INEQUALITY, DISCRIMINATION, AND VIOLENCE ‡ On the Returns to Invention within Firms: Evidence from Finland

In this paper we merge individual income data, firm-level data, patenting data, and IQ data in Finland over the period 1988–2012 to analyze the returns to invention for inventors and their coworkers or stakeholders within the same firm. We find that: (i) inventors collect only 8 percent of the total private return from invention; (ii) entrepreneurs get over 44 percent of the total gains; (iii) bluecollar workers get about 26 percent of the gains and the rest goes to white-collar workers. Moreover, entrepreneurs start with significant negative returns prior to the patent application, but their returns subsequently become highly positive.

and Väänänen (2012); Bell et al. (2017); and Akcigit, Grigsby, and Nicholas (2017). Toivanen and Väänänen (2012) use Finnish patent and income data to study the return to inventors of US patents. They find strong and long-lasting impacts, especially for the inventors of highly cited patents. Bell et al. (2017) merge US individual fiscal data, test score information and US individual patenting data over the recent period to look at the life cycle of inventors and the returns to invention. Akcigit, Grigsby, and Nicholas (2017) merge historical patent and individual census records to study, among other things, inventor compensation. We complement the existing literature by offering new evidence on the returns to inventors, but foremost by offering what to our knowledge is the first evidence on wage spillovers to non-innovating coworkers of different types. 2

I. Data
Our data come from the following sources. First, the Finnish Longitudinal Employer-Employee Data (FLEED) which we exploit for the period 1988-2012. FLEED is an annual panel constructed from administrative registers of individuals, firms, and establishments, maintained by Statistics Finland. It includes information on individuals' labor market status, salaries, and other sources of income extracted from tax and other administrative registers. It also includes information on other individual characteristics and employer and plant characteristics. Second, the European Patent Office data provide information on characteristics such as the inventor names and applicant names. 3 We have collected patent information on all patents with at least one inventor who registers Finland as his or her place of residence. We use data on all patents with a Finnish inventor up to and including 2012. Third, the Finnish Defence Force provides us with information on IQ test results for conscripts who did their military service in 1982 or later; all conscripts take the IQ test in the early stages of the service. These data contain the raw test scores of visuospatial, verbal, and quantitative IQ tests. We follow Aghion et al. (2017) and use the visuospatial IQ percentiles. 4 We limit our estimation sample to years 1994 -2010 to allow for pre-trends in the early part of the data sample and to ensure sufficient coverage of patent applications in the late parts of the data. In this paper we focus attention on male workers who did their military service in 1982 or later (meaning they were born in 1961 or later). To ensure sufficient labor market participation (individuals enter FLEED at age 15), we require positive wage income in preceding four years of the included observations. Finally, we restrict attention to private sector employees because we can only identify coworkers in the private sector.
We identify an individual as a coworker or stakeholder within the same firm if he: (i) works in the inventing firm in the year of the patent application, and (ii) is never an inventor himself. We study the following classes of coworkers or stakeholders within the same firm besides inventors: (i) entrepreneurs; 5 (ii) white-collar workers; 6 and (iii) blue-collar workers. 7 3 Here we want to thank the research project "Radical and Incremental Innovation in Industrial Renewal" by the VTT Research Centre (Hannes Toivanen, Olof Ejermo, and Olavi Lehtoranta) for granting us access to the patent-inventor data they compiled. 4 All the registry data is matched using individual identifiers. The matching of patent data to registry data is described in Aghion et al. (2017). 5 Individuals within the same firm are identified as entrepreneurs if: (i) they contribute to the entrepreneur pension system, and (ii) they own at least 50 percent of the company. 6 These and the remaining individuals' job status are identified through the socioeconomic status code contained in the FLEED. 7 The merged data contain 15 million observations on over 700,000 individuals who work in some 300,000 firms. 7,033 individuals invent at least once (conditional on inventing, avg. #applications = 3.08, median = 1). The annual number of observations varies between 340,000 (in 1988) and
Our specification includes: (i) individual fixed effects; (ii) treatment time fixed effects, with t = 0 denoting the year of patent application (baseline is t = − 5 ); (iii) calendar year fixed effects (baseline year 1994); and (iv) age fixed effects (baseline is a ≤ min (age) + 1 which may vary across estimation samples). The variable treate d i is an indicator variable taking value 1 if individual i belongs to the treatment group (inventor or coworker of type k = entrepreneur, blue-collar worker, white-collar worker) and 0 otherwise, and the α s denote the coefficients of the various fixed effects. We cluster standard errors at the individual level throughout.
We employ a conditional difference-indifference approach whereby we first match each treated individual with a control individual. 8 The matching is done without replacement on an annual basis, starting from 1994. Due to the 730,000 (from 2006 onwards). In the merged data, we have the following proportions of inventor and coworker observations: (i) inventors: 0.011; (ii) entrepreneurs: 0.048; (iii) white-collar workers: 0.270; (iv) blue-collar workers: 0.316; (v) others: 0.355. See Table A1 in the online Appendix for descriptive statistics on wage income. 8 For a similar approach, see Jaravel, Petkova, and Bell (forthcoming). We implement one-to-one matching using the coarsened exact matching of Iacus, King, and Porro (2012). small number of potential control individuals, we use a three-year period for entrepreneurs. We limit the potential control group to individuals who never invent and have never been coworkers of an inventor and who work in the private sector in the year of treatment. We use the following variables for matching: (i) having at least a master of science degree (MSc); (ii) having a STEM education; (iii) working in manufacturing; (iv) living in the southwest of Finland; (v) age ( < 30, 31-40, 41-50, > 50); (vi) quintiles of the annual firm size distribution; and (vii) having visuospatial IQ less than the fiftieth percentile, in the fifty-first to eightieth, in the eighty-first to ninetieth, or above the ninetieth percentile.
We execute the matching separately for each treated group (inventor, entrepreneur, blue-collar worker, white-collar worker). This choice means that apart from inventors, the matching is done within the same socioeconomic group. For white-collar workers, we perform the matching separately within the following subcategories: (i) senior managers; (ii) senior workers; (iii) junior managers; and (iv) junior workers. 9 Table 1 shows the results for our baseline regression where we constrain the treatment effect δ t to be constant both after the year of the patent application (i.e., δ t = δ post for t = 0, . . . , 10 ) and before that year (i.e., δ t = δ pre for t = − 4, . . . , −1 ). In other words, we allow for constant but different post-treatment and pre-treatment (or anticipation) effects.

III. Regression Results
We find that inventors earn on average a wage premium of 5 percent post invention, and earn on average 4 percent prior to invention starting 4 years before invention. This is similar in magnitude to what Toivanen and Väänänen (2012) report for annual returns a few years after the patent is granted. Next, we look at coworkers, and we find returns that are heterogeneous across the different types of coworkers. Entrepreneurs earn the highest returns with almost 28 percent post invention, but nothing pre-invention. White-collar worker earn a return of about 2 percent post invention but nothing pre-invention, and blue-collar workers earn 2.3 percent post invention but lose 1 percent pre-invention. Incidentally, it is interesting to see that bluecollar workers experience a post-invention return on par or slightly higher than those experienced by white-collar workers.
We then turn to the full specification of equation (1). We display the results for inventors, entrepreneurs, blue-collar, and white-collar workers in Figure 1. Inventors earn returns already in anticipation of the patent application; after the patent application, there is a slight (though statistically indistinguishable) decrease, but soon, the returns start again to increase.
The estimated returns to entrepreneurs display a markedly different path. They start with significant negative returns in anticipation of the patent application but already before the year of patent application the returns turn positive. Then the entrepreneurs' returns keep rising and reach a maximum above 20 percent between two and two and half years after the invention time, with some fluctuations year to year thereafter. A potential explanation for the negative anticipation returns is that these entrepreneurs in innovative (and small) companies are credit constrained, and they finance invention partly by foregoing own consumption.
We have checked the robustness of these results in several ways. These results are reported in Table A2  Our results are robust to these changes with two expected exceptions: first, the estimated returns to inventors are reduced when we do not allow for anticipation effects (Table A1, column 2) which were estimated to be positive (Table 1).
Second, the estimated returns to entrepreneurs are lower (0.13) and not statistically significant if we exclude observations with missing information on the number of employees (Table A3, column 8). With this rule, we lose 20 percent of the estimation sample of entrepreneurs as the rule excludes mainly observations from small, often entrepreneur-driven, firms. An important aspect of the returns to invention is an understanding of how the proceeds from invention are shared among the different types of workers within the innovating firm. To illustrate this, we use the coworker-type specific return estimates from Figure 1, the shares of different types of coworkers in innovating firms (we use 2003 data) and the wages of different types of coworkers in innovating firms before invention (we use mean wages in our base year, i.e., t = − 5 ). Using these numbers, we calculated both the total dollar-increase in the wage bill of an innovating firm, and how it is shared between these different types of workers. The result, displayed in Figure 2, reveals some interesting Notes: Standard errors in parentheses, and clustered at the individual level. Estimation samples are based on CEM one-to-one matching using annual data without replacement, starting from 1994 with the following matching criteria: (i) having a science education; (ii) having at least an MSc; (iii) working in manufacturing; (iv) region (2 regions); (v) firm size (quintiles); and (vi) visuospatial IQ (4 groups). For all groups but inventors, the matching is done within the socioeconomic group and for white-collar workers, within subgroups. The dependent variable is the natural log of the wage of the individual in a given year, measured in 2014 euros. Treated is an indicator variable that takes value 1 for each observation of an individual who belongs to the treatment group and is 0 otherwise, post is an indicator variable that takes value 1 in the year of receiving the treatment and thereafter and is 0 otherwise, and pre is an indicator variable that takes value 1 in the last 4 years preceding the year of treatment and 0 otherwise. All specifications include a full set of calendar year dummies (base year 1994), age dummies (base age ≤ min (age) + 1 ), and a set of treatment time dummies for treatment years, t = − 4, . . . , 10 (base year t = − 5 ). All specifications include the size of the firm (# employees) as a control variable and a dummy for missing employment information.
conclusions: First, inventors get only 8 percent of the total gains; second, entrepreneurs get over 44 percent of the total gains; and finally, blue-collar workers get about 26 percent of the gains and the rest goes to the white-collar workers.

IV. Conclusion
In this paper we start closing the gap on providing evidence on income spillovers from invention within the inventing firm. Using data from Finland 1988-2012 we found significant returns to inventors themselves. Moreover, we found significant spillover effects within the firm, with non-inventing coworkers and entrepreneurs in the same firm also benefiting from the invention. Both white-collar and blue-collar workers benefit from invention; after the invention, if anything, the latter more than the former. Entrepreneurs experience the highest percentage annual gains at over 20 percent. Gains for all groups are long-lasting.
Our findings show that inventors collect only less than 10 percent of the total private return. This result highlights the importance of taking into account the incentives of other actors in the firm (e.g., firm owner and coworkers) who also benefit from an invention both in modeling invention and in drawing policy conclusions (e.g., on taxation).