Searching for the existence of entrepreneurial ecosystems: a regional cross-section growth regression approach

The term entrepreneurial ecosystem was coined to refer to those elements in the entrepreneurs’ environment that help them to succeed (or not) in their efforts to grow a new venture. Even though the analogy with the natural and biological ecosystems cannot be upheld in all details (Holling 2001), the concept has proven very useful to signify those elements directly relevant to the entrepreneurial value chain (Adner and Kapoor 2010; Spigel 2015). The term “ecosystem,” however, also suggests a broader scope and extending the concept to the entire set of interdependent actors and factors coordinated in such a way that they enable productive entrepreneurship is in fact required (Stam 2015). This encompassing definition has proven useful in theorizing about the importance of the context in which entrepreneurs operate and led both academics and policy makers to zoom in on the local, regional, and national enablers of entrepreneurship (Audretsch et al. 2015). Measuring (the size and/or quality of) entrepreneurial ecosystems, however, continues to prove a challenge. A first problem with measuring the quality of an ecosystem is that we first have to establish the appropriate ecosystem boundaries.

In nature, ecosystems overlap. Climate is a global system that overlaps with more localized weather systems and the relevant ecosystem boundaries for the habitat of the eagle clearly span the ecosystems of many mice. Economic or entrepreneurial ecosystems arguably share that property. The local entrepreneurial ecosystem is part of the larger regional, national, international, and perhaps truly global ecosystem and although all economic activity is in the end local, there is no reason to assume that the most relevant ecosystem boundaries coincide with any administrative geographic units, however (dis)aggregated they may be. Instead, the relevant ecosystem boundaries depend on the question at hand. Do we want to study the eagles or the mice?

We follow Stam (2015), who puts forward that the entrepreneurial ecosystem is multileveled and proposed an approach building on Andersson and Henrekson (2015). They first define (local) competitiveness as the ability to continuously renew the economic base of the local economy. Then they argue that entrepreneurship (at the local level) is instrumental in maintaining that ability. Following (Baumol 1990) institutions, broadly defined as “the rules of the game” (North 1990), will influence the (local) supply and direction of entrepreneurship. The relevant institutions (Andersson and Henrekson 2015) then identified are clearly “located” at the national level (e.g., rule of law, ease of starting a firm, taxes, property rights, labor market institutions, social security) or the local level (e.g., local and regional taxes and regulations, presence of, e.g., universities and research institutes and local entrepreneurship cultures and skills). Andersson and Henrekson (2015) also argue that one should be able to identify the effects of local entrepreneurial ecosystems in the different effect entrepreneurial activity has on growth across localities, regions, and countries. They conclude that their arguments “suggest that the magnitude of local multipliers and growth effects associated with the local presence of entrepreneurial and knowledge-intensive activities are a function of the local institutional environment. This is a hypothesis that remains to be rigorously tested in empirical work.” (Andersson and Henrekson 2015, p.189)

In this paper, we aim to take a first step in empirically testing that hypothesis. More specifically, we test whether the contribution of entrepreneurship differs significantly across regions. If we may refer to the “local institutional environment” as the ecosystem, we focus on the predicted effect of local entrepreneurial ecosystems on the local multipliers/growth effects. If such effects exist, the marginal effect of entrepreneurial activity (measured in several standard ways) on local economic growth would vary systematically with the quality of the (unobserved) entrepreneurial ecosystem. By investigating at what level of geographic aggregation the marginal impact of entrepreneurial activity on growth shows this predicted variation, we can show whether entrepreneurial ecosystem quality matters most at the regional or national level when economic growth is our variable of interest.¹

As other researches suggest that the local context does matter and entrepreneurs may have a key role in shaping and developing local entrepreneurial ecosystems (see, e.g., Feldman 2014), we suspect that our data is geographically too aggregated to distinguish the differential impact of local entrepreneurial ecosystems. Our NUTS-1 and 2 level regions may simply be covering diverse local ecosystems, obscuring and confounding the differences that exist within these regions. Given the limited granularity available in the Global Entrepreneurship Monitor (GEM) data we use for our entrepreneurship variables, however, this is not something we can easily resolve in our cross-country study. The main contribution of this paper is therefore the empirical method we propose and we very much encourage those with access to more disaggregated data to repeat our analysis. In the current data set, we must conclude that we fail to find support for any theory that predicts different marginal effects of entrepreneurial activity on growth across (groups of) NUTS-1/2 EU regions. But the search should continue and we hope our method contributes to that.

The remainder of this paper is structured as follows. In the next section, we position our paper in the literature. Section 3 presents our empirical strategy and briefly derives the main equations we estimate. Section 4 presents a simulation exercise to show that our empirical strategy does uncover the pattern we are looking for in a dataset we have created for that specific purpose. Section 5 presents our real world data and shows the empirical results. We conclude with a discussion of these results and agenda for further research.

Theoretical interest in the role of entrepreneurship in creating economic growth goes far back. Schumpeter (1912) already put entrepreneurship center stage in his Theory of Economic Development by coining the term “creative destruction”. In more modern endogenous growth models, such as those developed by Romer (1986) and Lucas (1988), there is no role for entrepreneurship, as they passively commercialize all knowledge that is generated. Aghion and Howitt (1992) created a model where “entrepreneurs” both generate the knowledge and appropriate entrepreneurial rents through entry, while Michelacci (2003) and Acs and Sanders (2012) distinguished between invention and innovation and show that in such a model, entrepreneurship is complementary to knowledge creation and contributes to growth. The empirical evidence in support of such predictions, however, remains mixed (e.g. van Praag and Versloot 2007). As (Glaeser et al. 2010a, p.3) stated: “While it would be hard to imagine a world in which an abundance of entrepreneurs did not strengthen the local economy, the literature documenting this effect is still in its infancy”. Empirical results to date are inconclusive for three reasons. First, there is a measurement problem of entrepreneurship, which led to the use of many different proxies for entrepreneurship (Hechavarria and Reynolds 2009; Stenholm et al. 2011). Studies using varying proxies for entrepreneurship (e.g. Audretsch 1995; Carree and Thurik 1998) have found a positive relation overall, but often these results are multi-interpretable.² Second, and closely related, there is heterogeneity within entrepreneurial activity and the effect on growth is not equal for all types of entrepreneurship (Santarelli and Vivarelli 2007; Stam 2009). Research makes a distinction between ambitious entrepreneurship or gazelles (Henrekson and Johansson 2010; Stam et al. 2012) and businesses started at the local level that create little employment growth or economic expansion. For example, Wennekers et al. (2005) find a U-shaped relationship between entrepreneurial activity and economic development, proxied by GDP/capita levels. This finding was explained as labor moving from unproductive entrepreneurship into employment in early stages of development (higher GDP/cap reduces entrepreneurship) and from employment into growth enhancing innovative entrepreneurship (entrepreneurship causing GDP/cap growth) as development increases. Third, the two-way causality already referred to above already indicates that the effect of entrepreneurship on growth is hard to isolate empirically (e.g., Baumol 1990; Carree and Thurik 2010; Stephens and Partridge 2011; Fritsch and Wyrwich 2014). Entrepreneurship both causes and is caused by GDP (growth), creating an identification problem. In sum, the literature is only beginning to build a consensus on how to measure entrepreneurial activity, what outcomes to expect from what types of entrepreneurial activity, and how to properly identify such effects. But even if that consensus were achieved, the concept of “entrepreneurial ecosystems,” on which this special issue is focused, implies that the same entrepreneurship, accurately measured and properly instrumented, should be expected to affect economic growth differently in different contexts. The entrepreneurial ecosystem concept stresses the complex interactions between various environmental factors and entrepreneurial activity in creating economic dynamics. As such, the concept is closely related to literatures on the national (Lundvall 1992; Freeman 1987, 1995; Edquist 1997) and regional systems of innovation (Autio 1998; Cooke 2002; Doloreux 2002) as well as the more mainstream empirical growth literature that is linking economic growth and development to institutional quality (e.g., Easterly 2001; Acemoglu and Robinson 2000; Sokoloff and Engerman 2002; Nunn 2008). These strands of literature share a focus on the enabling conditions for Schumpeter (1942)’s “gales of creative destruction.” It is beyond the scope of this paper to survey all that literature here. Instead, we restrict ourselves to establishing the connection with more mainstream empirical growth literature. The main problem in this literature is that institutions develop endogenously with economic growth, and therefore, to identify and isolate the effect empirically, the researcher has to find an exogenous instrument for institutional quality (e.g., Acemoglu et al. 2001). But even if such a satisfactory instrument can be found at the regional and local levels, there is still the problem that no satisfactory measure of institutional quality exists. This problem, we argue, will also haunt empirical research into the impacts and relevance of the entrepreneurial ecosystem. In fact, both concepts are multidimensional complex systems that prove too elusive to be measured directly in a one-dimensional, cardinal number. The proxies that have been used for institutional quality, such as indices of property rights protection, economic freedom, and corruption, inevitably only “measure” a part of the relevant institutional complex and not even accurately so (Rodríguez-Pose 2013), based as they often are on expert opinion and surveys. The same holds for measuring the quality of the entrepreneurial ecosystem. Elaborate indices as for example proposed by Acs et al. (2014) and Szerb et al. (2015) do give us a useful first cross regional comparison of ecosystem quality. But such composite indices, by the way they have been constructed, have the disadvantage that they cannot be used as independent variables to properly identify the moderating effect of ecosystem quality. It will prove even more complicated to find suitable instruments for such complex multidimensional indices if we aim to use them as an independent variable in cross-sectional growth regressions. Here, we therefore propose an alternative strategy for “measuring” the effects of the entrepreneurial ecosystem. To do so, we exploit the assumption that the entrepreneurial ecosystem will affect output and growth through its effect on entrepreneurial activity, which we arguably can observe. That is, assuming the entrepreneurial ecosystem matters for growth, its quality will be systematically correlated with growth and works as a moderating variable on entrepreneurial activity. An example of such moderating mechanisms is as follows: better protection of property rights increases the future expected private gains from entrepreneurship. This may attract more talented entrepreneurs to the region (or country) to start their own business (see, e.g., Johnson et al. 1999; Sobel 2008). Another example is an improvement of financial institutions, enabling a flow of funds from investors to potential entrepreneurs such that they can undertake larger and more ambitious projects, thereby also increasing the effect of entrepreneurial activity on growth (e.g., King and Levine 1993). Informal institutions can also provide a moderating mechanism; an entrepreneurial culture, trust, social capital, and a dense local knowledge network are all likely to contribute to more productive entrepreneurship in a region or locality (Bosma 2013; Andersson and Henrekson 2015). Our approach thus distinguishes itself by researching a channel through which the moderating effect of the entrepreneurial ecosystem on growth will become visible in variation in the contribution of observed entrepreneurial activity to growth.

We conclude from the above review that entrepreneurship takes on many shapes and manifests itself in different contextual settings. Consequently, the relation between entrepreneurship and growth will be heterogeneous and not all types of entrepreneurship create economic growth to the same extent in different institutional contexts. As we are not primarily interested in the exact value or even sign of the marginal effect of entrepreneurship itself, however, we may sidestep these issues for now. We make the case that, even if the estimated marginal effect itself is inefficient or biased due to measurement error, impact heterogeneity, and reverse causation, possible differences in the marginal effect of entrepreneurship on growth are caused, by definition, by differences in the regional entrepreneurial ecosystem. In other words, we assume that the presence of an entrepreneurial ecosystem can be deduced from the way it transforms locally available entrepreneurial talent into activity and ultimately economic growth. Thus defined, the first step that is needed to prove the existence of an entrepreneurial ecosystem at the aggregation levels investigated is to show that the marginal effects differ significantly across geographies. Our proposition is that the entrepreneurial ecosystem is what (ever) moderates the effect of entrepreneurial activity on growth. Of course, we do need to consider relevant controls and specify our regional growth regression in such a way that we do not allow omitted variables that explain cross-regional variation in growth and might be correlated with entrepreneurial ecosystem quality to create the illusion of an ecosystem at work. Therefore, all the usual suspects and controls that the literature suggests should be included. If multiple regimes can be identified with regard to the relationship between entrepreneurship and economic growth, we can introduce variables to predict regime membership, in order to find out what constitutes the entrepreneurial ecosystem and to rule out other possible explanations. Capital formation as well as population growth obviously play a role, in the neoclassical growth model and in regional growth accounting (Ciccone 2002). For example, Sala-i Martin (1996) also finds evidence for beta-convergence between regions in the USA, Canada, Japan, and Europe, indicating that initial income should be controlled for in regional growth regressions. However, the largest share of variance in growth across regions has been attributed to productivity differences and innovation. In empirical work on regional growth, we can then distinguish (i) the linear model, (ii) the innovation-systems or learning-region model, and (iii) the knowledge spillovers model. Rodríguez-Pose and Crescenzi (2008) are the first to have nested these models in one empirical specification. They find that knowledge spillovers indeed happen. But they also show that the transmission of tacit knowledge strongly decays with distance, supporting the innovation systems approach. Furthermore, they find that R&D expenditures only lead to actual innovation and growth if (a) a region can also benefit from extra-regional spillovers, measured by R&D in neighboring regions, and (b) the territorially embedded institutional environment is conducive to the diffusion of innovation. These results can be related to the geographical clustering literature (Jaffe et al. 1993; Breschi and Lissoni 2003; Bosma et al. 2008) that links the success of certain regions to the formation of a cluster within which companies of the same sector thrive and benefit from agglomeration advantages and knowledge spillovers. Frenken et al. (2007) show that such spillovers are stronger for “related varieties” in “smartly specialized” regional clusters. Reynolds et al. (1994) find that urbanization promotes spillovers. A result that Bosma (2009) confirmed in a multilevel analysis for European regions is also presented in this paper. From the regional growth literature, we take away that it makes sense to control for regional population density, age structure, and industry diversification (Dogaru et al. 2011; Thissen et al. 2011). Other variables, such as the degree of urbanization, labor productivity, the level of human capital, the investment in R&D, or the level and type of competition, have been mentioned in the literature as possible moderating variables on the effect of entrepreneurship on (job) growth (Fritsch and Mueller 2008; Fritsch and Schroeter 2011; Fritsch and Noseleit 2013a, b; Unger et al. 2011). Our strategy, however, involves first estimating a clean growth residual. That is, we estimate growth that is not explained by entrepreneurship. Including variables that are complement to and therefore possibly endogenously correlated with entrepreneurship would then be undesirable here. It therefore seems inappropriate to include these variables in the first stage of our analysis. Rather, these variables would become relevant if we find different classes of regions as they would then potentially explain the class membership in the second stage. The above contributions led us to expect that entrepreneurship is positively correlated with growth at the regional level. There are also reasons to expect this effect to differ across regions. Helsley and Strange (2011) for example show that urbanized cultures are often more conducive to growth-oriented entrepreneurship. When looking at regional performance in general, Eriksson et al. (2013) have distinguished between regions with a “business climate” and regions with a “people climate,” finding that both matter for regional performance, but through different channels. Audretsch et al. (2012) classify regions in an explicit way, using the functional specialization theory. In a sample of German regions, they use regional characteristics to determine the industry structure and thereby the specialization into either production or innovation. They find five groups of regions: industrial districts, periphery, industrial agglomeration, urban agglomeration, and urban periphery. The propensity to start a business is highest in groups 1, 4, and 5, and these types of regions are thus said to have an entrepreneurial regime. These papers, however, focus entirely on the effects of specific aspects of the ecosystem on observed entrepreneurial activity. In this paper, we investigate if there are latent regional ecosystem characteristics that affect the contribution to growth of such early stage entrepreneurship (that is, conditional on observing such activity). We can summarize this in the following hypothesis:

We aim to localize the moderating effect of entrepreneurial ecosystems on the relationship between entrepreneurial activity and economic growth and employ a three-step empirical process. First, we need to account for economic growth caused by the “usual suspects” at the national level (Mankiw et al. 1992). We extend the standard growth model with entrepreneurship and the relevant regional controls in an OLS estimation to establish the effect of entrepreneurship on growth at the regional level. Next, we estimate this model in a multilevel specification. Leckie et al. (2010) and Paterson and Goldstein (1992) show that multilevel models are very suitable when a complex web of influences at different levels of aggregation are suspected. This is definitely the case when looking at entrepreneurial ecosystems and formal and informal institutions at the regional and national level in Europe. Estimating our regression in a multilevel model allows us to distinguish between variance in growth rates at the regional and national level, and the extent to which entrepreneurial activity can explain this variance. We then exclude the entrepreneurship variables in both growth regressions and, after obtaining the residuals in these regressions, regress them on our entrepreneurship variable in an unconditional latent class specification.³ That allows us to test the hypothesis that the marginal contribution of entrepreneurship to (unexplained) growth differs systematically between (groups of) regions that do not necessarily coincide with countries. This somewhat roundabout way implies we do not allow the coefficients on the usual suspects in the standard growth regression to differ across (classes of) regions. If we find significant latent classes in this approach, this suggests that it is the contribution of entrepreneurship to growth that differs significantly between endogenously determined groups of regions.

The growth literature has emphasized the importance of considering all relevant variables in a growth regression (Sala-i Martin 1994). We base our model on the Mankiw et al. (1992), hereafter MRW model, and we estimate a cross-regional growth regression including the “usual suspects,” adding also relevant regional controls. The MRW specification for a cross-sectional growth regression is given by where i indexes regions, \({s^{k}_{i}}\) and \({s^{h}_{i}}\) are the average shares of income invested in human and physical capital (in the period T0 − T), respectively, and n is the average growth rate of the population. By including initial income, regions are allowed to be out of the steady state and we can test for convergence. The disadvantage of a growth specification is that permanent productivity differences in the production function across countries and regions end up in the error term. We will control for these differences by estimating the country fixed effect on a multilevel model. where j indexes the country and the specification allows for country-specific intercepts α _0j . The MRW model was originally developed to model cross-country differences in growth rates. From the regional growth literature, we know that several other factors are important when explaining regional growth. These factors include population density, industry structure, and the demographic structure of the region. They are found to aid the conversion of innovation to growth at the regional level. Population density is a commonly used proxy to control for the effect of urbanization economies. The industry structure is important because related variety has been found to contribute to regional growth. Finally, the demographic structure controls for the share of working-age individuals and possible migration that is more relevant within than across countries. We estimated Eq. 3.2 with different combinations of these regional controls. Residual variation, 𝜖 _{i j} , can be decomposed in a fixed and a random part, which can be used to calculate between and within country variation using the intraclass correlation (ICC) (Snijders and Bosker 2012). This gives us insight into how similar regions within a country are in their growth process. The ICC is the ratio of the variance between countries to the total variance, so it tells us how much of the variance in economic growth is accounted for by clustering on the country-level. Note also that the error term now still contains any variation in growth caused by variation in quality of the entrepreneurial ecosystem, either at the regional or national level. But as we have argued above, quality to date has not been quantified and is likely to be inherently unmeasurable. We cannot include a proxy for that latent variable and test explicitly the hypothesis that its marginal effect on growth is positive. We can include, however, proxies for the level of entrepreneurial activity that are common in the literature. If we extend the OLS growth regression by adding the entrepreneurial variables of interest, the regression equation becomes where E _i is our proxy for entrepreneurship.⁴ By comparing the results from this regression to a multilevel specification (that is equivalent to the one in Eq. 3.2 but we do not write out to save on notation), we can test the hypothesis that the marginal effect of entrepreneurship on growth, α ₅, controlling for the usual suspects is different across countries and/or regions within countries.

We then take the residuals from regressions (3.1) and (3.2) (with or without regional controls) and regress these on regional entrepreneurship in a latent class specification that allows for endogenous sorting into regimes r(= 1,...R). where r = 1,...,R indicates the regime, and R refers to the (exogenous) total number of regimes. Each regime, r, has its own β ₀ r and β ₁ r. We can then test for the existence of latent classes in the data following Greene (2007) who suggests a “test-down” strategy to identify the correct number of regimes. By regressing the residuals on entrepreneurship only, we restrict this analysis to finding groups of regions that differ in the marginal contribution of entrepreneurship to unexplained, residual growth variation. We want to stress that the estimated coefficients are likely to be biased by endogeneity (entrepreneurship is endogenous to GDP growth) and omitted variable bias (our specification is parsimonious and variables, possibly correlated with the entrepreneurial ecosystem quality, such as religion and geography, are not included). For the purpose of this paper, however, we are only interested in the clustering of data points that the latent class model would suggest. As we have argued above, finding such a clustering in the data would indicate there is a significant variation in the quality of the entrepreneurial ecosystem across regions as it is that quality that moderates the contribution of entrepreneurial activity on regional growth. Before we turn to our results in Section 5, we first describe a calibration exercise we did to test our method.

To ensure our empirical strategy will in fact reveal the existence of variation in the quality of the entrepreneurial ecosystem, conceptualized here as a multidimensional set of unmeasured variables that positively affect or moderate the contribution of entrepreneurial activity on regional economic growth, we first created a fake database containing all the variables on the right-hand side of Eq. 3.3. We generated random numbers drawn from a uniform distribution between 0 and 0.1 for s _k , s _h , and n _i , where we drew l n Y _i (T0) from a uniform distribution between 0 and 1 to fill the database.⁵ We also generated entrepreneurial activity rate by taking a random number from a uniform distribution between 0 and 0.1. All independent variables are thus i.i.d. by construction. We defined our true entrepreneurial ecosystem in three classes by taking where class was set to 1 for the first third of the dataset, to 2 in the second, and to 3 in the third.⁶ We then set the parameters α ₁₋₅ = (0.01,0.3,0.3,0.4,−0.1,0.2) and computed where 𝜖 _{i j} was drawn from a normal distribution N(0,σ ²) and we tested for different levels of σ. We also added a more complex and general structure by adding a full set of interaction terms with E E S _{i j} (all with a coefficient of 0.1) to see if our empirical strategy would also pick up more complicated and interactive ways of conceptualizing the entrepreneurial ecosystem.⁷ The results of our simulation exercise are presented in Table 1.

It can be seen that after hiding a three-class multilevel structure in our data, the model actually reveals that structure, even if we complicate matters with interaction terms that would capture how the quality of the entrepreneurial ecosystem makes factor accumulation more productive (column 2). In both columns 1 and 2, we see the ICC indicates a high level of intraclass correlation reflecting the fact that class is correlated across countries. We also find, however, that our model fails to find anything when measurement error (the σ ² for the distribution of the error term) becomes large (columns 3 and 4). Our multilevel and latent class specifications can spot the multilevel and three class structure up to σ ² of about 0.015, which is about 0.5 to 0.33 times the mean growth rate in our dataset. This suggests that with measurement error below 33% of the mean value of growth, our model picks up any systematic variation in the quality of the entrepreneurial ecosystem if it indeed moderates the effect of entrepreneurial activity on regional growth. Our simulations show that this is regardless of whether this contribution to growth is direct or indirect through an interaction effect on the traditional correlates of growth. Having tested and found our method at least in principle and under moderate measurement error to be valid, we can now proceed to analyze the real world data.

In this paper, we argued that an entrepreneurial ecosystem must reveal itself in the data through affecting the impact of observed entrepreneurial activity on GDP growth. Then, if ecosystems differ in quality across space, we should be able to reveal the existence and relevance of entrepreneurial ecosystems in the heterogeneity of marginal impacts of entrepreneurial activity on GDP growth at the appropriate geographical level. Finding evidence of such heterogeneity would prove that something is affecting the contribution of entrepreneurship to economic growth differently in different places. We then collected data and designed an empirical strategy to reveal such heterogeneity. We find no support for our hypothesis at the NUTS1-2 aggregation level in our data. But not finding such heterogeneity in this dataset should certainly not be interpreted as proof that the entrepreneurial ecosystem is non-existent. Instead, we can only conclude it does not reveal itself at this level of aggregation.

We simply cannot reject the hypothesis of a homogeneous effect of entrepreneurship in our data. This, however, is not the same as rejecting the hypothesis of a heterogeneous effect. There may be a variety of reasons why our result was found. First, and we deem this most likely, our level of geographical disaggregation may simply be too high for the effects of local ecosystem differences to show up statistically significant. Second, the sample under study might not include enough countries to find a heterogeneous effect. The data is very noisy and from the literature, we already knew that the impact of entrepreneurship itself is hard to properly identify. This then holds a fortiori for any moderation effects on entrepreneurship. The regions we studied may simply not be different enough to find significantly different marginal effects. Third, the period under study is a short period when compared to other empirical growth studies, and the growth rates of European countries in this period have been severely affected by the impacts of the global financial crisis and the Eurocrisis. In an earlier version of this paper, we also ran the analysis on the average growth rate over a much longer period. There, however, we faced the problem that we regress long-run growth on a relatively short sample of entrepreneurial variables observed towards the end of the period. Changing this did not qualitatively change our results. We only find a significant effect of entrepreneurship on growth in our OLS model, where unobserved heterogeneity between countries is not taken into account. And fourth, the countries in the sample are all developed and very much integrated countries, which might imply entrepreneurship and other factors are highly mobile within and even across countries. We tried to control for that by taking into account spillover effects of factor accumulation in neighboring regions, but such mobility would compound the endogeneity problems and reduce the variation in the marginal productivity of entrepreneurial activity across regions. An interesting finding is that the only type of entrepreneurship that has an effect on growth in this study is from the group of entrepreneurs that also expect high job growth in the next years, indicating that the “gazelle hypothesis” (Henrekson and Johansson 2010) also holds for European regions and that entrepreneurs can actually identify themselves as gazelles.

We carried out several robustness checks on our estimations. We tried alternative specifications to control for the country effect, respectively a varying-coefficient multilevel model and a standard OLS regression with clustered standard errors by country or country dummies. Both models were rejected in favor of the varying-intercept model presented above. Following Gorard (2003) and Leckie et al. (2010) and Paterson and Goldstein (1992), we also recommend in general the use of multilevel models in studies on entrepreneurship and growth, as a multilayered web of influences at different levels of aggregation is at play.

Our results therefore only mean that the regional ecosystems considered here are either not that different or matter little for the contribution of entrepreneurship to regional growth in the period under study. Finally, during the sample period, most European regions have suffered from the global financial crisis, affecting growth rates more than entrepreneurship. This too may confound the long-run effects of ecosystem quality differences. But even as our data do not justify strong conclusions at this stage, we do feel strongly about the appropriateness of our method and consider this the main contribution of this paper. We feel that ecosystem quality proxies cannot (yet) capture the full richness of the entrepreneurial ecosystem, whereas more complex composite indices impose an assumed structure and are not suited as independent variables. Therefore, more studies should make use of latent variables and indirect relations to show the possible moderating effects of ecosystems on innovation and growth before pushing proxies and measures that are supposed to explain that variation directly. Moreover, such analyses may inform data collection efforts in the future. If anything, our analysis has shown that collecting entrepreneurial ecosystem data at the NUTS1-2 level may not be the right level of aggregation, and research should focus on national or more local levels of aggregation instead. Establishing where the relevant ecosystem boundaries (do not) lie is an important contribution of our paper also, as it is the first step in collecting better data efficiently in the future.

The challenge remains to make the entrepreneurial ecosystem visible in the data. With better and more complete data, our method may allow us and others to do so in the future.