Chapter 5. FIRM CHARACTERISTICS AND OUTCOMES

METHODOLOGY

In this section, we examine the firms within a consortium to answer the question: What kind of firm benefits the most from participation? Here, the unit of observation is a firm’s participation in an individual project. We seek to identify firm characteristics that are associated with measures of research success in the targeted areas during and after participation in a particular research project.

This component of our research faces one important data problem; we do not have good data on the division of individual firms’ R&D budgets across research projects, including that fraction of the R&D budget spent on consortia-related research. What we do know is each firm’s total R&D spending per year. We also know the total R&D budget for each consortium and how this total was divided between the government and the private sector. Using this information, we imputed a firm-level, project-specific R&D budget (inclusive of subsidies) by dividing the total annualized budget by the number of participants. While this will not necessarily be an accurate measure of the actual individual firm’s investment in project-related technology in every year, this is the best that can be done with the current data. Our firm-level variables are defined below.

Pre-project patenting in the targeted area. The same logic applies here as in the case of the more aggregated consortium-level data. In order to isolate the impact of the project on the firm, we need a quantitative measure of its research competence in the targeted area. To measure pre-existing patenting levels in the targeted area, we use an average measure of pre-project patenting in the targeted area taken over a five-year window prior to the official start date of the project, or as much of this window as the available data permit.

Total R&D spending. For a large number of firms, we have high-quality panel data on overall research and development spending. A positive association of this variable with the outcome measure would suggest that the technologically more progressive firms are the prime beneficiaries of the projects. A negative association would suggest that it is the technological followers that benefit rather than the technological leaders.

Industry effects. Due to large and persistent differences in the propensities of firms in different industries to patent, we include industry fixed effects¹ in our regression analysis. Each participating firm is classified into one of seven industry classes. The inclusion of industry effects enables us to determine whether consortia in some industries are substantially more productive than other industries.

Capital stock. This is included primarily as a measure of firm size. We also have data on employment for a large number of participating firms, which can serve as an alternative measure of size. This is included to get some insight into whether large firms or small firms benefit most from consortia participation and to partially control for other unmeasured firm characteristics that are correlated with size.

Effects of overlapping projects. Firms may have a particularly high output in a certain project, controlling for inputs, but this may simply reflect the firm’s simultaneous participation in another consortium that targets the same technologies. We control for overlapping projects by imputing the firm’s subsidies and private contributions to each overlapping project. This measure is included in the variable real_indirect_inputs.

RESULTS

Our baseline empirical specification for analysis at the consortium-firm level resembles that of the consortium-level analysis. The essential difference is, of course, that we have an additional dimension of variance: we observe multiple firms participating in the same projects and, conversely, multiple projects impacting the same firms. We initially estimate:

(4)

where j denotes the firm, i denotes the project, and t denotes the year. As in the previous analyses, we allow the project boost effect to have three components: an initial boost contemporaneous with the duration of firm j’s involvement in project i (project_duration_jit); an effect that captures the—possibly lagged—increase in patenting as years since the inception of the project increase (years_passed_jit); and a quadratic term in years elapsed since the inception of project i to allow for the decline in patenting that eventually sets in (years_passed²_jit). We also include as controls firm j’s share of the total budget (private outlays and public subsidy) for project i (budget_jit) and firm j’s pre-project level of patenting in the technologies targeted by project i (pre_project_patenting_jit).

The results of our initial regression are given in column 1 of Table 5. The coefficient on project_duration is positive but not statistically significant. The effects on years_passed and years_passed² are statistically significant and have the expected sign. The results suggest the boost to firm patenting stemming from participation is not immediate, but takes place with a lag. As in the consortium-level regressions, the regression coefficients suggest an effect that peaks about two years after the inception of the project and then declines.

Table 5. Firm-Consortium Level Analysis
                  Poisson regression
                  Dependent variable: firm patenting in the targeted area

Variables	(1)	(2)	(3)	(4)	(5)
Budget	-1.01e-06 (2.15e-08)	-1.56e-07 (7.75e-09)	-3.19e-07 (9.63e-09)	-.003 (.0000167)	-2.19e-07 (7.84e-09)
Pre-project patenting	.005 (.0000218)	.0042 (8.64e-06)	.004 (.0000142)	-1.24e-07 (8.41e-09)	.004 (9.77e-06)
Project duration dummy	11.86 (26.74)	—	—	—	—
Years passed	.566 (.019)	—	—	—	—
Years passed2	-.120 (.005)	—	—	—	—
Year 0 dummy(a)	—	.065 (.013)	.084 (.014)	.008 (.013)	.141 (.013)
Year 1 dummy	—	.772 (.015)	.574 (.016)	.469 (.016)	.930 (.016)
Year 2 dummy	—	.9998 (.017)	.901 (.018)	.777 (.017)	1.01 (.017)
Year 3 dummy	—	.745 (.02)	.558 (.020)	.474 (.020)	.915 (.020)
Year 4 dummy	—	.710 (.033)	.505 (.034)	.300 (.033)	.999 (.033)
Chemicals	—	—	—	-1.08 (.029)	—
Machinery	—	—	—	-.036 (.018)	—
Transportation	—	—	—	-1.169 (.021)	—
Precision instruments	—	—	—	.111 (.018)	—
Fabricated metals	—	—	—	-2.986 (.139)	—
Other manufacturing	—	—	—	-.423 (.030)	—
Average technological proximity	—	—	.987 (.020)	—	—
R&D spending	—	—	.0000966 (3.92e-06)	.000284 (3.79e-06)	—
Real net capital stock	—	—	2.41e-06 (3.04e-07)	—	—
Real indirect inputs	—	—	—	—	-1.28e-08 (2.66e-10)
Constant	-8.766 (26.738)	2.950 (.007)	2.771 (.012)	3.335 (.0130)	2.943 (.007)
R-squared	.6990	.6646	.7450	.7583	.6780

Note: Standard errors in parentheses. The R-squared measure for the Poisson regression given here is a pseudo-R-squared measure. The reference sector of industry dummies is electronics.
(a) Year 0 indicates the year of the inception of a consortium.

As a check on the robustness of the results, we estimate a more flexible “time path of benefits” structure using the set of year dummy variables. The results are given in column 2. These results are generally consistent with the picture we obtained from the results of column 1. Again, we have a boost that affects firms with a lag, peaks relatively early, and then declines.

Column 3 includes controls for the average technological proximity of firms within projects, overall R&D spending, and firm size as measured by net deflated capital stock. Average_technological_proximity is the same variable used in the previous section. R&D_spending measures firm j’s overall R&D spending in year t. This helps control for changes in the overall R&D intensity (and, potentially, R&D productivity) of firm j over time. Because we do not have R&D data for all firms in all years, our total number of observations declines in this specification. The results indicate that all three variables are positively associated with research outcomes. The results for R&D_spending and real_net_capital_stock imply that larger firms benefit more from consortium participation. However, the magnitude of the positive coefficient on the real_net_capital_stock variable is quite small, so it is not immediately obvious what its economic significance is.

Column 4 includes the industry dummy variables. A negative coefficient on an industry dummy variable suggests that, relative to the reference sector (electronics), firms generate fewer patents as a consequence of participation in a consortium. However, the coefficients on these industry dummy variables represent not only the differential effects of participation, but also the differential extent to which innovation resulting from participation is codified into patents.

We conducted two robustness tests, one that includes our measure of indirect inputs and a second that includes firm fixed effects. The former is presented in column 5 and indicates that our basic result survives this robustness check. The latter is shown in Table 6. The number of parameters needed to estimate firm fixed effects makes Poisson regression computationally impossible. Table 6 compares column 2 from Table 5 to a linear specification of the model with the firm effects added (but the fixed effects coefficients suppressed). The time path of benefits is essentially unchanged. These results are inconsistent with the view that project success is simply driven by the inclusion of “good” firms. Rather we find that, controlling for the unobserved research quality of firms within the targeted area, participation is associated with an increase in patenting in that area.²

In some cases, measuring the patent output of a particular firm, in a particular project, in a particular year is not disaggregated enough. A number of frequent participants in ATP-funded consortia were subsidiaries of large firms. The subsidiary’s participation may constitute a small part of the larger firm’s total research effort. Although the subsidiary’s participation may have little impact on the entire firm’s research effort, it may play a significant role in the subsidiary’s research effort. We thus sought to isolate the patenting of the participating subsidiary as our measure of research output. To do this, we took the patents assigned to the corporation and selected out that subset of patents invented by individuals residing in the same geographic area as the participating subsidiary. This, we reasoned, was as close to the subsidiary’s patents as the available data would allow us to get. The results are presented in Table 7 and are quite similar to the results presented in columns 2 and 3 of Table 5.

Table 6. Firm-Consortium Level Analysis: Comparison of Poisson and OLS Regression Models*
                  Dependent variable: Sum of patent grants by consortium participants in the targeted area

Variables	(1) Poisson	(2) OLS
Budget (7.75e-09)	-1.56e-07 (4.25e-06)	-9.25e-06
Pre-project patenting	.0042 (8.64e-06)	.961 (.019)
Year 0 dummy(a)	.065 (.013)	-1.651 (3.79)
Year 1 dummy	.772 (.015)	17.548 (7.095)
Year 2 dummy	.9998 (.017)	23.918 (7.899)
Year 3 dummy	.745 (.02)	3.445 (8.263)
Year 4 dummy	.710 (.033)	-31.758 (14.790)
Constant	2.950 (.007)	41.061 (28.229)
R-squared	.6646	.8831

Note: Standard errors in parentheses. The R-squared measure for the Poisson regression given here is a pseudo-R-squared measure.
* The OLS model is computed with the firm effects added, but the model is presented here with the coefficients suppressed.

IMPLICATIONS

In this section, we demonstrated that there is a statistical link between a firm’s participation in an ATP project and that firm’s patenting in the technologies targeted by the ATP consortium. This approach gets us as close as we can to causal identification between consortia participation and patenting outcomes without randomized experiments. We also demonstrated that this positive association between participation and patent output is not simply the result of “better” firms being systematically selected for more frequent participation. The patent boost from participation remains positive and statistically significant even when controlling for unobserved firm fixed effects, such as the firms’ research productivity in the targeted technologies.

Table 7. Subsidiary-Consortium Level Analysis*
                  Poisson Regression
                  Dependent variable: firm patented in the targeted area

Variables	(1)	(2)
Budget	2.07e-07 (1.07e-08)	1.70e-07 (1.52e-08)
Pre-project patenting	.009 (.0000276)	.007 (.0000471)
Year 0 dummy(a)	.105 (.020)	-.060 (.022)
Year 1 dummy	.811 (.025)	.446 (.026)
Year 2 dummy	.956 (.0259)	.734 (.027)
Year 3 dummy	.668 (.030)	.308 (.032)
Year 4 dummy	.664 (.049	.136 (.051)
Average technological proximity	—	1.160 (.030)
R&D spending	—	.0002905 (6.87e-06)
Real net capital stock	—	-9.34e-06 (6.04e-07)

* Measuring the subsidiary’s patents is not possible with the current data. This analysis uses patents invented by individuals residing in the same geographic area as the participating subsidiary as a proxy measure of the subsidiary’s patents.
Note: Standard errors in parentheses. The R-squared measure for the Poisson regression given here is a pseudo-R-squared measure.

We also began to address the kinds of firms that benefit most from consortia participation. We find that our measure of technological proximity is positively and significantly correlated with research outcomes in the presence of other control variables, suggesting that firms participating in consortia composed of other firms with similar patenting portfolios tend to do better. We also find some evidence that firms’ total R&D spending and firm size are also positively correlated with research outcomes. However, the economic significance of these coefficients is not clear given that our sample of ATP participants is not complete. In the absence of panel data on the research inputs and outputs of smaller firms, it is difficult to come to any definitive conclusions about the role of size or overall R&D spending in effecting research outcomes.

Return to Contents or go to next section.