Factors Impacting University–Industry Collaboration in European Countries

: This paper aims to examine the links between university-industry collaboration (UIC) predictors (inputs), and the results of UIC cooperation (outputs). The focus of the research is UIC within the European Union member states and the Western Balkan countries. The analysis was conducted using the partial least squares structural equation modeling (PLS-SEM). This method enabled examining the links between variables that are not directly observable. The authors used data for the period of three years, 2015 – 2018. The results prove that countries investing in UIC predictors (inputs) have better UIC performance (outputs). Based on the statistical analysis, the authors identified the investments in knowledge, networking, and research and development (R&D), in general, as the most significant that impact UIC performance.


Introduction
In recent decades, researchers have been intrigued by the ever-increasing importance of universityindustry collaboration (UIC), in particular by the factors of its success (Hillebrand and Biemans 2003;Parkhe 1993). UIC is emerging as a critical component of the innovation process. It is a vital component of the regional innovation policy-mix for regional policymakers as the successful exploitation of R&D results is fundamental for regional competitiveness (Morisson and Pattinson 2020).
A noticeable shift towards regionalization in innovation policy and technology is evident. One of its consequences is the expectation for universities to draft networks and establish regional ties (Koschatzky and Stahlecker 2010). Also of considerable importance is the construction of broad collaborative cluster networks, although most clusters focus on loosely defined local level networks. Notwithstanding, network members are responsible for more patent applications than others. What is more, there seems to be no loss of patent quality when collaborating with same cluster region universities (Nishimura and Okamuro 2010).
UIC is mostly based on knowledge and technology transfer. Philbin notes that there is much evidence for a strong correlation between technology transfer and practical knowledge on the one side, and successful collaboration on the other (Philbin 2010). An intense transfer can improve the technology novelty (Guan et al. 2005), encourage innovation performance (MingJi and Ping 2014), and/or boost product development (Fernandes and Ferreira 2013). Nevertheless, many barriers impact knowledge and technology transfer, and this is a topic of many pieces of research (see de Medeiros et al. 2012;Hong et al. 2010;Schofield 2013). In their study from 2009, Flores et al. state that technology transfer and knowledge are influenced by strategy and motivation (Flores et al. 2009), where utilization of adequate policies and incentives can enhance transfer activities in UIC (Schofield 2013). Knowledge transfers might be impacted differently by universities and companies. While the former initiate knowledge transfer through research, the latter take on more managerial positions afterward (Goel et al. 2017).
To investigate factors that impact UIC, the authors used data provided by the Global Innovation Index (GII). The Innovation Input Sub-Index and the Innovation Output Sub-Index -each constructed around pillars, are two sub-indices the GII relies on. In the Innovation Input Sub-Index, elements of the national economy that enable innovative activities are captured by five input pillars. The five input pillars are business sophistication, market sophistication, infrastructure, human capital and research, and institutions. The Innovation Output Sub-Index encompasses innovation outputs, which stem from innovative activities within the economy. The output pillars are creative outputs, and knowledge and technology outputs. The average of the Input and Output sub-indices marks the overall GII score. (Cornell University, INSEAD, andWIPO 2010-2020).
Based on the defined elements, the authors created groups of factors that influence UIC (impact factors). On the other hand, the UIC performance level in a specific country is measured by the set of output factors. The differences in approach to the support of UIC also lead to companies achieving different results. This motivated the authors to explore the main factors that impact UIC in European countries. Thus, this article intends to answer the following question: "What are the vital factors that influence UIC"?

Literature Review
Many articles deal with the factors that make UIC successful. Rybnicek and Königsgruber (2019) conducted a thorough, in-depth review of the published scholarly literature on industry-university collaboration. They performed an extensive analysis of UIC projects' research to distil factors that influence such partnerships' success. Given the comprehensiveness of this research, as the authors of this paper used it as a base for identifying factors that determine UIC. Many other authors also deal with this topic, and the most recent and relevant articles are presented in this part of the research.
Based on the review of the existing literature, the authors could organize the factors that determine UIC. They are marked as "input factors" in the article and deployed into four categories. Firstly, there are institutional factors, which refer to business environment and government effectiveness. Secondly, there are human factors, which refer to human capital and research. Thirdly, there are linkage factors, which refer to relationships between universities and companies. Fourthly, there are framework factors, which refer to the business infrastructure. Finally, the authors define "output factors" that presents the level of UIC in a specific country.

Institutional Factors
Institutional factors include the business environment, legal restrictions, and/or governmental support. Kozlinska defines the government as an influential power with the ability to either facilitate or harm collaboration (Kozlinska 2012). On the one hand, the governmental network (Rampersad 2015), public funding (e.g., Piva and Rossi-Lamastra 2013;Flores et al. 2009), or tax incentives (Bodas Freitas et al. 2013) can facilitate UIC. On the other hand, the absence of regional bracing structures (Şerbănică 2011), and/or regulations and legal restrictions (Hadjimanolis 2006;Attia 2015;Arvanitis et al. 2008) can have a negative impact on collaboration. As a rule, industryuniversity partnerships rely heavily on governmental support (e.g., Newberg and Dunn 2002;Myoken 2013;Collier et al. 2011;Sohal 2013;de Medeiros et al. 2012;Muscio and Vallanti 2014;Schofield 2013;Hemmert et al. 2014). Additional factors of business environmental success correspond to the market potential of research results (Barnes et al. 2002;Hadjimanolis 2006;Guan et al. 2005;Ankrah and AL-Tabbaa 2015) or market uncertainties (Hemmert et al. 2014). environmental factors. Nevertheless, with an overall improvement in R&D capability, R&D collaboration, and technology, commercialization will also be enhanced (Vea 2014).
The results demonstrate that increased attention to technology R&D and the protection of patents by governments and companies has led to immense improvements in the ICT sector. What is more, companies have now become the main body of technological innovation (Xia et al. 2014). Yarmouk University has developed dynamic programs of ICT enrichment and adopted an innovative partnership model, which in the form of an on-campus facility seeks to bridge the academia-industry gap. The facility offers a productive, internal collaborative environment for technical and business faculties, which together pursue projects to cultivate collaboration with industrial and business partners. Moreover, it facilitates the alignment of skills and knowledge of university staff and students to contemporary real industry needs, and updates the university's knowledge base with the latest industry developments (Al-Agtash and Al-Fahoum 2008).

Output Factors
In the literature, there are many performance indicators of UIC. The number of patents (Xia et al. 2014), scientific and technical articles (Salimi and Rezaei 2016), high-tech manufactures and exports (Aiello et al. 2019), intellectual property receipts (Valentin and Jensen 2007), ICT services exports (D'Costa 2006; Hwang 2020), cultural and creative services exports (Draghici et al. 2016), creative goods exports (Banal-Estañol et al. 2011), are frequently mentioned. In general, the output indicators can be divided into two groups: tangible and intangible.
As Rybnicek and Königsgruber (2019) stated, the compatibility of partner-to-partner goals is among the most discussed subjects in UICI literature. Failure to achieve the desired outcome is often the result of incompatibility (Henderson et al. 2006). For example, companies strive towards withholding the groundbreaking findings from universities from competitors, while the universities desire to publish them (Newberg and Dunn 2002). Lai and Lu provide similar results (2016). They state that companies and universities strive towards different goals. Consequently, it is paramount to seek a win-win situation with balanced benefits for both partners. This can be achieved only if both partners understand the other's interests. Also essential is for partners to agree upon achievable goals through a shared understanding of the objectives, and to materialize a precise strategy throughout the collaboration (Hong et al. 2010). It should also be added that partners more often than not hold unrealistic expectations about the outcome of cooperation, and/or have a different sense of urgency (Attia 2015).
For the purpose of finding the appropriate partner, a correct partner selection process is advised prior to collaboration. In this sense, confidence in one's own needs and requirements is also a prerequisite. Only then can the search for an adequate partner with concordant interests and goals begin (Arvanitis et al. 2008). Adequate search strategies can facilitate the search for a matching partner. Barnes et al. (2002) recommend a partner evaluation method with specific criteria.

Construct Measures
The authors have grounded the theoretical framework of this article in a systematic review of relevant scholarly literature while also considering the availability of the variables' data. At the outset, we elicited 36 measures. These were grouped into six constructs, four of which focus on input factors, the remaining two on output factors. We depicted the research model in Figure 1.  Figure 1 depicts the model' schematic diagram. Both the internal and external relationships were considered. 1 The estimation was conducted with the help of SmartPLS software.

Research Method
This research uses the structural equation modeling method (SEM). SEM encompasses a plethora of statistical methodologies by which a causal relationship's network can be approximated. A theoretical model defines such a network as one which links at least two latent complex concepts. Several observable indicators measure each of these concepts. In essence, we can study the complexity within a system by considering a causality network among latent concepts -"latent variables". Many observed indicators usually defined as "manifest variables" are used to measure each of these latent variables. According to this, structural equation models represent a joint-point between path analysis and confirmatory factor analysis (Esposito Vinzi et al. 2010). Among the methods of estimating SEM models, the covariance-based (CB) method 2 , invented by K. G. Jöreskog, enjoyed the greatest popularity for a long time. So universal was its recognition that in social sciences the phrases: SEM and covariance-based structural equation modeling (CB-SEM) were synonymous for many years (Chin et al. 1996). Meanwhile, H. Wold developed an alternative approachthe partial least square method (PLS). Its description and application for estimating models with latent variables were presented by Wold in, among others: (Wold 1979(Wold , 1980b(Wold , 1980a. Because the PLS method was an alternative to K. G. Jöreskog's 'hard' modelling, i.e. one based on strong assumptions regarding the normality of distributions and requiring large samples, Wold referred to his PLS approach as 'soft' modelling (Wold 1980b(Wold , 1982. After a time, the term 'PLS-path modelling'. 3 Came into use, and thenin order to emphasize that PLS was an alternative to CB, it began to be called 'PLS Structural Equation Modeling' (PLS-SEM).
PLS-SEM and CB-SEM were developed as distinct, though complementary, methods with specific purposes and requirements. This was clearly stressed by the authors of both approaches at the beginning of the 1980s (Jöreskog and Wold 1982). At present, the varying properties of PLS-SEM and CB-SEM are also noticed, with emphasis on the complementarity of the two methods instead of the competition between them. The advantages of the non-parameter, variance-based PLS-SEM modelling are, at the same time, the disadvantages of the parameter, covariance-based CB-SEMand the other way around. Therefore, the choice of method should depend on the empirical context and research purposes (Hair et al. 2019).
An SEM model consists of two sub-models: a structural one and a measurement one. In PLS-SEM terminology, the phrases' inner model' and 'outer model', respectively, are also used. A structural model describes the relationships among latent variables, whereas a measurement model the relationships among the latent variables and the indicators by which they are identified, also known as manifested variables (Wold 1980a).
When constructing a structural model, one must pay particular attention to two aspects: the nature of the analyzed latent variables and the associations which occur among them. It is important to distinguish between exogenous variables and endogenous ones. Furthermore, all the formulated elements of the conceptual framework should be derived from theory and logic. If a theoretical basis is lacking, or if the theory is inconsistent, one should rely on one's own judgment, experience, and intuition (Hair et al. 2017).
Specification of the measurement model is an equally important stage of the modeling process. Verification of the hypotheses reflected in the structural model's equations can be reliable when, and only when, the latent variables are correctly defined by means of indicators. And the choice of indicators is as crucial as the choice of the way in which they are defined (Hair et al. 2017). Definition of latent variables by means of indicators can be done either deductively or inductively (Rogowski 1990). Under the former approach, indicators reflect the defined latent variable and are then referred to as reflective indicators, while the measurement model is called a reflective measurement model. In the case of inductive definition, it is assumed that indicators make up the latent variables, hence the expressions formative indicators and formative measurement model. The type of definition (inductive or deductive) should follow from the assumed theoretical description (Rogowski 1990). Also, the choice of observable indicators should be preceded by an in-depth and thorough literature review, including the theory and empirical studies in measuring the latent variables present in the model.
Alongside examining latent variables' correlations, PLS-SEM modeling also helps approximate the values of said variables (weighted sums of indicators). For that reason, a synthetic measurement, with which we can obtain a linear ordering of the studied objects, is calculated for each of the model's latent variables.
Estimation of a PLS-SEM model is performed using the PLS method. The algorithm simultaneously estimates inner model parameterspath coefficientsand outer model parameters outer weights and outer loadings. The procedure also yields estimations of the values of all the latent variables included in the model. The estimation aims to maximize the explained variance of the latent dependent variables. The first stage involves the iterative estimation of measurement model weights and the values of latent variables. At the second stage, the loadings and path coefficients of the structural model are estimated. A detailed description of the PLS algorithm can be found, e.g. in (Henseler et al. 2012;Wold 1982), and its generalization in (Rogowski 1990).
Verification of a PLS-SEM model is a two-stage process. First, the structural model is assessed. Second, if the validity of the structural model has been confirmed, the structural model is tested. Table 2 lists the properties of the model, which should undergo evaluation. Source: authors' work based on Hair et al. (2017).
SEM using the PLS procedure used to be difficult due to the unavailability of software. Now the situation has greatly improved thanks to the wide range of user-friendly programs, which enable estimation and statistical verification of PLS-SEM models, e.g. WarpPLS (Kock 2020), ADANCO (Henseler and Dijkstra 2015), SmartPLS (Ringle et al. 2015). This study will use the SmartPLS software.

Specification of the Model
The model used for the realization of the research objective, i.e., proving the influence of UIC predictors on UIC performance, contains Eqs. 1a and 1b. where: OTItoutput tangible indicators in period t, OIItoutput intangible indicators in period t, IIFtinstitutional factors in period t IHFthuman factors in period t ILFtlinkage factors in period t IFFtframework factors in period t α0, α1structural parameters of the model, νtrandom component, t -Period of three years 2015-2018. 4 The authors used the deductive approach to defining latent variables in the model, i.e., each latent variable as a theoretical notion is a starting point for the search for empirical data. The indicators were selected based on substantive and statistical criteria. The following things were accounted for from the statistical perspective: indicator values' diversity, measured by the coefficient of variation 5 (the coefficient's critical value was calculated at 10%), and the quality of the estimated model (model evaluation measuresex-post analysis). Table 3 presents the indicators that passed substantive and statistical verification. The indicators of the input latent variables point to the most frequent and significant predictors of the UIC. Meanwhile, the OTI and OII measures reflect the outputs of UIC.  Figure 2 shows the PLS-SEM estimation results obtained in the SmartPLS software (Ringle et al. 2015). The results are interpreted in section "Results and Discussion".  Table 4 contains the results of the estimation of the outer sub model. Individual indicator reliability values significantly larger than the lowest acceptable level of 0.4 can be observed (Hulland 1999).

Estimation Results and Statistical Verification of the Model
Cronbach's alpha and composite reliability are shown to be larger than 0.6, which means that high levels of internal consistency reliability have been demonstrated among the latent variables.
In order to confirm convergent validity, the average variance extracted (AVE) is evaluated for each latent variable. Table 4 also shows that the AVE values are greater than the acceptable threshold of 0.5, which confirms convergent validity.
As suggested by Fornell and Larcker (1981), if this value exceeds other latent variables' correlation values, we can use the square root of AVE in each latent variable to establish discriminant validity. The latent variable OII's AVE is found to be 0.944 (from Table 4) hence its square root becomes 0.920 (Table 5). The result indicates that discriminant validity is well established.
Using a two-tailed t-test with a significance level of 5%, the path coefficient will be significant if the T-statistics is larger than 1.96. As presented in Table 6 all path coefficients in the inner model are statistically significant. The values of the Stone-Geisser test statistic, which verifies the model in terms of its predictive usefulness (see Table 7), are positive, which proves the model's high predictive quality. Both the measurement models and the structural models were positively assessed, therefore, in the next stage of modeling, the results can be interpreted. The following interpretation of the πij outer loading is assumed: -|πij | < 0.2-no correlation, -0.2 ≤ |πij | < 0.4-weak correlation, -0.4 ≤ |πij | < 0.7-moderate correlation, -0.7 ≤ |πij | < 0.9-strong correlation, -|πi j| ≥ 0.9 -very strong correlation.

Fig. 3a. Estimations of factor loadings of IIF latent variable
The latent variable IIF is very strongly reflected by two indicators: "Gross expenditure on R&D" (IIF5) and "Political and operational stability" (IIF1).

Fig. 3b. Estimations of factor loadings of IHF latent variable
The latent variable IHF is very strongly reflected by one indicator-"Researchers in R&D (per million people)" (IHF4)-and strongly reflected by three indicators: "Employment in knowledgeintensive services" (IHF5); "PISA scales in reading, math, and science" (IHF1); and "Research talent in business enterprise" (IHF6).

Fig. 3c. Estimations of factor loadings of ILF latent variable
Two indicators very strongly reflect the latent variable ILF: "GERD performed by business enterprise" (ILF1), and "GERD financed by business enterprise" (ILF2).

Fig. 3d. Estimations of factor loadings of IFF latent variable
The latent variable IFF is very strongly reflected by two indicators: "Logistics performance" (IFF4) and "ICT access" (IFF1). Also, the latent variable ILF is strongly reflected by one indicator, "Environmental performance" (IFF6).

Fig. 3e. Estimations of factor loadings of OII latent variable
The latent variable OII is very strongly reflected by two indicators: "PCT international applications by origin" (OII2) and "Patent applications by origin" (OII1). Moreover, the latent variable OII is strongly reflected by one indicator-"ICTs and business model creation" (OII6). The estimation of the internal model parameters indicates a significant positive correlation between UIC predictors and the level of UIC performance in the studied group of 33 European countries in the period 2015-2018. This means that those countries that reported a higher level of development of input factors also had a better UIC performance in the observed period. OIIt = 0.488*ILF2015-2018 + 3.76956 (2) Aside from investigating latent variables correlations, this way of modelling also helps estimate these variables' values (weighted sums of indicators). For that reason, a synthetic measurement, with which we can obtain a linear ordering of the studied objects, is calculated for each of the model's latent variables.
6 Parameter α0 was estimated in the PLS program (Rogowski 1990 Based on estimated values of the input and output variables, rankings of the studied countries have been compiled: a ranking of input and output variables. The results are shown in Table 8. The countries are also divided into typological groups, according to similar volumes of UIC inputs and outputs. The results of the grouping are presented in Figs. 4 a-e. The boundaries between the groups have been established based on the arithmetic means and standard deviations of the synthetic measure zi (equal to 0 and 1, respectively, for each of the latent variables). The groups are as follows: -Group I (very high level of latent variable): zi ≥ 1, -Group II (high level of latent variable): 0 < zi ≤ 1, -Group III (medium and low level of latent variable): −1 < zi ≤ 0, -Group IV (very low level of latent variable): zi ≤ −1.  When one considers the human factors, the classification looks somewhat different. Here, the top ranks are occupied by EU member states based on highly developed human capital, focusing on functional literacy, knowledge, and the R&D sector (Denmark, Finland, the Netherlands, Germany, Belgium, Luxemburg, and Austria). The group with a high level of human capital development comprised 11 countries: France, Ireland, United Kingdom, Sweden, Slovenia, Estonia, Spain, Czech Republic, Hungary, Malta, and Poland. Eight countries were qualified for the group of economies with a medium level of human capital development: Lithuania, Italy, Slovakia, Latvia, Portugal, Croatia, Bulgaria, and Greece.
Countries with weak indicators of functional literacy and of knowledge, and a small percentage of spending on R&D sectors (Cyprus, Serbia, Montenegro, North Macedonia, Romania, Bosnia and Herzegovina, Albania) are ranked at the bottom.
According to institutional factors, the ranking of countries demonstrates a dominance of North and Western European economies and those "catch-up" economies from Central Europe. Economies of Southeastern Europe occupy the lower ranks.  According to framework factors, the countries' ranking demonstrates a dominance of North and Western European economies and those "catch-up" economies from Central Europe. Economies of Southeastern Europe occupy the lower ranks.

Fig. 4e. Division of EU countries into typological groups according to intangible output indicators
As presented in the analysis of intangible output indicators, the top ranks are occupied by five EU member states based on highly developed human capital, focusing on functional literacy, innovation, ICT use, and the R&D sector (Finland, Germany, Luxemburg, the Netherlands and Denmark.). The group with a high level of intangible output indicators comprised nine countries: Sweden, France, Austria, United Kingdom, Malta, Spain, Belgium, Slovenia, and Ireland. Fifteen countries qualified for the group of economies with medium and low levels of intangible output indicators: Estonia, Italy, Portugal, Poland, Czech Republic, North Macedonia, Lithuania, Hungary, Slovakia, Latvia, Cyprus, Bulgaria, Serbia, Croatia, and Montenegro. Countries with weak indicators of patent applications by origin, PCT international applications by origin, and ICTs and business model creation (Romania, Greece, Albania and Bosnia and Herzegovina) were ranked at the bottom.
Within the EU-28, most countries treat UIC as a vital source of increasing the performance of companies. Nevertheless, the Balkan countries still lack cooperation between the universities and companies.

Conclusions
This article presents empirical studies' results of the relationship between UIC predictors and performances in selected European countries. The research involved developing a PLS-SEM model, measurement of the latent variables based on sets of observable variables, and the estimation and verification of the PLS-SME model. The modeling outcomes reveal a significant favorable influence of impact factors on the UIC performance in the analyzed European countries. Based on the literature review, the authors identified four main groups of factors that impact UIC performance. Firstly, there are institutional factors, which refer to the business environment and expenditures on R&D; secondly, there are human factors, which refer to functional literacy and research; thirdly, there is a linkage factor, which refers to relationships between universities and firms; and fourthly, there are framework factors, which refer to the business infrastructure. Finally, the authors defined the "output factors" that present the level of UIC in a specific country.
On average, predictors of successful UIC are led by EU-15 member states. A very high-level input indicator was observed in the following seven countries: Denmark, Finland, Germany, the Netherlands, Austria, Luxemburg, and the United Kingdom. The group with a high level of developed input factors comprised nine countries: Belgium, Ireland, France, Sweden, Spain, Estonia, Slovenia, Czech Republic, and Malta. Eleven countries qualified for the group of economies with a medium level of developed input factors: Italy, Poland, Lithuania, Slovakia, Hungary, Portugal, Latvia, Cyprus, Croatia, Bulgaria, and Greece. Six Balkan countries are ranked bottom in Europe, and these are countries with a low level of institutional factors development: North Macedonia, Serbia, Romania, Montenegro, Albania, and Bosnia and Herzegovina.
To improve the economic situation in the region, the Balkan countries should, in the long term, revise their education systems and invest more money in knowledge, networking, and R&D in general. Based on both public and private initiatives, this investment will have a long-term positive impact on companies' productivity, as well as their profit. Furthermore, they should follow successful examples from the EU member states and launch specialized programs for supporting innovation in the short term. The authors suggest Austria and Slovenia as the best model countries for the Balkan region since these two countries have been the best example for the Balkan economies in many fields throughout history. Moreover, universities must be more focused on the companies' real needs in the future if they want to justify their role in society.
This article provides new knowledge on how different factors accelerate UIC in Europe. To the