Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Firm Innovation and Financial Analysis: How Do They Interact? Jim Goldman∗ Joel Peress∗∗ 2 March 2015 ABSTRACT Entrepreneurs innovate more when financiers are better informed about their projects because they expect to receive more funding should their projects be successful. Conversely, financiers collect more information about projects when entrepreneurs innovate more because the opportunity cost of misinvesting, i.e. of missing out on successful projects, is higher. Thus, knowledge about technologies and technological knowledge are mutually reinforcing. We report evidence consistent with this mutual influence using two natural experiments that changed, respectively, the innovation incentives and the information environment for U.S. listed firms. We also estimate that a policy designed to stimulate innovation in these firms has an indirect effect through investors’ learning incentives that can account for a large proportion of the total effct of the policy on firms’ innovation effort. JEL classification codes : G20, O31, O4 Keywords: Financial development, growth, technological progress, innovation, capital allocation, learning. INSEAD, Department of Finance, Boulevard de Constance, 77305 Fontainebleau Cedex, France. ∗∗ CEPR and INSEAD, Department of Finance, Boulevard de Constance, 77305 Fontainebleau Cedex, France. Telephone: (33)-1 6072 4035. Fax: (33)-1 6074 5500. E-mail: [email protected]. 1 Introduction This paper studies the interplay between innovation carried out by firms (“Main Street research”) and the analysis investors make of its prospects (“Wall Street research”). The interplay we have in mind is the following. Entrepreneurs innovate more when financiers are better informed about the profitability of their projects because they expect to receive more capital should their projects be successful. Conversely, financiers collect more information about projects when entrepreneurs innovate more because the opportunity cost of misinvesting, i.e. of funding unsuccessful projects while missing out on successful projects, is higher. Thus, knowledge about technologies and technological knowledge are mutually reinforcing. We develop a dynamic model formalizing this insight, in which financial development and technological innovation feed on each other, thereby promoting economic growth through total factor productivity (TFP) growth. We then bring the model to the data, test whether such an interaction indeed is at work in reality and measure its economic magnitude. The model features competitive rational agents who conceive risky projects, learn about their prospects and invest in them. Costs have to be incurred both for innovating (what we call “research”) and for learning. Unlike other papers (the related literature is discussed below), the positive feedback between research and learning is not a consequence of risk sharing since risk is fully diversified away, nor moral hazard since efforts are contractible. Instead, it follows from the complementarity between productivity and capital. Expressing output as Y = AK where A and K denote respectively a project’s uncertain return and the amount of capital it attracts, shows that the return on financiers’ funds increases with A (every unit of capital yields a larger payoff) while the reward for research rises with K (an invention can be applied on a larger scale). The complementarity between A and K leads to a complementarity between research and learning. The model is consistent with stylized facts about the link between finance and growth (discussed in Section 2 below), to the extent that financial development is positively correlated with the quality of financiers’ knowledge 1 about projects.1 Better-informed agents—who live in economies that are more financially developed—invest capital more efficiently, thus stimulating research and TFP. This in turn makes them wealthier and more willing to learn about investment opportunities. Per capita income grows and its growth is entirely driven by TFP growth. We quantify the contributions of learning and research and their interplay, and show that the growth rate of income in the economy with both learning and research is larger than the sum of the growth rates in the no-learning and in the no-research economy—once again a reflection of the positive feedback between learning and research. This implies for example that an economy that would converge to a steady-state when learning and research do not interact can experience unbounded growth once they do. We evaluate empirically the model’s main predictions regarding the interplay between learning and technological innovation with two quasi-natural experiments affecting US listed firms. Specifically, the model predicts that (i) financiers learn more when entrepreneurs perform more research, and that (ii) entrepreneurs perform more research when financiers learn more. Exploiting plausibly exogenous shocks to the amount of resources (i) firms’ devote to innovating, and (ii) the financial sector allocates analysing firms, we present supporting evidence for those predictions. Empirically assessing these relationships requires proxies for research and learning, as well as a methodology to address the endogeneity bias that this two-way relationship generates.2 We measure firms’ research effort by their R&D expenditures, and financiers’ learning intensity about a firm as the number of financial analysts who follow the firm. To address the endogeneity of the relationships we resort to difference-in-difference estimates from two natural experiments that respectively shift firms’ innovation effort and financial sector learning between 1992 and 2006. The first experiment exploits the staggered implementation of R&D tax credits by US states over that period. After confirming the beneficial effect of these tax incentives to innovation, we show that, following their 1 To keep the model parsimonious, we do not model the financial sector explicitly. But we interpret the amount of resources devoted to analyzing investment opportunities as a proxy for the degree of financial development. 2 In the case of prediction (i) for example, a least squares regression of learning on research yields inconsistent estimates because the regression’s residual is correlated with the regressor — research, as implied by the second prediction. 2 passage, analysts significantly increase their coverage of firms subject to the tax credits compared to other firms in the country. Our second experiment uses the identification strategy pioneered by Hong and Kacperczyk (2010) and extended by other authors like Kelly and Ljungqvist (2012) and Derrien and Kecskes (2013)). These last authors consider closure of and mergers between brokerage houses which lead to the dismissal of analysts, and provide evidence that the resulting drop in analyst coverage is largely exogenous to firms’ policies. We document that firms that lose analysts as a result of broker events significantly reduce their R&D expenditures relative to unaffected firms. We find the feedback effect to be economically important in our sample of firms. Specifically, the indirect effect of an R&D tax credit, operating through analysts’ response is about the third of the size of its direct effect. For example, a 1% increase in R&D expenditures triggered by an R&D tax credit increases analyst coverage by around 6%, which in turn is responsible for up to 0.36% of the total increase in R&D expenditures. The model yields important insights on the effectiveness of policies aimed at promoting innovations (e.g., research subsidies or tax breaks). First, it suggests that such policies have a multiplier effect thanks to the induced improvement in capital efficiency. For example, an increase in R&D expenditures triggered by an R&D tax credit leads to more accurate financial analysis, which further stimulates R&D. Given our aforementioned estimates, the observed increase in R&D expenditures is for about two thirds attributable to the direct effect of the tax credit, and for one third attributable to the indirect effect of enhanced learning by the financial sector. Second, policies based on R&D incentives can be rendered more effective by coupling them with policies designed at increasing capital efficiency, such as reducing impediments to trading financial assets or improving accounting standards. We are not the first to emphasize the interaction between financial development and incentives to innovate under imperfect information. In Bhattacharya and Chiesa (1995), De la Fuente and Marin (1996), Acemoglu and Zilibotti (1999) and Acemoglu, Aghion and Zilibotti (2006), financiers supply capital to entrepreneurs whose effort they can only monitor at a cost.3 We assume away these moral hazard problems – information 3 Another strand of research centers on the adverse selection and signalling problems which occur when financiers do not observe 3 is symmetric in our setting, and focus instead on the evaluation of projects. In the model, it is the expectation that capital will be efficiently allocated that encourages innovation. Thus, the information that matters is produced ex ante rather than ex post. Both kinds of information play an important role in practice. The paper proceeds as follows. Section 2 presents the model. Section 3 discusses the empirical strategy. Section 4 describes the data. Section 5 displays the empirical results, and Section 6 concludes. The appendix contains proofs and extensions of the model. 2 The Model A simple model describes the interaction between technological innovations and investors’ information about them. It yields novel predictions, which are then tested using quasi-natural experiments, and allows to quantify this interaction. The model is motivated by the following empirical findings (see the appendix for more details and references). 1. Financial development promotes economic growth and its effect runs through TFP growth, rather than capital growth. The subsequent three observations shed light on the link from finance to TFP. 2. Financial development stimulates investments in R&D and R&D contributes to TFP. 3. Financial development also enhances TFP by improving capital efficiency. Countries with more developed financial sectors allocate capital more efficiently across industries and firms. A more efficient distribution of capital at the micro level translates into higher TFP at the macro level. 4. Financial development improves capital efficiency (among other ways) by alleviating informational problems. entrepreneurs’ ability (e.g. Bhattacharya and Ritter (1983), King and Levine (1993), Ueda (2004), Aghion, Howitt and MayerFoulkes (2005)). 4 The model is consistent with these observations. Its main results are presented in the text while the derivations are left for the appendix, together with some extensions. 2.1 Setup The economy is composed of two sectors — a final and an intermediate goods sector, and two types of agents: entrepreneurs who conceive the projects that compose the intermediate sector, and financiers who invest in them. 2.1.1 Agents The economy is populated by overlapping generations of risk neutral agents who live for two periods. There is no population growth. Agents only consume when old. Each generation consists of a representative entrepreneur (‘she’) and a representative financier (‘he’). The Entrepreneur. The entrepreneur creates the technologies that produce intermediate goods. In period t, she conceives a continuum of projects with unit mass indexed by n ∈ [0, 1]. Their output is determined by n ≡ A nt Ktn , where Y n is the quantity of intermediate good a technology with constant returns to capital, Yt+1 t+1 nt is its random productivity (or return) and produced in period t + 1 by project n net of capital depreciation, A nt independent Ktn is the amount of capital invested in the project. Projects are independent from one another (A m of A t for any m = n). They are liquidated immediately after production. Projects succeed with a 0.5 probability. Successful projects yield a return At , while unsuccessful projects yield At where At > At > 0. The entrepreneur sets projects’ return in case of success and failure, At and At , but has no influence on their probability of success. Creating a continuum of independent projects with returns A and A requires a research cost eA (A + A) where eA is continuous, increasing, convex, eA (0) = e′A (0) = 0 and e′A (+∞) = +∞, and where εA (A) ≡ 1 + Ae′′A (A)/e′A (A) > 1 denotes one plus the elasticity of e′A with respect to A. Under this formulation, returns in case of success and failure are perfect substitutes in terms of their cost. Moreover, there are no technological spillovers across generations, i.e. a generation’s technological achievements does not make it easier for the next generation to innovate. We refer to At and At as the research effort, or 5 productivity. The entrepreneur raises the capital required to operate her technologies from the financier. The Financier. The financier is employed in the final good sector, to which he supplies one unit of labor inelastically for a competitive wage wt . He invests his labor income in the projects set up by the entrepreneur. He allocates Ktn units of capital to project n in period t. At the time of investment, the financier does not know which projects will succeed. Instead, he receives an imperfect signal St that identifies the successful projects. The signal is right, i.e. accurately reveals successful projects, with a probability qt but wrong with a probability 1−qt . Observing a signal of precision q requires a learning cost eq (q), where eq is continuous, increasing, convex, eq (1/2) = e′q (1/2) = 0 and e′q (1) = +∞, and where εq (q) ≡ 1 + qe′′q (q)/e′q (q) > 1 denotes one plus the elasticity of e′q with respect to q. q = 1/2 corresponds to an uninformative signal, and q = 1 to a perfect signal. Unlike research, learning does not affect projects’ productivity. Instead, it allows to match capital with projects more efficiently. We refer to qt as the learning effort or the precision of information. 2.1.2 Technologies The economy is composed of two competitive sectors, a final and an intermediate good sector. The intermediate goods sector is made up of the projects conceived by the entrepreneur and funded by the financier. The intermediate goods are used as inputs in the production of the final good, according to a riskless technology, Gt ≡ Lt1−β Yt β , where Gt is final output (used as the numeraire), Lt is labor, Yt is the employment of intermediate good and 0 < β < 1 is the factor share of the intermediate good in the production of the final good. Many identical firms compete in the final good sector and aggregate to one representative firm. 2.1.3 Timing The timeline is depicted in Figure 1. An agent lives one period as a young agent (either as an entrepreneur or as a financier) and one period as an old agent (as a consumer). At the start of a period, the entrepreneur and the financier determine their research and learning efforts. After earning a wage and observing his signal, the financier distributes his wage across the projects. In the following period, the young become old, the successful projects are revealed, final goods are produced and agents consume their share of profits. 6 2.1.4 Discussion An assumption central to the model is that effort levels (in research and learning) are contractible. That is, they are determined ex ante cooperatively by the entrepreneur and the financier. As a result, the first-best outcome is achieved. Incomplete information rather than asymmetric information drives the findings of the paper.4 Many other assumptions can be relaxed without significantly altering our findings. First, risk neutrality is not essential and our results extend to any increasing concave utility function since there is no aggregate risk. Second, the assumption that the production of intermediary goods displays constant returns to scale makes our findings starker but is not crucial—our main findings continue to hold under decreasing and increasing returns to scale. Third, we can assume without loss of generality that the research and learning efforts are identical across projects as long as intermediate technologies are linear in capital and risk is fully diversified away. Their cost for the entire continuum is therefore adequately represented by eA (At +At ) and eq (qt ).5 Fourth, the assumption e′q (1/2) = 0 can be replaced with e′q (1/2) > 0, thus creating a no-learning regime for low levels of income, without altering investors’ learning decisions outside of this income range, as Appendix C.2. demonstrates. Finally, our findings obtain if, instead of controlling projects’ productivity, the entrepreneur sets their probability of success, as shown in Appendix C.1. They also obtain under a more general cost structure, eA (A, A), as long as eA is increasing and convex in each variable. 2.2 Equilibrium Three conditions define an equilibrium. 1. Market clearing in the intermediate goods sector. Final goods producers maximize their profit. Since labor 4 Initially, i.e. at the time they choose their efforts, the entrepreneur and the financier are equally ignorant about which projects will be successful. While information asymmetries may also be present or emerge later, the focus of the paper is on the interplay between ex ante efforts during the initial contracting phase. Moreover, the fact that efforts are contractible implies that multiple equilibria do not arise. n 5 n That is At = At , An t = At and qt = qt for all n ∈ [0, 1]. Indeed, the expected profit conditional on the signal is linear in the amount invested in each project so only projects that receive a success signal and have the highest expected conditional return, n n n n n E(A t | St ) = qt At + (1 − qt )At , are funded (problem 1 below). It follows that, given a set of projects to develop and evaluate, agents choose the same learning and research efforts across all projects (problem 2 below). If the financier learns about a subset of projects only, then the entrepreneur only researches projects in that subset and we simply assume these projects comprise the initial set of projects. The important assumption is that a large enough number of projects is conceived so that risk can be diversified away. Acemoglu and Zilibotti (1997) study what determines the optimal number of projects to develop. 7 and intermediate goods trade in competitive markets and aggregate labor supply equals one, the equilibrium factor prices in period t + 1 are wt+1 = (1 − β)(Yt+1 )β and ρt+1 = β(Yt+1 )β−1 , where ρt+1 denotes the price of intermediate goods in period t + 1, and Yt+1 = n n n At Kt sums up output over all projects. Since there is no aggregate risk in this economy, wt+1 , ρt+1 and Yt+1 are deterministic. 2. Capital allocation. After observing his signal St , the financier distributes his capital wt across the projects offered by the entrepreneur to maximize total expected profits, taking ρt+1 , At , At and qt as given: π(qt , At , At , wt ) ≡ max {Ktn }n∈[0,1] n n E ρt+1 At Kt | St subject to n n Ktn = wt . (1) As with output, πt is deterministic. 3. First-best effort levels. The entrepreneur and the financier determine cooperatively their effort levels (before the signal St is observed) to maximize the ex ante total surplus, which equals the expected profit net of effort costs, taking the price of intermediate goods ρt+1 as given (the model is agnostic about how this surplus is shared between the entrepreneur and the financier): max π(qt , At , At , wt ) − eq (qt ) − eA (At + At ). (2) qt ,At ,At The following proposition characterizes the equilibrium. P 1 : In equilibrium, the learning and research efforts, qt , At and At , are the unique solutions to the following system of equations: βqtβ wtβ 1−β At = e′A (At ) and At = 0 (3) β and βAt wtβ qt1−β = e′q (qt ) (4) The perfect substitutability of At and At in the cost of research (eA is a function of their sum) implies that the entrepreneur prefers to concentrate her effort on improving returns in case of success, setting At to 0. Equation 3 implies that the research effort At increases with the learning effort qt holding income fixed. Indeed, research is promoted when the financier is better-informed because the entrepreneur knows that she will receive 8 more funds for her successful projects. Conversely, equation 4 shows that the learning effort qt increases with the research effort At holding income fixed. Intuitively, a higher research effort encourages the financier to learn by magnifying the opportunity cost of misinvesting, i.e. of funding unsuccessful projects while missing out on successful projects. Thus, knowledge about technologies and technological knowledge are mutually reinforcing. The positive feedback between research and learning follows from the complementarity between productivity nt and capital Ktn in the production of intermediate goods. Since Y n ≡ A nt Ktn , the return on the financier’s A t+1 nt . Similarly, the reward nt because every unit of capital is more productive the larger A funds increases with A for innovating rises with Ktn because an invention is applied on a larger scale. Thus, the complementarity nt and Ktn leads to a complementarity between learning and research. Formally, the average quantity between A of goods produced by a project can be broken down into the contributions of projects’ average productivity, nt ) = (At + At )/2, of the average stock of capital per project, E(Ktn ) = wt , and of the quality of the match E(A nt , Ktn ) = (qt − 1/2)(At − At )wt : between projects and capital, captured by cov(A nt )E(Ktn ) + cov(A nt , Ktn ) = [qt At + (1 − qt )At ]wt = qt At wt . nt Ktn ) = E(A E(A Thus, average output increases with the product of the precision of the financier’s information qt with the return on the successful project At . 2.3 Dynamics We discuss next the dynamics of income, research and learning and their interplay along the economy’s growth path. 9 2.3.1 The Sources of Growth Income in period t+1, wt+1 , is related to the precision of information qt and productivity At in period t through the following equation: β wt+1 = (1 − β)Yt+1 = (1 − β) ∗ wtβ ∗ T F Pq (qt ) ∗ T F PA (At ) (5) where T F PA (At ) ≡ β At and T F Pq (qt ) ≡ qtβ . Together with equations 3 and 4 and an initial income level w0 , equation 5 describes fully the dynamics of the economy. It identifies the three forces that determine income growth. First, current income matters for next period’s income in the usual neoclassical way because it determines the aggregate capital stock: the marginal product of labor increases with current income but at a declining rate (the wtβ term). The other two forces operate through TFP, and can be broken down into the impact of research, T F PA , and learning, T F Pq . The first component, T F PA , is the focus of the endogenous growth literature, which acknowledges that technology can be improved by purposeful activity, such as R&D. In this context, growth stems from a scale effect, in that an expansion of either the size of the population, or the stocks of knowledge, physical or human capital improves the incentive to innovate (e.g., Jones (2005)). In our framework, this channel can be identified by freezing qt . Our model relates these incentives to the state of the financial sector, namely the quality of investment knowledge qt . The second component, T F Pq , is the focus of the financial development literature, which highlights the role of frictions that reduce the efficiency of investments. Examples include information limitations (e.g. Greenwood and Jovanovic (1990)) and investment indivisibilities (e.g. Acemoglu and Zilibotti (1997)). Since the cost of alleviating such frictions is (to a first approximation) unrelated to the size of the capital stock unlike its benefit, richer economies are less constrained and invest more efficiently. This effect can be identified in the model by freezing At . An alternative view of our model is that it shows how the incentive to mitigate investment inefficiencies depends on the level of the technology, i.e. how qt depends on At . Importantly, since learning and research influence each other, they each make both direct and indirect 10 contributions to economic growth. To capture the total effect of learning, one needs to take into account its positive influence on entrepreneurs’ incentive to innovate, i.e., its impact on T F PA . Conversely, the full benefit of research consists of its direct effect through T F PA and its indirect effect through T F Pq . This point has important implications for the effectiveness of policies aimed at stimulating innovations. First, it suggests that innovation policies, such as research subsidies or tax breaks, have a multiplier effect thanks to the improvement in capital efficiency. Second, innovations are also encouraged by policies designed at increasing capital efficiency, such as removing impediments to trading financial assets or improving accounting standards. 2.3.2 Dynamics A steady-state equilibrium satisfies equations 3, 4 and 5 together with the condition wt+1 = wt . A trivial solution to this system obtains when income equals zero and neither learning nor research take place (wt = At = 0 and qt = 1/2). A non-trivial steady-states also exists if 1 1 1 + = − 1 for all A > 0 and 1 > q > 1/2. β εA (A) εq (q) (6) ∗ We label with a ∗ steady-state quantities, denote ε∗A ≡ εA (A ) and ε∗q ≡ εq (q ∗ ), and define γ ≡ d ln(wt+1 /wt )/d ln(wt ) as the growth rate of income. The following proposition characterizes the dynamics of the economy. P 2 : Assume condition 6 holds. The economy admits two steady-state equilibria, 0 and w∗ > 0. The dynamics of the system in a neighborhood of the steady-state are governed by the following equations: ln(At ) ≈ and β [ln(qt ) + ln(wt )] + cst, ε∗A − β ln(qt ) ≈ ε∗q β ln(At ) + ln(wt ) + cst′ . −β ln(wt+1 /w∗ ) ≈ (γ + 1) ln(wt /w∗ ), where 1 1 1 1 = − ∗ − ∗ γ β εA εq (7) (8) (9) Three cases are possible: • Case 1 ( 1 1 1 + ∗ < − 1) : w∗ is a stable steady-state while 0 is not. Income converges to w∗ . ∗ εA εq β 1 1 1 1 > ∗ + ∗ > − 1) : 0 is a stable steady-state while w∗ is not. If w0 > w∗ , then the economy β εA εq β grows without bound. If instead w0 < w∗ , then the economy contracts towards 0. • Case 2 ( 11 • Case 3 ( 1 1 1 + > ) : the economy is unstable and oscillating. ε∗A ε∗q β In case 1, the economy converges to a steady-state in which income no longer grows. Thus, learning and research only have a transitory impact on growth. This is because their costs rise at a fast rate with effort levels (ε∗A or ε∗q large), while the marginal product of intermediate goods falls rapidly with its employment (β low). In case 2 instead, learning and research have a permanent impact and ongoing growth is possible. Income grows without bound if its initial value w0 exceeds w∗ , but shrinks towards 0 otherwise. If we focus on cases 1 and 2 and interpret the learning effort q and its cost eq (q) as measures of financial development, then the model predicts that the financial sector develops in tandem with the real economy.6 Finally in case 3, the system oscillates and is unstable. The following corollary breaks down the growth rate of income into its various components.7 C 3 : The growth rate of income in a neighborhood of the steady-state γ can be decomposed into γ = γ K + γ A + γ q + γ A−q where: • γ K ≡ −(1 − β) is the contribution of capital accumulation to growth, • γA ≡ β2 is the contribution of research in the absence of learning, ε∗A − β • γq ≡ β2 is the contribution of learning in the absence of research, ε∗q − β • γ A−q , the residual, is the contribution of the interaction between research and learning. Focusing on the non-oscillating dynamics (cases 1 and 2 in Proposition 2) and leaving aside the neoclassical effect on capital accumulation (represented by the γ K term), then the growth rate of income in the economy with both learning and research, γ A + γ q + γ A−q , is larger than the sum of the growth rates in the no-learning economy γ A and in the no-research economy γ q . This is again a reflection of the positive feedback between learning and research. 6 ′ eq (qt ) can represent the amount of resources devoted to analyzing investment opportunities. Alternatively, we could add to the economy a competitive intermediary who invests funds on behalf of the financier. The intermediary collects information about projects’ returns and is paid a fee to compensate for the disutility of learning. There is free entry in the intermediary sector. Equation 16 in the Appendix implies that d ln(qt /q ∗ )/d ln(wt /w∗ ) is positive if 1/ε∗A + 1/ε∗q < 1/β. 7 Corollary 3 nests the two polar cases we discussed above. When ε∗q = +∞, qt is frozen as in the “endogenous growth case” and γ q = 0, so the economy grows at the rate γ K + γ A . If instead ε∗A = +∞, then At is frozen as in the “financial development case” and γ A = 0, so the economy grows at the rate γ K + γ q . 12 2.4 Testable Predictions In our empirical work below, we test two specific hypotheses implied by Propositions 1 and 2. The first is that the entrepreneur performs more research when the financier learns more (equations 3 or 7). The second is that the financier learns more when the entrepreneur does more research (equations 4 or 8). We also test an auxiliary prediction of the model, namely that, as firms innovate, financiers learn more because investment mistakes become increasingly costly. That is, identifying the better firms becomes more important as the dispersion in firms’ profitability increases. 3 Empirical Strategy Testing the predictions of the model concerning the relationships between research and learning calls for a treatment of the biases that a direct OLS estimation of these relationships induce. First, any shock to capital (wt in the model) will stimulate learning and research independently, thereby generating a spurious correlation between them. Moreover, the two-way relationship generates an endogeneity bias. For example, a least squares regression of learning on research yields inconsistent estimates because the regression’s residual is correlated with the regressor (research) as equation 4 shows. Similarly for equation 3, the regressor (learning) is correlated with the residual from the regression of research on learning. Our strategy for addressing these issues is to exploit exogenous changes in public policies or regulations, as is commonly done in the finance and growth literature (see for example, Jayaratne and Strahan (1996) and Bertrand, Schoar and Thesmar (2007)). We measure firms’ innovation effort with the level of R&D expenditures and proxy for investors’ research effort about a firm’s prospects using the number of research analysts covering the firm. 3.1 More innovation leads to more financial analysis To test whether more innovation by firms leads to more learning by the financial sector, we examine whether analysts’ coverage of firms changed around the staggered implementation of R&D tax credits across US states 13 between 1992 and 2006.8 These policy changes provide a source of variation in firms’ innovation activities, which is plausibly exogenous to firms’ analyst coverage. States R&D tax credits proceeded from the implementation of federal tax credits in 1981. Minnesota introduced its own tax credit in 1982, followed by 32 other states as of 2006 (Wilson, 2009). These credits allow firms to reduce their state tax liability by deducting a portion of R&D expenditures from their state tax bill. State taxes are usually based on turnover or business activities (such as the presence of employees or real estate) in the state.9 We argue that increases in state R&D tax credits provide a plausibly exogenous source of variation in firms’ innovation effort. Indeed, from the point of view of an individual firm, changes in state R&D tax credits are likely to alter firms’ R&D behaviour in ways unrelated to variables (such as technological or market conditions) that could also affect the coverage decision of brokerage firms (Freeman, Marschke, and Wang (2010)). Previous analysis of R&D tax credits in the US and elsewhere show that tax credits do encourage changes in R&D expenditures (Hall and Van Reenen, 2000; Wilson, 2009). Table 1 provides information on state tax credits, including the year in which they were first introduced, their level and subsequent changes.10 Tax credit rates range from 3% in Nevada and South Carolina to 20% in Arizona and Hawaï. Few studies have assessed the impact of state tax credit on firms’ innovation effort. At the state-level, previous research suggests a positive impact of these credits on in-state innovation (Wilson, 2009) and on the number of high-tech establishments in the state (Wu, 2008). At the firm level, Paff (2005) focuses on the effect of the California tax credit and finds that firm R&D respond more to the state tax credit than to the federal tax credit. Recently, Bloom, Schankerman and Van Reenen (2013) use changes in states and federal tax credits to identify the effect of R&D spillovers of firms located in close geographic and product markets. 8 We start the R&D sample in 1992 to align it with our second experiment (brokerage house closures and mergers). The R&D sample stops in 2006 because it is the last year in which state tax credit information is available from Wilson (2009). 9 See Heider and Ljungqvist (2014) for more details on state corporate taxes. Some states allow loss-making firms to convert tax credits into cash, and/or to carry these credits forward. 10 We thank Daniel Wilson for making this data available (see http://www.frbsf.org/economic-research/economists/danielwilson/). 14 We first confirm that increases in tax credits are indeed associated with increases in R&D expenditures for firms headquartered in these states, and then compare the change in analyst coverage of firms located in a state that passed a tax credit with the change in coverage of comparable firms located in states without tax credit changes. The staggered implementation of tax credits across states allows to control for potential endogeneity problems thanks to a difference-in-difference estimation (e.g. Bertrand and Mullainathan (2003)). For example, aggregate shocks contemporaneous to the implementation of a tax credit may influence firms’ analyst coverage and confound the effect of innovation. To the extent that, absent treatment, analyst coverage of firms in different states follow similar trends, and that the passage of each state R&D tax credit is not contemporaneous to other changes in the state that correlated with the analyst coverage of firms headquartered in the state, our difference-in-difference estimation enables us to isolate the effect of innovation on analyst coverage. In effect, each year we use any change in analyst coverage of firms headquartered in states which do not experience a change in R&D tax credit as counterfactual for firms located in states passing an R&D tax credit in that year. By comparing the changes in coverage of treatment and control firms we obtain an estimate of the causal impact of innovation on analyst coverage. We conduct our analysis at the firm level, focusing on large manufacturing U.S. listed firms which consider research and development activities to be material to their business.11 Whenever a state implements a tax credit, we compare the change in coverage of firms affected by the state tax credit (treated firms) with the coverage of firms in other states (control firms). Following Heider and Ljungqvist (2014), to reduce any potential endogeneity of a state choosing a certain level of tax credit, we do not consider the actual level of the tax credit granted in the state, but an indicator variable that equals one in the year the state introduces or increases its R&D tax credit, and zero in other years.12 Firms’ locations are identified with the location of their headquarters as reported in Compact Disclosure.13 Since many firms locate in Delaware for reasons unrelated to their operations, we 11 In practice, we drop firms not reporting or reporting zero R&D expenditures. We do not consider tax credit cuts as very few states implement those over our sample period (60 firm-year observations). These are treated as untreated observations. 13 When such data is not available, we take the state of the latest headquarter as reported in Compustat. Acharya and Baghai (2014), citing Howells (1990) and Breschi (2008) suggest that large firms locate their R&D facilities close to the company’s head12 15 exclude all firms in that state. As in Heider and Ljungqvist (2014), we estimate our main regression in first differences, rather than in levels because our variables are serially correlated. All regressions include year dummies, as well as standard errors clustered at the firm level. This specification is the most appropriate in a panel with a large cross-section of firms but a small number of periods (Petersen, 2009 ; Wooldridge, 2011). In some specifications we include time varying controls such as the lagged logarithm of sales and a dummy variable indicating whether the firm made accounting losses (which affects the firm’s tax liability, and potentially the benefit of the tax credit). Formally, the regression takes the form: + ∆ ln(coveragei,s,t ) = βT Cs,t + ηt + j γ j ∆Xi,t−2 + εit , j + is a dummy variable which takes the values 1 if state s implemented or increased its R&D tax where T Cs,t j credit in year t − 1; η t are year dummies and Xi,t are the firm controls listed above. Our coefficient of interest is β which measures the difference between the change in analyst coverage for firms in the treated state relative to the change in coverage for firms in other states. That difference-in-difference estimate is robust to many potential confounds. Aggregate time-varying shocks as well as time-invariant firm attributes are captured by the year dummies and the differencing of the data. In addition, we control for time-varying changes in firm characteristics such as size and profitability by directly including these lagged variables in our specifications. One remaining potential concern with the methodology is the finding by Wilson (2009) that, at the state level, a portion of the increase in in-state R&D is due to a decrease in R&D in other states. In our context, it is possible that following the passage of an R&D tax credit firms relocate in high tax credit states at the expense of other states. For example, firms could hire more researchers in states that pass a tax credit, presumably by offering higher compensation and better work conditions to researchers from other states. However, the important point for our analysis is that R&D rises quarters and do not disperse geographically. 16 for the firms headquartered in the tax credit states, regardless of where the extra R&D comes from. We show empirically that this is the case. Further, if some firms were simply substituting R&D across states without increasing their overall R&D spending, this would bias our estimates towards not finding an effect of the tax credit on firm level innovation and analyst coverage in treated states. In sum, the change in state R&D tax credit provides a good setting to assess the effect that firm innovation has on the learning effort of the financial sector. 3.2 More financial analysis leads to more innovation The second prediction of our model is that more learning by the financial sector increases firms’ innovation effort. The ideal experiment for testing this prediction would be one in which the financial sector’s ability to learn about firms’ innovative projects changes for exogenous reasons. The identification strategy pioneered by Hong and Kacperczyk (2010), extended by Kelly and Ljungqvist (2012) as well as Derrien and Kecskes (2013) and followed by others, comes close to this. These last authors exploit closures of and mergers between brokerage houses which lead to the removal or dismissal of analysts. Indeed, closures often lead to the removal of analysts who are not re-hired by a new broker, while a number of mergers lead to the dismissal of redundant analysts who follow the same stocks as analysts working for the other merging entity. Kelly and Ljungqvist (2012) and Derrien and Kecskes (2013) provide convincing evidence that the drop in analyst coverage resulting from such events is largely exogenous to firms’ policies. What matters for our purpose is that these drops reflect a significant reduction in the resources allocated to financial analysis. To assess how a decrease in analyst coverage changes firms’ innovation effort, we study the firms affected by 52 events identified by Derrien and Kecskes (2013). But unlike them, we restrict our analysis to stocks reporting strictly positive R&D expenditures in order to focus on innovative firms. The firms affected by these exogenous variations in analyst coverage are those for which an analyst who was covering the firm before the closure of his brokerage house disappears from I/B/E/S, or for which a redundant analyst who was covering a firm also covered by the other merging entity disappears from the database. In our 17 study, we examine how the innovation effort of these firms evolves, compared to others firms, when they lose an analyst. Adopting the same specification as in our first experiment, our difference-in-differences estimator compares this change to the change experienced by control firms unaffected by the event, effectively controlling for changes (or overall downwards trends) in firms’ innovation effort.14 Overall, the broker closures and mergers events provide an excellent setting for assessing the relationship between the amount of information produced by the financial sector and firms’ innovation effort. Together, our two experiments allow us to study the impact innovation has on financial analysis, and conversely, the impact that financial analysis has on innovation. 4 Data We evaluate the feedback effects between analyst coverage and firms’ innovation effort on a unique set of innovative manufacturing firms consistent across both shocks (R&D tax credit changes and broker events). That way, our quantitative estimate of the feedback effect is not biased by differences in firm characteristics across the two experiments. In addition, in constructing our unique sample we pay attention to have, for both experiments, treatment and control firms that are sufficiently similar. This is particularly important for the second experiment, since, as reported by Hong and Kacperczyk (2010), brokerage closures primarily affect firms that are larger than the average Compustat firms. A lack of overlap on covariates (such as size) between treatment and control groups, can lead to imprecise estimates (Crump et al. (2009)). A practical solution recently suggested by Crump et al. (2009) to remedy to a potential lack of overlap is to first estimate a propensity score on all firms, and then restrict the analysis to firms with a score between 0.1 and 0.9. We adapt this nethodology to our setting and estimate the propensity score of all firms for each experiment. Our final sample firms include 1103 innovative firms are those with a score on the [0.1,0.9] interval for both experiments.15 14 As a robustness check we also implement the matching procedure of Derrien and Kecskes (2013) on our set of innovative manufacturing firms. We also find that the loss of an analyst decreases firms R&D expenditures. 15 The propensity score is estimated with a logit regression. The treated indicator takes the value one if the firm is treated at one point in the sample and zero otherwise. The covariates included in the logit regression are industry dummies and the time averages of the logarithm of sales and loss indicator over the pre-treatment period for firms that will eventually be treated, and over the 18 We focus on the firms that report strictly positive R&D expenditures. Firms typically report R&D expenditures in their financial statements when those expenditures are material to their business (Bound, Cummins, Griliches, Hall and Jaffe, 1984). Thus, keeping firms with strictly positive R&D expenditures ensures that our tests focus on firms for which our model is the most relevant. We exclude firms with year-on-year R&D growth of over 200% to reduce estimation noise introduced by mergers or radical strategic decisions that would have little to do with changes in analyst coverage or state R&D tax credits. We use the logarithm of R&D expenditures to measure firms’ innovation effort. To measure analyst coverage, we count the number of unique analysts making a yearly or a quarterly earnings forecast during the firm fiscal year. We then take the logarithm of that number as our measure of coverage.16 We deflate all accounting variables — taken from Compustat — at the Consumer Price Index. Table 2 presents the summary statistics for our sample. Our typical firm is large (the median amount of sales is $660m in the sample versus $63m for Compustat firms with positive R&D expenditures) and, by construction, is innovative. These larger firms have median R&D expenditures that represent a smaller proportion of their asset base than the typical Compustat firm that report positive R&D (4% vs. 8%). Figure 3 shows the time series of our measures of financial analysis (number of analysts following a firm) and innovation effort (level of R&D expenditures) adjusted for both industry effects and the amount of sales made by the firm.17 The figure illustrates that, as our theory suggests, the two series are positively correlated over time. 5 Results A visual examination of Figure 3 suggests a positive association between analyst coverage and R&D expenditures. To control for factors that could confound the relationship between those variables, Table 3 to 7 present the results of difference-in-difference estimations for states R&D tax credit (Tables 3 to 4), and for brokerage period in which they appear in the sample for the firms that will never be treated. 16 Coverage is strictly positive in our sample. That is, all our firms are followed by at least one analyst in all years. 17 That is, the variables plotted on the chart are the residuals of pooled regressions of each variable on industry dummies and logarithm of sales. 19 events (Tables 6 and 7). 5.1 More innovation leads to more financial analysis We are interested in the effect that changes in R&D tax credits have on learning by the financial sector measured by analyst coverage. We first confirm in Table 3 that an increase in a state’s R&D tax credit leads to an increase in R&D expenditures by firms located in that state. The coefficient of interest is that on the variable TC+, which shows the effect of a tax credit increase on the R&D of firms located in the treated state one year after the passage of the tax credit, compared to other firms which are not experiencing a change in their state’s tax credits in that year. After the state implements a R&D tax credit, treated firms increase their R&D expenditures by 5.6% relative to control firms. The change takes place in the year after the tax credit implementation, and doesn’t significantly revert the year after, as the unsignificant coefficient on the lagged tax credit variable, L.TC+, indicates. Thus, firms increase their innovation effort in response to tax credits increase, as expected. We turn to the first prediction of the model. Consistent with it, Table 4 shows that a greater innovation effort leads to more learning by the financial sector. After the passage of a state R&D tax credit, the coverage of firms in the treated states increases significantly by 6.2% compared to firms located in other states. Given that the average firm in the sample has 7 analysts, each firm gets an additional 0.4 analysts after the passage of a tax credit. Combining the numbers in Tables 3 and 4 indicates that the sensitivity of analyst coverage to R&D expenditures is about 0.0616/0.0558 = 1.1039, i.e., a 10% increase in R&D expenditures induces a 11% increase in analyst following. The effect of control variables is consistent with previous research. In particular, the coefficient on the change in firm size is positive and significant. We investigate next whether these findings can be explained by the mechanism outlined in our theory. The model suggests that as firms innovate, financiers learn more because investment mistakes become increasingly costly. That is, identifying the better firms becomes more important as the dispersion in firms’ profitability increases. We assess the change in the dispersion in firms’ profitability - measured as the return on assets 20 (ROA) — following the passage of state R&D tax credit.18 We expect to find that the cross-sectional variance of firms’ ROA in the treated state increases relative to the cross-sectional variances in control states. To test this, we adapt the methodology of Bertrand and Mullainathan (2003) for multiple treatment groups. We retain for each firm three years of observations before the tax credit change and three years of observations after. We proceed in two stages. First, we pool all observations (in both treated and control states), regress firms’ ROA on state and year dummies and extract the residuals. Then, for each treated firm (from here on, control firms do not play any role), we average the residuals before the treatment year on one hand, and after the treatment year on the other. Finally, we test whether the cross-sectional standard deviation of average residuals is equal before and after the treatment. We report both the results of a parametric F-test for equality of variances as well as the non-parametric version of the test developed by Levene (1960). As Table 5 indicates, the hypothesis of equality of variances is rejected: after the passage of an R&D tax credit, the cross-sectional dispersion in ROA significantly increases in the treatment state, compared to other states (p<0.05). This finding provides support for the model, suggesting that an increase in innovation leads to a wider dispersion in profits and hence to a greater need for financial analysis. 5.2 More financial analysis leads to more innovation Turning to our second equation, we use brokerage closures and mergers to gauge the effect of financial analysis on firms’ innovation effort. Table 6 confirms that treated firms (firms followed by analysts employed at a closing or merging brokers) experience a reduction in analyst coverage in the year following a merger. On average, a firm loses about 10% more analysts than control firms. Given that the average firm has about 7 analysts, treated firms lose on average 0.7 analyst compared to control firms. Table 7 presents the main results regarding the innovation effort. We find that treated firms’ R&D expenditures fall by 3.1%, relative to control firms, after brokers close or merge. Our findings confirm and strengthen those of Derrien and Kecskes (2013) who report a decrease in R&D expenditures over total assets of 0.21% in 18 We define return on asset as the ratio of EBIT to total assets. 21 a broader sample of firms which includes firms with nonmaterial R&D, and a comparable set of events. Our estimates indicate that a 10% drop in analyst coverage triggers a 10% × (−0.031)/(−0.096) = 3.230% decline in R&D expenditures.19 These results provide support for the second prediction of the model, namely that a more intense learning effort by the financial sector spurs innovation by the corporate sector. Together, our empirical investigations support the predictions at the core of our model: Wall Street research and Main Street innovation interact and reinforce each other. 5.3 Quantifying the indirect effect of learning on innovation Our estimates also allow us to assess the magnitude of the feedback effect between innovation and learning. The third column of Tables 6 yields an estimate of the sensitivity of R&D expenditures to analyst coverage: ∆ ln(rd) = −0.031 × ∆ ln(coverage) = 32.3% × ∆ ln(coverage). −0.096 From Table 4, the passage of a tax credit increases analyst coverage by 6.16% on average. Hence, the indirect effect of the tax credit, operating through analysts’ response, and denoted ∆ ln(rd)∗ , is given by ∆ ln(rd)∗ = 32.3% × 6.2% = 2.0%. To put this number into perspective, we compare it to the total effect of the tax credit on ∆ ln(rd), which equals 5.6% according to Table 3. Thus, the indirect effect of the tax credit through analysts’ response is about a third (36% = 2.0%/5.6%) of the size of its total effect. For example, for a policy that triggers a 1% increase in R&D expenditures (as a total effect), up to 36% of the increase indirectly comes from the catalysing effect of financial analysis. They illustrate how policies aimed at improving the functioning of financial markets can act as catalysts for other policies aimed at boosting firm investment. 19 Using patenting as a measure of innovation output, He and Tian (2013) find that a reduction in analyst coverage increases patenting activity in their sample of firms. This finding can be reconciled with ours and those of Derrien and Kesckes (2013) to the extent that a change in a firm’s informational environment (such as a reduction in analyst coverage) alters its patenting policy. For a given innovation effort, a firm could increase its patenting activity to compensate for the loss of information that the analyst no longer produces. 22 6 Conclusion We develop and test a model of financial development and technological progress. Its main insight is that knowledge about technologies and technological knowledge are mutually reinforcing. That is, entrepreneurs innovate more when financiers are better informed about their projects because they expect to receive more funding should their projects be successful. Conversely, financiers collect more information about projects when entrepreneurs innovate more because the opportunity cost of misinvesting, i.e. of allocating capital to unsuccessful projects and missing out on successful ones, is larger. This positive feedback promotes economic growth and leads to a variety of dynamic patterns. Its beneficial impact is permanent in some cases leading to unbounded income growth, but only transitory in others (when the marginal costs of learning and research are highly elastic and the factor share of capital is low). Poverty traps can also emerge in which only countries that are sufficiently wealthy enjoy the benefits of the feedback. New predictions are derived from the model. The predictions are supported empirically by two quasinatural experiments that uniquely permit to isolate the effect of innovation on learning from that of learning on innovation for a given set of firms. Moreover, we estimate that the feedback effect is about a third of the size of the total effect. For example, a 1% increase in R&D expenditures triggered by an R&D tax credit increases analyst coverage by 6%, which in turn is responsible for up to 0.36% of the total increase in R&D expenditures. An implication is that policies targeted on R&D incentives can be rendered more effective by coupling them with policies designed at increasing capital efficiency, such as reducing impediments to trading financial assets or improving accounting standards. Overall, our model and the empirical evidence strongly support the existence of positive feedback loops between learning by the financial sector and firms’ innovation: Wall Street research and Main Street innovation interact and reinforce each other. 23 A Empirical Motivation for the Model The model developped in the paper is motivated by the following observations on the link between finance and growth. 1. Financial development causes growth by improving TFP. A large literature, surveyed by Levine (1997, 2005) shows that financial development promotes economic growth. Country-level, industry-level, firm-level and event-study investigations suggest that financial intermediaries and markets have a large, causal impact on real GDP growth (e.g. Rajan and Zingales (1998), Jayaratne and Strahan (1996), Fisman and Love (2004), Levine and Zervos (1998), Beck, Levine and Loayza (2000)). Moreover, Levine and Zervos (1998) and Beck, Levine and Loayza (2000) find that their relation to TFP growth is strong whereas their link to capital growth is tenuous. Thus, it appears that financial development contributes to growth by improving TFP rather than capital accumulation. 2. Financial development stimulates R&D investments and R&D improves TFP. Carlin and Mayer (2003) examine a sample of advanced OECD countries. They show that industries dependent on equity finance invest more in R&D and grow faster in countries with better accounting standards. They do not find a similar increase for investment in fixed assets, or for countries with a large financial sector. This suggests that finance is associated with the funding of new technologies and that informational problems are a serious impediment to providing capital. Brown, Fazzari and Petersen (2008) also establish a link between equity financing and R&D by analyzing U.S. high tech firms. They estimate that improved access to finance explains most of the 1990’s R&D boom in the U.S.. Credit (Herrera and Minetti (2007)) and venture capital (Kortum and Lerner (2000), Ueda and Hirukawa (2003), Hellmann and Puri (2000)) are also essential to the funding of innovations. Moreover, there exists abundant evidence establishing that R&D is an important determinant of productivity (e.g. Griliches (1988), Coe and Helpman (1995)). 3. Financial development improves allocative efficiency and allocative efficiency improves TFP. Countries with more developed financial sectors allocate their capital more efficiently. In a cross-country study, Wurgler (2000) documents that they increase investments more in their growing industries, and decrease investments more in their declining industries, than countries with less developed markets.20 Event-studies report similar findings. Bekaert, Harvey and Lundblad (2001, 2005), Bertrand, Schoar and Thesmar (2007), Galindo, Schiantarelli and Weiss (2007), Chari and Henry (2008) show that countries that liberalize their financial sector allocate capital more efficiently. Henry (2003) and Henry and Sasson (2008) document in addition a rise in TFP. This is not surprising given that allocative efficiency is an important determinant of TFP (Caballero and Hammour (2000), Jeong and Townsend (2007), Restuccia and Rogerson (2008) and Hsieh and Klenow (2009)). For example, Hsieh and Klenow (2009) find that TFP would double in China and India if capital and labor were reallocated to equalize their marginal products across plants. 4. Financial development alleviates information imperfections. 20 Hartmann et al.(2007) find that the same pattern holds among OECD countries. 24 Rajan and Zingales (1998) and others show that the quality of information disclosure, proxied by accounting standards, enhances growth of industries dependent on external finance. Carlin and Mayer (2003) report that information disclosure is associated with more intense R&D in industries dependent on equity finance and that the relation of industry growth and R&D to information disclosure is more pronounced than to the size of the financial sector. Wurgler (2000) finds a positive cross-country relation between the efficiency of investments and the informativeness of stock prices. Herrera and Minetti (2007) show that the quality of banks’ information has a positive influence on the probability that Italian manufacturing firms innovate.21 B B.1 Model Proofs Proof of Proposition 1 We start with the financier’s investment decision. He maximizes a linear objective — he is a risk-neutral price-taker and the projects generate constant returns to scale, so he allocates all his capital to the projects deemed successful by his signal and none to the others. A fraction qt of successful projects are correctly identified yielding qt At wt units of intermediate goods in total, while a fraction 1 − qt of unsuccessful projects are incorrectly considered successful yielding (1 − qt )At wt in total. Because all risk is idiosyncratic, factor prices, wt+1 and ρt+1 , aggregate production, Yt+1 = [qt At + (1 − qt )At ]wt , and expected profits, πt = ρt+1 Yt+1 , that are deterministic. The financier allocates his wage wt across the continuum of projects conceived by the entrepreneur, guided by his signal St . At this stage, he takes as given the price of intermediate goods ρt+1 , the projects’ returns in case of success and failure, At and At , and the precision of his signal qt . We denote Kt+ the amount of capital allocated to a project deemed successful by the signal, and Ktj− the amount allocated to a project deemed unsuccessful. There are 1/2 projects in each category. For example, projects deemed successful consist of the qt /2 projects that are truly successful and correctly identified, and of the (1−qt )/2 projects that are unsuccessful but incorrectly identified, leading to qt /2 + (1 − qt )/2 = 1/2 projects in total. The budget constraint imposes that Kt+ /2 + Kt− /2 = wt . Output equals Yt = (qt /2)At Kt+ + [(1 − qt )/2]At Kt− + (qt /2)At Kt− + [(1 − qt )/2]At Kt+ , where the four terms represent respectively the production of successful projects that correctly and incorrectly identified, and the production of unsuccessful projects that correctly and incorrectly identified. Profits follow: πt = ρt+1 Yt = ρt+1 qt /2 At Kt+ + At Kt− + (1 − qt )/2 At Kt+ + At Kt− = ρt+1 (qt − 1/2)(At − At )Kt+ + qt At + (1 − qt )At wt , after substituting in the budget constraint. Maximizing this expression with respect to Kt+ (provided that the signal is informative, that is qt > 1/2) yields Kt+ = 2wt and Kt− = 0, i.e. the financier invests all his capital in 21 Wurgler (2000) uses a proxy for informativeness developed by Morck, Yeung and Yu (2000). They measure the extent to which stocks move together and argue that prices move in a more unsynchronized manner when they incorporate more firm-specific information. Examining the cross-section of U.S. firms, Durnev, Morck and Yeung (2004) and Chen, Goldstein and Jiang (2007) document that firms make more efficient capital budgeting decisions when their stock price is more informative. Herrera and Minetti (2007) use the duration of the credit relationship to proxy for the quality of a bank’s information about a firm. 25 the projects revealed by the signal and none to the others. The expected profit simplifies to πt = ρt+1 qt At + (1 − qt )At wt (10) once the optimal investment plan is set. The price of intermediate goods follows from their output: β−1 β−1 ρt+1 = β(Yt+1 )β−1 = β qt At + (1 − qt )At wt . (11) We turn to the determination of the learning and research efforts. The financier and the entrepreneur who exert efforts At , At and qt expect a surplus of πt − eq (qt ) − eA (At + At ) = ρt+1 qt At + (1 − qt )At wt − eq (qt ) − eA (At + At ). They maximize this expression with respect to At , At and qt , taking the price of intermediate goods ρt+1 as given. Since At and At are perfect substitutes in the cost of research, the optimum with respect to At is the corner solution At = 0 as long as qt > 1/2 (since qt > (1 − qt )). That is, the entrepreneur prefers to concentrate her effort on improving returns in case of success. The first-order conditions with respect to the learning and research efforts can be written as: ρt+1 At wt = e′q (qt ) and ρt+1 qt wt = e′A (At ). They admit a unique solution and it is interior because e′q (1/2) = e′A (0) = 0, e′q (1) = e′A (+∞) = +∞ and e′A and e′q are monotonic. The optimal learning effort equates the marginal benefit of learning (∂E(π t )/∂qt = ρt+1 At wt on the lefthand side of the first first-order condition above) to its marginal cost (e′q (qt ) on the right-hand side). Given that the former is constant in qt while the latter increases and spans the entire real line, there exist a unique solution and it is interior. Similarly, the entrepreneur’s research effort At is uniquely defined by the second first-order conition above which equates the marginal benefit of research (∂E(πt )/∂At = ρt+1 qt wt on the left-hand side) to its marginal cost (e′A (At ) on the right-hand side). Higher price and income, ρt+1 and wt , stimulate learning and research because they expand revenues without affecting costs (a higher wage implies larger investments and output, and a higher price larger revenues for a given output). Moreover, the precision of the financier’s information qt rises with the entrepreneur’s research effort At . Indeed, determining the successful projects is more useful when they outperform the others by a large amount. Conversely, the entrepreneur’s research effort is higher when the financier is better informed, i.e. At increases with qt . This is because successful projects, whichever they are, are expected to attract more capital, and hence to generate more revenues. In equilibrium, the price of intermediate goods ρt+1 is given by equation 11. Substituting this expression into the first-order conditions for At and qt leads to equations 4 and 3, which characterize the equilibrium effort levels. B.2 Proof of Proposition 2 and Corollary 3 We first prove the existence of steady-states, and then describe the transition thereto. A steady-state equilibrium is characterized by the following system of equations, obtained from setting wt+1 = wt = w∗ in equations 4, 3 and 5: ∗ ∗β ∗β ∗β w = (1 − β)w A q ∗1−β ′ ∗ (12) βq∗β w∗β = A eA (A ) ∗β ∗β ∗1−β ′ ∗ βA w = q eq (q ) 26 ∗ ∗ A trivial solution is w∗ = A = q ∗ = 0. Assuming w∗ , A and q ∗ are strictly positive, we can take logs and write the system as ∗ β β 1 ∗ ∗ − ln(w ) + 1−β ln(q ) + 1−β ln(A ) = − 1−β ln(β) ∗1−β ′ ∗ − ln(w∗ ) − ln(q∗ ) + β1 ln[A eA (A )] = β1 ln(β) ∗ − ln(w∗ ) + β1 ln[q∗1−β e′q (q ∗ )] − ln(A ) = β1 ln(β) The system’s Jacobian matrix, J, is defined as −1 J = −1 −1 ∗1−β ∗ β 1−β −1 1 ∗ β (εq − β) β 1−β 1 ∗ β (εA − β) −1 (13) ∗ where we use the fact that ∂ ln[A e′A (A )]/∂ ln(A ) = ε∗A − β and ∂ ln[q ∗1−β e′q (q ∗ )]/∂ ln(q ∗ ) = ε∗q − β. The determinant of J satisfies β 2 (1 − β)ε∗A ε∗q det J = 1 − β − β(1/ε∗A + 1/ε∗q ). Because we assume that 1/εA (A) + ∗ 1/εq (q) − 1/β + 1 never equals zero, det J = 0 for all A > 0 and 1 > q ∗ > 1/2. It follows that there exists a unique non-trivial steady-state. To study the dynamics, we log-linearize the system around its steady-state. Taylor-series expansions yield ∗ ∗ ′ ln[eq (qt )] ≈ ln[e′q (q ∗ )] + ε∗q ∗ [ln(qt ) − ln(q ∗ )] and ln[e′A (At )] ≈ ln[e′A (A )] + ε∗A ∗ [ln(At ) − ln(A )]. We substitute these expressions into equations 4 and 3 after taking logs and using conditions 12 characterizing a steady-state and obtain: β ∗ ln(At /A ) ≈ ∗ [ln(qt /q ∗ ) + ln(wt /w∗ )] , (14) εA − β β ∗ and ln(qt /q ∗ ) ≈ ∗ ln(At /A ) + ln(wt /w∗ ) . (15) εq − β ∗ Solving for ln(At /A ) and ln(qt /q ∗ ) yields ln(qt /q ∗ ) ≈ βε∗A ln(wt /w∗ ) and x ∗ ln(At /A ) ≈ βε∗q ln(wt /w∗ ), x (16) where x ≡ (ε∗A − β)(ε∗q − β) − β 2 . x is positive if and only if 1/ε∗A + 1/ε∗q < 1/β. Finally, we take the log of equation 5 and use conditions 12 to write: ∗ ln(wt+1 /w∗ ) ≈ β ln(qt /q ∗ ) + β ln(At /A ) + β ln(wt /w∗ ). (17) ∗ Substituting the expressions for ln(At /A ) and ln(qt /q∗ ) back into this equation leads to ln(wt+1 /w∗ ) ≈ (γ + 1) ln(wt /w∗ ), (18) where γ = βε∗A ε∗q /x is identical to the expression provided in Proposition 2. Income grows if γ > −1 (i.e. 1/ε∗A + 1/ε∗q < 1/β), at a declining rate if γ < 0 and at an expanding rate if γ > 0. The former occurs if 1/ε∗A + 1/ε∗q < 1/β − 1 and the latter if 1/ε∗A + 1/ε∗q > 1/β − 1. If instead γ < −1 (i.e. 1/ε∗A + 1/ε∗q > 1/β), then income oscillates. The cycles are unstable because γ < −1 implies that γ < −2. 27 C C.1 Model Extensions Controlling Projects’ Probability of Success We solve here a version of the model in which the entrepreneur controls projects’ probability of success rather than their return. Projects’ returns are exogenously set to 1 in case of success and 0 in case of failure. Creating projects with a success probability p requires a research effort ep (p) where ep is continuous, increasing, convex, ep (0) = e′p (0) = 0, e′p (1) = +∞ and εp (p) ≡ 1 + pe′′p (p)/e′p (p) > 0. The financier allocates Kt+ units of capital to a project deemed successful by his signal, and Kt− to a project deemed unsuccessful. The former consist of successful projects correctly identified — there are pt qt of them, and unsuccessful projects incorrectly identified — in number (1−pt )(1−qt ), leading to pt qt +(1−pt )(1−qt ) projects in total. Similarly, the number of projects deemed unsuccessful equals (1−pt )qt + pt (1 −qt ). The budget constraint imposes that [pt qt +(1−pt )(1−qt )]Kt+ +[(1−pt )qt +pt (1−qt )]Kt− = wt . Output equals Yt = pt qt Kt+ +pt (1−qt )Kt− , where the first term represents the production of correctly identified successful projects, and the second term that of incorrectly identified successful projects. Substituting in the budget constraint and maximizing this expression leads to Kt+ = wt /[pt qt + (1 − pt )(1 − qt )], Kt− = 0, Yt+1 wt ̟t and π(pt , qt , wt ) = ρt+1 wt ̟t where ̟t ≡ pt qt /[pt qt + (1 − pt )(1 − qt )]. We note that these expressions match those obtained in the main model if pt = 1/2 and At = 1 (equations 10 and 11). Maximizing the surplus, π(pt , qt , wt ) − eq (qt ) − ep (pt ), with respect to pt and qt , taking the price of intermediate goods ρt+1 as given yields the first-order conditions: ρt+1 wt (1 − ̟t )2 /(1 − pt )2 = e′q (qt ) and ρt+1 wt (1 − ̟t )2 /(1 − qt )2 = e′p (pt ). (19) Finally, substituting ρt+1 = β(̟t wt )β−1 into these equations leads to the conditions characterizing the equilibrium: βwtβ (1 − ̟t )2 ̟t1−β (1 − pt )2 = e′q (qt ) and βwtβ (1 − ̟t )2 ̟t1−β (1 − qt )2 = e′p (pt ). (20) We assume from now on that pt is small, i.e. projects are very risky (this happens in particular if ep is very strongly convex). The existence of an interior solution to equations 19 is guaranteed for any ρt+1 if eq is sufficiently convex, namely if εq (q) > (1 + q)/(1 − q) for 1 > q > 1/2. Equations 20 simplify to βwtβ pβt = qt1−β (1 − qt )1+β e′q (qt ) and βwtβ qtβ = pt1−β (1 − qt )β e′p (pt ). The equilibrium properties of learning and research are the same as in the main model (Corollary 2): the financier learns more when the entrepreneur carries out more research (qt increases with pt holding wt fixed), while the entrepreneur carries out more research when the financier learns more (pt increases with qt holding wt fixed). The dynamic system is characterized by equations 20 and 5, and the initial income w0 . It admits two steady-states, one of which is zero. The transition to the steady-states w ∗ > 0 is governed by the following 28 equation: ∗ ) ≈ ( γ + 1) ln(wt /w ∗ ), where ln(wt+1 /w 1 1 1 1 . = − ∗− ∗ γ +1 β εp εq − (1 + ε∗q )q ∗ If 1/ε∗p − 1/[ε∗q − (1 + ε∗q )q ∗ ] < 1/β, then income grows (note that the second term is positive since ε∗q > (1 + q∗ )/(1 − q ∗ ) by assumption). Its growth rate declines if γ < 0 (w ∗ is a stable steady-state) and expands if γ > 0 (unbounded growth). The former occurs if 1/ε∗p − 1/[ε∗q − (1 + ε∗q )q ∗ ] < 1/β − 1 and the latter if 1/ε∗p − 1/[ε∗q − (1 + ε∗q )q∗ ] > 1/β − 1. These dynamic patterns are similar to those derived in the main model. C.2 No-Learning Regime In the model, the financier always learns. This is because the cost of learning is assumed to satisfy e′q (1/2) = 0. Empirically however, financial institutions only emerge once a critical level of wealth or technological knowledge have been reached. In this extension, we assume that e′q (1/2) > 0 and show that learning only takes place for sufficiently developed economies. The following proposition shows how the financier’s learning decision is altered. P 4 : The financier learns if and only if his income exceeds the threshold w given by: wϕA (2−β βwβ ) = e′q (1/2)/(21−β β)1/β . (21) In this case, his learning effort qt solves equations 4 and 3. The proposition is easily understood by inspecting Figure 2. The optimal learning effort is located at the intersection of the declining solid curve (the marginal benefit of learning) and an increasing curve (its marginal cost). If e′q (1/2) = 0 (the situation considered in the previous sections), then the two curves intersect exactly once. In this case, w = 0 and the financier learns regardless of his income. If instead e′q (1/2) is large relative to income, the cost curve lies entirely above the benefit curve and the financier does not learn. As income increases, the benefit curve slides upward and eventually crosses the cost curve. Thus, when e′q (1/2) > 0, the financial sector develops only if the economy is sufficiently rich or technologically advanced. ∗ where w ∗ is the steady-state A variety of dynamic patterns are possible when e′q (1/2) > 0. If w0 < w < wA A level of income in the no-learning economy, then income grows initially at the rate γ K + γ A and accelerates ∗ < w, then a “no-learning trap” once learning takes place to grow at the higher rate γ. If instead w0 < wA ∗ but it is not high enough to make learning beneficial. Thus, poor appears: income grows until it reaches wA countries never learn and remain in a poverty trap, while rich countries can grow without bound. In this situation, policies designed to develop the financial sector and in particular the information environment such as improving accounting standards and corporate transparency, can help escape the no-learning trap. Policies ∗ beyond the threshold w. targeted at research can also work if they can push steady-state income wA The proof proceeds like that of Proposition 1. The first-order condition with respect to the learning effort, ρt+1 At wt = e′q (qt ), still admits a unique solution since e′q is monotonic but it is interior only if e′q (1/2) < j j ρt+1 At wt . Otherwise the financier does not learn: qtj = 1/2 if e′q (1/2) > ρt+1 At wt . The first-order condition with respect to the research effort is given by equation ρt+1 qt wt = e′A (At ). In equilibrium, the learning and 29 research efforts are the unique solutions to the following system of equations: q = 1/2 t β if e′q (1/2) > 2β−1 βAt wtβ β βAt wtβ = e′q (qt ) qt1−β and βqtβ wtβ 1−β At β if e′q (1/2) < 2β−1 βAt wtβ = e′A (At ) . β Thus learning only takes place if e′q (1/2) < 2β−1 βAt wtβ . The wealth threshold w above which learning takes place is obtained by plugging equation 3 into this condition, which leads to equation 21 presented in the Proposition. References Amore, M., Schneider, C. and A. Zaldokas, 2013, “Credit Supply and Corporate Innovation”, Journal of Financial Economics, 109(3), 835-855. Acemoglu, Daron, Philippe Aghion, and Fabrizio Zilibotti, 2006 “Distance to Frontier, Selection and Economic Growth.”, Journal of the European Economic Association, 4, 37-74. Acemoglu, Daron and Fabrizio Zilibotti, 1997, “Was Prometheus Unbound by Chance? Risk, Diversification, and Growth”, Journal of Political Economy, 105. Acemoglu, Daron, and Fabrizio Zilibotti, 1999, “Information Accumulation in Development,” Journal of Economic Growth, 4, 5-38. Aghion, P. and P. Howitt 1992, “ A Model of Growth through Creative Destruction”, Econometrica, 60(2), 323-51. Aghion, P., P. Howitt, and D. Mayer-Foulkes, 2005, “The Effect of Financial Development on Convergence: Theory and Evidence”, The Quarterly Journal of Economics 120, 173-222. Bertrand, M., Duflo, E. and S. Mullainathan, 2004, “How Much Should We Trust Difference-in-Difference Estimates?”, Quarterly Journal of Economics, 119(1), 249-275. Bertrand, M. and S. Mullainathan, 2003, “Enjoying the Quiet Life? Corporate Governance and Managerial Preferences”, Journal of Political Economy, 111(5), 1043-1075 Bertrand, Marianne, Antoinette Schoar and David Thesmar, 2007, “Banking Deregulation and Industry Structure: Evidence from the French Banking Reforms of 1985”, Journal of Finance, 62(2), 597-628. Bhattacharya, S. and G. Chiesa, 1995, ”Proprietary Information, Financial Intermediation, and Research Incentives ”, Journal of Financial Intermediation, 4(4), 328-357. 30 Bhattacharya, Sudipto, and Jay Ritter, 1983, “Innovation and Communication: Signaling with partial disclosure”, Review of Economic Studies 50, 331-346. Beck, Thorsten, Asli Demirgüç-Kunt, and Vojislav Maksimovic, 2005, “Financial and Legal Constraints to Firm Growth: Does Firm Size Matter?”, Journal of Finance, 60, 137-177. Beck, T., R. Levine and N. Loayza 2000, “Finance and the Sources of Growth”, Journal of Financial Economics, 58(1-2), 261-300. Bekaert, Geert, Campbell R. Harvey, and Christian Lundblad, 2001, “Emerging equity markets and economic development, Journal of Development Economics 66, 465-504. Bekaert, Geert, Campbell R. Harvey, and Christian Lundblad, 2005, “Does financial liberalization spur growth?,” Journal of Financial Economics 77, 3-55. Bloom, Nicholas, Schankerman, Mark, and John Van Reenen, 2013, ”Identifying technology spillovers and product market rivalry”, Econometrica 81, 1347-1393. Brown, James, Steven M. Fazzari, and Bruce B. Petersen, 2009 “Financing Innovation and Growth”, Journal of Finance, 64(1), 151-185. Bound, J., Cummins, C. Griliches, Z., Hall, B. and A. Jaffe, 1984, “Who Does R&D and Who Patents?”, in R & D, Patents, and Productivity, University of Chicago Press, 21-54. Caballero, Ricardo J., and Mohamad L. Hammour. 2000. “Creative Destruction and Development: Institutions, Crises, and Restructuring.” In Annual World Bank Conference on Development Economics, ed. Boris Pleskovic and Nicholas Stern, 213—41. Washington, DC: World Bank Publications. Carlin, Wendy, and Colin Mayer, 2003, “Finance, investment and growth”, Journal of Financial Economics, 69, 191—226. Cetorelli, N. and M. Gambera, 2001, “Banking Structure, Financial Dependence and Growth: International Evidence from Industry Data”, Journal of Finance, 56, 617—648. Chari, Anusha and Peter Blair Henry 2008, “Firm-Specific Information and the Efficiency of Investment,” Journal of Financial Economics, 87(3). Chen, Qi, Itay Goldstein, and Wei Jiang, 2007, “Price Informativeness and Investment Sensitivity to Stock Price,” Review of Financial Studies, 20, 619-650. Claessens, Stijn and Luc Laeven, 2003, “Financial Development, Property Rights, and Growth,” Journal of Finance, 58, 2401-2436. Coe, D. T. and Helpman, E., 1995, “International R&D spillovers”, European Economic Review, 39, 859-88. Crump, R.K., Hotz, V.J., Imbens, G.W. and O.A. Mitnik, 2009, ”Dealing with limited overlap in estimation of average treatment effect”, Biometrika, 96, 187-199. 31 De la Fuente, A. and J. M. Marin, 1996, “Innovation, Bank Monitoring, and Endogenous Financial Development”, Journal of Monetary Economics, 38, 269-301. Derrien, F. and A. Kecskes, 2013, “The Real Effects of Financial Shocks: Evidence from Exogenous Changes in Analyst Coverage”, Journal of Finance, 68(4), 1407-1440. Durnev, A., R. Morck and B. Yeung 2004, “Value Enhancing Capital Budgeting and Firm-Specific Stock Returns Variation”, Journal of Finance, 59(1), 65-106. Fisman, Raymond, and Inessa Love, 2004, “Financial development and intersector allocation: A new approach”, Journal of Finance, 59, 2785-2807. Freeman, R., Marschke, G. and A. Wang, 2010, “Estimating the payoff to R&D”, National Science Foundation, mimeo, available at http://www.nsf.gov/sbe/sosp/econ/freeman.pdf. Galindo, A., Schiantarelli, F., Weiss, A., 2007, “Does financial liberalization improve the allocation of investment? Micro-evidence from developing countries”. Journal of Development Economics, 83, 562-587. Goolsbee, A. , 1998, “Does Government R&D Policy mainly benefit Scientists and Engineers?”, American Economic Review, 298-302. Greenwood, J. and B. Jovanovic 1990. “Financial Development, Growth, and the Distribution of Income”, Journal of Political Economy, 98(5), 1076-1107. Griliches, Z., 1988, “Productivity puzzle and R&D: another nonexplanation”, Journal of Economic Perspectives, 2, 9- 21. Grossman, G. and E. Helpman, 1991, “Quality Ladders in the Theory of Growth”, Review of Economic Studies, 58(1). Hartmann, Philipp, Florian Heider, Elias Papaioannou and Marco Lo Duca, 2007, “The Role of Financial Markets and Innovation in Productivity and Growth in Europe”, ECB Occasional Paper 72. He, J. and X. Tian, 2013, ”The Dark Side of Analyst Coverage: The Case of Innovation”, Journal of Financial Economics, 109, 856-878. Heider, F. and A. Ljungqvist, 2014, “As Certain as Debt and Taxes: Estimating the Tax Sensitivity of Leverage from Exogenous State Tax Changes”, mimeo. Hellmann, T., Puri, M., 2000, “The interaction between product market and financing strategy: The role of venture capital”, Review of Financial Studies 13, 959—984. Henry, Peter Blair 2003, “Capital Account Liberalization, The Cost of Capital, and Economic Growth.” American Economic Review, 93 (2), 91-96. Henry, Peter Blair and Sasson, Diego, 2008,.“Capital Account Liberalization, Real Wages, and Productivity”,.NBER Working Paper 13880. 32 Herrera, A. M. and Minetti, R.: 2007, “Informed finance and technological change: Evidence from credit relationships”, Journal of Financial Economics 83(1), 223—269. Hope, O.-K, 2003, “Analyst Following and the lnfluence of Disclosure Components. IPOs and Ownership Concentration.” Asia-Pacific Journal of Accounting and Economics, 10(2). Hsieh, Chang-Tai and P. Klenow 2009, “Misallocation and Manufacturing TFP in China and India”, Quarterly Journal of Economics, 124, 1403-48. Irani, R.M. and D. Oesch, 2013, “Monitoring and Coporate Disclosure: Evidence from a Natural Experiment”, Journal of Financial Economics, 109(103), 398-418. Jayaratne, J. and P. E. Strahan, 1996, “The Finance-Growth Nexus: Evidence from Bank Branch Deregulation”, Quarterly Journal of Economics, 111, 639-670. Hyeok Jeong and Robert M. Townsend, 2007, “Sources of TFP Growth: Occupational Choice and Financial Deepening”, Economic Theory, 32(1), 179-221. Jones, Charles, 2005, “Growth and Ideas”in P. Aghion and S. Durlauf, eds.: Handbook of Economic Growth (North-Holland Elsevier Publishers, Amsterdam). Kanwar, S., Evenson, R., 2003, “Does intellectual property protection spur technical change?”, Oxford Economic Papers, 23(4), 568-587. Kelly, L. and A. Ljungqvist, 2012, “Testing Asymmetric-Information Asset Pricing Models”, Review of Financial Studies, 25(5), 1366-1413. King, R. G., and Levine, R. 1993, “Finance, entrepreneurship, and growth: Theory and evidence”, Journal of Monetary Economics, 32, 513—542. Kortum, S., Lerner, J., 2000, “Assessing the contribution of venture capital to innovation”, Rand Journal of Economics 31, 674—692. La Porta, R., Lopez-de-Silanes, F., Shleifer, A., Vishny, R., 1997, “Legal determinants of external finance”, Journal of Finance 52, 1131—1150. La Porta, R., Lopez-de-Silanes, F., Shleifer, A., Vishny, R., 1998, “Law and finance”, Journal of Political Economy 106, 1113—1155. Lang, M. H., and R. J . Lundholm, 1996, “Corporate Disclosure Policy and Analyst Behavior,” Accounting Review, 71(4), 467-92. Levine, R. 1997 “Financial Development and Economic Growth: Views and Agenda”, Journal of Economic Literature, 35(2), 688-726. Levine, Ross, 2005, “Finance and growth: Theory and evidence”, in P. Aghion and S. Durlauf, eds.: Handbook of Economic Growth (North-Holland Elsevier Publishers, Amsterdam). 33 Levene, H., 1960, “Robust tests for equality of variances”. In Contributions to Probability and Statistics: Essays in Honor of Harold Hotelling, ed. I. Olkin, S. G. Ghurye, W. Hoeffding, W. G. Madow, and H. B. Mann, 278-292. Menlo Park, CA: Stanford University Press. Levine, R. and S. Zervos 1998, “Stock Markets, Banks, and Economic Growth”, American Economic Review, 88(3), 537-58. Morck, R., B. Yeung and W. Yu 2000, “The Information Content of Stock Markets: Why Do Emerging Markets Have Synchronous Stock Price Movements?”, Journal of Financial Economics, 58(1-2), 215-60. Paff, L., 2005, “State-Level R&D Tax Credits: A Firm-Level Analysis”, Topics in Economic Analysis & Policy, 5(1). Petersen, M., 2009, “Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches”, Review of Financial Studies, 22(1), 435-480. Rajan, Raghuram G., and Luigi Zingales, 1998, Financial dependence and growth, American Economic Review 88, 559-587., D. and R. Rogerson 2008, “Policy Distortions and Aggregate Productivity with Heterogeneous Establishments”, Review of Economic Dynamics, 11, 707-20. Robinson, J. V. ,1952, “The model of an expanding economy”, Economic Journal , 62(245), 42-53. Romer, P. 1986, “Increasing Returns and Long-run Growth”, Journal of Political Economy, 94(5), 1002-37. Romer, P. 1990, “Endogenous Technological Change”, Journal of Political Economy, 98(5), S71-102. Ueda, M., 2004, “Banks versus Venture Capital: Project Evaluation, Screening, and Expropriation”, Journal of Finance 59(2), 601-621. Ueda, M., Hirukawa, M., 2003, “Venture capital and productivity”, Unpublished working paper. University of Wisconsin, Madison. Veldkamp, Laura, 2009, ”Information Choice in Macroeconomics and Finance”, forthcoming, Princeton University Press. Wilson, D., 2009, “Beggar Thy Neighbor? The In-State, Out-of-State, and Aggregate Effects of R&D Tax Credits”, Review of Economics and Statistics, 91(2), 431-36. Wooldridge, J., 2011, Cluster Samples and Clustering, http://www.eief.it/files/2011/10/slides_5_cluster.pdf. EIEF Labour Lectures, available at Wu, Y., 2008, “State R&D Tax Credits and High-Technology Establishments”, Economic Development Quarterly, 22, 136-148. Wurgler, Jeffrey, 2000, Financial markets and the allocation of capital, Journal of Financial Economics 58, 187-214. 34 Figure 1. Timing. 35 Figure 2. The Learning Effort. The picture displays the marginal benefit of learning (solid curve) and its marginal cost (dotted and dashed curves) as a function of the learning effort qt . The dotted curve assumes that e’q(q) = [(q − 1/2)/(1 − q)]0.4 and the dashed curve that e’q(q) = [(q − 1/2)/(1 − q)]0.4+1. The optimal learning effort lies at the intersection of the solid and dotted curves in the first case, and in the corner (q t = 1/2) in the second case. The other parameters are β = 1/2 and wt = At = 1. 36 Figure 3. Coverage and Innovation Measures. The figure shows the average number of analyst covering sample firms and the amount of R&D expenditures, both expressed in logarithm and adjusted for industry and size effects between 1992 and 2006. The series are constructed by taking the average of the residuals of pooled regressions of each variable on industry (sic2) dummies and the logarithm of sales. 37 Table 1. R&D Tax Credits Rate Changes Implemented by US States between 1992 and 2006. Data on states R&D tax credit is taken from Daniel Wilson’s website (http://www.frbsf.org/economic-research/economists/danielwilson/). In this table, given our focus on high-tech firms, we report the statutory tax credit for the highest tier of R&D spending, though for most states the tax credit rate does not vary with the level of R&D spending. Our regressions are based on the direction of the change in tax credit only, not the actual level. State AZ AZ CA CA CA CT DE GA HI ID IL IL IN LA ME MD MO MT NE NH NH NH NJ NC NC OH PA RI RI SC SC TX TX UT VT WV Year 1994 2001 1997 1999 2000 1993 2000 1998 2000 2001 2003 2004 2003 2003 1996 2000 1994 1999 2006 1993 1994 1995 1994 1996 2006 2004 1997 1994 1998 2001 2002 2001 2002 1999 2003 2003 Tax Credit 20.0% 11.0% 11.0% 12.0% 15.0% 6.0% 10.0% 10.0% 20.0% 5.0% 0.0% 7.0% 10.0% 8.0% 5.0% 10.0% 7.0% 5.0% 3.0% 7.0% 15.0% 0.0% 10.0% 5.0% 3.0% 7.0% 10.0% 5.0% 17.0% 3.0% 5.0% 4.0% 5.0% 6.0% 10.0% 3.0% 38 Direction of Tax Credit Rate Change + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - Table 2. Summary Statistics. This table presents the summary statistics on the full sample. The sample includes US firms reporting strictly positive R&D expenditures. The statistics are computed on one observation per firm (the time average of the variable). The last column reports the same statistics from the Compustat sample of firms with strictly positive R&D expenditures. coverage Rd ($m) rd/assets sales ($m) roa capx/assets mean sd 25th 50th 75th N 11.27 77.13 0.06 3170.34 0.12 0.06 8.96 279.90 0.07 10771.75 0.11 0.03 4.64 5.71 0.01 261.49 0.09 0.03 8.67 14.42 0.04 659.66 0.14 0.05 15.00 37.39 0.10 1989.64 0.18 0.07 1103 1103 1103 1103 1103 1097 39 50th in Compustat 2.29 3.98 0.08 63.02 0.05 0.04 Table 3. State R&D Tax Credits: Effect on Firm R&D Expenditures. The table presents the results of the difference-in-difference estimation for the effect of R&D tax credit implementation on firms’ R&D expenditures. The regressions are estimated in first difference, which control for firms’ time invariant characteristics (firms fixed effects). All regressions include year dummies to control for aggregate shocks in each year. Standard errors are clustered at the firm level. T-stats are in brackets. ***, **, * denote significance at 1%, 5%, and 10%, respectively. Δln(rd) Δln(rd) 0.0430*** (2.97) 0.0558*** (3.84) Yes 6795 0.1585*** (6.24) -0.0346 (-1.54) Yes 4889 F.TC+ TC+ L.TC+ L2.Δlnsales L2.Δloss Year FE N Δln(rd) 0.0136 (0.77) 0.0668*** (4.58) -0.0205 (-1.38) 0.1549*** (6.17) -0.0490** (-1.97) Yes 4099 Table 4. State R&D Tax Credits: Effect on Analyst Coverage. The table presents the results of the difference-in-difference estimation for the effect of R&D tax credit implementation on firms’ coverage by financial analysts. The regressions are estimated in first difference, which control for firms’ time invariant characteristics (firms fixed effects). All regressions include year dummies to control for aggregate shocks in each year. Standard errors are clustered at the firm level. T-stats are in brackets. ***, **, * denote significance at 1%, 5%, and 10%, respectively. Δln(coverage) Δln(coverage) 0.0464** (2.45) 0.0616*** (3.00) Yes 6795 0.0901*** (3.55) -0.0533* (-1.89) Yes 4889 F.TC+ TC+ L.TC+ L2.Δlnsales L2.Δloss Year FE N 40 Δln(coverage) -0.0029 (-0.15) 0.0640*** (2.83) -0.0071 (-0.36) 0.0911*** (3.38) -0.0764** (-2.45) Yes 4099 Table 5. State R&D Tax Credits: Effect on ROA Dispersion. The table tests whether the cross-section standard deviation of firm ROA in the state increases after the passage of an R&D tax credit. The test methodology is adapted from Bertrand and Mullainathan (2003). We retain all firms that are never treated as well as three observations before the tax credit change and three observations after the change for treated firms. In a first stage, we pool all observations and regress firms’ ROA on state and year dummies. We extract the residuals. Then, for each treated firm (from here on control firms do not play any role), we aggregate the residuals in two observations: we average the residuals before the treatment year and after the treatment year. The table reports the result of the test of equality of cross-sectional standard deviation of timeaveraged residuals before and after the treatment year. The test robust for non-normality is the Stata robvar command, implementing the test of Levene (1960). ROA – before ROA – after Variance test Parametric (F-test) Non-parametric (Levene test) 41 mean 0.005 -0.007 sd 0.158 0.179 statistic 0.78 5.22 p-value 0.000 0.022 Table 6. Brokerage Closures: Effect on Analyst Coverage. The table presents the results of the difference-in-difference estimation for the effect of R&D tax credit implementation on firms’ R&D expenditures. The regressions are estimated in first difference, which control for firms’ time invariant characteristics (firms fixed effects). All regressions include year dummies to control for aggregate shocks in each year. Standard errors are clustered at the firm level. T-stats are in brackets. ***, **, * denote significance at 1%, 5%, and 10%, respectively. Δln(coverage) Δln(coverage) -0.1148*** (-6.70) -0.0963*** (-5.38) Yes 6830 0.0897*** (3.56) -0.0502* (-1.78) Yes 4913 F.ANANL.ANL2.Δlnsales L2.Δloss Year FE N Δln(coverage) 0.0026 (0.17) -0.0917*** (-4.73) -0.0113 (-0.63) 0.0889*** (3.31) -0.0743** (-2.37) Yes 4119 Table 7. Brokerage Closures: Effect on Analyst Coverage. The table presents the results of the difference-in-difference estimation for the effect of R&D tax credit implementation on firms’ coverage by financial analysts. The regressions are estimated in first difference, which control for firms’ time invariant characteristics (firms fixed effects). All regressions include year dummies to control for aggregate shocks in each year. Standard errors are clustered at the firm level. T-stats are in brackets. ***, **, * denote significance at 1%, 5%, and 10%, respectively. Δln(rd) Δln(rd) F.ANAN- -0.0398*** (-2.61) -0.0308** (-2.04) Yes 6830 0.1615*** (6.31) -0.0317 (-1.42) Yes 4913 L.ANL2.Δlnsales L2.Δloss Year FE N 42 Δln(rd) -0.0164 (-0.95) -0.0396** (-2.27) -0.0137 (-0.97) 0.1568*** (6.13) -0.0468* (-1.88) Yes 4119