WPS4319 Policy ReseaRch WoRking PaPeR 4319 Product Innovation by Incumbent Firms in Developing Economies: The Roles of Research and Development Expenditures, Trade Policy, and the Investment Climate Daniel Lederman The World Bank Development Research Group Trade Team August 2007 Policy ReseaRch WoRking PaPeR 4319 Abstract A model of firm innovation illustrates the effects evidence suggests that the answers are yes and yes, of the threat of imitation and product varieties on but the investment climate affects product innovation a representative firm's decision to invest in research in a manner that is consistent with the presence of and development to produce new product varieties. market failures and state capture. National trade- The model motivates two empirical questions: (1) Is policy distortions appear to reduce the probability of research and development partially correlated with product innovation, and the density of exporting firms firms' propensity to introduce new products or product at the national level also seems to positively affect the innovation in developing countries? (2) Do trade propensity to introduce new products by individual policies and the national investment climate affect firms' firms. The paper discusses some policy implications. propensity for product innovation? The econometric This paper--a product of the Office of the Chief Economist for Latin America/Caribbean--is part of a larger effort in the department to understand the microeconomic determinants of economic growth. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at dlederman@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Product Innovation by Incumbent Firms in Developing Economies: The Roles of Research and Development Expenditures, Trade Policy, and the Investment Climate Daniel Lederman The World Bank Keywords: Product innovation, R&D, investment climate, developing countries JEL Classification: O14, L25 Lederman is Senior Economist with the Development Economics Research Group of the World Bank. Address correspondence to: dlederman@worldbank.org, 1818 H Street NW, Washington, DC 20433; (t) 202-473-9015, (f) 202-522-1159. The author gratefully acknowledges the impeccable research assistance provided by Daniel Chodos and Javier Cravino. Roberto Alvarez, Pablo Fajnzylber, Rodrigo Fuentes, Alvaro González, José L. Guasch, J. Humberto López, and Marcelo Olarreaga provided insightful comments on earlier drafts. Financial support from the World Bank's Latin America and Caribbean Regional Studies Program is gratefully acknowledged. The views expressed in this paper do not represent the views of the World Bank or of its member countries. All remaining errors are the responsibility of the author. 1. Introduction It is so widely recognized that innovation is a key driver of economic growth that it is almost cliché to say so.1 In spite of the extensive literature on the importance of expenditures in research and development (R&D) and science and technology policy to innovation, the distinction between adoption and invention in developing countries should lead us to explore numerous other areas that may pose barriers to the emergence of innovative firms. There is an emerging literature on what can be called "product" innovation, which focuses on the introduction of new products by firms. Hausmann and Rodrik (2003), for example, present a theoretical framework where market failures affecting the introduction of new export products in developing countries might be more severe than those affecting innovation in the developed countries, because in the latter most innovations can be patented, thus providing at least a partial institutional solution to the appropriability problem that inhibits private sector innovation. In developing countries, where most innovations are probably not patentable, other policy instruments would need to be devised to stimulate private-sector investments in product innovation. A related theoretical literature has emphasized the role of entrepreneurship that is responsible for commercializing research outputs, which are then reflected in the introduction of new products (Michelacci 2003). Even in the context of high-income countries, the determinants of product innovation across firms might be different from those of patentable innovation. Criscuolo, Haskel, and Slaughter (2005) find in a panel of firms from the United Kingdom that the correlates of patents and product innovation are different, particularly with respect to the role played by linkages between firms and universities, the latter being more important for patentable innovations. Another example is the study by Aghion et al. (2006) that found that the response of U.K. firms (measured by productivity changes and patenting) to increased competition (due to the regulatory 1Some studies reveal that much of the widening gap between rich and poor countries is due, not to differences in capital investment, but in technological progress. For example, according to Hall and Jones (1999) and Dollar and Wolf (1997), roughly half of cross-country differences in per capita income and growth are driven by differences in total factor productivity, generally associated with technological progress. Easterly and Levine (2003) also argue that productivity differences explain the lion's share of global income differentials. To the extent that productivity is driven by innovation, both patentable and non-patentable, then we can infer that innovation has become an important ingredient in the new growth agenda. Furthermore, there is empirical evidence that the rates of return to investments in R&D can be high (Jones and Williams 1998). At the level of the firm, Klette and Kortum (2004) provide an analytical framework for understanding widely recognized stylized facts linking productivity, firm size, R&D, and patenting across firms. The empirical literature on firm-level R&D, patenting, and productivity is enormous. 1 reforms of the Thatcher government) was different across firms, depending on their distance to the technological frontier (proxied by the productivity gap with respect to the most productive firms in each industry). Yet we still have much to learn about the empirical correlates of product innovation in developing countries. This paper examines the empirical determinants of firm-level innovation in a large sample of manufacturing firms, covering at least 36 and up to 60 developing economies, 8 manufacturing industries, and totaling thousands of firms, depending on the empirical model. More specifically, we address two questions: First, is there evidence of market failures that would justify government involvement to raise private-sector investments in product innovation? In the presence of market failures, aspects of the investment climate associated with the extent of market competition can have unexpected effects on the private firms' propensity to innovate, especially among firms that are farthest from the global technological frontier. For instance, regulatory reforms, as in the U.K. during the Thatcher era, might reduce private sector innovation as the enhanced entry of firms raises the prospects of imitation, thus leading entrepreneurs to reduce their innovation expenditures. Second, is R&D investment correlated with product innovation in developing-country firms? If so, then the "D" in R&D ­ investments in product development -- might be an important correlate of the propensity to innovate by firms in developing countries even when such innovations are not patentable. The evidence discussed herein, which is motivated by a simple model of firm behavior with respect to product innovation, suggests that market failures are empirically noticeable. Thus regulatory reforms can be beneficial for enhancing the diffusion of ideas and technology across firms and thus for productivity growth, but are not enough to stimulate product innovation.2 Also, we find that R&D expenditures, which are mis-measured in the firm data, are highly correlated with the propensity for product innovation by firms in developing countries, although our estimates do not prove that there is a causal effect. Nevertheless, in our view, the main policy implication from these findings is that in the context of reforms that improve the investment 2On the effects of entry regulations on new firm entry and incumbent-firm growth see, for example, Klapper, Laeven, and Rajan (2006). 2 climate, the public sector has an important role to play to stimulate private R&D expenditures associated with product innovation, even in developing economies where R&D is often viewed as a non-critical aspect of private-sector development and R&D is viewed as an important factor mostly for firms in developed economies that are close to the global technological frontier (e.g., Aghion and Howitt 2005). The rest of the paper is organized as follows. Section 2 briefly presents a model of a representative firm's decision to spend resources for product innovation. Section 3 presents the data. Section 4 focuses on the partial correlation between firm-level R&D expenditures and product innovation. Section 5 discusses the econometric strategy and presents the results concerning the role of the investment climate in determining firms' propensity for product innovation. Section 6 summarizes the main findings. 2. A Model of Product Innovation We follow Klette and Kortum (2004) by modeling firms' innovation behavior in terms of an innovation production function with product varieties. A discussion of the model follows its presentation. 2.a. The model Let the innovation production function I(.) depend on R&D expenditures, R, and knowledge capital embodied in the number of product varieties, n, produced by a representative firm: (1) I = g(R,n). If both research expenditures and product varieties (current and in the past) were observed, one could estimate an empirical counterpart of this equation directly. These variables are usually not observed in firm data. The corresponding research cost function, C(.), which is homogenous of degree 1 in I and n can be written as: 3 (2) R = C(I,n) = n ·c n . I This equation can be interpreted as the reverse function of (1). Again, it could be estimated only if product varieties were observed, and additional assumptions are thus required to find an empirical counterpart. Standard assumptions about constant returns to scale in g(.) and c(.) can are used. Equation (2) simply tells us that the total cost of research for product innovation is the product of the number of varieties times the research-cost intensity function, c(.), of each product variety.3 But we do need to introduce mechanisms through which a firm interacts with the market and competitors. Under the assumptions that the price of each variety is exogenous, and that each firm faces an exogenous probability that one of its product varieties will become obsolete or be replaced in the market by a competitor's newer or superior substitute product, the firm's expected profit at any point in time, E( ), can be written as: (3) E( ) = (1- )· n ·P - c + · (n -1)·P - c n , I I n where 0 1 represents the probability of losing one variety. P is the exogenously market- determined average price for the firm's product varieties. The model can be re-written for the general case where the firm faces the threat of losing multiple varieties. The research-cost intensity is the same under both scenarios, with or without losing a product variety, because the firm has already incurred all costs before a competitor's entrance. In order for the firm to incur further research costs, the expected profits would have to be larger than zero, and the corresponding research-cost intensity of the firm in the positive-profits state is: (4) c I P ·(n - ) n n + · (n -1) . 3Klette and Kortum (2004) use the term "innovation intensity" for I(n)/n, where I(.) is the innovation production function. 4 The relationship between research-cost intensity and the number of varieties is indeterminate, as n appears in both the numerator and the denominator. Market prices are expected to be positively correlated with research-cost intensity. Moreover, unlike Hausmann and Rodrik (2003), the appropriability problem does not affect prices, but can be thought as affecting the probability of losing a product variety. This model also predicts that the research-cost intensity of the firm will depend positively on the average price and negatively (and linearly) on the probability of losing one variety to competitors. The function is also inversely proportional to the probability of losing a variety. Since prices are given, innovation expenditures of firms as a share of sales would also depend on the probability of future imitation and on the number of varieties currently produced by the firm. We do observe this variable in firm data. Proxies for market conditions that are likely to be correlated with the average (relative) price of the firm's varieties, such as the growth of manufacturing value added can also be used to capture the effect of market prices. The probability of entry can be thought to be affected by policies, especially the regulatory environment, as suggested by Aghion et al. (2006). 2.b. Discussion The model is quite tractable and intuitive, but it does open the door to many questions of relevance for the empirical work. Although the empirics discussed in the following sections are done with cross-sections of firms, most of the interesting questions are related to dynamics. First, would the predictions of the model change with the introduction of financing costs or real interest rates? The answer is no as long as all firms face the same opportunity costs or real interest rates. Second, how would the firm respond when market demand for and the relative price of its product varieties fall? One response would be to reduce research-cost intensity, but another would be to increase research so as to enhance the chances of raising the number of varieties. The latter would be a "retooling" strategy, whereas the former could be called the "cost- reduction" strategy. Neither strategy is analyzed here, but the point is that either option could be viable, depending on an unspecified production function. One need only assume that firms' 5 overall production function (as opposed to the innovation function analyzed here) is positive with respect to I(.). Third, would the predictions of the model change if the probability of imitation is endogenous with respect to firms' number of varieties or research costs? Probably not, because any plausible strategic game among firms that would need to be modeled would not change the signs of the predictions. What would come out from such a modeling approach is an optimal dynamic path for the firm in terms of different combinations of research-cost intensity changes and number of varieties, rather than the impact of the probability of imitation imposed by other firms' strategic behavior. Thus far the discussion has been interpreting research-cost intensity as referring strictly to R&D expenditures. Since there is a substantial literature on international technology diffusion (e.g. Keller 2004), it is worth asking whether this model would apply to other forms of innovation expenditures. The answer is yes, as the model does not have sufficient structure to distinguish between innovation expenditures to import capital goods or licensing payments to use foreign technologies. Some of the empirical exercises presented later in this paper use data on R&D as well licensing payments as proxies for investments in product innovation. The simplicity of this setup, however, is attractive for empirical analysis, in spite of the open questions discussed above or many others.4 An important advantage is that it explicitly models the direct (equation 1) and the reverse (equation 4) models of innovation expenditures. The following section presents the data that are used to explore the partial correlation between firms' R&D as a share of sales ­ a proxy for research-cost intensity when product prices are exogenous ­ and firms' propensity to introduce a new product in developing countries, as well as to assess the role of the investment climate. The reverse model turns out to be an important tool to estimate the "true" partial correlation between product innovation and research-cost intensity. 3. Data 4One such additional question is whether the average price index in the model should be a relative price, for example, with respect to the price of alternative varieties not produced by the firm. Working with relative prices that are also exogenously given does not change the analysis. 6 We first discuss the data sources and definitions. We then present some descriptive statistics of relevance for the empirical analyses that follow. 3.a. Data definitions and sources The present study characterizes the role of the investment climate within which firms operate, and how it affects product innovation. This is done with data from the World Bank's numerous Investment Climate Surveys (ICS) and Business Environment and Enterprise Performance Surveys (BEEPS). There is substantial overlap between the ICS and BEEPS questionnaires, but there are, however, some differences in their sampling approaches. The ICS tend to focus on manufacturing firms; the BEEPS are drawn from a broad range of economic activities including services (actually the BEEPS database is slightly skew towards services firms). We restricted the coverage of the BEEPS data to firms in the manufacturing sectors. Three sets of variables are used in our regression analyses discussed below, namely firm-, sector- and country-level variables. The first set includes our product innovation proxy that is also de dependent variable: the introduction of a new product. The surveys asked managers whether the firm had introduced a new product during the past two years. Hence our dependent variable is dichotomous. Regarding explanatory variables, firm characteristics that may affect a firm's proclivity to innovate include firm size, measured by the natural logarithm of the average number of permanent and temporary workers and its squared term (to test for a nonlinear relationship); a firm's exporter status, measured by a dummy variable equal to 1 when a firm exports at least 10 percent of its sales; firm ownership, measured by a dummy variable equal to 1 for foreign ownership (when more than 10 percent of assets of the firm are owned by foreigners); and capacity utilization, measured as the average utilization of its productive capacity over the year preceding the survey. The surveys also provide information about the value of R&D expenditures and firm sales. From those data we calculated the share of R&D in sales. Since the literature on innovation has paid much attention to the adoption of foreign technologies, we also use data derived from a question in the surveys that asked managers whether the firm had paid licensing fees during the past two years. This variable is also dichotomous. 7 Sector-level variables include information on trade policies, namely a composite index of the average applied tariffs and its standard deviation. The index was estimated as the first principal components derived from factor analysis. The other trade policy indicator measure at the industry level is the share of tariff lines within each industry that faces one of the so-called "core" non-tariff barriers. These data were taken from Nicita and Olarreaga (2006). Country-level explanatory variables capture various aspects of the investment climate, besides trade policies. They include an index of infrastructure coverage (from the World Development Indicators (WDI) database), institutional quality (from Kaufmann et al. 2005), and real manufacturing GDP growth (also from WDI). In some models, we also include the level of development (GDP per capita from the Penn World Tables). An important explanatory variable for our analysis is a regulatory index capturing the ease of entry, which was calculated from data from the World Bank's Doing Business database. We use principal components analysis to calculate a composite index on infrastructure coverage (including paved roads per square kilometer and telephone lines per capita), the institutional index (including corruption, political stability, and rule of law), and the regulatory index (including Difficulty of Firing Index, Difficulty of Hiring Index and Days for Starting a Business). We also use patent-counts data from Lederman and Saenz (2005), namely the sum (stock) of utility patents granted to researchers in each country during 1963-2000 by the United States Patent and Trademark Office (USPTO) per person. The latter provides a measure of the density of innovative ideas available to firms operating in each country. Some studies, such as Criscuolo et al. (2005), also treat explanatory variables measured at a higher level of aggregation than the firm level as exogenous factors, but these are measured with data from the firm surveys themselves. Our approach is different in this regard, as we use objective data from other sources. As discussed elsewhere in this volume, the use of aggregate variables derived from the same dataset as the firm data can be assumed to be exogenous only under certain conditions, namely firms' deviations from the average must be orthogonal to the average and normally distributed with expected value of zero. We do not have to make any assumptions in this regard, because our data are objectively measured at the country level from data from other sources. The disadvantage of our approach is that we have fewer degrees of 8 freedom to estimate the relevant coefficients of the variables measure at the national level, which is limited by the number of countries. Missing data inevitably introduce ambiguity into the inferences that can be drawn from a study, so another caveat is in order. This section, as well as the regression analyses, relies on variables that were taken from firm-survey questions that are straightforward and with no ambiguity. That is, the question on whether a firm introduced a new product in the previous two years is straightforward. Therefore, it is safe to assume that firms that did not answer this question had not in fact introduced a new product. Although this change in the data is marginal, we do obtain more realistic estimates of the share of firms by country that introduced new products. For example, the percentage of firms reporting a New Product in China changes from 25 percent to 15 percent for all firms, which is a more reasonable share. In other countries, such as Turkey, there are no missing values. The data from the countries in Latin America and the Caribbean (LAC) are a mixed bag, but the same procedure was applied to all. R&D/Sales and licensing payments are likely to be measured with error, and this issue is discussed in the context of the econometric methodology in section 4. 3.b. Descriptive statistics Table 1 presents some descriptive statistics for a sample of 36 countries, which cover the sample used in the econometric analysis discussed in section 5 below. The sample used for the analyses of section 4 is larger, because that estimations of the partial correlation between R&D and product innovation uses country dummies to control for any country-level characteristic rather than specific aspects of the business environment. The number of firms by country and survey year in the various samples appears in the Data Appendix. Hence the restricted sample was due to the availability of data on the other country-level determinants of product innovation discussed above.5 The following paragraphs focus on the restricted sample, because it poses some issues about the representativeness of the sample of firms among developing countries. 5Please note, however, that the regressions reported in Table 2 include the average number of years of education of workers employed by each firm. China and Indonesia do not have these data and thus are not in the samples of 59 and 60 countries that are included in the regressions reported in Table 2. But they do appear in the sample of countries used for the estimation of the models reported in Table 3, and thus appear in Table 1. 9 The sample includes 6 countries that are high-income countries, namely Germany (Eastern after re-unification), Greece, Ireland, Portugal, Spain, and Korea. It includes 13 countries from LAC, 9 from Europe and Central Asia (ECA), 4 from East Asia and the Pacific (EAP) including China, 3 African countries, plus Egypt. Clearly, this sample, which is used for econometric analysis, is not representative across all regions of the world. However, the sample of firms might be representative of manufacturing firms from around the world, especially from developing countries. Since this might not be true, some of the relevant regressions use weights, based on each country's labor force (i.e., population aged 15-64 years). This is reasonable if the number of firms from each country is proportional to the labor force in each. Regarding the incidence of firms innovating via the introduction of a new product, the data show a wide range of country experiences, ranging from 15 percent of Chinese and Egyptian firms to 75 percent in Argentina. It is noteworthy that the percentages for the richer countries in the sample are not above the overall sample average of 43 percent. But LAC's average is above the sample average. In most countries, a large share of firms that reported new products also report R&D expenditures. In the total sample, 60 percent of firms with product innovation also report R&D, whereas only 15 percent of non-innovative firms report some R&D expenditures. This pattern holds for most countries individually for R&D, licensing, export status, and foreign ownership. China is the only exception. In this country, the percentage of non-innovative firms that report R&D expenditures, licensing payments, exporting, and foreign ownership is higher than among the innovative firms.6 Although the high correlation between innovation and the other firm characteristics is expected, it is clear that identifying the partial correlations between the propensity to introduce a new product and the other firm characteristics is important since high correlations among all the firm-level characteristics are also expected. There is no clear relationship between trade policies and the share of innovative firms across countries, however. For example, Argentina appears with 75 percent of firms being innovative, but it also utilizes numerous non-tariff barriers (NTBs) covering, on average across 6The data for China is consistent with product-level export data that suggests that mainland China introduced comparatively few new export products during 1994-2003 (Klinger and Lederman 2006). 10 the 8 manufacturing sectors, slightly over 29 percent of its tariff lines. In contrast, Bolivia has a low NTB coverage rate of about 3 percent, but only 43 percent of firms reported a product innovation. Hence it is possible that trade policy has little to do with product innovation, but econometric estimations might help to clarify this potential link between innovation and trade policy by controlling for other factors that might be correlated with both sets of variables. 4. Is R&D Related to Product Innovation? To answer this question we estimated reduced-form models of product innovation, but also considered the possibility that R&D expenditures and perhaps the sales variables that were recorded in the firm surveys are measured with error. If they are, the standard direct regression model with product innovation as the dependent variable and the R&D/Sales variable (our proxy for research-cost intensity derived from the theoretical model) might be biased, possibly suffering from attenuation bias if the measurement error is random. 4.a. Econometric strategy Due to the dichotomous nature of our variable of interest, the direct empirical model of product innovation can be written as: (5) P(yisc =1| Xisc, X sc, Xc ) = ( Xics +X sc + Xc + isc + c ) , where P is the probability of observing a value of one for product innovation, y. Subscript i represents firms, the s's are manufacturing sectors, and c's are countries. The Xs are matrices of the relevant explanatory variables, measure at the three levels of aggregation (firms, sectors, and countries). The betas, alphas, and deltas are the parameters to be estimated with a Probit estimator, which assumes a standard normal distribution of the relevant parameters with respect to the latent threshold variable. isc is, therefore, the standard white noise error. Below we report results that are robust to heteroskedasticity of regression errors clustered around the observations of each country, c . This correction becomes particularly important for the estimation of the delta parameters associated with industry and country variables when the dependent variable is a micro unit (see Moulton 1990). Allowing for error clusters around industries or industry- 11 countries provides identical coefficient estimates but with lower standard errors than those reported later in the paper. In this case, the variable of interest is R&D/Sales measured at the level of the firm. Since for this exercise we are not interested in uncovering industry- and country-level characteristics that might affect a firm's propensity to innovate, we can safely control for both by including industry/country dummy variables. Since the enterprise surveys for all countries were not implemented in the same year and global economic conditions might affect firm behavior, we also control for survey-year dummies. To assess the influence of measurement errors, we follow Leamer (1978, chapter 8) by estimating the reverse regression model. In this approach the dependent variable becomes R&D/Sales and the dummy variable for product innovation becomes an explanatory variable. If the innovation variable is measured accurately, whereas the R&D/Sales is measured with error, then the inverse of the estimated coefficient from the reverse regression is the "true" partial correlation between product innovation and R&D/Sales. The same strategy can be followed for assessing the partial correlation between product innovations and licensing payments. In this case, however, the licensing variable is also dichotomous, but that does not mean that all firms accurately report whether they made some licensing payments. 4.b. Results The results from the estimation of equation (5) with the appropriate set of dummy variables are presented under the first column of Table 2. The table reports the marginal coefficients, or the elasticities calculated at the sample mean. The R&D/Sales variable is not statistically significant and the point estimate of the elasticity is negligible. Is this unsatisfactory result due to measurement errors? The results from the reverse regression model are presented under the second column. In this case, the estimated Tobit coefficient is highly significant. Furthermore, its inverse implies a rather large partial correlation between R&D/Sales and product innovation. The elasticity of the 12 probability of introducing a new product with respect to R&D/Sales would be around 5, which is the inverse of the reverse-regression elasticity shown in Column 3. And the inclusion of country dummies (in Column 4) does not affect the Tobit coefficient. Columns 5-7 report the corresponding estimates for the licensing variable. In this case, the direct regression results suggest that licensing is positively correlated with product innovation, with and without country dummies. Nevertheless, the implied marginal effect estimated with the reverse regression model (Column 7) is significantly larger, thus also suggesting the there might be measurement errors in the licensing variable as well. We now turn to the analysis of the role of the investment climate. Since the evidence suggests that R&D/Sales and the dichotomous licensing variable are both measured with error, these variables are not included in the analyses. Furthermore, the results reported in Table 2 suggest that the coefficients of R&D and licensing are unaffected by the inclusion of country-dummy variables, thus also indicating that excluding these variables from the following estimations will not affect the coefficients on the investment-climate variables. The underlying assumption is that the introduction of a new product by firms reflects past research expenditures, as in Klette and Kortum (2004). 5. The Role of the Investment Climate The estimation strategy is similar to the one pursued in the previous section. But there are additional complications. 5.a. Estimation strategy As mentioned, we estimate partial correlations to help us characterize the relationship between firm-level probabilities of introducing a new product (i.e., a non-patentable innovation) and firm, sector, and country characteristics. While the estimated partial correlations among the firm-level variables could be due to endogeneity, the results concerning the sector- and country- level variables are less likely to be contaminated by this problem. That is, if each individual firm is too small to determine the level of a country's trade protection or its aggregate level of patents 13 accumulated since 1963, then the corresponding empirical relationships are likely to be due to causal effects.7 Within this framework, we estimate the stylized model of the probability of introducing a new product by firms in equation (5). To deal with the issue of whether our sample of firms is a representative sample of firms of operating in developing economies, we present both unweighted and weighted regressions. In the latter, each country's observations are weighted by the size of its labor force, and thus large countries such as China, Indonesia, and Brazil influence the coefficient estimates to a larger extent than smaller economies. To deal with one potential source of joint endogeneity of the firm- and sector-level variables and the probability of observing a product innovation, we can control for correlated country effects. Woolridge (2005) proposed modeling fixed effects in panel data by including the over-time averages of the unit of analysis as additional explanatory variables.8 In our case, we do not have a time dimension, but we do have the country dimension. Hence we can control for correlated country-specific effects by including the country averages of the variables that are measured at the firm and sector levels. We called these estimations "Quasi Fixed Effects." Finally, it is worth noting that identifying the effects on firms' product innovation of the country-level variables that capture different aspects of the national investment climate might be difficult due to the expected correlation among the relevant variables. For example, countries with good infrastructure coverage can also be expected to have higher incomes per capita, higher innovation densities, etc. If the point estimates seem to be stable across various specifications and the country-level variables of interest are jointly statistically significant, we can find some comfort in our estimates. Hence we also report F-tests for the joint significance of the firm-, sector- and country-level variables. 5.3. Results 7Omitted explanatory variables that are correlated with firm-level product innovation and the aggregate level of trade protection might bias the results. However, we showed in Table 2 that the coefficients of most firm-level variables are unaffected by the inclusion of country-dummy variables, thus suggesting that omitted variables measured at the level of the firm, which are correlated with the observed firm-level explanatory variables, might not be systematically correlated with country characteristics. 8Besides being concerned about unobserved heterogeneity, Woolridge (2005) is also concerned about dynamics; the article is about dynamic Probit models with fixed effects. Our case is simpler, since we do not have dynamics. If we had dynamics, to deal with the endogeneity of the lagged dependent variable, we would need to control for the initial value of the dependent variable. 14 Table 3 presents the regression results. The first two columns contain the results from our baseline models. The first shows the results from the unweighted estimation, the second has the weighted-regression results. The last four columns show the results with Quasi Country Fixed Effects, which, as mentioned above, are captured by the country-level averages of the firm- and sector-level variables. This allows us to distinguish between variables that are associated with product innovation at the various levels of aggregation. The p-values of specification tests for the joint endogeneity of our three sets of variables (measured at the level of the firm, sector and country) appear in the bottom three rows. Among the firm-level variables, the most robust results are associated with the size of the firm as captured by the number of employees (but not its squared term). It is highly significant and positive across all specifications. The magnitude of the coefficient is slightly lower in the weighted regressions, thus suggesting that scale is less important among the countries with the largest populations (e.g., China, Indonesia, and Brazil). Interestingly, foreign ownership is always negative and significant in the weighted regressions, but it also negative in all non- weighted estimations. This indicates that foreign owned firms might not settle in developing countries to undertake product innovations, although it is likely that multinational corporations can produce goods that local firms do not produce. Export status is always statistically significant and positive in the non-weighted regressions, but not in the weighted estimates with Quasi Country Fixed Effects. In those estimations, however, the national share of exporting firms does appear as positive and significant, thus suggesting the firms that operate in countries with a high share of exporters tend to have a higher propensity to undertake product innovations. These results are broadly consistent with emerging evidence concerning the effect of exporting on firm performance, as in the sample of Turkish firms in Yasar and Rejesus (2005). Regarding the trade policy variables measured at the sector level, the import tariff index is negative and significant in the baseline regressions, but it becomes non-significant and changes sign when controlling for the Quasi Country Fixed Effects. But the country-average import-tariff index is negative and significant. These results suggest that it is not the cross-sector variation in tariff policies that affects product innovation, but rather it is the cross-country variance that matters. The NTB coverage rate always appears with a negative sign, but it is never statistically different from zero. The test for joint significance of the trade policy variables does 15 suggest that they are highly significant determinants of product innovation. Thus we can conclude that trade-policy distortions matter for product innovation in general, but the most relevant aspect is probably the use of tariffs, and countries (not sectors) with high import tariffs and tariff dispersion tend to have firms with lower propensities to undertake product innovations. The country-level variables supposedly capture each country's investment climate. Unfortunately, there are very few variables that appear to be statistically significant. The most robust result concerns the regulatory index. It appears with a positive sign and it is statistically significant in all specifications except in the two weighted regressions with Quasi Country Fixed Effects. This result is consistent with the view that market failures affect the propensity of firms to invest in product innovation: As entry becomes more restricted by regulatory policies, the propensity of incumbent firms to introduce new products tends to rise. In terms of the theoretical model, the regulatory environment seems to affect the probability of imitation, thus reducing incentives for product innovation. The lack of significance of this variable in the weighted estimations with Quasi Country Fixed Effects is less worrisome when we look at the test for the joint significance of the country-level variables. They are always jointly highly significant, as reflected in the low p-values of the test of the null of lack of significance at the bottom of Table 3. Furthermore, as mentioned, when we allow for error clustering around industries or industries- countries, the regulatory index is statistically significant across all specifications. Regarding the other national variables, manufacturing GDP growth, the density of patenting activity, and the institutional index always appear with the same signs. The infrastructure variable changes sign in one specification. The positive effect of patent density could be interpreted as indicating the presence of knowledge spillovers, whereby firms that have access to a higher density of commercial ideas tend to have higher propensities to innovate than firms in countries with lower innovation densities. The negative coefficient of manufacturing GDP growth might suggest that product innovation is counter-cyclical, thus supporting the view that firms tend to choose the retooling strategy during downturns, which is consistent with the Schumpeterian view of creative destruction. Finally, the negative coefficients on the institutional quality variable can be interpreted as an indication that firms that reside in countries where governance is dysfunctional can find mechanisms to capture the state in order to impose barriers to competition that might not be reflected in the trade and regulatory policy variables. Again, the 16 discussion of the results pertaining to the national variables is worthwhile because they do seem to be jointly significant, even after controlling for Quasi Country Fixed Effects and the level of development. It is actually surprising that the investment climate variables appeared with consistently estimated signs, even when they do not appear to be individually statistically significant when the regression errors are assumed to be clustered around countries. 6. Concluding Remarks The theoretical model presented above motivated empirical analyses of the determinants of product innovation by firms. Both market conditions and the threat of entry by competitors were shown to be theoretical predictors of the research-cost intensity of firms, even when prices are exogenously determined by market conditions. The empirical analyses of Section 4 suggested that in fact data from 60 countries, covering thousands of firms, support the main prediction of the model: Research-cost intensity tends to be significantly associated with product innovation. The analysis also highlighted a potential pitfall in the firm data, as both R&D/Sales and licensing seem to be measured with error, thus shedding some doubt on the usefulness of direct regression estimates of the innovation function. The investment climate also seems to play an important role for product innovation. But the evidence also highlights market failures that hamper innovation. Of particular relevance in this regard were the results concerning the regulatory barriers to firm entry. Whereas deregulation is desirable to increase competition and knowledge diffusion, the results suggest that other policy instruments are needed to stimulate product innovation, especially after deregulation and trade liberalization. Also, the results concerning the density of patent counts suggest that knowledge spillovers might also be important. From a policy perspective, it is also worth noting that product innovation seems to be counter-cyclical, and thus the budgets of programs to stimulate product innovation need to be protected during downturns, so as to prevent the demise of firms that could have survived through retooling in terms of product innovation. The latter might have social benefits that greatly exceed the private returns, because private agents can benefit from the knowledge embodied in the product innovations of their competitors. 17 Finally, trade-policy distortions seem to hamper product innovation, although it seems that it is cross-national variation in tariff policies rather than inter-sector variation in tariffs that affects product innovation by individual firms. Moreover, the density of exporting firms (that is, the percentage of exporting firms at the national level) also affects positively the propensity to introduce new products by individual firms. Hence an open trade environment with a dynamic and dense export sector seems to be an important ingredient for maintaining an innovative private sector. 18 References Aghion, P., R. Blundell, R. Griffith, P. Howitt, and S. Prantl. 2006. "The Effects of Entry on Incumbent Innovation and Productivity." Harvard University, Cambridge, Massachusetts, http://post.economics.harvard.edu/faculty/aghion/papers.html. Aghion, P., and P. Howitt. 2005. "Appropriate Growth Policy: A Unifying Framework." Harvard University, Cambridge, Massachusetts, http://post.economics.harvard.edu/faculty/aghion/papers.html. Criscuolo, Chiara, Jonathan E. Haskel, and Matthew J. Slaughter. 2005. "Global Engagement and the Innovation Activities of Firms." NBER Working Paper 11479, Cambridge, Massachusetts. Dollar, David, and Edward N. Wolf. 1997. "Convergence of Industry Labor Productivity among Advanced Economies, 1963-1982." In Edward N. Wolff, ed., The Economics of Productivity. United Kingdom: Elgar. Easterly, William, and Ross Levine. 2001. "It's Not Factor Accumulation: Stylized Facts and Growth Models." World Bank Economic Review15(2): 177-219. Hall, Robert E, and Charles I. Jones. 1998. "Why Do Some Countries Produce So Much More Output than Others?" Quarterly Journal of Economics 114: 83-116. Hausmann, Ricardo, and Dani Rodrik. 2003. "Economic Development as Self-Discovery." Journal of Development Economics 72: 603-633. Jones, Charles I., and John Williams. 1998. "Measuring the Social Return to R&D." Quarterly Journal of Economics 113: 1119-35. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi. 2005. "Governance Matters IV: Governance Indicators for 1996­2004." World Bank Policy Research Working Paper 3630, Washington, DC. Keller, Wolfgang. 2004. "International Technology Diffusion." Journal of Economic Literature 42(3): 752-82. Klapper, Leora, Luc Laeven, and Raghuram Rajan. 2006. "Entry Regulation as a Barrier to Entrepreneurship." Journal of Financial Economics 82: 591-629. Klinger, Bailey, and Daniel Lederman. 2006. "Diversification, Innovation, and Imitation inside the Global Technological Frontier." Policy Research Working Paper 3872, World Bank, Washington, DC. Klette, Tor Jakob, and Samuel Kortum. 2004. "Innovating Firms and Aggregate Innovation." Journal of Political Economy 112(5): 986-1018. Leamer, Edward E. 1978. Specification Searches: Ad Hoc Inference with Nonexperimental Data. New York, John Wiley & Sons. Lederman, Daniel, and Laura Saenz. 2005. "Innovation around the World, 1960-2000." World Bank Policy Research Working Paper 3774, The World Bank, Washington DC. Michelacci, Claudio. 2003. "Low Returns in R&D Due to the Lack of Entrepreneurial Skills." Economic Journal 113: 207-225. Moulton, Brent R. 1990. "An Illustration of a Pitfall in Estimating the Effects of Aggregate Variables on Micro Units." The Review of Economics and Statistics 72(2): 334-338. Nicita, A. and M. Olarreaga. 2006. "Trade, Production and Protection 1976-2004." World Bank Policy Research Working Paper 2701, Washington, DC. 19 Woolridge, Jeffrey. 2005. "Simple Solutions to the Initial Conditions Problem in Dynamic, Nonlinear Panel Data Models with Unobserved Heterogeneity." Journal of Applied Econometrics 20(1): 39-54. Yasar, Mahmut, and Roderick M. Rejesus. 2005. "Exporting Status and Firm Performance: Evidence from a Matched Sample." Economics Letters 88: 397-402. 20 Data Appendix Table A1. Number of Firms by Countries and Survey Years ­ Samples of 59 and 60 Countries Used in the Regressions Reported in Table 2 I. Sample of 59 Countries Country Name 2002 2003 2004 2005 2006 Total Albania 0 0 0 42 0 42 Argentina 0 0 0 0 524 524 Armenia 0 0 0 177 0 177 Belarus 0 0 0 9 0 9 Benin 0 0 132 0 0 132 Bolivia 0 0 0 0 291 291 Bosnia and 0 0 0 27 0 27 Herzegovin Brazil 0 1,534 0 0 0 1,534 Bulgaria 0 0 0 35 0 35 Cambodia 0 10 0 0 0 10 Chile 0 0 620 0 0 620 Colombia 0 0 0 0 599 599 Costa Rica 0 0 0 268 0 268 Croatia 0 0 0 57 0 57 Czech Republic 0 0 0 73 0 73 Ecuador 0 86 0 0 0 86 Egypt, Arab Rep. 0 0 805 0 0 805 El Salvador 0 17 0 0 0 17 Estonia 0 0 0 19 0 19 Georgia 0 0 0 22 0 22 Germany 0 0 0 208 0 208 Greece 0 0 0 75 0 75 Guatemala 0 9 0 0 0 9 Guyana 0 0 124 0 0 124 Honduras 0 16 0 0 0 16 Hungary 0 0 0 215 0 215 Ireland 0 0 0 148 0 148 Kazakhstan 0 0 0 220 0 220 Korea, Rep. 0 0 0 173 0 173 Kyrgyz Republic 0 10 0 26 0 36 Latvia 0 0 0 11 0 11 Lithuania 0 0 0 34 0 34 Macedonia, FYR 0 0 0 28 0 28 Madagascar 0 0 0 163 0 163 Mali 0 45 0 0 0 45 Mauritius 0 0 0 101 0 101 Mexico 0 0 0 0 993 993 Moldova 0 18 0 77 0 95 Nicaragua 0 6 0 0 0 6 Oman 0 30 0 0 0 30 Panama 0 0 0 0 163 163 Paraguay 0 0 0 0 257 257 Peru 0 0 0 0 329 329 Poland 0 13 0 353 0 366 21 Portugal 0 0 0 109 0 109 Romania 0 0 0 226 0 226 Russian Federation 0 0 0 78 0 78 Serbia and 0 0 0 41 0 41 Montenegro Slovak Republic 0 0 0 22 0 22 Slovenia 0 0 0 41 0 41 South Africa 0 420 0 0 0 420 Spain 0 0 0 101 0 101 Tajikistan 0 7 0 44 0 51 Turkey 0 0 0 800 0 800 Ukraine 0 0 0 83 0 83 Uruguay 0 0 0 0 259 259 Uzbekistan 0 6 0 34 0 40 Vietnam 0 0 0 403 0 403 Zambia 58 0 0 0 0 58 Total 58 2,227 1,681 4,543 3,415 11,924 Sample of 60 Countries: Country Name 2002 2003 2004 2005 2006 Total Albania 60 0 0 75 0 135 Argentina 0 0 0 0 633 633 Armenia 63 0 0 227 0 290 Azerbaijan 49 0 0 205 0 254 Belarus 42 0 0 53 0 95 Benin 0 0 138 0 0 138 Bolivia 0 0 0 0 356 356 Bosnia and 60 0 0 66 0 126 Herzegovin Brazil 0 1,624 0 0 0 1,624 Bulgaria 47 0 0 58 0 105 Cambodia 0 55 0 0 0 55 Chile 0 0 668 0 0 668 Colombia 0 0 0 0 631 631 Costa Rica 0 0 0 339 0 339 Croatia 35 0 0 68 0 103 Czech Republic 65 0 0 82 0 147 Ecuador 0 111 0 0 0 111 Egypt, Arab Rep. 0 0 962 0 0 962 El Salvador 0 28 0 0 0 28 Estonia 26 0 0 34 0 60 Georgia 34 0 0 43 0 77 Germany 0 0 0 220 0 220 Greece 0 0 0 87 0 87 Guatemala 0 10 0 0 0 10 Guyana 0 0 141 0 0 141 Honduras 0 17 0 0 0 17 Hungary 46 0 0 350 0 396 Ireland 0 0 0 169 0 169 Kazakhstan 50 0 0 347 0 397 Korea, Rep. 0 0 0 197 0 197 Kyrgyz Republic 47 101 0 55 0 203 22 Latvia 23 0 0 28 0 51 Lithuania 40 0 141 43 0 224 Macedonia, FYR 42 0 0 56 0 98 Madagascar 0 0 0 192 0 192 Mali 0 58 0 0 0 58 Mauritius 0 0 0 116 0 116 Mexico 0 0 0 0 1,112 1,112 Moldova 50 103 0 183 0 336 Nicaragua 0 9 0 0 0 9 Oman 0 90 0 0 0 90 Panama 0 0 0 0 231 231 Paraguay 0 0 0 0 368 368 Peru 0 0 0 0 354 354 Poland 111 99 0 523 0 733 Portugal 0 0 0 128 0 128 Romania 82 0 0 381 0 463 Russian Federation 122 0 0 140 0 262 Serbia and 64 0 0 80 0 144 Montenegro Slovak Republic 26 0 0 37 0 63 Slovenia 47 0 0 57 0 104 South Africa 0 479 0 0 0 479 Spain 0 0 0 133 0 133 Tajikistan 47 89 0 59 0 195 Turkey 150 0 0 1,388 0 1,538 Ukraine 136 0 0 178 0 314 Uruguay 0 0 0 0 335 335 Uzbekistan 50 100 0 70 0 220 Vietnam 0 0 0 1,384 0 1,384 Zambia 79 0 0 0 0 79 Total 1,693 2,973 2,050 7,851 4,020 18,587 Source: Author's calculations based on data from the World Bank's Enterprise Surveys database, http://www.enterprisesurveys.org/ Table A2. Number of Firms by Countries and Survey Years ­ Sample of 36 Countries Used in the Regressions Reported in Table 3 Country Name 2002 2003 2004 2005 2006 Total Argentina 0 0 0 0 642 642 Bolivia 0 0 0 0 362 362 Brazil 0 1,628 0 0 0 1,628 Chile 0 0 684 0 0 684 China 0 1,324 0 0 0 1,324 Colombia 0 0 0 0 634 634 Costa Rica 0 0 0 339 0 339 Czech Republic 67 0 0 82 0 149 Ecuador 0 123 0 0 0 123 Egypt, Arab 0 0 965 0 0 965 Rep. El Salvador 0 462 0 0 0 462 Germany 0 0 0 220 0 220 Greece 0 0 0 94 0 94 23 Guatemala 0 435 0 0 0 435 Honduras 0 442 0 0 0 442 Hungary 47 0 0 358 0 405 Indonesia 0 595 0 0 0 595 Ireland 0 0 0 172 0 172 Korea, Rep. 0 0 0 213 0 213 Latvia 28 0 0 31 0 59 Lithuania 40 0 160 45 0 245 Mauritius 0 0 0 160 0 160 Mexico 0 0 0 0 1,117 1,117 Peru 0 0 0 0 360 360 Philippines 0 633 0 0 0 633 Poland 111 104 0 527 0 742 Portugal 0 0 0 130 0 130 Romania 82 0 0 386 0 468 Slovenia 47 0 0 57 0 104 South Africa 0 562 0 0 0 562 Spain 0 0 0 134 0 134 Tanzania 0 172 0 0 0 172 Thailand 0 0 1,377 0 0 1,377 Turkey 151 0 0 1,421 0 1,572 Ukraine 136 0 0 180 0 316 Uruguay 0 0 0 0 351 351 Total 709 6,480 3,186 4,549 3,466 18,390 Source: Author's calculations based on data from the World Bank's Enterprise Surveys database, http://www.enterprisesurveys.org/ Data Sources for Trade-Policy and Country-Level Variables Used in Regressions Reported in Table 3. Tariff Index: The composite index was estimated as the first principal component derived from factor analysis of the trade-weighted applied tariffs and their standard deviations within each manufacturing industry. The data come from Nicita and Olarreaga (2006). NTB Coverage Rate: The data come from Nicita and Olarreaga (2006). GDP per capita in 2000: Penn World Tables database. Average Real Manufacturing GDP growth 1998-2003: The data come from the World Bank's database on World Development Indicators (WDI). Stock of patents accumulated during 1960-2000 per worker: The variable is the total number of utility patents granted to the first innovator residing in each country and cited in the patent application at the U.S. Patent and Trademark Office (USPTO) during this period. The data come from Lederman and Saenz (2005). The number of workers is the population 15-64 years of age in 2000. These data come from the World Bank's WDI. Regulation Index: The composite index was estimated as the first principal component derived from factor analysis of the Difficulty of Firing Index, Difficulty of Hiring Index and Starting a Business Time (Days). These variables are the average by country for the years 2003, 2004, and 2005 (although some of these years may be missing for some variables for some countries). The data come from the World Bank's Doing Business database. Infrastructure Index: The composite index was estimated as the first principal component derived from factor analysis of the Total Road Length in 2001 (km) (per square km of 24 surf. area) and Main Telephone Lines (per 1000 habitants) in 2001. The data come from the World Bank's WDI. Institutional Index: The composite index was estimated as the first principal component derived from factor analysis of the Control of Corruption, Political Stability, and the Rule of Law. The data come from Kaufmann, Kraay, and Mastruzzi (2005). 25 Tables Table 1: New Product versus Non-Innovative Firms (% Firms with key Characteristics) and Trade Indicators Paid for Licensed Techn. Made R&D Expenditures From Foreign Firms Exported in previous years Foreign ownership Trade variables Mean Mean of Percentage of weighted standard percentage of firms Firms w/New Firms w/New Firms w/New Firms w/New average Other firms Other firms Other firms Other firms deviation of tariff lines reporting a Product Product Product Product applied tariff the applied with Core New Product rate tariff NTBs Argentina 75 88 17 83 10 86 19 78 9 15.1 5.4 29.47% Bolivia 43 73 7 85 1 83 16 49 7 9.4 1.3 2.64% Brazil 68 76 33 76 5 68 18 76 4 16.1 4.8 14.87% Chile 45 65 11 56 14 55 25 51 14 8.8 0.6 5.67% China 15 22 38 0 6 11 27 11 24 15.0 7.3 9.93% Colombia 69 81 26 76 5 77 17 79 2 16.0 3.6 28.63% Costa Rica 53 70 7 65 24 71 15 67 6 8.9 6.3 0.57% Czech Republic 42 59 14 N.A. N.A. 59 38 68 9 8.0 4.8 2.13% Ecuador 52 60 33 51 24 60 71 50 13 14.5 3.7 16.70% Egypt, Arab Rep. 15 35 6 37 7 35 15 26 3 32.2 41.8 6.37% El Salvador 62 78 10 69 11 65 34 68 7 11.8 6.6 19.81% Germany 35 51 35 N.A. N.A. 47 39 55 13 4.7 5.2 14.59% Greece 43 81 7 N.A. N.A. 65 20 17 9 5.2 5.6 18.28% Guatemala 54 66 27 62 16 59 30 50 10 13.4 5.8 0.20% Honduras 47 69 7 62 11 48 33 45 16 13.0 6.4 0.09% Hungary 34 55 8 N.A. N.A. 44 35 42 19 15.1 5.7 14.11% Indonesia 37 N.A. N.A. 57 12 46 33 53 12 12.7 9.4 4.29% Ireland 50 73 21 N.A. N.A. 68 29 68 11 3.2 3.6 10.35% Korea, Rep. 46 72 18 N.A. N.A. 56 31 72 8 8.3 3.4 0.06% Latvia 47 33 6 N.A. N.A. 55 45 50 39 5.5 4.2 6.10% Lithuania 46 90 1 45 14 59 35 54 14 6.3 6.9 1.91% Mauritius 49 61 28 64 15 52 53 33 13 30.1 22.7 3.15% Mexico 34 76 6 85 2 68 6 53 6 16.5 8.2 23.92% Peru 31 93 2 95 0 80 36 38 9 15.2 1.8 8.73% Philippines 47 61 15 63 11 46 36 57 19 14.9 5.4 0.61% Poland 45 57 10 N.A. N.A. 57 23 67 4 26.0 16.6 10.78% Portugal 28 51 18 N.A. N.A. 35 43 43 13 4.1 4.1 21.58% Romania 37 48 7 N.A. N.A. 47 23 48 12 22.7 12.5 17.64% Slovenia 38 41 34 N.A. N.A. 41 78 36 28 7.3 5.4 5.03% South Africa 68 81 30 76 17 71 41 65 20 9.5 9.1 4.78% Spain 43 64 18 N.A. N.A. 65 21 60 8 3.5 3.9 12.56% Tanzania 34 76 6 46 11 50 8 52 12 18.7 8.9 0.12% Thailand 50 62 17 N.A. N.A. 55 51 64 19 19.3 11.5 0.47% Turkey 36 51 12 48 11 43 49 47 5 10.0 9.1 1.04% Ukraine 65 94 1 N.A. N.A. 80 17 71 13 11.2 5.0 0.01% Uruguay 67 86 12 90 3 65 34 71 10 13.3 4.9 9.56% LAC Average 54 75 15 73 10 68 27 59 9 13.2 4.6 12.37% Total 43 60 17 60 6 52 29 47 12 14.9 8.4 9.26% Note: LAC = average of percentages of Argentina, Bolivia, Brazil, Chile, Colombia, Costa Rica, Ecuador, El Salvador, Guatemala, Honduras, Mexico, Peru, and Uruguay. N.A.= not available. Source: World Bank Investment Climate surveys, Business Environt and Enterprise Performance surveys, and Nicita and Olarreaga (2006). 26 Table 2. Are R&D Expenditures Related to Productive Innovation? Direct versus Reverse Regressions (1) (2) (3) (4) (5) (6) (7) Depedent Variables --> New Product R&D/Sales R&D/Sales R&D/Sales New Product New Product Licensing Estimation Method --> dProbit Tobit dTobit Tobit dProbit dProbit dProbit R&D/Sales 0.019 (0.71) Licensing 0.127 0.124 (3.41)** (3.43)** Newproduct 0.354 0.202 0.330 0.038 (2.32)* (2.32)* (2.34)* (3.64)** Employees (log) (firm-level) 0.084 0.286 1.404 0.316 0.085 0.087 0.05 (4.30)** (2.62)** (2.62)** (2.76)** (3.66)** (5.35)** (3.44)** Employees^2 (log) (firm-level) -0.003 -0.014 -0.295 -0.016 -0.005 -0.005 -0.002 (1.25) (2.50)* (2.50)* (.) (1.80) (2.00)* (1.40) Foreign ownership (firm-level) (d) 0.002 0.005 0.003 0.022 -0.009 -0.016 0.124 (0.12) (0.17) (0.17) (0.94) (0.54) (1.45) (9.31)** Capacity utilization (firm-level) 0.001 -0.002 -0.171 -0.002 0.001 0.001 0 (1.74) (.) (.) (.) (1.56) (1.79) (0.85) Export status (dummy) (firm-level) (d) 0.081 0.145 0.032 0.089 0.082 0.078 0.012 (3.87)** (2.01)* (2.01)* (1.78) (3.74)** (4.52)** (1.70) Average years of education of employees (log) 0.025 0.004 0.005 0.002 0.03 0.024 0.014 (1.86) (0.20) (0.20) (.) (1.76) (1.56) (2.66)** Observations 11924 11924 11924 11924 18587 18587 10898 Countries 59 59 59 59 60 60 60 Censored Observations --- 3876 3876 3876 --- --- --- Industry and Survey-Year Dummies Yes Yes Yes Yes Yes Yes Yes Country Dummies Yes No No Yes No Yes Yes Robust z statistics in parentheses; standard errors are clustered around countries. * significant at 5%; ** significant at 1% Country, industry, and survey-year dummies are not reported. Columns 1, 3, and 4-7 report marginal-effects coefficients (elasticities) calculated at the sample mean. 27 Table 3. The Role of the Investment Climate: Marginal Effects from Probit Estimations Baseline regressions Quasi Country Fixed Effects 1 2 3 4 5 6 Non Weighted Weighted Non Weighted Weighted Non Weighted Weighted Employees (log) (firm-level) 0.085 0.056 0.084 0.053 0.085 0.053 [0.000]*** [0.002]*** [0.000]*** [0.000]*** [0.000]*** [0.000]*** Employees^2 (log) (firm-level) -0.004 0.000 -0.003 0.001 -0.003 0.001 [0.016]** [0.916] [0.041]** [0.605] [0.038]** [0.619] Foreign ownership (firm-level) -0.007 -0.022 -0.014 -0.021 -0.014 -0.021 [0.719] [0.016]** [0.393] [0.008]*** [0.397] [0.008]*** Capacity utilization (firm-level) 0.000 0.000 0.000 0.000 0.000 0.000 [0.442] [0.001]*** [0.175] [0.000]*** [0.170] [0.000]*** Export status (dummy) (firm-level) 0.096 0.021 0.057 0.004 0.057 0.004 [0.000]*** [0.458] [0.005]*** [0.827] [0.005]*** [0.827] Import tariff index (by sector) -0.047 -0.059 0.005 0.011 0.004 0.011 [0.094]* [0.043]** [0.629] [0.394] [0.674] [0.391] NTB coverage (by sector) -0.141 -0.093 -0.041 -0.045 -0.043 -0.045 [0.176] [0.406] [0.506] [0.548] [0.491] [0.549] GDP per capita in 2000 (log, PPP) (by country) 0.065 0.137 [0.477] [0.395] Manufacturing GDP growth, 1998-2003 (by country) -0.472 -1.019 -0.376 -0.681 -0.398 -0.781 [0.272] [0.027]** [0.269] [0.177] [0.256] [0.109] Stock of patents per worker, 1960-2000 (log) (by country) 0.051 0.077 0.026 0.018 0.022 0.007 [0.015]** [0.001]*** [0.145] [0.481] [0.256] [0.801] Regulation index (by country) 0.114 0.155 0.111 0.037 0.106 0.035 [0.071]* [0.000]*** [0.026]** [0.411] [0.035]** [0.446] Infrastructure index (by country) -0.080 0.031 -0.086 -0.156 -0.128 -0.276 [0.376] [0.886] [0.345] [0.203] [0.218] [0.179] Institutional index (by country) -0.023 -0.089 -0.016 -0.066 -0.029 -0.075 [0.510] [0.094]* [0.583] [0.172] [0.452] [0.126] Quasi Country Fixed Effects Employees (log) (firm-level, country average) 0.219 0.150 0.187 0.074 [0.501] [0.699] [0.557] [0.853] Employees^2 (log) (firm-level, country average) -0.048 -0.057 -0.046 -0.053 [0.232] [0.224] [0.239] [0.260] Foreign ownership (firm-level, country average) 0.098 0.410 0.283 0.939 [0.809] [0.402] [0.587] [0.253] Capital utilization (firm-level, country average) 0.005 0.011 0.003 0.006 [0.481] [0.170] [0.645] [0.484] Export status (dummy) (firm-level , country average) 0.518 0.934 0.478 0.938 [0.042]** [0.001]*** [0.078]* [0.001]*** Import tariff index (by sector, country average) -0.145 -0.292 -0.149 -0.277 [0.061]* [0.007]*** [0.054]* [0.019]** NTB coverage (by sector, country average) -0.262 0.615 -0.211 0.817 [0.497] [0.400] [0.584] [0.276] R-Squared 0.068 0.096 0.086 0.110 0.087 0.110 Firms 18390 18390 18390 18390 18390 18390 Sectors 8 8 8 8 8 8 Countries 36 36 36 36 36 36 P-Value: Joint Significance of firm-level variables 0.000 0.000 0.000 0.000 0.000 0.000 P-Value: Joint Signifficance of Trade Variables 0.092 0.054 0.681 0.662 0.705 0.66 P-Value: Joint Signifficance of Country Level Variables 0.051 0.000 0.052 0.008 0.082 0.001 * p<0.1, ** p<0.05, *** p<0.01. P-values appear inside brackets. Errors are clustered around countries. Note: Industry and survey-year dummies are not reported. Test of joint significance of country variables excludes GDP per capita and the "Quasi Country Fixed Effects" variables. 28