticity of demand for the town good in the village is low so that most of the town good is sold outside the local economy, the labor market linkage predominates 1 The assumption that the town produces only one good is for the sake of simplicity. In the empirical analysis, we consider disaggregation of the town good, some of which are of relatively lower quality with lower income elasticity. 2 In a more general model with a positive labor supply curve at the household level, a positive shock to agriculture is expected to increase village income in general. 2 and a higher rural wage due to agricultural productivity increase reduces employment in small towns. In contrast, when demand for goods and services produced in small towns comes mainly from the adjacent rural areas, the demand effect can more than offset the negative wage effect, and lead to higher labor allocation to the production of the the town good. We test the predictions of the model using panel data compiled from the population and enterprise censuses from Bangladesh. Our data set covers the period between 2000 and 2010. Between 2000 and 2010, Bangladesh experienced substantial reduction in the incidence of poverty, from 48.9 percent to 31.5 percent. This decade also witnessed substantial expansion in non-farm employment as a result of which its share in total employment increased from 47 percent to 52 percent. The rice yield, which is taken as a measure of agricultural productivity, has grown at an annual rate of 3.6 percent.3 To understand the implications of agricultural productivity, we exploit variations in rainfall across upazilas and over time, and implement an approach that focuses on the effects of rainfall shocks in reduced form regressions on the outcome variables (employment in different types of non-farm activities) and also on the measure of agricultural productivity (crop yield). Our approach relies on the fact that rainfall variations can be interpreted as shifts in the production function, because rainfall is a major determinant of crop yield in Bangladesh (Sarkar et. al. (2012)). We also provide an instrumental variables interpretation of our estimates, using rainfall deviations (from long term average) across upazilas and over time as an instrument for crop yield (rice yield). Empirical estimation issues and strategies to deal with them are discussed in detail in Section 3. It is worth noting here that while rainfall shocks have been used for identification in a variety of contexts, agricultural productivity is arguably one of the most natural contexts where rainfall can provide reasonable identifying variations ((Foster and Rosenzweig (2004), Adhvaryu, Chari and Sharma (2013), Bruckner and Ciccone (2011)). The empirical results establish positive and statistically significant effects of rainfall on rice yield and agricultural wages. The positive effect of rainfall shock on agricultural productivity (as measured by rice 3 Rice is the predominant crop in Bangladesh. 3 yield) is similar across village and small-town samples but its effect on wage is larger in the village sample compared with the small-town sample. A positive rainfall shock increases employment in the production of informal village good (small-scale manufacturing and services) but has no statistically significant effect on the labor allocated to large-scale manufacturing employment in the rural sample. A positive productivity shock has large and positive impacts on employment in construction and transport, education, health and other services, and manufacturing employment in larger scale enterprises located in small towns. When interpreted as instrumental variables estimates of the effects of productivity increase, the empirical results suggest that agricultural productivity induces structural transformation within the services sector with employment in skilled services provided by larger firms growing at a much faster pace compared with services provided by smaller firms/individuals. Our analysis brings to focus a methodological point about the sample used in the empirical analysis. Most of the existing analysis of the effects of agricultural productivity relies on a sample of villages where all of the households are classified as rural. Our analysis shows that it is important to use the small towns in addition to the “village sample”. The rest of the paper is organized as follows. Section 2 provides a simple general equilibrium model to highlight the different channels of interactions between farm and non-farm sectors located in rural areas and towns. Section 3 describes the econometric strategy followed by a description of data used in empirical analysis in section 4. Section 5 organized in different sub-sections presents the main empirical results. Section 6 concludes the paper. 2. The Model We develop a simple general equilibrium model of a small-town economy (STE, for short) composed of both rural and urban households. The model is intended to capture the local economy at the upazila level as our empirical analysis is based on an upazila level panel data set. The representative rural household lives in a village, is endowed with A amount of agricultural land, and is engaged in the production of food (f ) and a composite village good (S ) (low quality manufacturing or services produced by small family enterprises). 4 The representative urban household lives in the small town located in the same upazila adjacent to the village, and is engaged in the production of a composite town good denoted as T − good, using capital and labor. The town has no agricultural land and can borrow at interest factor r for capital investment. The T −good represents relatively high quality manufacturing and services produced by large-scale formal enterprises. To incorporate the fact that the small town is also connected to the large city, we assume that part of the demand for T -good comes from the metropolitan city. To model this link in the simplest form, the income in the metropolitan city Y o is determined outside the model and the representative agent spends a constant share φ of her income on the T −good. Thus the only mechanism through which the metropolitan demand for T −good can be affected by agricultural productivity growth is changes in the price of T −good. There are L ¯ households in the small-town economy, each household supplying 1 unit ¯ . Labor is of labor inelastically, implying that the total labor endowment in the small-town economy is L assumed to be freely mobile between the village and the town at zero migration costs.4 In this model, a more rural upazila can be understood as a small-town economy with a large endowment of agricultural land, given the labor endowment. We assume food to be internationally traded and take it as the numeraire commodity, setting Pf = 1. The utility functions for households in the village and small town are: U v = U v (Cf , Cs , CT ); U T = U T (Cf , CT ) where superscript v refers to village households and T to town households, and Cf is consumption of food, Cs is consumption of the composite village good produced and consumed within the village, CT is the consumption of the town good. An important feature of this formulation is that it allows the village households to consume high quality goods and skilled services produced by formal firms in the town, but the low quality village good S is not consumed in the town. To provide an example, a health service 4 The zero migration costs assumption is not necessary for the results that follow. However, this is a good approximation in our context, because we are considering short-distance migration to the small town within a given upazilla. 5 available within the village (Cs ) could be a traditional healer whereas skilled service in the town could be a doctor trained in formal medical college. To model the Engel curve effect in a simple way, we assume that there is a subsistence requirement for food in the village and the utility function is of Stone-Geary form in food: ( v ) ∑ U v = υf ln Cf − γf + υs ln (Cs v v ) + υT ln (CT ), υi = 1, and i = f, s, T (1) where γf > 0 is the subsistence or minimum level of consumption of food. For simplicity, utility function in town is assumed to take the log linear form (log linearized Cobb-Douglas) with τk as the share ∑ of good/service i in town consumption ( k τk = 1; k = f, T ): ( T) ( T) U T = τf ln Cf + τT ln CT (2) We assume that υf < M in (υs , υT ) and τf < τT . This captures the idea that as income increases, consumers spend a larger share of the additional income on non-food items. We remain agnostic about the relative magnitudes of υs and υT , because they might differ across areas depending on the level of household income. For example, in a more urbanized (and high income) upazila, we would expect that the share of non-subsistence budget devoted to the better quality T −good is the largest. The budget constraint for the households can be stated as: Y v = Cf + Ps Cs + PT CT ; Y T = Cf + PT CT where Ps and PT are prices of village and town goods respectively, and Pf = 1 parametrically given from outside the local economy. Y v and Y T are total household incomes in village and town respectively. Given the assumptions about preference in equations (1) and (2) above, the demand functions in each 6 location can be expressed as:            Yv if Y v ≤ γf  0 , if Y v ≤ γf Village: Cf = ; Ci =     υi [ Y v − γ f ]      γf + υf [Y v − γf ] if Y v > γf  , if Y v > γf Pi τk Y T Town: Ck = , with k = f, T Pk Assuming that Y v > γf , the income elasticity of demand for different goods in the village can be expressed as: υf Y v Yv Ef Y = < 1; E iY = > 1, ∀i = s, T υf Y v + (1 − υf )γf Y v − γf For the rural households, demand for food is income inelastic, and demand for village and town goods are income elastic. The larger is the subsistence requirement (γf ), the higher is the income elasticity of village good S and town good T in the demand function of the village households. 2.1 Employment and Income Determination in the Village Households in the village produce food using land and labor under CRS technology. The optimization in food production is as follows: ( ) −α M axLf Πf = θ Aα L1 f − wLf where θ ≥ 1 represents total factor productivity in food production, A is the endowment of land, lf is the labor allocated to food production and w is the wage rate. A larger land endowment as represented by a higher value of A can be interpreted as a more rural (agricultural) local economy, given the labor ¯ . A higher agricultural productivity is captured by an increase in θ. We assume that the endowment L locally consumed village good (S ) is produced under a CRS technology using only labor: 7 M axLs Πs = Ps (ϕLs ) − wLs where ϕ is the technology in the production of village good (and service). This specification of the production function for the S good is appealing both in terms of simplicity and realism in the context of a developing country such as Bangladesh where small-scale family enterprises in villages use little capital or land.5 Given a wage rate w, the demand for labor in agriculture and village good can be derived as: [ ]α 1 (1 − α) υs v L∗ 1 f =θ A α , Ls = [Y − γf ] (3) w w w The equilibrium price of the village non-farm good is pinned down by the wage rate: Ps = . Denoting ϕ the return to land by µ, the total income from agriculture (Yf = Qf = wLf + µA) can be solved using the wLf marginal conditions for optimization as Yf = (1−α) . The total household income in the village (denoted as Y v ) thus becomes: wLf υs γ f Y v == Yf + Ys = + wLs = ∆1 wLf − (4) (1 − α) (1 − υs ) 1 where ∆1 = (1−α)(1−υs ) . Thus village income net of subsistence requirement and the labor allocated to the village non-farm production can be written as below: Y v − γf = ∆1 [wLf − (1 − α)γf ] (5) ∆1 υs L∗ s = [wLf − (1 − α)γf ] (6) w 5 The main implications of the model will hold if we use more general CRS production function using other factors (e.g. capital) as well. 8 2.2 Employment and Income Determination in the Town v T The demand for town good comes from three sources: village (CT ), town itself (CT ) and outside areas O 6 (e.g. metropolitan city or other countries) (CT ). The demand generated by the metropolitan city is o φY o CT = where d is the distance (transport cost) to the metropolitan city from the small town, PT (1 + d) and Y 0 is the exogenously given income in the metropolitan city. When all of the demand for town good comes from outside the local economy so that there is no demand linkage with the village (υT = 0), the only linkage between the town and the village within an upazila is due to the labor market linkage. On the o other hand, if (φ = 0) implying that (CT = 0), then the demand linkage between the town and the village will be at the maximum. The town good CT is produced under CRS technology using labor and capital: −β QT = K β L1 T We assume that the urban segment of the economy is populated by a single representative agent who invests K > 0 amounts of capital in the production of town good at an interest factor r which is determined outside the model. The optimization of the representative agent in the town is as follows: −β M axK,LT ΠT = PT K β L1 T − wLT − rK ∗ It is easy to check that the equilibrium price of the town good is a function of w and r , PT = ( ) 1−β ( ) β w r . Since the town borrows the capital from outside, its income equals the labor income, 1−β β i.e., Y T = wLT = (1 − β ) PT QT (w, r) where QT (w, r) is the optimal output chosen in the town facing the wage rate and interest factor.7 The demand for labor in town can be described as: 6 For simplicity, CTO is assumed to be nominal expenditure on the town good by consumers residing outside the village and town economies described here. 7 The outside here means that it is outside of the relevant metropolitan city. This keeps the model simple without losing any relevant insight. An alternative formulation is where the capital invested in small towns comes from the metropolitan city, and metropolitan city income in that case responds to agricultural productivity growth when the town income is influenced. 9 ∆3 ∆3 φY o L∗ T = [υT (Y v − γf ) + ] w w (1 + d) ∆3 ∆1 υT (wLf − (1 − α)γf ) ∆3 φY o = + (7) w w (1 + d) 1 where ∆3 = (1−τT ) . 2.3 Labor Market Equilibrium We assume that labor is perfectly mobile and migration cost is equal to zero, so that equilibrium wage (w) are the same regardless of location. The labor market equilibrium condition can be written as: [ ] υs + υT ∆3 { ∗ } ∆3 φY o L = L∗ ∗ ∗ ∗ f + Ls + LT = Lf + ∆1 wLf − (1 − α)γf + (8) w w (1 + d) 2.4 The Effects on Equilibrium Wage and Village Income We first analyze the effects of higher agricultural productivity as captured by a higher θ. Proposition (1) below summarizes the results on equilibrium wage w∗ and village income Y v . PROPOSITION 1 (1.1) In the small-town economy described above, a higher agricultural productivity (i.e., a higher θ) ∂w increases wage, i.e., ∂θ > 0. (1.2) A higher agricultural productivity increases (reduces) the village income when elasticity of wage w −1 with respect to θ (i.e., Ewθ ) is less (more) than (1 − α) . (1.3) The magnitude of the effects of an increase in agricultural productivity on wage depends on land endowment. A set of sufficient conditions for the effects of productivity on wage to rise with a larger land endowment are: (a) that the elasticity of wage with respect to θ is low enough at the initial equilibrium to { } −1 satisfy Ewθ < α (1 − α) , and (b) that the initial land/labor ratio is higher than a threshold. Proof: Omitted. Please see the (online) appendix. 10 Discussion Parts (1.1) and (1.2) show that although we would expect the wage rate to respond positively to a an increase in agricultural productivity, there is no guarantee that it will result in higher village income. To understand the uncertainty regarding the effects on village income, first note that the aggregate village ( ) income Y v = Y f + Y S can increase with a higher agricultural productivity only when agricultural income wL∗ increases as a result, because Y v = ∆1 wL∗ υ γ f f − (1−υs ) , and Y = s f f . The ambiguity in the income effect 1−α is the result of the tension that arises from the fact that a higher wage if strong enough may reduce the L∗ f so much that agricultural income falls after a positive productivity shock raising the crop yield. When −1 the elasticity of wage with respect to productivity Ewθ is equal to (1 − α) , the positive effect of a higher wage is precisely offset by the reduction in labor employed in the food production to keep Y f unchanged. −1 Thus when Ewθ < (1 − α) , the labor demand does not fall enough to wipe out the positive wage effect, leading to a higher agricultural income, and hence to a higher total village income. The focus on the elasticity of wage with respect to productivity change is partly motivated by the fact that in the empirical analysis, we provide an estimate of the elasticity of wage rate with respect to agricultural productivity (rice yield), and the theoretical results provide us a way to check if we can expect a positive demand effect from the observed wage response in the absence of reliable income data in the village. The result regarding the heterogeneous effects on wage rate (part (1.3) of proposition 1) guides our empirical work in that we split the sample into two sub-samples where the upazilas with low or no urbanization are put into a group to represent the purely village sample (large A), and a second sample consists of the upazila’s with relatively substantial urbanization. 2.5 The Effects on Intersectoral Labor Allocation We can utilize the labor demand functions along with the labor market equilibrium condition to investigate the impact of agricultural productivity shock on labor allocation among different subsectors. Proposition (2) below collects the results on the response of labor allocation to three different economic activities in the small-town economy. 11 PROPOSITION 2: (2.1) It is not possible to have structural change in employment that reallocates labor from agriculture to the S -goods sector in a village when there is no subsistence requirement for food (i.e., γf = 0), and hence the demand for food is not inelastic. However, if γf = 0 but φ > 0, a higher agricultural productivity ∂L∗T results in reverse structural change by lowering labor allocation to the town good, i.e., < 0. With ∂θ γf > 0, the strength of the reallocation of labor in response to higher agricultural productivity to non-farm sectors is a negative function of the magnitude of υf . (2.2) An increase in agricultural productivity (i.e., a higher θ) leads to structural change in employment in the small-town economy described above by reducing the labor allocated to agriculture, when the elasticity of wage with respect to θ is greater than 1, i.e., Ewθ > 1. A necessary condition for Ewθ > 1 is that the food production technology is such that the elasticity of agricultural output with respect to land is greater than 0.50, i.e., α > 0.50. (2.3) Assume that Ewθ > 1. Then a higher agricultural productivity increases the employment in the ( ) S −goods sector in the village, if Ewθ is less than a threshold Ewθ S > 1. Assuming that Ewθ ∈ 1, Ewθ S , a higher agricultural productivity increases employment in the S -good sector more when the value of υs is larger. (2.4) Assume that E wθ > 1,and the subsistence requirement is higher than a positive threshold γf > ( ) T γf > 0. Then there exists a cutoff elasticity value Ewθ T such that1 < Ewθ S < Ewθ . When Ewθ ∈ 1, Ewθ T , an increase in agricultural productivity increases labor allocated to town good production. Proof: Omitted. Please see the appendix. Discussion To see that it is not possible to have structural change in employment that reduces employment in agriculture when γf = 0, consider the labor demand functions in equations (3) and (4). With γf = 0, we L∗ f have = ∆1 υs which is not affected by agricultural productivity shock θ directly or indirectly through L∗ S 12 equilibrium wage w∗ . So without γf > 0, it is not possible for agricultural productivity shock to affect labor allocation between food and S -good production in an upazila. However, allocation of labor between agriculture (food) and the town good can still be affected. Using equation (7), setting γf = 0, we can rewrite the labor allocation between food and T -goods sectors as: L∗ T ∆3 φY o ∗ = ∆ 3 ∆ 1 υT + ∗ ∗ (9) Lf w Lf From equation (9), it is clear that labor allocation between food and the town good is affected by θ even when γf = 0. An intuitive explanation of this effect is that a higher wage caused by a positive productivity shock increases the price of the town good and given Y o reduces the demand for the town good in the metropolitan area. This leads to labor reallocation from the town good to food production, resulting in reverse structural change. We have ∗ ( )−β ( )β ∂PT w r ∂w∗ = (1 − β ) >0 ∂θ 1−β β ∂θ It is obvious that when both γf = φ = 0, then the allocation of labor across the three sectors (f, S, T ) is not altered by a productivity shock in agriculture. An important point that comes across from propositions (2.3) and (2.4) is that when part of the demand for the T -good comes from the metropolitan city, the effects of agricultural productivity on employment growth in this sector will be limited at best. In contrast, for the goods which are produced exclusively for local consumption (S -goods), it is more likely that the demand effect generated by productivity growth helps unleash significant employment growth. The theoretical analysis brings into focus the role played by the values of Ewθ , α, γf , and φ. In the empirical analysis, we estimate Ewθ , and available evidence provides us a plausible value for α. It is important to appreciate that land (A) in our model denotes all non-labor inputs including land, and thus 13 the value of α is likely to be significantly larger than 0.50, a plausible value is 0.70.8 If α = 0.70, then −1 −1 (1 − α) = 3.333 and α (1 − α) = 2.333. 3. Empirical Strategy To estimate the effects of agricultural productivity growth on non-farm employment, we use a panel data model: Oijt = ρj + ρt + πθjt + ΠZjt + εijt (10) where i indexes the outcome variables (e.g. employment, wage etc), and j upazila. Oijt is the outcome variable i in period t for upazila j . ρj and ρt denote upazila and year fixed effects respectively. θjt is the measure of agricultural productivity, Zjt is a vector of upazila characteristics and εijt is the error term. Estimation of the impact of agricultural productivity on non-farm employment however presents challenges. Unobserved upazila characteristics (e.g., proximity to a river) may attract firms into an upazila (due to ease of transport) and also affect agricultural productivity (due to availability of water for irrigation) positively. In cross-section OLS regressions, one might thus attribute this correlation to agricultural productivity growth. We use upazila fixed effects ρj remove any time invariant but unobserved regional effects, including geographic and persistent agro-climatic factors such as soil quality. The year fixed effects (ρt ) control for any macro economic and international shocks (including commodity price shocks) that may have affected both agricultural productivity and our outcomes of interest.9 For the estimation of equation (10), the upazila level unmeasured fixed factors are removed by de- meaning all variables in the regression. Such de-meaning however may lead to attenuation bias as it magnifies any measurement error in the agricultural productivity variable (rice yield). To remedy this potential source of bias, we implement an instrumental variable approach. Following a large literature (Foster and Rosenzweig (2004), Adhvaryu, Chari and Sharma (2013)), agricultural productivity is measured 8 In green revolution areas, A includes seed, fertilizer etc. in addition to land. The share-cropper usually gets one third of the output when the landlord bears the costs for seed and fertilizer, thus putting α = 0.6666. 9 The year fixed effects will control for any general equilibrium effect common to all households (e.g. prices) also. 14 by rice yield as rice is the predominant subsistence and cash crop in Bangladesh. We use rainfall shock as an instrument for rice yield. Rainfall is found to affect agricultural yields in both developed and developing countries and hence it is used widely as an instrument for agricultural yields (Foster and Rosenzweig (2004), Adhvaryu, Chari and Sharma (2013), Bruckner(2012)). To be precise, we use deviation of current rainfall from its long term average as an instrument. The long term average rainfall is defined as the average rainfall over a 30 year period from 1960 to 1990. In the empirical analysis, we follow a two step procedure: first, we run a reduced form regression of an outcome variable (for example, formal non-farm employment in T -good production) on the instrument, and second, a reduced form regression of the productivity measure (yield per acre) on rainfall deviation. This two-step procedure based on reduced form regressions has some important advantages in our application. First, given the focus of our analysis is on the effects of productivity increase in agriculture, one can plausibly interpret the variations in rainfall as variations in the parameter θ in the model. Second, with crop yield as the measure of agricultural productivity, and rainfall as an instrument, the reduced form estimates of rainfall on outcome (employment) provide evidence on the existence of a causal effect of higher crop yield on the relevant outcome which is not subject to weak instrument bias (Chernozhukov and Hansen(2008)).10 Third, by imposing an exclusion restriction on the rainfall shock, we can derive instrumental variables estimates of the causal effect as the ratio of the two reduced form estimates.11 We estimate the following reduced form regressions: Oijt = ρj + ρt + π1 Rjt + Π1 Zjt + εijt (11) Vjt = ηj + ηt + π2 Rjt + Π2 Zjt + υjt (12) where Rjt is the deviation of annual rainfall from its long term average in upazila j and Vjt is the 10 We emphasize that the main results of this paper do not depend on the exclusion restriction on rainfall; what we need is that rainfall affects productivity significantly. 11 Moreover, the reduced form estimates of the effects of rainfall on the outcome variables such as non-farm employment are of interest on their own; for example, they provide us evidence on the potential benefits of increased irrigation investment on the rural non-farm economy. 15 measure of productivity. The deviation is defined as the difference between log of current rainfall from log of its long term average. Thus our empirical model with upazila and year fixed effects provides estimates of the impact of rainfall shock on the growth of the outcome variables. A positive coefficient of rainfall (π2 > 0) for instance in the yield regression means that an increase in rainfall over its long term average level (a positive rain shock) increases rice yield. To ensure that rainfall primarily captures variation in agricultural productivity, we include an appro- priate set of controls in Zjt . A potential concern in our context is that rainfall may be correlated with growth in non-farm employment for factors unrelated to agricultural productivity, and hence may not be reflecting the impact of agricultural productivity. For example, upazilas located near large urban centers are likely to attract large firms producing manufacturing and services (Fafchamps and Shilpi (2003) and (2005), Emran and Shilpi (2012), Deichmann et al (2008)). When these areas happen to have higher rain- fall as well, then we may observe a positive correlation between rainfall and share of non-farm employment in an upazila which has little to do with agricultural productivity. To address this concern, we take two steps. First, all upazilas located in the two metropolitan areas (Dhaka and Chittagong) are excluded from the sample, Second, we provide strong evidence that in our data rice yield is positively affected by rainfall. In addition, we include travel time from the upazila center to the nearest of the two main metropolitan cities (Dhaka and Chittagong) as a control in the regressions. Travel distance to the nearest metropoli- tan city also serves a second purpose; it controls for heterogeneity in the price of T - good, faced by a metropolitan consumer (denoted as PT (1 + d) in the model). A second concern is that some non-farm activities may be susceptible to rainfall directly. For instance, construction employment may rise with higher rainfall if rainfall leads to flooding and destruction of housing and infrastructure. Flooding and destruction may also lead to a negative correlation between rainfall and non-farm employment growth if it disrupts production activities of firms. We include a dummy for floodplain in the regressions to control for such possible adverse effects. An important advantage of our empirical approach is that we can use upazila fixed effects to purge off the additive effects of all time-invariant confounding variables. Since both travel 16 time and the floodplain dummy are time-invariant, their additive effects are taken care of by upazila fixed effects. We take a conservative approach, and allow for time varying effects of these variables, and include interactions of these variables with the time trend in the empirical model. As additional control for time varying heterogeneity across upazilas, we include proportion of household with electricity as a control vari- able. To capture changes in labor endowment, we control for upazila population, total active labor force, and proportion of active labor force with secondary or above education (human capital). This makes the empirical model consistent with the theoretical assumption of a fixed labor endowment (i.e., insignificant labor supply response to productivity shock in agriculture), and enables us to focus on structural change in employment through labor reallocation.12 To account for any agglomeration externalities that may affect firm location, we control for upazila population in 1991 interacted with time trend (initial condition) and the share of urban households in total households in the upazila as an indicator of agglomeration economy. 13 A final issue for the empirical specification is that rainfall is expected to have a significant effect on rice yield only if a upazila has significant agricultural production (i.e., A is not too small in our model). Since we exclude upazilas located in two main metropolitan areas from our sample, this is not an issue with our samples. An important insight from the theoretical analysis is that the effects of rainfall on wage is likely to be stronger when the share of agriculture is higher in the upazila economy. Accordingly we carry out empirical analysis for two different samples: a “rural sample” and a “small town sample” which are defined in the data section. 4. Data To test the hypotheses regarding the effects of agricultural productivity growth on wage and the struc- ture of employment, we combine three different data sources to construct an upazila (subdistrict) level panel data set covering the period 2000 to 2009/2010. A primary source of employment data in Bangladesh is the population censuses (1990, 2000 and 2010). 12 In an economy with a growing labor force, a second source of structural change is allocation of new labor to the nonfarm activities, which is not the focus of this paper. 13 Our empirical results are robust is we instead use finer controls of agglomeration economies such as area share in total industry employment in 1991. 17 The census unit records are publicly available for 10 percent of the population in 1990 and 2000, and for 5 percent in 2010.14 The unit record data contain employment records of all household members.15 We define total employment level in each activity from these unit records using appropriate weights. The total employment defined from census data includes employment in both large and small firms, with the large firms in manufacturing and services producing the T -good and the small firms producing the S −good in our model. The population census contains a aggregated industrial classification of employment (top level). Specifically, the 2010 census distinguishes only between agriculture, manufacturing and services activities, and large and small firms cannot be distinguished. We thus use these data to understand the effects of agricultural productivity on all firms (farms) without any distinction between large and small firms. Data on employment in large-scale formal non-farm activities (corresponding to T -good in the model) are drawn from three rounds of economic censuses (2000, 2006 and 2009). The economic census provides a detailed list of all firms engaged in industrial and services activities (at 4-digit level) with number of employees more than 10. These relatively larger firms are mostly registered, and tend to produce better quality manufacturing and services to meet demand in larger markets. The larger is the firm size, the greater is the possibility that it produces goods and services for a wider geographical area. These data are useful in two ways. First, we use them to estimate the effects of agricultural productivity on large-scale manufacturing and services. A comparison of the estimates on all firms (farms) from the population census with the estimates for large firms from the economic census helps us to analyze possible differential effects across firm size. Given that these data are available at the 4-digit level, we rely on them to perform disaggregated analysis of employment (for example, in construction, education, health etc). The productivity growth in agriculture is measured by growth in crop yields. Rice (paddy) is the predominant crop in Bangladesh, of which three different types (Boro, Aman and Aus) are grown.16 The official source of agricultural statistics provides yield data at the district level, but the data are at aggregate 14 The sample sizes are 10.6, 12.4 and 7.2 million individuals in 1990, 2000 and 2010 censuses respectively. 15 Total numbers of employment records in census data are 3.1, 3.5 and 2.1 million for 1990, 2000 and 2010 censuses respectively. 16 High yielding variety of Boro rice now accounts for more than half of rice production (56%). Aman is the next important crop accounting for 44% of rice production. Yields of both of these varieties are much higher than Aus. 18 level with limited coverage.17 Another source of yield data is the community part of the Household Income and Expenditure Surveys (HIES) (2000, 2005, 2010). We define rice yield per acre as the average of yields of boro, aman and aus rice reported in the village/community part of HIES. The upazila (subdistrict) level rice yield is the average over villages surveyed within an upazila. Since the number of villages within an upazila is not large, these upazila level estimates of rice yield may involve significant measurement error. However, it is reassuring that the yield growth from the HIES data shows estimates comparable to that from the official estimates during the decade of 2000. The rainfall data are taken from Bandyopadhyay and Skoufias (2012). The original data on rainfall come from the Climate Research Unit (CRU) of the University of East Anglia. The CRU reports estimated monthly rainfall for most of the world at the half degree resolution from 1902 to 2009. The CRU method combines weather station data with other information to arrive at the estimates.18 To estimate the sub- district (upazila/thana) level rainfall from the CRU data, Bandyopadhyay and Skoufias (2012) uses area weighted averages.19 Travel times to metropolitan cities were computed using GIS software and road network from the mid-1990s. Data on agro-ecological zones are drawn from the Bangladesh Water Board database. Over the years, a number of larger upazilas were split to form new upazilas, thus increasing the total number of upazilas from 486 in 1990 to 507 in 2000 to 543 in 2010. We use upazila maps to identify the borders of upazilas over time and matched all upazilas in 2000 and 2010 to 1990 upazilas. The upazila level panel is defined using 1990 upazila boundaries. To define the estimation sample, we take out the upazilas that are located in the two main metropolitan cities (Dhaka and Chittagong), leaving us with a sample of 464 upazilas. We divide the upazilas into two sub-samples on the basis of share of urban in total 17 These data are reported at old (and much larger) district level – there are about 20 old districts. With newly created districts, there are now 64 districts in Bangladesh. These data are drawn from Statistical Yearbooks. 18 Previous versions of the CRU data were homogenized to reduce variability and provide more accurate estimation of mean rain at the cost of variability estimation. The version 3.1 data are not homogenized and thus allows for better variability estimates. The estimates of rainfall near international boundaries are not less reliable as compared with those in the interior of the country, as the CRU estimation utilizes data from all the weather stations in the region. 19 For example if an upazila/thana covers two half degree grid cells for which CRU has rainfall estimates, then upzila/thana rainfall is estimated as the average rainfall of the two grid-cells, where the weights are the proportion of the area of the upazila/thana in each grid-cell. For details, please see Bandyopadhyay and Skoufias(2012). 19 upazila population in 1991. The median share of urban population in upazila population was 11 percent in 1991. Our “rural sample” consists of 232 upazilas which had less than 11 percent of population classified as urban. The STE sample comprises of the other 232 upazilas with urban share higher than 11 percent. Table 1 provides the summary statistics for all upazilas in our sample over the years. The popula- tion censuses indicate strong employment growth in the overall non-farm sector.20 Total employment in non-farm activities grew at an annual rate of 8.6 percent between 2000 and 2010. Total manufacturing employment posted an annual growth of 18.5 percent during this period. Services employment grew by 6.8 percent. Between 2000 and 2009, employment in firms with more than 10 employees grew at an annual rate of 5.7 percent, according to the economic censuses. Among the large firms, growth in manufacturing employment has been quite robust at 5.1 percent compared with a 2 percent growth in services employment. A comparison of employment composition shows that employment in small firms (10 or fewer employees) accounts for more than three-quarters of total employment in non-farm sector. Interestingly, the share of employment in larger firms in total non-farm employment declined from 0.22 in 2000 to 0.17 in 2010. While about 25 percent of all employment is in manufacturing in 2010, the share of manufacturing in total employment of large firms is much higher (43 percent in 2009). This is expected as few services enterprises are large in size. Larger services enterprises appear to be more concentrated in the provision of health (hospitals, clinics etc) and education (schools) (Table 1). A comparison of summary statistics between our rural and STE samples confirms that most larger firms are located in small towns. These firms tend to serve larger markets compared with the firms in rural sample which are smaller and serve mainly local markets. The summary statistics in Table 1 indicate substantial growth in rice yield between 2000 and 2010. Average rice yield per acre has grown by an annual rate of 3.8 percent. This growth rate is consistent with about 3.7 percent growth in agricultural GDP during the same time.21 There has been substantial 20 According to the census data, the share of non-farm employment in total employment increased from 0.475 in 2000 to 0.53 percent in 2010. The total labor force during this period increased by an annual rate of 2.05 percent from 35 million to 42.5 million. 21 Crop agriculture accounts for 56 percent of agricultural GDP and rice is the single most important crop in Bangladesh not only as a subsistence but also as a cash crop. 20 expansion of irrigation during the decade as well -from 61 percent in 2000 to 69 percent in 2010. The standard deviation estimate (Table 1) shows that there are considerable variations in rice yields across upazilas. Among other variables, access to electricity by households improved considerably during the decade (6.5 percent annual growth rate). 5. Empirical Results In this section, the main empirical results are presented sequentially. All outcome variables as well as rainfall are expressed in logarithms. All regressions include upazila and year fixed effects. All standard errors are corrected for correlation in the error term within the upazila. 5.1 Rainfall, Agricultural Productivity, and Agricultural Wage We begin by presenting evidence on the impact of rainfall shock on agricultural productivity. Table 2 reports the results from OLS regressions where log of rice yield is regressed on deviation of rainfall after controlling for upazila and year fixed effects. Column (1) shows the estimate when no other explanatory variable is included in the regression. The specification in column (2) includes the full set of upazila level controls as discussed in the empirical strategy section (Section 3). The first panel reports results for the STE sample and the second for the rural sample. The evidence from both specifications shows statistically and numerically significant impact of rainfall shock on rice yield in the rural and STE samples. The estimated coefficients imply an increase in the yield growth due to a positive shock in rainfall over its mean, with the impact marginally larger in the rural sample compared with the STEs. This result is consistent with the findings from a rich body of evidence accumulated by the agronomists and crop scientists that shows that rainfall is a major determinant of rice yield growth in Bangladesh over the last few decades (see, for example, Sarkar et. al. (2012)). While a positive rainfall shock increases rice yield, does it increase the wage rate? We utilize HIES data on agricultural wage to examine the effects of rainfall shocks. Note that the sample size for the wage regression shown in column (3) of Table 2 is somewhat smaller due to the fact that HIES – though 21 nationally representative – does not cover all of the upazilas in our upazila level panel data sample used for columns (1) and (2). The results in column (3) show a statistically significant and positive impact of rainfall on agricultural wage in both samples. Interestingly, the wage response is larger in magnitude in the rural sample, consistent with the theoretical prediction that the wage response to productivity shocks, under plausible conditions, is a positive function of the endowment of agricultural land (i.e., a higher A) (proposition 1). This evidence is reassuring as it confirms that the central channel through which agricultural productivity works in the theoretical model is in fact operative in the data. The effects of rainfall on agricultural income cannot be estimated directly, as reliable data on income are unavailable. We can, however, estimate the elasticity of wage with respect to productivity (rice yield) and exploit the theoretical insights to understand if the data supports a positive income effect (a higher Y v in the model). Using the reduced form estimates from Table 2 (columns (1) and (4)), the implied wage elasticities with respect to productivity are: Ewθ = 1.47 in small town sample, and Ewθ = 1.82 in the rural sample. Although we do not have a precise estimate of α , the available evidence in the context of Bangladesh suggests that it is likely to belong to the interval α ∈ [0.60, 0.70] (Hossain and Bayes (2009)). −1 −1 Thus (1 − α) ∈ [2.5, 3.33], satisfying the condition specified in proposition 1 [Ewθ < (1 − α) ]. Even if we assume a conservative estimate of 2.33, the estimates for Ewθ in both the rural and STE samples imply that a positive productivity shock to rice yield would increase rural income, providing a basis for structural change in employment driven by the Engel curve effect. While a positive rainfall shock increases rice yield, for appropriate interpretation of the results, it is useful to understand whether this reflects only the impact of transitory weather shocks on farming. Though the rainfall variations across upazilas and over time are expected to affect the rice yield directly, they are likely to affect long-term productivity differences by influencing investment in irrigation. In a cross-section regression, irrigation investment is negatively correlated with rainfall, as they are substitutes.22 But over time, in the same upazila, irrigation expansion is positively correlated with rainfall, as irrigation investment 22 It is, however, important to appreciate that rainfall affects productivity even with irrigation. This is because of the simple fact that the marginal cost of irrigation water is positive, but it is zero for rainfall. 22 starts in the drought prone areas and expands to the areas with more rainfall over time. Since we use a panel data model with upazila fixed effect, the rainfall shocks in our case would thus pick up irrigation expansion over time. The final column reports the estimated effect of rainfall variations on the area irrigated in a specification with upazila fixed effects and other controls used in our main regressions. Though the coefficient estimates lack statistical precision, they are positive in both samples. The estimated coefficient is large in magnitude in the rural sample. A positive coefficient on the rainfall variable in this regression provides suggestive evidence that the irrigation expansion over our sample period has happened increasingly in areas with relatively higher rainfall. We interpret this as evidence that though rainfall variable in our panel regressions captures mostly transitory shock in agriculture, it may also be capturing to some extent the diffusion of modern technology in farming over time. Note also that modern farming technology such as irrigation may reduce risk by decreasing variability of yield even without increasing yields. The expansion of irrigation in Bangladesh allowed adoption of boro rice whose yields are significantly higher than other rice types (aman and aus). Another issue in the IV interpretation of rainfall is that it may capture not only agricultural productivity shocks but also the resulting price changes. In a completely segmented rice market at the upazila level, a rainfall shock would affect the equilibrium rice price through the income effect. However the rice market is the most developed and spatially integrated market in Bangladesh (see, for example, Hossain and Verbeke (2010)). A noted before, we control for the distance from an upazila to the metropolitan city markets, which captures spatial price dispersion due to transport costs. The theoretical model assumes that the rice price is pinned down by the international market, and the available evidence on rice markets in Bangladesh clearly supports this assumption. 5.2 Rainfall, Agricultural Productivity and Structural Change in Employment With the evidence that a positive rainfall shock increases agricultural productivity and wage leading to positive income effect, we now turn to an analysis of the effects of rainfall shocks on intersectoral employment. The dependent variables in the employment regressions are all expressed in logarithms. All 23 regressions include the full set of regressors and upazila and year fixed effects. The first three columns of Table 3 report the fixed effects regression results for employment by all firms (using population census data) and the last three for employment by the large firms (using data from economic censuses). In order to facilitate comparison between the economic census data on large firms and the population census data on all firms, the employment categories are kept at a fairly aggregate level, distinguishing only between manufacturing and services.23 The top panel reports the results for the STE sample and the bottom for the rural sample. The number of periods in the census panel data is two and economic census is three. This is because 1990 census data are used to define the initial characteristics of upazilas, and the economic census has an additional data point for 2006 which is a non-population census year. Evidence on All Firms (Irrespective of Size) The evidence in the first three columns on all firms shows an interesting pattern: the estimated effect of a positive rainfall shock is positive and statistically significant on the manufacturing and services sectors, while the effect is negative on agricultural employment, although it is not statistically significant at the 10 percent level. This pattern of estimates is observed in both the rural and the STE samples, with the magnitude of the effect larger for manufacturing and smaller for services in the case of the rural sample. The negative effect on agriculture, although imprecisely estimated, suggests labor reallocation away from ¯ , a higher agricultural farming. Note that according to the theoretical analysis, given a labor endowment L productivity results in labor reallocation away from agriculture when Ewθ > 1. Given that the estimates of Ewθ in the rural and STE samples are 1.82 and 1.47 respectively, the theory predicts a more robust negative effect on agricultural employment. One possible explanation is that the controls we use for changes on labor endowment over time are less than perfect, and thus the negative effect on agricultural employment is diluted by a growing labor force. The evidence from the population census data suggests that agricultural productivity growth leads to 23 Census data do not provide employment information at disaggregate level for 2010. 24 structural change in employment in favor of the manufacturing and services sectors, but we do not know if this structural change also entails employment growth in large (formal) enterprises. To understand whether agricultural productivity increase leads to labor reallocation in favor of large-scale (formal) firms, we take advantage of the economic census data. Evidence on Large Manufacturing and Services Firms (Economic Censuses Data) The estimates for manufacturing employment in large firms are reported in column (5) of Table 3, and they suggest a statistically significant positive effect of rainfall in the STE sample, but no effects in the rural sample. The fact that there is no significant effect on manufacturing in the rural sample is especially striking when compared to the estimate for all firms in column (1) where the effect is both statistically significant at the 1 percent level and numerically large. Taken together, the estimates in columns (1) and (5) suggest that agricultural productivity growth in predominantly rural upazilas reallocate labor from agriculture to small-scale manufacturing firms producing locally consumed low-quality S −good. Given the fact that the wage response to agricultural productivity growth is larger in magnitude in the rural sample, the results on the rural upazilas are consistent with a scenario where the demand for manufacturing S −goods is much stronger than that for manufacturing T −goods (i.e., υs > υT ), and the weak demand effect is more than offset by a strong wage response to productivity growth in the case of T −goods sector. In contrast, in the small town sample, a higher agricultural productivity results in a very strong growth in employment in the large-scale manufacturing. A comparison of the estimated effects of rainfall shock on all manufacturing firms and the large manufacturing firms in the STE sample shows that the employment growth in large firms is more than four times of that in all firms (i.e., large and small firms combined together). This implies that the employment growth in small firms in STE sample is weak at best. The contrasting effects on large-scale manufacturing in the rural and STE samples suggest that when the wage response is high enough, it may discourage large firms in locating production in an upazila, consistent with the findings of Rosenzweig and Foster (2004) in the context of India. The negative wage effect can dominate the choices of large manufacturing firms when the T −good is effectively a tradable good, but the 25 evidence that in the STE sample employment response in T −good (manufacturing) sector is strong rejects the assumption of a tradable T − good. For services employment in large firms, the estimated coefficient is positive and statistically significant in both the rural and STE samples, but the magnitude of the effect in the STE sample is almost twice that in the rural sample. Thus the effects of agricultural productivity growth seem to be especially potent in those upazilas where the small towns are located. In the rural sample, the estimate for large-scale (formal) services is in sharp contrast to that for large-scale manufacturing, when a comparison is made with the corresponding estimates for all firms in columns (1) and (2) using population census data. The estimate for employment in large-scale services firms in the rural sample is much larger in magnitude (1.50) compared to the estimate for all firms in the services sector in column (2), while the relative magnitudes are reversed in the case of manufacturing. The evidence shows that the growth in large-scale firms in the services sector is favorably affected by higher agricultural productivity irrespective of whether there is a small town in the upazila. This probably reflects the fact that formal health and education services are effectively nontradables and also enjoy more income elastic demand. We explore the effects on large-scale health and education services in greater depth below. Our results are partly consistent with findings from Foster and Rosenzweig (2004) for India. Using village level data, Foster and Rosenzweig (2004) find that agricultural productivity growth has a negative effect on the number of factory workers and a positive effect on the employment in services. Our evidence that there is no significant effect of agricultural productivity growth on large manufacturing employment in the rural sample, but the employment in large-scale services increases substantially is consistent with their findings. However, using the STE sample, we find that agricultural productivity growth can have a strong positive effect on the large-scale manufacturing when the villages are located in an upazila with a small town. Thus the conclusions depend critically on the sample used. In a related contribution, using district and state level data from India, Adhvaryu, Chari and Sharma (2013) on the other hand find factory employment to be affected negatively by a negative productivity (rainfall) shock which is 26 consistent with our evidence from the STE sample. The results in Table 3 thus highlight the need for making a distinction between large and small firms, and the importance of focusing on geographical areas beyond the villages. Indeed ignoring the small towns, as is done in most of the current literature, would lead to a gross underestimation of the positive effects of agricultural productivity growth on structural change in employment. 5.3 The Effects on Employment in Large Firms: Disaggregated Analysis The results discussed above imply that there are important differences in the way agricultural pro- ductivity shocks affect different types of economic activities located in different areas. In this section, we explore whether these differences are also observed at a more disaggregated sectoral levels of employment of larger firms for which we have data from the economic censuses. Within the manufacturing sector, we make a distinction between food processing and beverages, and other manufacturing. Among services activities, we distinguish between trade, education, health and other services. We also examine employ- ment in transport, construction and utilities as a separate category. The fixed effects regression results are reported in Table 4. As before, all regressions include the full set of controls along with upazila and time fixed effects. The results in columns 1 and 2 in Table 4 indicate significant and positive effects of rainfall shocks on both food precessing and other types of manufacturing employment in STEs. The effects on all different types of services are positive and statistically significant at the 5 percent level in the STE sample. In terms of magnitudes, the largest coefficient estimate is found for transport, construction and utilities, followed by education, other manufacturing, other services and health. The smallest impact is observed for the trading services.24 The estimates from the rural sample in general show a much weaker impact; only in two out of seven types of activities listed in Table 4, the estimated effect is statistically significant at the 10 percent level. The two exceptions are education, and transport and construction; but even in these cases the magnitudes of the effects are much smaller in the rural sample when compared to the corresponding 24 The weak impact on trading may in part be due to the fact that very few trading enterprises are large in size. 27 estimates for the STE sample. An important finding from the estimates in Tables 3 and 4 is that an increased agricultural productivity and the resulting higher rural income leads to employment transformation within the services sector in both the rural and STE samples: households appear to substitute away from traditional services provided by smaller enterprises, and into services provided by larger and more skilled enterprises, especially in education. Similar transformative effects are found for large-scale manufacturing located in the STEs. The effects on manufacturing in the rural areas are very different; a positive agricultural productivity shock results in a robust growth of small-scale manufacturing employment with no impacts on employment of relatively large-scale enterprises. 6. Conclusions This paper provides a theoretical and empirical analysis of the role of agricultural productivity in non- farm employment growth, and structural transformation in the labor market, using a broader framework that goes beyond rural-urban dualism, and focuses on the rural towns and smaller cities which we term as small-town economy (STE). We develop a simple model that formalizes the demand and labor market linkages between the STEs and the surrounding villages. A positive productivity shock to agriculture does not necessarily increase village income; given a labor endowment, the village income increases only when the response of wage with respect to productivity shocks (i.e, wage elasticity) is lower than a threshold.25 When the elasticity of demand for the town good in the village is low so that most of the town good is sold outside the local economy, the labor market linkage predominates and a higher rural wage due to agricultural productivity increase reduces employment in small towns. In contrast, when demand for goods and services produced in small towns comes mainly from the adjacent rural areas, the demand effect can more than offset the negative wage effect, and lead to higher labor allocation to the production of town good. Another key insight from the model is that when a significant part of village income is spent on goods and services in small towns, ignoring them will lead to serious underestimation of the employment 25 In a more general model with a positive labor supply curve at the household level, a positive shock to agriculture is expected to increase village income in general. 28 reallocation effects of agricultural productivity shocks. Following a large literature on the importance of rainfall shocks in agricultural productivity variations in Bangladesh, we exploit rainfall shocks (relative to the 30-year average level) across upazilas and over time to understand the effects of productivity increase on employment growth and diversification. We use a two-step empirical approach that focuses on the reduced form regressions of rainfall on the measure of productivity (rice yield per acre) and on the set of outcome variables noted above. The evidence from the reduced form regressions shows that a positive rainfall shock increases employment in all different types of non-farm activities in the small towns, but increases only services activities and small-scale manufacturing in rural areas. The largest impacts are found for construction and transport, education, health and other services, and manufacturing employment in larger scale enterprises located in the small towns, and for educational services provided by larger enterprises in the rural areas. When interpreted as instrumental variables estimates of the effects of productivity increase, the em- pirical results suggest that agricultural productivity induces structural transformation within the services sector with employment in skilled services provided by larger firms growing at a much faster pace compared with the services provided by smaller firms/individuals. We find much larger positive effects on non-farm employment growth and transformation in the small towns. In the rural areas, it increases employment mostly in small-scale manufacturing and skilled educational services. The estimates confirm the transfor- mative effect of agricultural productivity increase on nonfarm activities in the small towns. These results suggest that policies promoting agriculture would be beneficial to employment growth and transformation in the small towns in Bangladesh. References Adhvaryu, A., A. V. Chari and S. Sharma. 2013, “Firing Costs and Flexibility: Evidence from Firms’ Employment Responses to Shocks in India,” Review of Economics and Statistics, vol. 95, no.3. Bandyopadhyay, Sushenjit & Skoufias, Emmanuel, 2012. Rainfall variability, occupational choice, and 29 welfare in rural Bangladesh,Policy Research Working Paper Series 6134, The World Bank. Bhowmik, A, K, and Ana Cristina Costa (2012). A Geostatistical Approach to the Seasonal Pre- cipitation Effect on Boro Rice Production in Bangladesh, International Journal of Geosciences, 2012, 3, 443-462 uckner, Markus, 2012. Economic growth, size of the agricultural sector, and urbanization in Africa, Br¨ Journal of Urban Economics, vol. 71(1), pages 26-36. Bustos, Paula, Bruno Caprettini and Jacopo Ponticelli, 2016. ”Agricultural Productivity and Structural Transformation: Evidence from Brazil,” American Economic Review, American Economic Association, vol. 106(6), pages 1320-65, June. Chernozhukov, V and C, Hansen (2008), “The Reduced Form: A Simple Approach to Inference with Weak Instruments”, Economics Letters, 2008, 100(1), 68-71. Christiaensen, Luc & De Weerdt, Joachim & Todo, Yasuyuki, 2013.Urbanization and poverty reduction – the role of rural diversification and secondary towns, Policy Research Working Paper Series 6422, The World Bank. Deichmann, U., F. Shilpi, and R. Vakis. 2009. “Urban Proximity, Agricultural Potential and Rural Non-farm Employment: Evidence from Bangladesh,” World Development, 37(3), p. 645-660. Emran, M. Shahe and Forhad Shilpi, 2012.”The extent of the market and stages of agricultural special- ization,” Canadian Journal of Economics, Canadian Economics Association, vol. 45(3), pages 1125-1153, August. Fafchamps, M. and F. Shilpi, 2003, “Spatial Division of Labor in Nepal, Journal of Development Studies, Vol. 39, No. 3, pp. 23-66. Fafchamps, Marcel, and Forhad Shilpi, 2005. “Cities and Specialization: Evidence from South Asia.” Economic Journal,Vol. 115, April. Ferre, Celine, Francisco H.G. Ferreira and Peter Lanjouw, 2010.”Is there a metropolitan bias? The inverse relationship between poverty and city size in selected developing countries” Policy Research Working 30 Paper Series 5508, The World Bank. Foster, A. and M. Rosenzweig, 2004, “Agricultural development, industrialization and rural inequality, mimeographed, Harvard University. Haggblade, S., P. Hazell and T. Reardon, 2006, Transforming the Rural Nonfarm Economy, Johns Hopkins Press, Baltimore. Harris, John and Michael Todaro (1970). Migration, Unemployment, and Development: A Two Sector Analysis, American Economic Review, 40: 126-142. Hossain, M. and A. Bayes, 2009, Rural Economy and Livelihoods: Insights from Bangladesh, AH Devel- opment Publishing House. Hossain, Mohammad Ismail, and Wim Verbeke. 2010. “Evaluation of Rice Markets Integration in Bangladesh.” Lahore Journal of Economics 15 (2): 77–96. Hymer, Stephen and Stephen Resnick (1969), “A Model of an Agrarian Economy with Nonagricultural Activities”, American Economic Review, Vol. 59, No. 4, Part 1 (Sep., 1969), pp. 493-506 Kuznets, S. (1973) ‘Modern economic growth: findings and reflections’ (Nobel Lecture), American Economic Review 63, 247–58. Lanjouw, Peter and Rinku Murgai, 2009.Poverty decline, agricultural wages, and nonfarm employment in rural India: 1983-2004, Agricultural Economics, vol. 40(2), pages 243-263, 03. Lanjouw J. and P. Lanjouw, 2001, “Rural Non-farm Employment: Issues and Evidence from Developing Countries,” Agricultural Economics, vol. 26, no. 1, pp.1 –24. Lewis, A. (1954) ‘Economic development with unlimited supplies of labor,’ Manchester School 22, 139– 91 Mellor, J, 1976, The New Economics of Growth: A Strategy for India and the Developing World, Cornell University press, Ithaca. Ranis, G. and F. Stewart, 1973. “Rural Non-Agricultural Activities in Development: Theory and Application.” Journal of Development Economics, vol. 33, o.7 pp. 1045-1060. 31 Renkow, M., 2006, “Cities, towns and the rural nonfarm economy, in Haggblade, S., P. Hazell and T. Reardon (eds), Transforming the Rural Nonfarm Economy, Baltimore: Johns Hopkins University Press. Rodriguez-Clare, Andres, 1996.The division of labor and economic development, Journal of Develop- ment Economics, Elsevier, vol. 49(1), pages 3-32, April. Sarker, M. A. R., Alam, K., and Gow, J. (2012). Exploring the relationship between climate change and rice yield in Bangladesh: An analysis of time series data. Agricultural System, pp. 11-16. Online Appendix: Not for Publication PROOF OF PROPOSITION 1 (1.1) The labor market equilibrium in equation (8) in the text can simplified as: ( ) φY o wL = ∆5 w1− α θ α − ∆4 ∆1 (1 − α) γf + ∆3 1 1 (13) 1+d 1 where ∆4 = υT ∆3 + υs and ∆5 = A(1 − α) α (1 + ∆4 ∆1 ). The effect of agricultural productivity shock on wage cane be derived from equation (13) as: 1−a ∂w α ∆5 θ α w = 1 1 > 0 (14) ∂θ αw α L + (1 − α) ∆5 θ α ∂2w (1.2) We show that the stated restrictions on land/labor ratio and Ewθ are sufficient for > 0. ∂θ∂A Using equation (14), we get: [ { ¯ }] ∂2w w 1 1 L = (θ (1 − α)) (1 + ∆4 ∆1 ) {α − Ewθ (1 − α)} + EwA α (θ (1 − α)) (1 + ∆4 ∆1 ) − Ewθ α α ∂θ∂A Dθ A (15) where D is the denominator in the RHS of equation (14), and EwA is the elasticity of wage with respect 32 −1 to land endowment A. Now, {α − Ewθ (1 − α)} > 0 if Ewθ < α (1 − α) , and the first term in [.] in the RHS of equation (15) is positive. It is easy to check from labor market equilibrium condition (13) that EwA > 0. A sufficient condition for the second term in the RHS of equation (15) to be positive when we −1 set E wθ = α (1 − α) is that the labor/land ratio is small enough to satisfy the following condition: ¯ L 1+α 1 < (1 − α) α θ α (1 + ∆4 ∆1) (16) A So when inequality (16) is satisfied, it is over-sufficient for the RHS of equation (15) to be positive that −1 Ewθ < α (1 − α) . (1.3) From equation (4) in the text, the total village income is given by: υs γ f Y v = ∆1 wL∗ f − (1 − υs ) ( ) ∂ wL∗ f Thus a necessary and sufficient condition for Y v to increase with a higher θis that > 0. ∂θ Using the solution for L∗ f from equation (3) in the text, we have the following: ( ) ∂ wL∗ f 1 [Aθ (1 − α)] α α−1 = w α {1 − (1 − α) Ewθ } (17) ∂θ αθ ( ) ∂ wL∗ f −1 From equation (17), > 0 iff Ewθ < (1 − α) . ∂θ PROOF OF PROPOSITION 2 (2.1) From equation (3) in the text, we have: ∂L∗ 1 f A [θ (1 − α)] α = 1 {1 − Ewθ } (18) ∂θ αθw α ∂L∗f From equation (18), it is obvious that < 0 iff Ewθ > 1. ∂θ 33 (2.2) See Pp. 12-13 in the text. (2.3) Using equation (6) in the text, we have: [ 1−α ] ∂L∗ 1 ∆1 υs (1 − α) Aθ α (1 − α) α S = γf Ewθ + 1−α (1 − Ewθ ) (19) ∂θ θw αw α ∂L∗S It is obvious from equation (19) that > 0 when Ewθ = 1. By continuity, there exists a threshold ∂θ S ∂L∗S value Ewθ > 1 such that > 0 remains true. ∂θ (2.4) Using equation (7) in the text, we get: [ 1−α ] [ ] ∂L∗ 1 ∆3 ∆1 υT (1 − α) Aθ α (1 − α) α ∆3 Ewθ φY o T = 1−α (1 − Ewθ ) + ∆1 (1 − α) υT γf − (20) ∂θ θw αw α wθ (1 + d) A comparison of equations (19) and (20) makes it clear that as long as φ > 0, the effect of agricultural productivity growth on employment growth in the town goods sector is likely to be limited compared to that in the S -goods sector. If the subsistence requirement γf is large enough to satisfy γf > γf ≡ φY o , then the last term in equation (20) is positive and there exists a threshold value ∆1 (1 − α) υT (1 + d) T Ewθ > 1 such the RHS of equation (2) is positive. The stronger is the consumer preference for town good in the village (i.e., higher value of υT ), and the farther away is the metropolitan city (i.e., a higher d) , the T larger is the value of Ewθ , and the more likely it is that agricultural productivity has a positive effect on the employment in T -goods sector. Table 1: Summary Statistics: Sub-district level Employment, Yield and other indicators (Full Sample) Mean SD Mean SD Mean SD Economic Census 2000 2006 2009 Employment (Firm size>=10) 3128 4161 4256 6537 5158 9487 Manufacturing 1419 2925 2117 5806 2225 6256 Services 1062 1142 1319 1044 1267 991 of which Trade 64 178 47 127 40 110 Education 692 672 978 684 951 658 Health 151 208 169 196 162 183 Other (construction, transport, utility etc.) 1701 2052 2132 2010 2926 5343 Population Census 2000 2010 Employment (Total) 14054 4161 32214 31322 Manufacturing 1458 10871 7985 12043 Services 12595 2559 24229 22137 Proportion urban 0.13 0.13 0.14 0.11 Population 213274 108269 276587 153595 Proportion with Secondary or higher Education 0.11 0.04 0.13 0.04 Proportion of households with electricity 0.25 0.23 0.36 0.25 0.46 0.27 Household Income and Expenditure Survey 2000 2005 2010 Rice Yield (mt/acre) 0.94 0.11 1.01 0.13 1.36 0.48 % of land irrigated 61 30 60 33 69 26 Real Per Capita Expenditure 749 200 848 206 1098 266 Real Daily Wage 74 29 76 39 94 31 Annual Rainfall (mm) 1444 487 1653 410 1472 379 Table 2: Rainfall Shocks and Agricultural Yields Log (Rice Yield/acre) Log(agricultural % of Area wage) Irrigated (1) (2) (4) (5) Small Cities and Towns Sample Rainfall Deviation 0.387*** 0.448*** 0.567** 14.53 (7.316) (8.801) (2.459) (0.731) Observations 696 696 458 355 Number of Upazila 232 232 208 180 Rural Sample Rainfall Deviation 0.449*** 0.498*** 0.821*** 27.94 (8.472) (9.845) (2.792) (1.154) Observations 696 696 436 386 Number of Upazila 232 232 205 195 Year & Upazila Fixed Effects Yes Yes Yes Yes Area Characteristics No Yes Yes Yes Note: Rainfall Deviation is defined as the difference between log(current rainfall) and log(30 year average rainfall). The regressions in even numbered columns include a number of controls (log of travel time to Capital and/or Port cities interacted with time trend, flood prone dummy interacted with time trend, proportion of household with electricity, log of 1991 population interacted with trend, log of total active labor force, and proportion of labor force with above secondary education, share of urban households. All regressions include year and upazila fixed effects. Standard errors are clustered at upazila level. Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 3: Rainfall Shocks, Agricultural Productivity and Non-farm Employment Log (Employment) All Firms Large Firms Manufac- Services Total Log(Agri. Manufac- Services Total turing Non-Farm Employm.) turing Non-farm (1) (2) (3) (4) (5) (6) (7) Small Towns Sample Rainfall Deviation 0.656** 0.310*** 0.385*** -0.126 2.680*** 2.875*** 3.113*** (2.298) (3.168) (4.979) (-1.064) (2.812) (3.080) (3.098) Observations 464 464 464 464 696 696 696 Number of Upazilas 232 232 232 232 232 232 232 Rural Sample Rainfall Deviation 1.373*** 0.246** 0.337*** -0.0444 0.309 1.501** 1.064 (4.087) (2.138) (3.135) (-0.868) (0.386) (2.016) (1.449) Observations 464 464 464 464 696 696 696 Number of Upazilas 232 232 232 232 232 232 232 Year & Upazila Fixed Effects Yes Yes Yes Yes Yes Yes Yes Area Characteristics Yes Yes Yes Yes Yes Yes Yes Note: Rainfall Deviation is defined as the difference between log(current rainfall) and log(30 year average rainfall). The regressions include a number of controls (log of travel time to Capital and/or Port cities interacted with time trend, flood prone dummy interacted with time trend, proportion of household with electricity, log of 1991 population interacted with trend, log of total active labor force, and proportion of labor force with above secondary education, share of urban households. All regressions include year and upazila fixed effects. Standard errors are clustered at upazila level. Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 4: Rainfall Shocks, Agricultural Productivity and Employment in large firms Log (employment in manufacturing) Log(employment in services) Food & Other Transport Other Const. & Beverages Manfact. other Trade Education Health Services (1) (2) (3) (4) (5) (6) (7) Small Cities and Towns Sample Rainfall Deviation 1.769** 2.306** 2.699*** 1.215* 2.436*** 2.071*** 2.189*** (2.151) (2.486) (3.043) (1.691) (3.150) (3.005) (3.048) Observations 696 696 696 696 696 696 696 Number of Upazilas 232 232 232 232 232 232 232 Rural Sample Rainfall Deviation -0.249 -0.108 1.382* 0.189 1.430** 0.758 0.590 (-0.362) (-0.133) (1.970) (0.394) (2.307) (1.224) (0.944) Observations 696 696 696 696 696 696 696 Number of Upazilas 232 232 232 232 232 232 232 Year & Upazila FE Yes Yes Yes Yes Yes Yes Yes Area Characteristics Yes Yes Yes Yes Yes Yes Yes Note: Rainfall Deviation is defined as the difference between log (current rainfall) and log(30 year average rainfall). The regressions include a number of controls (log of travel time to Capital and/or Port cities interacted with time trend, flood prone dummy interacted with time trend, proportion of household with electricity, log of 1991 population interacted with trend, log of total active labor force, and proportion of labor force with above secondary education, share of urban households. All regressions include year and upazila fixed effects. Standard errors are clustered at upazila level. Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1