Import Competition, Multi-product Firm and Basic Innovation

Transcription

1 Import Competition, Multi-product Firm and Basic Innovation Runjuan Liu Carlos Rosell, January 30, 2009 Abstract The benefits of opening-up to international trade are without doubt; theoretical and empirical work consistently illustrate the benefits derived from specializing in industries of national comparative advantage, exploiting increasing returns to scale and widening the selection of goods. Yet, beyond these one-time improvements in efficiency what are the implications of trade on innovation, the true engine of long-run growth and welfare improvement? In this paper we explore this question and in so doing we illuminate a largely neglected but important area of economics, the intersection of international trade and the economics of innovation. We combine patent, firm level, and trade data to show that when confronted with higher import penetration, firms respond by producing innovations that are less likely to have spillovers and feed into long-run growth. That is, the data suggest a negative relationship between the basic nature of patented innovations and the degree to which patenting firms are exposed to competing imports. Underlying this result is the interplay between multi-product firms shedding peripheral product lines when exposed to greater import competition and the difficulties of inventors in capturing all the benefits of their basic innovations. In merging elements of the trade literature on multi-product firms and Nelson s diversification hypothesis (1959), this paper is novel and highlights significant considerations for both trade and innovation policy. JEL classification: F15, F19, O31, O32 Keywords: Import competition, Innovation basicness School of Business, University of Alberta. Address: 3-23 Business Building, University of Alberta, Edmonton, Alberta, T6G 2R6, Canada. runjuan.liu@ualberta.ca. Corresponding author: Department of Finance, Canada. carlos.rosell@fin.gc.ca We are grateful to Ajay Agrawal, Azim Essaji, Maryann Feldman, Nathan Nunn, Mark Schankerman and Paula Stephan for their helpful and insightful comments. Preliminary draft, all comments and suggestions are welcomed.

2 1 Introduction The benefits of international trade are well studied and accepted. Efficiency gains through specialization, the exploitation of increasing returns to scale and increased product variety are some usual channels through which trade has its positive impact. 1 Yet many of these benefits are largely static in nature. Once the effects of trade have fully worked themselves out they only provide level or one time improvements in welfare. A logical question to ask then is what other changes might trade cause that have an enduring impact on subsequent increases in productivity, welfare, and economic growth? In this paper we address this question by studying the relationship between trade and innovation. Specifically, we seek to determine whether the degree of import competition experienced by firms influences the basic nature of their innovations. As well, we seek to identify the channel through which this influence is exerted. In pursuing the these objectives we shed light on an important area of economics, the intersection of international trade and the economics of innovation. In light of the central role innovation has on long-run economic growth (Romer, 1986, 1990), this paper has obvious implications for long-run economic growth and welfare improvement. Basic innovation is widely considered as a major factor in economic growth. This type of innovation returns substantial private productivity increases for firms that undertake it (Mansfield, 1980; Griliches, 1986). More importantly, due to its general proximity to fundamental principles, basic innovation has broad spillovers as it serves as the foundation for many subsequent innovations (Bush, 1945). The positive externalities of basic innovation are thought to be so large that governments of the industrialized world invest billions of dollars annually to alleviate the under-provision of this public good and thus to promote economic growth. 2 Even though governments fund most basic research, as in the US, there has still been great interest in stimulating private investment in this type of innovation. As a consequence the factors behind private investment have been long understood (Nelson, 1959; Rosenberg, 1990). Prominently among factors inhibiting basic innovation is the uncertainty arising form undertaking a task which outcome is unpredictable. It is precisely this uncertainty that is at the heart of the relationship between trade and basicness that we study here and, as such, governments should be sensitive to 1 The case of immiserizing growth (Bhagwati, 1958) is an exception because it is a case in which opening-up to trade can leave a country worse-off. However, even in this case, not all trade partners will suffer; one trading partner will necessarily see its terms of trade improve thus allow it to benefit from trade. 2 For example, from 1991 to 2000 the US Federal Government alone has invested on average $18.9 billion dollars annually (1996 dollars) to fund basic research (?). 2

3 events or trade policy changes that affect this uncertainty. Our prior is that trade causes innovation to become less basic. 3 The influence that increased openness exerts on innovation is not direct however. Our hypothesis is instead that import competition has its impact by inducing multi-product firms in the importing country to return to their core; this entails a reorientation of firm productive activities by dropping peripheral product lines, products the firm is not best suited to make, and focusing its efforts on those products it is most competent in producing. It is this narrowing of firm activity that raises uncertainty and causes innovations to become less basic. The reason for this follow-on effect is highlighted by Nelson (1959). Intuitively, because, as Nelson (1959) notes, it is difficult to predict how or where the outcome of basic research will be applied, only firms that have their fingers in more pies (i.e. those involved in a wider set of product lines) will be best able to internalize the benefits of their innovations; simply, firms that do a lot of different things are in a better position to find some productive use for whatever emerges from their basic research. As a result then of greater import competition that narrows product scope and raises uncertainty, firm incentives to undertake basic research and innovation are ultimately reduced. Our hypotheses is derived from an original intertwining of two disparate literatures. On one hand is the trade literature on multi-product firms (e.g. Eckel and Neary (2006) and Nocke and Yeaple (2006)). This literature generally finds that multi-product firms respond to trade in a more sophisticated manner than one dimensional firms in standard theory. Of particular relevance is that multi-product firms react to increased import competition by dropping peripheral product lines (Bernard et al., 2006) and that the narrowing of firm scope is determined by how related each product line is to the firms main line of business, its core competency (Liu, 2006). On the other hand is the literature that followed from Nelson s diversity hypothesis. This literature tests empirically the importance of firm product diversity on its basic innovation. Specifically, Link and Long (1981) and Link (1982, 1985) confirm Nelson s original hypothesis. 4 The bringing together of these different literatures is reflected by the diverse data with which 3 By no means do we suggest firms become less innovative as measured by the level of research effort undertaken by firms or the number of innovations they produce. It is possible that the increased competition that opening-up can induce may actually raise the pace of innovation; this is possible as suggested by inverted-u relationships between competition and innovation Aghion et al. (2005). Our prior is instead that the nature of innovation changes as a result of international trade. 4 More recently, this literature has focused on the relationship between firm product diversity and innovation intensity. See Rodriguez-Duarte et al. (2007) for an example. 3

4 we test our hypotheses. First, we require firm level data to control for not only more conventional factors affecting innovation activity, such as research and development (R&D) expenditures, but also to assess the importance of firm product diversity. Second, we require industry level data to determine import penetration. Finally, data is needed that sheds light on the nature of firm innovative activity. This data is obtained from patents issued by the US Patent and Trademark Office (USPTO). Specifically, since actual firm data on basic R&D expenditures are difficult to obtain and subjective to determine, we instead use a measure of innovation basicness derived from patent data. This basicness measure is well documented by Trajtenberg et al. (1997) and accepted in the field of the economics of technology and innovation. An important caveat is necessary on the matter of using patent based measures to infer innovation basicness. We acknowledge that firms do not patent all their innovations and that firms may have various reasons for patenting. Nevertheless, some evidence suggests that not only does a large portion of R&D performers frequently patent but also that a large share of their innovations are patented. 5 Yet, as long as patented innovations generally reflect the innovation efforts conducted within each firm then, our patent based measures of basicness present a valuable means to address our research questions. 6 Using cross-sectional firm data we make four findings. First, we see a fairly robust negative relationship between the level of import penetration and the average basicness of patented firm innovations. Second, the data show that firms with more product lines patent on average more basic innovations. Third, the relationship between import penetration and basicness is substantially weakened once we cut the variation arising from the heterogeneity in firm product diversity. That is, once we segment the data according to the number of product lines firms operate, the negative trade-basicness relationship is weakened. This is consistent with the idea that import penetration has its effect by causing a reshaping of firm scope. Finally, we find further evidence that trade indirectly affects innovation through our hypothesized channel by finding that trade does not affect basicness in technology fields where firm product diversity may be inconsequential (i.e. in technology fields where patents reduce uncertainty by more effectively allowing firms to appropriate returns). 5 See Cohen et al. (2000) on this matter. 6 This assumption is not unlike that made by the literature studying the changing nature of university innovation. Henderson et al. (1998) and Mowery et al. (2004) utilize university patents to infer more broadly the orientation of university research even though patented university innovations represent only a small portion all academic innovation and research. 4

5 The remainder of the paper proceeds as follows. In Section 2 we describe the patent data we utilize to measure innovation basicness. In this same section we also introduce the basicness measure as well as test it to ensure the reader that the measure serves as a proxy of basicness. In the subsequent section, Section 3, we detail the firm and trade data. This section also introduces the metric for import penetration we use to stand in for the influence of international trade. Section 4 then describes how we link the firm and trade data to the patent data. Section 5 documents our main empirical result, the negative relationship between trade and innovation. We then follow with Section 6 where we present evidence suggesting that trade exerts its influence indirectly by affecting firm diversification. Finally, we conclude in Section 7 with a discussion of the implications for national innovation and trade policy as well as summarize the future work needed to round-out this paper. 2 Patent data and proxies for innovation basicness The patent data is obtained from two different sources. Our data is primarily from the NBER patent database described by Hall et al. (2002). 7 To augment the NBER database, we also use the NUS Patent Database. 8 In combination, both patent databases provide a substantial amount of information that reads on utility patents issued by the USPTO between 1963 and Of most relevance to our research questions, these data give us information on each patent s: Application year; 10 Primary technology field; 11 and, Assignee. 12 Our measure of innovation basicness is based on a proxy first proposed and tested by Trajtenberg et al. (1997). It was subsequently accepted as a standard proxy of basicness in the literature of the 7 Specifically, we use the updated NBER patent dataset ( ) available from Bronwyn Hall s website: bhhall/bhdata.html. 8 The NUS Patent database is available at: 9 A utility patent is a patent protecting a process, machine, composition of matter, or an improvement of any one of these things. 10 The year in which the innovation was submitted to the USPTO for the purposes of obtaining patent protection. 11 This is the main technology area to which the innovation belongs, it is analogous to the main industry category to which a firm might belong. 12 That is, for a large portion of patents the owner of the patent at the date of issue is identified. 5

6 economics of innovation. It has been used for example by Henderson et al. (1998) and Mowery et al. (2004) to test the extent to which patented university innovations have remained basic in nature since the passage of the Bay-Dole Act in The foundation of this proxy lies on the idea that innovations that are more basic will lend themselves more broadly to follow-up work in different areas. That is, due to an innovations closer proximity to fundamental principles and science more generally, an innovation that is basic will be able to be applied in many different ways in different fields rather than for a specific use in one particular area of use. In this respect, the metric lends itself well to measuring the basicness of patented innovations because we can utilize patent citations to identify the different technology fields to which an individual innovation is applied or contributes in follow-up inventions. Furthermore, given existing concordances linking technology fields to SIC industry fields, similar metrics of dispersion have been used by Hall and Trajtenberg (2004) and Feldman and Yoon (2008) to identify general purpose technologies, technologies which applications lend themselves to different industries rather than areas of technologies. In what follows in this section we illustrate the measure of basicness and test this proxy to assure the reader that the proxy conform to the expectations in the literature (i.e. that the proxy actually reflects basicness). 2.1 The proxies When a patent is granted it is classified into one or more technology fields. This classification system is intended to help patent examiners research the prior art when they evaluate the novelty and delineate the property rights of an innovation. In many respects, a patent classification system is very much like the Journal of Economic Literature (JEL) classification system. 13 Also, because patents need to cite prior art, we have the capability of identifying how patents contribute to subsequent patented innovations. Commonly, as exemplified by Jaffe et al. (1993) and Hall and Ziedonis (2001), cited patents are considered to be the foundation on which a citing patent builds. Consequently, the patent system creates a paper trail through which we can trace the different technology fields to which an innovation contributes. The metric Basicness p we use is essentially a herfindahl index measuring the dispersion of citations a patented innovations receives from subsequent inventions in different areas. Explicitly, 13 Specifically, articles can be assigned to multiple finely defined fields and they can be also aggregated up to broader fields. 6

7 the measure is define as, Basicness p = [ 1 T ec ( Xp,T ec X p ) 2 ][ ] Xp, (1) X p 1 where X p,t ec is the number of patents classified in technology field T ec that cite patent p and X p is the total number of patents that cite p. 14 As this measure is a herfindahl it only takes values between zero and one. Furthermore, innovations receiving basicness scores closer to one connote innovations that lend themselves more broadly to different areas and that would thus be considered to be more basic in nature. One way this measure can differ from study to study is in the different technology classification systems on which the metric is based. For example, one could base the measure on the USPTO US patent classification (USPC) system or the International Patent Classification (IPC). 15 Furthermore, one could base the measure on how fine each class is defined. For example, the USPTO assigns patents to six-digit technology class but also to more aggregated three-digit classes. Further still, Hall et al. (2001) creates a more aggregated technology fields by grouping three-digit USPC classes. Specifically, the 3-digit USPC classes can be considered to be aggregated to the 2-digit and 1-digit levels. Finally, we are also able to measure the dispersion of an innovation s use across different industries defined by the 1972 US Standard Industrial Classification (SIC). This is possible given the concordance linking USPC 6-digit technology fields to their associated 2-digit SIC industries that are provided and updated annually by the USPTO A test of proxies For this study, we test different forms of the basicness metric to determine which one(s) may be the most dependable proxies for our concept of basicness. Specifically, we determine whether any of the basicness measures based on the USPC 3, 2 and 1-digit technology fields and on the 1972 US SIC 2-digit industries act as a good proxy. As in the literature, our test rests on the prior that 14 Here: X p = T ec X p,t ec. 15 USPTO examiners assign all patents to classes within these two systems. These USPTO concordance can obtained free of charge from the following USPTO ftp site: ftp : //ftp.uspto.gov/pub/taf/sic conc/. We are personally grateful to Maryann Feldman for bringing this concordance to our attention. 7

8 universities produce on average more basic innovations than firms. Therefore empirical evidence that on average university related innovations carry with them greater basicness values will ratify conclusions reached by Trajtenberg et al. (1997) and reaffirm the use of these measures as a proxy for basicness. The procedure to test the different proxies is as follows. We first identify US university patents that were applied for between 1990 and 1996 and that were issued between 1990 and Next, as a benchmark, we match each of these university patents with a similar firm patent. In this respect, we follow Mowery et al. (2004) by ensuring that matched firm patents are assigned to the same 3-digit USPC technology field as their university partner and that application dates of paired patents are as close as possible. 17 We also ensure that the non-academic patents are assigned to US non-government organizations. This may be necessary as we would then include as benchmark patents innovations that are more basic in nature than the average. Examples of such innovations may be innovations patented by US government research laboratories or patents assigned to non-us firms. 18 Finally, each patent s basicness measure is calculated using only citations received no more than six years after the issue year of our sample patents. 19 We use the following regression specification to identify whether the basicness measure successfully ascribes higher values to university rather than firm patents, Basicness p = α 0 + t α t Issue year p,t + x β x Application year p,x University p + ε p. (2) In Equation 2, Issue year p,t and Application year p,x are a dummy variables that equal one when patent p is issued in year t and applied for in year x, respectively. Finally, University p is a dummy variable indicating whether p is assigned to a US university, it equals one when assigned to a university and zero otherwise. In this specification, if the coefficient {β x } is consistently estimated to be positive and statistically significant we will conclude as Trajtenberg et al. (1997) that the 17 In practical terms, matched patents are not allowed to have application dates that are more than three year apart. In cases where the application date of a firm patent is not within three year before or after a university patent s application, the university patent is excluded from the sample. 18 With respect to patents issued to non-us assignees, Kortum and Putnam (1997) find that these patents protect innovations that are of greater importance, this is why these foreign firms would incur the costs of patenting in the US. If importance translates to more basic innovations then, excluding these patents may also be necessary in order to have an appropriate benchmark. 19 This is exactly as in Mowery et al. (2004) and is done for technical reasons. Also, as with Mowery et al. (2004) we include self-citations when calculating the basicness measure. 8

9 metric is on average a good proxy for an innovation s level of basicness. Table 1 shows the tobit coefficient estimates for the interaction of the application year and university dummy of Equation 2 based on USPC technology fields at different levels of aggregation as well as when the proxy measures how broadly the innovation can be applied to different 2- digit SIC industries. 20 Generally, the table shows the basicness measure based on the 3-digit and 2-digit technology fields successfully indicate that university patents are more basic than similar firm patents. In all but in 1996 university patents are seen to be more basic than those from firms. Similarly, these same university patents are seen to be more broadly applicable to different industries defined by 2-digit technology fields. In this case, the coefficient is statistically significant in five of seven years and is nearly statistically significant in another. The only basicness measure unable to identify any difference between university and firm patents is when the metric measures citation dispersion across 1-digit technology fields. Overall, these results are consistent with Trajtenberg et al. (1997) and Mowery et al. (2004) and suggest that we can use these basicness proxies to investigate our main questions, at least if we base the measure on dispersion across 3-digit and 2-digit technology fields and 2-digit SIC industries. Consequently, our analysis to follow is be based on these three different flavors of the basicness measure. 3 Segment data and measures of import competition Information about firms production lines is obtained from Compustat (North America) segment data. Since the 1970s U.S. public firms are required to report information about their operations in different industries. This appears in the combined financial statements under FAS14 issued by the Financial Accounting Standards Board. Whenever a component of an enterprize engages in providing a product or service primarily to unaffiliated customers for a profit and whenever its sales, and operating profits or identifiable assets are over 10 percent of the combined figure of all industry segments, the firm s related information (including sales, operating profit, identifiable assets, etc.) must be reported in its financial statement. Compustat assigned a 4-digit SIC code to each reported industry segment of each firm. So the product (or equivalently the line of business 20 We use tobit regressions to address the fact that the support for our dependant variable is the unit interval. By using these regressions we also conform to Mowery et al. (2004) even though the distribution of the dependant variable is not truly censored from above by one and from below by zero. 9

10 or the segment of a firm) used in this paper is defined at the 4-digit SIC level. The accounting rule to report segment information has changed since Firms are required to report operation segments instead of industry segments. Therefore, the last year in our data is constrained to is a convenient year to start because it is the earliest year we can get the data from Wharton Research Data Services (WRDS). For multi-product firms, we define two sets of import competition measures. First, we define the average import penetration ratio for multi-product firm f at year t as, Average import pen f,t = i s f,i,t M i,t Q i,t + M i,t X i,t. In the definition of Average import pen f,t, s f,i,t is the share of firm f s revenues derived from industry product i at year t, M i,t is the value of imports in industry i at year t, X i,t is the value of exports of industry i, and Q i,t is the value added of shipments produced in industry i at year t. Average Import Pen f,t thus measures the weighted average import penetration that a multi-product firm f experiences across all its product lines at year t. Second, we define the core import penetration ratio for multi-product firm f at year t as, Core import pen f,t = M i,t Q i,t + M i,t X i,t. Core Import Pen f,t captures import penetration directed at the core product for firm f at year t. In defining Core import pen f,t, i indicates firm f s core industry product at year t (i.e., the product which generates the largest sales share for the firm). Furthermore, as in the above, M i,t is the value of imports in industry i at year t, X i,t is the value of exports of industry i, and Q i,t is the value of shipments produced in industry i at year t. On average, the core product generates over 50 percent of a firm s total sales, assets and operating profits (Bernard et al., 2006; Liu, 2006). Therefore, import competition directed at core products may be important enough to investigate as this alone may induce firm refocusing (Liu, 2006) that ultimately affects innovative strategy. Following the above definition of import competition for multi-product firms, it is straightforward to define an import competition measure for single-product firms. Let f denotes firm, i denotes the SIC industry product firm f produces, and t denotes year. We define the import 10

11 competition measure for firm f at year t as, Import Pen f,t = M i,t Q i,t + M i,t X i,t. In this definition, M i,t is the value of imports in industry i at year t, X i,t is the value of exports of industry i at year t, and Q i,t is the value of shipments produced in industry i at year t. The import and export data we use to calculate our import penetration measures is for the US and it comes from different sources. Annual import-export data for years prior to 1989 come from NBER Trade Database (Feenstra, 1996). We converted these data from 1982 SIC to 1987 SIC using the Bartelsman and Gray (1996) concordance (with weights) available from the NBER-CES Manufacturing Industry Database. Annual import-export data after 1989 is obtained from the US Department of Commerce official data as described in Feenstra et al. (2002). Annual industry shipment data comes from the NBER-CES Manufacturing Industry Database Bartelsman and Gray (1996). Other firm characteristics may affect firm s innovation strategy as well, especially R&D investment. To control for these firm characteristics, we use the firm-level information from Compustat industrial annual data downloaded from WRDS. Specifically, we use the share of R&D expenditure over sales, logarithm of sales and logarithm of operating profit to control for R&D intensity, firm size and firm profitability in the regressions respectively. 4 Linking segment data to patent data A crucial aspect of this paper is the ability to link patents to their assignees and ultimately to the import penetration experienced by these same assignees. In this respect the NBER patent database is invaluable. It links patents to assignees and, more importantly, it links many firm assignees to their Compustat data. Consequently, for a large number of patents we are able to associate them to their assigned firm s R&D expenditure, size and profits. Especially useful for our questions is that, because we can also identify the patent s firm assignee, we can identify the different industries in which the assignee is involved and, as a consequence, the related level of imports the assignee faces in each of its product lines. The Patent Name-Matching Project supervised by Bronwyn Hall links the universe of Com- 11

12 pustat company names that have existed up to 2005 to patent assignee identifiers. 21 The means by which this link is made is by standardizing the name of patent assignees and the names of Compustat firms. This process successfully bridges 7,000 Compustat firm names to 8,337 assignee identifiers. This mapping from a single firm name to potentially multiple assignee identifiers reflects, in part, that many firms patent under different names. For example, one can find that companies such as IBM patent under the name IBM as well as under International Business Machines. In total, 946 of compustat firms are found to have multiple assignee names under which they patent. In contrast, the number of Compustat firm names that have existed until 1997 for which we have at least partial information on segment, industry and trade data is more than twice the number of firms obtained from the patent matching. Specifically, the number of different Compustat firms is 16,797 versus the 7,000 obtained from NBER patent database. In all, the total number of firms for which we have both patent and complete Compustat data are shown in Table 2. This table shows, for example, that for 1996 our sample is composed of 689 different firms for which we have all the data described above. 22 The overall (i.e. Total) number of unique firms in our sample, 1,650, is considerably smaller than the universe of compustat firms identified in both the patent and Compustat data sets. There are several reasons for this. First, not all Compustat firms will patent and consequently the number of firms will be much smaller than 16,797. Second, our sample is shaped by whether patents receive sufficient citations to calculate our basicness measures. 23 Consequently, even if a firm patents it is not included in our sample if none of its patents receive at least two citations. Finally, our final sample is also shaped by whether we have the relevant industry and trade data for each firm. If we do not have this information then it makes analysis impossible and so we are unable to include these firms in our final sample. Table 2 also presents the number of patents (i.e. observations) included in our sample. For example, the total number of patents applied for from 1984 to 1996 that are included in our sample is 166,434. The general upward trend in the number of patents reflects a broader pattern noticed for example by Hall and Ziedonis (2001). 21 See Halls web site for further details on this name standardization project at bhhall/pat/namematch.html. 22 In the regressions that follow the number of firms in the sample and the number of observations may be higher than is shown in Table 2. This is because not all the data such as firm size, R&D intensity and operating profits are needed to estimate the some of the regressions. 23 Recall the definition of basicness given by Equation 1. Basicness is undefined for any patent receiving less than two citations. 12

13 Finally, since we use the NBER s data to link patents to Compustat firms the usual caveats of the literature must be mentioned. First, because of the constraints on the data, we can only speak to how imports affect the innovation activity of typically larger, publicly traded companies. And second, the NBER database does not perfectly link patents to Compustat owners. This is because often, patents are granted without an assignee and, at times, assignee names are misspelled or the names of subsidiaries are given. However, to the extent that enough patents are correctly linked or that errors do not affect one type of firm consistently more than an other, the use of patents for our research should not affect our results. 5 Empirical results In this section we specify the regression equation we use to establish a statistically significant relationship between innovation basicness and import penetration. As well we show the regression results. However, we begin by examining the simple correlation coefficients between our various measures of basicness and import penetration. 5.1 Summary statistics and correlation To begin identifying any relationship between import penetration and innovation basicness we present the Pearson correlations between these variables. These correlations are based on our entire sample. Table 3 shows that, consistent with our prior, firms exposed to greater import penetration produce innovations that are less basic on average. For example, the correlation between the average level of import penetration and innovation basicness based on 3-digit technology fields is equal to We also find that this negative relationship is statistically significant in general; all correlation coefficients are statistically significant at one percent. 24 Table 3 also provides the correlation coefficients between import penetration for each firm and the number of product segments each firm operates (Product segments). Both correlations are negative and statistically significant at one percent. For example, the correlation between the average import penetration a firm confronts and the number product lines the firm operates equals Consistent with Liu (2006), these results suggest that firms facing higher import 24 The correlation between the various forms of import penetration and basicness based on 1-digit technology fields is negative and statistically significant at one percent as well. 13

14 competition operate fewer product lines potentially as the result of refocusing induced by increased import penetration. Finally, it is import to note as well that Table 3 presents the correlation between basicness and the number of product lines a firm operates. In line with Nelson (1959), the positive and statistically significant correlation suggests that firms that have their fingers in more pies tend to produce innovation that are more basic in nature. 25 The importance of the correlations presented in Table 3 is that we are not able to reject our hypothesis that international trade induces firms to produce less basic innovations by causing them first to shed peripheral product lines. More thorough testing of the link between international trade and basicness is required. This is what is done in the following subsection. 5.2 Regression results To determine whether any relationship exists between import penetration and innovation basicness we pool the data for each firm in our sample and for each year they are available. Here we exploit cross-sectional variation between firms. Essentially we regress the basicness of a patented innovation p (Basicness p,f,t ) that was that was applied for by firm f in year t on the import penetration (Import Pen f,t 1 ) experienced by the same firm in the previous year, t 1. This implies that while our unit of observation is a the patent level, many of the dependant variables will vary by firm and year. Specifically, the regression equation we estimate is given by, Basicness p,f,t = γ + αimport pen f,t + βx f,t + T ech φ T ech I ( T ech ) + Y ear φ Y ear I ( Y ear ) + ε p,f,t. (3) In Equation 3 X f,t controls for firm specific factors that might affect innovative activity. In this respect we follow Link and Long (1981) and Link (1982, 1985) in controlling for the firms size in terms of sales (Size f,t ), R&D intensity defined as the fraction of contemporaneous R&D expenditures in sales (R&D Intensity f,t ) and the firm s operating profits (Profits f,t ). As suggested by Schumpeter (1950), these variables find their relevance by summarizing a firm s general openness to take on prolonged projects that have no certain outcome such as basic innovation. The summary statistics of our dependant and some independent variables are shown in Table This correlation is statistically significant at one percent. 14

15 Finally, in Equation 3 utilize two sets of different fixed effects. First, to soak-up any variation in basicness occurring because of something general to a particular technology field, we include technology field fixed effects; these fixed effects are represented by the set of dummy variables {I(T ech)}. These fixed effects are important because they control for such things as mature technologies where basic innovations may be less common. In addition, these fixed effects also control for institutional differences in the patent and legal system that vary by technology field and that would make comparison basicness across fields difficult. And, second, we control for application year effects with the set of dummy variables {I(Y ear)}. These dummy variables are needed to control for changes in the patent system or technology fields that occur over time that would otherwise effect measured basicness. In addition to testing separately the significance of import penetration on average and at the core product line, we tested the sensitivity of this relationship by varying the lag of these variable. We did this recognizing that the output of R&D activity occurs some time after the decision to undertake the R&D is made. Consequently, innovations that are obtained today may have emerged from a decision made in prior years and under the information then available. For this draft we provide the results based on lagging our import penetration measures by one year. Similar results hold when we use the import penetration as in the patent s application year and when we use lag penetration by 5 years. We estimate Equation 3 using fractional logit regressions. We use this method because OLS is not perfectly suited when the dependant variable takes on values between zero and one. Furthermore, it is also inadequate to regress the log-odds ratio of our dependant variable on the explanatory variables. This is because our dependant variable does not fall strictly between zero and one. It is common for the basicness measure to take on values of zero and one. For these observations the log-odds ratio would be undefined. Consequently, fractional logit regression is the most viable estimation procedure. 26,27 Table 5 presents coefficient estimates from our fractional logit regressions. The estimates confirm the statistical significance found in the correlations between basicness and import penetration 26 We could have used a tobit regression as we do above when testing our dependant variable for its ability to proxy for basicness during the 1990s. We utilized tobit regressions there to conform with the literature even though the estimation method is not quite correct. In our case, our dependant variable is not censored from above or below. Our measure is instead naturally bound between zero and one. 27 For further detail regarding fractional logit regressions see Papke and Wooldridge (1996). 15

16 in Table 3. Broadly, we find that after clustering standard errors by firms a negative and statistically significant relationship exists between basicness and import penetration. When we measure basicness according to dispersion across 2-digit technology fields or 2-digit SIC industries, this relationship is negative and statistically at no less than five percent. This is the case regardless of whether or not we control for fixed effects and or other explanatory variables. 28 Our results are only slightly weaker when basicness is based on 3-digit technology fields. In this case, depending on the specification, the negative coefficient on import penetration is statistically significant at no less than 10 percent. The regression results are also robust to the definition of our import penetration measure. When we focus on the impact import penetration on the firm s core line of business we find that import penetration continues to have a negative effect on basicness. Table 6 shows the coefficients estimates of our fractional logit regressions. In this table we again find evidence as in Table 5 that a statistically significant and negative relationship exists. In fact, not only are the qualitative results the same but the estimates are nearly identical. 6 Evidence of the indirect channel, firm refocusing In this section we go about determining whether any evidence exists in the data supporting our hypothesis that import penetration influences basicness indirectly through its affects on firm diversity. To this end, we investigate the data to show no obvious evidence exists that contradicts this hypothesis. We then investigate whether innovation basicness is affected by import penetration only when heterogeneity in firm diversity exists; this is intended to help link import penetration and diversity together as they possibly jointly affect basicness. Finally, we test our indirect channel hypothesis by studying whether import penetration loses its punch in cases where firm diversity is inconsequential to basic innovation. 6.1 Contrary evidence? A possible explanation why import competition affects innovation basicness is because the degree of this competition indirectly reflects the number of product lines operated by each firm. Con- 28 This is also the case when we measure basicness based on the dispersion of citations across different 1-digit technology fields. 16

17 vincing evidence already exists in the international trade literature in this respect. In general, Liu (2006) shows that, in response to increased import competition, firms are more likely to close down peripheral product lines and so narrow the scope of their operations. This relationship has already been observed in our data from the correlation between import penetration and the number of product segments operated by each firm (the correlations found on Table 3). This relationship is strong enough in the data that this correlation between imports and the number of segments is positive and statistically significant at no less than one percent when we parcel out the data by the application year of the patented innovations. 29 Consequently, in this respect we cannot find any obvious evidence that contradicts our indirect hypothesis. For the indirect channel we hypothesize to be correct we must find evidence of Nelson s diversification hypothesis. That is, the data must show that firms with broader sets of product lines produce innovations that are on average more basic. Our data indeed reflects this relationship. As noted above, the correlations found in Table 3 between the number of product segments and basicness is positive and statistically significant. Furthermore, when comparing the difference in mean basicness between single and multi-product firms we find that multi-product firms produce more basic innovations. For example, Table 7 shows that the difference in means between multi and single-product firm basicness that is based on SIC industries equals and is statistically significant at one percent. 30 This outcome is more as, for each of our basicness measures, the mean basicness of multi-product firms is statistically significantly greater than that of single product firms. Consequently, we again cannot find obvious evidence that rejects our hypothesis. 6.2 The joint effect of import penetration and diversity To test further whether the impact of import competition on basicness works through firm product diversity we segment the data according to number of product lines each firm operates. We then estimate Equation 3 independently using the data from each group. 31 The idea behind this is to see whether the variation in import penetration within each type of multi-product firm can explain basicness. If it is the case that import penetration directly impacts basicness then parceling out 29 These correlation are not shown in this version of the paper but are available from the authors on request. 30 This difference implies that multi-product firm basicness is approximately 28% greater than that of single product firms on average. 31 We do not carry out this analysis for groups of firms that operate more than six products because the numbers of observations in each of these groups is fairly small. 17

18 the data in this way should not affect the strength of the relationship. Column (I) in Table 8 shows only the flogit coefficient estimates on the average import penetration variable in a regression where we include year effects and 3-digit technology field effects. The coefficients in Column (III) differ from those in (I) in that only the former are based on regressions that also include the explanatory variables Size, R&D intensity, and Profits. Table 8 clearly shows that import penetration has a substantially weaker effect when we eliminate heterogeneity in firm diversity. In only a few cases do find evidence that import penetration reduces basicness. Essentially, only for firms operating three product lines does import penetration have any influence. Otherwise, the estimated coefficients are mostly not statistically different from zero. It should be noted that this lack of significance is not caused by a lack of degrees of freedom. As shown in Column (II) and (IV), the smallest number of observations for any group is 2,622 while only approximately 420 coefficients are estimated. Consequently, we can argue that import penetration and firm diversity jointly work to affect innovation basicness, something consistent with our indirect channel hypothesis. 6.3 Patent effectiveness In this subsection we further test whether the impact of greater import penetration on innovation basicness has its influence by affecting the product diversity of firms in the importing country. We do this by testing whether the effectiveness of patents in allowing innovation owners to appropriate the returns from their innovations mitigates the impact of trade on basicness. In this respect we exploit industry variation in the effectiveness of patent protection. The possibility that import competition reduces basicness indirectly by reducing product scope relies implicitly on the assumption that firms have a limited opportunity to sell or license innovations that they themselves cannot use. This is one of the implicit assumptions that underlie Nelson s diversity hypothesis. However, if firms were assured that intellectual property right protection effectively allowed them to transfer, license and, in general, defend their IP, firms would be able to appropriate much of their returns regardless of whether or not they utilized their innovations themselves. In such cases, the legal institution safeguarding IP acts much like product diversity in lessening the risks inherent in basic innovative activity. This would then imply that diversified firms would not necessarily stand out in terms of the basicness nature of their innovations This reasoning is consistent with casual observations that some of the most basic research conducted in the 18

19 Consequently, in these cases, the effect of imports on innovation would be mitigated because the indirect channel through which imports have an impact would be severed. As is well known the effectiveness of patents to appropriate returns varies substantially from industry to industry. Levin et al. (1988), Schankerman (1998), and Cohen et al. (2000) consistently find, for example, that patents are particularly effective in the pharmaceutical and chemicals industries. This is no doubt related to the ease with which the validity of these sort of innovations can be assessed and infringement can be successfully defended against. It is this heterogeneity in industry patent effectiveness that we use for our test. To implement our test we utilize Cohen et al. (2000) to identify manufacturing industries where patent protection is found to be relatively more effective in appropriating returns. This allows us to rank 18 2-digit ISIC industries according to how effective patents are regarded in protecting product innovations. 33 Table 9 shows the different ISIC industries we analyze. The table also shows the ranking of these industries according to how effective patents are in appropriating returns. This ranking is broadly consistent with common held views. In particular, the effectiveness of patents on product innovations in the chemicals and chemical products industry ranks high. Next, to implement our test we need to be able to identify what patents are most closely associated with the 18 different ISIC industries. This is done by creating a concordance between the ISIC industries and the 6-digit USPC fields. This concordance is however not direct. We use three different concordances for our work. Specifically, we use the USPTO s USPC-1972 SIC concordance, the Bartelsman and Gray (1996) 1972 SIC SIC concordance, and the Statistics Canada 1987 SIC - ISIC concordance. 34 In all, we are able to concord 26 ISIC 2-digit industries to 75,056 6-digit USPC technology fields. This concordance is sufficiently wide to allow us to concord the 18 ISIC industries for which we know the relative effectiveness of patents to the their associated technology fields. Finally, when determining the effectiveness of patents in a particular technology field we must be careful of when the field is associated to more than one of the 18 ISIC industries. In these situations we average the effectiveness score of the associated ISIC industry and assign it bio-tech industry, an industry where patents are most effective, is carried out by small and highly focused firms. 33 Cohen et al. (2000) is based on the opinions of surveyed industrial R&D practitioners. These practitioners identify generally the proportion of their innovations for which patents effectively allow them to appropriate the returns. It is the industry averages of these proportions that we use to rank the 18 different ISIC industries. 34 The location of where the USPC-1972 SIC concordance can be obtained is mentioned in a previous footnote. The latter two concordances are readily available from Jon Haveman s concordance web page: 19