National innovation and knowledge performance: the role of higher education teaching and training

This paper investigates the role of the higher education system (HES) in the production of national innovation. We focus on the issue of institutional diversity of HESs and its impact on national innovation systems. We identify four key HES characteristics and hypothesize their influence on the production of national innovation. The empirical evidence presented in this paper suggests several conclusions which have relevance for policy makers. In particular, there should be focus upon increasing access to and investment in higher education and lowering subsidies. The latter is found to have an adverse effect on patent production.


Introduction
The role of the higher education system (HES) as the main provider of human capital through teaching and training is well known. This 'first mission' has traditionally been seen as a function of the HES which is recognized as one of the pillars of any system of innovation (Fagerberg and Shrolec 2008). The innovation literature suggests that higher education institutions (HEIs) play a significant role within the National Innovation System (NIS) through two main channels. The first channel is the interdependency between the NIS and the research activities of HEIs (known as the 'second mission'). The second channel is the exploitation of HEI teaching and research activities (known as the 'third mission') (Etzkowitz and Leydesdorff 2000;Gulbrandsen and Slipersaeter 2007). While the second and third missions' role in the NIS is relatively well explored, the higher education (HE) policy issues related to the development of human capital through education (i.e. the first mission) and their relationship to the development of NIS remain under-explored, especially across developing and emerging economies.
There is a general consensus within the HE policy literature on the heterogeneity of HESs across countries (Ansell 2008). Policy choices relate to, for example: the extent of government control of HEIs (Etzkowitz 1998), the adoption of elitist versus open HESs (Ansell 2008), and the level of subsidies and public funding to HEIs. Given the institutional variety of HESs around the world, this paper intends to explore the relationships between the key institutional characteristics of a country's HES and its performance in innovation and knowledge production. The empirical evidence of this paper contributes to the policy debate on the appropriate HE model relating to national innovation and knowledge performance.
The remainder of this paper comprises six sections. Section 2 reviews the literature on the multiple roles of HEIs in national economic and innovation development. We also review the literature on national HES characteristics and propose four hypotheses on how these characteristics may be associated with the country's future innovation and knowledge performance. Section 3 describes the methodology, data and models adopted in this study. Section 4 presents the main empirical results; these results, and their implications for policy makers, are discussed in Section 5. The final section concludes.

Literature review The role of higher education institutions in innovation development
The HE sector contributes to the development of innovation through two main mechanisms. Firstly, HE provides a supply of human capital. Enhancing human capital leads to highly skilled human resources and better individual and organizational performance (Furman, Porter, and Stern 2002). Secondly, HE produces useful knowledge that supports innovation and economic and social development (Martin and Etzkowitz 2000).
According to the systemic approach of national innovation production, the innovation performance of a country depends on the existence of a highly effective NIS (Edquist 2005). This system consists of four interacting sub-systems, namely: the production; the research; the financial; and the teaching and training sub-systems. The HEIs' involvement in three of these sub-systems has been well explored (Perkmann and Walsh 2007). However, there is little research into the involvement of the teaching and training sub-systems in supporting innovation. Importantly, most of the economics and innovation system literature appears to have neglected the possible heterogeneities that may exist in the HESs around the world.
The characteristics of the HE teaching and training system and their effects on national innovation performance The education policy literature identifies a variety of HESs around the world (Conner and Rabovsky 2011). HESs differ in characteristics such as governance (Amaral, Jones, and Karseth 2002), accountability (Huisman and Currie 2004), funding (Titus 2009) and government intervention (Neave and van Vught 1994). These studies have shown that HESs are institutionally diverse and their effects on national innovation performance and knowledge production are not yet fully understood.
Most empirical studies have focused on the effect of human capital on economic growth. However, some have examined the effect of human capital on the innovation outputs; for example: the level of higher education attainment (Chi and Qian 2010); the mobility of human capital between regions (Simonen and McCann 2008) and quality education (Varsakelis 2006) all have a positive effect on innovation activity. However, none of these works has considered the 'institutional diversity' of HESs and their impact on national innovation performance.
Our work is closely related to Trow (2005) and Ansell (2008). Based on their studies, we identify four main HES characteristics: (i) the capacity of the HES; (ii) the level of investment in HE; (iii) the level of subsidies in the HES; and (iv) the matching between HES activity and societal needs.
The level of student enrolment and national innovation performance One important characteristic of a HES is its ability to absorb the growing demand for HE places (Trow 2005). Some systems are specifically designed to cope with a high influx of students, while others are more focused (Ansell 2008). The innovation literature suggests that the availability of skilled human resources can facilitate the production of a substantial amount of technological knowledge and thus enhance innovation performance (Nightingale 2000). In contrast, some scholars argue that, in a globalized world, size is not necessarily positively associated with performance (Murphy 1993). Dore (1976) posits that expanding higher education often leads to a poorer quality of that education, and to a quest for diploma rather than true skill.
The empirical evidence on the effect of HES scale on a country's performance is inconclusive. Seetanah (2009) found a positive effect of the primary and secondary education enrolment ratio on per capita GDP growth. Benhabib and Spiegel (1994) showed a positive association between the number of graduates and per capita gross domestic product (GDP) growth. Although these studies are not concerned directly with innovation, they do suggest a positive relationship between HES capacity and innovation. A more relevant study by Varsakelis (2006) found no significant correlation between student enrolment in HE with scientific orientation and the national innovation outputs of a sample of mainly Organisation for Economic Co-operation and Development (OECD) countries.
The above theoretical arguments and empirical results are clearly mixed. However, there seems to be a case that while the relationship between HE enrolment and national innovation performance is not strong in the developed world, there is empirical evidence in favour of a significant and positive association between enrolment and innovation performance in developing countries. Unlike developed economies, developing countries lag behind in terms of human capital and suffer from severe shortages in qualified and educated manpower. From this discussion we propose: H1: There is a positive relationship between HE size and national innovation performance.

The impact of funding
The effect of investment in education on economic growth has been widely researched. The literature suggests that the quality of, and improvements in, knowledge and research delivered by HEIs are fundamentally related to funding (Greenaway and Haynes 2003). However, while there is strong evidence in the developed world, the results for the developing countries are mixed. Some scholars found no significant relationship (Temple 2001), while others observed a strong relationship between education expenditure and economic growth (Neycheva 2010).
Improving the capacity and quality of teaching and learning require higher levels of investment per student (Greenaway and Haynes 2003;Barr 2004). Higher investment in HE is therefore expected to lead to better quality human capital and, hence, to higher levels of innovation performance. Thus, we propose: H2: There is a positive association between investment in higher education and national innovation performance.

Higher education subsidies
One important characteristic of HE is the level of subsidies. The extant literature reflects a significant divergence of views related to subsidizing HE (Greenaway and Haynes 2003;Archibald and Feldman 2008). There are two distinct schools of thought. The first school advocates a greater involvement of the public sector with high levels of subsidies to support HE (Cheslock and Gianneschi 2008). Subsidies are viewed as critical for the enhancement of access to HE equity (Heller 2001) and efficiency (Jongbloed 1998). However, a subsidized system is usually very expensive, putting increased pressure on government budgets (Barr 2004;Doyle, Delaney, and Naughton 2009). More importantly, Liefner (2003) claims that subsidized HEIs face the risk of running unsuccessful projects.
A second school of thought places a greater emphasis on market intervention mechanisms (Ansell 2008;Barr 2004). These new forms of funding from private actors (for example: fees, tuitions, gifts and grants) are aimed at enabling HEIs to enhance their performance and be more responsive to changes in demand (Greenaway and Haynes 2003). Liefner (2003) also suggests that the competitive approach can strengthen accountability and transform new knowledge into marketable products. However, the role of market orientation in HE is neither sufficiently investigated (Caruana, Ramaseshan, and Ewing 1998) nor well understood (Barr 2004). For Jongbloed and Vossensteyn (2001) the market in which HEIs operate is not perfect and can fail. The strong use of competitive and performance based approaches can generate an attitude of risk-avoidance that can be counterproductive as HEIs will be more interested in outputs and objectives that can be easily achieved (Jongbloed and Vossensteyn 2001).
In spite of their significant differences, these two types of funding mechanisms have both been successfully used by HEIs (Liefner 2003). Such a success is fundamentally dependent upon the effective use of these resources by HEIs, regardless of the type and source of funding. This view is supported by examples of OECD countries whose growth has been linked to the increased and satisfactory levels of funding allocated to their HES through the adoption of different policies and strategies of funding (Barr 2004;Liefner 2003). Based on the above discussion, we propose: H3: There is no association between the level of HE subsidies and national innovation performance.

Alignment of HES
Our final question relates to the relationship between HESs and their surrounding societal environment. There is an interesting debate on the relationship between knowledge and innovation developed by universities and their economic and social environment (Etzkowitz 2008). Some HESs are deliberately designed to directly serve the needs of their society while others are not particularly established to support any particular societal need other than the advancement of scientific knowledge and critical thinking. In the HE literature, the latter is sometimes called an 'ivory tower' system. Under this model, knowledge is mainly initiated and produced by disciplinary academics, separated from society and usually based on the autonomy of scientific investigation (Gibbons et al. 1994). In the alternative 'application driven' model, knowledge is created and developed for practical motives. This knowledge is usually interdisciplinary and initiated by practical, societal problems (Gibbons et al. 1994). Examples of this knowledge production are the modern 'entrepreneurial' universities (Etzkowitz and Leydesdorff 2000).
The merit of ivory tower versus application driven HESs for supporting national innovation performance has been the subject of intense debate. Nelson (2004) and Ziman (2000) argue that scientific knowledge produced by basic research and curiosity-driven knowledge production activities constitutes indispensable elements for future long term innovation performance. On the other hand, many scholars contend that having to serve societal needs will create new avenues for useful research which in turn will be useful to drive innovation performance (Schulte 2004).
Empirical studies in this area have reached conflicting conclusions. One group of scholars report positive impacts from engagement with industry. For example, Van Looy et al. (2004) noticed that researchers' involvement in contract research stimulates scientific productivity, and Jensen and Thursby (2001) showed that the active pursuit of commercialization could actually promote basic research and scholarly education. However, Mansfield (1991) estimated that 11% of new products and 9% of new processes could not have been developed without the support of basic academic research. Anselin and Varga (1997) found that firms' patenting activities were higher in a region where the universities were publishing a high amount of academic publications. Given the conflicting views discussed above, and based on the inconclusiveness of the empirical evidence, we propose our fourth hypothesis: H4: there is no difference between the 'ivory tower' and 'application driven' HESs in supporting national innovation performance.
Testing this last hypothesis is problematic as it is difficult to properly quantify and distinguish the two extreme higher education systems. Data constraints often compel researchers to use imperfect proxies. We provide several arguments for our choice of proxy for H4 in the next section.

Methodology
Our empirical investigation relies on several databases. Patent data was collected from the World International Patent Office. Country publication data was taken from the SCImago Journal and Country Rank database. The higher education data is from the World Bank Country Education Survey database, while the population data is obtained from US Bureau of Census. The distinction between developed and developing countries is based on the IMF World Economic Outlook 2004. The innovation performance of a country is measured by the annual average of patent family 1 between 2005 and 2008. Patents may not necessarily be an accurate measure of innovation as they measure inventiveness and do not always lead to innovation (Crosby 2000). More importantly, it could take a long time to transform an invention into a successful innovation. With this in mind, we nevertheless rely on patents as a proxy for innovation because of a lack of directly measured innovation data. Because patent output is correlated with innovation, we argue that patents are the best available indicator of the relative innovation performance of a nation (Pavitt 1998).
Knowledge performance is proxied by the annual average of academic publications between 2005 and 2008. Again, this is a rather noisy measurement of knowledge performance as publications have varying quality and impacts. Unfortunately, such a detailed data is not available across disciplines.
Data on education is generally fraught with missing and incomplete information. Apart from the mainly OECD countries, which have a full dataset available, the majority of the remaining countries have sparse or non-existent data. Our aim is to focus on the international variation in the education system rather than the variation within the OECD countries. We are therefore obliged to make use of an incomplete dataset. The literature has long considered methods on how to deal with missing data. 2 Rather than dropping the missing cases, or replacing missing observations with sample averages, we use instruments to approximate missing observations. Using sample averages biases the results in favour of OECD countries. However, using data from similar countries (based, for example, on per capita GDP) provides us with a more representative sample. In our empirical application we use five instruments to estimate missing observations, namely the 2005-8 average of GDP, per capita GDP, and GDP growth; as well as the 2001-4 average GDP growth and gross expenditure on research and development to GDP.
Proximity is based on the nearest neighbour. For each case, we rank the instrument in ascending order. Then, the country available on either side of the missing country is identified as neighbour. Then the missing observation of the missing country is equated with that of the neighbouring country. However, this method does not fill all missing observations because the nearest neighbours may themselves have missing observations (in which case we keep the observation as missing).
The estimation is carried out using simple ordinary least squares (OLS) as well as a censored regression model (TOBIT). Our dependent variables are the average number of patents and publications between 2005 and 2008. Both of these have possible values of zero as some countries may have no patents or publications during the period 2005 to 2008. In this situation a TOBIT model is more appropriate since this latter allows for variables with a mixture of discrete and continuous distributions. The TOBIT model is a nonlinear model, usually estimated via maximum likelihood method (Amemiya 1973), but the interpretation is the same as in ordinary regression.
We adopt a production function type where O * i is country i's output (patents or publications), 1 i is the disturbance term, and H 1i to H 4i are the main inputs to the production function and are related to the four hypotheses proposed in the previous section. All data related to these hypotheses are four year averages (2005)(2006)(2007)(2008). Expenditure figures are in constant 2005 international $.
For the first hypothesis, we proxy size by the gross enrolment ratio of students, which is the percentage of the number of the students enrolled in higher education to the total number of the population within the higher education age group. Thus, the gross enrolment ratio for country i is given by: This variable underpins the assumption that the greater the proportion of students enrolled in HES, the higher the potential to improve the innovation and knowledge performance of a country. We use relative, rather than absolute, size in order to avoid the scale effect problem (due to, say, population size).
For the second hypothesis we proxy investment in higher education by 'relative support'. Thus, the relative support for country i is given as follows: This proxy is the ratio of relative expenditure to relative enrolment. The numerator equals total education expenditure for higher education (tertiary educational level) as a percentage of GDP. The denominator represents relative enrolment (enrolment ratio). The relative support is a better measure than absolute funding measures (such as total education expenditure for HE) because different countries have different population sizes as well as different levels of economic output. The second hypothesis implies a significant and positive relation between relative education support and innovation output. The third hypothesis relates to the level of subsidies, which we measure as the ratio of public expenditure (in HE) to total expenditure (in HE), and is given as follows: Finally, the last hypothesis relates to the impact of HES alignment with its surrounding economic and social environment on innovation performance. We measure this variable by a mismatch measure, which is the ratio of unemployed HE graduates to total population. Thus, the mismatch for country i is given by: One major concern relates to our choice of unemployability of HE graduates as a proxy for HE alignment. While it is true that unemployability is not perfectly related to HE alignment, it is nevertheless expected to correlate with it. For example Gordon Brown stated that 'Given the substantial public investment in university students, it is particularly important that they are employable upon graduation' (cited in PISG/ HEFCE 1999). This implies that governments in general are keen to align the HE systems with the demands of the economic sector. It is obvious from the above statement that one main way to measure this alignment is through the employability of HE graduates. Other factors, such as political and economic stability, might obviously affect the employability of graduates. However, these factors are expected to be captured in our model by a set of control variables (see below). It is anticipated that these control variables together with our proposed measure should be a reasonable proxy for HE alignment.
Another concern relates to our choice of a continuous rather than discreet variable. Our choice of a continuous measure is based on three considerations. First, it is not obvious how to identify whether a specific country belongs to an application driven or an ivory tower system. Second, even if there was a way of identifying some characteristics by which one can make a judgement as to the nature of the adopted system, it may well be the case that a country chooses a combination of both. Thus, given the possibility that countries may not adopt pure systems, a continuous measure would be preferable. Finally, we expect the number of unemployed higher education graduates to be correlated with how deep a system is within one system or another. For example, with an extreme ivory tower system we expect H 4i to be at its highest value because none of the graduated students gets employment (unemployed higher graduates equal the number of graduates). On the other hand, an extreme application driven system will produce the lowest value for H 4i of zero. In this extreme case, all graduated students find employment.
The control variables are defined as follows. are dummies equalling 1 if country i is an OECD member, has high level of output, has medium level of output, or has missing observations, respectively. G i is the ratio of gross expenditure on research and development to GDP; and M i is a perception variable on the quality of math and science education in country i. Finally, to control for the possible effect of country size on output we use log population for country i (LogPop i ).
We define 'high' patent output countries as those exceeding 10,000 patents during the period of study, and 'medium' patent output countries as those between 1000 and 10,000 patents. High publication output countries are defined as those exceeding 50,000, while medium publication output is defined as between 10,000 and 50,000 publications. Although these cut-off points are necessarily arbitrary, our choice seems to fit the pattern shown in Table 2.

Summary of data
The initial dataset contains 208 countries. Of these, 11 countries have virtually no data and were removed. 3 The remaining 197 countries are mixed. Only 30 countries, of which 21 are OECD countries, have complete datasets while the remaining 167 countries have sparse data. Table 1 shows the main variables for a selection of countries with full and missing information. The non-missing dataset is dominated by OECD countries and shows some obvious outliers (US and Japan in terms of patents and USA, UK and Japan in terms of publications). The incomplete data countries are mostly developing countries. Within this group there also appear to be outliers (China, Canada and India).
The summary statistics are shown in Table 2. There are few missing observations for the dependent variables (patents and publications). The large standard errors for these two variables indicate the presence of outliers on both sides of the distribution. The information on enrolment (H1) is fairly acceptable with 142 countries having information on this dependent variable. However, the remaining independent variables (H2, H3 and H4) have less available information. Finally, the control variables show fewer missing observations.

Results
We begin our analysis with the full information sample. This sample, which is dominated by OECD and industrialized countries, is not representative of the rest of the world. Nevertheless, these results will help us gauge the possible difference between the developed and developing countries as illustrated in Table 3. Since there are no zero values (for the dependent variables) in this set, we use OLS.
The F probability values suggest that the overall fit is significant. The R-squared is very high, suggesting appropriateness of the control variables. It is clear that the two types of output have different outcomes. First, the high and medium output dummies are highly significant for patents. The scale of the coefficients of these two dummy variables indicates two clear (structural) breaks in innovation even within this mainly  developed nation sample. Roughly speaking, a typical high output nation produces e 3 .515 = 33.62 times more patents than an average low output nation, whereas a typical medium output nation produces 10.14 times more patents than a low output nation. These two dummies contribute about 6% to explaining patents. 4 Given that these are dummy variables, this contribution suggests that there is a substantial portion of patents that cannot be explained by the model. Population and expenditure on research and development are also highly significant and explain about 5% and 2% of the variation in patents output, respectively. Finally, given the high and medium dummies and population, the perception on quality of education does not contribute to explaining patents output. The OECD dummy is insignificant suggesting no difference in outputs between OECD and non-OECD countries once we control for relevant explanatory variables.

Analysis of the extended set
As our focus is international, we aim to assess our four hypotheses in light of the extended sample containing all possible countries. Table 4 shows the TOBIT results for the extended set. 5 The results for the extended set are generally different compared with the non-missing set. First, the control variables have more impact on patents. The gross expenditure on R&D is surprisingly negative, suggesting that, other things equal, countries with higher gross expenditure on R&D have marginally lower patent outputs. The remaining control variables are all highly significant. First, note that the coefficient of log population is significant but low (compared with the non-missing data case), suggesting that population size has a significant but less important role in determining the patent output. Second, the missing variable dummy is negative, implying that countries with missing variables are almost four times (e −1.276 = 0.28) less productive (in terms of patents) than a similar country without missing variables. Third, and most important, the OECD, medium and high dummies reveal a large unexplained gap between the high output and low output countries. The results suggest that on average, an OECD country has e 2 .623 = 13.78 times more patents than non-OECD ones. A high output country has 210 times more output than an average non-OECD and low output country; while a medium output country has more than 26 times more patents.
Our first hypothesis of a positive relationship between HE size and innovation is strongly supported. The coefficients for H1 are both positive and highly significant, suggesting that larger HE sectors produce higher levels of output, even after controlling for expenditure and population factors. There is also a strong support for our fourth hypothesis. Indeed, the coefficients for H4 are insignificant for both patents and publications, implying that there is no connection between mismatching and innovation output.
The results on our second and third hypotheses are less conclusive. The coefficient on relative support (H2) for patents is positive and highly significant, but the coefficient for publications is insignificant. So the second hypothesis is confirmed for patents, but rejected for publications. The roles are reversed when we consider the third hypothesis. This hypothesis is supported by the publication results but rejected by the patents results.

Discussion
Our findings indicate that our four hypotheses can be substantiated. The level of HE capacity and the level of funding to HE students are shown to be associated with the level of national innovation and knowledge performance. Our findings confirm that there is no difference between the ivory tower and application driven approaches, in supporting the national innovation performance of a country. These findings support the view conveyed by a large body of the literature, that the ability of a country to build up a knowledge society and to innovate is related to the development of its human capital (Chi and Qian 2010;Barr 2004).
The importance of the capacity of the HE sector for HE policy aimed at supporting national innovation and knowledge performance is perhaps somewhat unsurprising, as it is seems sensible to assume that an increase in educated human resource opens up the possibility for the creation of innovative undertakings or ventures. It is, however, worth noting that the scale of the capacity on its own can be counterproductive if it is not associated with the quality of education as suggested by an important branch of the literature (Hanushek and Woessmann 2008). The impact of funding of human capital production on national innovation performance is investigated through two hypotheses. Hypothesis 2 (H2) asserts that knowledge required for the production of innovation is dependent on the level of total funding invested in HE. H3 is focused on the impact of public funding (subsidies) on the national innovation performance. Although both of these hypotheses are less conclusive than H1 and H4, our empirical investigation still provides partial evidence of links between funding and national innovation performance, that is, relative support (H2) impacts patents. Moreover, our prediction that subsidies do not impact innovation is supported for publications only. On the other hand, countries with higher levels of subsidy tend to underperform in terms of patents.
Our results correspond to the ongoing debate on HEI funding systems and the lack of clear consensus with regards to their effectiveness and impact on national innovation performance (Muscio, Quaglione, and Vallanti 2013). This lack of consensus and the continuing emergence of contrasting views between the two main schools (public funding versus market mechanisms) can be explained in part by the heterogeneity of the HE system (Ansell 2008). This result provides evidence that both approaches and policies can be combined (Arcalean and Schiopu 2010) and when properly implemented, can lead to improved levels of innovation and knowledge performance (Muscio, Quaglione, and Vallanti 2013). This also explains the emerging view which places emphasis upon the complementarities between the different types of funding strategies (Arcalean and Schiopu 2010). It is also worth noting that expecting to secure immediate results from simply increasing HE funding can be misleading; an effective, high quality HE teaching and learning system is also required (Geuna and Muscio 2009) The objective of our last hypothesis (H4) was to investigate whether there were any differences between the ivory tower and application driven approaches in supporting the national innovation performance of a country. Our empirical findings show a strong support for H4, which unambiguously indicates a lack of relationship between the extent of matching graduates to labour demand and innovation output. Our results also confirm the absence of prior conclusive empirical evidence favouring one system over the other. Thus, we provide further evidence supporting the view that neither the ivory tower nor the application driven approach is a universally better approach in all applications.
The scale of HE capacity, the support per HE student and subsidy are likely to be amongst the main drivers for a country's future innovation performance. Thus, our findings suggest that these three elements should be the main constituents of HE policies aimed at supporting national innovation and knowledge performance. On the other hand, we find no difference between the ivory tower and application driven systems. Consequently, a greater emphasis should be placed on complementing the two systems.

Conclusions
The importance of knowledge and human capital and their impact on innovation and growth are widely recognized, and so is the role of the HES in the production of this knowledge. However, most of the innovation literature attributes this production to the second and third missions of HESs, which include the research and production functions of NIS. The role of the first mission has received noticeably less attention in the literature. This paper has been motivated by the existence of an apparent gap in the innovation literature concerning the appropriate HE policy to support national innovation and knowledge performance, specifically within the domain of the first mission.
In this paper, we show that the capacity of the HE sector and the support available to HE students are positively correlated with the level of national innovation and knowledge production. On the other hand subsidization is either neutral (for publications) or harmful (for patents). Finally, the adoption of any particular strategy with regards to the alignment between the HES and its external environment does not seem to be relevant. Indeed, once we control for other HES characteristics, we could not find any statistical discrepancy between ivory tower and application driven systems.
Our empirical evidence suggests these key institutional characteristics should form the foundation of HE policies aimed at supporting national innovation and knowledge performance. In particular, policy makers should focus on greater access to HE as it is important for promoting both innovation and knowledge production. They should also be aware of the trade-off that exists in other characteristics. For example, higher investment in HE is good for patents but higher subsidies have an adverse effect on patent production. On the other hand neither of these policy tools seems to impact publication production.
The results of this paper are based on a cross-country analysis. This leads to three main limitations. First, we view the relation between HE characteristics and innovation production as an aggregate phenomenon. Thus, we do not consider within-country heterogeneity of HESs. If such heterogeneity could be approximated, the empirical relevance of our results could be improved by including it as an additional HE characteristic. The second limitation stems from our choice to extend the sample beyond those countries that have complete datasets. This choice makes our study interesting because we avoid the limitation of drawing conclusions that are specific to the (mainly OECD) developed world. Unfortunately, this results in a thinner dataset, with a substantial number of missing observations. Although we partially mitigate this problem, we feel that until a fuller dataset is made available, conclusions drawn on partial datasets like ours should be treated with caution. Finally, because the proxy for our last hypothesis does not fully capture the distinguishing characteristics of the ivory tower and application driven systems, our conclusions regarding this hypothesis should also be treated with caution.