The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Sectoral Development Multipliers WP 24-02 Francisco Buera Washington University in St. Louis Nicholas Trachter Federal Reserve Bank of Richmond Sectoral Development Multipliers∗ Francisco Buera Washington University in St. Louis Nicholas Trachter Federal Reserve Bank of Richmond March 13, 2024 Abstract How should industrial policies be directed to reduce distortions and foster economic development? We study this question in a multi-sector model with technology adoption, where the production of goods and modern technologies features rich network structures. We provide simple formulas for the sectoral policy multipliers, and provide insights regarding the power of alternative policy instruments. We devise a simple procedure to estimate the model parameters and the distribution of technologies across sectors, which we apply to Indian data. We find that technology adoption greatly amplifies the multipliers’ magnitudes, and it changes the ranking of priority sectors for industrial policy. Further, we find that adoption subsidies are the most cost-effective instrument for promoting economic development. ∗ Buera: fjbuera@wustl.edu. Trachter: trachter@gmail.com. We thank participants at many seminars and conferences for comments and suggestions. We also thank Samira Gholami, Lindsay Li, Filip Milosavljevic, and Matthew Murphy for outstanding research assistance. Francisco J. Buera acknowledges support from National Science Foundation Grant SES2049674. The views expressed herein are those of the authors and do not necessarily represent the views of the Federal Reserve Bank of Richmond or the Federal Reserve System. 1 1 Introduction Economic development hinges on the adoption of modern, complementary, technologies. Distortions can hinder this process, resulting in underdevelopment. Industrial policies targeting keys sectors are called for alleviating these distortions and promoting the investment in modern technologies (Hirschman, 1958). Recent developments formalize the relevance of sectoral distortions in economies with input-output linkages (Baqaee and Farhi, 2020), and how the propagation of these distortions through the economy motivates the case for policies at the sector level to enhance production efficiency (Liu, 2019). However, these studies abstract from technology adoption. We extend the analysis to allow for technology adoption, and develop a laboratory to answer three key questions: Which sectors should be targeted in order to promote economic development? What is the role technology adoption, relative to production efficiency, in driving the impact of industrial policies? Which policy instrument is the most effective? We propose a model of technology adoption in a multisector economy, where sectors are interconnected through rich production and technology adoption networks. We provide analytical expressions for sectoral development multipliers—the elasticity of aggregate consumption to a sectoral subsidy divided by the fiscal cost of running the policy. We consider three set of sector specific policies: revenue, factor, and adoption subsidies. Plant level and input-output data for India, together with information on the investment network of developed economies, is combined with structural estimation and calibration methods to quantify the multipliers and measure which sectors should be targeted to promote economic development. We find that the introduction of technology adoption greatly amplify the effect of desirable policies and changes our understanding of which sectors are the most relevant for industrial policy, and that adoption subsidies are the most effective policy instruments at promoting economic development. Monopolistically competitive establishments across different sectors produce differentiated intermediate inputs, charging a markup over marginal cost. Intermediate inputs are combined to produce the final consumption bundle, bundles used by other establishments as a production input, and bundles to produce adoption goods. Establishments can operate a traditional technology or pay for the adoption goods and modernize their production techniques. Sectors are interconnected through heterogeneous networks. Intermediate input bundles used by different sectors combine intermediate inputs in different ways—production network, and the adoption good used by different sectors also combine intermediate inputs in different ways—investment network.1 These networks propagate the markup distortion through the economy, altering the production efficiency across sectors and, importantly, the technology adoption incentives. Correcting these distortions provide the grounds for sectoral industrial policy. We perform a local theoretical analysis of the effect of alternative sectoral policy instruments. We provide a decomposition of the sectoral development multipliers into a production efficiency channel, the focus on earlier studies, and a productivity channel. The productivity channel fully encodes the technology adoption margin of the economy. This second channel features a double-Leontieff inverse. This double inverse 1 We focus on modern technologies as being capital embedded (Caunedo and Kala, 2021; Jones and Liu, 2022). Thus, we interpret the investment network as the relevant object for modern technology adoption. 2 captures the fact that in a model with technology adoption there are further rounds of complementarities. Policies that subsidizes a sector trickles down through production and investment networks as the cost of intermediate inputs and adoption goods decline. The lower costs promote technology adoption, resulting in increases in the productivity of individual sectors, triggering additional rounds of cost reductions and technology adoption. We apply the theory to quantify sectoral development multipliers for the India economy. We devise a simple yet powerful estimation procedure to estimate the model parameters. With the estimated model at hand, we study quantitatively the extent by which technology adoption changes our understanding of which sectors are most relevant. We find that the adoption margin is highly relevant, greatly amplifying the size of the multiplier in some sectors—by a factor of two in key sectors, and reducing the size of the multiplier in other sectors. Further, accounting for technology adoption changes the relevance of sectors as engines of development. Intuitively, investment goods producing sectors are featured among the top 5-sectors in terms of the development multipliers when the technology adoption margin is taken into account. Revenue-based policies are useful as promote economic development by directly affecting inefficiencies and motivating firms to adopt modern techniques. However, they are relative ineffective vehicle for economic development as they require a high cost to run. We show that adoption subsides, while more modest in terms of their direct effects, provide more ’bang for the buck’ in the sense that the adoption-based subsidy elasticity gains are much larger than the fiscal cost of running the industrial policy. Intermediate input subsidies, which directly address the distortions, are the second most cost effective instrument, while labor subsidies are the least effective. The effect of labor subsidies are crowd-out in general equilibrium as labor is a fixed factor. Related literature Understanding the relevance of sectoral shocks and distortions in setting with sectoral linkages has been the focus of several important recent contributions. Acemoglu et al. (2012) shows how productivity shocks to particular sectors can have aggregate effects. Baqaee and Farhi (2020), Bigio and La’O (2020) and Caliendo et al. (2022) study how distortions get amplified through sectoral linkages and provide measures for the aggregate losses. Baqaee and Farhi (2021) explores the role of the entry margin. We contribute to this literature by studying technology adoption and the role of alternative sectoral policies. The analysis of industrial sectoral policies is also the subject of an active recent literature. Liu (2019) studies how to target labor subsidies as a tool to alleviate sectoral distortions, and Liu and Ma (2021) studies how to target subsidies for R&D when the process of innovation is also subject to sectoral interactions, with a natural focus on developed economies. Bartelme et al. (2019) considers the case in which sectors are subject to external economies of scale, and use trade data to infer economies of scale.2 We contribute to this literature by adding a technology adoption margin, and studying how distortions interact with technology choice, quantifying the gains from targeted sectoral policies aimed at generating economic development through the adotion channel. In addition, we analyze alternative policy instruments. 2 Other recent papers in this literature include Choi and Levchenko (2021) and Farrokhi et al. (2023). 3 Our model builds on the work by Buera et al. (2021), which studies the non-linear effects of distortions and policies, and the possibility of multiple equilibria, in a roundabout economy with complementarities in technology adoption. We generalize their theoretical framework by considering a multisector economy with rich production and investment interactions across sectors and a more general structure of firm heterogeneity that is more amenable to infer the distribution of technologies at the sector level. More broadly, our paper relates to a recent literature exploring the quantitative relevance of complementarities and technology choice, such as Alvarez et al. (2023), Boehm and Oberfield (2023), Crouzet et al. (2023), and Demir et al. (2024). Finally, a recent literature studies the role of investment networks in the propagation of sectoral productivity shocks. Foerster et al. (2022) and vom Lehn and Winberry (2022) show how investment and production networks interact to amplify trend sectoral shocks in the postwar US economy, while Casal and Caunedo (2023) provides novel and harmonized investment network data across countries and show how they change over the process of development. In our work, the investment network plays a central role in propagating the effect of sectoral policies through technology adoption, as the investment network affects the adoption cost of capital embedded technologies. 2 Framework We consider a multisector economy composed of S sectors. Within each sector s, establishments produce differentiated varieties, used in the production of intermediate inputs and the final consumption aggregate. The economy is populated by a measure L of individuals, each of whom supply a unit of labor inelastically, and a measure 1 of potential entrants in each sector. At each sector s, an active establishment has the ability to produce a differentiated variety j. Production of the differentiated varieties yjs is aggregated to the sector level output Ys using a Dixit-Stiglitz aggregator, Z Ys = η−1 η η η−1 yjs dj , η>1, where η denotes the elasticity of substitution across the differentiated varieties produced within sector s. The output of all sectors is then combined to create aggregate consumption and the intermediate input aggregate used by each sector s, C= Y (Cs )γs , Xs = Y (Xss′ )ωss′ , s′ s with γs ∈ [0, 1], Γ = [γs ]S×1 denoting a column vector collecting all γs , with 11×S Γ = 1. Likewise, let Ω = [ωss′ ]S×S denote a matrix collecting all ωss′ . Potential entrants in sector s are endowed with productivity z, where F (z) denotes the cumulative distribution of z within each sector and f (z) ≡ ∂F (z)/∂z denotes its density. We further assume that F (z) is Pareto, with parameter ζ > 0. Establishments can be either inactive or active. Inactive establishments 4 do not operate and remain out of the market. Active establishments face an extra productivity shock eε , which is realized after the entry and adoption decisions, but before the input choices. We assume that ε is normally distributed with mean µ and standard deviation χ. Further, we set µ = −(η −1)χ2 /2, a convenient normalization.3 An active establishment in sector s, endowed with productivity duple {z, ε} and using technology i ∈ {t, m}, produces output ys by combining labor l and the intermediate aggregate x according to the following Cobb-Douglas production function yjs = zAi eε xν l1−ν ν ν (1 − ν)(1−ν) , ν ∈ [0, 1] , where ν denotes the intermediate aggregate input elasticity, and where Ai is a technology shifter specific to i, with At < Am . We refer to i = t as the traditional technology and to i = m as the modern technology. An active establishment in sector s operating the traditional technology must incur in an entry cost κts . An active establishment in sector s operating the modern technology must incur in the entry cost κts and modern technology adoption cost Pms κms , where Pms denotes the price of the intermediate aggregate used to produce the adoption good used by sector s. The adoption good in sector s is produced by a representative competitive establishment that combines the output of the different sectors using the technology flow matrix Λ = [λss′ ]. Likewise, the entry cost in sector s is also produced by a representative establishment, converting labor one-to-one into the entry cost. Finally, we consider a variety of different instruments aimed at improving industrial outcomes: sector revenue subsidies rs , sector labor cost subsidies rsl , sector intermediate input cost subsidies rsx , and sector technology adoption subsidies rsa . We stack all revenue subsidies in a column vector, r = [rs ]S×1 . We stack all other subsidies in similar vectors. The government levies lump-sum taxes from consumers and runs a balanced budget. Demand and Price Indexes. The economy has two relevant layers. In the first layer, the inner layer, the output of differentiated establishments is aggregated to produce sector output Ys . Let Ps denote the sector price index corresponding to this output. Straightforward calculations provide that the demand for the output of a establishment producing variety j is given by yjs = (pjs /Ps )−η Ys , and the price index R 1/(1−η) . In the second layer, the outer layer, the output of the different sectors satisfies Ps = ( p1−η js dj) is aggregated to produce the consumption bundle C and the intermediate aggregate of each sector Xs , s = 1, ..., S. As before, it is immediate to obtain that final demand for the output of sector s satisfies Q Cs = γs (Pc /Ps ) C, where the price index of aggregate consumption is defined as Pc = s (Ps )γs . Likewise, we can obtain the intermediate demand by sector s for the output of sector s′ , Xss′ = ωss′ (Pxs /Ps′ ) Xs , Q Q where Pxs = s′ (Ps′ )ωss′ and Pms = s′ (Ps′ )λss′ . 3 This normalization for the mean of the ex-post distribution implies that ex-post heterogeneity does not affect equilibrium aggregates in the economy. Under this normalization, Eε [e(η−1)ε ] = 1. Ex-post heterogeneneity will be instrumental for the model to match the size distribution of establishments observed in the data. 5 The problem of a intermediate input producer. The operating profits of an active establishment using technology i in sector s are given by o πis (z, ε) p Ps −η Pxs 1 x− ll x p,x,l rs rs −η ε ν 1−ν p zAi e x l ≥ Ys , subject to (1−ν) Ps ν ν (1 − ν) = max rs p Ys − where we normalized the wage to one, i.e. w = 1. That is, for a given productivity duple {z, ε} and technology choice i, an active establishment in sector s chooses price p, intermediate inputs x and labor l demands in order to maximize profits, subject to revenue and input subsidies. Using the first order conditions with respect to p, x and l we obtain the following expressions for prices and factor demands, 1−ν 1 η 1 Pxs ν 1 pis (z, ε) = , rs η − 1 rsl rsx Ai eε z −η 1 η (Ai eε z)η−1 Psη Ys lis (z, ε) =(1 − ν) rsl ν η−1 , 1−ν rs η − 1 P 1 (2) xs rsx rsl xis (z, ε) = (1) ν rsx lis (z, ε) . 1 − ν Pxs rsl (3) Using the normalization of the distribution of ex-post heterogeneity, µ = −(η − 1)χ2 /2, expected operating profits for an active establishment in sector s, with ex-ante productivity z, operating technology i = t, m, are given by o πis (z) ≡ o Eε [πis (z, ε)] 1 =rsη η η−1 η η−1 (Ai z)η−1 Psη Ys ν η−1 . 1−ν 1 rsl (4) Pxs rsx Operating profits are continuous in establishment productivity z. Also, given that η > 1, operating profits o (z) are increasing in z. As a result, optimal entry and adoption decisions are given by threshold rules πis {zts , zms }∀s . Furthermore, provided that the modern technology is relatively expensive, both technologies would be used in each sector, zts < zms . A establishment is active and operates the traditional technology in sector s if and only if zts ≤ z < zms , and is active and operates the modern technology if and only z ≥ zms . For each sector s, the marginal entrant zts and marginal adopter zms satisfy o πts (zts ) =κts and, o o πms (zms ) − πts (zms ) = 6 Pms κms . rsa (5) (6) The next lemma summarizes an important, albeit evident, relationship between the thresholds. Lemma 1 For all s, zms ≥ zts . Further, if 1 + Pms κms /κts > (Ams /Ats )η−1 , zms > zts . In other words, for establishments that decide to be active in sector s, i.e. those with z ≥ zts , only a fraction ζ (zms ) zts of them operate the modern technology, aess = 1−F ≤ 1, where as and es denote the mass of = zms 1−F (zts ) adopters and entrants in sector s, respectively. Notice that there is selection in adoption: adopters are those with the highest productivity within a sector. This will have important implications for the distribution of activity across establishments. We now define a symmetric equilibrium, Definition 1 Given vector of subsidies r, r l , r x and r a , a symmetric equilibrium consists of thresholds {zts , zms }s ∈ S, {Ps , Pxs , Pms }s demand for labor and intermediate inputs by the different active establishments, prices ∈ S and Pc , and a level of aggregate consumption C, such that establishments maximize profits, markets clear, X X Pc Pxs′ Pms′ C+ ωs ′ s Xs′ + λs′ s as′ κms′ , Ps Ps Ps s′ s′ X X L= Ls + es κts , Ys =γs s where Xs ≡ R zms zts xts (z)df (z) + R∞ zms (7) (8) s xms (z)df (z) and Ls ≡ R zms zts lts (z)df (z) + R∞ zms lms (z)df (z), with xis (z) = Eε [xis (z, ε)] and lis (z) = Eε [lis (z, ε)]. Adoption and Sectoral Productivity. Aggregating the output of establishments in sector s we obtain an expression for the output of sector Ys , as a function of the sector’s inputs Xs and Ls , Ys = ν ν (1 1 Zs Xsν L1−ν , s − ν)1−ν where Zs is a neutral Total Factor Productivity (TFP) shifter in sector s, defined as Z Zs ≡ Aη−1 t zms z η−1 f (z)dz + Aη−1 m zts Z ∞ z η−1 f (z)dz 1 η−1 1 1− η−1 η−1 1− η−1 η−1 ζ ζ η−1 + A − A a = Aη−1 e . s s t t m (9) zms We let Z = [Zs ]S×1 . In turn, changes in sectoral productivity and policies affect sectoral and factor prices as these changes percolates through the production network. Using (1) on the definition of the price index Ps , we can obtain a condition relating the sector price Ps , the price of the intermediate aggregate Pxs , and sector productivity 7 Zs , 1−ν Ps = 1 η rs η − 1 1 rsl Zs Pxs rsx ν . (10) This is the sectoral couterpart of (1). The price index of a sector s is a markup over its marginal cost, which depends on the price of the sector’s intermediate aggregate Pxs , its productivity Zs , and subsidies. We combine (10) with the definition of the price index Pxs to obtain, ln P =1 1 η ln − (I − νΩ)−1 ln Z + ln r + (1 − ν) ln r l + ν ln r x . 1−ν η−1 (11) 1 where P = [Ps ]S×1 , and where we used that (I − νΩ)−1 1 = 1 1−ν . Because of the input-output network, sector prices compound all sector aggregate TFP levels and markup distortions η/(η − 1) through the Leontieff inverse (I − νΩ)−1 = I + νΩ + ν 2 Ω2 + ν 3 Ω3 + ... . From the definition of the consumption price index Pc and technology adoption price index Pms , we obtain the following relationships, d ln Pc =Γ′ d ln P , and d ln Pm = Λd ln P . (12) These expressions show how changes in sector prices translate into changes in the consumption bundle and technology adoption price indexes. A key feature of our economy is that, through adoption, sector productivity is endogenous. Higher adoption in sector s increases the share of value-added within the sector produced by modern establishments and, thus, the sector’s productivity. In particular, the elasticity of sector TFP, Zs , to a change in the mass of adopters, as , is given by η−1 d ln Zs 1 ζ + 1 − η Aη−1 m − At Ms ≡ βMs > 0 . = d ln as η − 1 ζ Aη−1 m (13) The value Ms represents the share of value-added in sector s attributed to modern establishments; M = [Ms ]S×1 collects these shares for all sectors. The first expression shows how changes in adoption are converted into changes in sector TFP. Importantly, the elasticity is higher (i) the larger the productivity wedge among the two technologies, (ii) the lower the elasticity of substitution and heterogeneity within a sector, as captured by η and 1/ζ, which are key determinants of complementarities across establishments within a sector (see Buera et al., 2021), and (iii) the higher is the share of value-added of modern establishments in the sector Ms . 8 3 Sectoral Development Multipliers In this section we study the mechanism by which technology adoption affects industrial policy. To keep things as simple as possible, our benchmark analysis for the characterization follows Baqaee and Farhi (2020), and only considers revenue subsidies r (that is, ln r l = ln r x = ln r a = 0). However, while we put special emphasis on revenue subsidies, as we advance through the section we provide insights on how the different subsidies operate as a tools for industrial policy. To maintain things as simple as possible in this section, we abstract from the entry margin, thus treating es as an exogenous object across all s and abstract from the entry costs. We do this to provide a more transparent analysis of the adoption margin, which is the main focus of the paper. The entry margin is relevant for matching the data, but it is somewhat orthogonal to endogenous technology adoption. In fact, in our quantitative analysis, as shown in Table IV, we show that the entry margin accounts for a very small fraction of the development multipliers, and thus little is lost by ignoring this margin.4 Still, we emphasize that we do take the entry margin into account when we quantify the multipliers using data. As in Liu (2019) we define the sectoral revenue development multiplier in sector s as subsidy elasticity ϵrs z }| d ln C ≡ d ln rs { /(Fiscal cost of policy) . (14) r=1 Basically, the multiplier accounts for the elasticity of aggregate consumption C to a revenue subsidy in sector s—–the subsidy elasticity, d ln C/d ln rs , around the equilibrium with no subsidies, ln r = 1, relative to the fiscal cost of the policy. A multiplier ϵr,s above one implies that an industrial policy in sector s increases aggregate consumption by more than the fiscal cost, a multiplier equal to one is neutral, a positive multiplier lower than one implies that the increase in aggregate consumption is lower than the fiscal costs, and a negative multiplier implies that aggregate consumption falls as result of the policy.5 3.1 Subsidy and Productivity Elasticity of Aggregate Consumption As a first step, we present a characterization of the elasticity of consumption with respect to a revenue subsidy and a sectoral productivity. Let δs = Pms as κms Pc C denote sector s adoption share of GDP, with ∆ = [δs ]S×1 a vector collecting these shares. We define the Domar and Forward Domar weights in the economy as −1 η−1 ′ Ψ = Γ +∆Λ I −ν Ω , Ψ̃ = Γ′ + ∆′ Λ (I − νΩ)−1 . η ′ ′ ′ 4 (15) Follows from comparing the columns for ϵ and ϵe of Table Table IV. While in our setting lump-sum taxation would imply that any positive subsidy elasticity would suffice to motivate the use of the instrument, one could consider that there are alternative uses for the income taxed away from consumers. In such a setting, taking into account the fiscal cost of the policy is appropriate. 5 9 The Forward Domar weight is usually referred in the literature as the Cost Domar weight. Our choice of language will be clear in the discussion of the following proposition. Proposition 1 Consider independent vectors of changes in revenue subsidies d log r and changes in sector TFP induced by changes in adoption d log Z = β diag (M ) d ln a, then TFP channel production channel d ln C = z }| { ′ Ψ̃ − Ψ′ z }| { η−1 ′ d ln r + Ψ̃ − Ψ d ln Z . η ′ (16) ′ The term Ψ̃ − Ψ′ gives the aggregate consumption gains that follow from changes in the allocation of resources across sectors. Ignoring the resources used, the change in the policy induces changes in the quantity produced by the different sectors. The first round effects of these changes are proportional to the final demand elasticities, which are measured by the consumption elasticities vector Γ′ and the adoption elasticities ∆′ Λ. The effects from the subsequent rounds are proportional to the intermediate input elasticities, which are measured by the matrix νΩ, and its subsequent powers, which themselves give the increase in the output of all sectors as a result of the greater availability of inputs from the promoted industries. The impact of these subsequent rounds are ultimately also proportional to the aggregate demand elasticities, thus ′ (Γ′ + ∆′ Λ) νΩ + (νΩ)2 + ... = (Γ′ + ∆′ Λ) (I − νΩ)−1 = Ψ̃ . The Forward Domar weight Ψ̃ accounts for all these cumulative effects.6 Increasing production requires reallocating labor across sectors. The cost of the reallocation is measured by the Domar weight Ψ, which accounts for the labor used by the sector directly and, indirectly, by all sectors supplying inputs to this sector. This cumulates to the labor or Gross Output share of the sector. Because of markup distortions, gross labor shares and cumulative final demand ′ elasticities differ, and thus Ψ̃ − Ψ′ can be positive. The larger is the markup, the larger the divergence between these vectors, and larger the gains. When the markup approaches zero, i.e. η/(η − 1) → 1, forward and backward Domar weights are equal and thus the production channel of the subsidy elasticity is zero. Overall, gains are large when a sector has large (forward) cumulative final demand elasticities and small (backward) cumulative resource shares. Endogenous technology adoption affects the subsidy elasticity by inducing a change in sector produc′ tivities Z, the TFP channel. The term Ψ̃ − η−1 ′ η Ψ regulates this margin. Notice that the intuition that we discussed for the production channel also applies here, with the only difference being that the vector of Domar weights, which accounts for the cost of reallocating labor, is now discounted by (η − 1)/η. This occurs as the total resources required to implement a given increase in TFP combines a direct labor reallocation cost, Ψ′ , with the way this labor is transformed into sector productivity through adoption. The reciprocal of the adoption elasticity of TFP, (∂ ln Zs /∂ ln as )−1 , accounts for how adoption is transformed into productivity, and it is proportional to (η−1). Generating this increase in adoption requires changing the identity of the marginal adopter. Thus, the cost per unit of adoption is given by the profit of the marginal 6 In deriving these expression it is convenient to work with the dual, and trace out the fall in sector and final consumption prices induced by the change in policies—see (11) and (12). 10 adopter, which are proportional to 1/η. The ratio (η − 1)/η combines the marginal rate of transformation from adoption to sector productivity with its marginal cost. As the elasticity of substitution declines, a lower increase in adoption is required to implement a given change in TFP, relative to the cost per unit of adoption. Therefore, as the elasticity of substitution declines, fewer resources are needed to increase sectoral productivity. ′ ′ ′ The entries of the vectors (Ψ̃ −Ψ′ ) and (Ψ̃ − η−1 η Ψ ) in (16) give the differential effects of sector-specific subsidies and TFP changes, respectively. The relative magnitude of each entry depends critically on the details of the production network, given by the matrix Ω. Aggregating these entries results in the effect of a uniform and unitary policy or change in TFP. The following remark gives a simple expression for the elasticity of aggregate consumption to an uniform unitary change in revenue subsidies and TFP. Remark 1 Consider a uniform unitary changes in revenue subsidies and TFP in all sectors, d ln r = 1 and d ln Z = 1, then the expression in (16) reduces to d ln C = 1 + δ̄ where δ̄ = ν 1 + 1−ν 1−ν 1 1 , η 1 − ν η−1 η P s δs . A notable feature about this result is that the expression shows that, under uniform policies, the complexities of the network structure of production do not play any role. Indeed, the gains accrued by the uniform policy and TFP gains are exactly those obtained in a one-sector economy with roundabout production. 3.2 Subsidy Elasticity of Adoption and Development Multiplier To find an expression relating how changes in the subsidy affect sector productivity, we need to find an expression for d ln a from the marginal adopter’s condition. In this case, it is convenient to consider both revenue and adoption subsidies, d ln r and d ln r a . Log-differentiating the vector of conditions for the marginal adopter in (6), and using (10), we obtain a condition linking adoption, sectoral productivity, gross output, and the price of the adoption good, with the vector of revenue and adoption subsidies, d ln r − (η − 1)d ln Z + d ln (P ◦ Y ) − η−1 d ln a = −d ln r a + d ln Pm , ζ (17) where the symbol ◦ denotes the Hadamard (element-wise) product. Notice how revenue r and adoption r a subsidies have the same direct effect on adoption. Revenue subsidies also affect adoption indirectly through its effect on sector prices, which affect the aggregate demand channel and, importantly, the vector of prices of the adoption good Pm . A simple case to study is when an adoption subsidy is applied to a sector where the modern share is zero, i.e. Ms = 0. The next Remark describes the result. 11 Remark 2 Assume that Mŝ = 0 for some sector ŝ. Then, d ln aŝ d ln rŝa = ζ η−1 , and d ln as d ln rŝa = 0 for all s ̸= ŝ. The Remark follows immediately after noticing that (13), given that Mŝ = 0, provides that the elasticity of the sector’s TFP to adoption is also zero, and thus d ln Z = d ln (P ◦ Y ) = d ln Pm = 0. Adoption increases in sector ŝ because of the direct effect of the policy: ζ measures the elasticity of the distribution of ex-ante heterogeneity, which is normalized by the curvature of the profit function with respect to this productivity, η − 1. But the percentage increase in adoption in this sector does not translate into a percentage increase in the sector’s TFP. As a result, the sector’s price index does not vary, and the same occurs to the price of adopting a modern technology in other sectors, and thus adoption in other sectors is unaffected by the policy. Remark 2 showcases the relevance of the modern share Ms : through its effect on the elasticity of TFP with respect to adoption, is a crucial ingredient for understanding how subsidies in one sector percolate to other sectors through adoption. We now follow to solve for d ln a from (17), together with the price feedbacks in (11) and (12). Combining the relationships between prices, policies and TFP in (11), (12) and (17), we obtain an expression for the elasticity of adoption d ln a with respect to an adoption subsidy d log r a . The next Proposition presents this result. Proposition 2 The elasticity adoption with respect to a adoption subsidy is given by direct effect amplification z }| z }| { −1{ i ζ ζ h ζ d ln a = d ln r a . I− Λ (I − νΩ)−1 − (η − 1)I βdiag (M ) − ∇P Y,a η−1 η−1 η−1 | {z } (18) ≡∇a,ra Details on the derivation are available in Appendix A.3. The operator diag(·) converts a column vector into a diagonal matrix, and ∇P Y,a denotes the elasticity of P ◦ Y with respect to a, which is presented in close-form in the appendix. To keep things simple, we will provide intuition abstracting from this last channel. The elasticity of adoption with respect to an adoption subsidy combines the direct effect of the policy, measured by the term ζ/(η −1), with the amplification effect that follows from the sector interactions. The degree of amplification depends on the structure of the production (Ω) and investment (Λ) networks, as the effect of policies percolates through them. A subsidy to a sector promotes adoption and TFP in that sector, which in turns lower the sector price, lowering the cost of adoption in all sectors. This leads to a further increase in adoption in other sectors, resulting in further feedback rounds. The magnitude of these multiplier effects depends on parameters governing the elasticity of productivity with respect to adoption given in (13), which are subsumed in β and the vector of modern shares M and, for given values of these quantities, the effects are increasing in the elasticity of substitution and heterogeneity within a sector, as captured by η and 1/ζ, which are key determinants of complementarities across establishments (see Buera et al., 2021). 12 In virtually all research exploring the relevance of input-output structures for economic aggregates, the key object is the Leontieff Inverse (I − νΩ)−1 , its backward counterpart, or linear combinations of them. This occurs as heterogeneity in the way sectors combine sectoral output to use as a intermediate aggregate in production results in heterogenenity in sector price indices. Because the intermediate aggregate used by any sector aggregates the output of all other sectors through Ω, so does sector prices P . Absent technology adoption, aggregate effects of policies are the result of comparisons of objects that directly map to the Leontieff Inverses. Technology adoption adds a double inverse to the loop. Adoption increases TFP, which lowers sector prices and thus the cost of adoption. This is the inner inverse. Also, adoption in a sector reduces the marginal cost of production in all sectors, thus increasing the profits accrued from adoption by the marginal adopter. This feedback accounts for the outer loop of the Double-Leontieff Inverse. To further understand the role of individual parameters and of the value-added share of modern establishments in shaping the Double-Leontieff Inverse, the next remark characterizes the impact of a uniform unitary change in adoption subsidies on the elasticity of adoption, under the additional restriction that there is no heterogeneity in the modern share of value-added across sectors. This expression is the same one would obtain in a one-sector economy with roundabout production. Remark 3 Assume M = mI with m ≥ 0, and d ln (P ◦ Y ) = 0. Consider a uniform unitary change in adoption subsidies in all sectors, d ln r a = 1, then the expression (18) reduces to 1 d ln a = 1− h 1 1−ν ζ i . ζ η−1 − (η − 1) βm η−1 (19) As briefly discussed earlier, the subsidy elasticity of adoption features two nested multipliers. Subsidies promote adoption which enhance productivity, mediated by the elasticity βmζ/(η − 1), which lowers sector prices and, therefore, the price of adoption. These effects get amplified as they percolate through the inputoutput structure. The inner multiplier 1/(1 − ν), which is the standard multiplier of a roundabout economy, encodes these effects. The increase in the productivity of competing producers, captured by the term η − 1, partially dampens this effect. The outer multiplier captures the infinite rounds of adoption lowering the marginal cost of production, further enhancing adoption. Finally, ζ/(η−1) is the direct elasticity of adoption with respect to the subsidy, which we described earlier. The expression in (19) also allows us to see ingredients that, through the adoption channel, can greatly amplify the effect of policies even under uniform policies. While in a simple production economy the multiplier becomes unbounded as ν approaches one, with adoption the multiplier becomes unbounded when ν approaches 1 − βmζ 1 η−1 1+βmζ . Under our calibration later on, this upper bound is always between zero and one. This implies that, through the adoption channel, the multiplier becomes unbounded for lower values of the intermediate input elasticity ν. In other words, through the adoption channel, even the effect of uniform policies can be greatly amplified, and it is possible to generate larger amplification than through the production channel for a wider set of parameter values. In our calibrated economy this is not the case, 13 i.e., technology adoption greatly amplifies the multiplier for key sectors, but not the multiplier associated with a uniform policy, which is consistent with the findings in Buera et al. (2021) for a symmetric economy with roundabout production. Putting together the results from Propositions 1 and 2, we obtain the following expression for the elasticity of aggregate consumption with respect to independent revenue and an adoption subsidies: h i η−1 ′ d ln C = Ψ̃ − Ψ d ln r + Ψ̃ − Ψ β diag (M ) ∇a,ra I + ∇P Y,r + Λ (I − νΩ)−1 d ln r η η−1 ′ ′ (20) Ψ β diag (M ) ∇a,ra d ln r a , + Ψ̃ − η ′ ′ ′ where ∇P Y,r denotes the elasticity of P ◦ Y with respect to r, which is provided in close-form in Appendix A.3. We note the relevance of the Double-Leontieff inverse, present in the key object ∇a,ra , as characterized in Proposition 2. The matrix ∇a,ra converts subsidies into adoption, which in turn is converted to TFP through η−1 d ln Z = βdiag (M ) d ln a. A key determinant of this conversion rate is the technology gap Aη−1 m − Am , as discussed in (13). Then, the change in TFP is converted to aggregate consumption through the elasticity ′ η−1 ′ Ψ̃ − η Ψ . We provided a discussion of this term in Proposition 1. In settings with exogenous productivity, Acemoglu et al. (2012) noted that Ψ̃ accounts for the Influence matrix. That is, the entry s of the Influence matrix provides the change in aggregate consumption resulting from a shock to exogenous sector productivity Zs .7 With endogenous sector TFP, we note that Ψ̃ accounts for Influence in production, while (∇a,ra )′ βdiag (M ) Ψ̃ accounts for influence in adoption. That is, the entry s of this object accounts for the resulting change in aggregate consumption following from an exogenous shock to the adoption technology in sector s. With elasticities at hand, we need expressions for the fiscal cost of the different policies to be able to produce the multipliers. Similarly to Liu (2019), the fiscal cost of the revenue subsidy is given by the Domar weight Ψ, as this weight provides the size of each sector, and thus, around the equilibrium with no subsidies, the fiscal cost of the subsidy. Therefore, obtaining an expression for ϵrs requires dividing the s entry of (20) by the s entry of vector of Domar weights Ψ. Similarly, the fiscal cost of the adoption subsidy in sector s is given by the adoption share of the sector, the s entry of ∆, the fiscal cost of a intermediate input subsidy in sector s is given by the s entry of νΨ, and the fiscal cost of a labor subsidy in sector s is given by the s entry of (1 − ν)Ψ. The following remark provide expressions for the development multipliers associated with uniform and unitary revenue and adoption subsidies, Remark 4 Suppose that M = mI and d ln P ◦ Y = 0 in the marginal adopter’s problem in each sector. 7 Exogenous sectoral productivity implies no adoption, and thus ∆ = 0. Then, under the restriction that Γ = 1, we get that Ψ̃ = 1′ (I − νΩ)−1 which is the influence measure described in Acemoglu et al. (2012). ′ 14 Then, under uniform and unitary revenue and adoption subsidies, i.e. d ln r = d ln r a = 1, we have that ϵur ν 1 η−1 1 u = , + βmϵra 1 + 1−νη η 1−ν where ϵura = 1 1 h η−11−ν1− 1 1 1−ν ζ i . ζ η−1 − (η − 1) βm η−1 These expressions highlight the counteracting forces driving the relative importance of the development multiplier associated with revenue and adoption subsidies. On the one hand, revenue subsidies have a direct effect enhancing production efficiency, as captured by the first term of ϵur . In addition, revenue subsidies affects the adoption incentive directly and indirectly, through their effect on the price of the adoption good, as capture by the last bracketed term. On the other hand, adoption subsidies have an advantage given their lower fiscal cost as the adoption share is smaller than the Domar weight, i.e. δs = ((η − 1)/η)βMs Ψs < Ψs . 3.3 Additional Results In this section we present two additional analysis. First, we study the sectoral development multipliers associated with input subsidies, and compare them to the multipliers associated with revenue subsidies. To simplify the analysis, we study them in the case with no adoption. This section sheds light of the relevance of different policy instruments as an engine of development, and will aid on our quantitative analysis contrasting the power of different instruments. Second, we consider a version of the model where the adoption good is produced with labor only. This section sheds light on the importance of using fixed or variable inputs in the production of the adoption good. 3.3.1 Alternative Policy Instruments Consider the case of revenue r, intermediate inputs r x and labor subsidies r l in the economy with no adoption. The results in this section will shed light on (i) the differential relevance of the adoption margin depending on whether the adoption good is produced with intermediate inputs or labor, and (ii) the potential of the different policy instruments as the engine of industrial modernization. In this case, the subsidy elasticity of aggregate consumption is given by ′ d ln C = Ψ̃ − Ψ ′ η−1 ′ η−1 ′ ′ x d ln r + Ψ̃ − Ψ νd ln r + (1 − ν)Ψ̃ − 1 − ν Ψ′ d ln r l . η η ′ Naturally, the contribution of the revenue subsidy (Ψ̃ − Ψ′ ) is the sum of the contributions of the intermediate input and labor subsidies. When considering intermediate input subsidies, the Forward Domar ′ weight Ψ̃ is weighted by the elasticity of intermediate inputs in production, ν, and the Backward Domar 15 weight Ψ′ is weighted by the intermediate input share, ν(η − 1)/η. A similar argument applies for labor subsidies. To further gain insights about the power of the different policy instruments we restrict to uniform and unitary subsidies, d ln r = d ln r x = d ln r l = 1, and we contrast the resulting multipliers. Under the uniform policy, ϵurl = 0 < ϵur = ν 1 1 1 < ϵurx = . 1−νη 1−νη While a labor subsidy to an individual sector has generically an effect on aggregate consumption, either positive or negative, these individual effects cancel out when aggregated up, implying that the multiplier of a uniform labor subsidy is zero. Key to this result is that labor is a fixed input. When aggregating the Backward Domar weight we obtain the gross output share of consumption, 1/(1 − ν(η − 1)/η), which −1 results from accumulating infinite rounds of the (variable) intermediate inputs share, I − ν η−1 Ω 1= η n P ∞ η−1 1. This share is then multiplied by the share of the (fixed) labor input, one minus the n=0 ν η Ω ′ share of the variable input. A similar argument applies to (1 − ν)Ψ̃ 1, resulting in zero consumption effect of a labor uniform subsidy. Intuitively, in the aggregate, nothing can be gained by promoting uniformly the use of a fixed input. This is reminiscent of the results in Liu (2019). The results of the uniform labor policy are useful as an input to understand the values of the uniform revenue and intermediate input multipliers. In fact, the revenue and intermediate input elasticities are the same, d ln C d ln r = r=1 d ln C d ln r x = (1 + δ̄) r x =1 ν 1 1 . 1 − ν η 1 − ν η−1 η They are equal because a uniform subsidy provides no gains in terms of labor reallocation and all gains follow from the intermediate input channel. This is why ν appears in the numerator. But, although the two uniform policies share the same value for the elasticity, they differ in the cost of the policy. The revenue uniform policy is ’paying’ for labor reallocation with no gains, while the intermediate inputs uniform policy circumnavigates paying for the added cost. Therefore, while both policies have the same subsidy elasticities, the intermediate input policy cost is only ν for every unit of cost in the revenue policy. As a result, ϵur /ϵurx = ν < 1. While the intermediate input subsidy dominates the revenue subsidy in the economy with no adoption, this is not necessary the case in the model with adoption. This occurs as, while both the revenue and intermediate input subsidies affect the marginal adopter’s condition indirectly through their effect on sector prices, the revenue subsidy also adds a direct effect on the marginal adopter’s profit. As a result, we resort to our quantitative analysis of multipliers to gauge the relative importance of the different subsidies as an engine of development. 16 3.3.2 Adoption goods produced with labor only While we see technologies in our model as embedded in intermediate inputs (closer to capital in a dynamic version of the model), it is useful to evaluate the extent to which sectoral multipliers are sensitive to this assumption. In this section we study the polar opposite case, where the adoption good is produced solely with labor instead of solely with goods. When labor is the sole input of production for the adoption good, the subsidy elasticity expression in (16) is given by ′ ′ d ln C = Ψ̃ − Ψ 1 η−1 ′ d ln r + (1 − ν)Ψ̃ − 1 − ν Ψ′ d ln Z , 1−ν η −1 ′ where, as before, Ψ̃ = Γ′ (I − νΩ)−1 , Ψ′ = Γ′ I − ν η−1 Ω , and d ln Z = βdiag(M )d ln a. η First, notice that when ν = 0 we have that d ln C = 0′ d ln r+0′ d ln Z. While there are markup distortions in the economy, they do not manifest in creating a wedge between labor shares and final demand elasticities, thus rendering industrial policy useless. Second, consider the effect of uniform subsidies in this economy. As before, we set d ln r = d ln Z = 1. The contribution of the production margin is analogous to the uniform policy in the baseline economy, 1 ν 1 η−1 . However, this analogy does not follow through for the contribution from productivity, 1−ν η 1−ν η " −1 # ′ ′ η−1 η−1 η−1 −1 ′ Ω 1=0. (1 − ν)Ψ̃ − 1 − ν Ψ̃ 1 = Γ (1 − ν) (I − νΩ) − 1 − ν I −ν η η η (21) This result is similar to the case of a labor subsidy in the economy without adoption. While the cumulative effect of the policy must be zero, aggregate consumption can increase or decrease by reallocating adoption labor across sectors. In fact, if a subsidy in one sector has a positive effect on aggregate consumption, there must be another sector with a negative effect on aggregate consumption. Finally, in a symmetric economy with roundabout production in each sector, i.e. Ω = I, we have that the vector of TFP contributions satisfy " Γ′ η−1 (1 − ν) (1 − ν)−1 I − 1 − ν η # η − 1 −1 1−ν I =0′ . η That is, each sector’s contribution through TFP must be zero, and thus there are no gains from reallocating adoption. This last result is reminiscent of that one in Atkeson and Burstein (2010). As a result, adoption reallocation can have relevant effects for aggregate multipliers if there is substantial heterogeneity in the network structure of production. 17 4 Estimation and Calibration We begin this section by presenting a simple identification strategy that allow us to structurally estimate the model and the distribution of technologies across sectors. Our strategy consists of three stages, and combines structural estimation and calibration methods. After describing our strategy, we introduce the data that we use and present the estimates. Finally, we perform a validation exercise of the structural estimates using independent proxies of technology adoption. 4.1 Identification While the model is able to capture complex sectoral production and adoption complementarities, it is parameterized in a relatively parsimonious fashion. We need to assign values to the following set of parameters (scalars, vectors and matrices): At , Am , ζ, χ, η, ν, {κts , κms }for all s , Γ, Ω, Λ. To do this, we follow a threestage procedure. In the first stage, we exploit the theory implications for the size of an establishment within a sector, and estimate a reduced-form representation of the parameters governing the distribution of technologies in the model, i.e. combination of the deep parameters of the model. In the second stage, we combine direct measurements from the data and estimates from other sources to calibrate the production and adoption networks, i.e. Ω and Λ, and the elasticity of substitution across varieties within a sector η. Finally, in the third stage, we use the equilibrium construct of the model to unbundle the reduced-form estimates obtained in the first stage and recover the deep parameters of the model governing technology adoption. We also show that this procedure provides a unique mapping from the reduced-form and deep parameters estimates. Further, following the working assumptions for multipliers, we assume no subsidies in the starting equilibirum. That is, ln r = ln r l = ln r x = ln r a = 1. Brute force estimation of the multisector economy is a challenging task. For example, given that entry and adoption in a sector depend on entry and adoption in other sectors, it is unclear how to operationalize the procedure for the estimation of entry and adoption costs. This problem gets exponentially more convoluted the more sectors are considered. Our insight to address these complications is the following: In order to explain variation in the size of establishments within a sector, a common estimation target in the literature, one does not need to understand how sectors are connected, nor the level of aggregate prices, nor entry and adoption rates in other sectors. This partial information procedure allows us to provide reduced-form estimates for entry and adoption thresholds for each sector, productivity parameters, and the distribution of ex-ante and ex-post heterogeneity. Taking logs to (2), using (10), and setting rs = rsl = rsx = 1, implies that ln lis (z, ε) = ln Ãi + ln z̃s + ln ε̃ , , z̃s ≡ (1−ν) η−1 where Ãi ≡ Aη−1 i η Ps Ys η−1 z , Zsη−1 (22) and ε̃ ≡ εη−1 . In this representation, the size of an establishment within sector s combines (i) a technology-specific component Ãi , (ii) a mix of a sector component and ex- 18 ante heterogeneity of the establishment, z̃s , and (iii) an idiosyncratic ex-post component, ε̃. Notice that the sector component cannot be identified by observing heterogeneity in size of establishments within a sector. Unbundling the sector and the ex-ante heterogeneity components is the main objective of the third stage in our devised procedure. Within a sector, heterogeneity in technology use manifests through its impact on the employment-size distribution. Let Hs (l) = Pr (ls (z̃, ε̃) ≤ l), with density hs (l) ≡ ∂Hs (l) ∂l , be the employment size distribution of establishments within sector s. ! 2 ζ̃ ln z̃ − ln l + µ̃ + χ̃ ms − e Φ hs (l) =l−ζ̃−1 −ζ̃ χ̃ z̃ts !# " !#) 2 ζ̃ ln z̃ts − ln l + µ̃ + χ̃2 ζ̃ ln z̃ − ln l + µ̃ + χ̃ ms Φ + Ãζ̃m 1 − Φ , χ̃ χ̃ ζ̃ where ζ̃ ≡ ζ η−1 , 2 µ̃ζ̃+ χ̃2 ζ̃ 2 ( " Ãζ̃t (23) χ̃ ≡ (η − 1) χ, µ̃ ≡ (η − 1)µ = −χ̃2 /2, and where z̃ts and z̃ms are the entry and adoption thresholds under the reduced-form representation. Details on the derivation are available in Appendix A.4. If one shuts down technology heterogeneity and the ex-post shock ε, i.e. Ãt = Ãm and χ̃ → 0, hs (l) is Pareto. Departures from Pareto are the result of heterogeneity in technology adoption and dispersion following from ε̃. Similarly, we can define the employment weighted size distribution across establishments R ˆ ˆ ˆ Gs (l), where gs (l) ≡ ∂Gs (l)/∂l = lhs (l) / lhs l dl. Inference using hs (l) or gs (l) in the reduced-form model provides estimates for Ãt , Ãm , χ̃, ζ̃, and for the reduced-form entry and adoption thresholds {z̃ts , z̃ms }for all s . The next two remarks show how the shape of hs (l) and gs (l) is informative about the extent of heterogeneity in technology use within sector s, which is encoded in the technology parameters Ãt and Ãm , and in the entry and adoption thresholds z̃ts and z̃ms . We explore the informativeness of the distributions regarding technology adoption by characterizing two extreme cases. In the first remark we show that if there is no variation in technology use within a sector the resulting distributions are unimodal. In the second remark we show that with no ex-post heterogeneity, heterogeneity in technology use results in bimodal distributions. Remark 5 Suppose only one technology is used in sector s. Then, the distributions Hs (l) and Gs (l) are unimodal. The proof is available in Appendix A.5. The proposition states that a combination of Pareto and Log-Normal shocks, both unimodal distributions, imply equilibrium distributions that are also unimodal. Because the ex-post shock ε̃ is iid, it shifts mass in such a way that does not generate two modes. Remark 6 Let χ → 0. Then, the distributions Hs (l) and Gs (l) have two modes if and only if z̃ms > z̃ts . In this case, the two modes are at l = z̃ts and l = z̃ms . If z̃ms = z̃ts , both distributions have one mode, located at l = z̃ts . 19 The proof is available in Appendix A.6. The proposition establishes that, by shifting mass from the middle of the distribution to the right of the distribution, technology adoption generates bimodality. Under the reduced-form representation the endogenous choices of establishments regarding entry and technology adoption are summarized by the thresholds z̃ts and z̃ms for all s. In order to compute development multipliers and perform counterfactual analysis we need to unbundle these thresholds and uncover the entry and adoption costs, i.e. κts and κms for all s, that rationalizes these thresholds. The next proposition uses the equilibrium conditions of the model and establishes that there is a unique mapping between the reduced-form thresholds and the underlying entry and adoption costs. Proposition 3 Given values for η, ν, Γ, Ω, Λ, and the reduced-form objects ζ̃, Ãt , Ãm , {z̃ts , z̃ms }for all s , then the vectors of entry and adoption costs, {κts , κms }for all s , are uniquely identified. The proof of this proposition is available in Appendix A.7, and follows from the fact that the equilibrium of the model can be described by a system of linear equations given the reduced-form thresholds, z̃ts and z̃ms for all s. 4.2 Data and calibration We focus our quantitative analysis on the India economy and combine three sources of data: (i) Indian Economic Census, (ii) the World Input-Output Database (WIOD) and, (iii) data on the investment network from vom Lehn and Winberry (2022). The Indian Economic Census is a complete count of all establishments/units located within the geographical boundaries, engaged in production or distribution of goods or services other than for the sole purpose of own consumption, crop production and plantation, public administration and defence, or activities of households as employers of domestic personnel. We further drop establishments in animal production, forestry and logging, and fishing and aquaculture. The sectoral data is given at the 3-digit level as per the National Industrial Classification (NIC 2004 and 2008, for the 2005 and 2012-13 Economic Censuses, respectively). We use the Sixth Indian Economic Census from 2012-13, which is our main source for data on the size distribution of establishment by sector, which is the data used to estimate the reduced-form representation of technology adoption by sector. We measure employment in an establishment as the sum of non-hired and hired workers, including family workers and owner.8 We also partially rely on the Fifth Indian Economic Census as it provides information about the type of power used by individual establishments, a proxy for the technology used at the establishment level. This data is used in the validation exercise at the end of this Section. The final dataset for the Sixth (Fifth) Indian Economic Census includes 45363786 (35171881) establishments employing 108411367 (84134947) individuals. 8 The question ask for the number of persons found working comprising, hired, non- hired (including family members; unpaid apprentice and owner), on the last working day in the establishment. Regular wage/salaried workers, owner/other family workers, who are temporarily absent on the last working day are also counted. 20 The WIOD (Timmer et al., 2015) contains international comparable information on the input-output linkages and sectoral intermediate and final uses, covering 43 countries, for the period 2000-2014. Data for 56 sectors are classified according to the International Standard Industrial Classification revision 4 (ISIC Rev. 4).9 We use the table for India in 2010. This data gives us information to directly calibrate Γ and Ω. There is no data available for India that can be easily mapped to the network of investment Λ. We overcome this limitation by using the network of investment produced in vom Lehn and Winberry (2022) for the United States in 2010. In particular, the investment network records the share of new tangible and intangible investment expenditures of sector s that were purchased from sector s′ for each pair of sectors. They construct the investment network using the BEA Fixed Assets and Input-Output databases for a sample of 37 private non-farm sectors from 1947-2018. We consider this as the best available proxy for the production network of modern technologies in India. After harmonizing the sectors across the three datasets, the India economy that we consider is composed of 30 2-digit sectors. Details on how we armonize the different data sources into these 30 cohesive sectors are available in Table III in the Appendix. Also, we follow Broda and Weinstein (2006) and Hsieh and Klenow (2009) and we set η = 3. Given the value of η, we set ν = 0.75 so that the intermediate input share of gross output equals 0.49, the value in the WIOD for India in 2010. 4.3 Structural Estimation In this section we structurally estimate the parameters of the model. We use the insights of previous sections and estimate reduced-form parameters for each sector within the Indian economy. Then, we use the system of equilibrium equations in the model to back out the fundamental parameters that make the partial information estimates that we obtained for each sector consistent across sectors. We assume that all sectors in the economy share the same distribution of (normalized) ex-ante and ex-post heterogeneity, {ζ̃, ε̃}, and gains from adoption Ãm . Without loss of generality, we set Ãt = 1. Thus, for each sector we estimate the (normalized) entry and adoption thresholds {z̃ts , z̃ms }s in S . While our approach may seem restrictive, we show that this parsimonious approach is able to provide a good fit of the data. Operationally, we estimate the parameters of the model by matching the binned employment share distribution for each sector—we use 21 bins per sector, where for the common parameters we weight equally all sectors. We want to notice that, unlike for other parameters, we estimate the tail parameter ζ̃ by using direct information from the tail of the size-distribution of establishments for the whole Indian economy. Table I presents the implied share of modern establishments, as /es , and the their value-added share, Ms , for the 30 sectors of the Indian economy, as well as the estimated entry cost κes and the entry cost relative to the cost of adoption, κes /(Pms κas ). The table shows large heterogeneity across sectors. For example, the estimation provides that modern establishments account for 60% of value-added share in the Mining sector, 0.4% in the Manufacturing of Wood sector (M-Wood), and 87% in the Manufacturing of Motor Vehicles 9 The industry codes of the ISIC Rev. 4 and the NIC 2008 classifications coinceide at 2-digit, which is the level of aggregation that we use to harmonize them with the investment network in vom Lehn and Winberry (2022). 21 Table I: Modern establishments in India Division as /es Ms κes κes Pms κas 1. Mining 2. M-Food 3. M-Textiles 4. M-Wood 5. M-Paper 6. M-Printing and Media 7. M-Petroleum 8. M-Chemicals and Pharma. 9. M-Plastics 10. M-Other Non-Metallic 11. M-Basic Metals 12. M-Metal Products 13. M-Computer and Electronic 14. M-Electrical Equipment 15. M-Machinery and Equipment 16. M-Motor Vehicles 17. M-Other Transport 18. M-Furniture and Other 19. Utilities 20. Construction 21. Trade 22. Transportation 23. Food and Accommodation 24. Information 25. Professional Services 26. Finance and Insurance 27. Real Estate 28. Education 29. Health and Social Work 30. Other Services 0.021 0.002 0.001 0.000 0.534 0.000 0.031 0.149 0.761 0.098 0.054 0.002 0.070 0.074 0.011 0.227 0.100 0.000 0.024 0.001 0.000 0.000 0.000 0.000 0.015 0.617 0.001 0.994 0.000 0.000 0.606 0.358 0.305 0.004 0.948 0.033 0.648 0.824 0.978 0.780 0.711 0.340 0.742 0.748 0.532 0.868 0.782 0.000 0.622 0.277 0.003 0.209 0.005 0.000 0.568 0.961 0.313 0.999 0.000 0.002 1.387 0.235 0.337 0.262 0.366 0.819 0.607 1.745 0.536 0.938 1.078 0.810 1.083 0.990 0.624 8.340 50.892 0.422 0.640 0.369 0.156 0.125 0.391 0.585 0.577 0.185 0.316 0.282 0.614 0.125 0.023 0.005 0.004 0.000 0.179 0.000 0.029 0.080 0.224 0.061 0.042 0.005 0.049 0.051 0.016 0.104 0.062 0.000 0.025 0.003 0.000 0.002 0.000 0.000 0.019 0.196 0.004 0.265 0.000 0.000 sector (M-Motor Vehicles). Heterogeneity in these shares is rationalized by heterogeneity in the relative cost of adoption: in Mining the entry cost is 2.3% of the added cost of adoption, in M-Wood it is basically negligible (in other words, the cost of adoption is very large), and in M-Motor Vehicles is 10% of the added cost. Figure 1 presents the empirical and the model implied employment share distributions for these three 22 Figure 1: Examples of employment share distributions, data vs. model 1. Mining 4. M-Wood 16. M-Motor Vehicles sectors.10 The model closely replicates the distributions in each sector, even though the only degrees of freedom at the sector level are the entry and adoption thresholds. The Mining sector presents an evident bimodal employment share distribution, implying that both small and large establishments are highly relevant. As a result, the estimation procedure matches this by having 60% of value-added share accounted for by modern establishments. Consistent with this, Table I shows that the cost of adoption in the sector, Pm sκms , is large relative to the entry cost, κes , but not extreme. Highly productive establishments end up adopting and, while accounting for a small share of establishments in the sector, they account for a relatively large share of value-added. In the M-Wood sector, the empirical distribution resembles a Pareto distribution and thus the estimation procedure considers the sector to have almost no establishments adopting modern technologies, with a value-added share of modern establishments close to zero, which, as shown in Table I, is obtained in the model by setting large adoption costs relative to entry costs.11 For the M-Motor Vehicles sector, the empirical distribution shows that most of employment is concentrated in large establishments, which the model matches with a high adoption rate. In fact, as previously discussed, the value-added share of modern establishments in the sector is almost 0.87. To generate high adoption rates in this sector, the estimation recovers entry and adoption costs that are relatively similar, with the adoption cost accounting for around 90% of total effective cost of adoption. While we estimate parameters to match individual sector employment share distributions, the model provides a employment share distribution for the whole India economy which we can contrast with the empirical one. This can be found in Figure 2. The model-implied distribution closely resembles the distribution observed in the data, although over-predicting the employment share of mid-size establishments, and under-predicting the employment share of the very large ones. Given that we calibrated the elasticity of substitution to be equal to 3, it is immediate to back out our estimates for ζ, χ and Am from the reduced-form estimates. We also note that we provide 95% bootstrap confidence intervals for the different objects of interest. Table II presents the estimates and confidence 10 Figure 11 in the Online Appendix presents the empirical and model implied employment share distributions for all sectors. It is also relevant that the M-Wood sector has relatively small establishments relative to the other sectors in the economy. Otherwise the estimation procedure would interpret the sector as having full adoption of modern technologies. 11 23 Figure 2: Employment share distribution in the aggregate Indian economy intervals for the common parameters. Table II: Common parameters, estimates Ex-ante heterogeneity, ζ Ex-post heterogeneity, χ Technology, Am 3.16 0.57 2.18 [3.03, 3.19] [0.49, 0.60] [2.01, 2.31] We also provide an off-sample test of the validity of our estimates. We do so by contrasting adoption measures implied by the estimated model with adoption proxies that can be observed in the data. In particular, the Fifth Indian Economic Census provides the type of power (i.e. electricity, horsepower, etc.) that a establishment employs. We then classify each power source as either traditional (without power, fire wood, animal power, non conventional, and others) or modern (electricity, coal/sift coke, petrol/diesel/kerosen, and liquefied petroleum gas/natural gas). Figure 3 contrasts the employment share of modern establishments within a sector using this proxy (x-axis) with that one implied by the structural estimation of the model (y-axis). The figure shows a strong correlation between the two measures. In fact, for many sectors we see that the values are close to the 45 degree line. We note that sectors that will exhibit high development multipliers in our quantitative exercises are among those exhibiting good fit in this here. This is reassuring of the lack of relevance of the bad fit we observe in Figure 3 for some sectors. 24 Figure 3: The employment share of modern establishments, proxy vs. estimated 4.4 Sectoral revenue multipliers We begin our quantitative analysis of sectoral development multipliers by presenting estimates of revenue multipliers for all sectors, ϵr . Figure 4 presents our estimates, with 95% bootstrap confidence intervals. As seen in the figure, estimates are ’tight’, in the sense that confidence intervals are small and they do not seem to affect the way one should interpret the ranking of the multipliers. Regarding the estimates, the figure showcases large heterogeneity in multipliers across sectors. Some sectors are cost-effective, i.e. ϵr > 1, some sectors are cost ineffective, i.e. ϵr < 1, and even some sectors are such that it is counterproductive to subsidize them, i.e. ϵr < 0. Among cost-effective sectors, there are a few that stand out as key engines for development, in decreasing order of relevance: 11. M-Basic Metals, 15. M-Machinery and Equipment, 1. Mining, 13. M-Computer and Electronic, and 14. M-Electrical Equipment. Multipliers are the result of normalizing the revenue subsidy elasticity of consumption by the cost of the subsidy, measured by the Domar weight. The left panel of Figure 5 provides a scatter plot with the subsidy elasticity in the y-axis and the revenue multiplier in the x-axis, while the right panel presents a scatter plot with the Domar weight in the y-axis and the revenue multiplier in the x-axis. As it is clear from the figure, there is a strong positive relationship between the subsidy elasticity and the multiplier, and a strong negative relationship between the Domar weight and the multiplier. The figure shows that sectors with high multipliers are those with intermediate-to-high values for the subsidy elasticity and low values for the 25 Figure 4: Sectoral revenue multipliers, ϵr Domar weight. The M-Basic Metals sector falls in this category, exhibiting the highest multiplier, a subsidy elasticity that is high, but 30% lower than that one in the Trade sector, and a Domar weight that is very low, 85% lower than in the Trade sector. Overall, sectors with high multipliers exhibit intermediate values for consumption elasticity and, relative to sectors with large values for the elasticity, are disproportionately smaller, thus severely reducing the fiscal cost of the policy. The relevance of adoption margin as a key determinant of sectoral multipliers can be measured by comparing ϵr with ϵea r , where this last one is the multiplier that would be obtained after shutting down entry and adoption margins. The left panel of Figure 6 presents the multipliers for these two cases. The effect of the adoption margin is highly heterogeneous across sectors. In some key sectors, the adoption margin greatly amplifies the degree of the sectoral multiplier. In some sectors, the adoption margin turns the multiplier negative. And in other sectors, the adoption margin barely affects the level of the multiplier. Heterogeneity of the effect of adoption for multipliers across sectors manifests on the way sectors are ranked in terms of multipliers with and without technology adoption. The right panel of Figure 6 presents in the x-axis the ranking of the sector revenue development multiplier in the economy with no entry and adoption, i.e. the ranking using ϵea , and in the y-axis the ranking of the sector revenue development multiplier in the economy with entry and adoption, i.e. the ranking using ϵ. While there is little variation for those sectors considered less relevant in the economy with no adoption—i.e. those sectors with low 26 Figure 5: Components of sectoral revenue multipliers Subsidy elasticity of C Domar weight, Ψ̃ Figure 6: The role of adoption ϵr and ϵea r Ranking multipliers, there is substantial variation in rankings among those exhibiting intermediate and high values for multipliers. For example, the top 5 sectors abstracting from adoption are, in descending order of relevance, 11. M-Basic Metals, 1. Mining, 19. Utilities, 12. M-Metal Products, and 10. M-Other NonMetallic, while the top 5 sectors in the economy with adoption are, in descending order of relevance, 11. M-Basic Metals, 15. M-Machinery and Equipment, 1. Mining, 13. M-Computer and Electronic, and 14. M-Electrical Equipment. With a finite amount of resources to be allocated to industrial policy, this result suggests that accounting for the adoption margin is crucial for the policy. 27 4.4.1 Determinants of amplification through adoption Figure 6 showcases the relevance of the adoption margin for multipliers: through adoption, there is substantial amplification, and this amplification is heterogeneous across sectors. In this section we study the determinants of amplification through adoption. To do so, we study the case with no entry as it is simpler and, as shown in Table IV in the Appendix, multipliers barely change when the entry channel is considered. Using (20) together with the definition of a sectoral revenue multiplier, we define the contribution of adoption to the multiplier as h i η−1 ′ ′ −1 A= Ψ̃ − ⊘ Ψ′ , Ψ βdiag (M ) ∇a,ra I + ∇P Y,r + Λ (I − νΩ) η where ⊘ denotes element-by-element division. We exploit the following decomposition of A to study the determinants of amplification, direct incentive z }| { η−1 ζ ′ ′ A = Ψ̃ ⊘ Ψ − 1 βdiag (M ) η η−1 price of adoption, Pm z }| { η−1 ′ ζ ′ + Ψ̃ − Ψ βdiag (M ) Λ (I − νΩ)−1 ⊘ Ψ′ η η−1 feedback through adoption z }| { h i η−1 ′ ζ ′ + Ψ̃ − Ψ βdiag (M ) I + Λ (I − νΩ)−1 ∇a,ra − I ⊘ Ψ′ η η−1 aggregate demand }| { z η−1 ′ ′ + Ψ̃ − Ψ βdiag (M ) ∇a,ra ∇P Y,a ⊘ Ψ′ , η (24) The direct channel in A relates to Remark 2, and considers the direct effect of a subsidy on adoption, as showcased in (18), abstracting from heterogeneity that stems from ∇a,ra . To assess the average relevance of this channel we consider the case where M = m̄I, where m̄ is equal to the average equally-weighted modern share across sectors. In this case, the direct effect on adoption implies a change in sectoral TFP, (d ln Z/d ln a) (d ln a/d ln ra ) = β m̄ζ/(η − 1). For a particular sector s, =0.115 =0.038 z }| { <3 z }| { z}|{ ζ ζ direct incentive = ϵea + β m̄ . rs β m̄ η−1 η(η − 1) (25) From Table IV in the Appendix, the multiplier ϵea rs is lower than 3. Thus, the direct channel of adoption can generate amplification raging from 0.038 to 0.383. Given that amplification is larger than 2 for many sectors, 28 these simple calculations imply that the average contribution of the direct margin is relatively small.12 The second term in the decomposition measures the importance of the revenue subsidy in lowering the price adopting modern technologies, Pm . The third term in the decomposition of A accounts for the relevance of feedback effects in adoption, as captured by ∇a,ra − ζ/(η − 1)I. This object captures the Double-Leontieff inverse described in detail in Proposition 2. Sectors showcasing large values for this term are those for which an increase in adoption generates large increases in adoption in other sectors. Finally, the fourth term in the decomposition of A accounts for the relevance of the feedback effects through the aggregate demand channel. Table V in the Appendix shows each term in this decomposition. Consistent with the discussion above, the direct incentive term is small relative to A. Furthermore, the aggregate demand term is also small. As a result, we now focus on the main two terms: the price Pm and the feedback through adoption terms. Figure 7: Determinants of amplification Figure 7 presents the contributions of the Pm channel and the feedback in adoption channel. We also ψ̃rs Notice that the term ϵea rs = ψrs − 1 in (25) is the same as the key statistic in Liu (2019). Here, because of adoption, the key statistic is ’transformed’ by the elasticity of TFP with respect to the subsidy. 12 29 add A to the figure to aid in the analysis. The other terms in the decomposition are not included as they are small relative to the level of amplification observed through adoption. The figure shows that the channel through the price of adoption Pm accounts for a large part of amplification through adoption A. While the price of adoption channel is relevant for all sectors, the feedback channel is only substantial for sectors exhibiting high values of A. For example, for the top 5 sectors in terms of multipliers, the feedback channel accounts for 20 to 30% of amplification through adoption. What are the forces driving the relevance of the price of adoption channel? Specifically, what is the relative importance of centrality of a sector in the production Ω and investment Λ networks? To gain insights, we further decompose the expression for the Pm term, exploiting the identity Λ (I − νΩ)−1 =Ω (I − νΩ)−1 + (Λ − Ω) (I − νΩ)−1 . (26) This identity allows us to study the relevance of the two different input-output networks in shaping up the price of adoption term in the decomposition above. We decompose the price of adoption effect into two terms. The first term accounts for the Pm channel under the counterfactual assumption that the adoption good is produced with the same input-output structure as the intermediate aggregate, i.e., Ω. The second term in the decomposition measures the relative importance of a sector in the actual investment network Λ relative to Ω.13 Figure 8 reproduces the contribution of the price of adoption term to amplification, and the difference between this term and a counterfactual price of adoption term where we replace the actual investment network Λ with Ω (labeled Λ − Ω in the figure). A sector exhibiting a high value for this term is one that is more central in the investment than in the production network. Figure 8 shows that the investment network is important for explaining amplification for some key sectors. For example, while for sector 11. M-Basic Metals, this channel is muted, for sector 15. M-Machinery and Equipment centrality in Λ is instrumental for amplification. 4.5 Alternative policy instruments Which policy instrument is the most appropriate to promote economic development? The different instruments differ in (i) the way they improve production efficiency and promote the adoption of modern technologies, and (ii) the fiscal cost of implementation (for example, see Remark 4 in Section 3.2 and results in Section 3.3.1). As there are clear trade-offs among the different policy instruments, we resort to a quantitative analysis to understand which instrument is the most cost-effective for development. Figure 9 presents, for each sector, the resulting multiplier under the four alternative policy instruments: 13 In particular, the price of adoption channel in (24) is decompose as follows: η−1 ′ ζ η−1 ′ ζ ′ ′ Ψ̃ − Ψ βdiag (M ) Ω (I − νΩ)−1 ⊘ Ψ′ + Ψ̃ − Ψ βdiag (M ) (Λ − Ω) (I − νΩ)−1 ⊘ Ψ′ . η η−1 η η−1 30 Figure 8: Investment Λ or production Ω networks? revenue subsidies r, intermediate input subsidies r x , labor subsidies r l , and adoption subsidies r a . Next we discuss the main takeaways. First, among all instruments, labor subsidies rl are the least effective tool to promote development, as evident from the fact that, for all sectors, labor subsidies provide the lowest multipliers. This is expected, as labor is in fixed supply, and consistent with the results in Section 3.3.1. Second, revenue multipliers ϵr are always above labor multipliers ϵrl and below intermediate input multipliers ϵrx . While consistent with the results in 3.3.1, it is surprising that there is no sector where mutlipliers under revenue subsidies are above multipliers under intermediate inputs, given that revenue subsidies have a direct effect on the marginal adopters’ condition and intermediate input subsidies do not. Key for this result is that the cost of implementing a intermediate input subsidy is a fraction ν of the cost of implementing a revenue subsidy. Third, and most importantly, adoption subsidies appear to be the most cost-effective way to promote economic development, in spite from the fact that they do not have a direct effect on the production 31 Figure 9: Multipliers under alternative policy instruments efficiency channel, as evident from (20). Overall, among all sectors and among all policy instruments, the top 5 multipliers are obtained through implementing adoption subsidies in the following sectors (in descending order of relevance): 11. M-Basic Metals, 1. Mining, 15. M-Machinery and Equipment, 12. M-Metal Products, and 13. M-Computer and Electronic. What does it make adoption subsidies the most cost-effective policy instrument? To shed light on this question, we perform a decomposition along the lines we did in (24). In particular, we decompose the adoption multiplier into a direct effect and a feedback through the adoption channel, feedback through adoption ϵr a direct incentive }| { z }| { z ′ η − 1 ζ η − 1 η ζ ′ = Ψ̃ ⊘ Ψ′ − 1 + Ψ̃ − Ψ′ βdiag (M ) ∇a,ra − I ⊘ ∆′ . (27) η η−1η−1 η η−1 As before, simple calculations can be used to gauge an idea of the magnitude of the direct incentive effect in a particular sector s, 32 =2.37 direct incentive, ra = ϵea rs }| =0.79 }| { ζ ζ η η + . η − 1 η − 1 η − 1 η(η − 1) z { z Different from the case with revenue subsidies in (25), the direct incentive effect is guaranteed to be large with adoption subsidies: It is over two times the size of the multiplier in a model with no entry and adoption, ϵea rs , and it is bounded below by 0.79. The larger magnitude results from the fact that the fiscal cost of an adoption subsidy δs is substantially lower than the fiscal cost of a revenue subsidy ψs . In fact, δs = ((η − 1)/η)βMs ψs < ψs . Indeed, the difference in the fiscal cost is the only difference between the direct incentive effect in (24) relative to (27). Figure 10: Adoption multipliers Figure 10 presents the contribution of the direct incentive and feedback through adoption effects for all sectors. As suggested by the simple calculations for the direct incentive under adoption subsidies, the contribution of the direct incentive effects is now dominant but, as with revenue subsidies, the feedback 33 through adoption effect is still substantial for the top sectors in terms of the multiplier. 5 Concluding remarks This paper provides a theoretical and quantitative analysis of the role of sectoral industrial policies as a tool to foster economic development. We find that the technology adoption margin has important effects on the magnitude and relative ranking of sectoral development multipliers, i.e., the effect of a sectoral policy on aggregate consumption, relative to its fiscal cost. In addition, we show that technology adoption subsidies are the most cost-effective policy instrument. More generally, the analysis in this paper can be seen as illustrating a framework that can be used to analyze policies aimed at promoting the adoption of new technologies. While we applied the framework to study technology adoption for economic development, it can also be applied to alternative questions. For example, we view the case of the adoption of green technologies as a natural application. In the current paper, traditional and modern technologies use factors with the same intensity, a counterfactual assumption (Boehm and Oberfield, 2023). This assumption limits the complementarities in technology adoption, which have been shown to be quantitatively important by Buera et al. (2021). While we have abstracted from this margin in order to be as transparent as possible, we conjecture that the conclusions would be strengthened by incorporating this heterogeneity. We also abstracted from dynamics. As such, the analysis should be understood as describing the effect of policies across steady-states. A dynamic version of the analysis, using the methods in Alvarez et al. (2023), is a natural next step. References Acemoglu, D., Carvalho, V. M., Ozdaglar, A., and Tahbaz-Salehi, A. (2012). The Network Origins of Aggregate Fluctuations. Econometrica, 80(5):1977–2016. Alvarez, F. E., Argente, D., Lippi, F., Méndez, E., and Van Patten, D. (2023). Strategic complementarities in a dynamic model of technology adoption: P2p digital payments. Working Paper 31280, National Bureau of Economic Research. Atkeson, A. and Burstein, A. (2010). Innovation, Firm Dynamics, and International Trade. Journal of Political Economy, 118(3):433–484. Baqaee, D. and Farhi, E. (2021). Entry vs. rents: Aggregation with economies of scale. Working Paper 27140, UCLA. Baqaee, D. R. and Farhi, E. (2020). Productivity and Misallocation in General Equilibrium. The Quarterly Journal of Economics, 135(1):105–163. Bartelme, D. G., Costinot, A., Donaldson, D., and Rodrı́guez-Clare, A. (2019). The textbook case for industrial policy: Theory meets data. Working Paper 26193, National Bureau of Economic Research. 34 Bigio, S. and La’O, J. (2020). Distortions in Production Networks. The Quarterly Journal of Economics, 135(4):2187–2253. Boehm, J. and Oberfield, E. (2023). Growth and the Fragmentation of Production. Technical report, Princeton University. Broda, C. and Weinstein, D. E. (2006). Globalization and the Gains From Variety. The Quarterly Journal of Economics, 121(2):541–585. Buera, F. J., Hopenhayn, H., Shin, Y., and Trachter, N. (2021). Big Push in Distorted Economies. Nber working papers, National Bureau of Economic Research, Inc. Caliendo, L., Parro, F., and Tsyvinski, A. (2022). Distortions and the structure of the world economy. American Economic Journal: Macroeconomics, 14(4):274–308. Casal, L. and Caunedo, J. (2023). On the Investment Network and Economic Development. Unpublished manuscript. Caunedo, J. and Kala, N. (2021). Mechanizing agriculture. Working Paper 29061, National Bureau of Economic Research. Choi, J. and Levchenko, A. A. (2021). The long-term effects of industrial policy. Working Paper 29263, National Bureau of Economic Research. Crouzet, N., Gupta, A., and Mezzanotti, F. (2023). Shocks and technology adoption: Evidence from electronic payment systems. Journal of Political Economy, 131(11):3003–3065. Demir, B., Fieler, A. C., Xu, D. Y., and Yang, K. K. (2024). O-ring production networks. Journal of Political Economy, 132(1):200–247. Farrokhi, F., Lashkaripour, A., and Pellegrina, H. S. (2023). Trade and technology adoption in distorted economies. Working paper. Foerster, A. T., Hornstein, A., Sarte, P.-D. G., and Watson, M. W. (2022). Aggregate Implications of Changing Sectoral Trends. Journal of Political Economy, 130(12):3286–3333. Hirschman, A. (1958). The Strategy of Economic Development. Yale paperbound. Yale University Press. Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and Manufacturing TFP in China and India. Quarterly Journal of Economics, 124(4):1403–1448. Jones, B. F. and Liu, X. (2022). A framework for economic growth with capital-embodied technical change. Working Paper 30459, National Bureau of Economic Research. Liu, E. (2019). Industrial Policies in Production Networks. 134(4):1883–1948. The Quarterly Journal of Economics, Liu, E. and Ma, S. (2021). Innovation networks and rd allocation. Working Paper 29607, National Bureau of Economic Research. 35 Timmer, M., Dietzenbacher, E., Los, B., Stehrer, R., and de Vries, G. (2015). An illustrated user guide to the world input–output database: the case of global automotive production. Review of International Economics, 23(3):575–605. vom Lehn, C. and Winberry, T. (2022). The Investment Network, Sectoral Comovement, and the Changing U.S. Business Cycle. The Quarterly Journal of Economics, 137(1):387–433. 36 Appendix A Proofs and derivations A.1 Domar weights Multiplying both sides of (7) by Ps and using that Ps Cs = γs Pc C provides that γs Pc C + X ωs′ s Pxs′ Xs′ + s′ X λs′ s Pms′ as′ κms′ =Ps Ys , s′ and by using the intermediate input demand in (3), the expression relating Ps and Pxs in (10), and the definition of sector productivity in (9) we obtain γs Pc C + ν X η−1X ωs′ s rs′ Ps′ Ys′ + λs′ s Pms′ as′ κms′ =Ps Ys . η ′ ′ s (28) s When rs = 1 for all s, and by dividing by Pc C we obtain for each s that X Ps Ys η−1X P s ′ Ys ′ Pms′ as′ κms′ −ν ωs′ s =γs + λs′ s , Pc C η Pc C Pc C ′ ′ s s X Ps Ys η−1X P s ′ Ys ′ −ν ωs′ s =γs + λs′ s δs′ . Pc C η Pc C ′ ′ s s We can stack these equations and obtain an expression for the Domar weigths Ψ, η−1 ′ Ψ= I −ν Ω η −1 Γ + Λ′ ∆ . Or, expressed in row vectors, −1 η−1 Ψ = Γ +∆Λ I −ν Ω . η ′ A.2 ′ ′ Derivation of equation (16) We begin by noting that we can combine labor input demand in (2), teh expression relating Ps and Pxs in (10), the definition of sector productivity in (9) with the labor market clearing condition in (8) to obtain the following expression, L =(1 − ν) η−1X rs Ps Ys . η s 37 Log-differentiating this expression around r = 1, X Ps Ys (d ln rs + d ln Ps Ys ) =0 . (29) s Note how, by dividing by Pc C, this expression provides that Ψ′ [d ln r + d ln (P ◦ Y )] = 0. That is, weighted by Domar weights, the total change in subsidies is offset by the change in gross product. Log-differentiating the goods market clearing condition in (28) around r = 1 provides, γs Pc C (d ln Pc + d ln C) + ν η−1X ωs′ s Ps′ Ys′ (d ln rs′ + d ln Ps′ Ys′ ) η ′ s + X λs′ s Pms′ as′ κms′ (d ln Pms′ + d ln as′ ) = Ps Ys d ln Ps Ys . s′ Adding across all sectors s, using (29), and dividing by Pc C, d ln C = − d ln Pc − X Pms′ as′ κms′ s′ Pc C (d ln Pms′ + d ln as′ ) − X Ps Ys s Pc C d ln rs , or, in matrix form, d ln C = − d ln Pc − ∆′ (d ln Pm + d ln a) − Ψ′ d ln r Using that from (11) we obtain that d ln P = − (I − νΩ)−1 (d ln Z + d ln r), and using (12), d ln C = Γ′ + ∆′ Λ′ (I − νΩ)−1 (d ln Z + d ln r) − ∆′ d ln a − Ψ′ d ln r . ′ Finally, noting that Ψ̃ = [Γ′ + ∆′ Λ′ ] (I − νΩ)−1 , ′ ′ d ln C = Ψ̃ − Ψ′ d ln r + Ψ̃ d ln Z − ∆′ d ln a . (30) Combining (4), (6), and (10), and specializing to the case with only revenue subsidies, we obtain rs η − η−1 η−1 Aη−1 as ζ m − At Zsη−1 Ps Ys = Pms κms or rs η−1 βMs Ps Ys = Pms as κms . η Finally, writing the conditions in vector form, dividing both sides of the expression by Pc C and setting 38 r = 1, there is a tight connection between the GDP share of a sector, or Domar weight, and its adoption share of GDP, η−1 β diag (M ) Ψ = ∆ . η Applying this expression to (30) completes the derivation. A.3 The elasticity of adoption to revenue and adoption subsidies We begin by computing an expression for d ln (P ◦ Y ). To this end, we first rewrite (28) in matrix form and log-differentiate to get, η−1 ′ Ω diag (Ψ) d ln (P ◦ Y ) =Γ (d ln Pc + d ln C) I −ν η d ln Pm ′ + Λ diag (∆) + I d ln a d ln a d ln Pm η−1 ′ ′ + Λ diag (∆) +ν Ω diag (Ψ) d ln r . d ln r η Also, we follow the same steps as in Appendix A.2 to obtain an expression for d ln Pc + d ln C, ′ d ln Pc + d ln C = −∆ d ln Pm ′ d ln Pm + I d ln a − ∆ + Ψ d ln r . d ln a d ln r Combining these two expressions provide that d ln (P ◦ Y ) =∇P Y,a d ln a + ∇P Y,r d ln r , h i−1 ′ where ∇P Y,a = I − ν η−1 Ω diag (Ψ) Λ′ diag (∆) ddlnlnPam + I − Γ∆′ ddlnlnPam + I , and ∇P Y,r = η h i−1 h i η−1 ′ d ln Pm ′ ′ ′ d ln Pm I − ν η−1 Ω diag (Ψ) Λ diag (∆) + ν Ω diag (Ψ) − Γ ∆ + Ψ , and where, usη d ln r η d ln r ing (11) and (12), d ln Pm d ln a = −Λ (I − νΩ)−1 d ln Z d ln a , and d ln Pm d ln r = −Λ (I − νΩ)−1 . Having obtained an expression for d ln (P ◦ Y ), we can now solve for d ln a using the marginal adopters’ conditions, d ln r − (η − 1)d ln Z + d ln (P ◦ Y ) − η−1 d ln a = −d ln r a + d ln Pm . ζ Substituting the expression for d ln (P ◦ Y ) produced in (37) and rearranging, η−1 d ln Pm d ln Z d ln Pm I + (η − 1)I − − ∇P Y,a d ln a = I + ∇P Y,r + d ln r + d ln r a , ζ d ln Z d ln a d ln r 39 so that h i d ln a =∇a,ra I + ∇P Y,r + Λ (I − νΩ)−1 d ln r + ∇a,ra d ln r a , where −1 h i η−1 −1 d ln Z = , I + (η − 1)I − Λ (I − νΩ) − ∇P Y,a ζ d ln a ( " )−1 # Λ (I − νΩ)−1 ζ ζ I− = − I ζβdiag (M ) − ∇P Y,a . η−1 η−1 η−1 ∇a,ra A.4 The density hs (l) We are interested in Hs (l) = Pr (lis (z̃, ε̃) < l) = Pr (ln lis (z̃, ε̃) < ln l). Using (22) this reduces to Hs (l) = Pr ln Ãi + ln z̃s + ε̃ < ln l = Pr ε̃ < ln l − ln Ãi − ln z̃s , ! Z z̃ms Z ∞ ln l − ln Ãt − ln z̃ − µ̃ 1 1 −ζ̃−1 Φ Φ ζ̃z dz̃ + = −ζ̃ −ζ̃ χ̃ z̃ts z̃ms z̃ts z̃ts ln l − ln Ãm − ln z̃ − µ̃ χ̃ where Φ(·) denotes the CDF of a standard normal. To compute hs (l) recall that hs (l) = hs (l) = Z 1 ln l − ln Ãt − ln z̃ − µ̃ χ̃ ϕ −ζ̃ χ̃lz̃ts + z̃ms z̃ts Z 1 −ζ̃ χ̃lz̃ts ∞ ϕ z̃ms ! ζ̃z −ζ̃−1 dz̃ , ∂Hs (l) ∂l . Then, ! ζ̃z −ζ̃−1 dz̃ ln l − ln Ãm − ln z̃ − µ̃ χ̃ ! ζ̃z −ζ̃−1 dz̃ . Integrating provides the expression in the text provided in (23). A.5 Proof of Remark 5 Without loss of generality let Ãt = 0. Because all entrants are adopters of the modern technology, z̃ts = z̃ms . Then (23) reduces to hs (l) =l −ζ̃−1 ζ̃ −ζ̃ z̃ms 2 µ̃ζ̃+ χ̃2 e " Ãζ̃m 1−Φ ln z̃ms − ln l + µ̃ + χ̃2 ζ̃ χ̃ !# . Notice that " ∂hs (l) hs (l) 1 = haz ∂l l χ̃ ln z̃ms − ln l + µ̃ + χ̃2 ζ̃ χ̃ 40 ! − ζ̃ + 1 # , ϕ where haz (·) ≡ ln z̃ms −ln l+µ̃+χ̃2 ζ̃ χ̃ 2 1−Φ ln z̃ms −lnχ̃l+µ̃+χ̃ ζ̃ is the hazard rate of the normal distribution, where haz ′ > 0 and haz ′′ > 0. Because hs (l) ≥ 0 and the hazard rate decreasing and convex in l, the distribution Hs (l) has one mode. χ̃2 ζ̃ ≤ χ̃ ζ̃ + 1 then the mode is at l = z̃ms . Otherwise, the mode is at some l > z̃ms . If haz ln z̃ms +µ̃+ χ̃ For the employment share distribution, recall that gs (l) = R lhs (l) , l̂hs (l̂)dl̂ so that ∂gs (l) ∂l ∝ hs (l) + l ∂h∂ls (l) . Then, " ∂gs (l) gs (l) 1 = haz ∂l l χ̃ As a result, again we have that if haz ln z̃ms − ln l + µ̃ + χ̃2 ζ̃ χ̃ ln z̃ms +µ̃+χ̃2 ζ̃ χ̃ ! # − ζ̃ . ≤ χ̃ζ̃ then the mode is at l = z̃ms . Otherwise, the mode is at some l > z̃ms . A.6 Proof of Remark 6 When χ → 0 we also have that χ̃ → 0. Then, when χ → 0, the expression for hs (l) in (23) reduces to ζ̃ l−ζ̃−1 z̃ −ζ̃ ζ̃ Ãt ts hs (l) = 0 l−ζ̃−1 ζ̃ Ãζ̃m −ζ̃ z̃ts if z̃ts ≤ l ≤ z̃ms , if z̃ms ≤ l < Ãm z̃ms , (31) if l ≥ z̃ms . Because ζ̃ > 0 and Ãm > Ãt , Hs (l) has two modes: one at l = z̃ts , and one at l = z̃ms . Likewise, because gs (l) ∝ lhs (l), Gs (l) has the same two modes. Further, if Ãm = Ãt both distributions have a single mode, and the mode is at l = z̃ts . A.7 Proof of Proposition 3 The data is given by values of η, ν, Γ, Ω, and Λ, the normalization Ats = 1, and the reduced form estimates 1 ζ̃ = ζ η−1 , Ãms = Aη−1 ms , and the following reduce-form thresholds z̃ts = (1 − ν) η − 1 Ps Ys η−1 z , η Zsη−1 ts (32) z̃ms = (1 − ν) η − 1 Ps Ys η−1 z . η Zsη−1 ms (33) 41 zts zms The ratio of (32) and (33) provides of two equations relating −ζ zts , −ζ zms = z̃ts z̃ms 1 η−1 . Combining this expression with (33) provides a system and a sector’s gross value added Ps Ys , (1 − ν) η−1 ζ −ζ η−1 Ps Ys = z̃ms zms Ats η ζ − (η − 1) and −ζ zts = z̃ts z̃ms − z̃ts z̃ms ζ η−1 − ζ−(η−1) η−1 η−1 + Aη−1 , ms − Ats −ζ zms . We next derive a system of equations determining the vector of gross output, given values for η, ν, Ω, Λ, the η−1 η−1 productivity parameters Ams and Ats , and the value of the reduced form thresholds. The good market clearing condition in sector s is given by γs X X Pxs′ Pms′ −ζ Pc C+ ωs′ ,s Xs′ + λs′ ,s z ′ κms′ = Ys , Ps Ps Ps ms ′ ′ s where Xs = ν η−1 η s η ν w 1−ν )(η−1) Pxs (Pxs Ps Psη Ys (Zs )η−1 = ν η−1 η Pxs Ys , where we used that Ps = ν w 1−ν η Pxs . η−1 Zs Then, the goods market clearing condition in sector s can be expressed as η X X Ps′ Pms′ −ζ Ys′ + λs′ ,s z ′ κms′ =Ys . Ps Ps ms ′ ωs′ ,s s′ (34) s Similarly, we can express the labor market clearing condition as follows, (1 − ν) X η−1X −ζ Ps Ys = L − κts zts . η s s (35) We now turn back to the goods market clearing condition in (34). Summing over all sectors s, Pc C + X −ζ Pms zms κms = s η−1 1−ν η X Ps Ys , s and using (35), Pc C = 1 − ν η−1 η (1 − ν) η−1 η " # L− X −ζ κts zts − X s −ζ Pms zms κms . s Substituting back into (34) ( γs 1 − ν η−1 η (1 − ν) η−1 η " L− # X s −ζ zts κts − ) X −ζ zms Pms κms +ν s X η−1X −ζ ωs′ ,s Ps′ Ys′ + λs′ ,s zms Pms′ κms′ = Ps Ys . η ′ ′ s 42 s The values of κts and Pms κms can be obtained as function of the reduced form thresholds and productivity, given values for η and ν. In particular, using the marginal entrant’s condition, κts = η−1 1 1 η−1 Ps Ys zts 1 Ats Aη−1 z̃ts , = η−1 η 1 − ν η − 1 ts Zs where we used (32). Similarly, using the marginal adopter’s condition, 1 η−1 1 Ams − Aη−1 z̃ms . ts 1−νη−1 Pms κms = Thus, given values for η, ν, Γ, Ω, Λ, and reduced form estimates ζ̃, Ãms , z̃ts , and z̃ms , the vector of the value of sector gross output,Ps Ys , solves ( γs 1 − ν η−1 η (1 − ν) η−1 η " L− # X −ζ zts κts s ) − X −ζ zms Pms κms +ν s X η−1X −ζ ωs′ ,s Ps′ Ys′ + λs′ ,s zms Pms′ κms′ = Ps Ys , η ′ ′ s s where 1 1 Aη−1 z̃ts , 1 − ν η − 1 ts 1 η−1 1 z̃ms , Ams − Aη−1 = ts 1−νη−1 κts = Pms κms −ζ zms = (1 − ν) η−1 η Ps Ys # , ζ−(η−1) − η−1 η−1 z̃ts η−1 η−1 Ats + Ams − Ats z̃ms " ζ z̃ms ζ−(η−1) and −ζ zts = z̃ts z̃ms − ζ η−1 −ζ zms . Thus, given the structural thresholds, the price of the adoption good in sector s Pms and the adoption cost κms can be calculated. B Additional tables and figures In this section we collect additional results. B.1 Detailed description of sectors and cross-walk In this section we present a table showing how we harmonize the different dataset sources into the 30 sectors we use to describe the Indian economy. The sectors in the Indian Economic Census of 2013 uses the 43 National Industrial Classification NIC 2008, which is based in the ISIC Revision 4. The WIOD data uses International Standard Industrial Classification ISIC revision 4. These two industry classification coincide at 2-digit, which is the lowest level of aggregation in our analysis. Finally, the investment network data from vom Lehn and Winberry (2022) is provided using a 37-sector partition, which is based on NAICS industry codes. We harmonize the sectors at the 2-digit. The cross-walk used is shown in the last two columns of Table III. 44 Table III: Detailed Description of Sectors, and Cross-Walk across Datasets Short Name Long Description 1. Mining 2. M-Food 3. M-Textiles 4. M-Wood 5. 6. M-Paper M-Printing and media 7. M-Petroleum 8. M-Chemicals and pharma 9. M-Plastics 10. M-Other Non-Metallic 11. M-Basic Metals 12. M-Metal Products 13. M-Computer and Electronic 14. M-Electrical Equipment 15. M-Machinery and Equipment 16. M-Motor Vehicles 17. 18. M-Other Transport M-Other 19. Utilities 20. Construction 21. Trade 22. Transportation 23. Food and accomodation 24. Information Mining and quarrying Manuf. of food products, beverages, and tobacco products Manuf. of textiles, apparel, leather and related products Manuf. of wood and products of wood, cork and straw, except furniture Manuf. of paper and paper products Printing and reproduction of recorded media Manufacture of coke and refined petroleum products Manuf. of chemicals and chemica, pharmaceuticals, and botanical products Manuf. of rubber and plastics products Manuf. of other non-metallic mineral products Manuf. of basic metals Manuf. of fabricated metal products, except machinery and equipment Manuf. of computer, electronic and optical products Manuf. of electrical equipment Manuf., repair and instalation of machinery and equipment n.e.c., Manuf. of motor vehicles, trailers and semi-trailers Manuf. of other transport equipment Manuf. of furniture and other manufacturing Electricity, gas, and water supply; sewerage, wast collection and other waste management services Construction of buildings, civil engineering, and specialized construction activities Wholesale and retail trade Land, water and air transportation, including transport via pipelines; warehousing and support activities for transportation; post and courier activities Foor and accommodation service activities Publishing activities; motion picture, video and television programme production, sound recording and45music publishing activities; broadcasting and programming activities Codes NIC 2008/ ISIC Rev. 4 Vom Lehn & Winberry 05-09 1 10-12 15 13-15 16-17 16 4 17 18 18 19 19 20 20-21 21 22 22 23 5 24 6 25 7 26 9 27 10 28-33 8 29 11 30 31-32 12 13-14 35-39 2 41-43 3 45-47 23-24 49-53 25 55-56 35-36 58-61 26 Short Name 25. Professional Services 26. Finance and insurance 27. Real estate 28. Education 29. Health and social work 30. Other services Long Description Computer programming and consultancy; accounting, advertising, architecture, engineering,legal, management consultancy, and market research activities; activities of head offices; technical testing and analysis; scientific research and development; veterinary activities Financial service, insurance, reinsurance and pension funding, except compulsory social security Real estate, including rental and leasing, services to building and landscaping; employment activities, security and investigation, office administrative, office support and other business support activities Education Health, residential care, social work activities without accommodation Other services 46 Codes NIC 2008/ ISIC Rev. 4 Vom Lehn & Winberry 62-63, 69-75 29 64-66 27, 30 68, 77-82 28 85 32 86-89 33 90-96 37 B.2 Sectoral revenue multipliers Table IV produces the revenue development multipliers for all sectors. The first column provides the (backward) Domar weight or influence measure Ψ̃ and the second column provides the (forward) Domar weight Ψ—the size of the sector. In the spirit of Baqaee and Farhi (2020), the third column provides the development multipliers fixing entry and adoption, ϵea . The fourth column presents the multiplier fixing only entry, ϵe , and the fifth column presents the multiplier with both entry and adoption being active. The fact that there are no substantial differences between column 4 and 5 provides reassurances of our analysis abstracting from the entry margin in Section 3. B.3 Decomposition of amplification under revenue multipliers In this section we present the results of the decomposition of the amplification term, as described in Section 4.4.1. We perform the decomposition abstracting from the entry channel, as accounting for this channel is barely relevant for multipliers as evident in Table IV. Each column of Table V accounts for a term in the decomposition of A. 47 Table IV: Sectoral (revenue) development multipliers Division 1. Mining 2. M-Food 3. M-Textiles 4. M-Wood 5. M-Paper 6. M-Printing and Media 7. M-Petroleum 8. M-Chemicals and Pharma. 9. M-Plastics 10. M-Other Non-Metallic 11. M-Basic Metals 12. M-Metal Products 13. M-Computer and Electronic 14. M-Electrical Equipment 15. M-Machinery and Equipment 16. M-Motor Vehicles 17. M-Other Transport 18. M-Furniture and Other 19. Utilities 20. Construction 21. Trade 22. Transportation 23. Food and Accommodation 24. Information 25. Professional Services 26. Finance and Insurance 27. Real Estate 28. Education 29. Health and Social Work 30. Other Services 48 Ψ̃ Ψ ϵea r ϵer ϵr 0.16 0.30 0.16 0.02 0.05 0.04 0.29 0.17 0.06 0.04 0.15 0.08 0.07 0.05 0.07 0.08 0.02 0.08 0.18 0.23 0.47 0.47 0.21 0.09 0.07 0.37 0.14 0.10 0.05 0.08 0.04 0.20 0.11 0.01 0.02 0.02 0.12 0.07 0.03 0.01 0.04 0.03 0.03 0.02 0.03 0.05 0.01 0.04 0.06 0.11 0.25 0.24 0.11 0.04 0.05 0.15 0.13 0.09 0.05 0.04 2.67 0.48 0.43 1.27 1.76 1.22 1.48 1.50 1.43 1.85 2.68 1.91 1.59 1.79 1.16 0.76 1.25 1.11 1.91 1.12 0.90 0.97 0.94 1.42 0.50 1.43 0.06 0.09 0.10 0.76 4.54 0.54 0.48 1.77 2.83 1.68 2.43 2.39 2.42 3.55 5.70 3.87 4.27 4.11 4.63 2.20 4.07 1.76 3.35 2.28 1.36 1.45 1.23 2.64 3.95 2.44 0.04 0.17 0.13 0.99 4.57 0.45 0.38 1.91 2.77 1.78 2.44 2.33 2.33 3.52 5.71 4.03 4.22 4.06 4.60 2.13 3.95 1.90 3.36 2.36 1.43 1.46 1.26 2.88 3.80 2.39 -0.06 0.10 -0.00 1.02 Table V: Decomposition of A Division Direct Pm Feedback Demand 0.45 0.07 0.06 0.00 0.47 0.01 0.29 0.37 0.40 0.42 0.53 0.20 0.35 0.39 0.20 0.23 0.31 0.00 0.35 0.10 0.00 0.07 0.00 0.00 0.12 0.40 0.03 0.10 0.00 0.00 1.18 0.12 0.12 0.53 0.57 0.44 0.51 0.49 0.54 1.07 1.95 1.35 1.73 1.47 2.48 0.93 1.91 0.65 0.89 0.88 0.47 0.40 0.40 1.04 2.50 0.59 0.04 0.09 0.12 0.31 0.39 0.05 0.05 0.15 0.15 0.13 0.17 0.15 0.15 0.38 0.72 0.48 0.74 0.55 0.98 0.34 0.78 0.18 0.32 0.30 0.13 0.13 0.12 0.28 1.17 0.18 0.02 0.04 0.04 0.09 -0.16 -0.18 -0.17 -0.19 -0.13 -0.12 -0.03 -0.11 -0.11 -0.16 -0.17 -0.06 -0.14 -0.09 -0.17 -0.06 -0.12 -0.18 -0.12 -0.13 -0.14 -0.13 -0.22 -0.09 -0.32 -0.16 -0.11 -0.14 -0.13 -0.16 1. Mining 2. M-Food 3. M-Textiles 4. M-Wood 5. M-Paper 6. M-Printing and Media 7. M-Petroleum 8. M-Chemicals and Pharma. 9. M-Plastics 10. M-Other Non-Metallic 11. M-Basic Metals 12. M-Metal Products 13. M-Computer and Electronic 14. M-Electrical Equipment 15. M-Machinery and Equipment 16. M-Motor Vehicles 17. M-Other Transport 18. M-Furniture and Other 19. Utilities 20. Construction 21. Trade 22. Transportation 23. Food and Accommodation 24. Information 25. Professional Services 26. Finance and Insurance 27. Real Estate 28. Education 29. Health and Social Work 30. Other Services 49 C Online Appendix - Not for publication C.1 The elasticity of aggregate consumption to revenue and adoption subsidies In this Appendix we derive an expression for the elasticity of aggregate consumption with respect to revenue and adoption subsidies where both entry and technology adoption decisions are considered. We begin by expanding the expression in (16), ′ ′ = βdiag (M ), d ln Z d ln e d ln C = Ψ̃ − Ψ where d ln Z d ln a η−1 ′ d ln r + Ψ̃ − Ψ η ′ d ln Z d ln Z d ln e + d ln a , d ln e d ln a = βe I − βdiag (M ) and βe ≡ 1 ζ−(η−1) . η−1 ζ (36) Obtaining an expression for the elasticity of aggregate consumption to revenue and adoption subsidies requires us to solve for d ln e and d ln a. We begin by computing an expression for d ln (P ◦ Y ). To this end, we first rewrite (28) in matrix form and log-differente to get, η−1 ′ I −ν Ω diag (Ψ) d ln (P ◦ Y ) =Γ (d ln Pc + d ln C) η d ln Pm ′ + Λ diag (∆) + I d ln a d ln a d ln Pm ′ + Λ diag (∆) d ln e d ln e d ln Pm η−1 ′ ′ + Λ diag (∆) +ν Ω diag (Ψ) d ln r . d ln r η Also, we follow the same steps as in Appendix A.2 to obtain an expression for d ln Pc + d ln C, d ln Pc + d ln C = −∆ " ′ d ln Pm + I d ln a d ln a ′ d ln Pm − ∆ − ∆′ d ln e d ln Pm d ln r where ∆e ≡ κt ◦e/(Pc C). Also, using (11) and (12), and d ln Pm d ln r + 1 − ν η−1 η (1 − ν) η−1 η + Ψ d ln r , d ln Pm d ln e # ∆e ′ d ln e = −Λ (I − νΩ)−1 d ln Z d ln Pm d ln e , d ln a = −Λ (I − νΩ)−1 = −Λ (I − νΩ)−1 . Combining these expressions provide d ln (P ◦ Y ) =∇P Y,a d ln a + ∇P Y,e d ln e + ∇P Y,r d ln r , 50 (37) d ln Z d ln a , where −1 ′ d ln Pm η−1 ′ I −ν Ω diag (Ψ) Λ diag (∆) − Λ∆′ +I , η d ln a # −1 " η−1 1 − ν η−1 ′ d ln P η m = I −ν Λ′ diag (∆) − Γ ∆′ ∆e ′ , Ω diag (Ψ) −Γ η d ln e (1 − ν) η−1 η −1 η−1 ′ η−1 ′ ′ ′ d ln Pm ′ = I −ν Ω diag (Ψ) Λ diag (∆) − Γ∆ +ν Ω diag (Ψ) − ΓΨ . η d ln r η ∇P Y,a = ∇P Y,e ∇P Y,r Having obtained an expression for d ln (P ◦ Y ), we can now solve for d ln e and d ln a using the marginal entrant and marginal adopters’ conditions, η−1 d ln e = 0 , ζ η−1 d ln r − (η − 1)d ln Z + d ln (P ◦ Y ) − d ln a = −d ln r a + d ln Pm . ζ d ln r − (η − 1)d ln Z + d ln (P ◦ Y ) − Substituting the expression for d ln (P ◦ Y ) produced in (37) and rearranging, ≡∇E,e z ≡∇E,a ≡∇E,r }| }| z { { z }| { η−1 d ln Z d ln Z I + (η − 1) − ∇P Y,e d ln e + (η − 1) − ∇P Y,a d ln a = (I + ∇P Y,r ) d ln r , ζ d ln e d ln a ≡∇A,a ≡∇A,e z }| z }| A,r { { z ≡∇ }| { d ln Z η−1 d ln Z (η − 1) − ∇P Y,e d ln e + I + (η − 1) − ∇P Y,a = (I + ∇P Y,r ) d ln r + d ln r a . d ln e ζ d ln a Notice that these equations are linear in the vectors d ln e and d ln a, and so there is a unique solution to the system. Stacking these equations in matrix form and solving for these two vectors provide, d ln e ! d ln a = ∇E,e ∇E,a !−1 " ∇A,e ∇A,a ∇E,r ∇A,r ! d ln r + 0 I ! # d ln r a . Using these expressions in (C.1) imply that we fully specified the elasticity of aggregate consumption with respect to revenue, r, and adoption subsidies, r a . C.2 Model fit, by sector 51 Figure 11: Employment share distributions, data vs. model 1 - Mining 2 - Food, beverages & Tabacco 3 - Textiles, wearing & Leather 4 - Wood and Cork 5 - Paper & Paper Products 6 - Printing & Media 7 - Coke & Petroleum 8 - Chemicals & Pharma. Products 9 - Rubber & Plastic Products 10 - Other Non-metallic Mineral 11 - Basic Metals 12 - Metal Products, except M&E 52 13 - Computer, Electronic & Optical 14 - Electrical Equip. 15 - Machinery & Equipment 16- Motor Vehicles 17 - Other Transport Equip. 18 - Furniture & Other 19 - Utilities 20 - Construction 21 - Trade 22 - Transportation 23 - Accomodation & Food Serv. 24 - Information 53 25 - Professional Services 26 - Finance & Insurance 27 - Real Estate 28- Education 29 - Human Heatlh & Social Work 30 - Repair & Other Serv. 54