94648 AUTHOR ACCEPTED MANUSCRIPT FINAL PUBLICATION INFORMATION The Risks of Innovation : Are Innovating Firms Less Likely to Die? The definitive version of the text was subsequently published in Review of Economics and Statistics, (Forthcoming), 2014-03-21 Published by MIT Press and found at http://dx.doi.org/10.1162/REST_a_00446 THE FINAL PUBLISHED VERSION OF THIS ARTICLE IS AVAILABLE ON THE PUBLISHER’S PLATFORM This Author Accepted Manuscript is copyrighted by the World Bank and published by MIT Press. It is posted here by agreement between them. Changes resulting from the publishing process—such as editing, corrections, structural formatting, and other quality control mechanisms—may not be reflected in this version of the text. You may download, copy, and distribute this Author Accepted Manuscript for noncommercial purposes. Your license is limited by the following restrictions: (1) You may use this Author Accepted Manuscript for noncommercial purposes only under a CC BY-NC-ND 3.0 IGO license http://creativecommons.org/licenses/by-nc-nd/3.0/igo. (2) The integrity of the work and identification of the author, copyright owner, and publisher must be preserved in any copy. (3) You must attribute this Author Accepted Manuscript in the following format: This is an Author Accepted Manuscript of an Article by Fernandes, Ana M.; Paunov, Caroline The Risks of Innovation : Are Innovating Firms Less Likely to Die? © World Bank, published in the Review of Economics and Statistics(Forthcoming) 2014-03-21 CC BY-NC-ND 3.0 IGO http://creativecommons.org/licenses/by-nc-nd/3.0/igo http:// dx.doi.org/10.1162/REST_a_00446 © 2015 The World Bank          446 The World Bank. The Risks of Innovation: Are Innovating Firms Less Likely to Die? Ana M. Fernandes a Caroline Paunov b The World Bank OECD January 19, 2014 Abstract While innovation matters for competitiveness it may expose firms to survival risks. Using plant- product data for Chile and discrete-time hazard models we show that innovating plants have a lower hazard of exit. However, risk impacts strongly on the innovation-exit relationship: only innovators that retain diversified sources of revenue or face lower market risk are less likely to die. Single-product innovators are at greater risk of exiting. Exposure to technical risk does not affect exit probabilities differentially. We provide tentative evidence that single-product innovators have higher profits which helps to rationalize their innovation decision despite the increased risk of exit. Keywords: Firm Exit, Firm Survival, Product Innovation, Multi-Product Firms, Chile. JEL Classification codes: D24, L16, L6, O31. a Ana Margarida Fernandes. The World Bank. Development Research Group. 1818 H Street NW, Washington DC, 20433. Email: afernandes@worldbank.org. 1          446 The World Bank. b Caroline Paunov, OECD, 2, rue André Pascal, 75 775 Paris Cedex 16, France. Email: caroline.paunov@oecd.org and caroline.paunov@gmail.com. The authors would like to thank two anonymous referees as well as Philippe Aghion (the editor), Dominique Guellec, Jonathan Haskel, Jacques Mairesse, Valentine Millot, Pierre Mohnen, Piotr Stryszowski, and seminar participants at the 2011 MEIDE Conference in San José, Costa Rica, the UNU-MERIT/School of Governance in Maastricht, the 2011 Zvi Griliches Seminar at the Barcelona Graduate School of Economics, the 2011 Meeting of the Latin American and Caribbean Economic and Econometric Associations in Santiago, Chile, the 11th CAED conference in Nürnberg, Germany, the 6th Annual Conference of the Academy of Innovation and Entrepreneurship in Oxford, United Kingdom and INSEAD Business School. We thank Asier Minondo for sharing do-files to calculate product proximity measures and Wolfgang Hess and Maria Persson for discussions on proportionality tests and survivor functions. The findings expressed in this paper are those of the authors and do not necessarily represent the views of the World Bank, the OECD or their member countries. 2          446 The World Bank. 1. Introduction Firm exit along with entry are crucial components of the growth and evolution of industries both in developed and in developing countries (Caves, 1998, Tybout, 2000). Models of industry dynamics emphasizing producer heterogeneity and market selection such as Jovanovic (1982) and Ericson and Pakes (1995) suggest that in reasonably efficient markets ‘superior’ firms have higher chances to survive and grow. While being innovative is a central characteristic of ‘superior’ firms it also is a risky venture due to the uncertainties inherent to the innovations themselves and their commercialization. The introduction of new products by a firm - an important type of innovation - involves high and often sunk development and production costs that may fail to bring a sufficient payoff to recover those costs.1 Demand for new products might not pick up or the products could be copied or replaced quickly by new products developed by competitors. The model proposed by Ericson and Pakes (1995) illustrates the risks associated with innovation as firms engage in R&D investments which may improve their efficiency, profits, and survival but can also lead to firm exit if the outcome is not successful. Given that failed product launches are frequent, innovators might ultimately face a higher probability of exit than other firms.2 In this paper, we examine the relationship between product innovation and plant exit focusing on the role of different types of risk for that relationship and on differences in performance payoffs for riskier types of innovation using a rich dataset on Chilean manufacturing plants and their products during the 1996-2003 period. 1 See OECD-Eurostat (2005) for a discussion of the different types of innovation: product innovation, process innovation, managerial innovation, and organizational innovation. 2 See Gourville (2006) on the failure of new product introductions. Famous examples of failed product launches include New Coke by Coca Cola and Sony’s Betamax. 3          446 The World Bank. Our paper makes several contributions to the empirical literature on the relationship between exit and firm characteristics, namely innovation-related variables, as a way to test the implications of industry dynamics models.3 First, our dataset allows constructing objective plant- level time-varying measures of product innovation - categorical and continuous - based on the observation of whether a product is newly manufactured by a plant in any year. This is a clear advantage relative to previous studies that mostly use measures of innovation based on subjective perceptions of managers for a cross-section of firms taken from innovation surveys (Mairesse and Mohnen, 2010). Second, our measures capture product innovations that are new to a plant but not necessarily new to the country nor the world. While these innovations may be considered ‘minor’, their cumulative effects are important drivers of growth (Puga and Trefler, 2010) and in emerging market economies such as Chile they account for the lion’s share of innovation activities, in contrast to the less frequent innovations associated with research and development (R&D) and patents. Third, our analysis goes beyond the innovation-exit relationship by focusing on the role of risk as a crucial determinant of that relationship and testing the hypothesis that a negative relationship is verified only for cautious innovators who are less exposed to risk. A first dimension of risk, inspired by the finance literature principles, relates to the lack of diversified sources of revenue that results from new products accounting for a very large proportion of 3 See Doms et al. (1995), Chen (2002), Disney et al. (2003), and Shiferaw (2009) and Manjon- Antolin and Arauzo-Carod (2008) on the determinants of firm survival and Esteve-Pérez et al. (2004), Hall (1987), Cefis and Marsili (2006), and Zhang and Mohnen (2013) on the role of innovation for firm survival. 4          446 The World Bank. plants’ revenues. A second dimension of risk relates to the technical challenges that innovators need to overcome in order to produce a substantially novel product that is better than the existing products at a competitive cost. A third dimension of risk relates to the market challenges faced by innovators, i.e., the market conditions and sales strategies required to get the new product to be successfully sold in the market.4 Fourth, we conduct a more rigorous test of the innovation-exit relationship than previous studies by applying for the first time discrete-time hazard models with random effects to plant exit rather than using the popular Cox hazard model for plant survival. In doing so, we address that model’s major shortcomings: the fact that it is adequate for continuous-time survival data only and does not allow controlling for unobserved plant heterogeneity. We estimate several discrete-time hazard models for plant exit - complementary log-log (cloglog), probit and logit - with plant random effects to correct for omitted variable biases as well as a linear exit probability model and a logit model for survival up to a fixed number of years with plant fixed effects. Finally, note that to the best of our knowledge, ours is the first study to examine the innovation- exit relationship for an emerging economy.5 Our main findings suggest that engaging in product innovation and introducing a larger number of new products reduces Chilean plants’ exit probabilities. These baseline results are obtained controlling for a large set of time-varying plant and industry characteristics as well as industry, region, and year fixed effects and are robust to a variety of alternative estimation 4 We thank an anonymous referee for pointing us to the relevance of technical and market risks for innovators and their potential impact on the innovation-exit relationship. 5 A more recent study by Zhang and Mohnen (2013) examines the innovation-exit relationship for Chinese firms. 5          446 The World Bank. methods and specifications. The benefits of innovation in terms of reducing the probability of plant death are significantly larger when the new products are exported and for plants engaged in prior investments in machinery or importers of intermediate inputs. We also show that innovation pays off for Chilean plants, in that innovators exhibit significantly higher labor productivity than non-innovators compared to private gains measured by profit rates. These findings suggest that the benefits do not accrue exclusively to the plant through a profit increase but are shared more widely in the economy by supporting higher efficiency. Regarding the role of risk for the innovation-exit relationship, we show that only innovators that retain diversified sources of revenues and are not too dependent on new products benefit in terms of lower death probabilities. In the extreme case of single-product plants, those that innovate are actually at a significantly higher risk of death than non-innovating plants. Our findings show that market risk captured by substantial innovation activities by competitors or higher sales volatility in the new products’ markets affects the innovation-exit relationship: the negative relationship holds only for the less risky types of innovation. By contrast, our results suggest that technical risk captured through the proximity of new products to the plants’ past expertise does not play a role for the innovation-exit relationship. One puzzling question that arises from those results is why would single-product plants engage in risky innovations if those put their survival at risk. This may be related to the expectation of higher payoffs which our evidence shows to be the case: single-product innovators exhibit significantly higher payoffs in terms of profits than multi-product innovators. For other types of risky innovations with a weaker (though not positive) effect on plant exit, the evidence suggests some compensation for risk in the form of higher performance payoffs but the results are weak pointing to explanations beyond differential payoffs. One possibility is that engaging in 6          446 The World Bank. risky innovation is not a choice: market failures such as limited access to finance or physical infrastructure force plants to take on the more risky innovations. The relationship between innovation and plant death is relevant for policy across several dimensions. Plant exit is a major cause of unemployment; thus, our findings are important for implicitly assessing the innovation-employment link. The implications of our findings are twofold. First, a striking implication for industry dynamics is that risk interferes in the firm selection process in terms of innovation, differently from what is known for productivity, i.e., that in well-functioning markets, firms with higher productivity tend to survive and grow under market competition while others exit at earlier stages. By contrast, we find that the survival of innovating firms is not necessarily ensured. Our evidence shows that only cautious innovators have lower exit probabilities. But despite the incentives, engaging in more cautious types of innovation may not be feasible for all plants in all industries. For example, small plants may be unable to add products - thus engage in cautious innovation - since they lack the capacity to maintain a large product range. Moreover, when a radical switch in production is required for innovation, the introduction of new products can occur only on a large scale. Hence, there could be a role for public policy in reducing exposure to risks by promoting investments that potentially result in cautious innovations for certain types of plants and providing guarantees or help to deal with failed innovations. Obviously, such policy interventions would need to be designed so as to set the right incentives ensuring that no moral hazard problems arise. The paper is organized as follows. Section 2 describes the data. Section 3 discusses the methodology. Our main results are discussed in Section 4 while Section 5 examines the role of risk. Section 6 concludes. 7          446 The World Bank. 2. Data and Descriptive Evidence on Plant Survival and Product Innovation in Chile We use a unique dataset on Chilean manufacturing plants and their products (ENIA) collected by the Chilean Statistical Institute (INE) and spanning the 1996-2003 period. The fact that the ENIA is a census of Chilean plants (with more than 10 employees) is crucial for our analysis of plant survival.6 The unique identifier included in the ENIA allows us to follow plants over time and identify exit of a plant in year t+1 if it is in the ENIA in year t but is not in the ENIA in year t+1 and after. Another advantage of our data is that we can identify multi-plant firms, which may exhibit important differences in terms of survival relative to single-plants (Disney et al., 2003).7 Table 1 shows that the average yearly exit rate in the Chilean manufacturing sector is about 9%. 6 Details on the ENIA are provided in Fernandes and Paunov (2012). Plant survival in the ENIA was studied by Lopez (2006) and Alvarez and Vergara (2010). The fact that the ENIA covers in principle only plants with more than 10 employees could pose a problem for our analysis in that plants might drop out of the sample due to their employment falling below 10 employees. However, that principle is in practice more flexible. In our estimating sample for the 1996-2003 period, out of 19439 plant-year observations, 939 plant-year observations (4.8%) have less than 10 employees and in most cases those plants remain in the ENIA multiple years reporting less than 10 employees. Hence, we are confident that plant exit from the ENIA indicates real failure. Nevertheless, we conduct a robustness check focusing on plants with more than 15 employees in Section 4. 7 Information on which ENIA plants are part of a multi-plant firm (with at least two plants responding to the survey) was kindly provided to us by INE for the purposes of this research project. During the 1997-2003 period on average 8.3% of firms are multi-plant firms. In the rest 8          446 The World Bank. The crucial feature of our dataset is that it includes for each plant and year information on all the products manufactured and sold classified at the 7-digit ISIC level (Rev. 2) allowing us to construct two novel measures of product innovation.8 Our first measure is a dummy variable that equals one in year t if the plant sells one or more new 7-digit products while our second measure explores the quantitative dimension of innovation and is the number of new 7-digit products sold by a plant in year t. For both measures a new 7-digit product is one that the plant has never sold prior to year t-1, but the product may not be new to the market or the world. Table 1 shows that the average percentage of plants introducing new products is 14%. The innovation rate is lowest for the food, beverage, and tobacco industries and well-above average for a diverse set of industries such as textiles, wearing apparel and leather, wood and wood products, chemicals, basic metals, and fabricated metal products. The largest numbers of new products are introduced in the basic metals, fabricated metal products and machinery and equipment industries. Most product innovators are multi-product plants. For most innovating plants, new products account for less than 50% of revenues. Figure 1 shows Kaplan-Meier survival functions for innovating plants versus non-innovating plants as preliminary evidence on the univariate relationship between innovation and plant survival, considering separately multi-product plants (Panel A) and single-product plants (Panel B). Innovating multi-product plants have higher survival odds, after five years, 72% survive of the paper we will denote single-unit establishments as ‘plants’ and refer to ‘firms’ when this corresponds to units with multiple plants. 8 See Navarro (2012) and Fernandes and Paunov (2013) for details on the products data. Due to a change in product classification from ISIC Rev. 2 to ISIC Rev. 3 classification in 2001, we omit that year in the econometric analysis. 9          446 The World Bank. while only 59% of non-innovating multi-product plants survive. In contrast, the relationship is not constant for single-product plants: innovating single-product plants have lower survival odds than non-innovating single-product plants after a certain time. 3. Model Specification In order to correctly identify the effects of innovation on plant survival, it is necessary to consider a hazard or duration model whose dependent variable is the time/spell between plant entry and exit (survival spell).9 A hazard model is required for plant survival analysis due to the incomplete nature of the duration information. The hazard function (or hazard rate) represents the conditional probability of a plant ending a survival spell (exiting) after t periods, given that it survived for t-1 periods, (the elapsed duration of the survival spell) and given plant characteristics. When conventional methods such as probit or OLS (linear probability) are used for the estimation of a plant exit model, they study in effect the unconditional probability of the event (e.g., the probability that a plant exits after 5 years in operation) rather than its conditional probability (e.g., the probability that a plant exits after 5 years in operation conditional on having survived until year 4). Hazard models account for the fact that the data contains not only information on plant exit in year t but also additional information that the plant survived until year t-1 before it was forced to exit. By using information on the duration of plant survival instead of focusing just on exit, hazard models do not impose the strong assumption that conditional survival rates are constant over time (i.e., that they are similar whether the plant exits in year 1 or year 4). Moreover, the use of hazard models avoids the biased estimates that probit or 9 See Kiefer (1988), Klein and Moeschberger (1997), and Hosmer et al. (2008) on hazard models and Manjon-Antolin and Arauzo-Carod (2008) on their use in the firm survival literature. 10          446 The World Bank. OLS regressions would obtain given that they ignore the right-censoring of observations (the fact that at the end of our sample period some of the plants are still in operation). Further, using OLS to estimate exit probabilities has the shortcomings that the resulting predicted probabilities are not meaningful as they may lie outside the [0,1] interval and the corresponding variances can be negative. Hence, the magnitude of the effect of innovation on exit probabilities cannot be assessed, which would introduce a substantial limitation to our empirical analysis. These aspects point to the use of hazard models for our analysis of the innovation-exit relationship. The most rigorous way of controlling for unobserved plant heterogeneity in these models is through plant random effects but this requires that those effects be orthogonal to plant characteristics, which is a condition that may not hold beyond experimental data.10 Thus, while hazard models with plant random effects will be our baseline approach - as described further below - we will also consider a more flexible approach to account for unobserved heterogeneity: a linear exit probability model with plant fixed effects. Moreover, we will estimate a logit specification for survival up to a fixed number of years, which also allows controlling for plant fixed effects. Regarding the specific hazard model to consider, the choice depends on the nature of the data and the identification requirements of the analysis. The continuous-time proportional hazards model proposed by Cox (1972) is popular in firm survival studies due to its convenient estimation of the effects of plant characteristics on survival as a proportional shifter of the baseline hazard 10 Hazard models with plant random effects are used by Bandick and Görg (2010) to study plant survival and foreign acquisition and by Brenton et al. (2010), Hess and Persson (2011, 2012), and Esteve-Pérez et al. (2013) to study the duration of trade flows at the product- or firm-level. 11          446 The World Bank. function making no assumptions on that function’s shape.11 A major caveat of the Cox model is that it requires survival time to be a continuous variable and plants to be ordered exactly regarding their failure time. In our data these requirements are not verified as plant survival times are grouped into discrete one-year intervals and while the identification of which plants and how many plants exit from year to year is possible, the ordering of plants’ failure times within a year is not, resulting in ‘ties’ among plants. In the presence of a sizeable fraction of tied survival times the coefficients and standard errors of the Cox model can be biased (Cox and Oakes, 1984).12 This caveat applies also to continuous-time hazard models with a parametric baseline hazard function. Another major caveat of the Cox model is that due to computational difficulties it does not allow controlling for unobserved plant heterogeneity. A failure to account for that heterogeneity leads to biases in the estimated effects of plant characteristics on the hazard of exit (Van den Berg, 2001) and to spurious negative duration dependence of the estimated Cox hazard function (Heckman and Singer, 1984).13 Discrete-time hazard models that can control for 11 See Audretsch and Mahmood (1995), Agarwal and Audretsch (2001), Chen (2002), Disney et al. (2003), and Girma et al. (2007) for the use of the Cox model in firm or industry survival studies. The baseline hazard function summarizes the duration dependence pattern and is estimated non-parametrically using a partial likelihood approach. 12 These biases are present even when correcting the Cox partial likelihood function for the existence of ‘ties’ using the method of Breslow (1974), as we do in Section 4.2. 13 The degree of negative duration dependence may be over-estimated when unobserved heterogeneity is not accounted for, because as time proceeds a selection process implies that only plants better suited to survive remain. A further caveat arises from the Co x model’s proportional hazards assumption that the effect of plant characteristics on the hazard rate does not depend on 12          446 The World Bank. unobserved plant heterogeneity and address the issue of tied failure times (Lancaster, 1990) are thus the more appropriate and preferred choice for our analysis of plant exit. In robustness checks we will obtain estimates based on continuous-time hazard models with a parametric baseline hazard function which can control for unobserved plant heterogeneity and estimates based on the Cox model correcting the partial likelihood function for ‘ties’ using the method of Breslow (1974) for comparability with previous studies. Let a plant-survival spell j be complete ( c j  1 ) or right-censored/incomplete ( c j  0 ) and the number of years a plant survives (i.e., the time to a failure event) T be used in the definition of the discrete-time survivor function which is the probability of plant survival at least m years: m S j (m)  Pr(T j  m)   (1  h jk ) (1) k 1   where T j  min T j , C j , T j is a latent failure time, C j is a latent censoring time for the plant * * * * survival spell j, and h is the discrete-time hazard rate of ending the survival spell, that is exiting, in m years, conditional on survival for m-1 years which is defined as: h j (m)  Pr(m  1  Ti  m / Ti  m  1)  Pr(m  1  Ti  m) / Pr(Ti  m  1) . (2) time duration (i.e., on plant age) which may fail due to unobserved heterogeneity and since the effect of some plant characteristics on the hazard is inherently non-proportional (e.g., initial plant size is likely to affect differently the hazard rate of a very young versus a relatively older plant). This caveat can however be addressed in the estimation by interacting variables with non- proportional effects with plant age, as we will do in Section 4.2. 13          446 The World Bank. Defining a binary dependent variable y jm to take a value of 1 if plant survival spell j ends in year m and 0 otherwise (i.e., a value of 1 in the year of exit for plants that exit and 0 otherwise), its log-likelihood function is given by:   J m log L   y jm log h jm  (1  y jm ) log(1  h jm ) (3) j 1 k 1 where the contribution to the log-likelihood (a) of a right-censored plant survival spell j is the discrete-time survivor function Eq. (1) and (b) of a completed plant survival spell j in interval m is the discrete-time density function (the probability of ending the spell in m years). Eq. (3) implies that discrete-time hazard models for grouped duration times can be estimated using standard regression models for binary choice panel data, as shown by Jenkins (1995). To be fully estimable, the log-likelihood function requires the specification of a functional form for the discrete-time hazard rate h jm that links exit probabilities to explanatory variables (time-varying plant and industry characteristics). We consider three functional forms - complementary log-log (cloglog) following Prentice and Gloecker (1978), probit, and logit - allowing in each case unobserved individual heterogeneity to be accounted for by plant random effects. For the cloglog model, our estimable equation is given by: c log log1  hm ( X / )  log(  log1  hm ( X / ))  X   m   (4) where X is a vector summarizing the characteristics of a plant survival spell (which are time- varying but constant within one-year survival spells) and  m is the baseline hazard. Unobserved plant random effects  are incorporated through the error term   log( ) assumed to be normally distributed. The baseline hazard in Eq. (4), which corresponds to all characteristics in X being equal to zero, varies over survival intervals but the effects of the characteristics are 14          446 The World Bank. constant over duration time, and represent a proportional shift of the baseline hazard function common to all survival spells. For the widely used probit and logit models, the discrete-time hazard rate h jm is distributed, respectively, as an inverse cumulative Gaussian (Normal) and a logistic function (the log of the odds ratio). In the estimation of Eq. (4) as well as the corresponding equations for probit and logit models, the baseline hazard is estimated non- parametrically by including year fixed effects which allow for unrestricted yearly changes in the hazard rates. The three exit probability models are estimated by maximum likelihood techniques using a quadrature approximation due to the inclusion of plant random effects. A stacked binary choice model using a cloglog link function with time-specific intercepts is the exact grouped duration (discrete-time) analogue of the continuous-time Cox proportional hazards model, while the probit and logit models not impose this proportionality assumption (Prentice and Gloeckler, 1978; Sueyoshi, 1995; Hess and Persson, 2011). Thus, the cloglog model assumes the impact of a regressor on survival is the same regardless of plant age, a caveat discussed in the context of the Cox model (footnote 13). We will test for this proportionality assumption for each regressor following the procedure proposed by McCall (1994) and modify the cloglog model accordingly to include regressors with non-proportional effects in levels and interacted with plant age.14 The vector of plant characteristics X includes one of the measures of product innovation defined in Section 2, 4-digit ISIC industry, region, and year fixed effects, and a rich set of plant 14 The test consists in estimating a variant of the cloglog model allowing each regressor to enter in levels and interacted with a duration trend. The proportionality assumption is rejected for regressors for which the coefficient on the interaction with the duration trend is significant. 15          446 The World Bank. and industry controls defined in Appendix Table 1. Following Dunne et al. (1989), Disney et al. (2003) and Bernard and Jensen (2007) we include plant-level time-varying capital intensity, current size and its square as well as the initial size at which the plant started operations and its square and an indicator for multi-plant firms.15 Controlling for size and size squared accounts for non-linearities in the innovation-size relationship while controlling for capital intensity ensures that the product innovation effect is not picking up a capital accumulation effect through process innovation. We also include a measure of plant sales growth to avoid capturing the effects of ‘desperate’ innovators, i.e., plants which switch products in a desperate last attempt to avoid an inevitable closure. This control is particularly crucial when studying the role of risk for the innovation-survival relationship. Following Audretsch and Mahmood (1995), Audretsch (1995), Mata and Portugal (1994), and Strotmann (2007) we include 6-digit ISIC industry-level time-varying sales growth, average innovation rates, and the Herfindahl index of concentration of market shares and its square to allow for non-linear effects of competition on plant exit.16 4. Effects of Innovation on Plant Death Table 2 presents our baseline estimates of the effect of product innovation on plant exit using the discrete-time hazard models cloglog, probit, and logit with plant random effects. Columns (1)-(3) show results for the innovation dummy while columns (4)-(6) show results for the 15 We use a complementary dataset with (non-product) information on all plants since their entry into the ENIA from 1979 onwards in order to compute plant age and initial plant size. 16 For multi-product plants, the 6-digit ISIC level used is that of the plant’s 7-digit product accounting for the largest share of total revenues. 16          446 The World Bank. continuous innovation measure. The significance of the estimated effects is assessed using heteroskedasticity-robust standard errors.17 Note that in what follows any unreported results that are discussed are available from the authors upon request. Unless otherwise noted, the tables report the marginal effects of each regressor on the probability of plant exit, evaluated at the means of the independent variables. Columns (1)-(3) show that engaging in product innovation reduces significantly the exit probability for Chilean plants. Columns (4)-(6) show that the higher is the number of new products introduced by a plant the lower is its hazard of exit, and the effects are significant at the 5% confidence level. For both innovation measures, the magnitude of the effects is close across the cloglog and logit specifications and is slightly lower in the probit specifications. The log- likelihood values in Table 2 suggest that the cloglog model provides the best fit to the data.18 The marginal effect in column (1) implies that a plant’s decision to engage in product innovation would decrease its death probability by 22%, keeping all other variables constant. The marginal effect in column (4) suggests that the introduction of one additional new product would decrease a plant’s death probability by 11%, keeping all other variables constant. The marginal effects of the plant-level controls reported in Table 2 show that plants with higher sales growth and multi-plant firms have a lower exit probability. In conformity with the 17 An alternative approach to the use of plant random effects which allows for a more general form of plant-level serial correlation is the estimation of robust standard errors clustered at the plant level. Our findings are maintained following that approach. 18 The proportionality assumption in the cloglog model is rejected for the multi-plant dummy, plant current size and capital intensity. Hence, these variables enter in levels and interacted with plant age in all cloglog specifications. 17          446 The World Bank. literature, we find a negative relationship between size and capital intensity and the probability of death of Chilean plants (Bernard and Jensen, 2007; Disney et al., 2003; Dunne et al., 1989; Hopenhayn, 1992; and Jovanovic, 1982). Higher average innovation in the industry has a significant positive effect on plant exit confirming previous findings for U.S. firms (Audretsch, 1991, 1995; Audretsch and Mahmood, 1995) and can be explained by the fact that industries with active innovation are more fast-paced.19 Importantly, note that the findings in Table 2 are not due to collinearity between our innovation measures and plant or industry controls since similar findings are obtained in regressions including just an innovation measure along with industry, region, and year fixed effects. Given the qualitatively unchanged effects of plant and industry controls on exit across specifications, we will omit them in the rest of the tables. In Table 3 we verify the robustness of our main findings to the use of other models following the discussion in Section 2. Columns (1) and (5) present the results from estimating a Cox proportional hazards model for plant survival. Columns (2) and (6) show the results from estimating a linear exit probability model including plant fixed effects that account for favorable demand or supply shocks or unobservable plant characteristics that might lead plants to innovate and remain in business. Columns (3) and (7) present the estimates from a continuous-time 19 Other industry controls do not affect survival significantly possibly due to theoretical ambiguity. On the one hand, fast-growing industries have lower exit probabilities as some plants’ growth does not necessarily result in market share losses of rivals, and may thus lead to less aggressive reactions by the latter. On the other hand, conditions are more unsettled in fast- growing industries and higher turnover rates may result. The lack of significance may also be linked to the inclusion of 4-digit industry fixed effects which already capture slow-moving industry characteristics. 18          446 The World Bank. parametric survival model where the distribution of the baseline hazard function is assumed to be a Weibull and the model allows for random plant effects to account for unobserved heterogeneity under the assumption that these are uncorrelated with other explanatory variables. Columns (4) and (8) show the estimates from a conditional logit model for the probability that the plant survives until 6 years including plant fixed effects.20 The negative and significant impact of innovation on plant death is maintained across all these models. The findings in columns (4) and (8) are particularly important since they confirm in a more appropriate setup that does not suffer from biases due to ignoring the right-censoring of observations and from predicted probabilities outside the [0,1] interval (as OLS with plant fixed effects does) that our results are robust to controlling for unobserved plant heterogeneity through fixed effects. Since the cloglog model provides the best fit to the data in Table 2, we will focus on this model for plant exit in all subsequent tables noting however that our findings are robust to the use of other models for exit.21 20 Since the dependent variable in this model is survival up to a certain number of years, the effect of innovation is expected to be of opposite sign to the effects in other columns of Table 3. The results are qualitatively similar when we consider survival until 3 years, 5 years, or 8 years. Due to convergence problems in the conditional maximum likelihood estimation, the logit specifications do not include plant nor industry controls, but rather a dummy variable equal to one for crisis year 1999 and 0 otherwise. 21 To further address the potential bias ensuing from favorable shocks leading plants to innovate and survive, we consider a plant fixed effects instrumental variables (IV) approach. Our instruments are defined at the industry level (which admittedly is not ideal) and capture either technological progress at the knowledge frontier which generates innovation opportunities for 19          446 The World Bank. The validity of our main results is checked further in a series of robustness exercises most of which are presented in Appendix Table 2. First, we estimate our specification for single-plant units only (where plants equal firms) to avoid possible biases related to this source of heterogeneity across plants. The effects of innovation hold for firm (rather than plant) exit. Second, to address the concern that exit could be mechanically due a plant reducing its workforce below 10 employees and being excluded from the ENIA, we estimate our specification for the sub-sample of plants with more than 15 employees throughout the sample period. The findings are qualitatively maintained for that sub-sample of plants hence that concern is not warranted. Third, we modify our measures of innovation to capture new products at the 6-digit ISIC level and find estimates that are qualitatively similar to those for new products at the 7-digit ISIC level.22 Fourth, we add to our specification plant or industry time-varying characteristics (not considered before for parsimony): dummies for whether plants are exporters or foreign-owned, plant productivity, industry advertising to sales ratio, industry capital intensity and entry rates (Audretsch and Mahmood, 1995; Geroski et al., 2007).23 The results confirm our earlier findings. Chilean plants (the number of patents filed under the Patent Cooperation Treaty and TFP growth in U.S. 2-digit ISIC industries) or the possibility of accessing foreign knowledge through spillover mechanisms (the share of plants with foreign licenses and the investment-capital ratio in Chilean 4-digit ISIC industries). The IV estimates show a significant negative effect of innovation on plant exit. 22 The evidence is also upheld for new products at the 5-digit ISIC level. 23 We use labor productivity instead of TFP due to the difficulties for inference that would arise from including an estimated TFP variable in our regression framework. By using labor 20          446 The World Bank. Fifth, our findings are also maintained when all industry controls are defined at the 4-digit or 5- digit level instead of the 6-digit ISIC level. Finally, we check for nonlinear effects for our continuous innovation variable but find none. To provide additional insights into what drives the negative innovation-exit relationship, Table 4 presents the results from estimating our specification splitting the innovation dummy into mutually exclusive groups according to whether the new products are exported or not, are preceded by machinery investments or not, or are accompanied by the use of imported inputs or not. Column (1) shows that whether they are exported or not, new products reduce Chilean plants’ exit probabilities. However the benefits are significantly larger for exported new products as indicated by the t-test for the difference in marginal effects, supporting the idea that the ability to send new products to highly demanding international markets is a strong indication of the success of a plant’s innovation. Column (2) shows that plant exit decreases with product innovation only when the plant also engages in prior machinery investments. A possible interpretation for this finding is that investments likely linked to process innovation are a natural early stage of innovation. Such investments also signal the expectation of a larger and more attractive market which in itself may reduce the plants’ death probability. Column (3) shows that product innovations reduce plant exit probabilities significantly more when accompanied by the use of imported inputs confirming potentially the role of imported inputs as a source of technical knowledge which complements product innovation and possibly reflecting higher quality products (e.g., Kasahara and Rodrigue, 2008; Paunov, 2011). This latter finding may reflect the productivity we avoid the problems associated with the measurement of TFP for multi-product plants highlighted by Bernard et al. (2009). 21          446 The World Bank. specific case of Chile as an emerging economy, as the link might not hold to the same extent for more advanced economies where plants can rely on domestic frontier technical knowledge. 5. Effects of Risk on the Innovation-Death Relationship and on Innovation Payoffs 5.1 Risk and the Innovation-Death Relationship The innovation process poses risks for plant survival along several dimensions. A first dimension of risk (or rather of the lack thereof) – inspired by the portfolio theory of finance – is the diversification associated with a larger number of sources of revenue for a plant. When new products account for a large share of plant revenues, the innovation strategy is more risky since their success and sustainability are more uncertain than those of more established products. A second dimension of risk relates to the technical difficulties faced by innovators, i.e., the fact that in order to produce a novel successful product never manufactured before, the plant needs to overcome substantial technical challenges, particularly if it aims at introducing a product beyond its expertise. A third dimension of risk relates to the market challenges faced by innovators, i.e., the competitive environment and sales strategies required to get the new product to be successfully sold on the market. We explore empirically via several proxies in Table 5 how each of these dimensions of risk feeds into the innovation-exit relationship. Our hypothesis is that the negative relationship shown in Section 4 would be verified only for innovators with less exposure to risk. We assess the first dimension of risk related to the diversification of plant revenue sources, expecting the risk of introducing a new product to be higher if that product accounts for a substantial share of plant revenues. That would be the case when a single-product plant introduces a new product which replaces the previous product it manufactured (thus remaining 22          446 The World Bank. single-product) and its only source of revenue is at stake as the market may not take up the new product, while that would clearly not be the case for a multi-product plant introducing a new product while retaining other more established sources of revenue. Allowing for a differential effect of the innovation dummy on exit for multi-product plants relative to single-product plants in our specification, column (1) of Table 5 shows that only multi-product innovators benefit while single-product innovators are actually at a significantly higher risk of death than non- innovators. The t-tests indicate that the differences across plant categories are statistically significant. The marginal effects suggest that, relative to non-innovators, the death probability is lower by 29% for innovative multi-product plants but higher by 40% for innovative single- product plants. Similar findings are obtained in specifications using the number of new products.24 Another specification examining how the innovation-exit relationship is affected by opportunities for revenue diversification considers explicitly the innovators’ dependence on revenues from new products, distinguishing across plants that introduce new products accounting for less versus more than 50% of revenues. The results reported in column (2) confirm the hypothesis that innovation reduces exit significantly only for cautious innovators that introduce new products on a small scale thus retaining diversified sources of revenue.25 24 Since single-product plants may be ‘special’ entities for a variety of reasons, we examine whether the results are maintained redefining multi-product plants to be those producing three or more products and find that innovation reduces the death probability of multi-product plants but not of ‘few-product’ plants (producing up to two products). 25 Similar results are obtained when changing the cutoff to 40% or 60% of revenues. Moreover, the negative innovation-exit relationship holds only for new products that add to the existing product scope of the plant, suggesting that new products that replace previous products (leaving 23          446 The World Bank. We examine the second dimension of technical risk by considering in Table 5 a measure that attempts to capture potential technical difficulties faced by innovators because the product differs significantly from the past production of the plant. We consider for each new product an index of its proximity to the plant’s past production capabilities following Hidalgo et al. (2007) and Boschma et al. (2012) to account for the relative closeness between the plants’ established industries and those of the new products.26 We estimate our specification allowing differential effects on plant exit for new products that are closer versus more distant from the plants’ past expertise and show the results in column (3). New products that are closer to the plants’ expertise and thus pose lower technical risks do not impact plant exit significantly differently than other new products.27 the product scope unchanged or reducing it) thus putting a plant’s sources of revenue at risk do not significantly reduce its death probability. 26 The proximity index is based on how often countries have comparative advantage in two products simultaneously: if countries with comparative advantage in product A also have comparative advantage in product B with high probability, this implies that products A and B demand the same production capabilities and hence are close to one another. The indices are obtained for each new 7-digit ISIC product relative to the weighted average of the past products of the plant, accounting for the relative weights of those past products in total plant revenues. See Appendix Section 2 in the working paper for a detailed description of the proximity index calculation. 27 We also consider a differential effect of less versus more sophisticated new products on plant exit, assuming that more sophisticated products are a more risky innovation. We use a trade- based proxy giving larger sophistication scores to products exported mostly by countries with 24          446 The World Bank. We examine the third dimension of market risk by considering in Table 5 two measures that capture the challenges associated with getting new products to be successfully sold on the market. First, the value of an innovation depends on the actions of competitors: it will be much lower if competitors also introduce novel products that are very close to the plants’ innovations. Column (4) shows that when distinguishing across new 7-digit products with more than versus less than 10 competitors who introduce innovations concurrently, Chilean plant death probabilities decrease significantly only when innovations face a lower degree of competition. Second, we consider the possibility that innovation is more risky for survival in industries with higher sales volatility where plants are less certain of how their innovations will perform in the market. Allowing the effects of product innovation to differ across industries with higher versus lower sales volatility over the sample period, column (5) suggests that innovation reduces Chilean plants’ death probabilities significantly only in industries with lower sales volatility. 28 higher GDP per capita, following Hausmann et al. (2007) and Jarreau and Poncet (2012). The negative innovation-exit relationship is significant only for plants introducing less sophisticated innovations, though the difference across marginal effects is insignificant. Moreover, we consider two simpler proxies for the risky technical nature of the innovation: (i) whether the new product is introduced in a new industry to the plant (where it may lack technical and market knowledge) relative to an industry in which the plant has manufacturing experience and (ii) whether the product is new to Chile relative to being new just to the plant. For both proxies the two types of innovations reduce the exit probability but the differences across marginal effects are insignificant. 28 While the p-test for the difference across innovation effects in column (5) is insignificant at conventional confidence levels, the p-tests from probit and logit specifications are significant. In 25          446 The World Bank. Overall, our findings show that only innovators that retain relatively diversified sources of revenues benefit in terms of lower death probabilities. In the extreme case of single-product plants, their death probability is actually raised by innovation. Our evidence suggests that technical risk does not play a substantial role for the negative innovation-exit relationship, but market risk does in that the relationship holds only for the less risky types of innovation. A caveat to the latter results is that while technical and market risk are conceptually distinct the proxies available sometimes confound these two types of risk: e.g., a new product more distant from the past expertise of the plant may embody both higher technical and market risk as the plant needs to overcome both production and marketing/distribution challenges for such products. 5.2 Comparing Performance Payoffs to Risky and Cautious Innovations with Survival Costs A question that arises based on the effects of risk on the innovation-exit relationship is why would plants - and single-product plants in particular - engage in risky innovations if these put their survival at risk? A straightforward rationale for the results in Section 5.1 would be that the payoffs from risky innovations are particularly high. Another fundamental question concerns the addition to those results, considering the innovating plant’s pricing strategy, we assume that a more risky type of innovation is charging higher prices for new products than competitors as this risks lower sales unless consumers are willing to pay extra for the new product relative to an old more cost-effective product. Allowing for a differential effect on exit of new products priced (in their introduction year) above versus below the median across all plants selling the product we find a clear reduction in death probabilities from introducing new products priced below the median but no effect for the other new products. 26          446 The World Bank. nature of the benefits from innovation and whether they are appropriated fully by the innovating plant or whether they also benefit the rest of the economy through higher efficiency. To address the nature of the benefits from innovation, we consider two measures of performance for Chilean plants: labor productivity and profit rates. The latter is a measure that reflects the private gains from innovation i.e., the ability of the plant to capture innovation rents, while the former reflects productive efficiency gains that result from innovative activities, thus indicating potential social gains from innovation.29 Columns (1)-(3) and (5)-(7) of Table 6 show the results from estimating the effect of product innovation on each of the two plant performance measures in levels using OLS with plant and year fixed effects, including sequentially no controls, plant and industry controls, and an additional variable identifying plant exit in the subsequent year. The estimates show that product innovation is linked to higher labor productivity but is not significantly linked to profit rates.30 Columns (4) and (8) account for possible dynamics by including a lagged dependent variable and estimating the model in first differences following the Anderson and Hsiao (1982) instrumental variables estimation method.31 The results from this stringent specification indicate that product innovation has a positive payoff 29 We thank a referee for suggesting this distinction. The performance measures are defined in Appendix Table 1. 30 This finding is in line with the conclusions from past studies that used the Crépon et al. (1998) or CDM model to estimate the innovation-performance relationship. This method was applied to Chile by Crespi and Zuniga (2012) using perception-based innovation measures for a cross- section of plants in the Chilean Innovation Survey. 31 The method consists in using the two-year lagged levels of the variables as instruments for the first differences. 27          446 The World Bank. across Chilean plants in terms of labor productivity. The finding of innovation payoffs in terms of efficiency gains being stronger than private returns to innovators - measured by profit rates - suggests that plants are not able to keep the gains from innovation and thus its benefits are reaching beyond the plants’ owners. Incidentally, note that the estimated effects of innovation in Table 6 may actually be affected by the role that risk plays for the innovation-death relationship. Since risky innovation can lead to a higher death probability it should logically bring higher performance payoffs, while cautious innovation reduces the exit probability hence its association with lower payoffs may be tolerated by plants. Due to attrition our sample may thus include relatively more plants that engaged in cautious innovation and fewer plants that engaged in risky innovation. As a result the estimated effect of innovation on performance in Table 6 may be underestimated. To partially mitigate this concern, columns (3), (4), (7) and (8) of Table 6 include a dummy variable to account for plant exit. Table 6 focuses on all plants and thus does not allow us to understand why single-product plants innovate even though that raises their death probability. Next, we focus on the issue of the payoffs to risk, by examining whether risky versus cautious types of innovation differentially impact the two measures of plant performance in Table 7. Our major interest is in exploring why single-product firms innovate in spite of finding their survival being at risk. Column (1) shows that both multi-product and single-product innovators derive positive but weak payoffs in terms of labor productivity while column (2) shows that the payoffs in terms of profit rates are positive and significant only for single-product innovators, and the t-test shows that the difference relative to multi-product innovators is significant. These results on payoffs help to explain why even single-product plants engage in innovation and put their survival at risk. 28          446 The World Bank. Regarding the other risky types of innovation, Table 5 has shown that none is associated with higher exit probabilities per se but those associated with lack of revenue diversification and higher market risk do not bring significant reductions in exit (relative to not innovating). We explore the payoffs to those risky types of innovation in columns (3)-(8) of Table 7. Product innovations accounting for less than 50% of plant revenues have a positive effect on labor productivity and no effect on profit rates. However, when market risk is captured either by the presence of many competitors introducing concurrent innovations in columns (5)-(6) or by the degree of sales volatility in the industry in which a new product is introduced in columns (7)-(8), the estimates show significantly higher payoffs in terms of labor productivity but weakly higher in terms of profit rates for risky innovations. These findings suggest some compensation for risk and constitute a rationale for plants to choose to engage in risky innovation strategies. However, since the results are somewhat weak, they point to the possibility that factors other than the prospects of higher payoffs force some plants to engage in risky innovation strategies. Possible factors well-known to apply in emerging economies such as Chile are market failures of various types and difficult framework conditions including, e.g., limited access to finance needed to expand production lines and invest in innovation activities or physical infrastructure constraints. While very far from a rigorous welfare analysis, our findings on exit and performance payoffs of product innovation suggest the following. Overall, product innovation improves performance across Chilean plants, with payoffs being stronger in terms of labor productivity than in terms of profit rates. The finding of efficiency gains being stronger than private returns to innovators suggests the presence of social payoffs from innovation reaching beyond the plants’ owners. On the cost side, product innovation increases the survival risk of some Chilean plants. For single-product plants, the death probability is significantly raised by innovation while for 29          446 The World Bank. plants engaged in innovations that reduce the diversification of their sources of revenue or are linked to higher market risk the death probability does not decrease, in contrast to those of plants engaged in cautious innovation. On the benefit side, product innovation brings payoffs in the form of significantly higher profit rates for single-product plants, rationalizing their innovation decision. The positive - albeit weak - performance payoffs accruing to plants engaged in other types of risky innovations, provide some evidence of a compensation for risk. However, the weakness in those results points to explanations other than differential payoffs in motivating that innovation. This analysis of the costs and benefits of innovation needs to be qualified by three limitations. First, some of the proxies for risk demonstrate clearly that a risky innovation strategy is not purely determined by plants’ own choices but rather it depends also on the actions undertaken by competitors and on the market conditions faced. Controlling for such conditions and choices, possibly within an industry dynamics model, would be necessary to fully understand the impact of different innovation strategies on plant performance and thus be equipped to conduct a more rigorous assessment of the costs and benefits of innovation. Second, to conduct a valuable fully-fledged welfare analysis additional modeling and data would be necessary. Third, performance payoffs may take longer to materialize than what can be captured by our six-year panel for Chile, thus future research using longer panel datasets will be better placed to measure the long-run impacts of different types of innovation on plant performance payoffs. 6. Conclusion Innovation can expose plants to significant survival risks as the launch of new products may result in lower than expected sales. At the same time, innovation is a potentially powerful source 30          446 The World Bank. to allow longer plant survival in the marketplace. Focusing on Chilean manufacturing plants, this paper shows that product innovation reduces the probability of plant death under certain circumstances. Risk plays an important role for the innovation-exit relationship in that only innovators that retain diversified sources of revenues or face lower market risk benefit in terms of lower exit probabilities. By contrast, we do not find the technical risk of innovation leads to differential impacts on plant death. A striking finding from our analysis is that single-product plants that innovate are at a significantly greater risk of exiting than non-innovating plants. But we also find that risk rewards single-product innovators with higher profits, rationalizing their innovation decision. The evidence for other types of risk points to the importance of explanations other than differential performance. Possible factors well-known to apply in emerging economies such as Chile are market failures of various types and difficult framework conditions including, e.g., limited access to finance needed to expand production lines and invest in innovation activities or physical infrastructure constraints. Our findings have several policy implications. First, a striking implication for industry dynamics is that risk interferes in the firm selection process in terms of innovation, differently from what is known for productivity, i.e., that in well-functioning markets, firms with higher productivity tend to survive and grow under market competition while others exit at earlier stages. By contrast, we find that the survival of innovating firms is not necessarily ensured. Second, while cautious innovation is the most desirable innovation strategy for reducing the probability of death, it may not be feasible for all types of plants nor in all types of industries. Policy actions may be required to improve for example the capacity of small plants to engage in cautious innovation given the desirability of small innovating plants’ survival in terms of securing employment. Where risky innovation is the only possibility, an adequate policy 31          446 The World Bank. mechanism that avoids moral hazard problems would be required providing a guarantee to help failed innovation while setting the right incentives. 32          446 The World Bank. References Agarwal, Rajshree and David Audretsch (2001). “Does Entry size Matter? The Impact of Life Cycle and Technology on Firm Survival,” Journal of Industrial Economics 49(1), 21-43. Alvarez, Roberto and Sebastian Vergara (2010). “Exit in Developing Countries: Economic Reforms and Plant Heterogeneity,” Economic Development and Cultural Change 58(3), 537-561. Anderson, T.W. and Cheng Hsiao (1982). “Formulation and Estimation of Dynamic Models Using Panel Data,” Journal of Econometrics 18(1), 47-82. Audretsch, David (1991). “New-Firm Survival and the Technological Regime,” Review of Economics and Statistics 73(3), 441-450. Audretsch, David (1995). “Innovation, Growth and Survival,” International Journal of Industrial Organization 13(4), 441-457. Audretsch, David and Talat Mahmood (1995). “New Firm Survival: New Results Using a Hazard Function,” Review of Economics and Statistics 77(1), 97-103. Bandick Roger and Holger Görg (2010). “Foreign Acquisition, Plant Survival, and employment Growth,” Canadian Journal of Economics 43(2), 547-573. Bernard, Andrew and J. Bradford Jensen (2007). “Firm Structure, Multinationals, and Manufacturing Plant Deaths,” Review of Economics and Statistics 89(2), 193-204. Bernard, Andrew, Redding, Stephen and Peter Schott (2009). “Products and Productivity,” Scandinavian Journal of Economics 111(4), 681-709. Brenton, Paul, Saborowski, Christian, and Erik von Uexkull (2010). “What Explains the Low Survival Rate of Developing Country Export Flows?” World Bank Economic Review 24(3), 474- 499. Breslow, N. (1974). “Covariance Analysis of Censored Survival Data,” Biometrics 30(1), 89-99. 33          446 The World Bank. Boschma, Ron, Minondo, Asier, and Mikel Navarro (2012). “Related Variety and Regional Growth in Spain,” Papers in Regional Science 91(2), 241-256. Caves, Richard (1998). “Industrial Organization and New Findings on the Turnover and Mobility of Firms,” Journal of Economic Literature 36(4), 1947-1982. Cefis, Elena and Orietta Marsili (2006). “Survivor: The Role of Innovation in Firms’ Survival,” Research Policy 35(5), 626-641. Chen, Ming-Yuan (2002). “Survival Duration of Plants: Evidence from the US Petroleum Refining Industry,” International Journal of Industrial Organization 20(4), 517-555. Cox, D. (1972). “Regression Models and Life Tables,” Journal of the Royal Statistical Society 34(2), 187-220. Cox, D. and David Oakes (1984). Analysis of Survival Data. London: Chapman & Hall/CRC. Crespi, Gustavo and Pluvia Zuniga (2012). “Innovation and Productivity: Evidence from Six Latin American Countries,” World Development 40(2), 273-290. Disney, Richard, Haskel, Jonathan, and Ylva Heden (2003). “Entry, Exit and Establishment Survival in UK Manufacturing,” Journal of Industrial Economics 51(1), 91-112. Doms, Mark, Dunne, Timothy, and Mark Roberts (1995). “The Role of Technology Use in the Survival and Growth of Manufacturing Plants,” International Journal of Industrial Organization 13(4), 523-542. Dunne, Timothy, Roberts, Mark and Larry Samuelson (1989). “The Growth and Failure of U.S. Manufacturing Plants,” Quarterly Journal of Economics, 104(4), 671-698. Ericson, Richard and Ariel Pakes (1995). “Markov-Perfect Industry Dynamics: A Framework for Empirical Work,” Review of Economic Studies 62(1), 53-82. 34          446 The World Bank. Esteve-Pérez, Silviano, Llopis, Amparo, and Juan Llopis (2004). “The Determinants of Survival of Spanish Manufacturing Firms,” Review of Industrial Organization 25(3), 251-273. Esteve-Pérez, Silviano, Requena-Silvente, Francisco, and Vicente Pallardó-Lopez (2013). “The Duration of Firm-Destination Export Relationships: Evidence from Spain 1997-2006,” Economic Inquiry 51(1), 159-180. Fernandes, Ana and Caroline Paunov (2012). “Foreign Direct Investment in Services and Manufacturing Productivity Growth: Evidence for Chile,” Journal of Development Economics 97(2), 305-321. Fernandes, Ana and Caroline Paunov (2013). “Does Tougher Import Competition Foster Product Quality Upgrading?,” Canadian Journal of Economics 46(4), 1232-1264. Girma, Sourafel, Görg, Holger, and Eric Strobl (2007). “The Effects of Government Grants on Plant Survival: A Micro-Econometric Analysis,” International Journal of Industrial Organization 25(4), 701-720. Gourville, John (2006). “The Curse of Innovation: A Theory of Why Innovative New Products Fail in the Marketplace,” Harvard Business School Marketing Research Papers No. 05-06. Hall, Bronwyn (1987). “The Relationship between Firm Size and Firm Growth in the US Manufacturing Sector,” Journal of Industrial Economics 35(4), 583–605. Hausmann, Ricardo, Hwang, Jason and Dani Rodrik (2007). “What you Export Matters,” Journal of Economic Growth 12(1), 1-25. Heckman, James and B. Singer (1984). “The Identifiability of the Proportional Hazard Model,” Review of Economic Studies 51(2), 231-241. Hess, Wolfgang and Maria Persson (2011). “Exploring the Duration of EU Imports,” Review of World Economics 147(4), 665-692. 35          446 The World Bank. Hess, Wolfgang and Maria Persson (2012). “The Duration of Trade Revisited: Continuous-Time vs. Discrete-Time Hazards,” Empirical Economics 43(3), 1083-1107. Hidalgo Cesar, Klinger Bailey, Barabási Albert-Laszlo, and Ricardo Hausmann (2007). “The Product Space Conditions the Development of Nations,” Science 317(5837) 482–487. Hopenhayn, Hugo (1992). “Entry, Exit, and Firm Dynamics in Long Run Equilibrium,” Econometrica 60(5), 1127-1150. Hosmer, David, Lemeshow, Stanley, and Susanne May (2008). Applied Survival Analysis: Regression Modeling of Time to Event Data, Wiley Series in Probability and Statistics. New Jersey. Jarreau, Joachim and Sandra Poncet (2012). “Export Sophistication and Economic Growth: Evidence from China,” Journal of Development Economics 97(2), 281-292. Jenkins, Stephen (1995). “Easy Estimation Methods for Discrete-Time Duration Models,” Oxford Bulletin of Economics and Statistics 57(1), 129-38. Jovanovic, Boyan (1982). “Selection and the Evolution of Industry,” Econometrica 50(3), 649- 670. Kasahara, Hiroyuki and Joel Rodrigue (2008). “Does the Use of Imported Intermediates Increase Productivity? Plant-Level Evidence,” Journal of Development Economics 87(1), 106-118. Kiefer, Nicholas (1988). “Economic Duration Data and Hazard Functions,” Journal of Economic Literature 26(2), 646-679. Klein, John and Melvin Moeschberger. (1997). Survival Analysis. Techniques for Censored and Truncated Data. New York: Springer. Lopez, Ricardo (2006). “Imports of Intermediate Inputs and Plant Survival,” Economics Letters 92(1), 58-62. 36          446 The World Bank. Mairesse, Jacques and Pierre Mohnen (2010). “Using Innovation Surveys for Econometric Analysis,” NBER Working Paper Number 15857. Manjon-Antolin, Miguel and Josep-Maria Arauzo-Carod (2008). “Firm Survival: Methods and Evidence,” Empirica 35(1), 1-24. Mata, Jose and Pedro Portugal (1994). “Life Duration of New Firms,” Journal of Industrial Economics 42(3), 227-245. McCall, Brian (1994). “Testing the Proportional Hazards Assumption in the Presence of Unmeasured Heterogeneity,” Journal of Applied Econometrics 9(3), 321-334. Navarro, Lucas (2012). “Plant Level Evidence on Product Mix Changes in Chilean Manufacturing,” Journal of International Trade and Economic Development 21(2), 165-195. OECD-Eurostat (2005). Oslo Manual: Guidelines for Collecting and Interpreting Innovation Data. Third Edition. OECD, Paris and Eurostat, Luxembourg. Paunov, Caroline. (2011). “Imports, Innovation and Employment after Crisis - Evidence from a Developing Country,” OECD Science, Technology and Industry Working Paper N° 2011/05. Prentice, R. and L. Gloeker (1978). “Regression Analysis of Grouped Survival Data with Application to Breast Cancer Data,” Biometrics 34(1), 57-67. Puga, Diego and Daniel Trefler (2010). “Wake Up and Smell the Ginseng: International Trade and the Rise of Incremental Innovation in Low-Wage Countries,” Journal of Development Economics 91(1), 64-76. Shiferaw, Admasu (2009). “Survival of Private Sector Manufacturing Establishments in Africa: The Role of Productivity and Ownership,” World Development 37(3), 572-584. Strotmann, Harald (2007). “Entrepreneurial Survival,” Small Business Economics 28(1), 87–104. 37          446 The World Bank. Sueyoshi, Glenn (1995). “A Class of Binary Response Models for Grouped Duration Data,” Journal of Applied Econometrics 10(4), 411-431. Tybout, James (2000). “Manufacturing Firms in Developing Countries: How Well Do They Do, and Why?,” Journal of Economic Literature 38(1), 11-44. van den Berg, Gerard (2011). “Duration Models: Specification, Identification, and Multiple Durations,” in James Heckman and Edward Leamer (Eds.) Handbook of Econometrics vol. 5 (pp. 3381-3460). Amsterdam: North-Holland. Zhang, Mingqian and Pierre Mohnen (2013). “Innovation and Survival of New Firms in Chinese Manufacturing, 2000-2006,” UNU-MERIT Working Paper Series 057. 38          446 The World Bank. Figure 1: Kaplan-Meier Survival Estimates Panel A. Multi-Product Plants 1.00 0.90 Multi-product Innovators 0.80 Proportion Surviving 0.70 Multi-product Non-innovators 0.60 0.50 0.40 0.30 0.20 0.10 0.00 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Number of Years Panel B. Single-Product Plants 1.00 0.90 Single-product Innovators 0.80 Proportion Surviving Single-product Non-innovators 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Number of Years Notes: each of the figures shows the Kaplan-Meier estimator for the survivor function that is the j probability of survival up to period t and after and is obtained as ˆ (t )  S i  (ni  hi ) / ni where n is the i 1 population alive in ti and h is the number of failures in ti (Kiefer, 1988). The number of years in the X-axis designates age categories: our sample includes plants that range in age from 1 to 25 years old (where age is measured relative to 1979). The graph shows for each age category the 39          446 The World Bank. probability of survival for innovators and for non-innovators focusing on multi-product plants in Panel A and on single-product plants in Panel B. 40          446 The World Bank. Table 1: Descriptive Statistics Plant Product Plant Product Number of Product Product Innovation Innovation Plant Exit Plant Product New Innovation by Innovation by Accounting Accounting for Rate (%) Innovation (%) Products by Multi-Product Single- Product More than 50% Less than 50% Plant Plants (%) Plants (%) of Revenues of Revenues (%) (%) Full Sample 0.09 0.14 0.22 0.13 0.01 0.04 0.10 Food, Beverages and Tobacco (ISIC 31) 0.08 0.08 0.12 0.08 0.00 0.01 0.07 Textile, Wearing Apparel and Leather Industries (ISIC 32) 0.11 0.15 0.26 0.14 0.01 0.03 0.12 Wood and Wood Products, Including Furniture (ISIC 33) 0.11 0.21 0.35 0.19 0.02 0.08 0.13 Paper and Paper Products, Printing and Publishing (ISIC 0.08 0.13 0.21 0.12 0.01 0.04 0.09 34) Chemicals and Chemical, Petroleum, Coal, Rubber and 0.07 0.15 0.25 0.14 0.01 0.04 0.11 Plastic Products (ISIC 35) Non-Metallic Mineral Products,except Products of 0.09 0.11 0.16 0.10 0.02 0.04 0.07 Petroleum and Coal (ISIC 36) Basic Metal Industries (ISIC 37) 0.06 0.22 0.31 0.18 0.04 0.10 0.12 Fabricated Metal Products, Machinery and Equipment 0.08 0.17 0.29 0.16 0.02 0.05 0.12 (ISIC 38) Other Manufacturing Industries (ISIC 39) 0.09 0.15 0.20 0.15 0.00 0.01 0.14 Notes: For the full sample and for each 2-digit industry, the numbers shown in the table are averages calculated across the sample period 1996-2003. Conditional on engaging in product innovation, the average number of new products is 1.6 per plant. 41          446 The World Bank. Table 2: Baseline Results on Innovation and Plant Death Discrete-Time Hazard Models for Plant Exit with Random Effects Cloglog Probit Logit Cloglog Probit Logit (1) (2) (3) (4) (5) (6) Product Innovation -0.216*** -0.135*** -0.240** (0.077) (0.051) (0.093) Number of New Products -0.110** -0.061** -0.113** (0.039) (0.025) (0.046) Plant Controls Plant Sales Growth -0.129 -0.441*** -0.838*** -0.127 -0.441*** -0.838*** (0.160) (0.042) (0.076) (0.160) (0.042) (0.076) Plant Capital Intensity -0.067*** -0.051*** -0.091*** -0.068*** -0.051*** -0.091*** (0.019) (0.015) (0.028) (0.020) (0.015) (0.028) Multi-Plant -3.047*** -0.675*** -1.294*** -3.045*** -0.675*** -1.295*** (0.636) (0.111) (0.206) (0.637) (0.111) (0.206) Plant Size -0.784*** -0.738*** -1.307*** -0.788*** -0.739*** -1.310*** (0.161) (0.133) (0.239) (0.161) (0.133) (0.239) Plant Size Squared 0.076*** 0.060*** 0.103*** 0.0760*** 0.0599*** 0.104*** (0.023) (0.018) (0.032) (0.023) (0.018) (0.032) Plant Initial Size -0.162 -0.076 -0.173 -0.161 -0.076 -0.173 (0.184) (0.144) (0.260) (0.184) (0.144) (0.260) Plant Initial Size Squared 0.045* 0.026 0.051 0.045* 0.026 0.052 (0.024) (0.019) (0.034) (0.024) (0.019) (0.034) Industry Controls Industry Sales Growth -0.030 -0.04 -0.058 -0.033 -0.041 -0.060 (0.084) (0.054) (0.099) (0.084) (0.054) (0.099) Industry Average Innovation 0.685** 0.473** 0.846** 0.650** 0.442** 0.798** (0.314) (0.207) (0.378) (0.312) (0.206) (0.375) Industry Herfindahl Index -0.576 -0.410 -0.770 -0.578 -0.410 -0.772 (0.525) (0.360) (0.656) (0.525) (0.360) (0.657) Industry Herfindahl Index Squared 0.748 0.537 1.004 0.746 0.534 1.001 (0.652) (0.447) (0.813) (0.652) (0.447) (0.814) 4-Digit Industry Fixed Effects Yes Yes Yes Yes Yes Yes Region Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 19,439 19,439 19,439 19,439 19,439 19,439 Log-Likelihood -5,496 -5,536 -5,531 -5,496 -5,537 -5,531 42          446 The World Bank. Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The specifications have a binary dependent variable is equal to 1 in the year of exit for plants that exit and 0 otherwise. The table shows marginal effects: for dummy variables the marginal effect is the change in the probability of exit associated with a change in the variable from 0 to 1 and for continuous variables the marginal effect is the marginal change in the probability of exit associated with a change in the variable evaluated at the means of other variables. The regressors are defined in Appendix Table 1. In addition to the controls shown, the specifications in columns (1) and (4) include also the multi-plant dummy, plant current size and capital intensity interacted with plant age. 43          446 The World Bank. Table 3: Robustness Results on Innovation and Plant Death Alternative Models for Plant Exit or Survival Weibull with Weibull with OLS with Logit with OLS with Logit with Cox Unobserved Cox Unobserved Fixed Effects Fixed Effects Fixed Effects Fixed Effects Heterogeneity Heterogeneity Model for Model for Model for Plant Model for Model for Plant Plant Model for Plant Model for Plant Plant Exit Survival for 6 Plant Exit Hazard of Exit Survival for 6 Hazard of Exit Hazard of Exit Hazard of Exit Years Years (1) (2) (3) (4) (5) (6) (7) (8) Product Innovation -0.200*** -0.177** -0.013* 0.079*** (0.073) (0.073) (0.007) (0.022) Number of New Products -0.091** -0.007** -0.081** 0.079* (0.040) (0.004) (0.039) (0.043) 4-Digit Industry Fixed Effects Yes Yes Yes Yes Region Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Plant Fixed Effects Yes Yes Yes Yes Crisis-Year Control Yes Yes Observations 19,439 19,439 19,439 5,088 19,439 19,439 19,439 5,088 Log-Pseudolikelihood -11,744 -1,797 -11,744 -1,797 44          446 The World Bank. Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The table shows in columns (1), (3), (5) and (7) hazard ratios (exponentiated coefficients) which represent the constant proportional effect of the innovation regressor on the conditional probability of completing a survival spell and negative values imply that the regressor is associated with a lower hazard rate (higher survival probability). The table shows coefficients in columns (2) and (6) and marginal effects in columns (4) and (8): for dummy variables the marginal effect is the change in the probability of exit associated with a change in the variable from 0 to 1. The innovation regressors are defined in Appendix Table 1. The specifications include also the plant controls and industry controls shown in Table 2. In columns (3) and (7) unobserved heterogeneity is accounted for through the use of a mixture model where the hazard function is multiplied by a plant-specific random variable assumed to follow a gamma distribution. 45          446 The World Bank. Table 4: Results on Characterization of Innovation and Plant Death Hazard Models for Plant Exit with Random Effects Cloglog Cloglog Cloglog (1) (2) (3) Product Innovation * Exported -0.614** (0.255) Product Innovation * Non-Exported -0.176** (0.080) Product Innovation * Prior Investments in Machinery -0.434*** (0.107) Product Innovation * No Prior Investments in Machinery 0.010 (0.110) Product Innovation * Imported Intermediate Inputs -0.566*** (0.191) Product Innovation * No Imported Intermediate Inputs -0.143* (0.083) 4-Digit Industry Fixed Effects Yes Yes Yes Region Fixed Effects Yes Yes Yes Year Fixed Effects Yes Yes Yes P-Value for Difference in Innovation Marginal Effects 0.10 0.00 0.04 Observations 19,439 19,216 19,439 Log-Likelihood -5,494 -5,332 -5,494 46          446 The World Bank. Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The specifications have a binary dependent variable is equal to 1 in the year of exit for plants that exit and 0 otherwise. The table shows marginal effects: for dummy variables the marginal effect is the change in the probability of exit associated with a change in the variable from 0 to 1. The innovation regressors are defined in Appendix Table 1. The specifications include also the plant controls and industry controls shown in Table 2 and the multi-plant dummy, plant current size and capital intensity interacted with plant age. The p-value shown in each column tests the null hypothesis that the difference in the marginal effects of the two innovation variables included as regressors in the column is statistically insignificant. 47          446 The World Bank. Table 5: Results on Diversification, Technical and Market Risk, Innovation and Plant Death Discrete-Time Hazard Models for Plant Exit with Random Effects Cloglog Cloglog Cloglog Cloglog Cloglog (1) (2) (3) (4) (5) Product Innovation * Multi-Product Plants -0.289*** (0.083) Product Innovation * Single-Product Plants 0.404** (0.190) Product Innovation Accounting for Less than 50% of -0.292*** Revenues (0.091) Product Innovation Accounting for More than 50% of -0.019 Revenues (0.132) Product Innovation Closer to Past Plant Expertize -0.268* (0.146) Product Innovation More Distant from Past Plant Expertize -0.244 (0.153) Innovation with Few Competitor Innovators -0.233** (0.094) Innovation with Many Competitor Innovators 0.124 (0.144) Innovation in Industry with Lower Sales Volatility -0.332*** (0.113) Innovation in Industry with Higher Sales Volatility -0.112 (0.103) 4-Digit Industry Fixed Effects Yes Yes Yes Yes Yes Region Fixed Effects Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes P-Value for Difference in Innovation Marginal Effects 0.00 0.08 0.91 0.05 0.14 Observations 19,439 19,439 18,029 15,887 19,439 Log-Likelihood -5,491 -5,495 -5,118 -4,505 -5,495 48          446 The World Bank. Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The specifications have a binary dependent variable is equal to 1 in the year of exit for plants that exit and 0 otherwise. The table shows marginal effects: for dummy variables the marginal effect is the change in the probability of exit associated with a change in the variable from 0 to 1. The innovation regressors are defined in Appendix Table 1. The specifications include also the plant controls and industry controls shown in Table 2 and the multi-plant dummy, plant current size and capital intensity interacted with plant age. The p-value shown in each column tests the null hypothesis that the difference in the marginal effects of the cautious innovation and risky innovation proxies included in the column is statistically insignificant. 49          446 The World Bank. Table 6: Results on Plant Performance and Innovation Dynamic Dynamic Estimation - Estimation - OLS Estimation - Levels OLS Estimation - Levels First First Differences Differences No With With controls With controls No With With controls With controls controls controls and exit and exit controls controls and exit and exit Plant Labor Productivity Plant Profit Rates (1) (2) (3) (4) (5) (6) (7) (8) Product Innovation 0.021** 0.014* 0.013* 0.011* 0.008 0.007 0.007 0.003 (0.009) (0.008) (0.008) (0.006) (0.010) (0.009) (0.009) (0.014) Plant Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Anderson canon. corr. LM statistic P-Value 0.00 0.00 (Underidentification test) Cragg-Donald Wald F statistic (Weak 635.65 581.28 identification test) Observations 20,480 18,861 18,861 5,881 21,153 19,423 19,423 6,023 R-squared 0.89 0.94 0.94 0.63 0.69 0.69 Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The dependent variables and the innovation regressors are defined in Appendix Table 1. All Anderson-Hsiao specifications reported are valid as indicated by the results of both the Anderson canon. corr. LM statistic and Cragg-Donald Wald F statistic test results reported. 50          446 The World Bank. Table 7: Results on Diversification, Technical Risk, Innovation and Performance Dynamic Estimation Plant Labor Plant Profit Plant Labor Plant Profit Plant Labor Plant Profit Plant Labor Plant Profit Productivity Rates Productivity Rates Productivity Rates Productivity Rates (1) (2) (3) (4) (5) (6) (7) (8) Product Innovation * Multi-Product Plants 0.010 -0.009 (0.006) (0.014) Product Innovation * Single-Product Plants 0.025 0.125*** (0.019) (0.043) Product Innovation Accounting for Less than 0.011* -0.004 50% of Revenues (0.007) (0.015) Product Innovation Accounting for More than 0.011 0.025 50% of Revenues (0.011) (0.025) Innovation with Few Competitor Innovators 0.011 -0.006 (0.009) (0.020) Innovation with Many Competitor Innovators 0.036* 0.001 (0.019) (0.042) Innovation in Industry with Lower Sales 0.005 -0.003 Volatility (0.008) (0.018) Innovation in Industry with Higher Sales 0.018** 0.010 Volatility (0.008) (0.019) P-Value for Difference in Innovation Marginal 0.45 0.00 0.97 0.29 0.24 0.88 0.25 0.62 Effects Plant Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Anderson canon. corr. LM statistic P-Value 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 (Underidentification test) Cragg-Donald Wald F statistic (Weak 629.52 581.20 630.70 580.77 301.93 170.80 629.75 581.15 identification test) Observations 5,881 6,023 5,881 6,023 3,118 3,223 5,881 6,023 R-squared 0.76 0.16 0.78 0.18 0.76 0.16 51          446 The World Bank. Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The dependent variables and the innovation regressors are defined in Appendix Table 1. All Anderson and Hsiao (1982) specifications reported are valid as indicated by the results of both the Anderson canon. corr. LM statistic and Cragg-Donald Wald F statistic test results reported. 52          446 The World Bank. Appendix Appendix Table 1. Variable Definitions and Summary Statistics Variables for Baseline Survival Mean Exit Variable equals 1 if the plant is in the sample in year t but not in year t+1, and 0 otherwise. 0.09 Variable equals 1 if the plant produces a 7-digit ISIC product in year t that it did not produce Product Innovation (Dummy) 0.14 in year t-1 nor in any earlier sample year year up to t-1, and 0 otherwise. Number of 7-digit ISIC products that the plant produces in year t that it did not produce in Number of New Products (Continuous) 0.14 year t-1 nor in any earlier sample year year up to t-1 . Variable equals 1 if the plant is part of a firm with multiple plants (establishments), and 0 Multi-Plant Dummy 0.08 otherwise. Plant Size Logarithm of the total number of workers of the plant. 3.56 Logarithm of the total number of workers of the plant in its initial year in the ENIA sample Plant Initial Size 3.91 (from 1979 onwards). Logarithm of the ratio of capital to the total number of workers of the plant. Capital is Plant Capital Intensity 8.64 constructed as defined in Fernandes and Paunov (2008). Plant Sales Growth Logarithm of the difference in plant sales between year t and year t-1 . 0.00 Logarithm of the difference in real sales of the 6-digit ISIC industry between year t and year Industry Sales Growth t 0.02 -1 . The deflators for nominal sales are described in Fernandes and Paunov (2012). H*=(H-1/N)/(1-1/N) where H is the Herfindahl index computed as the sum of the squares of Industry Normalized Herfindahl Index the market shares of all N plants in the 6-digit ISIC industry and year. H* ranges from 0 to 1 0.13 with larger values indicating higher concentration. Industry Average Innovation Average share of plants introducing new products in each 6-digit ISIC industry and year. 0.09 Variables for Payoffs Logarithm of the ratio of plant sales deflated by plant-specific price indices to the total number of workers. Plant-specific price indices are obtained as a weighted average of the growth in prices of each plant's products based on Tornquist indices as in Eslava et al. (2004). Prices of Plant Labor Productivity 9.54 each plant's products are obtained as the ratio of the value of sales of each product to the quantity sold of each product. Values above or below the top and bottom 1 percentile for every year are windsorized. Ratio of plant profits (equal to total plant sales minus materials costs, electricity costs, Plant Profit Rates expenditures on wages and wage benefits) to plant sales. Ratios above 0.8 and below -0.8 0.13 are windsorized. 53          446 The World Bank. Additional Variables for Robustness Variable equals 1 if the plant produces a 6-digit ISIC product in year t that it did not produce Product Innovation 6-digit (Dummy) 0.11 in year t-1 nor in any sample year year before t-1 , and 0 otherwise. Number of New Products 6-digit Number of 6-digit ISIC products that the plant produces in year t that it did not produce in 0.15 (Continuous) year t-1 nor in any sample year year before t-1 . Plant Foreign Ownership Status Variable equals 1 if the plant has a positive share of foreign capital, and 0 otherwise. 0.05 Plant Export Status Variable equals 1 if the plant exports a positive share of its output, and 0 otherwise. 0.22 Ratio of the number of new plants that operate in year t but not in year t-1 to the total Industry Entry Rate 4.90 number of plants in the 6-digit industry in year t (in percentage). Industry Capital Intensity Average across plants of the variable "Plant Capital Intensity" in each 6-digit industry. 8.63 Industry Advertising to Sales Ratio Average ratio of advertising expenditures to sales in each 6-digit industry (in percentage). 0.85 Additional Variables for Characterizing Innovation Product Innovation * Exported [Non- Variable equals 1 if at least one [none of ] of the new products of plant i is exported in year 0.02 [0.12 ] Exported ] t , and 0 otherwise. Product Innovation * Prior [No Prior ] Variable equals 1 if the plant invested [did not invest ] in machinery prior to its first product 0.08 [0.05 ] Investments in Machinery innovation and 0 otherwise. Product Innovation * Imported [No Variable equals 1 if the plant imported [did not import ] intermediate inputs prior to its first 0.04 [0.10 ] Imported ] Intermediate Inputs product innovation, and 0 otherwise. 54          446 The World Bank. Additional Variables for Risk Product Innovation * Multi-Product [Single- Variable equals 1 if the plant introduces a new 7-digit product in year t and is a multi-product 0.13 [0.01 ] Product ] Plants [single-product ] plant initially. Product Innovation Accounting for Less Variable equals 1 if the plant introduces in year t new products and these account for less 0.10 [0.04 ] [More ] than 50% of Revenues than [more than or equal to ] 50% of the total revenues of the plant. Variable equals 1 if the plant introduces in year t new products whose distance to the Product Innovation Closer to [More weighted average of the plant's past products measured by the product proximity index based 0.03 [0.03 ] Distant From ] Past Plant Expertise on Hidalgo et al. (2007) and Boschma et al. (2012) is above [below ] the median across all products in year t , and 0 otherwise. Variable equals 1 if the plant introduces in year t a new product with less than or equal to Product Innovation with Few [Many ] 0.11 [more than ] 10 other firms introducing the same product at the 7-digit level in year t or year [0.03 ] Competitor Innovators t+1 and 0 otherwise. Variable equals 1 if a plant introduces in year t a new product in a 3-digit industry with a Innovation in Industry with Larger 0.07 standard deviation of real sales during the period 1992-2004 that is above [below ] the [0.06 ] [Smaller ] Sales Volatility median value across all 3-digit industries and 0 otherwise. Number of New Products * Multi-Product Number of new products introduced in year t if the plant is a multi-product [single-product ] 0.21 plant initially. [0.01 ] [Single-Product ] Plants Variable equals 1 if the plant introduces in year t new products and as a consequence its total Product Innovation Adding to [Replacing ] 0.08 number of products increases [decreases or remains unchanged ] relative to year t-1 , and 0 [0.06 ] Existing Products otherwise. Variable equals 1 if the plant introduces in year t new products with a product sophistication Product Innovation More [Less ] 0.04 index based on Hausmann et al. (2007) and Jarreau and Poncet (2012) that is above [below ] [0.04 ] Sophisticated the median of the product sophistication index across all new products introduced in year t. Product Innovation in a New [Old ] 6-digit Variable equals 1 if the plant introduces in year t new products in a 6-digit industry that it did 0.02 [0.11 ] Industry not produce [it produced ] in any earlier sample year up to t-1 , and 0 otherwise. Variable equals 1 if the plant introduced in year t new products that have never been 0.01 Product Innovation New [Old ] to Chile produced by any plant [were already produced by some plant ] in Chile in any earlier [0.13 ] sample year year up to t-1 , and 0 otherwise. Price of New Product Above [Below ] the Variable equals 1 for new products introduced by the plant in year t with a unit value above 0.06 [0.06 ] Median [below ] the median unit value across all other plants producing the same product in in year t . 55          446 The World Bank. Appendix Table 2: Additional Robustness Results on Innovation and Plant Death Panel A. Product Innovation Discrete-Time Hazard Models for Plant Exit with Random Effects Excluding Plants Additional Single Plants Innovation 6- Additional Plant with Less than Industry Only Digit Controls 15 Employees Controls Cloglog Cloglog Cloglog Cloglog Cloglog (1) (2) (3) (4) (5) Product Innovation -0.187** -0.235** -0.218*** -0.099** (0.079) (0.092) (0.081) (0.039) Product Innovation 6-Digit -0.216** (0.086) 4-Digit Industry Fixed Effects Yes Yes Yes Yes Yes Region Fixed Effects Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Observations 17,906 18,862 19,439 19,439 19,439 Log-Likelihood -5,244 -4,006 -5,497 -5,332 -5,492 56          446 The World Bank. Panel B. Number of New Products Discrete-Time Hazard Models for Plant Exit with Random Effects Excluding Plants Additional Single Plants Innovation 6- Additional Plant with Less than Industry Only Digit Controls 15 Employees Controls Cloglog Cloglog Cloglog Cloglog Cloglog (1) (2) (3) (4) (5) Number of New Products -0.091** -0.087* -0.103** -0.099** (0.039) (0.045) (0.040) (0.039) Number of New Products 6-Digit -0.180*** (0.055) 4-Digit Industry Fixed Effects Yes Yes Yes Yes Yes Region Fixed Effects Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Observations 15,531 19439 17,906 18,862 19,439 Log-Likelihood -5,244 -4,008 -5,494 -5,332 -5,493 Notes: Robust standard errors in parentheses. *, **, and *** indicate significance at 10%, 5%, and 1% confidence levels, respectively. The specifications have a binary dependent variable is equal to 1 in the year of exit for plants that exit and 0 otherwise. The table shows marginal effects: for dummy variables the marginal effect is the change in the probability of exit associated with a change in the variable from 0 to 1. The innovation regressors are defined in Appendix Table 1. The specifications include also the plant controls and industry controls shown in Table 2 and the multi-plant dummy, plant current size and capital intensity interacted with plant age. The 57          446 The World Bank. estimating sample used in column (2) excludes all observations of a plant if the plant reports having less than 15 employees in any of the sample years. 58