Limitations of the Applied Empirical Model

Our thesis focuses on the CVC unit as the level of analysis. Whilst this ensures a suitable level of granularity for our research question, it also results in several limitations that would be interesting to further shed light on.

First of all, our cross-sectional analysis is based on the assumption of structural consistency over time.

In other words, we assumed that no CVC program changed from internal to external or vice versa during the investment period. However, we cannot fully eliminate the possibility of such an occurrence. It would be interesting to devote future research efforts to investigate such cases, as it could offer valuable additional insights on structural influences on performance. Another aspect that potentially influences the results of our empirical analysis is the uneven distribution between internal and external units but also across industries and geographies within our data sample which increases the risk of bias.

Furthermore, our level of analysis denotes that the dataset includes cases of investing corporations with multiple CVC programs. We cannot fully distinguish the total effect between those activities as some of these entries occur in overlapping investment periods. Additionally, patent related data occurs with a time lag, making it further difficult to assign it to particular knowledge creation activities.

Some attention shall be dedicated to our choice of using patent data to measure the innovativeness of investing corporations. While this is a widely accepted measure and frequently used by other scholars (Dushnitsky & Lenox, 2005a; Wadhwa & Kotha, 2006) it is exposed to several limitations. Patents do not seize knowledge creation to its full extent. This is the case because patents generally secure codifiable knowledge, whereas tacit knowledge such as know-how or software (which is typically protected by copyrights) is not represented. Therefore, a lower patent count does not necessarily depict a lower level of innovativeness. Furthermore, the usage of patents critically differs across industries (Chemmanur, Loutskina, & Tian, 2014). We consider a certain industry effect on innovativeness by using control variables for the largest underlying industries in our regression model but cannot be fully certain that the full effect is controlled for. Lastly, using the cumulative patent count as a measure for innovativeness might lead to an additional bias in the sense that it reflects the general innovativeness rather than the quality of such knowledge creation. An alternative would be to use forward citations of the granted patents, as it was done in previous studies (Trajtenberg, 1990; Wadhwa et al, 2016). However, we deemed

Discussion

this variable as inadequate for our study because it is subject to a truncation bias (Dushnitsky & Lenox, 2005b). Our data set only includes forward citations until 2015 and newer patents are likely to receive further forward citations beyond that cutoff date. Additionally, in our model we investigate the patent applications in the post-investment period of firms, which is operationalized as the number of years between the last CVC investment and 2015. This period varies significantly between the different CVC units and can be as short as one year for some entries. Whilst we can account for this timing effect when using the total patent count, this is not possible for forward citations because the longer the period we consider is, the more time there is to accumulate additional forward citations.

We have discussed the primary implications of and reasons for using a cross-sectional study in chapter 5.2 on our research design. However, we deem it necessary to highlight resulting limitations to give a comprehensive overview. Because we intend to investigate the effect of structure on the level of knowledge creation of the investing firm, we analyze the data at a certain point in time, thus avoiding complexity through restricting assumptions on time. In order to do so, we take the average for the financial data over time and aggregate the patent data to one line per CVC unit. While this naturally causes a less precise picture of the data, it allows us to use the sample for the cross-sectional analysis and smooth it for potential outliers over the years. However, using a cross-sectional analysis also forced us to exclude CVC units that were still active in 2015, the last year of our data set. CVC units active in 2015 (and potentially ongoing) therefore had a post-investment period of zero years and subsequently zero patents. These removed entries reflected some large companies that would have been valuable to include in our analysis. Furthermore, the use of a cross-sectional analysis prevents us from capturing the exact year of application of the individual patents as it only considers the cumulative count of patents in the full period. We therefore excluded patents applied for within active investment period from our dependent variable. A full picture would be shown if patents were already included when the parent organization was still active in CVC but considering a certain time lag of innovation (Dushnitsky &

Lenox, 2005a). However, we believe that the applied cumulative count of patents in the period after the last CVC investment resembles a representative picture of the effect of CVC on the rate of knowledge creation.

Our empirical model includes a vast selection of controlling variables in order to account for as many influential outside effects as possible. The control variables have been derived from existing CVC and

innovation literature. However, it is possible that further unobserved factors have potential to significantly impact our results.

Lastly, we conducted several robustness checks on our model to ensure a certain degree of validity.

Although we find robustness for our significant results, other robustness tests could add legitimacy.

Nonetheless, further tests were not deemed essential in the scope of our work, which is why we suggest additional tests with alternative proxies especially for our dependent variable as interesting approaches for future research. For instance, the innovativeness of firms could be operationalized as the R&D stock or forward citations of patents (Dushnitsky & Lenox, 2005a; Hall et al., 2005).

Concluding, this thesis does not claim to be without shortcomings but offers a comprehensive overview and introduction for the research on structural influences on parent firms’ innovativeness. The presented limitations offer several interesting opportunities for future research. We see our considerations and limitations on the classification of internal and external units as the most critical aspect for future research. In order to further investigate a potential structural impact and its orientation on innovation performance, additional studies including both quantitative and qualitative primary data would be necessary for a more granular categorization and thus more meaningful results.

Conclusion