Does mobility really raise inventiveness? Evidence from panel data
Lina Ahlin[1] and Olof Ejermo[2]
Abstract
We investigate how individuals’ patent productivity and patent quality are affected by moving between firms using a large panel of nearly 80% of the population of Swedish inventors linked to employer-employee data observed between 1987 and 2007. While we initially find a productivity effect from a move on the labor market, this effect arises primarily due to the first move on the labor market and seems to be related to moves to non-multinational firms with Swedish ownership. We further use labor market turnover variables as novel instrumental variables for mobility to infer causality, but this also turns the mobility effect insignificant. Our interpretation is that selection is a strong driver of general productivity effects. However, when we switch our dependent variables to citation-weighted patents to gauge quality, we find a strong positive effect on quality with a causal interpretation supported by instrumental variable regressions.
Key words: employer-employee data, inventors, mobility, patent productivity, selection, Sweden
1 Introduction
In recent years, labor mobility as a channel of knowledge transfer has received a high level of attention. If individuals’ mobility raises productivity without lowering it at the origin firm, the societal returns to knowledge should increase. Thus, if empirical evidence can be generalized to show such results, policy could simply be directed towards facilitating mobility in order to raise innovation. On the other hand, if for instance only seasoned inventors or newly educated individuals experience such effects, policy should perhaps be more targeted towards these groups. In this paper we try to uncover some of the factors that contribute to higher inventiveness through mobility.
Lacking other data on the innovativeness of individuals, many papers in the literature focus on inventors that are listed on patent records. Although inventors are a subset of the population of individuals contributing to innovative output, their activities are not peripheral in modern economies: in a small country like Sweden (9.5 million inhabitants), about 1,500 patents are granted yearly at the European Patent Office (EPO) where total patent filing costs can be estimated at 75 million euros and this does not even include costs at other patent bureaus such as the USPTO.
Despite the relative visibility of inventors in data compared to other definitions of innovative output, no large-scale analysis using population-wide data on inventors has been undertaken to estimate the effects of mobility on inventiveness. One of the reasons has been lacking complementary data that can be used to examine and control for individual characteristics and the possibility to observe how inventors’ patenting behavior changes over time, and as a consequence the inability to infer a solid general productivity-increasing effect from mobility.
We use a close to full population sample, consisting of 80 percent of Swedish inventors listed on applications to the European Patent Office (EPO) that are observed over the period 1990-2007. Inventors were linked with directories housed at Statistics Sweden on the whole population using their social security number. These data offer a number of advantages over datasets used previously concerning the definition of mobility and the time-patterns of inventive behavior. A first aim of our paper is therefore to uncover how general the productivity effect is following mobility using panel fixed effects methods. We also try to weed out a causal effect from mobility using novel instrumental variables based on national and regionally-varying turnover rates by industry that we show is correlated with inventor mobility, but not directly correlated with an inventor’s patent productivity.
Knowing if mobility is important for inventiveness is important, but disentangling how any effects come about is equally important. The current literature has taken several routes to examine mechanisms leading to productivity effects. Some of the recent studies focus on the learning effects that arise among the staff at the recipient firm (Singh and Agrawal 2011; Palomeras and Melero 2010; Song, Almeida, and Wu 2003), disagreement as a source of spillovers and labor mobility (Mokyr 2006; Klepper 2007, who studies spin-offs), the role of institutional aspects, such as noncompete agreements (Marx, Strumsky, and Fleming 2009), and the effect of a move on the productivity of the person who moves (Hoisl 2007, 2009; Lenzi 2009). Our study contributes to the estimation of how the move affects the individuals moving after their own productivity (inventiveness). Also, while existing studies has had an explicit or implicit focus on ‘star inventors’, our data allow us to draw conclusions for the population of inventors at large and in contrast to earlier studies we can employ panel regressions to eliminate time-invariant fixed effects.
In our investigation, we examine how productivity effects vary over the life-cycle of the inventor. Clearly, inventors may self-select into working for an employer who provides opportunities for learning. Rosen (1972) developed a model, where the main implication was that individuals seek to accumulate such experience especially early on in their career. While learning effects have been studied for wage formation (Møen 2005) we add evidence on career effects by separately analyzing whether productivity effects differ by the order of the move on the labor market.
Finally, we investigate mobility effects using an alternative productivity measure, citation-weighted patent counts, a method commonly used to quality-adjust patents. This may shed light on which individuals become ‘better’ inventors, in the sense of substituting quality for quantity.
Our estimates show that in non-instrumented regressions, there is a positive average productivity effect following a move that pertains to the first move made on the labor market, in line with Rosen’s prediction. However, this average effect quickly turns non-significant for two reasons. First, once we include control variables in addition to year dummy fixed effects we find that the move effect is absorbed by and explained by moves to Swedish-owned but not multinational firms.
Second, move-effects are also non-significant once we instrument mobility with turnover rates that vary by industry for the whole of Sweden and by region. On the other hand, when we instead employ citation-weighed patents as our dependent variable combined with instrumental variable methods, we find that the level of quality-adjusted patents substantially increases for inventors with earlier patent experience following a move. Moreover, this effect does not seem to differ for different move orders. Therefore, only for specific groups are there any significant mobility effects. Policy would therefore be misguided if it tried to stimulate mobility generally. Instead, there may be a role for policy in trying to stimulate mobility among more seasoned inventors.
2 Literature review
The study of spillovers has been of great interest in economics both empirically (Griliches 1957; Mansfield 1961) and theoretically (Nelson 1959; Romer 1986). However, pinning down the importance of spillovers numerically has proved to be difficult, due to its complexity and the variety of mechanisms involved (Feldman 1999).
In order to estimate the importance of spillovers many researchers use patent data (Griliches 1990). The first studies centered on the geographical (rather than organizational) mobility of knowledge workers. Jaffe, Trajtenberg, and Henderson (1993) investigated the geographical reach of spillovers through matched samples of patent citations, controlling for time patterns and technological similarity. They found that knowledge spillovers initially tended to be geographically localized, and then subsequently spread over large distances. Further work recognized that geographical distance may be mitigated by the social network created between firms from the loss or gain of an employee (Singh 2005; Agrawal, Cockburn, and McHale 2006; Corredoira and Rosenkopf 2010), although Breschi and Lissoni (2009) found that this knowledge diffusion is hampered by the reluctance of inventors to relocate spatially
Other studies focus on the individual level and the organizational integration of knowledge from spillovers at the firm level. The threat of interfirm mobility of inventors sometimes induces firms to patent strategically with the intention of appropriating as much of the returns to their innovations as possible, limiting the knowledge spillover that takes place between firms (Kim and Marschke 2005; Schankerman, Shalem, and Trajtenberg 2006). Other mechanisms include noncompete agreements and employee retention contracts to protect innovations (Marx et al. 2009) as well as a reputed toughness in patent litigation (Agarwal, Ganco, and Ziedonis 2009). These factors may hamper the mobility of inventors and the extent of technological progress (Cooper 2001).
Studies of worker inflows focus on firms’ ability to obtain knowledge and learn from an inventor recruited from another firm (Singh and Agrawal 2011; Palomeras and Melero 2010; Song et al. 2003). Song et al. (2003) conclude that knowledge transfers through labor mobility are more likely if the hiring firm is less technologically path dependent. Palomeras and Melero (2010) find that the quality of the inventor’s work and the level of complementarity with core competencies of a firm that an inventor is moving to has a positive impact on mobility. Singh and Agrawal (2011) identify the spillover effect from citations by combining Jaffe et al.’s (1993) matched sample approach with the change in citation patterns following a newly recruited inventor’s patents compared to pre-recruitment patents. Singh and Agrawal (2011) find that a large part of estimated knowledge spillovers is associated with the newly recruited inventor’s use of his/her own prior inventions, rather than organizational learning. Maliranta, Mohnen, and Rouvinen (2009) find that it is primarily workers who move from R&D to non-R&D activities who help boost the recipient firm’s productivity. Partly corroborating this, Kaiser, Kongsted, and Rønde (2015) investigate how mobile R&D workers affect the patenting activities of both the departed and the recipient firm. They find that the effect is stronger if the firm that the R&D worker leaves is already patent active.
Similar to us, others focus on the subsequent patent productivity of the mobile inventor. Trajtenberg and Shiff (2008) find that inventors who hold higher-quality patents are more mobile than inventors producing lesser-quality patents. Furthermore, the patents these inventors produce after moving are of higher quality. This suggests a positive relationship between patent quality and mobility, although no causal link is established. Team experience is further found to exert a negative impact on mobility, perhaps due to network effects or good matching. Schettino, Sterlacchini, and Venturini (2013) find that patent quality is associated with the age of the inventor, being male, having higher education as well as working in teams, although the individual characteristics are not found to influence patent productivity.
Trying to determine causal effects rather than just correlations, Hoisl (2007) investigates whether mobility and productivity are endogenously determined and looks at the causal relationship between the two for a survey dataset of German inventors. Using two sets of instrumental variable estimations, she finds a simultaneous relationship in which mobile inventors are more productive than their immobile counterparts but also that an increase in productivity reduces the likelihood of a move. As instruments for mobility, she includes incentives for inventive activities (taken from a questionnaire), the technical concentration of patents, and the size of the region in which the inventor works. The instruments for productivity are related to external sources of knowledge, that is, to the extent that inventors use patent documents and scientific literature to get input rather than through personal interaction, and are also taken from a questionnaire. However, difficulties in collecting comprehensive data place limitations on the generalizability of the study by Hoisl (2007). First, the study uses only inventors with at least two patent applications, although other studies indicate that most inventors have only one patent (Trajtenberg, Shiff, and Melamed 2006; Ejermo 2011). Second, the inventors analyzed by Hoisl (2007) have an average patent productivity of 14.7, which means it is highly skewed toward the more inventive part of the distribution.
In a follow-up study using a matched sample approach, Hoisl (2009) determines whether mobility affects high- and low-productivity inventors differently, using jointly estimated quantile regressions. She finds that inventors who are initially more productive benefit more, in terms of their patents receiving more citations from changing jobs than those who are initially less productive. Her results also suggest that inventors who are poorly matched to their employers tend to move to increase their productivity. Lenzi (2008) tries to overcome the potential high-productivity bias stemming from only using patent records to determine mobility by complementing patent documents with curriculum vitaes. She examines Italian inventors active in the pharmaceutical industry to determine whether interfirm mobility effects differ for these two types of records. In effect, important knowledge can be gained at one job although no patent is applied for or is kept secret by the inventor, and this knowledge is subsequently used to file a patent at another firm. She finds that patent documents often underestimate the number of moves that inventors actually undertake over the course of their careers as well as specify the affiliation of the inventors incorrectly. Using Poisson regressions, the results suggest a significant positive effect running from mobility on patent documents to both productivity and the number of citations that each patent of the inventor receives, as well as from productivity to patent document mobility (i.e., in the opposite direction). Basing instead mobility on CVs, she finds an insignificant effect on productivity and a positive significant effect on the number of citations each patent receives. Also, the effect runs from productivity toward mobility using CVs. These results cast doubt on the possibility to infer mobility from patent documents. Lenzi (2009) uses a duration analysis and incorporates more controls. The results indicate that life-cycle effects, inventive productivity, and the geographical location in which the inventor works are drivers of mobility. It is mainly the most productive inventors who are likely to move. Nevertheless, the results are largely consistent with Hoisl (2007), at least when only considering patent documents, in that mobility seems to be a mechanism that improves the matching between inventors and employers, which increases productivity. However, due both to the small sample size (106 inventors in one industry surveyed) and the absence of the possibility of causally establishing effects, the conclusions may not be generalizable.
Ge, Huang and Png (forthcoming) investigate the extent to which patent data alone misclassifies mobility by studying inventors’ LinkedIn profiles as well as conducting a small survey to verify their results. They find that the accuracy of patent data is 70 % or less, whereas LinkedIn provides 90 % or higher accuracy. Patent data tends to provide a high degree of false negatives, i.e. failing to observe mobility, and a smaller degree of false positives (i.e. observing mobility when there is none). Furthermore, their results suggest that the bias is larger for inventors who patent less frequently or have shorter careers. Running several regressions with mobility as the dependent variable, they show that the patent rate negatively affects mobility - i.e. the same results as found in Hoisl (2007).