Surviving in a high-tech manufacturing industry: the role of innovative environment and proximity to metropolitan industrial portfolio

As a result, increased efforts devoted to a more precise understanding of LKS have identified several pecuniary mechanisms that boost business performance, but because they are difficult to measure empirically, these mechanisms often get attributed to the catch-all category of knowledge spillovers. The most prominent of such mechanisms are inter-firm labor mobility (Braunerhjelm, Ding and Thulin 2018; Maliranta, Mohnen and Rouvinen 2009), graduates’ migration (Faggian and McCann 2008), as well as collaboration with universities/research organizations (Arvanitis and Woerter, 2009) and clients and suppliers (Andersson 2006). When the effects of these mechanisms are accounted for in empirical analyses, what is left is more likely to be the pure knowledge spillovers, i.e. externalities that allow companies that do not invest in knowledge production to benefit from it without covering the costs of knowledge discovery.

For example, industrial diversity is found to boost economic output (van Stel and Nieuwenhuijsen) and employment (Frenken et al., 2007) in Dutch regions. Feldman and Audretsch (1999) argue that diversification promotes innovation in U.S. cities. Van der Panne (2004), on the other hand, finds that the Dutch regions with high levels of specialization in specific industries are more innovative in these industries.

It includes NAICS 3341 (Computer and Peripheral Equipment Manufacturing), NAICS 3342 (Communications Equipment Manufacturing), NAICS 3343 (Audio and Video Equipment Manufacturing), NAICS 3344 (Semiconductor and Other Electronic Component Manufacturing), NAICS 3345 (Navigational, Measuring, Electromedical, and Control Instruments Manufacturing), and NAICS 3346 (Manufacturing and Reproducing Magnetic and Optical Media).

Our focus is also justified by the special role that the NAICS334 industry plays in the U.S. economy. In addition to being one of the most innovative industries, it contributes disproportionately to manufacturing GDP and productivity (Houseman, Bartik and Sturgeon 2015), creates high-paid jobs and substantial local multipliers (DeVol, Wong, Bedroussian, Hynek and Rice 2009; Helper, Krueger and Wial 2012).

Because we limit our analysis to stand-alone companies only, in what follows the terms “firm” and “establishment” are used interchangeably.

The NETS database contains data on all establishments that were ever recorded as active by Dun and Bradstreet. The establishment-level information includes location, the number of employees, sales, the first and the last years in operation, and other indicators.

The companies that went through merger and acquisition (M&A) are likely to be more successful. Such companies would be an interesting subject of a study aiming at understanding firm-level business performance determinants. Only 12 establishments in our sample were identified as having experienced an M&A process, which does not allow for a formal statistical analysis.

For every firm in our sample, we manually inserted its name and location indicated in the NETS database into the assignee name and assignee state (and then city) fields, respectively, in the USPTO Patent Full-Text and Image Database (https://​www.​uspto.​gov/​patents-application-process/​search-patents#heading-1).

These data were used previously to study the relationship between innovative environment and business survival in Tsvetkova, Thill and Strumsky (2014b) and (without patenting establishments and using non-parametric techniques) in Tsvetkova, Thill and Strumsky (2014a). Our study (although uses a measure of metropolitan innovative environment as one explanatory variable and business survival as the outcome) is considerably different from the cited papers in a number of ways. First, using insights from economic geography and evolutionary economics, we develop a framework that describes the way companies in a high-tech industry rely on and benefit from radical and base knowledge available in their regions. Next, we explore interactions in the effects of these knowledge types on business performance metrics.

The EMSI data is a proprietary dataset that provides yearly information on employment, earnings, and the number of establishments in the U.S. counties at the 4-digit NAICS code level (www.​economicmodeling​.​com). These data were recently used to study a range of local economic processes (Betz, Farren, Lobao and Partridge 2015; Tsvetkova and Partridge 2017; Tsvetkova, Partridge and Betz 2019).

Appendix Table 4 presents a brief description of the variables and their data sources.

\( LQ=\left({E}_{im}/{E}_m\right)/\left(\frac{{\boldsymbol{E}}_{\boldsymbol{in}}}{{\boldsymbol{E}}_{\boldsymbol{n}}}\right) \) where E is employment, i denotes an industry (agriculture and government excluded), m indicates an MSA and n is the nation (all economy).

The calculation approach follows previous research of He et al. 2018; Hidalgo, Klinger, Barabási and Hausmann 2007 and Poncet and de Waldemar 2013.

In compiling their data, EMSI uses an imputation algorithm to fill in values for county-industry pairs that are suppressed in publicly available data. This introduces a measurement error and employment counts at the 4-digit NAICS level in the EMSI data seem to be unreliable if a county employs up to 5 workers in an industry (also, small number of workers in an industry in a county as indicated by EMSI – often a fraction less than one – are entered in the data set to ensure that total counts across industries and states equal to the totals in the governmental data sources). Since MSAs encompass larger counties or are a combination of counties, we raised the threshold of considering an industry present in an MSA to 10 employees.

Although the measure of industrial proximity of a metropolitan economy conceptually sounds a lot like Porter-type clustering, these two notions are empirically different. The measure of proximity is based on the co-location patterns of all industries across U.S. MSAs and is by construction a broader measure. The distance to the industrial portfolio calculated for this research takes into account proximities among all industries present in a metro area, regardless of the presence or absence of clusters. To illustrate this point, Appendix Table 3 shows pairwise correlations between our measure of industrial proximity and other potential indicators of NAICS334 clustering. While industrial proximity for NAICS334 firms in our sample is positively related to the corresponding LQs, the correlation is very close to zero. All other potential measures of clustering for this industry are negatively (and significantly at the 0.001 level) correlated with proximity to the MSA industrial portfolio. NAICS334 location quotient, on the other hand, is positively related to all clustering metrics. Table 3, thus, demonstrates that the industrial proximity measure is not about clustering but about average proximity among all industries in an MSA and how these industries are close or far from NAICS334 (clusters will naturally be a special case of this scenario but the negative correlations reported in the appendix suggest that clustering is not a dominant story in our data set).

\( IndProximit{y}_{mt}(Centered)= IndProximit{y}_{mt}-\overline{IndProximit{y}_{mt}} \) where m is an MSA, t indicates year and the last term is an average of the industrial proximity measure for all MSAs and all years.

Some scholars argue that composite measures of innovation and knowledge should be used in order to capture the elusive nature of this phenomenon (Hagedoorn and Cloodt 2003; Nelson 2009). In the case of high-technology sectors, to which NAICS334 belongs, however, the correlation among alternative measures such as R&D inputs, patent counts, patent citations and new product announcements is very high. This makes using just one approximation an appropriate strategy (Hagedoorn and Cloodt 2003).

We thank Prof. Deborah Strumsky for providing data for this variable.

Another potential issue with studying business outcomes of patenting firms is that generally a researcher does not know their business strategy, i.e. if the company uses its patented knowledge or prefers to license it out. Although this is an important distinction and our data do not allow us to explore this issue, we hope that limiting our sample to stand-alone manufacturing establishments with less than 200 patents should mitigate this concern for the following reasons. First, the firms in our sample must be producers; otherwise they would be classified in a different NAICS code. Second, since stand-alone entities are likely to have fewer resources, they should be less likely to focus on both research for licensing out and manufacturing. Finally, the overwhelming majority of patenting firms in our sample have no more than two patents, which means that licensing out is likely to undermine own competitive advantages in production. For these reasons we assume (but cannot prove) that companies in our data set are producers who are unlikely to license their new knowledge out.

We thank anonymous referees for drawing our attention to these concerns.

Our specification differs from Tsvetkova et al. (2014b) in that we center our main variables that are expected to interact. This slightly changes the interpretation of the coefficients on the variables that were demeaned.

They are practically indistinguishable for the purpose of estimation.

Discussion in this subsection is based on Cleves et al. (2010).

One needs to keep in mind that the NAICS334 industry is a knowledge-intensive one, so the high/low absorptive capacity discussion applies only within this industry and is not generalizable to other, particularly low-tech, settings.

In previous research that relies on the same data set but uses uncentered variables, the effect of size on business survival was found to be statistically insignificant (Tsvetkova et al. 2014b).