«John R. Graham, Co-Supervisor David T. Robinson, Co-Supervisor Manuel Adelino Alon P. Brav Manju Puri Aaron K. Chatterji Dissertation submitted in ...»
Next, I standardize the names of the assignee and assignor in the raw patent assignment dataset, patent original assignee names reported in the USPTO databases, and inventor names in HBS inventor database. Speciﬁcally, I employ the name standardization algorithm developed by the NBER Patent Data Project. This algorithm standardizes common company preﬁxes and suﬃxes, strips names of punctuation and capitalization and it also isolates a company’s stem name (the main body of the company name), excluding these preﬁxes and suﬃxes. I keep only assignment records of which the assignment brief is included under “assignment of assignor’s interest” or “Merger,” that is, I remove cases when the reason for the assignment is clearly not transactions such as a “change of names.”
85D.2 Identifying Patent Transactions
The central part of the identiﬁcation of a patent transaction uses several basic principles that predict how patent transactions appear in the data. First, the ﬁrst assignment in a patent’s history is less likely to be a patent transaction. It is more likely to be an original assignment to the inventing ﬁrm. Note that this principle is more helpful on patents granted after 1980, when the raw dataset started to be systematically updated. Second, if an assignment record regards only one patent with the brief reason being “assignment of assignor’s interest,” it is less likely to be a transaction, as it should be rare that two parties transact only one patent in a deal (see Serrano (2010)). Third, if the assignor of an assignment is the inventor of the patent, it is less likely that this assignment is a transaction, but instead more likely to be an employee inventor who assigns the patent to her employer. Fourth, if both the assignor and assignee are corporations, it is likely that this assignment is a transaction, with the exception that the patent is transferred within a large corporation (from a subsidiary to the parent, or between subsidiaries). Based on these principles, the algorithm below is a process in which I remove cases which are
unlikely to be patent transactions. The steps I take are:
1. Check if the assignment record date coincides with the original grant date of the patent (the date when the patent was ﬁrst issued). If it does I label the assignment as a “non-transaction” and it is removed from the data set.
Otherwise, I move to step 2.
2. Check whether the patent assignment record contains only one patent, and is the ﬁrst record for this patent, with assignment of assignor’s interest as the assignment reason. If the answer is aﬃrmative I move to Step 3. Otherwise, the record is labeled as a potential transaction and I move to Step 4.
3. Compare the assignee in the assignment record with the assignee as of the original patent assignment in the USPTO. Similarly, compare the assignor in the assignment record with the inventor names in HBS patent database. If the assignee name coincides, or, the assignor is the patent inventor(s) plus the assignee is a ﬁrm, I then categorize the assignment as a “non-transaction” and it is removed from the dataset. This constraint covers cases when either the assignee or assignor have slightly diﬀerent names across diﬀerent databases.
Otherwise, the record is labeled as a potential transaction and I move to Step 4.
4. Perform the analysis described in step 3 on the “potential transactions” with one minor change: when comparing the assignee in the assignment record with the assignee as of the original patent assignment in USPTO, and when comparing the assignor in the assignment record with the inventor names in HBS patent database, I allow for spelling errors captured by Levenshtein edit distance less or equal to 10% of the average length of the two strings under comparison, and I denote these name as roughly equal to each other. Then, if the assignee name roughly coincides, or the assignor is roughly the patent inventor(s) plus the assignee is a ﬁrm, then assignment is categorized as a “non-transaction” and is removed from the data set. Otherwise, the record is kept as a “potential transaction” and I move to Step 5.
5. Compare the standardized names and stem names of the assignee and assignor of records in the “potential transactions.” If the names coincide, this is consistent with an internal transfer and the record is labeled as a “non-transaction.” If the names do not coincide the record is labeled as a “transaction.”
E.1 Additional Analysis on the CVC Initiation Stage The result that innovation deteriorations motivate CVC initiations raises a number of questions regarding the mechanisms behind it, which challenge the information acquisition interpretation.
One might worry that the result simply documents that desperate managers are more likely to conduct CVC investment to test their luck without carefully evaluating their abilities and potential beneﬁts from investing in CVC. I investigate this issue by studying the success rate of the portfolio companies invested by CVCs categorized by the severity of innovation declines at initiation. If the concern is correct, we would expect CVC parents that experienced the largest hit before initiating to have lower performance as they mostly make the decision under desperation. In Table E.2, I ﬁnd that those CVCs whose parents’ performance decline the most actually score a similar, if not higher success rate compared to other CVCs.
One might also worry that deteriorating ﬁrms choose to launch CVC units as a constrained optimal—that is—declining ﬁrms are more ﬁnancially constrained and 88 cannot conduct internal R&D or M&As, which are on average more costly than CVC.
In Table E.3, I show that the main result is robust on the subsample of ﬁrms whose KZ-index is below industry median or cash ﬂow ratio above industry median (less ﬁnancially constrained).
In order to conﬁrm the result is not driven by the sampling process or speciﬁcations, I conduct a vast of robustness checks. In Table E.4, I show that the result is not sensitive to the length used to capture innovation changes (τ “ 3 in the paper); in Table E.5, I show that the result is robust to removing ﬁrms that are large/small, that are from speciﬁc industries (such as IT or pharmaceuticals), or that are located in speciﬁc locations (in California). In Table E.6 I show that the result holds for deteriorations of product market performance, that is, ROA and growth rate in sales.
E.2 Internal Innovation during CVC
To set the stage, I ﬁrst examine how CVC investment inﬂuences internal innovation.
Under the information acquisition hypothesis, parent ﬁrms should be able to harvest the informational beneﬁt, particularly through improvements in information-sensitive activities such as innovation; moreover, newly gathered information should be reﬂected in those activities.
I assess this idea by characterizing the innovation dynamic of CVC parents around their CVC investment across several dimensions. The ﬁrst set of measurements is simply innovation quantity and quality as employed in Section 4. The second set of variables is New Cite Ratio and Explorativeness, which measure the proportion of new knowledge used in innovation. New knowledge is identiﬁed using patent citations referring to patents that never previously cited by the ﬁrm. Speciﬁcally, I ﬁrst deﬁne ﬁrm i’s existing knowledge in year t as all patents that are owned by i or that were cited by ﬁrm i’s patents ﬁled up to t; other patents are considered new knowledge to the ﬁrm. New Cite Ratio of a patent is calculated as the ratio between citations 89 made to new knowledge and the total number of citations made by the patent. Based on this measure, a patent is ﬂagged as Explorative if at least 80% of its citations are based on new knowledge (New Cite Ratioě 80%). I transform these patent-level measures to ﬁrm-year level by averaging across all patents produced by ﬁrm i in year t.1 Higher New Cite Ratio and Explorativeness suggest an innovation scheme focusing on exploring new ideas using new knowledge.
To construct a proper control group for CVC parents, I use a propensity score matching method and match each CVC parent ﬁrm that launches its CVC unit in year t with two non-CVC ﬁrms from the same year t and 2-digit SIC industry that have the closest propensity score estimated using ﬁrm size (the logarithm of total assets), market-to-book ratio, ∆Innovation, and patent stock,2 similar to the sample construction strategy in Bena and Li (2014). The CVC launching year for a CVC parent ﬁrm is also the “pseudo-CVC” year for its matched ﬁrms, and I include ﬁrm data beginning ﬁve years before the (pseudo-) event year through ﬁve years after the event.
I characterize corporate innovation dynamics around CVC investment under a
standard diﬀerence-in-diﬀerences (DiD) framework:
where dependent variables yi,t are innovation quantity, quality, new cite ratio, and explorativeness. IpCV CP arentqi is a dummy variable indicating whether ﬁrm i is a CVC parent or a matched control ﬁrm. IpP ostqi,t indicates whether the ﬁrm-year observation is within the rt ` 1, t ` 5s window after (pseudo-) CVC initiations. The 1 This measure is motivated by theoretical work on motivating innovation (e.g., Manso (2011)), and recently implemented in empirical studies (Almeida, Hsu, and Li, 2013; Cust´dio, Ferreira, and o Matos, 2013; Brav, Jiang, Ma, and Tian, 2016).
2 Patent stock is constructed as the total number of patents applied for by the ﬁrm up to year t ´ 1.
90 model includes industry-by-year ﬁxed-eﬀects αindustryˆt to absorb industry-speciﬁc technological trends.3 The coeﬃcient of interest β measures the incremental changes in innovation benchmarked by those of the matched ﬁrms.
Table E.7 reports the results. Columns (1) to (4) study the dynamics of patent quantity and quality. The β-coeﬃcients associated with the diﬀerence-in-diﬀerences term are positive and signiﬁcant across all columns, meaning that CVC parent ﬁrms’ innovation performance improves following CVC investment. The coeﬃcients should be interpreted in semi-elasticity terms. Following CVC investment, parent ﬁrms’ innovation quantity increase is 23.9% larger than the matched ﬁrms (column (1)), and these new innovations collect on average 21.7% more lifetime citations (column (3)) compared to the level before CVC investment.
Columns (5) and (6) study the ratio of new knowledge used in innovation. After CVC initiations, ﬁrms conduct innovation that involves more intense use of knowledge that they have not used before—the estimate of 0.097 in column (5) can be interpreted as a 9.7% increase in using new information (that is, one out of ten citations). Similarly, in columns (7) and (8), the proportion of explorative patents that are mainly (ě 80%) produced based on new knowledge increases by around 4%.4 3 The result is robust to controlling for ﬁrm ﬁxed eﬀects and year ﬁxed eﬀects.
4 Some might worry that this result merely means that CVC parent ﬁrms start to diversify and thus innovate in areas that they had not explored before. In unreported results, I ﬁnd that the increase in using new information concentrates on technological areas closer to the ﬁrm’s core expertise, which is inconsistent with the “diversiﬁcation” story.
91 Table E.1: Innovation Deterioration and CVC Initiation—The Role of Uncertainty This table documents the causal relation between innovation deterioration and CVC initiations across ﬁrms with heterogeneous informational environment. The analysis
is performed using extended speciﬁcations based on Table 4.2 and Table 4.3:
The panel sample is described in Table 4.1. Observations are categorized into two subgroups by the median of uncertainty level of the ﬁrm’s informational environment, indicated by Iuncertainty, which is measured using the average dispersion of patent quality in a technology class weighted by the technological distribution of the ﬁrm’s portfolio over technology classes. When estimating using 2SLS, I instrument ∆Innovationi,t´1 with Obsolescence, and the interaction term is instrumented by the interaction of Obsolescence with Iuncertainty,it.
Columns (1) and (4) report the ﬁrst-stage regression, which regresses the three-year change in innovation quantity (the natural logarithm of the number of new patents in each ﬁrm-year plus one) and innovation quality (the natural logarithm of average citations per new patent in each ﬁrm-year plus one) on the three-year knowledge obsolescence as deﬁned in (4.2) in the paper. Columns (2) and (5) report the OLS regression results, where IpCV Cqi,t is equal to one if ﬁrm i launches a Corporate Venture Capital unit in year t, and zero otherwise. Columns (3) and (6) report the second-stage regression. Firm-level controls Xi,t´1 include the ROA, size (logarithm of total assets), leverage, and R&D ratio (R&D expenditures scaled by total assets).
Industry-by-year dummies are included in the speciﬁcation to absorb industry-speciﬁc time trends in CVC activities and innovation. T-statistics are shown in parentheses and standard errors are clustered by ﬁrm. *, **, *** denote statistical signiﬁcance at the 10%, 5%, and 1% levels, respectively.
94 Table E.3: This table documents the relation between innovation deterioration and the initiation of Corporate Venture Capital on subsamples of ﬁnancially unconstrained ﬁrms. The analysis is performed using the following speciﬁcation: IpCV Cqi,t “ αindustryˆt ` β ˆ ∆Innovationi,t´1 ` γ ˆ Xi,t´1 ` εi,t, The original panel sample is described in Table III in the paper. In columns (1) and (2), the sample is restricted to ﬁrms of which the KZ-index is below industry medians (less ﬁnancially constrained);