Project Topics
Get Complete Project Material File(s) Now! »
Econometric approach
Our empirical strategy is to consider university supplier status as the “treatment”.12 According to the conceptual framework outlined in Section 2, this treatment should affect different aspects of the firms’ innovative behaviour, namely their product and process innovations. We proceed in two complementary steps: first, we estimate the effect of “being a university supplier” on the set of innovation variables using standard regression techniques; and, second, we adopt a quasi-experimental framework and employ propensity score matching (PSM) to obtain the impact of the treatment.
12 Note that some firms included in the overall sample might also be considered “treated” if, for example, they supply other universities. Unfortunately, we are not able to test this hypothesis but, as discussed in Section 5, our estimates would nevertheless reflect the lower bound of the treatment effect. In other words, we are confident that if a bias exists, it would not run counter to our conjectures.
where the dependent variable Inno represents the three innovation proxies considered (New Mkt; New Mkt Volume; Iproc), the main regressor Supplier is a binary variable taking a value of 1 if firm i is a university supplier, X is a vector of firm-level controls (as described in Section 3.3), and is an idiosyncratic error term. The model is estimated using Ordinary Least Squares (OLS) and the coefficient of interest in Equation (1) is α, representing the effect of the treatment on the innovative performance of firms.
Sources of selection bias
Although easily interpretable, the above econometric approach embeds an important assumption: namely, that the data come from randomised trials – i.e., the assignment of the treatment to firms (that is, being a supplier of the university or not) is completely random. However, here, we are dealing with non-randomized observational data as the university chooses its own suppliers and the latter are likely to differ substantially from other firms in many respects (see Table 6). This absence of randomly assigned treatment to firms introduces a bias in the regression estimates.
Indeed, there are two primary sources of bias. First, a university typically plays a “picking the best” strategy. As the university organises public procurement bids to choose its suppliers, it is reasonable to assume that it will pick “good companies”, essentially those characterized by the soundness of their financial conditions and a high degree of innovativeness. Second, it is also possible that firms self-select themselves to become suppliers. For instance, some companies may have better search capabilities, or other types of competitive advantage, that allow them to detect, and thus strategically apply for, a public procurement competition. In short, university suppliers are likely to be intrinsically different from non-suppliers even in the absence of the treatment, and we need to account for this possibility.
Propensity score matching
The goal is to estimate the expected value of the average treatment effect on the treated (ATT), defined as the difference between the expected outcome values with and without treatment for those who actually participated in the treatment. Formally:
where [ (1)| = 1] is the expected value of the outcome variable of the treated units and [ (0)| = 1] is the expected value of this variable when the units are not treated. As the counterfactual mean for the units treated is not observed, we have to choose a substitute for this value in order to estimate the ATT. We apply propensity score matching (PSM) to construct the pseudo-counterfactual or the control. Matching estimators are based on a comparison of the outcomes obtained by the treated units (i.e., university suppliers) and those obtained by a “comparable” control group (i.e., a subsample of other companies), conditional on a set of defined characteristics. Under certain assumptions, the difference in mean outcomes between the two groups can be attributed exclusively to the treatment.
The matching procedure requires the definition of a set of characteristics X, which leaves the estimate prone to the well-known “curse of dimensionality”. In short, this problem requires the estimation of a high-dimensional vector of exogenous covariates to find an exact twin for each treated unit. Rosenbaum and Rubin (1983) suggest it is possible to compress this vector into a single scalar index – that is, the propensity score – and to use this index to search for similar (in statistical terms) units. In our framework, the propensity score measures the probability of a firm becoming a supplier of scientific materials and equipment to UNISTRA based on a set of observable characteristics.
PSM requires three important methodological choices: i) the model to be estimated; ii) the variables to be included in the model; and iii) the matching algorithm to be applied. In the case of the first choice, because our treatment is a binary variable, we estimate a probit regression.
13 Two identifying conditions must be fulfilled: namely, unconfoundedness and common support. Unconfoundedness, or the conditional independence assumption, states that the outcome should be statistically independent of the treatment. For this condition to hold, all the variables likely to affect simultaneously the probability of receiving the treatment and the potential outcomes should be known and taken into consideration. The common support condition states that the control group should contain at least one sufficiently similar observation for each treated unit.
Caliendo and Kopeinig (2008) show that, in the case of binary treatments, probit and logit regressions generate very similar results. As regards the choice of variables, we exploit the entire set described in Section 3.3 to determine the probability of firms receiving the treatment. This choice was dictated by existing empirical evidence but, above all, by the idea of mimicking the practices adopted by UNISTRA’s public procurement office when selecting suppliers. As pointed out above, the university’s suppliers are selected via public procurement bids, a procedure that has a dual objective – to uphold competition and transparency during the selection process and to guarantee the effective spending of public money. Hence, we conducted three semi-structured interviews with the university’s public procurement managers to understand the implementation of the selection procedure, its various stages, selection criteria, and the role played by researchers in the process. The managers confirmed the appropriateness of our set of variables.14 Finally, regarding the choice of the matching algorithm, we opt for the bias-corrected nearest-neighbour
(NN) matching estimator proposed by Abadie and Imbens (2006). Given the large sample and the similar distribution of propensity scores between treated and control units, we apply a NN search without replacement and with oversampling – i.e., we match each treated unit with three untreated observations. As the results may be sensitive to these implementation choices (Caliendo and Kopeinig, 2008), in Chapter 2 we perform a series of robustness analyses implementing alternative specifications.
Thus, we proceed as follows. First, we obtain the propensity scores associated with the binary treatment via the estimation of the probit model (or selection equation) containing the original set of variables. Next, we apply the NN algorithm and use the estimated propensity scores to match the subsample of suppliers with the most similar group of firms in the sample. Finally, we compute the ATT to draw conclusions about the effect of university demand on the innovativeness of its suppliers.
14 The interviews took place in September 2017 at UNISTRA’s public procurement office, and lasted about one hour each. Specifically, the managers ranked the firms’ financial status and their receiving public support for innovation (acting as “reputation effect”) as being among the most important criteria for assessing candidates. Other important criteria were identified as the firms’ fiscal status and whether they respect the codes of ethics governing labour law.
As it is the applicants themselves that provide the information related to these last two criteria, it proves quite challenging to include a reliable proxy for them in our estimation.
Results
Regression analysis
Table 7 presents the results of the regression analysis. Two major findings merit discussion. First, the coefficients of the regressor Supplier associated with the two product innovation proxies (New Mkt and New Mkt Volume) are positive and statistically significant at the 1% level. These estimates imply that university suppliers exhibit a higher propensity to introduce new-to-the-market products and to enjoy higher sales from these products. Taken together, these results support our conjecture that university demand for goods and services affects the innovative performance of suppliers, and that this effect is positive in the case of product innovations. Indeed, as discussed in Section 2, university demand seems to act in two complementary ways: on the one hand, because of their quite specific needs, scientists can contribute to the emergence of new concepts and ideas, reducing the uncertainty and risk of failure that is inherent to the innovation process; while, on the other, scientists can act as lead-users of technologies, thus indirectly bearing the costs of learning and refining associated with their development.
Second, we observe that the regressor Supplier does not have any relevant effect on process innovations. In our conceptual framework, we argued that university demand might provide firms with a minimal market size, hence, providing incentives to improve production practices and achieve scale economies. A tentative explanation for the lack of effect found might be that scientists’ needs are highly specific and, as such, represent needs that are not yet common in the marketplace15. This is especially true of high-tech instrumentation that only serves very specific research aims. While idiosyncratic demand grants firms competitive advantages in specific solution types and fosters the development of new products, it may prevent the exploitation of economies of scale, at least in the short-term.
15 We would like to thank a lot to anonymous referee for his stimulating and intriguing remark on the subject. The hypothesis underlying our research is that tailor-made instruments can also be useful for other users beyond the scientific community. Although these items might not trigger particular volume effect for the firms involved, they might still have learning by doing effect on firms’ manufacturing capabilities, which is related to economies of scale The coefficients of the control variables conform, by and large, with those reported in previous studies (i.e., Laursen and Salter, 2006; Cohen, 2010; Beck et al., 2016; Scandura, 2016). The intensity of R&D investments (R&D) and public support for innovation (PubFund) positively affect the propensity to achieve product innovations. The breadth of openness of firms’ innovative cooperation (Breadth) also appears to be an important factor in explaining innovative performance. More efficient companies (LabProd_log) tend to innovate more in terms of new products, though not in terms of new processes. The lack of significance of the role of universities and PROs as collaboration partners in innovation (UniColl) is surprising yet consistent with the fact that the breadth measure could absorb this effect. Finally, it seems that large (Empl_log) but young (Age_log) companies belonging to industrial groups (Group) enjoy higher sales from their product innovations, whilst these demographic features do not present robust patterns across the other specifications.
Propensity score matching
We now turn to examine the results of the matching estimates. We first discuss the process of selection and the reliability of the control group. Next, we present the chapter’s main findings.
The results of the estimation are reported in Table 8 (left panel). The estimated coefficients represent the influence that each variable has on the probability of a firm becoming a university supplier. Note that the percentage of correctly predicted zeroes and ones implies a satisfactory goodness of fit. It emerges that larger firms have a higher probability of becoming suppliers of scientific material and equipment to UNISTRA. Moreover, we find that firms benefiting from public support for innovation and with a higher labour productivity index are also more likely to be selected as university suppliers. Firm profitability affects positively and significantly the probability of receiving the treatment, although at a low significance level. The other variables do not play any relevant role. Overall, these estimates suggest that a selection process is actually in place and that financial conditions and reputation are the most relevant factors.
Before discussing the final results, in Table 8 (right panel), we report a t-test for equality of means between treated and untreated units before and after the matching. Pre-matching comparisons (unmatched) show that the two groups present statistically significant differences in almost all the variables considered. If the matching procedure is effective, the sample of untreated firms should not differ in statistical terms from the sample of treated firms in any dimension. We find equality of means in the treated and control groups post-matching (matched), indicating that the matching procedure has generated a reliable counterfactual.
Finally, Table 9 shows the results of the propensity score matching. The first column reports the mean value of the outcome variables for the suppliers, the second column the mean values for the control group, while the third column represents the main parameter of interest, namely the ATT.
In line with the results presented in Table 7 above, we confirm the positive and statistically significant effect of the treatment in the case of product innovation. University suppliers are more innovative compared to other firms insofar as they show a higher propensity to introduce “radical” product innovations and to reap greater revenues from the sales of these products. It is worth stressing the actual magnitude of these effects. First, 90% of suppliers achieve product innovations compared to about 75% of firms operating in the same industries and with similar characteristics.
Second, the sales of suppliers’ market novelties are about 1.5 times higher than those of other firms. Again, we do not find any significant effect on process innovations.
Table of contents :
Chapter 1. University procurement of scientific equipment and corporate innovation: a literature review
1.1 Introduction
1.2 Some historical roots
1.3 Demand-side studies and innovative public procurement
1.4 Nature and varieties of scientific knowledge
1.5 Conclusion
Chapter 2. Demand-pull innovation in science: empirical evidence from a research university’s suppliers
2.1 Introduction
2.2 Conceptual framework
2.3 Data and measurement
2.3.1 The context
2.3.2 Data sources
2.3.2.1 University expenditures data
2.3.2.2 CIS and FARE datasets
2.3.2.3 The final dataset
2.3.3 Measures
2.3.3.1 Dependent variable
2.3.3.2 Other variables
2.3.3.3 Descriptive statistics
2.4 Econometric approach
2.4.1 Regression analysis
2.4.2 Sources of selection bias
2.4.3 Propensity score matching
2.5 Results
2.5.1 Regression analysis
2.5.2 Propensity score matching
2.6 Conclusions, policy implications and future extensions
Chapter 3. Alternative approaches for the analysis of universities’ impact on suppliers’ performance
3.1 Introduction
3.2 Sensitivity analysis
3.2.1 Controlling for University-Industry R&D collaborations
3.2.2 Applying a parsimonious selection equation
3.2.3 Restricting the treatment to scientific instruments
3.2.4 Restricting the treatment to regional suppliers
3.2.5 Applying different numbers of neighbours (NN) and matching algorithms
3.3 Sensitivity analysis results
3.4 CIS data limitations and suppliers’ patent analysis
3.4.1 The two methods of collecting data on industrial innovation before innovation surveys
3.4.2 A brief note on the history and origins of the CIS
3.4.3 CIS data harmonization across countries and waves
3.4.4 CIS data access
3.4.5 Cross-section data
3.5 Using alternative data sets
3.6 Empirical methodology and results from using patent data
3.7 Conclusions
Chapter 4. Opening the black box of university-suppliers’ co-invention: some field-study evidence
4.1 Introduction
4.2 Some theoretical landmarks
4.2.1 Formal and informal mechanisms of communication
4.2.2 Researcher-Supplier interactions as focal point in public procurement procedures
4.3 Field-study methodology
4.3.1 Case study as a tool of social sciences
4.3.2 Definition of the cases and sampling selection strategy
4.3.3 Data selection
4.3.4 Rival theories
4.4 Brief overview of the three cases
4.5 Presentation of the cases
4.5.1 Case 1. The development of an NMR measurement accessory
4.5.2 Case 2. The development of the Fluorescence Macroscope, the Confocal Scanner and The Correlative Light Electron Microscope 2.0
4.5.3 Case 3. The development of the hybrid device (LG-GC) for protein analysis in their native state
4.6 Putting the threads together: technology co-invention versus technology transfer
4.6.1 Universities as catalysts of suppliers’ technological capabilities
4.6.2 Embeddedness and mutual learning
4.6.3 Researchers-Suppliers’ interactions and public procurement procedures
4.6.4 From conduits of influence to innovation benefits
4.7 Conclusions
Chapter 5. Concluding remarks
