Chapter 3. OVERALL BENEFITS FROM RESEARCH CONFORTIA

METHODOLOGY

If ATP-funded consortia enhance the research productivity of member firms by promoting research spillovers among members, then we may observe a statistical relationship between the intensity of participation and the firm’s patent output in that year.¹ We seek to directly test this hypothesis using panel data on participating firms and a control group of non-participants.

We draw heavily from the methodology developed in Branstetter and Sakakibara (1998) and use the following simple log-linear equation, derived from a knowledge production function to estimate patenting output:

(1)

where p_it is the natural log of the number of patents generated by firm i in year t; r_it is the natural log of firm-level R&D spending; Cit is the intensity of participation in research consortia, measured as the count of concurrent projects in which firm i was involved in year t; d’s are the coefficients on our industry dummy variables (Ds), and µ is an error term. The d terms represent industry-level differences in the propensity to patent.

The process by which firms select one another as joint venture partners and the process by which ATP selects joint venture proposals for funding are not processes of random assignment. When seeking out joint venture partners, firms would logically seek to affiliate with firms that conduct high-quality research. Moreover, it is quite likely that ATP funds consortia consisting of high research quality firms. If research productivity is positively correlated with the intensity of participation in consortia, it may be that the chain of causality runs from research productivity to participation rather than the other way around.²

To deal with this problem, we could, in principle, take two approaches. One is to use a model with firm “fixed effects” (which also removes industry effects). A fixed-effects estimate gives us potentially unbiased and consistent estimates of all parameters, albeit at the cost of losing the cross-sectional variance in our data, which is most of the total variance. As Griliches and Hausman (1986) have shown, however, the fixed-effects estimator may itself be biased in the presence of measurement error. Given the imperfections of patents as indicators of innovative output and our measures of firm-level R&D spending as measures of innovative input, some level of measurement error is virtually certain.

Unfortunately, this is not the only shortcoming of this framework. At least some firms that participated in ATP consortia were large firms with large, diversified research portfolios. The technological focus of the ATP consortia may be only a small part of their total research agenda, and hence account for only a small or even trivial fraction of their total patenting. Thus, movements in the measure of output used here—total patents—could be partially or entirely unrelated to the actual outcomes of the firm’s research program that is involved in the ATP consortium. This argues for a more disaggregated approach to the data that focuses on patenting only in the areas targeted by the ATP consortia, which is the subject of the next two sections.

At this level of aggregation, the empirical alternative to a fixed-effects model is an instrumental variables approach. Such an approach assumes that changes in the intensity of participation in research consortia (C) is actually described by lagged values of Cs as well as other observable variables. This assumption implies that there is some “inertia” to the selection process. Perhaps firms that participated in the past would be more likely to participate in present projects regardless of their true research quality. Thus, we can achieve identification by using “predetermined” or k-lagged values of C_it as instruments, where k is a lag long enough to be exogenous with respect to research quality. It is not possible, however, to apply this alternative approach to our current U.S. data. At this stage, we do not possess long lags in the “participation” variable. Therefore, inference in the U.S. data will be based on fixed- and random-effects estimates.³

DATA

Our data consists of an unbalanced panel of 249 firms, 65 of which have participated in at least one ATP project. Data cover the years from 1985 through 1995. (Additional information regarding the data sources and construction of the panel data are explained in the Appendix.) The ATP provided information on 96 ATP-funded consortia, including the members of the consortia, the total budget, the time frame of the project, and the technological goals of each research joint venture.⁴

Information on total R&D spending, sales, and capital investment of participating and non-participating firms was obtained from Standard & Poor’s COMPUSTAT database. We do not have information on all participants for all years for several reasons: first, a number of smaller ATP project participants do not show up in the COMPUSTAT database in any year; second, some small firms only have data for the most recent years; and third, the data we obtained from COMPUSTAT only extend through 1995.

Information on the total patenting of firms was obtained from the REI Patent Database developed and maintained at the Case Western Reserve University Center for the Study of Regional Economic Issues. This database allows us to date patents by the date of application rather than the date of grant. This is important because the lag between the development of an idea by a firm (at which point the patent is applied for) and the granting of a patent by the USPTO can be as long as two to three years. At the time of this analysis, the REI Patent Database only contained information on patents granted through 1996. This means that our information on patents applied for effectively only goes up to 1994 or early 1995.

Thus, we confront two sources of “truncation” in the data. First, there is truncation in the cross-section dimension of our data. Many small firms are completely absent from our panel. Second, there is truncation in the time-series dimension of our data. Effectively, our data on the research inputs and outputs of the firm end in 1995 and do not capture the large number of ATP projects begun after 1995. Moreover, few projects that had begun before 1995 had actually ended by 1995. If the effects of participation continue long after the official end of a joint venture, we will likely underestimate the impact of consortia participation in our empirical analysis due to truncation of the data. To the extent that participation led to the strongest enhancement of research productivity in smaller firms, the lack of data on small firms could also lead to an underestimate of the total impact of consortia.

RESULTS

Table 1 presents some sample statistics of participants in ATP-funded research consortia versus non-participants. T-tests comparing these two groups indicate that firms in the participant sample are larger (as measured by nominal and inflation-adjusted sales), conduct more R&D, and generate more patents than firms in the non-participant sample. Given ATP’s emphasis on the inclusion of smaller firms as joint venture partners, these findings appear contrary to expectations. Larger firms, however, are often partners in research consortia even though they are rarely joint venture leaders, and participation of subsidiaries in a joint venture is “credited” to the larger parent firm. In addition, smaller firms that do not publicly report R&D expenditures or generate patents do not show up in our database because there is no publicly available data on them in our original sources.

Table 1. Summary Statistics for Non-Participants and Participants of ATP-Funded Consortia

	Non-participants			ATP participants
Variables	Mean (Std. dev.)	Min.	Max.	Mean (Std. dev.)	Min.	Max.
Patents	37.71 (89.22)	0.00	942.00	131.51 (204.31)	0.00	1,413.00
Real R&D	106.37 (215.10)	0.00	1.490.14	436.51 (912.65)	0.03	6,667.64
Sales	2,911.33 (6,996.35)	0.05	69,276.00	12,912.94 (24,036.08)	1.29	16,537.0 0
Real sales 2589.07	0.05 (6,156.09)	59,279.66	11,451.88	1.07 (21043.83)	127,501.9	0
Real net capital stock	2,273.38 (8,110.09)	0.00	161,717.30 0	8,835.99 (15,904.86)	0.37	95,607.24
Number of observations	1,898			684

Note: Units are millions of U.S. dollars. Patents are measured as total grants per year by date of application. These sample statistics are drawn from only one of several alternative approaches employed in this study.

Table 2 presents various specifications of equation (1). We add a full set of year dummies to our regression, which partly controls for the time-truncation effect in our patent data. Industries are classified into seven sectors. Given our earlier concerns about reverse causality, we estimate a random-effects specification (column 1) and a fixed-effects specification (column 2). In both cases, the measure of intensity of participation, C, is positively correlated with patenting. The coefficient is statistically significant in both specifications.

Columns 3 and 4 give the results when we include the log of the deflated net capital stock as a control for the size of the participating firms. This variable is included because previous research indicates a positive correlation between firm size and the propensity to patent. Controlling for firm size, the impact of participation in ATP-funded consortia on research output remains positive and significant in all specifications. These results are quite striking given the many imperfections of our data. It suggests that there may indeed be substantial spillover-enhancing effects achieved through these research consortia.

IMPLICATIONS

This section gives us a measure of the extent to which involvement in ATP projects impacts the entire research portfolio of participating firms. Since the participants in our data set include large firms with quite diversified research portfolios, we will only observe large, significant impacts in this analytical framework if the effect of participation is the generation of knowledge spillovers that transcend the generally narrow technical focus of the project itself. So how large is the effect of consortia participation on the overall patenting of participating firms? Perhaps it is most instructive to look at column 4 of Table 3—the results from the fixed-effects specification with all controls in place. The coefficient on the variable C is 0.075. This suggests that participation in one additional ATP-funded research consortium per year would generate an increase in patenting in that year of nearly 8%. We note that the causal interpretation that we give in this statistical relationship between research productivity and participation is not the only possible interpretation. In the absence of a randomized experiment in which ATP makes awards to research consortia without regard to research quality, it is difficult to prove that the statistical relationship we document is causal. Nevertheless, our best efforts within the context of this framework suggest that ATP has a positive impact on the research productivity of the firms participating in its research consortia. We are not able, within this framework, to undertake a full cost-benefit analysis of ATP’s funding; however, the statistical link we find between consortia participation and overall patenting is a necessary, though certainly not sufficient, condition of establishing that ATP’s investment is socially productive.

Table 2. Estimation of a Patent Productuion Function: Overall Effect of Participation (Dependent variable: Log of patents granted per firm classified by the year of application)

Variables	(1) Random Effects	(2) Fixed Effects	(3) Random Effects	(4) Fixed Effects
Real R&D	.567 (.019)	.424 (.027)	.315 (.029)	.265 (.034)
Real net capital stock	—	—	.351 (.030)	.311 (.04)
C	.081 (.026)	.066 (.026)	.088 (.026)	.075 (.026)
Chemicals	.353 (.148)	—	-.089 (.143)	—
Machinery	.185 (.160)	—	-.101 (.152)	—
Transportation	.031 (.202)	—	-.41 (.193)	—
Precision instruments	-.071 (.171)	—	.041 (.160)	—
Fabricated metals	-.286 (.435)	—	-.577 (.408)	—
Other manufacturing	.526 (.498)	—	-.011 (.469)	—
Year 1986 dummy	.028 (.049)	.036 (.048)	.0322 (.048)	.036 (.048)
Year 1987 dummy	.124 (.049)	.139 (.048)	.115 (.048)	.124 (.048)
Year 1988 dummy	.082 (.049)	.108 (.048)	.061 (.048)	.088 (.048)
Year 1989 dummy	.120 (.049)	.151 (.049)	.088 (.048)	.109 (.048)
Year 1990 dummy	.131 (.049)	.161 (.049)	.082 (.048)	.104 (.049)
Year 1991dummy	.121 (.049)	.152 (.049)	.077 (.048)	.100 (.049)
Year 1992 dummy	.115 (.05)	.154 (.05)	.066 (.049)	.093 (.049)
Year 1993 dummy	.051 (.05)	.093 (.05)	.009 (.049)	.038 (.05)
Year 1994 dummy	-.058 (.051)	-.012 (.050)	-.126 (.050)	-.093 (.051)
Year 1995 dummy	-.808 (.053)	-.746 (.053)	-.895 (.053)	-.849 (.054)
Constant	.594 (.120)	1.20 (.096)	-.473 (.145)	-.149 (.198)
R-squared	.7004	.6958	.7281	.7190

Note: Standard errors in parentheses. The reference sector of industry dummies is electronics. C is the count of concurrent projects in which firm i was involved in year t. Both R&D and net capital stock are measured in logs.

NOTES:

Return to Contents or go to next section.